While the NVIDIA PR engines are trumpeting now about an open-source CUDA compiler, the news isn't exactly new. Back in April is when I wrote about NVIDIA contributing a new PTX back-end for LLVM -- the news announcement today is about this contribution to LLVM as their "open-source CUDA compiler."
In January was also their announcement of an open-source LLVM CUDA compiler.
From within the community is also the Gdev open-source CUDA run-time. There's also StarPU for CPU/GPU task programming with GCC and some work this summer on automatic GPGPU code generation with LLVM.
Below is what they wrote in the e-mail news announcement this morning about the open-source CUDA contribution. More details are to be revealed next week from NVIDIA's GPU conference.
The NVIDIA Compiler team have worked with the LLVM developers to provide key CUDA compiler source code changes to the LLVM core and parallel thread execution backend. The latest LLVM compiler with NVIDIA GPU support is now available for download from the LLVM site.
At GPU Technology Conference (GTC), next week, we will be presenting more details about the new compiler and releasing a preview of the LLVM-based CUDA Compiler SDK that includes specifications, libraries, and code samples showing how to interface with LLVM to produce GPU accelerated binaries from domain-specific languages (DSLs).