[Phi]Delete kernel registry of elementwise_sub op in Fluid (#40039)

* delete elementwise_sub kernel registry * fix compile bugs in xpu ci * fix bugs when run inference ci

Author: YuanRisheng

paddle/fluid/operators/elementwise/elementwise_op_npu_test.cc

@@ -33,7 +33,7 @@ namespace p = paddle::platform;
 
 USE_OP_ITSELF(elementwise_add);
 USE_OP_DEVICE_KERNEL(elementwise_add, NPU);
-USE_OP(elementwise_sub);
+USE_OP_ITSELF(elementwise_sub);
 USE_OP_DEVICE_KERNEL(elementwise_sub, NPU);
 
 template <typename T>

paddle/fluid/operators/elementwise/elementwise_pow_op_xpu.cc

@@ -14,7 +14,6 @@ limitations under the License. */
 
 #ifdef PADDLE_WITH_XPU
 #include "paddle/fluid/operators/elementwise/elementwise_op.h"
-#include "paddle/fluid/operators/elementwise/elementwise_sub_op.h"
 #include "paddle/fluid/operators/elementwise/elementwise_xpu.h"
 #include "xpu/refactor/math.h"
 

paddle/fluid/operators/elementwise/elementwise_sub_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_sub_op.h"
-
 #include <string>
 
 #include "paddle/fluid/operators/elementwise/elementwise_op.h"
@@ -78,10 +76,16 @@ class ElementwiseSubDoubleGradMaker : public framework::SingleGradOpMaker<T> {
 } // namespace paddle
 
 REGISTER_ELEMWISE_GRAD_MAKER(elementwise_sub, Sub);
-REGISTER_ELEMWISE_EXPLICIT_OP_WITHOUT_GRAD(elementwise_sub, Sub);
 
 namespace ops = paddle::operators;
 
+REGISTER_OPERATOR(elementwise_sub, ::paddle::operators::ElementwiseOp,
+ ::paddle::operators::ElementwiseSubOpMaker,
+ ::paddle::operators::ElementwiseOpInferVarType,
+ elementwise_subGradMaker<::paddle::framework::OpDesc>,
+ elementwise_subGradMaker<::paddle::imperative::OpBase>,
+ ::paddle::operators::ElementwiseOpInplaceInferer);
+
 REGISTER_OPERATOR(
  elementwise_sub_grad, ops::ElementwiseOpGrad,
  ops::ElementwiseGradOpInplaceInferer, ops::ElementwiseGradNoBufVarsInferer,
@@ -92,51 +96,6 @@ REGISTER_OPERATOR(elementwise_sub_grad_grad,
  ops::ElementwiseDoubleGradOpInplaceInferer,
  ops::ElementwiseDoubleGradNoBufVarsInferer);
 
-REGISTER_OP_CPU_KERNEL(
- elementwise_sub,
- ops::ElementwiseSubKernel<paddle::platform::CPUDeviceContext, float>,
- ops::ElementwiseSubKernel<paddle::platform::CPUDeviceContext, double>,
- ops::ElementwiseSubKernel<paddle::platform::CPUDeviceContext, int16_t>,
- ops::ElementwiseSubKernel<paddle::platform::CPUDeviceContext, int>,
- ops::ElementwiseSubKernel<paddle::platform::CPUDeviceContext, int64_t>,
- ops::ElementwiseSubKernel<paddle::platform::CPUDeviceContext,
- paddle::platform::bfloat16>,
- ops::ElementwiseSubKernel<paddle::platform::CPUDeviceContext,
- paddle::platform::complex<float>>,
- ops::ElementwiseSubKernel<paddle::platform::CPUDeviceContext,
- paddle::platform::complex<double>>);
-REGISTER_OP_CPU_KERNEL(
- elementwise_sub_grad,
- ops::ElementwiseSubGradKernel<paddle::platform::CPUDeviceContext, float>,
- ops::ElementwiseSubGradKernel<paddle::platform::CPUDeviceContext, double>,
- ops::ElementwiseSubGradKernel<paddle::platform::CPUDeviceContext, int16_t>,
- ops::ElementwiseSubGradKernel<paddle::platform::CPUDeviceContext, int>,
- ops::ElementwiseSubGradKernel<paddle::platform::CPUDeviceContext, int64_t>,
- ops::ElementwiseSubGradKernel<paddle::platform::CPUDeviceContext,
- paddle::platform::bfloat16>,
- ops::ElementwiseSubGradKernel<paddle::platform::CPUDeviceContext,
- paddle::platform::complex<float>>,
- ops::ElementwiseSubGradKernel<paddle::platform::CPUDeviceContext,
- paddle::platform::complex<double>>);
-REGISTER_OP_CPU_KERNEL(
- elementwise_sub_grad_grad,
- ops::ElementwiseSubDoubleGradKernel<paddle::platform::CPUDeviceContext,
- float>,
- ops::ElementwiseSubDoubleGradKernel<paddle::platform::CPUDeviceContext,
- double>,
- ops::ElementwiseSubDoubleGradKernel<paddle::platform::CPUDeviceContext,
- int16_t>,
- ops::ElementwiseSubDoubleGradKernel<paddle::platform::CPUDeviceContext,
- int>,
- ops::ElementwiseSubDoubleGradKernel<paddle::platform::CPUDeviceContext,
- int64_t>,
- ops::ElementwiseSubDoubleGradKernel<paddle::platform::CPUDeviceContext,
- paddle::platform::bfloat16>,
- ops::ElementwiseSubDoubleGradKernel<paddle::platform::CPUDeviceContext,
- paddle::platform::complex<float>>,
- ops::ElementwiseSubDoubleGradKernel<paddle::platform::CPUDeviceContext,
- paddle::platform::complex<double>>);
-
 REGISTER_OP_VERSION(elementwise_sub)
  .AddCheckpoint(
  R"ROC(Register elementwise_sub for adding the attribute of Scale_y)ROC",

