Skip to content

P
Paddle

马赛克医院 / Paddle
与 Fork 源项目一致

Fork自 PaddlePaddle / Paddle

通知 1
Star 1
Fork 0
P
Paddle

体验新版 GitCode，发现更多精彩内容 >>

未验证 提交 c79fa1c3 编写于 作者: L Lijunhui 提交者： GitHub
浏览文件
操作

Set Tensor Core MathType for bfloat16 in conv using cudnn (#34409)

上级 56b7ebbc
隐藏空白更改
内联 并排
Showing with 39 addition and 68 deletion
paddle/fluid/operators/conv_cudnn_helper.h
浏览文件 @ c79fa1c3
...@@ -169,6 +169,35 @@ void ChooseAlgo(const std::vector<PerfType>& perf_results, ...@@ -169,6 +169,35 @@ void ChooseAlgo(const std::vector<PerfType>& perf_results,
using framework::ConvSearchCache; using framework::ConvSearchCache;
static void SetConvMathType(const framework::ExecutionContext& ctx,
cudnnDataType_t dtype,
const platform::ConvolutionDescriptor& cdesc) {
#if CUDA_VERSION >= 9000 && CUDNN_VERSION_MIN(7, 0, 1)
auto& dev_ctx = ctx.template device_context<platform::CUDADeviceContext>();
if (dev_ctx.GetComputeCapability() >= 70 && dtype == CUDNN_DATA_HALF) {
PADDLE_ENFORCE_CUDA_SUCCESS(platform::dynload::cudnnSetConvolutionMathType(
cdesc.desc(), CUDNN_TENSOR_OP_MATH));
VLOG(5) << "use cudnn_tensor_op_math";
#if CUDA_VERSION >= 11000
#if CUDNN_VERSION_MIN(8, 1, 0)
} else if (dev_ctx.GetComputeCapability() >= 80 &&
dtype == CUDNN_DATA_BFLOAT16) {
PADDLE_ENFORCE_CUDA_SUCCESS(platform::dynload::cudnnSetConvolutionMathType(
cdesc.desc(), CUDNN_TENSOR_OP_MATH));
#endif // CUDNN_VERSION_MIN(8, 1, 0)
} else if (dtype == CUDNN_DATA_FLOAT && !cdesc.allow_tf32_) {
PADDLE_ENFORCE_CUDA_SUCCESS(platform::dynload::cudnnSetConvolutionMathType(
cdesc.desc(), CUDNN_FMA_MATH));
#endif // CUDA_VERSION >= 11000
} else {
PADDLE_ENFORCE_CUDA_SUCCESS(platform::dynload::cudnnSetConvolutionMathType(
cdesc.desc(), CUDNN_DEFAULT_MATH));
VLOG(5) << "NOT use cudnn_tensor_op_math";
}
#endif
return;
}
struct ConvArgs { struct ConvArgs {
cudnnHandle_t handle; cudnnHandle_t handle;
platform::TensorDescriptor idesc, odesc; platform::TensorDescriptor idesc, odesc;
...@@ -208,36 +237,7 @@ struct SearchAlgorithm<cudnnConvolutionFwdAlgoPerf_t> { ...@@ -208,36 +237,7 @@ struct SearchAlgorithm<cudnnConvolutionFwdAlgoPerf_t> {
size_t workspace_size_limit = FLAGS_conv_workspace_size_limit * 1024 * 1024; size_t workspace_size_limit = FLAGS_conv_workspace_size_limit * 1024 * 1024;
size_t workspace_size = 0; size_t workspace_size = 0;
algo_t algo; algo_t algo;
SetConvMathType(ctx, dtype, args.cdesc);
#if CUDA_VERSION >= 9000 && CUDNN_VERSION_MIN(7, 0, 1)
auto& dev_ctx = ctx.template device_context<platform::CUDADeviceContext>();
if (dev_ctx.GetComputeCapability() >= 70 && dtype == CUDNN_DATA_HALF) {
PADDLE_ENFORCE_CUDA_SUCCESS(
platform::dynload::cudnnSetConvolutionMathType(args.cdesc.desc(),
CUDNN_TENSOR_OP_MATH));
VLOG(5) << "use cudnn_tensor_op_math";
#if CUDA_VERSION >= 11000
#if CUDNN_VERSION_MIN(8, 1, 0)
} else if (dev_ctx.GetComputeCapability() >= 80 &&
dtype == CUDNN_DATA_BFLOAT16) {
PADDLE_ENFORCE_CUDA_SUCCESS(
platform::dynload::cudnnSetConvolutionMathType(args.cdesc.desc(),
CUDNN_TENSOR_OP_MATH));
VLOG(5) << "use cudnn_tensor_op_math";
#endif // CUDNN_VERSION >= 8100
} else if (dtype == CUDNN_DATA_FLOAT && !args.cdesc.allow_tf32_) {
PADDLE_ENFORCE_CUDA_SUCCESS(
platform::dynload::cudnnSetConvolutionMathType(args.cdesc.desc(),
CUDNN_FMA_MATH));
VLOG(5) << "use cudnn_fma_math";
#endif // CUDA_VERSION >= 11000
} else {
PADDLE_ENFORCE_CUDA_SUCCESS(
platform::dynload::cudnnSetConvolutionMathType(args.cdesc.desc(),
CUDNN_DEFAULT_MATH));
VLOG(5) << "use cudnn_default_math";
}
#endif
if (!exhaustive_search && !deterministic) { if (!exhaustive_search && !deterministic) {
#if CUDNN_VERSION >= 7001 #if CUDNN_VERSION >= 7001
