From 011891abfae060903f76d5c7aee23208cd295e71 Mon Sep 17 00:00:00 2001 From: JRPAN <25518778+JRPan@users.noreply.github.com> Date: Tue, 1 Feb 2022 20:36:24 -0500 Subject: Update Copyrights --- src/gpgpu-sim/gpu-sim.h | 3 ++- 1 file changed, 2 insertions(+), 1 deletion(-) (limited to 'src/gpgpu-sim/gpu-sim.h') diff --git a/src/gpgpu-sim/gpu-sim.h b/src/gpgpu-sim/gpu-sim.h index 68b3dfa..de69ef8 100644 --- a/src/gpgpu-sim/gpu-sim.h +++ b/src/gpgpu-sim/gpu-sim.h @@ -1,5 +1,6 @@ // Copyright (c) 2009-2021, Tor M. Aamodt, Wilson W.L. Fung, Vijay Kandiah, Nikos Hardavellas -// The University of British Columbia, Northwestern University +// Mahmoud Khairy, Junrui Pan, Timothy G. Rogers +// The University of British Columbia, Northwestern University, Purdue University // All rights reserved. // // Redistribution and use in source and binary forms, with or without -- cgit v1.3 From e1afc53b51d24afcfd8b8aab15e4ba5d99b4a772 Mon Sep 17 00:00:00 2001 From: Aaron Barnes <42706182+barnes88@users.noreply.github.com> Date: Tue, 16 Jul 2024 12:33:16 -0600 Subject: Auto clang format (#74) * add automated clang formatter * Automated clang-format * use /bin/bash and add print * use default checkout ref * Format only after tests are success * Run CI on merge group --------- Co-authored-by: barnes88 Co-authored-by: JRPAN <25518778+JRPan@users.noreply.github.com> --- .github/workflows/cmake.yml | 3 + .github/workflows/main.yml | 26 + format-code.sh | 1 + libcuda/cuda_runtime_api.cc | 16 +- src/abstract_hardware_model.cc | 20 +- src/abstract_hardware_model.h | 81 +-- src/accelwattch/XML_Parse.cc | 130 ++--- src/accelwattch/XML_Parse.h | 40 +- src/accelwattch/gpgpu_sim_wrapper.cc | 927 ++++++++++++++++++++--------------- src/accelwattch/gpgpu_sim_wrapper.h | 69 +-- src/accelwattch/processor.cc | 4 +- src/accelwattch/xmlParser.cc | 28 +- src/cuda-sim/cuda-sim.cc | 251 +++++----- src/cuda-sim/instructions.cc | 15 +- src/cuda-sim/ptx_ir.cc | 15 +- src/cuda-sim/ptx_ir.h | 2 +- src/cuda-sim/ptx_sim.cc | 3 +- src/debug.cc | 4 +- src/gpgpu-sim/dram.cc | 104 ++-- src/gpgpu-sim/dram.h | 9 +- src/gpgpu-sim/gpu-cache.cc | 34 +- src/gpgpu-sim/gpu-cache.h | 40 +- src/gpgpu-sim/gpu-sim.cc | 216 ++++---- src/gpgpu-sim/gpu-sim.h | 17 +- src/gpgpu-sim/l2cache.cc | 14 +- src/gpgpu-sim/l2cache.h | 9 +- src/gpgpu-sim/local_interconnect.cc | 2 +- src/gpgpu-sim/power_interface.cc | 695 ++++++++++++++------------ src/gpgpu-sim/power_interface.h | 33 +- src/gpgpu-sim/power_stat.cc | 525 ++++++++++++-------- src/gpgpu-sim/power_stat.h | 869 ++++++++++++++++---------------- src/gpgpu-sim/shader.cc | 331 +++++++------ src/gpgpu-sim/shader.h | 470 +++++++++--------- 33 files changed, 2694 insertions(+), 2309 deletions(-) (limited to 'src/gpgpu-sim/gpu-sim.h') diff --git a/.github/workflows/cmake.yml b/.github/workflows/cmake.yml index ab9bfd0..c0a22eb 100644 --- a/.github/workflows/cmake.yml +++ b/.github/workflows/cmake.yml @@ -5,6 +5,9 @@ name: Short-Tests-CMake on: # Triggers the workflow on push or pull request events but only for the mydev branch push: + branches-ignore: + - "gh-readonly-queue**" + merge_group: pull_request: # Allows you to run this workflow manually from the Actions tab diff --git a/.github/workflows/main.yml b/.github/workflows/main.yml index c639ff3..39f65c9 100644 --- a/.github/workflows/main.yml +++ b/.github/workflows/main.yml @@ -6,6 +6,9 @@ name: Short-Tests on: # Triggers the workflow on push or pull request events but only for the mydev branch push: + branches-ignore: + - "gh-readonly-queue**" + merge_group: pull_request: # Allows you to run this workflow manually from the Actions tab @@ -82,3 +85,26 @@ jobs: - uses: actions/checkout@v4 - name: Run Simulation run: /bin/bash $GITHUB_WORKSPACE/short-tests.sh + format-code: + runs-on: ubuntu-latest + needs: [build-TITANV, build-TITANV-LOCALXBAR, build-QV100, build-2060, build-3070] + permissions: + # Give the default GITHUB_TOKEN write permission to commit and push the + # added or changed files to the repository. + contents: write + steps: + - uses: actions/checkout@v4 + # Other steps that change files in the repository go here + # … + - name: Run clang-format + run: | + sudo apt-get install -y clang-format + /bin/bash ./format-code.sh + - uses: stefanzweifel/git-auto-commit-action@v5 + with: + # Optional. Commit message for the created commit. + # Defaults to "Apply automatic changes" + commit_message: Automated clang-format + # Optional. Option used by `git-status` to determine if the repository is + # dirty. See https://git-scm.com/docs/git-status#_options + status_options: '--untracked-files=no' \ No newline at end of file diff --git a/format-code.sh b/format-code.sh index ac753f0..acd33ab 100755 --- a/format-code.sh +++ b/format-code.sh @@ -1,5 +1,6 @@ # This bash script formats GPGPU-Sim using clang-format THIS_DIR="$( cd "$( dirname "$BASH_SOURCE" )" && pwd )" +echo "Running clang-format on $THIS_DIR" clang-format -i ${THIS_DIR}/libcuda/*.h clang-format -i ${THIS_DIR}/libcuda/*.cc clang-format -i ${THIS_DIR}/src/*.h diff --git a/libcuda/cuda_runtime_api.cc b/libcuda/cuda_runtime_api.cc index 5866b36..b64c3d9 100644 --- a/libcuda/cuda_runtime_api.cc +++ b/libcuda/cuda_runtime_api.cc @@ -133,16 +133,16 @@ #if (CUDART_VERSION < 8000) #include "__cudaFatFormat.h" #endif -#include "gpgpu_context.h" -#include "cuda_api_object.h" -#include "../src/gpgpu-sim/gpu-sim.h" -#include "../src/cuda-sim/ptx_loader.h" +#include "../src/abstract_hardware_model.h" #include "../src/cuda-sim/cuda-sim.h" #include "../src/cuda-sim/ptx_ir.h" +#include "../src/cuda-sim/ptx_loader.h" #include "../src/cuda-sim/ptx_parser.h" +#include "../src/gpgpu-sim/gpu-sim.h" #include "../src/gpgpusim_entrypoint.h" #include "../src/stream_manager.h" -#include "../src/abstract_hardware_model.h" +#include "cuda_api_object.h" +#include "gpgpu_context.h" #include #include @@ -464,7 +464,7 @@ static int get_app_cuda_version() { " | grep libcudart.so | sed 's/.*libcudart.so.\\(.*\\) =>.*/\\1/' > " + fname; int res = system(app_cuda_version_command.c_str()); - if(res == -1){ + if (res == -1) { printf("Error - Cannot detect the app's CUDA version.\n"); exit(1); } @@ -3239,8 +3239,8 @@ char *readfile(const std::string filename) { // allocate and copy the entire ptx char *ret = (char *)malloc((filesize + 1) * sizeof(char)); int num = fread(ret, 1, filesize, fp); - if(num == 0){ - std::cout << "ERROR: Could not read data from file %s\n" + if (num == 0) { + std::cout << "ERROR: Could not read data from file %s\n" << filename << std::endl; assert(0); } diff --git a/src/abstract_hardware_model.cc b/src/abstract_hardware_model.cc index ed7347d..fd056c6 100644 --- a/src/abstract_hardware_model.cc +++ b/src/abstract_hardware_model.cc @@ -1,17 +1,18 @@ -// Copyright (c) 2009-2021, Tor M. Aamodt, Inderpreet Singh, Timothy Rogers, Vijay Kandiah, Nikos Hardavellas, -// Mahmoud Khairy, Junrui Pan, Timothy G. Rogers -// The University of British Columbia, Northwestern University, Purdue University -// All rights reserved. +// Copyright (c) 2009-2021, Tor M. Aamodt, Inderpreet Singh, Timothy Rogers, +// Vijay Kandiah, Nikos Hardavellas, Mahmoud Khairy, Junrui Pan, Timothy G. +// Rogers The University of British Columbia, Northwestern University, Purdue +// University All rights reserved. // // Redistribution and use in source and binary forms, with or without // modification, are permitted provided that the following conditions are met: // -// 1. Redistributions of source code must retain the above copyright notice, this +// 1. Redistributions of source code must retain the above copyright notice, +// this // list of conditions and the following disclaimer; // 2. Redistributions in binary form must reproduce the above copyright notice, // this list of conditions and the following disclaimer in the documentation // and/or other materials provided with the distribution; -// 3. Neither the names of The University of British Columbia, Northwestern +// 3. Neither the names of The University of British Columbia, Northwestern // University nor the names of their contributors may be used to // endorse or promote products derived from this software without specific // prior written permission. @@ -28,7 +29,6 @@ // ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE // POSSIBILITY OF SUCH DAMAGE. - #include "abstract_hardware_model.h" #include #include @@ -284,7 +284,7 @@ void warp_inst_t::broadcast_barrier_reduction( void warp_inst_t::generate_mem_accesses() { if (empty() || op == MEMORY_BARRIER_OP || m_mem_accesses_created) return; if (!((op == LOAD_OP) || (op == TENSOR_CORE_LOAD_OP) || (op == STORE_OP) || - (op == TENSOR_CORE_STORE_OP) )) + (op == TENSOR_CORE_STORE_OP))) return; if (m_warp_active_mask.count() == 0) return; // predicated off @@ -292,8 +292,8 @@ void warp_inst_t::generate_mem_accesses() { assert(is_load() || is_store()); - //if((space.get_type() != tex_space) && (space.get_type() != const_space)) - assert(m_per_scalar_thread_valid); // need address information per thread + // if((space.get_type() != tex_space) && (space.get_type() != const_space)) + assert(m_per_scalar_thread_valid); // need address information per thread bool is_write = is_store(); diff --git a/src/abstract_hardware_model.h b/src/abstract_hardware_model.h index ebf6535..e5f3b78 100644 --- a/src/abstract_hardware_model.h +++ b/src/abstract_hardware_model.h @@ -1,17 +1,18 @@ -// Copyright (c) 2009-2021, Tor M. Aamodt, Inderpreet Singh, Vijay Kandiah, Nikos Hardavellas, -// Mahmoud Khairy, Junrui Pan, Timothy G. Rogers -// The University of British Columbia, Northwestern University, Purdue University +// Copyright (c) 2009-2021, Tor M. Aamodt, Inderpreet Singh, Vijay Kandiah, +// Nikos Hardavellas, Mahmoud Khairy, Junrui Pan, Timothy G. Rogers The +// University of British Columbia, Northwestern University, Purdue University // All rights reserved. // // Redistribution and use in source and binary forms, with or without // modification, are permitted provided that the following conditions are met: // -// 1. Redistributions of source code must retain the above copyright notice, this +// 1. Redistributions of source code must retain the above copyright notice, +// this // list of conditions and the following disclaimer; // 2. Redistributions in binary form must reproduce the above copyright notice, // this list of conditions and the following disclaimer in the documentation // and/or other materials provided with the distribution; -// 3. Neither the names of The University of British Columbia, Northwestern +// 3. Neither the names of The University of British Columbia, Northwestern // University nor the names of their contributors may be used to // endorse or promote products derived from this software without specific // prior written permission. @@ -64,24 +65,24 @@ enum _memory_space_t { #ifndef COEFF_STRUCT #define COEFF_STRUCT -struct PowerscalingCoefficients{ - double int_coeff; - double int_mul_coeff; - double int_mul24_coeff; - double int_mul32_coeff; - double int_div_coeff; - double fp_coeff; - double dp_coeff; - double fp_mul_coeff; - double fp_div_coeff; - double dp_mul_coeff; - double dp_div_coeff; - double sqrt_coeff; - double log_coeff; - double sin_coeff; - double exp_coeff; - double tensor_coeff; - double tex_coeff; +struct PowerscalingCoefficients { + double int_coeff; + double int_mul_coeff; + double int_mul24_coeff; + double int_mul32_coeff; + double int_div_coeff; + double fp_coeff; + double dp_coeff; + double fp_mul_coeff; + double fp_div_coeff; + double dp_mul_coeff; + double dp_div_coeff; + double sqrt_coeff; + double log_coeff; + double sin_coeff; + double exp_coeff; + double tensor_coeff; + double tex_coeff; }; #endif @@ -974,18 +975,22 @@ class inst_t { memory_op == memory_store); } - bool is_fp() const { return ((sp_op == FP__OP));} //VIJAY - bool is_fpdiv() const { return ((sp_op == FP_DIV_OP));} - bool is_fpmul() const { return ((sp_op == FP_MUL_OP));} - bool is_dp() const { return ((sp_op == DP___OP));} - bool is_dpdiv() const { return ((sp_op == DP_DIV_OP));} - bool is_dpmul() const { return ((sp_op == DP_MUL_OP));} - bool is_imul() const { return ((sp_op == INT_MUL_OP));} - bool is_imul24() const { return ((sp_op == INT_MUL24_OP));} - bool is_imul32() const { return ((sp_op == INT_MUL32_OP));} - bool is_idiv() const { return ((sp_op == INT_DIV_OP));} - bool is_sfu() const {return ((sp_op == FP_SQRT_OP) || (sp_op == FP_LG_OP) || (sp_op == FP_SIN_OP) || (sp_op == FP_EXP_OP) || (sp_op == TENSOR__OP));} - bool is_alu() const {return (sp_op == INT__OP);} + bool is_fp() const { return ((sp_op == FP__OP)); } // VIJAY + bool is_fpdiv() const { return ((sp_op == FP_DIV_OP)); } + bool is_fpmul() const { return ((sp_op == FP_MUL_OP)); } + bool is_dp() const { return ((sp_op == DP___OP)); } + bool is_dpdiv() const { return ((sp_op == DP_DIV_OP)); } + bool is_dpmul() const { return ((sp_op == DP_MUL_OP)); } + bool is_imul() const { return ((sp_op == INT_MUL_OP)); } + bool is_imul24() const { return ((sp_op == INT_MUL24_OP)); } + bool is_imul32() const { return ((sp_op == INT_MUL32_OP)); } + bool is_idiv() const { return ((sp_op == INT_DIV_OP)); } + bool is_sfu() const { + return ((sp_op == FP_SQRT_OP) || (sp_op == FP_LG_OP) || + (sp_op == FP_SIN_OP) || (sp_op == FP_EXP_OP) || + (sp_op == TENSOR__OP)); + } + bool is_alu() const { return (sp_op == INT__OP); } unsigned get_num_operands() const { return num_operands; } unsigned get_num_regs() const { return num_regs; } @@ -1010,7 +1015,7 @@ class inst_t { operation_pipeline op_pipe; // code (uarch visible) identify the pipeline of // the operation (SP, SFU or MEM) mem_operation mem_op; // code (uarch visible) identify memory type - bool const_cache_operand; // has a load from constant memory as an operand + bool const_cache_operand; // has a load from constant memory as an operand _memory_op_t memory_op; // memory_op used by ptxplus unsigned num_operands; unsigned num_regs; // count vector operand as one register operand @@ -1057,7 +1062,7 @@ class warp_inst_t : public inst_t { m_empty = true; m_config = NULL; - // Ni: + // Ni: m_is_ldgsts = false; m_is_ldgdepbar = false; m_is_depbar = false; @@ -1077,7 +1082,7 @@ class warp_inst_t : public inst_t { m_is_cdp = 0; should_do_atomic = true; - // Ni: + // Ni: m_is_ldgsts = false; m_is_ldgdepbar = false; m_is_depbar = false; diff --git a/src/accelwattch/XML_Parse.cc b/src/accelwattch/XML_Parse.cc index eaec748..801de6f 100644 --- a/src/accelwattch/XML_Parse.cc +++ b/src/accelwattch/XML_Parse.cc @@ -30,14 +30,13 @@ ***************************************************************************/ /******************************************************************** * Modified by: - * Jingwen Leng, University of Texas, Austin - * Syed Gilani, University of Wisconsin–Madison + * Jingwen Leng, University of Texas, Austin + * Syed Gilani, University of Wisconsin–Madison * Tayler Hetherington, University of British Columbia * Ahmed ElTantawy, University of British Columbia * Vijay Kandiah, Northwestern University ********************************************************************/ - #include "XML_Parse.h" #include #include @@ -45,14 +44,18 @@ using namespace std; -const char * perf_count_label[] = { - "TOT_INST,", "FP_INT,", "IC_H,", "IC_M,", "DC_RH,", "DC_RM,", "DC_WH,", "DC_WM,", - "TC_H,", "TC_M,", "CC_H,", "CC_M,", "SHRD_ACC,", "REG_RD,", "REG_WR,", "NON_REG_OPs,", - "INT_ACC,", "FPU_ACC,", "DPU_ACC,", "INT_MUL24_ACC,", "INT_MUL32_ACC,", "INT_MUL_ACC,","INT_DIV_ACC,", - "FP_MUL_ACC,", "FP_DIV_ACC,", "FP_SQRT_ACC,", "FP_LG_ACC,", "FP_SIN_ACC,", "FP_EXP_ACC,", "DP_MUL_ACC,", - "DP_DIV_ACC,", "TENSOR_ACC,", "TEX_ACC,", "MEM_RD,","MEM_WR,", "MEM_PRE,", "L2_RH,", "L2_RM,", "L2_WH,", - "L2_WM,", "NOC_A,", "PIPE_A,", "IDLE_CORE_N,", "constant_power"}; - +const char* perf_count_label[] = { + "TOT_INST,", "FP_INT,", "IC_H,", "IC_M,", + "DC_RH,", "DC_RM,", "DC_WH,", "DC_WM,", + "TC_H,", "TC_M,", "CC_H,", "CC_M,", + "SHRD_ACC,", "REG_RD,", "REG_WR,", "NON_REG_OPs,", + "INT_ACC,", "FPU_ACC,", "DPU_ACC,", "INT_MUL24_ACC,", + "INT_MUL32_ACC,", "INT_MUL_ACC,", "INT_DIV_ACC,", "FP_MUL_ACC,", + "FP_DIV_ACC,", "FP_SQRT_ACC,", "FP_LG_ACC,", "FP_SIN_ACC,", + "FP_EXP_ACC,", "DP_MUL_ACC,", "DP_DIV_ACC,", "TENSOR_ACC,", + "TEX_ACC,", "MEM_RD,", "MEM_WR,", "MEM_PRE,", + "L2_RH,", "L2_RM,", "L2_WH,", "L2_WM,", + "NOC_A,", "PIPE_A,", "IDLE_CORE_N,", "constant_power"}; void ParseXML::parse(char* filepath) { unsigned int i, j, k, m, n; @@ -171,8 +174,6 @@ void ParseXML::parse(char* filepath) { continue; } - - if (strcmp(xNode2.getChildNode("param", i).getAttribute("name"), "static_cat1_flane") == 0) { sys.static_cat1_flane = @@ -355,7 +356,6 @@ void ParseXML::parse(char* filepath) { continue; } - if (strcmp(xNode2.getChildNode("param", i).getAttribute("name"), "target_chip_area") == 0) { sys.target_chip_area = @@ -615,106 +615,106 @@ void ParseXML::parse(char* filepath) { atof(xNode2.getChildNode("param", i).getAttribute("value")); continue; } - if (strcmp(xNode2.getChildNode("param",i).getAttribute("name"), - "INT_ACC")==0) { + if (strcmp(xNode2.getChildNode("param", i).getAttribute("name"), + "INT_ACC") == 0) { sys.scaling_coefficients[INT_ACC] = - atof(xNode2.getChildNode("param",i).getAttribute("value")); + atof(xNode2.getChildNode("param", i).getAttribute("value")); continue; } - if (strcmp(xNode2.getChildNode("param",i).getAttribute("name"), - "FP_ACC")==0) { + if (strcmp(xNode2.getChildNode("param", i).getAttribute("name"), + "FP_ACC") == 0) { sys.scaling_coefficients[FP_ACC] = - atof(xNode2.getChildNode("param",i).getAttribute("value")); + atof(xNode2.getChildNode("param", i).getAttribute("value")); continue; } - if (strcmp(xNode2.getChildNode("param",i).getAttribute("name"), - "DP_ACC")==0) { + if (strcmp(xNode2.getChildNode("param", i).getAttribute("name"), + "DP_ACC") == 0) { sys.scaling_coefficients[DP_ACC] = - atof(xNode2.getChildNode("param",i).getAttribute("value")); + atof(xNode2.getChildNode("param", i).getAttribute("value")); continue; } - if (strcmp(xNode2.getChildNode("param",i).getAttribute("name"), - "INT_MUL24_ACC")==0) { + if (strcmp(xNode2.getChildNode("param", i).getAttribute("name"), + "INT_MUL24_ACC") == 0) { sys.scaling_coefficients[INT_MUL24_ACC] = - atof(xNode2.getChildNode("param",i).getAttribute("value")); + atof(xNode2.getChildNode("param", i).getAttribute("value")); continue; } - if (strcmp(xNode2.getChildNode("param",i).getAttribute("name"), - "INT_MUL32_ACC")==0) { + if (strcmp(xNode2.getChildNode("param", i).getAttribute("name"), + "INT_MUL32_ACC") == 0) { sys.scaling_coefficients[INT_MUL32_ACC] = - atof(xNode2.getChildNode("param",i).getAttribute("value")); + atof(xNode2.getChildNode("param", i).getAttribute("value")); continue; } - if (strcmp(xNode2.getChildNode("param",i).getAttribute("name"), - "INT_MUL_ACC")==0) { + if (strcmp(xNode2.getChildNode("param", i).getAttribute("name"), + "INT_MUL_ACC") == 0) { sys.scaling_coefficients[INT_MUL_ACC] = - atof(xNode2.getChildNode("param",i).getAttribute("value")); + atof(xNode2.getChildNode("param", i).getAttribute("value")); continue; } - if (strcmp(xNode2.getChildNode("param",i).getAttribute("name"), - "INT_DIV_ACC")==0) { + if (strcmp(xNode2.getChildNode("param", i).getAttribute("name"), + "INT_DIV_ACC") == 0) { sys.scaling_coefficients[INT_DIV_ACC] = - atof(xNode2.getChildNode("param",i).getAttribute("value")); + atof(xNode2.getChildNode("param", i).getAttribute("value")); continue; } - if (strcmp(xNode2.getChildNode("param",i).getAttribute("name"), - "FP_MUL_ACC")==0) { + if (strcmp(xNode2.getChildNode("param", i).getAttribute("name"), + "FP_MUL_ACC") == 0) { sys.scaling_coefficients[FP_MUL_ACC] = - atof(xNode2.getChildNode("param",i).getAttribute("value")); + atof(xNode2.getChildNode("param", i).getAttribute("value")); continue; } - if (strcmp(xNode2.getChildNode("param",i).getAttribute("name"), - "FP_DIV_ACC")==0) { + if (strcmp(xNode2.getChildNode("param", i).getAttribute("name"), + "FP_DIV_ACC") == 0) { sys.scaling_coefficients[FP_DIV_ACC] = - atof(xNode2.getChildNode("param",i).getAttribute("value")); + atof(xNode2.getChildNode("param", i).getAttribute("value")); continue; } - if (strcmp(xNode2.getChildNode("param",i).getAttribute("name"), - "FP_SQRT_ACC")==0) { + if (strcmp(xNode2.getChildNode("param", i).getAttribute("name"), + "FP_SQRT_ACC") == 0) { sys.scaling_coefficients[FP_SQRT_ACC] = - atof(xNode2.getChildNode("param",i).getAttribute("value")); + atof(xNode2.getChildNode("param", i).getAttribute("value")); continue; } - if (strcmp(xNode2.getChildNode("param",i).getAttribute("name"), - "FP_LG_ACC")==0) { + if (strcmp(xNode2.getChildNode("param", i).getAttribute("name"), + "FP_LG_ACC") == 0) { sys.scaling_coefficients[FP_LG_ACC] = - atof(xNode2.getChildNode("param",i).getAttribute("value")); + atof(xNode2.getChildNode("param", i).getAttribute("value")); continue; } - if (strcmp(xNode2.getChildNode("param",i).getAttribute("name"), - "FP_SIN_ACC")==0) { + if (strcmp(xNode2.getChildNode("param", i).getAttribute("name"), + "FP_SIN_ACC") == 0) { sys.scaling_coefficients[FP_SIN_ACC] = - atof(xNode2.getChildNode("param",i).getAttribute("value")); + atof(xNode2.getChildNode("param", i).getAttribute("value")); continue; } - if (strcmp(xNode2.getChildNode("param",i).getAttribute("name"), - "FP_EXP_ACC")==0) { + if (strcmp(xNode2.getChildNode("param", i).getAttribute("name"), + "FP_EXP_ACC") == 0) { sys.scaling_coefficients[FP_EXP_ACC] = - atof(xNode2.getChildNode("param",i).getAttribute("value")); + atof(xNode2.getChildNode("param", i).getAttribute("value")); continue; } - if (strcmp(xNode2.getChildNode("param",i).getAttribute("name"), - "DP_MUL_ACC")==0) { + if (strcmp(xNode2.getChildNode("param", i).getAttribute("name"), + "DP_MUL_ACC") == 0) { sys.scaling_coefficients[DP_MUL_ACC] = - atof(xNode2.getChildNode("param",i).getAttribute("value")); + atof(xNode2.getChildNode("param", i).getAttribute("value")); continue; } - if (strcmp(xNode2.getChildNode("param",i).getAttribute("name"), - "DP_DIV_ACC")==0) { + if (strcmp(xNode2.getChildNode("param", i).getAttribute("name"), + "DP_DIV_ACC") == 0) { sys.scaling_coefficients[DP_DIV_ACC] = - atof(xNode2.getChildNode("param",i).getAttribute("value")); + atof(xNode2.getChildNode("param", i).getAttribute("value")); continue; } - if (strcmp(xNode2.getChildNode("param",i).getAttribute("name"), - "TENSOR_ACC")==0) { + if (strcmp(xNode2.getChildNode("param", i).getAttribute("name"), + "TENSOR_ACC") == 0) { sys.scaling_coefficients[TENSOR_ACC] = - atof(xNode2.getChildNode("param",i).getAttribute("value")); + atof(xNode2.getChildNode("param", i).getAttribute("value")); continue; } - if (strcmp(xNode2.getChildNode("param",i).getAttribute("name"), - "TEX_ACC")==0) { + if (strcmp(xNode2.getChildNode("param", i).getAttribute("name"), + "TEX_ACC") == 0) { sys.scaling_coefficients[TEX_ACC] = - atof(xNode2.getChildNode("param",i).getAttribute("value")); + atof(xNode2.getChildNode("param", i).getAttribute("value")); continue; } if (strcmp(xNode2.getChildNode("param", i).getAttribute("name"), diff --git a/src/accelwattch/XML_Parse.h b/src/accelwattch/XML_Parse.h index c82359f..176b82f 100644 --- a/src/accelwattch/XML_Parse.h +++ b/src/accelwattch/XML_Parse.h @@ -30,8 +30,8 @@ ***************************************************************************/ /******************************************************************** * Modified by: - * Jingwen Leng, University of Texas, Austin - * Syed Gilani, University of Wisconsin–Madison + * Jingwen Leng, University of Texas, Austin + * Syed Gilani, University of Wisconsin–Madison * Tayler Hetherington, University of British Columbia * Ahmed ElTantawy, University of British Columbia * Vijay Kandiah, Northwestern University @@ -70,7 +70,7 @@ ToXMLStringTool tx,tx2; extern const char* perf_count_label[]; enum perf_count_t { - TOT_INST=0, + TOT_INST = 0, FP_INT, IC_H, IC_M, @@ -86,23 +86,23 @@ enum perf_count_t { REG_RD, REG_WR, NON_REG_OPs, - INT_ACC, //SPU - FP_ACC, //FPU - DP_ACC, //FPU - INT_MUL24_ACC, //SFU - INT_MUL32_ACC, //SFU - INT_MUL_ACC, //SFU - INT_DIV_ACC, //SFU - FP_MUL_ACC, //SFU - FP_DIV_ACC, //SFU - FP_SQRT_ACC, //SFU - FP_LG_ACC, //SFU - FP_SIN_ACC, //SFU - FP_EXP_ACC, //SFU - DP_MUL_ACC, //SFU - DP_DIV_ACC, //SFU - TENSOR_ACC, //SFU - TEX_ACC, //SFU + INT_ACC, // SPU + FP_ACC, // FPU + DP_ACC, // FPU + INT_MUL24_ACC, // SFU + INT_MUL32_ACC, // SFU + INT_MUL_ACC, // SFU + INT_DIV_ACC, // SFU + FP_MUL_ACC, // SFU + FP_DIV_ACC, // SFU + FP_SQRT_ACC, // SFU + FP_LG_ACC, // SFU + FP_SIN_ACC, // SFU + FP_EXP_ACC, // SFU + DP_MUL_ACC, // SFU + DP_DIV_ACC, // SFU + TENSOR_ACC, // SFU + TEX_ACC, // SFU MEM_RD, MEM_WR, MEM_PRE, diff --git a/src/accelwattch/gpgpu_sim_wrapper.cc b/src/accelwattch/gpgpu_sim_wrapper.cc index 67d9daa..4883c7c 100644 --- a/src/accelwattch/gpgpu_sim_wrapper.cc +++ b/src/accelwattch/gpgpu_sim_wrapper.cc @@ -1,16 +1,17 @@ -// Copyright (c) 2009-2021, Tor M. Aamodt, Tayler Hetherington, Ahmed ElTantawy, Vijay Kandiah, Nikos Hardavellas -// The University of British Columbia, Northwestern University -// All rights reserved. +// Copyright (c) 2009-2021, Tor M. Aamodt, Tayler Hetherington, Ahmed ElTantawy, +// Vijay Kandiah, Nikos Hardavellas The University of British Columbia, +// Northwestern University All rights reserved. // // Redistribution and use in source and binary forms, with or without // modification, are permitted provided that the following conditions are met: // -// 1. Redistributions of source code must retain the above copyright notice, this +// 1. Redistributions of source code must retain the above copyright notice, +// this // list of conditions and the following disclaimer; // 2. Redistributions in binary form must reproduce the above copyright notice, // this list of conditions and the following disclaimer in the documentation // and/or other materials provided with the distribution; -// 3. Neither the names of The University of British Columbia, Northwestern +// 3. Neither the names of The University of British Columbia, Northwestern // University nor the names of their contributors may be used to // endorse or promote products derived from this software without specific // prior written permission. @@ -33,14 +34,16 @@ #define SFU_BASE_POWER 0 static const char* pwr_cmp_label[] = { - "IBP,", "ICP,", "DCP,", "TCP,", "CCP,", "SHRDP,", "RFP,", "INTP,", - "FPUP,", "DPUP,", "INT_MUL24P,", "INT_MUL32P,", "INT_MULP,", "INT_DIVP,", - "FP_MULP,", "FP_DIVP,", "FP_SQRTP,", "FP_LGP,", "FP_SINP,", "FP_EXP,", - "DP_MULP,", "DP_DIVP,", "TENSORP,", "TEXP,", "SCHEDP,", "L2CP,", "MCP,", "NOCP,", - "DRAMP,", "PIPEP,", "IDLE_COREP,", "CONSTP", "STATICP"}; + "IBP,", "ICP,", "DCP,", "TCP,", "CCP,", + "SHRDP,", "RFP,", "INTP,", "FPUP,", "DPUP,", + "INT_MUL24P,", "INT_MUL32P,", "INT_MULP,", "INT_DIVP,", "FP_MULP,", + "FP_DIVP,", "FP_SQRTP,", "FP_LGP,", "FP_SINP,", "FP_EXP,", + "DP_MULP,", "DP_DIVP,", "TENSORP,", "TEXP,", "SCHEDP,", + "L2CP,", "MCP,", "NOCP,", "DRAMP,", "PIPEP,", + "IDLE_COREP,", "CONSTP", "STATICP"}; enum pwr_cmp_t { - IBP=0, + IBP = 0, ICP, DCP, TCP, @@ -77,7 +80,8 @@ enum pwr_cmp_t { }; gpgpu_sim_wrapper::gpgpu_sim_wrapper(bool power_simulation_enabled, - char* xmlfile, int power_simulation_mode, bool dvfs_enabled) { + char* xmlfile, int power_simulation_mode, + bool dvfs_enabled) { kernel_sample_count = 0; total_sample_count = 0; @@ -142,7 +146,8 @@ bool gpgpu_sim_wrapper::sanity_check(double a, double b) { return false; } void gpgpu_sim_wrapper::init_mcpat_hw_mode(unsigned gpu_sim_cycle) { - p->sys.total_cycles = gpu_sim_cycle; //total simulated cycles for current kernel + p->sys.total_cycles = + gpu_sim_cycle; // total simulated cycles for current kernel } void gpgpu_sim_wrapper::init_mcpat( @@ -150,9 +155,9 @@ void gpgpu_sim_wrapper::init_mcpat( char* metric_trace_filename, char* steady_state_filename, bool power_sim_enabled, bool trace_enabled, bool steady_state_enabled, bool power_per_cycle_dump, double steady_power_deviation, - double steady_min_period, int zlevel, double init_val, - int stat_sample_freq, int power_sim_mode, bool dvfs_enabled, - unsigned clock_freq, unsigned num_shaders) { + double steady_min_period, int zlevel, double init_val, int stat_sample_freq, + int power_sim_mode, bool dvfs_enabled, unsigned clock_freq, + unsigned num_shaders) { // Write File Headers for (-metrics trace, -power trace) reset_counters(); @@ -382,10 +387,7 @@ void gpgpu_sim_wrapper::set_l2cache_power(double read_hits, double read_misses, sample_perf_counters[L2_WM] = write_misses; } -void gpgpu_sim_wrapper::set_num_cores(double num_core) { - - num_cores = num_core; -} +void gpgpu_sim_wrapper::set_num_cores(double num_core) { num_cores = num_core; } void gpgpu_sim_wrapper::set_idle_core_power(double num_idle_core) { p->sys.num_idle_cores = num_idle_core; @@ -411,29 +413,25 @@ void gpgpu_sim_wrapper::set_mem_ctrl_power(double reads, double writes, sample_perf_counters[MEM_PRE] = dram_precharge; } - void gpgpu_sim_wrapper::set_model_voltage(double model_voltage) { - modeled_chip_voltage = model_voltage; + modeled_chip_voltage = model_voltage; } - void gpgpu_sim_wrapper::set_exec_unit_power(double fpu_accesses, double ialu_accesses, double sfu_accesses) { p->sys.core[0].fpu_accesses = fpu_accesses; tot_fpu_accesses = fpu_accesses; - //Integer ALU (not present in Tesla) + // Integer ALU (not present in Tesla) p->sys.core[0].ialu_accesses = ialu_accesses; - //Sfu accesses + // Sfu accesses p->sys.core[0].mul_accesses = sfu_accesses; tot_sfu_accesses = sfu_accesses; } -PowerscalingCoefficients * gpgpu_sim_wrapper::get_scaling_coeffs() -{ - - PowerscalingCoefficients * scalingCoeffs = new PowerscalingCoefficients(); +PowerscalingCoefficients* gpgpu_sim_wrapper::get_scaling_coeffs() { + PowerscalingCoefficients* scalingCoeffs = new PowerscalingCoefficients(); scalingCoeffs->int_coeff = p->sys.scaling_coefficients[INT_ACC]; scalingCoeffs->int_mul_coeff = p->sys.scaling_coefficients[INT_MUL_ACC]; @@ -453,68 +451,55 @@ PowerscalingCoefficients * gpgpu_sim_wrapper::get_scaling_coeffs() scalingCoeffs->tensor_coeff = p->sys.scaling_coefficients[TENSOR_ACC]; scalingCoeffs->tex_coeff = p->sys.scaling_coefficients[TEX_ACC]; return scalingCoeffs; - } -void gpgpu_sim_wrapper::set_int_accesses(double ialu_accesses, - double imul24_accesses, - double imul32_accesses, - double imul_accesses, - double idiv_accesses) -{ - - sample_perf_counters[INT_ACC]=ialu_accesses; - sample_perf_counters[INT_MUL24_ACC]=imul24_accesses; - sample_perf_counters[INT_MUL32_ACC]=imul32_accesses; - sample_perf_counters[INT_MUL_ACC]=imul_accesses; - sample_perf_counters[INT_DIV_ACC]=idiv_accesses; +void gpgpu_sim_wrapper::set_int_accesses(double ialu_accesses, + double imul24_accesses, + double imul32_accesses, + double imul_accesses, + double idiv_accesses) { + sample_perf_counters[INT_ACC] = ialu_accesses; + sample_perf_counters[INT_MUL24_ACC] = imul24_accesses; + sample_perf_counters[INT_MUL32_ACC] = imul32_accesses; + sample_perf_counters[INT_MUL_ACC] = imul_accesses; + sample_perf_counters[INT_DIV_ACC] = idiv_accesses; } -void gpgpu_sim_wrapper::set_dp_accesses(double dpu_accesses, - double dpmul_accesses, - double dpdiv_accesses) -{ - sample_perf_counters[DP_ACC]=dpu_accesses; - sample_perf_counters[DP_MUL_ACC]=dpmul_accesses; - sample_perf_counters[DP_DIV_ACC]=dpdiv_accesses; +void gpgpu_sim_wrapper::set_dp_accesses(double dpu_accesses, + double dpmul_accesses, + double dpdiv_accesses) { + sample_perf_counters[DP_ACC] = dpu_accesses; + sample_perf_counters[DP_MUL_ACC] = dpmul_accesses; + sample_perf_counters[DP_DIV_ACC] = dpdiv_accesses; } -void gpgpu_sim_wrapper::set_fp_accesses(double fpu_accesses, - double fpmul_accesses, - double fpdiv_accesses) -{ - sample_perf_counters[FP_ACC]=fpu_accesses; - sample_perf_counters[FP_MUL_ACC]=fpmul_accesses; - sample_perf_counters[FP_DIV_ACC]=fpdiv_accesses; +void gpgpu_sim_wrapper::set_fp_accesses(double fpu_accesses, + double fpmul_accesses, + double fpdiv_accesses) { + sample_perf_counters[FP_ACC] = fpu_accesses; + sample_perf_counters[FP_MUL_ACC] = fpmul_accesses; + sample_perf_counters[FP_DIV_ACC] = fpdiv_accesses; } -void gpgpu_sim_wrapper::set_trans_accesses(double sqrt_accesses, - double log_accesses, - double sin_accesses, - double exp_accesses) -{ - - sample_perf_counters[FP_SQRT_ACC]=sqrt_accesses; - sample_perf_counters[FP_LG_ACC]=log_accesses; - sample_perf_counters[FP_SIN_ACC]=sin_accesses; - sample_perf_counters[FP_EXP_ACC]=exp_accesses; - +void gpgpu_sim_wrapper::set_trans_accesses(double sqrt_accesses, + double log_accesses, + double sin_accesses, + double exp_accesses) { + sample_perf_counters[FP_SQRT_ACC] = sqrt_accesses; + sample_perf_counters[FP_LG_ACC] = log_accesses; + sample_perf_counters[FP_SIN_ACC] = sin_accesses; + sample_perf_counters[FP_EXP_ACC] = exp_accesses; } -void gpgpu_sim_wrapper::set_tensor_accesses(double tensor_accesses) -{ - sample_perf_counters[TENSOR_ACC]=tensor_accesses; - +void gpgpu_sim_wrapper::set_tensor_accesses(double tensor_accesses) { + sample_perf_counters[TENSOR_ACC] = tensor_accesses; } -void gpgpu_sim_wrapper::set_tex_accesses(double tex_accesses) -{ - sample_perf_counters[TEX_ACC]=tex_accesses; - +void gpgpu_sim_wrapper::set_tex_accesses(double tex_accesses) { + sample_perf_counters[TEX_ACC] = tex_accesses; } -void gpgpu_sim_wrapper::set_avg_active_threads(float active_threads) -{ +void gpgpu_sim_wrapper::set_avg_active_threads(float active_threads) { avg_threads_per_warp = (unsigned)ceil(active_threads); avg_threads_per_warp_tot += active_threads; } @@ -536,7 +521,8 @@ void gpgpu_sim_wrapper::power_metrics_calculations() { kernel_sample_count++; // Current sample power - double sample_power = proc->rt_power.readOp.dynamic + sample_cmp_pwr[CONSTP] + sample_cmp_pwr[STATICP]; + double sample_power = proc->rt_power.readOp.dynamic + sample_cmp_pwr[CONSTP] + + sample_cmp_pwr[STATICP]; // double sample_power; // for(unsigned i=0; icores[0]->get_coefficient_fpint_insts(); - effpower_coeff[FP_INT]=initpower_coeff[FP_INT] * p->sys.scaling_coefficients[FP_INT]; - - initpower_coeff[TOT_INST]=proc->cores[0]->get_coefficient_tot_insts(); - effpower_coeff[TOT_INST]=initpower_coeff[TOT_INST] * p->sys.scaling_coefficients[TOT_INST]; - - initpower_coeff[REG_RD]=proc->cores[0]->get_coefficient_regreads_accesses()*(proc->cores[0]->exu->rf_fu_clockRate/proc->cores[0]->exu->clockRate); - initpower_coeff[REG_WR]=proc->cores[0]->get_coefficient_regwrites_accesses()*(proc->cores[0]->exu->rf_fu_clockRate/proc->cores[0]->exu->clockRate); - initpower_coeff[NON_REG_OPs]=proc->cores[0]->get_coefficient_noregfileops_accesses()*(proc->cores[0]->exu->rf_fu_clockRate/proc->cores[0]->exu->clockRate); - effpower_coeff[REG_RD]=initpower_coeff[REG_RD]*p->sys.scaling_coefficients[REG_RD]; - effpower_coeff[REG_WR]=initpower_coeff[REG_WR]*p->sys.scaling_coefficients[REG_WR]; - effpower_coeff[NON_REG_OPs]=initpower_coeff[NON_REG_OPs]*p->sys.scaling_coefficients[NON_REG_OPs]; - - initpower_coeff[IC_H]=proc->cores[0]->get_coefficient_icache_hits(); - initpower_coeff[IC_M]=proc->cores[0]->get_coefficient_icache_misses(); - effpower_coeff[IC_H]=initpower_coeff[IC_H]*p->sys.scaling_coefficients[IC_H]; - effpower_coeff[IC_M]=initpower_coeff[IC_M]*p->sys.scaling_coefficients[IC_M]; - - initpower_coeff[CC_H]=(proc->cores[0]->get_coefficient_ccache_readhits()+proc->get_coefficient_readcoalescing()); - initpower_coeff[CC_M]=(proc->cores[0]->get_coefficient_ccache_readmisses()+proc->get_coefficient_readcoalescing()); - effpower_coeff[CC_H]=initpower_coeff[CC_H]*p->sys.scaling_coefficients[CC_H]; - effpower_coeff[CC_M]=initpower_coeff[CC_M]*p->sys.scaling_coefficients[CC_M]; - - initpower_coeff[TC_H]=(proc->cores[0]->get_coefficient_tcache_readhits()+proc->get_coefficient_readcoalescing()); - initpower_coeff[TC_M]=(proc->cores[0]->get_coefficient_tcache_readmisses()+proc->get_coefficient_readcoalescing()); - effpower_coeff[TC_H]=initpower_coeff[TC_H]*p->sys.scaling_coefficients[TC_H]; - effpower_coeff[TC_M]=initpower_coeff[TC_M]*p->sys.scaling_coefficients[TC_M]; - - initpower_coeff[SHRD_ACC]=proc->cores[0]->get_coefficient_sharedmemory_readhits(); - effpower_coeff[SHRD_ACC]=initpower_coeff[SHRD_ACC]*p->sys.scaling_coefficients[SHRD_ACC]; - - initpower_coeff[DC_RH]=(proc->cores[0]->get_coefficient_dcache_readhits() + proc->get_coefficient_readcoalescing()); - initpower_coeff[DC_RM]=(proc->cores[0]->get_coefficient_dcache_readmisses() + proc->get_coefficient_readcoalescing()); - initpower_coeff[DC_WH]=(proc->cores[0]->get_coefficient_dcache_writehits() + proc->get_coefficient_writecoalescing()); - initpower_coeff[DC_WM]=(proc->cores[0]->get_coefficient_dcache_writemisses() + proc->get_coefficient_writecoalescing()); - effpower_coeff[DC_RH]=initpower_coeff[DC_RH]*p->sys.scaling_coefficients[DC_RH]; - effpower_coeff[DC_RM]=initpower_coeff[DC_RM]*p->sys.scaling_coefficients[DC_RM]; - effpower_coeff[DC_WH]=initpower_coeff[DC_WH]*p->sys.scaling_coefficients[DC_WH]; - effpower_coeff[DC_WM]=initpower_coeff[DC_WM]*p->sys.scaling_coefficients[DC_WM]; - - initpower_coeff[L2_RH]=proc->get_coefficient_l2_read_hits(); - initpower_coeff[L2_RM]=proc->get_coefficient_l2_read_misses(); - initpower_coeff[L2_WH]=proc->get_coefficient_l2_write_hits(); - initpower_coeff[L2_WM]=proc->get_coefficient_l2_write_misses(); - effpower_coeff[L2_RH]=initpower_coeff[L2_RH]*p->sys.scaling_coefficients[L2_RH]; - effpower_coeff[L2_RM]=initpower_coeff[L2_RM]*p->sys.scaling_coefficients[L2_RM]; - effpower_coeff[L2_WH]=initpower_coeff[L2_WH]*p->sys.scaling_coefficients[L2_WH]; - effpower_coeff[L2_WM]=initpower_coeff[L2_WM]*p->sys.scaling_coefficients[L2_WM]; - - initpower_coeff[IDLE_CORE_N]=p->sys.idle_core_power * proc->cores[0]->executionTime; - effpower_coeff[IDLE_CORE_N]=initpower_coeff[IDLE_CORE_N]*p->sys.scaling_coefficients[IDLE_CORE_N]; - - initpower_coeff[PIPE_A]=proc->cores[0]->get_coefficient_duty_cycle(); - effpower_coeff[PIPE_A]=initpower_coeff[PIPE_A]*p->sys.scaling_coefficients[PIPE_A]; - - initpower_coeff[MEM_RD]=proc->get_coefficient_mem_reads(); - initpower_coeff[MEM_WR]=proc->get_coefficient_mem_writes(); - initpower_coeff[MEM_PRE]=proc->get_coefficient_mem_pre(); - effpower_coeff[MEM_RD]=initpower_coeff[MEM_RD]*p->sys.scaling_coefficients[MEM_RD]; - effpower_coeff[MEM_WR]=initpower_coeff[MEM_WR]*p->sys.scaling_coefficients[MEM_WR]; - effpower_coeff[MEM_PRE]=initpower_coeff[MEM_PRE]*p->sys.scaling_coefficients[MEM_PRE]; - +void gpgpu_sim_wrapper::update_coefficients() { + initpower_coeff[FP_INT] = proc->cores[0]->get_coefficient_fpint_insts(); + effpower_coeff[FP_INT] = + initpower_coeff[FP_INT] * p->sys.scaling_coefficients[FP_INT]; + + initpower_coeff[TOT_INST] = proc->cores[0]->get_coefficient_tot_insts(); + effpower_coeff[TOT_INST] = + initpower_coeff[TOT_INST] * p->sys.scaling_coefficients[TOT_INST]; + + initpower_coeff[REG_RD] = + proc->cores[0]->get_coefficient_regreads_accesses() * + (proc->cores[0]->exu->rf_fu_clockRate / proc->cores[0]->exu->clockRate); + initpower_coeff[REG_WR] = + proc->cores[0]->get_coefficient_regwrites_accesses() * + (proc->cores[0]->exu->rf_fu_clockRate / proc->cores[0]->exu->clockRate); + initpower_coeff[NON_REG_OPs] = + proc->cores[0]->get_coefficient_noregfileops_accesses() * + (proc->cores[0]->exu->rf_fu_clockRate / proc->cores[0]->exu->clockRate); + effpower_coeff[REG_RD] = + initpower_coeff[REG_RD] * p->sys.scaling_coefficients[REG_RD]; + effpower_coeff[REG_WR] = + initpower_coeff[REG_WR] * p->sys.scaling_coefficients[REG_WR]; + effpower_coeff[NON_REG_OPs] = + initpower_coeff[NON_REG_OPs] * p->sys.scaling_coefficients[NON_REG_OPs]; + + initpower_coeff[IC_H] = proc->cores[0]->get_coefficient_icache_hits(); + initpower_coeff[IC_M] = proc->cores[0]->get_coefficient_icache_misses(); + effpower_coeff[IC_H] = + initpower_coeff[IC_H] * p->sys.scaling_coefficients[IC_H]; + effpower_coeff[IC_M] = + initpower_coeff[IC_M] * p->sys.scaling_coefficients[IC_M]; + + initpower_coeff[CC_H] = (proc->cores[0]->get_coefficient_ccache_readhits() + + proc->get_coefficient_readcoalescing()); + initpower_coeff[CC_M] = (proc->cores[0]->get_coefficient_ccache_readmisses() + + proc->get_coefficient_readcoalescing()); + effpower_coeff[CC_H] = + initpower_coeff[CC_H] * p->sys.scaling_coefficients[CC_H]; + effpower_coeff[CC_M] = + initpower_coeff[CC_M] * p->sys.scaling_coefficients[CC_M]; + + initpower_coeff[TC_H] = (proc->cores[0]->get_coefficient_tcache_readhits() + + proc->get_coefficient_readcoalescing()); + initpower_coeff[TC_M] = (proc->cores[0]->get_coefficient_tcache_readmisses() + + proc->get_coefficient_readcoalescing()); + effpower_coeff[TC_H] = + initpower_coeff[TC_H] * p->sys.scaling_coefficients[TC_H]; + effpower_coeff[TC_M] = + initpower_coeff[TC_M] * p->sys.scaling_coefficients[TC_M]; + + initpower_coeff[SHRD_ACC] = + proc->cores[0]->get_coefficient_sharedmemory_readhits(); + effpower_coeff[SHRD_ACC] = + initpower_coeff[SHRD_ACC] * p->sys.scaling_coefficients[SHRD_ACC]; + + initpower_coeff[DC_RH] = (proc->cores[0]->get_coefficient_dcache_readhits() + + proc->get_coefficient_readcoalescing()); + initpower_coeff[DC_RM] = + (proc->cores[0]->get_coefficient_dcache_readmisses() + + proc->get_coefficient_readcoalescing()); + initpower_coeff[DC_WH] = (proc->cores[0]->get_coefficient_dcache_writehits() + + proc->get_coefficient_writecoalescing()); + initpower_coeff[DC_WM] = + (proc->cores[0]->get_coefficient_dcache_writemisses() + + proc->get_coefficient_writecoalescing()); + effpower_coeff[DC_RH] = + initpower_coeff[DC_RH] * p->sys.scaling_coefficients[DC_RH]; + effpower_coeff[DC_RM] = + initpower_coeff[DC_RM] * p->sys.scaling_coefficients[DC_RM]; + effpower_coeff[DC_WH] = + initpower_coeff[DC_WH] * p->sys.scaling_coefficients[DC_WH]; + effpower_coeff[DC_WM] = + initpower_coeff[DC_WM] * p->sys.scaling_coefficients[DC_WM]; + + initpower_coeff[L2_RH] = proc->get_coefficient_l2_read_hits(); + initpower_coeff[L2_RM] = proc->get_coefficient_l2_read_misses(); + initpower_coeff[L2_WH] = proc->get_coefficient_l2_write_hits(); + initpower_coeff[L2_WM] = proc->get_coefficient_l2_write_misses(); + effpower_coeff[L2_RH] = + initpower_coeff[L2_RH] * p->sys.scaling_coefficients[L2_RH]; + effpower_coeff[L2_RM] = + initpower_coeff[L2_RM] * p->sys.scaling_coefficients[L2_RM]; + effpower_coeff[L2_WH] = + initpower_coeff[L2_WH] * p->sys.scaling_coefficients[L2_WH]; + effpower_coeff[L2_WM] = + initpower_coeff[L2_WM] * p->sys.scaling_coefficients[L2_WM]; + + initpower_coeff[IDLE_CORE_N] = + p->sys.idle_core_power * proc->cores[0]->executionTime; + effpower_coeff[IDLE_CORE_N] = + initpower_coeff[IDLE_CORE_N] * p->sys.scaling_coefficients[IDLE_CORE_N]; + + initpower_coeff[PIPE_A] = proc->cores[0]->get_coefficient_duty_cycle(); + effpower_coeff[PIPE_A] = + initpower_coeff[PIPE_A] * p->sys.scaling_coefficients[PIPE_A]; + + initpower_coeff[MEM_RD] = proc->get_coefficient_mem_reads(); + initpower_coeff[MEM_WR] = proc->get_coefficient_mem_writes(); + initpower_coeff[MEM_PRE] = proc->get_coefficient_mem_pre(); + effpower_coeff[MEM_RD] = + initpower_coeff[MEM_RD] * p->sys.scaling_coefficients[MEM_RD]; + effpower_coeff[MEM_WR] = + initpower_coeff[MEM_WR] * p->sys.scaling_coefficients[MEM_WR]; + effpower_coeff[MEM_PRE] = + initpower_coeff[MEM_PRE] * p->sys.scaling_coefficients[MEM_PRE]; + double fp_coeff = proc->cores[0]->get_coefficient_fpu_accesses(); double sfu_coeff = proc->cores[0]->get_coefficient_sfu_accesses(); - initpower_coeff[INT_ACC]= proc->cores[0]->get_coefficient_ialu_accesses()*(proc->cores[0]->exu->rf_fu_clockRate/proc->cores[0]->exu->clockRate); + initpower_coeff[INT_ACC] = + proc->cores[0]->get_coefficient_ialu_accesses() * + (proc->cores[0]->exu->rf_fu_clockRate / proc->cores[0]->exu->clockRate); + + if (tot_fpu_accesses != 0) { + initpower_coeff[FP_ACC] = + fp_coeff * sample_perf_counters[FP_ACC] / tot_fpu_accesses; + initpower_coeff[DP_ACC] = + fp_coeff * sample_perf_counters[DP_ACC] / tot_fpu_accesses; + } else { + initpower_coeff[FP_ACC] = 0; + initpower_coeff[DP_ACC] = 0; + } - if(tot_fpu_accesses != 0){ - initpower_coeff[FP_ACC]= fp_coeff * sample_perf_counters[FP_ACC]/tot_fpu_accesses; - initpower_coeff[DP_ACC]= fp_coeff * sample_perf_counters[DP_ACC]/tot_fpu_accesses; + if (tot_sfu_accesses != 0) { + initpower_coeff[INT_MUL24_ACC] = + sfu_coeff * sample_perf_counters[INT_MUL24_ACC] / tot_sfu_accesses; + initpower_coeff[INT_MUL32_ACC] = + sfu_coeff * sample_perf_counters[INT_MUL32_ACC] / tot_sfu_accesses; + initpower_coeff[INT_MUL_ACC] = + sfu_coeff * sample_perf_counters[INT_MUL_ACC] / tot_sfu_accesses; + initpower_coeff[INT_DIV_ACC] = + sfu_coeff * sample_perf_counters[INT_DIV_ACC] / tot_sfu_accesses; + initpower_coeff[DP_MUL_ACC] = + sfu_coeff * sample_perf_counters[DP_MUL_ACC] / tot_sfu_accesses; + initpower_coeff[DP_DIV_ACC] = + sfu_coeff * sample_perf_counters[DP_DIV_ACC] / tot_sfu_accesses; + initpower_coeff[FP_MUL_ACC] = + sfu_coeff * sample_perf_counters[FP_MUL_ACC] / tot_sfu_accesses; + initpower_coeff[FP_DIV_ACC] = + sfu_coeff * sample_perf_counters[FP_DIV_ACC] / tot_sfu_accesses; + initpower_coeff[FP_SQRT_ACC] = + sfu_coeff * sample_perf_counters[FP_SQRT_ACC] / tot_sfu_accesses; + initpower_coeff[FP_LG_ACC] = + sfu_coeff * sample_perf_counters[FP_LG_ACC] / tot_sfu_accesses; + initpower_coeff[FP_SIN_ACC] = + sfu_coeff * sample_perf_counters[FP_SIN_ACC] / tot_sfu_accesses; + initpower_coeff[FP_EXP_ACC] = + sfu_coeff * sample_perf_counters[FP_EXP_ACC] / tot_sfu_accesses; + initpower_coeff[TENSOR_ACC] = + sfu_coeff * sample_perf_counters[TENSOR_ACC] / tot_sfu_accesses; + initpower_coeff[TEX_ACC] = + sfu_coeff * sample_perf_counters[TEX_ACC] / tot_sfu_accesses; + } else { + initpower_coeff[INT_MUL24_ACC] = 0; + initpower_coeff[INT_MUL32_ACC] = 0; + initpower_coeff[INT_MUL_ACC] = 0; + initpower_coeff[INT_DIV_ACC] = 0; + initpower_coeff[DP_MUL_ACC] = 0; + initpower_coeff[DP_DIV_ACC] = 0; + initpower_coeff[FP_MUL_ACC] = 0; + initpower_coeff[FP_DIV_ACC] = 0; + initpower_coeff[FP_SQRT_ACC] = 0; + initpower_coeff[FP_LG_ACC] = 0; + initpower_coeff[FP_SIN_ACC] = 0; + initpower_coeff[FP_EXP_ACC] = 0; + initpower_coeff[TENSOR_ACC] = 0; + initpower_coeff[TEX_ACC] = 0; } - else{ - initpower_coeff[FP_ACC]= 0; - initpower_coeff[DP_ACC]= 0; + + effpower_coeff[INT_ACC] = initpower_coeff[INT_ACC]; + effpower_coeff[FP_ACC] = initpower_coeff[FP_ACC]; + effpower_coeff[DP_ACC] = initpower_coeff[DP_ACC]; + effpower_coeff[INT_MUL24_ACC] = initpower_coeff[INT_MUL24_ACC]; + effpower_coeff[INT_MUL32_ACC] = initpower_coeff[INT_MUL32_ACC]; + effpower_coeff[INT_MUL_ACC] = initpower_coeff[INT_MUL_ACC]; + effpower_coeff[INT_DIV_ACC] = initpower_coeff[INT_DIV_ACC]; + effpower_coeff[DP_MUL_ACC] = initpower_coeff[DP_MUL_ACC]; + effpower_coeff[DP_DIV_ACC] = initpower_coeff[DP_DIV_ACC]; + effpower_coeff[FP_MUL_ACC] = initpower_coeff[FP_MUL_ACC]; + effpower_coeff[FP_DIV_ACC] = initpower_coeff[FP_DIV_ACC]; + effpower_coeff[FP_SQRT_ACC] = initpower_coeff[FP_SQRT_ACC]; + effpower_coeff[FP_LG_ACC] = initpower_coeff[FP_LG_ACC]; + effpower_coeff[FP_SIN_ACC] = initpower_coeff[FP_SIN_ACC]; + effpower_coeff[FP_EXP_ACC] = initpower_coeff[FP_EXP_ACC]; + effpower_coeff[TENSOR_ACC] = initpower_coeff[TENSOR_ACC]; + effpower_coeff[TEX_ACC] = initpower_coeff[TEX_ACC]; + + initpower_coeff[NOC_A] = proc->get_coefficient_noc_accesses(); + effpower_coeff[NOC_A] = + initpower_coeff[NOC_A] * p->sys.scaling_coefficients[NOC_A]; + + // const_dynamic_power=proc->get_const_dynamic_power()/(proc->cores[0]->executionTime); + + for (unsigned i = 0; i < num_perf_counters; i++) { + initpower_coeff[i] /= (proc->cores[0]->executionTime); + effpower_coeff[i] /= (proc->cores[0]->executionTime); } +} - if(tot_sfu_accesses != 0){ - initpower_coeff[INT_MUL24_ACC]= sfu_coeff * sample_perf_counters[INT_MUL24_ACC]/tot_sfu_accesses; - initpower_coeff[INT_MUL32_ACC]= sfu_coeff * sample_perf_counters[INT_MUL32_ACC]/tot_sfu_accesses; - initpower_coeff[INT_MUL_ACC]= sfu_coeff * sample_perf_counters[INT_MUL_ACC]/tot_sfu_accesses; - initpower_coeff[INT_DIV_ACC]= sfu_coeff * sample_perf_counters[INT_DIV_ACC]/tot_sfu_accesses; - initpower_coeff[DP_MUL_ACC]= sfu_coeff * sample_perf_counters[DP_MUL_ACC]/tot_sfu_accesses; - initpower_coeff[DP_DIV_ACC]= sfu_coeff * sample_perf_counters[DP_DIV_ACC]/tot_sfu_accesses; - initpower_coeff[FP_MUL_ACC]= sfu_coeff * sample_perf_counters[FP_MUL_ACC]/tot_sfu_accesses; - initpower_coeff[FP_DIV_ACC]= sfu_coeff * sample_perf_counters[FP_DIV_ACC]/tot_sfu_accesses; - initpower_coeff[FP_SQRT_ACC]= sfu_coeff * sample_perf_counters[FP_SQRT_ACC]/tot_sfu_accesses; - initpower_coeff[FP_LG_ACC]= sfu_coeff * sample_perf_counters[FP_LG_ACC]/tot_sfu_accesses; - initpower_coeff[FP_SIN_ACC]= sfu_coeff * sample_perf_counters[FP_SIN_ACC]/tot_sfu_accesses; - initpower_coeff[FP_EXP_ACC]= sfu_coeff * sample_perf_counters[FP_EXP_ACC]/tot_sfu_accesses; - initpower_coeff[TENSOR_ACC]= sfu_coeff * sample_perf_counters[TENSOR_ACC]/tot_sfu_accesses; - initpower_coeff[TEX_ACC]= sfu_coeff * sample_perf_counters[TEX_ACC]/tot_sfu_accesses; +double gpgpu_sim_wrapper::calculate_static_power() { + double int_accesses = + initpower_coeff[INT_ACC] + initpower_coeff[INT_MUL24_ACC] + + initpower_coeff[INT_MUL32_ACC] + initpower_coeff[INT_MUL_ACC] + + initpower_coeff[INT_DIV_ACC]; + double int_add_accesses = initpower_coeff[INT_ACC]; + double int_mul_accesses = + initpower_coeff[INT_MUL24_ACC] + initpower_coeff[INT_MUL32_ACC] + + initpower_coeff[INT_MUL_ACC] + initpower_coeff[INT_DIV_ACC]; + double fp_accesses = initpower_coeff[FP_ACC] + initpower_coeff[FP_MUL_ACC] + + initpower_coeff[FP_DIV_ACC]; + double dp_accesses = initpower_coeff[DP_ACC] + initpower_coeff[DP_MUL_ACC] + + initpower_coeff[DP_DIV_ACC]; + double sfu_accesses = + initpower_coeff[FP_SQRT_ACC] + initpower_coeff[FP_LG_ACC] + + initpower_coeff[FP_SIN_ACC] + initpower_coeff[FP_EXP_ACC]; + double tensor_accesses = initpower_coeff[TENSOR_ACC]; + double tex_accesses = initpower_coeff[TEX_ACC]; + double total_static_power = 0.0; + double base_static_power = 0.0; + double lane_static_power = 0.0; + double per_active_core = (num_cores - num_idle_cores) / num_cores; + + double l1_accesses = initpower_coeff[DC_RH] + initpower_coeff[DC_RM] + + initpower_coeff[DC_WH] + initpower_coeff[DC_WM]; + double l2_accesses = initpower_coeff[L2_RH] + initpower_coeff[L2_RM] + + initpower_coeff[L2_WH] + initpower_coeff[L2_WM]; + double shared_accesses = initpower_coeff[SHRD_ACC]; + + if (avg_threads_per_warp == + 0) { // no functional unit threads, check for memory or a 'LIGHT_SM' + if (l1_accesses != 0.0) + return (p->sys.static_l1_flane * per_active_core); + else if (shared_accesses != 0.0) + return (p->sys.static_shared_flane * per_active_core); + else if (l2_accesses != 0.0) + return (p->sys.static_l2_flane * per_active_core); + else // LIGHT_SM + return (p->sys.static_light_flane * + per_active_core); // return LIGHT_SM base static power } - else{ - initpower_coeff[INT_MUL24_ACC]= 0; - initpower_coeff[INT_MUL32_ACC]= 0; - initpower_coeff[INT_MUL_ACC]= 0; - initpower_coeff[INT_DIV_ACC]= 0; - initpower_coeff[DP_MUL_ACC]= 0; - initpower_coeff[DP_DIV_ACC]= 0; - initpower_coeff[FP_MUL_ACC]= 0; - initpower_coeff[FP_DIV_ACC]= 0; - initpower_coeff[FP_SQRT_ACC]= 0; - initpower_coeff[FP_LG_ACC]= 0; - initpower_coeff[FP_SIN_ACC]= 0; - initpower_coeff[FP_EXP_ACC]= 0; - initpower_coeff[TENSOR_ACC]= 0; - initpower_coeff[TEX_ACC]= 0; + + /* using a linear model for thread divergence */ + if ((int_accesses != 0.0) && (fp_accesses != 0.0) && (dp_accesses != 0.0) && + (sfu_accesses == 0.0) && (tensor_accesses == 0.0) && + (tex_accesses == 0.0)) { + /* INT_FP_DP */ + base_static_power = p->sys.static_cat3_flane; + lane_static_power = p->sys.static_cat3_addlane; } - effpower_coeff[INT_ACC]= initpower_coeff[INT_ACC]; - effpower_coeff[FP_ACC]= initpower_coeff[FP_ACC]; - effpower_coeff[DP_ACC]= initpower_coeff[DP_ACC]; - effpower_coeff[INT_MUL24_ACC]= initpower_coeff[INT_MUL24_ACC]; - effpower_coeff[INT_MUL32_ACC]= initpower_coeff[INT_MUL32_ACC]; - effpower_coeff[INT_MUL_ACC]= initpower_coeff[INT_MUL_ACC]; - effpower_coeff[INT_DIV_ACC]= initpower_coeff[INT_DIV_ACC]; - effpower_coeff[DP_MUL_ACC]= initpower_coeff[DP_MUL_ACC]; - effpower_coeff[DP_DIV_ACC]= initpower_coeff[DP_DIV_ACC]; - effpower_coeff[FP_MUL_ACC]= initpower_coeff[FP_MUL_ACC]; - effpower_coeff[FP_DIV_ACC]= initpower_coeff[FP_DIV_ACC]; - effpower_coeff[FP_SQRT_ACC]= initpower_coeff[FP_SQRT_ACC]; - effpower_coeff[FP_LG_ACC]= initpower_coeff[FP_LG_ACC]; - effpower_coeff[FP_SIN_ACC]= initpower_coeff[FP_SIN_ACC]; - effpower_coeff[FP_EXP_ACC]= initpower_coeff[FP_EXP_ACC]; - effpower_coeff[TENSOR_ACC]= initpower_coeff[TENSOR_ACC]; - effpower_coeff[TEX_ACC]= initpower_coeff[TEX_ACC]; - - initpower_coeff[NOC_A]=proc->get_coefficient_noc_accesses(); - effpower_coeff[NOC_A]=initpower_coeff[NOC_A]*p->sys.scaling_coefficients[NOC_A]; - - //const_dynamic_power=proc->get_const_dynamic_power()/(proc->cores[0]->executionTime); - - for(unsigned i=0; icores[0]->executionTime); - effpower_coeff[i]/=(proc->cores[0]->executionTime); + else if ((int_accesses != 0.0) && (fp_accesses != 0.0) && + (dp_accesses == 0.0) && (sfu_accesses == 0.0) && + (tensor_accesses != 0.0) && (tex_accesses == 0.0)) { + /* INT_FP_TENSOR */ + base_static_power = p->sys.static_cat6_flane; + lane_static_power = p->sys.static_cat6_addlane; } -} -double gpgpu_sim_wrapper::calculate_static_power(){ - double int_accesses = initpower_coeff[INT_ACC] + initpower_coeff[INT_MUL24_ACC] + initpower_coeff[INT_MUL32_ACC] + initpower_coeff[INT_MUL_ACC] + initpower_coeff[INT_DIV_ACC]; - double int_add_accesses = initpower_coeff[INT_ACC]; - double int_mul_accesses = initpower_coeff[INT_MUL24_ACC] + initpower_coeff[INT_MUL32_ACC] + initpower_coeff[INT_MUL_ACC] + initpower_coeff[INT_DIV_ACC]; - double fp_accesses = initpower_coeff[FP_ACC] + initpower_coeff[FP_MUL_ACC] + initpower_coeff[FP_DIV_ACC]; - double dp_accesses = initpower_coeff[DP_ACC] + initpower_coeff[DP_MUL_ACC] + initpower_coeff[DP_DIV_ACC]; - double sfu_accesses = initpower_coeff[FP_SQRT_ACC] + initpower_coeff[FP_LG_ACC] + initpower_coeff[FP_SIN_ACC] + initpower_coeff[FP_EXP_ACC]; - double tensor_accesses = initpower_coeff[TENSOR_ACC]; - double tex_accesses = initpower_coeff[TEX_ACC]; - double total_static_power = 0.0; - double base_static_power = 0.0; - double lane_static_power = 0.0; - double per_active_core = (num_cores - num_idle_cores)/num_cores; - - - double l1_accesses = initpower_coeff[DC_RH] + initpower_coeff[DC_RM] + initpower_coeff[DC_WH] + initpower_coeff[DC_WM]; - double l2_accesses = initpower_coeff[L2_RH] + initpower_coeff[L2_RM] + initpower_coeff[L2_WH] + initpower_coeff[L2_WM]; - double shared_accesses = initpower_coeff[SHRD_ACC]; - - - if(avg_threads_per_warp == 0){ //no functional unit threads, check for memory or a 'LIGHT_SM' - if(l1_accesses != 0.0) - return (p->sys.static_l1_flane*per_active_core); - else if(shared_accesses != 0.0) - return (p->sys.static_shared_flane*per_active_core); - else if(l2_accesses != 0.0) - return (p->sys.static_l2_flane*per_active_core); - else //LIGHT_SM - return (p->sys.static_light_flane*per_active_core); //return LIGHT_SM base static power - } - - /* using a linear model for thread divergence */ - if((int_accesses != 0.0) && (fp_accesses != 0.0) && (dp_accesses != 0.0) && (sfu_accesses == 0.0) && (tensor_accesses == 0.0) && (tex_accesses == 0.0)){ - /* INT_FP_DP */ - base_static_power = p->sys.static_cat3_flane; - lane_static_power = p->sys.static_cat3_addlane; - } - - else if((int_accesses != 0.0) && (fp_accesses != 0.0) && (dp_accesses == 0.0) && (sfu_accesses == 0.0) && (tensor_accesses != 