From e46996bd9dd1134847e2a3dfb796f1e0ebc450a7 Mon Sep 17 00:00:00 2001
From: mhouston <mhouston@nvidia.com>
Date: Fri, 10 Jul 2015 16:05:48 -0700
Subject: [PATCH] Detect topology corner cases and improve broadcast order

- Start with distant nodes in broadcast
- Fix outside loop to loop for full tree depth
---
 src/caffe/parallel.cpp | 71 ++++++++++++++++++++++++------------------
 1 file changed, 40 insertions(+), 31 deletions(-)

diff --git a/src/caffe/parallel.cpp b/src/caffe/parallel.cpp
index ef04a21268e..4ba8015a865 100644
--- a/src/caffe/parallel.cpp
+++ b/src/caffe/parallel.cpp
@@ -119,18 +119,23 @@ void DevicePair::compute(const vector<int> devices, vector<DevicePair>* pairs) {
 #ifndef CPU_ONLY
   vector<int> remaining(devices);
 
+  // Depth for reduction tree
+  int remaining_depth = static_cast<int>(ceil(log2(remaining.size())));
+
   // Group GPUs by board
-  for (int i = 0; i < remaining.size(); ++i) {
-    for (int j = i + 1; j < remaining.size(); ++j) {
-      cudaDeviceProp a, b;
-      CUDA_CHECK(cudaGetDeviceProperties(&a, remaining[i]));
-      CUDA_CHECK(cudaGetDeviceProperties(&b, remaining[j]));
-      if (a.isMultiGpuBoard && b.isMultiGpuBoard) {
-        if (a.multiGpuBoardGroupID == b.multiGpuBoardGroupID) {
-          pairs->push_back(DevicePair(remaining[i], remaining[j]));
-          DLOG(INFO) << "GPU board: " << remaining[i] << ":" << remaining[j];
-          remaining.erase(remaining.begin() + j);
-          break;
+  for (int d = 0; d < remaining_depth; ++d) {
+    for (int i = 0; i < remaining.size(); ++i) {
+      for (int j = i + 1; j < remaining.size(); ++j) {
+        cudaDeviceProp a, b;
+        CUDA_CHECK(cudaGetDeviceProperties(&a, remaining[i]));
+        CUDA_CHECK(cudaGetDeviceProperties(&b, remaining[j]));
+        if (a.isMultiGpuBoard && b.isMultiGpuBoard) {
+          if (a.multiGpuBoardGroupID == b.multiGpuBoardGroupID) {
+            pairs->push_back(DevicePair(remaining[i], remaining[j]));
+            DLOG(INFO) << "GPU board: " << remaining[i] << ":" << remaining[j];
+            remaining.erase(remaining.begin() + j);
+            break;
+          }
         }
       }
     }
@@ -142,15 +147,19 @@ void DevicePair::compute(const vector<int> devices, vector<DevicePair>* pairs) {
   DLOG(INFO) << "GPUs paired by boards, remaining: " << s.str();
 
   // Group by P2P accessibility
-  for (int i = 0; i < remaining.size(); ++i) {
-    for (int j = i + 1; j < remaining.size(); ++j) {
-      int access;
-      CUDA_CHECK(cudaDeviceCanAccessPeer(&access, remaining[i], remaining[j]));
-      if (access) {
-        pairs->push_back(DevicePair(remaining[i], remaining[j]));
-        DLOG(INFO) << "P2P pair: " << remaining[i] << ":" << remaining[j];
-        remaining.erase(remaining.begin() + j);
-        break;
+  remaining_depth = ceil(log2(remaining.size()));
+  for (int d = 0; d < remaining_depth; ++d) {
+    for (int i = 0; i < remaining.size(); ++i) {
+      for (int j = i + 1; j < remaining.size(); ++j) {
+        int access;
+        CUDA_CHECK(
+            cudaDeviceCanAccessPeer(&access, remaining[i], remaining[j]));
+        if (access) {
+          pairs->push_back(DevicePair(remaining[i], remaining[j]));
+          DLOG(INFO) << "P2P pair: " << remaining[i] << ":" << remaining[j];
+          remaining.erase(remaining.begin() + j);
+          break;
+        }
       }
     }
   }
@@ -161,15 +170,19 @@ void DevicePair::compute(const vector<int> devices, vector<DevicePair>* pairs) {
   DLOG(INFO) << "GPUs paired by P2P access, remaining: " << s.str();
 
   // Group remaining
-  for (int i = 0; i < remaining.size(); ++i) {
-    for (int j = i + 1; j < remaining.size(); ++j) {
-      pairs->push_back(DevicePair(remaining[i], remaining[j]));
-      DLOG(INFO) << "Remaining pair: " << remaining[i] << ":" << remaining[j];
-      remaining.erase(remaining.begin() + j);
-      break;
+  remaining_depth = ceil(log2(remaining.size()));
+  for (int d = 0; d < remaining_depth; ++d) {
+    for (int i = 0; i < remaining.size(); ++i) {
+      pairs->push_back(DevicePair(remaining[i], remaining[i + 1]));
+      DLOG(INFO) << "Remaining pair: " << remaining[i] << ":"
+                 << remaining[i + 1];
+      remaining.erase(remaining.begin() + i + 1);
     }
   }
+
+  // Should only be the parent node remaining
   CHECK_EQ(remaining.size(), 1);
+
   pairs->insert(pairs->begin(), DevicePair(-1, remaining[0]));
 
   CHECK(pairs->size() == devices.size());
@@ -294,7 +307,7 @@ void P2PSync<Dtype>::on_start(Timer* timer, ostringstream* timing) {
   if (children_.size()) {
     timer->Start();
   }
-  for (int i = 0; i < children_.size(); ++i) {
+  for (int i = children_.size() - 1; i >= 0; i--) {
     Dtype* src = data_;
     Dtype* dst = children_[i]->data_;
 
@@ -308,11 +321,7 @@ void P2PSync<Dtype>::on_start(Timer* timer, ostringstream* timing) {
 
     CUDA_CHECK(cudaMemcpyAsync(dst, src, size_ * sizeof(Dtype),  //
         cudaMemcpyDeviceToDevice, cudaStreamDefault));
-  }
-  if (children_.size()) {
     CUDA_CHECK(cudaStreamSynchronize(cudaStreamDefault));
-  }
-  for (int i = 0; i < children_.size(); ++i) {
     children_[i]->queue_.push(this);
   }
   if (children_.size()) {