304 lines
13 KiB
Plaintext
304 lines
13 KiB
Plaintext
// 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.
|
|
|
|
#include "paddle/phi/kernels/layer_norm_grad_kernel.h"
|
|
|
|
#include "paddle/common/flags.h"
|
|
#include "paddle/phi/backends/gpu/gpu_context.h"
|
|
#include "paddle/phi/core/kernel_registry.h"
|
|
#include "paddle/phi/kernels/cast_kernel.h"
|
|
#include "paddle/phi/kernels/full_kernel.h"
|
|
#include "paddle/phi/kernels/funcs/layer_norm_impl.cu.h"
|
|
#include "paddle/phi/kernels/funcs/layer_norm_util.h"
|
|
#if defined(PADDLE_WITH_CUDA) && !defined(PADDLE_WITH_HIP) && !defined(_WIN32)
|
|
#include "paddle/phi/kernels/funcs/fast_ln_v2.h"
|
|
#endif
|
|
#ifdef PADDLE_WITH_CUDA
|
|
#include "paddle/phi/kernels/gpu/rms_norm_cuda_kernel.h"
|
|
#endif
|
|
COMMON_DECLARE_bool(use_apex_layer_norm_kernel);
|
|
|
|
namespace phi {
|
|
enum class LayerNormGadKernelVariant { FAST_LN_V2, GENERIC };
|
|
static inline LayerNormGadKernelVariant LayerNormGradKernelDispatch(
|
|
const DataType weight_type,
|
|
const DataType input_type,
|
|
const DataType output_type,
|
|
const DataType compute_type,
|
|
const uint32_t hidden_size,
|
|
const int64_t x_numel,
|
|
const DenseTensor* scale,
|
|
const DenseTensor* bias) {
|
|
#if defined(PADDLE_WITH_CUDA) && !defined(PADDLE_WITH_HIP) && !defined(_WIN32)
|
|
if (FLAGS_use_apex_layer_norm_kernel) {
|
|
if (funcs::fast_ln_v2::has_fast_ln_v2_bwd_kernel(
|
|
weight_type, input_type, output_type, compute_type, hidden_size)) {
|
|
return LayerNormGadKernelVariant::FAST_LN_V2;
|
|
}
|
|
PADDLE_THROW(common::errors::InvalidArgument(
|
|
"FLAGS_use_apex_layer_norm_kernel requires inputs supported by "
|
|
"fast_ln_v2 backward kernel."));
|
|
}
|
|
if (FLAGS_use_accuracy_compatible_kernel) {
|
|
return LayerNormGadKernelVariant::GENERIC;
|
|
}
|
|
if (scale != nullptr && bias != nullptr && input_type != DataType::FLOAT32 &&
|
|
hidden_size != 4096 && hidden_size > 1024 && hidden_size <= 10240 &&
|
|
x_numel <= std::numeric_limits<uint32_t>::max()) {
|
|
// using fast_ln_v2 only sm > 70 and x_numel <= uint32_max
|
|
auto prop = funcs::fast_ln_v2::GetDeviceProp();
|
|
if (prop->major > 7 &&
|
|
funcs::fast_ln_v2::has_fast_ln_v2_bwd_kernel(
|
|
weight_type, input_type, output_type, compute_type, hidden_size)) {
|
|
return LayerNormGadKernelVariant::FAST_LN_V2;
|
|
}
|
|
}
|
|
#endif
|
|
return LayerNormGadKernelVariant::GENERIC;
|
|
}
|
|
|
|
template <typename T, typename Context>
|
|
void LayerNormGradKernel(const Context& dev_ctx,
|
|
const DenseTensor& x,
|
|
const optional<DenseTensor>& scale_opt,
|
|
const optional<DenseTensor>& bias_opt,
|
|
const DenseTensor& mean,
|
|
const DenseTensor& variance,
|
|
const DenseTensor& out_grad,
|
|
double epsilon,
|
|
int begin_norm_axis,
|
|
DenseTensor* x_grad,
|
|
DenseTensor* scale_grad,
|
|
DenseTensor* bias_grad) {
|
|
if (x.numel() == 0) {
|
|
dev_ctx.template Alloc<T>(x_grad);
|
|
if (scale_grad) {
|
|
Full<T, Context>(dev_ctx, scale_grad->dims(), 0, scale_grad);
|
|
if (scale_opt.get_ptr() && x.dtype() != scale_opt.get().dtype()) {
