{"id":31791353,"url":"https://github.com/teambipartite/bipartite-gemm","last_synced_at":"2026-04-17T06:32:24.747Z","repository":{"id":260227173,"uuid":"871784180","full_name":"TeamBipartite/bipartite-gemm","owner":"TeamBipartite","description":"High throughput data-parallel GEMM implementations in Cuda using Cuda cores and Tensor cores","archived":false,"fork":false,"pushed_at":"2025-09-19T03:24:45.000Z","size":870,"stargazers_count":0,"open_issues_count":0,"forks_count":0,"subscribers_count":0,"default_branch":"master","last_synced_at":"2025-09-19T05:39:56.026Z","etag":null,"topics":["cuda","data-parallelism","gemm"],"latest_commit_sha":null,"homepage":"","language":"C++","has_issues":true,"has_wiki":null,"has_pages":null,"mirror_url":null,"source_name":null,"license":null,"status":null,"scm":"git","pull_requests_enabled":true,"icon_url":"https://github.com/TeamBipartite.png","metadata":{"files":{"readme":"README.md","changelog":null,"contributing":null,"funding":null,"license":null,"code_of_conduct":null,"threat_model":null,"audit":null,"citation":null,"codeowners":null,"security":null,"support":null,"governance":null,"roadmap":null,"authors":null,"dei":null,"publiccode":null,"codemeta":null,"zenodo":null,"notice":null,"maintainers":null,"copyright":null,"agents":null,"dco":null,"cla":null}},"created_at":"2024-10-12T23:31:30.000Z","updated_at":"2025-09-19T03:24:49.000Z","dependencies_parsed_at":"2025-05-22T06:09:57.020Z","dependency_job_id":"d65ff899-a12e-4ab5-bbbb-f8621f494a79","html_url":"https://github.com/TeamBipartite/bipartite-gemm","commit_stats":null,"previous_names":["teambipartite/csc485b-202409-a2","teambipartite/csc485b-202409-a4","teambipartite/bipartite-gemm"],"tags_count":2,"template":false,"template_full_name":null,"purl":"pkg:github/TeamBipartite/bipartite-gemm","repository_url":"https://repos.ecosyste.ms/api/v1/hosts/GitHub/repositories/TeamBipartite%2Fbipartite-gemm","tags_url":"https://repos.ecosyste.ms/api/v1/hosts/GitHub/repositories/TeamBipartite%2Fbipartite-gemm/tags","releases_url":"https://repos.ecosyste.ms/api/v1/hosts/GitHub/repositories/TeamBipartite%2Fbipartite-gemm/releases","manifests_url":"https://repos.ecosyste.ms/api/v1/hosts/GitHub/repositories/TeamBipartite%2Fbipartite-gemm/manifests","owner_url":"https://repos.ecosyste.ms/api/v1/hosts/GitHub/owners/TeamBipartite","download_url":"https://codeload.github.com/TeamBipartite/bipartite-gemm/tar.gz/refs/heads/master","sbom_url":"https://repos.ecosyste.ms/api/v1/hosts/GitHub/repositories/TeamBipartite%2Fbipartite-gemm/sbom","scorecard":null,"host":{"name":"GitHub","url":"https://github.com","kind":"github","repositories_count":279004689,"owners_count":26083753,"icon_url":"https://github.com/github.png","version":null,"created_at":"2022-05-30T11:31:42.601Z","updated_at":"2022-07-04T15:15:14.044Z","status":"online","status_checked_at":"2025-10-10T02:00:06.843Z","response_time":62,"last_error":null,"robots_txt_status":"success","robots_txt_updated_at":"2025-07-24T06:49:26.215Z","robots_txt_url":"https://github.com/robots.txt","online":true,"can_crawl_api":true,"host_url":"https://repos.ecosyste.ms/api/v1/hosts/GitHub","repositories_url":"https://repos.ecosyste.ms/api/v1/hosts/GitHub/repositories","repository_names_url":"https://repos.ecosyste.ms/api/v1/hosts/GitHub/repository_names","owners_url":"https://repos.ecosyste.ms/api/v1/hosts/GitHub/owners"}},"keywords":["cuda","data-parallelism","gemm"],"created_at":"2025-10-10T16:56:24.345Z","updated_at":"2025-10-10T16:56:26.534Z","avatar_url":"https://github.com/TeamBipartite.png","language":"C++","funding_links":[],"categories":[],"sub_categories":[],"readme":"# BipartiteGEMM\n\n High throughput data-parallel GEMM implementations in Cuda using Cuda cores and\n Tensor cores.