summaryrefslogtreecommitdiff
path: root/src/trace-driven
diff options
context:
space:
mode:
Diffstat (limited to 'src/trace-driven')
-rw-r--r--src/trace-driven/gpgpusim_trace_driven_main.cc3
-rw-r--r--src/trace-driven/kepler_opcode.h5
-rw-r--r--src/trace-driven/trace_driven.cc102
-rw-r--r--src/trace-driven/trace_driven.h28
4 files changed, 63 insertions, 75 deletions
diff --git a/src/trace-driven/gpgpusim_trace_driven_main.cc b/src/trace-driven/gpgpusim_trace_driven_main.cc
index bedac4c..5e07ace 100644
--- a/src/trace-driven/gpgpusim_trace_driven_main.cc
+++ b/src/trace-driven/gpgpusim_trace_driven_main.cc
@@ -48,6 +48,7 @@ int main ( int argc, const char **argv )
//prints stats
trace_parser tracer(m_gpgpu_sim->get_config().get_traces_filename(), m_gpgpu_sim, m_gpgpu_context);
+ trace_config config(m_gpgpu_sim);
std::vector<std::string> commandlist = tracer.parse_kernellist_file();
bool first_kernel=true;
@@ -69,7 +70,7 @@ int main ( int argc, const char **argv )
first_kernel = false;
continue;
}
- kernel_info = tracer.parse_kernel_info(commandlist[i]);
+ kernel_info = tracer.parse_kernel_info(commandlist[i], &config);
m_gpgpu_sim->launch(kernel_info);
}
diff --git a/src/trace-driven/kepler_opcode.h b/src/trace-driven/kepler_opcode.h
index f2bbc90..4aa8e0f 100644
--- a/src/trace-driven/kepler_opcode.h
+++ b/src/trace-driven/kepler_opcode.h
@@ -99,12 +99,13 @@ static const std::unordered_map<std::string,OpcodeChar> Kepler_OpcodeMap = {
//Load/Store Instructions
//For now, we ignore constant loads, consider it as ALU_OP, TO DO
{"LDC", OpcodeChar(OP_LDC, ALU_OP)},
- {"LD", OpcodeChar(OP_LD, LOAD_OP)},
+ //in Kepler, LD is load global so set it to LDG
+ {"LD", OpcodeChar(OP_LDG, LOAD_OP)},
{"LDG", OpcodeChar(OP_LDG, LOAD_OP)},
{"LDL", OpcodeChar(OP_LDL, LOAD_OP)},
{"LDS", OpcodeChar(OP_LDS, LOAD_OP)},
{"LDSLK", OpcodeChar(OP_LDSLK, LOAD_OP)},
- {"ST", OpcodeChar(OP_ST, STORE_OP)},
+ {"ST", OpcodeChar(OP_STG, STORE_OP)},
{"STL", OpcodeChar(OP_STL, STORE_OP)},
{"STS", OpcodeChar(OP_STS, STORE_OP)},
{"STSCUL", OpcodeChar(OP_STSCUL, STORE_OP)},
diff --git a/src/trace-driven/trace_driven.cc b/src/trace-driven/trace_driven.cc
index 35e953e..22c527e 100644
--- a/src/trace-driven/trace_driven.cc
+++ b/src/trace-driven/trace_driven.cc
@@ -98,7 +98,7 @@ void trace_parser::parse_memcpy_info(const std::string& memcpy_command, size_t&
ss>>std::dec>>count;
}
-trace_kernel_info_t* trace_parser::parse_kernel_info(const std::string& kerneltraces_filepath) {
+trace_kernel_info_t* trace_parser::parse_kernel_info(const std::string& kerneltraces_filepath, trace_config* config) {
ifs.open(kerneltraces_filepath.c_str());
@@ -167,7 +167,7 @@ trace_kernel_info_t* trace_parser::parse_kernel_info(const std::string& kerneltr
