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