fdo-mirrors/mesa
1
0
Fork 0
mirror of https://gitlab.freedesktop.org/mesa/mesa.git synced 2026-01-01 03:10:09 +01:00
Code Issues Projects Releases Packages Wiki Activity Actions 1
177217 commits 64 branches 914 tags 948 MiB
Commit graph

345 commits

Author SHA1 Message Date
Georg Lehmann
87d13ee73d aco: combine a | ~b to bfi(b, a, -1) 
Somehow I missed this when writing the a & ~b patch.

Foz-DB Navi21:
Totals from 1591 (1.20% of 132657) affected shaders:
Instrs: 2316379 -> 2315940 (-0.02%)
CodeSize: 12524240 -> 12528724 (+0.04%); split: -0.00%, +0.04%
Latency: 45393195 -> 45389285 (-0.01%); split: -0.01%, +0.00%
InvThroughput: 8658991 -> 8657944 (-0.01%); split: -0.01%, +0.00%
Copies: 135777 -> 135778 (+0.00%)

Reviewed-by: Daniel Schürmann <daniel@schuermann.dev>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/24505>
 2023-08-23 20:06:49 +00:00
Rhys Perry
9169fbf83c aco: clarify bpermute pseudo opcode names 
Signed-off-by: Rhys Perry <pendingchaos02@gmail.com>
Reviewed-by: Daniel Schürmann <daniel@schuermann.dev>
Reviewed-by: Timur Kristóf <timur.kristof@gmail.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/24693>
 2023-08-23 12:36:46 +01:00
Georg Lehmann
1659d982c3 aco: combine a & ~b to bfi(b, 0, a) 
Foz-DB Navi21:
Totals from 905 (0.68% of 132657) affected shaders:
Instrs: 1223583 -> 1221016 (-0.21%); split: -0.22%, +0.01%
CodeSize: 6567272 -> 6567064 (-0.00%); split: -0.04%, +0.03%
SpillSGPRs: 1231 -> 1223 (-0.65%)
SpillVGPRs: 829 -> 823 (-0.72%); split: -1.45%, +0.72%
Latency: 40952209 -> 40946230 (-0.01%); split: -0.02%, +0.01%
InvThroughput: 9411929 -> 9397932 (-0.15%); split: -0.17%, +0.02%
VClause: 29108 -> 29112 (+0.01%); split: -0.04%, +0.05%
Copies: 105272 -> 105221 (-0.05%); split: -0.28%, +0.23%
Branches: 29330 -> 29329 (-0.00%)

Reviewed-by: Rhys Perry <pendingchaos02@gmail.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/24383>
 2023-08-01 07:15:28 +00:00
Georg Lehmann
5099137612 aco/optimizer: delete s_bitcmp optimization 
This is now done in NIR.
No Foz-DB changes.

Reviewed-by: Rhys Perry <pendingchaos02@gmail.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/23298>
 2023-06-29 13:39:30 +00:00
Timur Kristóf
05928f4200 aco: Use ac_hw_stage instead of aco-specific HWStage. 
The new ac_hw_stage is going to be used by drivers as well.

Signed-off-by: Timur Kristóf <timur.kristof@gmail.com>
Reviewed-by: Qiang Yu <yuq825@gmail.com>
Reviewed-by: Daniel Schürmann <daniel@schuermann.dev>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/23597>
 2023-06-23 12:49:04 +00:00
Georg Lehmann
71d30bcede aco: combine scalar mul+pk_add to pk_fma 
Foz-DB Navi21:
Totals from 12 (0.01% of 134913) affected shaders:
CodeSize: 37860 -> 37668 (-0.51%)
Instrs: 6757 -> 6733 (-0.36%)
Latency: 25632 -> 25589 (-0.17%)
InvThroughput: 2637 -> 2622 (-0.57%)

