[libre-riscv-dev] Vulkan vs OpenCL and Alternatives

[Jacob Lifshay]

On Fri, Mar 6, 2020, 12:33 Immanuel, Yehowshua U <yimmanuel3 at gatech.edu>
wrote:

> So one question that I’ve been thinking about is “how good is Vulkan” for
> GPGPU?
>
> That is, how easy is it to express thing like matmul and dot in Vulkan?
>

basically, just as easy as OpenCL, though slightly different. If you only
want small matrices (2x2 through 4x4 -- includes stuff like 3x4 or 4x2),
it's built directly into SPIR-V, otherwise you will need a library.

>
> How does it compare against OpenCL. From a practical standpoint, I think
> it makes more sense for us to support Vulkan since Apple, the original
> proponents of OpenCL withdrew their support from the standard a few years
> ago anyways.
>
> Also, will the LibreSOC “GPU” be able to execute OpenCL code? I imagine
> that we could compile OpenCL into SPIR-V and then perhaps have some SPIR-V
> driver evaluate the SPIR-V?
>

there's clspv -- an OpenCL C to Vulkan SPIR-V compiler -- Adobe used it to
port Photoshop to Android (if I recall correctly)

there's also clvk -- an experimental OpenCL to Vulkan translator written on
top of clspv

>
> I bring all this up because, even though I favor Vulkan, many applications
> are still being written in OpenCL, and from a business customer
> perspective, I imagine that many of our potential customers are probably
> more familiar with OpenCL  - at l-east more so than Vulkan…
>
> @Jacob - can I get your input on this - especially since you are working
> on the Vulkan drivers?
>
> Yehowshua
> _______________________________________________
> libre-riscv-dev mailing list
> libre-riscv-dev at lists.libre-riscv.org
> http://lists.libre-riscv.org/mailman/listinfo/libre-riscv-dev
>