ROCm/composable_kernel
Sep 2025 – May 2026
6 Months active
Languages Used
C++PythonShellYAMLplaintext
Technical Skills
C++Code RefactoringGPU ProgrammingPerformance OptimizationBuild ProcessBuild System
Johannes Graner
PROFILE
Johannes Graner
Over six months, contributed to ROCm/composable_kernel by developing and optimizing high-performance GPU kernels for convolutional neural networks using C++ and CUDA. Work included refactoring grouped convolution examples with an Invoker pattern to improve modularity, implementing CI/CD enhancements for code quality, and extending kernel capabilities such as SplitN support and wavelet-based backward weight optimization. Addressed kernel robustness and accuracy through bug fixes and improved testing infrastructure, integrating GPU-based verification and profiling tools. Efforts focused on performance optimization, parallel computing, and reliability, with comprehensive testing and automation to ensure correctness and maintainability across diverse convolution workloads and hardware configurations.
Overall Statistics
Feature vs Bugs
83%Features
Repository Contributions
23Total
Bugs
2
Commits
23
Features
10
Lines of code
49,737
Activity Months6
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
May 2026
1 Commits • 1 Features
May 1, 2026May 2026 monthly summary for ROCm/composable_kernel: Delivered a wavelet-based backward weight convolution kernel optimization targeting VALU-heavy shapes. The solution separates load and math work across distinct waves, enabling improved throughput for conv-to-GEMM transformations and better utilization of VALU/MFMA units. Implemented core changes with new wavelet-enabled gridwise and device ops, updated scheduling, and corrected BlockSize logic to align with TileMathThreadGroupSize. Achieved end-to-end validation across F16/BF16 with 1D/2D/3D configurations and multiple G/N/K/C layouts. Demonstrated strong performance improvements on performance-critical workloads, including RetinaNet, while maintaining correctness. The work includes comprehensive testing and benchmarking to ensure reliability and CI readiness.
1 Commits • 1 Features
May 1, 2026May 2026 monthly summary for ROCm/composable_kernel: Delivered a wavelet-based backward weight convolution kernel optimization targeting VALU-heavy shapes. The solution separates load and math work across distinct waves, enabling improved throughput for conv-to-GEMM transformations and better utilization of VALU/MFMA units. Implemented core changes with new wavelet-enabled gridwise and device ops, updated scheduling, and corrected BlockSize logic to align with TileMathThreadGroupSize. Achieved end-to-end validation across F16/BF16 with 1D/2D/3D configurations and multiple G/N/K/C layouts. Demonstrated strong performance improvements on performance-critical workloads, including RetinaNet, while maintaining correctness. The work includes comprehensive testing and benchmarking to ensure reliability and CI readiness.
May 2026
January 2026
9 Commits • 3 Features
Jan 1, 2026Delivered three core outcomes for ROCm/composable_kernel in 2026-01, focusing on stability, robustness, and performance validation. Highlights include CI/test stability improvements for ROCm 7.1 compatibility, automatic split-k deduction for backward weight to support tensors larger than 2GB and improve robustness/accuracy in grouped convolution, and GPU-based verification/profiling enhancements that enable GPU verification on ckProfiler, GPU-side tensor initialization for profiling, and extended GPU references with richer error statistics.
January 2026
9 Commits • 3 Features
Jan 1, 2026Delivered three core outcomes for ROCm/composable_kernel in 2026-01, focusing on stability, robustness, and performance validation. Highlights include CI/test stability improvements for ROCm 7.1 compatibility, automatic split-k deduction for backward weight to support tensors larger than 2GB and improve robustness/accuracy in grouped convolution, and GPU-based verification/profiling enhancements that enable GPU verification on ckProfiler, GPU-side tensor initialization for profiling, and extended GPU references with richer error statistics.
December 2025
5 Commits • 1 Features
Dec 1, 2025December 2025 monthly summary for ROCm/composable_kernel focusing on delivering tangible business value through robust kernel fixes and expanded testing/verification tooling. Key features delivered - Verification and testing tooling enhancements for convolutions: parallelized dataset generation and tests for fwd, bwd data, and bwd weight across multiple convolution layouts; updated test orchestration and CI integration to improve coverage and reduce noise. - ckProfiler integration: GPU reference support integrated into ckProfiler for convolutions, enabling parity checks across CPU/GPU references for all directions and layouts; added profiler-based tests driven by GPU references. Major bugs fixed - GEMM kernel robustness and correctness: corrected stride calculations and behavior for grouped and two-stage GEMM kernels; addressed split-k configurations to prevent unsupported ops; improved workspace stride handling and kernel selection logic. - Deployed safety nets: disabled split-k kernel for split-k > 1 with non-contiguous strides; re-enabled the two-stage kernel only when alignment and loop conditions are satisfied, addressing regression PR feedback. Overall impact and accomplishments - Significantly improved kernel reliability and correctness, reducing risk of runtime failures in GEMM paths and strengthening verification for convolutions. - Expanded test coverage and parallelized test execution, accelerating validation cycles and enabling earlier detection of regressions. - Enhanced visibility into GPU vs CPU numerical parity through ckProfiler integration, strengthening confidence in performance and accuracy across platforms. Technologies and skills demonstrated - C++/CUDA kernel debugging and stride management, two-stage GEMM implementation and compatibility work. - Testing infrastructure design: dataset generation, parallel test execution, and test orchestration for convolutions. - Performance verification tooling: profiler integration and GPU reference workflows.
