Skip to content

Commit

Permalink
[Phi]Move Relu/Cos/Sin/Tan/Acos/Asin/Atan/Sinh/Cosh/Asinh/Acosh/Atanh…
Browse files Browse the repository at this point in the history 
… kernels in Activation to Phi (#40175)

* move activation op

* adjust code format

* fix compile bugs

* fix ci bugs

* code format adjust

* code format adjust2

* activate ci status

* modify according to comment
  • Loading branch information
@YuanRisheng
YuanRisheng authored Mar 8, 2022
1 parent f1fe2ad commit 975f99a
Show file tree
Hide file tree
Showing 25 changed files with 1,908 additions and 907 deletions.
2 changes: 1 addition & 1 deletion cmake/operators.cmake
View file
Original file line number Diff line number Diff line change
Expand Up @@ -478,7 +478,7 @@ function(op_library TARGET)
if (${pybind_flag} EQUAL 0)
# NOTE(*): activation use macro to regist the kernels, set use_op manually.
if(${TARGET} STREQUAL "activation")
file(APPEND ${pybind_file} "USE_OP(relu);\n")
file(APPEND ${pybind_file} "USE_OP_ITSELF(relu);\n")
elseif(${TARGET} STREQUAL "fake_dequantize")
file(APPEND ${pybind_file} "USE_OP(fake_dequantize_max_abs);\n")
elseif(${TARGET} STREQUAL "fake_quantize")
Expand Down
2 changes: 1 addition & 1 deletion paddle/fluid/framework/ir/mkldnn/mkldnn_inplace_pass_tester.cc
View file
Original file line number Diff line number Diff line change
Expand Up @@ -27,7 +27,7 @@ USE_OP_DEVICE_KERNEL(elementwise_add, MKLDNN);
USE_OP(leaky_relu);
USE_OP_DEVICE_KERNEL(leaky_relu, MKLDNN);
USE_OP(gelu);
USE_OP(relu);
USE_OP_ITSELF(relu);
USE_OP(tanh);
USE_OP_DEVICE_KERNEL(tanh, MKLDNN);

Expand Down
4 changes: 2 additions & 2 deletions paddle/fluid/framework/paddle2cinn/build_cinn_pass_test.cc
View file
Original file line number Diff line number Diff line change
Expand Up @@ -675,7 +675,7 @@ TEST(BuildCinnPassTest, NoNeedBufferInput) {

USE_PASS(build_cinn_pass);
USE_OP(mul);
USE_OP(relu);
USE_OP_ITSELF(relu);
USE_OP_ITSELF(elementwise_add);
USE_OP(relu_grad);
USE_OP_ITSELF(relu_grad);
USE_OP_ITSELF(elementwise_add_grad);
2 changes: 1 addition & 1 deletion paddle/fluid/framework/paddle2cinn/cinn_compiler_test.cc
View file
Original file line number Diff line number Diff line change
Expand Up @@ -301,5 +301,5 @@ TEST(CinnCompilerTest, Compile) {
USE_PASS(build_cinn_pass);
USE_PASS(graph_viz_pass);
USE_OP(mul);
USE_OP(relu);
USE_OP_ITSELF(relu);
USE_OP_ITSELF(elementwise_add);
2 changes: 1 addition & 1 deletion paddle/fluid/imperative/tests/test_prepare_op.cc
View file
Original file line number Diff line number Diff line change
Expand Up @@ -226,7 +226,7 @@ TEST(test_prepare_op, test_prepare_data_cpu_mkldnn) {
} // namespace paddle

USE_OP_ITSELF(split);
USE_OP(relu);
USE_OP_ITSELF(relu);
#ifdef PADDLE_WITH_MKLDNN
USE_OP_DEVICE_KERNEL(relu, MKLDNN);
#endif
2 changes: 1 addition & 1 deletion paddle/fluid/inference/tensorrt/convert/test_activation_op.cc
View file
Original file line number Diff line number Diff line change
Expand Up @@ -52,7 +52,7 @@ TEST(Relu6OpConverter, main) { test_activation("relu6"); }
} // namespace inference
} // namespace paddle

