Merge pull request #8270 from Yancey1989/improve_concat_split_op

Improve split and concat op

Author: Yancey

paddle/fluid/framework/ddim.cc

@@ -314,5 +314,15 @@ DDim stride(const DDim& ddim) {
   }
   return framework::make_ddim(strides);
 }
+
+DDim stride_numel(const framework::DDim& ddim) {
+  std::vector<int64_t> strides(ddim.size());
+  strides[ddim.size() - 1] = ddim[ddim.size() - 1];
+  for (int i = ddim.size() - 2; i >= 0; --i) {
+    strides[i] = strides[i + 1] * ddim[i];
+  }
+  return framework::make_ddim(strides);
+}
+
 }  // namespace framework
 }  // namespace paddle

paddle/fluid/framework/ddim.h

@@ -125,6 +125,8 @@ DDim flatten_to_2d(const DDim& src, int num_col_dims);
 DDim flatten_to_1d(const DDim& src);
 
 DDim stride(const DDim& ddim);
+
+DDim stride_numel(const DDim& ddim);
 }  // namespace framework
 }  // namespace paddle
 

paddle/fluid/operators/concat_op.h

@@ -28,17 +28,18 @@ class ConcatKernel : public framework::OpKernel<T> {
     auto ins = ctx.MultiInput<framework::Tensor>("X");
     auto* out = ctx.Output<framework::Tensor>("Out");
     int64_t axis = static_cast<int64_t>(ctx.Attr<int>("axis"));
-    const size_t n = ins.size();
+    auto place = ctx.GetPlace();
+    out->mutable_data<T>(place);
+
+    auto out_stride = framework::stride_numel(out->dims());
+
     size_t output_offset = 0;
-    out->mutable_data<T>(ctx.GetPlace());
-    auto out_stride = framework::stride(out->dims());
-    for (size_t i = 0; i < n; i++) {
-      auto& in = ins[i];
-      auto axis_dim = in->dims()[axis];
-      auto in_stride = framework::stride(in->dims());
-      StridedMemcpy<T>(ctx.device_context(), in->data<T>(), in_stride,
-                       in->dims(), out_stride, out->data<T>() + output_offset);
-      output_offset += axis_dim * in_stride[axis];
+    for (auto* in : ins) {
+      auto in_stride = framework::stride_numel(in->dims());
+      StridedNumelCopyWithAxis<T>(ctx.device_context(), axis,
+                                  out->data<T>() + output_offset, out_stride,
+                                  in->data<T>(), in_stride);
+      output_offset += in_stride[axis];
     }
   }
 };
@@ -50,17 +51,16 @@ class ConcatGradKernel : public framework::OpKernel<T> {
     auto* in = ctx.Input<framework::Tensor>(framework::GradVarName("Out"));
     auto outs = ctx.MultiOutput<framework::Tensor>(framework::GradVarName("X"));
     int64_t axis = static_cast<int64_t>(ctx.Attr<int>("axis"));
-    const size_t n = outs.size();
     size_t input_offset = 0;
-    auto in_stride = framework::stride(in->dims());
-    for (size_t i = 0; i < n; i++) {
-      auto& out = outs[i];
+    auto in_stride = framework::stride_numel(in->dims());
+
+    for (auto& out : outs) {
       out->mutable_data<T>(ctx.GetPlace());
-      size_t axis_dim = out->dims()[axis];
-      auto out_stride = framework::stride(out->dims());
-      StridedMemcpy<T>(ctx.device_context(), in->data<T>() + input_offset,
-                       in_stride, out->dims(), out_stride, out->data<T>());
-      input_offset += axis_dim * in_stride[axis];
+      auto out_stride = framework::stride_numel(out->dims());
+      StridedNumelCopyWithAxis<T>(ctx.device_context(), axis, out->data<T>(),
+                                  out_stride, in->data<T>() + input_offset,
+                                  in_stride);
+      input_offset += out_stride[axis];
     }
   }
 };

