-
Rhys Perry authored
Now that we try to create ffma from the beginning and are less reliant on inexact optimizations, we can enable this with much less regressions on GFX10.3. fossil-db (GFX10.3): Totals from 53274 (36.42% of 146267) affected shaders: VGPRs: 1979216 -> 1975944 (-0.17%); split: -0.67%, +0.50% SpillSGPRs: 1462 -> 1446 (-1.09%) CodeSize: 77389256 -> 79231200 (+2.38%); split: -0.06%, +2.44% MaxWaves: 1418040 -> 1419400 (+0.10%); split: +0.34%, -0.24% Instrs: 15015014 -> 15453979 (+2.92%); split: -0.14%, +3.06% Latency: 89229974 -> 89562172 (+0.37%); split: -0.76%, +1.13% InvThroughput: 15488878 -> 15990928 (+3.24%); split: -0.25%, +3.50% fossil-db (GFX10): Totals from 12563 (8.59% of 146267) affected shaders: VGPRs: 532880 -> 532200 (-0.13%); split: -0.42%, +0.29% SpillSGPRs: 1462 -> 1446 (-1.09%) CodeSize: 30125368 -> 30275508 (+0.50%); split: -0.06%, +0.56% MaxWaves: 327120 -> 327484 (+0.11%); split: +0.31%, -0.20% Instrs: 5783753 -> 5834581 (+0.88%); split: -0.02%, +0.89% Latency: 29271625 -> 29349407 (+0.27%); split: -0.14%, +0.41% InvThroughput: 6801730 -> 6855695 (+0.79%); split: -0.04%, +0.83% fossil-db (GFX9): Totals from 12563 (8.58% of 146401) affected shaders: SGPRs: 717840 -> 717888 (+0.01%); split: -0.16%, +0.16% VGPRs: 531464 -> 531892 (+0.08%); split: -0.23%, +0.31% SpillSGPRs: 1770 -> 1750 (-1.13%) CodeSize: 27658752 -> 27814028 (+0.56%); split: -0.05%, +0.61% MaxWaves: 81149 -> 81224 (+0.09%); split: +0.24%, -0.14% Instrs: 5489024 -> 5537939 (+0.89%); split: -0.02%, +0.91% Latency: 43480815 -> 43657720 (+0.41%); split: -0.08%, +0.49% InvThroughput: 19853380 -> 20017416 (+0.83%); split: -0.04%, +0.87% Signed-off-by: Rhys Perry <pendingchaos02@gmail.com>
15211031