(
A: T.Tensor(A_shape, in_dtype),
B: T.Tensor(B_shape, in_dtype),
C: T.Tensor((M, N), out_dtype),
)
| 25 | |
| 26 | @T.prim_func |
| 27 | def main( |
| 28 | A: T.Tensor(A_shape, in_dtype), |
| 29 | B: T.Tensor(B_shape, in_dtype), |
| 30 | C: T.Tensor((M, N), out_dtype), |
| 31 | ): |
| 32 | with T.Kernel(T.ceildiv(N, block_N), T.ceildiv(M, block_M), threads=threads) as (bx, by): |
| 33 | A_shared = T.alloc_shared(A_shared_shape, in_dtype) |
| 34 | B_shared = T.alloc_shared(B_shared_shape, in_dtype) |
| 35 | C_local = T.alloc_fragment((block_M, block_N), accum_dtype) |
| 36 | T.clear(C_local) |
| 37 | for k in T.Pipelined(T.ceildiv(K, block_K), num_stages=num_stages): |
| 38 | if trans_A: |
| 39 | T.copy(A[k * block_K, by * block_M], A_shared) |
| 40 | else: |
| 41 | T.copy(A[by * block_M, k * block_K], A_shared) |
| 42 | if trans_B: |
| 43 | T.copy(B[bx * block_N, k * block_K], B_shared) |
| 44 | else: |
| 45 | T.copy(B[k * block_K, bx * block_N], B_shared) |
| 46 | T.gemm(A_shared, B_shared, C_local, trans_A, trans_B) |
| 47 | T.copy(C_local, C[by * block_M, bx * block_N]) |
| 48 | |
| 49 | return main |
| 50 |
nothing calls this directly
no outgoing calls
no test coverage detected
searching dependent graphs…