[PTen] Add full kernel in pten (incomplete)

paddle/fluid/framework/operator.cc

@@ -1838,6 +1838,10 @@ pten::KernelContext OperatorWithKernel::BuildPtenKernelContext(
       if (std::type_index(attr.type()) == std::type_index(typeid(float))) {
         op_kernel_ctx.EmplaceBackAttr(
             std::move(pten::Scalar(BOOST_GET_CONST(float, attr))));
+      } else if (std::type_index(attr.type()) ==
+                 std::type_index(typeid(std::string))) {
+        op_kernel_ctx.EmplaceBackAttr(
+            std::move(pten::Scalar(BOOST_GET_CONST(std::string, attr))));
       } else {
         PADDLE_THROW(platform::errors::Unimplemented(
             "unsupported cast op attribute `%s` to Scalar when construct "

paddle/fluid/imperative/prepared_operator.cc

@@ -321,6 +321,10 @@ static pten::KernelContext BuildDygraphPtenKernelContext(
       if (std::type_index(attr.type()) == std::type_index(typeid(float))) {
         op_kernel_ctx.EmplaceBackAttr(
             std::move(pten::Scalar(BOOST_GET_CONST(float, attr))));
+      } else if (std::type_index(attr.type()) ==
+                 std::type_index(typeid(std::string))) {
+        op_kernel_ctx.EmplaceBackAttr(
+            std::move(pten::Scalar(BOOST_GET_CONST(std::string, attr))));
       } else {
         PADDLE_THROW(platform::errors::Unimplemented(
             "unsupported cast op attribute `%s` to Scalar when construct "

paddle/fluid/operators/fill_constant_op.cc

@@ -64,9 +64,51 @@ class FillConstantOp : public framework::OperatorWithKernel {
 
   framework::OpKernelType GetExpectedKernelType(
       const framework::ExecutionContext& ctx) const override {
-    return framework::OpKernelType(
+    framework::OpKernelType kt = framework::OpKernelType(
         framework::proto::VarType::Type(ctx.Attr<int>("dtype")),
         ctx.GetPlace());
+    // TODO(zyfncg) The force_cpu and place_type are conflicted, it's a issue
+    // lefted before, and we may merge them in the future.
+    // In order to invoke new fill_constant kernel, the place of OpKernelType
+    // will be setted by force_cpu and place_type here.
+    if (ctx.Attr<bool>("force_cpu")) {
+      kt.place_ = platform::CPUPlace();
+    }
+    auto place_type = ctx.Attr<int>("place_type");
+    if (place_type != -1) {
+      switch (place_type) {
+        case 0:
+          kt.place_ = platform::CPUPlace();
+          break;
+        case 1:
+        case 2:
+          kt.place_ = platform::CUDAPlace();
+          break;
+        case 3:
+          kt.place_ = platform::XPUPlace();
+          break;
+        default:
+          PADDLE_THROW(platform::errors::Unimplemented(
+              "Could NOT determine the place of variable, place_type = %d .",
+              place_type));
+      }
+    }
+
+    return kt;
+  }
+
+  framework::KernelSignature GetExpectedPtenKernelArgs(
+      const framework::ExecutionContext& ctx) const override {
+    if (!ctx.HasInput("ShapeTensor") &&
+        ctx.MultiInput<framework::Tensor>("ShapeTensorList").empty() &&
+        !ctx.HasInput("ValueTensor") &&
+        !ctx.OutputVar("Out")->IsType<framework::SelectedRows>()) {
+      const auto& str_value = ctx.Attr<std::string>("str_value");
+      std::string value = str_value.empty() ? "value" : "str_value";
+      return framework::KernelSignature("fill_constant.scalar", {}, {value},
+                                        {"Out"});
+    }
+    return framework::KernelSignature("fill_constant.unregistered", {}, {}, {});
   }
 };
 

