OpenCL is mandatory, yet CPU is used (solved)

kofa · January 8, 2022, 7:44am

opencl_scheduling_profile can be set to either very fast GPU or default. If it’s set to very fast GPU, priorities will be set as follows:

github.com

darktable-org/darktable/blob/master/src/common/opencl.c#L2606-L2607

    
      
          case OPENCL_PROFILE_VERYFAST_GPU:
            dt_opencl_update_priorities("+*/+*/+*/+*/+*");

If set to default, it’ll honour whatever is specified using in opencl_device_priority:

github.com

darktable-org/darktable/blob/master/src/common/opencl.c#L2610-L2612

    
      
          case OPENCL_PROFILE_DEFAULT:
          default:
            dt_opencl_update_priorities(dt_conf_get_string_const("opencl_device_priority"));

So one piece of the puzzle is either to use opencl_scheduling_profile=very fast GPU or opencl_scheduling_profile=default. The latter also requires that one specify a setting for opencl_device_priority that forces processing on the GPU. I used opencl_device_priority=+0/+0/+0/+0/+0 (use the 1st device for everything – I have a single GPU), but the one forced by very fast GPU, +*/+*/+*/+*/+* (force the use of any GPU device) is also a good option (especially if one has multiple high-performance GPUs).

The third piece is opencl_mandatory_timeout. No matter what profile is used, the timeout logic applies: if for one of the pipelines the GPU is mandatory (+ sign in the priority), darktable will wait for a GPU to become available, but that wait is limited by opencl_mandatory_timeout (which sets the number of attempts, each attempt meaning a 5-ms wait). Setting very fast GPU also sets pixelpipe_synchronization_timeout, but it does not set anything that affects the wait-then-fallback-to-CPU logic:

github.com

darktable-org/darktable/blob/master/src/common/opencl.c#L2613

    
      
                dt_opencl_update_priorities("*/*/*/*/*");
                dt_opencl_set_synchronization_timeout(20);
                break;
              case OPENCL_PROFILE_VERYFAST_GPU:
                dt_opencl_update_priorities("+*/+*/+*/+*/+*");
                dt_opencl_set_synchronization_timeout(0);
                break;
              case OPENCL_PROFILE_DEFAULT:
              default:
                dt_opencl_update_priorities(dt_conf_get_string_const("opencl_device_priority"));
                dt_opencl_set_synchronization_timeout(dt_conf_get_int("pixelpipe_synchronization_timeout"));
                break;
            }
            dt_pthread_mutex_unlock(&darktable.opencl->lock);
          }
          
          
/** get global memory of device */
          cl_ulong dt_opencl_get_max_global_mem(const int devid)
          {
            if(!darktable.opencl->inited || devid < 0) return 0;
            return darktable.opencl->dev[devid].max_global_mem;

So the simplest settings for my case are:

opencl_scheduling_profile=very fast GPU
opencl_mandatory_timeout=20000

With those, darktable will wait up to 100 s; opencl_device_priority specified in the file is ignored and overridden to +*/+*/+*/+*/+* (mandatory processing on any available GPU).