Re: [Intel-gfx] [igt-dev] [PATCH i-g-t 2/4] i915/gem_exec_balancer: Measure timeslicing fairness

[Chris Wilson <chris@chris-wilson.co.uk>]

Quoting Tvrtko Ursulin (2020-12-14 15:51:24)
> 
> On 14/12/2020 10:51, Chris Wilson wrote:
> > Oversaturate the virtual engines on the system and check that each
> > workload receives a fair share of the available GPU time.
> 
> A simpler one to start with, I support that.
> 
> > 
> > Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
> > ---
> >   tests/i915/gem_exec_balancer.c | 154 +++++++++++++++++++++++++++++++++
> >   1 file changed, 154 insertions(+)
> > 
> > diff --git a/tests/i915/gem_exec_balancer.c b/tests/i915/gem_exec_balancer.c
> > index 35a032ccb..5efd586ad 100644
> > --- a/tests/i915/gem_exec_balancer.c
> > +++ b/tests/i915/gem_exec_balancer.c
> > @@ -2763,6 +2763,157 @@ static void smoketest(int i915, int timeout)
> >       gem_close(i915, batch[0].handle);
> >   }
> >   
> > +static uint32_t read_ctx_timestamp(int i915, uint32_t ctx)
> > +{
> > +     struct drm_i915_gem_relocation_entry reloc;
> > +     struct drm_i915_gem_exec_object2 obj = {
> > +             .handle = gem_create(i915, 4096),
> > +             .offset = 32 << 20,
> > +             .relocs_ptr = to_user_pointer(&reloc),
> > +             .relocation_count = 1,
> > +     };
> > +     struct drm_i915_gem_execbuffer2 execbuf = {
> > +             .buffers_ptr = to_user_pointer(&obj),
> > +             .buffer_count = 1,
> > +             .rsvd1 = ctx,
> > +     };
> > +     uint32_t *map, *cs;
> > +     uint32_t ts;
> > +
> > +     cs = map = gem_mmap__device_coherent(i915, obj.handle,
> > +                                          0, 4096, PROT_WRITE);
> > +
> > +     *cs++ = 0x24 << 23 | 1 << 19 | 2; /* relative SRM */
> > +     *cs++ = 0x3a8; /* CTX_TIMESTAMP */
> > +     memset(&reloc, 0, sizeof(reloc));
> > +     reloc.target_handle = obj.handle;
> > +     reloc.presumed_offset = obj.offset;
> > +     reloc.offset = offset_in_page(cs);
> > +     reloc.delta = 4000;
> > +     *cs++ = obj.offset + 4000;
> > +     *cs++ = obj.offset >> 32;
> > +
> > +     *cs++ = MI_BATCH_BUFFER_END;
> > +
> > +     gem_execbuf(i915, &execbuf);
> > +     gem_sync(i915, obj.handle);
> > +     gem_close(i915, obj.handle);
> > +
> > +     ts = map[1000];
> > +     munmap(map, 4096);
> > +
> > +     return ts;
> > +}
> > +
> > +static int cmp_u32(const void *A, const void *B)
> > +{
> > +     const uint32_t *a = A, *b = B;
> > +
> > +     if (*a < *b)
> > +             return -1;
> > +     else if (*a > *b)
> > +             return 1;
> > +     else
> > +             return 0;
> > +}
> > +
> > +static int read_ctx_timestamp_frequency(int i915)
> > +{
> > +     int value = 12500000; /* icl!!! are you feeling alright? CTX vs CS */
> > +     drm_i915_getparam_t gp = {
> > +             .value = &value,
> > +             .param = I915_PARAM_CS_TIMESTAMP_FREQUENCY,
> > +     };
> > +     if (intel_gen(intel_get_drm_devid(i915)) != 11)
> > +             ioctl(i915, DRM_IOCTL_I915_GETPARAM, &gp);
> > +     return value;
> > +}
> > +
> > +static uint64_t div64_u64_round_up(uint64_t x, uint64_t y)
> > +{
> > +     return (x + y - 1) / y;
> > +}
> > +
