[Phi] Change the output format of C++ backward api (Part2) (#42545)

* change the output format of C++ backward api

* fix merge conflict

* fix sparse api code auto-gen

* fix eager_gen bug

* fix bug of output is null

* fix bug of conv2d_grad_impl

* fix optional grad

* fix bug of eager-gen double_grad

* fix bug

* fix multiply_double_grad bug

* fix bug of higher order derivative

* fix bug of FillZeroForEmptyGradInput

* remove redundant vector in grad_node

* fix bug of test_deformable_conv_v1_op

* fix bug of test_deformable_conv_v1_op

* some refacotr

paddle/fluid/eager/auto_code_generator/final_state_generator/codegen_utils.py

@@ -24,10 +24,11 @@ import os
 ops_to_fill_zero_for_empty_grads = set([
     "split_grad", "rnn_grad", "matmul_double_grad", "matmul_triple_grad",
     "sigmoid_double_grad", "sigmoid_triple_grad", "add_double_grad",
-    "add_triple_grad", "multiply_double_grad", "multiply_triple_grad",
-    "conv2d_grad_grad", "batch_norm_double_grad", "tanh_double_grad",
-    "tanh_triple_grad", "subtract_double_grad", "divide_double_grad",
-    "log_double_grad", "elu_double_grad", "leaky_relu_double_grad"
+    "add_triple_grad", "multiply_grad", "multiply_double_grad",
+    "multiply_triple_grad", "conv2d_grad_grad", "batch_norm_double_grad",
+    "tanh_double_grad", "tanh_triple_grad", "subtract_double_grad",
+    "divide_double_grad", "log_double_grad", "elu_double_grad",
+    "leaky_relu_double_grad"
 ])
 
 # For API dispatch used at python-level

paddle/fluid/eager/auto_code_generator/final_state_generator/eager_gen.py

@@ -337,14 +337,14 @@ AMP_LOGIC_TEMPLATE = \
 
 CREATE_PLAIN_OPTIONAL_TENSOR_TEMPLATE = \
 """
-    paddle::optional<const paddle::experimental::Tensor&> {}_optional = paddle::none;
-    if({}.initialized()) {}_optional = paddle::make_optional<const paddle::experimental::Tensor&>({});
+  paddle::optional<const paddle::experimental::Tensor&> {}_optional = paddle::none;
+  if({}.initialized()) {}_optional = paddle::make_optional<const paddle::experimental::Tensor&>({});
 """
 
 CREATE_RECOVER_OPTIONAL_TENSOR_TEMPLATE = \
 """
-    paddle::optional<const paddle::experimental::Tensor&> {}_optional = paddle::none;
-    if( {}.impl() ) {}_optional = paddle::make_optional<const paddle::experimental::Tensor&>({});
+  paddle::optional<const paddle::experimental::Tensor&> {}_optional = paddle::none;
+  if( {}.impl() ) {}_optional = paddle::make_optional<const paddle::experimental::Tensor&>({});
 """
 
 CHECK_NAN_AND_INF_TEMPLATE = \
@@ -738,9 +738,14 @@ class DygraphFunctionGeneratorBase(FunctionGeneratorBase):
         num_outputs = len(forward_outputs_position_map.keys())
         for name, (_, pos) in forward_outputs_position_map.items():
             output_autograd_meta_name = GetAutoGradMetaName(name)
+            set_out_rank = f"""{indent}if ({output_autograd_meta_name}) {{
+{indent}  egr::EagerUtils::SetOutRankWithSlot({output_autograd_meta_name}, {pos});
+{indent}}}"""
+
+            set_history = f"""{indent}if ({output_autograd_meta_name}) {{
+{indent}  egr::EagerUtils::SetHistory({output_autograd_meta_name}, grad_node);
+{indent}}}"""
 
