[pten]add pten conj kernel

paddle/fluid/operators/conj_op.h

@@ -14,11 +14,14 @@
 
 #pragma once
 
-#include "paddle/fluid/framework/eigen.h"
 #include "paddle/fluid/framework/op_registry.h"
 #include "paddle/fluid/framework/operator.h"
-#include "paddle/fluid/operators/math/complex_functors.h"
-#include "paddle/fluid/platform/for_range.h"
+
+// only can include the headers in paddle/pten/api dirs
+#include "paddle/pten/api/lib/utils/tensor_utils.h"
+#include "paddle/pten/include/core.h"
+#include "paddle/pten/kernels/cpu/conj_kernel.h"
+#include "paddle/pten/kernels/cuda/conj_kernel.h"
 
 namespace paddle {
 namespace operators {
@@ -30,16 +33,14 @@ class ConjKernel : public framework::OpKernel<T> {
   void Compute(const framework::ExecutionContext& context) const override {
     const Tensor* x = context.Input<Tensor>("X");
     Tensor* out = context.Output<Tensor>("Out");
+    out->mutable_data<T>(context.GetPlace(), size_t(x->numel() * sizeof(T)));
 
-    auto numel = x->numel();
-    auto* x_data = x->data<T>();
-    auto* out_data = out->mutable_data<T>(context.GetPlace(),
-                                          size_t(x->numel() * sizeof(T)));
+    auto& dev_ctx = context.device_context<DeviceContext>();
+    auto pt_x = paddle::experimental::MakePtenDenseTensor(*x);
+    auto pt_out = paddle::experimental::MakePtenDenseTensor(*out);
 
-    auto& dev_ctx = context.template device_context<DeviceContext>();
-    platform::ForRange<DeviceContext> for_range(dev_ctx, numel);
-    math::ConjFunctor<T> functor(x_data, numel, out_data);
-    for_range(functor);
+    // call new kernel
+    pten::Conj<T>(dev_ctx, *pt_x.get(), pt_out.get());
   }
 };
 

paddle/fluid/operators/spectral_op.cu

@@ -20,6 +20,7 @@
 #include <vector>
 
 #include "paddle/fluid/operators/conj_op.h"
+#include "paddle/fluid/operators/math/complex_functors.h"
 #include "paddle/fluid/operators/spectral_helper.h"
 #include "paddle/fluid/operators/spectral_op.h"
 #include "paddle/fluid/operators/transpose_op.h"

paddle/pten/CMakeLists.txt

@@ -25,10 +25,10 @@ add_subdirectory(tests)
 # make an unity target for compile deps
 set(PTEN_DEPS convert_utils dense_tensor pten_context kernel_factory kernel_context)
 set(PTEN_DEPS ${PTEN_DEPS} scale_kernel_eigen full_kernel_eigen)
-set(PTEN_DEPS ${PTEN_DEPS} math_cpu linalg_cpu manipulation_cpu)
+set(PTEN_DEPS ${PTEN_DEPS} math_cpu linalg_cpu manipulation_cpu conj_kernel_cpu)
 set(PTEN_DEPS ${PTEN_DEPS} nary unary binary)
 if(WITH_GPU OR WITH_ROCM)
-  set(PTEN_DEPS ${PTEN_DEPS} math_cuda linalg_cuda manipulation_cuda)
+  set(PTEN_DEPS ${PTEN_DEPS} math_cuda linalg_cuda manipulation_cuda conj_kernel_cuda)
 endif()
 if(WITH_XPU)
   set(PTEN_DEPS ${PTEN_DEPS} manipulation_xpu)

paddle/pten/api/include/kernel_signature.h

@@ -105,4 +105,8 @@ using subtract_kernel = void (*)(const DeviceContext&,
                                  int,
                                  DenseTensor*);
 
+using conj_kernel = void (*)(const DeviceContext&,
+                             const DenseTensor&,
+                             DenseTensor*);
+
 }  // namespace pten

paddle/pten/api/lib/kernel_declare.h

@@ -25,12 +25,14 @@ PT_DECLARE_KERNEL(dot, CPU, ALL_LAYOUT);
 PT_DECLARE_KERNEL(flatten, CPU, ALL_LAYOUT);
 PT_DECLARE_KERNEL(sign, CPU, ALL_LAYOUT);
 PT_DECLARE_KERNEL(scale, CPU, ALL_LAYOUT);
+PT_DECLARE_KERNEL(conj, CPU, ALL_LAYOUT);
 