dim3 blockDim(tb_dim_x, tb_dim_y, tb_dim_z);
trace_function_info* function_info = new trace_function_info(info, m_gpgpu_context);
function_info->set_name(kernel_name.c_str());
- trace_kernel_info_t* kernel_info = new trace_kernel_info_t(gridDim, blockDim, binary_verion, function_info, &ifs, m_gpgpu_sim, m_gpgpu_context);
+ trace_kernel_info_t* kernel_info = new trace_kernel_info_t(gridDim, blockDim, binary_verion, function_info, &ifs, m_gpgpu_sim, m_gpgpu_context, config);
return kernel_info;
}
@@ -211,11 +211,12 @@ address_type trace_shd_warp_t::get_pc(){
return warp_traces[trace_pc].pc;
}
-trace_kernel_info_t::trace_kernel_info_t(dim3 gridDim, dim3 blockDim, unsigned m_binary_verion, trace_function_info* m_function_info, std::ifstream* inputstream, gpgpu_sim * gpgpu_sim, gpgpu_context* gpgpu_context):kernel_info_t(gridDim, blockDim, m_function_info) {
+trace_kernel_info_t::trace_kernel_info_t(dim3 gridDim, dim3 blockDim, unsigned m_binary_verion, trace_function_info* m_function_info, std::ifstream* inputstream, gpgpu_sim * gpgpu_sim, gpgpu_context* gpgpu_context, class trace_config* config):kernel_info_t(gridDim, blockDim, m_function_info) {
ifs = inputstream;
m_gpgpu_sim = gpgpu_sim;
m_gpgpu_context = gpgpu_context;
binary_verion = m_binary_verion;
+ m_tconfig = config;
//resolve the binary version
if(m_binary_verion == VOLTA_BINART_VERSION)
@@ -285,8 +286,8 @@ bool trace_kernel_info_t::get_next_threadblock_traces(std::vector<std::vector<tr
}
else {
assert(start_of_tb_stream_found);
- trace_warp_inst_t inst(m_gpgpu_sim->getShaderCoreConfig(), m_gpgpu_context);
- inst.parse_from_string(line, OpcodeMap, binary_verion);
+ trace_warp_inst_t inst(m_gpgpu_sim->getShaderCoreConfig(), m_gpgpu_context, m_tconfig);
+ inst.parse_from_string(line, OpcodeMap);
threadblock_traces[warp_id]->push_back(inst);
}
}
@@ -323,7 +324,7 @@ unsigned trace_warp_inst_t::get_datawidth_from_opcode(const std::vector<std::str
return 4; //default is 4 bytes
}
-bool trace_warp_inst_t::parse_from_string(std::string trace, const std::unordered_map<std::string,OpcodeChar>* OpcodeMap, unsigned binary_verion){
+bool trace_warp_inst_t::parse_from_string(std::string trace, const std::unordered_map<std::string,OpcodeChar>* OpcodeMap){
std::stringstream ss;
ss.str(trace);
@@ -473,7 +474,7 @@ bool trace_warp_inst_t::parse_from_string(std::string trace, const std::unordere
//remove redundant registers
//fill latency and initl
- set_latency(op);
+ m_tconfig->set_latency(op, latency, initiation_interval);
//fill addresses
if(mem_width > 0) {
@@ -548,10 +549,7 @@ bool trace_warp_inst_t::parse_from_string(std::string trace, const std::unordere
//right now, we consider all loads are shared.
