Skip to content

Commit 545f23d

Browse files
slarenslaren
committed
refactor error checking
1 parent 4c0f300 commit 545f23d

File tree

1 file changed

+37
-57
lines changed

1 file changed

+37
-57
lines changed

‎ggml-cuda.cu

Lines changed: 37 additions & 57 deletions
Original file line numberDiff line numberDiff line change
@@ -201,63 +201,43 @@ static __device__ __forceinline__ int __dp4a(const int a, const int b, int c) {
201201

202202
static_assert(sizeof(half) == sizeof(ggml_fp16_t), "wrong fp16 size");
203203

204-
#define CUDA_CHECK(err) \
205-
do { \
206-
cudaError_t err_ = (err); \
207-
if (err_ != cudaSuccess) { \
208-
int id; \
209-
cudaGetDevice(&id); \
210-
fprintf(stderr, "\nCUDA error %d at %s:%d: %s\n", err_, __FILE__, __LINE__, \
211-
cudaGetErrorString(err_)); \
212-
fprintf(stderr, "current device: %d\n", id); \
213-
GGML_ASSERT(!"CUDA error"); \
214-
} \
215-
} while (0)
216-
217-
// driver API
218-
#define CU_CHECK(err) \
219-
do { \
220-
CUresult err_ = (err); \
221-
if (err_ != CUDA_SUCCESS) { \
222-
int id; \
223-
cuDeviceGet(&id, 0); \
224-
const char * err_str; \
225-
cuGetErrorString(err_, &err_str); \
226-
fprintf(stderr, "\nCUDA error %d at %s:%d: %s\n", err_, __FILE__, __LINE__, \
227-
err_str); \
228-
fprintf(stderr, "%s\n", #err); \
229-
fprintf(stderr, "current device: %d\n", id); \
230-
GGML_ASSERT(!"CUDA error"); \
231-
} \
232-
} while (0)
233-
234-
235204
#if CUDART_VERSION >= 12000
236-
#define CUBLAS_CHECK(err) \
237-
do { \
238-
cublasStatus_t err_ = (err); \
239-
if (err_ != CUBLAS_STATUS_SUCCESS) { \
240-
int id; \
241-
cudaGetDevice(&id); \
242-
fprintf(stderr, "\ncuBLAS error %d at %s:%d: %s\n", \
243-
err_, __FILE__, __LINE__, cublasGetStatusString(err_)); \
244-
fprintf(stderr, "current device: %d\n", id); \
245-
GGML_ASSERT(!"cuBLAS error"); \
246-
} \
247-
} while (0)
205+
static const char * cublas_get_error_str(const cublasStatus_t err) {
206+
return cublasGetStatusString(err);
207+
}
248208
#else
249-
#define CUBLAS_CHECK(err) \
250-
do { \
251-
cublasStatus_t err_ = (err); \
252-
if (err_ != CUBLAS_STATUS_SUCCESS) { \
253-
int id; \
254-
cudaGetDevice(&id); \
255-
fprintf(stderr, "\ncuBLAS error %d at %s:%d\n", err_, __FILE__, __LINE__); \
256-
fprintf(stderr, "current device: %d\n", id); \
257-
GGML_ASSERT(!"cuBLAS error"); \
258-
} \
259-
} while (0)
260-
#endif // CUDART_VERSION >= 11
209+
static const char * cublas_get_error_str(const cublasStatus_t err) {
210+
switch (err) {
211+
case CUBLAS_STATUS_SUCCESS: return "CUBLAS_STATUS_SUCCESS";
212+
case CUBLAS_STATUS_NOT_INITIALIZED: return "CUBLAS_STATUS_NOT_INITIALIZED";
213+
case CUBLAS_STATUS_ALLOC_FAILED: return "CUBLAS_STATUS_ALLOC_FAILED";
214+
case CUBLAS_STATUS_INVALID_VALUE: return "CUBLAS_STATUS_INVALID_VALUE";
215+
case CUBLAS_STATUS_ARCH_MISMATCH: return "CUBLAS_STATUS_ARCH_MISMATCH";
216+
case CUBLAS_STATUS_MAPPING_ERROR: return "CUBLAS_STATUS_MAPPING_ERROR";
217+
case CUBLAS_STATUS_EXECUTION_FAILED: return "CUBLAS_STATUS_EXECUTION_FAILED";
218+
case CUBLAS_STATUS_INTERNAL_ERROR: return "CUBLAS_STATUS_INTERNAL_ERROR";
219+
case CUBLAS_STATUS_NOT_SUPPORTED: return "CUBLAS_STATUS_NOT_SUPPORTED";
220+
case CUBLAS_STATUS_LICENSE_ERROR: return "CUBLAS_STATUS_LICENSE_ERROR";
221+
default: return "unknown error";
222+
}
223+
#endif // CUDART_VERSION >= 12000
224+
225+
static const char * cu_get_error_str(CUresult err) {
226+
const char * err_str;
227+
cuGetErrorString(err, &err_str);
228+
return err_str;
229+
}
230+
231+
[[noreturn]]
232+
static void ggml_cuda_error(const char * stmt, const char * func, const char * file, const int line, const char * msg) {
233+
fprintf(stderr, "CUDA error: %s: %s\n", stmt, msg);
234+
fprintf(stderr, " in function %s at %s:%d\n", func, file, line);
235+
GGML_ASSERT(!"CUDA error");
236+
}
237+
238+
#define CUDA_CHECK(err) do { auto err_ = (err); if (err_ != cudaSuccess) ggml_cuda_error(#err, __func__, __FILE__, __LINE__, cudaGetErrorString(err_)); } while (0)
239+
#define CUBLAS_CHECK(err) do { auto err_ = (err); if (err_ != CUBLAS_STATUS_SUCCESS) ggml_cuda_error(#err, __func__, __FILE__, __LINE__, cublas_get_error_str(err_)); } while (0)
240+
#define CU_CHECK(err) do { auto err_ = (err); if (err_ != CUDA_SUCCESS) ggml_cuda_error(#err, __func__, __FILE__, __LINE__, cu_get_error_str(err_)); } while (0)
261241

262242
#if CUDART_VERSION >= 11100
263243
#define GGML_CUDA_ASSUME(x) __builtin_assume(x)
@@ -537,13 +517,13 @@ static int g_device_count = -1;
537517
static int g_main_device = 0;
538518
static float g_tensor_split[GGML_CUDA_MAX_DEVICES] = {0};
539519

540-
struct device_capabilities {
520+
struct cuda_device_capabilities {
541521
int cc; // compute capability
542522
bool vmm; // virtual memory support
543523
size_t vmm_granularity; // granularity of virtual memory
544524
};
545525

546-
static device_capabilities g_device_caps[GGML_CUDA_MAX_DEVICES] = { {0, false, 0} };
526+
static cuda_device_capabilities g_device_caps[GGML_CUDA_MAX_DEVICES] = { {0, false, 0} };
547527

548528

549529
static void * g_scratch_buffer = nullptr;

0 commit comments

Comments
 (0)