Using 16F for internal compute doesn't speed anything up.

liuliu · Jan 2, 2024 · 9f4b282 · 9f4b282
1 parent f7e1aba
commit 9f4b282
Show file tree

Hide file tree

Showing 2 changed files with 6 additions and 1 deletion.
diff --git a/lib/nnc/cmd/blas/gpu/ccv_nnc_gemm_gpu_cublas.cu b/lib/nnc/cmd/blas/gpu/ccv_nnc_gemm_gpu_cublas.cu
@@ -23,6 +23,7 @@ static inline void _ccv_nnc_gbmm_and_bias(cublasHandle_t cublas, const void* con
 			one = &one_f16;
 			break;
 		case CUBLAS_COMPUTE_32F:
+		case CUBLAS_COMPUTE_32F_FAST_TF32:
 			one = &one_f32;
 			break;
 		case CUBLAS_COMPUTE_64F:
@@ -75,6 +76,7 @@ static inline void _ccv_nnc_gbmm(cublasHandle_t cublas, const unsigned char* con
 			one = &one_f16;
 			break;
 		case CUBLAS_COMPUTE_32F:
+		case CUBLAS_COMPUTE_32F_FAST_TF32:
 			one = &one_f32;
 			break;
 		case CUBLAS_COMPUTE_64F:
@@ -249,6 +251,7 @@ static inline void _ccv_nnc_gbmm_dbias(cublasHandle_t cublas, const int flags, c
 			one = &one_f16;
 			break;
 		case CUBLAS_COMPUTE_32F:
+		case CUBLAS_COMPUTE_32F_FAST_TF32:
 			one = &one_f32;
 			break;
 		case CUBLAS_COMPUTE_64F:
@@ -304,6 +307,7 @@ static inline void _ccv_nnc_gbmm_dw(cublasHandle_t cublas, const int flags, cons
 			one = &one_f16;
 			break;
 		case CUBLAS_COMPUTE_32F:
+		case CUBLAS_COMPUTE_32F_FAST_TF32:
 			one = &one_f32;
 			break;
 		case CUBLAS_COMPUTE_64F:
@@ -387,6 +391,7 @@ static inline void _ccv_nnc_gbmm_h(cublasHandle_t cublas, const int flags, const
 			one = &one_f16;
 			break;
 		case CUBLAS_COMPUTE_32F:
+		case CUBLAS_COMPUTE_32F_FAST_TF32:
 			one = &one_f32;
 			break;
 		case CUBLAS_COMPUTE_64F:

diff --git a/lib/nnc/gpu/ccv_nnc_compat.cu b/lib/nnc/gpu/ccv_nnc_compat.cu
@@ -767,7 +767,7 @@ cublasComputeType_t ccv_nnc_cuda_compute_datatype(int datatype)
 		case CCV_32S:
 			return CUBLAS_COMPUTE_32F;
 		case CCV_16F:
-			return CUBLAS_COMPUTE_16F;
+			return CUBLAS_COMPUTE_32F;
 		case CCV_32F:
 			return CUBLAS_COMPUTE_32F;
 		case CCV_64F: