[Intel-gfx] [RFC i-g-t 3/6] intel-gpu-overlay: Show 1s, 30s and 15m GPU load
Tvrtko Ursulin
tursulin at ursulin.net
Wed Oct 3 12:07:15 UTC 2018
From: Tvrtko Ursulin <tvrtko.ursulin at intel.com>
Show total GPU loads in the window banner.
Engine load is defined as total of runnable and running requests on an
engine.
Total, non-normalized, load is display. In other words if N engines are
busy with exactly one request, the load will be shown as N.
v2:
* Different flavour of load avg. (Chris Wilson)
* Simplify code. (Chris Wilson)
Signed-off-by: Tvrtko Ursulin <tvrtko.ursulin at intel.com>
---
overlay/gpu-top.c | 39 ++++++++++++++++++++++++++++++++++++++-
overlay/gpu-top.h | 11 ++++++++++-
overlay/overlay.c | 28 ++++++++++++++++++++++------
3 files changed, 70 insertions(+), 8 deletions(-)
diff --git a/overlay/gpu-top.c b/overlay/gpu-top.c
index 22e9badb22c1..501429b86379 100644
--- a/overlay/gpu-top.c
+++ b/overlay/gpu-top.c
@@ -28,6 +28,7 @@
#include <string.h>
#include <unistd.h>
#include <fcntl.h>
+#include <math.h>
#include <errno.h>
#include <assert.h>
@@ -126,6 +127,10 @@ static int perf_init(struct gpu_top *gt)
gt->ring[gt->num_rings++].name = d->name;
}
+ gt->have_load_avg = gt->have_queued &&
+ gt->have_runnable &&
+ gt->have_running;
+
return 0;
}
@@ -290,17 +295,32 @@ static void mmio_init(struct gpu_top *gt)
}
}
-void gpu_top_init(struct gpu_top *gt)
+void gpu_top_init(struct gpu_top *gt, unsigned int period_us)
{
+ const double period = (double)period_us / 1e6;
+ const double load_period[NUM_LOADS] = { 1.0, 30.0, 900.0 };
+ const char *load_names[NUM_LOADS] = { "1s", "30s", "15m" };
+ unsigned int i;
+
memset(gt, 0, sizeof(*gt));
gt->fd = -1;
+ for (i = 0; i < NUM_LOADS; i++) {
+ gt->load_name[i] = load_names[i];
+ gt->exp[i] = exp(-period / load_period[i]);
+ }
+
if (perf_init(gt) == 0)
return;
mmio_init(gt);
}
+static double update_load(double load, double exp, double val)
+{
+ return val + exp * (load - val);
+}
+
int gpu_top_update(struct gpu_top *gt)
{
uint32_t data[1024];
@@ -313,6 +333,8 @@ int gpu_top_update(struct gpu_top *gt)
struct gpu_top_stat *s = >->stat[gt->count++&1];
struct gpu_top_stat *d = >->stat[gt->count&1];
uint64_t *sample, d_time;
+ double gpu_qd = 0.0;
+ unsigned int i;
int n, m;
len = read(gt->fd, data, sizeof(data));
@@ -341,6 +363,8 @@ int gpu_top_update(struct gpu_top *gt)
d_time = s->time - d->time;
for (n = 0; n < gt->num_rings; n++) {
+ double qd = 0.0;
+
gt->ring[n].u.u.busy = (100 * (s->busy[n] - d->busy[n]) + d_time/2) / d_time;
if (gt->have_wait)
gt->ring[n].u.u.wait = (100 * (s->wait[n] - d->wait[n]) + d_time/2) / d_time;
@@ -353,6 +377,14 @@ int gpu_top_update(struct gpu_top *gt)
if (gt->have_running)
gt->ring[n].running = (double)((s->running[n] - d->running[n])) / I915_SAMPLE_RUNNING_DIVISOR * 1e9 / d_time;
+ qd = gt->ring[n].runnable + gt->ring[n].running;
+ gpu_qd += qd;
+
+ for (i = 0; i < NUM_LOADS; i++)
+ gt->ring[n].load[i] =
+ update_load(gt->ring[n].load[i],
+ gt->exp[i], qd);
+