 #if defined(PADDLE_WITH_CUDA) || defined(PADDLE_WITH_HIP)
 PT_DECLARE_KERNEL(full_like, CUDA, ALL_LAYOUT);
 PT_DECLARE_KERNEL(dot, CUDA, ALL_LAYOUT);
 PT_DECLARE_KERNEL(flatten, CUDA, ALL_LAYOUT);
 PT_DECLARE_KERNEL(sign, CUDA, ALL_LAYOUT);
+PT_DECLARE_KERNEL(conj, CUDA, ALL_LAYOUT);
 #endif
 
 #ifdef PADDLE_WITH_XPU

paddle/pten/include/math.h

@@ -17,7 +17,9 @@ limitations under the License. */
 // See Note: [ How do we organize the kernel directory ]
 #include "paddle/pten/api/lib/utils/storage.h"
 #include "paddle/pten/include/infermeta.h"
+#include "paddle/pten/kernels/cpu/conj_kernel.h"
 #include "paddle/pten/kernels/cpu/math.h"
+#include "paddle/pten/kernels/cuda/conj_kernel.h"
 #include "paddle/pten/kernels/cuda/math.h"
 #include "paddle/pten/kernels/scale_kernel.h"
 
@@ -139,4 +141,16 @@ DenseTensor Multiply(const ContextT& dev_ctx,
   Multiply<T>(dev_ctx, x, y, axis, &dense_out);
   return dense_out;
 }
+
+template <typename T, typename ContextT>
+DenseTensor Conj(const ContextT& dev_ctx, const DenseTensor& x) {
+  auto out_meta = UnchangedInferMeta(x.meta());
+  pten::DenseTensor dense_out(
+      pten::make_intrusive<paddle::experimental::SharedStorage>(
+          dev_ctx.GetPlace()),
+      std::move(out_meta));
+  Conj<T>(dev_ctx, x, &dense_out);
+  return dense_out;
+}
+
 }  // namespace pten

paddle/pten/kernels/cpu/CMakeLists.txt

@@ -2,3 +2,4 @@ cc_library(math_cpu SRCS math.cc DEPS dense_tensor kernel_context kernel_factory
 cc_library(linalg_cpu SRCS linalg.cc DEPS dense_tensor kernel_context kernel_factory)
 cc_library(utils_cpu SRCS utils.cc DEPS dense_tensor kernel_context kernel_factory memory convert_utils)
 cc_library(manipulation_cpu SRCS manipulation.cc DEPS dense_tensor kernel_context kernel_factory utils_cpu unary)
+cc_library(conj_kernel_cpu SRCS conj_kernel.cc DEPS dense_tensor kernel_context kernel_factory)

paddle/pten/kernels/cpu/conj_kernel.cc

@@ -0,0 +1,39 @@
+// 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.
+
+#include "paddle/pten/kernels/cpu/conj_kernel.h"
+
+#include "paddle/pten/backends/cpu/cpu_context.h"
+#include "paddle/pten/core/kernel_registry.h"
+#include "paddle/pten/kernels/hybird/math/conj_impl.h"
+
+namespace pten {
+
+template <typename T>
+void Conj(const CPUContext& dev_ctx, const DenseTensor& x, DenseTensor* out) {
+  ConjImpl<T, CPUContext>(dev_ctx, x, out);
+}
+
+}  // namespace pten
+
+PT_REGISTER_KERNEL(conj,
+                   CPU,
+                   ALL_LAYOUT,
+                   pten::Conj,
+                   paddle::platform::complex<float>,
+                   paddle::platform::complex<double>,
+                   float,
+                   double,
+                   int,
+                   int64_t) {}

