Skip to content

Commit

Permalink
OpenCL: remove scheduling profiles
Browse files Browse the repository at this point in the history 
Since 7751373, we enforce that only one darkroom pipe can run at a time, forcing pipes to wait for each other.

There is no reason anymore to maintain the "fast GPU" or "multiple GPU" profiles.

New behaviour:
1. On new devices and new drivers detection, run the auto-profiling again
2. On auto-profiling, find out the fastest device and write it everywhere in the priority config key for each pipeline
3. Don't re-run inconditionnally the auto-profiling on each application opening. It adds useless delays.
4. Fine-tuning by users can still be done in device priority config key, should the need arise.
5. Remove the cache synch option for OpenCL: synchronization with cache has been mandatorily enabled for some time.
6. Remove the user preference for OpenCL scheduling
  • Loading branch information
@aurelienpierre
aurelienpierre committed Jan 12, 2025
1 parent c15d92a commit 906ef3f
Show file tree
Hide file tree
Showing 3 changed files with 61 additions and 186 deletions.
28 changes: 1 addition & 27 deletions data/anselconfig.xml.in
View file
Original file line number Diff line number Diff line change
Expand Up @@ -255,7 +255,7 @@
<dtconfig>
<name>opencl_device_priority</name>
<type>string</type>
<default>*/!0,*/*/*</default>
<default>*/*/*/*</default>
<shortdescription>priority of OpenCL devices for each pixelpipe type</shortdescription>
<longdescription>defines priorities on how (multiple) OpenCL devices are allocated to the different types of pixelpipe (full, preview, export, thumbnail). for more details visit our usermanual (needs a restart).</longdescription>
</dtconfig>
Expand Down Expand Up @@ -327,19 +327,6 @@
<shortdescription>activate OpenCL support</shortdescription>
<longdescription>if found, use OpenCL runtime on your system for improved processing speed. can be switched on and off at any time.</longdescription>
</dtconfig>
<dtconfig prefs="processing" section="cpugpu" capability="opencl">
<name>opencl_scheduling_profile</name>
<type>
<enum>
<option>default</option>
<option>multiple GPUs</option>
<option>very fast GPU</option>
</enum>
</type>
<default>default</default>
<shortdescription>OpenCL scheduling profile</shortdescription>
<longdescription>defines how preview and full pixelpipe tasks are scheduled on OpenCL enabled systems. default - GPU processes full and CPU processes preview pipe (adaptable by config parameters); multiple GPUs - process both pixelpipes in parallel on two different GPUs; very fast GPU - process both pixelpipes sequentially on the GPU.</longdescription>
</dtconfig>
<dtconfig prefs="processing" section="cpugpu" capability="opencl">
<name>opencl_tuning_mode</name>
<type>
Expand All @@ -354,19 +341,6 @@
<shortdescription>tune OpenCL performance</shortdescription>
<longdescription>allows runtime tuning of OpenCL devices. 'memory size' tests for available graphics ram, 'memory transfer' tries a faster memory access mode (pinned memory) used for tiling.</longdescription>
</dtconfig>
<dtconfig>
<name>opencl_synch_cache</name>
<type>
<enum>
<option>true</option>
<option>active module</option>
<option>false</option>
</enum>
</type>
<default>active module</default>
<shortdescription>cache intermediate OpenCL output</shortdescription>
<longdescription>active module (default) - cache the input to the currently focused module, which allows for faster response time when making multiple adjustments to that module (though the whole pipeline may need to be reprocessed when another module is changed); true - cache the output after each module, which may improve speed, as the whole pixelpipe won't be reprocessed on every parameter change, though will require more memory transfers from the GPU; false - do not sync the pixelpipe cache from OpenCL, which avoids memory transfers from GPUs fast enough to smoothly reprocess the whole pixelpipe.</longdescription>
</dtconfig>
<dtconfig>
<name>opencl_library</name>
<type>string</type>
Expand Down
201 changes: 59 additions & 142 deletions src/common/opencl.c
View file
Original file line number Diff line number Diff line change
Expand Up @@ -59,12 +59,8 @@ static void dt_opencl_priority_parse(dt_opencl_t *cl, char *configstr, int *prio
static void dt_opencl_priorities_parse(dt_opencl_t *cl, const char *configstr);
/** set device priorities according to config string */
static void dt_opencl_update_priorities(const char *configstr);
/** read scheduling profile for config variables */
static dt_opencl_scheduling_profile_t dt_opencl_get_scheduling_profile(void);
/** read config of when/if to sync to cache */
static dt_opencl_sync_cache_t dt_opencl_get_sync_cache(void);
/** adjust opencl subsystem according to scheduling profile */
static void dt_opencl_apply_scheduling_profile(dt_opencl_scheduling_profile_t profile);
static void dt_opencl_apply_scheduling_profile();
/** set opencl specific synchronization timeout */
static void dt_opencl_set_synchronization_timeout(int value);

