Better debugging for GPU.

hpc · Dec 23, 2022 · 6aab769 · 6aab769
1 parent 58049c0
commit 6aab769
Show file tree

Hide file tree

Showing 4 changed files with 18 additions and 6 deletions.
diff --git a/src/ior-output.c b/src/ior-output.c
@@ -70,7 +70,7 @@ static void PrintKeyValEnd(){
   needNextToken = 1;
 }
 
-static void PrintKeyVal(char * key, char * value){
+void PrintKeyVal(char * key, char * value){
   if(value != NULL && value[0] != 0 && value[strlen(value) -1 ] == '\n'){
     // remove \n
     value[strlen(value) -1 ] = 0;
@@ -450,6 +450,11 @@ void ShowSetup(IOR_param_t *params)
   if (params->memoryPerNode != 0){
     PrintKeyVal("memoryPerNode", HumanReadable(params->memoryPerNode, BASE_TWO));
   }
+
+  PrintKeyVal("memoryBuffer", params->gpuMemoryFlags == IOR_MEMORY_TYPE_CPU ? "CPU" :  params->gpuMemoryFlags == IOR_MEMORY_TYPE_GPU_DEVICE_ONLY ? "GPU" : "Managed");
+  PrintKeyVal("dataAccess", params->gpuMemoryFlags > IOR_MEMORY_TYPE_GPU_MANAGED_CHECK_GPU ? "GPU" : "CPU");
+  PrintKeyVal("GPUDirect", params->gpuDirect ? "1" : "0");
+
   PrintKeyValInt("repetitions", params->repetitions);
   PrintKeyVal("xfersize", HumanReadable(params->transferSize, BASE_TWO));
   PrintKeyVal("blocksize", HumanReadable(params->blockSize, BASE_TWO));

diff --git a/src/ior.c b/src/ior.c
@@ -126,7 +126,17 @@ static int test_initialize(IOR_test_t * test){
   backend = test->params.backend;
 
 #ifdef HAVE_CUDA
-  cudaError_t cret = cudaSetDevice(test->params.gpuID);
+  int device_count;
+  cudaError_t cret = cudaGetDeviceCount(& device_count);
+  if(cret != cudaSuccess){
+    ERRF("cudaGetDeviceCount() error: %d %s", (int) cret, cudaGetErrorString(cret));
+  }  
+  {
+  char val[20];
+  sprintf(val, "%d", device_count);
+  PrintKeyVal("cudaDevices", val);
+  }
+  cret = cudaSetDevice(test->params.gpuID);
   if(cret != cudaSuccess){
     WARNF("cudaSetDevice(%d) error: %s", test->params.gpuID, cudaGetErrorString(cret));
   }

diff --git a/src/parse_options.c b/src/parse_options.c
@@ -64,10 +64,6 @@ static void CheckRunSettings(IOR_test_t *tests)
                 if(params->dualMount && !params->filePerProc) {
                   ERR("Dual Mount can only be used with File Per Process");
                 }
-
-                if(params->gpuDirect){
-                  params->gpuMemoryFlags = IOR_MEMORY_TYPE_GPU_DEVICE_ONLY;
-                }
         }
 }
 

diff --git a/src/utilities.h b/src/utilities.h
@@ -47,6 +47,7 @@ void invalidate_buffer_pattern(char * buf, size_t bytes, ior_memory_flags type);
 int verify_memory_pattern(uint64_t item, char * buffer, size_t bytes, int rand_seed, int pretendRank, ior_dataPacketType_e dataPacketType, ior_memory_flags type);
 int verify_memory_pattern_gpu(uint64_t item, char * buffer, size_t bytes, int rand_seed, int pretendRank, ior_dataPacketType_e dataPacketType);
 
+void PrintKeyVal(char * key, char * value);
 char *CurrentTimeString(void);
 int Regex(char *, char *);
 void ShowFileSystemSize(char * filename, const struct ior_aiori * backend, void * backend_options);