wehub-resource-sync
103 lines
3.4 KiB
Plaintext
103 lines
3.4 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.
|
|
|
|
// clang-format will try to sort headers according to google c++ style,
|
|
// and that cause compiling problems.
|
|
// clang-format off
|
|
#include "paddle/phi/kernels/gelu_kernel.h"
|
|
|
|
#include "paddle/phi/backends/gpu/gpu_context.h"
|
|
#include "paddle/phi/common/amp_type_traits.h"
|
|
#include "paddle/phi/core/dense_tensor.h"
|
|
#include "paddle/phi/core/kernel_registry.h"
|
|
#include "paddle/phi/kernels/funcs/broadcast_function.h"
|
|
#include "paddle/phi/kernels/gpu/gelu_funcs.h"
|
|
// clang-format on
|
|
|
|
COMMON_DECLARE_bool(use_fast_math);
|
|
COMMON_DECLARE_bool(use_accuracy_compatible_kernel);
|
|
|
|
namespace phi {
|
|
|
|
template <typename T>
|
|
struct GeluWithApproximateFunctor {
|
|
using MT = typename MPTypeTrait<T>::Type;
|
|
inline HOSTDEVICE T operator()(T arg_x) {
|
|
// this function is tanh approximation of gelu
|
|
MT x = static_cast<MT>(arg_x);
|
|
MT one = static_cast<MT>(1);
|
|
MT half = static_cast<MT>(0.5);
|
|
MT kAlpha = M_SQRT2 * M_2_SQRTPI * MT(0.5);
|
|
auto tanh_out =
|
|
tanh(kAlpha * (x + static_cast<MT>(GELU_CONSTANT) * (x * x * x)));
|
|
MT out = half * x * (one + tanh_out);
|
|
return static_cast<T>(out);
|
|
}
|
|
};
|
|
|
|
template <typename T>
|
|
struct GeluWithoutApproximateFunctor {
|
|
using MT = typename MPTypeTrait<T>::Type;
|
|
inline HOSTDEVICE T operator()(T arg_x) {
|
|
// actual gelu with approximation = false
|
|
MT x = static_cast<MT>(arg_x);
|
|
// return static_cast<T>(x * normcdf(x));
|
|
constexpr MT kAlpha = M_SQRT1_2;
|
|
return static_cast<T>(x * MT(0.5) * (MT(1) + std::erf(x * kAlpha)));
|
|
}
|
|
};
|
|
|
|
template <typename T, typename Context>
|
|
void GeluKernel(const Context& dev_ctx,
|
|
const DenseTensor& x,
|
|
bool approximate,
|
|
DenseTensor* out) {
|
|
dev_ctx.template Alloc<T>(out);
|
|
if (out && out->numel() == 0) {
|
|
return;
|
|
}
|
|
std::vector<const DenseTensor*> ins = {&x};
|
|
std::vector<DenseTensor*> outs = {out};
|
|
if (approximate) {
|
|
#if defined(__NVCC__) || defined(__HIPCC__)
|
|
if (std::is_same<T, dtype::float16>::value &&
|
|
!FLAGS_use_accuracy_compatible_kernel) {
|
|
size_t n = x.numel();
|
|
const auto* in_ptr = reinterpret_cast<const __half*>(x.data<T>());
|
|
auto* out_ptr = reinterpret_cast<__half*>(out->data<T>());
|
|
if (TryLaunchFP16FastGeluFwdVectorizeCUDAKernel(
|
|
dev_ctx, in_ptr, out_ptr, n)) {
|
|
return;
|
|
}
|
|
}
|
|
#endif
|
|
using Functor = GeluWithApproximateFunctor<T>;
|
|
funcs::ElementwiseKernel<T, Functor, 1>(dev_ctx, ins, &outs, Functor());
|
|
} else {
|
|
using Functor = GeluWithoutApproximateFunctor<T>;
|
|
funcs::ElementwiseKernel<T, Functor, 1>(dev_ctx, ins, &outs, Functor());
|
|
}
|
|
}
|
|
|
|
} // namespace phi
|
|
|
|
PD_REGISTER_KERNEL(gelu,
|
|
GPU,
|
|
ALL_LAYOUT,
|
|
phi::GeluKernel,
|
|
float,
|
|
double,
|
|
phi::float16,
|
|
phi::bfloat16) {}
|