December 2025 (2025-12) monthly summary for ROCm/composable_kernel. Focused on stability and architecture-specific compatibility. Delivered a targeted workaround to stabilize backward-ops on gfx90a for ROCm 7.1.1, preventing runtime errors due to insufficient wait-states between v_mfma_f32... and v_accvgpr_read_b32 when separated by s_cbranch. The change reduces production risk and improves reliability for workloads relying on gfx90a.

In 2025-11, the ROCm/composable_kernel team delivered a targeted optimization for the FMHA (softmax attention) forward pass with dropout. The changes reduce register spilling by vectorizing the storage of dropout random values, ensure the randvals are calculated and stored only once, and optimize memory traffic in dropout-enabled paths. A clang-22 CI workaround was implemented to improve CI stability, and the work was designed to be non-breaking for existing public APIs while delivering measurable throughput gains in attention kernels across transformer workloads.

October 2025 performance summary for ROCm/composable_kernel focusing on FMHA/WMMA on gfx12, multi-arch readiness, and reliability improvements. Delivered significant FMHA enhancements on gfx12, expanded arch-specific kernel generation, and validated cross-arch readiness to support a broader hardware base. Implemented critical build/test stability fixes and synchronization improvements to boost reliability in transformer workloads.

Monthly performance summary for Sep 2025 focused on ROCm/composable_kernel. Highlights include: extensive FMHA testing/validation suite, performance-oriented build-time reductions, synchronization/stability fixes across FP16/FP32 paths, and FP32 data-path support enabling broader precision coverage. The work enhances robustness, determinism, and business value by ensuring reliable FMHA kernels, faster CI feedback, and wider precision applicability.

July 2025 monthly summary for StreamHPC/rocm-libraries. Focused on strengthening numerical robustness in CK_TILE and stabilizing floating-point conversions, with emphasis on delivering reliable, production-ready math pathways and improving test coverage.

June 2025 monthly summary for StreamHPC/rocm-libraries. Focused on delivering a universal WMMA GEMM pipeline with mixed-precision and padding refinements, expanding data-type support, and optimizing test workflows. Key outcomes include faster validation cycles, broader hardware compatibility, and improved build reliability. The work aligns with business goals of accelerating ROCm library readiness and enabling more robust performance-critical workloads.

April 2025 highlights for StreamHPC/rocm-libraries: Delivered DeviceGemm_Wmma_CShuffleV3 GEMM with WMMA support (BlockGemmPipelineVersion::v3) across gfx11/gfx12, expanding data types to include FP8 variants (F8/BF8), introducing new layout variants and enhanced profiling capabilities. Implemented FP8 WMMA bug fixes to improve correctness and reliability of WMMA paths. This milestone is backed by the commit edd92fc546663094f42366e12a172701f18a2fd9 with message “DeviceGemm_Wmma_CShuffleV3 with BlockGemmPipelineVersion::v3 (#2096)”.

March 2025 performance summary for StreamHPC/rocm-libraries: Finalized Batch Normalization OpenCL kernel optimizations, enabling vectorization for forward and backward passes across NHWC and NCHW layouts. Achievements include improved workgroup sizing, enhanced memory access patterns, and robustness enhancements, all contributing to higher BN throughput and more stable performance in ROCm ML pipelines. Two commits under #3564 were landed to complete the work. No separate major bug fixes were required this month; the primary focus was delivering the optimization feature and its robustness improvements. This work strengthens production-ready BN performance and cross-layout support, enabling downstream frameworks to rely on more predictable BN behavior.