Fix TMA test: globalStrides in bytes not elements

This commit is contained in:
2026-05-28 16:46:56 +00:00
parent ad8050bbad
commit 0fd3e12a52

View File

@@ -83,8 +83,10 @@ int main() {
cudaMemcpy(d_src, h_src, DATA_BYTES, cudaMemcpyHostToDevice);
// Create CUtensorMap
// globalDim: sizes in ELEMENTS (x=cols, y=rows)
// globalStrides: strides in BYTES
uint64_t gdim[] = {(uint64_t)COLS, (uint64_t)ROWS};
uint64_t gstr[] = {1, (uint64_t)COLS};
uint64_t gstr[] = {2, (uint64_t)COLS * 2}; // BF16=2 bytes/element, row stride = COLS*2
uint32_t tdim[] = {TILE_COLS, TILE_ROWS};
uint32_t tstr[] = {1, TILE_COLS};