[Phi] Support setting size of vector<Tensor> for out in yaml (#41576)

* support setting vector out size in yaml

* support setting size of vector<tensor> for out in yaml

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

@@ -226,7 +226,7 @@ def ParseYamlReturns(string):
     returns = [x.strip() for x in string.strip().split(",")]
 
     for i in range(len(returns)):
-        ret = returns[i]
+        ret = returns[i].split("{")[0].strip()
 
         ret_name = ""
         if "(" in ret and ")" in ret:

paddle/phi/api/lib/api_custom_impl.cc

@@ -297,10 +297,10 @@ std::vector<Tensor> split_impl(const Tensor& x,
 
   // Calculate the number of out tensors
   size_t out_number;
-  if (num_or_sections.GetData().size() == 1) {
+  if (num_or_sections.size() == 1) {
     out_number = num_or_sections.GetData()[0];
   } else {
-    out_number = num_or_sections.GetData().size();
+    out_number = num_or_sections.size();
   }
 
   std::vector<Tensor> out;
@@ -475,54 +475,6 @@ std::tuple<Tensor, Tensor, Tensor> momentum_impl(
   return api_output;
 }
 
-std::vector<Tensor> unbind_impl(const Tensor& input, int axis) {
-  auto kernel_key_set = ParseKernelKeyByInputArgs(input);
-  auto kernel_key = kernel_key_set.GetHighestPriorityKernelKey();
-
-  Backend kernel_backend = kernel_key.backend();
-  DataLayout kernel_layout = kernel_key.layout();
-  DataType kernel_data_type = kernel_key.dtype();
-
-  auto kernel = phi::KernelFactory::Instance().SelectKernelOrThrowError(
-      "unbind", {kernel_backend, kernel_layout, kernel_data_type});
-  VLOG(6) << "unbind API kernel key: [" << kernel_backend << ", "
-          << kernel_layout << ", " << kernel_data_type << "]";
-  VLOG(6) << "unbind API kernel: " << kernel;
-
-  auto* dev_ctx = GetDeviceContextByBackend(kernel_backend);
-
-  auto dense_input = PrepareData(input, kernel.InputAt(0), {});
-
-  // Calculate the number of out tensors
-  auto input_shape = input.dims();
-  if (axis < 0) {
-    axis = input_shape.size() + axis;
-  }
-  auto out_num = input_shape[axis];
-
-  std::vector<Tensor> out;
-  auto dense_outs = SetKernelOutput(out_num, kernel_backend, &out);
-  std::vector<phi::MetaTensor> meta_outs;
-  meta_outs.reserve(out_num);
-  std::vector<phi::MetaTensor*> meta_out_ptrs;
-  meta_out_ptrs.reserve(out_num);
-  for (int64_t i = 0; i < out_num; ++i) {
-    meta_outs.push_back(dense_outs[i]);
-    meta_out_ptrs.push_back(&meta_outs.back());
-  }
-
-  phi::UnbindInferMeta(MakeMetaTensor(*dense_input), axis, meta_out_ptrs);
-
-  using kernel_signature = void (*)(const phi::DeviceContext&,
-                                    const phi::DenseTensor&,
-                                    int,
-                                    std::vector<phi::DenseTensor*>&);
-  auto* kernel_fn = kernel.GetVariadicKernelFn<kernel_signature>();
-  (*kernel_fn)(*dev_ctx, *dense_input, axis, dense_outs);
-
-  return out;
-}
-
 ////////////////// Backward(grad) api impls //////////////////////
 
 // TODO(chenweihang):  the original sum grad op can support higher-level
@@ -700,71 +652,6 @@ std::tuple<Tensor, Tensor, Tensor, Tensor, Tensor, Tensor> batch_norm_impl(
   return api_output;
 }
 
-std::vector<Tensor> concat_grad_impl(const std::vector<Tensor>& x,
-                                     const Tensor& out_grad,
-                                     const Scalar& axis) {
-  auto kernel_key_set = ParseKernelKeyByInputArgs(out_grad);
-  auto kernel_key = kernel_key_set.GetHighestPriorityKernelKey();
-
-  Backend kernel_backend = kernel_key.backend();
-  DataLayout kernel_layout = kernel_key.layout();
-  DataType kernel_data_type = kernel_key.dtype();
-
-  auto kernel = phi::KernelFactory::Instance().SelectKernelOrThrowError(
-      "concat_grad", {kernel_backend, kernel_layout, kernel_data_type});
-  VLOG(6) << "concat_grad API kernel key: [" << kernel_backend << ", "
-          << kernel_layout << ", " << kernel_data_type << "]";
-  VLOG(6) << "concat_grad API kernel: " << kernel;
-
-  auto* dev_ctx = GetDeviceContextByBackend(kernel_backend);
-
-  // std::unique_ptr<std::vector<phi::DenseTensor>>
-  auto dense_x = PrepareData(x, kernel.InputAt(0), {});
-  auto dense_out_grad = PrepareData(out_grad, kernel.InputAt(1), {});
-
-  // Calculate the number of out tensors
-  size_t out_number = x.size();
-  std::vector<Tensor> x_grad;
-  auto dense_x_grad = SetKernelOutput(out_number, kernel_backend, &x_grad);
-
-  std::vector<phi::MetaTensor> meta_x;
-  meta_x.reserve(x.size());
-  std::vector<phi::MetaTensor*> meta_x_ptrs;
-  meta_x_ptrs.reserve(x.size());
-  for (const auto& t : *dense_x) {
-    meta_x.push_back(t);
-    meta_x_ptrs.push_back(&meta_x.back());
-  }
-
-  std::vector<phi::MetaTensor> meta_x_grad;
-  meta_x_grad.reserve(x.size());
-  std::vector<phi::MetaTensor*> meta_x_grad_ptrs;
-  meta_x_grad_ptrs.reserve(x.size());
-  for (size_t i = 0; i < out_number; ++i) {
-    meta_x_grad.push_back(*dense_x_grad[i]);
-    meta_x_grad_ptrs.push_back(&meta_x_grad.back());
-  }
-
-  phi::UnchangedMultiInferMeta(meta_x_ptrs, meta_x_grad_ptrs);
-
-  std::vector<const phi::DenseTensor*> dense_x_ptr;
-  dense_x_ptr.reserve(x.size());
-  for (const auto& t : *dense_x) {
-    dense_x_ptr.push_back(&t);
-  }
-
-  using kernel_signature = void (*)(const platform::DeviceContext&,
-                                    const std::vector<const phi::DenseTensor*>&,
-                                    const phi::DenseTensor&,
-                                    const phi::Scalar&,
-                                    std::vector<phi::DenseTensor*>);
-  auto* kernel_fn = kernel.GetVariadicKernelFn<kernel_signature>();
-  (*kernel_fn)(
-      *dev_ctx, dense_x_ptr, *dense_out_grad, phi::Scalar(axis), dense_x_grad);
-
-  return x_grad;
-}
-
 Tensor imag_grad_impl(const Tensor& out_grad) {
   phi::KernelKey kernel_key{ParseBackend(out_grad),
                             out_grad.layout(),
@@ -821,328 +708,5 @@ Tensor real_grad_impl(const Tensor& out_grad) {
   return out;
 }
 
