Skip to content

[Phi]Move elementwise_div grad/double grad Kernel to Phi #40172

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
Mar 7, 2022
Merged

[Phi]Move elementwise_div grad/double grad Kernel to Phi #40172

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
089da30
move elementwise_div grad
YuanRisheng Mar 4, 2022
6fe0018
Merge branch 'develop' of github.com:YuanRisheng/Paddle into move_ele…
YuanRisheng Mar 4, 2022
7a85e74
change mutable_data to alloc
YuanRisheng Mar 4, 2022
1baeb3b
fix compile bugs
YuanRisheng Mar 4, 2022
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
2 changes: 1 addition & 1 deletion paddle/fluid/framework/new_executor/standalone_executor_test.cc
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -54,7 +54,7 @@ USE_OP(slice_grad);
USE_OP(lookup_table_grad);
USE_OP(sqrt);
USE_OP(elementwise_max);
USE_OP(elementwise_div);
USE_OP_ITSELF(elementwise_div);
USE_OP(sgd);
USE_OP(squared_l2_norm);
USE_OP(memcpy_h2d);
Expand Down
36 changes: 0 additions & 36 deletions paddle/fluid/operators/elementwise/elementwise_div_op.cc
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -102,42 +102,6 @@ REGISTER_OPERATOR(
REGISTER_OPERATOR(elementwise_div_grad_grad, ops::ElementwiseDivOpDoubleGrad,
ops::ElementwiseDoubleGradOpInplaceInferer);

REGISTER_OP_CPU_KERNEL(
elementwise_div,
ops::ElementwiseDivKernel<paddle::platform::CPUDeviceContext, float>,
ops::ElementwiseDivKernel<paddle::platform::CPUDeviceContext, double>,
ops::ElementwiseDivKernel<paddle::platform::CPUDeviceContext, int>,
ops::ElementwiseDivKernel<paddle::platform::CPUDeviceContext, int64_t>,
ops::ElementwiseDivKernel<paddle::platform::CPUDeviceContext,
paddle::platform::complex<float>>,
ops::ElementwiseDivKernel<paddle::platform::CPUDeviceContext,
paddle::platform::complex<double>>);
REGISTER_OP_CPU_KERNEL(
elementwise_div_grad,
ops::ElementwiseDivGradKernel<paddle::platform::CPUDeviceContext, float>,
ops::ElementwiseDivGradKernel<paddle::platform::CPUDeviceContext, double>,
ops::ElementwiseDivGradKernel<paddle::platform::CPUDeviceContext, int>,
ops::ElementwiseDivGradKernel<paddle::platform::CPUDeviceContext, int64_t>,
ops::ElementwiseDivGradKernel<paddle::platform::CPUDeviceContext,
paddle::platform::complex<float>>,
ops::ElementwiseDivGradKernel<paddle::platform::CPUDeviceContext,
paddle::platform::complex<double>>);

REGISTER_OP_CPU_KERNEL(
elementwise_div_grad_grad,
ops::ElementwiseDivDoubleGradKernel<paddle::platform::CPUDeviceContext,
float>,
ops::ElementwiseDivDoubleGradKernel<paddle::platform::CPUDeviceContext,
double>,
ops::ElementwiseDivDoubleGradKernel<paddle::platform::CPUDeviceContext,
int>,
ops::ElementwiseDivDoubleGradKernel<paddle::platform::CPUDeviceContext,
int64_t>,
ops::ElementwiseDivDoubleGradKernel<paddle::platform::CPUDeviceContext,
paddle::platform::complex<float>>,
ops::ElementwiseDivDoubleGradKernel<paddle::platform::CPUDeviceContext,
paddle::platform::complex<double>>);

REGISTER_OP_VERSION(elementwise_div)
.AddCheckpoint(
R"ROC(Register elementwise_div for adding the attribute of Scale_y)ROC",
Expand Down
96 changes: 0 additions & 96 deletions paddle/fluid/operators/elementwise/elementwise_div_op.cu
View file
Open in desktop

This file was deleted.

