oneapi-src/oneDNN
Oct 2024 – Oct 2025
13 Months active
Languages Used
C++CCMakeMarkdownPythonAssemblyOpenCLTOML
Technical Skills
BenchmarkingCPU ReorderingConvolutional Neural NetworksData Type ConversionGPU ProgrammingGraph Neural Networks
Dmitrii Zarukin
PROFILE
Dmitrii Zarukin
Dmitry Zarukin contributed to the oneapi-src/oneDNN repository, engineering high-performance CPU and GPU kernels for deep learning primitives such as convolution, matmul, and softmax. He focused on optimizing quantization, memory management, and in-kernel scale handling, refactoring core C++ and assembly code to improve numerical accuracy and runtime efficiency. Dmitry enhanced benchmarking infrastructure, expanded data-type support, and modernized build and testing workflows, addressing cross-platform compatibility and parallelism. His work included robust API design, low-level optimization, and debugging improvements, resulting in more reliable, maintainable, and scalable code. The depth of his contributions advanced both performance and developer productivity across the project.
Overall Statistics
Feature vs Bugs
68%Features
Repository Contributions
371Total
Bugs
63
Commits
371
Features
136
Lines of code
43,475
Activity Months13
Your Network
1970 people
Work History
October 2025
6 Commits • 3 Features
Oct 1, 2025October 2025 performance and stability enhancements for oneDNN (repo: oneapi-src/oneDNN). Delivered key features that boost compute efficiency, strengthen test coverage, and modernize parallelism support, with a focus on business value and technical excellence.
6 Commits • 3 Features
Oct 1, 2025October 2025 performance and stability enhancements for oneDNN (repo: oneapi-src/oneDNN). Delivered key features that boost compute efficiency, strengthen test coverage, and modernize parallelism support, with a focus on business value and technical excellence.
October 2025
September 2025
27 Commits • 6 Features
Sep 1, 2025September 2025 (2025-09) monthly summary for oneDNN: Key kernel and benchdnn improvements across CPU backends (AMX, AVX512) with a focus on performance, memory efficiency, and stability. Implemented in-kernel handling of scales for conv and matmul across CPU backends, enabling reduced host-side data movement and improved inference throughput. Refined benchdnn convolution reference kernel to lower memory traffic and optimize data layout (switching to an axb layout in src/weights, among other optimizations). Strengthened benchdnn memory tracking and scratchpad management with device memory accounting, mapped buffers, and improved utilities, boosting reproducibility and resource visibility. Introduced benchdnn implementation summary option support for easier benchmarking and reporting. Extended brgemm with single-scale data type support and addressed registry scratchpad usage to improve kernel robustness. Overall, these changes deliver measurable business value by accelerating CPU-based inference, increasing benchmarking reliability, and simplifying performance tuning across platforms.
September 2025
27 Commits • 6 Features
Sep 1, 2025September 2025 (2025-09) monthly summary for oneDNN: Key kernel and benchdnn improvements across CPU backends (AMX, AVX512) with a focus on performance, memory efficiency, and stability. Implemented in-kernel handling of scales for conv and matmul across CPU backends, enabling reduced host-side data movement and improved inference throughput. Refined benchdnn convolution reference kernel to lower memory traffic and optimize data layout (switching to an axb layout in src/weights, among other optimizations). Strengthened benchdnn memory tracking and scratchpad management with device memory accounting, mapped buffers, and improved utilities, boosting reproducibility and resource visibility. Introduced benchdnn implementation summary option support for easier benchmarking and reporting. Extended brgemm with single-scale data type support and addressed registry scratchpad usage to improve kernel robustness. Overall, these changes deliver measurable business value by accelerating CPU-based inference, increasing benchmarking reliability, and simplifying performance tuning across platforms.
August 2025
22 Commits • 7 Features
Aug 1, 20252025-08 performance summary for oneapi-src/oneDNN: Delivered API enhancement for precomputed reductions attribute, fixed critical kernel scaling bugs, overhauled benchdnn memory management, extended FP8 destination support for CPU and GPU softmax, and improved code quality and consistency across CPU kernels and JIT. These changes improve numerical correctness, runtime stability, and performance visibility, while providing stronger memory accounting and tooling for maintainability.
