[Intel-gfx] [PATCH v2 i-g-t] kms_atomic: Measure speed of some plane ioctls

[Chris Wilson]

On Mon, Apr 27, 2015 at 04:34:05PM +0100, Tvrtko Ursulin wrote:
> +static void prepare_crtc(data_t *data, igt_output_t *output, enum pipe pipe,
> +			 igt_plane_t *plane)
> +{
> +	drmModeModeInfo *mode;
> +	igt_display_t *display = &data->display;
> +	int fb_id, fb_modeset_id;
> +	unsigned int w, h;
> +	uint64_t tiling = LOCAL_DRM_FORMAT_MOD_NONE;
> +	uint32_t pixel_format = DRM_FORMAT_XRGB8888;
> +	enum igt_commit_style commit = COMMIT_LEGACY;
> +	igt_plane_t *primary;
> +
> +	igt_output_set_pipe(output, pipe);
> +
> +	mode = igt_output_get_mode(output);
> +
> +	w = mode->hdisplay;
> +	h = mode->vdisplay;
> +
> +	fb_modeset_id = igt_create_fb(data->gfx_fd,
> +				      w, h,
> +				      pixel_format,
> +				      tiling,
> +				      &data->fb_modeset);
> +	igt_assert(fb_modeset_id);
> +
> +	/*
> +	 * With igt_display_commit2 and COMMIT_UNIVERSAL, we call just the
> +	 * setplane without a modeset. So, to be able to call
> +	 * igt_display_commit and ultimately setcrtc to do the first modeset,
> +	 * we create an fb covering the crtc and call commit
> +	 */
> +
> +	primary = igt_output_get_plane(output, IGT_PLANE_PRIMARY);
> +	igt_plane_set_fb(primary, &data->fb_modeset);
> +	igt_display_commit(display);
> +
> +	fb_id = igt_create_fb(data->gfx_fd,
> +			      w, h,
> +			      pixel_format,
> +			      tiling,
> +			      &data->fb);
> +	igt_assert(fb_id);
> +
> +	igt_plane_set_fb(plane, &data->fb);
> +

Please do commit = COMMIT_LEGACY here so I don't have to scan backwads.
> +	if (plane->is_primary || plane->is_cursor) {
> +		igt_require(data->display.has_universal_planes);
> +		commit = COMMIT_UNIVERSAL;
> +	}
> +
> +	igt_display_commit2(display, commit);
> +}
> +
> +static void cleanup_crtc(data_t *data, igt_output_t *output, igt_plane_t *plane)
> +{
> +	igt_display_t *display = &data->display;
> +
> +	igt_remove_fb(data->gfx_fd, &data->fb);
> +	igt_remove_fb(data->gfx_fd, &data->fb_modeset);
> +
> +	/* XXX: see the note in prepare_crtc() */
> +	if (!plane->is_primary) {
> +		igt_plane_t *primary;
> +
> +		primary = igt_output_get_plane(output, IGT_PLANE_PRIMARY);
> +		igt_plane_set_fb(primary, NULL);
> +	}
> +
> +	igt_plane_set_fb(plane, NULL);
> +	igt_output_set_pipe(output, PIPE_ANY);
> +
> +	igt_display_commit(display);
> +}
> +
> +static double elapsed(const struct timeval *start,
> +		      const struct timeval *end,
> +		      int loop)
> +{
> +	return (1e6*(end->tv_sec - start->tv_sec) + (end->tv_usec - start->tv_usec))/loop;
> +}
> +
> +static void test_commit_speed(data_t *data, enum igt_plane plane_type)
> +{
> +	igt_display_t *display = &data->display;
> +	igt_output_t *output;
> +	enum pipe pipe;
> +	enum igt_commit_style commit = COMMIT_LEGACY;
> +	unsigned int i;
> +	const unsigned int loops = 10000;
> +	struct timeval start, end;
> +	double e;
> +

Same comment for commit = COMMIT_LEGACY here, mainly now for consistency
with before.
> +	if (plane_type == IGT_PLANE_PRIMARY || plane_type == IGT_PLANE_CURSOR) {
> +		igt_require(data->display.has_universal_planes);
> +		commit = COMMIT_UNIVERSAL;
> +	}
> +
> +	for_each_connected_output(display, output) {
> +		for_each_pipe(display, pipe) {
> +			igt_plane_t *plane;
> +
> +			igt_output_set_pipe(output, pipe);
> +			plane = igt_output_get_plane(output, plane_type);
> +
> +			prepare_crtc(data, output, pipe, plane);
> +
> +			igt_display_commit2(display, commit);
> +
> +			gettimeofday(&start, NULL);
> +			for (i = loops; i > 0; i--) {
> +				plane->position_changed = true;
> +				igt_plane_commit(plane, output, commit, true);
> +			}
> +			gettimeofday(&end, NULL);
> +			e = elapsed(&start, &end, loops);
> +			igt_info("Pipe: %u Plane: %u Time: %7.3fµs\n",
> +				 pipe, plane->index, e);
> +			igt_require(e < 1000); /* 1ms, just so. */
> +
> +			kmstest_restore_vt_mode();
> +			kmstest_set_vt_graphics_mode();

Do you really mean to switch VT between text/gfx on every display/pipe?
Why?

> +
> +			cleanup_crtc(data, output, plane);
> +		}
> +	}
> +}

Looks like magic to me. Is this enough variation to flush out future
bugs?
-Chris

-- 
Chris Wilson, Intel Open Source Technology Centre