> > +static uint64_t ticks_to_ns(int i915, uint64_t ticks)
> > +{
> > +     return div64_u64_round_up(ticks * NSEC_PER_SEC,
> > +                               read_ctx_timestamp_frequency(i915));
> > +}
> > +
> > +static void __fairslice(int i915,
> > +                     const struct i915_engine_class_instance *ci,
> > +                     unsigned int count)
> > +{
> > +     igt_spin_t *spin = NULL;
> > +     uint32_t ctx[count + 1];
> > +     uint32_t ts[count + 1];
> > +
> > +     igt_debug("Launching %zd spinners on %s\n",
> > +               ARRAY_SIZE(ctx), class_to_str(ci->engine_class));
> > +     igt_assert(ARRAY_SIZE(ctx) >= 3);
> > +
> > +     for (int i = 0; i < ARRAY_SIZE(ctx); i++) {
> > +             ctx[i] = load_balancer_create(i915, ci, count);
> > +             if (spin == NULL) {
> > +                     spin = __igt_spin_new(i915, .ctx = ctx[i]);
> > +             } else {
> > +                     struct drm_i915_gem_execbuffer2 eb = {
> > +                             .buffer_count = 1,
> > +                             .buffers_ptr = to_user_pointer(&spin->obj[IGT_SPIN_BATCH]),
> > +                             .rsvd1 = ctx[i],
> > +                     };
> > +                     gem_execbuf(i915, &eb);
> > +             }
> > +     }
> > +
> > +     sleep(2); /* over the course of many timeslices */
> > +
> > +     igt_assert(gem_bo_busy(i915, spin->handle));
> > +     igt_spin_end(spin);
> > +     igt_debug("Cancelled spinners\n");
> > +
> > +     for (int i = 0; i < ARRAY_SIZE(ctx); i++)
> > +             ts[i] = read_ctx_timestamp(i915, ctx[i]);
> 
> Now if we had context stats ioctl. :)
> 
> > +
> > +     for (int i = 0; i < ARRAY_SIZE(ctx); i++)
> > +             gem_context_destroy(i915, ctx[i]);
> > +     igt_spin_free(i915, spin);
> > +
> > +     qsort(ts, ARRAY_SIZE(ctx), sizeof(*ts), cmp_u32);
> > +     igt_info("%s: [%.1f, %.1f, %.1f] ms, expect %1.fms\n",
> > +              class_to_str(ci->engine_class),
> > +              1e-6 * ticks_to_ns(i915, ts[0]),
> > +              1e-6 * ticks_to_ns(i915, ts[(count + 1) / 2]),
> > +              1e-6 * ticks_to_ns(i915, ts[count]),
> > +              2e3 * count / ARRAY_SIZE(ctx));
> > +
> > +     igt_assert_f(ts[count], "CTX_TIMESTAMP not reported!\n");
> > +     igt_assert_f((ts[count] - ts[0]) * 6 < ts[(count + 1) / 2],
> > +                  "Range of timeslices greater than tolerable: %.2fms > %.2fms; unfair!\n",
> > +                  1e-6 * ticks_to_ns(i915, ts[count] - ts[0]),
> > +                  1e-6 * ticks_to_ns(i915, ts[(count  + 1) / 2]) / 6);
> 
> Just put a comment saying what is the criteria please. Six ranges vs 
> median, must be some statistical thing but that's not my forte.

The usual ballpark estimate for a test that should be reliable in CI.

I guess if we model it as a drunken walk, the range would scale with the
square-root of the number of timeslices. That would suggest we should not
expect deviations more than +-50ms. I plucked ~200ms out of the air,
which is turns out to be not too terrible a safety factor. :)
-Chris
_______________________________________________
Intel-gfx mailing list
Intel-gfx@lists.freedesktop.org
https://lists.freedesktop.org/mailman/listinfo/intel-gfx