Skip to content

P
Paddle

BaiXuePrincess / Paddle
与 Fork 源项目一致

Fork自 PaddlePaddle / Paddle

通知 1
Star 1
Fork 0
P
Paddle
提交 9171f737 编写于 作者: D danleifeng 提交者： gongweibao
浏览文件
操作

fix fp16 grid_size for size=1; test=develop (#20812)

上级 cd1c4043
隐藏空白更改
内联 并排
Showing with 24 addition and 20 deletion
paddle/fluid/operators/elementwise/elementwise_add_op.cu
浏览文件 @ 9171f737
...@@ -40,8 +40,9 @@ struct SameDimsElemwiseAdd<platform::CUDADeviceContext, platform::float16> { ...@@ -40,8 +40,9 @@ struct SameDimsElemwiseAdd<platform::CUDADeviceContext, platform::float16> {
const framework::Tensor* x, const framework::Tensor* y, const framework::Tensor* x, const framework::Tensor* y,
framework::Tensor* z) { framework::Tensor* z) {
auto size = x->numel(); auto size = x->numel();
dim3 gird_size = dim3( dim3 grid_size = dim3(((size + 1) / 2 + PADDLE_CUDA_THREAD_SIZE - 1) /
(size / 2 + PADDLE_CUDA_THREAD_SIZE - 1) / PADDLE_CUDA_THREAD_SIZE, 1); PADDLE_CUDA_THREAD_SIZE,
1);
dim3 block_size = dim3(PADDLE_CUDA_THREAD_SIZE, 1); dim3 block_size = dim3(PADDLE_CUDA_THREAD_SIZE, 1);
const half* x2 = const half* x2 =
reinterpret_cast<const half*>(x->data<platform::float16>()); reinterpret_cast<const half*>(x->data<platform::float16>());
...@@ -49,7 +50,7 @@ struct SameDimsElemwiseAdd<platform::CUDADeviceContext, platform::float16> { ...@@ -49,7 +50,7 @@ struct SameDimsElemwiseAdd<platform::CUDADeviceContext, platform::float16> {
reinterpret_cast<const half*>(y->data<platform::float16>()); reinterpret_cast<const half*>(y->data<platform::float16>());
half* z2 = reinterpret_cast<half*>(z->data<platform::float16>()); half* z2 = reinterpret_cast<half*>(z->data<platform::float16>());
SameDimsElemwiseAddCUDAKernel<<< SameDimsElemwiseAddCUDAKernel<<<
gird_size, block_size, 0, grid_size, block_size, 0,
ctx.template device_context<platform::CUDADeviceContext>().stream()>>>( ctx.template device_context<platform::CUDADeviceContext>().stream()>>>(
x2, y2, z2, size); x2, y2, z2, size);
} }
...@@ -78,10 +79,10 @@ elementwise_add_grad(const framework::ExecutionContext& ctx, ...@@ -78,10 +79,10 @@ elementwise_add_grad(const framework::ExecutionContext& ctx,
framework::Tensor* dy) { framework::Tensor* dy) {
dim3 block_size = dim3(PADDLE_CUDA_THREAD_SIZE, 1); dim3 block_size = dim3(PADDLE_CUDA_THREAD_SIZE, 1);
auto size = x->numel(); auto size = x->numel();
dim3 gird_size = dim3 grid_size =
dim3((size + PADDLE_CUDA_THREAD_SIZE - 1) / PADDLE_CUDA_THREAD_SIZE, 1); dim3((size + PADDLE_CUDA_THREAD_SIZE - 1) / PADDLE_CUDA_THREAD_SIZE, 1);
SimpleElemwiseAddGradCUDAKernel< SimpleElemwiseAddGradCUDAKernel<
T><<<gird_size, block_size, 0, T><<<grid_size, block_size, 0,
ctx.template device_context<plat::CUDADeviceContext>().stream()>>>( ctx.template device_context<plat::CUDADeviceContext>().stream()>>>(
dout->data<T>(), size, dx->mutable_data<T>(ctx.GetPlace()), dout->data<T>(), size, dx->mutable_data<T>(ctx.GetPlace()),
dy->mutable_data<T>(ctx.GetPlace())); dy->mutable_data<T>(ctx.GetPlace()));
......