paddle/pten/CMakeLists.txt

@@ -13,7 +13,7 @@ add_subdirectory(tests)
 # make an unity target for compile deps
 set(PTEN_DEPS convert_utils dense_tensor kernel_factory kernel_context)
 set(PTEN_DEPS ${PTEN_DEPS} math_cpu linalg_cpu creation_cpu manipulation_cpu)
-set(PTEN_DEPS ${PTEN_DEPS} unary binary)
+set(PTEN_DEPS ${PTEN_DEPS} nary unary binary)
 if(WITH_GPU OR WITH_ROCM)
   set(PTEN_DEPS ${PTEN_DEPS} math_cuda linalg_cuda creation_cuda manipulation_cuda)
 endif()

paddle/pten/api/include/creation.h

@@ -21,6 +21,12 @@
 namespace paddle {
 namespace experimental {
 
+Tensor full(const std::vector<int64_t>& shape,
+            const Scalar& value,
+            DataType dtype = DataType::FLOAT32,
+            Backend backend = Backend::CPU,
+            DataLayout layout = DataLayout::NCHW);
+
 Tensor full_like(const Tensor& x,
                  const Scalar& value,
                  DataType dtype = DataType::UNDEFINED);

paddle/pten/api/lib/creation.cc

@@ -26,14 +26,52 @@ limitations under the License. */
 namespace paddle {
 namespace experimental {
 
+Tensor full(const std::vector<int64_t>& shape,
+            const Scalar& value,
+            DataType dtype,
+            Backend backend,
+            DataLayout layout) {
+  // 1. Get kernel signature and kernel
+  pten::KernelKey kernel_key{backend, layout, dtype};
+  auto kernel = pten::KernelFactory::Instance().SelectKernelOrThrowError(
+      "fill_constant.scalar", kernel_key);
+
+  // 2. Get Device Context
+  auto* dev_ctx = GetDeviceContextByBackend(kernel_key.backend());
+  auto kernel_context = pten::KernelContext(*dev_ctx);
+
+  // 3. Auto data transform
+  kernel_context.EmplaceBackAttr(value);
+
+  // 4. InferShape
+  auto out_meta = pten::FullInferShape(shape, dtype, layout);
+
+  // 5. Prepare outputs
+  const auto allocator =
+      std::make_shared<paddle::experimental::DefaultAllocator>(
+          pten::TransToFluidPlace(kernel_key.backend()));
+  auto dense_out = std::make_shared<pten::DenseTensor>(allocator, out_meta);
+  kernel_context.EmplaceBackOutput(dense_out);
+  Tensor out;
+  out.set_impl(dense_out);
+
+  // 6. Call kernel
+  kernel(&kernel_context);
+
+  return out;
+}
+
 Tensor full_like(const Tensor& x,
                  const Scalar& value,
                  paddle::experimental::DataType dtype) {
   // 1. Get kernel signature and kernel
   auto kernel_key_set = ParseKernelKeyByInputArgs(x);
   auto kernel_key = kernel_key_set.GetHigestPriorityKernelKey();
   auto kernel = pten::KernelFactory::Instance().SelectKernelOrThrowError(
-      "fill_any_like", kernel_key);
+      "fill_any_like",
+      {kernel_key.backend(),
+       kernel_key.layout(),
+       dtype == DataType::UNDEFINED ? kernel_key.dtype() : dtype});
 
   // 2. Get Device Context
   auto* dev_ctx = GetDeviceContextByBackend(kernel_key.backend());

paddle/pten/common/scalar.h

@@ -34,6 +34,18 @@ class Scalar {
 
   Scalar(bool val) : tag(Tag::HAS_B) { data_.b = val; }  // NOLINT
 
+  Scalar(const std::string& str_value) : tag(Tag::HAS_D) {  // NOLINT
+    if (str_value == "inf") {
+      data_.d = std::numeric_limits<double>::infinity();
+    } else if (str_value == "-inf") {
+      data_.d = -std::numeric_limits<double>::infinity();
+    } else if (str_value == "nan") {
+      data_.d = std::numeric_limits<double>::quiet_NaN();
+    } else {
+      data_.d = std::stod(str_value);
+    }
+  }
+
   template <typename T>
   inline T to() const {
     switch (tag) {

paddle/pten/core/kernel_utils.h

@@ -207,6 +207,7 @@ struct KernelImpl<Return (*)(Args...), kernel_fn> {
   PT_SPECIALIZE_KernelCallHelper_FOR_ATTRIBUTE(int64_t);
   PT_SPECIALIZE_KernelCallHelper_FOR_ATTRIBUTE(paddle::platform::float16);
   PT_SPECIALIZE_KernelCallHelper_FOR_ATTRIBUTE(const Scalar&);
+  PT_SPECIALIZE_KernelCallHelper_FOR_ATTRIBUTE(const std::vector<int64_t>&);
 
