[Intel-gfx] [PATCH i-g-t 01/24] i915/gem_exec_latency: Measure the latency of context switching

[Chris Wilson]

Measure the baseline latency between contexts in order to directly
compare that with the additional cost of preemption.

Signed-off-by: Chris Wilson <chris at chris-wilson.co.uk>
---
 tests/i915/gem_exec_latency.c | 145 ++++++++++++++++++++++++++++++++++
 1 file changed, 145 insertions(+)

diff --git a/tests/i915/gem_exec_latency.c b/tests/i915/gem_exec_latency.c
index f308e0851..04ec7d9d2 100644
--- a/tests/i915/gem_exec_latency.c
+++ b/tests/i915/gem_exec_latency.c
@@ -629,6 +629,142 @@ rthog_latency_on_ring(int fd, unsigned int engine, const char *name, unsigned in
 	munmap(results, MMAP_SZ);
 }
 
+static void context_switch(int i915,
+			   unsigned int engine, const char *name,
+			   unsigned int flags)
+{
+	struct drm_i915_gem_exec_object2 obj[2];
+	struct drm_i915_gem_relocation_entry reloc[5];
+	struct drm_i915_gem_execbuffer2 eb;
+	uint32_t *cs, *bbe, *results, v;
+	unsigned int mmio_base;
+	struct igt_mean mean;
+	uint32_t ctx[2];
+
+	/* XXX i915_query()! */
+	switch (engine) {
+	case I915_EXEC_DEFAULT:
+	case I915_EXEC_RENDER:
+		mmio_base = 0x2000;
+		break;
+#if 0
+	case I915_EXEC_BSD:
+		mmio_base = 0x12000;
+		break;
+#endif
+	case I915_EXEC_BLT:
+		mmio_base = 0x22000;
+		break;
+
+	case I915_EXEC_VEBOX:
+		if (intel_gen(intel_get_drm_devid(i915)) >= 11)
+			mmio_base = 0x1d8000;
+		else
+			mmio_base = 0x1a000;
+		break;
+
+	default:
+		igt_skip("mmio base not known\n");
+	}
+
+	for (int i = 0; i < ARRAY_SIZE(ctx); i++)
+		ctx[i] = gem_context_create(i915);
+
+	if (flags & PREEMPT) {
+		gem_context_set_priority(i915, ctx[0], -1023);
+		gem_context_set_priority(i915, ctx[1], +1023);
+	}
+
+	memset(obj, 0, sizeof(obj));
+	obj[0].handle = gem_create(i915, 4096);
+	gem_set_caching(i915, obj[0].handle, 1);
+	results = gem_mmap__cpu(i915, obj[0].handle, 0, 4096, PROT_READ);
+	gem_set_domain(i915, obj[0].handle, I915_GEM_DOMAIN_CPU, 0);
+
+	obj[1].handle = gem_create(i915, 4096);
+	memset(reloc,0, sizeof(reloc));
+	obj[1].relocation_count = ARRAY_SIZE(reloc);
+	obj[1].relocs_ptr = to_user_pointer(reloc);
+	bbe = gem_mmap__wc(i915, obj[1].handle, 0, 4096, PROT_WRITE);
+	gem_set_domain(i915, obj[1].handle,
+		       I915_GEM_DOMAIN_WC, I915_GEM_DOMAIN_WC);
+
+	cs = bbe;
+	*cs++ = 0x5 << 23;
+	*cs++ = 0x24 << 23 | 2; /* SRM */
+	*cs++ = mmio_base + 0x358; /* TIMESTAMP */
+	reloc[0].target_handle = obj[0].handle;
+	reloc[0].offset = (cs - bbe) * sizeof(*cs);
+	*cs++ = 0;
+	*cs++ = 0;
+	*cs++ = MI_BATCH_BUFFER_START | 1;
+	reloc[1].target_handle = obj[1].handle;
+	reloc[1].offset = (cs - bbe) * sizeof(*cs);
+	*cs++ = 0;
+	*cs++ = 0;
+
+	cs = bbe + 64;
+	*cs++ = 0x24 << 23 | 2; /* SRM */
+	*cs++ = mmio_base + 0x358; /* TIMESTAMP */
+	reloc[2].target_handle = obj[0].handle;
+	reloc[2].offset = (cs - bbe) * sizeof(*cs);
+	*cs++ = reloc[2].delta = 4;
+	*cs++ = 0;
+	*cs++ = 0x29 << 23 | 2; /* LRM */
+	*cs++ = mmio_base + 0x600; /* GPR0 */
+	reloc[3].target_handle = obj[0].handle;
+	reloc[3].offset = (cs - bbe) * sizeof(*cs);
+	*cs++ = 0;
+	*cs++ = 0;
+	*cs++ = 0x24 << 23 | 2; /* SRM */
+	*cs++ = mmio_base + 0x600; /* GPR0 */
+	reloc[4].target_handle = obj[0].handle;
+	reloc[4].offset = (cs - bbe) * sizeof(*cs);
+	*cs++ = reloc[4].delta = 8;
+	*cs++ = 0;
+	*cs++ = 0xa << 23;
+
+	memset(&eb, 0, sizeof(eb));
+	eb.buffers_ptr = to_user_pointer(obj);
+	eb.buffer_count = ARRAY_SIZE(obj);
+	eb.flags = engine;
+	eb.flags |= LOCAL_I915_EXEC_NO_RELOC;
+
+	v = 0;
+	igt_mean_init(&mean);
+	igt_until_timeout(5) {
+		eb.rsvd1 = ctx[0];
+		eb.batch_start_offset = 0;
+		gem_execbuf(i915, &eb);
+
+		while (results[0] == v)
+			igt_assert(gem_bo_busy(i915, obj[1].handle));
+
+		eb.rsvd1 = ctx[1];
+		eb.batch_start_offset = 64 * sizeof(*cs);
+		gem_execbuf(i915, &eb);
+
+		*bbe = 0xa << 23;
+		gem_sync(i915, obj[1].handle);
+		*bbe = 0x5 << 23;
+
+		v = results[0];
+		igt_mean_add(&mean, (results[1] - results[2]) * rcs_clock);
+	}
+	igt_info("%s context switch latency%s: %.2f±%.2fus\n",
+		 name, flags & PREEMPT ? " (preempt)" : "",
+		 1e-3 * igt_mean_get(&mean),
+		 1e-3 * sqrt(igt_mean_get_variance(&mean)));
+	munmap(results, 4096);
+	munmap(bbe, 4096);
+
+	for (int i = 0; i < ARRAY_SIZE(obj); i++)
+		gem_close(i915, obj[i].handle);
+
+	for (int i = 0; i < ARRAY_SIZE(ctx); i++)
+		gem_context_destroy(i915, ctx[i]);
+}
+
 static double clockrate(int i915, int reg)
 {
 	volatile uint32_t *mmio;
@@ -754,12 +890,21 @@ igt_main
 							  e->exec_id | e->flags,
 							  e->name, CORK);
 
+				igt_subtest_f("%s-cs", e->name)
+					context_switch(device,
+						       e->exec_id | e->flags,
+						       e->name, 0);
 				igt_subtest_group {
 					igt_fixture {
 						gem_require_contexts(device);
 						igt_require(gem_scheduler_has_preemption(device));
 					}
 
+					igt_subtest_f("%s-cs-preempt", e->name)
+						context_switch(device,
+								e->exec_id | e->flags,
+								e->name, PREEMPT);
+
 					igt_subtest_f("%s-preemption", e->name)
 						latency_from_ring(device,
 								  e->exec_id | e->flags,
-- 
2.20.1