Merged in suchitapati/gpgpusim-cudnn-05-01-18 (pull request #7)

Support for -gpgpu_registers_per_block and addition of -save_embedded_ptx 1 to config files for mnistCUDNN

Approved-by: tgrogers-purdue <[email protected]>

10 files changed, 36 insertions, 19 deletions

diff --git a/configs/GTX480/gpgpusim.config b/configs/GTX480/gpgpusim.config
index 436cb41..580929f 100644
--- a/configs/GTX480/gpgpusim.config
+++ b/configs/GTX480/gpgpusim.config
@@ -1,3 +1,6 @@
+#Save the cuobjdump dump
+#-save_embedded_ptx 1
+
 # functional simulator specification
 -gpgpu_ptx_instruction_classification 0
 -gpgpu_ptx_sim_mode 0
diff --git a/configs/GeForceGTX1080Ti/gpgpusim.config b/configs/GeForceGTX1080Ti/gpgpusim.config
index a5ada34..6d8b43a 100644
--- a/configs/GeForceGTX1080Ti/gpgpusim.config
+++ b/configs/GeForceGTX1080Ti/gpgpusim.config
@@ -1,5 +1,8 @@
 # This config models the Pascal GP102 (GeForceGTX 1080Ti)
 
+#Save the cuobjdump dump
+#-save_embedded_ptx 1
+
 # functional simulator specification
 -gpgpu_ptx_instruction_classification 0
 -gpgpu_ptx_sim_mode 0
diff --git a/configs/GeForceGTX750Ti/gpgpusim.config b/configs/GeForceGTX750Ti/gpgpusim.config
index 8b030b6..b6870b0 100644
--- a/configs/GeForceGTX750Ti/gpgpusim.config
+++ b/configs/GeForceGTX750Ti/gpgpusim.config
@@ -1,3 +1,6 @@
+#Save the cuobjdump dump
+#-save_embedded_ptx 1
+
 # functional simulator specification
 -gpgpu_ptx_instruction_classification 0
 -gpgpu_ptx_sim_mode 0
diff --git a/configs/QuadroFX5600/gpgpusim.config b/configs/QuadroFX5600/gpgpusim.config
index cb87b65..6f6e032 100644
--- a/configs/QuadroFX5600/gpgpusim.config
+++ b/configs/QuadroFX5600/gpgpusim.config
@@ -1,3 +1,6 @@
+#Save the cuobjdump dump
+#-save_embedded_ptx 1
+
 # functional simulator specification
 -gpgpu_ptx_instruction_classification 0
 -gpgpu_ptx_sim_mode 0
diff --git a/configs/QuadroFX5800/gpgpusim.config b/configs/QuadroFX5800/gpgpusim.config
index 82243c2..77d9c6d 100644
--- a/configs/QuadroFX5800/gpgpusim.config
+++ b/configs/QuadroFX5800/gpgpusim.config
@@ -1,3 +1,6 @@
+#Save the cuobjdump dump
+#-save_embedded_ptx 1
+
 # functional simulator specification
 -gpgpu_ptx_instruction_classification 0
 -gpgpu_ptx_sim_mode 0
diff --git a/configs/TeslaC2050/gpgpusim.config b/configs/TeslaC2050/gpgpusim.config
index 442ab8b..480aaa4 100644
--- a/configs/TeslaC2050/gpgpusim.config
+++ b/configs/TeslaC2050/gpgpusim.config
@@ -1,3 +1,6 @@
+#Save the cuobjdump dump
+#-save_embedded_ptx 1
+
 # functional simulator specification
 -gpgpu_ptx_instruction_classification 0
 -gpgpu_ptx_sim_mode 0
diff --git a/libcuda/cuda_runtime_api.cc b/libcuda/cuda_runtime_api.cc
index ef46f00..e3c2542 100644
--- a/libcuda/cuda_runtime_api.cc
+++ b/libcuda/cuda_runtime_api.cc
@@ -349,24 +349,13 @@ class _cuda_device_id *GPGPUSim_Init()
 		prop->maxGridSize[2] = 0x40000000;
 		prop->totalConstMem = 0x40000000;
 		prop->textureAlignment = 0;
-       /*
-
-        * TODO: Update the .config and xml files of all GPU config files with new value of sharedMemPerBlock.
-
-        * Previously, this was thought as sharedMemPerMultiprocessor and is being used in many places.
-
-        * Check whether all the instances of shared_mem_size(), gpgpu_shmem_size or sharedMemPerBlock are meant to use sharedMemPerBlock or sharedMemPerMultiprocessor.
-
-       */
-
-	       prop->sharedMemPerBlock = the_gpu->shared_mem_per_block();
-
-		#if (CUDART_VERSION > 5000)
-
-	       prop->sharedMemPerMultiprocessor = the_gpu->shared_mem_size();
-
-		#endif	
-		
+//        * TODO: Update the .config and xml files of all GPU config files with new value of sharedMemPerBlock and regsPerBlock 
+	        prop->sharedMemPerBlock = the_gpu->shared_mem_per_block();
+#if (CUDART_VERSION > 5000)
+		prop->regsPerMultiprocessor = the_gpu->num_registers_per_core();
+  	        prop->sharedMemPerMultiprocessor = the_gpu->shared_mem_size();
+#endif	
+		prop->sharedMemPerBlock = the_gpu->shared_mem_per_block();
 		prop->regsPerBlock = the_gpu->num_registers_per_core();
 		prop->warpSize = the_gpu->wrp_size();
 		prop->clockRate = the_gpu->shader_clock();
@@ -912,7 +901,7 @@ __host__ cudaError_t CUDARTAPI cudaDeviceGetAttribute(int *value, enum cudaDevic
                         *value= prop->sharedMemPerMultiprocessor;
                         break;
                 case 82:
-                        *value= prop->regsPerBlock;
+                        *value= prop->regsPerMultiprocessor;
                         break;
 		default:
 			printf("ERROR: Attribute number %d unimplemented \n",attr);
diff --git a/src/gpgpu-sim/gpu-sim.cc b/src/gpgpu-sim/gpu-sim.cc
index 892b245..d0b362f 100644
--- a/src/gpgpu-sim/gpu-sim.cc
+++ b/src/gpgpu-sim/gpu-sim.cc
@@ -256,6 +256,9 @@ void shader_core_config::reg_options(class OptionParser * opp)
     option_parser_register(opp, "-gpgpu_shader_registers", OPT_UINT32, &gpgpu_shader_registers, 
                  "Number of registers per shader core. Limits number of concurrent CTAs. (default 8192)",
                  "8192");
+    option_parser_register(opp, "-gpgpu_registers_per_block", OPT_UINT32, &gpgpu_registers_per_block,
+                 "Maximum number of registers per CTA. (default 8192)",
+                 "8192");
     option_parser_register(opp, "-gpgpu_shader_cta", OPT_UINT32, &max_cta_per_core, 
                  "Maximum number of concurrent CTAs in shader (default 8)",
                  "8");
@@ -683,6 +686,11 @@ int gpgpu_sim::num_registers_per_core() const
    return m_shader_config->gpgpu_shader_registers;
 }
 
