MCPcopy Create free account
hub / github.com/tile-ai/tilelang / convolution

Function convolution

examples/convolution/example_convolution.py:20–50  ·  view source on GitHub ↗
(data, kernel, S, D, P, block_M, block_N, block_K, num_stages, threads, dtype=T.float16, accum_dtype=T.float32)

Source from the content-addressed store, hash-verified

18
19@tilelang.jit
20def 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
53def main(argv=None):

Callers 1

mainFunction · 0.70

Calls

no outgoing calls

Tested by

no test coverage detected

Used in the wild real call sites across dependent graphs

searching dependent graphs…