MLX-Cog-Gen

MLX-accelerated Cloud Optimized GeoTIFF generator for Apple Silicon.

About

• Replaces GDAL's CPU-based pyramid/overview generation with an MLX GPU implementation on Apple Silicon
• Everything else in the COG pipeline (tiling, compression, metadata) stays with GDAL
• GDAL is a required system dependency (links against the installed library; not bundled)

Why Apple Silicon

• On x86, GPU availability varies; GDAL never uses a GPU for overviews either way
• Every M-series Mac has a high-performance GPU in the same package as the CPU (shared memory)
• GDAL leaves that GPU idle for the entire overview step
• Optimising for Apple Silicon benefits all M-series machines, not a subset with discrete GPUs

Resampling methods

GDAL's default overview resampling is NEAREST:

• Picks one pixel from each 2×2 block and discards the rest
• Throws away 75% of the signal per level on continuous data
• Narrow features (e.g. a one-pixel ridge) can vanish at coarser levels

mlx_translate supports two methods that avoid this.

AVERAGE (default)

• Every pixel in a 2×2 block contributes equally:

out[i,j] = (src[2i,2j] + src[2i,2j+1] + src[2i+1,2j] + src[2i+1,2j+1]) / count_valid

• Box filter; preserves signal energy across zoom levels
• Standard choice for continuous rasters in geospatial workflows
• Maps directly onto GPU array ops with no approximation

BILINEAR

• Treats each output pixel as a point sample, not a fixed block average
• Sample position for output pixel i in source space:

x = (i + 0.5) * 2 - 0.5 = 2i + 0.5

• Lands halfway between source pixels 2i and 2i+1
• Tent filter (w = 1 - |distance|) assigns weight 0.5 to each neighbour
• Two 1D passes (horizontal, then vertical); matches GDAL's separable bilinear
• At 2× interior pixels, numerically the same as AVERAGE; model differs (interpolation vs integration)
• Most common alternative to AVERAGE; cubic/lanczos are higher-order and not defaults

Limitations

• Float32 only
  • Input bands must be Float32; other dtypes are rejected at the CLI
  • In scope: continuous elevation/analytic products (DSM, DTM, DEM, slope, aspect, similar)
  • Out of scope: Byte orthos, UInt16 imagery, Float16, integer elevation encodings, categorical maps
  • Use gdal_translate for non-Float32
• Memory
  • Each band is loaded fully into unified memory before GPU dispatch
  • Uncompressed raster must fit in available system memory
  • Size estimate: width × height × bands × bytes_per_pixel
  • Example: Float32 single-band 30000×30000 ≈ 3.4 GB
  • GDAL strips large rasters; if input will not fit, use gdal_translate

Build

Install dependencies:

brew install gdal cmake mlx

Build and test:

mkdir build && cd build
cmake .. -DCMAKE_PREFIX_PATH=/opt/homebrew
make
ctest --output-on-failure

Usage

build/mlx_translate input.tif output_cog.tif

• Default: COG with LZW compression and AVERAGE resampling
• -r selects resampling; -co KEY=VALUE overrides creation options

build/mlx_translate input.tif output_cog.tif -r BILINEAR
build/mlx_translate input.tif output_cog.tif -r AVERAGE -co COMPRESS=DEFLATE

• Supported resampling: AVERAGE (default), BILINEAR

Benchmarks

• Hardware: M1 Pro (16 GB)
• 5 runs per method
• Float32 single-band DEMs via TIN interpolation at six GSDs
• GDAL 1T: single-threaded default
• GDAL nT: ALL_CPUS (10 cores)
• MLX: parallel LZW tile compression (GDAL_NUM_THREADS=ALL_CPUS) on final COG write

AVERAGE

Raster	Dimensions	File size	GDAL 1T	GDAL nT	MLX	vs GDAL 1T	vs GDAL nT
dem_160cm	1873×1817	4.6 MB	0.345s	0.248s	0.325s	1.06× faster	1.31× slower
dem_80cm	3746×3634	14 MB	0.788s	0.393s	0.494s	1.60× faster	1.26× slower
dem_40cm	7491×7268	39 MB	2.310s	0.871s	0.999s	2.31× faster	1.15× slower
dem_20cm	14982×14536	128 MB	7.962s	2.780s	2.646s	3.01× faster	1.05× faster
dem_10cm	29967×29074	323 MB	29.309s	9.432s	5.130s	5.71× faster	1.84× faster
dem_5cm	59927×58141	928 MB	112.418s†	35.555s	13.012s	8.64× faster	2.73× faster

BILINEAR

Raster	Dimensions	File size	GDAL 1T	GDAL nT	MLX	vs GDAL 1T	vs GDAL nT
dem_160cm	1873×1817	4.6 MB	0.349s	0.248s	0.325s	1.07× faster	1.31× slower
dem_80cm	3746×3634	14 MB	0.831s	0.403s	0.491s	1.69× faster	1.22× slower
dem_40cm	7491×7268	39 MB	2.442s	0.903s	0.994s	2.46× faster	1.10× slower
dem_20cm	14982×14536	128 MB	8.798s	2.880s	2.652s	3.32× faster	1.09× faster
dem_10cm	29967×29074	323 MB	32.968s‡	9.815s	4.979s	6.62× faster	1.97× faster
dem_5cm	59927×58141	928 MB	126.848s	36.882s‡	13.055s	9.72× faster	2.83× faster

• † 1 of 5 runs excluded (144.8s outlier); average from 4 valid runs
• ‡ dem_10cm GDAL 1T: 1 of 5 runs excluded (293.9s outlier). dem_5cm GDAL nT: 1 of 5 runs excluded (57.4s outlier)

AVERAGE	BILINEAR

• MLX beats GDAL nT from dem_20cm (128 MB) upward
• At dem_5cm (~60k×58k): 2.73× (AVERAGE), 2.83× (BILINEAR) vs GDAL nT
• Below ~128 MB, Metal launch overhead can make MLX slower than GDAL nT
• MLX AVERAGE and BILINEAR wall times are nearly identical (GPU parallelises both uniformly)

Roadmap

• Additional resampling algorithms (cubic, lanczos)
• GPU-accelerated statistics (min, max, mean, stddev) across overview levels
• GPU-accelerated tile block creation (512×512; currently delegated to GDAL)
• OOM detection and graceful fallback to GDAL

Contributing

• Open an issue before raising a PR
• Describe what you want to do and why
• Open the PR once the approach is agreed

License

MIT