Expand Down Expand Up @@ -489,6 +485,9 @@ static int dt_opencl_device_init(dt_opencl_t *cl, const int dev, cl_device_id *d
goto end;
}

// take every detected device driver into account of checksum
cl->crc = crc32(cl->crc, (const unsigned char *)deviceversion, deviceversion_size);

(cl->dlocl->symbols->dt_clGetDeviceInfo)(devid, CL_DEVICE_TYPE, sizeof(cl_device_type), &type, NULL);
(cl->dlocl->symbols->dt_clGetDeviceInfo)(devid, CL_DEVICE_IMAGE_SUPPORT, sizeof(cl_bool), &image_support, NULL);
(cl->dlocl->symbols->dt_clGetDeviceInfo)(devid, CL_DEVICE_IMAGE2D_MAX_HEIGHT, sizeof(size_t),
Expand Down Expand Up @@ -850,7 +849,6 @@ void dt_opencl_init(dt_opencl_t *cl, const gboolean exclude_opencl, const gboole
char *locale = strdup(setlocale(LC_ALL, NULL));
setlocale(LC_ALL, "C");

cl->sync_cache = dt_opencl_get_sync_cache();
cl->crc = 5781;
cl->dlocl = NULL;
cl->dev_priority_image = NULL;
Expand All @@ -865,11 +863,6 @@ void dt_opencl_init(dt_opencl_t *cl, const gboolean exclude_opencl, const gboole
char *platform_vendor = calloc(DT_OPENCL_CBUFFSIZE, sizeof(char));

cl->cpubenchmark = dt_conf_get_float("dt_cpubenchmark");
if(cl->cpubenchmark <= 0.0f)
{
cl->cpubenchmark = dt_opencl_benchmark_cpu(1024, 1024, 5, 100.0f);
dt_conf_set_float("dt_cpubenchmark", cl->cpubenchmark);
}

if(exclude_opencl)
{
Expand All @@ -880,17 +873,14 @@ void dt_opencl_init(dt_opencl_t *cl, const gboolean exclude_opencl, const gboole

dt_print_nts(DT_DEBUG_OPENCL, "[opencl_init] opencl related configuration options:\n");
dt_print_nts(DT_DEBUG_OPENCL, "[opencl_init] opencl: %s\n", dt_conf_get_bool("opencl") ? "ON" : "OFF" );
const char *str = dt_conf_get_string_const("opencl_scheduling_profile");
dt_print_nts(DT_DEBUG_OPENCL, "[opencl_init] opencl_scheduling_profile: '%s'\n", str);
const char *str;
// look for explicit definition of opencl_runtime library in preferences
const char *library = dt_conf_get_string_const("opencl_library");
dt_print_nts(DT_DEBUG_OPENCL, "[opencl_init] opencl_library: '%s'\n", (strlen(library) == 0) ? "default path" : library);
str = dt_conf_get_string_const("opencl_device_priority");
dt_print_nts(DT_DEBUG_OPENCL, "[opencl_init] opencl_device_priority: '%s'\n", str);
dt_print_nts(DT_DEBUG_OPENCL, "[opencl_init] opencl_mandatory_timeout: %d\n",
dt_conf_get_int("opencl_mandatory_timeout"));
str = dt_conf_get_string_const("opencl_synch_cache");
dt_print_nts(DT_DEBUG_OPENCL, "[opencl_init] opencl_synch_cache: %s\n", str);

