On 23/05/2019 15:09, Chris Wilson wrote:
Quoting Chris Wilson (2019-05-23 15:05:05)
Quoting Tvrtko Ursulin (2019-05-22 16:57:12)
-static void
+static unsigned int
terminate_bb(struct w_step *w, unsigned int flags)
{
const uint32_t bbe = 0xa << 23;
unsigned long mmap_start, mmap_len;
unsigned long batch_start = w->bb_sz;
+ unsigned int r = 0;
uint32_t *ptr, *cs;
igt_assert(((flags & RT) && (flags & SEQNO)) || !(flags & RT));
@@ -838,6 +854,9 @@ terminate_bb(struct w_step *w, unsigned int flags)
if (flags & RT)
batch_start -= 12 * sizeof(uint32_t);
+ if (w->unbound_duration)
+ batch_start -= 4 * sizeof(uint32_t); /* MI_ARB_CHK + MI_BATCH_BUFFER_START */
+
mmap_start = rounddown(batch_start, PAGE_SIZE);
mmap_len = ALIGN(w->bb_sz - mmap_start, PAGE_SIZE);
@@ -847,8 +866,19 @@ terminate_bb(struct w_step *w, unsigned int flags)
ptr = gem_mmap__wc(fd, w->bb_handle, mmap_start, mmap_len, PROT_WRITE);
cs = (uint32_t *)((char *)ptr + batch_start - mmap_start);
+ if (w->unbound_duration) {
+ w->reloc[r++].offset = batch_start + 2 * sizeof(uint32_t);
+ batch_start += 4 * sizeof(uint32_t);
+
+ *cs++ = w->preempt_us ? 0x5 << 23 /* MI_ARB_CHK; */ : MI_NOOP;
+ w->recursive_bb_start = cs;
+ *cs++ = MI_BATCH_BUFFER_START | 1 << 8 | 1;
+ *cs++ = 0;
+ *cs++ = 0;
delta is zero, and mmap_len is consistent, so yup this gives a page of
nops before looping.
Waitasec... Only emitting ARB_CHK if preempt_us is set. You want an
unbounded unpreemptible batch?
I suppose it's not without use (although I plan for it to be killed
quickly and reset), but I would not advise for preempt_us to be 0 by
default then.
Default is 100us (search for "w->preempt_us = 100;") and then for
instance frame-split-60fps.wsim disables it to simulate media better
("X.1.0", "X.2.0"). AFAIR you actually asked for preemptable by default.
Now that I think of it.. I really need a command line switch to control
it globally. Or better per engine class to simulate media better.
Blast.. feature creep never ends.
Regards,
Tvrtko
_______________________________________________
Intel-gfx mailing list
Intel-gfx@xxxxxxxxxxxxxxxxxxxxx
https://lists.freedesktop.org/mailman/listinfo/intel-gfx
