ROCm/composable_kernel
Sep 2025 – Jul 2026
8 Months active
Languages Used
C++CMakePython
Technical Skills
C++Command Line ArgumentsGPU ProgrammingPerformance OptimizationSoftware DevelopmentExample Development
Michał Kulikowski
PROFILE
Michał Kulikowski
Worked on the ROCm/composable_kernel repository, delivering nine features over eight months focused on GPU kernel performance, robustness, and hardware compatibility. Developed enhancements such as configurable GEMM examples, instruction and data cache prefetching, and architecture-specific optimizations for RDNA3/4 and gfx1250. Leveraged C++, CMake, and Python scripting to implement command-line configurability, kernel parameter tuning, and automated testing. Addressed kernel execution reliability and improved developer experience by refining CLI parsing, return codes, and unit tests. The work enabled faster experimentation, improved throughput for ML and HPC workloads, and established a foundation for ongoing performance benchmarking and hardware validation within ROCm.
Overall Statistics
Feature vs Bugs
90%Features
Repository Contributions
15Total
Bugs
1
Commits
15
Features
9
Lines of code
4,988
Activity Months8
Your Network
1834 peopleSame Organization
@amd.com1733
7b30f3f5e26d48061f873d04cc7e1d1f_amdengMember
GunaShekar, AjayMember
aasbodduMember
Abdul Lateef AttarMember
Acim MaravicMember
Adam DickinMember
Pryor, AdamMember
Pryor, AdamMember
Adel JoharMember
Work History
July 2026
1 Commits • 1 Features
Jul 1, 2026Month: 2026-07 Concise monthly summary for ROCm/composable_kernel: Key features delivered: - MX GEMM: Added a hardware data cache prefetching example and alignment fixes for gfx1250 in the MX GEMM pipeline. Introduced a standalone example (mx_flatmm_data_cache_prefetch) that exercises MX GEMM through the compute TDM v1 pipeline with hardware prefetching enabled. Updated cacheline definitions to 128 bytes to align with hardware; adds per-operand prefetch controls via CLI flags. Major bugs fixed: - Fixed a bug where data cache prefetch was silently disabled in the scaled operator path of the GemmPipelineAgBgCrCompTDMV1. Defaulted prefetch to true in the scaled path, preserving behavior parity with the non-scaled path and ensuring proper emission when prefetch is enabled. Overall impact and accomplishments: - Technical: Delivered a hardware-aware prefetch enhancement in the MX GEMM path, with a validated test harness and realistic simulators. The changes culminate in improved data locality and reduced memory latency for gfx1250 workloads. - Business value: Enables higher-throughput GEMM workloads on ROCm-enabled GPUs by leveraging targeted data cache prefetching, contributing to faster ML/AI and HPC workloads that rely on MX GEMM performance. - Deliverables and traceability: Commit referenced in PR rocm/rocm-libraries #9752 (commit 5fa0bbe) with JIRA AICK-1670; tests named tile_example_mx_flatmm_mxgemm_data_cache_prefetch executed in simulator; performance impact observed in fp4 tests. Technologies/skills demonstrated: - GPU kernel optimization and data-path tuning (MX GEMM, TDM v1 pipeline) - Architecture-specific optimization for gfx1250 and alignment to 128-byte cachelines - Feature toggling with CLI flags and build gating (gfx125 gating) - Test-driven validation with simulator tests and performance metrics Client-facing summary statement: This work adds a hardware-aware prefetching path to MX GEMM, fixes a critical path bug in the scaled operator, and yields measurable throughput improvements on relevant workloads, aligning with ROCm’s goals of higher compute efficiency and broader hardware coverage.
1 Commits • 1 Features
Jul 1, 2026Month: 2026-07 Concise monthly summary for ROCm/composable_kernel: Key features delivered: - MX GEMM: Added a hardware data cache prefetching example and alignment fixes for gfx1250 in the MX GEMM pipeline. Introduced a standalone example (mx_flatmm_data_cache_prefetch) that exercises MX GEMM through the compute TDM v1 pipeline with hardware prefetching enabled. Updated cacheline definitions to 128 bytes to align with hardware; adds per-operand prefetch controls via CLI flags. Major bugs fixed: - Fixed a bug where data cache prefetch was silently disabled in the scaled operator path of the GemmPipelineAgBgCrCompTDMV1. Defaulted prefetch to true in the scaled path, preserving behavior parity with the non-scaled path and ensuring proper emission when prefetch is enabled. Overall impact and accomplishments: - Technical: Delivered a hardware-aware prefetch enhancement in the MX GEMM path, with a validated test harness and realistic simulators. The changes culminate in improved data locality and reduced memory latency for gfx1250 workloads. - Business value: Enables higher-throughput GEMM workloads on ROCm-enabled GPUs by leveraging targeted data cache prefetching, contributing to faster ML/AI and HPC workloads that rely on MX GEMM performance. - Deliverables and traceability: Commit referenced in PR rocm/rocm-libraries #9752 (commit 5fa0bbe) with JIRA AICK-1670; tests named tile_example_mx_flatmm_mxgemm_data_cache_prefetch executed in simulator; performance impact observed in fp4 tests. Technologies/skills demonstrated: - GPU kernel optimization and data-path tuning (MX GEMM, TDM v1 pipeline) - Architecture-specific optimization for gfx1250 and alignment to 128-byte cachelines - Feature toggling with CLI flags and build gating (gfx125 gating) - Test-driven validation with simulator tests and performance metrics Client-facing summary statement: This work adds a hardware-aware prefetching path to MX GEMM, fixes a critical path bug in the scaled operator, and yields measurable throughput improvements on relevant workloads, aligning with ROCm’s goals of higher compute efficiency and broader hardware coverage.
