fix: TMEM read/dealloc for 128-thread kernel

This commit is contained in:
2026-05-29 18:40:24 +00:00
parent bd1309ba88
commit c755e6fdde

View File

@@ -91,7 +91,7 @@ test_qk_minimal_kernel(float* __restrict__ out_s,
__syncthreads();
// Read from TMEM
if (wid == 0) {
if (lane < T) {
for (int n = 0; n < NUM_READS; n++) {
float tmp[8];
asm volatile("tcgen05.ld.sync.aligned.32x32b.x8.b32 {%0,%1,%2,%3,%4,%5,%6,%7},[%8];"
@@ -108,7 +108,7 @@ test_qk_minimal_kernel(float* __restrict__ out_s,
}
}
__syncthreads();
if (is_mma_warp) tmem_dealloc(tb, TMEM_N);
tmem_dealloc(tb, TMEM_N);
}
int main() {