Skip to content

[ILUVATAR_GPU] Fix bug of softmax kernel #2019

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

Open
wants to merge 1 commit into
base: develop
Choose a base branch
from
Open

[ILUVATAR_GPU] Fix bug of softmax kernel #2019

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
1 commit
Select commit
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
3 changes: 2 additions & 1 deletion backends/iluvatar_gpu/CMakeLists.txt
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -219,6 +219,7 @@ file(
${PADDLE_SOURCE_DIR}/paddle/phi/kernels/gpu/tril_triu_kernel.cu
${PADDLE_SOURCE_DIR}/paddle/phi/kernels/gpu/unbind_kernel.cu
${PADDLE_SOURCE_DIR}/paddle/phi/kernels/gpu/uniform_kernel.cu
${PADDLE_SOURCE_DIR}/paddle/phi/kernels/gpu/unique_consecutive_kernel.cu
${PADDLE_SOURCE_DIR}/paddle/phi/kernels/gpu/where_kernel.cu
# kernels/selected_rows
${PADDLE_SOURCE_DIR}/paddle/phi/kernels/selected_rows/gpu/adamw_kernel.cu
Expand Down Expand Up @@ -932,7 +933,7 @@ file(
${PADDLE_SOURCE_DIR}/paddle/phi/kernels/array_kernel.cc)

set(CUDA_SRCS ${CUDA_SRCS1} ${CUDA_SRCS2})
list(REMOVE_DUPLICATES CUDA_SRCS1)
list(REMOVE_DUPLICATES CUDA_SRCS)

list(
REMOVE_ITEM
Expand Down
5 changes: 3 additions & 2 deletions backends/iluvatar_gpu/kernels/cuda_kernels/cross_entropy_grad_kernel.cu
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -22,7 +22,7 @@ limitations under the License. */
namespace cub = hipcub;
#endif

#include "../gpudnn/softmax_gpudnn.h"
#include "kernels/gpudnn/softmax_gpudnn.h"
#include "paddle/phi/backends/gpu/gpu_device_function.h"
#include "paddle/phi/backends/gpu/gpu_dnn.h"
#include "paddle/phi/common/amp_type_traits.h"
Expand Down Expand Up @@ -277,4 +277,5 @@ PD_REGISTER_PLUGIN_KERNEL(cross_entropy_with_softmax_grad,
ALL_LAYOUT,
phi::CrossEntropyWithSoftmaxGradKernel,
float,
phi::dtype::float16) {}
phi::dtype::float16,
phi::dtype::bfloat16) {}
5 changes: 3 additions & 2 deletions backends/iluvatar_gpu/kernels/cuda_kernels/cross_entropy_kernel.cu
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -23,7 +23,7 @@ limitations under the License. */
namespace cub = hipcub;
#endif

#include "../gpudnn/softmax_gpudnn.h"
#include "kernels/gpudnn/softmax_gpudnn.h"
#include "paddle/phi/backends/gpu/gpu_device_function.h"
#include "paddle/phi/backends/gpu/gpu_dnn.h"
#include "paddle/phi/common/amp_type_traits.h"
Expand Down Expand Up @@ -1412,4 +1412,5 @@ PD_REGISTER_PLUGIN_KERNEL(cross_entropy_with_softmax,
ALL_LAYOUT,
phi::CrossEntropyWithSoftmaxKernel,
float,
phi::dtype::float16) {}
phi::dtype::float16,
phi::dtype::bfloat16) {}
24 changes: 19 additions & 5 deletions backends/iluvatar_gpu/kernels/cuda_kernels/index_elementwise_put_grad_kernel_register.cu
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -13,6 +13,7 @@
// limitations under the License.