// dynamically load opencl runtime
if((cl->dlocl = dt_dlopencl_init(library)) == NULL)
Expand Down Expand Up @@ -1037,7 +1027,7 @@ void dt_opencl_init(dt_opencl_t *cl, const gboolean exclude_opencl, const gboole
&& cl->dev_priority_export != NULL && cl->dev_priority_thumbnail != NULL);

dt_print_nts(DT_DEBUG_OPENCL, "[opencl_init] OpenCL successfully initialized.\n");
dt_print_nts(DT_DEBUG_OPENCL, "[opencl_init] here are the internal numbers and names of OpenCL devices available to darktable:\n");
dt_print_nts(DT_DEBUG_OPENCL, "[opencl_init] here are the internal numbers and names of OpenCL devices available to Ansel:\n");
for(int i = 0; i < dev; i++) dt_print_nts(DT_DEBUG_OPENCL, "[opencl_init]\t\t%d\t'%s'\n", i, cl->dev[i].name);
}
else
Expand All @@ -1050,6 +1040,16 @@ void dt_opencl_init(dt_opencl_t *cl, const gboolean exclude_opencl, const gboole
cl->inited ? "" : "NOT ");
dt_print_nts(DT_DEBUG_OPENCL, "[opencl_init] initial status of opencl enabled flag is %s.\n",
cl->enabled ? "ON" : "OFF");

// check if the list of existing OpenCL devices (indicated by checksum != oldchecksum) has changed
// If it has, reprofile and update config if needed.
// TODO: account for driver version too.
char checksum[64];
snprintf(checksum, sizeof(checksum), "%u", cl->crc);
const char *oldchecksum = dt_conf_get_string_const("opencl_checksum");
const gboolean manually = strcasecmp(oldchecksum, "OFF") == 0;
const gboolean newcheck = ((strcmp(oldchecksum, checksum) != 0) || (strlen(oldchecksum) < 1));

if(cl->inited)
{
dt_capabilities_add("opencl");
Expand All @@ -1062,6 +1062,15 @@ void dt_opencl_init(dt_opencl_t *cl, const gboolean exclude_opencl, const gboole
cl->heal = dt_heal_init_cl_global();
cl->colorspaces = dt_colorspaces_init_cl_global();
cl->guided_filter = dt_guided_filter_init_cl_global();
}

if(newcheck && !manually && cl->inited)
{
dt_print_nts(DT_DEBUG_OPENCL, "[opencl_init] OpenCL devices changed, we will update the profiling configuration.\n");
dt_conf_set_string("opencl_checksum", checksum);

cl->cpubenchmark = dt_opencl_benchmark_cpu(1024, 1024, 5, 100.0f);
dt_conf_set_float("dt_cpubenchmark", cl->cpubenchmark);

// make sure all active cl devices have a benchmark result
for(int n = 0; n < cl->num_devs; n++)
Expand All @@ -1073,69 +1082,42 @@ void dt_opencl_init(dt_opencl_t *cl, const gboolean exclude_opencl, const gboole
}
}

char checksum[64];
snprintf(checksum, sizeof(checksum), "%u", cl->crc);
const char *oldchecksum = dt_conf_get_string_const("opencl_checksum");

const gboolean manually = strcasecmp(oldchecksum, "OFF") == 0;
const gboolean newcheck = ((strcmp(oldchecksum, checksum) != 0) || (strlen(oldchecksum) < 1));

// check if the list of existing OpenCL devices (indicated by checksum != oldchecksum) has changed
if(newcheck && !manually)
// get minima and maxima of performance data of all active devices
const float tcpu = cl->cpubenchmark;
float tgpumin = INFINITY;
float tgpumax = -INFINITY;
int fastest_device = -1; // Device -1 is CPU
for(int n = 0; n < cl->num_devs; n++)
{
dt_conf_set_string("opencl_checksum", checksum);

// get minima and maxima of performance data of all active devices
const float tcpu = cl->cpubenchmark;
float tgpumin = INFINITY;
float tgpumax = -INFINITY;
for(int n = 0; n < cl->num_devs; n++)
if((cl->dev[n].benchmark > 0.0f) && (cl->dev[n].disabled == 0))
{
if((cl->dev[n].benchmark > 0.0f) && (cl->dev[n].disabled == 0))
if(cl->dev[n].benchmark < tgpumin)
{
tgpumin = fminf(cl->dev[n].benchmark, tgpumin);
tgpumax = fmaxf(cl->dev[n].benchmark, tgpumax);
tgpumin = cl->dev[n].benchmark;
fastest_device = n;
}
tgpumax = fmaxf(cl->dev[n].benchmark, tgpumax);
}
}