assert(mem_width>0);
data_size = get_datawidth_from_opcode(opcode_tokens);
- if(binary_verion == KEPLER_BINART_VERSION)
- space.set_type(global_space);
- else
- space.set_type(shared_space);
+ space.set_type(shared_space);
if(m_opcode == OP_LD)
memory_op = memory_load;
else
@@ -584,59 +582,29 @@ bool trace_warp_inst_t::parse_from_string(std::string trace, const std::unordere
return true;
}
-void trace_warp_inst_t::set_latency(unsigned category)
+trace_config::trace_config(gpgpu_sim* m_gpgpu_sim){
+
+ this->m_gpgpu_sim=m_gpgpu_sim;
+ parse_config();
+}
+
+void trace_config::parse_config()
{
- unsigned int_latency[5];
- unsigned fp_latency[5];
- unsigned dp_latency[5];
- unsigned sfu_latency;
- unsigned tensor_latency;
- unsigned int_init[5];
- unsigned fp_init[5];
- unsigned dp_init[5];
- unsigned sfu_init;
- unsigned tensor_init;
- /*
- * [0] ADD,SUB
- * [1] MAX,Min
- * [2] MUL
- * [3] MAD
- * [4] DIV
- */
- sscanf(m_gpgpu_context->func_sim->opcode_latency_int, "%u,%u,%u,%u,%u",
- &int_latency[0],&int_latency[1],&int_latency[2],
- &int_latency[3],&int_latency[4]);
- sscanf(m_gpgpu_context->func_sim->opcode_latency_fp, "%u,%u,%u,%u,%u",
- &fp_latency[0],&fp_latency[1],&fp_latency[2],
- &fp_latency[3],&fp_latency[4]);
- sscanf(m_gpgpu_context->func_sim->opcode_latency_dp, "%u,%u,%u,%u,%u",
- &dp_latency[0],&dp_latency[1],&dp_latency[2],
- &dp_latency[3],&dp_latency[4]);
- sscanf(m_gpgpu_context->func_sim->opcode_latency_sfu, "%u",
- &sfu_latency);
- sscanf(m_gpgpu_context->func_sim->opcode_latency_tensor, "%u",
- &tensor_latency);
- sscanf(m_gpgpu_context->func_sim->opcode_initiation_int, "%u,%u,%u,%u,%u",
- &int_init[0],&int_init[1],&int_init[2],
- &int_init[3],&int_init[4]);
- sscanf(m_gpgpu_context->func_sim->opcode_initiation_fp, "%u,%u,%u,%u,%u",
- &fp_init[0],&fp_init[1],&fp_init[2],
- &fp_init[3],&fp_init[4]);
- sscanf(m_gpgpu_context->func_sim->opcode_initiation_dp, "%u,%u,%u,%u,%u",
- &dp_init[0],&dp_init[1],&dp_init[2],
- &dp_init[3],&dp_init[4]);
- sscanf(m_gpgpu_context->func_sim->opcode_initiation_sfu, "%u",
- &sfu_init);
- sscanf(m_gpgpu_context->func_sim->opcode_initiation_tensor, "%u",
- &tensor_init);
- sscanf(m_gpgpu_context->func_sim->cdp_latency_str, "%u,%u,%u,%u,%u",
- &m_gpgpu_context->func_sim->cdp_latency[0],
- &m_gpgpu_context->func_sim->cdp_latency[1],
- &m_gpgpu_context->func_sim->cdp_latency[2],
- &m_gpgpu_context->func_sim->cdp_latency[3],
- &m_gpgpu_context->func_sim->cdp_latency[4]);
+ sscanf(m_gpgpu_sim->getShaderCoreConfig()->trace_opcode_latency_initiation_int, "%u,%u",
+ &int_latency,&int_init);
+ sscanf(m_gpgpu_sim->getShaderCoreConfig()->trace_opcode_latency_initiation_sp, "%u,%u",
+ &fp_latency,&fp_init);
+ sscanf(m_gpgpu_sim->getShaderCoreConfig()->trace_opcode_latency_initiation_dp, "%u,%u",
+ &dp_latency,&dp_init);
+ sscanf(m_gpgpu_sim->getShaderCoreConfig()->trace_opcode_latency_initiation_sfu, "%u,%u",
+ &sfu_latency,&sfu_init);
+ sscanf(m_gpgpu_sim->getShaderCoreConfig()->trace_opcode_latency_initiation_tensor, "%u,%u",
+ &tensor_latency,&tensor_init);
+}
+void trace_config::set_latency(unsigned category, unsigned& latency, unsigned& initiation_interval)
+{
initiation_interval = latency = 1;
switch(category){
@@ -645,16 +613,16 @@ void trace_warp_inst_t::set_latency(unsigned category)
case BRANCH_OP:
case CALL_OPS:
case RET_OPS:
