hodu_cuda_kernels
| Crates.io | hodu_cuda_kernels |
| lib.rs | hodu_cuda_kernels |
| version | 0.2.4 |
| created_at | 2025-11-10 18:58:51.391914+00 |
| updated_at | 2025-11-13 22:01:14.058028+00 |
| description | hodu cuda kernels |
| homepage | |
| repository | https://github.com/hodu-rs/hodu |
| max_upload_size | |
| id | 1925945 |
| size | 520,033 |
Han Damin (miniex)
documentation
README
hodu_cuda_kernels
High-performance CUDA kernels for tensor operations on NVIDIA GPUs.
cuBLAS Integration
Supported Operations
- matmul: Batched matrix multiplication with GEMM
- dot: 2D matrix multiplication with GEMM
Supported Data Types
- bf16: BFloat16 (compute in FP32, I/O in BF16)
- f16: Float16/Half (compute in FP32, I/O in FP16)
- f32: Float32 (native precision)
- f64: Float64 (native precision)
Features
- Automatic fallback to custom CUDA kernels for unsupported types or non-contiguous matrices
- Handles non-contiguous matrices via leading dimension parameters
- Transparent row-major to column-major layout conversion