-std::vector<Tensor> stack_grad_impl(const std::vector<Tensor>& x,
-                                    const Tensor& out_grad,
-                                    int axis) {
-  auto kernel_key_set = ParseKernelKeyByInputArgs(out_grad);
-  auto kernel_key = kernel_key_set.GetHighestPriorityKernelKey();
-
-  Backend kernel_backend = kernel_key.backend();
-  DataLayout kernel_layout = kernel_key.layout();
-  DataType kernel_data_type = kernel_key.dtype();
-
-  auto kernel = phi::KernelFactory::Instance().SelectKernelOrThrowError(
-      "stack_grad", {kernel_backend, kernel_layout, kernel_data_type});
-  VLOG(6) << "stack_grad API kernel key: [" << kernel_backend << ", "
-          << kernel_layout << ", " << kernel_data_type << "]";
-  VLOG(6) << "stack_grad API kernel: " << kernel;
-
-  auto* dev_ctx = GetDeviceContextByBackend(kernel_backend);
-
-  auto dense_out_grad = PrepareData(out_grad, kernel.InputAt(0), {});
-
-  size_t out_number = x.size();
-  std::vector<Tensor> x_grad;
-  auto dense_x_grad = SetKernelOutput(out_number, kernel_backend, &x_grad);
-  std::vector<phi::MetaTensor> meta_x_grad;
-  meta_x_grad.reserve(out_number);
-  std::vector<phi::MetaTensor*> meta_x_grad_ptrs;
-  meta_x_grad_ptrs.reserve(out_number);
-  for (size_t i = 0; i < out_number; ++i) {
-    meta_x_grad.push_back(dense_x_grad[i]);
-    meta_x_grad_ptrs.push_back(&meta_x_grad.back());
-  }
-
-  phi::StackGradInferMeta(
-      MakeMetaTensor(*dense_out_grad), axis, meta_x_grad_ptrs);
-
-  using kernel_signature = void (*)(const platform::DeviceContext&,
-                                    const phi::DenseTensor&,
-                                    int axis,
-                                    std::vector<phi::DenseTensor*>);
-  auto* kernel_fn = kernel.GetVariadicKernelFn<kernel_signature>();
-  (*kernel_fn)(*dev_ctx, *dense_out_grad, axis, dense_x_grad);
-
-  return x_grad;
-}
-
-std::vector<Tensor> meshgrid_impl(const std::vector<Tensor>& inputs) {
-  Backend kernel_backend = Backend::UNDEFINED;
-  DataLayout kernel_layout = DataLayout::UNDEFINED;
-  DataType kernel_data_type = DataType::UNDEFINED;
-
-  if (kernel_backend == Backend::UNDEFINED ||
-      kernel_layout == DataLayout::UNDEFINED ||
-      kernel_data_type == DataType::UNDEFINED) {
-    auto kernel_key_set = ParseKernelKeyByInputArgs(inputs);
-    auto kernel_key = kernel_key_set.GetHighestPriorityKernelKey();
-    if (kernel_backend == Backend::UNDEFINED) {
-      kernel_backend = kernel_key.backend();
-    }
-    if (kernel_layout == DataLayout::UNDEFINED) {
-      kernel_layout = kernel_key.layout();
-    }
-    if (kernel_data_type == DataType::UNDEFINED) {
-      kernel_data_type = kernel_key.dtype();
-    }
-  }
-
-  const auto& kernel = phi::KernelFactory::Instance().SelectKernelOrThrowError(
-      "meshgrid", {kernel_backend, kernel_layout, kernel_data_type});
-  VLOG(6) << "meshgrid API kernel key: [" << kernel_backend << ", "
-          << kernel_layout << ", " << kernel_data_type << "]";
-  VLOG(6) << "meshgrid API kernel: " << kernel;
-
-  auto* dev_ctx = GetDeviceContextByBackend(kernel_backend);
-
-  auto input_inputs_vec = PrepareData(inputs, kernel.InputAt(0), {});
-  std::vector<const phi::DenseTensor*> input_inputs(input_inputs_vec->size());
-  for (size_t i = 0; i < input_inputs.size(); ++i) {
-    input_inputs[i] = &input_inputs_vec->at(i);
-  }
-
-  auto x_meta_vec = MakeMetaTensor(input_inputs);
-  std::vector<phi::MetaTensor*> inputs_metas(x_meta_vec.size());
-  for (size_t i = 0; i < x_meta_vec.size(); ++i) {
-    inputs_metas[i] = &x_meta_vec[i];
-  }
-
-  // Calculate the number of out tensors
-  size_t out_number = inputs.size();
-
-  std::vector<Tensor> out;
-  auto dense_outs = SetKernelOutput(out_number, kernel_backend, &out);
-
-  std::vector<phi::MetaTensor> meta_outs;
-  meta_outs.reserve(out_number);
-  std::vector<phi::MetaTensor*> meta_out_ptrs;
-  meta_out_ptrs.reserve(out_number);
-  for (size_t i = 0; i < out_number; ++i) {
-    meta_outs.push_back(dense_outs[i]);
-    meta_out_ptrs.push_back(&meta_outs.back());
-  }
-  phi::MeshgridInferMeta(inputs_metas, meta_out_ptrs);
-
-  using kernel_signature = void (*)(const platform::DeviceContext&,
-                                    const std::vector<const phi::DenseTensor*>&,
-                                    std::vector<phi::DenseTensor*>&);
-  auto* kernel_fn = kernel.GetVariadicKernelFn<kernel_signature>();
