You are viewing a plain text version of this content. The canonical link for it is here.
Posted to commits@mxnet.apache.org by GitBox <gi...@apache.org> on 2020/01/29 06:18:59 UTC

[GitHub] [incubator-mxnet] eric-haibin-lin commented on a change in pull request #17466: Bert gemms true fp16

eric-haibin-lin commented on a change in pull request #17466: Bert gemms true fp16
URL: https://github.com/apache/incubator-mxnet/pull/17466#discussion_r372206029
 
 

 ##########
 File path: src/operator/linalg_impl.h
 ##########
 @@ -261,25 +262,43 @@ void linalg_gemm<gpu, mshadow::half::half_t>(const Tensor<gpu, 2, mshadow::half:
   auto previous_math_mode = SetCublasMathMode(blas_handle, cublas_math_mode);
 #endif
 
-  // pseudo-fp16 (fp32 math with fp16 I/O)
-  float alpha_f = float(alpha);  // NOLINT(*)
-  float beta_f = float(beta);  // NOLINT(*)
-
-  // As of cuda8, cublas adopted the cuda datatype, rather than maintaining its own datatype.
+// As of cuda8, cublas adopted the cuda datatype, rather than maintaining its own datatype.
 #if CUDA_VERSION >= 8000
   cudaDataType_t half_datatype = CUDA_R_16F;
 #else
   cublasDataType_t half_datatype = CUBLAS_DATA_HALF;
 #endif
-  CUBLAS_CALL(cublasSgemmEx(blas_handle,
-                            (tB ? CUBLAS_OP_T : CUBLAS_OP_N),
-                            (tA ? CUBLAS_OP_T : CUBLAS_OP_N),
-                            C.size(1), C.size(0), (tB ? B.size(1) : B.size(0)),
-                            &alpha_f,
-                            B.dptr_, half_datatype, B.stride_,
-                            A.dptr_, half_datatype, A.stride_,
-                            &beta_f,
-                            C.dptr_, half_datatype, C.stride_));
+  auto algo = CUBLAS_GEMM_DEFAULT_TENSOR_OP;
+  using TrueFP16Type = mshadow::half::half_t;
+  using PseudoFP16Type = typename CublasType<mshadow::half::half_t>::ScaleType;
+  TrueFP16Type trueFP16_alpha = static_cast<TrueFP16Type>(alpha);
+  TrueFP16Type trueFP16_beta = static_cast<TrueFP16Type>(beta);
+  PseudoFP16Type pseudoFP16_alpha = static_cast<PseudoFP16Type>(alpha);
+  PseudoFP16Type pseudoFP16_beta = static_cast<PseudoFP16Type>(beta);
+  const void *alpha_ptr;
+  const void *beta_ptr;
+  cudaDataType_t computeType;
+  bool use_true_fp16 = dmlc::GetEnv("MXNET_FC_TRUE_FP16", false);
 
 Review comment:
   For new env vars, please document in https://github.com/apache/incubator-mxnet/blob/master/docs/static_site/src/pages/api/faq/env_var.md 

----------------------------------------------------------------
This is an automated message from the Apache Git Service.
To respond to the message, please log on to GitHub and use the
URL above to go to the specific comment.
 
For queries about this service, please contact Infrastructure at:
users@infra.apache.org


With regards,
Apache Git Services