[PTen]Support XPU for Flatten Kernel

paddle/fluid/operators/flatten_op.cc

@@ -79,14 +79,6 @@ class FlattenOp : public framework::OperatorWithKernel {
       const framework::ExecutionContext &ctx) const override {
     auto input_data_type =
         framework::OperatorWithKernel::IndicateVarDataType(ctx, "X");
-
-    //#ifdef PADDLE_WITH_MKLDNN
-    //    if (this->CanMKLDNNBeUsed(ctx, input_data_type)) {
-    //      return framework::OpKernelType(input_data_type, ctx.GetPlace(),
-    //                                     framework::DataLayout::kMKLDNN,
-    //                                     framework::LibraryType::kMKLDNN);
-    //    }
-    //#endif
     return framework::OpKernelType(input_data_type, ctx.GetPlace());
   }
 };
@@ -157,14 +149,6 @@ class FlattenGradOp : public framework::OperatorWithKernel {
       const framework::ExecutionContext &ctx) const override {
     auto input_data_type = framework::OperatorWithKernel::IndicateVarDataType(
         ctx, framework::GradVarName("Out"));
-
-    //#ifdef PADDLE_WITH_MKLDNN
-    //    if (this->CanMKLDNNBeUsed(ctx, input_data_type)) {
-    //      return framework::OpKernelType(input_data_type, ctx.GetPlace(),
-    //                                     framework::DataLayout::kMKLDNN,
-    //                                     framework::LibraryType::kMKLDNN);
-    //    }
-    //#endif
     return framework::OpKernelType(input_data_type, ctx.GetPlace());
   }
 };
@@ -227,14 +211,6 @@ class Flatten2Op : public framework::OperatorWithKernel {
       const framework::ExecutionContext &ctx) const override {
     auto input_data_type =
         framework::OperatorWithKernel::IndicateVarDataType(ctx, "X");
-
-    //#ifdef PADDLE_WITH_MKLDNN
-    //    if (this->CanMKLDNNBeUsed(ctx, input_data_type)) {
-    //      return framework::OpKernelType(input_data_type, ctx.GetPlace(),
-    //                                     framework::DataLayout::kMKLDNN,
-    //                                     framework::LibraryType::kMKLDNN);
-    //    }
-    //#endif
     return framework::OpKernelType(input_data_type, ctx.GetPlace());
   }
 };
@@ -285,14 +261,6 @@ class Flatten2GradOp : public framework::OperatorWithKernel {
       const framework::ExecutionContext &ctx) const override {
     auto input_data_type = framework::OperatorWithKernel::IndicateVarDataType(
         ctx, framework::GradVarName("Out"));
-
-    //#ifdef PADDLE_WITH_MKLDNN
-    //    if (this->CanMKLDNNBeUsed(ctx, input_data_type)) {
-    //      return framework::OpKernelType(input_data_type, ctx.GetPlace(),
-    //                                     framework::DataLayout::kMKLDNN,
-    //                                     framework::LibraryType::kMKLDNN);
-    //    }
-    //#endif
     return framework::OpKernelType(input_data_type, ctx.GetPlace());
   }
 };
@@ -365,6 +333,18 @@ class FlattenContiguousRangeOp : public framework::OperatorWithKernel {
 
     return out_shape;
   }
+
+  framework::KernelSignature GetExpectedPtenKernelArgs(
+      const framework::ExecutionContext &ctx) const override {
+    if (ctx.HasOutput("XShape")) {
+      return framework::KernelSignature("flatten_contiguous_range.mid", {"X"},
+                                        {"start_axis", "stop_axis"},
+                                        {"Out", "XShape"});
+    } else {
+      return framework::KernelSignature("flatten_contiguous_range", {"X"},
+                                        {"start_axis", "stop_axis"}, {"Out"});
+    }
+  }
 };
 
 class FlattenContiguousRangeOpMaker : public FlattenOpMaker {

paddle/fluid/operators/flatten_op.h

@@ -15,10 +15,13 @@ limitations under the License. */
 #pragma once
 #include <vector>
 #include "paddle/fluid/framework/op_registry.h"
+#include "paddle/fluid/framework/pten_utils.h"
 #include "paddle/fluid/operators/math/blas.h"
 #include "paddle/fluid/operators/math/math_function.h"
 #include "paddle/fluid/operators/math/pooling.h"
 #include "paddle/fluid/platform/device_context.h"
+#include "paddle/pten/include/core.h"
+#include "paddle/pten/include/manipulation.h"
 
