From c4d1dff7816f11c82b588b683a893dce27ff2f5d Mon Sep 17 00:00:00 2001
From: Luka Stanisic <luka.stanisic@mpcdf.mpg.de>
Date: Fri, 16 Jun 2017 10:33:56 +0200
Subject: [PATCH] checking CUDA driver errors
---
bioem_cuda.cu | 8 ++--
include/bioem_cuda_internal.h | 71 +++++++++++++++++++++++++++++++++++
2 files changed, 75 insertions(+), 4 deletions(-)
diff --git a/bioem_cuda.cu b/bioem_cuda.cu
index aa70fb7..26535b3 100644
--- a/bioem_cuda.cu
+++ b/bioem_cuda.cu
@@ -297,13 +297,13 @@ int bioem_cuda::selectCudaDevice()
#else
unsigned int free, total;
#endif
- cuInit(0);
+ CU_ERROR_CHECK(cuInit(0));
CUdevice tmpDevice;
- cuDeviceGet(&tmpDevice, i);
+ CU_ERROR_CHECK(cuDeviceGet(&tmpDevice, i));
CUcontext tmpContext;
- cuCtxCreate(&tmpContext, 0, tmpDevice);
+ CU_ERROR_CHECK(cuCtxCreate(&tmpContext, 0, tmpDevice));
if(cuMemGetInfo(&free, &total)) exit(1);
- cuCtxDestroy(tmpContext);
+ CU_ERROR_CHECK(cuCtxDestroy(tmpContext));
checkCudaErrors(cudaGetDeviceProperties(&deviceProp, i));
if (DebugOutput >= 2 && mpi_rank == 0) printf("CUDA Device %2d: %s (Rev: %d.%d - Mem Avail %lld / %lld)\n", i, deviceProp.name, deviceProp.major, deviceProp.minor, (long long int) free, (long long int) deviceProp.totalGlobalMem);
diff --git a/include/bioem_cuda_internal.h b/include/bioem_cuda_internal.h
index 4d44c5e..60939da 100644
--- a/include/bioem_cuda_internal.h
+++ b/include/bioem_cuda_internal.h
@@ -72,5 +72,76 @@ private:
int maxRef;
};
+/* Handing CUDA Driver errors */
+/* Inspired from: https://github.com/garymacindoe/cuda-cholesky */
+
+// Expand and stringify argument
+#define STRINGx(x) #x
+#define STRING(x) STRINGx(x)
+
+const char * cuGetError(CUresult result) {
+ switch (result) {
+ case CUDA_SUCCESS: return "No errors";
+ case CUDA_ERROR_INVALID_VALUE: return "Invalid value";
+ case CUDA_ERROR_OUT_OF_MEMORY: return "Out of memory";
+ case CUDA_ERROR_NOT_INITIALIZED: return "Driver not initialized";
+ case CUDA_ERROR_DEINITIALIZED: return "Driver deinitialized";
+ case CUDA_ERROR_PROFILER_DISABLED: return "Profiler disabled";
+ case CUDA_ERROR_PROFILER_NOT_INITIALIZED: return "Profiler not initialized";
+ case CUDA_ERROR_PROFILER_ALREADY_STARTED: return "Profiler already started";
+ case CUDA_ERROR_PROFILER_ALREADY_STOPPED: return "Profiler already stopped";
+ case CUDA_ERROR_NO_DEVICE: return "No CUDA-capable device available";
+ case CUDA_ERROR_INVALID_DEVICE: return "Invalid device";
+ case CUDA_ERROR_INVALID_IMAGE: return "Invalid kernel image";
+ case CUDA_ERROR_INVALID_CONTEXT: return "Invalid context";
+ case CUDA_ERROR_CONTEXT_ALREADY_CURRENT: return "Context already current";
+ case CUDA_ERROR_MAP_FAILED: return "Map failed";
+ case CUDA_ERROR_UNMAP_FAILED: return "Unmap failed";
+ case CUDA_ERROR_ARRAY_IS_MAPPED: return "Array is mapped";
+ case CUDA_ERROR_ALREADY_MAPPED: return "Already mapped";
+ case CUDA_ERROR_NO_BINARY_FOR_GPU: return "No binary for GPU";
+ case CUDA_ERROR_ALREADY_ACQUIRED: return "Already acquired";
+ case CUDA_ERROR_NOT_MAPPED: return "Not mapped";
+ case CUDA_ERROR_NOT_MAPPED_AS_ARRAY: return "Not mapped as array";
+ case CUDA_ERROR_NOT_MAPPED_AS_POINTER: return "Not mapped as pointer";
+ case CUDA_ERROR_ECC_UNCORRECTABLE: return "Uncorrectable ECC error";
+ case CUDA_ERROR_UNSUPPORTED_LIMIT: return "Unsupported CUlimit";
+ case CUDA_ERROR_CONTEXT_ALREADY_IN_USE: return "Context already in use";
+ case CUDA_ERROR_INVALID_SOURCE: return "Invalid source";
+ case CUDA_ERROR_FILE_NOT_FOUND: return "File not found";
+ case CUDA_ERROR_SHARED_OBJECT_SYMBOL_NOT_FOUND: return "Shared object symbol not found";
+ case CUDA_ERROR_SHARED_OBJECT_INIT_FAILED: return "Shared object initialization failed";
+ case CUDA_ERROR_OPERATING_SYSTEM: return "Operating System call failed";
+ case CUDA_ERROR_INVALID_HANDLE: return "Invalid handle";
+ case CUDA_ERROR_NOT_FOUND: return "Not found";
+ case CUDA_ERROR_NOT_READY: return "CUDA not ready";
+ case CUDA_ERROR_LAUNCH_FAILED: return "Launch failed";
+ case CUDA_ERROR_LAUNCH_OUT_OF_RESOURCES: return "Launch exceeded resources";
+ case CUDA_ERROR_LAUNCH_TIMEOUT: return "Launch exceeded timeout";
+ case CUDA_ERROR_LAUNCH_INCOMPATIBLE_TEXTURING: return "Launch with incompatible texturing";
+ case CUDA_ERROR_PEER_ACCESS_ALREADY_ENABLED: return "Peer access already enabled";
+ case CUDA_ERROR_PEER_ACCESS_NOT_ENABLED: return "Peer access not enabled";
+ case CUDA_ERROR_PRIMARY_CONTEXT_ACTIVE: return "Primary context active";
+ case CUDA_ERROR_CONTEXT_IS_DESTROYED: return "Context is destroyed";
+ case CUDA_ERROR_ASSERT: return "Device assert failed";
+ case CUDA_ERROR_TOO_MANY_PEERS: return "Too many peers";
+ case CUDA_ERROR_HOST_MEMORY_ALREADY_REGISTERED: return "Host memory already registered";
+ case CUDA_ERROR_HOST_MEMORY_NOT_REGISTERED: return "Host memory not registered";
+ case CUDA_ERROR_UNKNOWN: return "Unknown error";
+ default: return "Unknown error code";
+ }
+}
+
+#define CU_ERROR_CHECK(call) \
+ do { \
+ CUresult __error__; \
+ if ((__error__ = (call)) != CUDA_SUCCESS) { \
+ printf(STRING(call), __func__, __FILE__, __LINE__, __error__, \
+ (const char * (*)(int))cuGetError); \
+ return __error__; \
+ } \
+ } while (false)
+
+
#endif
--
GitLab