wehub-resource-sync
411 lines
14 KiB
Plaintext
411 lines
14 KiB
Plaintext
// Copyright (c) 2025 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.
|
|
|
|
#include "paddle/phi/kernels/masked_fill_grad_kernel.h"
|
|
#include "paddle/phi/kernels/cast_kernel.h"
|
|
#include "paddle/phi/kernels/funcs/masked_fill_utils.h"
|
|
|
|
#include "paddle/phi/backends/gpu/gpu_context.h"
|
|
#include "paddle/phi/core/kernel_registry.h"
|
|
|
|
#include "paddle/phi/backends/gpu/gpu_launch_config.h"
|
|
#include "paddle/phi/backends/gpu/gpu_primitives.h"
|
|
#include "paddle/phi/common/memory_utils.h"
|
|
#include "paddle/phi/common/place.h"
|
|
#include "paddle/phi/core/dense_tensor.h"
|
|
#include "paddle/phi/kernels/empty_kernel.h"
|
|
#include "paddle/phi/kernels/expand_grad_kernel.h"
|
|
#include "paddle/phi/kernels/expand_kernel.h"
|
|
#include "paddle/phi/kernels/full_kernel.h"
|
|
#include "paddle/phi/kernels/funcs/aligned_vector.h"
|
|
#include "paddle/phi/kernels/funcs/common_infer_shape_functions.h"
|
|
#include "paddle/phi/kernels/primitive/kernel_primitives.h"
|
|
#include "paddle/phi/kernels/reduce_sum_kernel.h"
|
|
#include "paddle/phi/kernels/scale_kernel.h"
|
|
#include "paddle/phi/kernels/where_kernel.h"
|
|
|
|
namespace phi {
|
|
template <typename T, int VecSize>
|
|
__global__ void GPUMaskedFillXGradKernel(const T* out_grad,
|
|
const bool* mask,
|
|
const int64_t input_len,
|
|
const int64_t batch_size,
|
|
T* x_grad) {
|
|
int64_t idx = static_cast<int64_t>(blockIdx.x) * blockDim.x + threadIdx.x;
|
|
|
|
if (idx >= (input_len / VecSize)) {
|
|
return;
|
|
}
|
|
|
|
int64_t vec_idx = idx * VecSize;
|
|
int64_t mask_idx = vec_idx / batch_size;
|
|
using VecType = kps::details::VectorType<T, VecSize>;
|
|
const VecType* src = reinterpret_cast<const VecType*>(&out_grad[vec_idx]);
|
|
VecType* x_grad_dst = reinterpret_cast<VecType*>(&x_grad[vec_idx]);
|
|
T set_value[VecSize];
|
|
#pragma unroll
|
|
for (int i = 0; i < VecSize; i++) {
|
|
set_value[i] = 0;
|
|
}
|
|
const VecType* vec_value = reinterpret_cast<const VecType*>(&set_value[0]);
|
|
|
|
if (mask[mask_idx]) {
|
|
*x_grad_dst = *vec_value;
|
|
} else {
|
|
*x_grad_dst = *src;
|
|
}
|
|
}
|
|
|
|
template <typename T, int VecSize>
|
|
__global__ void GPUMaskedFillValueGradKernel(const T* out_grad,
|
|
const bool* mask,
|
|
const int64_t input_len,
|
|
const int64_t batch_size,
|
|
T* value_grad) {
|
|
int64_t idx = static_cast<int64_t>(blockIdx.x) * blockDim.x + threadIdx.x;
|
|
|
|
if (idx >= (input_len / VecSize)) {
|
|
return;
|
|
}
|
|
|
|
int64_t vec_idx = idx * VecSize;
|
|
int64_t mask_idx = vec_idx / batch_size;
|
|
using VecType = kps::details::VectorType<T, VecSize>;
|
|
const VecType* src = reinterpret_cast<const VecType*>(&out_grad[vec_idx]);
|
|
VecType* value_grad_dst = reinterpret_cast<VecType*>(&value_grad[vec_idx]);
|
|
T set_value[VecSize];
|
|
#pragma unroll
|
|
for (int i = 0; i < VecSize; i++) {
|
|
set_value[i] = 0;
|
|
}
|
|
