Blocksparse matmul error under specific block size

[LinjianMa]

I was looking at this test function: https://github.com/ptillet/triton/blob/master/python/test/test_blocksparse.py#L13,
when the blocksize is 16 or greater, it works fine. However, when the block size is set as 4 or 8, the error message looks like this:

Traceback (most recent call last):
  File "test_blocksparse.py", line 41, in <module>
    test_matmul("dds", TRANS_A=False, TRANS_B=False, BLOCK=8, DTYPE="float16", Z=3, H=2, M=512, N=384, K=256)
  File "test_blocksparse.py", line 30, in test_matmul
    rc = op(ra, rb)
  File "/home/ubuntu/triton/python/triton/ops/blocksparse/matmul.py", line 651, in __call__
    db_lut, db_num_locks, db_width, db_packs = self.make_lut(a.dtype, a.device)
  File "/home/ubuntu/triton/python/triton/ops/blocksparse/matmul.py", line 595, in make_lut
    c_lut, c_num_locks, c_width, c_packs = _matmul.make_dxx_lut(layout, block, step, self.trans_b, device)
  File "/home/ubuntu/triton/python/triton/ops/blocksparse/matmul.py", line 391, in make_dxx_lut
    xincs[:, 0] -= (div - 1) * step
IndexError: index 0 is out of bounds for dimension 1 with size 0

Any ideas about what's the reason? To reproduce the error, can run

test_matmul("dds", TRANS_A=False, TRANS_B=False, BLOCK=4, DTYPE="float16", Z=3, H=2, M=512, N=384, K=256)

[ptillet]

Hi! Blocksparse does not support BLOCK=4 or BLOCK=8. There is also no plan support it, as it wouldn't map well to tensor cores. We should have more explicit error message for it.

[LinjianMa]

I see, thanks for the quick response!