pub type DenseM16Bf16Gemm = fn(&dyn GpuBackend, KernelHandle, DevicePtr, &DenseWeight, DevicePtr, u32, u32, u32, u32, u32, u64) -> Result<()>;Expand description
The shared shape of both instantiations, so a caller that picks between them
(and between them and dense_gemv_batchm) can hold ONE function pointer and
the tile stays a dispatch choice rather than a code path.