perf: optimize SIMD loops with unsafe indexing#825
Draft
d-enk wants to merge 1 commit into
Draft
Conversation
Contributor
Author
│ old.txt │ new.txt │
│ sec/op │ sec/op vs base │
SumInt8/small/AVX-x16-4 11.74n ± 2% 13.22n ± 4% +12.60% (p=0.000 n=8)
SumInt8/small/AVX2-x32-4 19.36n ± 6% 17.55n ± 4% -9.35% (p=0.001 n=8)
SumInt8/medium/AVX-x16-4 13.26n ± 21% 12.47n ± 7% -5.99% (p=0.015 n=8)
SumInt8/medium/AVX2-x32-4 17.45n ± 5% 28.67n ± 2% +64.28% (p=0.000 n=8)
SumInt8/large/AVX-x16-4 48.85n ± 2% 45.22n ± 3% -7.44% (p=0.003 n=8)
SumInt8/large/AVX2-x32-4 33.16n ± 3% 26.35n ± 3% -20.52% (p=0.000 n=8)
SumInt8/xlarge/AVX-x16-4 305.8n ± 12% 256.2n ± 15% -16.22% (p=0.010 n=8)
SumInt8/xlarge/AVX2-x32-4 281.8n ± 28% 143.2n ± 11% -49.21% (p=0.000 n=8)
SumInt16/small/AVX-x8-4 7.346n ± 22% 5.850n ± 3% -20.37% (p=0.000 n=8)
SumInt16/small/AVX2-x16-4 13.84n ± 6% 14.15n ± 15% ~ (p=0.505 n=8)
SumInt16/medium/AVX-x8-4 15.35n ± 6% 12.58n ± 12% -17.99% (p=0.000 n=8)
SumInt16/medium/AVX2-x16-4 12.75n ± 4% 11.15n ± 7% -12.55% (p=0.006 n=8)
SumInt16/large/AVX-x8-4 89.76n ± 5% 72.66n ± 8% -19.06% (p=0.000 n=8)
SumInt16/large/AVX2-x16-4 58.96n ± 16% 48.37n ± 7% -17.97% (p=0.000 n=8)
SumInt16/xlarge/AVX-x8-4 612.5n ± 5% 474.8n ± 14% -22.47% (p=0.000 n=8)
SumInt16/xlarge/AVX2-x16-4 325.7n ± 24% 250.5n ± 5% -23.10% (p=0.000 n=8)
SumInt32/small/AVX-x4-4 7.254n ± 2% 6.150n ± 2% -15.22% (p=0.000 n=8)
SumInt32/small/AVX2-x8-4 6.651n ± 7% 6.084n ± 2% -8.53% (p=0.010 n=8)
SumInt32/medium/AVX-x4-4 30.08n ± 5% 24.48n ± 3% -18.59% (p=0.000 n=8)
SumInt32/medium/AVX2-x8-4 14.55n ± 10% 12.15n ± 4% -16.47% (p=0.000 n=8)
SumInt32/large/AVX-x4-4 171.8n ± 5% 124.6n ± 7% -27.44% (p=0.000 n=8)
SumInt32/large/AVX2-x8-4 95.56n ± 9% 71.80n ± 5% -24.87% (p=0.000 n=8)
SumInt32/xlarge/AVX-x4-4 1302.5n ± 11% 952.8n ± 6% -26.85% (p=0.000 n=8)
SumInt32/xlarge/AVX2-x8-4 631.6n ± 9% 463.2n ± 1% -26.65% (p=0.000 n=8)
SumInt64/small/AVX-x2-4 9.235n ± 14% 7.116n ± 2% -22.95% (p=0.000 n=8)
SumInt64/small/AVX2-x4-4 6.291n ± 1% 5.868n ± 2% -6.73% (p=0.001 n=8)
SumInt64/medium/AVX-x2-4 47.29n ± 13% 30.71n ± 2% -35.07% (p=0.000 n=8)
SumInt64/medium/AVX2-x4-4 23.45n ± 11% 26.09n ± 5% +11.21% (p=0.015 n=8)
SumInt64/large/AVX-x2-4 319.3n ± 2% 236.9n ± 3% -25.79% (p=0.000 n=8)
