CUTLASS / CuTe façade + optional TF32

CutlassAccel is the GPU linear-algebra entry for MathNN-sized work.

Full TF32 guide: TF32MatrixMultiply.md
VS configs: BuildConfigurations.md

Status

VS Configuration Manager ↔ TF32

msbuild /p:Configuration=Release-GPU-AVX2              # TF32 compiled in
msbuild /p:Configuration=Release-GPU-AVX2 /p:EnableTf32=false
msbuild /p:Configuration=Release /p:EnableTf32=true    # opt-in on plain Release

Default runtime precision remains fp32 even when TF32 is compiled.

Commands

cutlass status                 # tf32_compiled= / tf32_available=
cutlass bench [size] [repeats]
cutlass check [size]           # TF32 probe only if AI_USE_TF32
cutlass set precision fp32|tf32|fp16|bf16
cutlass gemm [M N K]

Alias: cute ….

API

_CutlassAccel::IsTf32Compiled();   // AI_USE_TF32 in this binary?
_CutlassAccel::IsTf32Available();  // compiled + GPU SM>=8?
_CutlassAccel::SetPrecision(Precision::Tf32); // no-ops to Fp32 if not compiled
_CutlassAccel::Gemm(A, B, C, M, N, K, detail);

Lua

cutlass_set_precision("tf32")  -- becomes fp32 if not compiled
cutlass_check(64)