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

Function ref_program

testing/python/language/test_tilelang_language_scan.py:61–75  ·  view source on GitHub ↗
(A)

Source from the content-addressed store, hash-verified

59 A = torch.randn(M, N, dtype=getattr(torch, dtype)).cuda()
60
61 def ref_program(A):
62 ref_b = torch.empty_like(A)
63 for i in range(M // block_M):
64 for j in range(N // block_N):
65 ref_b[i * block_M : (i + 1) * block_M, j * block_N : (j + 1) * block_N] = A[
66 i * block_M : (i + 1) * block_M, j * block_N : (j + 1) * block_N
67 ].cumsum(dim=dim)
68 if reverse:
69 ref_b[i * block_M : (i + 1) * block_M, j * block_N : (j + 1) * block_N] = (
70 A[i * block_M : (i + 1) * block_M, j * block_N : (j + 1) * block_N]
71 .flip(dims=[dim])
72 .cumsum(dim=dim)
73 .flip(dims=[dim])
74 )
75 return ref_b
76
77 tilelang_res = jit_kernel(A)
78 ref_res = ref_program(A)

Callers 6

run_cumsumFunction · 0.70
run_cumsum_1dFunction · 0.70
run_cumsum_region_1dFunction · 0.70
run_cumsum_region_2dFunction · 0.70
run_cummaxFunction · 0.70
run_cummax_1dFunction · 0.70

Calls 2

_torch_cummaxFunction · 0.85
minFunction · 0.50

Tested by

no test coverage detected

Used in the wild real call sites across dependent graphs

searching dependent graphs…