   /* Output Helpers */
 

paddle/pten/include/infershape.h

@@ -16,4 +16,5 @@ limitations under the License. */
 
 // See Note: [ How do we organize the kernel directory ]
 #include "paddle/pten/infershape/binary.h"
+#include "paddle/pten/infershape/nary.h"
 #include "paddle/pten/infershape/unary.h"

paddle/pten/infershape/CMakeLists.txt

@@ -1,2 +1,3 @@
+cc_library(nary SRCS nary.cc DEPS convert_utils)
 cc_library(unary SRCS unary.cc DEPS convert_utils)
 cc_library(binary SRCS binary.cc DEPS convert_utils)

paddle/pten/infershape/nary.cc

@@ -0,0 +1,27 @@
+/* Copyright (c) 2021 PaddlePaddle Authors. All Rights Reserved.
+
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+
+    http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+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. */
+
+// See Note [ Why still include the fluid headers? ]
+#include "paddle/pten/infershape/nary.h"
+
+namespace pten {
+
+DenseTensorMeta FullInferShape(const std::vector<int64_t>& shape,
+                               DataType dtype,
+                               DataLayout layout) {
+  const auto& out_dims = paddle::framework::make_ddim(shape);
+  return {dtype, out_dims, layout};
+}
+
+}  // namespace pten

paddle/pten/infershape/nary.h

@@ -0,0 +1,34 @@
+/* Copyright (c) 2021 PaddlePaddle Authors. All Rights Reserved.
+
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+
+    http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+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. */
+
+#pragma once
+
+// See Note [ Why still include the fluid headers? ]
+#include "paddle/pten/core/tensor_meta.h"
+
+namespace pten {
+
+// Common InferShape Functions for 0-nary operators(no input tensor), The format
+// like:
+//
+//   1. DenseTensorMeta [OpName]InferShape( ...)
+//  NOTE: The name "InferShape" may be not appropriate. "InferMeta" may be good.
+//  Because functions in this file
+//  not only can infer shape, but alse need infer lod or other useful data.
+
+DenseTensorMeta FullInferShape(const std::vector<int64_t>& shape,
+                               DataType dtype,
+                               DataLayout layout);
+
+}  // namespace pten

paddle/pten/kernels/cpu/creation.cc

@@ -24,7 +24,38 @@ void FillAnyLike(const CPUContext& dev_ctx,
                  const DenseTensor& x,
                  const Scalar& val,
                  DenseTensor* out) {
-  eigen::fill<CPUContext, T>(dev_ctx, out, val.to<float>());
+  auto value = val.to<float>();
+  using CommonType = typename std::common_type<
+      float,
+      typename std::conditional<
+          std::is_same<T, paddle::platform::float16>::value,
+          float,
+          T>::type>::type;
+
+  auto common_type_value = static_cast<CommonType>(value);
+
+  PADDLE_ENFORCE_EQ(
+      (common_type_value >=
+       static_cast<CommonType>(std::numeric_limits<T>::lowest())) &&
+          (common_type_value <=
+           static_cast<CommonType>(std::numeric_limits<T>::max())),
+      true,
+      paddle::platform::errors::InvalidArgument(
+          "The filled value is out of range for target type, "
+          "current kernel type is %s, the range should between %f "
+          "and %f, but now value is %f.",
+          typeid(T).name(),
+          static_cast<CommonType>(std::numeric_limits<T>::lowest()),
+          static_cast<CommonType>(std::numeric_limits<T>::max()),
+          static_cast<float>(value)));
+  eigen::fill<CPUContext, T>(dev_ctx, out, value);
+}
+
+template <typename T>
+void FillConstant(const CPUContext& dev_ctx,
+                  const Scalar& val,
+                  DenseTensor* out) {
+  eigen::fill<CPUContext, T>(dev_ctx, out, val.to<T>());
 }
 