July 2026
May 2026
1 Commits • 1 Features
May 1, 2026In May 2026, delivered a targeted performance/robustness enhancement for ROCm/composable_kernel: Instruction Cache Prefetching Enhancement with improved object file path handling and automatic llvm-objdump detection. Completed Copilot-review fixes for the Instruction Cache POC, and laid groundwork for more reliable prefetching behavior to boost kernel throughput.
May 2026
1 Commits • 1 Features
May 1, 2026In May 2026, delivered a targeted performance/robustness enhancement for ROCm/composable_kernel: Instruction Cache Prefetching Enhancement with improved object file path handling and automatic llvm-objdump detection. Completed Copilot-review fixes for the Instruction Cache POC, and laid groundwork for more reliable prefetching behavior to boost kernel throughput.
April 2026
1 Commits • 1 Features
Apr 1, 2026Month: 2026-04 | Repository: ROCm/composable_kernel Key accomplishments: Delivered a proof-of-concept for instruction cache prefetching to optimize GPU kernel performance. The PoC is implemented in commit 8de4cb72fba519d89948825783f5764de163d315 and demonstrates a potential path to reduce instruction fetch latency in compute kernels within the composable_kernel codebase. This work lays the groundwork for targeted performance measurements and future integration into the ROCm CK optimization stack. Major bugs fixed: No major bugs documented for this repository in this month. Overall impact and accomplishments: Establishes a concrete optimization direction for GPU kernels by exploring instruction cache prefetching. The PoC showcases the team’s capability to prototype performance-oriented changes within ROCm’s CK, setting the stage for benchmarking, validation, and potential scalability across workloads. Expected business value includes improved kernel throughput on memory-bound workloads and reduced latency in instruction fetch paths. Technologies/skills demonstrated: GPU kernel performance optimization, memory hierarchy awareness (instruction cache behavior), ROCm/CK codebase contribution, commit-driven development, proof-of-concept prototyping, and readiness for performance benchmarking.
1 Commits • 1 Features
Apr 1, 2026Month: 2026-04 | Repository: ROCm/composable_kernel Key accomplishments: Delivered a proof-of-concept for instruction cache prefetching to optimize GPU kernel performance. The PoC is implemented in commit 8de4cb72fba519d89948825783f5764de163d315 and demonstrates a potential path to reduce instruction fetch latency in compute kernels within the composable_kernel codebase. This work lays the groundwork for targeted performance measurements and future integration into the ROCm CK optimization stack. Major bugs fixed: No major bugs documented for this repository in this month. Overall impact and accomplishments: Establishes a concrete optimization direction for GPU kernels by exploring instruction cache prefetching. The PoC showcases the team’s capability to prototype performance-oriented changes within ROCm’s CK, setting the stage for benchmarking, validation, and potential scalability across workloads. Expected business value includes improved kernel throughput on memory-bound workloads and reduced latency in instruction fetch paths. Technologies/skills demonstrated: GPU kernel performance optimization, memory hierarchy awareness (instruction cache behavior), ROCm/CK codebase contribution, commit-driven development, proof-of-concept prototyping, and readiness for performance benchmarking.
April 2026
March 2026
1 Commits • 1 Features
Mar 1, 2026March 2026 monthly summary for ROCm/composable_kernel focusing on business value and technical achievements. Implemented command-line configurability for GEMM examples to enable verification, timing, and matrix dimension adjustments, improving experimentation speed and reproducibility. No major bugs fixed in available data. Overall impact includes improved usability for performance research, enabling faster iteration and benchmarking of GEMM variants. Technologies demonstrated include C++ development for CK, command-line parameter parsing, GEMM kernel variants (XDL, INT8 quantization), and integration with ROCm libraries.