22 Commits • 7 Features
Aug 1, 20252025-08 performance summary for oneapi-src/oneDNN: Delivered API enhancement for precomputed reductions attribute, fixed critical kernel scaling bugs, overhauled benchdnn memory management, extended FP8 destination support for CPU and GPU softmax, and improved code quality and consistency across CPU kernels and JIT. These changes improve numerical correctness, runtime stability, and performance visibility, while providing stronger memory accounting and tooling for maintainability.
August 2025
July 2025
18 Commits • 5 Features
Jul 1, 2025July 2025 (2025-07) monthly development summary for oneDNN. Delivered cross-cutting improvements across memory handling, GPU debugging, cache/primitive creation, and core brgemm/matmul workflows. The work enhances reliability, performance readiness, and developer efficiency, delivering concrete business value through more predictable behavior, faster triage, and a cleaner codebase.
July 2025
18 Commits • 5 Features
Jul 1, 2025July 2025 (2025-07) monthly development summary for oneDNN. Delivered cross-cutting improvements across memory handling, GPU debugging, cache/primitive creation, and core brgemm/matmul workflows. The work enhances reliability, performance readiness, and developer efficiency, delivering concrete business value through more predictable behavior, faster triage, and a cleaner codebase.
June 2025
8 Commits • 2 Features
Jun 1, 2025June 2025 (2025-06) highlights: Delivered high-impact improvements in oneDNN focused on quantization accuracy, zero-point handling, parser robustness, and cross-platform kernel enablement. Key features and fixes delivered across oneapi-src/oneDNN: - Quantization Scale Handling Improvements for brgemm Post-Ops: Consolidates and optimizes the application of scales in brgemm post-ops and related post-work, leveraging the with_scales flag for correct and efficient behavior. Commit trail includes 1a274be0d448f14a05ad570ae5374332119f58a7, eabd14913b84ef6f6aa3ccbff7f2c07ae9fa76dc, and 8820cfffb97ede7219bbd7d8de47183972574ccf. - Zero-point Handling Cleanup Across Convolution, MatMul, and Reorder: Refactors zero-point data access and propagation to a consistent CTX_IN_MEM-based approach, and removes redundant zero-point handling in f32 reorder. Commit trail includes 495dff5a1109a6f94575d6f0382b9e6ae87281c7 and 9e492c16e75cd89c05d674391448ae424ca75fe1. - Benchdnn Parser Robustness Against Undefined Behavior: Fixes string searching in benchdnn parser by replacing direct string::find with a robust matching helper to prevent UB. Commit 6bb6bba8522fcbed26b3f557bad37f8292cc5e7e. - PPC64 Macro Check Fixes to Enable Kernel Support: Corrects DNNL_PPC64 macro checks to ensure proper kernel enablement on ppc64 platforms, allowing gemm and reorder kernels to compile/run. Commits d42cbaab2c62dc27707ee624d76289678953389e and debc5f954b4be9b70d728d0a43646218567490b1. Overall impact: stronger correctness and performance potential in quantized paths, improved reliability and maintainability from reduced zero-point complexity, greater parser robustness, and expanded platform support through PPC64 kernel enablement.
8 Commits • 2 Features
Jun 1, 2025June 2025 (2025-06) highlights: Delivered high-impact improvements in oneDNN focused on quantization accuracy, zero-point handling, parser robustness, and cross-platform kernel enablement. Key features and fixes delivered across oneapi-src/oneDNN: - Quantization Scale Handling Improvements for brgemm Post-Ops: Consolidates and optimizes the application of scales in brgemm post-ops and related post-work, leveraging the with_scales flag for correct and efficient behavior. Commit trail includes 1a274be0d448f14a05ad570ae5374332119f58a7, eabd14913b84ef6f6aa3ccbff7f2c07ae9fa76dc, and 8820cfffb97ede7219bbd7d8de47183972574ccf. - Zero-point Handling Cleanup Across Convolution, MatMul, and Reorder: Refactors zero-point data access and propagation to a consistent CTX_IN_MEM-based approach, and removes redundant zero-point handling in f32 reorder. Commit trail includes 495dff5a1109a6f94575d6f0382b9e6ae87281c7 and 9e492c16e75cd89c05d674391448ae424ca75fe1. - Benchdnn Parser Robustness Against Undefined Behavior: Fixes string searching in benchdnn parser by replacing direct string::find with a robust matching helper to prevent UB. Commit 6bb6bba8522fcbed26b3f557bad37f8292cc5e7e. - PPC64 Macro Check Fixes to Enable Kernel Support: Corrects DNNL_PPC64 macro checks to ensure proper kernel enablement on ppc64 platforms, allowing gemm and reorder kernels to compile/run. Commits d42cbaab2c62dc27707ee624d76289678953389e and debc5f954b4be9b70d728d0a43646218567490b1. Overall impact: stronger correctness and performance potential in quantized paths, improved reliability and maintainability from reduced zero-point complexity, greater parser robustness, and expanded platform support through PPC64 kernel enablement.
