09:19 <pH5> According to compute_version(), we are missing GLSLVersion >= 130, MaxColorAttachments >= 4, ARB_depth_buffer_float, ARB_texture_float, ARB_texture_rg, EXT_draw_buffers2, EXT_framebuffer_sRGB, EXT_transform_feedback, and NV_conditional_render for OpenGL ES 3.0, at least on GC3000.

09:47 <mntirc> pH5: thx!

10:05 <lynxeye> mntirc: To debug the broken kicad rendering with TS sharing: could you attach gdb to xwayland and see if you hit glamor_egl_fd_from_pixmap?

10:07 <lynxeye> From reading the code it seems broken and might be the culprit, but I want to avoid chasing ghosts.

10:09 <mntirc> ok, i'll try!

10:21 <mntirc> lynxeye: as far as i can see it is not hit

10:22 <lynxeye> mntirc: Okay thanks. That rules out one theory, I'll keep looking.

10:27 <mntirc> lynxeye: i'm testing a bit with xgc. Copy Plane test makes things go really bad after it has run, for example

10:29 <mntirc> Copy Area too, as soon as there is a bit unset in "planemask"

10:30 <mntirc> lynxeye: the Copy Area test runs in 0.06 if the planemask is fully set (i.e. there is no mask). as soon as there is a mask bit, everything gets super slow and something is messed up in the gpu pipeline

10:31 <mntirc> same for Filled Rectangles test. so the problem has to do with masks.

10:32 <lynxeye> With the TS sharing we are apparently back at the "something doesn't take compression into account stage". Basically the sequence is something like 1. glClear makes TS valid 2. Something is rendered 3. glFlush 4. something samples from the rendered resource.

10:33 <lynxeye> Without TS sharing we'll decompress the resource at step 3, with TS sharing we skip the decompression and expect step 4 to take the compression into account and take appropriate action.

10:35 <mntirc> i see

10:37 <mntirc> it's not the planemask in the applications (eeschema from kicad, for example) case. i made glamor_pm_is_solid return true always; this "fixes" the xgc test but not the corruption in apps. so it's something else.

10:37 <lynxeye> This is probably the same things that caused https://gitlab.freedesktop.org/xorg/xserver/-/issues/941. https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/7603 just hid it due to decompression happening reliably at glFlush now. With TS sharing we try to avoid the overhead caused by decompression, but it seems that somewhere before step 4 the info is lost that the resource is still compressed.

10:39 <mntirc> as a sanity check i will remove my +glFlush() from glamor_render.c

10:41 <mntirc> (doesn't fix things... but it might also not be needed anymore)