0.0) && (tex_accesses == 0.0)){ - /* INT_FP_TENSOR */ - base_static_power = p->sys.static_cat6_flane; - lane_static_power = p->sys.static_cat6_addlane; - } - - else if((int_accesses != 0.0) && (fp_accesses != 0.0) && (dp_accesses == 0.0) && (sfu_accesses != 0.0) && (tensor_accesses == 0.0) && (tex_accesses == 0.0)){ - /* INT_FP_SFU */ - base_static_power = p->sys.static_cat4_flane; - lane_static_power = p->sys.static_cat4_addlane; - } - - else if((int_accesses != 0.0) && (fp_accesses != 0.0) && (dp_accesses == 0.0) && (sfu_accesses == 0.0) && (tensor_accesses == 0.0) && (tex_accesses != 0.0)){ - /* INT_FP_TEX */ - base_static_power = p->sys.static_cat5_flane; - lane_static_power = p->sys.static_cat5_addlane; - } - - else if((int_accesses != 0.0) && (fp_accesses != 0.0) && (dp_accesses == 0.0) && (sfu_accesses == 0.0) && (tensor_accesses == 0.0) && (tex_accesses == 0.0)){ - /* INT_FP */ - base_static_power = p->sys.static_cat2_flane; - lane_static_power = p->sys.static_cat2_addlane; - } - - else if((int_accesses != 0.0) && (fp_accesses == 0.0) && (dp_accesses == 0.0) && (sfu_accesses == 0.0) && (tensor_accesses == 0.0) && (tex_accesses == 0.0)){ - /* INT */ - /* Seperating INT_ADD only and INT_MUL only from mix of INT instructions */ - if((int_add_accesses != 0.0) && (int_mul_accesses == 0.0)){ //INT_ADD - base_static_power = p->sys.static_intadd_flane; - lane_static_power = p->sys.static_intadd_addlane; - } - else if((int_add_accesses == 0.0) && (int_mul_accesses != 0.0)){ //INT_MUL - base_static_power = p->sys.static_intmul_flane; - lane_static_power = p->sys.static_intmul_addlane; - } - else{ //INT_ADD+MUL - base_static_power = p->sys.static_cat1_flane; - lane_static_power = p->sys.static_cat1_addlane; - } - } - - else if((int_accesses == 0.0) && (fp_accesses == 0.0) && (dp_accesses == 0.0) && (sfu_accesses == 0.0) && (tensor_accesses == 0.0) && (tex_accesses == 0.0)){ - /* LIGHT_SM or memory only sample */ - lane_static_power = 0.0; //addlane static power is 0 for l1/l2/shared memory only accesses - if(l1_accesses != 0.0) - base_static_power = p->sys.static_l1_flane; - else if(shared_accesses != 0.0) - base_static_power = p->sys.static_shared_flane; - else if(l2_accesses != 0.0) - base_static_power = p->sys.static_l2_flane; - else{ - base_static_power = p->sys.static_light_flane; - lane_static_power = p->sys.static_light_addlane; - } - } - else{ - base_static_power = p->sys.static_geomean_flane; //GEOMEAN except LIGHT_SM if we don't fall into any of the categories above - lane_static_power = p->sys.static_geomean_addlane; - } - - total_static_power = base_static_power + (((double)avg_threads_per_warp-1.0)*lane_static_power); //Linear Model - return (total_static_power*per_active_core); -} + else if ((int_accesses != 0.0) && (fp_accesses != 0.0) && + (dp_accesses == 0.0) && (sfu_accesses != 0.0) && + (tensor_accesses == 0.0) && (tex_accesses == 0.0)) { + /* INT_FP_SFU */ + base_static_power = p->sys.static_cat4_flane; + lane_static_power = p->sys.static_cat4_addlane; + } -void gpgpu_sim_wrapper::update_components_power() -{ + else if ((int_accesses != 0.0) && (fp_accesses != 0.0) && + (dp_accesses == 0.0) && (sfu_accesses == 0.0) && + (tensor_accesses == 0.0) && (tex_accesses != 0.0)) { + /* INT_FP_TEX */ + base_static_power = p->sys.static_cat5_flane; + lane_static_power = p->sys.static_cat5_addlane; + } + else if ((int_accesses != 0.0) && (fp_accesses != 0.0) && + (dp_accesses == 0.0) && (sfu_accesses == 0.0) && + (tensor_accesses == 0.0) && (tex_accesses == 0.0)) { + /* INT_FP */ + base_static_power = p->sys.static_cat2_flane; + lane_static_power = p->sys.static_cat2_addlane; + } + + else if ((int_accesses != 0.0) && (fp_accesses == 0.0) && + (dp_accesses == 0.0) && (sfu_accesses == 0.0) && + (tensor_accesses == 0.0) && (tex_accesses == 0.0)) { + /* INT */ + /* Seperating INT_ADD only and INT_MUL only from mix of INT instructions */ + if ((int_add_accesses != 0.0) && (int_mul_accesses == 0.0)) { // INT_ADD + base_static_power = p->sys.static_intadd_flane; + lane_static_power = p->sys.static_intadd_addlane; + } else if ((int_add_accesses == 0.0) && + (int_mul_accesses != 0.0)) { // INT_MUL + base_static_power = p->sys.static_intmul_flane; + lane_static_power = p->sys.static_intmul_addlane; + } else { // INT_ADD+MUL + base_static_power = p->sys.static_cat1_flane; + lane_static_power = p->sys.static_cat1_addlane; + } + } + + else if ((int_accesses == 0.0) && (fp_accesses == 0.0) && + (dp_accesses == 0.0) && (sfu_accesses == 0.0) && + (tensor_accesses == 0.0) && (tex_accesses == 0.0)) { + /* LIGHT_SM or memory only sample */ + lane_static_power = + 0.0; // addlane static power is 0 for l1/l2/shared memory only accesses + if (l1_accesses != 0.0) + base_static_power = p->sys.static_l1_flane; + else if (shared_accesses != 0.0) + base_static_power = p->sys.static_shared_flane; + else if (l2_accesses != 0.0) + base_static_power = p->sys.static_l2_flane; + else { + base_static_power = p->sys.static_light_flane; + lane_static_power = p->sys.static_light_addlane; + } + } else { + base_static_power = + p->sys.static_geomean_flane; // GEOMEAN except LIGHT_SM if we don't + // fall into any of the categories above + lane_static_power = p->sys.static_geomean_addlane; + } + + total_static_power = + base_static_power + (((double)avg_threads_per_warp - 1.0) * + lane_static_power); // Linear Model + return (total_static_power * per_active_core); +} + +void gpgpu_sim_wrapper::update_components_power() { update_coefficients(); - proc_power=proc->rt_power.readOp.dynamic; - sample_cmp_pwr[IBP]=(proc->cores[0]->ifu->IB->rt_power.readOp.dynamic - +proc->cores[0]->ifu->IB->rt_power.writeOp.dynamic - +proc->cores[0]->ifu->ID_misc->rt_power.readOp.dynamic - +proc->cores[0]->ifu->ID_operand->rt_power.readOp.dynamic - +proc->cores[0]->ifu->ID_inst->rt_power.readOp.dynamic)/(proc->cores[0]->executionTime); + proc_power = proc->rt_power.readOp.dynamic; + sample_cmp_pwr[IBP] = + (proc->cores[0]->ifu->IB->rt_power.readOp.dynamic + + proc->cores[0]->ifu->IB->rt_power.writeOp.dynamic + + proc->cores[0]->ifu->ID_misc->rt_power.readOp.dynamic + + proc->cores[0]->ifu->ID_operand->rt_power.readOp.dynamic + + proc->cores[0]->ifu->ID_inst->rt_power.readOp.dynamic) / + (proc->cores[0]->executionTime); - sample_cmp_pwr[ICP]=proc->cores[0]->ifu->icache.rt_power.readOp.dynamic/(proc->cores[0]->executionTime); + sample_cmp_pwr[ICP] = proc->cores[0]->ifu->icache.rt_power.readOp.dynamic / + (proc->cores[0]->executionTime); - sample_cmp_pwr[DCP]=proc->cores[0]->lsu->dcache.rt_power.readOp.dynamic/(proc->cores[0]->executionTime); + sample_cmp_pwr[DCP] = proc->cores[0]->lsu->dcache.rt_power.readOp.dynamic / + (proc->cores[0]->executionTime); - sample_cmp_pwr[TCP]=proc->cores[0]->lsu->tcache.rt_power.readOp.dynamic/(proc->cores[0]->executionTime); + sample_cmp_pwr[TCP] = proc->cores[0]->lsu->tcache.rt_power.readOp.dynamic / + (proc->cores[0]->executionTime); - sample_cmp_pwr[CCP]=proc->cores[0]->lsu->ccache.rt_power.readOp.dynamic/(proc->cores[0]->executionTime); + sample_cmp_pwr[CCP] = proc->cores[0]->lsu->ccache.rt_power.readOp.dynamic / + (proc->cores[0]->executionTime); - sample_cmp_pwr[SHRDP]=proc->cores[0]->lsu->sharedmemory.rt_power.readOp.dynamic/(proc->cores[0]->executionTime); + sample_cmp_pwr[SHRDP] = + proc->cores[0]->lsu->sharedmemory.rt_power.readOp.dynamic / + (proc->cores[0]->executionTime); - sample_cmp_pwr[RFP]=(proc->cores[0]->exu->rfu->rt_power.readOp.dynamic/(proc->cores[0]->executionTime)) - *(proc->cores[0]->exu->rf_fu_clockRate/proc->cores[0]->exu->clockRate); + sample_cmp_pwr[RFP] = + (proc->cores[0]->exu->rfu->rt_power.readOp.dynamic / + (proc->cores[0]->executionTime)) * + (proc->cores[0]->exu->rf_fu_clockRate / proc->cores[0]->exu->clockRate); - double sample_fp_pwr = (proc->cores[0]->exu->fp_u->rt_power.readOp.dynamic/(proc->cores[0]->executionTime)); + double sample_fp_pwr = (proc->cores[0]->exu->fp_u->rt_power.readOp.dynamic / + (proc->cores[0]->executionTime)); - double sample_sfu_pwr = (proc->cores[0]->exu->mul->rt_power.readOp.dynamic/(proc->cores[0]->executionTime)); + double sample_sfu_pwr = (proc->cores[0]->exu->mul->rt_power.readOp.dynamic / + (proc->cores[0]->executionTime)); - sample_cmp_pwr[INTP]=(proc->cores[0]->exu->exeu->rt_power.readOp.dynamic/(proc->cores[0]->executionTime)) - *(proc->cores[0]->exu->rf_fu_clockRate/proc->cores[0]->exu->clockRate); + sample_cmp_pwr[INTP] = + (proc->cores[0]->exu->exeu->rt_power.readOp.dynamic / + (proc->cores[0]->executionTime)) * + (proc->cores[0]->exu->rf_fu_clockRate / proc->cores[0]->exu->clockRate); - - if(tot_fpu_accesses != 0){ - sample_cmp_pwr[FPUP]= sample_fp_pwr * sample_perf_counters[FP_ACC]/tot_fpu_accesses; - sample_cmp_pwr[DPUP]= sample_fp_pwr * sample_perf_counters[DP_ACC]/tot_fpu_accesses; - } - else{ - sample_cmp_pwr[FPUP]= 0; - sample_cmp_pwr[DPUP]= 0; - } - if(tot_sfu_accesses != 0){ - sample_cmp_pwr[INT_MUL24P]= sample_sfu_pwr * sample_perf_counters[INT_MUL24_ACC]/tot_sfu_accesses; - sample_cmp_pwr[INT_MUL32P]= sample_sfu_pwr * sample_perf_counters[INT_MUL32_ACC]/tot_sfu_accesses; - sample_cmp_pwr[INT_MULP]= sample_sfu_pwr * sample_perf_counters[INT_MUL_ACC]/tot_sfu_accesses; - sample_cmp_pwr[INT_DIVP]= sample_sfu_pwr * sample_perf_counters[INT_DIV_ACC]/tot_sfu_accesses; - sample_cmp_pwr[FP_MULP]= sample_sfu_pwr * sample_perf_counters[FP_MUL_ACC]/tot_sfu_accesses; - sample_cmp_pwr[FP_DIVP]= sample_sfu_pwr * sample_perf_counters[FP_DIV_ACC]/tot_sfu_accesses; - sample_cmp_pwr[FP_SQRTP]= sample_sfu_pwr * sample_perf_counters[FP_SQRT_ACC]/tot_sfu_accesses; - sample_cmp_pwr[FP_LGP]= sample_sfu_pwr * sample_perf_counters[FP_LG_ACC]/tot_sfu_accesses; - sample_cmp_pwr[FP_SINP]= sample_sfu_pwr * sample_perf_counters[FP_SIN_ACC]/tot_sfu_accesses; - sample_cmp_pwr[FP_EXP]= sample_sfu_pwr * sample_perf_counters[FP_EXP_ACC]/tot_sfu_accesses; - sample_cmp_pwr[DP_MULP]= sample_sfu_pwr * sample_perf_counters[DP_MUL_ACC]/tot_sfu_accesses; - sample_cmp_pwr[DP_DIVP]= sample_sfu_pwr * sample_perf_counters[DP_DIV_ACC]/tot_sfu_accesses; - sample_cmp_pwr[TENSORP]= sample_sfu_pwr * sample_perf_counters[TENSOR_ACC]/tot_sfu_accesses; - sample_cmp_pwr[TEXP]= sample_sfu_pwr * sample_perf_counters[TEX_ACC]/tot_sfu_accesses; + if (tot_fpu_accesses != 0) { + sample_cmp_pwr[FPUP] = + sample_fp_pwr * sample_perf_counters[FP_ACC] / tot_fpu_accesses; + sample_cmp_pwr[DPUP] = + sample_fp_pwr * sample_perf_counters[DP_ACC] / tot_fpu_accesses; + } else { + sample_cmp_pwr[FPUP] = 0; + sample_cmp_pwr[DPUP] = 0; } - else{ - sample_cmp_pwr[INT_MUL24P]= 0; - sample_cmp_pwr[INT_MUL32P]= 0; - sample_cmp_pwr[INT_MULP]= 0; - sample_cmp_pwr[INT_DIVP]= 0; - sample_cmp_pwr[FP_MULP]= 0; - sample_cmp_pwr[FP_DIVP]= 0; - sample_cmp_pwr[FP_SQRTP]= 0; - sample_cmp_pwr[FP_LGP]= 0; - sample_cmp_pwr[FP_SINP]= 0; - sample_cmp_pwr[FP_EXP]= 0; - sample_cmp_pwr[DP_MULP]= 0; - sample_cmp_pwr[DP_DIVP]= 0; - sample_cmp_pwr[TENSORP]= 0; - sample_cmp_pwr[TEXP]= 0; + if (tot_sfu_accesses != 0) { + sample_cmp_pwr[INT_MUL24P] = + sample_sfu_pwr * sample_perf_counters[INT_MUL24_ACC] / tot_sfu_accesses; + sample_cmp_pwr[INT_MUL32P] = + sample_sfu_pwr * sample_perf_counters[INT_MUL32_ACC] / tot_sfu_accesses; + sample_cmp_pwr[INT_MULP] = + sample_sfu_pwr * sample_perf_counters[INT_MUL_ACC] / tot_sfu_accesses; + sample_cmp_pwr[INT_DIVP] = + sample_sfu_pwr * sample_perf_counters[INT_DIV_ACC] / tot_sfu_accesses; + sample_cmp_pwr[FP_MULP] = + sample_sfu_pwr * sample_perf_counters[FP_MUL_ACC] / tot_sfu_accesses; + sample_cmp_pwr[FP_DIVP] = + sample_sfu_pwr * sample_perf_counters[FP_DIV_ACC] / tot_sfu_accesses; + sample_cmp_pwr[FP_SQRTP] = + sample_sfu_pwr * sample_perf_counters[FP_SQRT_ACC] / tot_sfu_accesses; + sample_cmp_pwr[FP_LGP] = + sample_sfu_pwr * sample_perf_counters[FP_LG_ACC] / tot_sfu_accesses; + sample_cmp_pwr[FP_SINP] = + sample_sfu_pwr * sample_perf_counters[FP_SIN_ACC] / tot_sfu_accesses; + sample_cmp_pwr[FP_EXP] = + sample_sfu_pwr * sample_perf_counters[FP_EXP_ACC] / tot_sfu_accesses; + sample_cmp_pwr[DP_MULP] = + sample_sfu_pwr * sample_perf_counters[DP_MUL_ACC] / tot_sfu_accesses; + sample_cmp_pwr[DP_DIVP] = + sample_sfu_pwr * sample_perf_counters[DP_DIV_ACC] / tot_sfu_accesses; + sample_cmp_pwr[TENSORP] = + sample_sfu_pwr * sample_perf_counters[TENSOR_ACC] / tot_sfu_accesses; + sample_cmp_pwr[TEXP] = + sample_sfu_pwr * sample_perf_counters[TEX_ACC] / tot_sfu_accesses; + } else { + sample_cmp_pwr[INT_MUL24P] = 0; + sample_cmp_pwr[INT_MUL32P] = 0; + sample_cmp_pwr[INT_MULP] = 0; + sample_cmp_pwr[INT_DIVP] = 0; + sample_cmp_pwr[FP_MULP] = 0; + sample_cmp_pwr[FP_DIVP] = 0; + sample_cmp_pwr[FP_SQRTP] = 0; + sample_cmp_pwr[FP_LGP] = 0; + sample_cmp_pwr[FP_SINP] = 0; + sample_cmp_pwr[FP_EXP] = 0; + sample_cmp_pwr[DP_MULP] = 0; + sample_cmp_pwr[DP_DIVP] = 0; + sample_cmp_pwr[TENSORP] = 0; + sample_cmp_pwr[TEXP] = 0; } - sample_cmp_pwr[SCHEDP]=proc->cores[0]->exu->scheu->rt_power.readOp.dynamic/(proc->cores[0]->executionTime); + sample_cmp_pwr[SCHEDP] = proc->cores[0]->exu->scheu->rt_power.readOp.dynamic / + (proc->cores[0]->executionTime); - sample_cmp_pwr[L2CP]=(proc->XML->sys.number_of_L2s>0)? proc->l2array[0]->rt_power.readOp.dynamic/(proc->cores[0]->executionTime):0; + sample_cmp_pwr[L2CP] = (proc->XML->sys.number_of_L2s > 0) + ? proc->l2array[0]->rt_power.readOp.dynamic / + (proc->cores[0]->executionTime) + : 0; - sample_cmp_pwr[MCP]=(proc->mc->rt_power.readOp.dynamic-proc->mc->dram->rt_power.readOp.dynamic)/(proc->cores[0]->executionTime); + sample_cmp_pwr[MCP] = (proc->mc->rt_power.readOp.dynamic - + proc->mc->dram->rt_power.readOp.dynamic) / + (proc->cores[0]->executionTime); - sample_cmp_pwr[NOCP]=proc->nocs[0]->rt_power.readOp.dynamic/(proc->cores[0]->executionTime); + sample_cmp_pwr[NOCP] = + proc->nocs[0]->rt_power.readOp.dynamic / (proc->cores[0]->executionTime); - sample_cmp_pwr[DRAMP]=proc->mc->dram->rt_power.readOp.dynamic/(proc->cores[0]->executionTime); + sample_cmp_pwr[DRAMP] = + proc->mc->dram->rt_power.readOp.dynamic / (proc->cores[0]->executionTime); - sample_cmp_pwr[PIPEP]=proc->cores[0]->Pipeline_energy/(proc->cores[0]->executionTime); + sample_cmp_pwr[PIPEP] = + proc->cores[0]->Pipeline_energy / (proc->cores[0]->executionTime); - sample_cmp_pwr[IDLE_COREP]=proc->cores[0]->IdleCoreEnergy/(proc->cores[0]->executionTime); + sample_cmp_pwr[IDLE_COREP] = + proc->cores[0]->IdleCoreEnergy / (proc->cores[0]->executionTime); - // This constant dynamic power (e.g., clock power) part is estimated via regression model. - sample_cmp_pwr[CONSTP]=0; - sample_cmp_pwr[STATICP]=0; - // double cnst_dyn = proc->get_const_dynamic_power()/(proc->cores[0]->executionTime); - // // If the regression scaling term is greater than the recorded constant dynamic power - // // then use the difference (other portion already added to dynamic power). Else, + // This constant dynamic power (e.g., clock power) part is estimated via + // regression model. + sample_cmp_pwr[CONSTP] = 0; + sample_cmp_pwr[STATICP] = 0; + // double cnst_dyn = + // proc->get_const_dynamic_power()/(proc->cores[0]->executionTime); + // // If the regression scaling term is greater than the recorded constant + // dynamic power + // // then use the difference (other portion already added to dynamic power). + // Else, // // all the constant dynamic power is accounted for, add nothing. // if(p->sys.scaling_coefficients[constant_power] > cnst_dyn) - // sample_cmp_pwr[CONSTP] = (p->sys.scaling_coefficients[constant_power]-cnst_dyn); + // sample_cmp_pwr[CONSTP] = + // (p->sys.scaling_coefficients[constant_power]-cnst_dyn); sample_cmp_pwr[CONSTP] = p->sys.scaling_coefficients[constant_power]; sample_cmp_pwr[STATICP] = calculate_static_power(); - if(g_dvfs_enabled){ - double voltage_ratio = modeled_chip_voltage/p->sys.modeled_chip_voltage_ref; - sample_cmp_pwr[IDLE_COREP] *= voltage_ratio; // static power scaled by voltage_ratio - sample_cmp_pwr[STATICP] *= voltage_ratio; // static power scaled by voltage_ratio - for(unsigned i=0; isys.modeled_chip_voltage_ref; + sample_cmp_pwr[IDLE_COREP] *= + voltage_ratio; // static power scaled by voltage_ratio + sample_cmp_pwr[STATICP] *= + voltage_ratio; // static power scaled by voltage_ratio + for (unsigned i = 0; i < num_pwr_cmps; i++) { + if ((i != IDLE_COREP) && (i != STATICP)) { + sample_cmp_pwr[i] *= + voltage_ratio * + voltage_ratio; // dynamic power scaled by square of voltage_ratio + } + } } - - proc_power+=sample_cmp_pwr[CONSTP]+sample_cmp_pwr[STATICP]; - if(!g_dvfs_enabled){ // sanity check will fail when voltage scaling is applied, fix later - double sum_pwr_cmp=0; - for(unsigned i=0; ilpXML[pXML->nIndex]; @@ -2193,15 +2191,15 @@ int XMLNode::CreateXMLStringR(XMLNodeData *pEntry, XMLSTR lpszMarker, nResult++; } } else - // If there are child nodes we need to terminate the start tag - if (nElementI) { - if (lpszMarker) lpszMarker[nResult - 1] = _CXML('>'); - if (nFormat >= 0) { - if (lpszMarker) lpszMarker[nResult] = _CXML('\n'); - nResult++; - } - } else - nResult--; + // If there are child nodes we need to terminate the start tag + if (nElementI) { + if (lpszMarker) lpszMarker[nResult - 1] = _CXML('>'); + if (nFormat >= 0) { + if (lpszMarker) lpszMarker[nResult] = _CXML('\n'); + nResult++; + } + } else + nResult--; } // Calculate the child format for when we recurse. This is used to diff --git a/src/cuda-sim/cuda-sim.cc b/src/cuda-sim/cuda-sim.cc index 888cf77..833d33f 100644 --- a/src/cuda-sim/cuda-sim.cc +++ b/src/cuda-sim/cuda-sim.cc @@ -1,18 +1,19 @@ // Copyright (c) 2009-2021, Tor M. Aamodt, Ali Bakhoda, Wilson W.L. Fung, -// George L. Yuan, Jimmy Kwa, Vijay Kandiah, Nikos Hardavellas, +// George L. Yuan, Jimmy Kwa, Vijay Kandiah, Nikos Hardavellas, // Mahmoud Khairy, Junrui Pan, Timothy G. Rogers -// The University of British Columbia, Northwestern University, Purdue University -// All rights reserved. +// The University of British Columbia, Northwestern University, Purdue +// University All rights reserved. // // Redistribution and use in source and binary forms, with or without // modification, are permitted provided that the following conditions are met: // -// 1. Redistributions of source code must retain the above copyright notice, this +// 1. Redistributions of source code must retain the above copyright notice, +// this // list of conditions and the following disclaimer; // 2. Redistributions in binary form must reproduce the above copyright notice, // this list of conditions and the following disclaimer in the documentation // and/or other materials provided with the distribution; -// 3. Neither the names of The University of British Columbia, Northwestern +// 3. Neither the names of The University of British Columbia, Northwestern // University nor the names of their contributors may be used to // endorse or promote products derived from this software without specific // prior written permission. @@ -591,118 +592,118 @@ void ptx_instruction::set_fp_or_int_archop() { } } -void ptx_instruction::set_mul_div_or_other_archop(){ - sp_op=OTHER_OP; - if((m_opcode != MEMBAR_OP) && (m_opcode != SSY_OP) && (m_opcode != BRA_OP) && (m_opcode != BAR_OP) && (m_opcode != EXIT_OP) && (m_opcode != NOP_OP) && (m_opcode != RETP_OP) && (m_opcode != RET_OP) && (m_opcode != CALLP_OP) && (m_opcode != CALL_OP)){ - if(get_type() == F64_TYPE || get_type() == FF64_TYPE){ - switch(get_opcode()){ - case MUL_OP: - case MAD_OP: - case FMA_OP: - sp_op=DP_MUL_OP; - break; - case DIV_OP: - case REM_OP: - sp_op=DP_DIV_OP; - break; - case RCP_OP: - sp_op=DP_DIV_OP; - break; - case LG2_OP: - sp_op=FP_LG_OP; - break; - case RSQRT_OP: - case SQRT_OP: - sp_op=FP_SQRT_OP; - break; - case SIN_OP: - case COS_OP: - sp_op=FP_SIN_OP; - break; - case EX2_OP: - sp_op=FP_EXP_OP; - break; - case MMA_OP: - sp_op=TENSOR__OP; - break; - case TEX_OP: - sp_op=TEX__OP; - break; - default: - if((op==DP_OP) || (op==ALU_OP)) - sp_op=DP___OP; - break; - } +void ptx_instruction::set_mul_div_or_other_archop() { + sp_op = OTHER_OP; + if ((m_opcode != MEMBAR_OP) && (m_opcode != SSY_OP) && (m_opcode != BRA_OP) && + (m_opcode != BAR_OP) && (m_opcode != EXIT_OP) && (m_opcode != NOP_OP) && + (m_opcode != RETP_OP) && (m_opcode != RET_OP) && (m_opcode != CALLP_OP) && + (m_opcode != CALL_OP)) { + if (get_type() == F64_TYPE || get_type() == FF64_TYPE) { + switch (get_opcode()) { + case MUL_OP: + case MAD_OP: + case FMA_OP: + sp_op = DP_MUL_OP; + break; + case DIV_OP: + case REM_OP: + sp_op = DP_DIV_OP; + break; + case RCP_OP: + sp_op = DP_DIV_OP; + break; + case LG2_OP: + sp_op = FP_LG_OP; + break; + case RSQRT_OP: + case SQRT_OP: + sp_op = FP_SQRT_OP; + break; + case SIN_OP: + case COS_OP: + sp_op = FP_SIN_OP; + break; + case EX2_OP: + sp_op = FP_EXP_OP; + break; + case MMA_OP: + sp_op = TENSOR__OP; + break; + case TEX_OP: + sp_op = TEX__OP; + break; + default: + if ((op == DP_OP) || (op == ALU_OP)) sp_op = DP___OP; + break; } - else if(get_type()==F16_TYPE || get_type()==F32_TYPE){ - switch(get_opcode()){ - case MUL_OP: - case MAD_OP: - case FMA_OP: - sp_op=FP_MUL_OP; - break; - case DIV_OP: - case REM_OP: - sp_op=FP_DIV_OP; - break; - case RCP_OP: - sp_op=FP_DIV_OP; - break; - case LG2_OP: - sp_op=FP_LG_OP; - break; - case RSQRT_OP: - case SQRT_OP: - sp_op=FP_SQRT_OP; - break; - case SIN_OP: - case COS_OP: - sp_op=FP_SIN_OP; - break; - case EX2_OP: - sp_op=FP_EXP_OP; - break; - case MMA_OP: - sp_op=TENSOR__OP; - break; - case TEX_OP: - sp_op=TEX__OP; - break; - default: - if((op==SP_OP) || (op==ALU_OP)) - sp_op=FP__OP; - break; - } - }else { - switch(get_opcode()){ - case MUL24_OP: - case MAD24_OP: - sp_op=INT_MUL24_OP; - break; - case MUL_OP: - case MAD_OP: - case FMA_OP: - if(get_type()==U32_TYPE || get_type()==S32_TYPE || get_type()==B32_TYPE) - sp_op=INT_MUL32_OP; - else - sp_op=INT_MUL_OP; - break; - case DIV_OP: - case REM_OP: - sp_op=INT_DIV_OP; - break; - case MMA_OP: - sp_op=TENSOR__OP; - break; - case TEX_OP: - sp_op=TEX__OP; - break; - default: - if((op==INTP_OP) || (op==ALU_OP)) - sp_op=INT__OP; - break; - } + } else if (get_type() == F16_TYPE || get_type() == F32_TYPE) { + switch (get_opcode()) { + case MUL_OP: + case MAD_OP: + case FMA_OP: + sp_op = FP_MUL_OP; + break; + case DIV_OP: + case REM_OP: + sp_op = FP_DIV_OP; + break; + case RCP_OP: + sp_op = FP_DIV_OP; + break; + case LG2_OP: + sp_op = FP_LG_OP; + break; + case RSQRT_OP: + case SQRT_OP: + sp_op = FP_SQRT_OP; + break; + case SIN_OP: + case COS_OP: + sp_op = FP_SIN_OP; + break; + case EX2_OP: + sp_op = FP_EXP_OP; + break; + case MMA_OP: + sp_op = TENSOR__OP; + break; + case TEX_OP: + sp_op = TEX__OP; + break; + default: + if ((op == SP_OP) || (op == ALU_OP)) sp_op = FP__OP; + break; } + } else { + switch (get_opcode()) { + case MUL24_OP: + case MAD24_OP: + sp_op = INT_MUL24_OP; + break; + case MUL_OP: + case MAD_OP: + case FMA_OP: + if (get_type() == U32_TYPE || get_type() == S32_TYPE || + get_type() == B32_TYPE) + sp_op = INT_MUL32_OP; + else + sp_op = INT_MUL_OP; + break; + case DIV_OP: + case REM_OP: + sp_op = INT_DIV_OP; + break; + case MMA_OP: + sp_op = TENSOR__OP; + break; + case TEX_OP: + sp_op = TEX__OP; + break; + default: + if ((op == INTP_OP) || (op == ALU_OP)) sp_op = INT__OP; + break; + } + } } } @@ -960,14 +961,16 @@ void ptx_instruction::set_opcode_and_latency() { break; } break; - case MUL24_OP: //MUL24 is performed on mul32 units (with additional instructions for bitmasking) on devices with compute capability >1.x - latency = int_latency[2]+1; - initiation_interval = int_init[2]+1; + case MUL24_OP: // MUL24 is performed on mul32 units (with additional + // instructions for bitmasking) on devices with compute + // capability >1.x + latency = int_latency[2] + 1; + initiation_interval = int_init[2] + 1; op = INTP_OP; break; case MAD24_OP: - latency = int_latency[3]+1; - initiation_interval = int_init[3]+1; + latency = int_latency[3] + 1; + initiation_interval = int_init[3] + 1; op = INTP_OP; break; case DIV_OP: @@ -1533,10 +1536,10 @@ void function_info::ptx_jit_config( filename_c.c_str()); assert(system(buff) != 0); FILE *fp = fopen(filename_c.c_str(), "r"); - char * ptr = fgets(buff, 1024, fp); - if(ptr == NULL ){ - printf("can't read file %s \n", filename_c.c_str()); - assert(0); + char *ptr = fgets(buff, 1024, fp); + if (ptr == NULL) { + printf("can't read file %s \n", filename_c.c_str()); + assert(0); } fclose(fp); std::string fn(buff); diff --git a/src/cuda-sim/instructions.cc b/src/cuda-sim/instructions.cc index 4981c99..4792efc 100644 --- a/src/cuda-sim/instructions.cc +++ b/src/cuda-sim/instructions.cc @@ -1,18 +1,19 @@ // Copyright (c) 2009-2021, Tor M. Aamodt, Wilson W.L. Fung, Ali Bakhoda, // Jimmy Kwa, George L. Yuan, Vijay Kandiah, Nikos Hardavellas, // Mahmoud Khairy, Junrui Pan, Timothy G. Rogers -// The University of British Columbia, Northwestern University, Purdue University -// All rights reserved. +// The University of British Columbia, Northwestern University, Purdue +// University All rights reserved. // // Redistribution and use in source and binary forms, with or without // modification, are permitted provided that the following conditions are met: // -// 1. Redistributions of source code must retain the above copyright notice, this +// 1. Redistributions of source code must retain the above copyright notice, +// this // list of conditions and the following disclaimer; // 2. Redistributions in binary form must reproduce the above copyright notice, // this list of conditions and the following disclaimer in the documentation // and/or other materials provided with the distribution; -// 3. Neither the names of The University of British Columbia, Northwestern +// 3. Neither the names of The University of British Columbia, Northwestern // University nor the names of their contributors may be used to // endorse or promote products derived from this software without specific // prior written permission. @@ -1948,7 +1949,7 @@ void mma_impl(const ptx_instruction *pI, core_t *core, warp_inst_t inst) { hex_val = (v[k / 2].s64 & 0xffff); else hex_val = ((v[k / 2].s64 & 0xffff0000) >> 16); - nw_v[k].f16 = *(reinterpret_cast(hex_val)); + nw_v[k].f16 = *(reinterpret_cast(hex_val)); } } if (!((operand_num == 3) && (type2 == F32_TYPE))) { @@ -3980,7 +3981,7 @@ void mad_def(const ptx_instruction *pI, ptx_thread_info *thread, fesetround(FE_TOWARDZERO); break; default: - //assert(0); + // assert(0); break; } d.f32 = a.f32 * b.f32 + c.f32; @@ -4326,7 +4327,7 @@ void mul_impl(const ptx_instruction *pI, ptx_thread_info *thread) { case S64_TYPE: t.s64 = a.s64 * b.s64; assert(!pI->is_wide()); - //assert(!pI->is_hi()); + // assert(!pI->is_hi()); d.s64 = t.s64; break; case U16_TYPE: diff --git a/src/cuda-sim/ptx_ir.cc b/src/cuda-sim/ptx_ir.cc index f25f1d5..d309542 100644 --- a/src/cuda-sim/ptx_ir.cc +++ b/src/cuda-sim/ptx_ir.cc @@ -1,18 +1,19 @@ // Copyright (c) 2009-2021, Tor M. Aamodt, Ali Bakhoda, Wilson W.L. Fung, // George L. Yuan, Vijay Kandiah, Nikos Hardavellas, // Mahmoud Khairy, Junrui Pan, Timothy G. Rogers -// The University of British Columbia, Northwestern University, Purdue University -// All rights reserved. +// The University of British Columbia, Northwestern University, Purdue +// University All rights reserved. // // Redistribution and use in source and binary forms, with or without // modification, are permitted provided that the following conditions are met: // -// 1. Redistributions of source code must retain the above copyright notice, this +// 1. Redistributions of source code must retain the above copyright notice, +// this // list of conditions and the following disclaimer; // 2. Redistributions in binary form must reproduce the above copyright notice, // this list of conditions and the following disclaimer in the documentation // and/or other materials provided with the distribution; -// 3. Neither the names of The University of British Columbia, Northwestern +// 3. Neither the names of The University of British Columbia, Northwestern // University nor the names of their contributors may be used to // endorse or promote products derived from this software without specific // prior written permission. @@ -1386,7 +1387,7 @@ ptx_instruction::ptx_instruction( case CS_OPTION: case LU_OPTION: case CV_OPTION: - case WB_OPTION: + case WB_OPTION: case WT_OPTION: m_cache_option = last_ptx_inst_option; break; @@ -1469,8 +1470,8 @@ std::string ptx_instruction::to_string() const { char buf[STR_SIZE]; unsigned used_bytes = 0; if (!is_label()) { - used_bytes += - snprintf(buf + used_bytes, STR_SIZE - used_bytes, " PC=0x%03llx ", m_PC); + used_bytes += snprintf(buf + used_bytes, STR_SIZE - used_bytes, + " PC=0x%03llx ", m_PC); } else { used_bytes += snprintf(buf + used_bytes, STR_SIZE - used_bytes, " "); diff --git a/src/cuda-sim/ptx_ir.h b/src/cuda-sim/ptx_ir.h index 7ba7171..8b1f19c 100644 --- a/src/cuda-sim/ptx_ir.h +++ b/src/cuda-sim/ptx_ir.h @@ -1248,7 +1248,7 @@ class function_info { const ptx_version &get_ptx_version() const { return m_symtab->get_ptx_version(); } - virtual ~function_info(){} + virtual ~function_info() {} unsigned get_sm_target() const { return m_symtab->get_sm_target(); } bool is_extern() const { return m_extern; } void set_name(const char *name) { m_name = name; } diff --git a/src/cuda-sim/ptx_sim.cc b/src/cuda-sim/ptx_sim.cc index 6503499..2a548ee 100644 --- a/src/cuda-sim/ptx_sim.cc +++ b/src/cuda-sim/ptx_sim.cc @@ -369,7 +369,8 @@ static void print_reg(FILE *fp, std::string name, ptx_reg_t value, fprintf(fp, ".u64 %llu [0x%llx]\n", value.u64, value.u64); break; case F16_TYPE: - fprintf(fp, ".f16 %f [0x%04x]\n", static_cast(value.f16), (unsigned)value.u16); + fprintf(fp, ".f16 %f [0x%04x]\n", static_cast(value.f16), + (unsigned)value.u16); break; case F32_TYPE: fprintf(fp, ".f32 %.15lf [0x%08x]\n", value.f32, value.u32); diff --git a/src/debug.cc b/src/debug.cc index e23ffd4..8cc5e1f 100644 --- a/src/debug.cc +++ b/src/debug.cc @@ -124,7 +124,7 @@ void gpgpu_sim::gpgpu_debug() { fflush(stdout); char line[1024]; - char * ptr = fgets(line, 1024, stdin); + char *ptr = fgets(line, 1024, stdin); char *tok = strtok(line, " \t\n"); if (!strcmp(tok, "dp")) { @@ -137,7 +137,7 @@ void gpgpu_sim::gpgpu_debug() { } else if (!strcmp(tok, "q") || !strcmp(tok, "quit")) { printf("\nreally quit GPGPU-Sim (y/n)?\n"); ptr = fgets(line, 1024, stdin); - if(ptr == NULL ){ + if (ptr == NULL) { printf("can't read input\n"); exit(0); } diff --git a/src/gpgpu-sim/dram.cc b/src/gpgpu-sim/dram.cc index 53c8238..80e20d7 100644 --- a/src/gpgpu-sim/dram.cc +++ b/src/gpgpu-sim/dram.cc @@ -1,18 +1,19 @@ // Copyright (c) 2009-2021, Tor M. Aamodt, Wilson W.L. Fung, Ali Bakhoda, -// Ivan Sham, George L. Yuan, Vijay Kandiah, Nikos Hardavellas, +// Ivan Sham, George L. Yuan, Vijay Kandiah, Nikos Hardavellas, // Mahmoud Khairy, Junrui Pan, Timothy G. Rogers -// The University of British Columbia, Northwestern University, Purdue University -// All rights reserved. +// The University of British Columbia, Northwestern University, Purdue +// University All rights reserved. // // Redistribution and use in source and binary forms, with or without // modification, are permitted provided that the following conditions are met: // -// 1. Redistributions of source code must retain the above copyright notice, this +// 1. Redistributions of source code must retain the above copyright notice, +// this // list of conditions and the following disclaimer; // 2. Redistributions in binary form must reproduce the above copyright notice, // this list of conditions and the following disclaimer in the documentation // and/or other materials provided with the distribution; -// 3. Neither the names of The University of British Columbia, Northwestern +// 3. Neither the names of The University of British Columbia, Northwestern // University nor the names of their contributors may be used to // endorse or promote products derived from this software without specific // prior written permission. @@ -592,39 +593,40 @@ bool dram_t::issue_col_command(int j) { bk[j]->mrq = NULL; } } else - // correct row activated for a WRITE - if (!issued && !CCDc && !bk[j]->RCDWRc && !(bkgrp[grp]->CCDLc) && - (bk[j]->curr_row == bk[j]->mrq->row) && (bk[j]->mrq->rw == WRITE) && - (RTWc == 0) && (bk[j]->state == BANK_ACTIVE) && !rwq->full()) { - if (rw == READ) { - rw = WRITE; - rwq->set_min_length(m_config->WL); - } - rwq->push(bk[j]->mrq); - - bk[j]->mrq->txbytes += m_config->dram_atom_size; - CCDc = m_config->tCCD; - bkgrp[grp]->CCDLc = m_config->tCCDL; - WTRc = m_config->tWTR; - bk[j]->WTPc = m_config->tWTP; - issued = true; - - if (bk[j]->mrq->data->get_access_type() == L2_WRBK_ACC) - n_wr_WB++; - else - n_wr++; - bwutil += m_config->BL / m_config->data_command_freq_ratio; - bwutil_partial += m_config->BL / m_config->data_command_freq_ratio; + // correct row activated for a WRITE + if (!issued && !CCDc && !bk[j]->RCDWRc && !(bkgrp[grp]->CCDLc) && + (bk[j]->curr_row == bk[j]->mrq->row) && (bk[j]->mrq->rw == WRITE) && + (RTWc == 0) && (bk[j]->state == BANK_ACTIVE) && !rwq->full()) { + if (rw == READ) { + rw = WRITE; + rwq->set_min_length(m_config->WL); + } + rwq->push(bk[j]->mrq); + + bk[j]->mrq->txbytes += m_config->dram_atom_size; + CCDc = m_config->tCCD; + bkgrp[grp]->CCDLc = m_config->tCCDL; + WTRc = m_config->tWTR; + bk[j]->WTPc = m_config->tWTP; + issued = true; + + if (bk[j]->mrq->data->get_access_type() == L2_WRBK_ACC) + n_wr_WB++; + else + n_wr++; + bwutil += m_config->BL / m_config->data_command_freq_ratio; + bwutil_partial += m_config->BL / m_config->data_command_freq_ratio; #ifdef DRAM_VERIFY - PRINT_CYCLE = 1; - printf("\tWR Bk:%d Row:%03x Col:%03x \n", j, bk[j]->curr_row, - bk[j]->mrq->col + bk[j]->mrq->txbytes - m_config->dram_atom_size); + PRINT_CYCLE = 1; + printf( + "\tWR Bk:%d Row:%03x Col:%03x \n", j, bk[j]->curr_row, + bk[j]->mrq->col + bk[j]->mrq->txbytes - m_config->dram_atom_size); #endif - // transfer done - if (!(bk[j]->mrq->txbytes < bk[j]->mrq->nbytes)) { - bk[j]->mrq = NULL; + // transfer done + if (!(bk[j]->mrq->txbytes < bk[j]->mrq->nbytes)) { + bk[j]->mrq = NULL; + } } - } } return issued; @@ -660,23 +662,23 @@ bool dram_t::issue_row_command(int j) { } else - // different row activated - if ((!issued) && (bk[j]->curr_row != bk[j]->mrq->row) && - (bk[j]->state == BANK_ACTIVE) && - (!bk[j]->RASc && !bk[j]->WTPc && !bk[j]->RTPc && - !bkgrp[grp]->RTPLc)) { - // make the bank idle again - bk[j]->state = BANK_IDLE; - bk[j]->RPc = m_config->tRP; - prio = (j + 1) % m_config->nbk; - issued = true; - n_pre++; - n_pre_partial++; + // different row activated + if ((!issued) && (bk[j]->curr_row != bk[j]->mrq->row) && + (bk[j]->state == BANK_ACTIVE) && + (!bk[j]->RASc && !bk[j]->WTPc && !bk[j]->RTPc && + !bkgrp[grp]->RTPLc)) { + // make the bank idle again + bk[j]->state = BANK_IDLE; + bk[j]->RPc = m_config->tRP; + prio = (j + 1) % m_config->nbk; + issued = true; + n_pre++; + n_pre_partial++; #ifdef DRAM_VERIFY - PRINT_CYCLE = 1; - printf("\tPRE BK:%d Row:%03x \n", j, bk[j]->curr_row); + PRINT_CYCLE = 1; + printf("\tPRE BK:%d Row:%03x \n", j, bk[j]->curr_row); #endif - } + } } return issued; } @@ -880,5 +882,5 @@ unsigned dram_t::get_bankgrp_number(unsigned i) { } else { assert(1); } - return 0; // we should never get here + return 0; // we should never get here } diff --git a/src/gpgpu-sim/dram.h b/src/gpgpu-sim/dram.h index 90ea3e4..9e9517b 100644 --- a/src/gpgpu-sim/dram.h +++ b/src/gpgpu-sim/dram.h @@ -1,18 +1,19 @@ // Copyright (c) 2009-2021, Tor M. Aamodt, Ivan Sham, Ali Bakhoda, // George L. Yuan, Wilson W.L. Fung, Vijay Kandiah, Nikos Hardavellas, // Mahmoud Khairy, Junrui Pan, Timothy G. Rogers -// The University of British Columbia, Northwestern University, Purdue University -// All rights reserved. +// The University of British Columbia, Northwestern University, Purdue +// University All rights reserved. // // Redistribution and use in source and binary forms, with or without // modification, are permitted provided that the following conditions are met: // -// 1. Redistributions of source code must retain the above copyright notice, this +// 1. Redistributions of source code must retain the above copyright notice, +// this // list of conditions and the following disclaimer; // 2. Redistributions in binary form must reproduce the above copyright notice, // this list of conditions and the following disclaimer in the documentation // and/or other materials provided with the distribution; -// 3. Neither the names of The University of British Columbia, Northwestern +// 3. Neither the names of The University of British Columbia, Northwestern // University nor the names of their contributors may be used to // endorse or promote products derived from this software without specific // prior written permission. diff --git a/src/gpgpu-sim/gpu-cache.cc b/src/gpgpu-sim/gpu-cache.cc index f4448d3..32cc56b 100644 --- a/src/gpgpu-sim/gpu-cache.cc +++ b/src/gpgpu-sim/gpu-cache.cc @@ -1,18 +1,19 @@ -// Copyright (c) 2009-2021, Tor M. Aamodt, Tayler Hetherington, +// Copyright (c) 2009-2021, Tor M. Aamodt, Tayler Hetherington, // Vijay Kandiah, Nikos Hardavellas, Mahmoud Khairy, Junrui Pan, // Timothy G. Rogers -// The University of British Columbia, Northwestern University, Purdue University -// All rights reserved. +// The University of British Columbia, Northwestern University, Purdue +// University All rights reserved. // // Redistribution and use in source and binary forms, with or without // modification, are permitted provided that the following conditions are met: // -// 1. Redistributions of source code must retain the above copyright notice, this +// 1. Redistributions of source code must retain the above copyright notice, +// this // list of conditions and the following disclaimer; // 2. Redistributions in binary form must reproduce the above copyright notice, // this list of conditions and the following disclaimer in the documentation // and/or other materials provided with the distribution; -// 3. Neither the names of The University of British Columbia, Northwestern +// 3. Neither the names of The University of British Columbia, Northwestern // University nor the names of their contributors may be used to // endorse or promote products derived from this software without specific // prior written permission. @@ -287,10 +288,11 @@ enum cache_request_status tag_array::probe(new_addr_type addr, unsigned &idx, // number of dirty lines / total lines in the cache float dirty_line_percentage = ((float)m_dirty / (m_config.m_nset * m_config.m_assoc)) * 100; - // If the cacheline is from a load op (not modified), + // If the cacheline is from a load op (not modified), // or the total dirty cacheline is above a specific value, - // Then this cacheline is eligible to be considered for replacement candidate - // i.e. Only evict clean cachelines until total dirty cachelines reach the limit. + // Then this cacheline is eligible to be considered for replacement + // candidate i.e. Only evict clean cachelines until total dirty cachelines + // reach the limit. if (!line->is_modified_line() || dirty_line_percentage >= m_config.m_wr_percent) { all_reserved = false; @@ -411,7 +413,7 @@ void tag_array::fill(new_addr_type addr, unsigned time, enum cache_request_status status = probe(addr, idx, mask, is_write); if (status == RESERVATION_FAIL) { - return; + return; } bool before = m_lines[idx]->is_modified_line(); @@ -437,7 +439,8 @@ void tag_array::fill(new_addr_type addr, unsigned time, void tag_array::fill(unsigned index, unsigned time, mem_fetch *mf) { assert(m_config.m_alloc_policy == ON_MISS); bool before = m_lines[index]->is_modified_line(); - m_lines[index]->fill(time, mf->get_access_sector_mask(), mf->get_access_byte_mask()); + m_lines[index]->fill(time, mf->get_access_sector_mask(), + mf->get_access_byte_mask()); if (m_lines[index]->is_modified_line() && !before) { m_dirty++; } @@ -1210,15 +1213,14 @@ void data_cache::update_m_readable(mem_fetch *mf, unsigned cache_index) { if (mf->get_access_sector_mask().test(i)) { bool all_set = true; for (unsigned k = i * SECTOR_SIZE; k < (i + 1) * SECTOR_SIZE; k++) { - // If any bit in the byte mask (within the sector) is not set, + // If any bit in the byte mask (within the sector) is not set, // the sector is unreadble if (!block->get_dirty_byte_mask().test(k)) { all_set = false; break; } } - if (all_set) - block->set_m_readable(true, mf->get_access_sector_mask()); + if (all_set) block->set_m_readable(true, mf->get_access_sector_mask()); } } } @@ -1239,7 +1241,7 @@ cache_request_status data_cache::wr_hit_wb(new_addr_type addr, } block->set_status(MODIFIED, mf->get_access_sector_mask()); block->set_byte_mask(mf); - update_m_readable(mf,cache_index); + update_m_readable(mf, cache_index); return HIT; } @@ -1263,7 +1265,7 @@ cache_request_status data_cache::wr_hit_wt(new_addr_type addr, } block->set_status(MODIFIED, mf->get_access_sector_mask()); block->set_byte_mask(mf); - update_m_readable(mf,cache_index); + update_m_readable(mf, cache_index); // generate a write-through send_write_request(mf, cache_event(WRITE_REQUEST_SENT), time, events); @@ -1559,7 +1561,7 @@ enum cache_request_status data_cache::wr_miss_wa_lazy_fetch_on_read( if (m_status == HIT_RESERVED) block->set_readable_on_fill(true, mf->get_access_sector_mask()); } - update_m_readable(mf,cache_index); + update_m_readable(mf, cache_index); if (m_status != RESERVATION_FAIL) { // If evicted block is modified and not a write-through diff --git a/src/gpgpu-sim/gpu-cache.h b/src/gpgpu-sim/gpu-cache.h index ad41320..5fd40a9 100644 --- a/src/gpgpu-sim/gpu-cache.h +++ b/src/gpgpu-sim/gpu-cache.h @@ -1,17 +1,18 @@ -// Copyright (c) 2009-2021, Tor M. Aamodt, Tayler Hetherington, Vijay Kandiah, Nikos Hardavellas, -// Mahmoud Khairy, Junrui Pan, Timothy G. Rogers -// The University of British Columbia, Northwestern University, Purdue University +// Copyright (c) 2009-2021, Tor M. Aamodt, Tayler Hetherington, Vijay Kandiah, +// Nikos Hardavellas, Mahmoud Khairy, Junrui Pan, Timothy G. Rogers The +// University of British Columbia, Northwestern University, Purdue University // All rights reserved. // // Redistribution and use in source and binary forms, with or without // modification, are permitted provided that the following conditions are met: // -// 1. Redistributions of source code must retain the above copyright notice, this +// 1. Redistributions of source code must retain the above copyright notice, +// this // list of conditions and the following disclaimer; // 2. Redistributions in binary form must reproduce the above copyright notice, // this list of conditions and the following disclaimer in the documentation // and/or other materials provided with the distribution; -// 3. Neither the names of The University of British Columbia, Northwestern +// 3. Neither the names of The University of British Columbia, Northwestern // University nor the names of their contributors may be used to // endorse or promote products derived from this software without specific // prior written permission. @@ -499,7 +500,7 @@ struct sector_cache_block : public cache_block_t { for (unsigned i = 0; i < SECTOR_CHUNCK_SIZE; ++i) { if (sector_mask.to_ulong() & (1 << i)) return i; } - return SECTOR_CHUNCK_SIZE; //error + return SECTOR_CHUNCK_SIZE; // error } }; @@ -564,12 +565,10 @@ class cache_config { char ct, rp, wp, ap, mshr_type, wap, sif; int ntok = - sscanf(config, "%c:%u:%u:%u,%c:%c:%c:%c:%c,%c:%u:%u,%u:%u,%u", - &ct, &m_nset, &m_line_sz, &m_assoc, - &rp, &wp, &ap, &wap, &sif, - &mshr_type, &m_mshr_entries, &m_mshr_max_merge, - &m_miss_queue_size, &m_result_fifo_entries, - &m_data_port_width); + sscanf(config, "%c:%u:%u:%u,%c:%c:%c:%c:%c,%c:%u:%u,%u:%u,%u", &ct, + &m_nset, &m_line_sz, &m_assoc, &rp, &wp, &ap, &wap, &sif, + &mshr_type, &m_mshr_entries, &m_mshr_max_merge, + &m_miss_queue_size, &m_result_fifo_entries, &m_data_port_width); if (ntok < 12) { if (!strcmp(config, "none")) { @@ -726,15 +725,14 @@ class cache_config { } if (m_cache_type == SECTOR) { - bool cond = - m_line_sz / SECTOR_SIZE == SECTOR_CHUNCK_SIZE && - m_line_sz % SECTOR_SIZE == 0; - if(!cond){ - std::cerr<<"error: For sector cache, the simulator uses hard-coded " - "SECTOR_SIZE and SECTOR_CHUNCK_SIZE. The line size " - "must be product of both values.\n"; - assert(0); - } + bool cond = m_line_sz / SECTOR_SIZE == SECTOR_CHUNCK_SIZE && + m_line_sz % SECTOR_SIZE == 0; + if (!cond) { + std::cerr << "error: For sector cache, the simulator uses hard-coded " + "SECTOR_SIZE and SECTOR_CHUNCK_SIZE. The line size " + "must be product of both values.\n"; + assert(0); + } } // default: port to data array width and granularity = line size diff --git a/src/gpgpu-sim/gpu-sim.cc b/src/gpgpu-sim/gpu-sim.cc index 47c0b4a..1cb8a25 100644 --- a/src/gpgpu-sim/gpu-sim.cc +++ b/src/gpgpu-sim/gpu-sim.cc @@ -1,18 +1,19 @@ // Copyright (c) 2009-2021, Tor M. Aamodt, Wilson W.L. Fung, George L. Yuan, -// Ali Bakhoda, Andrew Turner, Ivan Sham, Vijay Kandiah, Nikos Hardavellas, +// Ali Bakhoda, Andrew Turner, Ivan Sham, Vijay Kandiah, Nikos Hardavellas, // Mahmoud Khairy, Junrui Pan, Timothy G. Rogers -// The University of British Columbia, Northwestern University, Purdue University -// All rights reserved. +// The University of British Columbia, Northwestern University, Purdue +// University All rights reserved. // // Redistribution and use in source and binary forms, with or without // modification, are permitted provided that the following conditions are met: // -// 1. Redistributions of source code must retain the above copyright notice, this +// 1. Redistributions of source code must retain the above copyright notice, +// this // list of conditions and the following disclaimer; // 2. Redistributions in binary form must reproduce the above copyright notice, // this list of conditions and the following disclaimer in the documentation // and/or other materials provided with the distribution; -// 3. Neither the names of The University of British Columbia, Northwestern +// 3. Neither the names of The University of British Columbia, Northwestern // University nor the names of their contributors may be used to // endorse or promote products derived from this software without specific // prior written permission. @@ -80,7 +81,7 @@ class gpgpu_sim_wrapper {}; #include #include -// #define MAX(a, b) (((a) > (b)) ? (a) : (b)) //redefined +// #define MAX(a, b) (((a) > (b)) ? (a) : (b)) //redefined bool g_interactive_debugger_enabled = false; @@ -97,7 +98,6 @@ tr1_hash_map address_random_interleaving; #include "mem_latency_stat.h" - void power_config::reg_options(class OptionParser *opp) { option_parser_register(opp, "-accelwattch_xml_file", OPT_CSTR, &g_power_config_name, "AccelWattch XML file", @@ -111,91 +111,106 @@ void power_config::reg_options(class OptionParser *opp) { &g_power_per_cycle_dump, "Dump detailed power output each cycle", "0"); - - - option_parser_register(opp, "-hw_perf_file_name", OPT_CSTR, - &g_hw_perf_file_name, "Hardware Performance Statistics file", - "hw_perf.csv"); + &g_hw_perf_file_name, + "Hardware Performance Statistics file", "hw_perf.csv"); - option_parser_register(opp, "-hw_perf_bench_name", OPT_CSTR, - &g_hw_perf_bench_name, "Kernel Name in Hardware Performance Statistics file", - ""); + option_parser_register( + opp, "-hw_perf_bench_name", OPT_CSTR, &g_hw_perf_bench_name, + "Kernel Name in Hardware Performance Statistics file", ""); option_parser_register(opp, "-power_simulation_mode", OPT_INT32, &g_power_simulation_mode, - "Switch performance counter input for power simulation (0=Sim, 1=HW, 2=HW-Sim Hybrid)", "0"); + "Switch performance counter input for power " + "simulation (0=Sim, 1=HW, 2=HW-Sim Hybrid)", + "0"); - option_parser_register(opp, "-dvfs_enabled", OPT_BOOL, - &g_dvfs_enabled, + option_parser_register(opp, "-dvfs_enabled", OPT_BOOL, &g_dvfs_enabled, "Turn on DVFS for power model", "0"); option_parser_register(opp, "-aggregate_power_stats", OPT_BOOL, &g_aggregate_power_stats, "Accumulate power across all kernels", "0"); - //Accelwattch Hyrbid Configuration - - option_parser_register(opp, "-accelwattch_hybrid_perfsim_L1_RH", OPT_BOOL, - &accelwattch_hybrid_configuration[HW_L1_RH], - "Get L1 Read Hits for Accelwattch-Hybrid from Accel-Sim", "0"); - option_parser_register(opp, "-accelwattch_hybrid_perfsim_L1_RM", OPT_BOOL, - &accelwattch_hybrid_configuration[HW_L1_RM], - "Get L1 Read Misses for Accelwattch-Hybrid from Accel-Sim", "0"); - option_parser_register(opp, "-accelwattch_hybrid_perfsim_L1_WH", OPT_BOOL, - &accelwattch_hybrid_configuration[HW_L1_WH], - "Get L1 Write Hits for Accelwattch-Hybrid from Accel-Sim", "0"); - option_parser_register(opp, "-accelwattch_hybrid_perfsim_L1_WM", OPT_BOOL, - &accelwattch_hybrid_configuration[HW_L1_WM], - "Get L1 Write Misses for Accelwattch-Hybrid from Accel-Sim", "0"); - - option_parser_register(opp, "-accelwattch_hybrid_perfsim_L2_RH", OPT_BOOL, - &accelwattch_hybrid_configuration[HW_L2_RH], - "Get L2 Read Hits for Accelwattch-Hybrid from Accel-Sim", "0"); - option_parser_register(opp, "-accelwattch_hybrid_perfsim_L2_RM", OPT_BOOL, - &accelwattch_hybrid_configuration[HW_L2_RM], - "Get L2 Read Misses for Accelwattch-Hybrid from Accel-Sim", "0"); - option_parser_register(opp, "-accelwattch_hybrid_perfsim_L2_WH", OPT_BOOL, - &accelwattch_hybrid_configuration[HW_L2_WH], - "Get L2 Write Hits for Accelwattch-Hybrid from Accel-Sim", "0"); - option_parser_register(opp, "-accelwattch_hybrid_perfsim_L2_WM", OPT_BOOL, - &accelwattch_hybrid_configuration[HW_L2_WM], - "Get L2 Write Misses for Accelwattch-Hybrid from Accel-Sim", "0"); - - option_parser_register(opp, "-accelwattch_hybrid_perfsim_CC_ACC", OPT_BOOL, - &accelwattch_hybrid_configuration[HW_CC_ACC], - "Get Constant Cache Acesses for Accelwattch-Hybrid from Accel-Sim", "0"); - - option_parser_register(opp, "-accelwattch_hybrid_perfsim_SHARED_ACC", OPT_BOOL, - &accelwattch_hybrid_configuration[HW_SHRD_ACC], - "Get Shared Memory Acesses for Accelwattch-Hybrid from Accel-Sim", "0"); + // Accelwattch Hyrbid Configuration + + option_parser_register( + opp, "-accelwattch_hybrid_perfsim_L1_RH", OPT_BOOL, + &accelwattch_hybrid_configuration[HW_L1_RH], + "Get L1 Read Hits for Accelwattch-Hybrid from Accel-Sim", "0"); + option_parser_register( + opp, "-accelwattch_hybrid_perfsim_L1_RM", OPT_BOOL, + &accelwattch_hybrid_configuration[HW_L1_RM], + "Get L1 Read Misses for Accelwattch-Hybrid from Accel-Sim", "0"); + option_parser_register( + opp, "-accelwattch_hybrid_perfsim_L1_WH", OPT_BOOL, + &accelwattch_hybrid_configuration[HW_L1_WH], + "Get L1 Write Hits for Accelwattch-Hybrid from Accel-Sim", "0"); + option_parser_register( + opp, "-accelwattch_hybrid_perfsim_L1_WM", OPT_BOOL, + &accelwattch_hybrid_configuration[HW_L1_WM], + "Get L1 Write Misses for Accelwattch-Hybrid from Accel-Sim", "0"); + + option_parser_register( + opp, "-accelwattch_hybrid_perfsim_L2_RH", OPT_BOOL, + &accelwattch_hybrid_configuration[HW_L2_RH], + "Get L2 Read Hits for Accelwattch-Hybrid from Accel-Sim", "0"); + option_parser_register( + opp, "-accelwattch_hybrid_perfsim_L2_RM", OPT_BOOL, + &accelwattch_hybrid_configuration[HW_L2_RM], + "Get L2 Read Misses for Accelwattch-Hybrid from Accel-Sim", "0"); + option_parser_register( + opp, "-accelwattch_hybrid_perfsim_L2_WH", OPT_BOOL, + &accelwattch_hybrid_configuration[HW_L2_WH], + "Get L2 Write Hits for Accelwattch-Hybrid from Accel-Sim", "0"); + option_parser_register( + opp, "-accelwattch_hybrid_perfsim_L2_WM", OPT_BOOL, + &accelwattch_hybrid_configuration[HW_L2_WM], + "Get L2 Write Misses for Accelwattch-Hybrid from Accel-Sim", "0"); + + option_parser_register( + opp, "-accelwattch_hybrid_perfsim_CC_ACC", OPT_BOOL, + &accelwattch_hybrid_configuration[HW_CC_ACC], + "Get Constant Cache Acesses for Accelwattch-Hybrid from Accel-Sim", "0"); + + option_parser_register( + opp, "-accelwattch_hybrid_perfsim_SHARED_ACC", OPT_BOOL, + &accelwattch_hybrid_configuration[HW_SHRD_ACC], + "Get Shared Memory Acesses for Accelwattch-Hybrid from Accel-Sim", "0"); option_parser_register(opp, "-accelwattch_hybrid_perfsim_DRAM_RD", OPT_BOOL, &accelwattch_hybrid_configuration[HW_DRAM_RD], - "Get DRAM Reads for Accelwattch-Hybrid from Accel-Sim", "0"); - option_parser_register(opp, "-accelwattch_hybrid_perfsim_DRAM_WR", OPT_BOOL, - &accelwattch_hybrid_configuration[HW_DRAM_WR], - "Get DRAM Writes for Accelwattch-Hybrid from Accel-Sim", "0"); - - option_parser_register(opp, "-accelwattch_hybrid_perfsim_NOC", OPT_BOOL, - &accelwattch_hybrid_configuration[HW_NOC], - "Get Interconnect Acesses for Accelwattch-Hybrid from Accel-Sim", "0"); + "Get DRAM Reads for Accelwattch-Hybrid from Accel-Sim", + "0"); + option_parser_register( + opp, "-accelwattch_hybrid_perfsim_DRAM_WR", OPT_BOOL, + &accelwattch_hybrid_configuration[HW_DRAM_WR], + "Get DRAM Writes for Accelwattch-Hybrid from Accel-Sim", "0"); - option_parser_register(opp, "-accelwattch_hybrid_perfsim_PIPE_DUTY", OPT_BOOL, - &accelwattch_hybrid_configuration[HW_PIPE_DUTY], - "Get Pipeline Duty Cycle Acesses for Accelwattch-Hybrid from Accel-Sim", "0"); + option_parser_register( + opp, "-accelwattch_hybrid_perfsim_NOC", OPT_BOOL, + &accelwattch_hybrid_configuration[HW_NOC], + "Get Interconnect Acesses for Accelwattch-Hybrid from Accel-Sim", "0"); - option_parser_register(opp, "-accelwattch_hybrid_perfsim_NUM_SM_IDLE", OPT_BOOL, - &accelwattch_hybrid_configuration[HW_NUM_SM_IDLE], - "Get Number of Idle SMs for Accelwattch-Hybrid from Accel-Sim", "0"); + option_parser_register( + opp, "-accelwattch_hybrid_perfsim_PIPE_DUTY", OPT_BOOL, + &accelwattch_hybrid_configuration[HW_PIPE_DUTY], + "Get Pipeline Duty Cycle Acesses for Accelwattch-Hybrid from Accel-Sim", + "0"); - option_parser_register(opp, "-accelwattch_hybrid_perfsim_CYCLES", OPT_BOOL, - &accelwattch_hybrid_configuration[HW_CYCLES], - "Get Executed Cycles for Accelwattch-Hybrid from Accel-Sim", "0"); + option_parser_register( + opp, "-accelwattch_hybrid_perfsim_NUM_SM_IDLE", OPT_BOOL, + &accelwattch_hybrid_configuration[HW_NUM_SM_IDLE], + "Get Number of Idle SMs for Accelwattch-Hybrid from Accel-Sim", "0"); - option_parser_register(opp, "-accelwattch_hybrid_perfsim_VOLTAGE", OPT_BOOL, - &accelwattch_hybrid_configuration[HW_VOLTAGE], - "Get Chip Voltage for Accelwattch-Hybrid from Accel-Sim", "0"); + option_parser_register( + opp, "-accelwattch_hybrid_perfsim_CYCLES", OPT_BOOL, + &accelwattch_hybrid_configuration[HW_CYCLES], + "Get Executed Cycles for Accelwattch-Hybrid from Accel-Sim", "0"); + option_parser_register( + opp, "-accelwattch_hybrid_perfsim_VOLTAGE", OPT_BOOL, + &accelwattch_hybrid_configuration[HW_VOLTAGE], + "Get Chip Voltage for Accelwattch-Hybrid from Accel-Sim", "0"); // Output Data Formats option_parser_register( @@ -702,7 +717,8 @@ void gpgpu_sim_config::reg_options(option_parser_t opp) { option_parser_register( opp, "-gpgpu_max_concurrent_kernel", OPT_INT32, &max_concurrent_kernel, "maximum kernels that can run concurrently on GPU, set this value " - "according to max resident grids for your compute capability", "32"); + "according to max resident grids for your compute capability", + "32"); option_parser_register( opp, "-gpgpu_cflog_interval", OPT_INT32, &gpgpu_cflog_interval, "Interval between each snapshot in control flow logger", "0"); @@ -924,8 +940,9 @@ gpgpu_sim::gpgpu_sim(const gpgpu_sim_config &config, gpgpu_context *ctx) ptx_file_line_stats_create_exposed_latency_tracker(m_config.num_shader()); #ifdef GPGPUSIM_POWER_MODEL - m_gpgpusim_wrapper = new gpgpu_sim_wrapper(config.g_power_simulation_enabled, - config.g_power_config_name, config.g_power_simulation_mode, config.g_dvfs_enabled); + m_gpgpusim_wrapper = new