Quoting Tvrtko Ursulin (2018-04-23 17:52:54)
>
> On 23/04/2018 14:43, Chris Wilson wrote:
> > In the existing ABI, each engine operates its own timeline
> > (fence.context) and so should execute independently of any other. If we
> > install a blocker on all other engines, that should not affect execution
> > on the local engine.
> >
> > Signed-off-by: Chris Wilson <chris@xxxxxxxxxxxxxxxxxx>
> > Cc: Tvrtko Ursulin <tvrtko.ursulin@xxxxxxxxx>
> > ---
> > tests/gem_exec_schedule.c | 90 +++++++++++++++++++++++++++++++++++----
> > 1 file changed, 82 insertions(+), 8 deletions(-)
> >
> > diff --git a/tests/gem_exec_schedule.c b/tests/gem_exec_schedule.c
> > index 5d0f215b2..471275169 100644
> > --- a/tests/gem_exec_schedule.c
> > +++ b/tests/gem_exec_schedule.c
> > @@ -49,9 +49,9 @@
> >
> > IGT_TEST_DESCRIPTION("Check that we can control the order of execution");
> >
> > -static void store_dword(int fd, uint32_t ctx, unsigned ring,
> > - uint32_t target, uint32_t offset, uint32_t value,
> > - uint32_t cork, unsigned write_domain)
> > +static uint32_t __store_dword(int fd, uint32_t ctx, unsigned ring,
> > + uint32_t target, uint32_t offset, uint32_t value,
> > + uint32_t cork, unsigned write_domain)
> > {
> > const int gen = intel_gen(intel_get_drm_devid(fd));
> > struct drm_i915_gem_exec_object2 obj[3];
> > @@ -100,7 +100,17 @@ static void store_dword(int fd, uint32_t ctx, unsigned ring,
> > batch[++i] = MI_BATCH_BUFFER_END;
> > gem_write(fd, obj[2].handle, 0, batch, sizeof(batch));
> > gem_execbuf(fd, &execbuf);
> > - gem_close(fd, obj[2].handle);
> > +
> > + return obj[2].handle;
> > +}
> > +
> > +static void store_dword(int fd, uint32_t ctx, unsigned ring,
> > + uint32_t target, uint32_t offset, uint32_t value,
> > + uint32_t cork, unsigned write_domain)
> > +{
> > + gem_close(fd, __store_dword(fd, ctx, ring,
> > + target, offset, value,
> > + cork, write_domain));
> > }
> >
> > static uint32_t create_highest_priority(int fd)
> > @@ -161,6 +171,64 @@ static void fifo(int fd, unsigned ring)
> > munmap(ptr, 4096);
> > }
> >
> > +static void independent(int fd, unsigned int engine)
> > +{
> > + IGT_CORK_HANDLE(cork);
> > + uint32_t scratch, plug, batch;
> > + igt_spin_t *spin = NULL;
> > + unsigned int other;
> > + uint32_t *ptr;
> > +
> > + igt_require(engine != 0);
> > +
> > + scratch = gem_create(fd, 4096);
> > + plug = igt_cork_plug(&cork, fd);
> > +
> > + /* Check that we can submit to engine while all others are blocked */
> > + for_each_physical_engine(fd, other) {
> > + if (other == engine)
> > + continue;
> > +
> > + if (spin == NULL) {
> > + spin = __igt_spin_batch_new(fd, 0, other, 0);
> > + } else {
> > + struct drm_i915_gem_exec_object2 obj = {
> > + .handle = spin->handle,
> > + };
> > + struct drm_i915_gem_execbuffer2 eb = {
> > + .buffer_count = 1,
> > + .buffers_ptr = to_user_pointer(&obj),
> > + .flags = other,
> > + };
> > + gem_execbuf(fd, &eb);
> > + }
> > +
> > + store_dword(fd, 0, other, scratch, 0, other, plug, 0);
> > + }
> > + igt_require(spin);
> > +
> > + /* Same priority, but different timeline (as different engine) */
> > + batch = __store_dword(fd, 0, engine, scratch, 0, engine, plug, 0);
> > +
> > + unplug_show_queue(fd, &cork, engine);
> > + gem_close(fd, plug);
> > +
> > + gem_sync(fd, batch);
> > + gem_close(fd, batch);
>
> Strictly speaking I think you need to use the poll-able spinner and wait
> on it here, before the busy assert. It's unlikely, but spinners on
> 'other' engines are getting submitted async to the store dword batch on
> 'engine'.
We've waited for its completion, so we know batch is idle and the others
are still busy. We then check its seqno is written to the scratch; so
using pollable here is redundant. And then we check that the others are
run after.
-Chris
_______________________________________________
Intel-gfx mailing list
Intel-gfx@xxxxxxxxxxxxxxxxxxxxx
https://lists.freedesktop.org/mailman/listinfo/intel-gfx
