microsoft/mscclpp
1
0
Fork 0
mirror of https://github.com/microsoft/mscclpp.git synced 2026-04-20 14:59:29 +00:00
Code Issues Packages Projects Releases Wiki Activity

AlltoAll Test Support (#606)

Browse Source 
Co-authored-by: Binyang Li <binyli@microsoft.com>
This commit is contained in:
Caio Rocha
2025-08-15 16:00:41 -07:00
committed by GitHub
parent 2b40fe37b3
commit 9261b1d278
13 changed files with 778 additions and 23 deletions
Download Patch File Download Diff File Expand all files Collapse all files

24
python/test/executor_test_verifier.cu
View File

@@ -98,4 +98,26 @@ TEST_DATA_ALL_REDUCE(int32, int)
TEST_DATA_REDUCE_SCATTER(float16, __half)
TEST_DATA_REDUCE_SCATTER(float32, float)
TEST_DATA_REDUCE_SCATTER(int32, int)
TEST_DATA_REDUCE_SCATTER(int32, int)
#define TEST_DATA_ALL_TO_ALL(FuncNameType, DataType) \
extern "C" __global__ void __launch_bounds__(1024, 1) test_data_all_to_all_##FuncNameType( \
DataType* result_buf, DataType* test_buf, size_t num_elems, int num_ranks, int my_rank, int seq) { \
int nem_elems_per_rank = num_elems / num_ranks; \
int offset = nem_elems_per_rank * my_rank; \
for (int rank = 0; rank < num_ranks; rank++) { \
size_t rank_offset = rank * nem_elems_per_rank; \
unsigned int seed = (unsigned int)(blockIdx.x * blockDim.x + threadIdx.x + rank + seq); \
for (size_t i = blockIdx.x * blockDim.x + threadIdx.x; i < num_elems; i += blockDim.x * gridDim.x) { \
seed = ranqd1(seed); \
if (i >= my_rank * nem_elems_per_rank && i < (my_rank + 1) * nem_elems_per_rank) { \
test_buf[rank_offset + i - offset] = DataType(seed % blockDim.x) / DataType(blockDim.x); \
assert(result_buf[rank_offset + i - offset] == test_buf[rank_offset + i - offset]); \
} \
} \
} \
}
TEST_DATA_ALL_TO_ALL(float16, __half)
TEST_DATA_ALL_TO_ALL(float32, float)
TEST_DATA_ALL_TO_ALL(int32, int)