feat: vectorize image processing with Java Vector API

[Copilot]

Summary

• vectorize image processing with the Java Vector API
• upgrade the GraalVM native build plugin and improve OS detection for native build arguments

Testing

• Not run in this session

[crramirez]

@copilot you added a new bundle, remember to add it also to the native jsons for its native image inclusion

[Copilot]

@copilot you added a new bundle, remember to add it also to the native jsons for its native image inclusion

Fixed in 33b0771.

[Copilot]

Pull request overview

This PR introduces Java Vector API–accelerated kernels for image processing in Casciian’s bits package, and wires in a demo UI to compare vectorized vs scalar baselines. It also updates build/native-image settings so the incubator Vector module is available during compilation, tests, and GraalVM native builds.

Changes:

• Vectorized core image kernels (alpha blending and RGB distance-sum) and updated call sites to use the new SIMD kernel.
• Added a demo “Vector performance” window + localized resource bundles and native-image resource configuration.
• Updated module/build/native-image configuration to include jdk.incubator.vector.

Reviewed changes

Copilot reviewed 16 out of 16 changed files in this pull request and generated 4 comments.

Show a summary per file

File	Description
code/src/main/resources/META-INF/native-image/io.crramirez/casciian-demo/resource-config.json	Registers the new demo resource bundle for native-image builds.
code/src/main/resources/demo/DemoVectorPerformanceWindowBundle.properties	Adds English strings for the new vector performance demo window.
code/src/main/resources/demo/DemoVectorPerformanceWindowBundle_es.properties	Adds Spanish strings for the new vector performance demo window.
code/src/main/resources/demo/DemoMainWindowBundle.properties	Adds labels/buttons to open the new performance comparison window.
code/src/main/resources/demo/DemoMainWindowBundle_es.properties	Spanish localization for the new main-window entries.
code/src/main/resources/demo/DemoApplicationBundle.properties	Adds a menu item to launch the vector performance window.
code/src/main/resources/demo/DemoApplicationBundle_es.properties	Spanish localization for the new menu item.
code/src/main/java/module-info.java	Declares a module dependency on jdk.incubator.vector.
code/src/main/java/demo/DemoVectorPerformanceWindow.java	New demo UI window that benchmarks vector vs scalar kernels.
code/src/main/java/demo/DemoMainWindow.java	Hooks the new demo window into the main demo UI.
code/src/main/java/demo/DemoApplication.java	Adds a demo menu item and handler to open the new window.
code/src/main/java/casciian/bits/UnicodeGlyphImage.java	Switches stddev computation to use the new distance-sum kernel (with new flattening/allocation behavior).
code/src/main/java/casciian/bits/ScaleImageUtils.java	Introduces a Vector API path for convolution accumulation in resampling.
code/src/main/java/casciian/bits/Rgb.java	Adds distanceSquaredSum(...) SIMD kernel.
code/src/main/java/casciian/bits/ArrayImageRGB.java	Vectorizes alphaBlendOver(...) via a SIMD row kernel.
code/build.gradle	Adds --add-modules jdk.incubator.vector for compile/test/native-image.

Suppressed comments (5)

code/src/main/java/casciian/bits/ScaleImageUtils.java:112

• This block declares sumR/sumG/sumB/sumW and then immediately resets them to 0 again; the second assignment is redundant and makes the inner loop harder to read.

                double sumR = 0, sumG = 0, sumB = 0, sumW = 0;
                sumR = 0; sumG = 0; sumB = 0; sumW = 0;

code/src/main/java/casciian/bits/ScaleImageUtils.java:126

• The vectorized accumulation currently allocates new double[] arrays inside the tap loop (one per source sample). This turns the hot inner loop into an allocation-heavy path and can easily dominate runtime/GC, masking any SIMD benefit.

                        // [R, G, B, W] * weight  (pad with 0 for lanes > 4)
                        double[] contrib = new double[laneCount];
                        contrib[0] = ((pixel >>> 16) & 0xFF) * weight;

code/src/main/java/casciian/bits/ScaleImageUtils.java:194

• Same issue in the vertical pass: allocating a new contrib/result array inside the tap loop makes the SIMD path allocation-bound. Reusing a single scratch array per output pixel avoids O(kernelWidth) allocations per destination pixel.

                        double[] contrib = new double[laneCount];
                        contrib[0] = ((pixel >>> 16) & 0xFF) * weight;
                        contrib[1] = ((pixel >>>  8) & 0xFF) * weight;
                        contrib[2] = ( pixel         & 0xFF) * weight;
                        contrib[3] = weight;
                        acc = acc.add(DoubleVector.fromArray(DOUBLE_SPECIES, contrib, 0));

code/src/main/java/casciian/bits/ScaleImageUtils.java:49

• The comment claims AVX2 does a fused-multiply-add (FMA) per pixel, but the implementation uses scalar multiplies into a temporary array plus a vector add (no FMA). This is misleading for future maintainers trying to reason about the vectorization strategy.

     * Preferred {@code double} vector species for SIMD convolution.
     * On x86 with AVX2 this is 4 lanes; on ARM NEON it is 2 lanes.
     * The channel layout used in the kernel is [R, G, B, W] — exactly
     * 4 doubles — so AVX2 processes one pixel's full kernel contribution
     * in a single fused-multiply-add step.

code/src/main/java/demo/DemoVectorPerformanceWindow.java:335

• This file defines a private LongSupplier even though the JDK already provides java.util.function.LongSupplier. Using the JDK type would reduce custom surface area and avoid confusing readers who expect the standard functional interface.

    @FunctionalInterface
    private interface LongSupplier {
        long getAsLong();
    }

[crramirez]

@copilot Fix the code for all comments in this review thread.

When a review comment includes a suggested change, apply the suggestion exactly.

Do not make changes beyond what is described in the linked review thread.

[Copilot]

@copilot Fix the code for all comments in this review thread.

When a review comment includes a suggested change, apply t...

Addressed in cb4f313: removed the unused imports, added direct tests for Rgb.distanceSquaredSum across varied lengths/tails, and reworked UnicodeGlyphImage stddev paths to reuse buffers instead of allocating per call.

[crramirez]

@copilot this screenshot was taken from a native build

**

**

and this running the jar in java the first time

and this after ten times

[Copilot]

@copilot this screenshot was taken from a native build

**
...

Addressed in d05d930. I added native-image-aware scalar fallbacks for ArrayImageRGB.alphaBlendOver and Rgb.distanceSquaredSum while keeping the SIMD path on regular JVM runs, so native builds no longer pay Vector API overhead where it regresses.