What's Changed
MGPUSim now officially supports simulating HIP programs running on AMD GFX942 architecture (MI300a GPU)!
Here is a detailed list of merged PRs.
- single_core_example by @sajannat in #219
- Support HSACO V5 by @syifan in #241
- GFX942 Emulation Support by @syifan in #242
- Support multi warp emulation for gfx942 by @syifan in #243
- Timing simulation support for MI300A architecture by @syifan in #244
- Add MI300A unified multi-GPU support for vectoradd by @syifan in #245
- Gfx942 emu by @syifan in #249
- Scratchpad removal by @syifan in #250
- Fix MI300A timing model: DRAM controller, WfPoolSize, VGPR count by @syifan in #251
- Fix cache timing parameters for MI300A (L1V, VecMem, L2) by @syifan in #254
- Remove fixed latency, regression eval, HIP microbenchmarks, CI fixes by @syifan in #257
- Fix CI: pure-Go SQLite, no CGO/gcc needed by @syifan in #258
- Add simpleconvolution, FFT sub-MB, BFS sizes + timing model sync by @syifan in #259
- MI300A Calibration: 16.4% WMAPE Accuracy by @syifan in #260
Full Changelog: v4.1.5...v4.2.0
Contributors
syifan and sajannat