#include "paddle/phi/core/kernel_registry.h"
#include "paddle/phi/kernels/gpu/index_elementwise_put_grad_kernel.cu" // NOLINT
#include "paddle/phi/kernels/index_elementwise_put_grad_kernel.h"

PD_CUSTOM_KERNEL_REGISTER(index_elementwise_put_grad,
Expand All @@ -21,13 +22,26 @@ PD_CUSTOM_KERNEL_REGISTER(index_elementwise_put_grad,
phi::IndexElementwisePutGradKernel,
bool,
float,
double,
int,
int8_t,
int64_t,
int16_t,
uint8_t,
phi::dtype::float16,
phi::dtype::bfloat16,
phi::dtype::complex<float>,
phi::dtype::complex<double>) {}
phi::float16,
phi::bfloat16,
phi::complex64) {}

PD_CUSTOM_KERNEL_REGISTER(index_elementwise_put_with_tensor_grad,
iluvatar_gpu,
ALL_LAYOUT,
phi::IndexElementwisePutWithTensorGradKernel,
bool,
float,
int,
int8_t,
int64_t,
int16_t,
uint8_t,
phi::float16,
phi::bfloat16,
phi::complex64) {}
24 changes: 19 additions & 5 deletions backends/iluvatar_gpu/kernels/cuda_kernels/index_elementwise_put_kernel_register.cu
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -13,6 +13,7 @@
// limitations under the License.

#include "paddle/phi/core/kernel_registry.h"
#include "paddle/phi/kernels/gpu/index_elementwise_put_kernel.cu" // NOLINT
#include "paddle/phi/kernels/index_elementwise_put_kernel.h"

PD_CUSTOM_KERNEL_REGISTER(index_elementwise_put,
Expand All @@ -21,13 +22,26 @@ PD_CUSTOM_KERNEL_REGISTER(index_elementwise_put,
phi::IndexElementwisePutKernel,
bool,
float,
double,
int,
int8_t,
int64_t,
int16_t,
uint8_t,
phi::dtype::float16,
phi::dtype::bfloat16,
phi::dtype::complex<float>,
phi::dtype::complex<double>) {}
phi::float16,
phi::bfloat16,
phi::complex64) {}

PD_CUSTOM_KERNEL_REGISTER(index_elementwise_put_with_tensor,
iluvatar_gpu,
ALL_LAYOUT,
phi::IndexElementwisePutWithTensorKernel,
bool,
float,
int,
int8_t,
int64_t,
int16_t,
uint8_t,
phi::float16,
phi::bfloat16,
phi::complex64) {}
2 changes: 1 addition & 1 deletion backends/iluvatar_gpu/kernels/cuda_kernels/log_softmax_grad_kernel.cu
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -12,7 +12,7 @@
// See the License for the specific language governing permissions and
// limitations under the License.

#include "../gpudnn/softmax_gpudnn.h"
#include "kernels/gpudnn/softmax_gpudnn.h"
#include "paddle/phi/backends/gpu/gpu_context.h"
#include "paddle/phi/core/kernel_registry.h"
#include "paddle/phi/kernels/funcs/math_function.h"
Expand Down
2 changes: 1 addition & 1 deletion backends/iluvatar_gpu/kernels/cuda_kernels/log_softmax_kernel.cu
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -12,7 +12,7 @@
// See the License for the specific language governing permissions and
// limitations under the License.

#include "../gpudnn/softmax_gpudnn.h"
#include "kernels/gpudnn/softmax_gpudnn.h"
#include "paddle/phi/backends/gpu/gpu_context.h"
#include "paddle/phi/core/kernel_registry.h"
#include "paddle/phi/kernels/funcs/math_function.h"
Expand Down
2 changes: 1 addition & 1 deletion backends/iluvatar_gpu/kernels/cuda_kernels/softmax_grad_kernel.cu
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -12,7 +12,7 @@ WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
See the License for the specific language governing permissions and
limitations under the License. */

#include "../gpudnn/softmax_gpudnn.h"
#include "kernels/gpudnn/softmax_gpudnn.h"
#include "paddle/phi/backends/gpu/gpu_context.h"
#include "paddle/phi/core/kernel_registry.h"
#include "paddle/phi/kernels/funcs/math_function.h"
Expand Down
2 changes: 1 addition & 1 deletion backends/iluvatar_gpu/kernels/cuda_kernels/softmax_kernel.cu
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -12,7 +12,7 @@ WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
See the License for the specific language governing permissions and
limitations under the License. */

