move elementwise_max/min/mod into phi

Author: FlyingQianMM

paddle/fluid/framework/new_executor/standalone_executor_test.cc

@@ -54,7 +54,7 @@ USE_OP(sum);
 USE_OP_ITSELF(slice_grad);
 USE_OP_ITSELF(lookup_table_grad);
 USE_OP(sqrt);
-USE_OP(elementwise_max);
+USE_OP_ITSELF(elementwise_max);
 USE_OP_ITSELF(elementwise_div);
 USE_OP_ITSELF(sgd);
 USE_OP(squared_l2_norm);

paddle/fluid/operators/elementwise/elementwise_functor.h

@@ -70,75 +70,29 @@ struct InverseFloorDivFunctor {
 
 // Maximum
 template <typename T>
-struct MaxFunctor {
- inline HOSTDEVICE T operator()(const T a, const T b) const {
- return a > b ? a : b;
- }
-};
+using MaxFunctor = phi::funcs::MaximumFunctor<T>;
 
 // Minmum
 template <typename T>
-struct MinFunctor {
- inline HOSTDEVICE T operator()(const T a, const T b) const {
- return a < b ? a : b;
- }
-};
+using MinFunctor = phi::funcs::MinimumFunctor<T>;
 
 template <typename T>
 using Complex = paddle::platform::complex<T>;
 
+// Ternary compare
 template <typename T>
-struct MinGradXFunctor {
- inline HOSTDEVICE T operator()(const T x, const T y, const T dout) const {
- return dout * static_cast<T>(x < y);
- }
-};
+using MaxGradXFunctor = phi::funcs::MaxGradXFunctor<T>;
 template <typename T>
-struct MinGradYFunctor {
- inline HOSTDEVICE T operator()(const T x, const T y, const T dout) const {
- return dout * static_cast<T>(x >= y);
- }
-};
-
+using MaxGradYFunctor = phi::funcs::MaxGradYFunctor<T>;
 template <typename InT, typename OutT>
-struct MinGradXYFunctor {
- inline HOSTDEVICE phi::Array<OutT, 2> operator()(const InT x, const InT y,
- const InT dout) {
- phi::Array<OutT, 2> outs;
- // dx = dout * (x < y)
- outs[0] = static_cast<OutT>(dout * static_cast<InT>(x < y));
- // dy = dout * (x >= y)
- outs[1] = static_cast<OutT>(dout * static_cast<InT>(x >= y));
- return outs;
- }
-};
+using MaxGradXYFunctor = phi::funcs::MaxGradXYFunctor<InT, OutT>;
 
-// Ternary compare
 template <typename T>
-struct MaxGradXFunctor {
- inline HOSTDEVICE T operator()(const T x, const T y, const T dout) const {
- return dout * static_cast<T>(x > y);
- }
-};
+using MinGradXFunctor = phi::funcs::MinGradXFunctor<T>;
 template <typename T>
-struct MaxGradYFunctor {
- inline HOSTDEVICE T operator()(const T x, const T y, const T dout) const {
- return dout * static_cast<T>(x <= y);
- }
-};
-
+using MinGradYFunctor = phi::funcs::MinGradYFunctor<T>;
 template <typename InT, typename OutT>
-struct MaxGradXYFunctor {
- inline HOSTDEVICE phi::Array<OutT, 2> operator()(const InT x, const InT y,
- const InT dout) {
- phi::Array<OutT, 2> outs;
- // dx = dout * (x > y)
- outs[0] = static_cast<OutT>(dout * static_cast<InT>(x > y));
- // dy = dout * (x <= y)
- outs[1] = static_cast<OutT>(dout * static_cast<InT>(x <= y));
- return outs;
- }
-};
+using MinGradXYFunctor = phi::funcs::MinGradXYFunctor<InT, OutT>;
 
 } // namespace operators
 } // namespace paddle

paddle/fluid/operators/elementwise/elementwise_max_op.cc

@@ -12,8 +12,6 @@ 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/fluid/operators/elementwise/elementwise_max_op.h"
-
 #include <string>
 
