[Question] examples/gemm/example_gemm_intrinsic.py 在A5（PR）上会出现精度问题

[37ETs]

我把代码中的target改为‘pto’之后，运行后精度出现报错，请问是否是：
cid = T.use_swizzle(i * core_num + cid, M, N, K, block_M, block_N, off=3)
这个分核模式A5不支持？或者是其他问题？

[37ETs]

终端输出：

init successful!
[W723 08:15:37.556112904 ToKernelNpu.cpp:40] Warning: Device do not support double dtype now, dtype cast replace with float. (function operator())
tensor([[ 93.5000, -67.0000, 23.5469, ..., -95.4375, -108.4375,
62.2500],
[ 13.2656, 33.8750, 35.7500, ..., -35.9062, -113.5000,
0.9272],
[ 26.2812, 18.5938, 37.8438, ..., -14.6250, -81.2500,
-0.6689],
...,
[ -93.6250, -29.5781, -109.6250, ..., -17.6562, -71.1250,
-84.1875],
[-112.1875, 32.4688, 99.8750, ..., 152.8750, -55.3125,
141.7500],
[ 85.6250, -61.6875, 23.7344, ..., 13.6641, -130.8750,
126.6875]], device='npu:0', dtype=torch.float16) tensor([[ 93.5000, -67.0000, 23.5469, ..., 192.5000, 9.1875,
173.2500],
[ 13.2656, 33.8750, 35.7500, ..., 44.2812, 55.9062,
3.1660],
[ 26.2812, 18.5938, 37.8438, ..., -16.3906, 84.6875,
59.7188],
...,
[ -93.6250, -29.5781, -109.6250, ..., -14.9766, 66.6875,
8.4922],
[-112.1875, 32.4688, 99.8750, ..., 52.9062, 78.2500,
34.9375],
[ 85.6250, -61.6875, 23.7344, ..., 56.3438, -73.1875,
25.7656]], device='npu:0', dtype=torch.float16)
tilelang time: 0.6092355847358704 ms
torch time: 0.4492858052253723 ms
Traceback (most recent call last):
File "/data1/casperpan/tilelang-ascend/examples/gemm/example_gemm_intrinsic_pto.py", line 132, in
torch.testing.assert_close(c, ref_c, rtol=1e-2, atol=1e-2)
File "/root/miniconda3/envs/cy/lib/python3.10/site-packages/torch/testing/_comparison.py", line 1589, in assert_close
raise error_metas[0].to_error(msg)
AssertionError: Tensor-likes are not close!

Mismatched elements: 7838937 / 8388608 (93.4%)
Greatest absolute difference: 710.0 at index (4854, 17) (up to 0.01 allowed)
Greatest relative difference: inf at index (18, 74) (up to 0.01 allowed)
[ERROR] 2026-07-23-08:15:38 (PID:3753964, Device:0, RankID:-1) ERR99999 UNKNOWN applicaiton exception