paddle/pten/kernels/cpu/conj_kernel.h

@@ -0,0 +1,25 @@
+/* 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
+
+#include "paddle/pten/backends/cpu/cpu_context.h"
+#include "paddle/pten/core/dense_tensor.h"
+
+namespace pten {
+
+template <typename T>
+void Conj(const CPUContext& dev_ctx, const DenseTensor& x, DenseTensor* out);
+
+}  // namespace pten

paddle/pten/kernels/cuda/CMakeLists.txt

@@ -3,9 +3,11 @@ if(WITH_GPU)
   nv_library(linalg_cuda SRCS linalg.cu DEPS eigen_function dense_tensor kernel_context kernel_factory)
   nv_library(utils_cuda SRCS utils.cu DEPS dense_tensor kernel_context kernel_factory memory convert_utils)
   nv_library(manipulation_cuda SRCS manipulation.cu DEPS dense_tensor kernel_context kernel_factory utils_cuda unary)
+  nv_library(conj_kernel_cuda SRCS conj_kernel.cu DEPS dense_tensor kernel_context kernel_factory)
 elseif(WITH_ROCM)
   hip_library(math_cuda SRCS math.cu DEPS eigen_function dense_tensor convert_utils kernel_context kernel_factory pten_transpose_cuda)
   hip_library(linalg_cuda SRCS linalg.cu DEPS eigen_function dense_tensor kernel_context kernel_factory)
   hip_library(utils_cuda SRCS utils.cu DEPS dense_tensor kernel_context kernel_factory memory convert_utils)
   hip_library(manipulation_cuda SRCS manipulation.cu DEPS dense_tensor kernel_context kernel_factory utils_cuda unary)
+  hip_library(conj_kernel_cuda SRCS conj_kernel.cu DEPS dense_tensor kernel_context kernel_factory)
 endif()

paddle/pten/kernels/cuda/conj_kernel.cu

@@ -0,0 +1,39 @@
+// 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.
+
+#include "paddle/pten/kernels/cuda/conj_kernel.h"
+
+#include "paddle/pten/backends/cuda/cuda_context.h"
+#include "paddle/pten/core/kernel_registry.h"
+#include "paddle/pten/kernels/hybird/math/conj_impl.h"
+
+namespace pten {
+
+template <typename T>
+void Conj(const CUDAContext& dev_ctx, const DenseTensor& x, DenseTensor* out) {
+  ConjImpl<T, CUDAContext>(dev_ctx, x, out);
+}
+
+}  // namespace pten
+
+PT_REGISTER_KERNEL(conj,
+                   CUDA,
+                   ALL_LAYOUT,
+                   pten::Conj,
+                   paddle::platform::complex<float>,
+                   paddle::platform::complex<double>,
+                   float,
+                   double,
+                   int,
+                   int64_t) {}

paddle/pten/kernels/cuda/conj_kernel.h

@@ -0,0 +1,30 @@
+/* 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
+
+// CUDA and HIP use same api
+#if defined(PADDLE_WITH_CUDA) || defined(PADDLE_WITH_HIP)
+
+#include "paddle/pten/backends/cuda/cuda_context.h"
+#include "paddle/pten/core/dense_tensor.h"
+
+namespace pten {
+
+template <typename T>
+void Conj(const CUDAContext& dev_ctx, const DenseTensor& x, DenseTensor* out);
+
+}  // namespace pten
+
+#endif

paddle/pten/kernels/hybird/math/conj_impl.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
+
+#include "paddle/fluid/operators/math/complex_functors.h"
+#include "paddle/fluid/platform/complex.h"
+#include "paddle/fluid/platform/for_range.h"
+
+namespace pten {
+
+template <typename T, typename ContextT>
+void ConjImpl(const ContextT& dev_ctx, const DenseTensor& x, DenseTensor* out) {
+  auto numel = x.numel();
+  auto* x_data = x.data<T>();
+  auto* out_data = out->mutable_data<T>();
+
+  paddle::platform::ForRange<ContextT> for_range(dev_ctx, numel);
+  paddle::operators::math::ConjFunctor<T> functor(x_data, numel, out_data);
+  for_range(functor);
+}
+
+}  // namespace pten

