(data, kernel, S, D, P, block_M, block_N, block_K, num_stages, threads, dtype=T.float16, accum_dtype=T.float32)
| 18 | |
| 19 | @tilelang.jit |
| 20 | def convolution(data, kernel, S, D, P, block_M, block_N, block_K, num_stages, threads, dtype=T.float16, accum_dtype=T.float32): |
| 21 | N, C, H, W, F, K = T.const("N, C, H, W, F, K") |
| 22 | KH, KW = K, K |
| 23 | OH = (H + 2 * P - D * (K - 1) - 1) // S + 1 |
| 24 | OW = (W + 2 * P - D * (K - 1) - 1) // S + 1 |
| 25 | dtype = T.float16 |
| 26 | accum_dtype = T.float32 |
| 27 | |
| 28 | data: T.Tensor((N, H, W, C), dtype) |
| 29 | kernel: T.Tensor((KH, KW, C, F), dtype) |
| 30 | out = T.empty((N, OH, OW, F), dtype) |
| 31 | |
| 32 | with T.Kernel(T.ceildiv(F, block_N), T.ceildiv(N * OH * OW, block_M), threads=threads) as (bx, by): |
| 33 | data_shared = T.alloc_shared((block_M, block_K), dtype) |
| 34 | kernel_shared = T.alloc_shared((block_K, block_N), dtype) |
| 35 | out_local = T.alloc_fragment((block_M, block_N), accum_dtype) |
| 36 | out_shared = T.alloc_shared((block_M, block_N), dtype) |
| 37 | |
| 38 | kernel_flat = T.Tensor((KH * KW * C, F), dtype, kernel.data) |
| 39 | out_flat = T.Tensor((N * OH * OW, F), dtype, out.data) |
| 40 | |
| 41 | T.clear(out_local) |
| 42 | for k_iter in T.Pipelined(T.ceildiv(KH * KW * C, block_K), num_stages=num_stages): |
| 43 | T.im2col(data, data_shared, by, k_iter, KH, S, D, P) |
| 44 | T.copy(kernel_flat[k_iter * block_K, bx * block_N], kernel_shared) |
| 45 | T.gemm(data_shared, kernel_shared, out_local) |
| 46 | |
| 47 | T.copy(out_local, out_shared) |
| 48 | T.copy(out_shared, out_flat[by * block_M, bx * block_N]) |
| 49 | |
| 50 | return out |
| 51 | |
| 52 | |
| 53 | def main(argv=None): |
no outgoing calls
no test coverage detected
searching dependent graphs…