From fd3fa98fe4b504059235d3ac3d0eec687a387fe6 Mon Sep 17 00:00:00 2001 From: z00520135 Date: Fri, 15 May 2026 10:49:34 +0800 Subject: [PATCH 1/2] optimic elementwise pytest --- .../test_tilelang_ascend_language_elementwise.py | 10 +++++----- 1 file changed, 5 insertions(+), 5 deletions(-) diff --git a/testing/python/language/test_tilelang_ascend_language_elementwise.py b/testing/python/language/test_tilelang_ascend_language_elementwise.py index 958b0b258..99f94e33a 100644 --- a/testing/python/language/test_tilelang_ascend_language_elementwise.py +++ b/testing/python/language/test_tilelang_ascend_language_elementwise.py @@ -1298,7 +1298,7 @@ def cast(M, N, block_M, block_N, mode, count): m_num = M // block_M n_num = N // block_N - VEC_NUM = 1 + VEC_NUM = 2 @T.prim_func def main( @@ -1349,7 +1349,7 @@ def cast_slice(M, N, block_M, block_N, mode, count): m_num = M // block_M n_num = N // block_N - VEC_NUM = 1 + VEC_NUM = 2 @T.prim_func def main( @@ -1400,7 +1400,7 @@ def cast_scale(M, N, block_M, block_N, mode, count, scale): m_num = M // block_M n_num = N // block_N - VEC_NUM = 1 + VEC_NUM = 2 @T.prim_func def main( @@ -1991,7 +1991,7 @@ def createvecindex(M, N, block_M, block_N, firstValue, dtype="int32"): m_num = M // block_M n_num = N // block_N - VEC_NUM = 1 + VEC_NUM = 2 @T.prim_func def main( @@ -2003,7 +2003,7 @@ def main( c_ub = T.alloc_ub((block_M // VEC_NUM, block_N), dtype) - T.tile.createvecindex(c_ub, firstValue) + T.tile.createvecindex(c_ub, firstValue + vid * block_M // VEC_NUM) T.copy(c_ub, C[bx * block_M + vid * block_M // VEC_NUM, by * block_N]) From 6356092d7ee9db18cd835ddb6e114eada6387521 Mon Sep 17 00:00:00 2001 From: z00520135 Date: Mon, 18 May 2026 10:45:17 +0800 Subject: [PATCH 2/2] submit sync optimic --- src/transform/common/operation_config.h | 3 ++- .../test_tilelang_ascend_language_elementwise.py | 10 +++++----- 2 files changed, 7 insertions(+), 6 deletions(-) diff --git a/src/transform/common/operation_config.h b/src/transform/common/operation_config.h index 4f9bda62b..f7dd2177b 100644 --- a/src/transform/common/operation_config.h +++ b/src/transform/common/operation_config.h @@ -195,7 +195,8 @@ GetOperationConfig() { {{{0, "write"}, {1, "read"}, {2, "read"}}, "PIPE_V"}}, {"tl.ascend_sort32", {{{0, "write"}, {1, "read"}, {2, "read"}}, "PIPE_V"}}, - {"tl.ascend_createvecindex", {{{0, "write"}, {1, "read"}}, "PIPE_V"}}, + {"tl.ascend_createvecindex", {{{1, "write"}}, "PIPE_V"}}, + {"tl.ascend_arith_progression", {{{1, "write"}}, "PIPE_V"}}, {"tl.ascend_sin", {{{0, "write"}, {1, "read"}, {2, "read"}}, "PIPE_V"}}, {"tl.ascend_cos", {{{0, "write"}, {1, "read"}, {2, "read"}}, "PIPE_V"}}, {"tl.ascend_transpose", {{{0, "write"}, {1, "read"}}, "PIPE_V"}}, diff --git a/testing/python/language/test_tilelang_ascend_language_elementwise.py b/testing/python/language/test_tilelang_ascend_language_elementwise.py index 99f94e33a..958b0b258 100644 --- a/testing/python/language/test_tilelang_ascend_language_elementwise.py +++ b/testing/python/language/test_tilelang_ascend_language_elementwise.py @@ -1298,7 +1298,7 @@ def cast(M, N, block_M, block_N, mode, count): m_num = M // block_M n_num = N // block_N - VEC_NUM = 2 + VEC_NUM = 1 @T.prim_func def main( @@ -1349,7 +1349,7 @@ def cast_slice(M, N, block_M, block_N, mode, count): m_num = M // block_M n_num = N // block_N - VEC_NUM = 2 + VEC_NUM = 1 @T.prim_func def main( @@ -1400,7 +1400,7 @@ def cast_scale(M, N, block_M, block_N, mode, count, scale): m_num = M // block_M n_num = N // block_N - VEC_NUM = 2 + VEC_NUM = 1 @T.prim_func def main( @@ -1991,7 +1991,7 @@ def createvecindex(M, N, block_M, block_N, firstValue, dtype="int32"): m_num = M // block_M n_num = N // block_N - VEC_NUM = 2 + VEC_NUM = 1 @T.prim_func def main( @@ -2003,7 +2003,7 @@ def main( c_ub = T.alloc_ub((block_M // VEC_NUM, block_N), dtype) - T.tile.createvecindex(c_ub, firstValue + vid * block_M // VEC_NUM) + T.tile.createvecindex(c_ub, firstValue) T.copy(c_ub, C[bx * block_M + vid * block_M // VEC_NUM, by * block_N])