[Intel-gfx] [PATCH i-g-t 4/5] tests/perf_pmu: Add tests for engine queued/runnable/running stats

[Tvrtko Ursulin <tursulin@ursulin.net>]

From: Tvrtko Ursulin <tvrtko.ursulin@intel.com>

Simple tests to check reported queue depths are correct.

Signed-off-by: Tvrtko Ursulin <tvrtko.ursulin@intel.com>
---
 tests/perf_pmu.c | 224 +++++++++++++++++++++++++++++++++++++++++++++++++++++++
 1 file changed, 224 insertions(+)

diff --git a/tests/perf_pmu.c b/tests/perf_pmu.c
index 469b9becdbac..206c18960b7b 100644
--- a/tests/perf_pmu.c
+++ b/tests/perf_pmu.c
@@ -966,6 +966,196 @@ multi_client(int gem_fd, const struct intel_execution_engine2 *e)
 	assert_within_epsilon(val[1], perf_slept[1], tolerance);
 }
 
+static double calc_queued(uint64_t d_val, uint64_t d_ns)
+{
+	return (double)d_val * 1e9 / I915_SAMPLE_QUEUED_DIVISOR / d_ns;
+}
+
+static void
+queued(int gem_fd, const struct intel_execution_engine2 *e)
+{
+	const unsigned long engine = e2ring(gem_fd, e);
+	const unsigned int max_rq = 10;
+	double queued[max_rq + 1];
+	uint32_t bo[max_rq + 1];
+	unsigned int n, i;
+	uint64_t val[2];
+	uint64_t ts[2];
+	int fd;
+
+	memset(queued, 0, sizeof(queued));
+	memset(bo, 0, sizeof(bo));
+
+	fd = open_pmu(I915_PMU_ENGINE_QUEUED(e->class, e->instance));
+
+	for (n = 0; n <= max_rq; n++) {
+		int fence = -1;
+		struct igt_cork cork = { .fd = fence, .type = CORK_SYNC_FD };
+
+		gem_quiescent_gpu(gem_fd);
+
+		if (n)
+			fence = igt_cork_plug(&cork, -1);
+
+		for (i = 0; i < n; i++) {
+			struct drm_i915_gem_exec_object2 obj = { };
+			struct drm_i915_gem_execbuffer2 eb = { };
+
+			if (!bo[i]) {
+				const uint32_t bbe = MI_BATCH_BUFFER_END;
+
+				bo[i] = gem_create(gem_fd, 4096);
+				gem_write(gem_fd, bo[i], 4092, &bbe,
+					  sizeof(bbe));
+			}
+
+			obj.handle = bo[i];
+
+			eb.buffer_count = 1;
+			eb.buffers_ptr = to_user_pointer(&obj);
+
+			eb.flags = engine | I915_EXEC_FENCE_IN;
+			eb.rsvd2 = fence;
+
+			gem_execbuf(gem_fd, &eb);
+		}
+
+		val[0] = __pmu_read_single(fd, &ts[0]);
+		usleep(batch_duration_ns / 1000);
+		val[1] = __pmu_read_single(fd, &ts[1]);
+
+		queued[n] = calc_queued(val[1] - val[0], ts[1] - ts[0]);
+		igt_info("n=%u queued=%.2f\n", n, queued[n]);
+
+		if (fence >= 0)
+			igt_cork_unplug(&cork);
+
+		for (i = 0; i < n; i++)
+			gem_sync(gem_fd, bo[i]);
+	}
+
+	close(fd);
+
+	for (i = 0; i < max_rq; i++) {
+		if (bo[i])
+			gem_close(gem_fd, bo[i]);
+	}
+
+	for (i = 0; i <= max_rq; i++)
+		assert_within_epsilon(queued[i], i, tolerance);
+}
+
+static void
+runnable(int gem_fd, const struct intel_execution_engine2 *e)
+{
+	const unsigned long engine = e2ring(gem_fd, e);
+	const unsigned int max_rq = 10;
+	igt_spin_t *spin[max_rq + 1];
+	double runnable[max_rq + 1];
+	uint32_t ctx[max_rq];
+	unsigned int n, i;
+	uint64_t val[2];
+	uint64_t ts[2];
+	int fd;
+
+	memset(runnable, 0, sizeof(runnable));
+	memset(ctx, 0, sizeof(ctx));
+
+	fd = open_pmu(I915_PMU_ENGINE_RUNNABLE(e->class, e->instance));
+
+	for (n = 0; n <= max_rq; n++) {
+		gem_quiescent_gpu(gem_fd);
+
+		for (i = 0; i < n; i++) {
+			if (!ctx[i])
+				ctx[i] = gem_context_create(gem_fd);
+
+			if (i == 0)
+				spin[i] = __spin_poll(gem_fd, ctx[i], engine);
+			else
+				spin[i] = __igt_spin_batch_new(gem_fd, ctx[i],
+							       engine, 0);
+		}
+
+		if (n)
+			__spin_wait(gem_fd, spin[0]);
+
+		val[0] = __pmu_read_single(fd, &ts[0]);
+		usleep(batch_duration_ns / 1000);
+		val[1] = __pmu_read_single(fd, &ts[1]);
+
+		runnable[n] = calc_queued(val[1] - val[0], ts[1] - ts[0]);
+		igt_info("n=%u runnable=%.2f\n", n, runnable[n]);
+
+		for (i = 0; i < n; i++) {
+			end_spin(gem_fd, spin[i], FLAG_SYNC);
+			igt_spin_batch_free(gem_fd, spin[i]);
+		}
+	}
+
+	for (i = 0; i < max_rq; i++) {
+		if (ctx[i])
+			gem_context_destroy(gem_fd, ctx[i]);
+	}
+
+	close(fd);
+
+	assert_within_epsilon(runnable[0], 0, tolerance);
+	igt_assert(runnable[max_rq] > 0.0);
+	assert_within_epsilon(runnable[max_rq] - runnable[max_rq - 1], 1,
+			      tolerance);
+}
+
+static void
+running(int gem_fd, const struct intel_execution_engine2 *e)
+{
+	const unsigned long engine = e2ring(gem_fd, e);
+	const unsigned int max_rq = 10;
+	igt_spin_t *spin[max_rq + 1];
+	double running[max_rq + 1];
+	unsigned int n, i;
+	uint64_t val[2];
+	uint64_t ts[2];
+	int fd;
+
+	memset(running, 0, sizeof(running));
+	memset(spin, 0, sizeof(spin));
+
+	fd = open_pmu(I915_PMU_ENGINE_RUNNING(e->class, e->instance));
+
+	for (n = 0; n <= max_rq; n++) {
+		gem_quiescent_gpu(gem_fd);
+
+		for (i = 0; i < n; i++) {
+			if (i == 0)
+				spin[i] = __spin_poll(gem_fd, 0, engine);
+			else
+				spin[i] = __igt_spin_batch_new(gem_fd, 0,
+							       engine, 0);
+		}
+
+		if (n)
+			__spin_wait(gem_fd, spin[0]);
+
+		val[0] = __pmu_read_single(fd, &ts[0]);
+		usleep(batch_duration_ns / 1000);
+		val[1] = __pmu_read_single(fd, &ts[1]);
+
+		running[n] = calc_queued(val[1] - val[0], ts[1] - ts[0]);
+		igt_info("n=%u running=%.2f\n", n, running[n]);
+
+		for (i = 0; i < n; i++) {
+			end_spin(gem_fd, spin[i], FLAG_SYNC);
+			igt_spin_batch_free(gem_fd, spin[i]);
+		}
+	}
+
+	close(fd);
+
+	for (i = 0; i <= max_rq; i++)
+		assert_within_epsilon(running[i], i, tolerance);
+}
+
 /**
  * Tests that i915 PMU corectly errors out in invalid initialization.
  * i915 PMU is uncore PMU, thus:
@@ -1718,6 +1908,15 @@ igt_main
 		igt_subtest_f("init-sema-%s", e->name)
 			init(fd, e, I915_SAMPLE_SEMA);
 