SumInt64/large/AVX2-x4-4 176.2n ± 7% 127.9n ± 3% -27.46% (p=0.000 n=8)
SumInt64/xlarge/AVX-x2-4 2.726µ ± 8% 1.981µ ± 6% -27.33% (p=0.000 n=8)
SumInt64/xlarge/AVX2-x4-4 1.682µ ± 19% 1.082µ ± 2% -35.68% (p=0.000 n=8)
SumFloat32/small/AVX-x4-4 10.42n ± 27% 10.12n ± 4% ~ (p=0.645 n=8)
SumFloat32/small/AVX2-x8-4 11.475n ± 13% 9.527n ± 7% -16.97% (p=0.001 n=8)
SumFloat32/medium/AVX-x4-4 42.24n ± 24% 45.67n ± 2% +8.12% (p=0.002 n=8)
SumFloat32/medium/AVX2-x8-4 20.09n ± 5% 16.02n ± 1% -20.28% (p=0.000 n=8)
SumFloat32/large/AVX-x4-4 300.3n ± 6% 296.3n ± 5% ~ (p=0.938 n=8)
SumFloat32/large/AVX2-x8-4 139.3n ± 2% 127.8n ± 3% -8.26% (p=0.000 n=8)
SumFloat32/xlarge/AVX-x4-4 2.369µ ± 2% 2.354µ ± 1% ~ (p=0.721 n=8)
SumFloat32/xlarge/AVX2-x8-4 1.241µ ± 5% 1.154µ ± 6% -7.01% (p=0.007 n=8)
SumFloat64/small/AVX-x2-4 7.584n ± 27% 9.575n ± 4% +26.25% (p=0.050 n=8)
SumFloat64/small/AVX2-x4-4 7.571n ± 2% 5.891n ± 2% -22.18% (p=0.000 n=8)
SumFloat64/medium/AVX-x2-4 78.12n ± 28% 78.62n ± 6% ~ (p=0.574 n=8)
SumFloat64/medium/AVX2-x4-4 31.53n ± 6% 31.18n ± 4% ~ (p=1.000 n=8)
SumFloat64/large/AVX-x2-4 646.2n ± 44% 595.2n ± 3% -7.88% (p=0.003 n=8)
SumFloat64/large/AVX2-x4-4 309.2n ± 7% 267.3n ± 1% -13.54% (p=0.000 n=8)
SumFloat64/xlarge/AVX-x2-4 6.667µ ± 10% 4.859µ ± 7% -27.12% (p=0.000 n=8)
SumFloat64/xlarge/AVX2-x4-4 3.209µ ± 9% 2.372µ ± 2% -26.09% (p=0.000 n=8)
MeanInt32/small/AVX-x4-4 13.050n ± 10% 8.471n ± 1% -35.09% (p=0.000 n=8)
MeanInt32/small/AVX2-x8-4 14.51n ± 13% 10.42n ± 3% -28.22% (p=0.000 n=8)
MeanInt32/medium/AVX-x4-4 36.90n ± 19% 20.81n ± 1% -43.61% (p=0.000 n=8)
MeanInt32/medium/AVX2-x8-4 23.25n ± 11% 15.25n ± 3% -34.42% (p=0.000 n=8)
MeanInt32/large/AVX-x4-4 178.7n ± 5% 133.1n ± 1% -25.49% (p=0.000 n=8)
MeanInt32/large/AVX2-x8-4 109.90n ± 28% 79.89n ± 5% -27.30% (p=0.000 n=8)
MeanInt32/xlarge/AVX-x4-4 1.671µ ± 17% 1.010µ ± 8% -39.55% (p=0.000 n=8)
MeanInt32/xlarge/AVX2-x8-4 742.0n ± 16% 515.4n ± 4% -30.54% (p=0.000 n=8)
MeanFloat64/small/AVX-x2-4 14.12n ± 20% 10.26n ± 23% -27.28% (p=0.001 n=8)
MeanFloat64/small/AVX2-x4-4 11.220n ± 26% 9.695n ± 3% -13.59% (p=0.000 n=8)
MeanFloat64/medium/AVX-x2-4 85.54n ± 25% 65.62n ± 18% -23.29% (p=0.038 n=8)
MeanFloat64/medium/AVX2-x4-4 34.88n ± 7% 30.19n ± 7% -13.45% (p=0.001 n=8)
MeanFloat64/large/AVX-x2-4 665.3n ± 3% 555.7n ± 1% -16.47% (p=0.000 n=8)
