[PATCH i-g-t 08/19] i915/gem_sync: Make switch-default asymmetric

[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]

 



To make the demonstration of the cheeky preemption more impactful, make
the second context a nop to contrast the first being 1024
MI_STORE_DWORD_IMM. Then if we execute and wait on the second context
before executing the first, the client latency is even more drastically
reduced.

To more clearly show any effect on wait reordering, measure the
alternative path and present both.

Signed-off-by: Chris Wilson <chris@xxxxxxxxxxxxxxxxxx>
---
 tests/i915/gem_sync.c | 40 +++++++++++++++++++++++++++++-----------
 1 file changed, 29 insertions(+), 11 deletions(-)

diff --git a/tests/i915/gem_sync.c b/tests/i915/gem_sync.c
index fb209977d..3e4feff32 100644
--- a/tests/i915/gem_sync.c
+++ b/tests/i915/gem_sync.c
@@ -651,7 +651,7 @@ switch_ring(int fd, unsigned ring, int num_children, int timeout)
 			struct drm_i915_gem_relocation_entry reloc[1024];
 			struct drm_i915_gem_execbuffer2 execbuf;
 		} contexts[2];
-		double start, elapsed;
+		double elapsed, baseline;
 		unsigned long cycles;
 
 		for (int i = 0; i < ARRAY_SIZE(contexts); i++) {
@@ -679,7 +679,7 @@ switch_ring(int fd, unsigned ring, int num_children, int timeout)
 			c->object[1].handle = gem_create(fd, sz);
 
 			c->object[1].relocs_ptr = to_user_pointer(c->reloc);
-			c->object[1].relocation_count = 1024;
+			c->object[1].relocation_count = 1024 * i;
 
 			batch = gem_mmap__cpu(fd, c->object[1].handle, 0, sz,
 					PROT_WRITE | PROT_READ);
@@ -688,7 +688,7 @@ switch_ring(int fd, unsigned ring, int num_children, int timeout)
 
 			memset(c->reloc, 0, sizeof(c->reloc));
 			b = batch;
-			for (int r = 0; r < 1024; r++) {
+			for (int r = 0; r < c->object[1].relocation_count; r++) {
 				uint64_t offset;
 
 				c->reloc[r].presumed_offset = c->object[0].offset;
@@ -722,26 +722,44 @@ switch_ring(int fd, unsigned ring, int num_children, int timeout)
 		}
 
 		cycles = 0;
-		elapsed = 0;
-		start = gettime();
-		do {
+		baseline = 0;
+		igt_until_timeout(timeout) {
 			do {
 				double this;
 
-				gem_execbuf(fd, &contexts[0].execbuf);
 				gem_execbuf(fd, &contexts[1].execbuf);
+				gem_execbuf(fd, &contexts[0].execbuf);
 
 				this = gettime();
 				gem_sync(fd, contexts[1].object[1].handle);
-				elapsed += gettime() - this;
+				gem_sync(fd, contexts[0].object[1].handle);
+				baseline += gettime() - this;
+			} while (++cycles & 1023);
+		}
+		baseline /= cycles;
+
+		cycles = 0;
+		elapsed = 0;
+		igt_until_timeout(timeout) {
+			do {
+				double this;
 
+				gem_execbuf(fd, &contexts[1].execbuf);
+				gem_execbuf(fd, &contexts[0].execbuf);
+
+				this = gettime();
 				gem_sync(fd, contexts[0].object[1].handle);
+				elapsed += gettime() - this;
+
+				gem_sync(fd, contexts[1].object[1].handle);
 			} while (++cycles & 1023);
-		} while ((gettime() - start) < timeout);
-		igt_info("%s%sompleted %ld cycles: %.3f us\n",
+		}
+		elapsed /= cycles;
+
+		igt_info("%s%sompleted %ld cycles: %.3f us, baseline %.3f us\n",
 			 names[child % num_engines] ?: "",
 			 names[child % num_engines] ? " c" : "C",
-			 cycles, elapsed*1e6/cycles);
+			 cycles, elapsed*1e6, baseline*1e6);
 
 		for (int i = 0; i < ARRAY_SIZE(contexts); i++) {
 			gem_close(fd, contexts[i].object[1].handle);
-- 
2.20.1

_______________________________________________
Intel-gfx mailing list
Intel-gfx@xxxxxxxxxxxxxxxxxxxxx
https://lists.freedesktop.org/mailman/listinfo/intel-gfx




[Index of Archives]     [AMD Graphics]     [Linux USB Devel]     [Linux Audio Users]     [Yosemite News]     [Linux Kernel]     [Linux SCSI]

  Powered by Linux