An implementation of CUDA on top of the Vulkan Compute API.

• vudurt.so: Shared library that provides cuda-compatible symbols, such as cudaMalloc(), cudaFree(), cudaMemcpy(), etc.
• vuducc: Compiler which takes in .cu source files and does the following steps:
  • Generate PTX assembly for device code by invoking nvcc.
  • Split host code out and rewrite kernel invocation syntax foo<<<...>>>().
  • Transpile PTX assembly into SPIR-V compute code.
  • Generate kernel invocations which run the SPIR-V kernel using Vulkan.