gpgpu_sim_wrapper( + config.g_power_simulation_enabled, config.g_power_config_name, + config.g_power_simulation_mode, config.g_dvfs_enabled); #endif m_shader_stats = new shader_core_stats(m_shader_config); @@ -1157,8 +1174,7 @@ void gpgpu_sim::update_stats() { gpu_occupancy = occupancy_stats(); } -PowerscalingCoefficients *gpgpu_sim::get_scaling_coeffs() -{ +PowerscalingCoefficients *gpgpu_sim::get_scaling_coeffs() { return m_gpgpusim_wrapper->get_scaling_coeffs(); } @@ -1243,10 +1259,10 @@ std::string gpgpu_sim::executed_kernel_info_string() { } std::string gpgpu_sim::executed_kernel_name() { - std::stringstream statout; - if( m_executed_kernel_names.size() == 1) - statout << m_executed_kernel_names[0]; - else{ + std::stringstream statout; + if (m_executed_kernel_names.size() == 1) + statout << m_executed_kernel_names[0]; + else { for (unsigned int k = 0; k < m_executed_kernel_names.size(); k++) { statout << m_executed_kernel_names[k] << " "; } @@ -1433,20 +1449,23 @@ void gpgpu_sim::gpu_print_stat() { m_shader_stats->print(stdout); #ifdef GPGPUSIM_POWER_MODEL if (m_config.g_power_simulation_enabled) { - if(m_config.g_power_simulation_mode > 0){ - //if(!m_config.g_aggregate_power_stats) - mcpat_reset_perf_count(m_gpgpusim_wrapper); - calculate_hw_mcpat(m_config, getShaderCoreConfig(), m_gpgpusim_wrapper, - m_power_stats, m_config.gpu_stat_sample_freq, - gpu_tot_sim_cycle, gpu_sim_cycle, gpu_tot_sim_insn, - gpu_sim_insn, m_config.g_power_simulation_mode, m_config.g_dvfs_enabled, - m_config.g_hw_perf_file_name, m_config.g_hw_perf_bench_name, executed_kernel_name(), m_config.accelwattch_hybrid_configuration, m_config.g_aggregate_power_stats); + if (m_config.g_power_simulation_mode > 0) { + // if(!m_config.g_aggregate_power_stats) + mcpat_reset_perf_count(m_gpgpusim_wrapper); + calculate_hw_mcpat(m_config, getShaderCoreConfig(), m_gpgpusim_wrapper, + m_power_stats, m_config.gpu_stat_sample_freq, + gpu_tot_sim_cycle, gpu_sim_cycle, gpu_tot_sim_insn, + gpu_sim_insn, m_config.g_power_simulation_mode, + m_config.g_dvfs_enabled, m_config.g_hw_perf_file_name, + m_config.g_hw_perf_bench_name, executed_kernel_name(), + m_config.accelwattch_hybrid_configuration, + m_config.g_aggregate_power_stats); } m_gpgpusim_wrapper->print_power_kernel_stats( gpu_sim_cycle, gpu_tot_sim_cycle, gpu_tot_sim_insn + gpu_sim_insn, kernel_info_str, true); - //if(!m_config.g_aggregate_power_stats) - mcpat_reset_perf_count(m_gpgpusim_wrapper); + // if(!m_config.g_aggregate_power_stats) + mcpat_reset_perf_count(m_gpgpusim_wrapper); } #endif @@ -1810,7 +1829,8 @@ void shader_core_ctx::issue_block2core(kernel_info_t &kernel) { "GPGPU-Sim uArch: cta:%2u, start_tid:%4u, end_tid:%4u, " "initialized @(%lld,%lld), kernel_uid:%u, kernel_name:%s\n", free_cta_hw_id, start_thread, end_thread, m_gpu->gpu_sim_cycle, - m_gpu->gpu_tot_sim_cycle, kernel.get_uid(), kernel.get_name().c_str()); + m_gpu->gpu_tot_sim_cycle, kernel.get_uid(), + kernel.get_name().c_str()); } /////////////////////////////////////////////////////////////////////////////////////////// @@ -1987,11 +2007,11 @@ void gpgpu_sim::cycle() { // McPAT main cycle (interface with McPAT) #ifdef GPGPUSIM_POWER_MODEL if (m_config.g_power_simulation_enabled) { - if(m_config.g_power_simulation_mode == 0){ - mcpat_cycle(m_config, getShaderCoreConfig(), m_gpgpusim_wrapper, - m_power_stats, m_config.gpu_stat_sample_freq, - gpu_tot_sim_cycle, gpu_sim_cycle, gpu_tot_sim_insn, - gpu_sim_insn, m_config.g_dvfs_enabled); + if (m_config.g_power_simulation_mode == 0) { + mcpat_cycle(m_config, getShaderCoreConfig(), m_gpgpusim_wrapper, + m_power_stats, m_config.gpu_stat_sample_freq, + gpu_tot_sim_cycle, gpu_sim_cycle, gpu_tot_sim_insn, + gpu_sim_insn, m_config.g_dvfs_enabled); } } #endif diff --git a/src/gpgpu-sim/gpu-sim.h b/src/gpgpu-sim/gpu-sim.h index de69ef8..a24ffd3 100644 --- a/src/gpgpu-sim/gpu-sim.h +++ b/src/gpgpu-sim/gpu-sim.h @@ -1,17 +1,18 @@ -// Copyright (c) 2009-2021, Tor M. Aamodt, Wilson W.L. Fung, Vijay Kandiah, Nikos Hardavellas -// Mahmoud Khairy, Junrui Pan, Timothy G. Rogers -// The University of British Columbia, Northwestern University, Purdue University +// Copyright (c) 2009-2021, Tor M. Aamodt, Wilson W.L. Fung, Vijay Kandiah, +// Nikos Hardavellas Mahmoud Khairy, Junrui Pan, Timothy G. Rogers The +// University of British Columbia, Northwestern University, Purdue University // All rights reserved. // // Redistribution and use in source and binary forms, with or without // modification, are permitted provided that the following conditions are met: // -// 1. Redistributions of source code must retain the above copyright notice, this +// 1. Redistributions of source code must retain the above copyright notice, +// this // list of conditions and the following disclaimer; // 2. Redistributions in binary form must reproduce the above copyright notice, // this list of conditions and the following disclaimer in the documentation // and/or other materials provided with the distribution; -// 3. Neither the names of The University of British Columbia, Northwestern +// 3. Neither the names of The University of British Columbia, Northwestern // University nor the names of their contributors may be used to // endorse or promote products derived from this software without specific // prior written permission. @@ -28,7 +29,6 @@ // ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE // POSSIBILITY OF SUCH DAMAGE. - #ifndef GPU_SIM_H #define GPU_SIM_H @@ -72,7 +72,7 @@ extern tr1_hash_map address_random_interleaving; enum dram_ctrl_t { DRAM_FIFO = 0, DRAM_FRFCFS = 1 }; enum hw_perf_t { - HW_BENCH_NAME=0, + HW_BENCH_NAME = 0, HW_KERNEL_NAME, HW_L1_RH, HW_L1_RM, @@ -108,7 +108,7 @@ struct power_config { s++; } char buf1[1024]; - //snprintf(buf1, 1024, "accelwattch_power_report__%s.log", date); + // snprintf(buf1, 1024, "accelwattch_power_report__%s.log", date); snprintf(buf1, 1024, "accelwattch_power_report.log"); g_power_filename = strdup(buf1); char buf2[1024]; @@ -155,7 +155,6 @@ struct power_config { double gpu_steady_power_deviation; double gpu_steady_min_period; - char *g_hw_perf_file_name; char *g_hw_perf_bench_name; int g_power_simulation_mode; diff --git a/src/gpgpu-sim/l2cache.cc b/src/gpgpu-sim/l2cache.cc index 5b63765..8469453 100644 --- a/src/gpgpu-sim/l2cache.cc +++ b/src/gpgpu-sim/l2cache.cc @@ -1,17 +1,18 @@ -// Copyright (c) 2009-2021, Tor M. Aamodt, Vijay Kandiah, Nikos Hardavellas, +// Copyright (c) 2009-2021, Tor M. Aamodt, Vijay Kandiah, Nikos Hardavellas, // Mahmoud Khairy, Junrui Pan, Timothy G. Rogers -// The University of British Columbia, Northwestern University, Purdue University -// All rights reserved. +// The University of British Columbia, Northwestern University, Purdue +// University All rights reserved. // // Redistribution and use in source and binary forms, with or without // modification, are permitted provided that the following conditions are met: // -// 1. Redistributions of source code must retain the above copyright notice, this +// 1. Redistributions of source code must retain the above copyright notice, +// this // list of conditions and the following disclaimer; // 2. Redistributions in binary form must reproduce the above copyright notice, // this list of conditions and the following disclaimer in the documentation // and/or other materials provided with the distribution; -// 3. Neither the names of The University of British Columbia, Northwestern +// 3. Neither the names of The University of British Columbia, Northwestern // University nor the names of their contributors may be used to // endorse or promote products derived from this software without specific // prior written permission. @@ -390,7 +391,8 @@ void memory_partition_unit::set_done(mem_fetch *mf) { void memory_partition_unit::set_dram_power_stats( unsigned &n_cmd, unsigned &n_activity, unsigned &n_nop, unsigned &n_act, - unsigned &n_pre, unsigned &n_rd, unsigned &n_wr, unsigned &n_wr_WB, unsigned &n_req) const { + unsigned &n_pre, unsigned &n_rd, unsigned &n_wr, unsigned &n_wr_WB, + unsigned &n_req) const { m_dram->set_dram_power_stats(n_cmd, n_activity, n_nop, n_act, n_pre, n_rd, n_wr, n_wr_WB, n_req); } diff --git a/src/gpgpu-sim/l2cache.h b/src/gpgpu-sim/l2cache.h index 7fa1f29..ccf9b70 100644 --- a/src/gpgpu-sim/l2cache.h +++ b/src/gpgpu-sim/l2cache.h @@ -1,17 +1,18 @@ // Copyright (c) 2009-2021, Tor M. Aamodt, Vijay Kandiah, Nikos Hardavellas, // Mahmoud Khairy, Junrui Pan, Timothy G. Rogers -// The University of British Columbia, Northwestern University, Purdue University -// All rights reserved. +// The University of British Columbia, Northwestern University, Purdue +// University All rights reserved. // // Redistribution and use in source and binary forms, with or without // modification, are permitted provided that the following conditions are met: // -// 1. Redistributions of source code must retain the above copyright notice, this +// 1. Redistributions of source code must retain the above copyright notice, +// this // list of conditions and the following disclaimer; // 2. Redistributions in binary form must reproduce the above copyright notice, // this list of conditions and the following disclaimer in the documentation // and/or other materials provided with the distribution; -// 3. Neither the names of The University of British Columbia, Northwestern +// 3. Neither the names of The University of British Columbia, Northwestern // University nor the names of their contributors may be used to // endorse or promote products derived from this software without specific // prior written permission. diff --git a/src/gpgpu-sim/local_interconnect.cc b/src/gpgpu-sim/local_interconnect.cc index fe7bc74..7e1ab5b 100644 --- a/src/gpgpu-sim/local_interconnect.cc +++ b/src/gpgpu-sim/local_interconnect.cc @@ -148,7 +148,7 @@ void xbar_router::RR_Advance() { } } } - next_node_id = next_node_id + 1 ; + next_node_id = next_node_id + 1; next_node_id = (next_node_id % total_nodes); conflicts += conflict_sub; diff --git a/src/gpgpu-sim/power_interface.cc b/src/gpgpu-sim/power_interface.cc index 45a09bc..cddb6e9 100644 --- a/src/gpgpu-sim/power_interface.cc +++ b/src/gpgpu-sim/power_interface.cc @@ -1,17 +1,18 @@ -// Copyright (c) 2009-2021, Tor M. Aamodt, Ahmed El-Shafiey, Tayler Hetherington, Vijay Kandiah, Nikos Hardavellas, -// Mahmoud Khairy, Junrui Pan, Timothy G. Rogers -// The University of British Columbia, Northwestern University, Purdue University -// All rights reserved. +// Copyright (c) 2009-2021, Tor M. Aamodt, Ahmed El-Shafiey, Tayler +// Hetherington, Vijay Kandiah, Nikos Hardavellas, Mahmoud Khairy, Junrui Pan, +// Timothy G. Rogers The University of British Columbia, Northwestern +// University, Purdue University All rights reserved. // // Redistribution and use in source and binary forms, with or without // modification, are permitted provided that the following conditions are met: // -// 1. Redistributions of source code must retain the above copyright notice, this +// 1. Redistributions of source code must retain the above copyright notice, +// this // list of conditions and the following disclaimer; // 2. Redistributions in binary form must reproduce the above copyright notice, // this list of conditions and the following disclaimer in the documentation // and/or other materials provided with the distribution; -// 3. Neither the names of The University of British Columbia, Northwestern +// 3. Neither the names of The University of British Columbia, Northwestern // University nor the names of their contributors may be used to // endorse or promote products derived from this software without specific // prior written permission. @@ -28,10 +29,8 @@ // ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE // POSSIBILITY OF SUCH DAMAGE. - #include "power_interface.h" - void init_mcpat(const gpgpu_sim_config &config, class gpgpu_sim_wrapper *wrapper, unsigned stat_sample_freq, unsigned tot_inst, unsigned inst) { @@ -42,11 +41,9 @@ void init_mcpat(const gpgpu_sim_config &config, config.g_power_simulation_enabled, config.g_power_trace_enabled, config.g_steady_power_levels_enabled, config.g_power_per_cycle_dump, config.gpu_steady_power_deviation, config.gpu_steady_min_period, - config.g_power_trace_zlevel, tot_inst + inst, stat_sample_freq, - config.g_power_simulation_mode, - config.g_dvfs_enabled, - config.get_core_freq()/1000000, - config.num_shader()); + config.g_power_trace_zlevel, tot_inst + inst, stat_sample_freq, + config.g_power_simulation_mode, config.g_dvfs_enabled, + config.get_core_freq() / 1000000, config.num_shader()); } void mcpat_cycle(const gpgpu_sim_config &config, @@ -63,8 +60,9 @@ void mcpat_cycle(const gpgpu_sim_config &config, } if ((tot_cycle + cycle) % stat_sample_freq == 0) { - if(dvfs_enabled){ - wrapper->set_model_voltage(1); //performance model needs to support this. + if (dvfs_enabled) { + wrapper->set_model_voltage(1); // performance model needs to support + // this. } wrapper->set_inst_power( @@ -85,14 +83,17 @@ void mcpat_cycle(const gpgpu_sim_config &config, power_stats->get_inst_c_misses(0)); // Constant Cache, shared memory, texture cache - wrapper->set_ccache_power(power_stats->get_const_accessess(0), 0); //assuming all HITS in constant cache for now + wrapper->set_ccache_power( + power_stats->get_const_accessess(0), + 0); // assuming all HITS in constant cache for now wrapper->set_tcache_power(power_stats->get_texture_c_hits(), power_stats->get_texture_c_misses()); wrapper->set_shrd_mem_power(power_stats->get_shmem_access(0)); - wrapper->set_l1cache_power( - power_stats->get_l1d_read_hits(0), power_stats->get_l1d_read_misses(0), - power_stats->get_l1d_write_hits(0), power_stats->get_l1d_write_misses(0)); + wrapper->set_l1cache_power(power_stats->get_l1d_read_hits(0), + power_stats->get_l1d_read_misses(0), + power_stats->get_l1d_write_hits(0), + power_stats->get_l1d_write_misses(0)); wrapper->set_l2cache_power( power_stats->get_l2_read_hits(0), power_stats->get_l2_read_misses(0), @@ -120,24 +121,23 @@ void mcpat_cycle(const gpgpu_sim_config &config, // Execution pipeline accesses // FPU (SP) accesses, Integer ALU (not present in Tesla), Sfu accesses - wrapper->set_int_accesses(power_stats->get_ialu_accessess(0), - power_stats->get_intmul24_accessess(0), - power_stats->get_intmul32_accessess(0), - power_stats->get_intmul_accessess(0), + wrapper->set_int_accesses(power_stats->get_ialu_accessess(0), + power_stats->get_intmul24_accessess(0), + power_stats->get_intmul32_accessess(0), + power_stats->get_intmul_accessess(0), power_stats->get_intdiv_accessess(0)); - wrapper->set_dp_accesses(power_stats->get_dp_accessess(0), - power_stats->get_dpmul_accessess(0), - power_stats->get_dpdiv_accessess(0)); + wrapper->set_dp_accesses(power_stats->get_dp_accessess(0), + power_stats->get_dpmul_accessess(0), + power_stats->get_dpdiv_accessess(0)); - wrapper->set_fp_accesses(power_stats->get_fp_accessess(0), - power_stats->get_fpmul_accessess(0), - power_stats->get_fpdiv_accessess(0)); + wrapper->set_fp_accesses(power_stats->get_fp_accessess(0), + power_stats->get_fpmul_accessess(0), + power_stats->get_fpdiv_accessess(0)); - wrapper->set_trans_accesses(power_stats->get_sqrt_accessess(0), - power_stats->get_log_accessess(0), - power_stats->get_sin_accessess(0), - power_stats->get_exp_accessess(0)); + wrapper->set_trans_accesses( + power_stats->get_sqrt_accessess(0), power_stats->get_log_accessess(0), + power_stats->get_sin_accessess(0), power_stats->get_exp_accessess(0)); wrapper->set_tensor_accesses(power_stats->get_tensor_accessess(0)); @@ -154,23 +154,21 @@ void mcpat_cycle(const gpgpu_sim_config &config, (power_stats->get_sp_active_lanes()) / stat_sample_freq; float avg_sfu_active_lanes = (power_stats->get_sfu_active_lanes()) / stat_sample_freq; - if(avg_sp_active_lanes >32.0 ) - avg_sp_active_lanes = 32.0; - if(avg_sfu_active_lanes >32.0 ) - avg_sfu_active_lanes = 32.0; + if (avg_sp_active_lanes > 32.0) avg_sp_active_lanes = 32.0; + if (avg_sfu_active_lanes > 32.0) avg_sfu_active_lanes = 32.0; assert(avg_sp_active_lanes <= 32); assert(avg_sfu_active_lanes <= 32); wrapper->set_active_lanes_power(avg_sp_active_lanes, avg_sfu_active_lanes); - double n_icnt_simt_to_mem = - (double) - power_stats->get_icnt_simt_to_mem(0); // # flits from SIMT clusters - // to memory partitions - double n_icnt_mem_to_simt = - (double) - power_stats->get_icnt_mem_to_simt(0); // # flits from memory - // partitions to SIMT clusters - wrapper->set_NoC_power(n_icnt_mem_to_simt + n_icnt_simt_to_mem); // Number of flits traversing the interconnect + double n_icnt_simt_to_mem = (double)power_stats->get_icnt_simt_to_mem( + 0); // # flits from SIMT clusters + // to memory partitions + double n_icnt_mem_to_simt = (double)power_stats->get_icnt_mem_to_simt( + 0); // # flits from memory + // partitions to SIMT clusters + wrapper->set_NoC_power( + n_icnt_mem_to_simt + + n_icnt_simt_to_mem); // Number of flits traversing the interconnect wrapper->compute(); @@ -191,68 +189,77 @@ void mcpat_reset_perf_count(class gpgpu_sim_wrapper *wrapper) { wrapper->reset_counters(); } -bool parse_hw_file(char* hwpowerfile, bool find_target_kernel, vector &hw_data, char* benchname, std::string executed_kernelname){ +bool parse_hw_file(char *hwpowerfile, bool find_target_kernel, + vector &hw_data, char *benchname, + std::string executed_kernelname) { fstream hw_file; hw_file.open(hwpowerfile, ios::in); string line, word, temp; - while(!hw_file.eof()){ + while (!hw_file.eof()) { hw_data.clear(); getline(hw_file, line); stringstream s(line); - while (getline(s,word,',')){ + while (getline(s, word, ',')) { hw_data.push_back(word); } - if(hw_data[HW_BENCH_NAME] == std::string(benchname)){ - if(find_target_kernel){ - if(hw_data[HW_KERNEL_NAME] == ""){ + if (hw_data[HW_BENCH_NAME] == std::string(benchname)) { + if (find_target_kernel) { + if (hw_data[HW_KERNEL_NAME] == "") { hw_file.close(); return true; - } - else{ - if(hw_data[HW_KERNEL_NAME] == executed_kernelname){ + } else { + if (hw_data[HW_KERNEL_NAME] == executed_kernelname) { hw_file.close(); return true; } } - } - else{ + } else { hw_file.close(); return true; } - } + } } hw_file.close(); return false; } - -void calculate_hw_mcpat(const gpgpu_sim_config &config, - const shader_core_config *shdr_config, - class gpgpu_sim_wrapper *wrapper, - class power_stat_t *power_stats, unsigned stat_sample_freq, - unsigned tot_cycle, unsigned cycle, unsigned tot_inst, - unsigned inst, int power_simulation_mode, bool dvfs_enabled, char* hwpowerfile, - char* benchname, std::string executed_kernelname, - const bool *accelwattch_hybrid_configuration, bool aggregate_power_stats){ - +void calculate_hw_mcpat( + const gpgpu_sim_config &config, const shader_core_config *shdr_config, + class gpgpu_sim_wrapper *wrapper, class power_stat_t *power_stats, + unsigned stat_sample_freq, unsigned tot_cycle, unsigned cycle, + unsigned tot_inst, unsigned inst, int power_simulation_mode, + bool dvfs_enabled, char *hwpowerfile, char *benchname, + std::string executed_kernelname, + const bool *accelwattch_hybrid_configuration, bool aggregate_power_stats) { /* Reading HW data from CSV file */ vector hw_data; bool kernel_found = false; - kernel_found = parse_hw_file(hwpowerfile, true, hw_data, benchname, executed_kernelname); //Searching for matching executed_kernelname. - if(!kernel_found) - kernel_found = parse_hw_file(hwpowerfile, false, hw_data, benchname, executed_kernelname); //Searching for any kernel with same benchname. - assert("Could not find perf stats for the target benchmark in hwpowerfile.\n" && (kernel_found)); - unsigned perf_cycles = static_cast(std::stod(hw_data[HW_CYCLES]) + 0.5); - if((power_simulation_mode == 2) && (accelwattch_hybrid_configuration[HW_CYCLES])) + kernel_found = parse_hw_file( + hwpowerfile, true, hw_data, benchname, + executed_kernelname); // Searching for matching executed_kernelname. + if (!kernel_found) + kernel_found = parse_hw_file( + hwpowerfile, false, hw_data, benchname, + executed_kernelname); // Searching for any kernel with same benchname. + assert( + "Could not find perf stats for the target benchmark in hwpowerfile.\n" && + (kernel_found)); + unsigned perf_cycles = + static_cast(std::stod(hw_data[HW_CYCLES]) + 0.5); + if ((power_simulation_mode == 2) && + (accelwattch_hybrid_configuration[HW_CYCLES])) perf_cycles = cycle; - wrapper->init_mcpat_hw_mode(perf_cycles); //total PERF MODEL cycles for current kernel + wrapper->init_mcpat_hw_mode( + perf_cycles); // total PERF MODEL cycles for current kernel - if(dvfs_enabled){ - if((power_simulation_mode == 2) && (accelwattch_hybrid_configuration[HW_VOLTAGE])) - wrapper->set_model_voltage(1); //performance model needs to support this - else - wrapper->set_model_voltage(std::stod(hw_data[HW_VOLTAGE])); //performance model needs to support this + if (dvfs_enabled) { + if ((power_simulation_mode == 2) && + (accelwattch_hybrid_configuration[HW_VOLTAGE])) + wrapper->set_model_voltage(1); // performance model needs to support this + else + wrapper->set_model_voltage(std::stod( + hw_data[HW_VOLTAGE])); // performance model needs to support this } double l1_read_hits = std::stod(hw_data[HW_L1_RH]); @@ -260,266 +267,302 @@ void calculate_hw_mcpat(const gpgpu_sim_config &config, double l1_write_hits = std::stod(hw_data[HW_L1_WH]); double l1_write_misses = std::stod(hw_data[HW_L1_WM]); - if((power_simulation_mode == 2) && (accelwattch_hybrid_configuration[HW_L1_RH])) - l1_read_hits = power_stats->get_l1d_read_hits(1) - power_stats->l1r_hits_kernel; - if((power_simulation_mode == 2) && (accelwattch_hybrid_configuration[HW_L1_RM])) - l1_read_misses = power_stats->get_l1d_read_misses(1) - power_stats->l1r_misses_kernel; - if((power_simulation_mode == 2) && (accelwattch_hybrid_configuration[HW_L1_WH])) - l1_write_hits = power_stats->get_l1d_write_hits(1) - power_stats->l1w_hits_kernel; - if((power_simulation_mode == 2) && (accelwattch_hybrid_configuration[HW_L1_WM])) - l1_write_misses = power_stats->get_l1d_write_misses(1) - power_stats->l1w_misses_kernel; - - if(aggregate_power_stats){ - power_stats->tot_inst_execution += power_stats->get_total_inst(1); - power_stats->tot_int_inst_execution += power_stats->get_total_int_inst(1); - power_stats->tot_fp_inst_execution += power_stats->get_total_fp_inst(1); - power_stats->commited_inst_execution += power_stats->get_committed_inst(1); - wrapper->set_inst_power( - shdr_config->gpgpu_clock_gated_lanes, cycle, //TODO: core.[0] cycles counts don't matter, remove this + if ((power_simulation_mode == 2) && + (accelwattch_hybrid_configuration[HW_L1_RH])) + l1_read_hits = + power_stats->get_l1d_read_hits(1) - power_stats->l1r_hits_kernel; + if ((power_simulation_mode == 2) && + (accelwattch_hybrid_configuration[HW_L1_RM])) + l1_read_misses = + power_stats->get_l1d_read_misses(1) - power_stats->l1r_misses_kernel; + if ((power_simulation_mode == 2) && + (accelwattch_hybrid_configuration[HW_L1_WH])) + l1_write_hits = + power_stats->get_l1d_write_hits(1) - power_stats->l1w_hits_kernel; + if ((power_simulation_mode == 2) && + (accelwattch_hybrid_configuration[HW_L1_WM])) + l1_write_misses = + power_stats->get_l1d_write_misses(1) - power_stats->l1w_misses_kernel; + + if (aggregate_power_stats) { + power_stats->tot_inst_execution += power_stats->get_total_inst(1); + power_stats->tot_int_inst_execution += power_stats->get_total_int_inst(1); + power_stats->tot_fp_inst_execution += power_stats->get_total_fp_inst(1); + power_stats->commited_inst_execution += power_stats->get_committed_inst(1); + wrapper->set_inst_power( + shdr_config->gpgpu_clock_gated_lanes, + cycle, // TODO: core.[0] cycles counts don't matter, remove this cycle, power_stats->tot_inst_execution, power_stats->tot_int_inst_execution, power_stats->tot_fp_inst_execution, - l1_read_hits + l1_read_misses, - l1_write_hits + l1_write_misses, + l1_read_hits + l1_read_misses, l1_write_hits + l1_write_misses, power_stats->commited_inst_execution); - } - else{ - wrapper->set_inst_power( - shdr_config->gpgpu_clock_gated_lanes, cycle, //TODO: core.[0] cycles counts don't matter, remove this + } else { + wrapper->set_inst_power( + shdr_config->gpgpu_clock_gated_lanes, + cycle, // TODO: core.[0] cycles counts don't matter, remove this cycle, power_stats->get_total_inst(1), power_stats->get_total_int_inst(1), power_stats->get_total_fp_inst(1), - l1_read_hits + l1_read_misses, - l1_write_hits + l1_write_misses, + l1_read_hits + l1_read_misses, l1_write_hits + l1_write_misses, power_stats->get_committed_inst(1)); } - // Single RF for both int and fp ops -- activity factor set to 0 for Accelwattch HW and Accelwattch Hybrid because no HW Perf Stats for register files - wrapper->set_regfile_power(power_stats->get_regfile_reads(1), - power_stats->get_regfile_writes(1), - power_stats->get_non_regfile_operands(1)); + // Single RF for both int and fp ops -- activity factor set to 0 for + // Accelwattch HW and Accelwattch Hybrid because no HW Perf Stats for register + // files + wrapper->set_regfile_power(power_stats->get_regfile_reads(1), + power_stats->get_regfile_writes(1), + power_stats->get_non_regfile_operands(1)); + + // Instruction cache stats -- activity factor set to 0 for Accelwattch HW and + // Accelwattch Hybrid because no HW Perf Stats for instruction cache + wrapper->set_icache_power( + power_stats->get_inst_c_hits(1) - power_stats->l1i_hits_kernel, + power_stats->get_inst_c_misses(1) - power_stats->l1i_misses_kernel); + + // Constant Cache, shared memory, texture cache + if ((power_simulation_mode == 2) && + (accelwattch_hybrid_configuration[HW_CC_ACC])) + wrapper->set_ccache_power( + power_stats->get_const_accessess(1) - power_stats->cc_accesses_kernel, + 0); // assuming all HITS in constant cache for now + else + wrapper->set_ccache_power( + std::stod(hw_data[HW_CC_ACC]), + 0); // assuming all HITS in constant cache for now + + // wrapper->set_tcache_power(power_stats->get_texture_c_hits(), + // power_stats->get_texture_c_misses()); + + if ((power_simulation_mode == 2) && + (accelwattch_hybrid_configuration[HW_SHRD_ACC])) + wrapper->set_shrd_mem_power(power_stats->get_shmem_access(1) - + power_stats->shared_accesses_kernel); + else + wrapper->set_shrd_mem_power(std::stod(hw_data[HW_SHRD_ACC])); + + wrapper->set_l1cache_power(l1_read_hits, l1_read_misses, l1_write_hits, + l1_write_misses); + + double l2_read_hits = std::stod(hw_data[HW_L2_RH]); + double l2_read_misses = std::stod(hw_data[HW_L2_RM]); + double l2_write_hits = std::stod(hw_data[HW_L2_WH]); + double l2_write_misses = std::stod(hw_data[HW_L2_WM]); + + if ((power_simulation_mode == 2) && + (accelwattch_hybrid_configuration[HW_L2_RH])) + l2_read_hits = + power_stats->get_l2_read_hits(1) - power_stats->l2r_hits_kernel; + if ((power_simulation_mode == 2) && + (accelwattch_hybrid_configuration[HW_L2_RM])) + l2_read_misses = + power_stats->get_l2_read_misses(1) - power_stats->l2r_misses_kernel; + if ((power_simulation_mode == 2) && + (accelwattch_hybrid_configuration[HW_L2_WH])) + l2_write_hits = + power_stats->get_l2_write_hits(1) - power_stats->l2w_hits_kernel; + if ((power_simulation_mode == 2) && + (accelwattch_hybrid_configuration[HW_L2_WM])) + l2_write_misses = + power_stats->get_l2_write_misses(1) - power_stats->l2w_misses_kernel; + + wrapper->set_l2cache_power(l2_read_hits, l2_read_misses, l2_write_hits, + l2_write_misses); + + float active_sms = (*power_stats->m_active_sms) / stat_sample_freq; + float num_cores = shdr_config->num_shader(); + float num_idle_core = num_cores - active_sms; + wrapper->set_num_cores(num_cores); + if ((power_simulation_mode == 2) && + (accelwattch_hybrid_configuration[HW_NUM_SM_IDLE])) + wrapper->set_idle_core_power(num_idle_core); + else + wrapper->set_idle_core_power(std::stod(hw_data[HW_NUM_SM_IDLE])); - // Instruction cache stats -- activity factor set to 0 for Accelwattch HW and Accelwattch Hybrid because no HW Perf Stats for instruction cache - wrapper->set_icache_power(power_stats->get_inst_c_hits(1) - power_stats->l1i_hits_kernel, - power_stats->get_inst_c_misses(1) - power_stats->l1i_misses_kernel); + float pipeline_duty_cycle = + ((*power_stats->m_average_pipeline_duty_cycle / (stat_sample_freq)) < 0.8) + ? ((*power_stats->m_average_pipeline_duty_cycle) / stat_sample_freq) + : 0.8; - // Constant Cache, shared memory, texture cache - if((power_simulation_mode == 2) && (accelwattch_hybrid_configuration[HW_CC_ACC])) - wrapper->set_ccache_power(power_stats->get_const_accessess(1) - power_stats->cc_accesses_kernel, 0); //assuming all HITS in constant cache for now - else - wrapper->set_ccache_power(std::stod(hw_data[HW_CC_ACC]), 0); //assuming all HITS in constant cache for now - - - // wrapper->set_tcache_power(power_stats->get_texture_c_hits(), - // power_stats->get_texture_c_misses()); - - if((power_simulation_mode == 2) && (accelwattch_hybrid_configuration[HW_SHRD_ACC])) - wrapper->set_shrd_mem_power(power_stats->get_shmem_access(1) - power_stats->shared_accesses_kernel); - else - wrapper->set_shrd_mem_power(std::stod(hw_data[HW_SHRD_ACC])); - - wrapper->set_l1cache_power( l1_read_hits, l1_read_misses, l1_write_hits, l1_write_misses); - - double l2_read_hits = std::stod(hw_data[HW_L2_RH]); - double l2_read_misses = std::stod(hw_data[HW_L2_RM]); - double l2_write_hits = std::stod(hw_data[HW_L2_WH]); - double l2_write_misses = std::stod(hw_data[HW_L2_WM]); - - if((power_simulation_mode == 2) && (accelwattch_hybrid_configuration[HW_L2_RH])) - l2_read_hits = power_stats->get_l2_read_hits(1) - power_stats->l2r_hits_kernel; - if((power_simulation_mode == 2) && (accelwattch_hybrid_configuration[HW_L2_RM])) - l2_read_misses = power_stats->get_l2_read_misses(1) - power_stats->l2r_misses_kernel; - if((power_simulation_mode == 2) && (accelwattch_hybrid_configuration[HW_L2_WH])) - l2_write_hits = power_stats->get_l2_write_hits(1) - power_stats->l2w_hits_kernel; - if((power_simulation_mode == 2) && (accelwattch_hybrid_configuration[HW_L2_WM])) - l2_write_misses = power_stats->get_l2_write_misses(1) - power_stats->l2w_misses_kernel; - - wrapper->set_l2cache_power(l2_read_hits, l2_read_misses, l2_write_hits, l2_write_misses); - - float active_sms = (*power_stats->m_active_sms) / stat_sample_freq; - float num_cores = shdr_config->num_shader(); - float num_idle_core = num_cores - active_sms; - wrapper->set_num_cores(num_cores); - if((power_simulation_mode == 2) && (accelwattch_hybrid_configuration[HW_NUM_SM_IDLE])) - wrapper->set_idle_core_power(num_idle_core); - else - wrapper->set_idle_core_power(std::stod(hw_data[HW_NUM_SM_IDLE])); + if ((power_simulation_mode == 2) && + (accelwattch_hybrid_configuration[HW_PIPE_DUTY])) + wrapper->set_duty_cycle_power(pipeline_duty_cycle); + else + wrapper->set_duty_cycle_power(std::stod(hw_data[HW_PIPE_DUTY])); + + // Memory Controller + + double dram_reads = std::stod(hw_data[HW_DRAM_RD]); + double dram_writes = std::stod(hw_data[HW_DRAM_WR]); + double dram_pre = 0; + if ((power_simulation_mode == 2) && + (accelwattch_hybrid_configuration[HW_DRAM_RD])) + dram_reads = power_stats->get_dram_rd(1) - power_stats->dram_rd_kernel; + if ((power_simulation_mode == 2) && + (accelwattch_hybrid_configuration[HW_DRAM_WR])) + dram_writes = power_stats->get_dram_wr(1) - power_stats->dram_wr_kernel; + if ((power_simulation_mode == 2) && + (accelwattch_hybrid_configuration[HW_DRAM_RD])) + dram_pre = power_stats->get_dram_pre(1) - power_stats->dram_pre_kernel; + + wrapper->set_mem_ctrl_power(dram_reads, dram_writes, dram_pre); + + if (aggregate_power_stats) { + power_stats->ialu_acc_execution += power_stats->get_ialu_accessess(1); + power_stats->imul24_acc_execution += power_stats->get_intmul24_accessess(1); + power_stats->imul32_acc_execution += power_stats->get_intmul32_accessess(1); + power_stats->imul_acc_execution += power_stats->get_intmul_accessess(1); + power_stats->idiv_acc_execution += power_stats->get_intdiv_accessess(1); + power_stats->dp_acc_execution += power_stats->get_dp_accessess(1); + power_stats->dpmul_acc_execution += power_stats->get_dpmul_accessess(1); + power_stats->dpdiv_acc_execution += power_stats->get_dpdiv_accessess(1); + power_stats->fp_acc_execution += power_stats->get_fp_accessess(1); + power_stats->fpmul_acc_execution += power_stats->get_fpmul_accessess(1); + power_stats->fpdiv_acc_execution += power_stats->get_fpdiv_accessess(1); + power_stats->sqrt_acc_execution += power_stats->get_sqrt_accessess(1); + power_stats->log_acc_execution += power_stats->get_log_accessess(1); + power_stats->sin_acc_execution += power_stats->get_sin_accessess(1); + power_stats->exp_acc_execution += power_stats->get_exp_accessess(1); + power_stats->tensor_acc_execution += power_stats->get_tensor_accessess(1); + power_stats->tex_acc_execution += power_stats->get_tex_accessess(1); + power_stats->tot_fpu_acc_execution += power_stats->get_tot_fpu_accessess(1); + power_stats->tot_sfu_acc_execution += power_stats->get_tot_sfu_accessess(1); + power_stats->tot_threads_acc_execution += + power_stats->get_tot_threads_kernel(1); + power_stats->tot_warps_acc_execution += + power_stats->get_tot_warps_kernel(1); + + power_stats->sp_active_lanes_execution += + (power_stats->get_sp_active_lanes() * shdr_config->num_shader() * + shdr_config->gpgpu_num_sp_units); + power_stats->sfu_active_lanes_execution += + (power_stats->get_sfu_active_lanes() * shdr_config->num_shader() * + shdr_config->gpgpu_num_sp_units); + + wrapper->set_int_accesses( + power_stats->ialu_acc_execution, power_stats->imul24_acc_execution, + power_stats->imul32_acc_execution, power_stats->imul_acc_execution, + power_stats->idiv_acc_execution); + + wrapper->set_dp_accesses(power_stats->dp_acc_execution, + power_stats->dpmul_acc_execution, + power_stats->dpdiv_acc_execution); + + wrapper->set_fp_accesses(power_stats->fp_acc_execution, + power_stats->fpmul_acc_execution, + power_stats->fpdiv_acc_execution); + + wrapper->set_trans_accesses( + power_stats->sqrt_acc_execution, power_stats->log_acc_execution, + power_stats->sin_acc_execution, power_stats->exp_acc_execution); + + wrapper->set_tensor_accesses(power_stats->tensor_acc_execution); + + wrapper->set_tex_accesses(power_stats->tex_acc_execution); + + wrapper->set_exec_unit_power(power_stats->ialu_acc_execution, + power_stats->tot_fpu_acc_execution, + power_stats->tot_sfu_acc_execution); + + wrapper->set_avg_active_threads( + (double)((double)power_stats->tot_threads_acc_execution / + (double)power_stats->tot_warps_acc_execution)); - float pipeline_duty_cycle = - ((*power_stats->m_average_pipeline_duty_cycle / (stat_sample_freq)) < - 0.8) - ? ((*power_stats->m_average_pipeline_duty_cycle) / stat_sample_freq) - : 0.8; - - if((power_simulation_mode == 2) && (accelwattch_hybrid_configuration[HW_PIPE_DUTY])) - wrapper->set_duty_cycle_power(pipeline_duty_cycle); - else - wrapper->set_duty_cycle_power(std::stod(hw_data[HW_PIPE_DUTY])); + // Average active lanes for sp and sfu pipelines + float avg_sp_active_lanes = + (power_stats->sp_active_lanes_execution) / shdr_config->num_shader() / + shdr_config->gpgpu_num_sp_units / stat_sample_freq; + float avg_sfu_active_lanes = + (power_stats->sfu_active_lanes_execution) / shdr_config->num_shader() / + shdr_config->gpgpu_num_sp_units / stat_sample_freq; + if (avg_sp_active_lanes > 32.0) avg_sp_active_lanes = 32.0; + if (avg_sfu_active_lanes > 32.0) avg_sfu_active_lanes = 32.0; + assert(avg_sp_active_lanes <= 32); + assert(avg_sfu_active_lanes <= 32); + wrapper->set_active_lanes_power(avg_sp_active_lanes, avg_sfu_active_lanes); + } else { + wrapper->set_int_accesses(power_stats->get_ialu_accessess(1), + power_stats->get_intmul24_accessess(1), + power_stats->get_intmul32_accessess(1), + power_stats->get_intmul_accessess(1), + power_stats->get_intdiv_accessess(1)); - // Memory Controller - - double dram_reads = std::stod(hw_data[HW_DRAM_RD]); - double dram_writes = std::stod(hw_data[HW_DRAM_WR]); - double dram_pre = 0; - if((power_simulation_mode == 2) && (accelwattch_hybrid_configuration[HW_DRAM_RD])) - dram_reads = power_stats->get_dram_rd(1) - power_stats->dram_rd_kernel; - if((power_simulation_mode == 2) && (accelwattch_hybrid_configuration[HW_DRAM_WR])) - dram_writes = power_stats->get_dram_wr(1) - power_stats->dram_wr_kernel; - if((power_simulation_mode == 2) && (accelwattch_hybrid_configuration[HW_DRAM_RD])) - dram_pre = power_stats->get_dram_pre(1) - power_stats->dram_pre_kernel; - - - wrapper->set_mem_ctrl_power(dram_reads, dram_writes, dram_pre); - - if(aggregate_power_stats){ - power_stats->ialu_acc_execution += power_stats->get_ialu_accessess(1); - power_stats->imul24_acc_execution += power_stats->get_intmul24_accessess(1); - power_stats->imul32_acc_execution += power_stats->get_intmul32_accessess(1); - power_stats->imul_acc_execution += power_stats->get_intmul_accessess(1); - power_stats->idiv_acc_execution += power_stats->get_intdiv_accessess(1); - power_stats->dp_acc_execution += power_stats->get_dp_accessess(1); - power_stats->dpmul_acc_execution += power_stats->get_dpmul_accessess(1); - power_stats->dpdiv_acc_execution += power_stats->get_dpdiv_accessess(1); - power_stats->fp_acc_execution += power_stats->get_fp_accessess(1); - power_stats->fpmul_acc_execution += power_stats->get_fpmul_accessess(1); - power_stats->fpdiv_acc_execution += power_stats->get_fpdiv_accessess(1); - power_stats->sqrt_acc_execution += power_stats->get_sqrt_accessess(1); - power_stats->log_acc_execution += power_stats->get_log_accessess(1); - power_stats->sin_acc_execution += power_stats->get_sin_accessess(1); - power_stats->exp_acc_execution += power_stats->get_exp_accessess(1); - power_stats->tensor_acc_execution += power_stats->get_tensor_accessess(1); - power_stats->tex_acc_execution += power_stats->get_tex_accessess(1); - power_stats->tot_fpu_acc_execution += power_stats->get_tot_fpu_accessess(1); - power_stats->tot_sfu_acc_execution += power_stats->get_tot_sfu_accessess(1); - power_stats->tot_threads_acc_execution += power_stats->get_tot_threads_kernel(1); - power_stats->tot_warps_acc_execution += power_stats->get_tot_warps_kernel(1); - - power_stats->sp_active_lanes_execution += (power_stats->get_sp_active_lanes() * shdr_config->num_shader() * shdr_config->gpgpu_num_sp_units); - power_stats->sfu_active_lanes_execution += (power_stats->get_sfu_active_lanes() * shdr_config->num_shader() * shdr_config->gpgpu_num_sp_units); - - wrapper->set_int_accesses(power_stats->ialu_acc_execution, - power_stats->imul24_acc_execution, - power_stats->imul32_acc_execution, - power_stats->imul_acc_execution, - power_stats->idiv_acc_execution); - - wrapper->set_dp_accesses(power_stats->dp_acc_execution, - power_stats->dpmul_acc_execution, - power_stats->dpdiv_acc_execution); - - wrapper->set_fp_accesses(power_stats->fp_acc_execution, - power_stats->fpmul_acc_execution, - power_stats->fpdiv_acc_execution); - - wrapper->set_trans_accesses(power_stats->sqrt_acc_execution, - power_stats->log_acc_execution, - power_stats->sin_acc_execution, - power_stats->exp_acc_execution); - - wrapper->set_tensor_accesses(power_stats->tensor_acc_execution); - - wrapper->set_tex_accesses(power_stats->tex_acc_execution); - - wrapper->set_exec_unit_power(power_stats->ialu_acc_execution, - power_stats->tot_fpu_acc_execution, - power_stats->tot_sfu_acc_execution); - - wrapper->set_avg_active_threads((double)((double)power_stats->tot_threads_acc_execution / (double)power_stats->tot_warps_acc_execution)); - - // Average active lanes for sp and sfu pipelines - float avg_sp_active_lanes = - (power_stats->sp_active_lanes_execution) / shdr_config->num_shader() / shdr_config->gpgpu_num_sp_units / stat_sample_freq; - float avg_sfu_active_lanes = - (power_stats->sfu_active_lanes_execution) / shdr_config->num_shader() / shdr_config->gpgpu_num_sp_units / stat_sample_freq; - if(avg_sp_active_lanes >32.0 ) - avg_sp_active_lanes = 32.0; - if(avg_sfu_active_lanes >32.0 ) - avg_sfu_active_lanes = 32.0; - assert(avg_sp_active_lanes <= 32); - assert(avg_sfu_active_lanes <= 32); - wrapper->set_active_lanes_power(avg_sp_active_lanes, avg_sfu_active_lanes); - } - else{ - wrapper->set_int_accesses(power_stats->get_ialu_accessess(1), - power_stats->get_intmul24_accessess(1), - power_stats->get_intmul32_accessess(1), - power_stats->get_intmul_accessess(1), - power_stats->get_intdiv_accessess(1)); - - wrapper->set_dp_accesses(power_stats->get_dp_accessess(1), - power_stats->get_dpmul_accessess(1), - power_stats->get_dpdiv_accessess(1)); - - wrapper->set_fp_accesses(power_stats->get_fp_accessess(1), - power_stats->get_fpmul_accessess(1), - power_stats->get_fpdiv_accessess(1)); - - wrapper->set_trans_accesses(power_stats->get_sqrt_accessess(1), - power_stats->get_log_accessess(1), - power_stats->get_sin_accessess(1), - power_stats->get_exp_accessess(1)); - - wrapper->set_tensor_accesses(power_stats->get_tensor_accessess(1)); - - wrapper->set_tex_accesses(power_stats->get_tex_accessess(1)); - - wrapper->set_exec_unit_power(power_stats->get_tot_fpu_accessess(1), - power_stats->get_ialu_accessess(1), - power_stats->get_tot_sfu_accessess(1)); - - wrapper->set_avg_active_threads(power_stats->get_active_threads(1)); - - // Average active lanes for sp and sfu pipelines - float avg_sp_active_lanes = - (power_stats->get_sp_active_lanes()) / stat_sample_freq; - float avg_sfu_active_lanes = - (power_stats->get_sfu_active_lanes()) / stat_sample_freq; - if(avg_sp_active_lanes >32.0 ) - avg_sp_active_lanes = 32.0; - if(avg_sfu_active_lanes >32.0 ) - avg_sfu_active_lanes = 32.0; - assert(avg_sp_active_lanes <= 32); - assert(avg_sfu_active_lanes <= 32); - wrapper->set_active_lanes_power(avg_sp_active_lanes, avg_sfu_active_lanes); - } + wrapper->set_dp_accesses(power_stats->get_dp_accessess(1), + power_stats->get_dpmul_accessess(1), + power_stats->get_dpdiv_accessess(1)); - - double n_icnt_simt_to_mem = - (double) - (power_stats->get_icnt_simt_to_mem(1) - power_stats->noc_tr_kernel); // # flits from SIMT clusters - // to memory partitions - double n_icnt_mem_to_simt = - (double) - (power_stats->get_icnt_mem_to_simt(1)- power_stats->noc_rc_kernel); // # flits from memory - // partitions to SIMT clusters - if((power_simulation_mode == 2) && (accelwattch_hybrid_configuration[HW_NOC])) - wrapper->set_NoC_power(n_icnt_mem_to_simt + n_icnt_simt_to_mem); // Number of flits traversing the interconnect from Accel-Sim - else - wrapper->set_NoC_power(std::stod(hw_data[HW_NOC])); // Number of flits traversing the interconnect from HW - - wrapper->compute(); + wrapper->set_fp_accesses(power_stats->get_fp_accessess(1), + power_stats->get_fpmul_accessess(1), + power_stats->get_fpdiv_accessess(1)); - wrapper->update_components_power(); + wrapper->set_trans_accesses( + power_stats->get_sqrt_accessess(1), power_stats->get_log_accessess(1), + power_stats->get_sin_accessess(1), power_stats->get_exp_accessess(1)); - wrapper->power_metrics_calculations(); + wrapper->set_tensor_accesses(power_stats->get_tensor_accessess(1)); - wrapper->dump(); - power_stats->l1r_hits_kernel = power_stats->get_l1d_read_hits(1); - power_stats->l1r_misses_kernel = power_stats->get_l1d_read_misses(1); - power_stats->l1w_hits_kernel = power_stats->get_l1d_write_hits(1); - power_stats->l1w_misses_kernel = power_stats->get_l1d_write_misses(1); - power_stats->shared_accesses_kernel = power_stats->get_const_accessess(1); - power_stats->cc_accesses_kernel = power_stats->get_shmem_access(1); - power_stats->dram_rd_kernel = power_stats->get_dram_rd(1); - power_stats->dram_wr_kernel = power_stats->get_dram_wr(1); - power_stats->dram_pre_kernel = power_stats->get_dram_pre(1); - power_stats->l1i_hits_kernel = power_stats->get_inst_c_hits(1); - power_stats->l1i_misses_kernel = power_stats->get_inst_c_misses(1); - power_stats->l2r_hits_kernel = power_stats->get_l2_read_hits(1); - power_stats->l2r_misses_kernel = power_stats->get_l2_read_misses(1); - power_stats->l2w_hits_kernel = power_stats->get_l2_write_hits(1); - power_stats->l2w_misses_kernel = power_stats->get_l2_write_misses(1); - power_stats->noc_tr_kernel = power_stats->get_icnt_simt_to_mem(1); - power_stats->noc_rc_kernel = power_stats->get_icnt_mem_to_simt(1); - - - power_stats->clear(); + wrapper->set_tex_accesses(power_stats->get_tex_accessess(1)); + + wrapper->set_exec_unit_power(power_stats->get_tot_fpu_accessess(1), + power_stats->get_ialu_accessess(1), + power_stats->get_tot_sfu_accessess(1)); + + wrapper->set_avg_active_threads(power_stats->get_active_threads(1)); + + // Average active lanes for sp and sfu pipelines + float avg_sp_active_lanes = + (power_stats->get_sp_active_lanes()) / stat_sample_freq; + float avg_sfu_active_lanes = + (power_stats->get_sfu_active_lanes()) / stat_sample_freq; + if (avg_sp_active_lanes > 32.0) avg_sp_active_lanes = 32.0; + if (avg_sfu_active_lanes > 32.0) avg_sfu_active_lanes = 32.0; + assert(avg_sp_active_lanes <= 32); + assert(avg_sfu_active_lanes <= 32); + wrapper->set_active_lanes_power(avg_sp_active_lanes, avg_sfu_active_lanes); + } + + double n_icnt_simt_to_mem = + (double)(power_stats->get_icnt_simt_to_mem(1) - + power_stats->noc_tr_kernel); // # flits from SIMT clusters + // to memory partitions + double n_icnt_mem_to_simt = + (double)(power_stats->get_icnt_mem_to_simt(1) - + power_stats->noc_rc_kernel); // # flits from memory + // partitions to SIMT clusters + if ((power_simulation_mode == 2) && + (accelwattch_hybrid_configuration[HW_NOC])) + wrapper->set_NoC_power( + n_icnt_mem_to_simt + + n_icnt_simt_to_mem); // Number of flits traversing the interconnect + // from Accel-Sim + else + wrapper->set_NoC_power( + std::stod(hw_data[HW_NOC])); // Number of flits traversing the + // interconnect from HW + + wrapper->compute(); + + wrapper->update_components_power(); + + wrapper->power_metrics_calculations(); + + wrapper->dump(); + power_stats->l1r_hits_kernel = power_stats->get_l1d_read_hits(1); + power_stats->l1r_misses_kernel = power_stats->get_l1d_read_misses(1); + power_stats->l1w_hits_kernel = power_stats->get_l1d_write_hits(1); + power_stats->l1w_misses_kernel = power_stats->get_l1d_write_misses(1); + power_stats->shared_accesses_kernel = power_stats->get_const_accessess(1); + power_stats->cc_accesses_kernel = power_stats->get_shmem_access(1); + power_stats->dram_rd_kernel = power_stats->get_dram_rd(1); + power_stats->dram_wr_kernel = power_stats->get_dram_wr(1); + power_stats->dram_pre_kernel = power_stats->get_dram_pre(1); + power_stats->l1i_hits_kernel = power_stats->get_inst_c_hits(1); + power_stats->l1i_misses_kernel = power_stats->get_inst_c_misses(1); + power_stats->l2r_hits_kernel = power_stats->get_l2_read_hits(1); + power_stats->l2r_misses_kernel = power_stats->get_l2_read_misses(1); + power_stats->l2w_hits_kernel = power_stats->get_l2_write_hits(1); + power_stats->l2w_misses_kernel = power_stats->get_l2_write_misses(1); + power_stats->noc_tr_kernel = power_stats->get_icnt_simt_to_mem(1); + power_stats->noc_rc_kernel = power_stats->get_icnt_mem_to_simt(1); + + power_stats->clear(); } \ No newline at end of file diff --git a/src/gpgpu-sim/power_interface.h b/src/gpgpu-sim/power_interface.h index 1c6c510..3c043e6 100644 --- a/src/gpgpu-sim/power_interface.h +++ b/src/gpgpu-sim/power_interface.h @@ -1,17 +1,18 @@ -// Copyright (c) 2009-2021, Tor M. Aamodt, Ahmed El-Shafiey, Tayler Hetherington, Vijay Kandiah, Nikos Hardavellas, -// Mahmoud Khairy, Junrui Pan, Timothy G. Rogers -// The University of British Columbia, Northwestern University, Purdue University -// All rights reserved. +// Copyright (c) 2009-2021, Tor M. Aamodt, Ahmed El-Shafiey, Tayler +// Hetherington, Vijay Kandiah, Nikos Hardavellas, Mahmoud Khairy, Junrui Pan, +// Timothy G. Rogers The University of British Columbia, Northwestern +// University, Purdue University All rights reserved. // // Redistribution and use in source and binary forms, with or without // modification, are permitted provided that the following conditions are met: // -// 1. Redistributions of source code must retain the above copyright notice, this +// 1. Redistributions of source code must retain the above copyright notice, +// this // list of conditions and the following disclaimer; // 2. Redistributions in binary form must reproduce the above copyright notice, // this list of conditions and the following disclaimer in the documentation // and/or other materials provided with the distribution; -// 3. Neither the names of The University of British Columbia, Northwestern +// 3. Neither the names of The University of British Columbia, Northwestern // University nor the names of their contributors may be used to // endorse or promote products derived from this software without specific // prior written permission. @@ -47,16 +48,18 @@ void mcpat_cycle(const gpgpu_sim_config &config, unsigned tot_cycle, unsigned cycle, unsigned tot_inst, unsigned inst, bool dvfs_enabled); -void calculate_hw_mcpat(const gpgpu_sim_config &config, - const shader_core_config *shdr_config, - class gpgpu_sim_wrapper *wrapper, - class power_stat_t *power_stats, unsigned stat_sample_freq, - unsigned tot_cycle, unsigned cycle, unsigned tot_inst, - unsigned inst, int power_simulation_mode, bool dvfs_enabled, - char* hwpowerfile, char* benchname, std::string executed_kernelname, - const bool *accelwattch_hybrid_configuration, bool aggregate_power_stats); +void calculate_hw_mcpat( + const gpgpu_sim_config &config, const shader_core_config *shdr_config, + class gpgpu_sim_wrapper *wrapper, class power_stat_t *power_stats, + unsigned stat_sample_freq, unsigned tot_cycle, unsigned cycle, + unsigned tot_inst, unsigned inst, int power_simulation_mode, + bool dvfs_enabled, char *hwpowerfile, char *benchname, + std::string executed_kernelname, + const bool *accelwattch_hybrid_configuration, bool aggregate_power_stats); -bool parse_hw_file(char* hwpowerfile, bool find_target_kernel, vector &hw_data, char* benchname, std::string executed_kernelname); +bool parse_hw_file(char *hwpowerfile, bool find_target_kernel, + vector &hw_data, char *benchname, + std::string executed_kernelname); void mcpat_reset_perf_count(class gpgpu_sim_wrapper *wrapper); diff --git a/src/gpgpu-sim/power_stat.cc b/src/gpgpu-sim/power_stat.cc index d0e673c..dead4a0 100644 --- a/src/gpgpu-sim/power_stat.cc +++ b/src/gpgpu-sim/power_stat.cc @@ -1,17 +1,18 @@ -// Copyright (c) 2009-2021, Tor M. Aamodt, Ahmed El-Shafiey, Tayler Hetherington, Vijay Kandiah, Nikos Hardavellas, -// Mahmoud Khairy, Junrui Pan, Timothy G. Rogers -// The University of British Columbia, Northwestern University, Purdue University -// All rights reserved. +// Copyright (c) 2009-2021, Tor M. Aamodt, Ahmed El-Shafiey, Tayler +// Hetherington, Vijay Kandiah, Nikos Hardavellas, Mahmoud Khairy, Junrui Pan, +// Timothy G. Rogers The University of British Columbia, Northwestern +// University, Purdue University All rights reserved. // // Redistribution and use in source and binary forms, with or without // modification, are permitted provided that the following conditions are met: // -// 1. Redistributions of source code must retain the above copyright notice, this +// 1. Redistributions of source code must retain the above copyright notice, +// this // list of conditions and the following disclaimer; // 2. Redistributions in binary form must reproduce the above copyright notice, // this list of conditions and the following disclaimer in the documentation // and/or other materials provided with the distribution; -// 3. Neither the names of The University of British Columbia, Northwestern +// 3. Neither the names of The University of British Columbia, Northwestern // University nor the names of their contributors may be used to // endorse or promote products derived from this software without specific // prior written permission. @@ -56,60 +57,58 @@ power_mem_stat_t::power_mem_stat_t(const memory_config *mem_config, init(); } -void power_stat_t::clear(){ - for(unsigned i=0; i< NUM_STAT_IDX; ++i){ +void power_stat_t::clear() { + for (unsigned i = 0; i < NUM_STAT_IDX; ++i) { pwr_mem_stat->core_cache_stats[i].clear(); pwr_mem_stat->l2_cache_stats[i].clear(); - for(unsigned j=0; jnum_shader(); ++j){ - pwr_core_stat->m_pipeline_duty_cycle[i][j]=0; - pwr_core_stat->m_num_decoded_insn[i][j]=0; - pwr_core_stat->m_num_FPdecoded_insn[i][j]=0; - pwr_core_stat->m_num_INTdecoded_insn[i][j]=0; - pwr_core_stat->m_num_storequeued_insn[i][j]=0; - pwr_core_stat->m_num_loadqueued_insn[i][j]=0; - pwr_core_stat->m_num_tex_inst[i][j]=0; - pwr_core_stat->m_num_ialu_acesses[i][j]=0; - pwr_core_stat->m_num_fp_acesses[i][j]=0; - pwr_core_stat->m_num_imul_acesses[i][j]=0; - pwr_core_stat->m_num_imul24_acesses[i][j]=0; - pwr_core_stat->m_num_imul32_acesses[i][j]=0; - pwr_core_stat->m_num_fpmul_acesses[i][j]=0; - pwr_core_stat->m_num_idiv_acesses[i][j]=0; - pwr_core_stat->m_num_fpdiv_acesses[i][j]=0; - pwr_core_stat->m_num_dp_acesses[i][j]=0; - pwr_core_stat->m_num_dpmul_acesses[i][j]=0; - pwr_core_stat->m_num_dpdiv_acesses[i][j]=0; - pwr_core_stat->m_num_tensor_core_acesses[i][j]=0; - pwr_core_stat->m_num_const_acesses[i][j]=0; - pwr_core_stat->m_num_tex_acesses[i][j]=0; - pwr_core_stat->m_num_sp_acesses[i][j]=0; - pwr_core_stat->m_num_sfu_acesses[i][j]=0; - pwr_core_stat->m_num_sqrt_acesses[i][j]=0; - pwr_core_stat->m_num_log_acesses[i][j]=0; - pwr_core_stat->m_num_sin_acesses[i][j]=0; - pwr_core_stat->m_num_exp_acesses[i][j]=0; - pwr_core_stat->m_num_mem_acesses[i][j]=0; - pwr_core_stat->m_num_sp_committed[i][j]=0; - pwr_core_stat->m_num_sfu_committed[i][j]=0; - pwr_core_stat->m_num_mem_committed[i][j]=0; - pwr_core_stat->m_read_regfile_acesses[i][j]=0; - pwr_core_stat->m_write_regfile_acesses[i][j]=0; - pwr_core_stat->m_non_rf_operands[i][j]=0; - pwr_core_stat->m_active_sp_lanes[i][j]=0; - pwr_core_stat->m_active_sfu_lanes[i][j]=0; - pwr_core_stat->m_active_exu_threads[i][j]=0; - pwr_core_stat->m_active_exu_warps[i][j]=0; + for (unsigned j = 0; j < m_config->num_shader(); ++j) { + pwr_core_stat->m_pipeline_duty_cycle[i][j] = 0; + pwr_core_stat->m_num_decoded_insn[i][j] = 0; + pwr_core_stat->m_num_FPdecoded_insn[i][j] = 0; + pwr_core_stat->m_num_INTdecoded_insn[i][j] = 0; + pwr_core_stat->m_num_storequeued_insn[i][j] = 0; + pwr_core_stat->m_num_loadqueued_insn[i][j] = 0; + pwr_core_stat->m_num_tex_inst[i][j] = 0; + pwr_core_stat->m_num_ialu_acesses[i][j] = 0; + pwr_core_stat->m_num_fp_acesses[i][j] = 0; + pwr_core_stat->m_num_imul_acesses[i][j] = 0; + pwr_core_stat->m_num_imul24_acesses[i][j] = 0; + pwr_core_stat->m_num_imul32_acesses[i][j] = 0; + pwr_core_stat->m_num_fpmul_acesses[i][j] = 0; + pwr_core_stat->m_num_idiv_acesses[i][j] = 0; + pwr_core_stat->m_num_fpdiv_acesses[i][j] = 0; + pwr_core_stat->m_num_dp_acesses[i][j] = 0; + pwr_core_stat->m_num_dpmul_acesses[i][j] = 0; + pwr_core_stat->m_num_dpdiv_acesses[i][j] = 0; + pwr_core_stat->m_num_tensor_core_acesses[i][j] = 0; + pwr_core_stat->m_num_const_acesses[i][j] = 0; + pwr_core_stat->m_num_tex_acesses[i][j] = 0; + pwr_core_stat->m_num_sp_acesses[i][j] = 0; + pwr_core_stat->m_num_sfu_acesses[i][j] = 0; + pwr_core_stat->m_num_sqrt_acesses[i][j] = 0; + pwr_core_stat->m_num_log_acesses[i][j] = 0; + pwr_core_stat->m_num_sin_acesses[i][j] = 0; + pwr_core_stat->m_num_exp_acesses[i][j] = 0; + pwr_core_stat->m_num_mem_acesses[i][j] = 0; + pwr_core_stat->m_num_sp_committed[i][j] = 0; + pwr_core_stat->m_num_sfu_committed[i][j] = 0; + pwr_core_stat->m_num_mem_committed[i][j] = 0; + pwr_core_stat->m_read_regfile_acesses[i][j] = 0; + pwr_core_stat->m_write_regfile_acesses[i][j] = 0; + pwr_core_stat->m_non_rf_operands[i][j] = 0; + pwr_core_stat->m_active_sp_lanes[i][j] = 0; + pwr_core_stat->m_active_sfu_lanes[i][j] = 0; + pwr_core_stat->m_active_exu_threads[i][j] = 