-            set_out_rank = f"{indent}egr::EagerUtils::SetOutRankWithSlot({output_autograd_meta_name}, {pos});"
-            set_history = f"{indent}egr::EagerUtils::SetHistory({output_autograd_meta_name}, grad_node);"
             set_grad_in_meta = f"{indent}grad_node->SetGradInMeta({name}, {pos});"
             set_retain_grad = f"{indent}egr::EagerUtils::CheckAndRetainGrad({name});"
 
@@ -1265,7 +1270,17 @@ class DygraphNodeGenerator(DygraphFunctionGeneratorBase):
         # Fill Grad Ins with Zero
         fill_zero_str = ""
         if backward_api_name in ops_to_fill_zero_for_empty_grads:
-            fill_zero_str = f"{indent}egr::EagerUtils::FillZeroForEmptyGradInputs(&grads, this->InputMeta());\n"
+            fill_zero_str = f"{indent}const auto& input_metas = this->InputMeta();\n"
+            for name, (ttype, fwd_position,
+                       grad_api_position) in backward_grad_inputs_map.items():
+                if name in self.optional_inputs:
+                    if IsPlainTensorType(ttype):
+                        fill_zero_str += f"{indent}egr::EagerUtils::FillZeroForEmptyOptionalGradInput(&grads[{fwd_position}][0], input_metas[{fwd_position}][0]);\n"
+                else:
+                    if IsPlainTensorType(ttype):
+                        fill_zero_str += f"{indent}egr::EagerUtils::FillZeroForEmptyGradInput(&grads[{fwd_position}][0], input_metas[{fwd_position}][0]);\n"
+                    else:
+                        fill_zero_str += f"{indent}egr::EagerUtils::FillZeroForEmptyGradInput(&grads[{fwd_position}], input_metas[{fwd_position}]);\n"
 
         # Grad Ins from TensorWrappers
         for name, (_, is_fwd_input,
@@ -1322,40 +1337,46 @@ class DygraphNodeGenerator(DygraphFunctionGeneratorBase):
 
         get_grad_in_args_str = "\n".join(get_grad_in_args_list)
 
-        # Grad Outputs
-        for name, (ttype, fwd_position,
-                   grad_api_position) in backward_grad_outputs_map.items():
-            transformed_tensor_name = self.TransformToNextGradName(name)
-            if IsPlainTensorType(ttype):
-                grad_api_args.append(f"api_output[{fwd_position}][0]")
-            else:
-                assert IsVectorTensorType(ttype)
-                grad_api_args.append(f"api_output[{fwd_position}]")
-
-        grad_api_args_str = ", ".join(grad_api_args)
-
         # Grad Function Call String
         slot_num_bwd_outputs = len(self.forward_inputs_position_map.keys())
         grad_api_namespace = f"paddle::experimental::{namespace}"
         grad_function_call_str = f"""
   const auto& out_metas = OutputMeta();
   paddle::small_vector<std::vector<paddle::experimental::Tensor>, egr::kSlotSmallVectorSize> returns({slot_num_bwd_outputs});
-  paddle::small_vector<std::vector<paddle::experimental::Tensor*>, egr::kSlotSmallVectorSize> api_output({slot_num_bwd_outputs});
   for (int i = 0; i < {slot_num_bwd_outputs}; ++i) {{
     returns[i].resize(out_metas[i].size());
-    if(returns[i].size() == 0) {{
-      api_output[i].reserve(1);
-      api_output[i].push_back(nullptr);
-      continue;
-    }}
-    api_output[i].reserve(returns[i].size());
-    for (size_t j = 0; j < returns[i].size(); ++j) {{
-      api_output[i].push_back(&returns[i][j]);
-    }}
   }}
 """
 