USE_OP(relu);
USE_OP_ITSELF(relu);
USE_OP(sigmoid);
USE_OP(tanh);
USE_OP(relu6);
19 changes: 14 additions & 5 deletions paddle/fluid/operators/activation_cudnn_op.cu.cc
View file
Original file line number Diff line number Diff line change
Expand Up @@ -132,7 +132,9 @@ struct CudnnReluGradFunctor : public CudnnActivationGradFunctor<T> {
explicit CudnnReluGradFunctor(const CUDADeviceContext& ctx)
: CudnnActivationGradFunctor<T>(ctx, 0.0, GPUDNN_ACTIVATION_RELU) {}

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

template <typename T>
Expand All @@ -146,7 +148,9 @@ struct CudnnRelu6GradFunctor : public CudnnActivationGradFunctor<T> {
: CudnnActivationGradFunctor<T>(ctx, 6.0,
GPUDNN_ACTIVATION_CLIPPED_RELU) {}

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

template <typename T>
Expand All @@ -159,7 +163,9 @@ struct CudnnSigmoidGradFunctor : public CudnnActivationGradFunctor<T> {
explicit CudnnSigmoidGradFunctor(const CUDADeviceContext& ctx)
: CudnnActivationGradFunctor<T>(ctx, 0.0, GPUDNN_ACTIVATION_SIGMOID) {}

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

template <typename T>
Expand All @@ -172,7 +178,9 @@ struct CudnnTanhGradFunctor : public CudnnActivationGradFunctor<T> {
explicit CudnnTanhGradFunctor(const CUDADeviceContext& ctx)
: CudnnActivationGradFunctor<T>(ctx, 0.0, GPUDNN_ACTIVATION_TANH) {}

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

template <typename Functor>
Expand All @@ -197,7 +205,8 @@ class CudnnActivationGradKernel
public:
using T = typename Functor::ELEMENT_TYPE;
void Compute(const framework::ExecutionContext& context) const override {
static_assert(Functor::FwdDeps() == kDepOut, "Forward deps must be Out.");
static_assert(Functor::FwdDeps() == ActBwdOpFwdDeps::kDepOut,
"Forward deps must be Out.");

const framework::Tensor *X, *Out, *dOut;
X = Out = dOut = nullptr;
Expand Down
43 changes: 26 additions & 17 deletions paddle/fluid/operators/activation_op.cc
View file
Original file line number Diff line number Diff line change
Expand Up @@ -34,7 +34,8 @@ using paddle::framework::Tensor;

template <typename GradFunctor>
static constexpr bool CanInplaceAct() {
return GradFunctor::FwdDeps() == kDepOut || GradFunctor::FwdDeps() == kNoDeps;
return GradFunctor::FwdDeps() == ActBwdOpFwdDeps::kDepOut ||
GradFunctor::FwdDeps() == ActBwdOpFwdDeps::kNoDeps;
}

#define REGISTER_ACTIVATION_OP_MAKER(OP_NAME, OP_COMMENT) \
Expand Down Expand Up @@ -921,7 +922,8 @@ class ActivationOpDoubleGrad : public framework::OperatorWithKernel {
using framework::OperatorWithKernel::OperatorWithKernel;

void InferShape(framework::InferShapeContext* ctx) const override {
if (static_cast<int>(kDepValue) & static_cast<int>(kDepX)) {
if (static_cast<int>(kDepValue) &
static_cast<int>(ActBwdOpFwdDeps::kDepX)) {
if (ctx->HasOutput("DX")) {
ctx->ShareDim("X", "DX");
ctx->ShareLoD("X", "DX");
Expand All @@ -931,7 +933,8 @@ class ActivationOpDoubleGrad : public framework::OperatorWithKernel {
ctx->ShareLoD("X", "DDOut");
}
}
if (static_cast<int>(kDepValue) & static_cast<int>(kDepOut)) {
if (static_cast<int>(kDepValue) &
static_cast<int>(ActBwdOpFwdDeps::kDepOut)) {
if (ctx->HasOutput("DOut")) {
ctx->ShareDim("Out", "DOut");
ctx->ShareLoD("Out", "DOut");
Expand Down Expand Up @@ -960,13 +963,15 @@ class ActivationOpDoubleGrad2 : public framework::OperatorWithKernel {
using framework::OperatorWithKernel::OperatorWithKernel;