0; + pwr_core_stat->m_active_exu_warps[i][j] = 0; } for (unsigned j = 0; j < m_mem_config->m_n_mem; ++j) { - pwr_mem_stat->n_rd[i][j]=0; - pwr_mem_stat->n_wr[i][j]=0; - pwr_mem_stat->n_pre[i][j]=0; + pwr_mem_stat->n_rd[i][j] = 0; + pwr_mem_stat->n_wr[i][j] = 0; + pwr_mem_stat->n_pre[i][j] = 0; } } } - - void power_mem_stat_t::init() { shmem_access[CURRENT_STAT_IDX] = m_core_stats->gpgpu_n_shmem_bank_access; // Shared memory access @@ -175,7 +174,8 @@ void power_mem_stat_t::print(FILE *fout) const { unsigned total_mem_writes = 0; for (unsigned i = 0; i < m_config->m_n_mem; ++i) { total_mem_reads += n_rd[CURRENT_STAT_IDX][i]; - total_mem_writes += n_wr[CURRENT_STAT_IDX][i] + n_wr_WB[CURRENT_STAT_IDX][i]; + total_mem_writes += + n_wr[CURRENT_STAT_IDX][i] + n_wr_WB[CURRENT_STAT_IDX][i]; } fprintf(fout, "Total memory controller accesses: %u\n", total_mem_reads + total_mem_writes); @@ -205,165 +205,272 @@ void power_core_stat_t::print(FILE *fout) { // per core statistics fprintf(fout, "Power Metrics: \n"); for (unsigned i = 0; i < m_config->num_shader(); i++) { - fprintf(fout,"core %u:\n",i); - fprintf(fout,"\tpipeline duty cycle =%f\n",m_pipeline_duty_cycle[CURRENT_STAT_IDX][i]); - fprintf(fout,"\tTotal Deocded Instructions=%u\n",m_num_decoded_insn[CURRENT_STAT_IDX][i]); - fprintf(fout,"\tTotal FP Deocded Instructions=%u\n",m_num_FPdecoded_insn[CURRENT_STAT_IDX][i]); - fprintf(fout,"\tTotal INT Deocded Instructions=%u\n",m_num_INTdecoded_insn[CURRENT_STAT_IDX][i]); - fprintf(fout,"\tTotal LOAD Queued Instructions=%u\n",m_num_loadqueued_insn[CURRENT_STAT_IDX][i]); - fprintf(fout,"\tTotal STORE Queued Instructions=%u\n",m_num_storequeued_insn[CURRENT_STAT_IDX][i]); - fprintf(fout,"\tTotal IALU Acesses=%f\n",m_num_ialu_acesses[CURRENT_STAT_IDX][i]); - fprintf(fout,"\tTotal FP Acesses=%f\n",m_num_fp_acesses[CURRENT_STAT_IDX][i]); - fprintf(fout,"\tTotal DP Acesses=%f\n",m_num_dp_acesses[CURRENT_STAT_IDX][i]); - fprintf(fout,"\tTotal IMUL Acesses=%f\n",m_num_imul_acesses[CURRENT_STAT_IDX][i]); - fprintf(fout,"\tTotal IMUL24 Acesses=%f\n",m_num_imul24_acesses[CURRENT_STAT_IDX][i]); - fprintf(fout,"\tTotal IMUL32 Acesses=%f\n",m_num_imul32_acesses[CURRENT_STAT_IDX][i]); - fprintf(fout,"\tTotal IDIV Acesses=%f\n",m_num_idiv_acesses[CURRENT_STAT_IDX][i]); - fprintf(fout,"\tTotal FPMUL Acesses=%f\n",m_num_fpmul_acesses[CURRENT_STAT_IDX][i]); - fprintf(fout,"\tTotal DPMUL Acesses=%f\n",m_num_dpmul_acesses[CURRENT_STAT_IDX][i]); - fprintf(fout,"\tTotal SQRT Acesses=%f\n",m_num_sqrt_acesses[CURRENT_STAT_IDX][i]); - fprintf(fout,"\tTotal LOG Acesses=%f\n",m_num_log_acesses[CURRENT_STAT_IDX][i]); - fprintf(fout,"\tTotal SIN Acesses=%f\n",m_num_sin_acesses[CURRENT_STAT_IDX][i]); - fprintf(fout,"\tTotal EXP Acesses=%f\n",m_num_exp_acesses[CURRENT_STAT_IDX][i]); - fprintf(fout,"\tTotal FPDIV Acesses=%f\n",m_num_fpdiv_acesses[CURRENT_STAT_IDX][i]); - fprintf(fout,"\tTotal DPDIV Acesses=%f\n",m_num_dpdiv_acesses[CURRENT_STAT_IDX][i]); - fprintf(fout,"\tTotal TENSOR Acesses=%f\n",m_num_tensor_core_acesses[CURRENT_STAT_IDX][i]); - fprintf(fout,"\tTotal CONST Acesses=%f\n",m_num_const_acesses[CURRENT_STAT_IDX][i]); - fprintf(fout,"\tTotal TEX Acesses=%f\n",m_num_tex_acesses[CURRENT_STAT_IDX][i]); - fprintf(fout,"\tTotal SFU Acesses=%f\n",m_num_sfu_acesses[CURRENT_STAT_IDX][i]); - fprintf(fout,"\tTotal SP Acesses=%f\n",m_num_sp_acesses[CURRENT_STAT_IDX][i]); - fprintf(fout,"\tTotal MEM Acesses=%f\n",m_num_mem_acesses[CURRENT_STAT_IDX][i]); - fprintf(fout,"\tTotal SFU Commissions=%u\n",m_num_sfu_committed[CURRENT_STAT_IDX][i]); - fprintf(fout,"\tTotal SP Commissions=%u\n",m_num_sp_committed[CURRENT_STAT_IDX][i]); - fprintf(fout,"\tTotal MEM Commissions=%u\n",m_num_mem_committed[CURRENT_STAT_IDX][i]); - fprintf(fout,"\tTotal REG Reads=%u\n",m_read_regfile_acesses[CURRENT_STAT_IDX][i]); - fprintf(fout,"\tTotal REG Writes=%u\n",m_write_regfile_acesses[CURRENT_STAT_IDX][i]); - fprintf(fout,"\tTotal NON REG=%u\n",m_non_rf_operands[CURRENT_STAT_IDX][i]); + fprintf(fout, "core %u:\n", i); + fprintf(fout, "\tpipeline duty cycle =%f\n", + m_pipeline_duty_cycle[CURRENT_STAT_IDX][i]); + fprintf(fout, "\tTotal Deocded Instructions=%u\n", + m_num_decoded_insn[CURRENT_STAT_IDX][i]); + fprintf(fout, "\tTotal FP Deocded Instructions=%u\n", + m_num_FPdecoded_insn[CURRENT_STAT_IDX][i]); + fprintf(fout, "\tTotal INT Deocded Instructions=%u\n", + m_num_INTdecoded_insn[CURRENT_STAT_IDX][i]); + fprintf(fout, "\tTotal LOAD Queued Instructions=%u\n", + m_num_loadqueued_insn[CURRENT_STAT_IDX][i]); + fprintf(fout, "\tTotal STORE Queued Instructions=%u\n", + m_num_storequeued_insn[CURRENT_STAT_IDX][i]); + fprintf(fout, "\tTotal IALU Acesses=%f\n", + m_num_ialu_acesses[CURRENT_STAT_IDX][i]); + fprintf(fout, "\tTotal FP Acesses=%f\n", + m_num_fp_acesses[CURRENT_STAT_IDX][i]); + fprintf(fout, "\tTotal DP Acesses=%f\n", + m_num_dp_acesses[CURRENT_STAT_IDX][i]); + fprintf(fout, "\tTotal IMUL Acesses=%f\n", + m_num_imul_acesses[CURRENT_STAT_IDX][i]); + fprintf(fout, "\tTotal IMUL24 Acesses=%f\n", + m_num_imul24_acesses[CURRENT_STAT_IDX][i]); + fprintf(fout, "\tTotal IMUL32 Acesses=%f\n", + m_num_imul32_acesses[CURRENT_STAT_IDX][i]); + fprintf(fout, "\tTotal IDIV Acesses=%f\n", + m_num_idiv_acesses[CURRENT_STAT_IDX][i]); + fprintf(fout, "\tTotal FPMUL Acesses=%f\n", + m_num_fpmul_acesses[CURRENT_STAT_IDX][i]); + fprintf(fout, "\tTotal DPMUL Acesses=%f\n", + m_num_dpmul_acesses[CURRENT_STAT_IDX][i]); + fprintf(fout, "\tTotal SQRT Acesses=%f\n", + m_num_sqrt_acesses[CURRENT_STAT_IDX][i]); + fprintf(fout, "\tTotal LOG Acesses=%f\n", + m_num_log_acesses[CURRENT_STAT_IDX][i]); + fprintf(fout, "\tTotal SIN Acesses=%f\n", + m_num_sin_acesses[CURRENT_STAT_IDX][i]); + fprintf(fout, "\tTotal EXP Acesses=%f\n", + m_num_exp_acesses[CURRENT_STAT_IDX][i]); + fprintf(fout, "\tTotal FPDIV Acesses=%f\n", + m_num_fpdiv_acesses[CURRENT_STAT_IDX][i]); + fprintf(fout, "\tTotal DPDIV Acesses=%f\n", + m_num_dpdiv_acesses[CURRENT_STAT_IDX][i]); + fprintf(fout, "\tTotal TENSOR Acesses=%f\n", + m_num_tensor_core_acesses[CURRENT_STAT_IDX][i]); + fprintf(fout, "\tTotal CONST Acesses=%f\n", + m_num_const_acesses[CURRENT_STAT_IDX][i]); + fprintf(fout, "\tTotal TEX Acesses=%f\n", + m_num_tex_acesses[CURRENT_STAT_IDX][i]); + fprintf(fout, "\tTotal SFU Acesses=%f\n", + m_num_sfu_acesses[CURRENT_STAT_IDX][i]); + fprintf(fout, "\tTotal SP Acesses=%f\n", + m_num_sp_acesses[CURRENT_STAT_IDX][i]); + fprintf(fout, "\tTotal MEM Acesses=%f\n", + m_num_mem_acesses[CURRENT_STAT_IDX][i]); + fprintf(fout, "\tTotal SFU Commissions=%u\n", + m_num_sfu_committed[CURRENT_STAT_IDX][i]); + fprintf(fout, "\tTotal SP Commissions=%u\n", + m_num_sp_committed[CURRENT_STAT_IDX][i]); + fprintf(fout, "\tTotal MEM Commissions=%u\n", + m_num_mem_committed[CURRENT_STAT_IDX][i]); + fprintf(fout, "\tTotal REG Reads=%u\n", + m_read_regfile_acesses[CURRENT_STAT_IDX][i]); + fprintf(fout, "\tTotal REG Writes=%u\n", + m_write_regfile_acesses[CURRENT_STAT_IDX][i]); + fprintf(fout, "\tTotal NON REG=%u\n", + m_non_rf_operands[CURRENT_STAT_IDX][i]); } } void power_core_stat_t::init() { - m_pipeline_duty_cycle[CURRENT_STAT_IDX]=m_core_stats->m_pipeline_duty_cycle; - m_num_decoded_insn[CURRENT_STAT_IDX]=m_core_stats->m_num_decoded_insn; - m_num_FPdecoded_insn[CURRENT_STAT_IDX]=m_core_stats->m_num_FPdecoded_insn; - m_num_INTdecoded_insn[CURRENT_STAT_IDX]=m_core_stats->m_num_INTdecoded_insn; - m_num_storequeued_insn[CURRENT_STAT_IDX]=m_core_stats->m_num_storequeued_insn; - m_num_loadqueued_insn[CURRENT_STAT_IDX]=m_core_stats->m_num_loadqueued_insn; - m_num_ialu_acesses[CURRENT_STAT_IDX]=m_core_stats->m_num_ialu_acesses; - m_num_fp_acesses[CURRENT_STAT_IDX]=m_core_stats->m_num_fp_acesses; - m_num_imul_acesses[CURRENT_STAT_IDX]=m_core_stats->m_num_imul_acesses; - m_num_imul24_acesses[CURRENT_STAT_IDX]=m_core_stats->m_num_imul24_acesses; - m_num_imul32_acesses[CURRENT_STAT_IDX]=m_core_stats->m_num_imul32_acesses; - m_num_fpmul_acesses[CURRENT_STAT_IDX]=m_core_stats->m_num_fpmul_acesses; - m_num_idiv_acesses[CURRENT_STAT_IDX]=m_core_stats->m_num_idiv_acesses; - m_num_fpdiv_acesses[CURRENT_STAT_IDX]=m_core_stats->m_num_fpdiv_acesses; - m_num_dp_acesses[CURRENT_STAT_IDX]=m_core_stats->m_num_dp_acesses; - m_num_dpmul_acesses[CURRENT_STAT_IDX]=m_core_stats->m_num_dpmul_acesses; - m_num_dpdiv_acesses[CURRENT_STAT_IDX]=m_core_stats->m_num_dpdiv_acesses; - m_num_sp_acesses[CURRENT_STAT_IDX]=m_core_stats->m_num_sp_acesses; - m_num_sfu_acesses[CURRENT_STAT_IDX]=m_core_stats->m_num_sfu_acesses; - m_num_sqrt_acesses[CURRENT_STAT_IDX]=m_core_stats->m_num_sqrt_acesses; - m_num_log_acesses[CURRENT_STAT_IDX]=m_core_stats->m_num_log_acesses; - m_num_sin_acesses[CURRENT_STAT_IDX]=m_core_stats->m_num_sin_acesses; - m_num_exp_acesses[CURRENT_STAT_IDX]=m_core_stats->m_num_exp_acesses; - m_num_tensor_core_acesses[CURRENT_STAT_IDX]=m_core_stats->m_num_tensor_core_acesses; - m_num_const_acesses[CURRENT_STAT_IDX]=m_core_stats->m_num_const_acesses; - m_num_tex_acesses[CURRENT_STAT_IDX]=m_core_stats->m_num_tex_acesses; - m_num_mem_acesses[CURRENT_STAT_IDX]=m_core_stats->m_num_mem_acesses; - m_num_sp_committed[CURRENT_STAT_IDX]=m_core_stats->m_num_sp_committed; - m_num_sfu_committed[CURRENT_STAT_IDX]=m_core_stats->m_num_sfu_committed; - m_num_mem_committed[CURRENT_STAT_IDX]=m_core_stats->m_num_mem_committed; - m_read_regfile_acesses[CURRENT_STAT_IDX]=m_core_stats->m_read_regfile_acesses; - m_write_regfile_acesses[CURRENT_STAT_IDX]=m_core_stats->m_write_regfile_acesses; - m_non_rf_operands[CURRENT_STAT_IDX]=m_core_stats->m_non_rf_operands; - m_active_sp_lanes[CURRENT_STAT_IDX]=m_core_stats->m_active_sp_lanes; - m_active_sfu_lanes[CURRENT_STAT_IDX]=m_core_stats->m_active_sfu_lanes; - m_active_exu_threads[CURRENT_STAT_IDX]=m_core_stats->m_active_exu_threads; - m_active_exu_warps[CURRENT_STAT_IDX]=m_core_stats->m_active_exu_warps; - m_num_tex_inst[CURRENT_STAT_IDX]=m_core_stats->m_num_tex_inst; - - m_pipeline_duty_cycle[PREV_STAT_IDX]=(float*)calloc(m_config->num_shader(),sizeof(float)); - m_num_decoded_insn[PREV_STAT_IDX]=(unsigned *)calloc(m_config->num_shader(),sizeof(unsigned)); - m_num_FPdecoded_insn[PREV_STAT_IDX]=(unsigned *)calloc(m_config->num_shader(),sizeof(unsigned)); - m_num_INTdecoded_insn[PREV_STAT_IDX]=(unsigned *)calloc(m_config->num_shader(),sizeof(unsigned)); - m_num_storequeued_insn[PREV_STAT_IDX]=(unsigned *)calloc(m_config->num_shader(),sizeof(unsigned)); - m_num_loadqueued_insn[PREV_STAT_IDX]=(unsigned *)calloc(m_config->num_shader(),sizeof(unsigned)); - m_num_tex_inst[PREV_STAT_IDX]=(unsigned *)calloc(m_config->num_shader(),sizeof(unsigned)); - - m_num_ialu_acesses[PREV_STAT_IDX]=(double *)calloc(m_config->num_shader(),sizeof(double)); - m_num_fp_acesses[PREV_STAT_IDX]=(double *)calloc(m_config->num_shader(),sizeof(double)); - m_num_imul_acesses[PREV_STAT_IDX]=(double *)calloc(m_config->num_shader(),sizeof(double)); - m_num_imul24_acesses[PREV_STAT_IDX]=(double *)calloc(m_config->num_shader(),sizeof(double)); - m_num_imul32_acesses[PREV_STAT_IDX]=(double *)calloc(m_config->num_shader(),sizeof(double)); - m_num_fpmul_acesses[PREV_STAT_IDX]=(double *)calloc(m_config->num_shader(),sizeof(double)); - m_num_idiv_acesses[PREV_STAT_IDX]=(double *)calloc(m_config->num_shader(),sizeof(double)); - m_num_fpdiv_acesses[PREV_STAT_IDX]=(double *)calloc(m_config->num_shader(),sizeof(double)); - m_num_dp_acesses[PREV_STAT_IDX]=(double *)calloc(m_config->num_shader(),sizeof(double)); - m_num_dpmul_acesses[PREV_STAT_IDX]=(double *)calloc(m_config->num_shader(),sizeof(double)); - m_num_dpdiv_acesses[PREV_STAT_IDX]=(double *)calloc(m_config->num_shader(),sizeof(double)); - m_num_tensor_core_acesses[PREV_STAT_IDX]=(double *)calloc(m_config->num_shader(),sizeof(double)); - m_num_const_acesses[PREV_STAT_IDX]=(double *)calloc(m_config->num_shader(),sizeof(double)); - m_num_tex_acesses[PREV_STAT_IDX]=(double *)calloc(m_config->num_shader(),sizeof(double)); - m_num_sp_acesses[PREV_STAT_IDX]=(double *)calloc(m_config->num_shader(),sizeof(double)); - m_num_sfu_acesses[PREV_STAT_IDX]=(double *)calloc(m_config->num_shader(),sizeof(double)); - m_num_sqrt_acesses[PREV_STAT_IDX]=(double *)calloc(m_config->num_shader(),sizeof(double)); - m_num_log_acesses[PREV_STAT_IDX]=(double *)calloc(m_config->num_shader(),sizeof(double)); - m_num_sin_acesses[PREV_STAT_IDX]=(double *)calloc(m_config->num_shader(),sizeof(double)); - m_num_exp_acesses[PREV_STAT_IDX]=(double *)calloc(m_config->num_shader(),sizeof(double)); - m_num_mem_acesses[PREV_STAT_IDX]=(double *)calloc(m_config->num_shader(),sizeof(double)); - m_num_sp_committed[PREV_STAT_IDX]=(unsigned *)calloc(m_config->num_shader(),sizeof(unsigned)); - m_num_sfu_committed[PREV_STAT_IDX]=(unsigned *)calloc(m_config->num_shader(),sizeof(unsigned)); - m_num_mem_committed[PREV_STAT_IDX]=(unsigned *)calloc(m_config->num_shader(),sizeof(unsigned)); - m_read_regfile_acesses[PREV_STAT_IDX]=(unsigned *)calloc(m_config->num_shader(),sizeof(unsigned)); - m_write_regfile_acesses[PREV_STAT_IDX]=(unsigned *)calloc(m_config->num_shader(),sizeof(unsigned)); - m_non_rf_operands[PREV_STAT_IDX]=(unsigned *)calloc(m_config->num_shader(),sizeof(unsigned)); - m_active_sp_lanes[PREV_STAT_IDX]=(unsigned *)calloc(m_config->num_shader(),sizeof(unsigned)); - m_active_sfu_lanes[PREV_STAT_IDX]=(unsigned *)calloc(m_config->num_shader(),sizeof(unsigned)); - m_active_exu_threads[PREV_STAT_IDX]=(double *)calloc(m_config->num_shader(),sizeof(double)); - m_active_exu_warps[PREV_STAT_IDX]=(double *)calloc(m_config->num_shader(),sizeof(double)); + m_pipeline_duty_cycle[CURRENT_STAT_IDX] = m_core_stats->m_pipeline_duty_cycle; + m_num_decoded_insn[CURRENT_STAT_IDX] = m_core_stats->m_num_decoded_insn; + m_num_FPdecoded_insn[CURRENT_STAT_IDX] = m_core_stats->m_num_FPdecoded_insn; + m_num_INTdecoded_insn[CURRENT_STAT_IDX] = m_core_stats->m_num_INTdecoded_insn; + m_num_storequeued_insn[CURRENT_STAT_IDX] = + m_core_stats->m_num_storequeued_insn; + m_num_loadqueued_insn[CURRENT_STAT_IDX] = m_core_stats->m_num_loadqueued_insn; + m_num_ialu_acesses[CURRENT_STAT_IDX] = m_core_stats->m_num_ialu_acesses; + m_num_fp_acesses[CURRENT_STAT_IDX] = m_core_stats->m_num_fp_acesses; + m_num_imul_acesses[CURRENT_STAT_IDX] = m_core_stats->m_num_imul_acesses; + m_num_imul24_acesses[CURRENT_STAT_IDX] = m_core_stats->m_num_imul24_acesses; + m_num_imul32_acesses[CURRENT_STAT_IDX] = m_core_stats->m_num_imul32_acesses; + m_num_fpmul_acesses[CURRENT_STAT_IDX] = m_core_stats->m_num_fpmul_acesses; + m_num_idiv_acesses[CURRENT_STAT_IDX] = m_core_stats->m_num_idiv_acesses; + m_num_fpdiv_acesses[CURRENT_STAT_IDX] = m_core_stats->m_num_fpdiv_acesses; + m_num_dp_acesses[CURRENT_STAT_IDX] = m_core_stats->m_num_dp_acesses; + m_num_dpmul_acesses[CURRENT_STAT_IDX] = m_core_stats->m_num_dpmul_acesses; + m_num_dpdiv_acesses[CURRENT_STAT_IDX] = m_core_stats->m_num_dpdiv_acesses; + m_num_sp_acesses[CURRENT_STAT_IDX] = m_core_stats->m_num_sp_acesses; + m_num_sfu_acesses[CURRENT_STAT_IDX] = m_core_stats->m_num_sfu_acesses; + m_num_sqrt_acesses[CURRENT_STAT_IDX] = m_core_stats->m_num_sqrt_acesses; + m_num_log_acesses[CURRENT_STAT_IDX] = m_core_stats->m_num_log_acesses; + m_num_sin_acesses[CURRENT_STAT_IDX] = m_core_stats->m_num_sin_acesses; + m_num_exp_acesses[CURRENT_STAT_IDX] = m_core_stats->m_num_exp_acesses; + m_num_tensor_core_acesses[CURRENT_STAT_IDX] = + m_core_stats->m_num_tensor_core_acesses; + m_num_const_acesses[CURRENT_STAT_IDX] = m_core_stats->m_num_const_acesses; + m_num_tex_acesses[CURRENT_STAT_IDX] = m_core_stats->m_num_tex_acesses; + m_num_mem_acesses[CURRENT_STAT_IDX] = m_core_stats->m_num_mem_acesses; + m_num_sp_committed[CURRENT_STAT_IDX] = m_core_stats->m_num_sp_committed; + m_num_sfu_committed[CURRENT_STAT_IDX] = m_core_stats->m_num_sfu_committed; + m_num_mem_committed[CURRENT_STAT_IDX] = m_core_stats->m_num_mem_committed; + m_read_regfile_acesses[CURRENT_STAT_IDX] = + m_core_stats->m_read_regfile_acesses; + m_write_regfile_acesses[CURRENT_STAT_IDX] = + m_core_stats->m_write_regfile_acesses; + m_non_rf_operands[CURRENT_STAT_IDX] = m_core_stats->m_non_rf_operands; + m_active_sp_lanes[CURRENT_STAT_IDX] = m_core_stats->m_active_sp_lanes; + m_active_sfu_lanes[CURRENT_STAT_IDX] = m_core_stats->m_active_sfu_lanes; + m_active_exu_threads[CURRENT_STAT_IDX] = m_core_stats->m_active_exu_threads; + m_active_exu_warps[CURRENT_STAT_IDX] = m_core_stats->m_active_exu_warps; + m_num_tex_inst[CURRENT_STAT_IDX] = m_core_stats->m_num_tex_inst; + m_pipeline_duty_cycle[PREV_STAT_IDX] = + (float *)calloc(m_config->num_shader(), sizeof(float)); + m_num_decoded_insn[PREV_STAT_IDX] = + (unsigned *)calloc(m_config->num_shader(), sizeof(unsigned)); + m_num_FPdecoded_insn[PREV_STAT_IDX] = + (unsigned *)calloc(m_config->num_shader(), sizeof(unsigned)); + m_num_INTdecoded_insn[PREV_STAT_IDX] = + (unsigned *)calloc(m_config->num_shader(), sizeof(unsigned)); + m_num_storequeued_insn[PREV_STAT_IDX] = + (unsigned *)calloc(m_config->num_shader(), sizeof(unsigned)); + m_num_loadqueued_insn[PREV_STAT_IDX] = + (unsigned *)calloc(m_config->num_shader(), sizeof(unsigned)); + m_num_tex_inst[PREV_STAT_IDX] = + (unsigned *)calloc(m_config->num_shader(), sizeof(unsigned)); + m_num_ialu_acesses[PREV_STAT_IDX] = + (double *)calloc(m_config->num_shader(), sizeof(double)); + m_num_fp_acesses[PREV_STAT_IDX] = + (double *)calloc(m_config->num_shader(), sizeof(double)); + m_num_imul_acesses[PREV_STAT_IDX] = + (double *)calloc(m_config->num_shader(), sizeof(double)); + m_num_imul24_acesses[PREV_STAT_IDX] = + (double *)calloc(m_config->num_shader(), sizeof(double)); + m_num_imul32_acesses[PREV_STAT_IDX] = + (double *)calloc(m_config->num_shader(), sizeof(double)); + m_num_fpmul_acesses[PREV_STAT_IDX] = + (double *)calloc(m_config->num_shader(), sizeof(double)); + m_num_idiv_acesses[PREV_STAT_IDX] = + (double *)calloc(m_config->num_shader(), sizeof(double)); + m_num_fpdiv_acesses[PREV_STAT_IDX] = + (double *)calloc(m_config->num_shader(), sizeof(double)); + m_num_dp_acesses[PREV_STAT_IDX] = + (double *)calloc(m_config->num_shader(), sizeof(double)); + m_num_dpmul_acesses[PREV_STAT_IDX] = + (double *)calloc(m_config->num_shader(), sizeof(double)); + m_num_dpdiv_acesses[PREV_STAT_IDX] = + (double *)calloc(m_config->num_shader(), sizeof(double)); + m_num_tensor_core_acesses[PREV_STAT_IDX] = + (double *)calloc(m_config->num_shader(), sizeof(double)); + m_num_const_acesses[PREV_STAT_IDX] = + (double *)calloc(m_config->num_shader(), sizeof(double)); + m_num_tex_acesses[PREV_STAT_IDX] = + (double *)calloc(m_config->num_shader(), sizeof(double)); + m_num_sp_acesses[PREV_STAT_IDX] = + (double *)calloc(m_config->num_shader(), sizeof(double)); + m_num_sfu_acesses[PREV_STAT_IDX] = + (double *)calloc(m_config->num_shader(), sizeof(double)); + m_num_sqrt_acesses[PREV_STAT_IDX] = + (double *)calloc(m_config->num_shader(), sizeof(double)); + m_num_log_acesses[PREV_STAT_IDX] = + (double *)calloc(m_config->num_shader(), sizeof(double)); + m_num_sin_acesses[PREV_STAT_IDX] = + (double *)calloc(m_config->num_shader(), sizeof(double)); + m_num_exp_acesses[PREV_STAT_IDX] = + (double *)calloc(m_config->num_shader(), sizeof(double)); + m_num_mem_acesses[PREV_STAT_IDX] = + (double *)calloc(m_config->num_shader(), sizeof(double)); + m_num_sp_committed[PREV_STAT_IDX] = + (unsigned *)calloc(m_config->num_shader(), sizeof(unsigned)); + m_num_sfu_committed[PREV_STAT_IDX] = + (unsigned *)calloc(m_config->num_shader(), sizeof(unsigned)); + m_num_mem_committed[PREV_STAT_IDX] = + (unsigned *)calloc(m_config->num_shader(), sizeof(unsigned)); + m_read_regfile_acesses[PREV_STAT_IDX] = + (unsigned *)calloc(m_config->num_shader(), sizeof(unsigned)); + m_write_regfile_acesses[PREV_STAT_IDX] = + (unsigned *)calloc(m_config->num_shader(), sizeof(unsigned)); + m_non_rf_operands[PREV_STAT_IDX] = + (unsigned *)calloc(m_config->num_shader(), sizeof(unsigned)); + m_active_sp_lanes[PREV_STAT_IDX] = + (unsigned *)calloc(m_config->num_shader(), sizeof(unsigned)); + m_active_sfu_lanes[PREV_STAT_IDX] = + (unsigned *)calloc(m_config->num_shader(), sizeof(unsigned)); + m_active_exu_threads[PREV_STAT_IDX] = + (double *)calloc(m_config->num_shader(), sizeof(double)); + m_active_exu_warps[PREV_STAT_IDX] = + (double *)calloc(m_config->num_shader(), sizeof(double)); } void power_core_stat_t::save_stats() { for (unsigned i = 0; i < m_config->num_shader(); ++i) { - m_pipeline_duty_cycle[PREV_STAT_IDX][i]=m_pipeline_duty_cycle[CURRENT_STAT_IDX][i]; - m_num_decoded_insn[PREV_STAT_IDX][i]= m_num_decoded_insn[CURRENT_STAT_IDX][i]; - m_num_FPdecoded_insn[PREV_STAT_IDX][i]=m_num_FPdecoded_insn[CURRENT_STAT_IDX][i]; - m_num_INTdecoded_insn[PREV_STAT_IDX][i]=m_num_INTdecoded_insn[CURRENT_STAT_IDX][i]; - m_num_storequeued_insn[PREV_STAT_IDX][i]=m_num_storequeued_insn[CURRENT_STAT_IDX][i]; - m_num_loadqueued_insn[PREV_STAT_IDX][i]=m_num_loadqueued_insn[CURRENT_STAT_IDX][i]; - m_num_ialu_acesses[PREV_STAT_IDX][i]=m_num_ialu_acesses[CURRENT_STAT_IDX][i]; - m_num_fp_acesses[PREV_STAT_IDX][i]=m_num_fp_acesses[CURRENT_STAT_IDX][i]; - m_num_tex_inst[PREV_STAT_IDX][i]=m_num_tex_inst[CURRENT_STAT_IDX][i]; - m_num_imul_acesses[PREV_STAT_IDX][i]=m_num_imul_acesses[CURRENT_STAT_IDX][i]; - m_num_imul24_acesses[PREV_STAT_IDX][i]=m_num_imul24_acesses[CURRENT_STAT_IDX][i]; - m_num_imul32_acesses[PREV_STAT_IDX][i]=m_num_imul32_acesses[CURRENT_STAT_IDX][i]; - m_num_fpmul_acesses[PREV_STAT_IDX][i]=m_num_fpmul_acesses[CURRENT_STAT_IDX][i]; - m_num_idiv_acesses[PREV_STAT_IDX][i]=m_num_idiv_acesses[CURRENT_STAT_IDX][i]; - m_num_fpdiv_acesses[PREV_STAT_IDX][i]=m_num_fpdiv_acesses[CURRENT_STAT_IDX][i]; - m_num_sp_acesses[PREV_STAT_IDX][i]=m_num_sp_acesses[CURRENT_STAT_IDX][i]; - m_num_sfu_acesses[PREV_STAT_IDX][i]=m_num_sfu_acesses[CURRENT_STAT_IDX][i]; - m_num_sqrt_acesses[PREV_STAT_IDX][i]=m_num_sqrt_acesses[CURRENT_STAT_IDX][i]; - m_num_log_acesses[PREV_STAT_IDX][i]=m_num_log_acesses[CURRENT_STAT_IDX][i]; - m_num_sin_acesses[PREV_STAT_IDX][i]=m_num_sin_acesses[CURRENT_STAT_IDX][i]; - m_num_exp_acesses[PREV_STAT_IDX][i]=m_num_exp_acesses[CURRENT_STAT_IDX][i]; - m_num_dp_acesses[PREV_STAT_IDX][i]=m_num_dp_acesses[CURRENT_STAT_IDX][i]; - m_num_dpmul_acesses[PREV_STAT_IDX][i]=m_num_dpmul_acesses[CURRENT_STAT_IDX][i]; - m_num_dpdiv_acesses[PREV_STAT_IDX][i]=m_num_dpdiv_acesses[CURRENT_STAT_IDX][i]; - m_num_tensor_core_acesses[PREV_STAT_IDX][i]=m_num_tensor_core_acesses[CURRENT_STAT_IDX][i]; - m_num_const_acesses[PREV_STAT_IDX][i]=m_num_const_acesses[CURRENT_STAT_IDX][i]; - m_num_tex_acesses[PREV_STAT_IDX][i]=m_num_tex_acesses[CURRENT_STAT_IDX][i]; - m_num_mem_acesses[PREV_STAT_IDX][i]=m_num_mem_acesses[CURRENT_STAT_IDX][i]; - m_num_sp_committed[PREV_STAT_IDX][i]=m_num_sp_committed[CURRENT_STAT_IDX][i]; - m_num_sfu_committed[PREV_STAT_IDX][i]=m_num_sfu_committed[CURRENT_STAT_IDX][i]; - m_num_mem_committed[PREV_STAT_IDX][i]=m_num_mem_committed[CURRENT_STAT_IDX][i]; - m_read_regfile_acesses[PREV_STAT_IDX][i]=m_read_regfile_acesses[CURRENT_STAT_IDX][i]; - m_write_regfile_acesses[PREV_STAT_IDX][i]=m_write_regfile_acesses[CURRENT_STAT_IDX][i]; - m_non_rf_operands[PREV_STAT_IDX][i]=m_non_rf_operands[CURRENT_STAT_IDX][i]; - m_active_sp_lanes[PREV_STAT_IDX][i]=m_active_sp_lanes[CURRENT_STAT_IDX][i]; - m_active_sfu_lanes[PREV_STAT_IDX][i]=m_active_sfu_lanes[CURRENT_STAT_IDX][i]; - m_active_exu_threads[PREV_STAT_IDX][i]=m_active_exu_threads[CURRENT_STAT_IDX][i]; - m_active_exu_warps[PREV_STAT_IDX][i]=m_active_exu_warps[CURRENT_STAT_IDX][i]; + m_pipeline_duty_cycle[PREV_STAT_IDX][i] = + m_pipeline_duty_cycle[CURRENT_STAT_IDX][i]; + m_num_decoded_insn[PREV_STAT_IDX][i] = + m_num_decoded_insn[CURRENT_STAT_IDX][i]; + m_num_FPdecoded_insn[PREV_STAT_IDX][i] = + m_num_FPdecoded_insn[CURRENT_STAT_IDX][i]; + m_num_INTdecoded_insn[PREV_STAT_IDX][i] = + m_num_INTdecoded_insn[CURRENT_STAT_IDX][i]; + m_num_storequeued_insn[PREV_STAT_IDX][i] = + m_num_storequeued_insn[CURRENT_STAT_IDX][i]; + m_num_loadqueued_insn[PREV_STAT_IDX][i] = + m_num_loadqueued_insn[CURRENT_STAT_IDX][i]; + m_num_ialu_acesses[PREV_STAT_IDX][i] = + m_num_ialu_acesses[CURRENT_STAT_IDX][i]; + m_num_fp_acesses[PREV_STAT_IDX][i] = m_num_fp_acesses[CURRENT_STAT_IDX][i]; + m_num_tex_inst[PREV_STAT_IDX][i] = m_num_tex_inst[CURRENT_STAT_IDX][i]; + m_num_imul_acesses[PREV_STAT_IDX][i] = + m_num_imul_acesses[CURRENT_STAT_IDX][i]; + m_num_imul24_acesses[PREV_STAT_IDX][i] = + m_num_imul24_acesses[CURRENT_STAT_IDX][i]; + m_num_imul32_acesses[PREV_STAT_IDX][i] = + m_num_imul32_acesses[CURRENT_STAT_IDX][i]; + m_num_fpmul_acesses[PREV_STAT_IDX][i] = + m_num_fpmul_acesses[CURRENT_STAT_IDX][i]; + m_num_idiv_acesses[PREV_STAT_IDX][i] = + m_num_idiv_acesses[CURRENT_STAT_IDX][i]; + m_num_fpdiv_acesses[PREV_STAT_IDX][i] = + m_num_fpdiv_acesses[CURRENT_STAT_IDX][i]; + m_num_sp_acesses[PREV_STAT_IDX][i] = m_num_sp_acesses[CURRENT_STAT_IDX][i]; + m_num_sfu_acesses[PREV_STAT_IDX][i] = + m_num_sfu_acesses[CURRENT_STAT_IDX][i]; + m_num_sqrt_acesses[PREV_STAT_IDX][i] = + m_num_sqrt_acesses[CURRENT_STAT_IDX][i]; + m_num_log_acesses[PREV_STAT_IDX][i] = + m_num_log_acesses[CURRENT_STAT_IDX][i]; + m_num_sin_acesses[PREV_STAT_IDX][i] = + m_num_sin_acesses[CURRENT_STAT_IDX][i]; + m_num_exp_acesses[PREV_STAT_IDX][i] = + m_num_exp_acesses[CURRENT_STAT_IDX][i]; + m_num_dp_acesses[PREV_STAT_IDX][i] = m_num_dp_acesses[CURRENT_STAT_IDX][i]; + m_num_dpmul_acesses[PREV_STAT_IDX][i] = + m_num_dpmul_acesses[CURRENT_STAT_IDX][i]; + m_num_dpdiv_acesses[PREV_STAT_IDX][i] = + m_num_dpdiv_acesses[CURRENT_STAT_IDX][i]; + m_num_tensor_core_acesses[PREV_STAT_IDX][i] = + m_num_tensor_core_acesses[CURRENT_STAT_IDX][i]; + m_num_const_acesses[PREV_STAT_IDX][i] = + m_num_const_acesses[CURRENT_STAT_IDX][i]; + m_num_tex_acesses[PREV_STAT_IDX][i] = + m_num_tex_acesses[CURRENT_STAT_IDX][i]; + m_num_mem_acesses[PREV_STAT_IDX][i] = + m_num_mem_acesses[CURRENT_STAT_IDX][i]; + m_num_sp_committed[PREV_STAT_IDX][i] = + m_num_sp_committed[CURRENT_STAT_IDX][i]; + m_num_sfu_committed[PREV_STAT_IDX][i] = + m_num_sfu_committed[CURRENT_STAT_IDX][i]; + m_num_mem_committed[PREV_STAT_IDX][i] = + m_num_mem_committed[CURRENT_STAT_IDX][i]; + m_read_regfile_acesses[PREV_STAT_IDX][i] = + m_read_regfile_acesses[CURRENT_STAT_IDX][i]; + m_write_regfile_acesses[PREV_STAT_IDX][i] = + m_write_regfile_acesses[CURRENT_STAT_IDX][i]; + m_non_rf_operands[PREV_STAT_IDX][i] = + m_non_rf_operands[CURRENT_STAT_IDX][i]; + m_active_sp_lanes[PREV_STAT_IDX][i] = + m_active_sp_lanes[CURRENT_STAT_IDX][i]; + m_active_sfu_lanes[PREV_STAT_IDX][i] = + m_active_sfu_lanes[CURRENT_STAT_IDX][i]; + m_active_exu_threads[PREV_STAT_IDX][i] = + m_active_exu_threads[CURRENT_STAT_IDX][i]; + m_active_exu_warps[PREV_STAT_IDX][i] = + m_active_exu_warps[CURRENT_STAT_IDX][i]; } } @@ -390,12 +497,12 @@ power_stat_t::power_stat_t(const shader_core_config *shader_config, dram_rd_kernel = 0; dram_wr_kernel = 0; dram_pre_kernel = 0; - l1i_hits_kernel =0; - l1i_misses_kernel =0; - l2r_hits_kernel =0; - l2r_misses_kernel =0; - l2w_hits_kernel =0; - l2w_misses_kernel =0; + l1i_hits_kernel = 0; + l1i_misses_kernel = 0; + l2r_hits_kernel = 0; + l2r_misses_kernel = 0; + l2w_hits_kernel = 0; + l2w_misses_kernel = 0; noc_tr_kernel = 0; noc_rc_kernel = 0; diff --git a/src/gpgpu-sim/power_stat.h b/src/gpgpu-sim/power_stat.h index d40f1d9..13f144a 100644 --- a/src/gpgpu-sim/power_stat.h +++ b/src/gpgpu-sim/power_stat.h @@ -1,17 +1,18 @@ -// Copyright (c) 2009-2021, Tor M. Aamodt, Ahmed El-Shafiey, Tayler Hetherington, Vijay Kandiah, Nikos Hardavellas, -// Mahmoud Khairy, Junrui Pan, Timothy G. Rogers -// The University of British Columbia, Northwestern University, Purdue University -// All rights reserved. +// Copyright (c) 2009-2021, Tor M. Aamodt, Ahmed El-Shafiey, Tayler +// Hetherington, Vijay Kandiah, Nikos Hardavellas, Mahmoud Khairy, Junrui Pan, +// Timothy G. Rogers The University of British Columbia, Northwestern +// University, Purdue University All rights reserved. // // Redistribution and use in source and binary forms, with or without // modification, are permitted provided that the following conditions are met: // -// 1. Redistributions of source code must retain the above copyright notice, this +// 1. Redistributions of source code must retain the above copyright notice, +// this // list of conditions and the following disclaimer; // 2. Redistributions in binary form must reproduce the above copyright notice, // this list of conditions and the following disclaimer in the documentation // and/or other materials provided with the distribution; -// 3. Neither the names of The University of British Columbia, Northwestern +// 3. Neither the names of The University of British Columbia, Northwestern // University nor the names of their contributors may be used to // endorse or promote products derived from this software without specific // prior written permission. @@ -53,40 +54,40 @@ struct shader_core_power_stats_pod { unsigned *m_num_INTdecoded_insn[NUM_STAT_IDX]; // number of instructions committed // by this shader core - unsigned *m_num_storequeued_insn[NUM_STAT_IDX]; - unsigned *m_num_loadqueued_insn[NUM_STAT_IDX]; - unsigned *m_num_tex_inst[NUM_STAT_IDX]; - double *m_num_ialu_acesses[NUM_STAT_IDX]; - double *m_num_fp_acesses[NUM_STAT_IDX]; - double *m_num_imul_acesses[NUM_STAT_IDX]; - double *m_num_imul32_acesses[NUM_STAT_IDX]; - double *m_num_imul24_acesses[NUM_STAT_IDX]; - double *m_num_fpmul_acesses[NUM_STAT_IDX]; - double *m_num_idiv_acesses[NUM_STAT_IDX]; - double *m_num_fpdiv_acesses[NUM_STAT_IDX]; - double *m_num_dp_acesses[NUM_STAT_IDX]; - double *m_num_dpmul_acesses[NUM_STAT_IDX]; - double *m_num_dpdiv_acesses[NUM_STAT_IDX]; - double *m_num_sp_acesses[NUM_STAT_IDX]; - double *m_num_sfu_acesses[NUM_STAT_IDX]; - double *m_num_sqrt_acesses[NUM_STAT_IDX]; - double *m_num_log_acesses[NUM_STAT_IDX]; - double *m_num_sin_acesses[NUM_STAT_IDX]; - double *m_num_exp_acesses[NUM_STAT_IDX]; - double *m_num_tensor_core_acesses[NUM_STAT_IDX]; - double *m_num_const_acesses[NUM_STAT_IDX]; - double *m_num_tex_acesses[NUM_STAT_IDX]; - double *m_num_mem_acesses[NUM_STAT_IDX]; - unsigned *m_num_sp_committed[NUM_STAT_IDX]; - unsigned *m_num_sfu_committed[NUM_STAT_IDX]; - unsigned *m_num_mem_committed[NUM_STAT_IDX]; - unsigned *m_active_sp_lanes[NUM_STAT_IDX]; - unsigned *m_active_sfu_lanes[NUM_STAT_IDX]; - double *m_active_exu_threads[NUM_STAT_IDX]; - double *m_active_exu_warps[NUM_STAT_IDX]; - unsigned *m_read_regfile_acesses[NUM_STAT_IDX]; - unsigned *m_write_regfile_acesses[NUM_STAT_IDX]; - unsigned *m_non_rf_operands[NUM_STAT_IDX]; + unsigned *m_num_storequeued_insn[NUM_STAT_IDX]; + unsigned *m_num_loadqueued_insn[NUM_STAT_IDX]; + unsigned *m_num_tex_inst[NUM_STAT_IDX]; + double *m_num_ialu_acesses[NUM_STAT_IDX]; + double *m_num_fp_acesses[NUM_STAT_IDX]; + double *m_num_imul_acesses[NUM_STAT_IDX]; + double *m_num_imul32_acesses[NUM_STAT_IDX]; + double *m_num_imul24_acesses[NUM_STAT_IDX]; + double *m_num_fpmul_acesses[NUM_STAT_IDX]; + double *m_num_idiv_acesses[NUM_STAT_IDX]; + double *m_num_fpdiv_acesses[NUM_STAT_IDX]; + double *m_num_dp_acesses[NUM_STAT_IDX]; + double *m_num_dpmul_acesses[NUM_STAT_IDX]; + double *m_num_dpdiv_acesses[NUM_STAT_IDX]; + double *m_num_sp_acesses[NUM_STAT_IDX]; + double *m_num_sfu_acesses[NUM_STAT_IDX]; + double *m_num_sqrt_acesses[NUM_STAT_IDX]; + double *m_num_log_acesses[NUM_STAT_IDX]; + double *m_num_sin_acesses[NUM_STAT_IDX]; + double *m_num_exp_acesses[NUM_STAT_IDX]; + double *m_num_tensor_core_acesses[NUM_STAT_IDX]; + double *m_num_const_acesses[NUM_STAT_IDX]; + double *m_num_tex_acesses[NUM_STAT_IDX]; + double *m_num_mem_acesses[NUM_STAT_IDX]; + unsigned *m_num_sp_committed[NUM_STAT_IDX]; + unsigned *m_num_sfu_committed[NUM_STAT_IDX]; + unsigned *m_num_mem_committed[NUM_STAT_IDX]; + unsigned *m_active_sp_lanes[NUM_STAT_IDX]; + unsigned *m_active_sfu_lanes[NUM_STAT_IDX]; + double *m_active_exu_threads[NUM_STAT_IDX]; + double *m_active_exu_warps[NUM_STAT_IDX]; + unsigned *m_read_regfile_acesses[NUM_STAT_IDX]; + unsigned *m_write_regfile_acesses[NUM_STAT_IDX]; + unsigned *m_non_rf_operands[NUM_STAT_IDX]; }; class power_core_stat_t : public shader_core_power_stats_pod { @@ -97,7 +98,6 @@ class power_core_stat_t : public shader_core_power_stats_pod { void print(FILE *fout); void init(); void save_stats(); - private: shader_core_stats *m_core_stats; @@ -206,35 +206,37 @@ class power_stat_t { double get_total_inst(bool aggregate_stat) { double total_inst = 0; for (unsigned i = 0; i < m_config->num_shader(); i++) { - if(aggregate_stat) + if (aggregate_stat) total_inst += (pwr_core_stat->m_num_decoded_insn[CURRENT_STAT_IDX][i]); else total_inst += (pwr_core_stat->m_num_decoded_insn[CURRENT_STAT_IDX][i]) - - (pwr_core_stat->m_num_decoded_insn[PREV_STAT_IDX][i]); + (pwr_core_stat->m_num_decoded_insn[PREV_STAT_IDX][i]); } return total_inst; } double get_total_int_inst(bool aggregate_stat) { double total_inst = 0; for (unsigned i = 0; i < m_config->num_shader(); i++) { - if(aggregate_stat) - total_inst += - (pwr_core_stat->m_num_INTdecoded_insn[CURRENT_STAT_IDX][i]); - else + if (aggregate_stat) total_inst += - (pwr_core_stat->m_num_INTdecoded_insn[CURRENT_STAT_IDX][i]) - - (pwr_core_stat->m_num_INTdecoded_insn[PREV_STAT_IDX][i]); + (pwr_core_stat->m_num_INTdecoded_insn[CURRENT_STAT_IDX][i]); + else + total_inst += + (pwr_core_stat->m_num_INTdecoded_insn[CURRENT_STAT_IDX][i]) - + (pwr_core_stat->m_num_INTdecoded_insn[PREV_STAT_IDX][i]); } return total_inst; } double get_total_fp_inst(bool aggregate_stat) { double total_inst = 0; for (unsigned i = 0; i < m_config->num_shader(); i++) { - if(aggregate_stat) - total_inst += (pwr_core_stat->m_num_FPdecoded_insn[CURRENT_STAT_IDX][i]); - else - total_inst += (pwr_core_stat->m_num_FPdecoded_insn[CURRENT_STAT_IDX][i]) - - (pwr_core_stat->m_num_FPdecoded_insn[PREV_STAT_IDX][i]); + if (aggregate_stat) + total_inst += + (pwr_core_stat->m_num_FPdecoded_insn[CURRENT_STAT_IDX][i]); + else + total_inst += + (pwr_core_stat->m_num_FPdecoded_insn[CURRENT_STAT_IDX][i]) - + (pwr_core_stat->m_num_FPdecoded_insn[PREV_STAT_IDX][i]); } return total_inst; } @@ -283,43 +285,45 @@ class power_stat_t { double get_committed_inst(bool aggregate_stat) { double total_inst = 0; for (unsigned i = 0; i < m_config->num_shader(); i++) { - if(aggregate_stat) - total_inst += (pwr_core_stat->m_num_mem_committed[CURRENT_STAT_IDX][i]) + - (pwr_core_stat->m_num_sfu_committed[CURRENT_STAT_IDX][i]) + - (pwr_core_stat->m_num_sp_committed[CURRENT_STAT_IDX][i]); + if (aggregate_stat) + total_inst += + (pwr_core_stat->m_num_mem_committed[CURRENT_STAT_IDX][i]) + + (pwr_core_stat->m_num_sfu_committed[CURRENT_STAT_IDX][i]) + + (pwr_core_stat->m_num_sp_committed[CURRENT_STAT_IDX][i]); else - total_inst += (pwr_core_stat->m_num_mem_committed[CURRENT_STAT_IDX][i]) - - (pwr_core_stat->m_num_mem_committed[PREV_STAT_IDX][i]) + - (pwr_core_stat->m_num_sfu_committed[CURRENT_STAT_IDX][i]) - - (pwr_core_stat->m_num_sfu_committed[PREV_STAT_IDX][i]) + - (pwr_core_stat->m_num_sp_committed[CURRENT_STAT_IDX][i]) - - (pwr_core_stat->m_num_sp_committed[PREV_STAT_IDX][i]); + total_inst += + (pwr_core_stat->m_num_mem_committed[CURRENT_STAT_IDX][i]) - + (pwr_core_stat->m_num_mem_committed[PREV_STAT_IDX][i]) + + (pwr_core_stat->m_num_sfu_committed[CURRENT_STAT_IDX][i]) - + (pwr_core_stat->m_num_sfu_committed[PREV_STAT_IDX][i]) + + (pwr_core_stat->m_num_sp_committed[CURRENT_STAT_IDX][i]) - + (pwr_core_stat->m_num_sp_committed[PREV_STAT_IDX][i]); } return total_inst; } double get_regfile_reads(bool aggregate_stat) { double total_inst = 0; for (unsigned i = 0; i < m_config->num_shader(); i++) { - if(aggregate_stat) - total_inst += - (pwr_core_stat->m_read_regfile_acesses[CURRENT_STAT_IDX][i]); + if (aggregate_stat) + total_inst += + (pwr_core_stat->m_read_regfile_acesses[CURRENT_STAT_IDX][i]); else total_inst += - (pwr_core_stat->m_read_regfile_acesses[CURRENT_STAT_IDX][i]) - - (pwr_core_stat->m_read_regfile_acesses[PREV_STAT_IDX][i]); + (pwr_core_stat->m_read_regfile_acesses[CURRENT_STAT_IDX][i]) - + (pwr_core_stat->m_read_regfile_acesses[PREV_STAT_IDX][i]); } return total_inst; } double get_regfile_writes(bool aggregate_stat) { double total_inst = 0; for (unsigned i = 0; i < m_config->num_shader(); i++) { - if(aggregate_stat) + if (aggregate_stat) total_inst += - (pwr_core_stat->m_write_regfile_acesses[CURRENT_STAT_IDX][i]); + (pwr_core_stat->m_write_regfile_acesses[CURRENT_STAT_IDX][i]); else total_inst += - (pwr_core_stat->m_write_regfile_acesses[CURRENT_STAT_IDX][i]) - - (pwr_core_stat->m_write_regfile_acesses[PREV_STAT_IDX][i]); + (pwr_core_stat->m_write_regfile_acesses[CURRENT_STAT_IDX][i]) - + (pwr_core_stat->m_write_regfile_acesses[PREV_STAT_IDX][i]); } return total_inst; } @@ -337,11 +341,11 @@ class power_stat_t { double get_non_regfile_operands(bool aggregate_stat) { double total_inst = 0; for (unsigned i = 0; i < m_config->num_shader(); i++) { - if(aggregate_stat) - total_inst += (pwr_core_stat->m_non_rf_operands[CURRENT_STAT_IDX][i]); + if (aggregate_stat) + total_inst += (pwr_core_stat->m_non_rf_operands[CURRENT_STAT_IDX][i]); else total_inst += (pwr_core_stat->m_non_rf_operands[CURRENT_STAT_IDX][i]) - - (pwr_core_stat->m_non_rf_operands[PREV_STAT_IDX][i]); + (pwr_core_stat->m_non_rf_operands[PREV_STAT_IDX][i]); } return total_inst; } @@ -364,45 +368,49 @@ class power_stat_t { return total_inst; } - double get_sqrt_accessess(bool aggregate_stat){ - double total_inst=0; - for(unsigned i=0; inum_shader();i++){ - if(aggregate_stat) - total_inst+=(pwr_core_stat->m_num_sqrt_acesses[CURRENT_STAT_IDX][i]); - else - total_inst+=(pwr_core_stat->m_num_sqrt_acesses[CURRENT_STAT_IDX][i]) - (pwr_core_stat->m_num_sqrt_acesses[PREV_STAT_IDX][i]); - } - return total_inst; - } - double get_log_accessess(bool aggregate_stat){ - double total_inst=0; - for(unsigned i=0; inum_shader();i++){ - if(aggregate_stat) - total_inst+=(pwr_core_stat->m_num_log_acesses[CURRENT_STAT_IDX][i]); - else - total_inst+=(pwr_core_stat->m_num_log_acesses[CURRENT_STAT_IDX][i]) - (pwr_core_stat->m_num_log_acesses[PREV_STAT_IDX][i]); - } - return total_inst; - } - double get_sin_accessess(bool aggregate_stat){ - double total_inst=0; - for(unsigned i=0; inum_shader();i++){ - if(aggregate_stat) - total_inst+=(pwr_core_stat->m_num_sin_acesses[CURRENT_STAT_IDX][i]); - else - total_inst+=(pwr_core_stat->m_num_sin_acesses[CURRENT_STAT_IDX][i]) - (pwr_core_stat->m_num_sin_acesses[PREV_STAT_IDX][i]); - } - return total_inst; - } - double get_exp_accessess(bool aggregate_stat){ - double total_inst=0; - for(unsigned i=0; inum_shader();i++){ - if(aggregate_stat) - total_inst+=(pwr_core_stat->m_num_exp_acesses[CURRENT_STAT_IDX][i]); - else - total_inst+=(pwr_core_stat->m_num_exp_acesses[CURRENT_STAT_IDX][i]) - (pwr_core_stat->m_num_exp_acesses[PREV_STAT_IDX][i]); - } - return total_inst; + double get_sqrt_accessess(bool aggregate_stat) { + double total_inst = 0; + for (unsigned i = 0; i < m_config->num_shader(); i++) { + if (aggregate_stat) + total_inst += (pwr_core_stat->m_num_sqrt_acesses[CURRENT_STAT_IDX][i]); + else + total_inst += (pwr_core_stat->m_num_sqrt_acesses[CURRENT_STAT_IDX][i]) - + (pwr_core_stat->m_num_sqrt_acesses[PREV_STAT_IDX][i]); + } + return total_inst; + } + double get_log_accessess(bool aggregate_stat) { + double total_inst = 0; + for (unsigned i = 0; i < m_config->num_shader(); i++) { + if (aggregate_stat) + total_inst += (pwr_core_stat->m_num_log_acesses[CURRENT_STAT_IDX][i]); + else + total_inst += (pwr_core_stat->m_num_log_acesses[CURRENT_STAT_IDX][i]) - + (pwr_core_stat->m_num_log_acesses[PREV_STAT_IDX][i]); + } + return total_inst; + } + double get_sin_accessess(bool aggregate_stat) { + double total_inst = 0; + for (unsigned i = 0; i < m_config->num_shader(); i++) { + if (aggregate_stat) + total_inst += (pwr_core_stat->m_num_sin_acesses[CURRENT_STAT_IDX][i]); + else + total_inst += (pwr_core_stat->m_num_sin_acesses[CURRENT_STAT_IDX][i]) - + (pwr_core_stat->m_num_sin_acesses[PREV_STAT_IDX][i]); + } + return total_inst; + } + double get_exp_accessess(bool aggregate_stat) { + double total_inst = 0; + for (unsigned i = 0; i < m_config->num_shader(); i++) { + if (aggregate_stat) + total_inst += (pwr_core_stat->m_num_exp_acesses[CURRENT_STAT_IDX][i]); + else + total_inst += (pwr_core_stat->m_num_exp_acesses[CURRENT_STAT_IDX][i]) - + (pwr_core_stat->m_num_exp_acesses[PREV_STAT_IDX][i]); + } + return total_inst; } double get_mem_accessess() { @@ -417,11 +425,11 @@ class power_stat_t { double get_intdiv_accessess(bool aggregate_stat) { double total_inst = 0; for (unsigned i = 0; i < m_config->num_shader(); i++) { - if(aggregate_stat) + if (aggregate_stat) total_inst += (pwr_core_stat->m_num_idiv_acesses[CURRENT_STAT_IDX][i]); else total_inst += (pwr_core_stat->m_num_idiv_acesses[CURRENT_STAT_IDX][i]) - - (pwr_core_stat->m_num_idiv_acesses[PREV_STAT_IDX][i]); + (pwr_core_stat->m_num_idiv_acesses[PREV_STAT_IDX][i]); } return total_inst; } @@ -429,11 +437,12 @@ class power_stat_t { double get_fpdiv_accessess(bool aggregate_stat) { double total_inst = 0; for (unsigned i = 0; i < m_config->num_shader(); i++) { - if(aggregate_stat) + if (aggregate_stat) total_inst += (pwr_core_stat->m_num_fpdiv_acesses[CURRENT_STAT_IDX][i]); - else - total_inst += (pwr_core_stat->m_num_fpdiv_acesses[CURRENT_STAT_IDX][i]) - - (pwr_core_stat->m_num_fpdiv_acesses[PREV_STAT_IDX][i]); + else + total_inst += + (pwr_core_stat->m_num_fpdiv_acesses[CURRENT_STAT_IDX][i]) - + (pwr_core_stat->m_num_fpdiv_acesses[PREV_STAT_IDX][i]); } return total_inst; } @@ -441,11 +450,13 @@ class power_stat_t { double get_intmul32_accessess(bool aggregate_stat) { double total_inst = 0; for (unsigned i = 0; i < m_config->num_shader(); i++) { - if(aggregate_stat) - total_inst += (pwr_core_stat->m_num_imul32_acesses[CURRENT_STAT_IDX][i]); - else - total_inst += (pwr_core_stat->m_num_imul32_acesses[CURRENT_STAT_IDX][i]) - - (pwr_core_stat->m_num_imul32_acesses[PREV_STAT_IDX][i]); + if (aggregate_stat) + total_inst += + (pwr_core_stat->m_num_imul32_acesses[CURRENT_STAT_IDX][i]); + else + total_inst += + (pwr_core_stat->m_num_imul32_acesses[CURRENT_STAT_IDX][i]) - + (pwr_core_stat->m_num_imul32_acesses[PREV_STAT_IDX][i]); } return total_inst; } @@ -453,118 +464,126 @@ class power_stat_t { double get_intmul24_accessess(bool aggregate_stat) { double total_inst = 0; for (unsigned i = 0; i < m_config->num_shader(); i++) { - if(aggregate_stat) - total_inst += (pwr_core_stat->m_num_imul24_acesses[CURRENT_STAT_IDX][i]); - else - total_inst += (pwr_core_stat->m_num_imul24_acesses[CURRENT_STAT_IDX][i]) - - (pwr_core_stat->m_num_imul24_acesses[PREV_STAT_IDX][i]); + if (aggregate_stat) + total_inst += + (pwr_core_stat->m_num_imul24_acesses[CURRENT_STAT_IDX][i]); + else + total_inst += + (pwr_core_stat->m_num_imul24_acesses[CURRENT_STAT_IDX][i]) - + (pwr_core_stat->m_num_imul24_acesses[PREV_STAT_IDX][i]); } return total_inst; } - double get_intmul_accessess(bool aggregate_stat){ - double total_inst=0; - for(unsigned i=0; inum_shader();i++){ - if(aggregate_stat) - total_inst+= (pwr_core_stat->m_num_imul_acesses[CURRENT_STAT_IDX][i]); - else - total_inst+= (pwr_core_stat->m_num_imul_acesses[CURRENT_STAT_IDX][i]) - - (pwr_core_stat->m_num_imul_acesses[PREV_STAT_IDX][i]); - } - return total_inst; + double get_intmul_accessess(bool aggregate_stat) { + double total_inst = 0; + for (unsigned i = 0; i < m_config->num_shader(); i++) { + if (aggregate_stat) + total_inst += (pwr_core_stat->m_num_imul_acesses[CURRENT_STAT_IDX][i]); + else + total_inst += (pwr_core_stat->m_num_imul_acesses[CURRENT_STAT_IDX][i]) - + (pwr_core_stat->m_num_imul_acesses[PREV_STAT_IDX][i]); + } + return total_inst; } - double get_fpmul_accessess(bool aggregate_stat){ - double total_inst=0; - for(unsigned i=0; inum_shader();i++){ - if(aggregate_stat) - total_inst += (pwr_core_stat->m_num_fpmul_acesses[CURRENT_STAT_IDX][i]); - else - total_inst += (pwr_core_stat->m_num_fpmul_acesses[CURRENT_STAT_IDX][i]) - - (pwr_core_stat->m_num_fpmul_acesses[PREV_STAT_IDX][i]); + double get_fpmul_accessess(bool aggregate_stat) { + double total_inst = 0; + for (unsigned i = 0; i < m_config->num_shader(); i++) { + if (aggregate_stat) + total_inst += (pwr_core_stat->m_num_fpmul_acesses[CURRENT_STAT_IDX][i]); + else + total_inst += + (pwr_core_stat->m_num_fpmul_acesses[CURRENT_STAT_IDX][i]) - + (pwr_core_stat->m_num_fpmul_acesses[PREV_STAT_IDX][i]); } return total_inst; } - double get_fp_accessess(bool aggregate_stat){ - double total_inst=0; - for(unsigned i=0; inum_shader();i++){ - if(aggregate_stat) - total_inst += (pwr_core_stat->m_num_fp_acesses[CURRENT_STAT_IDX][i]); - else - total_inst += (pwr_core_stat->m_num_fp_acesses[CURRENT_STAT_IDX][i]) - + double get_fp_accessess(bool aggregate_stat) { + double total_inst = 0; + for (unsigned i = 0; i < m_config->num_shader(); i++) { + if (aggregate_stat) + total_inst += (pwr_core_stat->m_num_fp_acesses[CURRENT_STAT_IDX][i]); + else + total_inst += (pwr_core_stat->m_num_fp_acesses[CURRENT_STAT_IDX][i]) - (pwr_core_stat->m_num_fp_acesses[PREV_STAT_IDX][i]); } return total_inst; } - double get_dp_accessess(bool aggregate_stat){ - double total_inst=0; - for(unsigned i=0; inum_shader();i++){ - if(aggregate_stat) - total_inst += (pwr_core_stat->m_num_dp_acesses[CURRENT_STAT_IDX][i]); - else - total_inst += (pwr_core_stat->m_num_dp_acesses[CURRENT_STAT_IDX][i]) - + double get_dp_accessess(bool aggregate_stat) { + double total_inst = 0; + for (unsigned i = 0; i < m_config->num_shader(); i++) { + if (aggregate_stat) + total_inst += (pwr_core_stat->m_num_dp_acesses[CURRENT_STAT_IDX][i]); + else + total_inst += (pwr_core_stat->m_num_dp_acesses[CURRENT_STAT_IDX][i]) - (pwr_core_stat->m_num_dp_acesses[PREV_STAT_IDX][i]); } return total_inst; } - double get_dpmul_accessess(bool aggregate_stat){ - double total_inst=0; - for(unsigned i=0; inum_shader();i++){ - if(aggregate_stat) + double get_dpmul_accessess(bool aggregate_stat) { + double total_inst = 0; + for (unsigned i = 0; i < m_config->num_shader(); i++) { + if (aggregate_stat) total_inst += (pwr_core_stat->m_num_dpmul_acesses[CURRENT_STAT_IDX][i]); - else - total_inst += (pwr_core_stat->m_num_dpmul_acesses[CURRENT_STAT_IDX][i]) - - (pwr_core_stat->m_num_dpmul_acesses[PREV_STAT_IDX][i]); + else + total_inst += + (pwr_core_stat->m_num_dpmul_acesses[CURRENT_STAT_IDX][i]) - + (pwr_core_stat->m_num_dpmul_acesses[PREV_STAT_IDX][i]); } return total_inst; } - double get_dpdiv_accessess(bool aggregate_stat){ - double total_inst=0; - for(unsigned i=0; inum_shader();i++){ - if(aggregate_stat) + double get_dpdiv_accessess(bool aggregate_stat) { + double total_inst = 0; + for (unsigned i = 0; i < m_config->num_shader(); i++) { + if (aggregate_stat) total_inst += (pwr_core_stat->m_num_dpdiv_acesses[CURRENT_STAT_IDX][i]); - else - total_inst += (pwr_core_stat->m_num_dpdiv_acesses[CURRENT_STAT_IDX][i]) - - (pwr_core_stat->m_num_dpdiv_acesses[PREV_STAT_IDX][i]); + else + total_inst += + (pwr_core_stat->m_num_dpdiv_acesses[CURRENT_STAT_IDX][i]) - + (pwr_core_stat->m_num_dpdiv_acesses[PREV_STAT_IDX][i]); } return total_inst; } - double get_tensor_accessess(bool aggregate_stat){ - double total_inst=0; - for(unsigned i=0; inum_shader();i++){ - if(aggregate_stat) - total_inst += (pwr_core_stat->m_num_tensor_core_acesses[CURRENT_STAT_IDX][i]); - else - total_inst += (pwr_core_stat->m_num_tensor_core_acesses[CURRENT_STAT_IDX][i]) - - (pwr_core_stat->m_num_tensor_core_acesses[PREV_STAT_IDX][i]); + double get_tensor_accessess(bool aggregate_stat) { + double total_inst = 0; + for (unsigned i = 0; i < m_config->num_shader(); i++) { + if (aggregate_stat) + total_inst += + (pwr_core_stat->m_num_tensor_core_acesses[CURRENT_STAT_IDX][i]); + else + total_inst += + (pwr_core_stat->m_num_tensor_core_acesses[CURRENT_STAT_IDX][i]) - + (pwr_core_stat->m_num_tensor_core_acesses[PREV_STAT_IDX][i]); } return total_inst; } - double get_const_accessess(bool aggregate_stat){ - double total_inst=0; - for(unsigned i=0; inum_shader();i++){ - if(aggregate_stat) - total_inst += pwr_core_stat->m_num_const_acesses[CURRENT_STAT_IDX][i]; - else - total_inst += (pwr_core_stat->m_num_const_acesses[CURRENT_STAT_IDX][i]) - - (pwr_core_stat->m_num_const_acesses[PREV_STAT_IDX][i]); + double get_const_accessess(bool aggregate_stat) { + double total_inst = 0; + for (unsigned i = 0; i < m_config->num_shader(); i++) { + if (aggregate_stat) + total_inst += pwr_core_stat->m_num_const_acesses[CURRENT_STAT_IDX][i]; + else + total_inst += + (pwr_core_stat->m_num_const_acesses[CURRENT_STAT_IDX][i]) - + (pwr_core_stat->m_num_const_acesses[PREV_STAT_IDX][i]); } return (total_inst); } - double get_tex_accessess(bool aggregate_stat){ - double total_inst=0; - for(unsigned i=0; inum_shader();i++){ - if(aggregate_stat) + double get_tex_accessess(bool aggregate_stat) { + double total_inst = 0; + for (unsigned i = 0; i < m_config->num_shader(); i++) { + if (aggregate_stat) total_inst += (pwr_core_stat->m_num_tex_acesses[CURRENT_STAT_IDX][i]); - else - total_inst += (pwr_core_stat->m_num_tex_acesses[CURRENT_STAT_IDX][i]) - + else + total_inst += (pwr_core_stat->m_num_tex_acesses[CURRENT_STAT_IDX][i]) - (pwr_core_stat->m_num_tex_acesses[PREV_STAT_IDX][i]); } return total_inst; @@ -590,23 +609,24 @@ class power_stat_t { m_config->gpgpu_num_sfu_units; } - float get_active_threads(bool aggregate_stat) { unsigned total_threads = 0; unsigned total_warps = 0; for (unsigned i = 0; i < m_config->num_shader(); i++) { - if(aggregate_stat){ - total_threads += (pwr_core_stat->m_active_exu_threads[CURRENT_STAT_IDX][i]) ; + if (aggregate_stat) { + total_threads += + (pwr_core_stat->m_active_exu_threads[CURRENT_STAT_IDX][i]); total_warps += (pwr_core_stat->m_active_exu_warps[CURRENT_STAT_IDX][i]); + } else { + total_threads += + (pwr_core_stat->m_active_exu_threads[CURRENT_STAT_IDX][i]) - + (pwr_core_stat->m_active_exu_threads[PREV_STAT_IDX][i]); + total_warps += + (pwr_core_stat->m_active_exu_warps[CURRENT_STAT_IDX][i]) - + (pwr_core_stat->m_active_exu_warps[PREV_STAT_IDX][i]); } - else{ - total_threads += (pwr_core_stat->m_active_exu_threads[CURRENT_STAT_IDX][i]) - - (pwr_core_stat->m_active_exu_threads[PREV_STAT_IDX][i]); - total_warps += (pwr_core_stat->m_active_exu_warps[CURRENT_STAT_IDX][i]) - - (pwr_core_stat->m_active_exu_warps[PREV_STAT_IDX][i]); - } } - if(total_warps != 0) + if (total_warps != 0) return (float)((float)total_threads / (float)total_warps); else return 0; @@ -615,98 +635,98 @@ class power_stat_t { unsigned long long get_tot_threads_kernel(bool aggregate_stat) { unsigned total_threads = 0; for (unsigned i = 0; i < m_config->num_shader(); i++) { - if(aggregate_stat){ - total_threads += (pwr_core_stat->m_active_exu_threads[CURRENT_STAT_IDX][i]) ; + if (aggregate_stat) { + total_threads += + (pwr_core_stat->m_active_exu_threads[CURRENT_STAT_IDX][i]); + } else { + total_threads += + (pwr_core_stat->m_active_exu_threads[CURRENT_STAT_IDX][i]) - + (pwr_core_stat->m_active_exu_threads[PREV_STAT_IDX][i]); } - else{ - total_threads += (pwr_core_stat->m_active_exu_threads[CURRENT_STAT_IDX][i]) - - (pwr_core_stat->m_active_exu_threads[PREV_STAT_IDX][i]); - } } - return total_threads; + return total_threads; } unsigned long long get_tot_warps_kernel(bool aggregate_stat) { unsigned long long total_warps = 0; for (unsigned i = 0; i < m_config->num_shader(); i++) { - if(aggregate_stat){ + if (aggregate_stat) { total_warps += (pwr_core_stat->m_active_exu_warps[CURRENT_STAT_IDX][i]); + } else { + total_warps += + (pwr_core_stat->m_active_exu_warps[CURRENT_STAT_IDX][i]) - + (pwr_core_stat->m_active_exu_warps[PREV_STAT_IDX][i]); } - else{ - total_warps += (pwr_core_stat->m_active_exu_warps[CURRENT_STAT_IDX][i]) - - (pwr_core_stat->m_active_exu_warps[PREV_STAT_IDX][i]); - } } - return total_warps; + return total_warps; } - - double get_tot_fpu_accessess(bool aggregate_stat){ - double total_inst=0; - for(unsigned i=0; inum_shader();i++){ - if(aggregate_stat) - total_inst += (pwr_core_stat->m_num_fp_acesses[CURRENT_STAT_IDX][i])+ - (pwr_core_stat->m_num_dp_acesses[CURRENT_STAT_IDX][i]); + double get_tot_fpu_accessess(bool aggregate_stat) { + double total_inst = 0; + for (unsigned i = 0; i < m_config->num_shader(); i++) { + if (aggregate_stat) + total_inst += (pwr_core_stat->m_num_fp_acesses[CURRENT_STAT_IDX][i]) + + (pwr_core_stat->m_num_dp_acesses[CURRENT_STAT_IDX][i]); else - total_inst += (pwr_core_stat->m_num_fp_acesses[CURRENT_STAT_IDX][i]) - - (pwr_core_stat->m_num_fp_acesses[PREV_STAT_IDX][i]) + - (pwr_core_stat->m_num_dp_acesses[CURRENT_STAT_IDX][i]) - - (pwr_core_stat->m_num_dp_acesses[PREV_STAT_IDX][i]); + total_inst += (pwr_core_stat->m_num_fp_acesses[CURRENT_STAT_IDX][i]) - + (pwr_core_stat->m_num_fp_acesses[PREV_STAT_IDX][i]) + + (pwr_core_stat->m_num_dp_acesses[CURRENT_STAT_IDX][i]) - + (pwr_core_stat->m_num_dp_acesses[PREV_STAT_IDX][i]); } - //total_inst += get_total_load_inst()+get_total_store_inst()+get_tex_inst(); + // total_inst += + // get_total_load_inst()+get_total_store_inst()+get_tex_inst(); return total_inst; } - - - double get_tot_sfu_accessess(bool aggregate_stat){ - double total_inst=0; - for(unsigned i=0; inum_shader();i++){ - if(aggregate_stat) - total_inst += (pwr_core_stat->m_num_idiv_acesses[CURRENT_STAT_IDX][i])+ - (pwr_core_stat->m_num_imul32_acesses[CURRENT_STAT_IDX][i])+ - (pwr_core_stat->m_num_sqrt_acesses[CURRENT_STAT_IDX][i])+ - (pwr_core_stat->m_num_log_acesses[CURRENT_STAT_IDX][i])+ - (pwr_core_stat->m_num_sin_acesses[CURRENT_STAT_IDX][i])+ - (pwr_core_stat->m_num_exp_acesses[CURRENT_STAT_IDX][i])+ - (pwr_core_stat->m_num_fpdiv_acesses[CURRENT_STAT_IDX][i])+ - (pwr_core_stat->m_num_fpmul_acesses[CURRENT_STAT_IDX][i])+ - (pwr_core_stat->m_num_dpmul_acesses[CURRENT_STAT_IDX][i])+ - (pwr_core_stat->m_num_dpdiv_acesses[CURRENT_STAT_IDX][i])+ - (pwr_core_stat->m_num_imul24_acesses[CURRENT_STAT_IDX][i]) + - (pwr_core_stat->m_num_imul_acesses[CURRENT_STAT_IDX][i])+ - (pwr_core_stat->m_num_tensor_core_acesses[CURRENT_STAT_IDX][i])+ - (pwr_core_stat->m_num_tex_acesses[CURRENT_STAT_IDX][i]); - else - total_inst += (pwr_core_stat->m_num_idiv_acesses[CURRENT_STAT_IDX][i]) - - (pwr_core_stat->m_num_idiv_acesses[PREV_STAT_IDX][i]) + - (pwr_core_stat->m_num_imul32_acesses[CURRENT_STAT_IDX][i]) - - (pwr_core_stat->m_num_imul32_acesses[PREV_STAT_IDX][i]) + - (pwr_core_stat->m_num_sqrt_acesses[CURRENT_STAT_IDX][i]) - - (pwr_core_stat->m_num_sqrt_acesses[PREV_STAT_IDX][i]) + - (pwr_core_stat->m_num_log_acesses[CURRENT_STAT_IDX][i]) - - (pwr_core_stat->m_num_log_acesses[PREV_STAT_IDX][i]) + - (pwr_core_stat->m_num_sin_acesses[CURRENT_STAT_IDX][i]) - - (pwr_core_stat->m_num_sin_acesses[PREV_STAT_IDX][i]) + - (pwr_core_stat->m_num_exp_acesses[CURRENT_STAT_IDX][i]) - - (pwr_core_stat->m_num_exp_acesses[PREV_STAT_IDX][i]) + - (pwr_core_stat->m_num_fpdiv_acesses[CURRENT_STAT_IDX][i]) - - (pwr_core_stat->m_num_fpdiv_acesses[PREV_STAT_IDX][i]) + - (pwr_core_stat->m_num_fpmul_acesses[CURRENT_STAT_IDX][i]) - - (pwr_core_stat->m_num_fpmul_acesses[PREV_STAT_IDX][i]) + - (pwr_core_stat->m_num_dpmul_acesses[CURRENT_STAT_IDX][i]) - - (pwr_core_stat->m_num_dpmul_acesses[PREV_STAT_IDX][i]) + - (pwr_core_stat->m_num_dpdiv_acesses[CURRENT_STAT_IDX][i]) - - (pwr_core_stat->m_num_dpdiv_acesses[PREV_STAT_IDX][i]) + - (pwr_core_stat->m_num_imul24_acesses[CURRENT_STAT_IDX][i]) - - (pwr_core_stat->m_num_imul24_acesses[PREV_STAT_IDX][i]) + - (pwr_core_stat->m_num_imul_acesses[CURRENT_STAT_IDX][i]) - - (pwr_core_stat->m_num_imul_acesses[PREV_STAT_IDX][i]) + - (pwr_core_stat->m_num_tensor_core_acesses[CURRENT_STAT_IDX][i]) - - (pwr_core_stat->m_num_tensor_core_acesses[PREV_STAT_IDX][i]) + - (pwr_core_stat->m_num_tex_acesses[CURRENT_STAT_IDX][i]) - - (pwr_core_stat->m_num_tex_acesses[PREV_STAT_IDX][i]); - + double