Merge commit 'd8ae7893e056ebf4e7a5e96bab2c3b69f196ddfd'

Author: soumith

torch/lib/THC/THCSortUtils.cuh

@@ -65,12 +65,9 @@ __device__ inline void bitonicSort(K keys[Power2SortSize],
 
 #pragma unroll
  for (unsigned int stride = size / 2; stride > 0; stride /= 2) {
-
- // Single warp per slice is completely synchronous
- if (Power2SortSize > 64) {
- __syncthreads();
- }
-
+ 
+ __syncthreads();
+ 
  unsigned int pos = 2 * threadIdx.x - (threadIdx.x & (stride - 1));
  bitonicSwap<Comparator, K, V>(
  keys[pos], values[pos], valid[pos],
@@ -81,22 +78,18 @@ __device__ inline void bitonicSort(K keys[Power2SortSize],
 
 #pragma unroll
  for (unsigned int stride = Power2SortSize / 2; stride > 0; stride /= 2) {
- // Single warp per slice is completely synchronous
- if (Power2SortSize > 64) {
- __syncthreads();
- }
-
+ 
+ __syncthreads();
+ 
  unsigned int pos = 2 * threadIdx.x - (threadIdx.x & (stride - 1));
  bitonicSwap<Comparator, K, V>(
  keys[pos], values[pos], valid[pos],
  keys[pos + stride], values[pos + stride], valid[pos + stride],
  false, comp);
  }
 
- // Single warp per slice is completely synchronous
- if (Power2SortSize > 64) {
- __syncthreads();
- }
+ __syncthreads();
+ 
 }
 
 template <typename Comparator, typename K,
@@ -111,11 +104,8 @@ __device__ inline void bitonicSortKeys(K keys[Power2SortSize],
 #pragma unroll
  for (unsigned int stride = size / 2; stride > 0; stride /= 2) {
 
- // Single warp per slice is completely synchronous
- if (Power2SortSize > 64) {
- __syncthreads();
- }
-
+ __syncthreads();
+ 
  unsigned int pos = 2 * threadIdx.x - (threadIdx.x & (stride - 1));
  bitonicSwapKeys<Comparator, K>(
  keys[pos], valid[pos],
@@ -126,22 +116,17 @@ __device__ inline void bitonicSortKeys(K keys[Power2SortSize],
 
 #pragma unroll
  for (unsigned int stride = Power2SortSize / 2; stride > 0; stride /= 2) {
- // Single warp per slice is completely synchronous
- if (Power2SortSize > 64) {
- __syncthreads();
- }
-
+ __syncthreads();
+ 
  unsigned int pos = 2 * threadIdx.x - (threadIdx.x & (stride - 1));
  bitonicSwapKeys<Comparator, K>(
  keys[pos], valid[pos],
  keys[pos + stride], valid[pos + stride],
  false, comp);
  }
 
- // Single warp per slice is completely synchronous
- if (Power2SortSize > 64) {
- __syncthreads();
- }
+ __syncthreads();
+ 
 }
 
 // Sorts (key, value) pairs (in different tensors) in-place; i.e.,