- latency = int_latency[0];
- initiation_interval = int_init[0];
+ latency = int_latency;
+ initiation_interval = int_init;
break;
case SP_OP:
- latency = fp_latency[0];
- initiation_interval = fp_init[0];
+ latency = fp_latency;
+ initiation_interval = fp_init;
break;
case DP_OP:
- latency = dp_latency[0];
- initiation_interval = dp_init[0];
+ latency = dp_latency;
+ initiation_interval = dp_init;
break;
case SFU_OP:
latency = sfu_latency;
diff --git a/src/trace-driven/trace_driven.h b/src/trace-driven/trace_driven.h
index 9539e6d..2888f86 100644
--- a/src/trace-driven/trace_driven.h
+++ b/src/trace-driven/trace_driven.h
@@ -37,18 +37,20 @@ public:
trace_warp_inst_t() {
m_gpgpu_context=NULL;
m_opcode=0;
+ m_tconfig=NULL;
}
- trace_warp_inst_t(const class core_config *config, gpgpu_context* gpgpu_context ):warp_inst_t(config) {
+ trace_warp_inst_t(const class core_config *config, gpgpu_context* gpgpu_context, class trace_config* tconfig ):warp_inst_t(config) {
m_gpgpu_context = gpgpu_context;
m_opcode=0;
+ m_tconfig=tconfig;
}
- bool parse_from_string(std::string trace, const std::unordered_map<std::string,OpcodeChar>* OpcodeMap, unsigned binary_verion);
+ bool parse_from_string(std::string trace, const std::unordered_map<std::string,OpcodeChar>* OpcodeMap);
private:
- void set_latency(unsigned cat);
gpgpu_context* m_gpgpu_context;
+ class trace_config* m_tconfig;
unsigned m_opcode;
bool check_opcode_contain(const std::vector<std::string>& opcode, std::string param);
unsigned get_datawidth_from_opcode(const std::vector<std::string>& opcode);
@@ -56,7 +58,7 @@ private:
class trace_kernel_info_t: public kernel_info_t {
public:
- trace_kernel_info_t(dim3 gridDim, dim3 blockDim, unsigned m_binary_verion, trace_function_info* m_function_info, std::ifstream* inputstream, gpgpu_sim * gpgpu_sim, gpgpu_context* gpgpu_context);
+ trace_kernel_info_t(dim3 gridDim, dim3 blockDim, unsigned m_binary_verion, trace_function_info* m_function_info, std::ifstream* inputstream, gpgpu_sim * gpgpu_sim, gpgpu_context* gpgpu_context, class trace_config* config);
bool get_next_threadblock_traces(std::vector<std::vector<trace_warp_inst_t>*> threadblock_traces);
@@ -64,19 +66,35 @@ private:
std::ifstream* ifs;
gpgpu_sim * m_gpgpu_sim;
gpgpu_context* m_gpgpu_context;
+ trace_config* m_tconfig;
unsigned binary_verion;
const std::unordered_map<std::string,OpcodeChar>* OpcodeMap;
};
+class trace_config {
+public:
+ trace_config(gpgpu_sim * m_gpgpu_sim);
+
+ void set_latency(unsigned category, unsigned& latency, unsigned& initiation_interval);
+ void parse_config();
+
+
+private:
+
+ unsigned int_latency, fp_latency, dp_latency, sfu_latency, tensor_latency;
+ unsigned int_init, fp_init, dp_init, sfu_init, tensor_init;
+ gpgpu_sim* m_gpgpu_sim;
+
+};
class trace_parser {
public:
trace_parser(const char* kernellist_filepath, gpgpu_sim * m_gpgpu_sim, gpgpu_context* m_gpgpu_context);
std::vector<std::string> parse_kernellist_file();
- trace_kernel_info_t* parse_kernel_info(const std::string& kerneltraces_filepath);
+ trace_kernel_info_t* parse_kernel_info(const std::string& kerneltraces_filepath, trace_config* config);
void parse_memcpy_info(const std::string& memcpy_command, size_t& add, size_t& count);
void kernel_finalizer(trace_kernel_info_t* kernel_info);
generated by cgit v1.3 (git 2.53.0) at 2026-07-11 02:57:15 +0000