paddle/fluid/operators/elementwise/elementwise_div_op.cu
浏览文件 @ 9171f737
...@@ -41,8 +41,9 @@ struct SameDimsElemwiseDiv<platform::CUDADeviceContext, platform::float16> { ...@@ -41,8 +41,9 @@ struct SameDimsElemwiseDiv<platform::CUDADeviceContext, platform::float16> {
const framework::Tensor* x, const framework::Tensor* y, const framework::Tensor* x, const framework::Tensor* y,
framework::Tensor* z) { framework::Tensor* z) {
auto size = x->numel(); auto size = x->numel();
dim3 gird_size = dim3( dim3 grid_size = dim3(((size + 1) / 2 + PADDLE_CUDA_THREAD_SIZE - 1) /
(size / 2 + PADDLE_CUDA_THREAD_SIZE - 1) / PADDLE_CUDA_THREAD_SIZE, 1); PADDLE_CUDA_THREAD_SIZE,
1);
dim3 block_size = dim3(PADDLE_CUDA_THREAD_SIZE, 1); dim3 block_size = dim3(PADDLE_CUDA_THREAD_SIZE, 1);
const half* x2 = const half* x2 =
reinterpret_cast<const half*>(x->data<platform::float16>()); reinterpret_cast<const half*>(x->data<platform::float16>());
...@@ -50,7 +51,7 @@ struct SameDimsElemwiseDiv<platform::CUDADeviceContext, platform::float16> { ...@@ -50,7 +51,7 @@ struct SameDimsElemwiseDiv<platform::CUDADeviceContext, platform::float16> {
reinterpret_cast<const half*>(y->data<platform::float16>()); reinterpret_cast<const half*>(y->data<platform::float16>());
half* z2 = reinterpret_cast<half*>(z->data<platform::float16>()); half* z2 = reinterpret_cast<half*>(z->data<platform::float16>());
SameDimsElemwiseDivCUDAKernel<<< SameDimsElemwiseDivCUDAKernel<<<
gird_size, block_size, 0, grid_size, block_size, 0,
ctx.template device_context<platform::CUDADeviceContext>().stream()>>>( ctx.template device_context<platform::CUDADeviceContext>().stream()>>>(
x2, y2, z2, size); x2, y2, z2, size);
} }
...@@ -82,10 +83,10 @@ elementwise_div_grad(const framework::ExecutionContext& ctx, ...@@ -82,10 +83,10 @@ elementwise_div_grad(const framework::ExecutionContext& ctx,
framework::Tensor* dy) { framework::Tensor* dy) {
dim3 block_size = dim3(PADDLE_CUDA_THREAD_SIZE, 1); dim3 block_size = dim3(PADDLE_CUDA_THREAD_SIZE, 1);
auto size = x->numel(); auto size = x->numel();
dim3 gird_size = dim3 grid_size =
dim3((size + PADDLE_CUDA_THREAD_SIZE - 1) / PADDLE_CUDA_THREAD_SIZE, 1); dim3((size + PADDLE_CUDA_THREAD_SIZE - 1) / PADDLE_CUDA_THREAD_SIZE, 1);
SimpleElemwiseDivGradCUDAKernel< SimpleElemwiseDivGradCUDAKernel<
T><<<gird_size, block_size, 0, T><<<grid_size, block_size, 0,
ctx.template device_context<plat::CUDADeviceContext>().stream()>>>( ctx.template device_context<plat::CUDADeviceContext>().stream()>>>(
x->data<T>(), y->data<T>(), out->data<T>(), dout->data<T>(), size, x->data<T>(), y->data<T>(), out->data<T>(), dout->data<T>(), size,
dx->mutable_data<T>(ctx.GetPlace()), dy->mutable_data<T>(ctx.GetPlace())); dx->mutable_data<T>(ctx.GetPlace()), dy->mutable_data<T>(ctx.GetPlace()));
......
