LeiWang1999 commented on issue #16566:
URL: https://github.com/apache/tvm/issues/16566#issuecomment-1946764691
> Thanks, that definitely goes in the right direction! I have ported this
over to my local branch. I still get an error though:
>
> ```
> Error message: The stmt tir.For#0 doesn't match the tensor intrin
> The pattern attempting to be matched:
> for j in range(512):
> k = T.int32()
> with T.block("res_update"):
> v_j_i = T.axis.spatial(512, j)
> v_k_i = T.axis.reduce(1024, k)
> res = T.Buffer((1, 512), "int8")
> v_i_o = T.int32()
> a_in = T.Buffer((1, 1024), "int8")
> b_in = T.Buffer((1024, 512), "int8")
> T.reads(res[v_i_o, v_j_i], a_in[v_i_o, v_k_i], b_in[v_k_i, v_j_i])
> T.writes(res[v_i_o, v_j_i])
> res[v_i_o, v_j_i] = res[v_i_o, v_j_i] + a_in[v_i_o, v_k_i] *
b_in[v_k_i, v_j_i]
> Does not match the tensorize description:
> for j in range(512):
> k = T.int32()
> with T.block(""):
> vii = T.axis.spatial(1, 0)
> vjj = T.axis.spatial(512, j)
> vkk = T.axis.reduce(1024, k)
> C = T.Buffer((1, 512), "int8", offset_factor=1)
> A = T.Buffer((1, 1024), "int8", offset_factor=1)
> B = T.Buffer((1024, 512), "int8", offset_factor=1)
> T.reads(C[0, vjj], A[0, vkk], B[vkk, vjj])
> T.writes(C[0, vjj])
> C[0, vjj] = C[0, vjj] + A[0, vkk] * B[vkk, vjj]
> CompareArray array size mismatch. lhs.size()=2 vs rhs.size()=3
> BlockRealizeNode iter_values do not match: op->iter_values=[j, k] vs
rhs->iter_values=[0, j, k]
> ```
>
> It seems like the unit iterator is not actually deleted, but is still
preserved.
btw, should we keep vii after doing simplification even though vii is not
used in this block? @Hzfengsy
--
This is an automated message from the Apache Git Service.
To respond to the message, please log on to GitHub and use the
URL above to go to the specific comment.
To unsubscribe, e-mail: [email protected]
For queries about this service, please contact Infrastructure at:
[email protected]