+		igt_subtest_f("init-queued-%s", e->name)
+			init(fd, e, I915_SAMPLE_QUEUED);
+
+		igt_subtest_f("init-runnable-%s", e->name)
+			init(fd, e, I915_SAMPLE_RUNNABLE);
+
+		igt_subtest_f("init-running-%s", e->name)
+			init(fd, e, I915_SAMPLE_RUNNING);
+
 		igt_subtest_group {
 			igt_fixture {
 				gem_require_engine(fd, e->class, e->instance);
@@ -1823,6 +2022,24 @@ igt_main
 
 			igt_subtest_f("busy-hang-%s", e->name)
 				single(fd, e, TEST_BUSY | FLAG_HANG);
+
+			/**
+			 * Test that queued metric works.
+			 */
+			igt_subtest_f("queued-%s", e->name)
+				queued(fd, e);
+
+			/**
+			 * Test that runnable metric works.
+			 */
+			igt_subtest_f("runnable-%s", e->name)
+				runnable(fd, e);
+
+			/**
+			 * Test that running metric works.
+			 */
+			igt_subtest_f("running-%s", e->name)
+				running(fd, e);
 		}
 
 		/**
@@ -1915,6 +2132,13 @@ igt_main
 					      e->name)
 					single(fd, e,
 					       TEST_BUSY | TEST_TRAILING_IDLE);
+				igt_subtest_f("render-node-queued-%s", e->name)
+					queued(fd, e);
+				igt_subtest_f("render-node-runnable-%s",
+					      e->name)
+					runnable(fd, e);
+				igt_subtest_f("render-node-running-%s", e->name)
+					running(fd, e);
 			}
 		}
 
-- 
2.14.1

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