MeanFloat64/large/AVX2-x4-4 327.1n ± 7% 283.4n ± 3% -13.34% (p=0.000 n=8)
MeanFloat64/xlarge/AVX-x2-4 5.418µ ± 23% 4.731µ ± 3% -12.68% (p=0.003 n=8)
MeanFloat64/xlarge/AVX2-x4-4 2.870µ ± 17% 2.542µ ± 3% -11.43% (p=0.001 n=8)
MinInt32/small/AVX-x4-4 5.732n ± 3% 4.532n ± 1% -20.93% (p=0.000 n=8)
MinInt32/small/AVX2-x8-4 5.675n ± 9% 4.627n ± 1% -18.48% (p=0.000 n=8)
MinInt32/medium/AVX-x4-4 33.31n ± 2% 27.84n ± 4% -16.42% (p=0.000 n=8)
MinInt32/medium/AVX2-x8-4 19.57n ± 30% 16.53n ± 3% -15.56% (p=0.000 n=8)
MinInt32/large/AVX-x4-4 254.6n ± 10% 216.3n ± 3% -15.01% (p=0.000 n=8)
MinInt32/large/AVX2-x8-4 139.2n ± 10% 110.0n ± 1% -21.01% (p=0.000 n=8)
MinInt32/xlarge/AVX-x4-4 2.245µ ± 13% 1.648µ ± 4% -26.61% (p=0.000 n=8)
MinInt32/xlarge/AVX2-x8-4 1217.0n ± 12% 847.6n ± 5% -30.35% (p=0.000 n=8)
MinFloat64/small/AVX-x2-4 11.92n ± 32% 11.66n ± 2% ~ (p=0.627 n=8)
MinFloat64/small/AVX2-x4-4 8.325n ± 19% 5.962n ± 2% -28.39% (p=0.000 n=8)
MinFloat64/medium/AVX-x2-4 150.4n ± 24% 120.2n ± 1% -20.08% (p=0.015 n=8)
MinFloat64/medium/AVX2-x4-4 47.40n ± 14% 35.02n ± 4% -26.10% (p=0.000 n=8)
MinFloat64/large/AVX-x2-4 931.6n ± 37% 928.8n ± 2% ~ (p=1.000 n=8)
MinFloat64/large/AVX2-x4-4 384.0n ± 12% 286.9n ± 9% -25.28% (p=0.000 n=8)
MinFloat64/xlarge/AVX-x2-4 8.061µ ± 15% 7.908µ ± 5% ~ (p=0.645 n=8)
MinFloat64/xlarge/AVX2-x4-4 3.059µ ± 8% 2.347µ ± 0% -23.29% (p=0.000 n=8)
MaxInt32/small/AVX-x4-4 5.939n ± 7% 5.189n ± 4% -12.64% (p=0.000 n=8)
MaxInt32/small/AVX2-x8-4 6.117n ± 11% 4.543n ± 1% -25.72% (p=0.000 n=8)
MaxInt32/medium/AVX-x4-4 34.40n ± 7% 28.20n ± 4% -18.02% (p=0.000 n=8)
MaxInt32/medium/AVX2-x8-4 20.04n ± 9% 16.86n ± 4% -15.87% (p=0.000 n=8)
MaxInt32/large/AVX-x4-4 256.7n ± 17% 208.7n ± 1% -18.70% (p=0.000 n=8)
MaxInt32/large/AVX2-x8-4 154.8n ± 37% 111.1n ± 3% -28.21% (p=0.000 n=8)
MaxInt32/xlarge/AVX-x4-4 2.323µ ± 18% 1.769µ ± 7% -23.85% (p=0.000 n=8)
MaxInt32/xlarge/AVX2-x8-4 969.5n ± 7% 811.1n ± 4% -16.33% (p=0.000 n=8)
MaxFloat64/small/AVX-x2-4 10.46n ± 33% 12.77n ± 3% +22.07% (p=0.000 n=8)
MaxFloat64/small/AVX2-x4-4 12.27n ± 3% 12.16n ± 2% ~ (p=0.978 n=8)
MaxFloat64/medium/AVX-x2-4 130.6n ± 18% 120.8n ± 5% ~ (p=0.083 n=8)
MaxFloat64/medium/AVX2-x4-4 45.92n ± 7% 38.05n ± 4% -17.16% (p=0.000 n=8)
MaxFloat64/large/AVX-x2-4 1003.1n ± 21% 926.0n ± 1% ~ (p=0.130 n=8)