Reviewed-by: Daniel Schürmann <daniel@schuermann.dev>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/21819>
 2023-06-20 14:29:24 +00:00
Georg Lehmann
6db61d0dc0 aco: use uses helpers for pk_fma opt 
Reviewed-by: Daniel Schürmann <daniel@schuermann.dev>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/21819>
 2023-06-20 14:29:24 +00:00
Eric Engestrom
6b21653ab4 aco: reformat according to its .clang-format 
Signed-off-by: Eric Engestrom <eric@igalia.com>
Reviewed-by: Daniel Schürmann <daniel@schuermann.dev>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/23253>
 2023-06-16 19:59:52 +00:00
Georg Lehmann
dfb6d3e443 aco: use v_fma_mix for f2f32 and f2f16 on gfx11 if wave64 
v_fma_mix can be dual issued, trade some code size for throughput.

Foz-DB GFX1100:
Totals from 8204 (6.08% of 134864) affected shaders:
CodeSize: 89608584 -> 89693968 (+0.10%)
Latency: 160744811 -> 160699309 (-0.03%); split: -0.03%, +0.00%
InvThroughput: 19737977 -> 19678308 (-0.30%)

Reviewed-by: Rhys Perry <pendingchaos02@gmail.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/21402>
 2023-06-07 12:30:11 +00:00
Georg Lehmann
177dba62a1 aco: use v_add_f{16,32} with clamp for fsat 
v_add can be dual issued on gfx11, v_med3 cannot.
Don't use v_add directly to still optimize omod(fsat(x)).

Foz-DB GFX1100:
Totals from 32702 (24.24% of 134913) affected shaders:
Latency: 475008203 -> 474928037 (-0.02%); split: -0.02%, +0.00%
InvThroughput: 59226198 -> 59140787 (-0.14%); split: -0.14%, +0.00%

Reviewed-by: Rhys Perry <pendingchaos02@gmail.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/21402>
 2023-06-07 12:30:11 +00:00
Georg Lehmann
a7cef01db1 aco/optimizer: allow DPP to use VOP3 on GFX11 
Reviewed-by: Rhys Perry <pendingchaos02@gmail.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/23059>
 2023-05-18 22:57:24 +00:00
Georg Lehmann
476149f90d aco: use can_use_input_modifiers helper 
Foz-DB GFX1100:
Totals from 80 (0.06% of 132657) affected shaders:
CodeSize: 504500 -> 503660 (-0.17%)
Instrs: 95033 -> 94824 (-0.22%)
Latency: 629695 -> 629235 (-0.07%)
InvThroughput: 97105 -> 97008 (-0.10%)
VClause: 1779 -> 1777 (-0.11%)
Copies: 3233 -> 3236 (+0.09%); split: -0.03%, +0.12%

Reviewed-by: Rhys Perry <pendingchaos02@gmail.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/23059>
 2023-05-18 22:57:24 +00:00
Georg Lehmann
644c5e95a0 aco: use get_operand_size for dpp opt 
This matters now that v_fma_mixlo_f16/v_fma_mixhi_f16 can use dpp.

Reviewed-by: Rhys Perry <pendingchaos02@gmail.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/23059>
 2023-05-18 22:57:24 +00:00
Georg Lehmann
12b28d64ab aco/gfx11: use fmamk/fmaak with opsel 
Reviewed-by: Rhys Perry <pendingchaos02@gmail.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/23059>
 2023-05-18 22:57:23 +00:00
Georg Lehmann
6a53af3fc8 aco: introduce helper to swap valu operands with modifiers 
Reviewed-by: Rhys Perry <pendingchaos02@gmail.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/23059>
 2023-05-18 22:57:23 +00:00
Georg Lehmann
04976beac7 aco: don't apply dpp if the alu instr uses the operand twice 
CP77 has a ton of fma(dpp(a), dpp(a), b).

