I decided to work on adding OpenMP support to the `graphics/dartktable` starting at version `3.6.1`.
Here's the original revision with OpenMP disabled. http://cvsweb.openbsd.org/cgi-bin/cvsweb/ports/graphics/darktable/Makefile?rev=1.113&content-type=text/x-cvsweb-markup Initially I've rebuild gcc8 from ports enabling OpenMP for -main and -libs packages. This gcc with omp support seemed to work fine at first glance but I quickly realized the makefiles generated for Darktable were incompatible. - OpenMP was detected, good part - Issues with `external/rawspeed` cmake checks for l1d cache line size and page size querying, partially fixed (workaround) `_SC_LEVEL1_DCACHE_LINESIZE` being undefined on OpenBSD - base `as(1)` and the ports-gcc where conflicting on instruction sets giving `Error: no such instruction: vmovlps [...]` and others for some C and C++ Darktables sources. After seeing those `as(1)` conflicts, I decided to rebuild gcc-8 with CONFIGURE_ARGS += --with-as="${LOCALBASE}/bin/gas" and errors were gone. Others followed. Missing `graphite` and `ifunc`. This was too much time spent on GCC, so I decided to switch back the integration to base- clang. With the help of two OpenBSD 6.8 unofficial ports for `stlibomp` and `libomp` found there: https://j-bm.github.io/on/onp.html I upgraded them to llvm 13.0.0 locally and good results started to show. In `graphic/darktable`, -DUSE_OPENMP=ON with COMPILER = base-clang made a working package. I locally added two temporary patches covering the issues encountered initially. The first one is for the l1d cache line size detection which calls this for `defined(__OpenBSD__)` and others: if (sysctlbyname("hw.cachelinesize", &val, &size, NULL, 0) This is not declared for OpenBSD. Maybe a good approach would be to implement a `sysctl(2)` constant for `_SC_LEVEL1_DCACHE_LINESIZE` as seen around, see: https://github.com/darktable-org/rawspeed/blob/88e2b92a204f2d404d156b0ba11a66604685d5a0/cmake/Modules/cpu-cache-line-size.cpp As I said on GH, sysctl(2) would be perfect except that this CPUID information is not available on OpenBSD currently. There's no `_SC_LEVEL1_DCACHE_LINESIZE` as well as others 0x8000005/6 cache informations defined. For information, I did not see such definitions for other BSD's. Maybe I've not looked so deep, but at first glance, there's no hw.cachelinesize as well... I then scratched this to get the required value: #include <stdio.h> // This returns 64 on my AMD Ryzen 7 PRO 3700U // I'm not 100% sure this is a portable solution across Intel and AMD // Also I'm no 100% sure how to get the right L1D value rn, still studying it. // Seems that on AMD I shall return ECX from bits 0-7, I think a proper __get_cpuid // (cpuid.h from gcc, also available for clang afaik) would be better. int main() { uint32_t linesize; __asm__("cpuid" : "=a" (linesize) : "0" (5)); printf("%d\n",linesize); return 0; } Thanks to these references: https://en.wikipedia.org/wiki/CPUID#CPU-specific_information_outside_x86 https://github.com/gcc-mirror/gcc/blob/master/gcc/config/i386/cpuid.h http://www.flounder.com/cpuid_explorer2.htm#CPUID(0x80000005) This allowed me to produce this _dirty_ local workaround: --- src/external/rawspeed/cmake/Modules/cpu-cache-line-size.cpp.orig Thu Jan 27 03:13:57 2022 +++ src/external/rawspeed/cmake/Modules/cpu-cache-line-size.cpp Thu Jan 27 03:16:04 2022 @@ -14,7 +14,7 @@ int main() { return 0; } -#elif defined(__FreeBSD__) || defined(__NetBSD__) || defined(__OpenBSD__) || \ +#elif defined(__FreeBSD__) || defined(__NetBSD__) || \ defined(__DragonFly__) || defined(__APPLE__) #include <stddef.h> // for size_t @@ -27,6 +27,13 @@ int main() { return 1; std::cout << val << std::endl; return 0; +} + +#elif defined(__OpenBSD__) +int main() { + size_t val = 64; + std::cout << val << std::endl; + return 0; } #elif defined(_WIN32) || defined(_WIN64) The second one was simpler as it was just a missing `defined(__OpenBSD__)` for getting the page size constant. --- src/external/rawspeed/cmake/Modules/cpu-page-size.cpp.orig Thu Jan 27 03:08:25 2022 +++ src/external/rawspeed/cmake/Modules/cpu-page-size.cpp Thu Jan 27 03:10:39 2022 @@ -4,7 +4,8 @@ #include <unistd.h> // for _POSIX_C_SOURCE, sysconf, _SC_PAGESIZE #endif -#if (defined(_POSIX_C_SOURCE) && _POSIX_C_SOURCE >= 1) || defined(__APPLE__) +#if (defined(_POSIX_C_SOURCE) && _POSIX_C_SOURCE >= 1) || \ + defined(__APPLE__) || defined(__OpenBSD__) int main() { long val = ::sysconf(_SC_PAGESIZE); At the end of the day, Darktable is now working _a lot_ faster with OpenMP. I wish to be able to get `stlibomp` and `libomp` starting at 13.0.0 in ports and a proper workaround or the missing CPUID definitions in sys. Thanks for your reading, please let me know your thoughts.
