(
A: T.Tensor((M, K), dtype),
B: T.Tensor((K, N), dtype),
C: T.Tensor((M, N), dtype),
)
| 7 | def matmul(M, N, K, block_M, block_N, block_K, dtype=T.float16, accum_dtype=T.float32): |
| 8 | @T.prim_func |
| 9 | def main( |
| 10 | A: T.Tensor((M, K), dtype), |
| 11 | B: T.Tensor((K, N), dtype), |
| 12 | C: T.Tensor((M, N), dtype), |
| 13 | ): |
| 14 | # Initialize Kernel Context |
| 15 | with T.Kernel(T.ceildiv(N, block_N), T.ceildiv(M, block_M), threads=128) as (bx, by): |
| 16 | A_shared = T.alloc_shared((block_M, block_K), dtype) |
| 17 | B_shared = T.alloc_shared((block_K, block_N), dtype) |
| 18 | C_local = T.alloc_fragment((block_M, block_N), accum_dtype) |
| 19 | |
| 20 | T.clear(C_local) |
| 21 | |
| 22 | for ko in T.Pipelined(T.ceildiv(K, block_K), num_stages=0): |
| 23 | T.copy(A[by * block_M, ko * block_K], A_shared, coalesced_width=2) |
| 24 | T.copy(B[ko * block_K, bx * block_N], B_shared, coalesced_width=2) |
| 25 | |
| 26 | for i, j, k in T.grid(block_M, block_N, block_K): |
| 27 | C_local[i, j] += A_shared[i, k] * B_shared[k, j] |
| 28 | |
| 29 | T.copy(C_local, C[by * block_M, bx * block_N], coalesced_width=2) |
| 30 | |
| 31 | return main |
| 32 |
nothing calls this directly
no outgoing calls
no test coverage detected
searching dependent graphs…