 }  // namespace pten
@@ -41,3 +72,19 @@ PT_REGISTER_KERNEL("fill_any_like",
                    int64_t,
                    bool,
                    paddle::platform::float16) {}
+
+PT_REGISTER_KERNEL("fill_constant.scalar",
+                   CPU,
+                   ANY,
+                   pten::FillConstant,
+                   float,
+                   double,
+                   uint8_t,
+                   int16_t,
+                   int,
+                   int64_t,
+                   bool,
+                   paddle::platform::float16,
+                   paddle::platform::bfloat16,
+                   paddle::platform::complex<float>,
+                   paddle::platform::complex<double>) {}

paddle/pten/kernels/cpu/creation.h

@@ -29,4 +29,9 @@ void FillAnyLike(const CPUContext& dev_ctx,
                  const Scalar& val,
                  DenseTensor* out);
 
+template <typename T>
+void FillConstant(const CPUContext& dev_ctx,
+                  const Scalar& val,
+                  DenseTensor* out);
+
 }  // namespace pten

paddle/pten/kernels/cuda/creation.cu

@@ -24,9 +24,41 @@ void FillAnyLike(const CUDAContext& dev_ctx,
                  const DenseTensor& x,
                  const Scalar& val,
                  DenseTensor* out) {
+  auto value = val.to<float>();
+  using CommonType = typename std::common_type<
+      float,
+      typename std::conditional<
+          std::is_same<T, paddle::platform::float16>::value,
+          float,
+          T>::type>::type;
+
+  auto common_type_value = static_cast<CommonType>(value);
+
+  PADDLE_ENFORCE_EQ(
+      (common_type_value >=
+       static_cast<CommonType>(std::numeric_limits<T>::lowest())) &&
+          (common_type_value <=
+           static_cast<CommonType>(std::numeric_limits<T>::max())),
+      true,
+      paddle::platform::errors::InvalidArgument(
+          "The filled value is out of range for target type, "
+          "current kernel type is %s, the range should between %f "
+          "and %f, but now value is %f.",
+          typeid(T).name(),
+          static_cast<CommonType>(std::numeric_limits<T>::lowest()),
+          static_cast<CommonType>(std::numeric_limits<T>::max()),
+          static_cast<float>(value)));
+
   eigen::fill<CUDAContext, T>(dev_ctx, out, val.to<float>());
 }
 
+template <typename T>
+void FillConstant(const CUDAContext& dev_ctx,
+                  const Scalar& val,
+                  DenseTensor* out) {
+  eigen::fill<CUDAContext, T>(dev_ctx, out, val.to<T>());
+}
+
 }  // namespace pten
 
 PT_REGISTER_MODULE(CreationCUDA);
@@ -41,3 +73,18 @@ PT_REGISTER_KERNEL("fill_any_like",
                    int64_t,
                    bool,
                    paddle::platform::float16) {}
+
+PT_REGISTER_KERNEL("fill_constant.scalar",
+                   CUDA,
+                   ANY,
+                   pten::FillConstant,
+                   float,
+                   double,
+                   uint8_t,
+                   int16_t,
+                   int,
+                   int64_t,
+                   bool,
+                   paddle::platform::float16,
+                   paddle::platform::complex<float>,
+                   paddle::platform::complex<double>) {}

paddle/pten/kernels/cuda/creation.h

@@ -32,6 +32,11 @@ void FillAnyLike(const CUDAContext& dev_ctx,
                  const Scalar& val,
                  DenseTensor* out);
 
+template <typename T>
+void FillConstant(const CUDAContext& dev_ctx,
+                  const Scalar& val,
+                  DenseTensor* out);
+
 }  // namespace pten
 
 #endif