Skip to content

[XPU] add complex,real,imag,conj op for xpu #72649

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

Merged
merged 4 commits into from
May 15, 2025
Merged

[XPU] add complex,real,imag,conj op for xpu #72649

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
4 commits
Select commit Hold shift + click to select a range
06cae11
add complex_op_xpu
fangfangssj May 9, 2025
47b4612
fix
fangfangssj May 12, 2025
6b14709
Merge branch 'develop' into complexop
fangfangssj May 12, 2025
e91411a
fix
fangfangssj May 14, 2025
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
8 changes: 8 additions & 0 deletions cmake/phi.cmake
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -110,6 +110,13 @@ function(kernel_declare TARGET_LIST)
set(first_registry "")
endif()
endif()
# The kernel related to xpufft must have WITH_XPU_FFT enabled.
if(WITH_XPU AND NOT WITH_XPU_FFT)
string(FIND "${first_registry}" "xpufft" pos)
if(pos GREATER 1)
set(first_registry "")
endif()
endif()

if(NOT first_registry STREQUAL "")
string(
Expand Down Expand Up @@ -141,6 +148,7 @@ function(kernel_declare TARGET_LIST)
string(REPLACE "," ";" kernel_msg "${kernel_msg}")
string(REGEX REPLACE "[ \\\t\r\n]+" "" kernel_msg "${kernel_msg}")
string(REGEX REPLACE "//cuda_only" "" kernel_msg "${kernel_msg}")
string(REGEX REPLACE "//xpufft" "" kernel_msg "${kernel_msg}")

list(GET kernel_msg 0 kernel_name)
if(NOT is_all_backend STREQUAL "")
Expand Down
9 changes: 9 additions & 0 deletions paddle/phi/backends/xpu/xpu3_op_list.cc
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -1822,6 +1822,15 @@ XPUOpMap& get_kl3_ops() {
phi::DataType::INT16,
phi::DataType::INT64,
phi::DataType::INT32})},
#ifdef PADDLE_WITH_XPU_FFT
{"conj", XPUKernelSet({phi::DataType::COMPLEX64})},
{"real", XPUKernelSet({phi::DataType::COMPLEX64})},
{"real_grad", XPUKernelSet({phi::DataType::COMPLEX64})},
{"imag", XPUKernelSet({phi::DataType::COMPLEX64})},
{"imag_grad", XPUKernelSet({phi::DataType::COMPLEX64})},
{"complex", XPUKernelSet({phi::DataType::FLOAT32})},
{"complex_grad", XPUKernelSet({phi::DataType::FLOAT32})},
#endif
};

return s_xpu3_kernels;
Expand Down
143 changes: 143 additions & 0 deletions paddle/phi/kernels/xpu/complex_grad_kernel.cc
View file
Open in desktop
Original file line number Diff line number Diff line change
@@ -0,0 +1,143 @@
// 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.

#ifdef PADDLE_WITH_XPU_FFT
#include "paddle/phi/kernels/complex_grad_kernel.h"

#include "fft/cuComplex.h"
Copy link
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

这里是不是需要开启WITH_XPU_XFT才能编译？那就要考虑以下几件事：
1、现在这个配置，默认是OFF，所以默认情况下就编不过？
2、是否需要像其它宏定义一样，加个判断？比如，如果没定义这个，就走到另外的分支上，报not implemented或者precondition not met。

Copy link
Contributor Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

是的，已修改为打开WITH_XPU_FFT才编译新增加的fft相关的算子，未打开的情况上不编译进去
这样的话不需要报not implemented或者precondition not met，当未编译时使用到相关的算子时，框架会自动fallback回cpu上执行，这样不会报错，使用体验更好一些

#include "paddle/phi/backends/xpu/enforce_xpu.h"
#include "paddle/phi/common/complex.h"
#include "paddle/phi/common/type_traits.h"
#include "paddle/phi/core/kernel_registry.h"
#include "paddle/phi/kernels/expand_grad_kernel.h"
#include "paddle/phi/kernels/funcs/math_function.h"

namespace xfft_internal::xpu {
int combine_as_complex(int N, float* real, float* imag, float2* out);
int complex_spilt_float(int N, float2* in, float* real, float* imag);
} // namespace xfft_internal::xpu

namespace phi {

template <class T, class Context>
static DenseTensor Fill(const Context& ctx,
std::vector<int> shape,
T fill_value) {
DenseTensor ret;
ret.Resize(common::make_ddim(shape));
ctx.template Alloc<T>(&ret);
funcs::SetConstant<Context, T>()(ctx, &ret, fill_value);
return ret;
}

template <typename T, typename Context>
void RealGradKernel(const Context& dev_ctx,
const DenseTensor& dout,
DenseTensor* dx) {
auto numel = dout.numel();
auto* dx_data =
dev_ctx.template Alloc<T>(dx, static_cast<size_t>(numel * sizeof(T)));
DenseTensor imag = Fill<phi::dtype::Real<T>, Context>(
dev_ctx, common::vectorize<int>(dout.dims()), phi::dtype::Real<T>(0.0));
int r = xfft_internal::xpu::combine_as_complex(
numel,
const_cast<phi::dtype::Real<T>*>(dout.data<phi::dtype::Real<T>>()),
imag.data<phi::dtype::Real<T>>(),
reinterpret_cast<cuFloatComplex*>(dx_data));
PADDLE_ENFORCE_XPU_SUCCESS(r);
}