-  (*kernel_fn)(*dev_ctx, input_inputs, dense_outs);
-
-  return out;
-}
-
-std::vector<Tensor> meshgrid_grad_impl(
-    const std::vector<Tensor>& inputs,
-    const std::vector<Tensor>& outputs_grad) {
-  Backend kernel_backend = Backend::UNDEFINED;
-  DataLayout kernel_layout = DataLayout::UNDEFINED;
-  DataType kernel_data_type = DataType::UNDEFINED;
-
-  if (kernel_backend == Backend::UNDEFINED ||
-      kernel_layout == DataLayout::UNDEFINED ||
-      kernel_data_type == DataType::UNDEFINED) {
-    auto kernel_key_set = ParseKernelKeyByInputArgs(inputs, outputs_grad);
-    auto kernel_key = kernel_key_set.GetHighestPriorityKernelKey();
-    if (kernel_backend == Backend::UNDEFINED) {
-      kernel_backend = kernel_key.backend();
-    }
-    if (kernel_layout == DataLayout::UNDEFINED) {
-      kernel_layout = kernel_key.layout();
-    }
-    if (kernel_data_type == DataType::UNDEFINED) {
-      kernel_data_type = kernel_key.dtype();
-    }
-  }
-
-  const auto& kernel = phi::KernelFactory::Instance().SelectKernelOrThrowError(
-      "meshgrid_grad", {kernel_backend, kernel_layout, kernel_data_type});
-  VLOG(6) << "meshgrid_grad API kernel key: [" << kernel_backend << ", "
-          << kernel_layout << ", " << kernel_data_type << "]";
-  VLOG(6) << "meshgrid_grad API kernel: " << kernel;
-
-  auto* dev_ctx = GetDeviceContextByBackend(kernel_backend);
-
-  auto input_inputs_vec = PrepareData(inputs, kernel.InputAt(0), {});
-  std::vector<const phi::DenseTensor*> input_inputs(input_inputs_vec->size());
-  for (size_t i = 0; i < input_inputs.size(); ++i) {
-    input_inputs[i] = &input_inputs_vec->at(i);
-  }
-  auto input_outputs_grad_vec =
-      PrepareData(outputs_grad, kernel.InputAt(1), {});
-  std::vector<const phi::DenseTensor*> input_outputs_grad(
-      input_outputs_grad_vec->size());
-  for (size_t i = 0; i < input_outputs_grad.size(); ++i) {
-    input_outputs_grad[i] = &input_outputs_grad_vec->at(i);
-  }
-
-  size_t out_number = inputs.size();
-  std::vector<Tensor> api_output;
-  auto kernel_out = SetKernelOutput(out_number, kernel_backend, &api_output);
-
-  auto inputs_meta_vec = MakeMetaTensor(input_inputs);
-  std::vector<phi::MetaTensor*> inputs_metas(inputs_meta_vec.size());
-  for (size_t i = 0; i < inputs_meta_vec.size(); ++i) {
-    inputs_metas[i] = &inputs_meta_vec[i];
-  }
-
-  auto outputs_grad_meta_vec = MakeMetaTensor(input_outputs_grad);
-  std::vector<phi::MetaTensor*> outputs_grad_metas(
-      outputs_grad_meta_vec.size());
-  for (size_t i = 0; i < outputs_grad_meta_vec.size(); ++i) {
-    outputs_grad_metas[i] = &outputs_grad_meta_vec[i];
-  }
-
-  std::vector<phi::MetaTensor> meta_outs;
-  meta_outs.reserve(out_number);
-  std::vector<phi::MetaTensor*> meta_out_ptrs;
-  meta_out_ptrs.reserve(out_number);
-  for (size_t i = 0; i < out_number; ++i) {
-    meta_outs.push_back(kernel_out[i]);
-    meta_out_ptrs.push_back(&meta_outs.back());
-  }
-
-  phi::MeshgridGradInferMeta(inputs_metas, outputs_grad_metas, meta_out_ptrs);
-
-  using kernel_signature = void (*)(const platform::DeviceContext&,
-                                    const std::vector<const phi::DenseTensor*>&,
-                                    const std::vector<const phi::DenseTensor*>&,
-                                    std::vector<phi::DenseTensor*>&);
-  auto* kernel_fn = kernel.GetVariadicKernelFn<kernel_signature>();
-  (*kernel_fn)(*dev_ctx, input_inputs, input_outputs_grad, kernel_out);
-
-  return api_output;
-}
-
-std::vector<Tensor> multi_dot_grad_impl(const std::vector<Tensor>& x,
-                                        const Tensor& out_grad) {
-  Backend kernel_backend = Backend::UNDEFINED;
-  DataLayout kernel_layout = DataLayout::UNDEFINED;
-  DataType kernel_data_type = DataType::UNDEFINED;
-
-  if (kernel_backend == Backend::UNDEFINED ||
-      kernel_layout == DataLayout::UNDEFINED ||
-      kernel_data_type == DataType::UNDEFINED) {
-    auto kernel_key_set = ParseKernelKeyByInputArgs(x, out_grad);
-    auto kernel_key = kernel_key_set.GetHighestPriorityKernelKey();
-    if (kernel_backend == Backend::UNDEFINED) {
-      kernel_backend = kernel_key.backend();
-    }
-    if (kernel_layout == DataLayout::UNDEFINED) {
-      kernel_layout = kernel_key.layout();
-    }
-    if (kernel_data_type == DataType::UNDEFINED) {
-      kernel_data_type = kernel_key.dtype();
-    }
-  }
-
-  VLOG(6) << "multi_dot_grad API kernel key: [" << kernel_backend << ", "