|
|
CastKernel<T, Context>(
|
|
dev_ctx, *scale_grad, scale_opt.get().dtype(), scale_grad);
|
|
}
|
|
}
|
|
if (bias_grad) {
|
|
Full<T, Context>(dev_ctx, bias_grad->dims(), 0, bias_grad);
|
|
if (bias_opt.get_ptr() && x.dtype() != bias_opt.get().dtype()) {
|
|
CastKernel<T, Context>(
|
|
dev_ctx, *bias_grad, bias_opt.get().dtype(), bias_grad);
|
|
}
|
|
}
|
|
return;
|
|
}
|
|
using U = funcs::LayerNormParamType<T>;
|
|
// d_x, d_scale, d_bias may be nullptr
|
|
auto* d_x = x_grad;
|
|
auto* d_scale = scale_grad;
|
|
auto* d_bias = bias_grad;
|
|
|
|
auto* scale = scale_opt.get_ptr();
|
|
auto* bias = bias_opt.get_ptr();
|
|
auto* d_y = &out_grad;
|
|
|
|
const auto& x_dims = x.dims();
|
|
auto matrix_dim = common::flatten_to_2d(x_dims, begin_norm_axis);
|
|
int64_t batch_size = static_cast<int64_t>(matrix_dim[0]);
|
|
int64_t feature_size = static_cast<int64_t>(matrix_dim[1]);
|
|
|
|
auto* x_data = x.data<T>();
|
|
auto* d_y_data = d_y->data<T>();
|
|
|
|
auto* mean_data = mean.data<U>();
|
|
auto* var_data = variance.data<U>();
|
|
|
|
auto* d_x_data = (d_x == nullptr ? nullptr : dev_ctx.template Alloc<T>(d_x));
|
|
|
|
auto x_dtype = x.dtype();
|
|
|
|
DataType scale_bias_dtype;
|
|
if (scale != nullptr) {
|
|
scale_bias_dtype = scale->dtype();
|
|
} else {
|
|
// FIXME(zengjinle): do not find a better way to get the right
|
|
// data type of the d_scale and d_bias if scale == nullptr.
|
|
if (bias != nullptr) {
|
|
scale_bias_dtype = bias->dtype();
|
|
} else {
|
|
scale_bias_dtype = x_dtype;
|
|
}
|
|
}
|
|
|
|
#define PADDLE_LAUNCH_LAYERNORM_BWD(ScaleBiasT, IsScaleBiasSameDTypeWithX) \
|
|
do { \
|
|
auto* scale_data = \
|
|
(scale == nullptr ? nullptr : scale->data<ScaleBiasT>()); \
|
|
auto* d_scale_data = \
|
|
(d_scale == nullptr ? nullptr \
|
|
: dev_ctx.template Alloc<ScaleBiasT>(d_scale)); \
|
|
auto* d_bias_data = \
|
|
(d_bias == nullptr ? nullptr \
|
|
: dev_ctx.template Alloc<ScaleBiasT>(d_bias)); \
|
|
auto* d_x_data = \
|
|
(d_x == nullptr ? nullptr : dev_ctx.template Alloc<T>(d_x)); \
|
|
funcs::LayerNormBackward<T, U, IsScaleBiasSameDTypeWithX>(x_data, \
|
|
d_y_data, \
|
|
scale_data, \
|
|
mean_data, \
|
|
var_data, \
|
|
d_x_data, \
|
|
d_scale_data, \
|
|
d_bias_data, \
|
|
epsilon, \
|
|
batch_size, \
|
|
feature_size, \
|
|
dev_ctx); \
|
|
} while (0)
|
|
|
|
#define PADDLE_LAUNCH_FAST_LAYERNORM_V2_BWD(ScaleBiasT) \
|
|
do { \
|
|
auto stream = dev_ctx.stream(); \
|
|
auto place = x.place(); \
|
|
auto* scale_data = \
|
|
(scale == nullptr ? nullptr : scale->data<ScaleBiasT>()); \
|
|
auto* d_scale_data = \
|
|
(d_scale == nullptr ? nullptr \
|
|
: dev_ctx.template Alloc<ScaleBiasT>(d_scale)); \
|
|
auto* d_bias_data = \
|
|
(d_bias == nullptr ? nullptr \
|
|
: dev_ctx.template Alloc<ScaleBiasT>(d_bias)); \
|
|
auto* d_x_data = \
|
|
(d_x == nullptr ? nullptr : dev_ctx.template Alloc<T>(d_x)); \
|
|
funcs::fast_ln_v2::LaunchNormBwd<T, Context>(dev_ctx, \
|
|
stream, \
|
|
place, \
|
|
x_data, \
|
|
scale_data, \
|
|
mean_data, \
|
|
var_data, \
|
|
