Reduce memory allocations by reusing pixel buffers #32
Conversation
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
```
cpu: Apple M1 Pro
│ master.bench │ feat-perf.bench │
│ sec/op │ sec/op vs base │
Filter/Resize_Lanczos-10 4.578m ± 1% 4.583m ± 2% ~ (p=0.818 n=6)
Filter/Resize_Cubic-10 3.671m ± 0% 3.687m ± 1% ~ (p=0.240 n=6)
Filter/Resize_Linear-10 2.934m ± 1% 3.005m ± 2% +2.44% (p=0.004 n=6)
Filter/Resize_Box-10 2.665m ± 1% 2.758m ± 2% +3.48% (p=0.002 n=6)
Filter/Resize_Nearest-10 79.28µ ± 0% 79.13µ ± 1% ~ (p=0.240 n=6)
Filter/Crop-10 101.1µ ± 0% 101.5µ ± 0% +0.37% (p=0.002 n=6)
Filter/CropToSize-10 91.19µ ± 1% 91.96µ ± 1% +0.85% (p=0.002 n=6)
Filter/FlipHorizontal-10 3.104m ± 1% 3.123m ± 0% ~ (p=0.065 n=6)
Filter/FlipVertical-10 3.213m ± 1% 3.216m ± 0% ~ (p=0.699 n=6)
Filter/Transpose-10 3.202m ± 1% 3.202m ± 1% ~ (p=0.818 n=6)
Filter/Transverse-10 3.294m ± 1% 3.283m ± 3% ~ (p=0.240 n=6)
Filter/Rotate90-10 3.117m ± 1% 3.109m ± 1% ~ (p=0.310 n=6)
Filter/Rotate180-10 3.096m ± 1% 3.096m ± 3% ~ (p=0.818 n=6)
Filter/Rotate270-10 3.215m ± 1% 3.205m ± 0% ~ (p=0.240 n=6)
Filter/Rotate-10 45.00m ± 3% 44.89m ± 2% ~ (p=0.937 n=6)
Filter/Brightness-10 3.894m ± 1% 3.819m ± 1% -1.93% (p=0.002 n=6)
Filter/Contrast-10 4.415m ± 1% 4.417m ± 1% ~ (p=0.937 n=6)
Filter/Saturation-10 5.698m ± 1% 5.680m ± 3% ~ (p=0.310 n=6)
Filter/Gamma-10 3.413m ± 0% 3.406m ± 0% ~ (p=0.394 n=6)
Filter/GaussianBlur-10 14.06m ± 2% 14.03m ± 1% ~ (p=0.485 n=6)
Filter/UnsharpMask-10 23.32m ± 1% 23.24m ± 2% ~ (p=0.394 n=6)
Filter/Sigmoid-10 3.457m ± 1% 3.453m ± 1% ~ (p=0.818 n=6)
Filter/Pixelate-10 3.516m ± 1% 3.504m ± 1% -0.34% (p=0.041 n=6)
Filter/Colorize-10 5.793m ± 1% 5.814m ± 1% ~ (p=0.485 n=6)
Filter/ColorBalance-10 3.209m ± 2% 3.207m ± 0% ~ (p=0.589 n=6)
Filter/Threshold-10 3.334m ± 0% 3.343m ± 0% ~ (p=0.093 n=6)
Filter/Hue-10 6.423m ± 2% 6.423m ± 1% ~ (p=0.699 n=6)
Filter/Grayscale-10 3.176m ± 1% 3.178m ± 1% ~ (p=0.818 n=6)
Filter/Sepia-10 3.290m ± 1% 3.290m ± 4% ~ (p=0.818 n=6)
Filter/Invert-10 3.810m ± 0% 3.726m ± 1% -2.21% (p=0.002 n=6)
Filter/ColorFunc-10 3.339m ± 0% 3.345m ± 1% ~ (p=0.589 n=6)
Filter/ColorspaceSRGBToLinear-10 3.421m ± 1% 3.422m ± 1% ~ (p=0.699 n=6)
Filter/ColorspaceLinearToSRGB-10 3.424m ± 1% 3.422m ± 1% ~ (p=0.394 n=6)
Filter/Mean-10 9.886m ± 2% 9.745m ± 1% ~ (p=0.065 n=6)
Filter/Median-10 112.8m ± 1% 112.4m ± 4% ~ (p=0.818 n=6)
Filter/Minimum-10 37.59m ± 6% 43.18m ± 13% +14.86% (p=0.002 n=6)
Filter/Maximum-10 37.53m ± 2% 45.63m ± 7% +21.60% (p=0.002 n=6)
Filter/Convolution-10 7.051m ± 1% 6.991m ± 1% -0.85% (p=0.041 n=6)
Filter/Sobel-10 22.40m ± 1% 22.56m ± 1% ~ (p=0.394 n=6)
geomean 4.239m 4.275m +0.86%
│ master.bench │ feat-perf.bench │
│ B/op │ B/op vs base │
