Module spirv_subgroup 

Sub-group optimized SPIR-V kernel generators for Intel GPUs.

This module provides SPIR-V generators that leverage Intel GPU sub-group operations (analogous to CUDA warps) for efficient intra-sub-group communication:

• reduction_subgroup_spirv — Two-phase sub-group reduction
• scan_subgroup_spirv — Inclusive prefix sum via sub-group scan
• gemm_subgroup_spirv — GEMM with sub-group shuffle for A-row broadcast

All kernels use the OpenCL SPIR-V execution model (Kernel) with Physical64/OpenCL memory model and require GroupNonUniform family capabilities.

Functions

gemm_subgroup_spirv

Generate an OpenCL SPIR-V compute kernel for GEMM with sub-group shuffle.

reduction_subgroup_spirv

Generate an OpenCL SPIR-V compute kernel for sub-group optimized reduction.

scan_subgroup_spirv

Generate an OpenCL SPIR-V compute kernel for sub-group scan (prefix sum).