Reviewed-by: Rhys Perry <pendingchaos02@gmail.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/22698>
 2023-05-12 13:31:16 +00:00
Georg Lehmann
151bcc1e8b aco: use VOP3+DPP 
Reviewed-by: Rhys Perry <pendingchaos02@gmail.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/22698>
 2023-05-12 13:31:16 +00:00
Georg Lehmann
d0e73cb313 aco/optimizer: copy pass flags for newly created valu instructions 
Reviewed-by: Rhys Perry <pendingchaos02@gmail.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/22698>
 2023-05-12 13:31:15 +00:00
Georg Lehmann
d27e03d719 aco/optimizer: don't use pass_flags for mad idx 
fma can use DPP on GFX11+, so we want to keep the exec id in pass_flags

Reviewed-by: Rhys Perry <pendingchaos02@gmail.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/22698>
 2023-05-12 13:31:15 +00:00
Timur Kristóf
38447b3f63 aco: Disallow constant propagation on SOPP and fixed operands. 
Cc: mesa-stable
Signed-off-by: Timur Kristóf <timur.kristof@gmail.com>
Reviewed-by: Rhys Perry <pendingchaos02@gmail.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/22690>
 2023-05-04 19:08:58 +00:00
Rhys Perry
1a6095b36e aco: remove SMEM_instruction::prevent_overflow 
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>
 2023-04-19 19:29:48 +00:00
Harri Nieminen
aea48a4ff1 amd: fix typos 
Reviewed-by: Timur Kristóf <timur.kristof@gmail.com>
Reviewed-by: Marek Olšák <marek.olsak@amd.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/22432>
 2023-04-13 23:08:22 +00:00
Timur Kristóf
cff02468c6 aco: Fix optimization of v_cmp with subgroup invocation. 
There was a typo in this optimization which went unnoticed.

Fixes: 2c40215ab9
Signed-off-by: Timur Kristóf <timur.kristof@gmail.com>
Reviewed-by: Georg Lehmann <dadschoorse@gmail.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/22393>
 2023-04-10 19:15:27 +00:00
Rhys Perry
0f60c18f29 aco: don't optimize s_or_b64(v_cmp_u_f32(a, b), cmp(a, a)) 
Signed-off-by: Rhys Perry <pendingchaos02@gmail.com>
Reviewed-by: Georg Lehmann <dadschoorse@gmail.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/22214>
 2023-03-31 19:41:54 +00:00
Georg Lehmann
bb7c2b70c1 aco/optimizer: remove to_SDWA 
Reviewed-by: Rhys Perry <pendingchaos02@gmail.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/22069>
 2023-03-30 03:34:34 +00:00
Georg Lehmann
e699a4181c aco: keep label_mul/usedef/minmax in apply_extract 
16bit int mad/fma/minmax combining can work with opsel set.

All other optimizations should already check if the instruction uses sdwa,
because we don't check this when applying the label initially.

Reviewed-by: Rhys Perry <pendingchaos02@gmail.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/22069>
 2023-03-30 03:34:34 +00:00
Georg Lehmann
7014145ab2 aco/optimizer: use opsel for VOP12C 
Foz-DB GFX1100:
Totals from 11759 (8.72% of 134864) affected shaders:
VGPRs: 848288 -> 844556 (-0.44%); split: -0.44%, +0.00%
SpillSGPRs: 8527 -> 8543 (+0.19%)
SpillVGPRs: 1411 -> 1423 (+0.85%); split: -0.21%, +1.06%
CodeSize: 114337120 -> 113882472 (-0.40%); split: -0.40%, +0.01%
Scratch: 128768 -> 129024 (+0.20%); split: -0.20%, +0.40%
MaxWaves: 250962 -> 252014 (+0.42%)
Instrs: 22187426 -> 22062378 (-0.56%); split: -0.57%, +0.00%
Latency: 232655375 -> 232376977 (-0.12%); split: -0.20%, +0.08%
InvThroughput: 28292530 -> 28217699 (-0.26%); split: -0.45%, +0.18%
VClause: 352463 -> 352364 (-0.03%); split: -0.12%, +0.10%
SClause: 659282 -> 659354 (+0.01%); split: -0.02%, +0.04%
Copies: 1371369 -> 1342340 (-2.12%); split: -2.30%, +0.19%
Branches: 495903 -> 495941 (+0.01%); split: -0.00%, +0.01%
PreSGPRs: 867295 -> 863664 (-0.42%)
PreVGPRs: 793480 -> 790549 (-0.37%)