 namespace paddle {
 namespace operators {
@@ -122,13 +125,16 @@ class FlattenContiguousRangeKernel : public framework::OpKernel<T> {
   void Compute(const framework::ExecutionContext &context) const override {
     auto *in = context.Input<framework::LoDTensor>("X");
     auto *out = context.Output<framework::LoDTensor>("Out");
-    auto out_dims = out->dims();
-
     out->mutable_data(context.GetPlace(), in->type());
-    framework::TensorCopy(
-        *in, context.GetPlace(),
-        context.template device_context<platform::DeviceContext>(), out);
-    out->Resize(out_dims);
+    auto &start_axis = context.Attr<int>("start_axis");
+    auto &stop_axis = context.Attr<int>("stop_axis");
+    auto &dev_ctx = context.device_context<DeviceContext>();
+
+    auto pt_x = paddle::experimental::MakePtenDenseTensor(*in);
+    auto pt_out = paddle::experimental::MakePtenDenseTensor(*out);
+
+    // call new kernel
+    pten::Flatten<T>(dev_ctx, *pt_x.get(), start_axis, stop_axis, pt_out.get());
   }
 };
 

paddle/pten/CMakeLists.txt

@@ -17,5 +17,7 @@ set(PTEN_DEPS ${PTEN_DEPS} unary binary)
 if(WITH_GPU OR WITH_ROCM)
   set(PTEN_DEPS ${PTEN_DEPS} math_cuda linalg_cuda creation_cuda manipulation_cuda)
 endif()
-
+if(WITH_XPU)
+  set(PTEN_DEPS ${PTEN_DEPS} manipulation_xpu)
+endif()
 cc_library(pten SRCS all.cc DEPS ${PTEN_DEPS})

paddle/pten/include/manipulation.h

@@ -19,6 +19,7 @@
 #include "paddle/pten/include/infershape.h"
 #include "paddle/pten/kernels/cpu/manipulation.h"
 #include "paddle/pten/kernels/cuda/manipulation.h"
+#include "paddle/pten/kernels/xpu/manipulation.h"
 
 namespace pten {
 

paddle/pten/kernels/xpu/CMakeLists.txt

@@ -0,0 +1,2 @@
+cc_library(utils_xpu SRCS utils.cc DEPS dense_tensor kernel_context kernel_factory memory convert_utils)
+cc_library(manipulation_xpu SRCS manipulation.cc DEPS dense_tensor kernel_context kernel_factory utils_xpu unary)

paddle/pten/kernels/xpu/manipulation.cc

@@ -0,0 +1,82 @@
+//   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/xpu/manipulation.h"
+#include "paddle/pten/infershape/unary.h"
+#include "paddle/pten/kernels/xpu/utils.h"
+
+namespace pten {
+
+template <typename T>
+void Flatten(const XPUContext& dev_ctx,
+             const DenseTensor& x,
+             int start_axis,
+             int stop_axis,
+             DenseTensor* out) {
+  auto out_dims = out->dims();
+  pten::Copy(dev_ctx, x, out);
+  out->Resize(out_dims);
+}
+
+// TODO(yuanrisheng): this kernel is for training and xshape is a Intermediate
+// Output Tensor，
+// is there a more flexible way to deal with this case?
+template <typename T>
+void FlattenWithXShape(const XPUContext& dev_ctx,
+                       const DenseTensor& x,
+                       int start_axis,
+                       int stop_axis,
+                       DenseTensor* out,
+                       DenseTensor* xshape) {
+  Flatten<T>(dev_ctx, x, start_axis, stop_axis, out);
+  const auto& in_dims = x.dims();
+  std::vector<int64_t> xshape_dims(in_dims.size() + 1);
+  xshape_dims[0] = 0;
+  for (int i = 0; i < in_dims.size(); ++i) {
+    xshape_dims[i + 1] = in_dims[i];
+  }
+  xshape->Resize(paddle::framework::make_ddim(xshape_dims));
+  xshape->set_lod(x.lod());
+}
+
+}  // namespace pten
+
+// TODO(chenweihang): replace by better impl
+PT_REGISTER_MODULE(ManipulationXPU);
+
+// TODO(yuanrisheng): "flatten_contiguous_range" is compatible with old kernel
+// architecture, kernel_name should be "flatten".
+PT_REGISTER_KERNEL("flatten_contiguous_range",
+                   XPU,
+                   ANY,
+                   pten::Flatten,
+                   float,
+                   paddle::platform::float16,
+                   double,
+                   uint8_t,
+                   int8_t,
+                   int,
+                   int64_t) {}
+
+PT_REGISTER_KERNEL("flatten_contiguous_range.mid",
+                   XPU,
+                   ANY,
+                   pten::FlattenWithXShape,
+                   float,
+                   paddle::platform::float16,
+                   double,
+                   uint8_t,
+                   int8_t,
+                   int,
+                   int64_t) {}