-          << kernel_layout << ", " << kernel_data_type << "]";
-  const auto& kernel = phi::KernelFactory::Instance().SelectKernelOrThrowError(
-      "multi_dot_grad", {kernel_backend, kernel_layout, kernel_data_type});
-  VLOG(6) << "multi_dot_grad API kernel: " << kernel;
-
-  auto* dev_ctx = GetDeviceContextByBackend(kernel_backend);
-
-  auto input_x_vec = PrepareData(x, kernel.InputAt(0), {});
-  std::vector<const phi::DenseTensor*> input_x(input_x_vec->size());
-  for (size_t i = 0; i < input_x.size(); ++i) {
-    input_x[i] = &input_x_vec->at(i);
-  }
-  auto input_out_grad = PrepareData(out_grad, kernel.InputAt(1), {});
-
-  size_t out_number = input_x.size();
-  std::vector<Tensor> api_output;
-  auto kernel_out = SetKernelOutput(out_number, kernel_backend, &api_output);
-
-  auto x_meta_vec = MakeMetaTensor(input_x);
-  std::vector<phi::MetaTensor*> x_metas(x_meta_vec.size());
-  for (size_t i = 0; i < x_meta_vec.size(); ++i) {
-    x_metas[i] = &x_meta_vec[i];
-  }
-
-  std::vector<phi::MetaTensor> meta_outs;
-  meta_outs.reserve(out_number);
-  std::vector<phi::MetaTensor*> meta_out_ptrs;
-  meta_out_ptrs.reserve(out_number);
-  for (size_t i = 0; i < out_number; ++i) {
-    meta_outs.push_back(kernel_out[i]);
-    meta_out_ptrs.push_back(&meta_outs.back());
-  }
-
-  phi::MultiDotGradInferMeta(
-      x_metas, MakeMetaTensor(*input_out_grad), meta_out_ptrs);
-
-  using kernel_signature = void (*)(const platform::DeviceContext&,
-                                    const std::vector<const phi::DenseTensor*>&,
-                                    const phi::DenseTensor&,
-                                    std::vector<phi::DenseTensor*>&);
-  auto* kernel_fn = kernel.GetVariadicKernelFn<kernel_signature>();
-  (*kernel_fn)(*dev_ctx, input_x, *input_out_grad, kernel_out);
-
-  return api_output;
-}
-
-std::vector<Tensor> multiplex_grad_impl(const std::vector<Tensor>& inputs,
-                                        const Tensor& ids,
-                                        const Tensor& out_grad) {
-  Backend kernel_backend = Backend::UNDEFINED;
-  DataLayout kernel_layout = DataLayout::UNDEFINED;
-  DataType kernel_data_type = DataType::UNDEFINED;
-
-  if (kernel_backend == Backend::UNDEFINED ||
-      kernel_layout == DataLayout::UNDEFINED ||
-      kernel_data_type == DataType::UNDEFINED) {
-    auto kernel_key_set = ParseKernelKeyByInputArgs(out_grad);
-    auto kernel_key = kernel_key_set.GetHighestPriorityKernelKey();
-    if (kernel_backend == Backend::UNDEFINED) {
-      kernel_backend = kernel_key.backend();
-    }
-    if (kernel_layout == DataLayout::UNDEFINED) {
-      kernel_layout = kernel_key.layout();
-    }
-    if (kernel_data_type == DataType::UNDEFINED) {
-      kernel_data_type = kernel_key.dtype();
-    }
-  }
-
-  VLOG(6) << "multiplex_grad API kernel key: [" << kernel_backend << ", "
-          << kernel_layout << ", " << kernel_data_type << "]";
-  const auto& kernel = phi::KernelFactory::Instance().SelectKernelOrThrowError(
-      "multiplex_grad", {kernel_backend, kernel_layout, kernel_data_type});
-  VLOG(6) << "multiplex_grad API kernel: " << kernel;
-
-  auto* dev_ctx = GetDeviceContextByBackend(kernel_backend);
-
-  auto input_ids = PrepareData(ids, kernel.InputAt(0), {});
-  auto input_out_grad = PrepareData(out_grad, kernel.InputAt(1), {});
-
-  auto out_number = inputs.size();
-  std::vector<Tensor> api_output;
-  auto kernel_out = SetKernelOutput(out_number, kernel_backend, &api_output);
-
-  std::vector<phi::MetaTensor> meta_outs;
-  meta_outs.reserve(out_number);
-  std::vector<phi::MetaTensor*> meta_out_ptrs;
-  meta_out_ptrs.reserve(out_number);
-  for (size_t i = 0; i < out_number; ++i) {
-    meta_outs.push_back(kernel_out[i]);
-    meta_out_ptrs.push_back(&meta_outs.back());
-  }
-
-  phi::MultiplexGradInferMeta(MakeMetaTensor(*input_ids),
-                              MakeMetaTensor(*input_out_grad),
-                              meta_out_ptrs);
-
-  using kernel_signature = void (*)(const platform::DeviceContext&,
-                                    const phi::DenseTensor&,
-                                    const phi::DenseTensor&,
-                                    std::vector<phi::DenseTensor*>&);
-  auto* kernel_fn = kernel.GetVariadicKernelFn<kernel_signature>();
-  (*kernel_fn)(*dev_ctx, *input_ids, *input_out_grad, kernel_out);
-
-  return api_output;
-}
-
 }  // namespace experimental
 }  // namespace paddle