if(tcpu < tgpumin / 3.0f)
{
// de-activate opencl for darktable in case the cpu is three times faster than the fastest GPU.
// FIXME the problem here is that the benchmark might not reflect real-world performance.
// user can always manually overrule this later.
cl->enabled = FALSE;
dt_conf_set_bool("opencl", FALSE);
dt_print_nts(DT_DEBUG_OPENCL, "[opencl_init] due to a slow GPU the opencl flag has been set to OFF.\n");
dt_control_log(_("due to a slow GPU hardware acceleration via opencl has been de-activated"));
}
else if((cl->num_devs >= 2) && ((tgpumax / tgpumin) < 1.1f))
{
// set scheduling profile to "multiple GPUs" if more than one device has been found and they are equally fast
dt_conf_set_string("opencl_scheduling_profile", "multiple GPUs");
dt_print_nts(DT_DEBUG_OPENCL, "[opencl_init] set scheduling profile for multiple GPUs.\n");
dt_control_log(_("multiple GPUs detected - opencl scheduling profile has been set accordingly"));
}
else if((tcpu >= 2.0f * tgpumin) && (cl->num_devs == 1))
{
// set scheduling profile to "very fast GPU" if fastest GPU is at least 2 times better than CPU and there is just one device
// We might want a better benchmark but even with the current result (underestimates real world performance) this is safe.
dt_conf_set_string("opencl_scheduling_profile", "very fast GPU");
dt_print_nts(DT_DEBUG_OPENCL, "[opencl_init] set scheduling profile for very fast GPU.\n");
dt_control_log(_("very fast GPU detected - opencl scheduling profile has been set accordingly"));
}
else
{
// set scheduling profile to "default"
dt_conf_set_string("opencl_scheduling_profile", "default");
dt_print_nts(DT_DEBUG_OPENCL, "[opencl_init] set scheduling profile to default.\n");
dt_control_log(_("opencl scheduling profile set to default"));
}
if(tcpu < tgpumin / 1.5f)
{
// de-activate opencl for darktable in case the cpu is faster than the fastest GPU.
// FIXME the problem here is that the benchmark might not reflect real-world performance.
// user can always manually overrule this later.
cl->enabled = FALSE;
dt_conf_set_bool("opencl", FALSE);
dt_print_nts(DT_DEBUG_OPENCL, "[opencl_init] due to a slow GPU the opencl flag has been set to OFF.\n");
dt_control_log(_("due to a slow GPU hardware acceleration via opencl has been de-activated"));
}

// apply config settings for scheduling profile: sets device priorities and pixelpipe synchronization timeout
dt_opencl_scheduling_profile_t profile = dt_opencl_get_scheduling_profile();
dt_opencl_apply_scheduling_profile(profile);
dt_conf_set_string("opencl_device_priority", g_strdup_printf("%i/%i/%i/%i", fastest_device, fastest_device, fastest_device, fastest_device));
}
else // initialization failed

dt_opencl_apply_scheduling_profile();

if(!cl->inited)// initialization failed
{
for(int i = 0; cl->dev && i < cl->num_devs; i++)
{
Expand Down Expand Up @@ -1686,10 +1668,11 @@ static void dt_opencl_update_priorities(const char *configstr)
dt_opencl_priorities_parse(cl, configstr);