 #include "paddle/fluid/operators/elementwise/elementwise_op.h"
@@ -119,23 +117,6 @@ REGISTER_OPERATOR(elementwise_max, ops::ElementwiseOp,
 
 REGISTER_OPERATOR(elementwise_max_grad, ops::ElementwiseOpGrad);
 
-REGISTER_OP_CPU_KERNEL(
- elementwise_max,
- ops::ElementwiseMaxKernel<paddle::platform::CPUDeviceContext, float>,
- ops::ElementwiseMaxKernel<paddle::platform::CPUDeviceContext, double>,
- ops::ElementwiseMaxKernel<paddle::platform::CPUDeviceContext, int>,
- ops::ElementwiseMaxKernel<paddle::platform::CPUDeviceContext, int64_t>,
- ops::ElementwiseMaxKernel<paddle::platform::CPUDeviceContext,
- paddle::platform::bfloat16>);
-REGISTER_OP_CPU_KERNEL(
- elementwise_max_grad,
- ops::ElementwiseMaxGradKernel<paddle::platform::CPUDeviceContext, float>,
- ops::ElementwiseMaxGradKernel<paddle::platform::CPUDeviceContext, double>,
- ops::ElementwiseMaxGradKernel<paddle::platform::CPUDeviceContext, int>,
- ops::ElementwiseMaxGradKernel<paddle::platform::CPUDeviceContext, int64_t>,
- ops::ElementwiseMaxGradKernel<paddle::platform::CPUDeviceContext,
- paddle::platform::bfloat16>);
-
 REGISTER_OP_VERSION(elementwise_max)
  .AddCheckpoint(
  R"ROC(Register elementwise_max for adding the attribute of Scale_y)ROC",

paddle/fluid/operators/elementwise/elementwise_max_op.cu

@@ -12,7 +12,6 @@ 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/fluid/operators/elementwise/elementwise_max_op.h"
 #include "paddle/fluid/operators/elementwise/elementwise_npu.h"
 #include "paddle/fluid/platform/device/npu/npu_op_runner.h"
 

paddle/fluid/operators/elementwise/elementwise_max_op.h

@@ -14,7 +14,6 @@ limitations under the License. */
 
 #ifdef PADDLE_WITH_XPU
 
-#include "paddle/fluid/operators/elementwise/elementwise_max_op.h"
 #include "paddle/fluid/operators/elementwise/elementwise_op.h"
 #include "paddle/fluid/operators/elementwise/elementwise_xpu.h"
 namespace paddle {

paddle/fluid/operators/elementwise/elementwise_max_op_npu.cc

@@ -12,8 +12,6 @@ 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/fluid/operators/elementwise/elementwise_min_op.h"
-
 #include <string>
 
 #include "paddle/fluid/operators/elementwise/elementwise_op.h"
@@ -119,19 +117,6 @@ REGISTER_OPERATOR(elementwise_min, ops::ElementwiseOp,
 
 REGISTER_OPERATOR(elementwise_min_grad, ops::ElementwiseOpGrad);
 
-REGISTER_OP_CPU_KERNEL(
- elementwise_min,
- ops::ElementwiseMinKernel<paddle::platform::CPUDeviceContext, float>,
- ops::ElementwiseMinKernel<paddle::platform::CPUDeviceContext, double>,
- ops::ElementwiseMinKernel<paddle::platform::CPUDeviceContext, int>,
- ops::ElementwiseMinKernel<paddle::platform::CPUDeviceContext, int64_t>);
-REGISTER_OP_CPU_KERNEL(
- elementwise_min_grad,
- ops::ElementwiseMinGradKernel<paddle::platform::CPUDeviceContext, float>,
- ops::ElementwiseMinGradKernel<paddle::platform::CPUDeviceContext, double>,
- ops::ElementwiseMinGradKernel<paddle::platform::CPUDeviceContext, int>,
- ops::ElementwiseMinGradKernel<paddle::platform::CPUDeviceContext, int64_t>);
-
 REGISTER_OP_VERSION(elementwise_min)
  .AddCheckpoint(
  R"ROC(Register elementwise_min for adding the attribute of Scale_y)ROC",