Support rusticl runtime #276
Description
I'm trying to run gpuowl on rusticl instead of rocm, it fails with this:
[f40@p95 gpuowl]$ RUSTICL_ENABLE=radeonsi RUSTICL_FEATURES=fp64 ./build-debug/gpuowl -prp 77936867
20240218 14:10:06 GpuOwl VERSION v7.5-2-gba227ce
20240218 14:10:06 GpuOwl VERSION v7.5-2-gba227ce
20240218 14:10:06 config: -prp 77936867
20240218 14:10:06 device 0, unique id ''
20240218 14:10:06 77936867 FFT: 4M 1K:8:256 (18.58 bpw)
20240218 14:10:06 77936867 OpenCL args "-DEXP=77936867u -DWIDTH=1024u -DSMALL_HEIGHT=256u -DMIDDLE=8u -DAMDGPU=1 -DMM_CHAIN=1u -DMM2_CHAIN=2u -DWEIGHT_STEP=0.33644726404543274 -DIWEIGHT_STEP=-0.25174750481886216 -DIWEIGHTS={0,-0.44011820345520131,-0.37306474779553728,-0.29798072935699788,-0.21390437908665341,-0.11975874301407295,-0.014337887291734644,-0.44814572555075455,} -DFWEIGHTS={0,0.78609128957452257,0.5950610473469905,0.42446232150303748,0.2721098723818392,0.1360521812214803,0.014546452690911484,0.81207258201996746,} -cl-std=CL2.0 -cl-finite-math-only "
20240218 14:10:06 77936867 ASM compilation failed, retrying compilation using NO_ASM
20240218 14:10:06 77936867 OpenCL compilation error -11 (args -DEXP=77936867u -DWIDTH=1024u -DSMALL_HEIGHT=256u -DMIDDLE=8u -DAMDGPU=1 -DMM_CHAIN=1u -DMM2_CHAIN=2u -DWEIGHT_STEP=0.33644726404543274 -DIWEIGHT_STEP=-0.25174750481886216 -DIWEIGHTS={0,-0.44011820345520131,-0.37306474779553728,-0.29798072935699788,-0.21390437908665341,-0.11975874301407295,-0.014337887291734644,-0.44814572555075455,} -DFWEIGHTS={0,0.78609128957452257,0.5950610473469905,0.42446232150303748,0.2721098723818392,0.1360521812214803,0.014546452690911484,0.81207258201996746,} -cl-std=CL2.0 -cl-finite-math-only -DNO_ASM=1)
20240218 14:10:06 77936867 input.cl:44:26: warning: unsupported OpenCL extension 'cl_khr_int64_base_atomics' - ignoring [-Wignored-pragmas]
input.cl:1494:29: error: call to 'atom_add' is ambiguous
Error executing LLVM compilation action.
20240218 14:10:06 Exception gpu_error: BUILD_PROGRAM_FAILURE clBuildProgram at src/clwrap.cpp:245 build
20240218 14:10:06 Bye
gpuowl.cl says this:
// 64-bit atomics used in kernel sum64
// If 64-bit atomics aren't available, sum64() can be implemented with 32-bit
#pragma OPENCL EXTENSION cl_khr_int64_base_atomics : enable
and the sum64 function looks simple enough:
KERNEL(256) sum64(global ulong* out, u32 sizeBytes, global ulong* in) {
if (get_global_id(0) == 0) { out[0] = 0; }
ulong sum = 0;
for (i32 p = get_global_id(0); p < sizeBytes / sizeof(u64); p += get_global_size(0)) {
sum += in[p];
}
sum = work_group_reduce_add(sum);
if (get_local_id(0) == 0) { atom_add(&out[0], sum); }
}
If implementing sum64 with 32 bit atomics and sorting the atom_add ambiguity (which appears to be related) is all it takes to get gpuowl working on rusticl then cool. Rusticl is cross-platform and built into mesa (should be in OOTB for Ubuntu 24.04) and should be the way forwards for opencl on Linux (also mfakto is quicker under rusticl, for my hardware). I have no idea if atomic 64 bit int will ever be supported with rusticl, or if there are other hidden or vendor issues to be uncovered (can only test with RDNA3 iGPU 780M).