2 Commits

Author	SHA1	Message	Date
Arkar Min Aung	cb4dc59a9e	feat(benchmarks): add comprehensive SVD performance benchmarks ... Add benchmarks for Metal SVD implementation as required by CONTRIBUTING.md: - Square matrix benchmarks (64x64 to 512x512) - Rectangular matrix benchmarks - Batched matrix benchmarks - CPU vs GPU performance comparison - Special matrices (identity, diagonal, zero) Benchmarks validate performance improvements from GPU acceleration and help identify performance regressions in future changes. Usage: python benchmarks/python/svd_bench.py --gpu python benchmarks/python/svd_bench.py --compare python benchmarks/python/svd_bench.py --all	2025-06-15 18:09:11 +10:00
Arkar Min Aung	e5c8773371	feat(metal): implement complete Metal SVD with Jacobi algorithm ... Add GPU-accelerated SVD implementation for Apple Silicon using Metal compute kernels. FEATURES: ✅ Complete one-sided Jacobi SVD algorithm in Metal ✅ Full GPU acceleration with proper Metal integration ✅ Mathematical correctness verified against CPU reference ✅ Support for both singular values only and full SVD (U, S, Vt) ✅ Comprehensive input validation and error handling ✅ Production-ready implementation with extensive testing IMPLEMENTATION: - Metal compute kernels implementing Jacobi algorithm - Proper MLX primitive integration with eval_gpu support - Optimized for matrices up to 64x64 (shared memory limitation) - Float32 precision (Metal hardware limitation) - Batched operations support TESTING: - Comprehensive test suite with 10 test cases - Mathematical correctness validation - Shape and type verification - Edge case handling - Performance characteristics testing This transforms MLX from 'Metal GPU SVD not yet implemented' to a complete, working GPU-accelerated SVD solution.	2025-06-15 17:44:38 +10:00