Skip to content
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

【Paddle Tensor 规范化第二期】pow support complex #71230

Open
wants to merge 1 commit into
base: develop
Choose a base branch
from
Open

【Paddle Tensor 规范化第二期】pow support complex #71230

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
1 commit
Select commit
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
36 changes: 33 additions & 3 deletions paddle/phi/kernels/cpu/activation_grad_kernel.cc
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -248,6 +248,30 @@ void HardSwishGradKernel(const Context& dev_ctx,
dev_ctx, &x, nullptr, &dout, dx, functor);
}

template <typename T, typename Context>
void PowGradKernel(const Context& dev_ctx,
const DenseTensor& x,
const DenseTensor& dout,
const Scalar& factor,
DenseTensor* dx) {
PADDLE_ENFORCE_NOT_NULL(
dx, errors::NotFound("The output DenseTensor dX can not be nullptr"));
if (dx) {
dev_ctx.template Alloc<T>(dx);
}
auto dout_flatten = EigenVector<T>::Flatten(
GET_DATA_SAFELY(&dout, "Input", "Out@GRAD", "PowGrad"));
auto dx_flatten = EigenVector<T>::Flatten(
GET_DATA_SAFELY(dx, "Output", "X@GRAD", "PowGrad"));
auto x_flatten =
EigenVector<T>::Flatten(GET_DATA_SAFELY(&x, "Input", "X", "PowGrad"));
auto* place = dev_ctx.eigen_device();
phi::funcs::PowGradFunctor<T> functor;
auto attrs = functor.GetAttrs();
*(attrs[0].second) = factor.to<float>();
functor(*place, x_flatten, nullptr, dout_flatten, dx_flatten);
}

} // namespace phi

PD_REGISTER_KERNEL(
Expand Down Expand Up @@ -462,20 +486,26 @@ PD_REGISTER_KERNEL(pow_grad,
float,
double,
int,
int64_t) {}
int64_t,
phi::dtype::complex<float>,
phi::dtype::complex<double>) {}
PD_REGISTER_KERNEL(pow_double_grad,
CPU,
ALL_LAYOUT,
phi::PowDoubleGradKernel,
float,
double,
int,
int64_t) {}
int64_t,
phi::dtype::complex<float>,
phi::dtype::complex<double>) {}
PD_REGISTER_KERNEL(pow_triple_grad,
CPU,
ALL_LAYOUT,
phi::PowTripleGradKernel,
float,
double,
int,
int64_t) {}
int64_t,
phi::dtype::complex<float>,
phi::dtype::complex<double>) {}
59 changes: 41 additions & 18 deletions paddle/phi/kernels/cpu/activation_kernel.cc
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -173,6 +173,25 @@ void RoundKernel(const Context& dev_ctx,
dev_ctx, x, out, functor);
}

template <typename T, typename Context>
void PowKernel(const Context& dev_ctx,
const DenseTensor& x,
const Scalar& factor,
DenseTensor* out) {
PADDLE_ENFORCE_NOT_NULL(out,
errors::NotFound("Output Out should not be nullptr"));
dev_ctx.template Alloc<T>(out);
auto x_flatten = phi::EigenVector<T>::Flatten(
GET_DATA_SAFELY(&x, "Input", "X", "Activation"));
auto out_flatten = phi::EigenVector<T>::Flatten(
GET_DATA_SAFELY(out, "Output", "Out", "Activation"));
auto* place = dev_ctx.eigen_device();
phi::funcs::PowFunctor<T> functor;
auto attrs = functor.GetAttrs();
*(attrs[0].second) = factor.to<float>();
functor(*place, x_flatten, out_flatten);
}

} // namespace phi
PD_REGISTER_KERNEL(relu, CPU, ALL_LAYOUT, phi::ReluKernel, float, double) {}