get_tot_sfu_accessess(bool aggregate_stat) { + double total_inst = 0; + for (unsigned i = 0; i < m_config->num_shader(); i++) { + if (aggregate_stat) + total_inst += + (pwr_core_stat->m_num_idiv_acesses[CURRENT_STAT_IDX][i]) + + (pwr_core_stat->m_num_imul32_acesses[CURRENT_STAT_IDX][i]) + + (pwr_core_stat->m_num_sqrt_acesses[CURRENT_STAT_IDX][i]) + + (pwr_core_stat->m_num_log_acesses[CURRENT_STAT_IDX][i]) + + (pwr_core_stat->m_num_sin_acesses[CURRENT_STAT_IDX][i]) + + (pwr_core_stat->m_num_exp_acesses[CURRENT_STAT_IDX][i]) + + (pwr_core_stat->m_num_fpdiv_acesses[CURRENT_STAT_IDX][i]) + + (pwr_core_stat->m_num_fpmul_acesses[CURRENT_STAT_IDX][i]) + + (pwr_core_stat->m_num_dpmul_acesses[CURRENT_STAT_IDX][i]) + + (pwr_core_stat->m_num_dpdiv_acesses[CURRENT_STAT_IDX][i]) + + (pwr_core_stat->m_num_imul24_acesses[CURRENT_STAT_IDX][i]) + + (pwr_core_stat->m_num_imul_acesses[CURRENT_STAT_IDX][i]) + + (pwr_core_stat->m_num_tensor_core_acesses[CURRENT_STAT_IDX][i]) + + (pwr_core_stat->m_num_tex_acesses[CURRENT_STAT_IDX][i]); + else + total_inst += + (pwr_core_stat->m_num_idiv_acesses[CURRENT_STAT_IDX][i]) - + (pwr_core_stat->m_num_idiv_acesses[PREV_STAT_IDX][i]) + + (pwr_core_stat->m_num_imul32_acesses[CURRENT_STAT_IDX][i]) - + (pwr_core_stat->m_num_imul32_acesses[PREV_STAT_IDX][i]) + + (pwr_core_stat->m_num_sqrt_acesses[CURRENT_STAT_IDX][i]) - + (pwr_core_stat->m_num_sqrt_acesses[PREV_STAT_IDX][i]) + + (pwr_core_stat->m_num_log_acesses[CURRENT_STAT_IDX][i]) - + (pwr_core_stat->m_num_log_acesses[PREV_STAT_IDX][i]) + + (pwr_core_stat->m_num_sin_acesses[CURRENT_STAT_IDX][i]) - + (pwr_core_stat->m_num_sin_acesses[PREV_STAT_IDX][i]) + + (pwr_core_stat->m_num_exp_acesses[CURRENT_STAT_IDX][i]) - + (pwr_core_stat->m_num_exp_acesses[PREV_STAT_IDX][i]) + + (pwr_core_stat->m_num_fpdiv_acesses[CURRENT_STAT_IDX][i]) - + (pwr_core_stat->m_num_fpdiv_acesses[PREV_STAT_IDX][i]) + + (pwr_core_stat->m_num_fpmul_acesses[CURRENT_STAT_IDX][i]) - + (pwr_core_stat->m_num_fpmul_acesses[PREV_STAT_IDX][i]) + + (pwr_core_stat->m_num_dpmul_acesses[CURRENT_STAT_IDX][i]) - + (pwr_core_stat->m_num_dpmul_acesses[PREV_STAT_IDX][i]) + + (pwr_core_stat->m_num_dpdiv_acesses[CURRENT_STAT_IDX][i]) - + (pwr_core_stat->m_num_dpdiv_acesses[PREV_STAT_IDX][i]) + + (pwr_core_stat->m_num_imul24_acesses[CURRENT_STAT_IDX][i]) - + (pwr_core_stat->m_num_imul24_acesses[PREV_STAT_IDX][i]) + + (pwr_core_stat->m_num_imul_acesses[CURRENT_STAT_IDX][i]) - + (pwr_core_stat->m_num_imul_acesses[PREV_STAT_IDX][i]) + + (pwr_core_stat->m_num_tensor_core_acesses[CURRENT_STAT_IDX][i]) - + (pwr_core_stat->m_num_tensor_core_acesses[PREV_STAT_IDX][i]) + + (pwr_core_stat->m_num_tex_acesses[CURRENT_STAT_IDX][i]) - + (pwr_core_stat->m_num_tex_acesses[PREV_STAT_IDX][i]); } return total_inst; } @@ -714,11 +734,11 @@ class power_stat_t { double get_ialu_accessess(bool aggregate_stat) { double total_inst = 0; for (unsigned i = 0; i < m_config->num_shader(); i++) { - if(aggregate_stat) + if (aggregate_stat) total_inst += (pwr_core_stat->m_num_ialu_acesses[CURRENT_STAT_IDX][i]); - else + else total_inst += (pwr_core_stat->m_num_ialu_acesses[CURRENT_STAT_IDX][i]) - - (pwr_core_stat->m_num_ialu_acesses[PREV_STAT_IDX][i]); + (pwr_core_stat->m_num_ialu_acesses[PREV_STAT_IDX][i]); } return total_inst; } @@ -805,17 +825,16 @@ class power_stat_t { sizeof(access_type) / sizeof(enum mem_access_type); unsigned num_request_status = sizeof(request_status) / sizeof(enum cache_request_status); - if(aggregate_stat) + if (aggregate_stat) return (pwr_mem_stat->core_cache_stats[CURRENT_STAT_IDX].get_stats( - access_type, num_access_type, request_status, - num_request_status)); + access_type, num_access_type, request_status, num_request_status)); else return (pwr_mem_stat->core_cache_stats[CURRENT_STAT_IDX].get_stats( - access_type, num_access_type, request_status, - num_request_status)) - - (pwr_mem_stat->core_cache_stats[PREV_STAT_IDX].get_stats( - access_type, num_access_type, request_status, - num_request_status)); + access_type, num_access_type, request_status, + num_request_status)) - + (pwr_mem_stat->core_cache_stats[PREV_STAT_IDX].get_stats( + access_type, num_access_type, request_status, + num_request_status)); } double get_inst_c_misses(bool aggregate_stat) { enum mem_access_type access_type[] = {INST_ACC_R}; @@ -824,46 +843,45 @@ class power_stat_t { sizeof(access_type) / sizeof(enum mem_access_type); unsigned num_request_status = sizeof(request_status) / sizeof(enum cache_request_status); - if(aggregate_stat) + if (aggregate_stat) return (pwr_mem_stat->core_cache_stats[CURRENT_STAT_IDX].get_stats( - access_type, num_access_type, request_status, - num_request_status)); + access_type, num_access_type, request_status, num_request_status)); else return (pwr_mem_stat->core_cache_stats[CURRENT_STAT_IDX].get_stats( - access_type, num_access_type, request_status, - num_request_status)) - - (pwr_mem_stat->core_cache_stats[PREV_STAT_IDX].get_stats( - access_type, num_access_type, request_status, - num_request_status)); + access_type, num_access_type, request_status, + num_request_status)) - + (pwr_mem_stat->core_cache_stats[PREV_STAT_IDX].get_stats( + access_type, num_access_type, request_status, + num_request_status)); } double get_inst_c_hits(bool aggregate_stat) { - return (get_inst_c_accesses(aggregate_stat) - get_inst_c_misses(aggregate_stat)); + return (get_inst_c_accesses(aggregate_stat) - + get_inst_c_misses(aggregate_stat)); } double get_l1d_read_accesses(bool aggregate_stat) { enum mem_access_type access_type[] = {GLOBAL_ACC_R, LOCAL_ACC_R}; - enum cache_request_status request_status[] = {HIT, MISS, SECTOR_MISS}; + enum cache_request_status request_status[] = {HIT, MISS, SECTOR_MISS}; unsigned num_access_type = sizeof(access_type) / sizeof(enum mem_access_type); unsigned num_request_status = sizeof(request_status) / sizeof(enum cache_request_status); - if(aggregate_stat){ + if (aggregate_stat) { return (pwr_mem_stat->core_cache_stats[CURRENT_STAT_IDX].get_stats( - access_type, num_access_type, request_status, - num_request_status)); - } - else{ + access_type, num_access_type, request_status, num_request_status)); + } else { return (pwr_mem_stat->core_cache_stats[CURRENT_STAT_IDX].get_stats( - access_type, num_access_type, request_status, - num_request_status)) - - (pwr_mem_stat->core_cache_stats[PREV_STAT_IDX].get_stats( - access_type, num_access_type, request_status, - num_request_status)); - } + access_type, num_access_type, request_status, + num_request_status)) - + (pwr_mem_stat->core_cache_stats[PREV_STAT_IDX].get_stats( + access_type, num_access_type, request_status, + num_request_status)); + } } double get_l1d_read_misses(bool aggregate_stat) { - return (get_l1d_read_accesses(aggregate_stat) - get_l1d_read_hits(aggregate_stat)); + return (get_l1d_read_accesses(aggregate_stat) - + get_l1d_read_hits(aggregate_stat)); } double get_l1d_read_hits(bool aggregate_stat) { enum mem_access_type access_type[] = {GLOBAL_ACC_R, LOCAL_ACC_R}; @@ -873,19 +891,17 @@ class power_stat_t { unsigned num_request_status = sizeof(request_status) / sizeof(enum cache_request_status); - if(aggregate_stat){ - return (pwr_mem_stat->core_cache_stats[CURRENT_STAT_IDX].get_stats( - access_type, num_access_type, request_status, - num_request_status)); - } - else{ + if (aggregate_stat) { return (pwr_mem_stat->core_cache_stats[CURRENT_STAT_IDX].get_stats( - access_type, num_access_type, request_status, - num_request_status)) - - (pwr_mem_stat->core_cache_stats[PREV_STAT_IDX].get_stats( - access_type, num_access_type, request_status, - num_request_status)); - } + access_type, num_access_type, request_status, num_request_status)); + } else { + return (pwr_mem_stat->core_cache_stats[CURRENT_STAT_IDX].get_stats( + access_type, num_access_type, request_status, + num_request_status)) - + (pwr_mem_stat->core_cache_stats[PREV_STAT_IDX].get_stats( + access_type, num_access_type, request_status, + num_request_status)); + } } double get_l1d_write_accesses(bool aggregate_stat) { enum mem_access_type access_type[] = {GLOBAL_ACC_W, LOCAL_ACC_W}; @@ -895,22 +911,21 @@ class power_stat_t { unsigned num_request_status = sizeof(request_status) / sizeof(enum cache_request_status); - if(aggregate_stat){ - return (pwr_mem_stat->core_cache_stats[CURRENT_STAT_IDX].get_stats( - access_type, num_access_type, request_status, - num_request_status)); - } - else{ + if (aggregate_stat) { return (pwr_mem_stat->core_cache_stats[CURRENT_STAT_IDX].get_stats( - access_type, num_access_type, request_status, - num_request_status)) - - (pwr_mem_stat->core_cache_stats[PREV_STAT_IDX].get_stats( - access_type, num_access_type, request_status, - num_request_status)); - } + access_type, num_access_type, request_status, num_request_status)); + } else { + return (pwr_mem_stat->core_cache_stats[CURRENT_STAT_IDX].get_stats( + access_type, num_access_type, request_status, + num_request_status)) - + (pwr_mem_stat->core_cache_stats[PREV_STAT_IDX].get_stats( + access_type, num_access_type, request_status, + num_request_status)); + } } double get_l1d_write_misses(bool aggregate_stat) { - return (get_l1d_write_accesses(aggregate_stat) - get_l1d_write_hits(aggregate_stat)); + return (get_l1d_write_accesses(aggregate_stat) - + get_l1d_write_hits(aggregate_stat)); } double get_l1d_write_hits(bool aggregate_stat) { enum mem_access_type access_type[] = {GLOBAL_ACC_W, LOCAL_ACC_W}; @@ -920,19 +935,17 @@ class power_stat_t { unsigned num_request_status = sizeof(request_status) / sizeof(enum cache_request_status); - if(aggregate_stat){ - return (pwr_mem_stat->core_cache_stats[CURRENT_STAT_IDX].get_stats( - access_type, num_access_type, request_status, - num_request_status)); - } - else{ + if (aggregate_stat) { return (pwr_mem_stat->core_cache_stats[CURRENT_STAT_IDX].get_stats( - access_type, num_access_type, request_status, - num_request_status)) - - (pwr_mem_stat->core_cache_stats[PREV_STAT_IDX].get_stats( - access_type, num_access_type, request_status, - num_request_status)); - } + access_type, num_access_type, request_status, num_request_status)); + } else { + return (pwr_mem_stat->core_cache_stats[CURRENT_STAT_IDX].get_stats( + access_type, num_access_type, request_status, + num_request_status)) - + (pwr_mem_stat->core_cache_stats[PREV_STAT_IDX].get_stats( + access_type, num_access_type, request_status, + num_request_status)); + } } double get_cache_misses() { return get_l1d_read_misses(0) + get_constant_c_misses() + @@ -949,111 +962,107 @@ class power_stat_t { double get_shmem_access(bool aggregate_stat) { unsigned total_inst = 0; for (unsigned i = 0; i < m_config->num_shader(); i++) { - if(aggregate_stat) + if (aggregate_stat) total_inst += (pwr_mem_stat->shmem_access[CURRENT_STAT_IDX][i]); else total_inst += (pwr_mem_stat->shmem_access[CURRENT_STAT_IDX][i]) - - (pwr_mem_stat->shmem_access[PREV_STAT_IDX][i]); + (pwr_mem_stat->shmem_access[PREV_STAT_IDX][i]); } return total_inst; } - unsigned long long get_l2_read_accesses(bool aggregate_stat) { + unsigned long long get_l2_read_accesses(bool aggregate_stat) { enum mem_access_type access_type[] = { GLOBAL_ACC_R, LOCAL_ACC_R, CONST_ACC_R, TEXTURE_ACC_R, INST_ACC_R}; - enum cache_request_status request_status[] = {HIT, HIT_RESERVED, MISS, SECTOR_MISS}; + enum cache_request_status request_status[] = {HIT, HIT_RESERVED, MISS, + SECTOR_MISS}; unsigned num_access_type = sizeof(access_type) / sizeof(enum mem_access_type); unsigned num_request_status = sizeof(request_status) / sizeof(enum cache_request_status); - if(aggregate_stat){ - return (pwr_mem_stat->l2_cache_stats[CURRENT_STAT_IDX].get_stats( - access_type, num_access_type, request_status, - num_request_status)); - } - else{ + if (aggregate_stat) { return (pwr_mem_stat->l2_cache_stats[CURRENT_STAT_IDX].get_stats( - access_type, num_access_type, request_status, - num_request_status)) - - (pwr_mem_stat->l2_cache_stats[PREV_STAT_IDX].get_stats( - access_type, num_access_type, request_status, - num_request_status)); + access_type, num_access_type, request_status, num_request_status)); + } else { + return (pwr_mem_stat->l2_cache_stats[CURRENT_STAT_IDX].get_stats( + access_type, num_access_type, request_status, + num_request_status)) - + (pwr_mem_stat->l2_cache_stats[PREV_STAT_IDX].get_stats( + access_type, num_access_type, request_status, + num_request_status)); } } unsigned long long get_l2_read_misses(bool aggregate_stat) { - return (get_l2_read_accesses(aggregate_stat) - get_l2_read_hits(aggregate_stat)); + return (get_l2_read_accesses(aggregate_stat) - + get_l2_read_hits(aggregate_stat)); } unsigned long long get_l2_read_hits(bool aggregate_stat) { - enum mem_access_type access_type[] = { + enum mem_access_type access_type[] = { GLOBAL_ACC_R, LOCAL_ACC_R, CONST_ACC_R, TEXTURE_ACC_R, INST_ACC_R}; - enum cache_request_status request_status[] = {HIT, HIT_RESERVED}; + enum cache_request_status request_status[] = {HIT, HIT_RESERVED}; unsigned num_access_type = sizeof(access_type) / sizeof(enum mem_access_type); unsigned num_request_status = sizeof(request_status) / sizeof(enum cache_request_status); - if(aggregate_stat){ - return (pwr_mem_stat->l2_cache_stats[CURRENT_STAT_IDX].get_stats( - access_type, num_access_type, request_status, - num_request_status)); - } - else{ + if (aggregate_stat) { return (pwr_mem_stat->l2_cache_stats[CURRENT_STAT_IDX].get_stats( - access_type, num_access_type, request_status, - num_request_status)) - - (pwr_mem_stat->l2_cache_stats[PREV_STAT_IDX].get_stats( - access_type, num_access_type, request_status, - num_request_status)); + access_type, num_access_type, request_status, num_request_status)); + } else { + return (pwr_mem_stat->l2_cache_stats[CURRENT_STAT_IDX].get_stats( + access_type, num_access_type, request_status, + num_request_status)) - + (pwr_mem_stat->l2_cache_stats[PREV_STAT_IDX].get_stats( + access_type, num_access_type, request_status, + num_request_status)); } } unsigned long long get_l2_write_accesses(bool aggregate_stat) { enum mem_access_type access_type[] = {GLOBAL_ACC_W, LOCAL_ACC_W, L1_WRBK_ACC}; - enum cache_request_status request_status[] = {HIT, HIT_RESERVED, MISS, SECTOR_MISS}; + enum cache_request_status request_status[] = {HIT, HIT_RESERVED, MISS, + SECTOR_MISS}; unsigned num_access_type = sizeof(access_type) / sizeof(enum mem_access_type); unsigned num_request_status = sizeof(request_status) / sizeof(enum cache_request_status); - if(aggregate_stat){ + if (aggregate_stat) { return (pwr_mem_stat->l2_cache_stats[CURRENT_STAT_IDX].get_stats( - access_type, num_access_type, request_status, - num_request_status)); - } - else{ + access_type, num_access_type, request_status, num_request_status)); + } else { return (pwr_mem_stat->l2_cache_stats[CURRENT_STAT_IDX].get_stats( - access_type, num_access_type, request_status, - num_request_status)) - - (pwr_mem_stat->l2_cache_stats[PREV_STAT_IDX].get_stats( - access_type, num_access_type, request_status, - num_request_status)); + access_type, num_access_type, request_status, + num_request_status)) - + (pwr_mem_stat->l2_cache_stats[PREV_STAT_IDX].get_stats( + access_type, num_access_type, request_status, + num_request_status)); } } unsigned long long get_l2_write_misses(bool aggregate_stat) { - return (get_l2_write_accesses(aggregate_stat) - get_l2_write_hits(aggregate_stat)); + return (get_l2_write_accesses(aggregate_stat) - + get_l2_write_hits(aggregate_stat)); } unsigned long long get_l2_write_hits(bool aggregate_stat) { - enum mem_access_type access_type[] = {GLOBAL_ACC_W, LOCAL_ACC_W, + enum mem_access_type access_type[] = {GLOBAL_ACC_W, LOCAL_ACC_W, L1_WRBK_ACC}; enum cache_request_status request_status[] = {HIT, HIT_RESERVED}; unsigned num_access_type = sizeof(access_type) / sizeof(enum mem_access_type); unsigned num_request_status = sizeof(request_status) / sizeof(enum cache_request_status); - if(aggregate_stat){ + if (aggregate_stat) { return (pwr_mem_stat->l2_cache_stats[CURRENT_STAT_IDX].get_stats( - access_type, num_access_type, request_status, - num_request_status)); - } - else{ + access_type, num_access_type, request_status, num_request_status)); + } else { return (pwr_mem_stat->l2_cache_stats[CURRENT_STAT_IDX].get_stats( - access_type, num_access_type, request_status, - num_request_status)) - - (pwr_mem_stat->l2_cache_stats[PREV_STAT_IDX].get_stats( - access_type, num_access_type, request_status, - num_request_status)); + access_type, num_access_type, request_status, + num_request_status)) - + (pwr_mem_stat->l2_cache_stats[PREV_STAT_IDX].get_stats( + access_type, num_access_type, request_status, + num_request_status)); } } double get_dram_cmd() { @@ -1091,12 +1100,11 @@ class power_stat_t { double get_dram_pre(bool aggregate_stat) { unsigned total = 0; for (unsigned i = 0; i < m_mem_config->m_n_mem; ++i) { - if(aggregate_stat){ + if (aggregate_stat) { total += pwr_mem_stat->n_pre[CURRENT_STAT_IDX][i]; - } - else{ + } else { total += (pwr_mem_stat->n_pre[CURRENT_STAT_IDX][i] - - pwr_mem_stat->n_pre[PREV_STAT_IDX][i]); + pwr_mem_stat->n_pre[PREV_STAT_IDX][i]); } } return total; @@ -1104,12 +1112,11 @@ class power_stat_t { double get_dram_rd(bool aggregate_stat) { unsigned total = 0; for (unsigned i = 0; i < m_mem_config->m_n_mem; ++i) { - if(aggregate_stat){ + if (aggregate_stat) { total += pwr_mem_stat->n_rd[CURRENT_STAT_IDX][i]; - } - else{ + } else { total += (pwr_mem_stat->n_rd[CURRENT_STAT_IDX][i] - - pwr_mem_stat->n_rd[PREV_STAT_IDX][i]); + pwr_mem_stat->n_rd[PREV_STAT_IDX][i]); } } return total; @@ -1117,15 +1124,14 @@ class power_stat_t { double get_dram_wr(bool aggregate_stat) { unsigned total = 0; for (unsigned i = 0; i < m_mem_config->m_n_mem; ++i) { - if(aggregate_stat){ - total += pwr_mem_stat->n_wr[CURRENT_STAT_IDX][i] + - pwr_mem_stat->n_wr_WB[CURRENT_STAT_IDX][i]; - } - else{ - total += (pwr_mem_stat->n_wr[CURRENT_STAT_IDX][i] - - pwr_mem_stat->n_wr[PREV_STAT_IDX][i]) + - (pwr_mem_stat->n_wr_WB[CURRENT_STAT_IDX][i] - - pwr_mem_stat->n_wr_WB[PREV_STAT_IDX][i]); + if (aggregate_stat) { + total += pwr_mem_stat->n_wr[CURRENT_STAT_IDX][i] + + pwr_mem_stat->n_wr_WB[CURRENT_STAT_IDX][i]; + } else { + total += (pwr_mem_stat->n_wr[CURRENT_STAT_IDX][i] - + pwr_mem_stat->n_wr[PREV_STAT_IDX][i]) + + (pwr_mem_stat->n_wr_WB[CURRENT_STAT_IDX][i] - + pwr_mem_stat->n_wr_WB[PREV_STAT_IDX][i]); } } return total; @@ -1141,13 +1147,12 @@ class power_stat_t { unsigned long long get_icnt_simt_to_mem(bool aggregate_stat) { long total = 0; - for (unsigned i = 0; i < m_config->n_simt_clusters; ++i){ - if(aggregate_stat){ + for (unsigned i = 0; i < m_config->n_simt_clusters; ++i) { + if (aggregate_stat) { total += pwr_mem_stat->n_simt_to_mem[CURRENT_STAT_IDX][i]; - } - else{ + } else { total += (pwr_mem_stat->n_simt_to_mem[CURRENT_STAT_IDX][i] - - pwr_mem_stat->n_simt_to_mem[PREV_STAT_IDX][i]); + pwr_mem_stat->n_simt_to_mem[PREV_STAT_IDX][i]); } } return total; @@ -1156,13 +1161,13 @@ class power_stat_t { unsigned long long get_icnt_mem_to_simt(bool aggregate_stat) { long total = 0; for (unsigned i = 0; i < m_config->n_simt_clusters; ++i) { - if(aggregate_stat){ + if (aggregate_stat) { total += pwr_mem_stat->n_mem_to_simt[CURRENT_STAT_IDX][i]; } - - else{ + + else { total += (pwr_mem_stat->n_mem_to_simt[CURRENT_STAT_IDX][i] - - pwr_mem_stat->n_mem_to_simt[PREV_STAT_IDX][i]); + pwr_mem_stat->n_mem_to_simt[PREV_STAT_IDX][i]); } } return total; diff --git a/src/gpgpu-sim/shader.cc b/src/gpgpu-sim/shader.cc index 855aa1c..9fe4c09 100644 --- a/src/gpgpu-sim/shader.cc +++ b/src/gpgpu-sim/shader.cc @@ -1,18 +1,19 @@ // Copyright (c) 2009-2021, Tor M. Aamodt, Wilson W.L. Fung, Ali Bakhoda, -// George L. Yuan, Andrew Turner, Inderpreet Singh, Vijay Kandiah, Nikos Hardavellas, -// Mahmoud Khairy, Junrui Pan, Timothy G. Rogers -// The University of British Columbia, Northwestern University, Purdue University -// All rights reserved. +// George L. Yuan, Andrew Turner, Inderpreet Singh, Vijay Kandiah, Nikos +// Hardavellas, Mahmoud Khairy, Junrui Pan, Timothy G. Rogers The University of +// British Columbia, Northwestern University, Purdue University All rights +// reserved. // // Redistribution and use in source and binary forms, with or without // modification, are permitted provided that the following conditions are met: // -// 1. Redistributions of source code must retain the above copyright notice, this +// 1. Redistributions of source code must retain the above copyright notice, +// this // list of conditions and the following disclaimer; // 2. Redistributions in binary form must reproduce the above copyright notice, // this list of conditions and the following disclaimer in the documentation // and/or other materials provided with the distribution; -// 3. Neither the names of The University of British Columbia, Northwestern +// 3. Neither the names of The University of British Columbia, Northwestern // University nor the names of their contributors may be used to // endorse or promote products derived from this software without specific // prior written permission. @@ -187,20 +188,16 @@ void shader_core_ctx::create_schedulers() { // must currently occur after all inputs have been initialized. std::string sched_config = m_config->gpgpu_scheduler_string; const concrete_scheduler scheduler = - sched_config.find("lrr") != std::string::npos - ? CONCRETE_SCHEDULER_LRR - : sched_config.find("two_level_active") != std::string::npos - ? CONCRETE_SCHEDULER_TWO_LEVEL_ACTIVE - : sched_config.find("gto") != std::string::npos - ? CONCRETE_SCHEDULER_GTO - : sched_config.find("rrr") != std::string::npos - ? CONCRETE_SCHEDULER_RRR - : sched_config.find("old") != std::string::npos - ? CONCRETE_SCHEDULER_OLDEST_FIRST - : sched_config.find("warp_limiting") != - std::string::npos - ? CONCRETE_SCHEDULER_WARP_LIMITING - : NUM_CONCRETE_SCHEDULERS; + sched_config.find("lrr") != std::string::npos ? CONCRETE_SCHEDULER_LRR + : sched_config.find("two_level_active") != std::string::npos + ? CONCRETE_SCHEDULER_TWO_LEVEL_ACTIVE + : sched_config.find("gto") != std::string::npos ? CONCRETE_SCHEDULER_GTO + : sched_config.find("rrr") != std::string::npos ? CONCRETE_SCHEDULER_RRR + : sched_config.find("old") != std::string::npos + ? CONCRETE_SCHEDULER_OLDEST_FIRST + : sched_config.find("warp_limiting") != std::string::npos + ? CONCRETE_SCHEDULER_WARP_LIMITING + : NUM_CONCRETE_SCHEDULERS; assert(scheduler != NUM_CONCRETE_SCHEDULERS); for (unsigned i = 0; i < m_config->gpgpu_num_sched_per_core; i++) { @@ -487,8 +484,8 @@ shader_core_ctx::shader_core_ctx(class gpgpu_sim *gpu, m_sid = shader_id; m_tpc = tpc_id; - if(get_gpu()->get_config().g_power_simulation_enabled){ - scaling_coeffs = get_gpu()->get_scaling_coeffs(); + if (get_gpu()->get_config().g_power_simulation_enabled) { + scaling_coeffs = get_gpu()->get_scaling_coeffs(); } m_last_inst_gpu_sim_cycle = 0; @@ -641,7 +638,8 @@ void shader_core_stats::print(FILE *fout) const { fprintf(fout, "gpgpu_n_param_mem_insn = %d\n", gpgpu_n_param_insn); fprintf(fout, "gpgpu_n_shmem_bkconflict = %d\n", gpgpu_n_shmem_bkconflict); - fprintf(fout, "gpgpu_n_l1cache_bkconflict = %d\n", gpgpu_n_l1cache_bkconflict); + fprintf(fout, "gpgpu_n_l1cache_bkconflict = %d\n", + gpgpu_n_l1cache_bkconflict); fprintf(fout, "gpgpu_n_intrawarp_mshr_merge = %d\n", gpgpu_n_intrawarp_mshr_merge); @@ -893,7 +891,9 @@ void shader_core_ctx::decode() { m_warp[m_inst_fetch_buffer.m_warp_id]->inc_inst_in_pipeline(); if (pI1) { m_stats->m_num_decoded_insn[m_sid]++; - if ((pI1->oprnd_type == INT_OP) || (pI1->oprnd_type == UN_OP)) { //these counters get added up in mcPat to compute scheduler power + if ((pI1->oprnd_type == INT_OP) || + (pI1->oprnd_type == UN_OP)) { // these counters get added up in mcPat + // to compute scheduler power m_stats->m_num_INTdecoded_insn[m_sid]++; } else if (pI1->oprnd_type == FP_OP) { m_stats->m_num_FPdecoded_insn[m_sid]++; @@ -904,7 +904,9 @@ void shader_core_ctx::decode() { m_warp[m_inst_fetch_buffer.m_warp_id]->ibuffer_fill(1, pI2); m_warp[m_inst_fetch_buffer.m_warp_id]->inc_inst_in_pipeline(); m_stats->m_num_decoded_insn[m_sid]++; - if ((pI1->oprnd_type == INT_OP) || (pI1->oprnd_type == UN_OP)) { //these counters get added up in mcPat to compute scheduler power + if ((pI1->oprnd_type == INT_OP) || + (pI1->oprnd_type == UN_OP)) { // these counters get added up in + // mcPat to compute scheduler power m_stats->m_num_INTdecoded_insn[m_sid]++; } else if (pI2->oprnd_type == FP_OP) { m_stats->m_num_FPdecoded_insn[m_sid]++; @@ -950,7 +952,8 @@ void shader_core_ctx::fetch() { m_threadState[tid].m_active = false; unsigned cta_id = m_warp[warp_id]->get_cta_id(); if (m_thread[tid] == NULL) { - register_cta_thread_exit(cta_id, m_warp[warp_id]->get_kernel_info()); + register_cta_thread_exit(cta_id, + m_warp[warp_id]->get_kernel_info()); } else { register_cta_thread_exit(cta_id, &(m_thread[tid]->get_kernel())); @@ -987,11 +990,10 @@ void shader_core_ctx::fetch() { m_gpu->gpu_tot_sim_cycle + m_gpu->gpu_sim_cycle); std::list events; enum cache_request_status status; - if (m_config->perfect_inst_const_cache){ + if (m_config->perfect_inst_const_cache) { status = HIT; shader_cache_access_log(m_sid, INSTRUCTION, 0); - } - else + } else status = m_L1I->access( (new_addr_type)ppc, mf, m_gpu->gpu_sim_cycle + m_gpu->gpu_tot_sim_cycle, events); @@ -1050,14 +1052,13 @@ void shader_core_ctx::issue_warp(register_set &pipe_reg_set, if (next_inst->m_is_ldgsts) { if (m_warp[warp_id]->m_ldgdepbar_buf.size() == ldgdepbar_id + 1) { m_warp[warp_id]->m_ldgdepbar_buf[ldgdepbar_id].push_back(*next_inst); - } - else { + } else { assert(m_warp[warp_id]->m_ldgdepbar_buf.size() < ldgdepbar_id + 1); std::vector l; l.push_back(*next_inst); m_warp[warp_id]->m_ldgdepbar_buf.push_back(l); } - // If the mask of the instruction is all 0, then the address is also 0, + // If the mask of the instruction is all 0, then the address is also 0, // so that there's no need to check through the writeback if (next_inst->get_active_mask() == 0) { (m_warp[warp_id]->m_ldgdepbar_buf.back()).back().pc = -1; @@ -1071,7 +1072,7 @@ void shader_core_ctx::issue_warp(register_set &pipe_reg_set, } else if (next_inst->op == MEMORY_BARRIER_OP) { m_warp[warp_id]->set_membar(); - } else if (next_inst->m_is_ldgdepbar) { // Add for LDGDEPBAR + } else if (next_inst->m_is_ldgdepbar) { // Add for LDGDEPBAR m_warp[warp_id]->m_ldgdepbar_id++; // If there are no added LDGSTS, insert an empty vector if (m_warp[warp_id]->m_ldgdepbar_buf.size() != ldgdepbar_id + 1) { @@ -1082,16 +1083,20 @@ void shader_core_ctx::issue_warp(register_set &pipe_reg_set, } else if (next_inst->m_is_depbar) { // Add for DEPBAR // Set to true immediately when a DEPBAR instruction is met m_warp[warp_id]->m_waiting_ldgsts = true; - m_warp[warp_id]->m_depbar_group = next_inst->m_depbar_group_no; // set in trace_driven.cc + m_warp[warp_id]->m_depbar_group = + next_inst->m_depbar_group_no; // set in trace_driven.cc - // Record the last group that's possbily being monitored by this DEPBAR instr + // Record the last group that's possbily being monitored by this DEPBAR + // instr m_warp[warp_id]->m_depbar_start_id = m_warp[warp_id]->m_ldgdepbar_id - 1; - - // Record the last group that's actually being monitored by this DEPBAR instr - unsigned int end_group = m_warp[warp_id]->m_ldgdepbar_id - m_warp[warp_id]->m_depbar_group; - // Check for the case that the LDGSTSs monitored have finished when encountering the - // DEPBAR instruction + // Record the last group that's actually being monitored by this DEPBAR + // instr + unsigned int end_group = + m_warp[warp_id]->m_ldgdepbar_id - m_warp[warp_id]->m_depbar_group; + + // Check for the case that the LDGSTSs monitored have finished when + // encountering the DEPBAR instruction bool done_flag = true; for (int i = 0; i < end_group; i++) { for (int j = 0; j < m_warp[warp_id]->m_ldgdepbar_buf[i].size(); j++) { @@ -1101,7 +1106,7 @@ void shader_core_ctx::issue_warp(register_set &pipe_reg_set, } } } - + UpdateDEPBAR: if (done_flag) { if (m_warp[warp_id]->m_waiting_ldgsts) { @@ -1184,11 +1189,12 @@ void scheduler_unit::order_rrr( if (m_num_issued_last_cycle > 0 || warp(m_current_turn_warp).done_exit() || warp(m_current_turn_warp).waiting()) { std::vector::const_iterator iter = - (last_issued_from_input == input_list.end()) ? - input_list.begin() : last_issued_from_input + 1; + (last_issued_from_input == input_list.end()) + ? input_list.begin() + : last_issued_from_input + 1; for (unsigned count = 0; count < num_warps_to_add; ++iter, ++count) { if (iter == input_list.end()) { - iter = input_list.begin(); + iter = input_list.begin(); } unsigned warp_id = (*iter)->get_warp_id(); if (!(*iter)->done_exit() && !(*iter)->waiting()) { @@ -1854,33 +1860,38 @@ void ldst_unit::get_L1T_sub_stats(struct cache_sub_stats &css) const { // Add this function to unset depbar void shader_core_ctx::unset_depbar(const warp_inst_t &inst) { bool done_flag = true; - unsigned int end_group = m_warp[inst.warp_id()]->m_depbar_start_id == 0 ? - m_warp[inst.warp_id()]->m_ldgdepbar_buf.size() : - (m_warp[inst.warp_id()]->m_depbar_start_id - m_warp[inst.warp_id()]->m_depbar_group + 1); + unsigned int end_group = m_warp[inst.warp_id()]->m_depbar_start_id == 0 + ? m_warp[inst.warp_id()]->m_ldgdepbar_buf.size() + : (m_warp[inst.warp_id()]->m_depbar_start_id - + m_warp[inst.warp_id()]->m_depbar_group + 1); - if (inst.m_is_ldgsts) { + if (inst.m_is_ldgsts) { for (int i = 0; i < m_warp[inst.warp_id()]->m_ldgdepbar_buf.size(); i++) { - for (int j = 0; j < m_warp[inst.warp_id()]->m_ldgdepbar_buf[i].size(); j++) { + for (int j = 0; j < m_warp[inst.warp_id()]->m_ldgdepbar_buf[i].size(); + j++) { if (m_warp[inst.warp_id()]->m_ldgdepbar_buf[i][j].pc == inst.pc) { - // Handle the case that same pc results in multiple LDGSTS instructions - if (m_warp[inst.warp_id()]->m_ldgdepbar_buf[i][j].get_addr(0) == inst.get_addr(0)) { + // Handle the case that same pc results in multiple LDGSTS + // instructions + if (m_warp[inst.warp_id()]->m_ldgdepbar_buf[i][j].get_addr(0) == + inst.get_addr(0)) { m_warp[inst.warp_id()]->m_ldgdepbar_buf[i][j].pc = -1; goto DoneWB; } - } + } } } DoneWB: for (int i = 0; i < end_group; i++) { - for (int j = 0; j < m_warp[inst.warp_id()]->m_ldgdepbar_buf[i].size(); j++) { + for (int j = 0; j < m_warp[inst.warp_id()]->m_ldgdepbar_buf[i].size(); + j++) { if (m_warp[inst.warp_id()]->m_ldgdepbar_buf[i][j].pc != -1) { done_flag = false; goto UpdateDEPBAR; } } } - + UpdateDEPBAR: if (done_flag) { if (m_warp[inst.warp_id()]->m_waiting_ldgsts) { @@ -2000,7 +2011,7 @@ mem_stage_stall_type ldst_unit::process_cache_access( if (inst.is_load()) { for (unsigned r = 0; r < MAX_OUTPUT_VALUES; r++) if (inst.out[r] > 0) m_pending_writes[inst.warp_id()][inst.out[r]]--; - + // release LDGSTS if (inst.m_is_ldgsts) { m_pending_ldgsts[inst.warp_id()][inst.pc][inst.get_addr(0)]--; @@ -2139,8 +2150,12 @@ void ldst_unit::L1_latency_queue_cycle() { // release LDGSTS if (mf_next->get_inst().m_is_ldgsts) { - m_pending_ldgsts[mf_next->get_inst().warp_id()][mf_next->get_inst().pc][mf_next->get_inst().get_addr(0)]--; - if (m_pending_ldgsts[mf_next->get_inst().warp_id()][mf_next->get_inst().pc][mf_next->get_inst().get_addr(0)] == 0) { + m_pending_ldgsts[mf_next->get_inst().warp_id()] + [mf_next->get_inst().pc] + [mf_next->get_inst().get_addr(0)]--; + if (m_pending_ldgsts[mf_next->get_inst().warp_id()] + [mf_next->get_inst().pc] + [mf_next->get_inst().get_addr(0)] == 0) { m_core->unset_depbar(mf_next->get_inst()); } } @@ -2207,7 +2222,8 @@ bool ldst_unit::constant_cycle(warp_inst_t &inst, mem_stage_stall_type &rc_fail, while (inst.accessq_count() > 0) inst.accessq_pop_back(); if (inst.is_load()) { for (unsigned r = 0; r < MAX_OUTPUT_VALUES; r++) - if (inst.out[r] > 0) m_pending_writes[inst.warp_id()][inst.out[r]] -= access_count; + if (inst.out[r] > 0) + m_pending_writes[inst.warp_id()][inst.out[r]] -= access_count; } } else { fail = process_memory_access_queue(m_L1C, inst); @@ -2395,7 +2411,7 @@ void sp_unit::active_lanes_in_pipeline() { void dp_unit::active_lanes_in_pipeline() { unsigned active_count = pipelined_simd_unit::get_active_lanes_in_pipeline(); assert(active_count <= m_core->get_config()->warp_size); - //m_core->incspactivelanes_stat(active_count); + // m_core->incspactivelanes_stat(active_count); m_core->incfuactivelanes_stat(active_count); m_core->incfumemactivelanes_stat(active_count); } @@ -2527,9 +2543,9 @@ void pipelined_simd_unit::cycle() { if (!m_dispatch_reg->dispatch_delay()) { int start_stage = m_dispatch_reg->latency - m_dispatch_reg->initiation_interval; - if(m_pipeline_reg[start_stage]->empty()) { - move_warp(m_pipeline_reg[start_stage], m_dispatch_reg); - active_insts_in_pipeline++; + if (m_pipeline_reg[start_stage]->empty()) { + move_warp(m_pipeline_reg[start_stage], m_dispatch_reg); + active_insts_in_pipeline++; } } } @@ -2682,10 +2698,12 @@ void ldst_unit::writeback() { m_next_wb.out[r]); insn_completed = true; } - } - else if (m_next_wb.m_is_ldgsts) { // for LDGSTS instructions where no output register is used - m_pending_ldgsts[m_next_wb.warp_id()][m_next_wb.pc][m_next_wb.get_addr(0)]--; - if (m_pending_ldgsts[m_next_wb.warp_id()][m_next_wb.pc][m_next_wb.get_addr(0)] == 0) { + } else if (m_next_wb.m_is_ldgsts) { // for LDGSTS instructions where no + // output register is used + m_pending_ldgsts[m_next_wb.warp_id()][m_next_wb.pc] + [m_next_wb.get_addr(0)]--; + if (m_pending_ldgsts[m_next_wb.warp_id()][m_next_wb.pc] + [m_next_wb.get_addr(0)] == 0) { insn_completed = true; } break; @@ -2923,7 +2941,8 @@ void ldst_unit::cycle() { // release LDGSTS if (m_dispatch_reg->m_is_ldgsts) { // m_pending_ldgsts[m_dispatch_reg->warp_id()][m_dispatch_reg->pc][m_dispatch_reg->get_addr(0)]--; - if (m_pending_ldgsts[m_dispatch_reg->warp_id()][m_dispatch_reg->pc][m_dispatch_reg->get_addr(0)] == 0) { + if (m_pending_ldgsts[m_dispatch_reg->warp_id()][m_dispatch_reg->pc] + [m_dispatch_reg->get_addr(0)] == 0) { m_core->unset_depbar(*m_dispatch_reg); } } @@ -3223,68 +3242,68 @@ void warp_inst_t::print(FILE *fout) const { m_config->gpgpu_ctx->func_sim->ptx_print_insn(pc, fout); fprintf(fout, "\n"); } -void shader_core_ctx::incexecstat(warp_inst_t *&inst) -{ - // Latency numbers for next operations are used to scale the power values - // for special operations, according observations from microbenchmarking - // TODO: put these numbers in the xml configuration - if(get_gpu()->get_config().g_power_simulation_enabled){ - switch(inst->sp_op){ - case INT__OP: - incialu_stat(inst->active_count(), scaling_coeffs->int_coeff); - break; - case INT_MUL_OP: - incimul_stat(inst->active_count(), scaling_coeffs->int_mul_coeff); - break; - case INT_MUL24_OP: - incimul24_stat(inst->active_count(), scaling_coeffs->int_mul24_coeff); - break; - case INT_MUL32_OP: - incimul32_stat(inst->active_count(), scaling_coeffs->int_mul32_coeff); - break; - case INT_DIV_OP: - incidiv_stat(inst->active_count(), scaling_coeffs->int_div_coeff); - break; - case FP__OP: - incfpalu_stat(inst->active_count(),scaling_coeffs->fp_coeff); - break; - case FP_MUL_OP: - incfpmul_stat(inst->active_count(), scaling_coeffs->fp_mul_coeff); - break; - case FP_DIV_OP: - incfpdiv_stat(inst->active_count(), scaling_coeffs->fp_div_coeff); - break; - case DP___OP: - incdpalu_stat(inst->active_count(), scaling_coeffs->dp_coeff); - break; - case DP_MUL_OP: - incdpmul_stat(inst->active_count(), scaling_coeffs->dp_mul_coeff); - break; - case DP_DIV_OP: - incdpdiv_stat(inst->active_count(), scaling_coeffs->dp_div_coeff); - break; - case FP_SQRT_OP: - incsqrt_stat(inst->active_count(), scaling_coeffs->sqrt_coeff); - break; - case FP_LG_OP: - inclog_stat(inst->active_count(), scaling_coeffs->log_coeff); - break; - case FP_SIN_OP: - incsin_stat(inst->active_count(), scaling_coeffs->sin_coeff); - break; - case FP_EXP_OP: - incexp_stat(inst->active_count(), scaling_coeffs->exp_coeff); - break; - case TENSOR__OP: - inctensor_stat(inst->active_count(), scaling_coeffs->tensor_coeff); - break; - case TEX__OP: - inctex_stat(inst->active_count(), scaling_coeffs->tex_coeff); - break; - default: - break; +void shader_core_ctx::incexecstat(warp_inst_t *&inst) { + // Latency numbers for next operations are used to scale the power values + // for special operations, according observations from microbenchmarking + // TODO: put these numbers in the xml configuration + if (get_gpu()->get_config().g_power_simulation_enabled) { + switch (inst->sp_op) { + case INT__OP: + incialu_stat(inst->active_count(), scaling_coeffs->int_coeff); + break; + case INT_MUL_OP: + incimul_stat(inst->active_count(), scaling_coeffs->int_mul_coeff); + break; + case INT_MUL24_OP: + incimul24_stat(inst->active_count(), scaling_coeffs->int_mul24_coeff); + break; + case INT_MUL32_OP: + incimul32_stat(inst->active_count(), scaling_coeffs->int_mul32_coeff); + break; + case INT_DIV_OP: + incidiv_stat(inst->active_count(), scaling_coeffs->int_div_coeff); + break; + case FP__OP: + incfpalu_stat(inst->active_count(), scaling_coeffs->fp_coeff); + break; + case FP_MUL_OP: + incfpmul_stat(inst->active_count(), scaling_coeffs->fp_mul_coeff); + break; + case FP_DIV_OP: + incfpdiv_stat(inst->active_count(), scaling_coeffs->fp_div_coeff); + break; + case DP___OP: + incdpalu_stat(inst->active_count(), scaling_coeffs->dp_coeff); + break; + case DP_MUL_OP: + incdpmul_stat(inst->active_count(), scaling_coeffs->dp_mul_coeff); + break; + case DP_DIV_OP: + incdpdiv_stat(inst->active_count(), scaling_coeffs->dp_div_coeff); + break; + case FP_SQRT_OP: + incsqrt_stat(inst->active_count(), scaling_coeffs->sqrt_coeff); + break; + case FP_LG_OP: + inclog_stat(inst->active_count(), scaling_coeffs->log_coeff); + break; + case FP_SIN_OP: + incsin_stat(inst->active_count(), scaling_coeffs->sin_coeff); + break; + case FP_EXP_OP: + incexp_stat(inst->active_count(), scaling_coeffs->exp_coeff); + break; + case TENSOR__OP: + inctensor_stat(inst->active_count(), scaling_coeffs->tensor_coeff); + break; + case TEX__OP: + inctex_stat(inst->active_count(), scaling_coeffs->tex_coeff); + break; + default: + break; } - if(inst->const_cache_operand) //warp has const address space load as one operand + if (inst->const_cache_operand) // warp has const address space load as one + // operand inc_const_accesses(1); } } @@ -4034,7 +4053,9 @@ void shader_core_ctx::get_icnt_power_stats(long &n_simt_to_mem, n_mem_to_simt += m_stats->n_mem_to_simt[m_sid]; } -kernel_info_t* shd_warp_t::get_kernel_info() const { return m_shader->get_kernel_info(); } +kernel_info_t *shd_warp_t::get_kernel_info() const { + return m_shader->get_kernel_info(); +} bool shd_warp_t::functional_done() const { return get_n_completed() == m_warp_size; @@ -4069,7 +4090,8 @@ bool shd_warp_t::waiting() { void shd_warp_t::print(FILE *fout) const { if (!done_exit()) { - fprintf(fout, "w%02u npc: 0x%04llx, done:%c%c%c%c:%2u i:%u s:%u a:%u (done: ", + fprintf(fout, + "w%02u npc: 0x%04llx, done:%c%c%c%c:%2u i:%u s:%u a:%u (done: ", m_warp_id, m_next_pc, (functional_done() ? 'f' : ' '), (stores_done() ? 's' : ' '), (inst_in_pipeline() ? ' ' : 'i'), (done_exit() ? 'e' : ' '), n_completed, m_inst_in_pipeline, @@ -4156,18 +4178,18 @@ void opndcoll_rfu_t::init(unsigned num_banks, shader_core_ctx *shader) { unsigned cusPerSched = m_cu.size() / m_num_warp_scheds; reg_id = j / cusPerSched; } - m_cu[j]->init(j, num_banks, shader->get_config(), this, - sub_core_model, reg_id, m_num_banks_per_sched); + m_cu[j]->init(j, num_banks, shader->get_config(), this, sub_core_model, + reg_id, m_num_banks_per_sched); } for (unsigned j = 0; j < m_dispatch_units.size(); j++) { - m_dispatch_units[j].init(sub_core_model,m_num_warp_scheds); + m_dispatch_units[j].init(sub_core_model, m_num_warp_scheds); } m_initialized = true; } unsigned register_bank(int regnum, int wid, unsigned num_banks, - bool sub_core_model, - unsigned banks_per_sched, unsigned sched_id) { + bool sub_core_model, unsigned banks_per_sched, + unsigned sched_id) { int bank = regnum; bank += wid; if (sub_core_model) { @@ -4186,14 +4208,13 @@ bool opndcoll_rfu_t::writeback(warp_inst_t &inst) { int reg_num = inst.arch_reg.dst[op]; // this math needs to match that used // in function_info::ptx_decode_inst if (reg_num >= 0) { // valid register - unsigned bank = register_bank(reg_num, inst.warp_id(), m_num_banks, - sub_core_model, - m_num_banks_per_sched, inst.get_schd_id()); + unsigned bank = + register_bank(reg_num, inst.warp_id(), m_num_banks, sub_core_model, + m_num_banks_per_sched, inst.get_schd_id()); if (m_arbiter.bank_idle(bank)) { m_arbiter.allocate_bank_for_write( - bank, - op_t(&inst, reg_num, m_num_banks, sub_core_model, - m_num_banks_per_sched, inst.get_schd_id())); + bank, op_t(&inst, reg_num, m_num_banks, sub_core_model, + m_num_banks_per_sched, inst.get_schd_id())); inst.arch_reg.dst[op] = -1; } else { return false; @@ -4301,9 +4322,8 @@ void opndcoll_rfu_t::allocate_reads() { const op_t &rr = *r; unsigned reg = rr.get_reg(); unsigned wid = rr.get_wid(); - unsigned bank = - register_bank(reg, wid, m_num_banks, sub_core_model, - m_num_banks_per_sched, rr.get_sid()); + unsigned bank = register_bank(reg, wid, m_num_banks, sub_core_model, + m_num_banks_per_sched, rr.get_sid()); m_arbiter.allocate_for_read(bank, rr); read_ops[bank] = rr; } @@ -4353,10 +4373,12 @@ void opndcoll_rfu_t::collector_unit_t::dump( } } -void opndcoll_rfu_t::collector_unit_t::init( - unsigned n, unsigned num_banks, - const core_config *config, opndcoll_rfu_t *rfu, bool sub_core_model, - unsigned reg_id, unsigned banks_per_sched) { +void opndcoll_rfu_t::collector_unit_t::init(unsigned n, unsigned num_banks, + const core_config *config, + opndcoll_rfu_t *rfu, + bool sub_core_model, + unsigned reg_id, + unsigned banks_per_sched) { m_rfu = rfu; m_cuid = n; m_num_banks = num_banks; @@ -4376,7 +4398,7 @@ bool opndcoll_rfu_t::collector_unit_t::allocate(register_set *pipeline_reg_set, warp_inst_t **pipeline_reg = pipeline_reg_set->get_ready(); if ((pipeline_reg) and !((*pipeline_reg)->empty())) { m_warp_id = (*pipeline_reg)->warp_id(); - std::vector prev_regs; // remove duplicate regs within same instr + std::vector prev_regs; // remove duplicate regs within same instr for (unsigned op = 0; op < MAX_REG_OPERANDS; op++) { int reg_num = (*pipeline_reg) @@ -4384,14 +4406,13 @@ bool opndcoll_rfu_t::collector_unit_t::allocate(register_set *pipeline_reg_set, // function_info::ptx_decode_inst bool new_reg = true; for (auto r : prev_regs) { - if (r == reg_num) - new_reg = false; + if (r == reg_num) new_reg = false; } - if (reg_num >= 0 && new_reg) { // valid register + if (reg_num >= 0 && new_reg) { // valid register prev_regs.push_back(reg_num); - m_src_op[op] = op_t(this, op, reg_num, m_num_banks, - m_sub_core_model, m_num_banks_per_sched, - (*pipeline_reg)->get_schd_id()); + m_src_op[op] = + op_t(this, op, reg_num, m_num_banks, m_sub_core_model, + m_num_banks_per_sched, (*pipeline_reg)->get_schd_id()); m_not_ready.set(op); } else m_src_op[op] = op_t(); diff --git a/src/gpgpu-sim/shader.h b/src/gpgpu-sim/shader.h index 95e142e..b1f904f 100644 --- a/src/gpgpu-sim/shader.h +++ b/src/gpgpu-sim/shader.h @@ -1,18 +1,19 @@ // Copyright (c) 2009-2021, Tor M. Aamodt, Wilson W.L. Fung, Andrew Turner, -// Ali Bakhoda, Vijay Kandiah, Nikos Hardavellas, +// Ali Bakhoda, Vijay Kandiah, Nikos Hardavellas, // Mahmoud Khairy, Junrui Pan, Timothy G. Rogers -// The University of British Columbia, Northwestern University, Purdue University -// All rights reserved. +// The University of British Columbia, Northwestern University, Purdue +// University All rights reserved. // // Redistribution and use in source and binary forms, with or without // modification, are permitted provided that the following conditions are met: // -// 1. Redistributions of source code must retain the above copyright notice, this +// 1. Redistributions of source code must retain the above copyright notice, +// this // list of conditions and the following disclaimer; // 2. Redistributions in binary form must reproduce the above copyright notice, // this list of conditions and the following disclaimer in the documentation // and/or other materials provided with the distribution; -// 3. Neither the names of The University of British Columbia, Northwestern +// 3. Neither the names of The University of British Columbia, Northwestern // University nor the names of their contributors may be used to // endorse or promote products derived from this software without specific // prior written permission. @@ -199,7 +200,7 @@ class shd_warp_t { void clear_membar() { m_membar = false; } bool get_membar() const { return m_membar; } virtual address_type get_pc() const { return m_next_pc; } - virtual kernel_info_t* get_kernel_info() const; + virtual kernel_info_t *get_kernel_info() const; void set_next_pc(address_type pc) { m_next_pc = pc; } void store_info_of_last_inst_at_barrier(const warp_inst_t *pI) { @@ -318,12 +319,14 @@ class shd_warp_t { bool m_cdp_dummy; // Ni: LDGDEPBAR barrier support - public: - unsigned int m_ldgdepbar_id; // LDGDEPBAR barrier ID - std::vector> m_ldgdepbar_buf; // LDGDEPBAR barrier buffer - unsigned int m_depbar_start_id; - unsigned int m_depbar_group; - bool m_waiting_ldgsts; // Ni: Whether the warp is waiting for the LDGSTS instrs to finish + public: + unsigned int m_ldgdepbar_id; // LDGDEPBAR barrier ID + std::vector> + m_ldgdepbar_buf; // LDGDEPBAR barrier buffer + unsigned int m_depbar_start_id; + unsigned int m_depbar_group; + bool m_waiting_ldgsts; // Ni: Whether the warp is waiting for the LDGSTS + // instrs to finish }; inline unsigned hw_tid_from_wid(unsigned wid, unsigned warp_size, unsigned i) { @@ -337,8 +340,8 @@ const unsigned WARP_PER_CTA_MAX = 64; typedef std::bitset warp_set_t; unsigned register_bank(int regnum, int wid, unsigned num_banks, - bool sub_core_model, - unsigned banks_per_sched, unsigned sched_id); + bool sub_core_model, unsigned banks_per_sched, + unsigned sched_id); class shader_core_ctx; class shader_core_config; @@ -681,28 +684,26 @@ class opndcoll_rfu_t { // operand collector based register file unit public: op_t() { m_valid = false; } op_t(collector_unit_t *cu, unsigned op, unsigned reg, unsigned num_banks, - bool sub_core_model, - unsigned banks_per_sched, unsigned sched_id) { + bool sub_core_model, unsigned banks_per_sched, unsigned sched_id) { m_valid = true; m_warp = NULL; m_cu = cu; m_operand = op; m_register = reg; m_shced_id = sched_id; - m_bank = register_bank(reg, cu->get_warp_id(), num_banks, - sub_core_model, banks_per_sched, sched_id); + m_bank = register_bank(reg, cu->get_warp_id(), num_banks, sub_core_model, + banks_per_sched, sched_id); } op_t(const warp_inst_t *warp, unsigned reg, unsigned num_banks, - bool sub_core_model, - unsigned banks_per_sched, unsigned sched_id) { + bool sub_core_model, unsigned banks_per_sched, unsigned sched_id) { m_valid = true; m_warp = warp; m_register = reg; m_cu = NULL; m_operand = -1; m_shced_id = sched_id; - m_bank = register_bank(reg, warp->warp_id(), num_banks, - sub_core_model, banks_per_sched, sched_id); + m_bank = register_bank(reg, warp->warp_id(), num_banks, sub_core_model, + banks_per_sched, sched_id); } // accessors @@ -950,9 +951,8 @@ class opndcoll_rfu_t { // operand collector based register file unit unsigned get_reg_id() const { return m_reg_id; } // modifiers - void init(unsigned n, unsigned num_banks, - const core_config *config, opndcoll_rfu_t *rfu, - bool m_sub_core_model, unsigned reg_id, + void init(unsigned n, unsigned num_banks, const core_config *config, + opndcoll_rfu_t *rfu, bool m_sub_core_model, unsigned reg_id, unsigned num_banks_per_sched); bool allocate(register_set *pipeline_reg, register_set *output_reg); @@ -996,8 +996,8 @@ class opndcoll_rfu_t { // operand collector based register file unit // With sub-core enabled round robin starts with the next cu assigned to a // different sub-core than the one that dispatched last unsigned cusPerSched = m_num_collectors / m_num_warp_scheds; - unsigned rr_increment = m_sub_core_model ? - cusPerSched - (m_last_cu % cusPerSched) : 1; + unsigned rr_increment = + m_sub_core_model ? cusPerSched - (m_last_cu % cusPerSched) : 1; for (unsigned n = 0; n < m_num_collectors; n++) { unsigned c = (m_last_cu + n + rr_increment) % m_num_collectors; if ((*m_collector_units)[c].ready()) { @@ -1317,8 +1317,8 @@ class sp_unit : public pipelined_simd_unit { class specialized_unit : public pipelined_simd_unit { public: specialized_unit(register_set *result_port, const shader_core_config *config, - shader_core_ctx *core, int supported_op, - char *unit_name, unsigned latency, unsigned issue_reg_id); + shader_core_ctx *core, int supported_op, char *unit_name, + unsigned latency, unsigned issue_reg_id); virtual bool can_issue(const warp_inst_t &inst) const { if (inst.op != m_supported_op) { return false; @@ -1348,13 +1348,14 @@ class ldst_unit : public pipelined_simd_unit { unsigned sid, unsigned tpc); // Add a structure to record the LDGSTS instructions, - // similar to m_pending_writes, but since LDGSTS does not have a output register - // to write to, so a new structure needs to be added - /* A multi-level map: unsigned (warp_id) -> unsigned (pc) -> unsigned (addr) -> unsigned (count) + // similar to m_pending_writes, but since LDGSTS does not have a output + // register to write to, so a new structure needs to be added + /* A multi-level map: unsigned (warp_id) -> unsigned (pc) -> unsigned (addr) + * -> unsigned (count) */ std::map>> + std::map>> m_pending_ldgsts; // modifiers virtual void issue(register_set &inst); @@ -1766,8 +1767,8 @@ struct shader_core_stats_pod { unsigned *m_active_tensor_core_lanes; unsigned *m_active_fu_lanes; unsigned *m_active_fu_mem_lanes; - double *m_active_exu_threads; //For power model - double *m_active_exu_warps; //For power model + double *m_active_exu_threads; // For power model + double *m_active_exu_warps; // For power model unsigned *m_n_diverge; // number of divergence occurring in this shader unsigned gpgpu_n_load_insn; unsigned gpgpu_n_store_insn; @@ -1838,56 +1839,41 @@ class shader_core_stats : public shader_core_stats_pod { (unsigned *)calloc(config->num_shader(), sizeof(unsigned)); m_num_loadqueued_insn = (unsigned *)calloc(config->num_shader(), sizeof(unsigned)); - m_num_tex_inst = - (unsigned *)calloc(config->num_shader(), sizeof(unsigned)); + m_num_tex_inst = (unsigned *)calloc(config->num_shader(), sizeof(unsigned)); m_num_INTdecoded_insn = (unsigned *)calloc(config->num_shader(), sizeof(unsigned)); - m_num_ialu_acesses = - (double *)calloc(config->num_shader(), sizeof(double)); - m_num_fp_acesses = - (double *)calloc(config->num_shader(), sizeof(double)); - m_num_imul_acesses = - (double *)calloc(config->num_shader(), sizeof(double)); + m_num_ialu_acesses = (double *)calloc(config->num_shader(), sizeof(double)); + m_num_fp_acesses = (double *)calloc(config->num_shader(), sizeof(double)); + m_num_imul_acesses = (double *)calloc(config->num_shader(), sizeof(double)); m_num_imul24_acesses = (double *)calloc(config->num_shader(), sizeof(double)); m_num_imul32_acesses = (double *)calloc(config->num_shader(), sizeof(double)); m_num_fpmul_acesses = (double *)calloc(config->num_shader(), sizeof(double)); - m_num_idiv_acesses = - (double *)calloc(config->num_shader(), sizeof(double)); + m_num_idiv_acesses = (double *)calloc(config->num_shader(), sizeof(double)); m_num_fpdiv_acesses = (double *)calloc(config->num_shader(), sizeof(double)); - m_num_dp_acesses = - (double*) calloc(config->num_shader(),sizeof(double)); - m_num_dpmul_acesses = - (double*) calloc(config->num_shader(),sizeof(double)); - m_num_dpdiv_acesses = - (double*) calloc(config->num_shader(),sizeof(double)); - m_num_sp_acesses = + m_num_dp_acesses = (double *)calloc(config->num_shader(), sizeof(double)); + m_num_dpmul_acesses = (double *)calloc(config->num_shader(), sizeof(double)); - m_num_sfu_acesses = + m_num_dpdiv_acesses = (double *)calloc(config->num_shader(), sizeof(double)); - m_num_tensor_core_acesses = + m_num_sp_acesses = (double *)calloc(config->num_shader(), sizeof(double)); + m_num_sfu_acesses = (double *)calloc(config->num_shader(), sizeof(double)); + m_num_tensor_core_acesses = (double *)calloc(config->num_shader(), sizeof(double)); m_num_const_acesses = (double *)calloc(config->num_shader(), sizeof(double)); - m_num_tex_acesses = - (double *)calloc(config->num_shader(), sizeof(double)); - m_num_sqrt_acesses = - (double*) calloc(config->num_shader(),sizeof(double)); - m_num_log_acesses = - (double*) calloc(config->num_shader(),sizeof(double)); - m_num_sin_acesses = - (double*) calloc(config->num_shader(),sizeof(double)); - m_num_exp_acesses = - (double*) calloc(config->num_shader(),sizeof(double)); - m_num_mem_acesses = - (double *)calloc(config->num_shader(), sizeof(double)); + m_num_tex_acesses = (double *)calloc(config->num_shader(), sizeof(double)); + m_num_sqrt_acesses = (double *)calloc(config->num_shader(), sizeof(double)); + m_num_log_acesses = (double *)calloc(config->num_shader(), sizeof(double)); + m_num_sin_acesses = (double *)calloc(config->num_shader(), sizeof(double)); + m_num_exp_acesses = (double *)calloc(config->num_shader(), sizeof(double)); + m_num_mem_acesses = (double *)calloc(config->num_shader(), sizeof(double)); m_num_sp_committed = (unsigned *)calloc(config->num_shader(), sizeof(unsigned)); - m_num_tlb_hits = - (unsigned *)calloc(config->num_shader(), sizeof(unsigned)); + m_num_tlb_hits = (unsigned *)calloc(config->num_shader(), sizeof(unsigned)); m_num_tlb_accesses = (unsigned *)calloc(config->num_shader(), sizeof(unsigned)); m_active_sp_lanes = @@ -1900,8 +1886,7 @@ class shader_core_stats : public shader_core_stats_pod { (unsigned *)calloc(config->num_shader(), sizeof(unsigned)); m_active_exu_threads = (double *)calloc(config->num_shader(), sizeof(double)); - m_active_exu_warps = - (double *)calloc(config->num_shader(), sizeof(double)); + m_active_exu_warps = (double *)calloc(config->num_shader(), sizeof(double)); m_active_fu_mem_lanes = (unsigned *)calloc(config->num_shader(), sizeof(unsigned)); m_num_sfu_committed = @@ -1916,8 +1901,7 @@ class shader_core_stats : public shader_core_stats_pod { (unsigned *)calloc(config->num_shader(), sizeof(unsigned)); m_non_rf_operands = (unsigned *)calloc(config->num_shader(), sizeof(unsigned)); - m_n_diverge = - (unsigned *)calloc(config->num_shader(), sizeof(unsigned)); + m_n_diverge = (unsigned *)calloc(config->num_shader(), sizeof(unsigned)); shader_cycle_distro = (unsigned *)calloc(config->warp_size + 3, sizeof(unsigned)); last_shader_cycle_distro = @@ -2153,206 +2137,244 @@ class shader_core_ctx : public core_t { void incload_stat() { m_stats->m_num_loadqueued_insn[m_sid]++; } void incstore_stat() { m_stats->m_num_storequeued_insn[m_sid]++; } - void incialu_stat(unsigned active_count,double latency) { - if(m_config->gpgpu_clock_gated_lanes==false){ - m_stats->m_num_ialu_acesses[m_sid]=m_stats->m_num_ialu_acesses[m_sid]+(double)active_count*latency - + inactive_lanes_accesses_nonsfu(active_count, latency); - }else { - m_stats->m_num_ialu_acesses[m_sid]=m_stats->m_num_ialu_acesses[m_sid]+(double)active_count*latency; - } - m_stats->m_active_exu_threads[m_sid]+=active_count; + void incialu_stat(unsigned active_count, double latency) { + if (m_config->gpgpu_clock_gated_lanes == false) { + m_stats->m_num_ialu_acesses[m_sid] = + m_stats->m_num_ialu_acesses[m_sid] + (double)active_count * latency + + inactive_lanes_accesses_nonsfu(active_count, latency); + } else { + m_stats->m_num_ialu_acesses[m_sid] = + m_stats->m_num_ialu_acesses[m_sid] + (double)active_count * latency; + } + m_stats->m_active_exu_threads[m_sid] += active_count; m_stats->m_active_exu_warps[m_sid]++; } - void