paddle/phi/api/lib/api_custom_impl.h

@@ -30,6 +30,20 @@ namespace experimental {
 
 ////////////////// Forward api impls //////////////////////
 
+std::tuple<Tensor, Tensor, Tensor, Tensor, Tensor, Tensor> batch_norm_impl(
+    const Tensor& x,
+    const Tensor& scale,
+    const Tensor& bias,
+    const Tensor& mean,
+    const Tensor& variance,
+    float momentum,
+    float epsilon,
+    const std::string& data_layout,
+    bool is_test,
+    bool use_global_stats,
+    bool trainable_statistics,
+    bool fuse_with_relu);
+
 Tensor conv2d_impl(const Tensor& input,
                    const Tensor& filter,
                    const std::vector<int>& strides,
@@ -62,8 +76,6 @@ std::vector<Tensor> split_impl(const Tensor& x,
                                const IntArray& num_or_sections,
                                const Scalar& axis);
 
-std::vector<Tensor> meshgrid_impl(const std::vector<Tensor>& inputs);
-
 std::tuple<Tensor, Tensor, Tensor> momentum_impl(
     const Tensor& param,
     const Tensor& grad,
@@ -77,49 +89,14 @@ std::tuple<Tensor, Tensor, Tensor> momentum_impl(
     bool multi_precision,
     float rescale_grad);
 
-std::vector<Tensor> unbind_impl(const Tensor& input, int axis);
-
 ////////////////// Backward(grad) api impls //////////////////////
 
 std::vector<Tensor> add_n_grad_impl(const std::vector<Tensor>& x,
                                     const Tensor& out_grad);
 
-std::tuple<Tensor, Tensor, Tensor, Tensor, Tensor, Tensor> batch_norm_impl(
-    const Tensor& x,
-    const Tensor& scale,
-    const Tensor& bias,
-    const Tensor& mean,
-    const Tensor& variance,
-    float momentum,
-    float epsilon,
-    const std::string& data_layout,
-    bool is_test,
-    bool use_global_stats,
-    bool trainable_statistics,
-    bool fuse_with_relu);
-
-/************************   backward api impl   ***************************/
-
-std::vector<Tensor> concat_grad_impl(const std::vector<Tensor>& x,
-                                     const Tensor& out_grad,
-                                     const Scalar& axis);
-
 Tensor imag_grad_impl(const Tensor& x);
 
 Tensor real_grad_impl(const Tensor& x);
 
-std::vector<Tensor> stack_grad_impl(const std::vector<Tensor>& x,
-                                    const Tensor& out_grad,
-                                    int axis);
-std::vector<Tensor> meshgrid_grad_impl(const std::vector<Tensor>& inputs,
-                                       const std::vector<Tensor>& outputs_grad);
-
-std::vector<Tensor> multi_dot_grad_impl(const std::vector<Tensor>& x,
-                                        const Tensor& out_grad);
-
-std::vector<Tensor> multiplex_grad_impl(const std::vector<Tensor>& inputs,
-                                        const Tensor& ids,
-                                        const Tensor& out_grad);
-
 }  // namespace experimental
 }  // namespace paddle

paddle/phi/api/lib/api_gen_utils.cc

@@ -76,6 +76,16 @@ std::vector<phi::MetaTensor> MakeMetaTensor(
   return meta_tensors;
 }
 
+std::vector<phi::MetaTensor> MakeMetaTensor(
+    const std::vector<phi::DenseTensor*>& tensors) {
+  std::vector<phi::MetaTensor> meta_tensors;
+  meta_tensors.reserve(tensors.size());
+  for (auto* t : tensors) {
+    meta_tensors.emplace_back(*t);
+  }
+  return meta_tensors;
+}
+
 phi::MetaTensor MakeMetaTensor(const phi::SelectedRows& tensor) {
   return phi::MetaTensor(tensor);
 }

paddle/phi/api/lib/api_gen_utils.h

@@ -53,6 +53,9 @@ phi::MetaTensor MakeMetaTensor(const phi::DenseTensor& tensor);
 std::vector<phi::MetaTensor> MakeMetaTensor(
     const std::vector<const phi::DenseTensor*>& tensors);
 
