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

Full TF32 / matmul guide: TF32MatrixMultiply.md (TF32MatrixMultiply.md)

Lua bindings: LuaBrainstormAPI.md (LuaBrainstormAPI.md)

Chat feature index: ChatSession-Features.md (ChatSession-Features.md)

Status

Mode	Behavior
Default GPU (`USE_CUDA`)	cuBLAS `GemmEx` — FP32 or TF32 (CUBLAS_COMPUTE_32F_FAST_TF32, SM≥8)
No GPU / cuBLAS fail	Host MatrixAcceleration AVX/AVX-512 GEMM
`AI_USE_CUTLASS` + device .cu	Optional future MMA kernels (same API)

GPU MathNNDevice::MatMul calls _CutlassAccel::Gemm first.

Commands

[text]
cutlass status
cutlass bench [size] [repeats]
cutlass check [size]
cutlass set precision fp32|tf32|fp16|bf16
cutlass gemm [M N K]

Alias: cute ….

API

[cpp]
_CutlassAccel::Gemm(A, B, C, M, N, K, detail);
_CutlassAccel::LinearFused(W, x, bias, y, batch, in, out, relu, detail);
_CutlassAccel::SetPrecision(Precision::Tf32);

Lua

[lua]
cutlass_set_precision(“tf32”)
cutlass_gemm(128, 128, 128)
cutlass_check(64)
cutlass_bench(256, 3)

Correctness

• cutlass check: host golden vs façade; FP32 tol ~1e-3; TF32 probe tol ~5e-2. 
• Detail strings report cublas_tf32, cublas_fp32, or host fallback.

Next steps for real CUTLASS MMA

1. Add CutlassGemm.cu including CUTLASS/CuTe headers. 
2. Define AI_USE_CUTLASS on Tensor or a GPU config. 
3. Keep the same Gemm API; leave cuBLAS as default path. 
4. Keep cutlass check as golden vs AVX.

---

Filed under: Uncategorized - @ July 18, 2026 6:28 am