feat: implement WASM SIMD optimizations for PNG filter operations

[superstructor]

This commit adds comprehensive WASM SIMD support for PNG filter operations, targeting 4x overall decoding speedup for typical PNG images.

Performance Targets

• Sub filter: ≥4x speedup for RGBA images
• Up filter: ≥5x speedup (simplest filter)
• Average filter: ≥3x speedup
• Overall PNG decode: ≥3x speedup for typical images

Changes

New Files

• wasm/png_wasm_filter_init.c: Integration layer that hooks SIMD filter functions into libpng's PNG_FILTER_OPTIMIZATIONS dispatch system
• bench/png-simd-bench.ts: Comprehensive benchmark suite for validating SIMD performance across different image sizes and filter types
• SIMD_IMPLEMENTATION.md: Complete documentation of the SIMD implementation, including technical details, usage instructions, and performance notes

Modified Files

• wasm/filter_wasm_simd.c: Fixed critical bugs in existing SIMD implementations:

  • Sub filter: Simplified to rely on loading updated values from previous iteration
  • Average filter: Fixed to use floor division instead of rounding up (wasm_u8x16_avgr rounds up, but PNG spec requires floor division)
  • Added proper comments explaining dependency handling

• meson.build: Updated build configuration to enable SIMD compilation:

  • Added wasm_simd_sources file list
  • Conditional compilation based on 'simd' option
  • Added -DPNG_WASM_SIMD_OPT=1 and PNG_FILTER_OPTIMIZATIONS macro
  • Added -msimd128 compiler and linker flags
  • Updated all three build targets (main, side, static library)

• deno.json: Updated build tasks to enable SIMD by default:

  • Added -Dsimd=true to build:main and build:side tasks
  • Added bench:simd task for running SIMD performance benchmarks

Technical Implementation

Up Filter (png_read_filter_row_up_wasm_simd)

Fully vectorized - processes 16 bytes per iteration with no dependencies. Expected 5-6x speedup over scalar implementation.

Sub Filter (png_read_filter_row_sub{3,4,6,8}_wasm_simd) Vectorized with dependency chain handling. Each iteration loads updated values from previous iteration to maintain correctness. Expected 4x speedup for RGBA (bpp=4).

Average Filter (png_read_filter_row_avg{3,4,6,8}_wasm_simd) Vectorized with corrected floor division. Uses XOR trick to convert wasm_u8x16_avgr's rounding up to proper floor division required by PNG spec. Expected 3x speedup.

Paeth Filter (png_read_filter_row_paeth{3,4,6,8}_wasm_simd) Currently uses scalar implementation due to complex predictor logic. Future optimization opportunity using SIMD min/max operations.

Building

Enable SIMD:

deno task build:wasm  # Builds with -Dsimd=true by default

Disable SIMD (scalar fallback):

meson setup build-main --cross-file=scripts/emscripten.cross -Dsimd=false

Testing

Run standard tests:

deno task test

Run SIMD benchmarks:

deno task bench:simd

Browser Compatibility

WASM SIMD supported in Chrome 91+, Edge 91+, Firefox 89+, Safari 16.4+. Automatic fallback to scalar implementation for older browsers.

Closes: Performance optimization for PNG decoding