5 Commits • 1 Features
Dec 1, 2025December 2025 monthly summary for ROCm/composable_kernel focusing on delivering tangible business value through robust kernel fixes and expanded testing/verification tooling. Key features delivered - Verification and testing tooling enhancements for convolutions: parallelized dataset generation and tests for fwd, bwd data, and bwd weight across multiple convolution layouts; updated test orchestration and CI integration to improve coverage and reduce noise. - ckProfiler integration: GPU reference support integrated into ckProfiler for convolutions, enabling parity checks across CPU/GPU references for all directions and layouts; added profiler-based tests driven by GPU references. Major bugs fixed - GEMM kernel robustness and correctness: corrected stride calculations and behavior for grouped and two-stage GEMM kernels; addressed split-k configurations to prevent unsupported ops; improved workspace stride handling and kernel selection logic. - Deployed safety nets: disabled split-k kernel for split-k > 1 with non-contiguous strides; re-enabled the two-stage kernel only when alignment and loop conditions are satisfied, addressing regression PR feedback. Overall impact and accomplishments - Significantly improved kernel reliability and correctness, reducing risk of runtime failures in GEMM paths and strengthening verification for convolutions. - Expanded test coverage and parallelized test execution, accelerating validation cycles and enabling earlier detection of regressions. - Enhanced visibility into GPU vs CPU numerical parity through ckProfiler integration, strengthening confidence in performance and accuracy across platforms. Technologies and skills demonstrated - C++/CUDA kernel debugging and stride management, two-stage GEMM implementation and compatibility work. - Testing infrastructure design: dataset generation, parallel test execution, and test orchestration for convolutions. - Performance verification tooling: profiler integration and GPU reference workflows.
December 2025
November 2025
2 Commits • 1 Features
Nov 1, 2025November 2025 monthly summary for ROCm/composable_kernel highlighting key feature delivery and bug fixes that improve developer experience and model accuracy, with a focus on business value and technical excellence.
November 2025
2 Commits • 1 Features
Nov 1, 2025November 2025 monthly summary for ROCm/composable_kernel highlighting key feature delivery and bug fixes that improve developer experience and model accuracy, with a focus on business value and technical excellence.
October 2025
5 Commits • 3 Features
Oct 1, 2025October 2025 monthly summary for ROCm/composable_kernel: Delivered key CI/QC improvements, extended kernel capabilities, and a practical example to demonstrate new features. Implemented CI pre-commit hooks and remod formatting automation to ensure consistent code style and faster PR validation. Added SplitN support in Grouped Convolution Backward Data for greater flexibility and performance, and introduced a forward convolution with bias and clamp example to showcase end-to-end usage. These efforts reduce lint/format drift, improve reliability, and provide tangible examples for users and contributors, strengthening the business value of the project.
5 Commits • 3 Features
Oct 1, 2025October 2025 monthly summary for ROCm/composable_kernel: Delivered key CI/QC improvements, extended kernel capabilities, and a practical example to demonstrate new features. Implemented CI pre-commit hooks and remod formatting automation to ensure consistent code style and faster PR validation. Added SplitN support in Grouped Convolution Backward Data for greater flexibility and performance, and introduced a forward convolution with bias and clamp example to showcase end-to-end usage. These efforts reduce lint/format drift, improve reliability, and provide tangible examples for users and contributors, strengthening the business value of the project.
October 2025
September 2025
1 Commits • 1 Features
Sep 1, 2025Sep 2025 monthly summary for ROCm/composable_kernel: Delivered the Grouped Convolution Examples: Invoker Pattern Refactor. Introduced Invoker classes to encapsulate forward and backward data passes for grouped convolutions, significantly improving code organization, modularity, and maintainability of the example code. Implemented Invoker for the remaining grouped_conv_fwd and grouped_conv_bwd_data paths and fixed an incorrect out layout identifier as part of this effort. This work aligns with CK Tile architecture, lays groundwork for broader reuse across primitives, and reduces future maintenance effort.
September 2025
1 Commits • 1 Features
Sep 1, 2025Sep 2025 monthly summary for ROCm/composable_kernel: Delivered the Grouped Convolution Examples: Invoker Pattern Refactor. Introduced Invoker classes to encapsulate forward and backward data passes for grouped convolutions, significantly improving code organization, modularity, and maintainability of the example code. Implemented Invoker for the remaining grouped_conv_fwd and grouped_conv_bwd_data paths and fixed an incorrect out layout identifier as part of this effort. This work aligns with CK Tile architecture, lays groundwork for broader reuse across primitives, and reduces future maintenance effort.
Activity
Loading activity data...
Quality Metrics
Correctness87.8%
Maintainability82.6%
Architecture84.4%
Performance81.2%
AI Usage30.4%
Skills & Technologies
Programming Languages
C++PythonShellYAMLplaintext
Technical Skills
Algorithm optimizationBuild ProcessBuild SystemC++C++ DevelopmentC++ developmentC++17CI/CDCUDACode FormattingCode RefactoringContinuous IntegrationConvolution AlgorithmsConvolutional Neural NetworksDevOps
Repositories Contributed To
Overview of all repositories you've contributed to across your timeline