[XPU] add complex,real,imag,conj op for xpu
#72649
There are no files selected for viewing
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
| Original file line number | Diff line number | Diff line change |
|---|---|---|
| @@ -0,0 +1,138 @@ | ||
| // 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/complex_grad_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_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( | ||
fangfangssj marked this conversation as resolved.
Show resolved
Hide resolved
|
||
| 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_XDNN_SUCCESS(r, "real_grad"); | ||
| } | ||
|
|
||
| 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( | ||
|
|
||
| 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_XDNN_SUCCESS(r, "imag_grad"); | ||
| } | ||
|
|
||
| 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_XDNN_SUCCESS(r, "complex_grad"); | ||
|
|
||
| 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, | ||
| XPU, | ||
| ALL_LAYOUT, | ||
| phi::ImagGradKernel, | ||
| phi::dtype::complex<float>) { | ||
| kernel->InputAt(0).SetDataType(phi::dtype::ToReal(kernel_key.dtype())); | ||
| } | ||
|
|
||
| PD_REGISTER_KERNEL(real_grad, | ||
| XPU, | ||
| ALL_LAYOUT, | ||
| phi::RealGradKernel, | ||
| phi::dtype::complex<float>) { | ||
| kernel->InputAt(0).SetDataType(phi::dtype::ToReal(kernel_key.dtype())); | ||
| } | ||
|
|
||
| PD_REGISTER_KERNEL( | ||
| complex_grad, XPU, ALL_LAYOUT, phi::ComplexGradKernel, float) { | ||
| kernel->InputAt(2).SetDataType(phi::dtype::ToComplex(kernel_key.dtype())); | ||
| } | ||
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
| Original file line number | Diff line number | Diff line change |
|---|---|---|
| @@ -0,0 +1,137 @@ | ||
| // 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/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" | ||
|
|
||
| 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); | ||
| 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_XDNN_SUCCESS(r, "conj"); | ||
| } | ||
|
|
||
| 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); | ||
| 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_XDNN_SUCCESS(r, "real"); | ||
fangfangssj marked this conversation as resolved.
Show resolved
Hide resolved
|
||
| } | ||
|
|
||
| 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); | ||
| 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_XDNN_SUCCESS(r, "imag"); | ||
| } | ||
|
|
||
| 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_XDNN_SUCCESS(r, "complex"); | ||
| } | ||
| } // namespace phi | ||
|
|
||
| PD_REGISTER_KERNEL( | ||
| conj, XPU, ALL_LAYOUT, phi::ConjKernel, float, phi::dtype::complex<float>) { | ||
| } | ||
|
|
||
| PD_REGISTER_KERNEL( | ||
| real, XPU, ALL_LAYOUT, phi::RealKernel, phi::dtype::complex<float>) { | ||
| kernel->OutputAt(0).SetDataType(phi::dtype::ToReal(kernel_key.dtype())); | ||
| } | ||
|
|
||
| PD_REGISTER_KERNEL( | ||
| imag, XPU, ALL_LAYOUT, phi::ImagKernel, phi::dtype::complex<float>) { | ||
| kernel->OutputAt(0).SetDataType(phi::dtype::ToReal(kernel_key.dtype())); | ||
| } | ||
|
|
||
| PD_REGISTER_KERNEL(complex, XPU, ALL_LAYOUT, phi::ComplexKernel, float) { | ||
| kernel->OutputAt(0).SetDataType(phi::dtype::ToComplex(kernel_key.dtype())); | ||
| } | ||
Oops, something went wrong.
Oops, something went wrong.
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
There was a problem hiding this comment.
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。
There was a problem hiding this comment.
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