Re: Horribly slow rendering using Quartz surfaces

On Tue, Aug 6, 2013 at 4:34 PM, Bobby Salazar <[email protected]> wrote:

> I am testing some code on iOS and have run into a rendering problem
> using Quartz surfaces that does not happen by merely switching the
> surface out to an image surface. Here is the code:
>
> void TestRadialRender()
> {
>         cairo_surface_t* pSurface;
>         cairo_t* pContext;
>         cairo_matrix_t Matrix;
>
>         // Setup surface
>         //pSurface = cairo_image_surface_create(CAIRO_FORMAT_ARGB32, 640,
> 1136);
>         pSurface = cairo_quartz_surface_create(CAIRO_FORMAT_ARGB32, 640,
> 1136);
>         pContext = cairo_create(pSurface);
>
>         // Setup matrix
>         cairo_matrix_init(&Matrix, 3.6, 0.0, 0.0, 1.35, 568, 66);
>         cairo_set_matrix(pContext, &Matrix);
>
>         // Setup path
>         cairo_new_path(pContext);
>         double cx = -45.84799;
>         double cy = 7.394;
>         double rx = 3;
>         double ry = 3;
>
>         cairo_move_to (pContext, cx + rx, cy);
>         cairo_curve_to (pContext, cx + rx, cy + ry * 0.5522847498, cx + rx
> *
> 0.5522847498, cy + ry, cx, cy + ry);
>         cairo_curve_to (pContext, cx - rx * 0.5522847498, cy + ry, cx - rx,
> cy + ry * 0.5522847498, cx - rx, cy);
>         cairo_curve_to (pContext, cx - rx, cy - ry * 0.5522847498, cx - rx
> *
> 0.5522847498, cy - ry, cx, cy - ry);
>         cairo_curve_to (pContext, cx + rx * 0.5522847498, cy - ry, cx + rx,
> cy - ry * 0.5522847498, cx + rx, cy);
>         cairo_close_path (pContext);
>
>
>         // Setup pattern
>         cairo_pattern_t*        pat;
>         pat = cairo_pattern_create_radial(-45.84799, 7.394, 0, -45.84799,
> 7.394, 3);
>         cairo_pattern_set_extend(pat, CAIRO_EXTEND_REFLECT);
>
>         // Setup brush color
>         cairo_pattern_add_color_stop_rgba(pat, 0, 0.00, 0.00, 1.00, 0.125);
>
>         // Setup color stops
>         cairo_pattern_add_color_stop_rgba(pat, 0.0000, 0.54, 0.55, 0.56,
> 1.00);
>         cairo_pattern_add_color_stop_rgba(pat, 0.1272, 0.67, 0.68, 0.69,
> 1.00);
>         cairo_pattern_add_color_stop_rgba(pat, 0.2750, 0.78, 0.79, 0.80,
> 1.00);
>         cairo_pattern_add_color_stop_rgba(pat, 0.4307, 0.88, 0.88, 0.89,
> 0.50);
>         cairo_pattern_add_color_stop_rgba(pat, 0.5955, 0.95, 0.95, 0.95,
> 0.25);
>         cairo_pattern_add_color_stop_rgba(pat, 0.7757, 0.98, 0.98, 0.98,
> 0.125);
>         cairo_pattern_add_color_stop_rgba(pat, 1.0000, 1.00, 1.00, 1.00,
> 0.00);
>
>         // Render it
>         cairo_set_source(pContext, pat);
>
>         //// START TIMER ////
>         cairo_fill_preserve(pContext);
>         //// END TIMER ////
>
>         cairo_pattern_destroy(pat);
>
>         // Cleanup
>         cairo_surface_destroy(pSurface);
>         cairo_destroy(pContext);
> }
>
>
> In that test function above, if the code is using
> cairo_quartz_surface_create() then the call to cairo_fill_preserve()
> takes around 600ms to complete. If you comment out
> cairo_quartz_surface_create() and use cairo_image_surface_create()
> instead, then the call to cairo_fill_preserve() takes around 2ms to
> complete.
>
> In the case where it renders really slow, it is usually stuck down in
> the Pixman code calculating radial gradient information.


My profiling seems to show that the code path for the Quartz rendering does
not involve Pixman.


> However, I
> would expect that it uses the same Pixman code for the image surface
> as it does for the Quartz surface type. Does anyone have any ideas
> what could be going wrong here?
>

Instead of using Pixman, cairo-quartz is using the native Quartz radial
gradient implementation.
The radial gradient implementation in Pixman has been tuned quite a lot to
achieve high performance.


>
> This testing was done on an iPod touch 5th generation device (ARM
> based processor) using Cairo v1.8.0 and Pixman v0.24.0. I realize this
> version of Cairo may be a little old now, is anyone able to test this
> function using a newer build to see if it may have already been fixed?
>
> As I type this I realize it may be helpful to get numbers on an x86
> processor. I ran the same function under the iPhone simulator on a
> Core i7 MacBook Pro and it produced 74ms for the Quartz surface and
> less than 1 millisecond for the image surface. As you can see, the
> Quartz surface is still an order of magnitude slower.
>
> Any ideas?
>

Apparently the main slow-down comes from the evaluation of
ComputeGradientValue.
It should be possible to improve its performance by precomputing some more
information (example: the 1/bx coefficient), by caching the current
position with respect to the gradient stops (to reduce the lookup time for
the correct position) and by avoiding calls to fmod/floor/... function when
it is not actually needed (this is hit by your gradient, in particular).

As you might have noticed, recently I did not have much time to work on
cairo.
I will try to prepare a patch in the next days, but if my guess is
incorrect (that is, if the slowdown does not just depends on the things I
pointed out above), I might be unable to get it done before September.

Andrea



> --
> cairo mailing list
> [email protected]
> http://lists.cairographics.org/mailman/listinfo/cairo
>

-- 
cairo mailing list
[email protected]
http://lists.cairographics.org/mailman/listinfo/cairo