dt_print_nts(DT_DEBUG_OPENCL, "[dt_opencl_update_priorities] these are your device priorities:\n");
dt_print_nts(DT_DEBUG_OPENCL, "[dt_opencl_update_priorities] \t\timage\tpreview\texport\tthumbs\n");
dt_print_nts(DT_DEBUG_OPENCL, "[dt_opencl_update_priorities] \tid |\t\timage\tpreview\texport\tthumbs\n");
for(int i = 0; i < cl->num_devs; i++)
dt_print_nts(DT_DEBUG_OPENCL, "[dt_opencl_update_priorities]\t\t%d\t%d\t%d\t%d\n", cl->dev_priority_image[i],
cl->dev_priority_preview[i], cl->dev_priority_export[i], cl->dev_priority_thumbnail[i]);
dt_print_nts(DT_DEBUG_OPENCL, "[dt_opencl_update_priorities]\t%i |\t\t%d\t%d\t%d\t%d\n",
i, cl->dev_priority_image[i],
cl->dev_priority_preview[i], cl->dev_priority_export[i], cl->dev_priority_thumbnail[i]);
dt_print_nts(DT_DEBUG_OPENCL, "[dt_opencl_update_priorities] show if opencl use is mandatory for a given pixelpipe:\n");
dt_print_nts(DT_DEBUG_OPENCL, "[dt_opencl_update_priorities] \t\timage\tpreview\texport\tthumbs\n");
dt_print_nts(DT_DEBUG_OPENCL, "[dt_opencl_update_priorities]\t\t%d\t%d\t%d\t%d\n", cl->mandatory[0],
Expand Down Expand Up @@ -2970,43 +2953,9 @@ int dt_opencl_update_settings(void)
dt_print(DT_DEBUG_OPENCL, "[opencl_update_enabled] enabled flag set to %s\n", prefs ? "ON" : "OFF");
}

dt_opencl_scheduling_profile_t profile = dt_opencl_get_scheduling_profile();

if(cl->scheduling_profile != profile)
{
const char *pstr = dt_conf_get_string_const("opencl_scheduling_profile");
dt_print(DT_DEBUG_OPENCL, "[opencl_update_scheduling_profile] scheduling profile set to %s\n", pstr);
dt_opencl_apply_scheduling_profile(profile);
}

dt_opencl_sync_cache_t sync = dt_opencl_get_sync_cache();

if(cl->sync_cache != sync)
{
const char *pstr = dt_conf_get_string_const("opencl_synch_cache");
dt_print(DT_DEBUG_OPENCL, "[opencl_update_synch_cache] sync cache set to %s\n", pstr);
cl->sync_cache = sync;
}

return (cl->enabled && !cl->stopped);
}

/** read scheduling profile for config variables */
static dt_opencl_scheduling_profile_t dt_opencl_get_scheduling_profile(void)
{
const char *pstr = dt_conf_get_string_const("opencl_scheduling_profile");
if(!pstr) return OPENCL_PROFILE_DEFAULT;

dt_opencl_scheduling_profile_t profile = OPENCL_PROFILE_DEFAULT;

if(!strcmp(pstr, "multiple GPUs"))
profile = OPENCL_PROFILE_MULTIPLE_GPUS;
else if(!strcmp(pstr, "very fast GPU"))
profile = OPENCL_PROFILE_VERYFAST_GPU;

return profile;
}

int dt_opencl_get_tuning_mode(void)
{
int res = DT_OPENCL_TUNE_NOTHING;
Expand All @@ -3020,22 +2969,6 @@ int dt_opencl_get_tuning_mode(void)
return res;
}

/** read config of when/if to synch to cache */
static dt_opencl_sync_cache_t dt_opencl_get_sync_cache(void)
{
const char *pstr = dt_conf_get_string_const("opencl_synch_cache");
if(!pstr) return OPENCL_SYNC_ACTIVE_MODULE;

dt_opencl_sync_cache_t sync = OPENCL_SYNC_ACTIVE_MODULE;

if(!strcmp(pstr, "true"))
sync = OPENCL_SYNC_TRUE;
else if(!strcmp(pstr, "false"))
sync = OPENCL_SYNC_FALSE;

return sync;
}

/** set opencl specific synchronization timeout */
static void dt_opencl_set_synchronization_timeout(int value)
{
Expand All @@ -3044,27 +2977,11 @@ static void dt_opencl_set_synchronization_timeout(int value)
}

/** adjust opencl subsystem according to scheduling profile */
static void dt_opencl_apply_scheduling_profile(dt_opencl_scheduling_profile_t profile)
static void dt_opencl_apply_scheduling_profile()
{
dt_pthread_mutex_lock(&darktable.opencl->lock);
darktable.opencl->scheduling_profile = profile;

switch(profile)
{
case OPENCL_PROFILE_MULTIPLE_GPUS:
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_opencl_update_priorities(dt_conf_get_string_const("opencl_device_priority"));
dt_opencl_set_synchronization_timeout(dt_conf_get_int("pixelpipe_synchronization_timeout"));
dt_pthread_mutex_unlock(&darktable.opencl->lock);
}

Expand Down
Loading

0 comments on commit 906ef3f

Please sign in to comment.