[igt-dev] [PATCH i-g-t 3/3] test/perf: Pass context id for gen12 mi rpc test

[Umesh Nerlige Ramappa]

On Mon, Nov 18, 2019 at 03:24:49PM +0200, Lionel Landwerlin wrote:
>On 12/11/2019 00:16, Umesh Nerlige Ramappa wrote:
>>On Gen12, MI RPC uses OAR. OAR is configured only for the render context
>>that wants to measure the performance. Hence a context must be passed to
>>perf in the gen12 MI RPC when compared to previous gens.
>>
>>Signed-off-by: Umesh Nerlige Ramappa <umesh.nerlige.ramappa at intel.com>
>>---
>>  tests/perf.c | 109 ++++++++++++++++++++++++++++++++++++++++++++++++++-
>>  1 file changed, 108 insertions(+), 1 deletion(-)
>>
>>diff --git a/tests/perf.c b/tests/perf.c
>>index b439e5bd..b6bc1428 100644
>>--- a/tests/perf.c
>>+++ b/tests/perf.c
>>@@ -2816,6 +2816,106 @@ test_disabled_read_error(void)
>>  	__perf_close(stream_fd);
>>  }
>>+static void
>>+gen12_test_mi_rpc(void)
>>+{
>>+	uint64_t properties[] = {
>>+		/* On Gen12, MI RPC uses OAR. OAR is configured only for the
>>+		 * render context that wants to measure the performance. Hence a
>>+		 * context must be specified in the gen12 MI RPC when compared
>>+		 * to previous gens.
>>+		 *
>>+		 * Have a random value here for the context id, but initialize
>>+		 * it once you figure out the context ID for the work to be
>>+		 * measured
>>+		 */
>>+		DRM_I915_PERF_PROP_CTX_HANDLE, UINT64_MAX,
>>+
>>+		/* OA unit configuration:
>>+		 * DRM_I915_PERF_PROP_SAMPLE_OA is no longer required for Gen12
>>+		 * because the OAR unit increments counters only for the
>>+		 * relevant context. No other parameters are needed since we do
>>+		 * not rely on the OA buffer anymore to normalize the counter
>>+		 * values.
>>+		 */
>>+		DRM_I915_PERF_PROP_OA_METRICS_SET, test_metric_set_id,
>>+		DRM_I915_PERF_PROP_OA_FORMAT, test_oa_format,
>>+	};
>>+	struct drm_i915_perf_open_param param = {
>>+		.flags = I915_PERF_FLAG_FD_CLOEXEC,
>>+		.num_properties = ARRAY_SIZE(properties) / 2,
>>+		.properties_ptr = to_user_pointer(properties),
>>+	};
>>+	drm_intel_bo *bo;
>>+	drm_intel_bufmgr *bufmgr;
>>+	drm_intel_context *context;
>>+	struct intel_batchbuffer *batch;
>>+#define INVALID_CTX_ID 0xffffffff
>>+	uint32_t ctx_id = INVALID_CTX_ID;
>>+	uint32_t *report32;
>>+	int ret;
>>+	size_t format_size_32;
>>+
>>+	/* Ensure perf_stream_paranoid is set to 1 by default */
>>+	write_u64_file("/proc/sys/dev/i915/perf_stream_paranoid", 1);
>>+
>>+	bufmgr = drm_intel_bufmgr_gem_init(drm_fd, 4096);
>>+	igt_assert(bufmgr);
>>+
>>+	drm_intel_bufmgr_gem_enable_reuse(bufmgr);
>>+
>>+	context = drm_intel_gem_context_create(bufmgr);
>>+	igt_assert(context);
>>+
>>+	ret = drm_intel_gem_context_get_id(context, &ctx_id);
>>+	igt_assert_eq(ret, 0);
>>+	igt_assert_neq(ctx_id, INVALID_CTX_ID);
>>+	properties[1] = ctx_id;
>>+
>>+	batch = intel_batchbuffer_alloc(bufmgr, devid);
>>+	bo = drm_intel_bo_alloc(bufmgr, "mi_rpc dest bo", 4096, 64);
>>+
>>+	ret = drm_intel_bo_map(bo, true);
>>+	igt_assert_eq(ret, 0);
>>+	memset(bo->virtual, 0x80, 4096);
>>+	drm_intel_bo_unmap(bo);
>>+
>>+	stream_fd = __perf_open(drm_fd, &param, false);
>>+
>>+#define REPORT_ID 0xdeadbeef
>>+#define REPORT_OFFSET 0
>>+	emit_report_perf_count(batch,
>>+			       bo,
>>+			       REPORT_OFFSET,
>>+			       REPORT_ID);
>>+	intel_batchbuffer_flush_with_context(batch, context);
>>+
>>+	ret = drm_intel_bo_map(bo, false);
>>+	igt_assert_eq(ret, 0);
>>+
>>+	report32 = bo->virtual;
>>+	format_size_32 = get_oa_format(test_oa_format).size >> 2;
>>+	dump_report(report32, format_size_32, "mi-rpc");
>>+
>>+	/* Sanity check reports
>>+	 * reportX_32[0]: report id passed with mi-rpc
>>+	 * reportX_32[1]: timestamp
>>+	 * reportX_32[format_size_32 - 1]: end of report
>>+	 * reportX_32[format_size_32]: outside report
>>+	 */
>>+	igt_assert_eq(report32[0], REPORT_ID);
>>+	igt_assert_neq(report32[1], 0);
>
>
>In theory you can get very unlucky and get a timestamp value of 0 once 
>every ~6minutes.
>

Good point. For now, slowest perf test runs for around 10 seconds and 
all tests seem to start fresh, so we may never run into this. I think I 
will add a note in the comment so that we don't miss it out if future 
tests run longer or we do not reinitialize the hw on every test. 

Thanks,
Umesh
>
>>+	igt_assert_neq(report32[format_size_32 - 1], 0x80808080);
>
>
>Again I don't know what the Test config is supposed to generate on the 
>last counter.
>
>I would pick Counter B0, it's defined as 0.

>
>
>>+	igt_assert_eq(report32[format_size_32], 0x80808080);
>>+
>>+	drm_intel_bo_unmap(bo);
>>+	drm_intel_bo_unreference(bo);
>>+	intel_batchbuffer_free(batch);
>>+	drm_intel_gem_context_destroy(context);
>>+	drm_intel_bufmgr_destroy(bufmgr);
>>+	__perf_close(stream_fd);
>>+}
>>+
>>  static void
>>  test_mi_rpc(void)
>>  {
>>@@ -4533,8 +4633,15 @@ igt_main
>>  	igt_subtest("short-reads")
>>  		test_short_reads();
>>-	igt_subtest("mi-rpc")
>>+	igt_subtest("mi-rpc") {
>>+		igt_require(intel_gen(devid) < 12);
>>  		test_mi_rpc();
>>+	}
>>+
>>+	igt_subtest("gen12-mi-rpc") {
>>+		igt_require(intel_gen(devid) >= 12);
>>+		gen12_test_mi_rpc();
>>+	}
>>  	igt_subtest("unprivileged-single-ctx-counters") {
>>  		igt_require(IS_HASWELL(devid));
>
>