paddle/fluid/operators/elementwise/elementwise_mul_op.cu
浏览文件 @ 9171f737
...@@ -41,8 +41,9 @@ struct SameDimsElemwiseMul<platform::CUDADeviceContext, platform::float16> { ...@@ -41,8 +41,9 @@ struct SameDimsElemwiseMul<platform::CUDADeviceContext, platform::float16> {
const framework::Tensor* x, const framework::Tensor* y, const framework::Tensor* x, const framework::Tensor* y,
framework::Tensor* z) { framework::Tensor* z) {
auto size = x->numel(); auto size = x->numel();
dim3 gird_size = dim3( dim3 grid_size = dim3(((size + 1) / 2 + PADDLE_CUDA_THREAD_SIZE - 1) /
(size / 2 + PADDLE_CUDA_THREAD_SIZE - 1) / PADDLE_CUDA_THREAD_SIZE, 1); PADDLE_CUDA_THREAD_SIZE,
1);
dim3 block_size = dim3(PADDLE_CUDA_THREAD_SIZE, 1); dim3 block_size = dim3(PADDLE_CUDA_THREAD_SIZE, 1);
const half* x2 = const half* x2 =
reinterpret_cast<const half*>(x->data<platform::float16>()); reinterpret_cast<const half*>(x->data<platform::float16>());
...@@ -50,7 +51,7 @@ struct SameDimsElemwiseMul<platform::CUDADeviceContext, platform::float16> { ...@@ -50,7 +51,7 @@ struct SameDimsElemwiseMul<platform::CUDADeviceContext, platform::float16> {
reinterpret_cast<const half*>(y->data<platform::float16>()); reinterpret_cast<const half*>(y->data<platform::float16>());
half* z2 = reinterpret_cast<half*>(z->data<platform::float16>()); half* z2 = reinterpret_cast<half*>(z->data<platform::float16>());
SameDimsElemwiseMulCUDAKernel<<< SameDimsElemwiseMulCUDAKernel<<<
gird_size, block_size, 0, grid_size, block_size, 0,
ctx.template device_context<platform::CUDADeviceContext>().stream()>>>( ctx.template device_context<platform::CUDADeviceContext>().stream()>>>(
x2, y2, z2, size); x2, y2, z2, size);
} }
...@@ -82,10 +83,10 @@ elementwise_mul_grad(const framework::ExecutionContext& ctx, ...@@ -82,10 +83,10 @@ elementwise_mul_grad(const framework::ExecutionContext& ctx,
framework::Tensor* dy) { framework::Tensor* dy) {
dim3 block_size = dim3(PADDLE_CUDA_THREAD_SIZE, 1); dim3 block_size = dim3(PADDLE_CUDA_THREAD_SIZE, 1);
auto size = x->numel(); auto size = x->numel();
dim3 gird_size = dim3 grid_size =
dim3((size + PADDLE_CUDA_THREAD_SIZE - 1) / PADDLE_CUDA_THREAD_SIZE, 1); dim3((size + PADDLE_CUDA_THREAD_SIZE - 1) / PADDLE_CUDA_THREAD_SIZE, 1);
SimpleElemwiseMulGradCUDAKernel< SimpleElemwiseMulGradCUDAKernel<
T><<<gird_size, block_size, 0, T><<<grid_size, block_size, 0,
ctx.template device_context<plat::CUDADeviceContext>().stream()>>>( ctx.template device_context<plat::CUDADeviceContext>().stream()>>>(
x->data<T>(), y->data<T>(), out->data<T>(), dout->data<T>(), size, x->data<T>(), y->data<T>(), out->data<T>(), dout->data<T>(), size,
dx->mutable_data<T>(ctx.GetPlace()), dy->mutable_data<T>(ctx.GetPlace())); dx->mutable_data<T>(ctx.GetPlace()), dy->mutable_data<T>(ctx.GetPlace()));
......