June 2025
May 2025
22 Commits • 6 Features
May 1, 2025May 2025 monthly summary for oneapi-src/oneDNN. This month focused on delivering robust convolution features, stabilizing benchdnn benchmarking, and hardening core APIs and tests to improve reliability, performance, and scalability across CPU backends.
May 2025
22 Commits • 6 Features
May 1, 2025May 2025 monthly summary for oneapi-src/oneDNN. This month focused on delivering robust convolution features, stabilizing benchdnn benchmarking, and hardening core APIs and tests to improve reliability, performance, and scalability across CPU backends.
April 2025
53 Commits • 24 Features
Apr 1, 2025April 2025 (2025-04) performance/engineering summary for oneDNN (oneapi-src/oneDNN): Key features delivered: - CPU: conv_list: add x8:s8:f16 combination to enable higher-performance convolution paths on CPU. - BenchDNN matmul: allow a single scale/zp group of any size, with associated non-dense output handling for improved kernel flexibility and accuracy. - BenchDNN benchmarking improvements: prim_ref benchmarking now uses tag::any for binary and f32 for sum pointwise ops; quant_entry_t introduced with refactored arg_scales_t for clearer typedefs. - BenchDNN memory and utilities: added fast-access get/set interfaces for f32 memories, support for prefilling underlying buffers, and consolidation of utility behaviors (e.g., extra reorder reliance for non-f32 ref mem in comparisons). - CPU/x64 and build reliability improvements: added verbose messages to jit_reorder for easier debugging; brgemm_reorder path now reports status; strengthened macro guards to avoid Wundef hits and enabled Wundef in build; numerous readability/modernization commits across CPU code. - Miscellaneous quality improvements: GTests performance tidies, GPU/Graph tidy efforts, and overall cleanup of build-system issues (clang-tidy cleanup, header fixme removal). Major bugs fixed: - BenchDNN mem_check: account second compare tensor when prim_ref is used. - CPU x64: matmul: brgemm_reorder: add status check to catch errors early. - BenchDNN graph: fix false positives by reducing the matmul range. - Macro-related reliability: fix undefined macro hits across components (third_party ITT, dnnl_thread workaround, ukernel, aarch64). - General declaration/scope issues: moved declarations to proper spots to improve compile-time reliability. Overall impact and accomplishments: - Expanded kernel support and benchmarking fidelity, enabling more accurate performance evaluation and broader data-type coverage. - Improved reliability, debuggability, and maintainability through targeted fixes, modernization, and build-system hygiene. - Reduced risk of runtime and compile-time issues via explicit status checks and macro guards, contributing to more robust releases. Technologies/skills demonstrated: - C/C++ performance engineering, CPU/x64 kernel tuning, and BenchDNN optimization. - Memory interface design (fast-access get/set) and memory management strategies. - Build-system hygiene (clang-tidy cleanup, header fixme removal) and macro guard strategies (Wundef). - Debugging/monitoring enhancements (verbose jit_reorder messages, status checks) to improve maintainability.
