Use emplace_back for more speed

r-barnes · facebook-github-bot · commit 950c0e238197 · 2022-06-24T23:45:17.000-07:00
Summary: Prevents a copy

Reviewed By: rohan-varma

Differential Revision: D37418718

fbshipit-source-id: 9475a5817cc461066a2b462eea1242fdee175042
diff --git a/gloo/cuda_collectives_nccl.h b/gloo/cuda_collectives_nccl.h
@@ -23,12 +23,11 @@ std::vector<nccl::NCCLElement<T> > toDeviceElements(
   std::vector<nccl::NCCLElement<T> > elements;
   elements.reserve(ptrs.size());
   for (auto i = 0; i < ptrs.size(); i++) {
-    elements.push_back(
-        nccl::NCCLElement<T>(
+    elements.emplace_back(
             ptrs[i].range(offset, count),
             streams[i],
             ptrs[i].range(offset, count),
-            streams[i]));
+            streams[i]);
   }
   return elements;
 }
diff --git a/gloo/cuda_private.h b/gloo/cuda_private.h
@@ -148,7 +148,7 @@ class CudaDeviceStreams {
     const int numDevices = getDeviceCount();
     streams_.reserve(numDevices);
     for (auto i = 0; i < numDevices; i++) {
-      streams_.push_back(CudaStream(i));
+      streams_.emplace_back(i);
     }
   }
   cudaStream_t operator[](const int i) {

Original file line number	Diff line number	Diff line change
`@@ -148,7 +148,7 @@ class CudaDeviceStreams {`
`148`	`148`	`const int numDevices = getDeviceCount();`
`149`	`149`	`streams_.reserve(numDevices);`
`150`	`150`	`for (auto i = 0; i < numDevices; i++) {`
`151`		`- streams_.push_back(CudaStream(i));`
	`151`	`+ streams_.emplace_back(i);`
`152`	`152`	`}`
`153`	`153`	`}`
`154`	`154`	`cudaStream_t operator[](const int i) {`