fdo-mirrors/mesa
1
0
Fork 0
mirror of https://gitlab.freedesktop.org/mesa/mesa.git synced 2026-01-21 13:00:22 +01:00
Code Issues Projects Releases Packages Wiki Activity Actions 2

aco: remove SMEM_instruction::prevent_overflow

Browse source 
This doesn't seem useful anymore, and it seems we forgot to set it in a
few places.

This commit changes the behaviour of the optimizer so that
prevent_overflow is always true.

fossil-db (navi21):
Totals from 7421 (5.47% of 135636) affected shaders:
Instrs: 5402823 -> 5440126 (+0.69%); split: -0.00%, +0.69%
CodeSize: 28731300 -> 28974152 (+0.85%); split: -0.00%, +0.85%
VGPRs: 317528 -> 317552 (+0.01%)
SpillSGPRs: 419 -> 415 (-0.95%)
Latency: 40712478 -> 40783115 (+0.17%); split: -0.01%, +0.19%
InvThroughput: 7612708 -> 7616751 (+0.05%); split: -0.00%, +0.06%
VClause: 123824 -> 123848 (+0.02%); split: -0.09%, +0.11%
SClause: 161915 -> 172741 (+6.69%); split: -0.03%, +6.71%
Copies: 393015 -> 394429 (+0.36%); split: -0.20%, +0.56%
PreSGPRs: 288658 -> 289603 (+0.33%); split: -0.04%, +0.36%

Signed-off-by: Rhys Perry <pendingchaos02@gmail.com>
Reviewed-by: Qiang Yu <yuq825@gmail.com>
Reviewed-by: Timur Kristóf <timur.kristof@gmail.com>
Closes: https://gitlab.freedesktop.org/mesa/mesa/-/issues/8864
Cc: mesa-stable
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/22553>
(cherry picked from commit 1a6095b36e)
This commit is contained in:
Rhys Perry 2023-04-18 14:50:18 +01:00 • committed by Dylan Baker
parent c67e28251e
commit 20b252c4c8
5 changed files with 5 additions and 9 deletions
Show stats Download patch file Download diff file Expand all files Collapse all files

2
.pick_status.json
View file

@ -3174,7 +3174,7 @@
"description": "aco: remove SMEM_instruction::prevent_overflow",
"nominated": true,
"nomination_type": 0,
"resolution": 0,
"resolution": 1,
"main_sha": null,
"because_sha": null,
"notes": null

2
src/amd/compiler/aco_instruction_selection.cpp
View file

@ -6020,7 +6020,7 @@ visit_load_push_constant(isel_context* ctx, nir_intrinsic_instr* instr)
default: unreachable("unimplemented or forbidden load_push_constant.");
}
bld.smem(op, Definition(vec), ptr, index)->smem().prevent_overflow = true;
bld.smem(op, Definition(vec), ptr, index);
if (!aligned) {
Operand byte_offset = index_cv ? Operand::c32((offset + index_cv->u32) % 4) : Operand(index);

3
src/amd/compiler/aco_ir.h
View file

@ -1428,8 +1428,7 @@ struct SMEM_instruction : public Instruction {
bool dlc : 1; /* NAVI: device level coherent */
bool nv : 1; /* VEGA only: Non-volatile */
bool disable_wqm : 1;
bool prevent_overflow : 1; /* avoid overflow when combining additions */
uint8_t padding : 3;
uint8_t padding : 4;
};
static_assert(sizeof(SMEM_instruction) == sizeof(Instruction) + 4, "Unexpected padding");

4
src/amd/compiler/aco_opt_value_numbering.cpp
View file

@ -218,10 +218,8 @@ struct InstrPred {
case Format::SMEM: {
SMEM_instruction& aS = a->smem();
SMEM_instruction& bS = b->smem();
/* isel shouldn't be creating situations where this assertion fails */
assert(aS.prevent_overflow == bS.prevent_overflow);
return aS.sync == bS.sync && aS.glc == bS.glc && aS.dlc == bS.dlc && aS.nv == bS.nv &&
aS.disable_wqm == bS.disable_wqm && aS.prevent_overflow == bS.prevent_overflow;
aS.disable_wqm == bS.disable_wqm;
}
case Format::VINTRP: {
VINTRP_instruction& aI = a->vintrp();

3
src/amd/compiler/aco_optimizer.cpp
View file

@ -879,13 +879,12 @@ smem_combine(opt_ctx& ctx, aco_ptr<Instruction>& instr)
Temp base;
uint32_t offset;
bool prevent_overflow = smem.operands[0].size() > 2 || smem.prevent_overflow;
if (info.is_constant_or_literal(32) &&
((ctx.program->gfx_level == GFX6 && info.val <= 0x3FF) ||
(ctx.program->gfx_level == GFX7 && info.val <= 0xFFFFFFFF) ||
(ctx.program->gfx_level >= GFX8 && info.val <= 0xFFFFF))) {
instr->operands[1] = Operand::c32(info.val);
} else if (parse_base_offset(ctx, instr.get(), 1, &base, &offset, prevent_overflow) &&
} else if (parse_base_offset(ctx, instr.get(), 1, &base, &offset, true) &&
base.regClass() == s1 && offset <= 0xFFFFF && ctx.program->gfx_level >= GFX9 &&
offset % 4u == 0) {
bool soe = smem.operands.size() >= (!smem.definitions.empty() ? 3 : 4);