Module: Mesa Branch: main Commit: 4d02543853eb86c6c3cb6dd2a84e9a673b44001f URL: http://cgit.freedesktop.org/mesa/mesa/commit/?id=4d02543853eb86c6c3cb6dd2a84e9a673b44001f
Author: Konstantin Seurer <konstantin.seu...@gmail.com> Date: Wed Dec 27 15:19:12 2023 +0100 vtn: Remove transpose(m0)*m1 fast path This is broken for games that rely on invariant geometry since the usage of matrices can affect how gl_Position is computed. The fdot fastpath relied on if and how fdot is lowered for correctness. Totals from 6578 (7.73% of 85071) affected shaders: MaxWaves: 147190 -> 147170 (-0.01%) Instrs: 4451406 -> 4438140 (-0.30%); split: -0.31%, +0.01% CodeSize: 23553020 -> 23541772 (-0.05%); split: -0.07%, +0.03% VGPRs: 302304 -> 302328 (+0.01%) SpillSGPRs: 1309 -> 1329 (+1.53%) Latency: 22509985 -> 22177164 (-1.48%); split: -1.48%, +0.00% InvThroughput: 4862795 -> 4842951 (-0.41%); split: -0.41%, +0.01% VClause: 85035 -> 84998 (-0.04%); split: -0.06%, +0.02% SClause: 131008 -> 131055 (+0.04%); split: -0.02%, +0.05% Copies: 298935 -> 298060 (-0.29%); split: -0.71%, +0.41% PreSGPRs: 266833 -> 267292 (+0.17%); split: -0.85%, +1.03% PreVGPRs: 249511 -> 249601 (+0.04%) Closes: https://gitlab.freedesktop.org/mesa/mesa/-/issues/9562 cc: mesa-stable Reviewed-by: Georg Lehmann <dadschoo...@gmail.com> Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/26821> --- src/compiler/spirv/vtn_alu.c | 40 +++++++++------------------------------- 1 file changed, 9 insertions(+), 31 deletions(-) diff --git a/src/compiler/spirv/vtn_alu.c b/src/compiler/spirv/vtn_alu.c index 4fec751540d..8f8a355dbda 100644 --- a/src/compiler/spirv/vtn_alu.c +++ b/src/compiler/spirv/vtn_alu.c @@ -94,38 +94,16 @@ matrix_multiply(struct vtn_builder *b, transpose_result = true; } - if (src0_transpose && !src1_transpose && - glsl_get_base_type(src0->type) == GLSL_TYPE_FLOAT) { - /* We already have the rows of src0 and the columns of src1 available, - * so we can just take the dot product of each row with each column to - * get the result. - */ - - for (unsigned i = 0; i < src1_columns; i++) { - nir_def *vec_src[4]; - for (unsigned j = 0; j < src0_rows; j++) { - vec_src[j] = nir_fdot(&b->nb, src0_transpose->elems[j]->def, - src1->elems[i]->def); - } - dest->elems[i]->def = nir_vec(&b->nb, vec_src, src0_rows); - } - } else { - /* We don't handle the case where src1 is transposed but not src0, since - * the general case only uses individual components of src1 so the - * optimizer should chew through the transpose we emitted for src1. - */ - - for (unsigned i = 0; i < src1_columns; i++) { - /* dest[i] = sum(src0[j] * src1[i][j] for all j) */ + for (unsigned i = 0; i < src1_columns; i++) { + /* dest[i] = sum(src0[j] * src1[i][j] for all j) */ + dest->elems[i]->def = + nir_fmul(&b->nb, src0->elems[src0_columns - 1]->def, + nir_channel(&b->nb, src1->elems[i]->def, src0_columns - 1)); + for (int j = src0_columns - 2; j >= 0; j--) { dest->elems[i]->def = - nir_fmul(&b->nb, src0->elems[src0_columns - 1]->def, - nir_channel(&b->nb, src1->elems[i]->def, src0_columns - 1)); - for (int j = src0_columns - 2; j >= 0; j--) { - dest->elems[i]->def = - nir_ffma(&b->nb, src0->elems[j]->def, - nir_channel(&b->nb, src1->elems[i]->def, j), - dest->elems[i]->def); - } + nir_ffma(&b->nb, src0->elems[j]->def, + nir_channel(&b->nb, src1->elems[i]->def, j), + dest->elems[i]->def); } }
