Skip to content

Update the cuda API and enable tensor core for GEMM #9622

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
merged 2 commits into from
Apr 6, 2018
Merged

Update the cuda API and enable tensor core for GEMM #9622

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
42b6f75
change from hgemm to gemmEx
kexinzhao Apr 4, 2018
96ea7b7
fix cpplint
kexinzhao Apr 4, 2018
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
33 changes: 24 additions & 9 deletions paddle/fluid/operators/math/math_function.cu
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -39,18 +39,33 @@ void gemm<platform::CUDADeviceContext, float16>(
cublasOperation_t cuTransB =
(transB == CblasNoTrans) ? CUBLAS_OP_N : CUBLAS_OP_T;

const half h_alpha = static_cast<const half>(alpha);
const half h_beta = static_cast<const half>(beta);
const half* h_A = reinterpret_cast<const half*>(A);
const half* h_B = reinterpret_cast<const half*>(B);
half* h_C = reinterpret_cast<half*>(C);
float h_alpha = static_cast<float>(alpha);
float h_beta = static_cast<float>(beta);

// TODO(kexinzhao): add processing code for compute capability < 53 case
PADDLE_ENFORCE_GE(context.GetComputeCapability(), 53,
"cublas Hgemm requires GPU compute capability >= 53");
PADDLE_ENFORCE(platform::dynload::cublasHgemm(
context.cublas_handle(), cuTransB, cuTransA, N, M, K, &h_alpha, h_B, ldb,
h_A, lda, &h_beta, h_C, N));
"cublas fp16 gemm requires GPU compute capability >= 53");

cublasGemmAlgo_t algo = CUBLAS_GEMM_DFALT;
#if CUDA_VERSION >= 9000
if (context.GetComputeCapability() >= 70) {
PADDLE_ENFORCE(platform::dynload::cublasSetMathMode(context.cublas_handle(),
CUBLAS_TENSOR_OP_MATH));
algo = CUBLAS_GEMM_DFALT_TENSOR_OP;
} else {
PADDLE_ENFORCE(platform::dynload::cublasSetMathMode(context.cublas_handle(),
CUBLAS_DEFAULT_MATH));
}
#endif

// cublasHgemm does true FP16 computation which is slow for non-Volta
// GPUs. So use cublasGemmEx instead which does pesudo FP16 computation:
// input/output in fp16, computation in fp32, which can also be accelerated
// using tensor cores in volta GPUs.
PADDLE_ENFORCE(platform::dynload::cublasGemmEx(
context.cublas_handle(), cuTransB, cuTransA, N, M, K, &h_alpha, B,
CUDA_R_16F, ldb, A, CUDA_R_16F, lda, &h_beta, C, CUDA_R_16F, N,
CUDA_R_32F, algo));
}

template <>
Expand Down
4 changes: 4 additions & 0 deletions paddle/fluid/platform/dynload/cublas.cc
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -24,6 +24,10 @@ void *cublas_dso_handle = nullptr;

CUBLAS_BLAS_ROUTINE_EACH(DEFINE_WRAP);

#ifdef CUBLAS_BLAS_ROUTINE_EACH_R2
CUBLAS_BLAS_ROUTINE_EACH_R2(DEFINE_WRAP);
#endif

} // namespace dynload
} // namespace platform
} // namespace paddle
14 changes: 11 additions & 3 deletions paddle/fluid/platform/dynload/cublas.h
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -15,8 +15,9 @@ limitations under the License. */
#pragma once

#include <cublas_v2.h>
#include <cuda.h>
#include <dlfcn.h>
#include <mutex>
#include <mutex> // NOLINT
#include "paddle/fluid/platform/dynload/dynamic_loader.h"

namespace paddle {
Expand Down Expand Up @@ -70,6 +71,7 @@ extern void *cublas_dso_handle;
__macro(cublasDgemm_v2); \
__macro(cublasHgemm); \
__macro(cublasSgemmEx); \
__macro(cublasGemmEx); \
__macro(cublasSgeam_v2); \
__macro(cublasDgeam_v2); \
__macro(cublasCreate_v2); \
Expand All @@ -89,9 +91,15 @@ extern void *cublas_dso_handle;
__macro(cublasSgetrfBatched); \
__macro(cublasSgetriBatched); \
__macro(cublasDgetrfBatched); \
__macro(cublasDgetriBatched)
__macro(cublasDgetriBatched);

CUBLAS_BLAS_ROUTINE_EACH(DECLARE_DYNAMIC_LOAD_CUBLAS_WRAP);
CUBLAS_BLAS_ROUTINE_EACH(DECLARE_DYNAMIC_LOAD_CUBLAS_WRAP)

// APIs available after CUDA 9.0
#if CUDA_VERSION >= 9000
#define CUBLAS_BLAS_ROUTINE_EACH_R2(__macro) __macro(cublasSetMathMode);
CUBLAS_BLAS_ROUTINE_EACH_R2(DECLARE_DYNAMIC_LOAD_CUBLAS_WRAP)
#endif

#undef DECLARE_DYNAMIC_LOAD_CUBLAS_WRAP
} // namespace dynload
Expand Down