const VecType* vec_value = reinterpret_cast<const VecType*>(&set_value[0]);
|
|
|
|
if (mask[mask_idx]) {
|
|
*value_grad_dst = *src;
|
|
} else {
|
|
*value_grad_dst = *vec_value;
|
|
}
|
|
}
|
|
|
|
template <typename T, int VecSize>
|
|
__global__ void GPUMaskedFillGradKernel(const T* out_grad,
|
|
const bool* mask,
|
|
const int64_t input_len,
|
|
const int64_t batch_size,
|
|
T* x_grad,
|
|
T* value_grad) {
|
|
int64_t idx = static_cast<int64_t>(blockIdx.x) * blockDim.x + threadIdx.x;
|
|
if (idx >= (input_len / VecSize)) {
|
|
return;
|
|
}
|
|
|
|
int64_t vec_idx = idx * VecSize;
|
|
int64_t mask_idx = vec_idx / batch_size;
|
|
using VecType = kps::details::VectorType<T, VecSize>;
|
|
const VecType* src = reinterpret_cast<const VecType*>(&out_grad[vec_idx]);
|
|
VecType* x_grad_dst = reinterpret_cast<VecType*>(&x_grad[vec_idx]);
|
|
VecType* value_grad_dst = reinterpret_cast<VecType*>(&value_grad[vec_idx]);
|
|
T set_value[VecSize];
|
|
#pragma unroll
|
|
for (int i = 0; i < VecSize; i++) {
|
|
set_value[i] = 0;
|
|
}
|
|
const VecType* vec_value = reinterpret_cast<const VecType*>(&set_value[0]);
|
|
|
|
if (mask[mask_idx]) {
|
|
*x_grad_dst = *vec_value;
|
|
*value_grad_dst = *src;
|
|
} else {
|
|
*x_grad_dst = *src;
|
|
*value_grad_dst = *vec_value;
|
|
}
|
|
}
|
|
|
|
template <typename T>
|
|
void DispatchMaskFillGradKernel(const GPUContext& dev_ctx,
|
|
const T* input,
|
|
const bool* mask,
|
|
const int64_t input_len,
|
|
const int64_t batch_size,
|
|
T* x_grad,
|
|
T* value_grad,
|
|
int vec_size,
|
|
const backends::gpu::GpuLaunchConfig& config) {
|
|
auto stream = dev_ctx.stream();
|
|
if (x_grad && value_grad) {
|
|
switch (vec_size) {
|
|
#define CASE_VECSIZE(__Vs) \
|
|
case __Vs: \
|
|
GPUMaskedFillGradKernel<T, __Vs> \
|
|
<<<config.block_per_grid, config.thread_per_block, 0, stream>>>( \
|
|
input, mask, input_len, batch_size, x_grad, value_grad); \
|
|
break;
|
|
CASE_VECSIZE(1)
|
|
CASE_VECSIZE(2)
|
|
CASE_VECSIZE(4)
|
|
CASE_VECSIZE(8)
|
|
#undef CASE_VECSIZE
|
|
default:
|
|
PADDLE_THROW(common::errors::Unimplemented(
|
|
"Unsupported vectorized size: %d", vec_size));
|
|
}
|
|
} else if (x_grad) {
|
|
switch (vec_size) {
|
|
#define CASE_VECSIZE(__Vs) \
|
|
case __Vs: \
|
|
GPUMaskedFillXGradKernel<T, __Vs> \
|
|
<<<config.block_per_grid, config.thread_per_block, 0, stream>>>( \
|
|
input, mask, input_len, batch_size, x_grad); \
|
|
break;
|
|
CASE_VECSIZE(1)
|
|
CASE_VECSIZE(2)
|
|
CASE_VECSIZE(4)
|
|
CASE_VECSIZE(8)
|
|
#undef CASE_VECSIZE
|
|
default:
|
|
PADDLE_THROW(common::errors::Unimplemented(
|
|
"Unsupported vectorized size: %d", vec_size));
|
|
}
|
|
} else if (value_grad) {
|
|
switch (vec_size) {
|
|
#define CASE_VECSIZE(__Vs) \
|
|
case __Vs: \
|
|
GPUMaskedFillValueGradKernel<T, __Vs> \
|
|
<<<config.block_per_grid, config.thread_per_block, 0, stream>>>( \
|
|
input, mask, input_len, batch_size, value_grad); \
|
|
break;
|
|
CASE_VECSIZE(1)
|
|
CASE_VECSIZE(2)
|
|
CASE_VECSIZE(4)
|