Reviewed-by: Rhys Perry <pendingchaos02@gmail.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/22069>
 2023-03-30 03:34:34 +00:00
Georg Lehmann
3907c54443 aco: don't label mul with opsel as abs/neg 
Reviewed-by: Rhys Perry <pendingchaos02@gmail.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/22069>
 2023-03-30 03:34:34 +00:00
Georg Lehmann
a60b9313d3 aco: swap opsel when swapping VOP2/C operands 
Reviewed-by: Rhys Perry <pendingchaos02@gmail.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/22069>
 2023-03-30 03:34:34 +00:00
Georg Lehmann
82f7b3acfa aco: support neg(mul)/abs(mul) optimization in more cases 
Reviewed-by: Rhys Perry <pendingchaos02@gmail.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/22069>
 2023-03-30 03:34:34 +00:00
Georg Lehmann
9d841507e1 aco: support v_cvt_f32_f16 with opsel in combine_mad_mix 
Reviewed-by: Rhys Perry <pendingchaos02@gmail.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/22069>
 2023-03-30 03:34:34 +00:00
Georg Lehmann
9d6e223a7a aco: update match_op3_for_vop3 for VOP12C opsel 
Reviewed-by: Rhys Perry <pendingchaos02@gmail.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/22069>
 2023-03-30 03:34:34 +00:00
Georg Lehmann
0896ecec9a aco: handle opsel in combine_constant_comparison_ordering 
Reviewed-by: Rhys Perry <pendingchaos02@gmail.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/22069>
 2023-03-30 03:34:34 +00:00
Georg Lehmann
d8f07a0ddc aco: handle opsel in combine_ordering_test 
Reviewed-by: Rhys Perry <pendingchaos02@gmail.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/22069>
 2023-03-30 03:34:34 +00:00
Georg Lehmann
4db43415e5 aco: handle opsel in combine_comparison_ordering 
Reviewed-by: Rhys Perry <pendingchaos02@gmail.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/22069>
 2023-03-30 03:34:34 +00:00
Georg Lehmann
8e6d79d10d aco/optimizer: preserve opsel when fusing fma 
Reviewed-by: Rhys Perry <pendingchaos02@gmail.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/22069>
 2023-03-30 03:34:34 +00:00
Georg Lehmann
fd3ea4ffc2 aco: clean up to_mad_mix 
These instructions are 32bit, so they don't support opsel anyway.

Reviewed-by: Rhys Perry <pendingchaos02@gmail.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/22103>
 2023-03-28 23:30:08 +00:00
Georg Lehmann
b1668aedaf aco: don't check usesModifiers for pseudo instructions 
This can't happen.

Reviewed-by: Rhys Perry <pendingchaos02@gmail.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/22106>
 2023-03-27 14:22:07 +00:00
Timur Kristóf
b3933ffe60 aco: Don't add soffset to swizzled MUBUF base. 
No Fossil DB changes on Rembrandt (GFX10.3).