template <typename T, typename Context>
void ImagGradKernel(const Context& dev_ctx,
const DenseTensor& dout,
DenseTensor* dx) {
auto numel = dout.numel();
auto* dx_data =
dev_ctx.template Alloc<T>(dx, static_cast<size_t>(numel * sizeof(T)));
DenseTensor real = Fill<phi::dtype::Real<T>, Context>(
dev_ctx, common::vectorize<int>(dout.dims()), phi::dtype::Real<T>(0.0));
int r = xfft_internal::xpu::combine_as_complex(
Copy link
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

xpufftPR：#72477

numel,
real.data<phi::dtype::Real<T>>(),
const_cast<phi::dtype::Real<T>*>(dout.data<phi::dtype::Real<T>>()),
reinterpret_cast<cuFloatComplex*>(dx_data));
PADDLE_ENFORCE_XPU_SUCCESS(r);
}

template <typename T, typename Context>
void ComplexGradKernel(const Context& dev_ctx,
const DenseTensor& x,
const DenseTensor& y,
const DenseTensor& dout,
DenseTensor* dx,
DenseTensor* dy) {
using C = phi::dtype::complex<T>;
auto numel = dout.numel();
DenseTensor real_dout, imag_dout;
real_dout.Resize(dout.dims());
imag_dout.Resize(dout.dims());
T* real_data = dev_ctx.template Alloc<T>(&real_dout);
T* imag_data = dev_ctx.template Alloc<T>(&imag_dout);

int r = xfft_internal::xpu::complex_spilt_float(
numel,
reinterpret_cast<cuFloatComplex*>(const_cast<C*>(dout.data<C>())),
real_data,
imag_data);
PADDLE_ENFORCE_XPU_SUCCESS(r);

if (dx) {
if (x.dims() == dout.dims()) {
dx->ShareDataWith(real_dout);
} else {
ExpandGradKernel<T, Context>(
dev_ctx, x, real_dout, phi::IntArray(phi::vectorize(x.dims())), dx);
}
}

if (dy) {
if (y.dims() == dout.dims()) {
dy->ShareDataWith(imag_dout);
} else {
ExpandGradKernel<T, Context>(
dev_ctx, y, imag_dout, phi::IntArray(phi::vectorize(y.dims())), dy);
}
}
}
} // namespace phi

PD_REGISTER_KERNEL(imag_grad, // xpufft
XPU,
ALL_LAYOUT,
phi::ImagGradKernel,
phi::dtype::complex<float>) {
kernel->InputAt(0).SetDataType(phi::dtype::ToReal(kernel_key.dtype()));
}

PD_REGISTER_KERNEL(real_grad, // xpufft
XPU,
ALL_LAYOUT,
phi::RealGradKernel,
phi::dtype::complex<float>) {
kernel->InputAt(0).SetDataType(phi::dtype::ToReal(kernel_key.dtype()));
}

PD_REGISTER_KERNEL(complex_grad, // xpufft
XPU,
ALL_LAYOUT,
phi::ComplexGradKernel,
float) {
kernel->InputAt(2).SetDataType(phi::dtype::ToComplex(kernel_key.dtype()));
}
#endif // PADDLE_WITH_XPU_FFT
171 changes: 171 additions & 0 deletions paddle/phi/kernels/xpu/complex_kernel.cc
View file
Open in desktop
Original file line number Diff line number Diff line change
@@ -0,0 +1,171 @@
// 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.

#ifdef PADDLE_WITH_XPU_FFT
#include "paddle/phi/kernels/complex_kernel.h"

#include "fft/cuComplex.h"
#include "paddle/phi/backends/xpu/enforce_xpu.h"
#include "paddle/phi/common/complex.h"
#include "paddle/phi/common/type_traits.h"
#include "paddle/phi/core/kernel_registry.h"
#include "paddle/phi/kernels/expand_kernel.h"
#include "paddle/phi/kernels/funcs/common_infer_shape_functions.h"
#include "paddle/phi/kernels/xpu/xpu_api_wrapper.h"

namespace xfft_internal::xpu {
int combine_as_complex(int N, float* real, float* imag, float2* out);
int complex_spilt_float(int N, float2* in, float* real, float* imag);
int Conj(int N, float2* input, float2* output);
} // namespace xfft_internal::xpu

namespace phi {
template <typename T, typename Context>
void ConjKernel(const Context& dev_ctx,
const DenseTensor& x,
DenseTensor* out) {
dev_ctx.template Alloc<T>(out);
if (std::is_same<T, phi::dtype::complex<float>>::value) {
Copy link
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