\n\n# Prerequisites\nYou should have a CUDA environment installed with a GPU of compute\ncapability 7.5 or higher. Our code has been tested on both the Turing and Ampere\narchitectures.\n\nIf using the included benchmark/verification program, OpenBLAS is helpful for checking correctness of output, but not necessary.\nSee the 'Build' section for more details.\n\n# Build\nA makefile is provided in  the top-level directory which handles building the application. \nThe default target is `sm_75` (eg, Compute Capability 7.5 graphics cards such as the Turing T4).\n\n```bash\n$ make\n```\nis sufficient to build the library with default options.\n\n## Build options\nTo specify the `arch` string for your target, use the `TARGET` variable when\ncalling `make`. For example, for a Compute Capability 8.6 graphics card such as the RTX 3060, use:\n```bash\n$ make TARGET=sm_86\n```\n\nTo configure the tests ran by `bench`, the following options may also be passed:\n\n * `NUM_SMS=n` For accurate GFLOPs per SM calculations, this defines the number\n    of SMs to consider. default=28\n * `TEST_N=n` Input size to use for test calculations. Note that this may be\n   padded as required by the individual library functions. default=4096\n * `TEST_MAX_ELEMENT=n` For library functions that operate on numerical data,\n   specifies the maximum element to (randomly) generate for test data.\n   default=1\n * `CXXVERSION=20`: if compiling with CUDA \u003c 12, you can set this to `17` to\n   compile using only `C++17` features.\n\nThe following options may also be specified to configure use of the OpenBLAS\nlibrary in `bench`:\n\n* `USE_OPENBLAS=yes`: By default, the OpenBLAS library is called to\n  perform a CPU matrix multiplication to serve as a baseline to check for\n  correctness. If you do not have OpenBLAS on your system, set this variable\n  to `no` to use a naive provided n^3 CPU implementation.\n* `OPENBLAS_NUM_THREADS=$(nproc)`: If `USE_OPENBLAS=yes`, this option can be\n  specified to reduce the number of threads used by OpenBLAS.\n\nA clean is required before switching configurations.\n\n# Install \u0026 run\nSimply copy the `bipartite-gemm` directory to your system's `include`\ndirectory to install the library for use in other applications. The interface is\nprovided within the `bipartite` namespace.\n\nA `bench` executable is generated in the `build` directory for testing and verification purposes. \nSimply run this file to both verify output correctness and benchmark\nthe library - it does not need to be installed to run. Note that the test parameters must be passed at compile time using\nthe options specified in the `build` section of this document - this is to\nensure best performance possible through hte use of compile-time constants.\n\nFor debugging and output inspection, `bench` provides a `-p` argument. When this\nargument is provided, both the expected and actual outputs is printed.\n\n**PRECISION NOTE:** We provide an FP16 version of our tensor matrix multiplication,\nbut since FP16 only has a 10-bit mantissa, it can be quite inaccurate for larger\nmatrices (or matrices with large values). We use a fixed epsilon (currently set\nto `0.00001`) for output checking. This works on the provided example, but if the\nparameters in `main.cu` are changed, **the test may report a fail for the FP16 runs\neven though the output is as accurate as practical for FP16**.\n","project_url":"https://awesome.ecosyste.ms/api/v1/projects/github.com%2Fteambipartite%2Fbipartite-gemm","html_url":"https://awesome.ecosyste.ms/projects/github.com%2Fteambipartite%2Fbipartite-gemm","lists_url":"https://awesome.ecosyste.ms/api/v1/projects/github.com%2Fteambipartite%2Fbipartite-gemm/lists"}