Expand Down Expand Up @@ -215,6 +234,18 @@ PD_REGISTER_ACTIVATION_KERNEL_WITH_COMPLEX(reciprocal, ReciprocalKernel)
PD_REGISTER_ACTIVATION_KERNEL_WITH_COMPLEX(sqrt, SqrtKernel)
PD_REGISTER_ACTIVATION_KERNEL(rsqrt, RsqrtKernel)
PD_REGISTER_ACTIVATION_KERNEL_WITH_COMPLEX(softplus, SoftplusKernel)
PD_REGISTER_ACTIVATION_KERNEL(logit, LogitKernel)
PD_REGISTER_ACTIVATION_KERNEL_WITH_COMPLEX(softsign, SoftsignKernel)
PD_REGISTER_ACTIVATION_KERNEL_WITH_COMPLEX(sigmoid, SigmoidKernel)
PD_REGISTER_ACTIVATION_KERNEL_WITH_COMPLEX(logsigmoid, LogSigmoidKernel)
PD_REGISTER_ACTIVATION_KERNEL(hardsigmoid, HardSigmoidKernel)
PD_REGISTER_ACTIVATION_KERNEL(swish, SwishKernel)
PD_REGISTER_ACTIVATION_KERNEL(relu6, Relu6Kernel)
PD_REGISTER_ACTIVATION_KERNEL_WITH_COMPLEX(hardswish, HardSwishKernel)
PD_REGISTER_ACTIVATION_KERNEL(round, RoundKernel)
PD_REGISTER_ACTIVATION_KERNEL(floor, FloorKernel)
PD_REGISTER_ACTIVATION_KERNEL(ceil, CeilKernel)
PD_REGISTER_ACTIVATION_KERNEL(celu, CeluKernel)

PD_REGISTER_KERNEL(exp,
CPU,
Expand All @@ -227,7 +258,6 @@ PD_REGISTER_KERNEL(exp,
phi::dtype::float16,
phi::dtype::complex<float>,
phi::dtype::complex<double>) {}

PD_REGISTER_KERNEL(expm1,
CPU,
ALL_LAYOUT,
Expand All @@ -239,8 +269,6 @@ PD_REGISTER_KERNEL(expm1,
phi::dtype::float16,
phi::dtype::complex<float>,
phi::dtype::complex<double>) {}

PD_REGISTER_KERNEL(logit, CPU, ALL_LAYOUT, phi::LogitKernel, float, double) {}
PD_REGISTER_KERNEL(square,
CPU,
ALL_LAYOUT,
Expand All @@ -251,13 +279,6 @@ PD_REGISTER_KERNEL(square,
int64_t,
phi::dtype::complex<float>,
phi::dtype::complex<double>) {}
PD_REGISTER_ACTIVATION_KERNEL_WITH_COMPLEX(softsign, SoftsignKernel)
PD_REGISTER_ACTIVATION_KERNEL_WITH_COMPLEX(sigmoid, SigmoidKernel)
PD_REGISTER_ACTIVATION_KERNEL_WITH_COMPLEX(logsigmoid, LogSigmoidKernel)
PD_REGISTER_ACTIVATION_KERNEL(hardsigmoid, HardSigmoidKernel)
PD_REGISTER_ACTIVATION_KERNEL(swish, SwishKernel)
PD_REGISTER_ACTIVATION_KERNEL(relu6, Relu6Kernel)

PD_REGISTER_KERNEL(log,
CPU,
ALL_LAYOUT,
Expand Down Expand Up @@ -306,11 +327,6 @@ PD_REGISTER_KERNEL(log1p,
phi::dtype::bfloat16,
phi::dtype::complex<float>,
phi::dtype::complex<double>) {}

PD_REGISTER_ACTIVATION_KERNEL_WITH_COMPLEX(hardswish, HardSwishKernel)
PD_REGISTER_ACTIVATION_KERNEL(round, RoundKernel)
PD_REGISTER_ACTIVATION_KERNEL(floor, FloorKernel)
PD_REGISTER_ACTIVATION_KERNEL(ceil, CeilKernel)
PD_REGISTER_KERNEL(negative,
CPU,
ALL_LAYOUT,
Expand All @@ -322,6 +338,13 @@ PD_REGISTER_KERNEL(negative,
int64_t,
phi::dtype::complex<float>,
phi::dtype::complex<double>) {}
PD_REGISTER_ACTIVATION_KERNEL(celu, CeluKernel)
PD_REGISTER_KERNEL(
pow, CPU, ALL_LAYOUT, phi::PowKernel, float, double, int, int64_t) {}
PD_REGISTER_KERNEL(pow,
CPU,
ALL_LAYOUT,
phi::PowKernel,
float,
double,
int,
int64_t,
phi::dtype::complex<float>,
phi::dtype::complex<double>) {}
99 changes: 99 additions & 0 deletions paddle/phi/kernels/funcs/activation_functor.h
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -31,6 +31,7 @@