+int gpgpu_sim::num_registers_per_block() const
+{
+   return m_shader_config->gpgpu_registers_per_block;
+}
+
 int gpgpu_sim::wrp_size() const
 {
    return m_shader_config->warp_size;
diff --git a/src/gpgpu-sim/gpu-sim.h b/src/gpgpu-sim/gpu-sim.h
index 8d1c4fc..5908fd5 100644
--- a/src/gpgpu-sim/gpu-sim.h
+++ b/src/gpgpu-sim/gpu-sim.h
@@ -391,6 +391,7 @@ public:
    int shared_mem_size() const;
    int shared_mem_per_block() const;
    int num_registers_per_core() const;
+   int num_registers_per_block() const;
    int wrp_size() const;
    int shader_clock() const;
    const struct cudaDeviceProp *get_prop() const;
diff --git a/src/gpgpu-sim/shader.h b/src/gpgpu-sim/shader.h
index fbddd18..6201840 100644
--- a/src/gpgpu-sim/shader.h
+++ b/src/gpgpu-sim/shader.h
@@ -1286,6 +1286,7 @@ struct shader_core_config : public core_config
     unsigned max_barriers_per_cta;
     char * gpgpu_scheduler_string;
     unsigned gpgpu_shmem_per_block;
+    unsigned gpgpu_registers_per_block;
     char* pipeline_widths_string;
     int pipe_widths[N_PIPELINE_STAGES];