Benchmark Variants

Each benchmark should have the following variants with different optimizations and parallelization schemes. Each benchmark should also have file README_fpga, describing some more details of the variants.

Version 0 (v0)

• Minimal changes to be compatible with AOCL
  • Ahead-of-time compilation
  • Device selection
  • No struct kernel parameter
  • Aggregating multiple kernel files into a single file
  • restrict
• No optimization
• Originally developed for GPUs with NDRange kernels

Version 1 (v1)

• Unoptimized single work-item kernels with restrict and ivdep if necessary to allow correct pipelining

Version 2 (v2)

• NDRange
• Basic optimizations
  • work-group size (reqd_work_group_size or max_work_group_size)
  • simd lanes (num_simd_work_items)
  • compute units (num_compute_units)

Version 3 (v3)

• Single work-item
• Basic optimizations
  • Shift register for floating-point reduction
  • Unrolling

Version 4 (v4)

• NDRange
• Advanced optimizations
  • Kernel rewrite
  • Local memory access reduction

Version 5 (v5)

• Single work-item
• Advanced optimizations
  • Kernel rewrite
  • Shift register as on-chip buffer
  • Loop blocking/tiling
  • Loop collapse
  • Exit condition optimization
  • Systolic array
  • etc.