gluon-mm-inst-opt

Optimize Gluon GEMM memory access on CDNA GPUs by replacing gl.load/gl.store with buffer_load/buffer_store to reduce branches and address pressure. On CDNA4, it uses async_copy to shared memory through LDS to enable prefetch-ready pipelines.

Async_copy to shared memory does not work on CDNA3 GPUs. CDNA3 uses buffer_load and buffer_store to optimize memory paths, while async_copy to LDS for streaming DMA is exclusively a CDNA4 gfx950 feature.

Use buffer_load instead of gl.load for large GEMM tile workloads to reduce branches and address pressure. Replacing global loads with AMD-native memory instructions improves MFMA throughput and reduces latency on MI300X and MI350 GPUs.

Enable prefetch-ready pipelines for GEMM kernels on CDNA4 by introducing shared memory and using async_copy.buffer_load_to_shared. This establishes a streaming DMA path from global memory to LDS to optimize large GEMM workloads.

This optimization focuses on large GEMM tile workloads and performance-driven workflows. Applying it to small GEMM tiles may not yield the intended MFMA throughput improvements or latency reductions.

Apply Step A and Step B memory optimizations by running the instruction replacement process. Step A replaces global loads with buffer instructions, and Step B applies CDNA4 async_copy to LDS if the target hardware supports it.