incimul_stat(unsigned active_count,double latency) { - if(m_config->gpgpu_clock_gated_lanes==false){ - m_stats->m_num_imul_acesses[m_sid]=m_stats->m_num_imul_acesses[m_sid]+(double)active_count*latency - + inactive_lanes_accesses_nonsfu(active_count, latency); - }else { - m_stats->m_num_imul_acesses[m_sid]=m_stats->m_num_imul_acesses[m_sid]+(double)active_count*latency; + void incimul_stat(unsigned active_count, double latency) { + if (m_config->gpgpu_clock_gated_lanes == false) { + m_stats->m_num_imul_acesses[m_sid] = + m_stats->m_num_imul_acesses[m_sid] + (double)active_count * latency + + inactive_lanes_accesses_nonsfu(active_count, latency); + } else { + m_stats->m_num_imul_acesses[m_sid] = + m_stats->m_num_imul_acesses[m_sid] + (double)active_count * latency; } - m_stats->m_active_exu_threads[m_sid]+=active_count; + m_stats->m_active_exu_threads[m_sid] += active_count; m_stats->m_active_exu_warps[m_sid]++; } - void incimul24_stat(unsigned active_count,double latency) { - if(m_config->gpgpu_clock_gated_lanes==false){ - m_stats->m_num_imul24_acesses[m_sid]=m_stats->m_num_imul24_acesses[m_sid]+(double)active_count*latency - + inactive_lanes_accesses_nonsfu(active_count, latency); - }else { - m_stats->m_num_imul24_acesses[m_sid]=m_stats->m_num_imul24_acesses[m_sid]+(double)active_count*latency; - } - m_stats->m_active_exu_threads[m_sid]+=active_count; - m_stats->m_active_exu_warps[m_sid]++; - } - void incimul32_stat(unsigned active_count,double latency) { - if(m_config->gpgpu_clock_gated_lanes==false){ - m_stats->m_num_imul32_acesses[m_sid]=m_stats->m_num_imul32_acesses[m_sid]+(double)active_count*latency - + inactive_lanes_accesses_sfu(active_count, latency); - }else{ - m_stats->m_num_imul32_acesses[m_sid]=m_stats->m_num_imul32_acesses[m_sid]+(double)active_count*latency; - } - m_stats->m_active_exu_threads[m_sid]+=active_count; + void incimul24_stat(unsigned active_count, double latency) { + if (m_config->gpgpu_clock_gated_lanes == false) { + m_stats->m_num_imul24_acesses[m_sid] = + m_stats->m_num_imul24_acesses[m_sid] + + (double)active_count * latency + + inactive_lanes_accesses_nonsfu(active_count, latency); + } else { + m_stats->m_num_imul24_acesses[m_sid] = + m_stats->m_num_imul24_acesses[m_sid] + (double)active_count * latency; + } + m_stats->m_active_exu_threads[m_sid] += active_count; m_stats->m_active_exu_warps[m_sid]++; } - void incidiv_stat(unsigned active_count,double latency) { - if(m_config->gpgpu_clock_gated_lanes==false){ - m_stats->m_num_idiv_acesses[m_sid]=m_stats->m_num_idiv_acesses[m_sid]+(double)active_count*latency - + inactive_lanes_accesses_sfu(active_count, latency); - }else { - m_stats->m_num_idiv_acesses[m_sid]=m_stats->m_num_idiv_acesses[m_sid]+(double)active_count*latency; - } - m_stats->m_active_exu_threads[m_sid]+=active_count; - m_stats->m_active_exu_warps[m_sid]++; - } - void incfpalu_stat(unsigned active_count,double latency) { - if(m_config->gpgpu_clock_gated_lanes==false){ - m_stats->m_num_fp_acesses[m_sid]=m_stats->m_num_fp_acesses[m_sid]+(double)active_count*latency - + inactive_lanes_accesses_nonsfu(active_count, latency); - }else { - m_stats->m_num_fp_acesses[m_sid]=m_stats->m_num_fp_acesses[m_sid]+(double)active_count*latency; - } - m_stats->m_active_exu_threads[m_sid]+=active_count; - m_stats->m_active_exu_warps[m_sid]++; - } - void incfpmul_stat(unsigned active_count,double latency) { - // printf("FP MUL stat increament\n"); - if(m_config->gpgpu_clock_gated_lanes==false){ - m_stats->m_num_fpmul_acesses[m_sid]=m_stats->m_num_fpmul_acesses[m_sid]+(double)active_count*latency - + inactive_lanes_accesses_nonsfu(active_count, latency); - }else { - m_stats->m_num_fpmul_acesses[m_sid]=m_stats->m_num_fpmul_acesses[m_sid]+(double)active_count*latency; - } - m_stats->m_active_exu_threads[m_sid]+=active_count; + void incimul32_stat(unsigned active_count, double latency) { + if (m_config->gpgpu_clock_gated_lanes == false) { + m_stats->m_num_imul32_acesses[m_sid] = + m_stats->m_num_imul32_acesses[m_sid] + + (double)active_count * latency + + inactive_lanes_accesses_sfu(active_count, latency); + } else { + m_stats->m_num_imul32_acesses[m_sid] = + m_stats->m_num_imul32_acesses[m_sid] + (double)active_count * latency; + } + m_stats->m_active_exu_threads[m_sid] += active_count; m_stats->m_active_exu_warps[m_sid]++; - } - void incfpdiv_stat(unsigned active_count,double latency) { - if(m_config->gpgpu_clock_gated_lanes==false){ - m_stats->m_num_fpdiv_acesses[m_sid]=m_stats->m_num_fpdiv_acesses[m_sid]+(double)active_count*latency - + inactive_lanes_accesses_sfu(active_count, latency); - }else { - m_stats->m_num_fpdiv_acesses[m_sid]=m_stats->m_num_fpdiv_acesses[m_sid]+(double)active_count*latency; - } - m_stats->m_active_exu_threads[m_sid]+=active_count; + } + void incidiv_stat(unsigned active_count, double latency) { + if (m_config->gpgpu_clock_gated_lanes == false) { + m_stats->m_num_idiv_acesses[m_sid] = + m_stats->m_num_idiv_acesses[m_sid] + (double)active_count * latency + + inactive_lanes_accesses_sfu(active_count, latency); + } else { + m_stats->m_num_idiv_acesses[m_sid] = + m_stats->m_num_idiv_acesses[m_sid] + (double)active_count * latency; + } + m_stats->m_active_exu_threads[m_sid] += active_count; m_stats->m_active_exu_warps[m_sid]++; - } - void incdpalu_stat(unsigned active_count,double latency) { - if(m_config->gpgpu_clock_gated_lanes==false){ - m_stats->m_num_dp_acesses[m_sid]=m_stats->m_num_dp_acesses[m_sid]+(double)active_count*latency - + inactive_lanes_accesses_nonsfu(active_count, latency); - }else { - m_stats->m_num_dp_acesses[m_sid]=m_stats->m_num_dp_acesses[m_sid]+(double)active_count*latency; - } - m_stats->m_active_exu_threads[m_sid]+=active_count; - m_stats->m_active_exu_warps[m_sid]++; - } - void incdpmul_stat(unsigned active_count,double latency) { - // printf("FP MUL stat increament\n"); - if(m_config->gpgpu_clock_gated_lanes==false){ - m_stats->m_num_dpmul_acesses[m_sid]=m_stats->m_num_dpmul_acesses[m_sid]+(double)active_count*latency - + inactive_lanes_accesses_nonsfu(active_count, latency); - }else { - m_stats->m_num_dpmul_acesses[m_sid]=m_stats->m_num_dpmul_acesses[m_sid]+(double)active_count*latency; - } - m_stats->m_active_exu_threads[m_sid]+=active_count; + } + void incfpalu_stat(unsigned active_count, double latency) { + if (m_config->gpgpu_clock_gated_lanes == false) { + m_stats->m_num_fp_acesses[m_sid] = + m_stats->m_num_fp_acesses[m_sid] + (double)active_count * latency + + inactive_lanes_accesses_nonsfu(active_count, latency); + } else { + m_stats->m_num_fp_acesses[m_sid] = + m_stats->m_num_fp_acesses[m_sid] + (double)active_count * latency; + } + m_stats->m_active_exu_threads[m_sid] += active_count; m_stats->m_active_exu_warps[m_sid]++; - } - void incdpdiv_stat(unsigned active_count,double latency) { - if(m_config->gpgpu_clock_gated_lanes==false){ - m_stats->m_num_dpdiv_acesses[m_sid]=m_stats->m_num_dpdiv_acesses[m_sid]+(double)active_count*latency - + inactive_lanes_accesses_sfu(active_count, latency); - }else { - m_stats->m_num_dpdiv_acesses[m_sid]=m_stats->m_num_dpdiv_acesses[m_sid]+(double)active_count*latency; - } - m_stats->m_active_exu_threads[m_sid]+=active_count; + } + void incfpmul_stat(unsigned active_count, double latency) { + // printf("FP MUL stat increament\n"); + if (m_config->gpgpu_clock_gated_lanes == false) { + m_stats->m_num_fpmul_acesses[m_sid] = + m_stats->m_num_fpmul_acesses[m_sid] + (double)active_count * latency + + inactive_lanes_accesses_nonsfu(active_count, latency); + } else { + m_stats->m_num_fpmul_acesses[m_sid] = + m_stats->m_num_fpmul_acesses[m_sid] + (double)active_count * latency; + } + m_stats->m_active_exu_threads[m_sid] += active_count; m_stats->m_active_exu_warps[m_sid]++; - } - - void incsqrt_stat(unsigned active_count,double latency) { - if(m_config->gpgpu_clock_gated_lanes==false){ - m_stats->m_num_sqrt_acesses[m_sid]=m_stats->m_num_sqrt_acesses[m_sid]+(double)active_count*latency - + inactive_lanes_accesses_sfu(active_count, latency); - }else{ - m_stats->m_num_sqrt_acesses[m_sid]=m_stats->m_num_sqrt_acesses[m_sid]+(double)active_count*latency; + } + void incfpdiv_stat(unsigned active_count, double latency) { + if (m_config->gpgpu_clock_gated_lanes == false) { + m_stats->m_num_fpdiv_acesses[m_sid] = + m_stats->m_num_fpdiv_acesses[m_sid] + (double)active_count * latency + + inactive_lanes_accesses_sfu(active_count, latency); + } else { + m_stats->m_num_fpdiv_acesses[m_sid] = + m_stats->m_num_fpdiv_acesses[m_sid] + (double)active_count * latency; + } + m_stats->m_active_exu_threads[m_sid] += active_count; + m_stats->m_active_exu_warps[m_sid]++; + } + void incdpalu_stat(unsigned active_count, double latency) { + if (m_config->gpgpu_clock_gated_lanes == false) { + m_stats->m_num_dp_acesses[m_sid] = + m_stats->m_num_dp_acesses[m_sid] + (double)active_count * latency + + inactive_lanes_accesses_nonsfu(active_count, latency); + } else { + m_stats->m_num_dp_acesses[m_sid] = + m_stats->m_num_dp_acesses[m_sid] + (double)active_count * latency; + } + m_stats->m_active_exu_threads[m_sid] += active_count; + m_stats->m_active_exu_warps[m_sid]++; + } + void incdpmul_stat(unsigned active_count, double latency) { + // printf("FP MUL stat increament\n"); + if (m_config->gpgpu_clock_gated_lanes == false) { + m_stats->m_num_dpmul_acesses[m_sid] = + m_stats->m_num_dpmul_acesses[m_sid] + (double)active_count * latency + + inactive_lanes_accesses_nonsfu(active_count, latency); + } else { + m_stats->m_num_dpmul_acesses[m_sid] = + m_stats->m_num_dpmul_acesses[m_sid] + (double)active_count * latency; + } + m_stats->m_active_exu_threads[m_sid] += active_count; + m_stats->m_active_exu_warps[m_sid]++; + } + void incdpdiv_stat(unsigned active_count, double latency) { + if (m_config->gpgpu_clock_gated_lanes == false) { + m_stats->m_num_dpdiv_acesses[m_sid] = + m_stats->m_num_dpdiv_acesses[m_sid] + (double)active_count * latency + + inactive_lanes_accesses_sfu(active_count, latency); + } else { + m_stats->m_num_dpdiv_acesses[m_sid] = + m_stats->m_num_dpdiv_acesses[m_sid] + (double)active_count * latency; } - m_stats->m_active_exu_threads[m_sid]+=active_count; + m_stats->m_active_exu_threads[m_sid] += active_count; m_stats->m_active_exu_warps[m_sid]++; - } + } - void inclog_stat(unsigned active_count,double latency) { - if(m_config->gpgpu_clock_gated_lanes==false){ - m_stats->m_num_log_acesses[m_sid]=m_stats->m_num_log_acesses[m_sid]+(double)active_count*latency - + inactive_lanes_accesses_sfu(active_count, latency); - }else{ - m_stats->m_num_log_acesses[m_sid]=m_stats->m_num_log_acesses[m_sid]+(double)active_count*latency; + void incsqrt_stat(unsigned active_count, double latency) { + if (m_config->gpgpu_clock_gated_lanes == false) { + m_stats->m_num_sqrt_acesses[m_sid] = + m_stats->m_num_sqrt_acesses[m_sid] + (double)active_count * latency + + inactive_lanes_accesses_sfu(active_count, latency); + } else { + m_stats->m_num_sqrt_acesses[m_sid] = + m_stats->m_num_sqrt_acesses[m_sid] + (double)active_count * latency; } - m_stats->m_active_exu_threads[m_sid]+=active_count; + m_stats->m_active_exu_threads[m_sid] += active_count; m_stats->m_active_exu_warps[m_sid]++; - } + } - void incexp_stat(unsigned active_count,double latency) { - if(m_config->gpgpu_clock_gated_lanes==false){ - m_stats->m_num_exp_acesses[m_sid]=m_stats->m_num_exp_acesses[m_sid]+(double)active_count*latency - + inactive_lanes_accesses_sfu(active_count, latency); - }else{ - m_stats->m_num_exp_acesses[m_sid]=m_stats->m_num_exp_acesses[m_sid]+(double)active_count*latency; + void inclog_stat(unsigned active_count, double latency) { + if (m_config->gpgpu_clock_gated_lanes == false) { + m_stats->m_num_log_acesses[m_sid] = + m_stats->m_num_log_acesses[m_sid] + (double)active_count * latency + + inactive_lanes_accesses_sfu(active_count, latency); + } else { + m_stats->m_num_log_acesses[m_sid] = + m_stats->m_num_log_acesses[m_sid] + (double)active_count * latency; } - m_stats->m_active_exu_threads[m_sid]+=active_count; + m_stats->m_active_exu_threads[m_sid] += active_count; m_stats->m_active_exu_warps[m_sid]++; } - void incsin_stat(unsigned active_count,double latency) { - if(m_config->gpgpu_clock_gated_lanes==false){ - m_stats->m_num_sin_acesses[m_sid]=m_stats->m_num_sin_acesses[m_sid]+(double)active_count*latency - + inactive_lanes_accesses_sfu(active_count, latency); - }else{ - m_stats->m_num_sin_acesses[m_sid]=m_stats->m_num_sin_acesses[m_sid]+(double)active_count*latency; + void incexp_stat(unsigned active_count, double latency) { + if (m_config->gpgpu_clock_gated_lanes == false) { + m_stats->m_num_exp_acesses[m_sid] = + m_stats->m_num_exp_acesses[m_sid] + (double)active_count * latency + + inactive_lanes_accesses_sfu(active_count, latency); + } else { + m_stats->m_num_exp_acesses[m_sid] = + m_stats->m_num_exp_acesses[m_sid] + (double)active_count * latency; } - m_stats->m_active_exu_threads[m_sid]+=active_count; + m_stats->m_active_exu_threads[m_sid] += active_count; m_stats->m_active_exu_warps[m_sid]++; } + void incsin_stat(unsigned active_count, double latency) { + if (m_config->gpgpu_clock_gated_lanes == false) { + m_stats->m_num_sin_acesses[m_sid] = + m_stats->m_num_sin_acesses[m_sid] + (double)active_count * latency + + inactive_lanes_accesses_sfu(active_count, latency); + } else { + m_stats->m_num_sin_acesses[m_sid] = + m_stats->m_num_sin_acesses[m_sid] + (double)active_count * latency; + } + m_stats->m_active_exu_threads[m_sid] += active_count; + m_stats->m_active_exu_warps[m_sid]++; + } - void inctensor_stat(unsigned active_count,double latency) { - if(m_config->gpgpu_clock_gated_lanes==false){ - m_stats->m_num_tensor_core_acesses[m_sid]=m_stats->m_num_tensor_core_acesses[m_sid]+(double)active_count*latency - + inactive_lanes_accesses_sfu(active_count, latency); - }else{ - m_stats->m_num_tensor_core_acesses[m_sid]=m_stats->m_num_tensor_core_acesses[m_sid]+(double)active_count*latency; + void inctensor_stat(unsigned active_count, double latency) { + if (m_config->gpgpu_clock_gated_lanes == false) { + m_stats->m_num_tensor_core_acesses[m_sid] = + m_stats->m_num_tensor_core_acesses[m_sid] + + (double)active_count * latency + + inactive_lanes_accesses_sfu(active_count, latency); + } else { + m_stats->m_num_tensor_core_acesses[m_sid] = + m_stats->m_num_tensor_core_acesses[m_sid] + + (double)active_count * latency; } - m_stats->m_active_exu_threads[m_sid]+=active_count; + m_stats->m_active_exu_threads[m_sid] += active_count; m_stats->m_active_exu_warps[m_sid]++; } - void inctex_stat(unsigned active_count,double latency) { - if(m_config->gpgpu_clock_gated_lanes==false){ - m_stats->m_num_tex_acesses[m_sid]=m_stats->m_num_tex_acesses[m_sid]+(double)active_count*latency - + inactive_lanes_accesses_sfu(active_count, latency); - }else{ - m_stats->m_num_tex_acesses[m_sid]=m_stats->m_num_tex_acesses[m_sid]+(double)active_count*latency; + void inctex_stat(unsigned active_count, double latency) { + if (m_config->gpgpu_clock_gated_lanes == false) { + m_stats->m_num_tex_acesses[m_sid] = + m_stats->m_num_tex_acesses[m_sid] + (double)active_count * latency + + inactive_lanes_accesses_sfu(active_count, latency); + } else { + m_stats->m_num_tex_acesses[m_sid] = + m_stats->m_num_tex_acesses[m_sid] + (double)active_count * latency; } - m_stats->m_active_exu_threads[m_sid]+=active_count; + m_stats->m_active_exu_threads[m_sid] += active_count; m_stats->m_active_exu_warps[m_sid]++; } void inc_const_accesses(unsigned active_count) { - m_stats->m_num_const_acesses[m_sid]=m_stats->m_num_const_acesses[m_sid]+active_count; + m_stats->m_num_const_acesses[m_sid] = + m_stats->m_num_const_acesses[m_sid] + active_count; } void incsfu_stat(unsigned active_count, double latency) { m_stats->m_num_sfu_acesses[m_sid] = - m_stats->m_num_sfu_acesses[m_sid] + (double)active_count*latency; + m_stats->m_num_sfu_acesses[m_sid] + (double)active_count * latency; } void incsp_stat(unsigned active_count, double latency) { m_stats->m_num_sp_acesses[m_sid] = - m_stats->m_num_sp_acesses[m_sid] + (double)active_count*latency; + m_stats->m_num_sp_acesses[m_sid] + (double)active_count * latency; } void incmem_stat(unsigned active_count, double latency) { if (m_config->gpgpu_clock_gated_lanes == false) { m_stats->m_num_mem_acesses[m_sid] = - m_stats->m_num_mem_acesses[m_sid] + (double)active_count*latency + + m_stats->m_num_mem_acesses[m_sid] + (double)active_count * latency + inactive_lanes_accesses_nonsfu(active_count, latency); } else { m_stats->m_num_mem_acesses[m_sid] = - m_stats->m_num_mem_acesses[m_sid] + (double)active_count*latency; + m_stats->m_num_mem_acesses[m_sid] + (double)active_count * latency; } } void incexecstat(warp_inst_t *&inst); -- cgit v1.3 From 42a0cde4b463794d041b544309afb69c315f78bc Mon Sep 17 00:00:00 2001 From: JRPan <25518778+JRPan@users.noreply.github.com> Date: Fri, 16 Aug 2024 11:16:06 -0400 Subject: remove unused & unintilized variables & format automation (#69) * run formatter only on PR * remove unused & unintilized variables * fix signed & unsigned comparison warning * enable merge queue * resolve conflict * in formatter, checkout the forked repo, not the base repo in PR * Try to use jenkins for formatter * Automated Format --------- Co-authored-by: purdue-jenkins --- .github/workflows/main.yml | 29 ++++++++++++----------------- src/cuda-sim/instructions.cc | 16 ++++++++-------- src/cuda-sim/ptx_ir.h | 2 +- src/gpgpu-sim/gpu-sim.h | 6 +++--- src/gpgpu-sim/shader.h | 4 ++-- 5 files changed, 26 insertions(+), 31 deletions(-) (limited to 'src/gpgpu-sim/gpu-sim.h') diff --git a/.github/workflows/main.yml b/.github/workflows/main.yml index 39f65c9..8e0ae23 100644 --- a/.github/workflows/main.yml +++ b/.github/workflows/main.yml @@ -4,7 +4,6 @@ name: Short-Tests # Controls when the workflow will run on: - # Triggers the workflow on push or pull request events but only for the mydev branch push: branches-ignore: - "gh-readonly-queue**" @@ -86,25 +85,21 @@ jobs: - name: Run Simulation run: /bin/bash $GITHUB_WORKSPACE/short-tests.sh format-code: - runs-on: ubuntu-latest + if: github.event_name == 'pull_request' + runs-on: tgrogers-raid needs: [build-TITANV, build-TITANV-LOCALXBAR, build-QV100, build-2060, build-3070] - permissions: - # Give the default GITHUB_TOKEN write permission to commit and push the - # added or changed files to the repository. - contents: write steps: - uses: actions/checkout@v4 - # Other steps that change files in the repository go here - # … + with: + ref: ${{github.event.pull_request.head.ref}} + repository: ${{github.event.pull_request.head.repo.full_name}} + ssh-key: '' + - name: Run clang-format run: | - sudo apt-get install -y clang-format + git config user.name "purdue-jenkins" + git config user.email "purdue-jenkins@users.noreply.github.com" + git remote set-url origin git@github.com:${{github.event.pull_request.head.repo.full_name}} + git remote -v /bin/bash ./format-code.sh - - uses: stefanzweifel/git-auto-commit-action@v5 - with: - # Optional. Commit message for the created commit. - # Defaults to "Apply automatic changes" - commit_message: Automated clang-format - # Optional. Option used by `git-status` to determine if the repository is - # dirty. See https://git-scm.com/docs/git-status#_options - status_options: '--untracked-files=no' \ No newline at end of file + if git status --untracked-files=no | grep -q "nothing to commit"; then echo "No changes to commit."; else git commit -a -m "Automated Format"; git push; fi diff --git a/src/cuda-sim/instructions.cc b/src/cuda-sim/instructions.cc index 2314bef..108de97 100644 --- a/src/cuda-sim/instructions.cc +++ b/src/cuda-sim/instructions.cc @@ -5442,33 +5442,33 @@ void shfl_impl(const ptx_instruction *pI, core_t *core, warp_inst_t inst) { } void shf_impl(const ptx_instruction *pI, ptx_thread_info *thread) { - ptx_reg_t a,b,c,d; + ptx_reg_t a, b, c, d; const operand_info &dst = pI->dst(); const operand_info &src1 = pI->src1(); const operand_info &src2 = pI->src2(); const operand_info &src3 = pI->src3(); - // Only b32 is allowed + // Only b32 is allowed unsigned i_type = pI->get_type(); a = thread->get_operand_value(src1, dst, i_type, thread, 1); b = thread->get_operand_value(src2, dst, i_type, thread, 1); c = thread->get_operand_value(src3, dst, i_type, thread, 1); - if(i_type != B32_TYPE) + if (i_type != B32_TYPE) printf("Only the b32 data_type is allowed per the ISA\n"); unsigned clamp_mode = pI->clamp_mode(); unsigned n = c.u32 & 0x1f; - if(clamp_mode) { - if(c.u32 < 32) + if (clamp_mode) { + if (c.u32 < 32) n = c; else n = 32; } - if(pI->left_mode()) - d.u32 = (b.u32 << n) | (a.u32 >> (32-n)); + if (pI->left_mode()) + d.u32 = (b.u32 << n) | (a.u32 >> (32 - n)); else - d.u32 = (b.u32 << (32-n)) | (a.u32 >> n); + d.u32 = (b.u32 << (32 - n)) | (a.u32 >> n); thread->set_operand_value(dst, d, i_type, thread, pI); } diff --git a/src/cuda-sim/ptx_ir.h b/src/cuda-sim/ptx_ir.h index 46f183b..d253866 100644 --- a/src/cuda-sim/ptx_ir.h +++ b/src/cuda-sim/ptx_ir.h @@ -1085,7 +1085,7 @@ class ptx_instruction : public warp_inst_t { unsigned cache_option() const { return m_cache_option; } unsigned rounding_mode() const { return m_rounding_mode; } unsigned saturation_mode() const { return m_saturation_mode; } - unsigned clamp_mode() const {return m_clamp_mode;} + unsigned clamp_mode() const { return m_clamp_mode; } unsigned left_mode() const { return m_left_mode; } unsigned dimension() const { return m_geom_spec; } unsigned barrier_op() const { return m_barrier_op; } diff --git a/src/gpgpu-sim/gpu-sim.h b/src/gpgpu-sim/gpu-sim.h index a24ffd3..d43b399 100644 --- a/src/gpgpu-sim/gpu-sim.h +++ b/src/gpgpu-sim/gpu-sim.h @@ -132,9 +132,9 @@ struct power_config { // NOTE: After changing the nonlinear model to only scaling idle core, // NOTE: The min_inc_per_active_sm is not used any more - if (g_use_nonlinear_model) - sscanf(gpu_nonlinear_model_config, "%lf:%lf", &gpu_idle_core_power, - &gpu_min_inc_per_active_sm); + // if (g_use_nonlinear_model) + // sscanf(gpu_nonlinear_model_config, "%lf:%lf", &gpu_idle_core_power, + // &gpu_min_inc_per_active_sm); } void reg_options(class OptionParser *opp); diff --git a/src/gpgpu-sim/shader.h b/src/gpgpu-sim/shader.h index b1f904f..92691d3 100644 --- a/src/gpgpu-sim/shader.h +++ b/src/gpgpu-sim/shader.h @@ -134,7 +134,7 @@ class shd_warp_t { m_waiting_ldgsts = false; // Ni: Clear m_ldgdepbar_buf - for (int i = 0; i < m_ldgdepbar_buf.size(); i++) { + for (unsigned i = 0; i < m_ldgdepbar_buf.size(); i++) { m_ldgdepbar_buf[i].clear(); } m_ldgdepbar_buf.clear(); @@ -165,7 +165,7 @@ class shd_warp_t { m_waiting_ldgsts = false; // Ni: Clear m_ldgdepbar_buf - for (int i = 0; i < m_ldgdepbar_buf.size(); i++) { + for (unsigned i = 0; i < m_ldgdepbar_buf.size(); i++) { m_ldgdepbar_buf[i].clear(); } m_ldgdepbar_buf.clear(); -- cgit v1.3 From 38b4df5653ecbd9907a3d39b125640cd4fb7d012 Mon Sep 17 00:00:00 2001 From: JRPan <25518778+JRPan@users.noreply.github.com> Date: Tue, 20 Aug 2024 20:43:32 -0400 Subject: Stream stats (#71) * Temp commit for Justin and Cassie to sync on code changes for adding per-stream status. * Resolved compile errors. * Removed redundant parameter * Passed cuda_stream_id from accelsim to gpgpusim * Cleaned up unused changes * Changed vector to map, having operator problems. * StreamID defaults to zero * Implemented streams to inc_stats and so on * Fixed TOTAL_ACCESS counts * Implemented GLOBAL_TIMER. * Fixed m_shader->get_kernel SEGFAULT issue in shader.cc. * Use warp_init to track streamID instead of issue_warp * Removed temp debug print * Modified cache_stats to only print data from latest finished stream Added optional arg to cache_stats::print_stats, cache_stats::print_fail_stats and their upstream functions. When streamID is specified, print stats from that stream. When not specified, print all stats. NOTE: current implementation depending on streamid never equals -1 * Removed default arg values of streamID * modified constructor of mem_fetch to pass in streamID * changed get_streamid to get_streamID * Added TODO to gpgpusim_entrypoint.cc and power_stat.cc * Only collect power stats when enabled * print last finished stream in PTX mode using last_streamID * take out additional printf * Add a field to baseline cache to indicate cache level * save gpu object in cache * Print stream ID only once per kernel * rm test print * use -1 for default stream id * cleanup debug prints * remove GLOABL_TIMER * Automated clang-format * Should be correct to print everything in power model * addressing concerns & errors * Automated clang-format * add m_stats_pw in operator+ * Automated Format --------- Co-authored-by: Justin Qiao Co-authored-by: Justin Qiao <71228724+ShichenQiao@users.noreply.github.com> Co-authored-by: Tim Rogers Co-authored-by: JRPan Co-authored-by: purdue-jenkins --- libcuda/cuda_runtime_api.cc | 4 +- src/abstract_hardware_model.cc | 8 +- src/abstract_hardware_model.h | 21 +- src/gpgpu-sim/gpu-cache.cc | 533 +++++++++++++++++++++++++++++------------ src/gpgpu-sim/gpu-cache.h | 77 ++++-- src/gpgpu-sim/gpu-sim.cc | 72 ++++-- src/gpgpu-sim/gpu-sim.h | 18 +- src/gpgpu-sim/l2cache.cc | 29 +-- src/gpgpu-sim/l2cache.h | 7 +- src/gpgpu-sim/mem_fetch.cc | 9 +- src/gpgpu-sim/mem_fetch.h | 10 +- src/gpgpu-sim/power_stat.cc | 6 +- src/gpgpu-sim/shader.cc | 42 ++-- src/gpgpu-sim/shader.h | 19 +- src/gpgpusim_entrypoint.cc | 6 +- src/stream_manager.cc | 12 +- 16 files changed, 613 insertions(+), 260 deletions(-) (limited to 'src/gpgpu-sim/gpu-sim.h') diff --git a/libcuda/cuda_runtime_api.cc b/libcuda/cuda_runtime_api.cc index b64c3d9..b540ffd 100644 --- a/libcuda/cuda_runtime_api.cc +++ b/libcuda/cuda_runtime_api.cc @@ -1421,7 +1421,9 @@ cudaOccupancyMaxActiveBlocksPerMultiprocessorWithFlagsInternal( dim3 gridDim(context->get_device()->get_gpgpu()->max_cta_per_core() * context->get_device()->get_gpgpu()->get_config().num_shader()); dim3 blockDim(blockSize); - kernel_info_t result(gridDim, blockDim, entry); + // because this fuction is only checking for resource requirements, we do + // not care which stream this kernel runs at, just picked -1 + kernel_info_t result(gridDim, blockDim, entry, -1); // if(entry == NULL){ // *numBlocks = 1; // return g_last_cudaError = cudaErrorUnknown; diff --git a/src/abstract_hardware_model.cc b/src/abstract_hardware_model.cc index fd056c6..e8ddf95 100644 --- a/src/abstract_hardware_model.cc +++ b/src/abstract_hardware_model.cc @@ -49,12 +49,14 @@ void mem_access_t::init(gpgpu_context *ctx) { m_addr = 0; m_req_size = 0; } + void warp_inst_t::issue(const active_mask_t &mask, unsigned warp_id, unsigned long long cycle, int dynamic_warp_id, - int sch_id) { + int sch_id, unsigned long long streamID) { m_warp_active_mask = mask; m_warp_issued_mask = mask; m_uid = ++(m_config->gpgpu_ctx->warp_inst_sm_next_uid); + m_streamID = streamID; m_warp_id = warp_id; m_dynamic_warp_id = dynamic_warp_id; issue_cycle = cycle; @@ -755,7 +757,8 @@ void warp_inst_t::completed(unsigned long long cycle) const { } kernel_info_t::kernel_info_t(dim3 gridDim, dim3 blockDim, - class function_info *entry) { + class function_info *entry, + unsigned long long streamID) { m_kernel_entry = entry; m_grid_dim = gridDim; m_block_dim = blockDim; @@ -765,6 +768,7 @@ kernel_info_t::kernel_info_t(dim3 gridDim, dim3 blockDim, m_next_tid = m_next_cta; m_num_cores_running = 0; m_uid = (entry->gpgpu_ctx->kernel_info_m_next_uid)++; + m_streamID = streamID; m_param_mem = new memory_space_impl<8192>("param", 64 * 1024); // Jin: parent and child kernel management for CDP diff --git a/src/abstract_hardware_model.h b/src/abstract_hardware_model.h index e5f3b78..98a4039 100644 --- a/src/abstract_hardware_model.h +++ b/src/abstract_hardware_model.h @@ -233,7 +233,8 @@ class kernel_info_t { // m_num_cores_running=0; // m_param_mem=NULL; // } - kernel_info_t(dim3 gridDim, dim3 blockDim, class function_info *entry); + kernel_info_t(dim3 gridDim, dim3 blockDim, class function_info *entry, + unsigned long long streamID); kernel_info_t( dim3 gridDim, dim3 blockDim, class function_info *entry, std::map nameToCudaArray, @@ -292,6 +293,7 @@ class kernel_info_t { m_next_tid.x < m_block_dim.x; } unsigned get_uid() const { return m_uid; } + unsigned long long get_streamID() const { return m_streamID; } std::string get_name() const { return name(); } std::string name() const; @@ -325,7 +327,8 @@ class kernel_info_t { class function_info *m_kernel_entry; - unsigned m_uid; + unsigned m_uid; // Kernel ID + unsigned long long m_streamID; // These maps contain the snapshot of the texture mappings at kernel launch std::map m_NameToCudaArray; @@ -900,8 +903,8 @@ class mem_fetch_interface { class mem_fetch_allocator { public: virtual mem_fetch *alloc(new_addr_type addr, mem_access_type type, - unsigned size, bool wr, - unsigned long long cycle) const = 0; + unsigned size, bool wr, unsigned long long cycle, + unsigned long long streamID) const = 0; virtual mem_fetch *alloc(const class warp_inst_t &inst, const mem_access_t &access, unsigned long long cycle) const = 0; @@ -911,7 +914,8 @@ class mem_fetch_allocator { const mem_access_sector_mask_t §or_mask, unsigned size, bool wr, unsigned long long cycle, unsigned wid, unsigned sid, unsigned tpc, - mem_fetch *original_mf) const = 0; + mem_fetch *original_mf, + unsigned long long streamID) const = 0; }; // the maximum number of destination, source, or address uarch operands in a @@ -1059,6 +1063,7 @@ class warp_inst_t : public inst_t { // constructors warp_inst_t() { m_uid = 0; + m_streamID = (unsigned long long)-1; m_empty = true; m_config = NULL; @@ -1071,6 +1076,7 @@ class warp_inst_t : public inst_t { } warp_inst_t(const core_config *config) { m_uid = 0; + m_streamID = (unsigned long long)-1; assert(config->warp_size <= MAX_WARP_SIZE); m_config = config; m_empty = true; @@ -1098,7 +1104,8 @@ class warp_inst_t : public inst_t { void clear() { m_empty = true; } void issue(const active_mask_t &mask, unsigned warp_id, - unsigned long long cycle, int dynamic_warp_id, int sch_id); + unsigned long long cycle, int dynamic_warp_id, int sch_id, + unsigned long long streamID); const active_mask_t &get_active_mask() const { return m_warp_active_mask; } void completed(unsigned long long cycle) @@ -1226,11 +1233,13 @@ class warp_inst_t : public inst_t { void print(FILE *fout) const; unsigned get_uid() const { return m_uid; } + unsigned long long get_streamID() const { return m_streamID; } unsigned get_schd_id() const { return m_scheduler_id; } active_mask_t get_warp_active_mask() const { return m_warp_active_mask; } protected: unsigned m_uid; + unsigned long long m_streamID; bool m_empty; bool m_cache_hit; unsigned long long issue_cycle; diff --git a/src/gpgpu-sim/gpu-cache.cc b/src/gpgpu-sim/gpu-cache.cc index 32cc56b..cd3c880 100644 --- a/src/gpgpu-sim/gpu-cache.cc +++ b/src/gpgpu-sim/gpu-cache.cc @@ -634,14 +634,6 @@ void mshr_table::display(FILE *fp) const { /***************************************************************** Caches * *****************************************************************/ cache_stats::cache_stats() { - m_stats.resize(NUM_MEM_ACCESS_TYPE); - m_stats_pw.resize(NUM_MEM_ACCESS_TYPE); - m_fail_stats.resize(NUM_MEM_ACCESS_TYPE); - for (unsigned i = 0; i < NUM_MEM_ACCESS_TYPE; ++i) { - m_stats[i].resize(NUM_CACHE_REQUEST_STATUS, 0); - m_stats_pw[i].resize(NUM_CACHE_REQUEST_STATUS, 0); - m_fail_stats[i].resize(NUM_CACHE_RESERVATION_FAIL_STATUS, 0); - } m_cache_port_available_cycles = 0; m_cache_data_port_busy_cycles = 0; m_cache_fill_port_busy_cycles = 0; @@ -651,11 +643,10 @@ void cache_stats::clear() { /// /// Zero out all current cache statistics /// - for (unsigned i = 0; i < NUM_MEM_ACCESS_TYPE; ++i) { - std::fill(m_stats[i].begin(), m_stats[i].end(), 0); - std::fill(m_stats_pw[i].begin(), m_stats_pw[i].end(), 0); - std::fill(m_fail_stats[i].begin(), m_fail_stats[i].end(), 0); - } + m_stats.clear(); + m_stats_pw.clear(); + m_fail_stats.clear(); + m_cache_port_available_cycles = 0; m_cache_data_port_busy_cycles = 0; m_cache_fill_port_busy_cycles = 0; @@ -665,35 +656,67 @@ void cache_stats::clear_pw() { /// /// Zero out per-window cache statistics /// - for (unsigned i = 0; i < NUM_MEM_ACCESS_TYPE; ++i) { - std::fill(m_stats_pw[i].begin(), m_stats_pw[i].end(), 0); - } + m_stats_pw.clear(); } -void cache_stats::inc_stats(int access_type, int access_outcome) { +void cache_stats::inc_stats(int access_type, int access_outcome, + unsigned long long streamID) { /// /// Increment the stat corresponding to (access_type, access_outcome) by 1. /// if (!check_valid(access_type, access_outcome)) assert(0 && "Unknown cache access type or access outcome"); - m_stats[access_type][access_outcome]++; + if (m_stats.find(streamID) == m_stats.end()) { + std::vector> new_val; + new_val.resize(NUM_MEM_ACCESS_TYPE); + for (unsigned j = 0; j < NUM_MEM_ACCESS_TYPE; ++j) { + new_val[j].resize(NUM_CACHE_REQUEST_STATUS, 0); + } + m_stats.insert(std::pair>>( + streamID, new_val)); + } + m_stats.at(streamID)[access_type][access_outcome]++; } -void cache_stats::inc_stats_pw(int access_type, int access_outcome) { +void cache_stats::inc_stats_pw(int access_type, int access_outcome, + unsigned long long streamID) { /// /// Increment the corresponding per-window cache stat /// if (!check_valid(access_type, access_outcome)) assert(0 && "Unknown cache access type or access outcome"); - m_stats_pw[access_type][access_outcome]++; + + if (m_stats_pw.find(streamID) == m_stats_pw.end()) { + std::vector> new_val; + new_val.resize(NUM_MEM_ACCESS_TYPE); + for (unsigned j = 0; j < NUM_MEM_ACCESS_TYPE; ++j) { + new_val[j].resize(NUM_CACHE_REQUEST_STATUS, 0); + } + m_stats_pw.insert(std::pair>>( + streamID, new_val)); + } + m_stats_pw.at(streamID)[access_type][access_outcome]++; } -void cache_stats::inc_fail_stats(int access_type, int fail_outcome) { +void cache_stats::inc_fail_stats(int access_type, int fail_outcome, + unsigned long long streamID) { if (!check_fail_valid(access_type, fail_outcome)) assert(0 && "Unknown cache access type or access fail"); - m_fail_stats[access_type][fail_outcome]++; + if (m_fail_stats.find(streamID) == m_fail_stats.end()) { + std::vector> new_val; + new_val.resize(NUM_MEM_ACCESS_TYPE); + for (unsigned j = 0; j < NUM_MEM_ACCESS_TYPE; ++j) { + new_val[j].resize(NUM_CACHE_RESERVATION_FAIL_STATUS, 0); + } + m_fail_stats.insert(std::pair>>( + streamID, new_val)); + } + m_fail_stats.at(streamID)[access_type][fail_outcome]++; } enum cache_request_status cache_stats::select_stats_status( @@ -712,7 +735,8 @@ enum cache_request_status cache_stats::select_stats_status( } unsigned long long &cache_stats::operator()(int access_type, int access_outcome, - bool fail_outcome) { + bool fail_outcome, + unsigned long long streamID) { /// /// Simple method to read/modify the stat corresponding to (access_type, /// access_outcome) Used overloaded () to avoid the need for separate @@ -722,17 +746,18 @@ unsigned long long &cache_stats::operator()(int access_type, int access_outcome, if (!check_fail_valid(access_type, access_outcome)) assert(0 && "Unknown cache access type or fail outcome"); - return m_fail_stats[access_type][access_outcome]; + return m_fail_stats.at(streamID)[access_type][access_outcome]; } else { if (!check_valid(access_type, access_outcome)) assert(0 && "Unknown cache access type or access outcome"); - return m_stats[access_type][access_outcome]; + return m_stats.at(streamID)[access_type][access_outcome]; } } unsigned long long cache_stats::operator()(int access_type, int access_outcome, - bool fail_outcome) const { + bool fail_outcome, + unsigned long long streamID) const { /// /// Const accessor into m_stats. /// @@ -740,12 +765,12 @@ unsigned long long cache_stats::operator()(int access_type, int access_outcome, if (!check_fail_valid(access_type, access_outcome)) assert(0 && "Unknown cache access type or fail outcome"); - return m_fail_stats[access_type][access_outcome]; + return m_fail_stats.at(streamID)[access_type][access_outcome]; } else { if (!check_valid(access_type, access_outcome)) assert(0 && "Unknown cache access type or access outcome"); - return m_stats[access_type][access_outcome]; + return m_stats.at(streamID)[access_type][access_outcome]; } } @@ -754,15 +779,74 @@ cache_stats cache_stats::operator+(const cache_stats &cs) { /// Overloaded + operator to allow for simple stat accumulation /// cache_stats ret; - for (unsigned type = 0; type < NUM_MEM_ACCESS_TYPE; ++type) { - for (unsigned status = 0; status < NUM_CACHE_REQUEST_STATUS; ++status) { - ret(type, status, false) = - m_stats[type][status] + cs(type, status, false); + for (auto iter = m_stats.begin(); iter != m_stats.end(); ++iter) { + unsigned long long streamID = iter->first; + ret.m_stats.insert(std::pair>>( + streamID, m_stats.at(streamID))); + } + for (auto iter = m_stats_pw.begin(); iter != m_stats_pw.end(); ++iter) { + unsigned long long streamID = iter->first; + ret.m_stats_pw.insert( + std::pair>>( + streamID, m_stats_pw.at(streamID))); + } + for (auto iter = m_fail_stats.begin(); iter != m_fail_stats.end(); ++iter) { + unsigned long long streamID = iter->first; + ret.m_fail_stats.insert( + std::pair>>( + streamID, m_fail_stats.at(streamID))); + } + for (auto iter = cs.m_stats.begin(); iter != cs.m_stats.end(); ++iter) { + unsigned long long streamID = iter->first; + if (ret.m_stats.find(streamID) == ret.m_stats.end()) { + ret.m_stats.insert( + std::pair>>( + streamID, cs.m_stats.at(streamID))); + } else { + for (unsigned type = 0; type < NUM_MEM_ACCESS_TYPE; ++type) { + for (unsigned status = 0; status < NUM_CACHE_REQUEST_STATUS; ++status) { + ret.m_stats.at(streamID)[type][status] += + cs(type, status, false, streamID); + } + } } - for (unsigned status = 0; status < NUM_CACHE_RESERVATION_FAIL_STATUS; - ++status) { - ret(type, status, true) = - m_fail_stats[type][status] + cs(type, status, true); + } + for (auto iter = cs.m_stats_pw.begin(); iter != cs.m_stats_pw.end(); ++iter) { + unsigned long long streamID = iter->first; + if (ret.m_stats_pw.find(streamID) == ret.m_stats_pw.end()) { + ret.m_stats_pw.insert( + std::pair>>( + streamID, cs.m_stats_pw.at(streamID))); + } else { + for (unsigned type = 0; type < NUM_MEM_ACCESS_TYPE; ++type) { + for (unsigned status = 0; status < NUM_CACHE_REQUEST_STATUS; ++status) { + ret.m_stats_pw.at(streamID)[type][status] += + cs(type, status, false, streamID); + } + } + } + } + for (auto iter = cs.m_fail_stats.begin(); iter != cs.m_fail_stats.end(); + ++iter) { + unsigned long long streamID = iter->first; + if (ret.m_fail_stats.find(streamID) == ret.m_fail_stats.end()) { + ret.m_fail_stats.insert( + std::pair>>( + streamID, cs.m_fail_stats.at(streamID))); + } else { + for (unsigned type = 0; type < NUM_MEM_ACCESS_TYPE; ++type) { + for (unsigned status = 0; status < NUM_CACHE_RESERVATION_FAIL_STATUS; + ++status) { + ret.m_fail_stats.at(streamID)[type][status] += + cs(type, status, true, streamID); + } + } } } ret.m_cache_port_available_cycles = @@ -778,16 +862,52 @@ cache_stats &cache_stats::operator+=(const cache_stats &cs) { /// /// Overloaded += operator to allow for simple stat accumulation /// - for (unsigned type = 0; type < NUM_MEM_ACCESS_TYPE; ++type) { - for (unsigned status = 0; status < NUM_CACHE_REQUEST_STATUS; ++status) { - m_stats[type][status] += cs(type, status, false); + for (auto iter = cs.m_stats.begin(); iter != cs.m_stats.end(); ++iter) { + unsigned long long streamID = iter->first; + if (m_stats.find(streamID) == m_stats.end()) { + m_stats.insert(std::pair>>( + streamID, cs.m_stats.at(streamID))); + } else { + for (unsigned type = 0; type < NUM_MEM_ACCESS_TYPE; ++type) { + for (unsigned status = 0; status < NUM_CACHE_REQUEST_STATUS; ++status) { + m_stats.at(streamID)[type][status] += + cs(type, status, false, streamID); + } + } } - for (unsigned status = 0; status < NUM_CACHE_REQUEST_STATUS; ++status) { - m_stats_pw[type][status] += cs(type, status, false); + } + for (auto iter = cs.m_stats_pw.begin(); iter != cs.m_stats_pw.end(); ++iter) { + unsigned long long streamID = iter->first; + if (m_stats_pw.find(streamID) == m_stats_pw.end()) { + m_stats_pw.insert(std::pair>>( + streamID, cs.m_stats_pw.at(streamID))); + } else { + for (unsigned type = 0; type < NUM_MEM_ACCESS_TYPE; ++type) { + for (unsigned status = 0; status < NUM_CACHE_REQUEST_STATUS; ++status) { + m_stats_pw.at(streamID)[type][status] += + cs(type, status, false, streamID); + } + } } - for (unsigned status = 0; status < NUM_CACHE_RESERVATION_FAIL_STATUS; - ++status) { - m_fail_stats[type][status] += cs(type, status, true); + } + for (auto iter = cs.m_fail_stats.begin(); iter != cs.m_fail_stats.end(); + ++iter) { + unsigned long long streamID = iter->first; + if (m_fail_stats.find(streamID) == m_fail_stats.end()) { + m_fail_stats.insert( + std::pair>>( + streamID, cs.m_fail_stats.at(streamID))); + } else { + for (unsigned type = 0; type < NUM_MEM_ACCESS_TYPE; ++type) { + for (unsigned status = 0; status < NUM_CACHE_RESERVATION_FAIL_STATUS; + ++status) { + m_fail_stats.at(streamID)[type][status] += + cs(type, status, true, streamID); + } + } } } m_cache_port_available_cycles += cs.m_cache_port_available_cycles; @@ -796,46 +916,65 @@ cache_stats &cache_stats::operator+=(const cache_stats &cs) { return *this; } -void cache_stats::print_stats(FILE *fout, const char *cache_name) const { +void cache_stats::print_stats(FILE *fout, unsigned long long streamID, + const char *cache_name) const { /// - /// Print out each non-zero cache statistic for every memory access type and - /// status "cache_name" defaults to "Cache_stats" when no argument is - /// provided, otherwise the provided name is used. The printed format is + /// For a given CUDA stream, print out each non-zero cache statistic for every + /// memory access type and status "cache_name" defaults to "Cache_stats" when + /// no argument is provided, otherwise the provided name is used. The printed + /// format is /// "[][] = " - /// + /// Specify streamID to be -1 to print every stream. + std::vector total_access; - total_access.resize(NUM_MEM_ACCESS_TYPE, 0); std::string m_cache_name = cache_name; - for (unsigned type = 0; type < NUM_MEM_ACCESS_TYPE; ++type) { - for (unsigned status = 0; status < NUM_CACHE_REQUEST_STATUS; ++status) { - fprintf(fout, "\t%s[%s][%s] = %llu\n", m_cache_name.c_str(), - mem_access_type_str((enum mem_access_type)type), - cache_request_status_str((enum cache_request_status)status), - m_stats[type][status]); + for (auto iter = m_stats.begin(); iter != m_stats.end(); ++iter) { + unsigned long long streamid = iter->first; + // when streamID is specified, skip stats for all other streams, otherwise, + // print stats from all streams + if ((streamID != -1) && (streamid != streamID)) continue; + total_access.clear(); + total_access.resize(NUM_MEM_ACCESS_TYPE, 0); + for (unsigned type = 0; type < NUM_MEM_ACCESS_TYPE; ++type) { + for (unsigned status = 0; status < NUM_CACHE_REQUEST_STATUS; ++status) { + fprintf(fout, "\t%s[%s][%s] = %llu\n", m_cache_name.c_str(), + mem_access_type_str((enum mem_access_type)type), + cache_request_status_str((enum cache_request_status)status), + m_stats.at(streamid)[type][status]); - if (status != RESERVATION_FAIL && status != MSHR_HIT) - // MSHR_HIT is a special type of SECTOR_MISS - // so its already included in the SECTOR_MISS - total_access[type] += m_stats[type][status]; + if (status != RESERVATION_FAIL && status != MSHR_HIT) + // MSHR_HIT is a special type of SECTOR_MISS + // so its already included in the SECTOR_MISS + total_access[type] += m_stats.at(streamid)[type][status]; + } + } + for (unsigned type = 0; type < NUM_MEM_ACCESS_TYPE; ++type) { + if (total_access[type] > 0) + fprintf(fout, "\t%s[%s][%s] = %u\n", m_cache_name.c_str(), + mem_access_type_str((enum mem_access_type)type), "TOTAL_ACCESS", + total_access[type]); } - } - for (unsigned type = 0; type < NUM_MEM_ACCESS_TYPE; ++type) { - if (total_access[type] > 0) - fprintf(fout, "\t%s[%s][%s] = %u\n", m_cache_name.c_str(), - mem_access_type_str((enum mem_access_type)type), "TOTAL_ACCESS", - total_access[type]); } } -void cache_stats::print_fail_stats(FILE *fout, const char *cache_name) const { +void cache_stats::print_fail_stats(FILE *fout, unsigned long long streamID, + const char *cache_name) const { std::string m_cache_name = cache_name; - for (unsigned type = 0; type < NUM_MEM_ACCESS_TYPE; ++type) { - for (unsigned fail = 0; fail < NUM_CACHE_RESERVATION_FAIL_STATUS; ++fail) { - if (m_fail_stats[type][fail] > 0) { - fprintf(fout, "\t%s[%s][%s] = %llu\n", m_cache_name.c_str(), - mem_access_type_str((enum mem_access_type)type), - cache_fail_status_str((enum cache_reservation_fail_reason)fail), - m_fail_stats[type][fail]); + for (auto iter = m_fail_stats.begin(); iter != m_fail_stats.end(); ++iter) { + unsigned long long streamid = iter->first; + // when streamID is specified, skip stats for all other streams, otherwise, + // print stats from all streams + if ((streamID != -1) && (streamid != streamID)) continue; + for (unsigned type = 0; type < NUM_MEM_ACCESS_TYPE; ++type) { + for (unsigned fail = 0; fail < NUM_CACHE_RESERVATION_FAIL_STATUS; + ++fail) { + if (m_fail_stats.at(streamid)[type][fail] > 0) { + fprintf( + fout, "\t%s[%s][%s] = %llu\n", m_cache_name.c_str(), + mem_access_type_str((enum mem_access_type)type), + cache_fail_status_str((enum cache_reservation_fail_reason)fail), + m_fail_stats.at(streamid)[type][fail]); + } } } } @@ -866,11 +1005,14 @@ unsigned long long cache_stats::get_stats( /// cache_request_statuses. /// unsigned long long total = 0; - for (unsigned type = 0; type < num_access_type; ++type) { - for (unsigned status = 0; status < num_access_status; ++status) { - if (!check_valid((int)access_type[type], (int)access_status[status])) - assert(0 && "Unknown cache access type or access outcome"); - total += m_stats[access_type[type]][access_status[status]]; + for (auto iter = m_stats.begin(); iter != m_stats.end(); ++iter) { + unsigned long long streamID = iter->first; + for (unsigned type = 0; type < num_access_type; ++type) { + for (unsigned status = 0; status < num_access_status; ++status) { + if (!check_valid((int)access_type[type], (int)access_status[status])) + assert(0 && "Unknown cache access type or access outcome"); + total += m_stats.at(streamID)[access_type[type]][access_status[status]]; + } } } return total; @@ -883,18 +1025,23 @@ void cache_stats::get_sub_stats(struct cache_sub_stats &css) const { struct cache_sub_stats t_css; t_css.clear(); - for (unsigned type = 0; type < NUM_MEM_ACCESS_TYPE; ++type) { - for (unsigned status = 0; status < NUM_CACHE_REQUEST_STATUS; ++status) { - if (status == HIT || status == MISS || status == SECTOR_MISS || - status == HIT_RESERVED) - t_css.accesses += m_stats[type][status]; + for (auto iter = m_stats.begin(); iter != m_stats.end(); ++iter) { + unsigned long long streamID = iter->first; + for (unsigned type = 0; type < NUM_MEM_ACCESS_TYPE; ++type) { + for (unsigned status = 0; status < NUM_CACHE_REQUEST_STATUS; ++status) { + if (status == HIT || status == MISS || status == SECTOR_MISS || + status == HIT_RESERVED) + t_css.accesses += m_stats.at(streamID)[type][status]; - if (status == MISS || status == SECTOR_MISS) - t_css.misses += m_stats[type][status]; + if (status == MISS || status == SECTOR_MISS) + t_css.misses += m_stats.at(streamID)[type][status]; - if (status == HIT_RESERVED) t_css.pending_hits += m_stats[type][status]; + if (status == HIT_RESERVED) + t_css.pending_hits += m_stats.at(streamID)[type][status]; - if (status == RESERVATION_FAIL) t_css.res_fails += m_stats[type][status]; + if (status == RESERVATION_FAIL) + t_css.res_fails += m_stats.at(streamID)[type][status]; + } } } @@ -912,41 +1059,48 @@ void cache_stats::get_sub_stats_pw(struct cache_sub_stats_pw &css) const { struct cache_sub_stats_pw t_css; t_css.clear(); - for (unsigned type = 0; type < NUM_MEM_ACCESS_TYPE; ++type) { - for (unsigned status = 0; status < NUM_CACHE_REQUEST_STATUS; ++status) { - if (status == HIT || status == MISS || status == SECTOR_MISS || - status == HIT_RESERVED) - t_css.accesses += m_stats_pw[type][status]; - - if (status == HIT) { - if (type == GLOBAL_ACC_R || type == CONST_ACC_R || type == INST_ACC_R) { - t_css.read_hits += m_stats_pw[type][status]; - } else if (type == GLOBAL_ACC_W) { - t_css.write_hits += m_stats_pw[type][status]; + for (auto iter = m_stats_pw.begin(); iter != m_stats_pw.end(); ++iter) { + unsigned long long streamID = iter->first; + for (unsigned type = 0; type < NUM_MEM_ACCESS_TYPE; ++type) { + for (unsigned status = 0; status < NUM_CACHE_REQUEST_STATUS; ++status) { + if (status == HIT || status == MISS || status == SECTOR_MISS || + status == HIT_RESERVED) + t_css.accesses += m_stats_pw.at(streamID)[type][status]; + + if (status == HIT) { + if (type == GLOBAL_ACC_R || type == CONST_ACC_R || + type == INST_ACC_R) { + t_css.read_hits += m_stats_pw.at(streamID)[type][status]; + } else if (type == GLOBAL_ACC_W) { + t_css.write_hits += m_stats_pw.at(streamID)[type][status]; + } } - } - if (status == MISS || status == SECTOR_MISS) { - if (type == GLOBAL_ACC_R || type == CONST_ACC_R || type == INST_ACC_R) { - t_css.read_misses += m_stats_pw[type][status]; - } else if (type == GLOBAL_ACC_W) { - t_css.write_misses += m_stats_pw[type][status]; + if (status == MISS || status == SECTOR_MISS) { + if (type == GLOBAL_ACC_R || type == CONST_ACC_R || + type == INST_ACC_R) { + t_css.read_misses += m_stats_pw.at(streamID)[type][status]; + } else if (type == GLOBAL_ACC_W) { + t_css.write_misses += m_stats_pw.at(streamID)[type][status]; + } } - } - if (status == HIT_RESERVED) { - if (type == GLOBAL_ACC_R || type == CONST_ACC_R || type == INST_ACC_R) { - t_css.read_pending_hits += m_stats_pw[type][status]; - } else if (type == GLOBAL_ACC_W) { - t_css.write_pending_hits += m_stats_pw[type][status]; + if (status == HIT_RESERVED) { + if (type == GLOBAL_ACC_R || type == CONST_ACC_R || + type == INST_ACC_R) { + t_css.read_pending_hits += m_stats_pw.at(streamID)[type][status]; + } else if (type == GLOBAL_ACC_W) { + t_css.write_pending_hits += m_stats_pw.at(streamID)[type][status]; + } } - } - if (status == RESERVATION_FAIL) { - if (type == GLOBAL_ACC_R || type == CONST_ACC_R || type == INST_ACC_R) { - t_css.read_res_fails += m_stats_pw[type][status]; - } else if (type == GLOBAL_ACC_W) { - t_css.write_res_fails += m_stats_pw[type][status]; + if (status == RESERVATION_FAIL) { + if (type == GLOBAL_ACC_R || type == CONST_ACC_R || + type == INST_ACC_R) { + t_css.read_res_fails += m_stats_pw.at(streamID)[type][status]; + } else if (type == GLOBAL_ACC_W) { + t_css.write_res_fails += m_stats_pw.at(streamID)[type][status]; + } } } } @@ -1139,6 +1293,50 @@ void baseline_cache::display_state(FILE *fp) const { fprintf(fp, "\n"); } +void baseline_cache::inc_aggregated_stats(cache_request_status status, + cache_request_status cache_status, + mem_fetch *mf, + enum cache_gpu_level level) { + if (level == L1_GPU_CACHE) { + m_gpu->aggregated_l1_stats.inc_stats( + mf->get_streamID(), mf->get_access_type(), + m_gpu->aggregated_l1_stats.select_stats_status(status, cache_status)); + } else if (level == L2_GPU_CACHE) { + m_gpu->aggregated_l2_stats.inc_stats( + mf->get_streamID(), mf->get_access_type(), + m_gpu->aggregated_l2_stats.select_stats_status(status, cache_status)); + } +} + +void baseline_cache::inc_aggregated_fail_stats( + cache_request_status status, cache_request_status cache_status, + mem_fetch *mf, enum cache_gpu_level level) { + if (level == L1_GPU_CACHE) { + m_gpu->aggregated_l1_stats.inc_fail_stats( + mf->get_streamID(), mf->get_access_type(), + m_gpu->aggregated_l1_stats.select_stats_status(status, cache_status)); + } else if (level == L2_GPU_CACHE) { + m_gpu->aggregated_l2_stats.inc_fail_stats( + mf->get_streamID(), mf->get_access_type(), + m_gpu->aggregated_l2_stats.select_stats_status(status, cache_status)); + } +} + +void baseline_cache::inc_aggregated_stats_pw(cache_request_status status, + cache_request_status cache_status, + mem_fetch *mf, + enum cache_gpu_level level) { + if (level == L1_GPU_CACHE) { + m_gpu->aggregated_l1_stats.inc_stats_pw( + mf->get_streamID(), mf->get_access_type(), + m_gpu->aggregated_l1_stats.select_stats_status(status, cache_status)); + } else if (level == L2_GPU_CACHE) { + m_gpu->aggregated_l2_stats.inc_stats_pw( + mf->get_streamID(), mf->get_access_type(), + m_gpu->aggregated_l2_stats.select_stats_status(status, cache_status)); + } +} + /// Read miss handler without writeback void baseline_cache::send_read_request(new_addr_type addr, new_addr_type block_addr, @@ -1170,7 +1368,7 @@ void baseline_cache::send_read_request(new_addr_type addr, m_tag_array->access(block_addr, time, cache_index, wb, evicted, mf); m_mshrs.add(mshr_addr, mf); - m_stats.inc_stats(mf->get_access_type(), MSHR_HIT); + m_stats.inc_stats(mf->get_access_type(), MSHR_HIT, mf->get_streamID()); do_miss = true; } else if (!mshr_hit && mshr_avail && @@ -1191,9 +1389,11 @@ void baseline_cache::send_read_request(new_addr_type addr, do_miss = true; } else if (mshr_hit && !mshr_avail) - m_stats.inc_fail_stats(mf->get_access_type(), MSHR_MERGE_ENRTY_FAIL); + m_stats.inc_fail_stats(mf->get_access_type(), MSHR_MERGE_ENRTY_FAIL, + mf->get_streamID()); else if (!mshr_hit && !mshr_avail) - m_stats.inc_fail_stats(mf->get_access_type(), MSHR_ENRTY_FAIL); + m_stats.inc_fail_stats(mf->get_access_type(), MSHR_ENRTY_FAIL, + mf->get_streamID()); else assert(0); } @@ -1253,7 +1453,8 @@ cache_request_status data_cache::wr_hit_wt(new_addr_type addr, std::list &events, enum cache_request_status status) { if (miss_queue_full(0)) { - m_stats.inc_fail_stats(mf->get_access_type(), MISS_QUEUE_FULL); + m_stats.inc_fail_stats(mf->get_access_type(), MISS_QUEUE_FULL, + mf->get_streamID()); return RESERVATION_FAIL; // cannot handle request this cycle } @@ -1281,7 +1482,8 @@ cache_request_status data_cache::wr_hit_we(new_addr_type addr, std::list &events, enum cache_request_status status) { if (miss_queue_full(0)) { - m_stats.inc_fail_stats(mf->get_access_type(), MISS_QUEUE_FULL); + m_stats.inc_fail_stats(mf->get_access_type(), MISS_QUEUE_FULL, + mf->get_streamID()); return RESERVATION_FAIL; // cannot handle request this cycle } @@ -1330,11 +1532,14 @@ enum cache_request_status data_cache::wr_miss_wa_naive( (m_miss_queue.size() < m_config.m_miss_queue_size)))) { // check what is the exactly the failure reason if (miss_queue_full(2)) - m_stats.inc_fail_stats(mf->get_access_type(), MISS_QUEUE_FULL); + m_stats.inc_fail_stats(mf->get_access_type(), MISS_QUEUE_FULL, + mf->get_streamID()); else if (mshr_hit && !mshr_avail) - m_stats.inc_fail_stats(mf->get_access_type(), MSHR_MERGE_ENRTY_FAIL); + m_stats.inc_fail_stats(mf->get_access_type(), MSHR_MERGE_ENRTY_FAIL, + mf->get_streamID()); else if (!mshr_hit && !mshr_avail) - m_stats.inc_fail_stats(mf->get_access_type(), MSHR_ENRTY_FAIL); + m_stats.inc_fail_stats(mf->get_access_type(), MSHR_ENRTY_FAIL, + mf->get_streamID()); else assert(0); @@ -1353,10 +1558,10 @@ enum cache_request_status data_cache::wr_miss_wa_naive( mf->get_access_warp_mask(), mf->get_access_byte_mask(), mf->get_access_sector_mask(), m_gpu->gpgpu_ctx); - mem_fetch *n_mf = - new mem_fetch(*ma, NULL, mf->get_ctrl_size(), mf->get_wid(), - mf->get_sid(), mf->get_tpc(), mf->get_mem_config(), - m_gpu->gpu_tot_sim_cycle + m_gpu->gpu_sim_cycle); + mem_fetch *n_mf = new mem_fetch( + *ma, NULL, mf->get_streamID(), mf->get_ctrl_size(), mf->get_wid(), + mf->get_sid(), mf->get_tpc(), mf->get_mem_config(), + m_gpu->gpu_tot_sim_cycle + m_gpu->gpu_sim_cycle); bool do_miss = false; bool wb = false; @@ -1378,7 +1583,7 @@ enum cache_request_status data_cache::wr_miss_wa_naive( evicted.m_block_addr, m_wrbk_type, mf->get_access_warp_mask(), evicted.m_byte_mask, evicted.m_sector_mask, evicted.m_modified_size, true, m_gpu->gpu_tot_sim_cycle + m_gpu->gpu_sim_cycle, -1, -1, -1, - NULL); + NULL, mf->get_streamID()); // the evicted block may have wrong chip id when advanced L2 hashing is // used, so set the right chip address from the original mf wb->set_chip(mf->get_tlx_addr().chip); @@ -1404,7 +1609,8 @@ enum cache_request_status data_cache::wr_miss_wa_fetch_on_write( // reserve mshr if (miss_queue_full(0)) { - m_stats.inc_fail_stats(mf->get_access_type(), MISS_QUEUE_FULL); + m_stats.inc_fail_stats(mf->get_access_type(), MISS_QUEUE_FULL, + mf->get_streamID()); return RESERVATION_FAIL; // cannot handle request this cycle } @@ -1431,7 +1637,7 @@ enum cache_request_status data_cache::wr_miss_wa_fetch_on_write( evicted.m_block_addr, m_wrbk_type, mf->get_access_warp_mask(), evicted.m_byte_mask, evicted.m_sector_mask, evicted.m_modified_size, true, m_gpu->gpu_tot_sim_cycle + m_gpu->gpu_sim_cycle, -1, -1, -1, - NULL); + NULL, mf->get_streamID()); // the evicted block may have wrong chip id when advanced L2 hashing is // used, so set the right chip address from the original mf wb->set_chip(mf->get_tlx_addr().chip); @@ -1451,11 +1657,14 @@ enum cache_request_status data_cache::wr_miss_wa_fetch_on_write( (m_miss_queue.size() < m_config.m_miss_queue_size)))) { // check what is the exactly the failure reason if (miss_queue_full(1)) - m_stats.inc_fail_stats(mf->get_access_type(), MISS_QUEUE_FULL); + m_stats.inc_fail_stats(mf->get_access_type(), MISS_QUEUE_FULL, + mf->get_streamID()); else if (mshr_hit && !mshr_avail) - m_stats.inc_fail_stats(mf->get_access_type(), MSHR_MERGE_ENRTY_FAIL); + m_stats.inc_fail_stats(mf->get_access_type(), MSHR_MERGE_ENRTY_FAIL, + mf->get_streamID()); else if (!mshr_hit && !mshr_avail) - m_stats.inc_fail_stats(mf->get_access_type(), MSHR_ENRTY_FAIL); + m_stats.inc_fail_stats(mf->get_access_type(), MSHR_ENRTY_FAIL, + mf->get_streamID()); else assert(0); @@ -1468,7 +1677,8 @@ enum cache_request_status data_cache::wr_miss_wa_fetch_on_write( if (m_mshrs.probe(mshr_addr) && m_mshrs.is_read_after_write_pending(mshr_addr) && mf->is_write()) { // assert(0); - m_stats.inc_fail_stats(mf->get_access_type(), MSHR_RW_PENDING); + m_stats.inc_fail_stats(mf->get_access_type(), MSHR_RW_PENDING, + mf->get_streamID()); return RESERVATION_FAIL; } @@ -1479,8 +1689,8 @@ enum cache_request_status data_cache::wr_miss_wa_fetch_on_write( mf->get_access_sector_mask(), m_gpu->gpgpu_ctx); mem_fetch *n_mf = new mem_fetch( - *ma, NULL, mf->get_ctrl_size(), mf->get_wid(), mf->get_sid(), - mf->get_tpc(), mf->get_mem_config(), + *ma, NULL, mf->get_streamID(), mf->get_ctrl_size(), mf->get_wid(), + mf->get_sid(), mf->get_tpc(), mf->get_mem_config(), m_gpu->gpu_tot_sim_cycle + m_gpu->gpu_sim_cycle, NULL, mf); new_addr_type block_addr = m_config.block_addr(addr); @@ -1504,7 +1714,7 @@ enum cache_request_status data_cache::wr_miss_wa_fetch_on_write( evicted.m_block_addr, m_wrbk_type, mf->get_access_warp_mask(), evicted.m_byte_mask, evicted.m_sector_mask, evicted.m_modified_size, true, m_gpu->gpu_tot_sim_cycle + m_gpu->gpu_sim_cycle, -1, -1, -1, - NULL); + NULL, mf->get_streamID()); // the evicted block may have wrong chip id when advanced L2 hashing is // used, so set the right chip address from the original mf wb->set_chip(mf->get_tlx_addr().chip); @@ -1528,7 +1738,8 @@ enum cache_request_status data_cache::wr_miss_wa_lazy_fetch_on_read( // mshr if (miss_queue_full(0)) { - m_stats.inc_fail_stats(mf->get_access_type(), MISS_QUEUE_FULL); + m_stats.inc_fail_stats(mf->get_access_type(), MISS_QUEUE_FULL, + mf->get_streamID()); return RESERVATION_FAIL; // cannot handle request this cycle } @@ -1571,7 +1782,7 @@ enum cache_request_status data_cache::wr_miss_wa_lazy_fetch_on_read( evicted.m_block_addr, m_wrbk_type, mf->get_access_warp_mask(), evicted.m_byte_mask, evicted.m_sector_mask, evicted.m_modified_size, true, m_gpu->gpu_tot_sim_cycle + m_gpu->gpu_sim_cycle, -1, -1, -1, - NULL); + NULL, mf->get_streamID()); // the evicted block may have wrong chip id when advanced L2 hashing is // used, so set the right chip address from the original mf wb->set_chip(mf->get_tlx_addr().chip); @@ -1589,7 +1800,8 @@ enum cache_request_status data_cache::wr_miss_no_wa( new_addr_type addr, unsigned cache_index, mem_fetch *mf, unsigned time, std::list &events, enum cache_request_status status) { if (miss_queue_full(0)) { - m_stats.inc_fail_stats(mf->get_access_type(), MISS_QUEUE_FULL); + m_stats.inc_fail_stats(mf->get_access_type(), MISS_QUEUE_FULL, + mf->get_streamID()); return RESERVATION_FAIL; // cannot handle request this cycle } @@ -1634,7 +1846,8 @@ enum cache_request_status data_cache::rd_miss_base( if (miss_queue_full(1)) { // cannot handle request this cycle // (might need to generate two requests) - m_stats.inc_fail_stats(mf->get_access_type(), MISS_QUEUE_FULL); + m_stats.inc_fail_stats(mf->get_access_type(), MISS_QUEUE_FULL, + mf->get_streamID()); return RESERVATION_FAIL; } @@ -1653,7 +1866,7 @@ enum cache_request_status data_cache::rd_miss_base( evicted.m_block_addr, m_wrbk_type, mf->get_access_warp_mask(), evicted.m_byte_mask, evicted.m_sector_mask, evicted.m_modified_size, true, m_gpu->gpu_tot_sim_cycle + m_gpu->gpu_sim_cycle, -1, -1, -1, - NULL); + NULL, mf->get_streamID()); // the evicted block may have wrong chip id when advanced L2 hashing is // used, so set the right chip address from the original mf wb->set_chip(mf->get_tlx_addr().chip); @@ -1693,16 +1906,20 @@ enum cache_request_status read_only_cache::access( cache_status = RESERVATION_FAIL; } else { cache_status = RESERVATION_FAIL; - m_stats.inc_fail_stats(mf->get_access_type(), MISS_QUEUE_FULL); + m_stats.inc_fail_stats(mf->get_access_type(), MISS_QUEUE_FULL, + mf->get_streamID()); } } else { - m_stats.inc_fail_stats(mf->get_access_type(), LINE_ALLOC_FAIL); + m_stats.inc_fail_stats(mf->get_access_type(), LINE_ALLOC_FAIL, + mf->get_streamID()); } m_stats.inc_stats(mf->get_access_type(), - m_stats.select_stats_status(status, cache_status)); + m_stats.select_stats_status(status, cache_status), + mf->get_streamID()); m_stats.inc_stats_pw(mf->get_access_type(), - m_stats.select_stats_status(status, cache_status)); + m_stats.select_stats_status(status, cache_status), + mf->get_streamID()); return cache_status; } @@ -1730,7 +1947,8 @@ enum cache_request_status data_cache::process_tag_probe( } else { // the only reason for reservation fail here is LINE_ALLOC_FAIL (i.e all // lines are reserved) - m_stats.inc_fail_stats(mf->get_access_type(), LINE_ALLOC_FAIL); + m_stats.inc_fail_stats(mf->get_access_type(), LINE_ALLOC_FAIL, + mf->get_streamID()); } } else { // Read if (probe_status == HIT) { @@ -1742,7 +1960,8 @@ enum cache_request_status data_cache::process_tag_probe( } else { // the only reason for reservation fail here is LINE_ALLOC_FAIL (i.e all // lines are reserved) - m_stats.inc_fail_stats(mf->get_access_type(), LINE_ALLOC_FAIL); + m_stats.inc_fail_stats(mf->get_access_type(), LINE_ALLOC_FAIL, + mf->get_streamID()); } } @@ -1767,9 +1986,11 @@ enum cache_request_status data_cache::access(new_addr_type addr, mem_fetch *mf, enum cache_request_status access_status = process_tag_probe(wr, probe_status, addr, cache_index, mf, time, events); m_stats.inc_stats(mf->get_access_type(), - m_stats.select_stats_status(probe_status, access_status)); - m_stats.inc_stats_pw(mf->get_access_type(), m_stats.select_stats_status( - probe_status, access_status)); + m_stats.select_stats_status(probe_status, access_status), + mf->get_streamID()); + m_stats.inc_stats_pw(mf->get_access_type(), + m_stats.select_stats_status(probe_status, access_status), + mf->get_streamID()); return access_status; } @@ -1831,9 +2052,11 @@ enum cache_request_status tex_cache::access(new_addr_type addr, mem_fetch *mf, cache_status = HIT_RESERVED; } m_stats.inc_stats(mf->get_access_type(), - m_stats.select_stats_status(status, cache_status)); + m_stats.select_stats_status(status, cache_status), + mf->get_streamID()); m_stats.inc_stats_pw(mf->get_access_type(), - m_stats.select_stats_status(status, cache_status)); + m_stats.select_stats_status(status, cache_status), + mf->get_streamID()); return cache_status; } diff --git a/src/gpgpu-sim/gpu-cache.h b/src/gpgpu-sim/gpu-cache.h index 5fd40a9..c07695f 100644 --- a/src/gpgpu-sim/gpu-cache.h +++ b/src/gpgpu-sim/gpu-cache.h @@ -72,6 +72,13 @@ enum cache_event_type { WRITE_ALLOCATE_SENT }; +enum cache_gpu_level { + L1_GPU_CACHE = 0, + L2_GPU_CACHE, + OTHER_GPU_CACHE, + NUM_CACHE_GPU_LEVELS +}; + struct evicted_block_info { new_addr_type m_block_addr; unsigned m_modified_size; @@ -1200,20 +1207,26 @@ class cache_stats { void clear(); // Clear AerialVision cache stats after each window void clear_pw(); - void inc_stats(int access_type, int access_outcome); + void inc_stats(int access_type, int access_outcome, + unsigned long long streamID); // Increment AerialVision cache stats - void inc_stats_pw(int access_type, int access_outcome); - void inc_fail_stats(int access_type, int fail_outcome); + void inc_stats_pw(int access_type, int access_outcome, + unsigned long long streamID); + void inc_fail_stats(int access_type, int fail_outcome, + unsigned long long streamID); enum cache_request_status select_stats_status( enum cache_request_status probe, enum cache_request_status access) const; unsigned long long &operator()(int access_type, int access_outcome, - bool fail_outcome); + bool fail_outcome, + unsigned long long streamID); unsigned long long operator()(int access_type, int access_outcome, - bool fail_outcome) const; + bool fail_outcome, + unsigned long long streamID) const; cache_stats operator+(const cache_stats &cs); cache_stats &operator+=(const cache_stats &cs); - void print_stats(FILE *fout, const char *cache_name = "Cache_stats") const; - void print_fail_stats(FILE *fout, + void print_stats(FILE *fout, unsigned long long streamID, + const char *cache_name = "Cache_stats") const; + void print_fail_stats(FILE *fout, unsigned long long streamID, const char *cache_name = "Cache_fail_stats") const; unsigned long long get_stats(enum mem_access_type *access_type, @@ -1231,10 +1244,14 @@ class cache_stats { bool check_valid(int type, int status) const; bool check_fail_valid(int type, int fail) const; - std::vector > m_stats; + // CUDA streamID -> cache stats[NUM_MEM_ACCESS_TYPE] + std::map>> + m_stats; // AerialVision cache stats (per-window) - std::vector > m_stats_pw; - std::vector > m_fail_stats; + std::map>> + m_stats_pw; + std::map>> + m_fail_stats; unsigned long long m_cache_port_available_cycles; unsigned long long m_cache_data_port_busy_cycles; @@ -1264,11 +1281,14 @@ class baseline_cache : public cache_t { public: baseline_cache(const char *name, cache_config &config, int core_id, int type_id, mem_fetch_interface *memport, - enum mem_fetch_status status) + enum mem_fetch_status status, enum cache_gpu_level level, + gpgpu_sim *gpu) : m_config(config), m_tag_array(new tag_array(config, core_id, type_id)), m_mshrs(config.m_mshr_entries, config.m_mshr_max_merge), - m_bandwidth_management(config) { + m_bandwidth_management(config), + m_level(level), + m_gpu(gpu) { init(name, config, memport, status); } @@ -1336,6 +1356,15 @@ class baseline_cache : public cache_t { bool fill_port_free() const { return m_bandwidth_management.fill_port_free(); } + void inc_aggregated_stats(cache_request_status status, + cache_request_status cache_status, mem_fetch *mf, + enum cache_gpu_level level); + void inc_aggregated_fail_stats(cache_request_status status, + cache_request_status cache_status, + mem_fetch *mf, enum cache_gpu_level level); + void inc_aggregated_stats_pw(cache_request_status status, + cache_request_status cache_status, mem_fetch *mf, + enum cache_gpu_level level); // This is a gapping hole we are poking in the system to quickly handle // filling the cache on cudamemcopies. We don't care about anything other than @@ -1367,6 +1396,8 @@ class baseline_cache : public cache_t { std::list m_miss_queue; enum mem_fetch_status m_miss_queue_status; mem_fetch_interface *m_memport; + cache_gpu_level m_level; + gpgpu_sim *m_gpu; struct extra_mf_fields { extra_mf_fields() { m_valid = false; } @@ -1453,8 +1484,10 @@ class read_only_cache : public baseline_cache { public: read_only_cache(const char *name, cache_config &config, int core_id, int type_id, mem_fetch_interface *memport, - enum mem_fetch_status status) - : baseline_cache(name, config, core_id, type_id, memport, status) {} + enum mem_fetch_status status, enum cache_gpu_level level, + gpgpu_sim *gpu) + : baseline_cache(name, config, core_id, type_id, memport, status, level, + gpu) {} /// Access cache for read_only_cache: returns RESERVATION_FAIL if request /// could not be accepted (for any reason) @@ -1478,8 +1511,10 @@ class data_cache : public baseline_cache { data_cache(const char *name, cache_config &config, int core_id, int type_id, mem_fetch_interface *memport, mem_fetch_allocator *mfcreator, enum mem_fetch_status status, mem_access_type wr_alloc_type, - mem_access_type wrbk_type, class gpgpu_sim *gpu) - : baseline_cache(name, config, core_id, type_id, memport, status) { + mem_access_type wrbk_type, class gpgpu_sim *gpu, + enum cache_gpu_level level) + : baseline_cache(name, config, core_id, type_id, memport, status, level, + gpu) { init(mfcreator); m_wr_alloc_type = wr_alloc_type; m_wrbk_type = wrbk_type; @@ -1668,9 +1703,10 @@ class l1_cache : public data_cache { public: l1_cache(const char *name, cache_config &config, int core_id, int type_id, mem_fetch_interface *memport, mem_fetch_allocator *mfcreator, - enum mem_fetch_status status, class gpgpu_sim *gpu) + enum mem_fetch_status status, class gpgpu_sim *gpu, + enum cache_gpu_level level) : data_cache(name, config, core_id, type_id, memport, mfcreator, status, - L1_WR_ALLOC_R, L1_WRBK_ACC, gpu) {} + L1_WR_ALLOC_R, L1_WRBK_ACC, gpu, level) {} virtual ~l1_cache() {} @@ -1693,9 +1729,10 @@ class l2_cache : public data_cache { public: l2_cache(const char *name, cache_config &config, int core_id, int type_id, mem_fetch_interface *memport, mem_fetch_allocator *mfcreator, - enum mem_fetch_status status, class gpgpu_sim *gpu) + enum mem_fetch_status status, class gpgpu_sim *gpu, + enum cache_gpu_level level) : data_cache(name, config, core_id, type_id, memport, mfcreator, status, - L2_WR_ALLOC_R, L2_WRBK_ACC, gpu) {} + L2_WR_ALLOC_R, L2_WRBK_ACC, gpu, level) {} virtual ~l2_cache() {} diff --git a/src/gpgpu-sim/gpu-sim.cc b/src/gpgpu-sim/gpu-sim.cc index 1cb8a25..0c922bd 100644 --- a/src/gpgpu-sim/gpu-sim.cc +++ b/src/gpgpu-sim/gpu-sim.cc @@ -788,6 +788,22 @@ void increment_x_then_y_then_z(dim3 &i, const dim3 &bound) { } void gpgpu_sim::launch(kernel_info_t *kinfo) { + unsigned kernelID = kinfo->get_uid(); + unsigned long long streamID = kinfo->get_streamID(); + + kernel_time_t kernel_time = {gpu_tot_sim_cycle + gpu_sim_cycle, 0}; + if (gpu_kernel_time.find(streamID) == gpu_kernel_time.end()) { + std::map new_val; + new_val.insert(std::pair(kernelID, kernel_time)); + gpu_kernel_time.insert( + std::pair>( + streamID, new_val)); + } else { + gpu_kernel_time.at(streamID).insert( + std::pair(kernelID, kernel_time)); + ////////// assume same kernel ID do not appear more than once + } + unsigned cta_size = kinfo->threads_per_cta(); if (cta_size > m_shader_config->n_thread_per_shader) { printf( @@ -893,7 +909,10 @@ kernel_info_t *gpgpu_sim::select_kernel() { } unsigned gpgpu_sim::finished_kernel() { - if (m_finished_kernel.empty()) return 0; + if (m_finished_kernel.empty()) { + last_streamID = -1; + return 0; + } unsigned result = m_finished_kernel.front(); m_finished_kernel.pop_front(); return result; @@ -901,6 +920,11 @@ unsigned gpgpu_sim::finished_kernel() { void gpgpu_sim::set_kernel_done(kernel_info_t *kernel) { unsigned uid = kernel->get_uid(); + last_uid = uid; + unsigned long long streamID = kernel->get_streamID(); + last_streamID = streamID; + gpu_kernel_time.at(streamID).at(uid).end_cycle = + gpu_tot_sim_cycle + gpu_sim_cycle; m_finished_kernel.push_back(uid); std::vector::iterator k; for (k = m_running_kernels.begin(); k != m_running_kernels.end(); k++) { @@ -971,6 +995,9 @@ gpgpu_sim::gpgpu_sim(const gpgpu_sim_config &config, gpgpu_context *ctx) gpu_tot_sim_cycle_parition_util = 0; partiton_replys_in_parallel = 0; partiton_replys_in_parallel_total = 0; + last_streamID = -1; + + gpu_kernel_time.clear(); m_memory_partition_unit = new memory_partition_unit *[m_memory_config->m_n_mem]; @@ -1178,9 +1205,9 @@ PowerscalingCoefficients *gpgpu_sim::get_scaling_coeffs() { return m_gpgpusim_wrapper->get_scaling_coeffs(); } -void gpgpu_sim::print_stats() { +void gpgpu_sim::print_stats(unsigned long long streamID) { gpgpu_ctx->stats->ptx_file_line_stats_write_file(); - gpu_print_stat(); + gpu_print_stat(streamID); if (g_network_mode) { printf( @@ -1363,12 +1390,15 @@ void gpgpu_sim::clear_executed_kernel_info() { m_executed_kernel_names.clear(); m_executed_kernel_uids.clear(); } -void gpgpu_sim::gpu_print_stat() { + +void gpgpu_sim::gpu_print_stat(unsigned long long streamID) { FILE *statfout = stdout; std::string kernel_info_str = executed_kernel_info_string(); fprintf(statfout, "%s", kernel_info_str.c_str()); + printf("kernel_stream_id = %llu\n", streamID); + printf("gpu_sim_cycle = %lld\n", gpu_sim_cycle); printf("gpu_sim_insn = %lld\n", gpu_sim_insn); printf("gpu_ipc = %12.4f\n", (float)gpu_sim_insn / gpu_sim_cycle); @@ -1440,9 +1470,10 @@ void gpgpu_sim::gpu_print_stat() { m_cluster[i]->get_cache_stats(core_cache_stats); } printf("\nTotal_core_cache_stats:\n"); - core_cache_stats.print_stats(stdout, "Total_core_cache_stats_breakdown"); + core_cache_stats.print_stats(stdout, streamID, + "Total_core_cache_stats_breakdown"); printf("\nTotal_core_cache_fail_stats:\n"); - core_cache_stats.print_fail_stats(stdout, + core_cache_stats.print_fail_stats(stdout, streamID, "Total_core_cache_fail_stats_breakdown"); shader_print_scheduler_stat(stdout, false); @@ -1510,9 +1541,10 @@ void gpgpu_sim::gpu_print_stat() { printf("L2_total_cache_reservation_fails = %llu\n", total_l2_css.res_fails); printf("L2_total_cache_breakdown:\n"); - l2_stats.print_stats(stdout, "L2_cache_stats_breakdown"); + l2_stats.print_stats(stdout, streamID, "L2_cache_stats_breakdown"); printf("L2_total_cache_reservation_fail_breakdown:\n"); - l2_stats.print_fail_stats(stdout, "L2_cache_stats_fail_breakdown"); + l2_stats.print_fail_stats(stdout, streamID, + "L2_cache_stats_fail_breakdown"); total_l2_css.print_port_stats(stdout, "L2_cache"); } } @@ -1955,8 +1987,10 @@ void gpgpu_sim::cycle() { if (mf) partiton_reqs_in_parallel_per_cycle++; } m_memory_sub_partition[i]->cache_cycle(gpu_sim_cycle + gpu_tot_sim_cycle); - m_memory_sub_partition[i]->accumulate_L2cache_stats( - m_power_stats->pwr_mem_stat->l2_cache_stats[CURRENT_STAT_IDX]); + if (m_config.g_power_simulation_enabled) { + m_memory_sub_partition[i]->accumulate_L2cache_stats( + m_power_stats->pwr_mem_stat->l2_cache_stats[CURRENT_STAT_IDX]); + } } } partiton_reqs_in_parallel += partiton_reqs_in_parallel_per_cycle; @@ -1978,14 +2012,16 @@ void gpgpu_sim::cycle() { *active_sms += m_cluster[i]->get_n_active_sms(); } // Update core icnt/cache stats for AccelWattch - m_cluster[i]->get_icnt_stats( - m_power_stats->pwr_mem_stat->n_simt_to_mem[CURRENT_STAT_IDX][i], - m_power_stats->pwr_mem_stat->n_mem_to_simt[CURRENT_STAT_IDX][i]); - m_cluster[i]->get_cache_stats( - m_power_stats->pwr_mem_stat->core_cache_stats[CURRENT_STAT_IDX]); - m_cluster[i]->get_current_occupancy( - gpu_occupancy.aggregate_warp_slot_filled, - gpu_occupancy.aggregate_theoretical_warp_slots); + if (m_config.g_power_simulation_enabled) { + m_cluster[i]->get_icnt_stats( + m_power_stats->pwr_mem_stat->n_simt_to_mem[CURRENT_STAT_IDX][i], + m_power_stats->pwr_mem_stat->n_mem_to_simt[CURRENT_STAT_IDX][i]); + m_cluster[i]->get_cache_stats( + m_power_stats->pwr_mem_stat->core_cache_stats[CURRENT_STAT_IDX]); + m_cluster[i]->get_current_occupancy( + gpu_occupancy.aggregate_warp_slot_filled, + gpu_occupancy.aggregate_theoretical_warp_slots); + } } float temp = 0; for (unsigned i = 0; i < m_shader_config->num_shader(); i++) { diff --git a/src/gpgpu-sim/gpu-sim.h b/src/gpgpu-sim/gpu-sim.h index d43b399..8e81451 100644 --- a/src/gpgpu-sim/gpu-sim.h +++ b/src/gpgpu-sim/gpu-sim.h @@ -539,7 +539,7 @@ class gpgpu_sim : public gpgpu_t { (m_config.gpu_max_completed_cta_opt && (gpu_completed_cta >= m_config.gpu_max_completed_cta_opt)); } - void print_stats(); + void print_stats(unsigned long long streamID); void update_stats(); void deadlock_check(); void inc_completed_cta() { gpu_completed_cta++; } @@ -568,7 +568,7 @@ class gpgpu_sim : public gpgpu_t { void decrement_kernel_latency(); const gpgpu_sim_config &get_config() const { return m_config; } - void gpu_print_stat(); + void gpu_print_stat(unsigned long long streamID); void dump_pipeline(int mask, int s, int m) const; void perf_memcpy_to_gpu(size_t dst_start_addr, size_t count); @@ -685,6 +685,17 @@ class gpgpu_sim : public gpgpu_t { occupancy_stats gpu_occupancy; occupancy_stats gpu_tot_occupancy; + typedef struct { + unsigned long long start_cycle; + unsigned long long end_cycle; + } kernel_time_t; + std::map> + gpu_kernel_time; + unsigned long long last_streamID; + unsigned long long last_uid; + cache_stats aggregated_l1_stats; + cache_stats aggregated_l2_stats; + // performance counter for stalls due to congestion. unsigned int gpu_stall_dramfull; unsigned int gpu_stall_icnt2sh; @@ -712,6 +723,9 @@ class gpgpu_sim : public gpgpu_t { public: bool is_functional_sim() { return m_functional_sim; } kernel_info_t *get_functional_kernel() { return m_functional_sim_kernel; } + std::vector get_running_kernels() { + return m_running_kernels; + } void functional_launch(kernel_info_t *k) { m_functional_sim = true; m_functional_sim_kernel = k; diff --git a/src/gpgpu-sim/l2cache.cc b/src/gpgpu-sim/l2cache.cc index 8469453..52eed0e 100644 --- a/src/gpgpu-sim/l2cache.cc +++ b/src/gpgpu-sim/l2cache.cc @@ -51,12 +51,12 @@ mem_fetch *partition_mf_allocator::alloc(new_addr_type addr, mem_access_type type, unsigned size, - bool wr, - unsigned long long cycle) const { + bool wr, unsigned long long cycle, + unsigned long long streamID) const { assert(wr); mem_access_t access(type, addr, size, wr, m_memory_config->gpgpu_ctx); - mem_fetch *mf = new mem_fetch(access, NULL, WRITE_PACKET_SIZE, -1, -1, -1, - m_memory_config, cycle); + mem_fetch *mf = new mem_fetch(access, NULL, streamID, WRITE_PACKET_SIZE, -1, + -1, -1, m_memory_config, cycle); return mf; } @@ -65,12 +65,12 @@ mem_fetch *partition_mf_allocator::alloc( const mem_access_byte_mask_t &byte_mask, const mem_access_sector_mask_t §or_mask, unsigned size, bool wr, unsigned long long cycle, unsigned wid, unsigned sid, unsigned tpc, - mem_fetch *original_mf) const { + mem_fetch *original_mf, unsigned long long streamID) const { mem_access_t access(type, addr, size, wr, active_mask, byte_mask, sector_mask, m_memory_config->gpgpu_ctx); - mem_fetch *mf = - new mem_fetch(access, NULL, wr ? WRITE_PACKET_SIZE : READ_PACKET_SIZE, - wid, sid, tpc, m_memory_config, cycle, original_mf); + mem_fetch *mf = new mem_fetch(access, NULL, streamID, + wr ? WRITE_PACKET_SIZE : READ_PACKET_SIZE, wid, + sid, tpc, m_memory_config, cycle, original_mf); return mf; } memory_partition_unit::memory_partition_unit(unsigned partition_id, @@ -436,9 +436,9 @@ memory_sub_partition::memory_sub_partition(unsigned sub_partition_id, m_mf_allocator = new partition_mf_allocator(config); if (!m_config->m_L2_config.disabled()) - m_L2cache = - new l2_cache(L2c_name, m_config->m_L2_config, -1, -1, m_L2interface, - m_mf_allocator, IN_PARTITION_L2_MISS_QUEUE, gpu); + m_L2cache = new l2_cache(L2c_name, m_config->m_L2_config, -1, -1, + m_L2interface, m_mf_allocator, + IN_PARTITION_L2_MISS_QUEUE, gpu, L2_GPU_CACHE); unsigned int icnt_L2; unsigned int L2_dram; @@ -733,7 +733,7 @@ memory_sub_partition::breakdown_request_to_sector_requests(mem_fetch *mf) { mf->get_access_warp_mask(), mf->get_access_byte_mask() & mask, std::bitset().set(i), SECTOR_SIZE, mf->is_write(), m_gpu->gpu_tot_sim_cycle + m_gpu->gpu_sim_cycle, mf->get_wid(), - mf->get_sid(), mf->get_tpc(), mf); + mf->get_sid(), mf->get_tpc(), mf, mf->get_streamID()); result.push_back(n_mf); } @@ -756,7 +756,7 @@ memory_sub_partition::breakdown_request_to_sector_requests(mem_fetch *mf) { mf->get_access_byte_mask() & mask, std::bitset().set(i), SECTOR_SIZE, mf->is_write(), m_gpu->gpu_tot_sim_cycle + m_gpu->gpu_sim_cycle, mf->get_wid(), - mf->get_sid(), mf->get_tpc(), mf); + mf->get_sid(), mf->get_tpc(), mf, mf->get_streamID()); result.push_back(n_mf); } @@ -772,7 +772,8 @@ memory_sub_partition::breakdown_request_to_sector_requests(mem_fetch *mf) { mf->get_access_warp_mask(), mf->get_access_byte_mask() & mask, std::bitset().set(i), SECTOR_SIZE, mf->is_write(), m_gpu->gpu_tot_sim_cycle + m_gpu->gpu_sim_cycle, - mf->get_wid(), mf->get_sid(), mf->get_tpc(), mf); + mf->get_wid(), mf->get_sid(), mf->get_tpc(), mf, + mf->get_streamID()); result.push_back(n_mf); } diff --git a/src/gpgpu-sim/l2cache.h b/src/gpgpu-sim/l2cache.h index ccf9b70..65c9c38 100644 --- a/src/gpgpu-sim/l2cache.h +++ b/src/gpgpu-sim/l2cache.h @@ -52,15 +52,16 @@ class partition_mf_allocator : public mem_fetch_allocator { return NULL; } virtual mem_fetch *alloc(new_addr_type addr, mem_access_type type, - unsigned size, bool wr, - unsigned long long cycle) const; + unsigned size, bool wr, unsigned long long cycle, + unsigned long long streamID) const; virtual mem_fetch *alloc(new_addr_type addr, mem_access_type type, const active_mask_t &active_mask, const mem_access_byte_mask_t &byte_mask, const mem_access_sector_mask_t §or_mask, unsigned size, bool wr, unsigned long long cycle, unsigned wid, unsigned sid, unsigned tpc, - mem_fetch *original_mf) const; + mem_fetch *original_mf, + unsigned long long streamID) const; private: const memory_config *m_memory_config; diff --git a/src/gpgpu-sim/mem_fetch.cc b/src/gpgpu-sim/mem_fetch.cc index 0d86046..7211a7d 100644 --- a/src/gpgpu-sim/mem_fetch.cc +++ b/src/gpgpu-sim/mem_fetch.cc @@ -35,10 +35,10 @@ unsigned mem_fetch::sm_next_mf_request_uid = 1; mem_fetch::mem_fetch(const mem_access_t &access, const warp_inst_t *inst, - unsigned ctrl_size, unsigned wid, unsigned sid, - unsigned tpc, const memory_config *config, - unsigned long long cycle, mem_fetch *m_original_mf, - mem_fetch *m_original_wr_mf) + unsigned long long streamID, unsigned ctrl_size, + unsigned wid, unsigned sid, unsigned tpc, + const memory_config *config, unsigned long long cycle, + mem_fetch *m_original_mf, mem_fetch *m_original_wr_mf) : m_access(access) { @@ -48,6 +48,7 @@ mem_fetch::mem_fetch(const mem_access_t &access, const warp_inst_t *inst, m_inst = *inst; assert(wid == m_inst.warp_id()); } + m_streamID = streamID; m_data_size = access.get_size(); m_ctrl_size = ctrl_size; m_sid = sid; diff --git a/src/gpgpu-sim/mem_fetch.h b/src/gpgpu-sim/mem_fetch.h index 283fe80..7704218 100644 --- a/src/gpgpu-sim/mem_fetch.h +++ b/src/gpgpu-sim/mem_fetch.h @@ -54,9 +54,10 @@ class memory_config; class mem_fetch { public: mem_fetch(const mem_access_t &access, const warp_inst_t *inst, - unsigned ctrl_size, unsigned wid, unsigned sid, unsigned tpc, - const memory_config *config, unsigned long long cycle, - mem_fetch *original_mf = NULL, mem_fetch *original_wr_mf = NULL); + unsigned long long streamID, unsigned ctrl_size, unsigned wid, + unsigned sid, unsigned tpc, const memory_config *config, + unsigned long long cycle, mem_fetch *original_mf = NULL, + mem_fetch *original_wr_mf = NULL); ~mem_fetch(); void set_status(enum mem_fetch_status status, unsigned long long cycle); @@ -105,6 +106,7 @@ class mem_fetch { unsigned get_timestamp() const { return m_timestamp; } unsigned get_return_timestamp() const { return m_timestamp2; } unsigned get_icnt_receive_time() const { return m_icnt_receive_time; } + unsigned long long get_streamID() const { return m_streamID; } enum mem_access_type get_access_type() const { return m_access.get_type(); } const active_mask_t &get_access_warp_mask() const { @@ -163,6 +165,8 @@ class mem_fetch { // requesting instruction (put last so mem_fetch prints nicer in gdb) warp_inst_t m_inst; + unsigned long long m_streamID; + static unsigned sm_next_mf_request_uid; const memory_config *m_mem_config; diff --git a/src/gpgpu-sim/power_stat.cc b/src/gpgpu-sim/power_stat.cc index dead4a0..764652b 100644 --- a/src/gpgpu-sim/power_stat.cc +++ b/src/gpgpu-sim/power_stat.cc @@ -181,11 +181,11 @@ void power_mem_stat_t::print(FILE *fout) const { total_mem_reads + total_mem_writes); fprintf(fout, "Total memory controller reads: %u\n", total_mem_reads); fprintf(fout, "Total memory controller writes: %u\n", total_mem_writes); - + // TODO: print_stats(require stream ID input) fprintf(fout, "Core cache stats:\n"); - core_cache_stats->print_stats(fout); + core_cache_stats->print_stats(fout, -1); fprintf(fout, "L2 cache stats:\n"); - l2_cache_stats->print_stats(fout); + l2_cache_stats->print_stats(fout, -1); } power_core_stat_t::power_core_stat_t(const shader_core_config *shader_config, diff --git a/src/gpgpu-sim/shader.cc b/src/gpgpu-sim/shader.cc index 9fe4c09..4d4f112 100644 --- a/src/gpgpu-sim/shader.cc +++ b/src/gpgpu-sim/shader.cc @@ -57,11 +57,11 @@ mem_fetch *shader_core_mem_fetch_allocator::alloc( new_addr_type addr, mem_access_type type, unsigned size, bool wr, - unsigned long long cycle) const { + unsigned long long cycle, unsigned long long streamID) const { mem_access_t access(type, addr, size, wr, m_memory_config->gpgpu_ctx); - mem_fetch *mf = - new mem_fetch(access, NULL, wr ? WRITE_PACKET_SIZE : READ_PACKET_SIZE, -1, - m_core_id, m_cluster_id, m_memory_config, cycle); + mem_fetch *mf = new mem_fetch( + access, NULL, streamID, wr ? WRITE_PACKET_SIZE : READ_PACKET_SIZE, -1, + m_core_id, m_cluster_id, m_memory_config, cycle); return mf; } @@ -70,12 +70,12 @@ mem_fetch *shader_core_mem_fetch_allocator::alloc( const mem_access_byte_mask_t &byte_mask, const mem_access_sector_mask_t §or_mask, unsigned size, bool wr, unsigned long long cycle, unsigned wid, unsigned sid, unsigned tpc, - mem_fetch *original_mf) const { + mem_fetch *original_mf, unsigned long long streamID) const { mem_access_t access(type, addr, size, wr, active_mask, byte_mask, sector_mask, m_memory_config->gpgpu_ctx); mem_fetch *mf = new mem_fetch( - access, NULL, wr ? WRITE_PACKET_SIZE : READ_PACKET_SIZE, wid, m_core_id, - m_cluster_id, m_memory_config, cycle, original_mf); + access, NULL, streamID, wr ? WRITE_PACKET_SIZE : READ_PACKET_SIZE, wid, + m_core_id, m_cluster_id, m_memory_config, cycle, original_mf); return mf; } ///////////////////////////////////////////////////////////////////////////// @@ -178,7 +178,7 @@ void shader_core_ctx::create_front_pipeline() { snprintf(name, STRSIZE, "L1I_%03d", m_sid); m_L1I = new read_only_cache(name, m_config->m_L1I_config, m_sid, get_shader_instruction_cache_id(), m_icnt, - IN_L1I_MISS_QUEUE); + IN_L1I_MISS_QUEUE, OTHER_GPU_CACHE, m_gpu); } void shader_core_ctx::create_schedulers() { @@ -447,7 +447,7 @@ void shader_core_ctx::create_exec_pipeline() { m_ldst_unit = new ldst_unit(m_icnt, m_mem_fetch_allocator, this, &m_operand_collector, m_scoreboard, m_config, - m_memory_config, m_stats, m_sid, m_tpc); + m_memory_config, m_stats, m_sid, m_tpc, m_gpu); m_fu.push_back(m_ldst_unit); m_dispatch_port.push_back(ID_OC_MEM); m_issue_port.push_back(OC_EX_MEM); @@ -567,7 +567,8 @@ void shader_core_ctx::init_warps(unsigned cta_id, unsigned start_thread, start_pc = pc; } - m_warp[i]->init(start_pc, cta_id, i, active_threads, m_dynamic_warp_id); + m_warp[i]->init(start_pc, cta_id, i, active_threads, m_dynamic_warp_id, + kernel.get_streamID()); ++m_dynamic_warp_id; m_not_completed += n_active; ++m_active_warps; @@ -985,8 +986,8 @@ void shader_core_ctx::fetch() { // mem_fetch *mf = m_mem_fetch_allocator->alloc() mem_access_t acc(INST_ACC_R, ppc, nbytes, false, m_gpu->gpgpu_ctx); mem_fetch *mf = new mem_fetch( - acc, NULL /*we don't have an instruction yet*/, READ_PACKET_SIZE, - warp_id, m_sid, m_tpc, m_memory_config, + acc, NULL, m_warp[warp_id]->get_kernel_info()->get_streamID(), + READ_PACKET_SIZE, warp_id, m_sid, m_tpc, m_memory_config, m_gpu->gpu_tot_sim_cycle + m_gpu->gpu_sim_cycle); std::list events; enum cache_request_status status; @@ -1040,10 +1041,10 @@ void shader_core_ctx::issue_warp(register_set &pipe_reg_set, m_warp[warp_id]->ibuffer_free(); assert(next_inst->valid()); **pipe_reg = *next_inst; // static instruction information - (*pipe_reg)->issue(active_mask, warp_id, - m_gpu->gpu_tot_sim_cycle + m_gpu->gpu_sim_cycle, - m_warp[warp_id]->get_dynamic_warp_id(), - sch_id); // dynamic instruction information + (*pipe_reg)->issue( + active_mask, warp_id, m_gpu->gpu_tot_sim_cycle + m_gpu->gpu_sim_cycle, + m_warp[warp_id]->get_dynamic_warp_id(), sch_id, + m_warp[warp_id]->get_streamID()); // dynamic instruction information m_stats->shader_cycle_distro[2 + (*pipe_reg)->active_count()]++; func_exec_inst(**pipe_reg); @@ -2597,7 +2598,7 @@ void ldst_unit::init(mem_fetch_interface *icnt, IN_SHADER_L1T_ROB); m_L1C = new read_only_cache(L1C_name, m_config->m_L1C_config, m_sid, get_shader_constant_cache_id(), icnt, - IN_L1C_MISS_QUEUE); + IN_L1C_MISS_QUEUE, OTHER_GPU_CACHE, m_gpu); m_L1D = NULL; m_mem_rc = NO_RC_FAIL; m_num_writeback_clients = @@ -2613,9 +2614,10 @@ ldst_unit::ldst_unit(mem_fetch_interface *icnt, shader_core_ctx *core, opndcoll_rfu_t *operand_collector, Scoreboard *scoreboard, const shader_core_config *config, const memory_config *mem_config, shader_core_stats *stats, - unsigned sid, unsigned tpc) + unsigned sid, unsigned tpc, gpgpu_sim *gpu) : pipelined_simd_unit(NULL, config, config->smem_latency, core, 0), - m_next_wb(config) { + m_next_wb(config), + m_gpu(gpu) { assert(config->smem_latency > 1); init(icnt, mf_allocator, core, operand_collector, scoreboard, config, mem_config, stats, sid, tpc); @@ -2624,7 +2626,7 @@ ldst_unit::ldst_unit(mem_fetch_interface *icnt, snprintf(L1D_name, STRSIZE, "L1D_%03d", m_sid); m_L1D = new l1_cache(L1D_name, m_config->m_L1D_config, m_sid, get_shader_normal_cache_id(), m_icnt, m_mf_allocator, - IN_L1D_MISS_QUEUE, core->get_gpu()); + IN_L1D_MISS_QUEUE, core->get_gpu(), L1_GPU_CACHE); l1_latency_queue.resize(m_config->m_L1D_config.l1_banks); assert(m_config->m_L1D_config.l1_latency > 0); diff --git a/src/gpgpu-sim/shader.h b/src/gpgpu-sim/shader.h index 92691d3..e658a14 100644 --- a/src/gpgpu-sim/shader.h +++ b/src/gpgpu-sim/shader.h @@ -120,6 +120,7 @@ class shd_warp_t { m_done_exit = true; m_last_fetch = 0; m_next = 0; + m_streamID = (unsigned long long)-1; // Jin: cdp support m_cdp_latency = 0; @@ -140,8 +141,9 @@ class shd_warp_t { m_ldgdepbar_buf.clear(); } void init(address_type start_pc, unsigned cta_id, unsigned wid, - const std::bitset &active, - unsigned dynamic_warp_id) { + const std::bitset &active, unsigned dynamic_warp_id, + unsigned long long streamID) { + m_streamID = streamID; m_cta_id = cta_id; m_warp_id = wid; m_dynamic_warp_id = dynamic_warp_id; @@ -265,6 +267,7 @@ class shd_warp_t { m_inst_in_pipeline--; } + unsigned long long get_streamID() const { return m_streamID; } unsigned get_cta_id() const { return m_cta_id; } unsigned get_dynamic_warp_id() const { return m_dynamic_warp_id; } @@ -277,6 +280,7 @@ class shd_warp_t { private: static const unsigned IBUFFER_SIZE = 2; class shader_core_ctx *m_shader; + unsigned long long m_streamID; unsigned m_cta_id; unsigned m_warp_id; unsigned m_warp_size; @@ -1345,7 +1349,7 @@ class ldst_unit : public pipelined_simd_unit { shader_core_ctx *core, opndcoll_rfu_t *operand_collector, Scoreboard *scoreboard, const shader_core_config *config, const memory_config *mem_config, class shader_core_stats *stats, - unsigned sid, unsigned tpc); + unsigned sid, unsigned tpc, gpgpu_sim *gpu); // Add a structure to record the LDGSTS instructions, // similar to m_pending_writes, but since LDGSTS does not have a output @@ -1435,6 +1439,7 @@ class ldst_unit : public pipelined_simd_unit { warp_inst_t &inst); mem_stage_stall_type process_memory_access_queue_l1cache(l1_cache *cache, warp_inst_t &inst); + gpgpu_sim *m_gpu; const memory_config *m_memory_config; class mem_fetch_interface *m_icnt; @@ -2025,18 +2030,20 @@ class shader_core_mem_fetch_allocator : public mem_fetch_allocator { m_memory_config = config; } mem_fetch *alloc(new_addr_type addr, mem_access_type type, unsigned size, - bool wr, unsigned long long cycle) const; + bool wr, unsigned long long cycle, + unsigned long long streamID) const; mem_fetch *alloc(new_addr_type addr, mem_access_type type, const active_mask_t &active_mask, const mem_access_byte_mask_t &byte_mask, const mem_access_sector_mask_t §or_mask, unsigned size, bool wr, unsigned long long cycle, unsigned wid, - unsigned sid, unsigned tpc, mem_fetch *original_mf) const; + unsigned sid, unsigned tpc, mem_fetch *original_mf, + unsigned long long streamID) const; mem_fetch *alloc(const warp_inst_t &inst, const mem_access_t &access, unsigned long long cycle) const { warp_inst_t inst_copy = inst; mem_fetch *mf = new mem_fetch( - access, &inst_copy, + access, &inst_copy, inst.get_streamID(), access.is_write() ? WRITE_PACKET_SIZE : READ_PACKET_SIZE, inst.warp_id(), m_core_id, m_cluster_id, m_memory_config, cycle); return mf; diff --git a/src/gpgpusim_entrypoint.cc b/src/gpgpusim_entrypoint.cc index f4287d8..42c6981 100644 --- a/src/gpgpusim_entrypoint.cc +++ b/src/gpgpusim_entrypoint.cc @@ -57,7 +57,8 @@ void *gpgpu_sim_thread_sequential(void *ctx_ptr) { ctx->the_gpgpusim->g_the_gpu->cycle(); ctx->the_gpgpusim->g_the_gpu->deadlock_check(); } - ctx->the_gpgpusim->g_the_gpu->print_stats(); + ctx->the_gpgpusim->g_the_gpu->print_stats( + ctx->the_gpgpusim->g_the_gpu->last_streamID); ctx->the_gpgpusim->g_the_gpu->update_stats(); ctx->print_simulation_time(); } @@ -144,7 +145,8 @@ void *gpgpu_sim_thread_concurrent(void *ctx_ptr) { fflush(stdout); } if (sim_cycles) { - ctx->the_gpgpusim->g_the_gpu->print_stats(); + ctx->the_gpgpusim->g_the_gpu->print_stats( + ctx->the_gpgpusim->g_the_gpu->last_streamID); ctx->the_gpgpusim->g_the_gpu->update_stats(); ctx->print_simulation_time(); } diff --git a/src/stream_manager.cc b/src/stream_manager.cc index 0ce3c6a..72f8bb0 100644 --- a/src/stream_manager.cc +++ b/src/stream_manager.cc @@ -302,6 +302,14 @@ bool stream_manager::register_finished_kernel(unsigned grid_uid) { void stream_manager::stop_all_running_kernels() { pthread_mutex_lock(&m_lock); + std::vector finished_streams; + std::vector running_kernels = m_gpu->get_running_kernels(); + for (kernel_info_t *k : running_kernels) { + if (k != NULL) { + finished_streams.push_back(k->get_streamID()); + } + } + // Signal m_gpu to stop all running kernels m_gpu->stop_all_running_kernels(); @@ -312,7 +320,9 @@ void stream_manager::stop_all_running_kernels() { } // If any kernels completed, print out the current stats - if (count > 0) m_gpu->print_stats(); + for (unsigned long long streamID : finished_streams) { + m_gpu->print_stats(streamID); + } pthread_mutex_unlock(&m_lock); } -- cgit v1.3 From 752d4e5bf622b0d7c730e3eb2f1b3e3cf91e81fa Mon Sep 17 00:00:00 2001 From: WilliamMTK Date: Wed, 11 Dec 2024 16:52:40 -0500 Subject: Add SST integration into gpgpusim (#44) * Add accommodations to run gpgpusim with SST simulation framework through balar * Output setup_environment options when sourcing * Add SST directive check when creating sim thread * Add sst side test for jenkins * sst-integration: update Jenkinsfile with offical sst-elements repo and fix bugs in pipeline script * sst-integration: direct jenkins to rebuild gpgpusim before testing for sst * sst-integration: fix bugs in sst repos config * sst-integration: let Jenkins rebuilds simulator Since the simulator needs to be configured with both normal mode and sst mode, need to rebuild make target to clean prior runs. * sst-integration: Update Jenkinsfile to source env vars when running balar test * sst-integration: refactor code to remove __SST__ flag * sst-integration: fix a bug that init cluster twice for sst * sst-integration: fix a bug of not sending mem packets to SST * sst-integration: remove sst flags from makefiles and setup_env * sst-integration: add comments to SST changes * sst-integration: remove rebuilding simulator in jenkins when testing for SST * sst-integration: revert simulator build script * Add a function to support querying function argument info for SST * sst-integration: add version detection for vanadis binary * Automated Format * add version detection support for gcc 10+ * sst-integration: add cudaMallocHost for SST * sst-integration: fix a compilation bug * sst-integration: add sst balar unittest CI * sst-integration: specify GPU_ARCH for CI test * sst-integration: use bash for github actions * sst-integration: use https links for sst repos * sst-integration: add SST dependencies to CI config * sst-integration: remove sudo * sst-integration: default to yes for apt install * sst-integration: add manual trigger for github action * sst-integration: remove wrong on event * sst-integration: limit CPU usage for compilation * sst-integration: fix wrong path * sst-integration: use personal repo for testing * sst-integration: remove sst-core source in CI to free space * sst-integration: SST_Cycle use print stats with stream id * Automated Format * sst-integration: check for diskspace and try to clean it * sst-integration: move out of docker image * sst-integration: testing for ci path * sst-integration: fix syntax * sst-integration: pass env vars * sst-integration: set env properly * sst-integration: merge LLVM build and test into same job * sst-integration: fix step order * sst-integration: checkout correct branch for env-setup * sst-integration: remove resourcing gpu apps * sst-integration: revert back to docker github action * sst-integration: enable debug trace for sst testing * sst-integration: resourcing gpu app for env vars * sst-integration: use GPUAPPS_ROOT for path for gpu app * sst-integration: use GPUAPPS_ROOT for path for gpu app * sst-integration: enable parallel ci tests and fix not returning with cudaMallocHostSST * sst-integration: using debug flag for CI run * sst-integration: revert debug ci run * sst-integration: CI skips cuda sdk download and launch multiple jobs * sst-integration: reenable parallel tests * sst-integration: reduce concurrent test thread count * sst-integration: skip long test for github runner * sst-integration: try running CI with single core * sst-integrtion: add callback to SST to check thread sync is done in SST_Cycle() * sst-integration: ignore lookup if already found and add callbacks to SST * Automated Format * sst-integration: add support for indirect texture access * Automated Format * sste-integration: fix up for PR * Automated Format --------- Co-authored-by: purdue-jenkins --- .github/workflows/sst_integration.yml | 80 ++++++++++ Jenkinsfile | 35 +++++ Makefile | 2 + libcuda/cuda_api_object.h | 17 +++ libcuda/cuda_runtime_api.cc | 274 +++++++++++++++++++++++++++++++--- libcuda/gpgpu_context.h | 4 + setup_environment | 18 ++- src/cuda-sim/cuda-sim.cc | 45 ++++-- src/cuda-sim/instructions.cc | 11 ++ src/cuda-sim/ptx_ir.cc | 16 ++ src/cuda-sim/ptx_ir.h | 2 + src/cuda-sim/ptx_sim.h | 3 + src/gpgpu-sim/gpu-cache.cc | 1 + src/gpgpu-sim/gpu-sim.cc | 200 ++++++++++++++++++++++--- src/gpgpu-sim/gpu-sim.h | 139 ++++++++++++++++- src/gpgpu-sim/mem_fetch.cc | 12 +- src/gpgpu-sim/mem_latency_stat.cc | 17 ++- src/gpgpu-sim/shader.cc | 135 +++++++++++++++-- src/gpgpu-sim/shader.h | 101 +++++++++++++ src/gpgpusim_entrypoint.cc | 138 +++++++++++++++-- src/stream_manager.cc | 12 +- 21 files changed, 1176 insertions(+), 86 deletions(-) create mode 100644 .github/workflows/sst_integration.yml (limited to 'src/gpgpu-sim/gpu-sim.h') diff --git a/.github/workflows/sst_integration.yml b/.github/workflows/sst_integration.yml new file mode 100644 index 0000000..03635db --- /dev/null +++ b/.github/workflows/sst_integration.yml @@ -0,0 +1,80 @@ +# Workflow with cmake build system +name: SST Integration Test + +# Controls when the workflow will run +on: + # Triggers the workflow on push or pull request events but only for the mydev branch + push: + branches-ignore: + - "gh-readonly-queue**" + pull_request: + + # Allows you to run this workflow manually from the Actions tab + workflow_dispatch: + +# A workflow run is made up of one or more jobs that can run sequentially or in parallel +jobs: + build-QV100: + runs-on: ubuntu-latest + defaults: + run: + shell: bash + strategy: + matrix: + # test_type: [simple, medium, long] + test_type: [simple, medium] + container: + image: tgrogers/accel-sim_regress:SST-Integration-Ubuntu-22.04-cuda-11.7-llvm-18.1.8-riscv-gnu-2024.08.06-nightly + env: + CONFIG: QV100 + GPU_ARCH: sm_70 + + # Steps represent a sequence of tasks that will be executed as part of the job + steps: + # Checks-out your repository under $GITHUB_WORKSPACE, so your job can access it + - uses: actions/checkout@v4 + - name: Build GPGPU-Sim in SST mode + run: | + source ./setup_environment + make -j4 + - name: Prepare SST dependencies + run: | + apt install -y openmpi-bin openmpi-common libtool libtool-bin autoconf python3 python3-dev automake build-essential git + # Use personal repo for now + - name: Build SST-Core + run: | + git clone https://github.com/William-An/sst-core.git + cd sst-core + git pull + git checkout devel + ./autogen.sh + ./configure --prefix=`realpath ../sstcore-install` --disable-mpi --disable-mem-pools + make -j4 + make install + cd .. + rm -rf ./sst-core + # Use personal repo for now + - name: Build SST-Elements + run: | + git clone https://github.com/William-An/sst-elements.git + source ./setup_environment + cd sst-elements + git pull + git checkout balar-mmio-vanadis-llvm + ./autogen.sh + ./configure --prefix=`realpath ../sstelements-install` --with-sst-core=`realpath ../sstcore-install` --with-cuda=$CUDA_INSTALL_PATH --with-gpgpusim=$GPGPUSIM_ROOT + make -j4 + make install + # Have to resource the gpu app + # Also fake a SDK since rodinia 2.0 does not need this, speed things up on github + - name: Balar Test + run: | + pip install testtools blessings pygments + source ./setup_environment + mkdir 4.2 + mkdir fake_sdk + export NVIDIA_COMPUTE_SDK_LOCATION=$(readlink -f ./fake_sdk) + source $GPUAPPS_ROOT/src/setup_environment sst + rm -rf 4.2 + rm -f gpucomputingsdk_4.2.9_linux.run + ./sstcore-install/bin/sst-test-elements -w "*balar*${{ matrix.test_type }}*" \ No newline at end of file diff --git a/Jenkinsfile b/Jenkinsfile index f6676bf..4ef467b 100644 --- a/Jenkinsfile +++ b/Jenkinsfile @@ -87,6 +87,41 @@ pipeline { ssh tgrogers@dynamo.ecn.purdue.edu "cd $PLOTDIR && rm -rf latest && cp -r ${BUILD_NUMBER} latest"' } } + stage('sst-core-build') { + steps { + sh 'rm -rf sstcore-install' + sh 'rm -rf sst-core && git clone git@github.com:sstsimulator/sst-core.git' + sh '''#!/bin/bash + cd sst-core + ./autogen.sh + ./configure --prefix=`realpath ../sstcore-install` --disable-mpi --disable-mem-pools + make -j 10 + make install''' + } + } + stage('sst-elements-build') { + steps { + sh 'rm -rf sstelements-install' + sh 'rm -rf sst-elements && git clone git@github.com:sstsimulator/sst-elements.git' + // First sourcing the env_setup and setup_environment script for env vars + sh '''#!/bin/bash + source ./env-setup/11.0_env_setup.sh + source `pwd`/setup_environment + cd sst-elements + ./autogen.sh + ./configure --prefix=`realpath ../sstelements-install` --with-sst-core=`realpath ../sstcore-install` --with-cuda=$CUDA_INSTALL_PATH --with-gpgpusim=$GPGPUSIM_ROOT + make -j 10 + make install''' + } + } + stage('sst balar test') { + steps { + sh '''#!/bin/bash + source ./env-setup/11.0_env_setup.sh + source `pwd`/setup_environment sst + ./sstcore-install/bin/sst-test-elements -p ./sst-elements/src/sst/elements/balar/tests''' + } + } } post { success { diff --git a/Makefile b/Makefile index 82ea399..37dba01 100644 --- a/Makefile +++ b/Makefile @@ -34,6 +34,7 @@ INTERSIM ?= intersim2 include version_detection.mk +# Check for debug ifeq ($(GPGPUSIM_CONFIG), gcc-$(CC_VERSION)/cuda-$(CUDART_VERSION)/debug) export DEBUG=1 else @@ -168,6 +169,7 @@ $(SIM_LIB_DIR)/libcudart.so: makedirs $(LIBS) cudalib if [ ! -f $(SIM_LIB_DIR)/libcudart.so.10.0 ]; then ln -s libcudart.so $(SIM_LIB_DIR)/libcudart.so.10.0; fi if [ ! -f $(SIM_LIB_DIR)/libcudart.so.10.1 ]; then ln -s libcudart.so $(SIM_LIB_DIR)/libcudart.so.10.1; fi if [ ! -f $(SIM_LIB_DIR)/libcudart.so.11.0 ]; then ln -s libcudart.so $(SIM_LIB_DIR)/libcudart.so.11.0; fi + if [ ! -f $(SIM_LIB_DIR)/libcudart_mod.so ]; then ln -s libcudart.so $(SIM_LIB_DIR)/libcudart_mod.so; fi $(SIM_LIB_DIR)/libcudart.dylib: makedirs $(LIBS) cudalib g++ -dynamiclib -Wl,-headerpad_max_install_names,-undefined,dynamic_lookup,-compatibility_version,1.1,-current_version,1.1\ diff --git a/libcuda/cuda_api_object.h b/libcuda/cuda_api_object.h index d292e22..e620e57 100644 --- a/libcuda/cuda_api_object.h +++ b/libcuda/cuda_api_object.h @@ -1,6 +1,7 @@ #ifndef __cuda_api_object_h__ #define __cuda_api_object_h__ +#include #include #include #include @@ -193,9 +194,25 @@ class cuda_runtime_api { // backward pointer class gpgpu_context *gpgpu_ctx; // member function list + + // For SST and other potential simulator interface + void cuobjdumpInit(const char *fn); + void extract_code_using_cuobjdump(const char *fn); + void extract_ptx_files_using_cuobjdump(CUctx_st *context, const char *fn); + + // For running GPGPUSim alone void cuobjdumpInit(); void extract_code_using_cuobjdump(); void extract_ptx_files_using_cuobjdump(CUctx_st *context); + + // Internal functions for the above public methods + void cuobjdumpInit_internal(std::function ctx_extract_code_func); + void extract_code_using_cuobjdump_internal( + CUctx_st *context, std::string &app_binary, + std::function ctx_extract_ptx_func); + void extract_ptx_files_using_cuobjdump_internal(CUctx_st *context, + std::string &app_binary); + std::list pruneSectionList(CUctx_st *context); std::list mergeMatchingSections(std::string identifier); std::list mergeSections(); diff --git a/libcuda/cuda_runtime_api.cc b/libcuda/cuda_runtime_api.cc index b540ffd..5dfd3fc 100644 --- a/libcuda/cuda_runtime_api.cc +++ b/libcuda/cuda_runtime_api.cc @@ -109,6 +109,7 @@ #include #include #include +#include #include #include #include @@ -151,6 +152,9 @@ #include #endif +// SST cycle +extern bool SST_Cycle(); + /*DEVICE_BUILTIN*/ struct cudaArray { void *devPtr; @@ -412,6 +416,13 @@ void setCuobjdumpsassfilename( //! processes (e.g. cuobjdump) reading /proc//exe will see the emulator //! executable instead of the application binary. //! +// In SST need the string to pass the binary information +// as we cannot get it from /proc/self/exe +std::string get_app_binary(const char *fn) { + printf("self exe links to: %s\n", fn); + return fn; +} + std::string get_app_binary() { char self_exe_path[1025]; #ifdef __APPLE__ @@ -453,19 +464,25 @@ char *get_app_binary_name(std::string abs_path) { return self_exe_path; } -static int get_app_cuda_version() { +static int get_app_cuda_version_internal(std::string app_binary) { int app_cuda_version = 0; char fname[1024]; snprintf(fname, 1024, "_app_cuda_version_XXXXXX"); int fd = mkstemp(fname); close(fd); + // Weili: Add way to extract CUDA version information from Balar Vanadis + // binary (stored as a const string) std::string app_cuda_version_command = - "ldd " + get_app_binary() + + "ldd " + app_binary + " | grep libcudart.so | sed 's/.*libcudart.so.\\(.*\\) =>.*/\\1/' > " + + fname + " && strings " + app_binary + + " | grep libcudart_vanadis.a | sed " + "'s/.*libcudart_vanadis.a.\\(.*\\)/\\1/' >> " + fname; int res = system(app_cuda_version_command.c_str()); if (res == -1) { - printf("Error - Cannot detect the app's CUDA version.\n"); + printf("Error - Cannot detect the app's CUDA version. Command: %s\n", + app_cuda_version_command.c_str()); exit(1); } FILE *cmd = fopen(fname, "r"); @@ -476,12 +493,24 @@ static int get_app_cuda_version() { } fclose(cmd); if (app_cuda_version == 0) { - printf("Error - Cannot detect the app's CUDA version.\n"); + printf("Error - Cannot detect the app's CUDA version. Command: %s\n", + app_cuda_version_command.c_str()); exit(1); } return app_cuda_version; } +static int get_app_cuda_version(const char *fn) { + // Use for other simulator integration + std::string app_binary = get_app_binary(fn); + return get_app_cuda_version_internal(app_binary); +} + +static int get_app_cuda_version() { + std::string app_binary = get_app_binary(); + return get_app_cuda_version_internal(app_binary); +} + //! Keep track of the association between filename and cubin handle void cuda_runtime_api::cuobjdumpRegisterFatBinary(unsigned int handle, const char *filename, @@ -574,8 +603,11 @@ __host__ cudaError_t CUDARTAPI cudaDeviceGetLimitInternal( return g_last_cudaError = cudaSuccess; } -void **cudaRegisterFatBinaryInternal(void *fatCubin, - gpgpu_context *gpgpu_ctx = NULL) { +// Internal implementation for cudaRegisterFatBiaryInternal +void **cudaRegisterFatBiaryInternal_impl( + void *fatCubin, gpgpu_context *gpgpu_ctx, std::string &app_binary_path, + int app_cuda_version, + std::function ctx_cuobjdumpInit_func) { gpgpu_context *ctx; if (gpgpu_ctx) { ctx = gpgpu_ctx; @@ -606,11 +638,9 @@ void **cudaRegisterFatBinaryInternal(void *fatCubin, // compiled with a newer version of CUDA to run apps compiled with older // versions of CUDA. This is especially useful for PTXPLUS execution. // Skip cuda version check for pytorch application - std::string app_binary_path = get_app_binary(); int pos = app_binary_path.find("python"); if (pos == std::string::npos) { // Not pytorch app : checking cuda version - int app_cuda_version = get_app_cuda_version(); assert( app_cuda_version == CUDART_VERSION / 1000 && "The app must be compiled with same major version as the simulator."); @@ -661,7 +691,7 @@ void **cudaRegisterFatBinaryInternal(void *fatCubin, * then for next calls, only returns the appropriate number */ assert(fat_cubin_handle >= 1); - if (fat_cubin_handle == 1) ctx->api->cuobjdumpInit(); + if (fat_cubin_handle == 1) ctx_cuobjdumpInit_func(ctx); ctx->api->cuobjdumpRegisterFatBinary(fat_cubin_handle, filename, context); return (void **)fat_cubin_handle; @@ -753,6 +783,28 @@ void **cudaRegisterFatBinaryInternal(void *fatCubin, #endif } +void **cudaRegisterFatBinaryInternal(const char *fn, void *fatCubin, + gpgpu_context *gpgpu_ctx = NULL) { + std::string app_binary_path = get_app_binary(fn); + int app_cuda_version = get_app_cuda_version(fn); + auto ctx_cuobjdumpInit = [=](gpgpu_context *ctx) { + ctx->api->cuobjdumpInit(fn); + }; + return cudaRegisterFatBiaryInternal_impl(fatCubin, gpgpu_ctx, app_binary_path, + app_cuda_version, ctx_cuobjdumpInit); +} + +void **cudaRegisterFatBinaryInternal(void *fatCubin, + gpgpu_context *gpgpu_ctx = NULL) { + std::string app_binary_path = get_app_binary(); + int app_cuda_version = get_app_cuda_version(); + auto ctx_cuobjdumpInit = [](gpgpu_context *ctx) { + ctx->api->cuobjdumpInit(); + }; + return cudaRegisterFatBiaryInternal_impl(fatCubin, gpgpu_ctx, app_binary_path, + app_cuda_version, ctx_cuobjdumpInit); +} + void cudaRegisterFunctionInternal(void **fatCubinHandle, const char *hostFun, char *deviceFun, const char *deviceName, int thread_limit, uint3 *tid, uint3 *bid, @@ -1057,6 +1109,24 @@ cudaError_t cudaMallocHostInternal(void **ptr, size_t size, } } +// SST malloc done by vanadis, we just need to record the memory addr +cudaError_t CUDARTAPI cudaMallocHostSSTInternal( + void *addr, size_t size, gpgpu_context *gpgpu_ctx = NULL) { + gpgpu_context *ctx; + if (gpgpu_ctx) { + ctx = gpgpu_ctx; + } else { + ctx = GPGPU_Context(); + } + if (g_debug_execution >= 3) { + announce_call(__my_func__); + } + // track pinned memory size allocated in the host so that same amount of + // memory is also allocated in GPU. + ctx->api->pinned_memory_size[addr] = size; + return g_last_cudaError = cudaSuccess; +} + __host__ cudaError_t CUDARTAPI cudaMallocPitchInternal(void **devPtr, size_t *pitch, size_t width, size_t height, gpgpu_context *gpgpu_ctx = NULL) { @@ -2301,13 +2371,77 @@ cudaDeviceSynchronizeInternal(gpgpu_context *gpgpu_ctx = NULL) { * * *******************************************************************************/ -extern "C" { - /******************************************************************************* * * - * * + * SST Specific functions, used by Balar * * * *******************************************************************************/ + +/** + * @brief Custom function to get CUDA function parameter size and offset + * from PTX parsing result + * + * @param hostFun + * @param index + * @return std::tuple + */ +std::tuple SST_cudaGetParamConfig( + uint64_t hostFun, unsigned index) { + if (g_debug_execution >= 3) { + announce_call(__my_func__); + } + CUctx_st *context = GPGPUSim_Context(GPGPU_Context()); + function_info *entry = context->get_kernel((char *)hostFun); + cudaError_t result = cudaSuccess; + size_t size = 0; + unsigned alignment = 0; + if (index >= entry->num_args()) { + result = cudaErrorAssert; + } else { + std::pair p = entry->get_param_config(index); + size = p.first; + alignment = p.second; + } + return std::tuple(result, size, alignment); +} + +extern "C" { +void SST_receive_mem_reply(unsigned core_id, void *mem_req) { + CUctx_st *context = GPGPUSim_Context(GPGPU_Context()); + static_cast(context->get_device()->get_gpgpu()) + ->SST_receive_mem_reply(core_id, mem_req); + // printf("GPGPU-sim: Recived Request\n"); +} + +bool SST_gpu_core_cycle() { return SST_Cycle(); } + +void SST_gpgpusim_numcores_equal_check(unsigned sst_numcores) { + CUctx_st *context = GPGPUSim_Context(GPGPU_Context()); + static_cast(context->get_device()->get_gpgpu()) + ->SST_gpgpusim_numcores_equal_check(sst_numcores); +} + +uint64_t cudaMallocSST(void **devPtr, size_t size) { + if (g_debug_execution >= 3) { + announce_call(__my_func__); + } + void *test_malloc; + test_malloc = (void *)malloc(size); + void **test_malloc2 = &test_malloc; + CUctx_st *context = GPGPUSim_Context(GPGPU_Context()); + *test_malloc2 = context->get_device()->get_gpgpu()->gpu_malloc(size); + printf("GPGPU-Sim PTX: cudaMallocing %zu bytes starting at 0x%llx..