-        grad_function_call_str = grad_function_call_str + f"{indent}{grad_api_namespace}{backward_api_name}({grad_api_args_str});"
+        # Grad Outputs
+        out_index = -1
+        for name, (ttype, fwd_position,
+                   grad_api_position) in backward_grad_outputs_map.items():
+            transformed_tensor_name = self.TransformToNextGradName(name)
+            out_index = out_index + 1
+            grad_api_args.append(f"api_output_{out_index}")
+
+            if IsPlainTensorType(ttype):
+                grad_function_call_str += f"""
+  auto* api_output_{out_index} = (out_metas[{fwd_position}].empty() || out_metas[{fwd_position}][0].IsStopGradient()) ? nullptr : &returns[{fwd_position}][0];"""
+
+            else:
+                assert IsVectorTensorType(ttype)
+                grad_function_call_str += f"""
+  std::vector<paddle::experimental::Tensor*> api_output_{out_index};
+  api_output_{out_index}.reserve(returns[{fwd_position}].size());
+  for (size_t i = 0; i < returns[{fwd_position}].size(); ++i) {{
+    if (out_metas[{fwd_position}].empty() || out_metas[{fwd_position}][i].IsStopGradient()) {{
+      api_output_{out_index}.push_back(nullptr);
+    }} else {{
+      api_output_{out_index}.push_back(&returns[{fwd_position}][i]);
+    }}
+  }}"""
+
+        grad_api_args_str = ", ".join(grad_api_args)
+
+        grad_function_call_str = grad_function_call_str + f"""
+{indent}{grad_api_namespace}{backward_api_name}({grad_api_args_str});"""
 
         # Check Nan and Inf
         check_nan_inf_str = CHECK_NAN_AND_INF_TEMPLATE.format(backward_api_name,
@@ -1425,7 +1446,7 @@ class DygraphNodeGenerator(DygraphFunctionGeneratorBase):
                 if IsPlainTensorType(rtype):
                     output_autograd_meta = f"""
   auto& {transformed_tensor_name} = returns[{pos}][0];
-  egr::AutogradMeta* {output_autograd_meta_name} = egr::EagerUtils::autograd_meta(&{transformed_tensor_name});"""
+  egr::AutogradMeta* {output_autograd_meta_name} = returns[{pos}][0].initialized() ? egr::EagerUtils::autograd_meta(&{transformed_tensor_name}) : nullptr;"""
 
                 else:
                     assert IsVectorTensorType(rtype)

paddle/fluid/eager/grad_node_info.cc

@@ -218,6 +218,8 @@ void GradNodeBase::SetGradOutMeta(const paddle::experimental::Tensor& fwd_in,
   // Set Stop_gradient
   if (fwd_in_meta) {
     meta.SetStopGradient(fwd_in_meta->StopGradient());
+  } else {
+    meta.SetStopGradient(true);
   }
   // Set Adj Edges
   if (fwd_in_meta && !fwd_in_meta->StopGradient()) {

paddle/fluid/eager/utils.cc

@@ -454,16 +454,48 @@ void EagerUtils::FillZeroForEmptyGradInputs(
             grad_in_meta.HasTensorMeta(),
             paddle::platform::errors::Fatal(
                 "Unable to fill empty grad inputs due to empty GradSlotMeta"));
-
         const auto& tensor_meta = grad_in_meta.GetTensorMeta();
-        phi::Place place = grad_in_meta.GetPlace();
-
         auto tensor_with_zero = paddle::experimental::full(
-            phi::vectorize(tensor_meta.dims), 0.0, tensor_meta.dtype, place);
+            phi::vectorize(tensor_meta.dims), 0.0, tensor_meta.dtype,
+            grad_in_meta.GetPlace());
         grad.set_impl(tensor_with_zero.impl());
       }
     }
   }
 }
 