Filter/Resize_Lanczos-10 1.832Mi ± 0% 1.401Mi ± 0% -23.52% (p=0.002 n=6)
Filter/Resize_Cubic-10 1.754Mi ± 0% 1.364Mi ± 0% -22.22% (p=0.002 n=6)
Filter/Resize_Linear-10 1.692Mi ± 0% 1.295Mi ± 0% -23.44% (p=0.002 n=6)
Filter/Resize_Box-10 1.647Mi ± 0% 1.260Mi ± 1% -23.51% (p=0.002 n=6)
Filter/Resize_Nearest-10 896.0 ± 0% 896.0 ± 0% ~ (p=1.000 n=6) ¹
Filter/Crop-10 880.0 ± 0% 880.0 ± 0% ~ (p=1.000 n=6) ¹
Filter/CropToSize-10 880.0 ± 0% 880.0 ± 0% ~ (p=1.000 n=6) ¹
Filter/FlipHorizontal-10 880.0 ± 0% 880.0 ± 0% ~ (p=1.000 n=6) ¹
Filter/FlipVertical-10 880.0 ± 0% 880.0 ± 0% ~ (p=1.000 n=6) ¹
Filter/Transpose-10 880.0 ± 0% 880.0 ± 0% ~ (p=1.000 n=6) ¹
Filter/Transverse-10 880.0 ± 0% 880.0 ± 0% ~ (p=1.000 n=6) ¹
Filter/Rotate90-10 880.0 ± 0% 880.0 ± 0% ~ (p=1.000 n=6) ¹
Filter/Rotate180-10 880.0 ± 0% 880.0 ± 0% ~ (p=1.000 n=6) ¹
Filter/Rotate270-10 880.0 ± 0% 880.0 ± 0% ~ (p=1.000 n=6) ¹
Filter/Rotate-10 928.0 ± 0% 928.0 ± 0% ~ (p=1.000 n=6)
Filter/Brightness-10 912.0 ± 0% 912.0 ± 2% ~ (p=1.000 n=6)
Filter/Contrast-10 912.0 ± 0% 912.0 ± 0% ~ (p=1.000 n=6) ¹
Filter/Saturation-10 880.0 ± 0% 880.0 ± 0% ~ (p=1.000 n=6) ¹
Filter/Gamma-10 1.891Ki ± 0% 1.891Ki ± 0% ~ (p=1.000 n=6) ¹
Filter/GaussianBlur-10 12.17Mi ± 0% 11.51Mi ± 0% -5.38% (p=0.002 n=6)
Filter/UnsharpMask-10 23.55Mi ± 0% 22.84Mi ± 0% -3.02% (p=0.002 n=6)
Filter/Sigmoid-10 1.891Ki ± 0% 1.891Ki ± 0% ~ (p=1.000 n=6) ¹
Filter/Pixelate-10 896.0 ± 0% 896.0 ± 0% ~ (p=1.000 n=6) ¹
Filter/Colorize-10 880.0 ± 0% 880.0 ± 0% ~ (p=1.000 n=6) ¹
Filter/ColorBalance-10 880.0 ± 0% 880.0 ± 0% ~ (p=1.000 n=6) ¹
Filter/Threshold-10 880.0 ± 0% 880.0 ± 0% ~ (p=1.000 n=6) ¹
Filter/Hue-10 880.0 ± 0% 880.0 ± 0% ~ (p=1.000 n=6)
Filter/Grayscale-10 880.0 ± 0% 880.0 ± 0% ~ (p=1.000 n=6) ¹
Filter/Sepia-10 880.0 ± 0% 880.0 ± 0% ~ (p=1.000 n=6) ¹
Filter/Invert-10 912.0 ± 0% 912.0 ± 0% ~ (p=1.000 n=6) ¹
Filter/ColorFunc-10 880.0 ± 0% 880.0 ± 0% ~ (p=1.000 n=6) ¹
Filter/ColorspaceSRGBToLinear-10 1.891Ki ± 0% 1.891Ki ± 0% ~ (p=1.000 n=6) ¹
Filter/ColorspaceLinearToSRGB-10 1.891Ki ± 0% 1.891Ki ± 0% ~ (p=1.000 n=6) ¹
Filter/Mean-10 1.175Mi ± 0% 1.175Mi ± 0% ~ (p=0.455 n=6)
Filter/Median-10 13.672Ki ± 0% 5.562Ki ± 4% -59.31% (p=0.002 n=6)
Filter/Minimum-10 10.391Ki ± 0% 1.503Ki ± 3% -85.54% (p=0.002 n=6)
Filter/Maximum-10 10.391Ki ± 0% 1.470Ki ± 4% -85.85% (p=0.002 n=6)
Filter/Convolution-10 722.0Ki ± 0% 722.0Ki ± 0% ~ (p=0.987 n=6)
Filter/Sobel-10 24.18Mi ± 0% 24.18Mi ± 0% ~ (p=0.937 n=6)
geomean 7.735Ki 6.644Ki -14.10%
¹ all samples are equal
│ master.bench │ feat-perf.bench │
│ allocs/op │ allocs/op vs base │
Filter/Resize_Lanczos-10 74.00 ± 0% 35.00 ± 3% -52.70% (p=0.002 n=6)
Filter/Resize_Cubic-10 74.00 ± 0% 40.00 ± 3% -45.95% (p=0.002 n=6)
Filter/Resize_Linear-10 74.00 ± 0% 39.00 ± 3% -47.30% (p=0.002 n=6)