d_y_data, \
|
|
d_x_data, \
|
|
d_scale_data, \
|
|
d_bias_data, \
|
|
scale_bias_dtype, \
|
|
x_dtype, \
|
|
x_grad->dtype(), \
|
|
compute_dtype, \
|
|
feature_size, \
|
|
batch_size, \
|
|
feature_size, \
|
|
epsilon); \
|
|
} while (0)
|
|
|
|
auto compute_dtype = CppTypeToDataType<U>::Type();
|
|
auto kernel_variant = LayerNormGradKernelDispatch(scale_bias_dtype,
|
|
x_dtype,
|
|
x_dtype,
|
|
compute_dtype,
|
|
feature_size,
|
|
x.numel(),
|
|
scale,
|
|
bias);
|
|
switch (kernel_variant) {
|
|
#if defined(PADDLE_WITH_CUDA) && !defined(PADDLE_WITH_HIP) && !defined(_WIN32)
|
|
case LayerNormGadKernelVariant::FAST_LN_V2:
|
|
if (scale_bias_dtype == x_dtype) {
|
|
PADDLE_LAUNCH_FAST_LAYERNORM_V2_BWD(T);
|
|
} else {
|
|
PADDLE_LAUNCH_FAST_LAYERNORM_V2_BWD(U);
|
|
}
|
|
break;
|
|
#endif
|
|
case LayerNormGadKernelVariant::GENERIC:
|
|
default:
|
|
#ifdef PADDLE_WITH_CUDA
|
|
if ((FLAGS_use_accuracy_compatible_kernel ||
|
|
(!isPowerOfTwo(feature_size) && feature_size > 1024)) &&
|
|
scale_bias_dtype == x_dtype) {
|
|
auto* scale_data = (scale == nullptr ? nullptr : scale->data<T>());
|
|
auto* d_scale_data =
|
|
(d_scale == nullptr ? nullptr : dev_ctx.template Alloc<T>(d_scale));
|
|
auto* d_bias_data =
|
|
(d_bias == nullptr ? nullptr : dev_ctx.template Alloc<T>(d_bias));
|
|
auto* d_x_data =
|
|
(d_x == nullptr ? nullptr : dev_ctx.template Alloc<T>(d_x));
|
|
LayerNormBwdCompatKernel<T, Context>(dev_ctx,
|
|
d_y_data,
|
|
x_data,
|
|
scale_data,
|
|
mean_data,
|
|
var_data,
|
|
d_x_data,
|
|
d_scale_data,
|
|
d_bias_data,
|
|
epsilon,
|
|
batch_size,
|
|
feature_size);
|
|
} else {
|
|
#endif
|
|
if (scale_bias_dtype == x_dtype) {
|
|
PADDLE_LAUNCH_LAYERNORM_BWD(T, true);
|
|
} else {
|
|
PADDLE_LAUNCH_LAYERNORM_BWD(U, false);
|
|
}
|
|
#ifdef PADDLE_WITH_CUDA
|
|
}
|
|
#endif
|
|
}
|
|
|
|
#undef PADDLE_LAUNCH_LAYERNORM_BWD
|
|
#undef PADDLE_LAUNCH_FAST_LAYERNORM_V2_BWD
|
|
}
|
|
|
|
} // namespace phi
|
|
|
|
#ifdef PADDLE_WITH_HIP
|
|
// MIOPEN do not support double
|
|
PD_REGISTER_KERNEL(layer_norm_grad,
|
|
GPU,
|
|
ALL_LAYOUT,
|
|
phi::LayerNormGradKernel,
|
|
float,
|
|
phi::float16) {
|
|
if (kernel_key.dtype() == phi::DataType::FLOAT16) {
|
|
kernel->OutputAt(1).SetDataType(phi::DataType::FLOAT32);
|
|
kernel->OutputAt(2).SetDataType(phi::DataType::FLOAT32);
|
|
}
|
|
}
|
|
#elif CUDNN_VERSION_MIN(8, 1, 0)
|
|
PD_REGISTER_KERNEL(layer_norm_grad,
|
|
GPU,
|
|
ALL_LAYOUT,
|
|
phi::LayerNormGradKernel,
|
|
float,
|
|
double,
|
|
phi::float16,
|
|
phi::bfloat16) {
|
|
if (kernel_key.dtype() == phi::DataType::FLOAT16) {
|
|
kernel->OutputAt(1).SetDataType(phi::DataType::FLOAT32);
|
|
kernel->OutputAt(2).SetDataType(phi::DataType::FLOAT32);
|
|
}
|
|
}
|
|
#else
|
|
PD_REGISTER_KERNEL(layer_norm_grad,
|
|
GPU,
|
|
ALL_LAYOUT,
|
|
phi::LayerNormGradKernel,
|
|
float,
|
|
double,
|
|
phi::float16) {
|
|
if (kernel_key.dtype() == phi::DataType::FLOAT16) {
|
|
kernel->OutputAt(1).SetDataType(phi::DataType::FLOAT32);
|
|
kernel->OutputAt(2).SetDataType(phi::DataType::FLOAT32);
|
|
}
|
|
}
|
|
#endif
|