Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Jump to bottom

Introduce balanced OpenCL vs CPU tiling #12545

Merged
merged 3 commits into from Sep 28, 2022
Merged

Introduce balanced OpenCL vs CPU tiling #12545

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
3 commits
Select commit Hold shift + click to select a range
9439836
Introduce public functions estimating used memory for tiling
jenshannoschwalm Sep 23, 2022
5dbb2ba
Add cldevice specific advantage option
jenshannoschwalm Sep 24, 2022
0f0f35e
Introduce check for cpu advantage over cl tiled code
jenshannoschwalm Sep 24, 2022
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Jump to
Jump to file
Failed to load files.
Diff view
Diff view
21 changes: 14 additions & 7 deletions src/common/opencl.c
View file
Open in desktop
Expand Up @@ -226,7 +226,7 @@ void dt_opencl_write_device_config(const int devid)
gchar key[256] = { 0 };
gchar dat[512] = { 0 };
g_snprintf(key, 254, "%s%s", DT_CLDEVICE_HEAD, cl->dev[devid].cname);
g_snprintf(dat, 510, "%i %i %i %i %i %i %i %i %f",
g_snprintf(dat, 510, "%i %i %i %i %i %i %i %i %f %.3f",
cl->dev[devid].avoid_atomics,
cl->dev[devid].micro_nap,
cl->dev[devid].pinned_memory & (DT_OPENCL_PINNING_ON | DT_OPENCL_PINNING_DISABLED),
Expand All @@ -235,7 +235,8 @@ void dt_opencl_write_device_config(const int devid)
cl->dev[devid].event_handles,
cl->dev[devid].asyncmode & 1,
cl->dev[devid].disabled & 1,
cl->dev[devid].benchmark);
cl->dev[devid].benchmark,
cl->dev[devid].advantage);
dt_vprint(DT_DEBUG_OPENCL, "[dt_opencl_write_device_config] writing data '%s' for '%s'\n", dat, key);
dt_conf_set_string(key, dat);

Expand Down Expand Up @@ -267,9 +268,10 @@ gboolean dt_opencl_read_device_config(const int devid)
int event_handles;
int asyncmode;
int disabled;
float benchmark;
sscanf(dat, "%i %i %i %i %i %i %i %i %f",
&avoid_atomics, &micro_nap, &pinned_memory, &wd, &ht, &event_handles, &asyncmode, &disabled, &benchmark);
float benchmark;
float advantage;
sscanf(dat, "%i %i %i %i %i %i %i %i %f %f",
&avoid_atomics, &micro_nap, &pinned_memory, &wd, &ht, &event_handles, &asyncmode, &disabled, &benchmark, &advantage);

// some rudimentary safety checking if string seems to be ok
safety_ok = (wd > 1) && (wd < 513) && (ht > 1) && (ht < 513);
Expand All @@ -285,6 +287,7 @@ gboolean dt_opencl_read_device_config(const int devid)
cl->dev[devid].asyncmode = asyncmode;
cl->dev[devid].disabled = disabled;
cl->dev[devid].benchmark = benchmark;
cl->dev[devid].advantage = advantage;
}
else // if there is something wrong with the found conf key reset to defaults
{
Expand All @@ -308,6 +311,8 @@ gboolean dt_opencl_read_device_config(const int devid)
cl->dev[devid].asyncmode &= 1;
cl->dev[devid].disabled &= 1;

cl->dev[devid].advantage = fmaxf(0.0f, cl->dev[devid].advantage);

// Also take care of extended device data, these are not only device specific but also depend on the devid
g_snprintf(key, 254, "%s%s_id%i", DT_CLDEVICE_HEAD, cl->dev[devid].cname, devid);
if(dt_conf_key_not_empty(key))
Expand Down Expand Up @@ -369,6 +374,7 @@ static int dt_opencl_device_init(dt_opencl_t *cl, const int dev, cl_device_id *d
cl->dev[dev].clroundup_wd = 16;
cl->dev[dev].clroundup_ht = 16;
cl->dev[dev].benchmark = 0.0f;
cl->dev[dev].advantage = 0.0f;
cl->dev[dev].use_events = 1;
cl->dev[dev].event_handles = 128;
cl->dev[dev].asyncmode = 0;
Expand Down Expand Up @@ -624,7 +630,8 @@ static int dt_opencl_device_init(dt_opencl_t *cl, const int dev, cl_device_id *d
dt_print_nts(DT_DEBUG_OPENCL, " ROUNDUP HEIGHT: %i\n", cl->dev[dev].clroundup_ht);
dt_print_nts(DT_DEBUG_OPENCL, " CHECK EVENT HANDLES: %i\n", cl->dev[dev].event_handles);
if(cl->dev[dev].benchmark > 0.0f)
dt_print_nts(DT_DEBUG_OPENCL, " PERFORMANCE: %f\n", dt_opencl_device_perfgain(dev));
dt_print_nts(DT_DEBUG_OPENCL, " PERFORMANCE: %.3f\n", dt_opencl_device_perfgain(dev));
dt_print_nts(DT_DEBUG_OPENCL, " TILING ADVANTAGE: %.3f\n", cl->dev[dev].advantage);
dt_print_nts(DT_DEBUG_OPENCL, " DEFAULT DEVICE: %s\n", (type & CL_DEVICE_TYPE_DEFAULT) ? "YES" : "NO");

if(cl->dev[dev].disabled)
Expand Down Expand Up @@ -2789,7 +2796,7 @@ void dt_opencl_check_tuning(const int devid)
cl->dev[devid].tuneactive |= DT_OPENCL_TUNE_PINNED;

int level = res->level;
const gboolean info = ((oldlevel != level) || (oldtuned != tunemode) || (darktable.unmuted & DT_DEBUG_VERBOSE));
const gboolean info = ((oldlevel != level) || (oldtuned != tunemode));
oldlevel = level;
oldtuned = tunemode;

Expand Down
5 changes: 4 additions & 1 deletion src/common/opencl.h
View file
Open in desktop
Expand Up @@ -194,8 +194,11 @@ typedef struct dt_opencl_device_t
int disabled;

// Some devices are known to be unused by other apps so there is no need to test for available memory at all.
// Also some devices might behave badly with the checking code, in this case we could enforce a headroom here.
int forced_headroom;

// As the benchmarks are not good enough to calculate tiled-gpu vs untiled-cpu we have a parameter exposed
// in the cldevice conf key to balance this
float advantage;
} dt_opencl_device_t;

struct dt_bilateral_cl_global_t;
Expand Down