MaxFloat64/large/AVX2-x4-4 325.9n ± 7% 299.4n ± 2% -8.15% (p=0.000 n=8)
MaxFloat64/xlarge/AVX-x2-4 8.442µ ± 15% 7.522µ ± 1% ~ (p=0.105 n=8)
MaxFloat64/xlarge/AVX2-x4-4 2.527µ ± 3% 2.490µ ± 5% ~ (p=0.664 n=8)
SumInt8ByWidth/AVX-x16-4 150.8n ± 10% 126.7n ± 1% -16.01% (p=0.000 n=8)
SumInt8ByWidth/AVX2-x32-4 97.53n ± 3% 78.44n ± 4% -19.58% (p=0.000 n=8)
SumInt64SteadyState/AVX-x2-4 3.049µ ± 9% 1.901µ ± 2% -37.67% (p=0.000 n=8)
SumInt64SteadyState/AVX2-x4-4 1.441µ ± 9% 1.062µ ± 6% -26.24% (p=0.000 n=8)
geomean 120.3n 102.7n -14.60% |
d-enk
force-pushed
the
perf-simd-no-bounds-check
branch
2 times, most recently
from
d0b5c52 to
c987df8
Compare
Contributor
Author
|
For some reason, there's degradation for these. I couldn't figure out why. Some kind of tricky caching... |
Replace slice-based access (base[i:i+lanes]) with unsafeIndexVec to eliminate runtime.panicBounds checks in critical SIMD loops. Changes: - Replace unsafeSliceInt8/16/32/64 and unsafeSliceUint* with unsafeIndexVec - Add unsafeIndexBase and unsafeIndexOffset for optimal performance - Update Sum*, Mean*, Min*, Max*, Contains*, Clamp* functions - Remove unused unsafeSlice* helper functions Generated code size: ~96 bytes / ~12 instructions less per function Better cache locality without slice header overhead
d-enk
force-pushed
the
perf-simd-no-bounds-check
branch
from
c987df8 to
babf7ad
Compare
Codecov Report✅ All modified and coverable lines are covered by tests. Additional details and impacted files@@ Coverage Diff @@
## master #825 +/- ##
==========================================
+ Coverage 92.11% 92.48% +0.37%
==========================================
Files 32 32
Lines 4259 5019 +760
==========================================
+ Hits 3923 4642 +719
- Misses 252 274 +22
- Partials 84 103 +19
Flags with carried forward coverage won't be shown. Click here to find out more. ☔ View full report in Codecov by Sentry. 🚀 New features to boost your workflow:
|
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Replace slice-based access (base[i:i+lanes]) with unsafeIndexVec to eliminate runtime.panicBounds checks in critical SIMD loops.
Changes:
Generated code size: ~96 bytes / ~12 instructions less per function
Better cache locality without slice header overhead