Collection of highly optimized ROCm HIP kernels for common GPU operations on AMD CDNA architecture.
- GEMM — Matrix multiplication with 95%+ TFLOPS utilization
- Flash Attention — Memory-efficient attention for CDNA3
- Softmax — Fused online softmax kernel
- LayerNorm — Fused layer normalization + residual
- RoPE — Rotary positional embeddings
- TopK — Parallel top-k selection
| Kernel | TFLOPS | Peak % |
|---|---|---|
| GEMM FP16 | 5.02 | 94.7% |
| Flash Attn | 4.81 | 90.8% |
| Softmax | 3.20 | 60.4% |
mkdir build && cd build
cmake -DROCM_PATH=/opt/rocm ..
make -j$(nproc)
make test