Filter/Resize_Box-10 74.00 ± 0% 40.00 ± 3% -45.95% (p=0.002 n=6)
Filter/Resize_Nearest-10 14.00 ± 0% 14.00 ± 0% ~ (p=1.000 n=6) ¹
Filter/Crop-10 14.00 ± 0% 14.00 ± 0% ~ (p=1.000 n=6) ¹
Filter/CropToSize-10 14.00 ± 0% 14.00 ± 0% ~ (p=1.000 n=6) ¹
Filter/FlipHorizontal-10 14.00 ± 0% 14.00 ± 0% ~ (p=1.000 n=6) ¹
Filter/FlipVertical-10 14.00 ± 0% 14.00 ± 0% ~ (p=1.000 n=6) ¹
Filter/Transpose-10 14.00 ± 0% 14.00 ± 0% ~ (p=1.000 n=6) ¹
Filter/Transverse-10 14.00 ± 0% 14.00 ± 0% ~ (p=1.000 n=6) ¹
Filter/Rotate90-10 14.00 ± 0% 14.00 ± 0% ~ (p=1.000 n=6) ¹
Filter/Rotate180-10 14.00 ± 0% 14.00 ± 0% ~ (p=1.000 n=6) ¹
Filter/Rotate270-10 14.00 ± 0% 14.00 ± 0% ~ (p=1.000 n=6) ¹
Filter/Rotate-10 14.00 ± 0% 14.00 ± 0% ~ (p=1.000 n=6) ¹
Filter/Brightness-10 14.00 ± 0% 14.00 ± 0% ~ (p=1.000 n=6) ¹
Filter/Contrast-10 14.00 ± 0% 14.00 ± 0% ~ (p=1.000 n=6) ¹
Filter/Saturation-10 14.00 ± 0% 14.00 ± 0% ~ (p=1.000 n=6) ¹
Filter/Gamma-10 15.00 ± 0% 15.00 ± 0% ~ (p=1.000 n=6) ¹
Filter/GaussianBlur-10 78.00 ± 0% 47.50 ± 3% -39.10% (p=0.002 n=6)
Filter/UnsharpMask-10 95.00 ± 0% 62.50 ± 6% -34.21% (p=0.002 n=6)
Filter/Sigmoid-10 15.00 ± 0% 15.00 ± 0% ~ (p=1.000 n=6) ¹
Filter/Pixelate-10 14.00 ± 0% 14.00 ± 0% ~ (p=1.000 n=6) ¹
Filter/Colorize-10 14.00 ± 0% 14.00 ± 0% ~ (p=1.000 n=6) ¹
Filter/ColorBalance-10 14.00 ± 0% 14.00 ± 0% ~ (p=1.000 n=6) ¹
Filter/Threshold-10 14.00 ± 0% 14.00 ± 0% ~ (p=1.000 n=6) ¹
Filter/Hue-10 14.00 ± 0% 14.00 ± 0% ~ (p=1.000 n=6) ¹
Filter/Grayscale-10 14.00 ± 0% 14.00 ± 0% ~ (p=1.000 n=6) ¹
Filter/Sepia-10 14.00 ± 0% 14.00 ± 0% ~ (p=1.000 n=6) ¹
Filter/Invert-10 14.00 ± 0% 14.00 ± 0% ~ (p=1.000 n=6) ¹
Filter/ColorFunc-10 14.00 ± 0% 14.00 ± 0% ~ (p=1.000 n=6) ¹
Filter/ColorspaceSRGBToLinear-10 15.00 ± 0% 15.00 ± 0% ~ (p=1.000 n=6) ¹
Filter/ColorspaceLinearToSRGB-10 15.00 ± 0% 15.00 ± 0% ~ (p=1.000 n=6) ¹
Filter/Mean-10 81.00 ± 0% 81.00 ± 0% ~ (p=1.000 n=6) ¹
Filter/Median-10 85.00 ± 0% 52.50 ± 3% -38.24% (p=0.002 n=6)
Filter/Minimum-10 55.00 ± 0% 18.00 ± 6% -67.27% (p=0.002 n=6)
Filter/Maximum-10 55.00 ± 0% 17.00 ± 6% -69.09% (p=0.002 n=6)
Filter/Convolution-10 58.00 ± 0% 58.00 ± 0% ~ (p=1.000 n=6) ¹
Filter/Sobel-10 137.0 ± 0% 137.0 ± 0% ~ (p=1.000 n=6) ¹
geomean 23.72 20.18 -14.93%
```
There are some lines above that actually runs slower in this, e.g. `Filter/Maximum`, but:
* Running those benchmarks in isolation gets the diff down to ~7%.
* And the allocation saving should far outweigh this in real world usage.
I also added a .gitignore to avoid committing temporary files again.
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.
There are some lines above that actually runs slower in this, e.g.
Filter/Maximum, but:I also added a .gitignore to avoid committing temporary files again.