...@@ -353,24 +353,7 @@ struct SearchAlgorithm<cudnnConvolutionBwdDataAlgoPerf_t> { ...@@ -353,24 +353,7 @@ struct SearchAlgorithm<cudnnConvolutionBwdDataAlgoPerf_t> {
size_t workspace_size = 0; size_t workspace_size = 0;
bool has_got_workspace_size = true; bool has_got_workspace_size = true;
algo_t algo; algo_t algo;
#if CUDA_VERSION >= 9000 && CUDNN_VERSION_MIN(7, 0, 1) SetConvMathType(ctx, dtype, args.cdesc);
auto& dev_ctx = ctx.template device_context<platform::CUDADeviceContext>();
PADDLE_ENFORCE_CUDA_SUCCESS(platform::dynload::cudnnSetConvolutionMathType(
args.cdesc.desc(), CUDNN_DEFAULT_MATH));
VLOG(5) << "NOT use cudnn_tensor_op_math";
if (dev_ctx.GetComputeCapability() >= 70 && dtype == CUDNN_DATA_HALF) {
PADDLE_ENFORCE_CUDA_SUCCESS(
platform::dynload::cudnnSetConvolutionMathType(args.cdesc.desc(),
CUDNN_TENSOR_OP_MATH));
VLOG(5) << "use cudnn_tensor_op_math";
} else if (dtype == CUDNN_DATA_FLOAT && !args.cdesc.allow_tf32_) {
#if CUDA_VERSION >= 11000
PADDLE_ENFORCE_CUDA_SUCCESS(
platform::dynload::cudnnSetConvolutionMathType(args.cdesc.desc(),
CUDNN_FMA_MATH));
#endif // CUDA_VERSION >= 11000
}
#endif
if (!exhaustive_search && !deterministic) { if (!exhaustive_search && !deterministic) {
#if CUDNN_VERSION >= 7001 #if CUDNN_VERSION >= 7001
...@@ -501,25 +484,7 @@ struct SearchAlgorithm<cudnnConvolutionBwdFilterAlgoPerf_t> { ...@@ -501,25 +484,7 @@ struct SearchAlgorithm<cudnnConvolutionBwdFilterAlgoPerf_t> {
size_t workspace_size_limit = FLAGS_conv_workspace_size_limit * 1024 * 1024; size_t workspace_size_limit = FLAGS_conv_workspace_size_limit * 1024 * 1024;
size_t workspace_size = 0; size_t workspace_size = 0;
bool has_got_workspace_size = true; bool has_got_workspace_size = true;
SetConvMathType(ctx, dtype, args.cdesc);
#if CUDA_VERSION >= 9000 && CUDNN_VERSION_MIN(7, 0, 1)
auto& dev_ctx = ctx.template device_context<platform::CUDADeviceContext>();
PADDLE_ENFORCE_CUDA_SUCCESS(platform::dynload::cudnnSetConvolutionMathType(
args.cdesc.desc(), CUDNN_DEFAULT_MATH));
VLOG(5) << "NOT use cudnn_tensor_op_math";
if (dev_ctx.GetComputeCapability() >= 70 && dtype == CUDNN_DATA_HALF) {
PADDLE_ENFORCE_CUDA_SUCCESS(
platform::dynload::cudnnSetConvolutionMathType(args.cdesc.desc(),
CUDNN_TENSOR_OP_MATH));
VLOG(5) << "use cudnn_tensor_op_math";
} else if (dtype == CUDNN_DATA_FLOAT && !args.cdesc.allow_tf32_) {
#if CUDA_VERSION >= 11000
PADDLE_ENFORCE_CUDA_SUCCESS(
platform::dynload::cudnnSetConvolutionMathType(args.cdesc.desc(),
CUDNN_FMA_MATH));
#endif // CUDA_VERSION >= 11000
}
#endif
algo_t algo; algo_t algo;
if (!exhaustive_search && !deterministic) { if (!exhaustive_search && !deterministic) {
......
paddle/fluid/platform/cudnn_desc.h
浏览文件 @ c79fa1c3
...@@ -253,8 +253,14 @@ class ConvolutionDescriptor { ...@@ -253,8 +253,14 @@ class ConvolutionDescriptor {
PADDLE_ENFORCE_CUDA_SUCCESS( PADDLE_ENFORCE_CUDA_SUCCESS(
platform::dynload::cudnnSetConvolutionMathType(desc, platform::dynload::cudnnSetConvolutionMathType(desc,
CUDNN_TENSOR_OP_MATH)); CUDNN_TENSOR_OP_MATH));
} else if (dtype == CUDNN_DATA_FLOAT && !allow_tf32) {
#if CUDA_VERSION >= 11000 #if CUDA_VERSION >= 11000
#if CUDNN_VERSION_MIN(8, 1, 0)
} else if (dtype == CUDNN_DATA_BFLOAT16) {
PADDLE_ENFORCE_CUDA_SUCCESS(
platform::dynload::cudnnSetConvolutionMathType(desc,
CUDNN_TENSOR_OP_MATH));
#endif // CUDNN_VERSION_MIN(8,1,0)
} else if (dtype == CUDNN_DATA_FLOAT && !allow_tf32) {
PADDLE_ENFORCE_CUDA_SUCCESS( PADDLE_ENFORCE_CUDA_SUCCESS(
platform::dynload::cudnnSetConvolutionMathType(desc, CUDNN_FMA_MATH)); platform::dynload::cudnnSetConvolutionMathType(desc, CUDNN_FMA_MATH));
#endif // CUDA_VERSION >= 11000 #endif // CUDA_VERSION >= 11000
......
Markdown is supported
0% 或 .
You are about to add 0 people to the discussion. Proceed with caution.
先完成此消息的编辑！
想要评论请 注册 或