Signed-off-by: Timur Kristóf <timur.kristof@gmail.com>
Reviewed-by: Rhys Perry <pendingchaos02@gmail.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/21930>
 2023-03-17 00:34:20 +00:00
Georg Lehmann
13ff4a5f64 aco: use bitfield_array for temporary neg/abs/opsel 
Reviewed-by: Rhys Perry <pendingchaos02@gmail.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/21766>
 2023-03-09 14:15:14 +00:00
Georg Lehmann
828aff2a2d aco: use array indexing for opsel/opsel_lo/opsel_hi 
Reviewed-by: Rhys Perry <pendingchaos02@gmail.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/21766>
 2023-03-09 14:15:13 +00:00
Georg Lehmann
a47c3f84fb aco: use integer access for neg_lo/neg_hi 
Reviewed-by: Rhys Perry <pendingchaos02@gmail.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/21766>
 2023-03-09 14:15:13 +00:00
Georg Lehmann
60cd3ba39f aco: copy abs/neg with assignment 
Reviewed-by: Rhys Perry <pendingchaos02@gmail.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/21766>
 2023-03-09 14:15:13 +00:00
Georg Lehmann
0614c2e8bd aco: don't reallocate fma{mk,ak,_mix} instruction 
Reviewed-by: Daniel Schürmann <daniel@schuermann.dev>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/21762>
 2023-03-08 18:42:21 +00:00
Georg Lehmann
a4873071e6 aco/optimizer: don't reallocate instruction when converting to VOP3 
Reviewed-by: Daniel Schürmann <daniel@schuermann.dev>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/21762>
 2023-03-08 18:42:21 +00:00
Georg Lehmann
de4805f25f aco: use bitfield array helpers for valu modifiers 
Reviewed-by: Daniel Schürmann <daniel@schuermann.dev>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/21023>
 2023-03-07 11:53:23 +00:00
Georg Lehmann
097a97cc42 aco: remove VOP[123C]P? structs 
Reviewed-by: Daniel Schürmann <daniel@schuermann.dev>
Reviewed-by: Timur Kristóf <timur.kristof@gmail.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/21023>
 2023-03-07 11:53:23 +00:00
Georg Lehmann
08542318e7 aco/optimizer: simplify using VALU instruction 
Reviewed-by: Daniel Schürmann <daniel@schuermann.dev>
Reviewed-by: Timur Kristóf <timur.kristof@gmail.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/21023>
 2023-03-07 11:53:23 +00:00
Georg Lehmann
ede0630f9e aco: use v_fma_mix_f32 for v_fma_f32 with 2 fp16 representable, different literals 
We can pack two fp16 literals into one 32bit literal and use opsel to select
the correct value. Note that LLVM currently disassembles these instructions
incorrectly.

Foz-DB Navi21:
Totals from 13365 (9.91% of 134913) affected shaders:
VGPRs: 840880 -> 840016 (-0.10%); split: -0.11%, +0.01%
SpillSGPRs: 724 -> 722 (-0.28%)
CodeSize: 82439364 -> 82451336 (+0.01%); split: -0.06%, +0.08%
MaxWaves: 244858 -> 244980 (+0.05%)
Instrs: 15265976 -> 15247201 (-0.12%); split: -0.13%, +0.01%
Latency: 223316180 -> 223272495 (-0.02%); split: -0.03%, +0.02%
InvThroughput: 41981375 -> 41969917 (-0.03%); split: -0.04%, +0.01%
VClause: 266775 -> 266558 (-0.08%); split: -0.14%, +0.06%
SClause: 646602 -> 645996 (-0.09%); split: -0.16%, +0.07%
Copies: 794703 -> 776075 (-2.34%); split: -2.46%, +0.12%
Branches: 296317 -> 296316 (-0.00%)
PreSGPRs: 658796 -> 656479 (-0.35%); split: -0.35%, +0.00%
PreVGPRs: 744014 -> 743679 (-0.05%)

Reviewed-by: Rhys Perry <pendingchaos02@gmail.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/20587>
 2023-03-02 10:59:05 +00:00
Georg Lehmann
ed349951cb aco: mark mad definition as precise if the mul/add were precise 
Reviewed-by: Rhys Perry <pendingchaos02@gmail.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/20587>
 2023-03-02 10:59:05 +00:00