When Apple switched to Intel in 2006, all manufacturers started writing EFI firmwares rather than OpenFirmware. OpenCL support started with the GF 8XXX series and Radeon HD 4XXX. Having OpenCL for PPC SL would have meant keeping on writing OF roms on top of EFI, which probably proved not worth it. EFI being little-endian, this probably proved easier than big-endian OF which limited the availability of models on the PM platform. So overall porting OpenCL to PPC would have made little sense.
Ironically, stock Radeon HD cards are supported under ppc Linux. I have one at home (HD 5770), and it works pretty well on Lubuntu 14.04 powerpc (strangely not 16.04 in my hands). And ironically, there are ppc/ppc64 opencl Linux ports. But these builds came later, otherwise Apple could have done it easily. That said, it could have been done, even for AGP. The Radeon HD 4670 AGP for example supports OpenCL (I used it on x86 with BOINC for OpenCL computations). Great card, incidentally.
Does anyone know if 10A96 or 10A190 implement OpenCL? If the 10A428 and 10A430 releases do exist, it is very likely that they would have supported OpenCL, being so close to 10A432. If not, it would probably be some undertaking to port ppc/ppc64 opencl from Linux.
In 10A190, via Pacifist, the version of OpenCL was more of a placeholder by Apple than anything functional. In 2008, OpenCL was still being finalized as an industry standard, and Apple was prepared for this eventuality:
In 10A96, there are very early elements of OpenCL in a Framework which, as memory serves, was very incomplete (my test build of 10A96 includes a later version of OpenCL — a function of the testing which helped me arrive to the conclusion about OpenCL being Intel-only and applicable to only newer GPUs whose drivers were not written in BrookGPU).
At the time of Build 10A96, 4–8 June 2008, Apple had not yet submitted its OpenCL proposal to the industry working group, Khronos, and wouldn’t do so until
16 June 2008, per Wikipedia. When Apple released Build 10A190 in October 2008, the version 1.0 draft of OpenCL was still being worked on and wouldn’t be published until 18 November 2008.
So in short: the OpenCL framework bundled with either of 10A96 or 10A190 are placeholders and they are likely non-functional for OpenCL-aware GPUs. More specifically, Apple developed OpenCL alongside Snow Leopard.
And this is indeed where the buck stops. It would have been far easier to port OpenCL to Intel OSX (e.g. from Linux x86 where very little work would have been required) for Apple than to PPC. Besides, none of the PCIe GPUs on the PM G5 supported OpenCL (e.g. GF 7800 GTX, Radeon X1900 GT with OpenFirmware roms), let alone AGP ones on earlier ones and on G4s.
In practice, it was probably too great a request (at any price) for Apple to compel AMD or NVIDIA to re-write drivers for their older GPUs to be OpenCL-aware, as the hardware itself (and, by extension, the firmware) was designed for BrookGPU-written drivers. They would probably have needed to write some kind of layer or shim to translate to BrookGPU, which was probably too much work for the limited outcome (and which might have slowed GPU performance in high-demand circumstances, such as gaming).