#include "paddle/phi/common/amp_type_traits.h"
#include "paddle/phi/common/bfloat16.h"
#include "paddle/phi/common/complex.h"
#include "paddle/phi/common/float16.h"
#include "paddle/phi/core/dense_tensor.h"
#include "paddle/phi/core/enforce.h"
Expand Down Expand Up @@ -2936,6 +2937,19 @@ struct PowFunctor : public BaseActivationFunctor<T> {
}
};

template <typename T>
struct PowFunctor<ComplexType<T>>
: public BaseActivationFunctor<ComplexType<T>> {
float factor;
typename BaseActivationFunctor<ComplexType<T>>::AttrPair GetAttrs() {
return {{"factor", &factor}};
}
template <typename Device, typename X, typename Out>
void operator()(Device d, X x, Out out) const {
out.device(d) = x.pow(static_cast<ComplexType<T>>(factor)); // NOLINT
}
};

template <typename T>
struct PowGradFunctor : public BaseActivationFunctor<T> {
float factor;
Expand All @@ -2955,6 +2969,27 @@ struct PowGradFunctor : public BaseActivationFunctor<T> {
static constexpr ActBwdOpFwdDeps FwdDeps() { return ActBwdOpFwdDeps::kDepX; }
};

template <typename T>
struct PowGradFunctor<ComplexType<T>>
: public BaseActivationFunctor<ComplexType<T>> {
float factor;
typename BaseActivationFunctor<ComplexType<T>>::AttrPair GetAttrs() {
return {{"factor", &factor}};
}
template <typename Device,
typename X,
typename Out,
typename dOut,
typename dX>
void operator()(Device d, X x, Out out UNUSED, dOut dout, dX dx) const {
dx.device(d) =
dout * static_cast<ComplexType<T>>(factor) *
x.pow(static_cast<ComplexType<T>>(factor - 1)).unaryExpr(Conj<T>());
}

static constexpr ActBwdOpFwdDeps FwdDeps() { return ActBwdOpFwdDeps::kDepX; }
};

// floor(x) = flooring(x)
template <typename T>
struct FloorFunctor : public BaseActivationFunctor<T> {
Expand Down Expand Up @@ -5195,6 +5230,70 @@ struct CudaCeilFunctor : public BaseActivationFunctor<T> {
}
};

template <typename T>
struct CudaPowFunctor : public BaseActivationFunctor<T> {
using MT = typename phi::dtype::MPTypeTrait<T>::Type;
using MPType =
typename std::conditional<std::is_integral<T>::value, float, MT>::type;
float factor;
typename BaseActivationFunctor<T>::AttrPair GetAttrs() {
return {{"factor", &factor}};
}
__device__ __forceinline__ T operator()(const T x) const {
MPType x_t = static_cast<MPType>(x);
MPType factor_t = static_cast<MPType>(factor);
return static_cast<T>(pow(x_t, factor_t));
}
};

template <typename T>
struct CudaPowGradFunctor : public BaseActivationFunctor<T> {
using MT = typename phi::dtype::MPTypeTrait<T>::Type;
using MPType =
typename std::conditional<std::is_integral<T>::value, float, MT>::type;
float factor;
typename BaseActivationFunctor<T>::AttrPair GetAttrs() {
return {{"factor", &factor}};
}
// dx = dout * n * pow(x, n - 1)
__device__ __forceinline__ T operator()(const T dout, const T x) const {
MPType x_t = static_cast<MPType>(x);
MPType dout_t = static_cast<MPType>(dout);
MPType factor_t = static_cast<MPType>(factor);
return static_cast<T>(dout_t * factor_t * pow(x_t, factor - 1));
}
static constexpr ActBwdOpFwdDeps FwdDeps() { return ActBwdOpFwdDeps::kDepX; }
};

template <typename T>
struct CudaPowFunctor<ComplexType<T>>
: public BaseActivationFunctor<ComplexType<T>> {
float factor;
typename BaseActivationFunctor<T>::AttrPair GetAttrs() {
return {{"factor", &factor}};
}
__device__ __forceinline__ ComplexType<T> operator()(
const ComplexType<T> x) const {
return pow(x, static_cast<ComplexType<T>>(factor));
}
};