211 changes: 0 additions & 211 deletions paddle/fluid/operators/elementwise/elementwise_div_op.h
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -20,142 +20,6 @@ limitations under the License. */
namespace paddle {
namespace operators {

template <typename DeviceContext, typename T>
void default_elementwise_sub(const framework::ExecutionContext& ctx,
const framework::Tensor* x,
const framework::Tensor* y, framework::Tensor* z) {
int axis = ctx.Attr<int>("axis");
auto x_dims = x->dims();
auto y_dims = y->dims();
if (x_dims.size() >= y_dims.size()) {
ElementwiseComputeEx<SubFunctor<T>, DeviceContext, T>(ctx, x, y, axis,
SubFunctor<T>(), z);
} else {
ElementwiseComputeEx<InverseSubFunctor<T>, DeviceContext, T>(
ctx, x, y, axis, InverseSubFunctor<T>(), z);
}
}

template <typename DeviceContext, typename T>
void default_elementwise_div(const framework::ExecutionContext& ctx,
const framework::Tensor* x,
const framework::Tensor* y, framework::Tensor* z) {
int axis = ctx.Attr<int>("axis");
auto x_dims = x->dims();
auto y_dims = y->dims();
if (x_dims.size() >= y_dims.size()) {
ElementwiseComputeEx<DivFunctor<T>, DeviceContext, T>(ctx, x, y, axis,
DivFunctor<T>(), z);
} else {
ElementwiseComputeEx<InverseDivFunctor<T>, DeviceContext, T>(
ctx, x, y, axis, InverseDivFunctor<T>(), z);
}
}

template <typename DeviceContext, typename T>
class ElementwiseDivKernel : public framework::OpKernel<T> {
public:
void Compute(const framework::ExecutionContext& ctx) const override {
auto* x = ctx.Input<framework::LoDTensor>("X");
auto* y = ctx.Input<framework::LoDTensor>("Y");
auto* z = ctx.Output<framework::LoDTensor>("Out");
z->mutable_data<T>(ctx.GetPlace());

auto& dev_ctx = ctx.device_context<DeviceContext>();
int axis = ctx.Attr<int>("axis");
auto pt_x = paddle::experimental::MakePhiDenseTensor(*x);
auto pt_y = paddle::experimental::MakePhiDenseTensor(*y);
auto pt_z = paddle::experimental::MakePhiDenseTensor(*z);
phi::DivideRawKernel<T>(
static_cast<const typename framework::ConvertToPhiContext<
DeviceContext>::TYPE&>(dev_ctx),
*pt_x.get(), *pt_y.get(), axis, pt_z.get());
}
};

template <typename T>
struct DivGradDX {
HOSTDEVICE T operator()(T x, T y, T out, T dout) const { return dout / y; }
};

template <typename T>
struct DivGradDX<paddle::platform::complex<T>> {
HOSTDEVICE paddle::platform::complex<T> operator()(
paddle::platform::complex<T> x, paddle::platform::complex<T> y,
paddle::platform::complex<T> out,
paddle::platform::complex<T> dout) const {
paddle::platform::complex<T> y_conj(y.real, -y.imag);
return dout / y_conj;
}
};

template <typename T>
struct DivGradDY {
HOSTDEVICE T operator()(T x, T y, T out, T dout) const {
return -dout * out / y;
}
};

template <typename T>
struct DivGradDY<paddle::platform::complex<T>> {
HOSTDEVICE paddle::platform::complex<T> operator()(
paddle::platform::complex<T> x, paddle::platform::complex<T> y,
paddle::platform::complex<T> out,
paddle::platform::complex<T> dout) const {
paddle::platform::complex<T> out_div_y_conj((out / y).real,
-(out / y).imag);
return -dout * out_div_y_conj;
}
};

template <typename T>
struct DivDoubleDY {
HOSTDEVICE T operator()(T x, T y, T out, T dout) const {
return y * out * dout - x * dout;
}
};

template <typename DeviceContext, typename T>
typename std::enable_if<
std::is_same<DeviceContext, platform::CPUDeviceContext>::value>::type
ElementwiseDivGrad(const framework::ExecutionContext& ctx,
const framework::Tensor* x, const framework::Tensor* y,
const framework::Tensor* out, const framework::Tensor* dout,
framework::Tensor* dx, framework::Tensor* dy) {
int axis = ctx.Attr<int>("axis");

ElemwiseGradCompute<DeviceContext, T, DivGradDX<T>, DivGradDY<T>>(
ctx, *x, *y, *out, *dout, axis, dx, dy, DivGradDX<T>(), DivGradDY<T>());
}

#if defined(PADDLE_WITH_CUDA) || defined(PADDLE_WITH_HIP)
template <typename DeviceContext, typename T>
typename std::enable_if<
std::is_same<DeviceContext, platform::CUDADeviceContext>::value>::type
ElementwiseDivGrad(const framework::ExecutionContext& ctx,
const framework::Tensor* x, const framework::Tensor* y,
const framework::Tensor* out, const framework::Tensor* dout,
framework::Tensor* dx, framework::Tensor* dy);
#endif

template <typename DeviceContext, typename T>
class ElementwiseDivGradKernel : public ElemwiseGradKernel<T> {
public:
void Compute(const framework::ExecutionContext& ctx) const override {
ElemwiseGradKernel<T>::Compute(ctx);
using Tensor = framework::Tensor;

auto* x = ctx.Input<Tensor>("X");
auto* y = ctx.Input<Tensor>("Y");
auto* out = ctx.Input<Tensor>("Out");
auto* dout = ctx.Input<Tensor>(framework::GradVarName("Out"));
auto* dx = ctx.Output<Tensor>(framework::GradVarName("X"));
auto* dy = ctx.Output<Tensor>(framework::GradVarName("Y"));

