From c3864eab994ffacfe52c5c4477019268263f473e Mon Sep 17 00:00:00 2001
From: chengduoZH <zhaochengduo@163.com>
Date: Mon, 5 Mar 2018 23:50:36 +0800
Subject: [PATCH] if axis == 0; directly copy D->D

---
 paddle/fluid/operators/concat_op.h | 60 +++++++++++++++++++++---------
 1 file changed, 43 insertions(+), 17 deletions(-)

diff --git a/paddle/fluid/operators/concat_op.h b/paddle/fluid/operators/concat_op.h
index 6ac70eacaf9..92c8ab6d9ff 100644
--- a/paddle/fluid/operators/concat_op.h
+++ b/paddle/fluid/operators/concat_op.h
@@ -33,14 +33,26 @@ class ConcatKernel : public framework::OpKernel<T> {
     auto place = ctx.GetPlace();
     out->mutable_data<T>(place);
 
-    // TODO(zcd): Sometimes direct copies will be faster
-    std::vector<framework::Tensor> inputs(ins.size());
-    for (size_t j = 0; j < ins.size(); ++j) {
-      inputs[j] = *ins[j];
+    // Sometimes direct copies will be faster, this maybe need deeply analysis.
+    if (axis == 0 && ins.size() < 10) {
+      size_t output_offset = 0;
+      for (auto* in : ins) {
+        auto in_stride = framework::stride_numel(in->dims());
+        auto out_stride = framework::stride_numel(out->dims());
+        StridedNumelCopyWithAxis<T>(ctx.device_context(), axis,
+                                    out->data<T>() + output_offset, out_stride,
+                                    in->data<T>(), in_stride, in_stride[axis]);
+        output_offset += in_stride[axis];
+      }
+    } else {
+      std::vector<framework::Tensor> inputs(ins.size());
+      for (size_t j = 0; j < ins.size(); ++j) {
+        inputs[j] = *ins[j];
+      }
+      auto& dev_ctx = ctx.template device_context<DeviceContext>();
+      paddle::operators::math::ConcatFunctor<DeviceContext, T> concat_functor;
+      concat_functor(dev_ctx, inputs, static_cast<int>(axis), out);
     }
-    auto& dev_ctx = ctx.template device_context<DeviceContext>();
-    paddle::operators::math::ConcatFunctor<DeviceContext, T> concat_functor;
-    concat_functor(dev_ctx, inputs, static_cast<int>(axis), out);
   }
 };
 
@@ -52,17 +64,31 @@ class ConcatGradKernel : public framework::OpKernel<T> {
     auto outs = ctx.MultiOutput<framework::Tensor>(framework::GradVarName("X"));
     int64_t axis = static_cast<int64_t>(ctx.Attr<int>("axis"));
 
-    // TODO(zcd): Sometimes direct copies will be faster
-    std::vector<framework::Tensor> outputs(outs.size());
-    for (size_t j = 0; j < outs.size(); ++j) {
-      outs[j]->mutable_data<T>(ctx.GetPlace());
-      outputs[j] = *outs[j];
-    }
+    // Sometimes direct copies will be faster, this maybe need deeply analysis.
+    if (axis == 0 && outs.size() < 10) {
+      size_t input_offset = 0;
+      auto in_stride = framework::stride_numel(in->dims());
+
+      for (auto& out : outs) {
+        out->mutable_data<T>(ctx.GetPlace());
+        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, out_stride[axis]);
+        input_offset += out_stride[axis];
+      }
+    } else {
+      std::vector<framework::Tensor> outputs(outs.size());
+      for (size_t j = 0; j < outs.size(); ++j) {
+        outs[j]->mutable_data<T>(ctx.GetPlace());
+        outputs[j] = *outs[j];
+      }
 
-    auto& dev_ctx = ctx.template device_context<DeviceContext>();
-    paddle::operators::math::ConcatGradFunctor<DeviceContext, T>
-        concat_grad_functor;
-    concat_grad_functor(dev_ctx, *in, static_cast<int>(axis), outputs);
+      auto& dev_ctx = ctx.template device_context<DeviceContext>();
+      paddle::operators::math::ConcatGradFunctor<DeviceContext, T>
+          concat_grad_functor;
+      concat_grad_functor(dev_ctx, *in, static_cast<int>(axis), outputs);
+    }
   }
 };
 
-- 
GitLab