Unable to run on macOS (x86-64) AMDGPU: no atomic_uint

[Artoria2e5]

20251120 17:12:06  PRPLL 0.15-154-g6c16f7e-dirty starting
20251120 17:12:06  config: -tune -d 1
20251120 17:12:06  device 1, OpenCL 1.2 (Jul 14 2025 21:13:32), unique id ''
20251120 17:12:07  In file included from <program source>:3:
carryfused.cl:144:34: error: use of undeclared identifier 'atomic_uint'
    if (me == 0) { atomic_store((atomic_uint *) &ready[gr], 1); }
                                 ^
carryfused.cl:144:47: error: expected expression
    if (me == 0) { atomic_store((atomic_uint *) &ready[gr], 1); }
                                              ^
carryfused.cl:172:62: error: use of undeclared identifier 'atomic_uint'
  if (me == 0) { do { spin(); } while(!atomic_load_explicit((atomic_uint *) &ready[gr - 1], memory_order_relaxed, memory_scope_device)); }
                                                             ^
carryfused.cl:172:75: error: expected expression
  if (me == 0) { do { spin(); } while(!atomic_load_explicit((atomic_uint *) &ready[gr - 1], memory_order_relaxed, memory_scope_device)); }
                                                                          ^

20251120 17:12:07  Compiling 'carryfused.cl' error BUILD_PROGRAM_FAILURE (-11) (args -w -cl-finite-math-only -cl-std=CL2.0  -DIN_SIZEX=8 -DIN_WG=64 -DEXP=4980727u -DWIDTH=256u -DSMALL_HEIGHT=256u -DMIDDLE=2u -DCARRY_LEN=8u -DNW=4u -DNH=4u -DWEIGHT_STEP=2.3797603047682418e-05 -DIWEIGHT_STEP=-2.3797036735248459e-05 -DFFT_VARIANT=101u -DTAILT=U2(-7.5298160855459062e-05,0.012271538285719925) -DTRIG_SCALE=9 -DTRIG_SIN={5.3263221801584762e-06,2.9767495243664746e-22,-2.5184367464740274e-17,3.5723657515182318e-29,-2.4130231682795684e-41,9.5078818943081101e-54,-2.4520163974664508e-66,4.4149667939391264e-79,} -DTRIG_COS={1,-1.4184853983424073e-11,3.3535013755177081e-23,-3.1712618229987226e-35,1.6065673539660592e-47,-5.0642051365037344e-60,1.0883867404577732e-72,-1.6846025309095688e-85,} -DFRAC_BPW_HI=4294819839u -DFRAC_BPW_LO=4294967295u -DFRAC_BITS_BIGSTEP=3221225471u )
20251120 17:12:07  Can't compile carryfused.cl
20251120 17:12:07  Exiting because "Can't compile carryfused.cl"
20251120 17:12:07  Bye

Device list

 N  : PCI BDF |   UID            |   Driver                 |    Device
 0  :         |                  | 1.1                      | Intel(R) Core(TM) i7-8086K CPU @ 4.00GHz |
 1  :         |                  | 1.2 (Jul 14 2025 21:13:32) | AMD Radeon RX 6800 XT Compute Engine |

The original output was filled with warnings with no sign of the error (macOS truncates the output by itself). I had to edit src/KernelCompiler.cpp to add a -w option to show the actual error.

I suppose this is kinda expected due to how badly Apple handles OpenCL (officially stops at 1.2), but #292 did make me expect that it would somewhat work. It didn't stop at -cl-std=CL2.0 at least.

---

Also, using -d 0 causes a different compiler to be used (the GPU compiler is some ancient clang 4.2 or something; the CPU appears more recent), with the different error message of INVALID_WORK_GROUP_SIZE.

[Artoria2e5]

I am terrible at reasoning about atomics and any kind of concurrency. OpenCL 1.x did have a "legacy atomics" thing, but like https://registry.khronos.org/OpenCL/specs/3.0-unified/html/OpenCL_C.html#atomic-legacy says:

These were equivalent to atomic operations with memory_order_relaxed consistency, and memory_scope_work_group scope.

And I'm prrreetty sure memory_scope_work_group is very different from memory_scope_device.

[Artoria2e5]

I think since we've established that OpenCL 1.2 is the culprit, this one can be safely closed like #279. It's interesting how these not-2.0 issues each manifest: this one doesn't complain about any global variables! In fact, if you use -Datomic_uint=uint -Dmemory_order_relaxed=0 -Dmemory_scope_device=0, it survives compilation to instead give:

20251120 17:29:51  Linking 'carryfused.cl' error DEVICE_NOT_AVAILABLE (-2) (args -cl-finite-math-only )
terminate called after throwing an instance of 'gpu_error'
  what():  INVALID_PROGRAM_EXECUTABLE (-45) carryFused at src/clwrap.cpp:282 loadKernel

I guess this time it's looking for a missing symbol, but without the error output I can't be sure.

[Artoria2e5]

https://github.com/philipturner/MoltenCL might be useful, despite being deprecated. Among other thing, it has relaxed atomics in device scope!