|
CASE_VECSIZE(8)
|
|
#undef CASE_VECSIZE
|
|
default:
|
|
PADDLE_THROW(common::errors::Unimplemented(
|
|
"Unsupported vectorized size: %d", vec_size));
|
|
}
|
|
}
|
|
}
|
|
|
|
template <typename T>
|
|
void DispatchMaskFillOneValueGradKernel(
|
|
const GPUContext& dev_ctx,
|
|
const T* input,
|
|
const bool* mask,
|
|
const int64_t input_len,
|
|
const int64_t batch_size,
|
|
T* x_grad,
|
|
int vec_size,
|
|
const backends::gpu::GpuLaunchConfig& config) {
|
|
auto stream = dev_ctx.stream();
|
|
if (x_grad) {
|
|
switch (vec_size) {
|
|
#define CASE_VECSIZE(__Vs) \
|
|
case __Vs: \
|
|
GPUMaskedFillXGradKernel<T, __Vs> \
|
|
<<<config.block_per_grid, config.thread_per_block, 0, stream>>>( \
|
|
input, mask, input_len, batch_size, x_grad); \
|
|
break;
|
|
CASE_VECSIZE(1)
|
|
CASE_VECSIZE(2)
|
|
CASE_VECSIZE(4)
|
|
CASE_VECSIZE(8)
|
|
#undef CASE_VECSIZE
|
|
default:
|
|
PADDLE_THROW(common::errors::Unimplemented(
|
|
"Unsupported vectorized size: %d", vec_size));
|
|
}
|
|
}
|
|
}
|
|
|
|
template <typename T>
|
|
void GPUMaskedFillGrad(const GPUContext& dev_ctx,
|
|
const DenseTensor& out_grad,
|
|
const DenseTensor& mask,
|
|
DenseTensor* x_grad,
|
|
DenseTensor* value_grad) {
|
|
const T* out_grad_data = out_grad.data<T>();
|
|
const bool* mask_data = mask.data<bool>();
|
|
|
|
T* x_grad_data = nullptr;
|
|
T* value_grad_data = nullptr;
|
|
|
|
int64_t input_len = out_grad.numel();
|
|
int64_t mask_len = mask.numel();
|
|
int64_t batch_size = input_len / mask_len;
|
|
|
|
int vec_size = 8;
|
|
vec_size = std::min(GetVectorizedSize(out_grad_data), vec_size);
|
|
if (x_grad && x_grad->initialized()) {
|
|
x_grad_data = x_grad->data<T>();
|
|
vec_size = std::min(GetVectorizedSize(x_grad_data), vec_size);
|
|
}
|
|
|
|
if (value_grad && value_grad->initialized()) {
|
|
value_grad_data = value_grad->data<T>();
|
|
vec_size = std::min(GetVectorizedSize(value_grad_data), vec_size);
|
|
}
|
|
|
|
while (vec_size > 1 && batch_size % vec_size != 0) {
|
|
vec_size /= 2;
|
|
}
|
|
|
|
auto config =
|
|
backends::gpu::GetGpuLaunchConfig1D(dev_ctx, input_len, vec_size);
|
|
|
|
if (value_grad && value_grad->numel() == 1) {
|
|
DispatchMaskFillOneValueGradKernel<T>(dev_ctx,
|
|
out_grad_data,
|
|
mask_data,
|
|
input_len,
|
|
batch_size,
|
|
x_grad_data,
|
|
vec_size,
|
|
config);
|
|
if (value_grad) {
|
|
DenseTensor zero_tensor;
|
|
Full<T, GPUContext>(dev_ctx, out_grad.dims(), T(0.0), &zero_tensor);
|
|
DenseTensor value_grad_tensor;
|
|
value_grad_tensor.set_meta(out_grad.meta());
|
|
WhereKernel<T, GPUContext>(
|
|
dev_ctx, mask, out_grad, zero_tensor, &value_grad_tensor);
|
|
std::vector<int> v_dims(value_grad_tensor.dims().size());
|
|
std::iota(v_dims.begin(), v_dims.end(), 0);
|
|
IntArray v_axis(v_dims);
|
|
SumKernel<T, GPUContext>(dev_ctx,
|
|
value_grad_tensor,
|
|
v_axis,
|
|
value_grad->dtype(),
|
|
false,
|
|
value_grad);
|
|
}
|
|
|
|
} else {
|
|
DispatchMaskFillGradKernel<T>(dev_ctx,
|
|
out_grad_data,
|
|