+void EagerUtils::FillZeroForEmptyGradInput(
+    paddle::experimental::Tensor* in_grad, const GradSlotMeta& grad_in_meta) {
+  if (!in_grad->initialized()) {
+    PADDLE_ENFORCE(
+        grad_in_meta.HasTensorMeta(),
+        paddle::platform::errors::Fatal(
+            "Unable to fill empty grad inputs due to empty GradSlotMeta"));
+    const auto& tensor_meta = grad_in_meta.GetTensorMeta();
+    auto tensor_with_zero =
+        paddle::experimental::full(phi::vectorize(tensor_meta.dims), 0.0,
+                                   tensor_meta.dtype, grad_in_meta.GetPlace());
+    in_grad->set_impl(tensor_with_zero.impl());
+  }
+}
+
+void EagerUtils::FillZeroForEmptyOptionalGradInput(
+    paddle::experimental::Tensor* in_grad, const GradSlotMeta& grad_in_meta) {
+  if (!in_grad->initialized() && grad_in_meta.HasTensorMeta()) {
+    const auto& tensor_meta = grad_in_meta.GetTensorMeta();
+    auto tensor_with_zero =
+        paddle::experimental::full(phi::vectorize(tensor_meta.dims), 0.0,
+                                   tensor_meta.dtype, grad_in_meta.GetPlace());
+    in_grad->set_impl(tensor_with_zero.impl());
+  }
+}
+
+void EagerUtils::FillZeroForEmptyGradInput(
+    std::vector<paddle::experimental::Tensor>* in_grads,
+    const std::vector<GradSlotMeta>& grad_in_metas) {
+  for (size_t i = 0; i < in_grads->size(); i++) {
+    FillZeroForEmptyGradInput(&in_grads->at(i), grad_in_metas[i]);
+  }
+}
+
 }  // namespace egr

paddle/fluid/eager/utils.h

@@ -238,6 +238,13 @@ class EagerUtils {
                            kSlotSmallVectorSize>* out_grads,
       const paddle::small_vector<std::vector<GradSlotMeta>,
                                  kSlotSmallVectorSize>& grad_out_metas);
+  static void FillZeroForEmptyGradInput(paddle::experimental::Tensor* in_grad,
+                                        const GradSlotMeta& grad_in_meta);
+  static void FillZeroForEmptyOptionalGradInput(
+      paddle::experimental::Tensor* in_grad, const GradSlotMeta& grad_in_meta);
+  static void FillZeroForEmptyGradInput(
+      std::vector<paddle::experimental::Tensor>* in_grads,
+      const std::vector<GradSlotMeta>& grad_in_metas);
 };
 
 }  // namespace egr

paddle/phi/api/lib/kernel_dispatch.h

@@ -109,7 +109,12 @@ struct KernelKeyParser : ArgsIterator<KernelKeyParser> {
     }
   }
 
-  void operator()(const Tensor& x) { AssignKernelKeySet(*x.impl()); }
+  void operator()(const Tensor& x) {
+    const auto* tensor = x.impl().get();
+    if (tensor) {
+      AssignKernelKeySet(*tensor);
+    }
+  }
 
   void operator()(const std::vector<Tensor>& x) {
     const phi::TensorBase& tensor = *x.at(0).impl();

paddle/phi/kernels/activation_grad_kernel.h

@@ -17,6 +17,7 @@ limitations under the License. */
 #include "paddle/phi/common/scalar.h"
 #include "paddle/phi/core/dense_tensor.h"
 #include "paddle/phi/infermeta/unary.h"
+#include "paddle/utils/optional.h"
 
 namespace phi {
 
@@ -136,7 +137,7 @@ void SigmoidTripleGradKernel(const Context& dev_ctx,
                              const DenseTensor& dout,
                              const DenseTensor& ddx,
                              const DenseTensor& d_dout_new,
-                             const DenseTensor& d_ddout,
+                             paddle::optional<const DenseTensor&> d_ddout,
                              DenseTensor* d_out_new,
                              DenseTensor* d_dout,
                              DenseTensor* d_ddx);

paddle/phi/kernels/funcs/activation_functor.h

@@ -1428,16 +1428,19 @@ struct SigmoidTripleGradFunctor : public BaseActivationFunctor<T> {
         GET_DATA_SAFELY(Out, "Input", "Out", "SigmoidTripleGrad"));
     auto dout = EigenVector<T>::Flatten(
         GET_DATA_SAFELY(dOut, "Input", "DOut", "SigmoidTripleGrad"));
-    auto d_ddOut = EigenVector<T>::Flatten(
-        GET_DATA_SAFELY(d_DDOut, "Input", "D_DDOut", "SigmoidTripleGrad"));
     auto d_dOutNew = EigenVector<T>::Flatten(GET_DATA_SAFELY(
         d_dOut_New, "Input", "D_DOut_New", "SigmoidTripleGrad"));
 