paddle/fluid/operators/split_op.h

@@ -14,6 +14,7 @@ limitations under the License. */
 
 #pragma once
 
+#include <chrono>
 #include <vector>
 #include "paddle/fluid/framework/op_registry.h"
 #include "paddle/fluid/operators/strided_memcpy.h"
@@ -27,18 +28,18 @@ class SplitOpKernel : public framework::OpKernel<T> {
   void Compute(const framework::ExecutionContext& ctx) const override {
     auto* in = ctx.Input<framework::Tensor>("X");
     auto outs = ctx.MultiOutput<framework::Tensor>("Out");
-    auto in_stride = framework::stride(in->dims());
+    auto in_stride = framework::stride_numel(in->dims());
     int64_t axis = static_cast<int64_t>(ctx.Attr<int>("axis"));
-    const size_t n = outs.size();
+    auto place = ctx.GetPlace();
+
     size_t input_offset = 0;
-    for (size_t i = 0; i < n; i++) {
-      auto& out = outs[i];
+    for (auto& out : outs) {
       out->mutable_data<T>(ctx.GetPlace());
-      size_t axis_dim = out->dims()[axis];
-      auto out_stride = framework::stride(out->dims());
-      StridedMemcpy<T>(ctx.device_context(), in->data<T>() + input_offset,
-                       in_stride, out->dims(), out_stride, out->data<T>());
-      input_offset += axis_dim * in_stride[axis];
+      auto out_stride = framework::stride_numel(out->dims());
+      StridedNumelCopyWithAxis<T>(ctx.device_context(), axis, out->data<T>(),
+                                  out_stride, in->data<T>() + input_offset,
+                                  in_stride);
+      input_offset += out_stride[axis];
     }
   }
 };

paddle/fluid/operators/strided_memcpy.h

@@ -41,5 +41,62 @@ inline void StridedMemcpy(const platform::DeviceContext& dev_ctx, const T* src,
   StridedCopyDimVisitor<T> func(dev_ctx, src, src_stride, dst_stride, dst);
   boost::apply_visitor(func, dst_dim);
 }
+
+// Strided numel memory copy from src to dst by the specified axis
+//
+// For example, for a tensor dims [4, 20, 100], the strieded numel is
+// [8000, 2000, 100]
+//
+// NOTE: The src and dst tensor should have the same elements
+// except the specified axis.
+template <typename T>
+inline void StridedNumelCopyWithAxis(const platform::DeviceContext& ctx,
+                                     int64_t axis, T* dst,
+                                     const framework::DDim& dst_stride_numel,
+                                     const T* src,
+                                     const framework::DDim& src_stride_numel) {
+  int64_t before = dst_stride_numel[0] / dst_stride_numel[axis];
+  int64_t src_after = src_stride_numel[axis];
+  int64_t dst_after = dst_stride_numel[axis];
+  auto place = ctx.GetPlace();
+
+  PADDLE_ENFORCE_EQ(src_stride_numel.size(), dst_stride_numel.size(),
+                    "src and dst tensor should have the same dims size.");
+
+  for (int64_t i = 0; i < axis; ++i) {
+    if (i < axis) {
+      PADDLE_ENFORCE_EQ(src_stride_numel[i] / src_stride_numel[axis],
+                        dst_stride_numel[i] / dst_stride_numel[axis],
+                        "src and dst should have the same elements "
+                        "except the specified axis.");
+    } else if (i == axis) {
+      continue;
+    } else {
+      PADDLE_ENFORCE_EQ(src_stride_numel[i], dst_stride_numel[i],
+                        "src and dst should have the same elements "
+                        "except the specified axis.");
+    }
+  }
+
+  for (int64_t i = 0; i < before; ++i) {
+    if (platform::is_cpu_place(place)) {
+      auto& cpu_place = boost::get<platform::CPUPlace>(place);
+      memory::Copy(cpu_place, dst + i * dst_after, cpu_place,
+                   src + i * src_after, sizeof(T) * src_after);
+    } else {
+#ifdef PADDLE_WITH_CUDA
+      auto& gpu_place = boost::get<platform::CUDAPlace>(place);
+      auto& cuda_ctx =
+          reinterpret_cast<const platform::CUDADeviceContext&>(ctx);
+      memory::Copy(gpu_place, dst + i * dst_after, gpu_place,
+                   src + i * src_after, sizeof(T) * src_after,
+                   cuda_ctx.stream());
+#else
+      PADDLE_THROW("Paddle is not compiled with GPU");
+#endif
+    }
+  }
+}
+
 }  // namespace operators
 }  // namespace paddle

python/paddle/v2/fluid/tests/test_split_op.py

@@ -20,11 +20,11 @@
 class TestSplitOp(OpTest):
     def setUp(self):
         self.op_type = "split"
-        axis = 0
-        x = np.random.random((4, 2, 5)).astype('float32')
-        out = np.split(x, [1, 3], axis)
+        axis = 1
+        x = np.random.random((4, 5, 6)).astype('float32')
+        out = np.split(x, [2, 3], axis)
         self.inputs = {'X': x}
-        self.attrs = {'axis': axis, 'sections': [1, 2, 1]}
+        self.attrs = {'axis': axis, 'sections': [2, 1, 2]}
         self.outputs = {'Out': [('out%d' % i, out[i]) \
             for i in xrange(len(out))]}