mask_data,
|
|
input_len,
|
|
batch_size,
|
|
x_grad_data,
|
|
value_grad_data,
|
|
vec_size,
|
|
config);
|
|
}
|
|
}
|
|
|
|
template <typename T, typename Context>
|
|
void MaskedFillGradKernel(const Context& dev_ctx,
|
|
const DenseTensor& x,
|
|
const DenseTensor& mask,
|
|
const DenseTensor& value UNUSED,
|
|
const DenseTensor& out_grad,
|
|
DenseTensor* x_grad,
|
|
DenseTensor* v_grad) {
|
|
if (out_grad.numel() == 0 || mask.numel() == 0) {
|
|
// x shape [2, 1, 3], mask shape [2, 0, 3], x_grad shape [2, 1, 3]
|
|
if (x_grad) {
|
|
Full<T, Context>(dev_ctx, x_grad->dims(), 0, x_grad);
|
|
}
|
|
if (v_grad) {
|
|
Full<T, Context>(dev_ctx, v_grad->dims(), 0, v_grad);
|
|
}
|
|
return;
|
|
}
|
|
auto out_grad_dims = out_grad.dims();
|
|
auto x_dims = x.dims();
|
|
auto mask_dims = mask.dims();
|
|
DenseTensor mask_expand;
|
|
DenseTensor x_grad_expand;
|
|
DenseTensor v_grad_expand;
|
|
bool expand_x = false;
|
|
bool expand_v = false;
|
|
auto expanded_size =
|
|
vectorize(funcs::BroadcastTwoDims(x_dims, mask_dims, -1));
|
|
auto expanded_dims = make_ddim(expanded_size);
|
|
bool flag = funcs::CanDispatchMaskFillShortcut(out_grad_dims, mask_dims);
|
|
if (expanded_dims != x_dims) flag = false;
|
|
if (v_grad && v_grad->dims() != expanded_dims && v_grad->numel() != 1)
|
|
flag = false;
|
|
|
|
if (x_grad) {
|
|
dev_ctx.template Alloc<T>(x_grad);
|
|
}
|
|
if (v_grad) {
|
|
dev_ctx.template Alloc<T>(v_grad);
|
|
}
|
|
if (flag) {
|
|
GPUMaskedFillGrad<T>(dev_ctx, out_grad, mask, x_grad, v_grad);
|
|
return;
|
|
}
|
|
|
|
if (mask.dims() != expanded_dims) {
|
|
ExpandKernel<bool, Context>(
|
|
dev_ctx, mask, IntArray(expanded_size), &mask_expand);
|
|
} else {
|
|
mask_expand = mask;
|
|
}
|
|
|
|
auto mask_size = mask_expand.numel();
|
|
if (mask_size <= 0) return;
|
|
|
|
if (x_grad) {
|
|
if (x_grad->dims() != expanded_dims) {
|
|
x_grad_expand = Empty<T, Context>(dev_ctx, IntArray(expanded_size));
|
|
expand_x = true;
|
|
} else {
|
|
x_grad_expand = *x_grad;
|
|
}
|
|
}
|
|
|
|
if (v_grad) {
|
|
if (v_grad->dims() != expanded_dims && v_grad->numel() != 1) {
|
|
v_grad_expand = Empty<T, Context>(dev_ctx, IntArray(expanded_size));
|
|
expand_v = true;
|
|
} else {
|
|
v_grad_expand = *v_grad;
|
|
}
|
|
}
|
|
|
|
GPUMaskedFillGrad<T>(
|
|
dev_ctx, out_grad, mask_expand, &x_grad_expand, &v_grad_expand);
|
|
|
|
if (expand_x) {
|
|
ExpandGradKernel<T, Context>(
|
|
dev_ctx, x, x_grad_expand, IntArray(expanded_size), x_grad);
|
|
}
|
|
if (expand_v) {
|
|
ExpandGradKernel<T, Context>(
|
|
dev_ctx, value, v_grad_expand, IntArray(expanded_size), v_grad);
|
|
}
|
|
}
|
|
} // namespace phi
|
|
|
|
PD_REGISTER_KERNEL(masked_fill_grad,
|
|
GPU,
|
|
ALL_LAYOUT,
|
|
phi::MaskedFillGradKernel,
|
|
bool,
|
|
float,
|
|
double,
|
|
int,
|
|
int8_t,
|
|
int64_t,
|
|
int16_t,
|
|
uint8_t,
|
|
phi::float16,
|
|
phi::bfloat16,
|
|
phi::complex64,
|
|
phi::complex128) {
|
|
kernel->InputAt(1).SetDataType(phi::DataType::BOOL);
|
|
}
|