可以尝试把is_same::value换成is_same_v

Copy link
Contributor Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

在下个PR中修改，同时conj在xpu3_op_list.cc中少添加的类型也在下个PR中添加

Copy link
Contributor Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

#72734

int r = xfft_internal::xpu::Conj(
x.numel(),
reinterpret_cast<cuFloatComplex*>(const_cast<T*>(x.data<T>())),
reinterpret_cast<cuFloatComplex*>(out->data<T>()));
PADDLE_ENFORCE_XPU_SUCCESS(r);
} else {
using XPUType = typename XPUCopyTypeTrait<T>::Type;
const auto* input_data = x.data<T>();
int r = xpu::copy<XPUType>(dev_ctx.x_context(),
reinterpret_cast<const XPUType*>(input_data),
reinterpret_cast<XPUType*>(out->data<T>()),
x.numel());
PADDLE_ENFORCE_XDNN_SUCCESS(r, "copy");
}
}

template <typename T, typename Context>
void RealKernel(const Context& dev_ctx,
const DenseTensor& x,
DenseTensor* out) {
dev_ctx.template Alloc<phi::dtype::Real<T>>(out);
// The allocation of imag here is redundant and could be optimized.
phi::DenseTensor imag;
imag.Resize(x.dims());
dev_ctx.template Alloc<phi::dtype::Real<T>>(&imag);
int r = xfft_internal::xpu::complex_spilt_float(
out->numel(),
reinterpret_cast<cuFloatComplex*>(const_cast<T*>(x.data<T>())),
out->data<phi::dtype::Real<T>>(),
imag.data<phi::dtype::Real<T>>());
PADDLE_ENFORCE_XPU_SUCCESS(r);
}

template <typename T, typename Context>
void ImagKernel(const Context& dev_ctx,
const DenseTensor& x,
DenseTensor* out) {
dev_ctx.template Alloc<phi::dtype::Real<T>>(out);
// The allocation of ‘real’ here is redundant and could be optimized.
phi::DenseTensor real;
real.Resize(x.dims());
dev_ctx.template Alloc<phi::dtype::Real<T>>(&real);
int r = xfft_internal::xpu::complex_spilt_float(
out->numel(),
reinterpret_cast<cuFloatComplex*>(const_cast<T*>(x.data<T>())),
real.data<phi::dtype::Real<T>>(),
out->data<phi::dtype::Real<T>>());
PADDLE_ENFORCE_XPU_SUCCESS(r);
}

template <typename T, typename Context>
void ComplexKernel(const Context& dev_ctx,
const DenseTensor& x,
const DenseTensor& y,
DenseTensor* out) {
using C = phi::dtype::complex<T>;
auto x_dims = x.dims();
auto y_dims = y.dims();
auto out_dims = phi::funcs::BroadcastTwoDims(x_dims, y_dims);
std::vector<int64_t> out_dims_vec = phi::vectorize(out_dims);

DenseTensor broadcasted_x, broadcasted_y;
T* x_data = nullptr;
T* y_data = nullptr;

if (x_dims == out_dims) {
x_data = const_cast<T*>(x.data<T>());
} else {
broadcasted_x.Resize(out_dims);
dev_ctx.template Alloc<T>(&broadcasted_x);
ExpandKernel<T, Context>(
dev_ctx, x, phi::IntArray(out_dims_vec), &broadcasted_x);
x_data = broadcasted_x.data<T>();
}

if (y_dims == out_dims) {
y_data = const_cast<T*>(y.data<T>());
} else {
broadcasted_y.Resize(out_dims);
dev_ctx.template Alloc<T>(&broadcasted_y);
ExpandKernel<T, Context>(
dev_ctx, y, phi::IntArray(out_dims_vec), &broadcasted_y);
y_data = broadcasted_y.data<T>();
}

dev_ctx.template Alloc<C>(out);
int r = xfft_internal::xpu::combine_as_complex(
out->numel(),
x_data,
y_data,
reinterpret_cast<cuFloatComplex*>(out->data<C>()));
PADDLE_ENFORCE_XPU_SUCCESS(r);
}
} // namespace phi

PD_REGISTER_KERNEL(conj, // xpufft
XPU,
ALL_LAYOUT,
phi::ConjKernel,
bool,
int,
int64_t,
float,
double,
phi::dtype::float16,
phi::dtype::bfloat16,
phi::dtype::complex<float>) {}

PD_REGISTER_KERNEL(real, // xpufft
XPU,
ALL_LAYOUT,
phi::RealKernel,
phi::dtype::complex<float>) {
kernel->OutputAt(0).SetDataType(phi::dtype::ToReal(kernel_key.dtype()));
}

PD_REGISTER_KERNEL(imag, // xpufft
XPU,
ALL_LAYOUT,
phi::ImagKernel,
phi::dtype::complex<float>) {
kernel->OutputAt(0).SetDataType(phi::dtype::ToReal(kernel_key.dtype()));
}

PD_REGISTER_KERNEL(complex, // xpufft
XPU,
ALL_LAYOUT,
phi::ComplexKernel,
float) {
kernel->OutputAt(0).SetDataType(phi::dtype::ToComplex(kernel_key.dtype()));
}
#endif // PADDLE_WITH_XPU_FFT
Loading
Loading