AlexIndustrial/mesa
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

aco/gfx10+: set lateKill for sgprs used by wave64 VALU writing a mask

Browse Source 
RDNA2 ISA doc, 6.2.4. Wave64 Destination Restrictions:
The first pass of a wave64 VALU instruction may not overwrite a scalar value
used by the second half.

Foz-DB Navi31:
Totals from 5221 (6.58% of 79395) affected shaders:
Instrs: 9751484 -> 9752179 (+0.01%); split: -0.01%, +0.01%
CodeSize: 50624072 -> 50626088 (+0.00%); split: -0.00%, +0.01%
Latency: 85646450 -> 85647419 (+0.00%); split: -0.00%, +0.00%
InvThroughput: 15039160 -> 15039277 (+0.00%); split: -0.00%, +0.00%
VClause: 200275 -> 200204 (-0.04%)
SClause: 248645 -> 248607 (-0.02%); split: -0.03%, +0.01%
Copies: 640802 -> 641413 (+0.10%); split: -0.01%, +0.11%
PreSGPRs: 236297 -> 236735 (+0.19%)
VALU: 5666449 -> 5666440 (-0.00%)
SALU: 967482 -> 968111 (+0.07%); split: -0.01%, +0.07%

Cc: mesa-stable

Reviewed-by: Rhys Perry <pendingchaos02@gmail.com>
Reviewed-by: Daniel Schürmann <daniel@schuermann.dev>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/30368>
This commit is contained in:
Georg Lehmann
2024-07-25 17:01:37 +02:00
committed by Marge Bot
parent 9d0491c12b
commit 510f5e55be
1 changed files with 15 additions and 0 deletions
Download Patch File Download Diff File Expand all files Collapse all files
src/amd/compiler/aco_live_var_analysis.cpp
+15
View File
@@ -199,6 +199,21 @@ process_live_temps_per_block(live_ctx& ctx, Block* block)
}
}
if (ctx.program->gfx_level >= GFX10 && insn->isVALU() &&
insn->definitions.back().regClass() == s2) {
/* RDNA2 ISA doc, 6.2.4. Wave64 Destination Restrictions:
* The first pass of a wave64 VALU instruction may not overwrite a scalar value used by
* the second half.
*/
bool carry_in = insn->opcode == aco_opcode::v_addc_co_u32 ||
insn->opcode == aco_opcode::v_subb_co_u32 ||
insn->opcode == aco_opcode::v_subbrev_co_u32;
for (unsigned op_idx = 0; op_idx < (carry_in ? 2 : insn->operands.size()); op_idx++) {
if (insn->operands[op_idx].isOfType(RegType::sgpr))
insn->operands[op_idx].setLateKill(true);
}
}
/* we need to do this in a separate loop because the next one can
* setKill() for several operands at once and we don't want to
* overwrite that in a later iteration */