#include "../gpudnn/softmax_gpudnn.h"
#include "kernels/gpudnn/softmax_gpudnn.h"
#include "paddle/phi/backends/gpu/gpu_context.h"
#include "paddle/phi/core/kernel_registry.h"
#include "paddle/phi/kernels/funcs/math_function.h"
Expand Down
29 changes: 29 additions & 0 deletions backends/iluvatar_gpu/kernels/cuda_kernels/unique_consecutive_kernel_register.cc
View file
Open in desktop
Original file line number Diff line number Diff line change
@@ -0,0 +1,29 @@
// Copyright (c) 2022 PaddlePaddle Authors. All Rights Reserved.
//
// Licensed under the Apache License, Version 2.0 (the "License");
// you may not use this file except in compliance with the License.
// You may obtain a copy of the License at
//
// http://www.apache.org/licenses/LICENSE-2.0
//
// Unless required by applicable law or agreed to in writing, software
// distributed under the License is distributed on an "AS IS" BASIS,
// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
// See the License for the specific language governing permissions and
// limitations under the License.

#pragma once

#include "paddle/phi/core/kernel_registry.h"
#include "paddle/phi/kernels/unique_consecutive_kernel.h"

PD_CUSTOM_KERNEL_REGISTER(unique_consecutive,
iluvatar_gpu,
ALL_LAYOUT,
phi::UniqueConsecutiveKernel,
float,
int32_t,
int64_t) {
kernel->OutputAt(1).SetDataType(kernel_key.dtype());
kernel->OutputAt(2).SetDataType(kernel_key.dtype());
}
11 changes: 9 additions & 2 deletions backends/iluvatar_gpu/kernels/gpudnn/softmax_gpudnn.h
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -30,6 +30,9 @@ limitations under the License. */

#define MATRIX_SOFTMAX_ALIGN_BYTES 16
#define MATRIX_SOFTMAX_THRESHOLD 100000
#ifdef PADDLE_WITH_COREX
#define MAX_YZ_DIM_SIZE 65535
#endif

namespace phi {

Expand Down Expand Up @@ -845,6 +848,10 @@ static void GetGridDim(
grid_x = std::min(grid_x, max_num_blocks);
int grid_y = (max_num_blocks + grid_x - 1) / grid_x;
grid_y = std::min(grid_y, high_dim);
#ifdef PADDLE_WITH_COREX
grid_y = std::min(grid_y,
std::max(MAX_YZ_DIM_SIZE / static_cast<int>(block.y), 1));
#endif
grid->x = grid_x;
grid->y = grid_y;
}
Expand Down Expand Up @@ -1211,7 +1218,7 @@ void SoftmaxForwardCUDAKernelDriverImpl(const GPUContext& dev_ctx,
IndexType dim = tensor_dims[1];
int D = tensor_dims[2];

if (D == 1) {
if (D == 1 && x.dtype() != phi::DataType::BFLOAT16) {
if (!UseCudnnSoftmax<T>(dev_ctx, dim, true)) {
int dim_log2 = static_cast<int>(Log2Ceil(dim));
IndexType dim_ceil = 1 << dim_log2;
Expand Down Expand Up @@ -1278,7 +1285,7 @@ void SoftmaxBackwardCUDAKernelDriver(const GPUContext& dev_ctx,
int dim = tensor_dims[1];
int D = tensor_dims[2];

if (D == 1) {
if (D == 1 && out.dtype() != phi::DataType::BFLOAT16) {
if (!UseCudnnSoftmax<T>(dev_ctx, dim, true)) {
int dim_log2 = Log2Ceil(dim);
int dim_ceil = 1 << dim_log2;
Expand Down
4 changes: 4 additions & 0 deletions backends/iluvatar_gpu/runtime/runtime.cc
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -555,6 +555,10 @@ C_Status Allocate(const C_Device device, void **ptr, size_t size) {
err = cudaMalloc(ptr, size);
if (err != cudaSuccess) {
*ptr = NULL;
if (err == cudaErrorMemoryAllocation) {
VLOG(0) << "[RUNTIME] Failed to alloc hbm, size: " << size
<< ", out of memory.";
}
return C_ERROR;
}

Expand Down