matmul, the mesh's exact-mode GPU GEMM and GEMV (bit-identical on every GPU), as PTX from C
$ sexc add @ajax/mesh-prog-matmul- 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 ^0.1.0, @ajax/mesh-rt-linalg ^0.1.0
- dependents
- none yet
- keywords
- mesh alphaos gpu ptx program gemm deterministic
- homepage
- https://github.com/thomasdavis/alpha-c-mesh/tree/main/programs/matmul
- repository
- https://github.com/thomasdavis/alpha-c-mesh
1# @ajax/mesh-prog-matmul 2 3The exact-mode GPU GEMM and GEMV: 64 x 64 tiles of C on 8-warp nodes (@ajax/mesh-rt-linalg), y = A x a 4warp a row, the matrices generated on the device in the program's scratch and copied out through the 5view's window. Every output a sequential sum in k from +0.0, unfused and FTZ, so the bits equal the CPU 6reference on every GPU. 7 8- **Ported from:** alpha-rust-mesh `programs/matmul/src/lib.rs`. 9- **Tests:** in @ajax/mesh-prog-matmul-check (the reference is host-only): on the simulator against 10 @ajax/mesh-ref-matmul, bitwise, for dimensions off the tile grid, and bad arguments and node shapes refused.