void InferShape(framework::InferShapeContext* ctx) const override {
if (static_cast<int>(kDepValue) & static_cast<int>(kDepX)) {
if (static_cast<int>(kDepValue) &
static_cast<int>(ActBwdOpFwdDeps::kDepX)) {
if (ctx->HasOutput("DDOut")) {
ctx->ShareDim("X", "DDOut");
ctx->ShareLoD("X", "DDOut");
}
}
if (static_cast<int>(kDepValue) & static_cast<int>(kDepOut)) {
if (static_cast<int>(kDepValue) &
static_cast<int>(ActBwdOpFwdDeps::kDepOut)) {
if (ctx->HasOutput("DDOut")) {
ctx->ShareDim("Out", "DDOut");
ctx->ShareLoD("Out", "DDOut");
Expand All @@ -987,7 +992,8 @@ class ActivationOpTripleGrad : public framework::OperatorWithKernel {
using framework::OperatorWithKernel::OperatorWithKernel;

void InferShape(framework::InferShapeContext* ctx) const override {
if (static_cast<int>(kDepValue) & static_cast<int>(kDepX)) {
if (static_cast<int>(kDepValue) &
static_cast<int>(ActBwdOpFwdDeps::kDepX)) {
if (ctx->HasOutput("DX")) {
ctx->ShareDim("X", "DX");
ctx->ShareLoD("X", "DX");
Expand All @@ -997,7 +1003,8 @@ class ActivationOpTripleGrad : public framework::OperatorWithKernel {
ctx->ShareLoD("X", "DDOut");
}
}
if (static_cast<int>(kDepValue) & static_cast<int>(kDepOut)) {
if (static_cast<int>(kDepValue) &
static_cast<int>(ActBwdOpFwdDeps::kDepOut)) {
if (ctx->HasOutput("D_DOut")) {
ctx->ShareDim("Out", "D_DOut");
ctx->ShareLoD("Out", "D_DOut");
Expand Down Expand Up @@ -1464,6 +1471,18 @@ namespace plat = paddle::platform;
FOR_EACH_ACTIVATION_OP(REGISTER_ACTIVATION_OP);
FOR_EACH_ACTIVATION_OP(REGISTER_ACTIVATION_CPU_KERNEL);

REGISTER_ACTIVATION_OP(cos, Cos, CosFunctor, CosGradFunctor)
REGISTER_ACTIVATION_OP(tan, Tan, TanFunctor, TanGradFunctor);
REGISTER_ACTIVATION_OP(acos, Acos, AcosFunctor, AcosGradFunctor);
REGISTER_ACTIVATION_OP(sin, Sin, SinFunctor, SinGradFunctor);
REGISTER_ACTIVATION_OP(asin, Asin, AsinFunctor, AsinGradFunctor);
REGISTER_ACTIVATION_OP(atan, Atan, AtanFunctor, AtanGradFunctor);
REGISTER_ACTIVATION_OP(sinh, Sinh, SinhFunctor, SinhGradFunctor);
REGISTER_ACTIVATION_OP(cosh, Cosh, CoshFunctor, CoshGradFunctor);
REGISTER_ACTIVATION_OP(asinh, Asinh, AsinhFunctor, AsinhGradFunctor);
REGISTER_ACTIVATION_OP(acosh, Acosh, AcoshFunctor, AcoshGradFunctor);
REGISTER_ACTIVATION_OP(atanh, Atanh, AtanhFunctor, AtanhGradFunctor);

/* ========================== sigmoid register =============================
*/
// 1. Register Sigmoid Operator
Expand Down Expand Up @@ -1584,16 +1603,6 @@ REGISTER_OPERATOR(
ops::ActivationOpDoubleGrad2<ops::ReluGradFunctor<float>::FwdDeps()>,
ops::ActivationDoubleGradOpInplaceInferer);

REGISTER_ACTIVATION_CPU_KERNEL(relu, Relu, ReluCPUFunctor, ReluGradFunctor);

REGISTER_OP_CPU_KERNEL(
relu_grad_grad,
ops::ActivationDoubleGradKernel<plat::CPUDeviceContext,
ops::ReluGradGradFunctor<float>>,
ops::ActivationDoubleGradKernel<plat::CPUDeviceContext,
ops::ReluGradGradFunctor<double>>,
ops::ActivationDoubleGradKernel<plat::CPUDeviceContext,
ops::ReluGradGradFunctor<plat::float16>>);
/* ========================================================================== */

/* ======================== leaky relu register ============================ */
Expand Down
Loading

0 comments on commit 975f99a

Please sign in to comment.