[cherry-pick] [PHI] transpose2_grad op migration (#46139) (#46873)

* Revert pool+grad oneDNN kernel conversion (#45989)

* [PHI] transpose2_grad op migration (#46139)

* op migrated, Copy(OneDNNContext, ...) added

* mutable_data & op registration in fluid removed

* refactoring

* OneDNNGetDataType to uppercase

* missing cpu check added, handler moved to .h file

* name changed to transpose_grad

* Copy changed back to TensorCopy

* Resizing corrected, Copy(OneDNNContext) removed
Co-authored-by: NPiotr Paturej <48731682+piotrekobi@users.noreply.github.com>
Co-authored-by: NPaulina Gacek <paulina.gacek@intel.com>

paddle/fluid/operators/mkldnn/transpose_mkldnn_op.cc

@@ -224,8 +224,3 @@ REGISTER_OP_KERNEL(transpose_grad,
                    MKLDNN,
                    ::paddle::platform::CPUPlace,
                    ops::TransposeMKLDNNGradOpKernel<float>);
-
-REGISTER_OP_KERNEL(transpose2_grad,
-                   MKLDNN,
-                   ::paddle::platform::CPUPlace,
-                   ops::TransposeMKLDNNGradOpKernel<float>);

paddle/phi/backends/onednn/onednn_reuse.h

@@ -1046,5 +1046,69 @@ class ClipOneDNNHandler
                                             to_void_cast<T>(input_data));
   }
 };
+template <typename T>
+class TransposeOneDNNHandler {
+ public:
+  TransposeOneDNNHandler(const OneDNNContext& dev_ctx,
+                         std::vector<int64_t>& dims,  // NOLINT
+                         std::vector<int>& axis,      // NOLINT
+                         dnnl::engine engine)
+      : dev_ctx_(dev_ctx),
+        dims_(dims),
+        axis_(axis),
+        logical_axis_(dims.size(), 0),
+        engine_(engine) {}
+
+  std::shared_ptr<dnnl::memory> AcquireSrcMemory(const OneDNNMemoryFormat& fmt,
+                                                 void* ptr) {
+    // Make memory descriptor using input format, unless it
+    // cannot be trusted (nchw) then make up memory fmt manually
+    for (size_t i = 0; i < this->logical_axis_.size(); ++i) {
+      this->logical_axis_[i] = i;
+    }
+
+    auto src_md = fmt != OneDNNMemoryFormat::nchw
+                      ? OneDNNMemDesc(dims_, OneDNNGetDataType<T>(), fmt)
+                      : Axis2MemoryDesc(dims_, logical_axis_);
+    return std::make_shared<dnnl::memory>(src_md, engine_, ptr);
+  }
+
+  std::shared_ptr<dnnl::memory> AcquireDstMemory(DenseTensor* output,
+                                                 Place place) {
+    auto dst_md = Axis2MemoryDesc(dims_, axis_);
+    auto dst_data = dev_ctx_.Alloc<T>(output);
+    return std::make_shared<dnnl::memory>(dst_md, engine_, dst_data);
+  }
+
+  std::shared_ptr<dnnl::reorder> AcquireTranspose(
+      std::shared_ptr<dnnl::memory> dst_memory_p,
+      std::shared_ptr<dnnl::memory> src_memory_p) {
+    return std::make_shared<dnnl::reorder>(*(src_memory_p), *(dst_memory_p));
+  }
+
+ protected:
+  dnnl::memory::desc Axis2MemoryDesc(std::vector<int64_t>& nchw_tz,  // NOLINT
+                                     std::vector<int>& axis          // NOLINT
+  ) {
+    size_t ndims = axis.size();
+
+    std::vector<int64_t> strides(ndims);
+    unsigned int total_stride = 1;
+    for (int i = ndims - 1; i >= 0; --i) {
+      strides[axis[i]] = total_stride;
+      total_stride *= nchw_tz[axis[i]];
+    }
+    dnnl::memory::desc mem_d(nchw_tz, OneDNNGetDataType<T>(), strides);
+
+    return mem_d;
+  }
+
+ private:
+  const OneDNNContext& dev_ctx_;
+  std::vector<int64_t> dims_;
+  std::vector<int> axis_;
+  std::vector<int> logical_axis_;
+  dnnl::engine engine_;
+};
 }  // namespace funcs
 }  // namespace phi

paddle/phi/kernels/onednn/transpose_grad_kernel.cc

+// Copyright (c) 2022 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/phi/kernels/transpose_grad_kernel.h"
+
+#include "paddle/fluid/framework/tensor_util.h"
+#include "paddle/phi/backends/onednn/onednn_reuse.h"
+#include "paddle/phi/core/kernel_registry.h"
+
+namespace phi {
+template <typename T, typename Context>
+void TransposeGradKernel(const Context& dev_ctx,
+                         const DenseTensor& out_grad,
+                         const std::vector<int>& axis,
+                         DenseTensor* x_grad) {
+  PADDLE_ENFORCE_EQ(dev_ctx.GetPlace().GetType() == phi::AllocationType::CPU,
+                    true,
+                    errors::PreconditionNotMet(
+                        "Operator DNNL TransposeGrad must use CPUPlace"));
+  if (!x_grad) return;
+
+  const auto& onednn_engine = dev_ctx.GetEngine();
+  std::vector<int> reversed_axis(axis);
+  if (axis.size() == 1) {
+    paddle::framework::TensorCopy(out_grad, out_grad.place(), x_grad);
+    x_grad->set_format(out_grad.format());
+    return;
+  }
+
+  for (size_t i = 0; i < axis.size(); i++) {
+    reversed_axis[axis[i]] = i;
+  }
+
+  const T* out_grad_data = out_grad.data<T>();
+  dev_ctx.template Alloc<T>(x_grad);
+  auto nchw_tz = vectorize<int64_t>(out_grad.dims());
+
+  funcs::TransposeOneDNNHandler<T> handler(
+      dev_ctx, nchw_tz, reversed_axis, onednn_engine);
+
+  auto transpose_src_memory_p = handler.AcquireSrcMemory(
+      out_grad.format(), funcs::to_void_cast<T>(out_grad_data));
+  auto transpose_dst_memory_p =
+      handler.AcquireDstMemory(x_grad, dev_ctx.GetPlace());
+  auto transpose_p =
+      handler.AcquireTranspose(transpose_dst_memory_p, transpose_src_memory_p);
+
+  auto& astream = OneDNNContext::tls().get_stream();
+  transpose_p->execute(
+      astream, *transpose_src_memory_p, *transpose_dst_memory_p);
+  astream.wait();
+}
+
+}  // namespace phi
+
+PD_REGISTER_KERNEL(
+    transpose_grad, OneDNN, ALL_LAYOUT, phi::TransposeGradKernel, float) {}

python/paddle/fluid/tests/unittests/test_transpose_op.py

@@ -16,7 +16,7 @@ from __future__ import print_function
 
 import unittest
 import numpy as np
-from op_test import OpTest, convert_float_to_uint16
+from paddle.fluid.tests.unittests.op_test import OpTest, convert_float_to_uint16
 import paddle
 import paddle.fluid as fluid
 from paddle.fluid import Program, program_guard