March 2026
1 Commits • 1 Features
Mar 1, 2026March 2026 monthly summary for ROCm/composable_kernel focusing on business value and technical achievements. Implemented command-line configurability for GEMM examples to enable verification, timing, and matrix dimension adjustments, improving experimentation speed and reproducibility. No major bugs fixed in available data. Overall impact includes improved usability for performance research, enabling faster iteration and benchmarking of GEMM variants. Technologies demonstrated include C++ development for CK, command-line parameter parsing, GEMM kernel variants (XDL, INT8 quantization), and integration with ROCm libraries.
January 2026
3 Commits • 2 Features
Jan 1, 2026January 2026 monthly summary for ROCm/composable_kernel: Delivered key kernel execution robustness and readability improvements in CK examples and extended RDNA3/4 GEMM support with performance-focused parameter tuning. The work enhances reliability, maintainability, and hardware readiness for next-gen GPUs, directly supporting faster developer onboarding and improved business value.
3 Commits • 2 Features
Jan 1, 2026January 2026 monthly summary for ROCm/composable_kernel: Delivered key kernel execution robustness and readability improvements in CK examples and extended RDNA3/4 GEMM support with performance-focused parameter tuning. The work enhances reliability, maintainability, and hardware readiness for next-gen GPUs, directly supporting faster developer onboarding and improved business value.
January 2026
November 2025
2 Commits • 1 Features
Nov 1, 2025Month 2025-11 || ROCm/composable_kernel: Delivered AMD GPU Scalar Buffer Prefetching Optimization and Testing. Implemented s_prefetch functionality with inline assembly for s_buffer_load_b32/64, and refactored/extended unit tests to validate prefetch paths. Fixed unit-test issues in the prefetch path to improve reliability. Key commits include f3ef7acca07a12a25c9a33279423cf617cbe27f8 and cd8af997e6d1fde6bc4397bd6ab4fca46510e776.
November 2025
2 Commits • 1 Features
Nov 1, 2025Month 2025-11 || ROCm/composable_kernel: Delivered AMD GPU Scalar Buffer Prefetching Optimization and Testing. Implemented s_prefetch functionality with inline assembly for s_buffer_load_b32/64, and refactored/extended unit tests to validate prefetch paths. Fixed unit-test issues in the prefetch path to improve reliability. Key commits include f3ef7acca07a12a25c9a33279423cf617cbe27f8 and cd8af997e6d1fde6bc4397bd6ab4fca46510e776.
October 2025
4 Commits • 1 Features
Oct 1, 2025October 2025 monthly summary for ROCm/composable_kernel: - Focused on extending architecture coverage (RDNA3/4) in examples, improving robustness, and enhancing developer experience across the composable_kernel suite. - Work centered on delivering architecture compatibility, reliability improvements, and CLI/parameter usability for CK examples.
4 Commits • 1 Features
Oct 1, 2025October 2025 monthly summary for ROCm/composable_kernel: - Focused on extending architecture coverage (RDNA3/4) in examples, improving robustness, and enhancing developer experience across the composable_kernel suite. - Work centered on delivering architecture compatibility, reliability improvements, and CLI/parameter usability for CK examples.
October 2025
September 2025
2 Commits • 1 Features
Sep 1, 2025September 2025: ROCm/composable_kernel delivered key GEMM enhancements and stability improvements, focusing on performance and hardware readiness for RDNA3/4. Implemented configurable grouping for grouped GEMM examples and extended RDNA3/4 support across a broad set of GEMM paths. Also fixed a command parser issue in grouped_conv_bwd_weight, improving usability and reliability of the example suite. The work strengthens business value by enabling better utilization of newer GPUs, faster experimentation, and more predictable performance in user workflows.
September 2025
2 Commits • 1 Features
Sep 1, 2025September 2025: ROCm/composable_kernel delivered key GEMM enhancements and stability improvements, focusing on performance and hardware readiness for RDNA3/4. Implemented configurable grouping for grouped GEMM examples and extended RDNA3/4 support across a broad set of GEMM paths. Also fixed a command parser issue in grouped_conv_bwd_weight, improving usability and reliability of the example suite. The work strengthens business value by enabling better utilization of newer GPUs, faster experimentation, and more predictable performance in user workflows.
Activity
Loading activity data...
Quality Metrics
Correctness82.6%
Maintainability81.4%
Architecture80.0%
Performance77.4%
AI Usage29.4%
Skills & Technologies
Programming Languages
C++CMakePython
Technical Skills
Build system configurationC++C++ DevelopmentCMakeCUDACommand Line ArgumentsExample DevelopmentGPGPUGPU ComputingGPU ProgrammingGPU programmingHIPHigh-Performance ComputingKernel OptimizationLow-Level Programming
Repositories Contributed To
Overview of all repositories you've contributed to across your timeline