The llama.cpp b10306 release introduces performance optimizations for the SYCL backend, specifically targeting Gated Linear Unit (GLU) operations. The update adds a contiguous fast path for fused GLU ops and consolidates previously distinct kernels to share a common launcher.
- Adds SWIGLU performance test cases covering 17408 columns with 512 and 2048 tokens in f16 and f32.
- Consolidates fused-GLU kernels by taking the operation as an argument, dropping unused SYCL_GELU_BLOCK_SIZE and SYCL_SILU_BLOCK_SIZE constants.
- Implements a flat kernel dispatch for contiguous operands where de-interleave index math collapses to identity.
- Achieves a +14% performance improvement in f16 and +4% in f32 for split GLU on an Arc Pro B70, with fused operations remaining unchanged.
These changes improve SYCL backend efficiency by reducing code duplication and optimizing memory access patterns for specific tensor configurations.