+std::vector<phi::MetaTensor> MakeMetaTensor(
+    const std::vector<phi::DenseTensor*>& tensors);
+
 phi::MetaTensor MakeMetaTensor(const phi::SelectedRows& tensor);
 
 phi::MetaTensor MakeMetaTensor(const phi::StringTensor& tensor);

paddle/phi/common/int_array.h

@@ -96,6 +96,8 @@ class IntArrayBase {
   template <typename OtherT>
   IntArrayBase(const IntArrayBase<OtherT>& other) : array_(other.GetData()) {}
 
+  size_t size() const { return array_.size(); }
+
   const std::vector<int64_t>& GetData() const { return array_; }
 
  private:

python/paddle/utils/code_gen/api.yaml

@@ -1290,8 +1290,11 @@
 
 - api : meshgrid
   args : (Tensor[] inputs)
-  output : Tensor[]
-  invoke : meshgrid_impl(inputs)
+  output : Tensor[]{inputs.size()}
+  infer_meta :
+    func : MeshgridInferMeta
+  kernel :
+    func : meshgrid
   backward : meshgrid_grad
 
 - api : min
@@ -2059,8 +2062,11 @@
 
 - api : unbind
   args : (Tensor input, int axis)
-  output : Tensor[]
-  invoke : unbind_impl(input, axis)
+  output : Tensor[] {axis<0 ? input.dims()[input.dims().size()+axis]:input.dims()[axis]}
+  infer_meta :
+    func : UnbindInferMeta
+  kernel :
+    func : unbind
   backward : unbind_grad
 
 # unfold

python/paddle/utils/code_gen/api_base.py

@@ -31,6 +31,7 @@ class BaseAPI(object):
         # outputs:
         #     names : [], list of output names
         #     types : [], list of output types
+        #     out_size_expr : [], expression for getting size of vector<Tensor>
         #     return_type : Tensor, vector<Tensor>, ..., the return type of api
         # args_str:
         #     args_declare : "str" // str of function params with default value. Example: (..., bool flag=false)
@@ -67,11 +68,12 @@ class BaseAPI(object):
             ]
         inputs, attrs, args_str = self.parse_input_and_attr(
             api_name, api_item_yaml['args'], optional_vars)
-        output_type_list, output_names, return_type = self.parse_output(
+        output_type_list, output_names, out_size_expr, return_type = self.parse_output(
             api_name, api_item_yaml['output'])
         return inputs, attrs, {
             'names': output_names,
             'types': output_type_list,
+            'out_size_expr': out_size_expr,
             'return_type': return_type
         }, args_str, optional_vars
 
@@ -184,39 +186,36 @@ class BaseAPI(object):
                 'Tensor': 'Tensor',
                 'Tensor[]': 'std::vector<Tensor>'
             }
-            if re.search(r'\([a-zA-Z0-9_@]*\)', output_item):
-                result = re.search(
-                    r"(?P<out_type>[a-zA-Z0-9_[\]]+)\s*\((?P<name>[a-zA-Z0-9_@]+)\)",
-                    output_item)
-                out_type = result.group('out_type')
-                assert out_type in output_type_map, \
-                    f"{api_name} : Output type error: the output type only support Tensor and Tensor[], \
-                      but now is {out_type}."
-
-                return output_type_map[out_type], result.group('name')
-
-            else:
-                if output_item.strip() in output_type_map:
-                    return output_type_map[output_item.strip()], 'out'
-                else:
-                    raise ValueError(
-                        "{} : Output type error: the output type only support Tensor and Tensor[], \
-                      but now is {}.".format(api_name, output_item.strip()))
+            result = re.search(
+                r"(?P<out_type>[a-zA-Z0-9_[\]]+)\s*(?P<name>\([a-zA-Z0-9_@]+\))?\s*(?P<expr>\{[^\}]+\})?",
+                output_item)
+            assert result is not None, f"{api_name} : the output config parse error."
+            out_type = result.group('out_type')
+            assert out_type in output_type_map, \
+                f"{api_name} : Output type error: the output type only support Tensor and Tensor[], \
+                  but now is {out_type}."
+
+            out_name = 'out' if result.group('name') is None else result.group(
+                'name')[1:-1]
+            out_size_expr = None if result.group(
+                'expr') is None else result.group('expr')[1:-1]
+            return output_type_map[out_type], out_name, out_size_expr
 
         temp_list = output_config.split(',')
 
         if len(temp_list) == 1:
-            out_type, out_name = parse_output_item(temp_list[0])
-            return [out_type], [out_name], self.get_return_type([out_type])
+            out_type, out_name, size_expr = parse_output_item(temp_list[0])
+            return [out_type], [out_name], size_expr, self.get_return_type(
+                [out_type])
         else:
             out_type_list = []
             out_name_list = []
             for output_item in temp_list:
-                out_type, out_name = parse_output_item(output_item)
+                out_type, out_name, size_expr = parse_output_item(output_item)
                 out_type_list.append(out_type)
                 out_name_list.append(out_name)
 
-            return out_type_list, out_name_list, self.get_return_type(
+            return out_type_list, out_name_list, size_expr, self.get_return_type(
                 out_type_list)
 
     def parse_infer_meta(self, infer_meta_config):
@@ -462,9 +461,8 @@ PADDLE_API {self.gene_return_type_code()} {self.get_api_func_name() + '_'}({self
         attr_names = self.attrs['names']
         infer_meta = self.infer_meta
 
