gluon-beyond-loop-opt

To reduce register pressure in Gluon GEMM kernels, apply epilogue M-slicing. This technique slices the final accumulator into smaller live portions, overlapping MFMA operations with buffer_store instructions to hide store latency and lower VGPR pressure.

XCD remapping is an optimization technique for GEMM kernels that improves L2 cache efficiency on CDNA GPUs. It distributes program IDs across XCDs using a grouped swizzle pattern, maximizing matrix B tile reuse across M-block groups for higher throughput.

To reduce register pressure in Gluon GEMM kernels, apply epilogue M-slicing. This technique slices the final accumulator into smaller live portions, overlapping MFMA operations with buffer_store instructions to hide store latency and lower VGPR pressure.

Yes, the Gluon GEMM optimization requires the kernel to expose NUM_XCDS and GROUP_SIZE_M controls. These parameters are necessary for tuning XCD remapping and ensuring compatibility with Gluon kernels on MI300X and MI350 GPUs.

Apply beyond-loop GEMM optimizations when targeting MI300X or MI350 GPUs with large M, N, and K shapes where L2 bandwidth is a bottleneck. These optimizations maximize MFMA utilization by improving cache efficiency and reducing register spills.

To verify correctness after applying XCD remapping and epilogue slicing, run the optimization skill to execute the modified Gluon GEMM kernel and compare its output against a reference matmul implementation to ensure numerical accuracy.