Exact-mode GEMM tile and GEMV at mesh node scope: sequential sums, unfused, bit-identical on every GPU
$ sexc add @ajax/mesh-rt-linalg- total 3
- drwxr-xr-x-../
- drwxr-xr-x-Oct 27 03:26v/# 1 version
- drwxr-xr-x-Oct 27 03:26files/# the files of 0.1.0
- latest
- 0.1.0, published Oct 27 03:26 by ajax
- owners
- ajax
- license
- MIT
- dependencies
- @ajax/mesh-rt-block ^0.1.0, @ajax/mesh-rt-ctx ^0.1.0
- dependents
- @ajax/mesh-prog-gpt2x, @ajax/mesh-prog-matmul
- keywords
- mesh alphaos gpu linear-algebra gemm deterministic
- homepage
- https://github.com/thomasdavis/alpha-c-mesh/tree/main/packages/mesh-rt-linalg
- repository
- https://github.com/thomasdavis/alpha-c-mesh
1# @ajax/mesh-rt-linalg 2 3Dense linear algebra at node scope in the exact mode: `mesh_linalg_gemm_tile` (a 64 x 64 tile of C on a 4256-thread node, k in slabs of 16 through shared memory) and `mesh_linalg_gemv_cols` (64 columns of one 5row, slabs of B staged through registers into shared memory). Every output is the sequential sum over k 6from +0.0 of a correctly rounded multiply and add, never fused, so the bits do not depend on tile shape, 7node or GPU. 8 9- **Ported from:** alpha-rust-mesh `crates/mesh-rt/src/linalg.rs`. 10- **Tests:** programs/matmul compares its tiles and rows with @ajax/mesh-ref-matmul on the simulator.