nv-tegra-mirror/linux-nvgpu
1
0
Fork 0
mirror of git://nv-tegra.nvidia.com/linux-nvgpu.git synced 2025-12-22 09:12:24 +03:00
Code Issues Packages Projects Releases Wiki Activity

gpu: nvgpu: using pmu counters for load estimate

Browse Source 
PMU counters #0 and #4 are used to count total cycles and busy cycles.
These counts are used by podgov to estimate GPU load.

PMU idle intr status register is used to monitor overflow. Overflow
rarely occurs because frequency governor reads and resets the counters
at a high cadence. When overflow occurs, 100% work load is reported to
frequency governor.

Bug 1963732

Change-Id: I046480ebde162e6eda24577932b96cfd91b77c69
Signed-off-by: Peng Liu <pengliu@nvidia.com>
Reviewed-on: https://git-master.nvidia.com/r/1939547
Reviewed-by: mobile promotions <svcmobile_promotions@nvidia.com>
Tested-by: mobile promotions <svcmobile_promotions@nvidia.com>
This commit is contained in:
Peng Liu
2018-10-30 13:45:43 -07:00
committed by mobile promotions
parent 75ff0feeff
commit 34df003519
21 changed files with 449 additions and 10 deletions
Download Patch File Download Diff File Expand all files Collapse all files

18
drivers/gpu/nvgpu/os/linux/scale.c
View File

@@ -211,18 +211,18 @@ static int gk20a_scale_target(struct device *dev, unsigned long *freq,
}
/*
* update_load_estimate_gpmu(profile)
* update_load_estimate_busy_cycles(dev)
*
* Update load estimate using gpmu. The gpmu value is normalised
* Update load estimate using pmu idle counters. Result is normalised
* based on the time it was asked last time.
*/
static void update_load_estimate_gpmu(struct device *dev)
static void update_load_estimate_busy_cycles(struct device *dev)
{
struct gk20a *g = get_gk20a(dev);
struct gk20a_scale_profile *profile = g->scale_profile;
unsigned long dt;
u32 busy_time;
u32 busy_cycles_norm;
ktime_t t;
t = ktime_get();
@@ -230,8 +230,9 @@ static void update_load_estimate_gpmu(struct device *dev)
profile->dev_stat.total_time = dt;
profile->last_event_time = t;
nvgpu_pmu_load_norm(g, &busy_time);
profile->dev_stat.busy_time = (busy_time * dt) / 1000;
nvgpu_pmu_busy_cycles_norm(g, &busy_cycles_norm);
profile->dev_stat.busy_time =
(busy_cycles_norm * dt) / PMU_BUSY_CYCLES_NORM_MAX;
}
/*
@@ -284,9 +285,6 @@ static int gk20a_scale_get_dev_status(struct device *dev,
struct gk20a_scale_profile *profile = g->scale_profile;
struct gk20a_platform *platform = dev_get_drvdata(dev);
/* update the software shadow */
nvgpu_pmu_load_update(g);
/* inform edp about new constraint */
if (platform->prescale)
platform->prescale(dev);
@@ -296,7 +294,7 @@ static int gk20a_scale_get_dev_status(struct device *dev,
g->ops.clk.get_rate(g, CTRL_CLK_DOMAIN_GPCCLK);
/* Update load estimate */
update_load_estimate_gpmu(dev);
update_load_estimate_busy_cycles(dev);
/* Copy the contents of the current device status */
*stat = profile->dev_stat;