On Tue, 3 Nov 2026 18:51:13 GMT, Vladimir Kozlov <[email protected]> wrote:
>> Hi @aprilivy, here are the latest changes results on Intel Xeon 6980P >> (Granite Rapids) >> >> #### default flags >> Size | baseline | 563c276 | 3c2d1f | baseline -> 563c276 | baseline -> >> 3c2d1f | 563c276-> 3c2d1f >> -- | -- | -- | -- | -- | -- | -- >> 1 | 104174.5 | 166971.3 | 166521.8 | +60.28% | +59.85% | -0.27% >> 2 | 101667.1 | 159447.0 | 161647.5 | +56.83% | +59.00% | +1.38% >> 4 | 94600.9 | 145313.9 | 149690.5 | +53.61% | +58.23% | +3.01% >> 8 | 84098.3 | 122873.1 | 122949.7 | +46.11% | +46.20% | +0.06% >> 15 | 69281.9 | 91600.4 | 93480.1 | +32.21% | +34.93% | +2.05% >> 16 | 66885.7 | 161567.4 | 161586.9 | +141.56% | +141.59% | +0.01% >> 24 | 56272.5 | 111253.1 | 108333.6 | +97.70% | +92.52% | -2.62% >> 31 | 50305.8 | 84461.0 | 83596.0 | +67.90% | +66.18% | -1.02% >> 32 | 54057.7 | 71378.4 | 73284.2 | +32.04% | +35.57% | +2.67% >> 64 | 50015.4 | 70696.5 | 69268.8 | +41.35% | +38.50% | -2.02% >> 128 | 44439.3 | 67154.2 | 65542.9 | +51.11% | +47.49% | -2.40% >> 256 | 33044.8 | 32366.4 | 32388.8 | -2.05% | -1.99% | +0.07% >> 512 | 22974.7 | 23041.7 | 23051.1 | +0.29% | +0.33% | +0.04% >> 1024 | 15342.5 | 15654.1 | 15644.6 | +2.03% | +1.97% | -0.06% >> 2048 | 9214.4 | 9245.4 | 9287.7 | +0.34% | +0.80% | +0.46% >> 5012 | 4006.5 | 4072.2 | 4090.7 | +1.64% | +2.10% | +0.45% >> 8192 | 2451.9 | 2537.7 | 2513.3 | +3.50% | +2.51% | -0.96% >> 16384 | 1264.5 | 1307.4 | 1301.5 | +3.39% | +2.93% | -0.45% >> 32768 | 640.9 | 664.3 | 661.5 | +3.65% | +3.22% | -0.42% >> 65536 | 325.9 | 337.3 | 335.9 | +3.51% | +3.07% | -0.43% >> >> #### CopyAVX3Threshold=8192 >> >> Size | baseline | 563c276 | 3c2d1f | baseline -> 563c276 | baseline -> >> 3c2d1f | 563c276-> 3c2d1f >> -- | -- | -- | -- | -- | -- | -- >> 1 | 104297.1 | 167000.5 | 166537.6 | +60.12% | +59.68% | -0.28% >> 2 | 101604.8 | 159411.4 | 162135.8 | +56.89% | +59.57% | +1.71% >> 4 | 94604.6 | 145315.8 | 149903.8 | +53.60% | +58.45% | +3.16% >> 8 | 84103.2 | 122871.6 | 123033.9 | +46.10% | +46.29% | +0.13% >> 15 | 69272.8 | 91502.3 | 93531.9 | +32.09% | +35.02% | +2.22% >> 16 | 66867.8 | 161560.4 | 161574.3 | +141.61% | +141.63% | +0.01% >> 24 | 56272.4 | 111257.8 | 108339.8 | +97.71% | +92.53% | -2.62% >> 31 | 50339.5 | 84471.0 | 83601.5 | +67.80% | +66.08% | -1.03% >> 32 | 54059.7 | 71390.6 | 73286.0 | +32.06% | +35.57% | +2.66% >> 64 | 50013.4 | 70701.1 | 69264.2 | +41.36% | +38.49% | -2.03% >> 128 | 44439.2 | 67165.6 | 65547.5 | +51.14% | +47.50% | -2.41% >> 256 | 30051.0 | 59941.4 | 58363.0 | +99.47% | +94.21% | -2.63% >> 512 | 22989.3 | 45839.8 | ... > > @LangruiZhou I will review and test it later. @vnkozlov @jatin-bhateja @vpaprotsk @aprilivy Thanks again for all your reviews and suggestions. Sorry for the slight detour, while analyzing SPECjvm performance, I found that DESCrypt.cipherBlock and the DESede block encrypt/decrypt paths do not appear to have corresponding HotSpot intrinsics. I built x86-64 C2 intrinsic prototypes for both paths and saw encouraging gains. I’d like to move this work forward as a separate change. Would any of you be willing to help get two Enhancement JBS issues filed, one for DES and one for DESede, or point me to the right person to ask? I can put together a short scope note with the benchmark setup, results, and test plan first, so it’s easier to judge whether this is worth pursuing. Thanks! ------------- PR Comment: https://git.openjdk.org/jdk/pull/33338#issuecomment-6654305548