paddle/fluid/operators/elementwise/elementwise_sub_op.cu

@@ -15,7 +15,7 @@ limitations under the License. */
 #include <memory>
 #include <string>
 
-#include "paddle/fluid/operators/elementwise/elementwise_sub_op.h"
+#include "paddle/fluid/operators/elementwise/elementwise_op.h"
 #include "paddle/fluid/platform/device/npu/npu_op_runner.h"
 
 namespace paddle {

paddle/fluid/operators/elementwise/elementwise_sub_op.h

@@ -13,7 +13,6 @@ See the License for the specific language governing permissions and
 limitations under the License. */
 
 #ifdef PADDLE_WITH_XPU
-#include "paddle/fluid/operators/elementwise/elementwise_sub_op.h"
 #include "paddle/fluid/operators/elementwise/elementwise_op.h"
 #include "paddle/fluid/operators/elementwise/elementwise_xpu.h"
 #include "xpu/refactor/math.h"

paddle/fluid/operators/elementwise/elementwise_sub_op_npu.cc

@@ -197,7 +197,8 @@ PD_REGISTER_KERNEL(subtract,
  int64_t,
  phi::dtype::float16,
  complex64,
- complex128) {}
+ complex128,
+ phi::dtype::bfloat16) {}
 PD_REGISTER_KERNEL(divide,
  GPU,
  ALL_LAYOUT,

paddle/fluid/operators/elementwise/elementwise_sub_op_xpu.cc

@@ -100,6 +100,12 @@ KernelSignature ElementwiseSubGradOpArgumentMapping(
  return KernelSignature("unregistered", {}, {}, {});
 }
 
+KernelSignature ElementwiseSubDoubleGradOpArgumentMapping(
+ const ArgumentMappingContext& ctx) {
+ return KernelSignature(
+ "subtract_double_grad", {"Y", "DDX", "DDY", "DOut"}, {"axis"}, {"DDOut"});
+}
+
 } // namespace phi
 
 PD_REGISTER_BASE_KERNEL_NAME(elementwise_add, add);
@@ -110,6 +116,7 @@ PD_REGISTER_BASE_KERNEL_NAME(elementwise_add_grad, add_grad);
 PD_REGISTER_BASE_KERNEL_NAME(elementwise_add_grad_grad, add_double_grad);
 PD_REGISTER_BASE_KERNEL_NAME(elementwise_add_triple_grad, add_triple_grad);
 PD_REGISTER_BASE_KERNEL_NAME(elementwise_sub_grad, subtract_grad);
+PD_REGISTER_BASE_KERNEL_NAME(elementwise_sub_grad_grad, subtract_double_grad);
 
 PD_REGISTER_ARG_MAPPING_FN(elementwise_add,
  phi::ElementwiseAddOpArgumentMapping);
@@ -127,3 +134,5 @@ PD_REGISTER_ARG_MAPPING_FN(elementwise_add_triple_grad,
  phi::ElementwiseAddTripleGradOpArgumentMapping);
 PD_REGISTER_ARG_MAPPING_FN(elementwise_sub_grad,
  phi::ElementwiseSubGradOpArgumentMapping);
+PD_REGISTER_ARG_MAPPING_FN(elementwise_sub_grad_grad,
+ phi::ElementwiseSubDoubleGradOpArgumentMapping);