-        infer_meta_params = infer_meta[
-            'param'] + kernel_output_names if infer_meta[
-                'param'] is not None else input_names + attr_names + kernel_output_names
+        infer_meta_params = infer_meta['param'] if infer_meta[
+            'param'] is not None else input_names + attr_names
         # generate meta tensors
         meta_tensor_code = ""
         param_code = ""
@@ -500,11 +498,6 @@ PADDLE_API {self.gene_return_type_code()} {self.get_api_func_name() + '_'}({self
                     raise ValueError(
                         f"{self.api} : Param of infer_meta error : {self.inputs['input_info'][param]} type is not supported."
                     )
-            elif param in kernel_output_names:
-                meta_tensor_code = meta_tensor_code + code_indent + "  phi::MetaTensor " + param.replace(
-                    'kernel_', PREFIX_META_TENSOR_NAME) + "(" + param + ");\n"
-                param_code = param_code + "&" + param.replace(
-                    'kernel_', PREFIX_META_TENSOR_NAME) + ", "
             elif param in attr_names:
                 param_code = param_code + param + ", "
             elif isinstance(param, str):
@@ -514,6 +507,23 @@ PADDLE_API {self.gene_return_type_code()} {self.get_api_func_name() + '_'}({self
             else:
                 param_code = param_code + str(param) + ", "
 
+        for i, out_name in enumerate(kernel_output_names):
+            if self.outputs['types'][i] == 'std::vector<Tensor>':
+                meta_tensor_code = meta_tensor_code + f"""
+{code_indent}  auto {out_name}_{PREFIX_META_TENSOR_NAME}vec = MakeMetaTensor({out_name});
+{code_indent}  std::vector<phi::MetaTensor*> {out_name}_metas({out_name}_{PREFIX_META_TENSOR_NAME}vec.size());
+{code_indent}  for (size_t i = 0; i < {out_name}_{PREFIX_META_TENSOR_NAME}vec.size(); ++i) {{
+{code_indent}    {out_name}_metas[i] = &{out_name}_{PREFIX_META_TENSOR_NAME}vec[i];
+{code_indent}  }}"""
+
+                param_code = param_code + out_name + '_metas, '
+            else:
+                meta_tensor_code = meta_tensor_code + code_indent + "  phi::MetaTensor " + out_name.replace(
+                    'kernel_',
+                    PREFIX_META_TENSOR_NAME) + "(" + out_name + ");\n"
+                param_code = param_code + "&" + out_name.replace(
+                    'kernel_', PREFIX_META_TENSOR_NAME) + ", "
+
         param_code = param_code[:-2]
         return f"""{meta_tensor_code}
 {code_indent}  phi::{infer_meta['func']}({param_code});

python/paddle/utils/code_gen/api_gen.py

@@ -91,7 +91,16 @@ class ForwardAPI(BaseAPI):
                 0]] if inplace_flag and self.inplace_map is not None and self.outputs[
                     'names'][0] in self.inplace_map else ""
             output_create = f"""
-{code_indent}  {self.outputs['return_type']} api_output{inplace_assign};
+{code_indent}  {self.outputs['return_type']} api_output{inplace_assign};"""
+
+            if self.outputs['return_type'] == 'std::vector<Tensor>':
+                assert self.outputs['out_size_expr'] is not None, \
+                     f"{api_name}: The out size expr : '{{expr}}' should be set when output has Tensor[]. You can refer 'split' api."
+                output_create = output_create + f"""
+{code_indent}  auto kernel_out = {set_out_func}({self.outputs['out_size_expr']}, kernel_backend, &api_output);"""
+
+            else:
+                output_create = output_create + f"""
 {code_indent}  auto kernel_out = {set_out_func}(kernel_backend, &api_output);"""
 
             if not inplace_flag and self.view_map is not None and self.outputs[
@@ -113,7 +122,14 @@ class ForwardAPI(BaseAPI):
                     output_create = output_create + f"""
 {code_indent}  std::get<{i}>(api_output) = {self.inplace_map[self.outputs['names'][i]]};"""
 
-                output_create = output_create + f"""
+                if output_type_list[i] == 'std::vector<Tensor>':
+                    assert self.outputs['out_size_expr'][i] is not None, \
+                        f"{api_name}: The out size expr : '{{expr}}' should be set when output has Tensor[]. You can refer 'split' api."
+                    output_create = output_create + f"""
+{code_indent}  auto kernel_out_{i} = {set_out_func}({self.outputs['out_size_expr'][i]}, kernel_backend, &std::get<{i}>(api_output));"""
+
+                else:
+                    output_create = output_create + f"""
 {code_indent}  auto kernel_out_{i} = {set_out_func}(kernel_backend, &std::get<{i}>(api_output));"""
 
                 if not inplace_flag and self.view_map is not None and self.outputs[

python/paddle/utils/code_gen/backward.yaml

@@ -44,7 +44,7 @@
 - backward_api : add_n_grad
   forward : add_n (Tensor[] x) -> Tensor(out)
   args : (Tensor[] x, Tensor out_grad)
-  output : Tensor[](x_grad)
+  output : Tensor[](x_grad){x.size()}
   invoke : add_n_grad_impl(x, out_grad)
   no_need_buffer : x
 
@@ -215,8 +215,12 @@
 - backward_api : concat_grad
   forward : concat (Tensor[] x, Scalar axis) -> Tensor(out)
   args : (Tensor[] x, Tensor out_grad, Scalar axis = 0)
-  output : Tensor[](x_grad)
-  invoke : concat_grad_impl(x, out_grad, axis)
+  output : Tensor[](x_grad){x.size()}
+  infer_meta :
+    func : UnchangedMultiInferMeta
+    param : [x]
+  kernel :
+    func : concat_grad
   no_need_buffer : x
 
 - backward_api : conj_grad
@@ -944,8 +948,11 @@
 - backward_api : meshgrid_grad
   forward : meshgrid (Tensor[] inputs) -> Tensor[](outputs)
   args : (Tensor[] inputs, Tensor[] outputs_grad)
-  output : Tensor[](inputs_grad)
-  invoke : meshgrid_grad_impl(inputs, outputs_grad)
+  output : Tensor[](inputs_grad){inputs.size()}
+  infer_meta :
+    func : MeshgridGradInferMeta
+  kernel :
+    func : meshgrid_grad
 