paddle/pten/tests/api/CMakeLists.txt

@@ -22,3 +22,4 @@ cc_test(test_slice_api SRCS test_slice_api.cc DEPS pten_tensor pten_api pten_api
 cc_test(test_sum_api SRCS test_sum_api.cc DEPS pten_tensor pten_api pten_api_utils)
 cc_test(test_scale_api SRCS test_scale_api.cc DEPS pten_tensor pten_api pten_api_utils)
 cc_test(test_scale_benchmark SRCS test_scale_benchmark.cc DEPS pten_tensor pten_api pten_api_utils scale_kernel_eigen)
+cc_test(test_conj_api SRCS test_conj_api.cc DEPS pten_tensor pten_api pten_api_utils)

paddle/pten/tests/api/test_conj_api.cc

@@ -0,0 +1,76 @@
+/* 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. */
+
+#include <gtest/gtest.h>
+#include <memory>
+
+#include "paddle/pten/api/include/api.h"
+
+#include "paddle/pten/api/lib/utils/allocator.h"
+#include "paddle/pten/core/dense_tensor.h"
+#include "paddle/pten/core/kernel_registry.h"
+
+namespace paddle {
+namespace tests {
+
+namespace framework = paddle::framework;
+using DDim = paddle::framework::DDim;
+
+// TODO(chenweihang): Remove this test after the API is used in the dygraph
+TEST(API, conj) {
+  // 1. create tensor
+  const auto alloc = std::make_shared<paddle::experimental::DefaultAllocator>(
+      paddle::platform::CPUPlace());
+  auto dense_x = std::make_shared<pten::DenseTensor>(
+      alloc,
+      pten::DenseTensorMeta(pten::DataType::COMPLEX64,
+                            framework::make_ddim({3, 10}),
+                            pten::DataLayout::NCHW));
+  auto* dense_x_data = dense_x->mutable_data<paddle::complex64>();
+
+  for (size_t i = 0; i < 3; ++i) {
+    for (size_t j = 0; j < 10; ++j) {
+      dense_x_data[i * 10 + j] = paddle::complex64(i * 10 + j, i * 10 + j);
+    }
+  }
+
+  paddle::experimental::Tensor x(dense_x);
+
+  // 2. test API
+  auto out = paddle::experimental::conj(x);
+
+  // 3. check result
+  ASSERT_EQ(out.dims().size(), 2);
+  ASSERT_EQ(out.dims()[0], 3);
+  ASSERT_EQ(out.dims()[1], 10);
+  ASSERT_EQ(out.numel(), 30);
+  ASSERT_EQ(out.is_cpu(), true);
+  ASSERT_EQ(out.type(), pten::DataType::COMPLEX64);
+  ASSERT_EQ(out.layout(), pten::DataLayout::NCHW);
+  ASSERT_EQ(out.initialized(), true);
+
+  auto dense_out = std::dynamic_pointer_cast<pten::DenseTensor>(out.impl());
+  auto actual_result = dense_out->data<paddle::complex64>();
+
+  for (size_t i = 0; i < 3; ++i) {
+    for (size_t j = 0; j < 10; ++j) {
+      dense_x_data[i * 10 + j] = paddle::complex64(i * 10 + j, i * 10 + j);
+      ASSERT_NEAR(actual_result[i * 10 + j].real, 1.0 * (i * 10 + j), 1e-6f);
+      ASSERT_NEAR(actual_result[i * 10 + j].imag, -1.0 * (i * 10 + j), 1e-6f);
+    }
+  }
+}
+
+}  // namespace tests
+}  // namespace paddle

paddle/pten/tests/kernels/CMakeLists.txt

@@ -8,3 +8,4 @@ cc_test(test_cast_dev_api SRCS test_cast_dev_api.cc DEPS pten pten_api_utils)
 cc_test(test_elementwise_dev_api SRCS test_elementwise_dev_api.cc DEPS pten pten_api_utils)
 cc_test(test_reshape_dev_api SRCS test_reshape_dev_api.cc DEPS pten pten_api_utils)
 cc_test(test_sum_dev_api SRCS test_sum_dev_api.cc DEPS pten pten_api_utils)
+cc_test(test_conj_dev_api SRCS test_conj_dev_api.cc DEPS pten pten_api_utils)