template <typename T>
struct CudaPowGradFunctor<ComplexType<T>>
: public BaseActivationFunctor<ComplexType<T>> {
float factor;
typename BaseActivationFunctor<T>::AttrPair GetAttrs() {
return {{"factor", &factor}};
}
// dx = dout * n * pow(x, n - 1)
__device__ __forceinline__ ComplexType<T> operator()(
const ComplexType<T> dout, const ComplexType<T> x) const {
return dout * conj(static_cast<ComplexType<T>>(factor) *
pow(x, static_cast<ComplexType<T>>(factor - 1)));
}
static constexpr ActBwdOpFwdDeps FwdDeps() { return ActBwdOpFwdDeps::kDepX; }
};

template <typename T>
struct CudaFloorFunctor : public BaseActivationFunctor<T> {
using MPType = typename phi::dtype::MPTypeTrait<T>::Type;
Expand Down
7 changes: 2 additions & 5 deletions paddle/phi/kernels/funcs/elementwise_functor.h
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -959,21 +959,18 @@ inline HOSTDEVICE typename std::enable_if<std::is_integral<T>::value, T>::type
compute_pow(const T a, const T b) {
// TODO(wujionghao): A potential speed improvement is supporting different
// types in C++.
// On CUDAPlace, std::pow(3, 1) calls pow(float, float), and
// On CUDAPlace, pow(3, 1) calls pow(float, float), and
// it will return a float number like 2.99... , which floor to 2
// when cast to int by default and it is wrong.
// Use llrint to cast it to the nearest integer, which is 3.
return std::llrint(std::pow(static_cast<double>(a), static_cast<double>(b)));
return llrint(pow(static_cast<double>(a), static_cast<double>(b)));
}
template <typename T, typename MPType>
inline HOSTDEVICE typename std::enable_if<!std::is_integral<T>::value, T>::type
compute_pow(const T a, const T b) {
MPType a_val = static_cast<MPType>(a);
MPType b_val = static_cast<MPType>(b);
#ifdef PADDLE_WITH_XPU_KP
return static_cast<T>(pow(a_val, b_val));
#endif
return static_cast<T>(std::pow(a_val, b_val));
}
#else
template <typename T, typename MPType>
Expand Down
25 changes: 22 additions & 3 deletions paddle/phi/kernels/gpu/activation_grad_kernel.cu
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -297,6 +297,19 @@ void HardSwishGradKernel(const Context& dev_ctx,
dev_ctx, &x, nullptr, &dout, dx, functor);
}

template <typename T, typename Context>
void PowGradKernel(const Context& dev_ctx,
const DenseTensor& x,
const DenseTensor& dout,
const Scalar& factor,
DenseTensor* dx) {
funcs::CudaPowGradFunctor<T> functor;
auto attrs = functor.GetAttrs();
*(attrs[0].second) = factor.to<float>();
ActivationGradGPUImpl<T, Context, funcs::CudaPowGradFunctor<T>>(
dev_ctx, &x, nullptr, &dout, dx, functor);
}

} // namespace phi

#ifdef PADDLE_WITH_HIP
Expand Down Expand Up @@ -542,7 +555,9 @@ PD_REGISTER_KERNEL(pow_grad,
int,
int64_t,
phi::dtype::float16,
phi::dtype::bfloat16) {}
phi::dtype::bfloat16,
phi::dtype::complex<float>,
phi::dtype::complex<double>) {}
PD_REGISTER_KERNEL(pow_double_grad,
GPU,
ALL_LAYOUT,
Expand All @@ -552,7 +567,9 @@ PD_REGISTER_KERNEL(pow_double_grad,
int,
int64_t,
phi::dtype::float16,
phi::dtype::bfloat16) {}
phi::dtype::bfloat16,
phi::dtype::complex<float>,
phi::dtype::complex<double>) {}
PD_REGISTER_KERNEL(pow_triple_grad,
GPU,
ALL_LAYOUT,
Expand All @@ -562,4 +579,6 @@ PD_REGISTER_KERNEL(pow_triple_grad,
int,
int64_t,
phi::dtype::float16,
phi::dtype::bfloat16) {}
phi::dtype::bfloat16,
phi::dtype::complex<float>,
phi::dtype::complex<double>) {}
Loading
Loading