Adding additional fixes for the texture bugs

1 files changed, 12 insertions, 10 deletions

diff --git a/libcuda/cuda_runtime_api.cc b/libcuda/cuda_runtime_api.cc
index 7569b20..f6e5e6f 100644
--- a/libcuda/cuda_runtime_api.cc
+++ b/libcuda/cuda_runtime_api.cc
@@ -1121,10 +1121,10 @@ __host__ cudaError_t CUDARTAPI cudaDeviceGetAttribute(int *value, enum cudaDevic
                         *value= 0;
                         break;
                 case 75:
-                        *value= 9 ;
+                        *value= 7 ;
                         break;
                 case 76:
-                        *value= 3 ;
+                        *value= 0 ;
                         break;
                 case 78:
                         *value= 0 ; //TODO: as of now, we dont support stream priorities.
@@ -1199,22 +1199,23 @@ __host__ cudaError_t CUDARTAPI cudaGetDevice(int *device)
 	*device = g_active_device;
 	return g_last_cudaError = cudaSuccess;
 }
+
 __host__ cudaError_t CUDARTAPI cudaDeviceGetLimit ( size_t* pValue, cudaLimit limit )
 {
 	if(g_debug_execution >= 3){
             announce_call(__my_func__);
    	 }
         switch(limit) {
-        case 0:
+        case 0:  // cudaLimitStackSize
                                  *pValue=1024;
                                  break;
-        case 2:
+        case 2:  // cudaLimitMallocHeapSize
                                  *pValue=8388608;
                                  break;
-        case 3:
+        case 3: // cudaLimitDevRuntimeSyncDepth
                                  *pValue=2;
                                  break;
-        case 4:
+        case 4: // cudaLimitDevRuntimePendingLaunchCount
                                  *pValue=2048;
                                  break;
         default:
@@ -1225,7 +1226,6 @@ __host__ cudaError_t CUDARTAPI cudaDeviceGetLimit ( size_t* pValue, cudaLimit li
 
 }
 
-
 __host__ cudaError_t CUDARTAPI cudaStreamGetPriority ( cudaStream_t hStream, int* priority )
 {
         if(g_debug_execution >= 3){
@@ -1585,7 +1585,7 @@ __host__ cudaError_t CUDARTAPI cudaStreamDestroy(cudaStream_t stream)
 	    announce_call(__my_func__);
     }
 #if (CUDART_VERSION >= 3000)
-	//synchronization required for application using external libraries without explicit synchronization in the code to 
+	//per-stream synchronization required for application using external libraries without explicit synchronization in the code to 
 	//avoid the stream_manager from spinning forever to destroy non-empty streams without making any forward progress. 
 	stream->synchronize();
 	g_stream_manager->destroy_stream(stream);
@@ -2714,7 +2714,6 @@ cudaError_t cudaDeviceReset ( void ) {
 	return g_last_cudaError = cudaSuccess;
 }
 cudaError_t CUDARTAPI cudaDeviceSynchronize(void){
-	// I don't know what this should do
 	if(g_debug_execution >= 3){
 	    announce_call(__my_func__);
     }
@@ -2722,7 +2721,6 @@ cudaError_t CUDARTAPI cudaDeviceSynchronize(void){
 	return g_last_cudaError = cudaSuccess;
 }
 
-
 void CUDARTAPI __cudaRegisterFunction(
 		void   **fatCubinHandle,
 		const char    *hostFun,
@@ -3333,6 +3331,10 @@ kernel_info_t *gpgpu_cuda_ptx_sim_init_grid( const char *hostFun,
     }
 	function_info *entry = context->get_kernel(hostFun);
 	gpgpu_t* gpu= context->get_device()->get_gpgpu();
+	/*
+	Passing a snapshot of the GPU's current texture mapping to the kernel's info
+	as kernels should use texture bindings present at the time of their launch.
+	*/
 	kernel_info_t *result = new kernel_info_t(gridDim,blockDim,entry,gpu->getNameArrayMapping(),gpu->getNameInfoMapping());
 	if( entry == NULL ) {
 		printf("GPGPU-Sim PTX: ERROR launching kernel -- no PTX implementation found for %p\n", hostFun);