added config -gpgpu_shmem_per_block and fixed cudaGetDeviceAttributes function

4 files changed, 37 insertions, 5 deletions

diff --git a/libcuda/cuda_runtime_api.cc b/libcuda/cuda_runtime_api.cc
index bcc3e14..d6ce4e0 100644
--- a/libcuda/cuda_runtime_api.cc
+++ b/libcuda/cuda_runtime_api.cc
@@ -349,7 +349,24 @@ class _cuda_device_id *GPGPUSim_Init()
 		prop->maxGridSize[2] = 0x40000000;
 		prop->totalConstMem = 0x40000000;
 		prop->textureAlignment = 0;
-		prop->sharedMemPerBlock = the_gpu->shared_mem_size();
+       /*
+
+        * 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	
+		
 		prop->regsPerBlock = the_gpu->num_registers_per_core();
 		prop->warpSize = the_gpu->wrp_size();
 		prop->clockRate = the_gpu->shader_clock();
@@ -861,6 +878,9 @@ __host__ cudaError_t CUDARTAPI cudaDeviceGetAttribute(int *value, enum cudaDevic
 		case 8:
                         *value= prop->sharedMemPerBlock;
                         break;
+		case 9:
+                        *value= prop->totalConstMem;
+                        break;
                 case 10:
                         *value= prop->warpSize;
                         break;
@@ -873,11 +893,14 @@ __host__ cudaError_t CUDARTAPI cudaDeviceGetAttribute(int *value, enum cudaDevic
                 case 16:
                         *value= prop->multiProcessorCount ;
                         break;
+		case 34:
+                        *value= 0;
+                        break;
                 case 39:
                         *value= dev->get_gpgpu()->threads_per_core();
                         break;
                 case 75:
-                        *value= 8 ;
+                        *value= 9 ;
                         break;
                 case 76:
                         *value= 3 ;
@@ -886,13 +909,13 @@ __host__ cudaError_t CUDARTAPI cudaDeviceGetAttribute(int *value, enum cudaDevic
                         *value= 0 ; //TODO: as of now, we dont support stream priorities.
                         break;
 		case 81:
-                        *value= prop->sharedMemPerBlock;
+                        *value= prop->sharedMemPerMultiprocessor;
                         break;
                 case 82:
                         *value= prop->regsPerBlock;
                         break;
 		default:
-			printf("ERROR: implement the attribute numbered %d \n",attr);
+			printf("ERROR: Attribute number %d unimplemented \n",attr);
 			abort();
                 }
                 return g_last_cudaError = cudaSuccess;
diff --git a/src/gpgpu-sim/gpu-sim.cc b/src/gpgpu-sim/gpu-sim.cc
index 3829861..5d63ab7 100644
--- a/src/gpgpu-sim/gpu-sim.cc
+++ b/src/gpgpu-sim/gpu-sim.cc
@@ -265,6 +265,9 @@ void shader_core_config::reg_options(class OptionParser * opp)
     option_parser_register(opp, "-gpgpu_n_ldst_response_buffer_size", OPT_UINT32, &ldst_unit_response_queue_size, 
                  "number of response packets in ld/st unit ejection buffer",
                  "2");
+    option_parser_register(opp, "-gpgpu_shmem_per_block", OPT_UINT32, &gpgpu_shmem_per_block,
+                 "Size of shared memory per thread block or CTA (default 48kB)",
+                 "49152");
     option_parser_register(opp, "-gpgpu_shmem_size", OPT_UINT32, &gpgpu_shmem_size,
                  "Size of shared memory per shader core (default 16kB)",
                  "16384");
@@ -661,6 +664,11 @@ int gpgpu_sim::shared_mem_size() const
    return m_shader_config->gpgpu_shmem_size;
 }
 
+int gpgpu_sim::shared_mem_per_block() const
+{
+   return m_shader_config->gpgpu_shmem_per_block;
+}
+
 int gpgpu_sim::num_registers_per_core() const
 {
    return m_shader_config->gpgpu_shader_registers;
diff --git a/src/gpgpu-sim/gpu-sim.h b/src/gpgpu-sim/gpu-sim.h
index 7d92c66..8d1c4fc 100644
--- a/src/gpgpu-sim/gpu-sim.h
+++ b/src/gpgpu-sim/gpu-sim.h
@@ -389,6 +389,7 @@ public:
    void get_pdom_stack_top_info( unsigned sid, unsigned tid, unsigned *pc, unsigned *rpc );
 
    int shared_mem_size() const;
+   int shared_mem_per_block() const;
    int num_registers_per_core() const;
    int wrp_size() const;
    int shader_clock() const;
diff --git a/src/gpgpu-sim/shader.h b/src/gpgpu-sim/shader.h
index bdd8dbe..fbddd18 100644
--- a/src/gpgpu-sim/shader.h
+++ b/src/gpgpu-sim/shader.h
@@ -1285,7 +1285,7 @@ struct shader_core_config : public core_config
     unsigned max_cta_per_core; //Limit on number of concurrent CTAs in shader core
     unsigned max_barriers_per_cta;
     char * gpgpu_scheduler_string;
-
+    unsigned gpgpu_shmem_per_block;
     char* pipeline_widths_string;
     int pipe_widths[N_PIPELINE_STAGES];