(M, N, K, block_M, block_N, block_K, dtype=T.float16, accum_dtype=T.float32, with_bias=False)
| 7 | |
| 8 | @tl.jit |
| 9 | def tensor_null_test(M, N, K, block_M, block_N, block_K, dtype=T.float16, accum_dtype=T.float32, with_bias=False): |
| 10 | @T.prim_func |
| 11 | def main( |
| 12 | A: T.Tensor((M, K), dtype), |
| 13 | B: T.Tensor((K, N), dtype), |
| 14 | C: T.Tensor((M, N), accum_dtype), |
| 15 | Bias: T.Tensor((N), accum_dtype), |
| 16 | ): |
| 17 | # Initialize Kernel Context |
| 18 | with T.Kernel(T.ceildiv(N, block_N), T.ceildiv(M, block_M), threads=128) as (bx, by): |
| 19 | A_shared = T.alloc_shared((block_M, block_K), dtype) |
| 20 | B_shared = T.alloc_shared((block_N, block_K), dtype) |
| 21 | C_local = T.alloc_fragment((block_M, block_N), accum_dtype) |
| 22 | |
| 23 | T.clear(C_local) |
| 24 | |
| 25 | for ko in T.Pipelined(T.ceildiv(K, block_K), num_stages=3): |
| 26 | # Copy tile of A |
| 27 | T.copy(A[by * block_M, ko * block_K], A_shared) |
| 28 | T.copy(B[bx * block_N, ko * block_K], B_shared) |
| 29 | T.gemm(A_shared, B_shared, C_local, transpose_B=True) |
| 30 | |
| 31 | if with_bias: |
| 32 | for i, j in T.Parallel(block_M, block_N): |
| 33 | C_local[i, j] += Bias[bx * block_N + j] |
| 34 | |
| 35 | T.copy(C_local, C[by * block_M, bx * block_N]) |
| 36 | |
| 37 | return main |
| 38 | |
| 39 | |
| 40 | def run_test(M, N, K, block_M, block_N, block_K, dtype=T.float16, accum_dtype=T.float32): |
no outgoing calls
no test coverage detected
searching dependent graphs…