github
/
ColossalAI
mirror of https://github.com/hpcaitech/ColossalAI
1
0
Fork 0
Code Issues Projects Releases Wiki Activity

block_reduce.h fix format (#581)

pull/673/head
DouJS 2022-03-31 17:13:09 +08:00 committed by binmakeswell
parent d2dc6049b5
commit f08fc17f2b
1 changed files with 20 additions and 21 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

41
colossalai/kernel/cuda_native/csrc/kernels/include/block_reduce.h
View File

@ -13,23 +13,22 @@ 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;
@ -57,10 +56,10 @@ __inline__ __device__ void warpReduce<ReduceType::kMax, 1>(float *pval) {
template <>
__inline__ __device__ void warpReduce<ReduceType::kMax, 2>(float *pval) {
float val0_tmp, val1_tmp;
#define WarpReduceMaxOneStep(a, b) \
val0_tmp = __shfl_xor_sync(WARP_REDUCE_MASK, *(pval), a, b); \
val1_tmp = __shfl_xor_sync(WARP_REDUCE_MASK, *(pval + 1), a, b); \
*(pval) = max(val0_tmp, *(pval)); \
#define WarpReduceMaxOneStep(a, b) \
val0_tmp = __shfl_xor_sync(WARP_REDUCE_MASK, *(pval), a, b); \
val1_tmp = __shfl_xor_sync(WARP_REDUCE_MASK, *(pval + 1), a, b); \
*(pval) = max(val0_tmp, *(pval)); \
*(pval + 1) = max(val1_tmp, *(pval + 1));
WarpReduceMaxOneStep(16, 32);
@ -89,10 +88,10 @@ __inline__ __device__ void warpReduce<ReduceType::kSum, 1>(float *pval) {
template <>
__inline__ __device__ void warpReduce<ReduceType::kSum, 2>(float *pval) {
float val0_tmp, val1_tmp;
#define WarpReduceSumOneStep(a, b) \
val0_tmp = __shfl_xor_sync(WARP_REDUCE_MASK, *(pval + 0), a, b); \
val1_tmp = __shfl_xor_sync(WARP_REDUCE_MASK, *(pval + 1), a, b); \
*(pval + 0) += val0_tmp; \
#define WarpReduceSumOneStep(a, b) \
val0_tmp = __shfl_xor_sync(WARP_REDUCE_MASK, *(pval + 0), a, b); \
val1_tmp = __shfl_xor_sync(WARP_REDUCE_MASK, *(pval + 1), a, b); \
*(pval + 0) += val0_tmp; \
*(pval + 1) += val1_tmp
WarpReduceSumOneStep(16, 32);
@ -107,14 +106,14 @@ __inline__ __device__ void warpReduce<ReduceType::kSum, 2>(float *pval) {
template <>
__inline__ __device__ void warpReduce<ReduceType::kSum, 4>(float *pval) {
float val0_tmp, val1_tmp, val2_tmp, val3_tmp;
#define WarpReduceSumOneStep(a, b) \
val0_tmp = __shfl_xor_sync(WARP_REDUCE_MASK, *(pval + 0), a, b); \
val1_tmp = __shfl_xor_sync(WARP_REDUCE_MASK, *(pval + 1), a, b); \
val2_tmp = __shfl_xor_sync(WARP_REDUCE_MASK, *(pval + 2), a, b); \
val3_tmp = __shfl_xor_sync(WARP_REDUCE_MASK, *(pval + 3), a, b); \
*(pval + 0) += val0_tmp; \
*(pval + 1) += val1_tmp; \
*(pval + 2) += val2_tmp; \
#define WarpReduceSumOneStep(a, b) \
val0_tmp = __shfl_xor_sync(WARP_REDUCE_MASK, *(pval + 0), a, b); \
val1_tmp = __shfl_xor_sync(WARP_REDUCE_MASK, *(pval + 1), a, b); \
val2_tmp = __shfl_xor_sync(WARP_REDUCE_MASK, *(pval + 2), a, b); \
val3_tmp = __shfl_xor_sync(WARP_REDUCE_MASK, *(pval + 3), a, b); \
*(pval + 0) += val0_tmp; \
*(pval + 1) += val1_tmp; \
*(pval + 2) += val2_tmp; \
*(pval + 3) += val3_tmp
WarpReduceSumOneStep(16, 32);