53 Commits • 24 Features
Apr 1, 2025April 2025 (2025-04) performance/engineering summary for oneDNN (oneapi-src/oneDNN): Key features delivered: - CPU: conv_list: add x8:s8:f16 combination to enable higher-performance convolution paths on CPU. - BenchDNN matmul: allow a single scale/zp group of any size, with associated non-dense output handling for improved kernel flexibility and accuracy. - BenchDNN benchmarking improvements: prim_ref benchmarking now uses tag::any for binary and f32 for sum pointwise ops; quant_entry_t introduced with refactored arg_scales_t for clearer typedefs. - BenchDNN memory and utilities: added fast-access get/set interfaces for f32 memories, support for prefilling underlying buffers, and consolidation of utility behaviors (e.g., extra reorder reliance for non-f32 ref mem in comparisons). - CPU/x64 and build reliability improvements: added verbose messages to jit_reorder for easier debugging; brgemm_reorder path now reports status; strengthened macro guards to avoid Wundef hits and enabled Wundef in build; numerous readability/modernization commits across CPU code. - Miscellaneous quality improvements: GTests performance tidies, GPU/Graph tidy efforts, and overall cleanup of build-system issues (clang-tidy cleanup, header fixme removal). Major bugs fixed: - BenchDNN mem_check: account second compare tensor when prim_ref is used. - CPU x64: matmul: brgemm_reorder: add status check to catch errors early. - BenchDNN graph: fix false positives by reducing the matmul range. - Macro-related reliability: fix undefined macro hits across components (third_party ITT, dnnl_thread workaround, ukernel, aarch64). - General declaration/scope issues: moved declarations to proper spots to improve compile-time reliability. Overall impact and accomplishments: - Expanded kernel support and benchmarking fidelity, enabling more accurate performance evaluation and broader data-type coverage. - Improved reliability, debuggability, and maintainability through targeted fixes, modernization, and build-system hygiene. - Reduced risk of runtime and compile-time issues via explicit status checks and macro guards, contributing to more robust releases. Technologies/skills demonstrated: - C/C++ performance engineering, CPU/x64 kernel tuning, and BenchDNN optimization. - Memory interface design (fast-access get/set) and memory management strategies. - Build-system hygiene (clang-tidy cleanup, header fixme removal) and macro guard strategies (Wundef). - Debugging/monitoring enhancements (verbose jit_reorder messages, status checks) to improve maintainability.
April 2025
March 2025
60 Commits • 17 Features
Mar 1, 2025March 2025 monthly summary for oneDNN benchdnn contributions (oneapi-src/oneDNN). Focused on stability for sparse workloads, memory safety, and maintainability enhancements, alongside performance/measurement improvements to support optimization efforts. Key outcomes include bug fixes in matmul, enhanced memory handling and protection, and refactors across CPU x64 ukernel and API layers.
March 2025
60 Commits • 17 Features
Mar 1, 2025March 2025 monthly summary for oneDNN benchdnn contributions (oneapi-src/oneDNN). Focused on stability for sparse workloads, memory safety, and maintainability enhancements, alongside performance/measurement improvements to support optimization efforts. Key outcomes include bug fixes in matmul, enhanced memory handling and protection, and refactors across CPU x64 ukernel and API layers.
February 2025
56 Commits • 27 Features
Feb 1, 2025February 2025 oneDNN monthly summary focusing on quantization API improvements, benchdnn robustness, CI automation, and cross-platform stability. Key work included introducing quant_entry_t and refactoring arg_scales_t to support a unified quantization path, broad benchdnn improvements (regression tests, memory protections, shape handling), and several stability fixes across CPU x64, AMD, and non-x64 platforms. These changes enhance quantization accuracy and consistency, kernel reliability, CI maturity, and developer productivity, delivering tangible business value through improved performance, easier maintenance, and faster integration cycles.
56 Commits • 27 Features
Feb 1, 2025February 2025 oneDNN monthly summary focusing on quantization API improvements, benchdnn robustness, CI automation, and cross-platform stability. Key work included introducing quant_entry_t and refactoring arg_scales_t to support a unified quantization path, broad benchdnn improvements (regression tests, memory protections, shape handling), and several stability fixes across CPU x64, AMD, and non-x64 platforms. These changes enhance quantization accuracy and consistency, kernel reliability, CI maturity, and developer productivity, delivering tangible business value through improved performance, easier maintenance, and faster integration cycles.
February 2025
January 2025
60 Commits • 22 Features
Jan 1, 2025January 2025 monthly summary for oneDNN: key features delivered, major bugs fixed, overall impact and technologies demonstrated. Delivered AVX512_CORE/AVX2 BRGEMM/Matmul f32:f16 and f32:bf16 support; BenchDNN usability and graph enhancements (graph styling, F32 path for custom ops, bia-dt support, binary filling updates, option removal, and DNNL_VERBOSE OFF early exit); introduced quant_entry_t and relocated zero_points to stabilize quantization flow; strengthened testing and safety with GTest and signature improvements; and improved reliability and governance with ISA fixes, ldb check fix, default_attr relocation, and governance tooling.
