github/ColossalAI
1
0
Fork 0
mirror of https://github.com/hpcaitech/ColossalAI.git synced 2025-06-23 06:00:44 +00:00
Code Issues Packages Projects Releases Wiki Activity

[NFC] polish colossalai/kernel/cuda_native/csrc/kernels/include/block_reduce.h code style (#938)

Browse Source
This commit is contained in:
Maruyama_Aya 2022-05-13 15:38:19 +08:00 committed by binmakeswell
parent b6cc9313ef
commit 7aa35eae6a
1 changed files with 21 additions and 20 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

9
colossalai/kernel/cuda_native/csrc/kernels/include/block_reduce.h
View File

@ -13,22 +13,23 @@ const float REDUCE_FLOAT_INF_NEG = -100000000.f;
const float REDUCE_FLOAT_INF_POS = 100000000.f;
const unsigned int WARP_REDUCE_SIZE = 32;
template <typename T> __forceinline__ __device__ T warpReduceSum(T val) {
template <typename T>
__forceinline__ __device__ T warpReduceSum(T val) {
for (int mask = (WARP_REDUCE_SIZE >> 1); mask > 0; mask >>= 1)
val += __shfl_xor_sync(WARP_REDUCE_MASK, val, mask, WARP_REDUCE_SIZE);
return val;
}
/* Calculate the sum of all elements in a block */
template <typename T> __forceinline__ __device__ T blockReduceSum(T val) {
template <typename T>
__forceinline__ __device__ T blockReduceSum(T val) {
static __shared__ T shared[32];
int lane = threadIdx.x & 0x1f;
int wid = threadIdx.x >> 5;
val = warpReduceSum<T>(val);
if (lane == 0)
shared[wid] = val;
if (lane == 0) shared[wid] = val;
__syncthreads();
val = (threadIdx.x < (blockDim.x >> 5)) ? shared[lane] : (T)0.0f;