https://gcc.gnu.org/bugzilla/show_bug.cgi?id=115640
--- Comment #14 from Andrew Stubbs <ams at gcc dot gnu.org> --- On 26/06/2024 13:34, rguenth at gcc dot gnu.org wrote: > https://gcc.gnu.org/bugzilla/show_bug.cgi?id=115640 > > --- Comment #13 from Richard Biener <rguenth at gcc dot gnu.org> --- > (In reply to Richard Biener from comment #12) >> (In reply to Andrew Stubbs from comment #10) >>> GFX10 has more limited permutation capabilities than GFX9 because it >>> only has 32-lane vectors natively, even though we're using the 64-lane >>> "compatibility" mode. >>> >>> However, in theory, the permutation capabilities on V32 and below should >>> be the same, and some permutations on V64 are allowed, so I don't know >>> why it doesn't use it. It's possible I broke the logic in >>> gcn_vectorize_vec_perm_const: >>> >>> /* RDNA devices can only do permutations within each group of 32-lanes. >>> Reject permutations that cross the boundary. */ >>> if (TARGET_RDNA2_PLUS) >>> for (unsigned int i = 0; i < nelt; i++) >>> if (i < 31 ? perm[i] > 31 : perm[i] < 32) >>> return false; >>> >>> It looks right to me though? >> >> nelt == 32 so I think the last element has the wrong check applied? >> >> It should be >> >>> if (i < 32 ? perm[i] > 31 : perm[i] < 32) >> >> I think. With that the vectorization happens in a similar way but the >> failure still doesn't reproduce (without the patch, of course). Oops, I think you're right. > Btw, the above looks quite odd for nelt == 32 anyway - we are permuting > two vectors src0 and src1 into one 32 element dst vector (it's no longer > required that src0 and src1 line up with the dst vector size btw, they > might have different nelt). So the loop would reject interleaving > the low parts of two 32 element vectors, a permute that would look like > { 0, 32, 1, 33, 2, 34 ... } so does "within each group of 32-lanes" > mean you can never mix the two vector inputs? Or does GCN not have > a two-to-one vector permute instruction? GCN does not have two-to-one vector permute in hardware, so we do two permutes and a vec_merge to get the same effect. GFX9 can permute all the elements within a 64 lane vector arbitrarily. GFX10 and GFX11 can permute the low-32 and high-32 elements freely, but no value may cross the boundary. AFAIK there's no way to do that via any vector instruction (i.e. without writing to memory, or extracting values element-wise). In theory, we could implement permutes with different sized inputs and outputs, but right now those are rejected early. The interleave example wouldn't work in hardware, for GFX10, but we could have it for GFX9. However, I think you might be right about the numbering of the "perm" array; we probably need to be testing "(perm[i] % nelt) > 31" if we are to support two-to-one permutations. Thanks for looking at this. Andrew