paddle/pten/kernels/xpu/manipulation.h

@@ -0,0 +1,38 @@
+/* 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
+
+#ifdef PADDLE_WITH_XPU
+
+#include "paddle/pten/core/dense_tensor.h"
+#include "paddle/pten/core/kernel_registry.h"
+
+// See Note [ Why still include the fluid headers? ]
+#include "paddle/fluid/platform/device_context.h"
+
+namespace pten {
+
+using XPUContext = paddle::platform::XPUDeviceContext;
+
+template <typename T>
+void Flatten(const XPUContext& dev_ctx,
+             const DenseTensor& x,
+             int start_axis,
+             int stop_axis,
+             DenseTensor* out);
+
+}  // namespace pten
+
+#endif

paddle/pten/kernels/xpu/utils.cc

@@ -0,0 +1,81 @@
+/* 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/xpu/utils.h"
+#include "paddle/fluid/memory/memcpy.h"
+#include "paddle/pten/common/data_type.h"
+#include "paddle/pten/core/convert_utils.h"
+
+namespace pten {
+
+void Copy(const XPUDeviceContext& dev_ctx,
+          const DenseTensor& src,
+          DenseTensor* dst) {
+  auto* src_ptr = src.data();
+  auto* dst_ptr = dst->mutable_data();
+  const auto& src_place = src.place();
+  const auto& dst_place = dst->place();
+
+  if (src_ptr == dst_ptr && src_place == dst_place) {
+    VLOG(3) << "Skip copy the same data async from " << src_place << " to "
+            << dst_place;
+    return;
+  }
+  VLOG(4) << "src:" << src_ptr << ", dst:" << dst_ptr;
+
+  VLOG(3) << "TensorCopy " << src.dims() << " from " << src.place() << " to "
+          << dst_place;
+  dst->Resize(src.dims());
+  CHECK(dst->layout() == src.layout());
+  auto size = src.numel() * paddle::framework::SizeOfType(
+                                TransToProtoVarType(src.data_type()));
+
+  if (paddle::platform::is_xpu_place(src_place) &&  // NOLINT
+      paddle::platform::is_cpu_place(dst_place)) {
+    paddle::memory::Copy(BOOST_GET_CONST(paddle::platform::CPUPlace, dst_place),
+                         dst_ptr,
+                         BOOST_GET_CONST(paddle::platform::XPUPlace, src_place),
+                         src_ptr,
+                         size);
+  } else if (paddle::platform::is_cpu_place(src_place) &&
+             paddle::platform::is_xpu_place(dst_place)) {
+    paddle::memory::Copy(BOOST_GET_CONST(paddle::platform::XPUPlace, dst_place),
+                         dst_ptr,
+                         BOOST_GET_CONST(paddle::platform::CPUPlace, src_place),
+                         src_ptr,
+                         size);
+  } else if (paddle::platform::is_xpu_place(src_place) &&
+             paddle::platform::is_xpu_place(dst_place)) {
+    if (src_ptr == dst_ptr) {
+      VLOG(3) << "Skip copy the same data async from " << src_place << " to "
+              << dst_place;
+      return;
+    }
+    paddle::memory::Copy(BOOST_GET_CONST(paddle::platform::XPUPlace, dst_place),
+                         dst_ptr,
+                         BOOST_GET_CONST(paddle::platform::XPUPlace, src_place),
+                         src_ptr,
+                         size);
+  } else {
+    PADDLE_THROW(paddle::platform::errors::Unimplemented(
+        "Copy from %s to %s is not supported.", src_place, dst_place));
+  }
+}
+
+}  // namespace pten
+
+// TODO(chenweihang): replace by better impl
+PT_REGISTER_MODULE(UtilsXPU);
+
+PT_REGISTER_KERNEL_WITH_NO_TYPE("copy", XPU, ANY, pten::Copy) {}

paddle/pten/kernels/xpu/utils.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
+
+#ifdef PADDLE_WITH_XPU
+
+#include "paddle/pten/core/dense_tensor.h"
+#include "paddle/pten/core/kernel_registry.h"
+
+// See Note [ Why still include the fluid headers? ]
+#include "paddle/fluid/platform/device_context.h"
+namespace pten {
+
+using XPUDeviceContext = paddle::platform::XPUDeviceContext;
+
+void Copy(const XPUDeviceContext& dev_ctx,
+          const DenseTensor& src,
+          DenseTensor* dst);
+
+}  // namespace pten
+
+#endif