WebGL does automatically fall back to the CPU in Chrome, via SwiftShader[1]. WebGPU hasn't shipped yet but I think it's likely that a CPU backend with SwiftShader will be available eventually. And SwiftShader implements Vulkan now too, so while it's not automatic for Vulkan you can certainly have a fallback if you want it.
So you spend a lot of time in the talk massaging your sequential CPU task into an appropriately GPU task. It's clearly tricky and involves reworking your algorithms to leverage the parallelism and the memory "features" of the GPU. But through rewriting, have you actually substantively hurt the performance of the sequential program?
The big picture question is : is the reverse problem of going from a GPU program to a CPU problem ever problematic? I mean in a algorithmic sense - without looking at micro optimizations for leveraging CPU SIMD instructions or whatever. Or are you always going to be pretty okay with running your shaders sequentially one after another on your big luxurious CPU memory pool?
and ultimately is there anything stopping you from compiling SPIR-V for a CPU? Could you not autgenerate in the kernel dispatcher just a switch that'll branch and run a precompiled on-CPU kernel if no driver is found? Then you'd finally really get compile-once-run-everywhere GPU code
I guess since it's not being done then I'm missing something haha :) Maybe you are going to often hit scenarios where you'd say "No, if I'm going to run this on the CPU I need to fundamentally rewrite the algorithm"
If you have a workload that runs really well on GPU, then it can be adapted to run on CPU. There are automated tools for this, but it's not mainstream. To answer one of your questions, spirv-cross has a C++ generator.
There are a bunch of things that are in the space between GPU and traditional scalar CPU. One of the most interesting is Intel's ispc. The Larabee project was also an attempt to build some GPU features into a CPU, and that is evolving into AVX-512. The new mask features are particularly useful for emulating the bitmap of active threads, as this is something that's hard to emulate with traditional SIMD.
I think it would be a very interesting project to build a compiler from SPIR-V compute workloads to optimized multithreaded SIMD, and there are projects exploring that: https://software.intel.com/en-us/articles/spir-v-to-ispc-con...
The main reason this hasn't happened is that when doing optimization it's always better to target the actual hardware than go through layers of translation. If you want to run a neural network on CPU, you're definitely going to get better performance out of matrix multiply code tuned for the specific SIMD implementation than something adapted from GPU. But I think there may still be a niche, especially if it's possible to get "pretty good" results.
For machine learning and imaging workloads in particular, there's probably a lot more juice in having a high level (Halide) or medium level (MLIR) representation, and targeting both GPU and CPU as backends that do optimization specific to their targets.
I'm really interested to see how this space evolves, it feels like there are major opportunities, while the scalar CPU side feels largely mined out to me.
Intel GPU has three hidden strengths:
* CPU <-> GPU communication is cheaper, because they can actually share memory (a separate copy to staging buffers is not needed).
* Latency is potentially lower, although it's not clear to me yet that driver software can take advantage of the potential offered by the hardware. (More empirical measurement is needed)
* Subgroup operations on Intel appear to be wicked-fast, shuffle in particular.
Long story short, I think there is opportunity here that most people aren't exploiting.
Looking at SwiftShader - it looks like you can bundle it with your application? Is that correct? I don't see a way to bundle POCL though