On 04/23/2015 11:39 AM, Chris Wilson wrote:
On Thu, Apr 23, 2015 at 11:08:35AM +0100, Tvrtko Ursulin wrote:
+/*
+ * Test that tiled page flips work.
Worth saying that we don't always expect linear->tiled flips to work, so
the mo is to first set the CRTC with a tiled fb, then flip to the
second.
I didn't have that in mind for this test - here it only concerns itself
with skl_do_mmio_flip bug (2ebef630fd283642a11c48c0e0f054c3c5c59e86).
+ */
+static void
+test_tiled_flip(data_t *data, igt_output_t *output, uint64_t tiling)
+{
+ drmModeModeInfo *mode;
+ igt_plane_t *primary;
+ igt_pipe_crc_t *pipe_crc;
+ igt_crc_t reference_crc, crc;
+ struct igt_fb fb[2];
+ int fb_id[2], pipe, ret, width, height;
+
+ pipe = output->config.pipe;
+ pipe_crc = igt_pipe_crc_new(pipe, INTEL_PIPE_CRC_SOURCE_AUTO);
+ igt_output_set_pipe(output, pipe);
+
+ mode = igt_output_get_mode(output);
+ primary = igt_output_get_plane(output, 0);
+
+ width = mode->hdisplay;
+ height = mode->vdisplay;
+
+ /* create two identical fbs */
+ fb_id[0] = igt_create_fb(data->drm_fd, width, height,
+ DRM_FORMAT_XRGB8888, tiling, &fb[0]);
+ igt_assert(fb_id[0]);
+
+ fb_id[1] = igt_create_fb(data->drm_fd, width, height,
+ DRM_FORMAT_XRGB8888, tiling, &fb[1]);
+ igt_assert(fb_id[1]);
+
+ fill_fb(&fb[0], data, mode);
+ fill_fb(&fb[1], data, mode);
Pretty please can we have two different patterns?
+ /* set the crtc and generate a reference crc */
+ igt_plane_set_fb(primary, &fb[0]);
+ igt_display_commit(&data->display);
+ igt_pipe_crc_collect_crc(pipe_crc, &reference_crc);
+
+ /* flip to the other fb */
+ ret = drmModePageFlip(data->drm_fd, output->config.crtc->crtc_id,
+ fb_id[1], 0, NULL);
+ igt_assert_eq(ret, 0);
+
+ igt_wait_for_vblank(data->drm_fd, pipe);
I'd be more comfortable waiting for the flip-completion event (and then
a vblank for the CRC to be computed? idk)
I can look into that but shame on you for ruining my copy&paste party! ;)
+ /* get a crc and compare with the reference */
+ igt_pipe_crc_collect_crc(pipe_crc, &crc);
+ igt_assert_crc_equal(&reference_crc, &crc);
With 2 patterns you can even test the second flip back again!
So for detecting the current breakage, we should just a
test_tiled_change that if PageFlip returns 0, the crc matches the
reference? And iterate over the modes to find one that causes an
underrun?
Underruns and watermarks I intended for another test case. This one is a
really simple one - flip with the same framebuffer and check that it has
really remained the same.
And the pre-existent test_flip_changes_tiling works in the opposite
direction so won't cause an underrun. But will start failing if we merge
(drm/i915/skl: Disallow tiling changes during page flip).
So third test case could be what you say, flip from linear to tiled and
check for underrun if it works.
Regards,
Tvrtko
_______________________________________________
Intel-gfx mailing list
Intel-gfx@xxxxxxxxxxxxxxxxxxxxx
http://lists.freedesktop.org/mailman/listinfo/intel-gfx