paddle/fluid/operators/elementwise/elementwise_sub_op.cu
浏览文件 @ 9171f737
...@@ -41,8 +41,9 @@ struct SameDimsElemwiseSub<platform::CUDADeviceContext, platform::float16> { ...@@ -41,8 +41,9 @@ struct SameDimsElemwiseSub<platform::CUDADeviceContext, platform::float16> {
const framework::Tensor* x, const framework::Tensor* y, const framework::Tensor* x, const framework::Tensor* y,
framework::Tensor* z) { framework::Tensor* z) {
auto size = x->numel(); auto size = x->numel();
dim3 gird_size = dim3( dim3 grid_size = dim3(((size + 1) / 2 + PADDLE_CUDA_THREAD_SIZE - 1) /
(size / 2 + PADDLE_CUDA_THREAD_SIZE - 1) / PADDLE_CUDA_THREAD_SIZE, 1); PADDLE_CUDA_THREAD_SIZE,
1);
dim3 block_size = dim3(PADDLE_CUDA_THREAD_SIZE, 1); dim3 block_size = dim3(PADDLE_CUDA_THREAD_SIZE, 1);
const half* x2 = const half* x2 =
reinterpret_cast<const half*>(x->data<platform::float16>()); reinterpret_cast<const half*>(x->data<platform::float16>());
...@@ -50,7 +51,7 @@ struct SameDimsElemwiseSub<platform::CUDADeviceContext, platform::float16> { ...@@ -50,7 +51,7 @@ struct SameDimsElemwiseSub<platform::CUDADeviceContext, platform::float16> {
reinterpret_cast<const half*>(y->data<platform::float16>()); reinterpret_cast<const half*>(y->data<platform::float16>());
half* z2 = reinterpret_cast<half*>(z->data<platform::float16>()); half* z2 = reinterpret_cast<half*>(z->data<platform::float16>());
SameDimsElemwiseSubCUDAKernel<<< SameDimsElemwiseSubCUDAKernel<<<
gird_size, block_size, 0, grid_size, block_size, 0,
ctx.template device_context<platform::CUDADeviceContext>().stream()>>>( ctx.template device_context<platform::CUDADeviceContext>().stream()>>>(
x2, y2, z2, size); x2, y2, z2, size);
} }
...@@ -79,10 +80,10 @@ elementwise_sub_grad(const framework::ExecutionContext& ctx, ...@@ -79,10 +80,10 @@ elementwise_sub_grad(const framework::ExecutionContext& ctx,
framework::Tensor* dy) { framework::Tensor* dy) {
dim3 block_size = dim3(PADDLE_CUDA_THREAD_SIZE, 1); dim3 block_size = dim3(PADDLE_CUDA_THREAD_SIZE, 1);
auto size = x->numel(); auto size = x->numel();
dim3 gird_size = dim3 grid_size =
dim3((size + PADDLE_CUDA_THREAD_SIZE - 1) / PADDLE_CUDA_THREAD_SIZE, 1); dim3((size + PADDLE_CUDA_THREAD_SIZE - 1) / PADDLE_CUDA_THREAD_SIZE, 1);
SimpleElemwiseSubGradCUDAKernel< SimpleElemwiseSubGradCUDAKernel<
T><<<gird_size, block_size, 0, T><<<grid_size, block_size, 0,
ctx.template device_context<plat::CUDADeviceContext>().stream()>>>( ctx.template device_context<plat::CUDADeviceContext>().stream()>>>(
dout->data<T>(), size, dx->mutable_data<T>(ctx.GetPlace()), dout->data<T>(), size, dx->mutable_data<T>(ctx.GetPlace()),
dy->mutable_data<T>(ctx.GetPlace())); dy->mutable_data<T>(ctx.GetPlace()));
......
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