 - backward_api : min_grad
   forward: min (Tensor x,  int64_t[] dims={},  bool keep_dim=false) -> Tensor(out)
@@ -1001,14 +1008,22 @@
 - backward_api : multi_dot_grad
   forward : multi_dot (Tensor[] x) -> Tensor(out)
   args : (Tensor[] x, Tensor out_grad)
-  output : Tensor[](x_grad)
-  invoke : multi_dot_grad_impl(x, out_grad)
+  output : Tensor[](x_grad) {x.size()}
+  infer_meta :
+    func : MultiDotGradInferMeta
+  kernel :
+    func : multi_dot_grad
 
 - backward_api : multiplex_grad
   forward : multiplex (Tensor[] ins, Tensor ids) -> Tensor(out)
   args : (Tensor[] ins, Tensor ids, Tensor out_grad)
-  output : Tensor[](ins_grad)
-  invoke : multiplex_grad_impl(ins, ids, out_grad)
+  output : Tensor[](ins_grad){ins.size()}
+  infer_meta :
+    func : MultiplexGradInferMeta
+    param : [ids, out_grad]
+  kernel :
+    func : multiplex_grad
+    param : [ids, out_grad]
 
 - backward_api : multiply_grad
   forward : multiply (Tensor x, Tensor y) -> Tensor(out)
@@ -1448,8 +1463,13 @@
 - backward_api : stack_grad
   forward : stack (Tensor[] x, int axis) -> Tensor(out)
   args : (Tensor[] x, Tensor out_grad, int axis)
-  output : Tensor[](x_grad)
-  invoke : stack_grad_impl(x, out_grad, axis)
+  output : Tensor[](x_grad){x.size()}
+  infer_meta :
+    func : StackGradInferMeta
+    param: [out_grad, axis]
+  kernel :
+    func : stack_grad
+    param : [out_grad, axis]
   no_need_buffer : x
 
 - backward_api : strided_slice_grad

python/paddle/utils/code_gen/backward_api_gen.py

@@ -35,7 +35,7 @@ class BackwardAPI(BaseAPI):
             r"(?P<api>[a-z][a-z0-9_]+)\s*(?P<args>\([^\)]+\))\s*->\s*(?P<outputs>.+)",
             forward_config)
         api = result.group('api')
-        _, outputs, _ = self.parse_output(self.api, result.group('outputs'))
+        _, outputs, _, _ = self.parse_output(self.api, result.group('outputs'))
         outputs = [item.split('@')[0] for item in outputs]
         fw_inputs, fw_attrs, _, = self.parse_input_and_attr(
             api, result.group('args'))
@@ -110,7 +110,16 @@ class BackwardAPI(BaseAPI):
                 0]] if inplace_flag and self.inplace_map is not None and self.outputs[
                     'names'][0] in self.inplace_map else ""
             output_create = f"""
-{code_indent}  {self.outputs['return_type']} api_output{inplace_assign};
+{code_indent}  {self.outputs['return_type']} api_output{inplace_assign};"""
+
+            if output_type_list[0] == 'std::vector<Tensor>':
+                assert self.outputs['out_size_expr'] is not None, \
+                     f"{api_name}: The out size expr : '{{expr}}' should be set when output has Tensor[]. You can refer 'split' api."
+                output_create = output_create + f"""
+{code_indent}  auto kernel_out = {set_out_func}({self.outputs['out_size_expr']}, kernel_backend, &api_output);"""
+
+            else:
+                output_create = output_create + f"""
 {code_indent}  auto kernel_out = {set_out_func}(kernel_backend, &api_output);"""
 
         elif len(output_type_list) > 1:
@@ -121,7 +130,6 @@ class BackwardAPI(BaseAPI):
                 kernel_output = kernel_output + f'kernel_out_{i}, '
                 output_names.append(f'kernel_out_{i}')
                 if out_type_item == 'Tensor':
-                    get_out_code = f'&api_output[{i}][0]'
                     if inplace_flag and self.inplace_map is not None and self.outputs[
                             'names'][i] in self.inplace_map:
                         output_create = output_create + f"""
@@ -131,6 +139,9 @@ class BackwardAPI(BaseAPI):
                         output_create = output_create + f"""
 {code_indent}  api_output[{i}].emplace_back();"""
 
+                    output_create = output_create + f"""
+{code_indent}  auto kernel_out_{i} = {set_out_func}(kernel_backend, &api_output[{i}][0]);"""
+
                 else:
                     get_out_code = f'&api_output[{i}]'
                     if inplace_flag and self.inplace_map is not None and self.outputs[
@@ -138,8 +149,10 @@ class BackwardAPI(BaseAPI):
                         output_create = output_create + f"""
 {code_indent}  api_output[{i}] = {self.inplace_map[self.outputs['names'][i]]};"""
 
-                output_create = output_create + f"""
-{code_indent}  auto kernel_out_{i} = {set_out_func}(kernel_backend, {get_out_code});"""
+                    assert self.outputs['out_size_expr'][i] is not None, \
+                        f"{api_name}: The out size expr : '{{expr}}' should be set when output has Tensor[]. You can refer 'split' api."
+                    output_create = output_create + f"""
+{code_indent}  auto kernel_out_{i} = {set_out_func}({self.outputs['out_size_expr'][i]}, kernel_backend, &api_output[{i}]);"""
 
             kernel_output = kernel_output[:-2]
         else: