Update of PHI transpose_grad (#47311)

* - halfway transforming transpose grad - Fixes - buildable * - lint * rerunning the process
PaddlePaddle · Oct 27, 2022 · 493fbfd · 493fbfd
1 parent 77dbb31
commit 493fbfd
Show file tree

Hide file tree

Showing 2 changed files with 16 additions and 82 deletions.
diff --git a/paddle/phi/backends/onednn/onednn_reuse.h b/paddle/phi/backends/onednn/onednn_reuse.h
@@ -1085,69 +1085,5 @@ 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
diff --git a/paddle/phi/kernels/onednn/transpose_grad_kernel.cc b/paddle/phi/kernels/onednn/transpose_grad_kernel.cc
@@ -31,35 +31,33 @@ void TransposeGradKernel(const Context& dev_ctx,
   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());
+    x_grad->set_mem_desc(out_grad.mem_desc());
     return;
   }
 
-  for (size_t i = 0; i < axis.size(); i++) {
-    reversed_axis[axis[i]] = i;
-  }
+  std::vector<int64_t> out_grad_tz = vectorize(out_grad.dims());
+  funcs::ReorderOneDNNHandler reorder_handler(
+      out_grad_tz,
+      out_grad.dtype(),
+      funcs::ToOneDNNDataType(out_grad.dtype()),
+      onednn_engine);
 
-  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());
+  auto reorder_src_memory_p = reorder_handler.AcquireSrcMemory(
+      out_grad.mem_desc(), funcs::to_void_cast(out_grad.data<T>()));
 
-  funcs::TransposeOneDNNHandler<T> handler(
-      dev_ctx, nchw_tz, reversed_axis, onednn_engine);
+  auto reorder_dst_memory_p = reorder_handler.AcquireDstMemory(
+      x_grad, out_grad.mem_desc(), dev_ctx.GetPlace());
 
-  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 reorder_p = reorder_handler.AcquireReorder(reorder_dst_memory_p,
+                                                  reorder_src_memory_p);
 
   auto& astream = OneDNNContext::tls().get_stream();
-  transpose_p->execute(
-      astream, *transpose_src_memory_p, *transpose_dst_memory_p);
+  reorder_p->execute(astream, *reorder_src_memory_p, *reorder_dst_memory_p);
   astream.wait();
+  x_grad->set_mem_desc(reorder_dst_memory_p->get_desc().permute_axes(axis));
 }
 
 }  // namespace phi