/* in case of rounding + sampling errors, fudge */
if (gt->ring[n].u.u.busy > 100)
gt->ring[n].u.u.busy = 100;
@@ -362,6 +394,11 @@ int gpu_top_update(struct gpu_top *gt)
gt->ring[n].u.u.sema = 100;
}
+ for (i = 0; i < NUM_LOADS; i++) {
+ gt->load[i] = update_load(gt->load[i], gt->exp[i],
+ gpu_qd);
+ gt->norm_load[i] = gt->load[i] / gt->num_rings;
+ }
update = 1;
} else {
while ((len = read(gt->fd, data, sizeof(data))) > 0) {
diff --git a/overlay/gpu-top.h b/overlay/gpu-top.h
index cb4310c82a94..115ce8c482c1 100644
--- a/overlay/gpu-top.h
+++ b/overlay/gpu-top.h
@@ -26,6 +26,7 @@
#define GPU_TOP_H
#define MAX_RINGS 16
+#define NUM_LOADS 3
#include <stdint.h>
@@ -39,6 +40,12 @@ struct gpu_top {
int have_queued;
int have_runnable;
int have_running;
+ int have_load_avg;
+
+ double exp[NUM_LOADS];
+ double load[NUM_LOADS];
+ double norm_load[NUM_LOADS];
+ const char *load_name[NUM_LOADS];
struct gpu_top_ring {
const char *name;
@@ -54,6 +61,8 @@ struct gpu_top {
double queued;
double runnable;
double running;
+
+ double load[NUM_LOADS];
} ring[MAX_RINGS];
struct gpu_top_stat {
@@ -69,7 +78,7 @@ struct gpu_top {
int count;
};
-void gpu_top_init(struct gpu_top *gt);
+void gpu_top_init(struct gpu_top *gt, unsigned int period_us);
int gpu_top_update(struct gpu_top *gt);
#endif /* GPU_TOP_H */
diff --git a/overlay/overlay.c b/overlay/overlay.c
index 7087d5e5b24e..1a1d01db082b 100644
--- a/overlay/overlay.c
+++ b/overlay/overlay.c
@@ -141,7 +141,8 @@ struct overlay_context {
};
static void init_gpu_top(struct overlay_context *ctx,
- struct overlay_gpu_top *gt)
+ struct overlay_gpu_top *gt,
+ unsigned int period_us)
{
const double rgba[][4] = {
{ 1, 0.25, 0.25, 1 },
@@ -153,7 +154,7 @@ static void init_gpu_top(struct overlay_context *ctx,
int n;
cpu_top_init(>->cpu_top);
- gpu_top_init(>->gpu_top);
+ gpu_top_init(>->gpu_top, period_us);
chart_init(>->cpu, "CPU", 120);
chart_set_position(>->cpu, PAD, PAD);
@@ -928,13 +929,13 @@ int main(int argc, char **argv)
debugfs_init();
- init_gpu_top(&ctx, &ctx.gpu_top);
+ sample_period = get_sample_period(&config);
+
+ init_gpu_top(&ctx, &ctx.gpu_top, sample_period);
init_gpu_perf(&ctx, &ctx.gpu_perf);
init_gpu_freq(&ctx, &ctx.gpu_freq);
init_gem_objects(&ctx, &ctx.gem_objects);
- sample_period = get_sample_period(&config);
-
i = 0;
while (1) {
ctx.time = time(NULL);
@@ -950,9 +951,24 @@ int main(int argc, char **argv)
show_gem_objects(&ctx, &ctx.gem_objects);
{
- char buf[80];
+ struct gpu_top *gt = &ctx.gpu_top.gpu_top;
cairo_text_extents_t extents;
+ char buf[256];
+
gethostname(buf, sizeof(buf));
+
+ if (gt->have_load_avg) {
+ int len = strlen(buf);
+
+ snprintf(buf + len, sizeof(buf) - len,
+ "%s; %u engines; load %s %.2f, %s %.2f, %s %.2f",
+ buf,
+ gt->num_rings,
+ gt->load_name[0], gt->load[0],
+ gt->load_name[1], gt->load[1],
+ gt->load_name[2], gt->load[2]);
+ }
+
cairo_set_source_rgb(ctx.cr, .5, .5, .5);
cairo_set_font_size(ctx.cr, PAD-2);
cairo_text_extents(ctx.cr, buf, &extents);
--
2.17.1
More information about the Intel-gfx
mailing list