ElementwiseDivGrad<DeviceContext, T>(ctx, x, y, out, dout, dx, dy);
}
};

class ElementwiseDivOpDoubleGrad : public framework::OperatorWithKernel {
public:
using framework::OperatorWithKernel::OperatorWithKernel;
Expand Down Expand Up @@ -206,80 +70,5 @@ class ElementwiseDivOpDoubleGrad : public framework::OperatorWithKernel {
}
};

template <typename DeviceContext, typename T>
class ElementwiseDivDoubleGradKernel : public framework::OpKernel<T> {
using Tensor = framework::Tensor;

public:
void Compute(const framework::ExecutionContext& ctx) const override {
auto* Y = ctx.Input<Tensor>("Y");
auto* Out = ctx.Input<Tensor>("Out");
auto* ddX = ctx.Input<Tensor>("DDX");
auto* ddY = ctx.Input<Tensor>("DDY");
auto* dX = ctx.Input<Tensor>("DX");

auto* dY = ctx.Output<Tensor>(framework::GradVarName("Y"));
auto* dOut = ctx.Output<Tensor>("DOut");
auto* ddOut = ctx.Output<Tensor>("DDOut");

int axis = ctx.Attr<int>("axis");

if (dY) dY->mutable_data<T>(Y->dims(), ctx.GetPlace());
if (dOut) dOut->mutable_data<T>(Out->dims(), ctx.GetPlace());
if (ddOut) ddOut->mutable_data<T>(Out->dims(), ctx.GetPlace());

// ddX_safe == null ? 0 : ddX
// ddY_safe == null ? 0 : ddY
Tensor ddX_safe, ddY_safe;
GetDoubleGradSafeTensor<DeviceContext, T>(ctx, dX, ddX, &ddX_safe);
GetDoubleGradSafeTensor<DeviceContext, T>(ctx, Y, ddY, &ddY_safe);

// ddOut = ddX / Y - Out * ddY / Y = (ddX - Out * ddY) / Y
// dY = Out * dX * ddY / Y - dX * ddX / Y
// dOut = - dX * ddY
// To save memory, (1) dout can be used as 'tmp' tensor, (2) ddout can
// inplace ddx
Tensor tmp;
if (dOut) {
tmp = *dOut;
} else {
auto& dev_ctx = ctx.template device_context<DeviceContext>();
tmp = ctx.AllocateTmpTensor<T, DeviceContext>(Out->dims(), dev_ctx);
}
if (dY) {
// dX_div_Y = dX / Y;
Tensor dX_div_Y = tmp;
default_elementwise_div<DeviceContext, T>(ctx, dX, Y, &dX_div_Y);

// NOTE(dengkaipeng): in the following ElemwiseGradCompute, for the
// first output tensor is nullptr, the branch to calculate first
// output tensor will not be activated, DivGradDx function will not
// be called and can be ignored, the first branch has little effect
// on running speed.

// dY = Out * dX * ddY / Y - dX * ddX / Y
ElemwiseGradCompute<DeviceContext, T, DivGradDX<T>, DivDoubleDY<T>>(
ctx, ddX_safe, ddY_safe, *Out, dX_div_Y, axis, nullptr, dY,
DivGradDX<T>(), DivDoubleDY<T>());
}

if (ddOut) {
// ddOut = ddX / Y - Out * ddY / Y = (ddX - Out * ddY) / Y
default_elementwise_mul<DeviceContext, T>(ctx, Out, &ddY_safe, &tmp);
default_elementwise_sub<DeviceContext, T>(ctx, &ddX_safe, &tmp, &tmp);
default_elementwise_div<DeviceContext, T>(ctx, &tmp, Y, ddOut);
}

if (dOut) {
// dOut = - dX * ddY
default_elementwise_mul<DeviceContext, T>(ctx, dX, &ddY_safe, dOut);
auto& place =
*ctx.template device_context<DeviceContext>().eigen_device();
auto dout = framework::EigenVector<T>::Flatten(*dOut);
dout.device(place) = static_cast<T>(-1) * dout;
}
}
};

} // namespace operators
} // namespace paddle
Loading