OptimizedSparseGEMM

This repo now contains four progressively optimized BCSR SGEMM kernels:

1. Masked Core (previous)

   • On-the-fly AVX2 masked loads/stores for non-multiple-of-8 tails.

2. Padded Core (previous)

   • Zero-padded blocks to the next multiple of 8, pure 8-wide AVX2.

---

New Advanced Versions

3. bcsr_sgemm_masked_advanced.h

• Register‐blocked over 4 rows at a time.
• Aligned loads/stores (_mm256_load_ps / _store_ps).
• Software prefetch inserted for weight and output.
• Single mask computed once and reused.
• Minimizes branching in inner loops.

4. bcsr_sgemm_padded_advanced.h

• Offline block packing: all weight blocks pre-padded and aligned.
• Aligned 32-byte buffers for prefetch and load/store.
• Register‐block of 4 rows per iteration.
• Eliminates memcpy/memset in hot loop by moving to preload.
• Simplest inner loop: only _mm256_fmadd_ps on aligned data.

---

Benchmarks & Tuning

• Both versions target x64 AVX2; tune r (4) and c (8) for your CPU.
• Measure with std::chrono::high_resolution_clock, capture GFLOPS.
• Compare memory bandwidth using tools like perf or Intel VTune.

Choose the variant that best fits your memory‐compute tradeoffs:

• Masked for minimal extra storage & dynamic tails.
• Padded for best pure‐vector throughput at slight memory overhead.