     if (d_Out_New) {
       auto d_OutNew = EigenVector<T>::Flatten(GET_DATA_SAFELY(
           d_Out_New, "Output", "D_OutNew", "SigmoidTripleGrad"));
-      d_OutNew.device(*d) = (ddx - static_cast<T>(2) * out * ddx) * d_ddOut -
-                            static_cast<T>(2) * dout * ddx * d_dOutNew;
+      d_OutNew.device(*d) = -static_cast<T>(2) * dout * ddx * d_dOutNew;
+      if (d_DDOut) {
+        auto d_ddOut = EigenVector<T>::Flatten(
+            GET_DATA_SAFELY(d_DDOut, "Input", "D_DDOut", "SigmoidTripleGrad"));
+        d_OutNew.device(*d) =
+            (ddx - static_cast<T>(2) * out * ddx) * d_ddOut + d_OutNew;
+      }
     }
     if (d_d_Out) {
       auto d_dOut = EigenVector<T>::Flatten(
@@ -1449,8 +1452,12 @@ struct SigmoidTripleGradFunctor : public BaseActivationFunctor<T> {
       auto d_ddx = EigenVector<T>::Flatten(
           GET_DATA_SAFELY(d_DDx, "Output", "D_DDx", "SigmoidTripleGrad"));
       d_ddx.device(*d) =
-          (static_cast<T>(1) - out) * out * d_ddOut +
           (static_cast<T>(1) - static_cast<T>(2) * out) * dout * d_dOutNew;
+      if (d_DDOut) {
+        auto d_ddOut = EigenVector<T>::Flatten(
+            GET_DATA_SAFELY(d_DDOut, "Input", "D_DDOut", "SigmoidTripleGrad"));
+        d_ddx.device(*d) = d_ddx + (static_cast<T>(1) - out) * out * d_ddOut;
+      }
     }
   }
   static constexpr ActBwdOpFwdDeps FwdDeps() {

paddle/phi/kernels/impl/activation_grad_impl.h

@@ -265,7 +265,7 @@ void SigmoidTripleGradKernel(const Context& dev_ctx,
                              const DenseTensor& dout,
                              const DenseTensor& ddx,
                              const DenseTensor& d_dout_new,
-                             const DenseTensor& d_ddout,
+                             paddle::optional<const DenseTensor&> d_ddout,
                              DenseTensor* d_out_new,
                              DenseTensor* d_dout,
                              DenseTensor* d_ddx) {
@@ -274,11 +274,11 @@ void SigmoidTripleGradKernel(const Context& dev_ctx,
     dev_ctx.template Alloc<T>(d_dout);
   }
   if (d_out_new) {
-    d_dout->Resize(out.dims());
+    d_out_new->Resize(out.dims());
     dev_ctx.template Alloc<T>(d_out_new);
   }
   if (d_ddx) {
-    d_dout->Resize(ddx.dims());
+    d_ddx->Resize(ddx.dims());
     dev_ctx.template Alloc<T>(d_ddx);
   }
   funcs::SigmoidTripleGradFunctor<T> functor;
@@ -286,7 +286,7 @@ void SigmoidTripleGradKernel(const Context& dev_ctx,
           &out,
           &ddx,
           &dout,
-          &d_ddout,
+          d_ddout.get_ptr(),
           &d_dout_new,
           d_dout,
           d_out_new,

python/paddle/utils/code_gen/backward.yaml

@@ -1655,6 +1655,7 @@
     param : [out, fwd_grad_out, grad_grad_x]
   kernel :
     func : sigmoid_triple_grad
+  optional : grad_grad_out_grad
 
 - backward_api : silu_grad
   forward : silu (Tensor x) -> Tensor(out)