Mirror of forum.nim-lang.org

3245 :: "the upcoming OpenCL target"

• monster (orginal) [2017-10-13T20:09:34+02:00] view original

  The Nim manual mentions "the upcoming OpenCL target". Is that still a coming? And what about Vulkan, which seems to be it's successor?

---

• mratsim (orginal) [2017-10-13T21:48:24+02:00] view original

  I bet 2 cents on Valve time ;)

---

• mratsim (orginal) [2017-10-14T00:30:04+02:00] view original

  Just stumbled upon this while looking for completely unrelated stuff:

  https://github.com/nim-lang/Nim/wiki/GSoC-2016-Ideas#add-a-code-generator-for-opencl

---

• monster (orginal) [2017-10-14T13:50:16+02:00] view original

  I assume that means nobody is working on it atm. Maybe using one of those two APIs would simplify things?

  • SYCL C++ Single-source Heterogeneous Programming for OpenCL
  • HIP Convert CUDA to Portable C++ Code (where "portable" means that it runs fast on AMD too)

  If I remember well, basically "everyone" is on the Vulkan bandwagon (NVidia, AMD, Microsoft, ...), except Apple(?). Vulkan will also be better at working on a wide variety of platforms (PC, consoles, mobile, ...) So it would probably make more sense to work on Vulkan instead, unless one is also targeting "older platforms". Not that I'm volunteering; I barely made it through my compiler course...

---

• mratsim (orginal) [2017-10-15T20:12:57+02:00] view original

  I will write OpenCL code and also HIP code in my deep learning library.

  Vulkan however is different from OpenCL and Cuda. Vulkan provides Graphics and Physics related primitives like OpenGL or GLSL shaders while OpenCL and Cuda provides you with a subset of C/C++ working on the GPU (loops, if/else, pointers, math operations ...)

---

• Stefan_Salewski (orginal) [2017-10-15T21:14:18+02:00] view original

  OpenCL and Cuda provides you with a subset

  From

  https://news.ycombinator.com/item?id=14383296

  I got the impression that Vulkan also supports that and may merge with OpenCL.

---

• monster (orginal) [2017-10-15T22:01:23+02:00] view original

  Exactly, Vulkan is meant to replace both OpenGL and OpenCL, in a single language. I suspect, in many case, someone using OpenGL is likely to use OpenCL (or CUDA) too. And they need to learn two different languages to do that. Vulkan will make the life of such devs easier (in theory).

---

• mratsim (orginal) [2017-10-16T09:28:37+02:00] view original

  Actually in the top comment of that article it says that the article author had a misundertanding.

  What will potentially be merged is the intermediate representation SPIR-V, not the languages on top.

  I did find something called "Vulkan Compute". However OpenCL can also run on CPU while Vulkan cannot (yet?).

  For now, neither Vulkan nor HIP are mature enough, in my opinion it would probably be easier to target LLVM-IR and let LLVM do the SPIR-V compilation.

---

• Krux02 (orginal) [2017-10-16T13:52:23+02:00] view original

  Vulkan introduced the Standard Portable Immediate Representation (SPIR-V). You can imagine this as some sort of assembler instructions that will be compiled and optimized to the final GPU instructions from the GPU vendor driver. SPIR-V is through extensions and the latest core version also available in OpenGL and OpenCL. So I think it is a good idea to target SPIR-V.

---