OpenCL on Metal #2 – what if clvk works?

For April Fools day, an alternate approach is discussed for the OpenCL on Metal project.

Writing an OpenCL implementation from scratch would take a long time – what if we can leverage existing bricks for a proof of concept?

CLVK is an experimental, limited OpenCL implementation that runs on top of Vulkan. It uses clspv, a compiler to transform OpenCL C code into Vulkan compute shaders, with some important limitations.

MoltenVK is a Vulkan (Portability) implementation targeting Apple operating systems, allowing to run Vulkan implementations on top of Metal.

$ cd ~/VulkanSDK/1.3.204.1/ && source setup-env.sh && cd -
~/devlab/opencl/clvk
$ git clone --recurse https://github.com/kpet/clvk
[...]
$ cd clvk/external/clspv
$ git diff # small change to use python3
diff --git a/utils/fetch_sources.py b/utils/fetch_sources.py
index f0ba884..a0d6a35 100755
--- a/utils/fetch_sources.py
+++ b/utils/fetch_sources.py
@@ -1,4 +1,4 @@
-#!/usr/bin/env python
+#!/usr/bin/env python3
 
 # Copyright 2017 The Clspv Authors. All rights reserved.
 #
$ cd ../.. ; ./external/clspv/utils/fetch_sources.py --deps llvm
[...]
$ mkdir build && cd build
$ cmake ..
$ make -j8

Does it work at first glance? Yes it does.

$ ./clpeak # linking to libOpenCL.dylib instead of OpenCL.framework
Platform: clvk
  Device: Apple M1
    Driver version  : 1.2 CLVK on Vulkan v1.1.189 driver 10105 (Macintosh)
    Compute units   : 1
    Clock frequency : 0 MHz

    Global memory bandwidth (GBPS)
      float   : 55.63
      float2  : 57.66
      float4  : 57.50
      float8  : 57.88
      float16 : 62.41

    Single-precision compute (GFLOPS)
      float   : 1267.18
      float2  : 1502.12
      float4  : 1525.60
      float8  : 892.10
      float16 : 1470.11

    Half-precision compute (GFLOPS)
      half   : 1333.62
      half2  : 1508.59
      half4  : 1591.51
      half8  : 1525.85
      half16 : 1423.71

    No double precision support! Skipped

    Integer compute (GIOPS)
      int   : 472.07
      int2  : 467.69
      int4  : 469.41
      int8  : 476.42
      int16 : 461.98

    Integer compute Fast 24bit (GIOPS)
      int   : 480.22
      int2  : 478.74
      int4  : 437.78
      int8  : 474.67
      int16 : 475.28

    Transfer bandwidth (GBPS)
      enqueueWriteBuffer              : 28.62
      enqueueReadBuffer               : 28.93
      enqueueWriteBuffer non-blocking : 28.99
      enqueueReadBuffer non-blocking  : 28.97
      enqueueMapBuffer(for read)      : 550636.56
        memcpy from mapped ptr        : 28.75
      enqueueUnmap(after write)       : 727960.19
        memcpy to mapped ptr          : 28.96

    Kernel launch latency : 6.92 us