\n", size, + (unsigned long long)*test_malloc2); + if (g_debug_execution >= 3) + printf("GPGPU-Sim PTX: cudaMallocing %zu bytes starting at 0x%llx..\n", + size, (unsigned long long)*test_malloc2); + return (uint64_t)*test_malloc2; +} + +__host__ cudaError_t CUDARTAPI cudaMallocHostSST(void *addr, size_t size) { + return cudaMallocHostSSTInternal(addr, size); +} + cudaError_t cudaPeekAtLastError(void) { return g_last_cudaError; } __host__ cudaError_t CUDARTAPI cudaMalloc(void **devPtr, size_t size) { @@ -2534,6 +2668,7 @@ cudaError_t CUDARTAPI cudaOccupancyMaxActiveBlocksPerMultiprocessorWithFlags( * * * * *******************************************************************************/ + __host__ cudaError_t CUDARTAPI cudaMemset(void *mem, int c, size_t count) { return cudaMemsetInternal(mem, c, count); } @@ -2754,11 +2889,32 @@ __host__ const char *CUDARTAPI cudaGetErrorString(cudaError_t error) { return strdup(buf); } +// SST specific cuda apis +__host__ cudaError_t CUDARTAPI cudaSetupArgumentSST(uint64_t arg, + uint8_t value[200], + size_t size, + size_t offset) { + void *local_value; + local_value = (void *)malloc(size); + + if (arg) { + memcpy(local_value, (void *)&arg, size); + } else { + memcpy(local_value, value, size); + } + return cudaSetupArgumentInternal(local_value, size, offset); +} + __host__ cudaError_t CUDARTAPI cudaSetupArgument(const void *arg, size_t size, size_t offset) { return cudaSetupArgumentInternal(arg, size, offset); } +// SST specific cuda apis +__host__ cudaError_t CUDARTAPI cudaLaunchSST(uint64_t hostFun) { + return cudaLaunchInternal((char *)hostFun); +} + __host__ cudaError_t CUDARTAPI cudaLaunch(const char *hostFun) { return cudaLaunchInternal(hostFun); } @@ -2933,6 +3089,27 @@ __host__ cudaError_t CUDARTAPI cudaThreadSynchronize(void) { return cudaThreadSynchronizeInternal(); } +__host__ cudaError_t CUDARTAPI cudaThreadSynchronizeSST(void) { + // For SST, perform a one-time check and let SST_Cycle() + // do the polling test and invoke callback to SST + // to signal ThreadSynchonize done + gpgpu_context *ctx = GPGPU_Context(); + if (g_debug_execution >= 3) { + announce_call(__my_func__); + } + + // Called on host side + bool thread_sync_done = ctx->synchronize_check(); + g_last_cudaError = cudaSuccess; + if (thread_sync_done) { + // We are already done, so no need to poll for sync done + ctx->requested_synchronize = false; + return cudaSuccess; + } else { + return cudaErrorNotReady; + } +} + int CUDARTAPI __cudaSynchronizeThreads(void **, void *) { if (g_debug_execution >= 3) { announce_call(__my_func__); @@ -2992,10 +3169,10 @@ __host__ cudaError_t CUDARTAPI cudaGetExportTable( // extracts all ptx files from binary and dumps into // prog_name.unique_no.sm_<>.ptx files -void cuda_runtime_api::extract_ptx_files_using_cuobjdump(CUctx_st *context) { +void cuda_runtime_api::extract_ptx_files_using_cuobjdump_internal( + CUctx_st *context, std::string &app_binary) { char command[1000]; char *pytorch_bin = getenv("PYTORCH_BIN"); - std::string app_binary = get_app_binary(); char ptx_list_file_name[1024]; snprintf(ptx_list_file_name, 1024, "_cuobjdump_list_ptx_XXXXXX"); @@ -3062,6 +3239,17 @@ void cuda_runtime_api::extract_ptx_files_using_cuobjdump(CUctx_st *context) { } } +void cuda_runtime_api::extract_ptx_files_using_cuobjdump(CUctx_st *context, + const char *fn) { + std::string app_binary = get_app_binary(fn); + this->extract_ptx_files_using_cuobjdump_internal(context, app_binary); +} + +void cuda_runtime_api::extract_ptx_files_using_cuobjdump(CUctx_st *context) { + std::string app_binary = get_app_binary(); + this->extract_ptx_files_using_cuobjdump_internal(context, app_binary); +} + //! Call cuobjdump to extract everything (-elf -sass -ptx) /*! * This Function extract the whole PTX (for all the files) using cuobjdump @@ -3069,13 +3257,12 @@ void cuda_runtime_api::extract_ptx_files_using_cuobjdump(CUctx_st *context) { *with each binary in its own file It is also responsible for extracting the *libraries linked to the binary if the option is enabled * */ -void cuda_runtime_api::extract_code_using_cuobjdump() { - CUctx_st *context = GPGPUSim_Context(gpgpu_ctx); - +void cuda_runtime_api::extract_code_using_cuobjdump_internal( + CUctx_st *context, std::string &app_binary, + std::function ctx_extract_ptx_func) { // prevent the dumping by cuobjdump everytime we execute the code! const char *override_cuobjdump = getenv("CUOBJDUMP_SIM_FILE"); char command[1000]; - std::string app_binary = get_app_binary(); // Running cuobjdump using dynamic link to current process snprintf(command, 1000, "md5sum %s ", app_binary.c_str()); printf("Running md5sum using \"%s\"\n", command); @@ -3090,7 +3277,7 @@ void cuda_runtime_api::extract_code_using_cuobjdump() { // used by ptxas. int result = 0; #if (CUDART_VERSION >= 6000) - extract_ptx_files_using_cuobjdump(context); + ctx_extract_ptx_func(context); return; #endif // TODO: redundant to dump twice. how can it be prevented? @@ -3222,6 +3409,26 @@ void cuda_runtime_api::extract_code_using_cuobjdump() { } } +void cuda_runtime_api::extract_code_using_cuobjdump(const char *fn) { + CUctx_st *context = GPGPUSim_Context(gpgpu_ctx); + std::string app_binary = get_app_binary(fn); + auto ctx_extract_ptx_func = [=](CUctx_st *context) { + extract_ptx_files_using_cuobjdump(context, fn); + }; + extract_code_using_cuobjdump_internal(context, app_binary, + ctx_extract_ptx_func); +} + +void cuda_runtime_api::extract_code_using_cuobjdump() { + CUctx_st *context = GPGPUSim_Context(gpgpu_ctx); + std::string app_binary = get_app_binary(); + auto ctx_extract_ptx_func = [=](CUctx_st *context) { + extract_ptx_files_using_cuobjdump(context); + }; + extract_code_using_cuobjdump_internal(context, app_binary, + ctx_extract_ptx_func); +} + //! Read file into char* // TODO: convert this to C++ streams, will be way cleaner char *readfile(const std::string filename) { @@ -3466,10 +3673,11 @@ cuobjdumpPTXSection *cuda_runtime_api::findPTXSection( } //! Extract the code using cuobjdump and remove unnecessary sections -void cuda_runtime_api::cuobjdumpInit() { +void cuda_runtime_api::cuobjdumpInit_internal( + std::function ctx_extract_code_func) { CUctx_st *context = GPGPUSim_Context(gpgpu_ctx); - extract_code_using_cuobjdump(); // extract all the output of cuobjdump to - // _cuobjdump_*.* + ctx_extract_code_func(); // extract all the output of cuobjdump to + // _cuobjdump_*.* const char *pre_load = getenv("CUOBJDUMP_SIM_FILE"); if (pre_load == NULL || strlen(pre_load) == 0) { cuobjdumpSectionList = pruneSectionList(context); @@ -3477,6 +3685,16 @@ void cuda_runtime_api::cuobjdumpInit() { } } +void cuda_runtime_api::cuobjdumpInit(const char *fn) { + auto ctx_extract_code_func = [=]() { extract_code_using_cuobjdump(fn); }; + cuobjdumpInit_internal(ctx_extract_code_func); +} + +void cuda_runtime_api::cuobjdumpInit() { + auto ctx_extract_code_func = [=]() { extract_code_using_cuobjdump(); }; + cuobjdumpInit_internal(ctx_extract_code_func); +} + //! Either submit PTX for simulation or convert SASS to PTXPlus and submit it void gpgpu_context::cuobjdumpParseBinary(unsigned int handle) { CUctx_st *context = GPGPUSim_Context(this); @@ -3587,6 +3805,10 @@ void gpgpu_context::cuobjdumpParseBinary(unsigned int handle) { extern "C" { +void **CUDARTAPI __cudaRegisterFatBinarySST(const char *fn) { + return cudaRegisterFatBinaryInternal(fn, NULL); +} + void **CUDARTAPI __cudaRegisterFatBinary(void *fatCubin) { if (g_debug_execution >= 3) { announce_call(__my_func__); @@ -3619,6 +3841,14 @@ cudaError_t CUDARTAPI __cudaPopCallConfiguration(dim3 *gridDim, dim3 *blockDim, return g_last_cudaError = cudaSuccess; } +void CUDARTAPI __cudaRegisterFunctionSST(unsigned fatCubinHandle, + uint64_t hostFun, + char deviceFun[512]) { + cudaRegisterFunctionInternal((void **)fatCubinHandle, (const char *)hostFun, + (char *)deviceFun, NULL, NULL, NULL, NULL, NULL, + NULL); +} + void CUDARTAPI __cudaRegisterFunction(void **fatCubinHandle, const char *hostFun, char *deviceFun, const char *deviceName, int thread_limit, diff --git a/libcuda/gpgpu_context.h b/libcuda/gpgpu_context.h index d0cd7c4..5ef2117 100644 --- a/libcuda/gpgpu_context.h +++ b/libcuda/gpgpu_context.h @@ -44,6 +44,9 @@ class gpgpu_context { s_g_pc_to_insn; // a direct mapping from PC to instruction bool debug_tensorcore; + // SST related + bool requested_synchronize = false; + // objects pointers for each file cuda_runtime_api *api; ptxinfo_data *ptxinfo; @@ -54,6 +57,7 @@ class gpgpu_context { ptx_stats *stats; // member function list void synchronize(); + bool synchronize_check(); void exit_simulation(); void print_simulation_time(); int gpgpu_opencl_ptx_sim_main_perf(kernel_info_t *grid); diff --git a/setup_environment b/setup_environment index 3428101..2fac1b9 100644 --- a/setup_environment +++ b/setup_environment @@ -46,7 +46,6 @@ fi CC_VERSION=$(gcc --version | head -1 | awk '{for(i=1;i<=NF;i++){ if(match($i,/^[0-9]+\.[0-9]+\.[0-9]+$/)) {print $i; exit 0}}}') - CUDA_VERSION_STRING=`$CUDA_INSTALL_PATH/bin/nvcc --version | awk '/release/ {print $5;}' | sed 's/,//'`; export CUDA_VERSION_NUMBER=`echo $CUDA_VERSION_STRING | sed 's/\./ /' | awk '{printf("%02u%02u", 10*int($1), 10*$2);}'` if [ $CUDA_VERSION_NUMBER -gt 11100 -o $CUDA_VERSION_NUMBER -lt 2030 ]; then @@ -61,13 +60,18 @@ if [ $CUDA_VERSION_NUMBER -ge 6000 ]; then export CUOBJDUMP_SIM_FILE=jj fi +# Simple configure, loop through all positional arguments +# Default config +export GPGPUSIM_CONFIG=gcc-$CC_VERSION/cuda-$CUDA_VERSION_NUMBER/release -if [ $# = '1' ] ; -then - export GPGPUSIM_CONFIG=gcc-$CC_VERSION/cuda-$CUDA_VERSION_NUMBER/$1 -else - export GPGPUSIM_CONFIG=gcc-$CC_VERSION/cuda-$CUDA_VERSION_NUMBER/release -fi +for opt in $@ +do + if [[ $opt == 'debug' ]] ; then + # Debug mode + echo -n "enabled debug mode " + export GPGPUSIM_CONFIG=gcc-$CC_VERSION/cuda-$CUDA_VERSION_NUMBER/$1 + fi +done export QTINC=/usr/include diff --git a/src/cuda-sim/cuda-sim.cc b/src/cuda-sim/cuda-sim.cc index 833d33f..2fd90c0 100644 --- a/src/cuda-sim/cuda-sim.cc +++ b/src/cuda-sim/cuda-sim.cc @@ -1305,7 +1305,12 @@ void function_info::add_param_name_type_size(unsigned index, std::string name, void function_info::add_param_data(unsigned argn, struct gpgpu_ptx_sim_arg *args) { const void *data = args->m_start; - + if (g_debug_execution >= 3) { + if (args->m_nbytes == 4) + printf("ADD_PARAM_DATA %d\n", *((uint32_t *)data)); + else + printf("ADD_PARAM_DATA %p\n", *((void **)data)); + } bool scratchpad_memory_param = false; // Is this parameter in CUDA shared memory or OpenCL local memory @@ -1746,6 +1751,17 @@ static unsigned get_tex_datasize(const ptx_instruction *pI, ptx_thread_info *thread) { const operand_info &src1 = pI->src1(); // the name of the texture std::string texname = src1.name(); + // If indirect access, use register's value as address + // to find the symbol + if (src1.is_reg()) { + const operand_info &dst = pI->dst(); + ptx_reg_t src1_data = + thread->get_operand_value(src1, dst, pI->get_type(), thread, 1); + addr_t sym_addr = src1_data.u64; + symbol *texRef = thread->get_symbol_table()->lookup_by_addr(sym_addr); + assert(texRef != NULL); + texname = texRef->name(); + } /* For programs with many streams, textures can be bound and unbound @@ -2285,15 +2301,24 @@ void cuda_sim::gpgpu_ptx_sim_memcpy_symbol(const char *hostVar, const void *src, sym_name = g->second; mem_region = global_space; } - if (g_globals.find(hostVar) != g_globals.end()) { - found_sym = true; - sym_name = hostVar; - mem_region = global_space; - } - if (g_constants.find(hostVar) != g_constants.end()) { - found_sym = true; - sym_name = hostVar; - mem_region = const_space; + + // Weili: Only attempt to find symbol as it is a string + // if we could not find it in previously registered variable. + // This will avoid constructing std::string() from hostVar address + // where it is not a string as + // Use of a string naming a variable as the symbol parameter was deprecated in + // CUDA 4.1 and removed in CUDA 5.0. + if (!found_sym) { + if (g_globals.find(hostVar) != g_globals.end()) { + found_sym = true; + sym_name = hostVar; + mem_region = global_space; + } + if (g_constants.find(hostVar) != g_constants.end()) { + found_sym = true; + sym_name = hostVar; + mem_region = const_space; + } } if (!found_sym) { diff --git a/src/cuda-sim/instructions.cc b/src/cuda-sim/instructions.cc index 108de97..843bf0b 100644 --- a/src/cuda-sim/instructions.cc +++ b/src/cuda-sim/instructions.cc @@ -6055,6 +6055,17 @@ void tex_impl(const ptx_instruction *pI, ptx_thread_info *thread) { // to be fetched std::string texname = src1.name(); + // If indirect access, use register's value as address + // to find the symbol + if (src1.is_reg()) { + ptx_reg_t src1_data = + thread->get_operand_value(src1, dst, pI->get_type(), thread, 1); + addr_t sym_addr = src1_data.u64; + symbol *texRef = thread->get_symbol_table()->lookup_by_addr(sym_addr); + assert(texRef != NULL); + texname = texRef->name(); + } + unsigned to_type = pI->get_type(); unsigned c_type = pI->get_type2(); fflush(stdout); diff --git a/src/cuda-sim/ptx_ir.cc b/src/cuda-sim/ptx_ir.cc index 1399209..4e500cc 100644 --- a/src/cuda-sim/ptx_ir.cc +++ b/src/cuda-sim/ptx_ir.cc @@ -139,6 +139,22 @@ symbol *symbol_table::lookup(const char *identifier) { return NULL; } +symbol *symbol_table::lookup_by_addr(addr_t addr) { + for (auto it = m_symbols.begin(); it != m_symbols.end(); ++it) { + symbol *sym = it->second; + + // check if symbol has the addr to be found + if ((!sym->is_reg()) && (sym->has_valid_address()) && + (sym->get_address() == addr)) { + return sym; + } + } + if (m_parent) { + return m_parent->lookup_by_addr(addr); + } + return NULL; +} + symbol *symbol_table::add_variable(const char *identifier, const type_info *type, unsigned size, const char *filename, unsigned line) { diff --git a/src/cuda-sim/ptx_ir.h b/src/cuda-sim/ptx_ir.h index d253866..b08a692 100644 --- a/src/cuda-sim/ptx_ir.h +++ b/src/cuda-sim/ptx_ir.h @@ -205,6 +205,7 @@ class symbol { const std::string &name() const { return m_name; } const std::string &decl_location() const { return m_decl_location; } const type_info *type() const { return m_type; } + bool has_valid_address() const { return m_address_valid; } addr_t get_address() const { assert(m_is_label || !m_type->get_key().is_reg()); // todo : other assertions @@ -310,6 +311,7 @@ class symbol_table { void set_ptx_version(float ver, unsigned ext); void set_sm_target(const char *target, const char *ext, const char *ext2); symbol *lookup(const char *identifier); + symbol *lookup_by_addr(addr_t addr); std::string get_scope_name() const { return m_scope_name; } symbol *add_variable(const char *identifier, const type_info *type, unsigned size, const char *filename, unsigned line); diff --git a/src/cuda-sim/ptx_sim.h b/src/cuda-sim/ptx_sim.h index f0c26ef..8eec922 100644 --- a/src/cuda-sim/ptx_sim.h +++ b/src/cuda-sim/ptx_sim.h @@ -459,6 +459,9 @@ class ptx_thread_info { // Jin: get corresponding kernel grid for CDP purpose kernel_info_t &get_kernel() { return m_kernel; } + // Weili: access symbol_table + symbol_table *get_symbol_table() { return m_symbol_table; } + public: addr_t m_last_effective_address; bool m_branch_taken; diff --git a/src/gpgpu-sim/gpu-cache.cc b/src/gpgpu-sim/gpu-cache.cc index cd3c880..0ea9ff6 100644 --- a/src/gpgpu-sim/gpu-cache.cc +++ b/src/gpgpu-sim/gpu-cache.cc @@ -2062,6 +2062,7 @@ enum cache_request_status tex_cache::access(new_addr_type addr, mem_fetch *mf, void tex_cache::cycle() { // send next request to lower level of memory + // TODO: Use different full() for sst_mem_interface? if (!m_request_fifo.empty()) { mem_fetch *mf = m_request_fifo.peek(); if (!m_memport->full(mf->get_ctrl_size(), false)) { diff --git a/src/gpgpu-sim/gpu-sim.cc b/src/gpgpu-sim/gpu-sim.cc index 5bd4180..b92494b 100644 --- a/src/gpgpu-sim/gpu-sim.cc +++ b/src/gpgpu-sim/gpu-sim.cc @@ -319,6 +319,9 @@ void memory_config::reg_options(class OptionParser *opp) { "elimnate_rw_turnaround i.e set tWTR and tRTW = 0", "0"); option_parser_register(opp, "-icnt_flit_size", OPT_UINT32, &icnt_flit_size, "icnt_flit_size", "32"); + // SST mode activate + option_parser_register(opp, "-SST_mode", OPT_BOOL, &SST_mode, "SST mode", + "0"); m_address_mapping.addrdec_setoption(opp); } @@ -955,6 +958,16 @@ void exec_gpgpu_sim::createSIMTCluster() { m_shader_stats, m_memory_stats); } +// SST get its own simt_cluster +void sst_gpgpu_sim::createSIMTCluster() { + m_cluster = new simt_core_cluster *[m_shader_config->n_simt_clusters]; + for (unsigned i = 0; i < m_shader_config->n_simt_clusters; i++) + m_cluster[i] = + new sst_simt_core_cluster(this, i, m_shader_config, m_memory_config, + m_shader_stats, m_memory_stats); + SST_gpgpu_reply_buffer.resize(m_shader_config->n_simt_clusters); +} + gpgpu_sim::gpgpu_sim(const gpgpu_sim_config &config, gpgpu_context *ctx) : gpgpu_t(config, ctx), m_config(config) { gpgpu_ctx = ctx; @@ -999,26 +1012,29 @@ gpgpu_sim::gpgpu_sim(const gpgpu_sim_config &config, gpgpu_context *ctx) gpu_kernel_time.clear(); - m_memory_partition_unit = - new memory_partition_unit *[m_memory_config->m_n_mem]; - m_memory_sub_partition = - new memory_sub_partition *[m_memory_config->m_n_mem_sub_partition]; - for (unsigned i = 0; i < m_memory_config->m_n_mem; i++) { - m_memory_partition_unit[i] = - new memory_partition_unit(i, m_memory_config, m_memory_stats, this); - for (unsigned p = 0; - p < m_memory_config->m_n_sub_partition_per_memory_channel; p++) { - unsigned submpid = - i * m_memory_config->m_n_sub_partition_per_memory_channel + p; - m_memory_sub_partition[submpid] = - m_memory_partition_unit[i]->get_sub_partition(p); + // TODO: somehow move this logic to the sst_gpgpu_sim constructor? + if (!m_config.is_SST_mode()) { + // Init memory if not in SST mode + m_memory_partition_unit = + new memory_partition_unit *[m_memory_config->m_n_mem]; + m_memory_sub_partition = + new memory_sub_partition *[m_memory_config->m_n_mem_sub_partition]; + for (unsigned i = 0; i < m_memory_config->m_n_mem; i++) { + m_memory_partition_unit[i] = + new memory_partition_unit(i, m_memory_config, m_memory_stats, this); + for (unsigned p = 0; + p < m_memory_config->m_n_sub_partition_per_memory_channel; p++) { + unsigned submpid = + i * m_memory_config->m_n_sub_partition_per_memory_channel + p; + m_memory_sub_partition[submpid] = + m_memory_partition_unit[i]->get_sub_partition(p); + } } - } - - icnt_wrapper_init(); - icnt_create(m_shader_config->n_simt_clusters, - m_memory_config->m_n_mem_sub_partition); + icnt_wrapper_init(); + icnt_create(m_shader_config->n_simt_clusters, + m_memory_config->m_n_mem_sub_partition); + } time_vector_create(NUM_MEM_REQ_STAT); fprintf(stdout, "GPGPU-Sim uArch: performance model initialization complete.\n"); @@ -1037,6 +1053,22 @@ gpgpu_sim::gpgpu_sim(const gpgpu_sim_config &config, gpgpu_context *ctx) m_functional_sim_kernel = NULL; } +void sst_gpgpu_sim::SST_receive_mem_reply(unsigned core_id, void *mem_req) { + assert(core_id < m_shader_config->n_simt_clusters); + mem_fetch *mf = (mem_fetch *)mem_req; + + (SST_gpgpu_reply_buffer[core_id]).push_back(mf); +} + +mem_fetch *sst_gpgpu_sim::SST_pop_mem_reply(unsigned core_id) { + if (SST_gpgpu_reply_buffer[core_id].size() > 0) { + mem_fetch *temp = SST_gpgpu_reply_buffer[core_id].front(); + SST_gpgpu_reply_buffer[core_id].pop_front(); + return temp; + } else + return NULL; +} + int gpgpu_sim::shared_mem_size() const { return m_shader_config->gpgpu_shmem_size; } @@ -1132,6 +1164,26 @@ bool gpgpu_sim::active() { return false; } +bool sst_gpgpu_sim::active() { + if (m_config.gpu_max_cycle_opt && + (gpu_tot_sim_cycle + gpu_sim_cycle) >= m_config.gpu_max_cycle_opt) + return false; + if (m_config.gpu_max_insn_opt && + (gpu_tot_sim_insn + gpu_sim_insn) >= m_config.gpu_max_insn_opt) + return false; + if (m_config.gpu_max_cta_opt && + (gpu_tot_issued_cta >= m_config.gpu_max_cta_opt)) + return false; + if (m_config.gpu_max_completed_cta_opt && + (gpu_completed_cta >= m_config.gpu_max_completed_cta_opt)) + return false; + if (m_config.gpu_deadlock_detect && gpu_deadlock) return false; + for (unsigned i = 0; i < m_shader_config->n_simt_clusters; i++) + if (m_cluster[i]->get_not_completed() > 0) return true; + if (get_more_cta_left()) return true; + return false; +} + void gpgpu_sim::init() { // run a CUDA grid on the GPU microarchitecture simulator gpu_sim_cycle = 0; @@ -2157,6 +2209,11 @@ void gpgpu_sim::cycle() { } } +void sst_gpgpu_sim::cycle() { + SST_cycle(); + return; +} + void shader_core_ctx::dump_warp_state(FILE *fout) const { fprintf(fout, "\n"); fprintf(fout, "per warp functional simulation status:\n"); @@ -2236,3 +2293,110 @@ const shader_core_config *gpgpu_sim::getShaderCoreConfig() { const memory_config *gpgpu_sim::getMemoryConfig() { return m_memory_config; } simt_core_cluster *gpgpu_sim::getSIMTCluster() { return *m_cluster; } + +void sst_gpgpu_sim::SST_gpgpusim_numcores_equal_check(unsigned sst_numcores) { + if (m_shader_config->n_simt_clusters != sst_numcores) { + assert( + "\nSST core is not equal the GPGPU-sim cores. Open gpgpu-sim.config " + "file and ensure n_simt_clusters" + "is the same as SST gpu cores.\n" && + 0); + } else { + printf("\nSST GPU core is equal the GPGPU-sim cores = %d\n", sst_numcores); + } +} + +void sst_gpgpu_sim::SST_cycle() { + // shader core loading (pop from ICNT into core) follows CORE clock + for (unsigned i = 0; i < m_shader_config->n_simt_clusters; i++) + static_cast(m_cluster[i])->icnt_cycle_SST(); + + // L1 cache + shader core pipeline stages + m_power_stats->pwr_mem_stat->core_cache_stats[CURRENT_STAT_IDX].clear(); + for (unsigned i = 0; i < m_shader_config->n_simt_clusters; i++) { + if (m_cluster[i]->get_not_completed() || get_more_cta_left()) { + m_cluster[i]->core_cycle(); + *active_sms += m_cluster[i]->get_n_active_sms(); + } + // Update core icnt/cache stats for GPUWattch + m_cluster[i]->get_icnt_stats( + m_power_stats->pwr_mem_stat->n_simt_to_mem[CURRENT_STAT_IDX][i], + m_power_stats->pwr_mem_stat->n_mem_to_simt[CURRENT_STAT_IDX][i]); + m_cluster[i]->get_cache_stats( + m_power_stats->pwr_mem_stat->core_cache_stats[CURRENT_STAT_IDX]); + } + float temp = 0; + for (unsigned i = 0; i < m_shader_config->num_shader(); i++) { + temp += m_shader_stats->m_pipeline_duty_cycle[i]; + } + temp = temp / m_shader_config->num_shader(); + *average_pipeline_duty_cycle = ((*average_pipeline_duty_cycle) + temp); + // cout<<"Average pipeline duty cycle: "<<*average_pipeline_duty_cycle<= g_single_step)) { + asm("int $03"); + } + gpu_sim_cycle++; + if (g_interactive_debugger_enabled) gpgpu_debug(); + + // McPAT main cycle (interface with McPAT) +#ifdef GPGPUSIM_POWER_MODEL + if (m_config.g_power_simulation_enabled) { + mcpat_cycle(m_config, getShaderCoreConfig(), m_gpgpusim_wrapper, + m_power_stats, m_config.gpu_stat_sample_freq, gpu_tot_sim_cycle, + gpu_sim_cycle, gpu_tot_sim_insn, gpu_sim_insn, + m_config.g_dvfs_enabled); + } +#endif + + issue_block2core(); + + if (!(gpu_sim_cycle % m_config.gpu_stat_sample_freq)) { + time_t days, hrs, minutes, sec; + time_t curr_time; + time(&curr_time); + unsigned long long elapsed_time = + MAX(curr_time - gpgpu_ctx->the_gpgpusim->g_simulation_starttime, 1); + if ((elapsed_time - last_liveness_message_time) >= + m_config.liveness_message_freq) { + days = elapsed_time / (3600 * 24); + hrs = elapsed_time / 3600 - 24 * days; + minutes = elapsed_time / 60 - 60 * (hrs + 24 * days); + sec = elapsed_time - 60 * (minutes + 60 * (hrs + 24 * days)); + + last_liveness_message_time = elapsed_time; + } + visualizer_printstat(); + m_memory_stats->memlatstat_lat_pw(); + if (m_config.gpgpu_runtime_stat && (m_config.gpu_runtime_stat_flag != 0)) { + if (m_config.gpu_runtime_stat_flag & GPU_RSTAT_BW_STAT) { + for (unsigned i = 0; i < m_memory_config->m_n_mem; i++) + m_memory_partition_unit[i]->print_stat(stdout); + printf("maxmrqlatency = %d \n", m_memory_stats->max_mrq_latency); + printf("maxmflatency = %d \n", m_memory_stats->max_mf_latency); + } + if (m_config.gpu_runtime_stat_flag & GPU_RSTAT_SHD_INFO) + shader_print_runtime_stat(stdout); + if (m_config.gpu_runtime_stat_flag & GPU_RSTAT_L1MISS) + shader_print_l1_miss_stat(stdout); + if (m_config.gpu_runtime_stat_flag & GPU_RSTAT_SCHED) + shader_print_scheduler_stat(stdout, false); + } + } + + if (!(gpu_sim_cycle % 20000)) { + // deadlock detection + if (m_config.gpu_deadlock_detect && gpu_sim_insn == last_gpu_sim_insn) { + gpu_deadlock = true; + } else { + last_gpu_sim_insn = gpu_sim_insn; + } + } + try_snap_shot(gpu_sim_cycle); + spill_log_to_file(stdout, 0, gpu_sim_cycle); + +#if (CUDART_VERSION >= 5000) + // launch device kernel + gpgpu_ctx->device_runtime->launch_one_device_kernel(); +#endif +} diff --git a/src/gpgpu-sim/gpu-sim.h b/src/gpgpu-sim/gpu-sim.h index 8e81451..d0c2a17 100644 --- a/src/gpgpu-sim/gpu-sim.h +++ b/src/gpgpu-sim/gpu-sim.h @@ -69,6 +69,38 @@ class gpgpu_context; extern tr1_hash_map address_random_interleaving; +// SST communication functions +/** + * @brief Check if SST requests buffer is full + * + * @param core_id + * @return true + * @return false + */ +extern bool is_SST_buffer_full(unsigned core_id); + +/** + * @brief Send loads to SST memory backend + * + * @param core_id + * @param address + * @param size + * @param mem_req + */ +extern void send_read_request_SST(unsigned core_id, uint64_t address, + size_t size, void *mem_req); + +/** + * @brief Send stores to SST memory backend + * + * @param core_id + * @param address + * @param size + * @param mem_req + */ +extern void send_write_request_SST(unsigned core_id, uint64_t address, + size_t size, void *mem_req); + enum dram_ctrl_t { DRAM_FIFO = 0, DRAM_FRFCFS = 1 }; enum hw_perf_t { @@ -274,6 +306,14 @@ class memory_config { } void reg_options(class OptionParser *opp); + /** + * @brief Check if the config script is in SST mode + * + * @return true + * @return false + */ + bool is_SST_mode() const { return SST_mode; } + bool m_valid; mutable l2_cache_config m_L2_config; bool m_L2_texure_only; @@ -351,7 +391,7 @@ class memory_config { unsigned write_low_watermark; bool m_perf_sim_memcpy; bool simple_dram_model; - + bool SST_mode; gpgpu_context *gpgpu_ctx; }; @@ -398,6 +438,15 @@ class gpgpu_sim_config : public power_config, unsigned num_shader() const { return m_shader_config.num_shader(); } unsigned num_cluster() const { return m_shader_config.n_simt_clusters; } unsigned get_max_concurrent_kernel() const { return max_concurrent_kernel; } + + /** + * @brief Check if we are in SST mode + * + * @return true + * @return false + */ + bool is_SST_mode() const { return m_memory_config.SST_mode; } + unsigned checkpoint_option; size_t stack_limit() const { return stack_size_limit; } @@ -462,6 +511,7 @@ class gpgpu_sim_config : public power_config, unsigned long long liveness_message_freq; friend class gpgpu_sim; + friend class sst_gpgpu_sim; }; struct occupancy_stats { @@ -600,10 +650,18 @@ class gpgpu_sim : public gpgpu_t { void hit_watchpoint(unsigned watchpoint_num, ptx_thread_info *thd, const ptx_instruction *pI); + /** + * @brief Check if we are in SST mode + * + * @return true + * @return false + */ + bool is_SST_mode() { return m_config.is_SST_mode(); } + // backward pointer class gpgpu_context *gpgpu_ctx; - private: + protected: // clocks void reinit_clock_domains(void); int next_clock_domain(void); @@ -715,7 +773,7 @@ class gpgpu_sim : public gpgpu_t { void set_cache_config(std::string kernel_name); // Jin: functional simulation for CDP - private: + protected: // set by stream operation every time a functoinal simulation is done bool m_functional_sim; kernel_info_t *m_functional_sim_kernel; @@ -748,4 +806,79 @@ class exec_gpgpu_sim : public gpgpu_sim { virtual void createSIMTCluster(); }; +/** + * @brief A GPGPUSim class customized to SST Balar interfacing + * + */ +class sst_gpgpu_sim : public gpgpu_sim { + public: + sst_gpgpu_sim(const gpgpu_sim_config &config, gpgpu_context *ctx) + : gpgpu_sim(config, ctx) { + createSIMTCluster(); + } + + // SST memory handling + std::vector> + SST_gpgpu_reply_buffer; /** SST mem response queue */ + + /** + * @brief Receive mem request's response from SST and put + * it in a buffer (SST_gpgpu_reply_buffer) + * + * @param core_id + * @param mem_req + */ + void SST_receive_mem_reply(unsigned core_id, void *mem_req); + + /** + * @brief Pop the head of the buffer queue to get the + * memory response + * + * @param core_id + * @return mem_fetch* + */ + mem_fetch *SST_pop_mem_reply(unsigned core_id); + + virtual void createSIMTCluster(); + + // SST Balar interfacing + /** + * @brief Advance core and collect stats + * + */ + void SST_cycle(); + + /** + * @brief Wrapper of SST_cycle() + * + */ + void cycle(); + + /** + * @brief Whether the GPU is active, removed test for + * memory system since that is handled in SST + * + * @return true + * @return false + */ + bool active(); + + /** + * @brief SST mode use SST memory system instead, so the memcpy + * is empty here + * + * @param dst_start_addr + * @param count + */ + void perf_memcpy_to_gpu(size_t dst_start_addr, size_t count){}; + + /** + * @brief Check if the SST config matches up with the + * gpgpusim.config in core number + * + * @param sst_numcores SST core count + */ + void SST_gpgpusim_numcores_equal_check(unsigned sst_numcores); +}; + #endif diff --git a/src/gpgpu-sim/mem_fetch.cc b/src/gpgpu-sim/mem_fetch.cc index 7211a7d..809c920 100644 --- a/src/gpgpu-sim/mem_fetch.cc +++ b/src/gpgpu-sim/mem_fetch.cc @@ -54,9 +54,15 @@ mem_fetch::mem_fetch(const mem_access_t &access, const warp_inst_t *inst, m_sid = sid; m_tpc = tpc; m_wid = wid; - config->m_address_mapping.addrdec_tlx(access.get_addr(), &m_raw_addr); - m_partition_addr = - config->m_address_mapping.partition_address(access.get_addr()); + + if (!config->is_SST_mode()) { + // In SST memory model, the SST memory hierarchy is + // responsible to generate the correct address mapping + config->m_address_mapping.addrdec_tlx(access.get_addr(), &m_raw_addr); + m_partition_addr = + config->m_address_mapping.partition_address(access.get_addr()); + } + m_type = m_access.is_write() ? WRITE_REQUEST : READ_REQUEST; m_timestamp = cycle; m_timestamp2 = 0; diff --git a/src/gpgpu-sim/mem_latency_stat.cc b/src/gpgpu-sim/mem_latency_stat.cc index 63d7ee8..c77a686 100644 --- a/src/gpgpu-sim/mem_latency_stat.cc +++ b/src/gpgpu-sim/mem_latency_stat.cc @@ -203,7 +203,15 @@ unsigned memory_stats_t::memlatstat_done(mem_fetch *mf) { } void memory_stats_t::memlatstat_read_done(mem_fetch *mf) { - if (m_memory_config->gpgpu_memlatency_stat) { + if (m_memory_config->SST_mode) { + // in SST mode, we just calculate mem latency + unsigned mf_latency; + mf_latency = + (m_gpu->gpu_sim_cycle + m_gpu->gpu_tot_sim_cycle) - mf->get_timestamp(); + num_mfs++; + mf_total_lat += mf_latency; + if (mf_latency > max_mf_latency) max_mf_latency = mf_latency; + } else if (m_memory_config->gpgpu_memlatency_stat) { unsigned mf_latency = memlatstat_done(mf); if (mf_latency > mf_max_lat_table[mf->get_tlx_addr().chip][mf->get_tlx_addr().bk]) @@ -273,7 +281,12 @@ void memory_stats_t::memlatstat_print(unsigned n_mem, unsigned gpu_mem_n_bk) { unsigned max_bank_accesses, min_bank_accesses, max_chip_accesses, min_chip_accesses; - if (m_memory_config->gpgpu_memlatency_stat) { + if (m_memory_config->SST_mode) { + // in SST mode, we just calculate mem latency + printf("max_mem_SST_latency = %d \n", max_mf_latency); + if (num_mfs) + printf("average_mf_SST_latency = %lld \n", mf_total_lat / num_mfs); + } else if (m_memory_config->gpgpu_memlatency_stat) { printf("maxmflatency = %d \n", max_mf_latency); printf("max_icnt2mem_latency = %d \n", max_icnt2mem_latency); printf("maxmrqlatency = %d \n", max_mrq_latency); diff --git a/src/gpgpu-sim/shader.cc b/src/gpgpu-sim/shader.cc index 4d4f112..7482e0e 100644 --- a/src/gpgpu-sim/shader.cc +++ b/src/gpgpu-sim/shader.cc @@ -162,7 +162,10 @@ void shader_core_ctx::create_front_pipeline() { } // m_icnt = new shader_memory_interface(this,cluster); - if (m_config->gpgpu_perfect_mem) { + if (m_memory_config->SST_mode) { + m_icnt = new sst_memory_interface( + this, static_cast(m_cluster)); + } else if (m_config->gpgpu_perfect_mem) { m_icnt = new perfect_memory_interface(this, m_cluster); } else { m_icnt = new shader_memory_interface(this, m_cluster); @@ -2281,7 +2284,15 @@ bool ldst_unit::memory_cycle(warp_inst_t &inst, inst.is_store() ? WRITE_PACKET_SIZE : READ_PACKET_SIZE; unsigned size = access.get_size() + control_size; // printf("Interconnect:Addr: %x, size=%d\n",access.get_addr(),size); - if (m_icnt->full(size, inst.is_store() || inst.isatomic())) { + if (m_memory_config->SST_mode && + (static_cast(m_icnt)->full( + size, inst.is_store() || inst.isatomic(), access.get_type()))) { + // SST need mf type here + // Cast it to sst_memory_interface pointer first as this full() method + // is not a virtual method in parent class + stall_cond = ICNT_RC_FAIL; + } else if (!m_memory_config->SST_mode && + (m_icnt->full(size, inst.is_store() || inst.isatomic()))) { stall_cond = ICNT_RC_FAIL; } else { mem_fetch *mf = @@ -2846,7 +2857,10 @@ void ldst_unit::cycle() { } } else { if (mf->get_type() == WRITE_ACK || - (m_config->gpgpu_perfect_mem && mf->get_is_write())) { + ((m_config->gpgpu_perfect_mem || m_memory_config->SST_mode) && + mf->get_is_write())) { + // SST memory is handled by SST mem hierarchy + // Perfect mem m_core->store_ack(mf); m_response_fifo.pop_front(); delete mf; @@ -4020,7 +4034,8 @@ void shader_core_ctx::accept_ldst_unit_response(mem_fetch *mf) { void shader_core_ctx::store_ack(class mem_fetch *mf) { assert(mf->get_type() == WRITE_ACK || - (m_config->gpgpu_perfect_mem && mf->get_is_write())); + ((m_config->gpgpu_perfect_mem || m_memory_config->SST_mode) && + mf->get_is_write())); unsigned warp_id = mf->get_wid(); m_warp[warp_id]->dec_store_req(); } @@ -4573,7 +4588,46 @@ bool simt_core_cluster::icnt_injection_buffer_full(unsigned size, bool write) { return !::icnt_has_buffer(m_cluster_id, request_size); } +bool sst_simt_core_cluster::SST_injection_buffer_full(unsigned size, bool write, + mem_access_type type) { + switch (type) { + case CONST_ACC_R: + case INST_ACC_R: { + return response_queue_full(); + break; + } + default: { + return ::is_SST_buffer_full(m_cluster_id); + break; + } + } +} + void simt_core_cluster::icnt_inject_request_packet(class mem_fetch *mf) { + // Update stats based on mf type + update_icnt_stats(mf); + + // The packet size varies depending on the type of request: + // - For write request and atomic request, the packet contains the data + // - For read request (i.e. not write nor atomic), the packet only has control + // metadata + unsigned int packet_size = mf->size(); + if (!mf->get_is_write() && !mf->isatomic()) { + packet_size = mf->get_ctrl_size(); + } + m_stats->m_outgoing_traffic_stats->record_traffic(mf, packet_size); + unsigned destination = mf->get_sub_partition_id(); + mf->set_status(IN_ICNT_TO_MEM, + m_gpu->gpu_sim_cycle + m_gpu->gpu_tot_sim_cycle); + if (!mf->get_is_write() && !mf->isatomic()) + ::icnt_push(m_cluster_id, m_config->mem2device(destination), (void *)mf, + mf->get_ctrl_size()); + else + ::icnt_push(m_cluster_id, m_config->mem2device(destination), (void *)mf, + mf->size()); +} + +void simt_core_cluster::update_icnt_stats(class mem_fetch *mf) { // stats if (mf->get_is_write()) m_stats->made_write_mfs++; @@ -4618,6 +4672,12 @@ void simt_core_cluster::icnt_inject_request_packet(class mem_fetch *mf) { default: assert(0); } +} + +void sst_simt_core_cluster::icnt_inject_request_packet_to_SST( + class mem_fetch *mf) { + // Update stats + update_icnt_stats(mf); // The packet size varies depending on the type of request: // - For write request and atomic request, the packet contains the data @@ -4628,15 +4688,25 @@ void simt_core_cluster::icnt_inject_request_packet(class mem_fetch *mf) { packet_size = mf->get_ctrl_size(); } m_stats->m_outgoing_traffic_stats->record_traffic(mf, packet_size); - unsigned destination = mf->get_sub_partition_id(); mf->set_status(IN_ICNT_TO_MEM, m_gpu->gpu_sim_cycle + m_gpu->gpu_tot_sim_cycle); - if (!mf->get_is_write() && !mf->isatomic()) - ::icnt_push(m_cluster_id, m_config->mem2device(destination), (void *)mf, - mf->get_ctrl_size()); - else - ::icnt_push(m_cluster_id, m_config->mem2device(destination), (void *)mf, - mf->size()); + switch (mf->get_access_type()) { + case CONST_ACC_R: + case INST_ACC_R: { + push_response_fifo(mf); + break; + } + default: { + if (!mf->get_is_write() && !mf->isatomic()) + ::send_read_request_SST(m_cluster_id, mf->get_addr(), + mf->get_data_size(), (void *)mf); + else + ::send_write_request_SST(m_cluster_id, mf->get_addr(), + mf->get_data_size(), (void *)mf); + + break; + } + } } void simt_core_cluster::icnt_cycle() { @@ -4678,6 +4748,49 @@ void simt_core_cluster::icnt_cycle() { } } +void sst_simt_core_cluster::icnt_cycle_SST() { + if (!m_response_fifo.empty()) { + mem_fetch *mf = m_response_fifo.front(); + unsigned cid = m_config->sid_to_cid(mf->get_sid()); + if (mf->get_access_type() == INST_ACC_R) { + // instruction fetch response + if (!m_core[cid]->fetch_unit_response_buffer_full()) { + m_response_fifo.pop_front(); + m_core[cid]->accept_fetch_response(mf); + } + } else { + // data response + if (!m_core[cid]->ldst_unit_response_buffer_full()) { + m_response_fifo.pop_front(); + m_memory_stats->memlatstat_read_done(mf); + m_core[cid]->accept_ldst_unit_response(mf); + } + } + } + + // pop from SST buffers + if (m_response_fifo.size() < m_config->n_simt_ejection_buffer_size) { + mem_fetch *mf = (mem_fetch *)(static_cast(get_gpu()) + ->SST_pop_mem_reply(m_cluster_id)); + if (!mf) return; + assert(mf->get_tpc() == m_cluster_id); + + // do atomic here + // For now, we execute atomic when the mem reply comes back + // This needs to be validated + if (mf && mf->isatomic()) mf->do_atomic(); + + unsigned int packet_size = + (mf->get_is_write()) ? mf->get_ctrl_size() : mf->size(); + m_stats->m_incoming_traffic_stats->record_traffic(mf, packet_size); + mf->set_status(IN_CLUSTER_TO_SHADER_QUEUE, + m_gpu->gpu_sim_cycle + m_gpu->gpu_tot_sim_cycle); + // m_memory_stats->memlatstat_read_done(mf,m_shader_config->max_warps_per_shader); + m_response_fifo.push_back(mf); + m_stats->n_mem_to_simt[m_cluster_id] += mf->get_num_flits(false); + } +} + void simt_core_cluster::get_pdom_stack_top_info(unsigned sid, unsigned tid, unsigned *pc, unsigned *rpc) const { diff --git a/src/gpgpu-sim/shader.h b/src/gpgpu-sim/shader.h index e658a14..ee10af6 100644 --- a/src/gpgpu-sim/shader.h +++ b/src/gpgpu-sim/shader.h @@ -2015,6 +2015,7 @@ class shader_core_stats : public shader_core_stats_pod { friend class shader_core_ctx; friend class ldst_unit; friend class simt_core_cluster; + friend class sst_simt_core_cluster; friend class scheduler_unit; friend class TwoLevelScheduler; friend class LooseRoundRobbinScheduler; @@ -2624,6 +2625,7 @@ class simt_core_cluster { void cache_invalidate(); bool icnt_injection_buffer_full(unsigned size, bool write); void icnt_inject_request_packet(class mem_fetch *mf); + void update_icnt_stats(class mem_fetch *mf); // for perfect memory interface bool response_queue_full() { @@ -2685,6 +2687,50 @@ class exec_simt_core_cluster : public simt_core_cluster { virtual void create_shader_core_ctx(); }; +/** + * @brief SST cluster class + * + */ +class sst_simt_core_cluster : public exec_simt_core_cluster { + public: + sst_simt_core_cluster(class gpgpu_sim *gpu, unsigned cluster_id, + const shader_core_config *config, + const memory_config *mem_config, + class shader_core_stats *stats, + class memory_stats_t *mstats) + : exec_simt_core_cluster(gpu, cluster_id, config, mem_config, stats, + mstats) {} + + /** + * @brief Check if SST memory request injection + * buffer is full by using extern + * function is_SST_buffer_full() + * defined in Balar + * + * @param size + * @param write + * @param type + * @return true + * @return false + */ + bool SST_injection_buffer_full(unsigned size, bool write, + mem_access_type type); + + /** + * @brief Send memory request packets to SST + * memory + * + * @param mf + */ + void icnt_inject_request_packet_to_SST(class mem_fetch *mf); + + /** + * @brief Advance ICNT between core and SST + * + */ + void icnt_cycle_SST(); +}; + class shader_memory_interface : public mem_fetch_interface { public: shader_memory_interface(shader_core_ctx *core, simt_core_cluster *cluster) { @@ -2725,6 +2771,61 @@ class perfect_memory_interface : public mem_fetch_interface { simt_core_cluster *m_cluster; }; +/** + * @brief SST memory interface + * + */ +class sst_memory_interface : public mem_fetch_interface { + public: + sst_memory_interface(shader_core_ctx *core, sst_simt_core_cluster *cluster) { + m_core = core; + m_cluster = cluster; + } + /** + * @brief For constant, inst, tex cache access + * + * @param size + * @param write + * @return true + * @return false + */ + virtual bool full(unsigned size, bool write) const { + assert(false && "Use the full() method with access type instead!"); + return true; + } + + /** + * @brief With SST, the core will direct all mem access except for + * constant, tex, and inst reads to SST mem system + * (i.e. not modeling constant mem right now), thus + * requiring the mem_access_type information to be passed in + * + * @param size + * @param write + * @param type + * @return true + * @return false + */ + bool full(unsigned size, bool write, mem_access_type type) const { + return m_cluster->SST_injection_buffer_full(size, write, type); + } + + /** + * @brief Push memory request to SST memory system and + * update stats + * + * @param mf + */ + virtual void push(mem_fetch *mf) { + m_core->inc_simt_to_mem(mf->get_num_flits(true)); + m_cluster->icnt_inject_request_packet_to_SST(mf); + } + + private: + shader_core_ctx *m_core; + sst_simt_core_cluster *m_cluster; +}; + inline int scheduler_unit::get_sid() const { return m_shader->get_sid(); } #endif /* SHADER_H */ diff --git a/src/gpgpusim_entrypoint.cc b/src/gpgpusim_entrypoint.cc index 42c6981..839fef6 100644 --- a/src/gpgpusim_entrypoint.cc +++ b/src/gpgpusim_entrypoint.cc @@ -43,6 +43,20 @@ static int sg_argc = 3; static const char *sg_argv[] = {"", "-config", "gpgpusim.config"}; +// Help funcs to avoid multiple '->' for SST +GPGPUsim_ctx *GPGPUsim_ctx_ptr() { return GPGPU_Context()->the_gpgpusim; } + +class sst_gpgpu_sim *g_the_gpu() { + return static_cast(GPGPUsim_ctx_ptr()->g_the_gpu); +} + +class stream_manager *g_stream_manager() { + return GPGPUsim_ctx_ptr()->g_stream_manager; +} + +// SST callback +extern void SST_callback_cudaThreadSynchronize_done(); + void *gpgpu_sim_thread_sequential(void *ctx_ptr) { gpgpu_context *ctx = (gpgpu_context *)ctx_ptr; // at most one kernel running at a time @@ -169,6 +183,75 @@ void *gpgpu_sim_thread_concurrent(void *ctx_ptr) { return NULL; } +bool sst_sim_cycles = false; + +bool SST_Cycle() { + // Check if Synchronize is done when SST previously requested + // cudaThreadSynchronize + if (GPGPU_Context()->requested_synchronize && + ((g_stream_manager()->empty() && !GPGPUsim_ctx_ptr()->g_sim_active) || + GPGPUsim_ctx_ptr()->g_sim_done)) { + SST_callback_cudaThreadSynchronize_done(); + GPGPU_Context()->requested_synchronize = false; + } + + if (g_stream_manager()->empty_protected() && + !GPGPUsim_ctx_ptr()->g_sim_done && !g_the_gpu()->active()) { + GPGPUsim_ctx_ptr()->g_sim_active = false; + // printf("stream is empty %d \n", g_stream_manager->empty()); + return false; + } + + if (g_stream_manager()->operation(&sst_sim_cycles) && + !g_the_gpu()->active()) { + if (sst_sim_cycles) { + sst_sim_cycles = false; + } + return false; + } + + // printf("GPGPU-Sim: Give GPU Cycle\n"); + GPGPUsim_ctx_ptr()->g_sim_active = true; + + // functional simulation + if (g_the_gpu()->is_functional_sim()) { + kernel_info_t *kernel = g_the_gpu()->get_functional_kernel(); + assert(kernel); + GPGPUsim_ctx_ptr()->gpgpu_ctx->func_sim->gpgpu_cuda_ptx_sim_main_func( + *kernel); + g_the_gpu()->finish_functional_sim(kernel); + } + + // performance simulation + if (g_the_gpu()->active()) { + g_the_gpu()->SST_cycle(); + sst_sim_cycles = true; + g_the_gpu()->deadlock_check(); + } else { + if (g_the_gpu()->cycle_insn_cta_max_hit()) { + g_stream_manager()->stop_all_running_kernels(); + GPGPUsim_ctx_ptr()->g_sim_done = true; + GPGPUsim_ctx_ptr()->g_sim_active = false; + GPGPUsim_ctx_ptr()->break_limit = true; + } + } + + if (!g_the_gpu()->active()) { + g_the_gpu()->print_stats(GPGPUsim_ctx_ptr()->g_the_gpu->last_streamID); + g_the_gpu()->update_stats(); + GPGPU_Context()->print_simulation_time(); + } + + if (GPGPUsim_ctx_ptr()->break_limit) { + printf( + "GPGPU-Sim: ** break due to reaching the maximum cycles (or " + "instructions) **\n"); + return true; + } + + return false; +} + void gpgpu_context::synchronize() { printf("GPGPU-Sim: synchronize waiting for inactive GPU simulation\n"); the_gpgpusim->g_stream_manager->print(stdout); @@ -187,6 +270,27 @@ void gpgpu_context::synchronize() { // sem_post(&g_sim_signal_start); } +bool gpgpu_context::synchronize_check() { + // printf("GPGPU-Sim: synchronize checking for inactive GPU simulation\n"); + requested_synchronize = true; + the_gpgpusim->g_stream_manager->print(stdout); + fflush(stdout); + // sem_wait(&g_sim_signal_finish); + bool done = false; + pthread_mutex_lock(&(the_gpgpusim->g_sim_lock)); + done = (the_gpgpusim->g_stream_manager->empty() && + !the_gpgpusim->g_sim_active) || + the_gpgpusim->g_sim_done; + pthread_mutex_unlock(&(the_gpgpusim->g_sim_lock)); + if (done) { + printf( + "GPGPU-Sim: synchronize checking: detected inactive GPU simulation " + "thread\n"); + } + fflush(stdout); + return done; +} + void gpgpu_context::exit_simulation() { the_gpgpusim->g_sim_done = true; printf("GPGPU-Sim: exit_simulation called\n"); @@ -220,8 +324,14 @@ gpgpu_sim *gpgpu_context::gpgpu_ptx_sim_init_perf() { assert(setlocale(LC_NUMERIC, "C")); the_gpgpusim->g_the_gpu_config->init(); - the_gpgpusim->g_the_gpu = - new exec_gpgpu_sim(*(the_gpgpusim->g_the_gpu_config), this); + if (the_gpgpusim->g_the_gpu_config->is_SST_mode()) { + // Create SST specific GPGPUSim + the_gpgpusim->g_the_gpu = + new sst_gpgpu_sim(*(the_gpgpusim->g_the_gpu_config), this); + } else { + the_gpgpusim->g_the_gpu = + new exec_gpgpu_sim(*(the_gpgpusim->g_the_gpu_config), this); + } the_gpgpusim->g_stream_manager = new stream_manager( (the_gpgpusim->g_the_gpu), func_sim->g_cuda_launch_blocking); @@ -237,12 +347,17 @@ gpgpu_sim *gpgpu_context::gpgpu_ptx_sim_init_perf() { void gpgpu_context::start_sim_thread(int api) { if (the_gpgpusim->g_sim_done) { the_gpgpusim->g_sim_done = false; - if (api == 1) { - pthread_create(&(the_gpgpusim->g_simulation_thread), NULL, - gpgpu_sim_thread_concurrent, (void *)this); + if (the_gpgpusim->g_the_gpu_config->is_SST_mode()) { + // Do not create concurrent thread in SST mode + g_the_gpu()->init(); } else { - pthread_create(&(the_gpgpusim->g_simulation_thread), NULL, - gpgpu_sim_thread_sequential, (void *)this); + if (api == 1) { + pthread_create(&(the_gpgpusim->g_simulation_thread), NULL, + gpgpu_sim_thread_concurrent, (void *)this); + } else { + pthread_create(&(the_gpgpusim->g_simulation_thread), NULL, + gpgpu_sim_thread_sequential, (void *)this); + } } } } @@ -266,8 +381,13 @@ void gpgpu_context::print_simulation_time() { const unsigned cycles_per_sec = (unsigned)(the_gpgpusim->g_the_gpu->gpu_tot_sim_cycle / difference); printf("gpgpu_simulation_rate = %u (cycle/sec)\n", cycles_per_sec); - printf("gpgpu_silicon_slowdown = %ux\n", - the_gpgpusim->g_the_gpu->shader_clock() * 1000 / cycles_per_sec); + + if (cycles_per_sec == 0) { + printf("gpgpu_silicon_slowdown = Nan\n"); + } else { + printf("gpgpu_silicon_slowdown = %ux\n", + the_gpgpusim->g_the_gpu->shader_clock() * 1000 / cycles_per_sec); + } fflush(stdout); } diff --git a/src/stream_manager.cc b/src/stream_manager.cc index 72f8bb0..b974791 100644 --- a/src/stream_manager.cc +++ b/src/stream_manager.cc @@ -34,6 +34,12 @@ unsigned CUstream_st::sm_next_stream_uid = 0; +// SST memcpy callbacks +extern void SST_callback_memcpy_H2D_done(); +extern void SST_callback_memcpy_D2H_done(); +extern void SST_callback_memcpy_to_symbol_done(); +extern void SST_callback_memcpy_from_symbol_done(); + CUstream_st::CUstream_st() { m_pending = false; m_uid = sm_next_stream_uid++; @@ -122,11 +128,13 @@ bool stream_operation::do_operation(gpgpu_sim *gpu) { if (g_debug_execution >= 3) printf("memcpy host-to-device\n"); gpu->memcpy_to_gpu(m_device_address_dst, m_host_address_src, m_cnt); m_stream->record_next_done(); + if (gpu->is_SST_mode()) SST_callback_memcpy_H2D_done(); break; case stream_memcpy_device_to_host: if (g_debug_execution >= 3) printf("memcpy device-to-host\n"); gpu->memcpy_from_gpu(m_host_address_dst, m_device_address_src, m_cnt); m_stream->record_next_done(); + if (gpu->is_SST_mode()) SST_callback_memcpy_D2H_done(); break; case stream_memcpy_device_to_device: if (g_debug_execution >= 3) printf("memcpy device-to-device\n"); @@ -138,12 +146,14 @@ bool stream_operation::do_operation(gpgpu_sim *gpu) { gpu->gpgpu_ctx->func_sim->gpgpu_ptx_sim_memcpy_symbol( m_symbol, m_host_address_src, m_cnt, m_offset, 1, gpu); m_stream->record_next_done(); + if (gpu->is_SST_mode()) SST_callback_memcpy_to_symbol_done(); break; case stream_memcpy_from_symbol: if (g_debug_execution >= 3) printf("memcpy from symbol\n"); gpu->gpgpu_ctx->func_sim->gpgpu_ptx_sim_memcpy_symbol( m_symbol, m_host_address_dst, m_cnt, m_offset, 0, gpu); m_stream->record_next_done(); + if (gpu->is_SST_mode()) SST_callback_memcpy_from_symbol_done(); break; case stream_kernel_launch: if (m_sim_mode) { // Functional Sim @@ -472,7 +482,7 @@ void stream_manager::push(stream_operation op) { } if (g_debug_execution >= 3) print_impl(stdout); pthread_mutex_unlock(&m_lock); - if (m_cuda_launch_blocking || stream == NULL) { + if (!m_gpu->is_SST_mode() && (m_cuda_launch_blocking || stream == NULL)) { unsigned int wait_amount = 100; unsigned int wait_cap = 100000; // 100ms while (!empty()) { -- cgit v1.3 From 3844f7559e69f206a80ae9b970b19e7e609e949e Mon Sep 17 00:00:00 2001 From: WilliamMTK Date: Tue, 14 Jan 2025 12:33:11 -0500 Subject: fix_sst_callbacks: add weak definitions for sst callbacks (#81) * fix_sst_callbacks: add weak definitions for sst callbacks * Automated Format --------- Co-authored-by: purdue-jenkins --- src/gpgpu-sim/gpu-sim.h | 10 +++++++++- src/gpgpusim_entrypoint.cc | 1 + src/stream_manager.cc | 4 ++++ 3 files changed, 14 insertions(+), 1 deletion(-) (limited to 'src/gpgpu-sim/gpu-sim.h') diff --git a/src/gpgpu-sim/gpu-sim.h b/src/gpgpu-sim/gpu-sim.h index d0c2a17..5b253ab 100644 --- a/src/gpgpu-sim/gpu-sim.h +++ b/src/gpgpu-sim/gpu-sim.h @@ -78,6 +78,9 @@ extern tr1_hash_map address_random_interleaving; * @return false */ extern bool is_SST_buffer_full(unsigned core_id); +__attribute__((weak)) bool is_SST_buffer_full(unsigned core_id) { + return false; +} /** * @brief Send loads to SST memory backend @@ -89,7 +92,9 @@ extern bool is_SST_buffer_full(unsigned core_id); */ extern void send_read_request_SST(unsigned core_id, uint64_t address, size_t size, void *mem_req); - +__attribute__((weak)) void send_read_request_SST(unsigned core_id, + uint64_t address, size_t size, + void *mem_req) {} /** * @brief Send stores to SST memory backend * @@ -100,6 +105,9 @@ extern void send_read_request_SST(unsigned core_id, uint64_t address, */ extern void send_write_request_SST(unsigned core_id, uint64_t address, size_t size, void *mem_req); +__attribute__((weak)) void send_write_request_SST(unsigned core_id, + uint64_t address, size_t size, + void *mem_req) {} enum dram_ctrl_t { DRAM_FIFO = 0, DRAM_FRFCFS = 1 }; diff --git a/src/gpgpusim_entrypoint.cc b/src/gpgpusim_entrypoint.cc index 839fef6..e2b711e 100644 --- a/src/gpgpusim_entrypoint.cc +++ b/src/gpgpusim_entrypoint.cc @@ -56,6 +56,7 @@ class stream_manager *g_stream_manager() { // SST callback extern void SST_callback_cudaThreadSynchronize_done(); +__attribute__((weak)) void SST_callback_cudaThreadSynchronize_done() {} void *gpgpu_sim_thread_sequential(void *ctx_ptr) { gpgpu_context *ctx = (gpgpu_context *)ctx_ptr; diff --git a/src/stream_manager.cc b/src/stream_manager.cc index b974791..58c2ec4 100644 --- a/src/stream_manager.cc +++ b/src/stream_manager.cc @@ -39,6 +39,10 @@ extern void SST_callback_memcpy_H2D_done(); extern void SST_callback_memcpy_D2H_done(); extern void SST_callback_memcpy_to_symbol_done(); extern void SST_callback_memcpy_from_symbol_done(); +__attribute__((weak)) void SST_callback_memcpy_H2D_done() {} +__attribute__((weak)) void SST_callback_memcpy_D2H_done() {} +__attribute__((weak)) void SST_callback_memcpy_to_symbol_done() {} +__attribute__((weak)) void SST_callback_memcpy_from_symbol_done() {} CUstream_st::CUstream_st() { m_pending = false; -- cgit v1.3