Optimize radix sort for multi-work-group performance

[danhoeflinger]

This PR refactors the SYCL radix sort implementation to improve performance across multiple work-groups. The optimization focuses on kernel simplification, better memory access patterns, and improved work distribution.

Key Changes

Kernel Consolidation

• Removed the separate __radix_sort_reorder_peer_kernel
• Unified even/odd iteration logic to reduce kernel variants
• Eliminated compile-time kernel specialization overhead

Count Kernel Optimization

• Reorganized to support larger work-group sizes, previously limited to a single subgroup which resulted in poor occupancy
• Changed from _CountT (uint32) to uint8 accumulation in shared local memory (SLM) for bucket counters
• Improved SLM indexing with __index_views helper for better cache locality during counting phase
• Implemented 8-way unrolling with subgroup-strided memory access for better coalescing
• Optimized reduction with tree-based aggregation where all work-items participate in parallel
• Reduced register pressure by having each work-item handle 4 radix states during reduction

Reorder Kernel Improvements

• Redesigned to support larger work-group sizes, previously limited to a single subgroup per work-group which resulted in bad utilization of the hardware

Code Simplification

• Removed complex peer prefix algorithm variants (subgroup_ballot, atomic_fetch_or, scan_then_broadcast)

[Copilot]

Pull request overview

This PR optimizes the SYCL radix sort implementation to improve multi-work-group performance through kernel simplification and better memory access patterns. The optimization removes compile-time kernel specialization, unifies iteration logic, and improves work distribution across larger work-group sizes.

Changes:

• Consolidated kernel variants by removing separate peer reorder kernel and even/odd iteration specializations
• Optimized count kernel with uint8 accumulation, improved SLM indexing, 8-way unrolling, and tree-based reduction
• Redesigned reorder kernel to support larger work-group sizes with better subgroup coordination

Reviewed changes

Copilot reviewed 2 out of 2 changed files in this pull request and generated 4 comments.

File	Description
parallel_backend_sycl_utils.h	Added __min_sub_group_size helper function to query minimum supported subgroup size
parallel_backend_sycl_radix_sort.h	Major refactoring of radix sort kernels with simplified templates, new count/reorder implementations, and unified iteration logic

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.