January 2025
60 Commits • 22 Features
Jan 1, 2025January 2025 monthly summary for oneDNN: key features delivered, major bugs fixed, overall impact and technologies demonstrated. Delivered AVX512_CORE/AVX2 BRGEMM/Matmul f32:f16 and f32:bf16 support; BenchDNN usability and graph enhancements (graph styling, F32 path for custom ops, bia-dt support, binary filling updates, option removal, and DNNL_VERBOSE OFF early exit); introduced quant_entry_t and relocated zero_points to stabilize quantization flow; strengthened testing and safety with GTest and signature improvements; and improved reliability and governance with ISA fixes, ldb check fix, default_attr relocation, and governance tooling.
December 2024
7 Commits • 2 Features
Dec 1, 2024Monthly summary for December 2024 (oneDNN repo): Delivered two major feature enhancements that improve CPU path performance, robustness, and observability, laying a stronger foundation for future optimizations. Key outcomes include enabling relaxed accumulation mode for CPU softmax with improved dispatch and verbose debugging outputs, along with clarified error messages and initialization information display. In addition, internal hashing and debuginfo foundations were strengthened through a refactor to op_desc_t::to_desc, verbose seed logging, and richer debuginfo (filename and line number) with zero-initialization of opdesc members. These changes reduce debugging time, improve numerical robustness, and enhance maintainability for long-term performance work. Note: No explicit bug fixes were required in this period; the focus was on feature delivery and code quality improvements that improve observability and correctness of the CPU backend.
7 Commits • 2 Features
Dec 1, 2024Monthly summary for December 2024 (oneDNN repo): Delivered two major feature enhancements that improve CPU path performance, robustness, and observability, laying a stronger foundation for future optimizations. Key outcomes include enabling relaxed accumulation mode for CPU softmax with improved dispatch and verbose debugging outputs, along with clarified error messages and initialization information display. In addition, internal hashing and debuginfo foundations were strengthened through a refactor to op_desc_t::to_desc, verbose seed logging, and richer debuginfo (filename and line number) with zero-initialization of opdesc members. These changes reduce debugging time, improve numerical robustness, and enhance maintainability for long-term performance work. Note: No explicit bug fixes were required in this period; the focus was on feature delivery and code quality improvements that improve observability and correctness of the CPU backend.
December 2024
November 2024
29 Commits • 14 Features
Nov 1, 2024November 2024 (oneapi-src/oneDNN) delivered a focused set of performance improvements, safety fixes, API/kernel refinements, and maintainability upgrades across Benchdnn and CPU x64 paths. This work increases benchmarking accuracy, stabilizes perf-mode paths, enhances correctness in matmul/brgemm, and improves build hygiene, delivering tangible business value through faster, more reliable benchmarks and easier maintenance for contributors.
November 2024
29 Commits • 14 Features
Nov 1, 2024November 2024 (oneapi-src/oneDNN) delivered a focused set of performance improvements, safety fixes, API/kernel refinements, and maintainability upgrades across Benchdnn and CPU x64 paths. This work increases benchmarking accuracy, stabilizes perf-mode paths, enhances correctness in matmul/brgemm, and improves build hygiene, delivering tangible business value through faster, more reliable benchmarks and easier maintenance for contributors.
October 2024
3 Commits • 1 Features
Oct 1, 2024October 2024: Focused improvements on oneDNN to extend data-type support, strengthen correctness for low-precision paths, and stabilize GPU/CPU workflows. Key work delivered expands data-type transformation, improves post-ops and fusion correctness, and enhances reliability for int4/bf16/f16 workloads across CPU and GPU.
3 Commits • 1 Features
Oct 1, 2024October 2024: Focused improvements on oneDNN to extend data-type support, strengthen correctness for low-precision paths, and stabilize GPU/CPU workflows. Key work delivered expands data-type transformation, improves post-ops and fusion correctness, and enhances reliability for int4/bf16/f16 workloads across CPU and GPU.
October 2024
Activity
Loading activity data...
Quality Metrics
Correctness89.6%
Maintainability88.4%
Architecture85.2%
Performance81.4%
AI Usage20.2%
Skills & Technologies
Programming Languages
AssemblyCC++CMakeCmakeMarkdownOpenCLPythonShellTOML
Technical Skills
AMXAMX InstructionsAPI DesignAPI DevelopmentAPI IntegrationAPI designAPI integrationAVX InstructionsAVX-512AVX2 VNNIAVX512AbstractionAlgorithm ImplementationAssemblyAssembly (JIT)
Repositories Contributed To
Overview of all repositories you've contributed to across your timeline