I built a lightweight (93KB) CUDA→AMD translation layer using LD_PRELOAD.
It intercepts CUDA API calls at runtime and translates them to HIP/rocBLAS/MIOpen.
No source code needed. No recompilation. Just:
LD_PRELOAD=./libapex_hip_bridge.so ./your_cuda_app
Currently supports:
- 38 CUDA Runtime functions
- 15+ cuBLAS operations (matrix multiply, etc)
- 8+ cuDNN operations (convolutions, pooling, batch norm)
- PyTorch training and inference
Built in ~10 hours using dlopen/dlsym for dynamic loading. 100% test pass rate.
The goal: break NVIDIA's CUDA vendor lock-in and make AMD GPUs viable for
existing CUDA workloads without months of porting effort.
It intercepts CUDA API calls at runtime and translates them to HIP/rocBLAS/MIOpen.
No source code needed. No recompilation. Just:
Currently supports:- 38 CUDA Runtime functions
- 15+ cuBLAS operations (matrix multiply, etc)
- 8+ cuDNN operations (convolutions, pooling, batch norm)
- PyTorch training and inference
Built in ~10 hours using dlopen/dlsym for dynamic loading. 100% test pass rate.
The goal: break NVIDIA's CUDA vendor lock-in and make AMD GPUs viable for
existing CUDA workloads without months of porting effort.