Gen9 hardware has this annoying little corner where CCS_E is not allowed for any sRGB formats. This is fixed on gen10 but on gen9 there's nothing we can do; it just doesn't work. The old approach to working around this was to just disable CCS_E the moment we saw sRGB. This is bad because GLX gives out sRGB-capable visuals by default and you can easily get one through EGL as well even if you never enable sRGB encode. This means that users who have sRGB visuals and don't care about sRGB encode are getting unnecessarily punished. This isn't a huge problem today because you also can't do CCS_E on X-tiled images but it will be a problem the moment we start seeing the Y-tiling modifier through the window system.
Also, I think sRGB + CCS was plain broken for the less likely case of rendering to a texture-backed framebuffer. Our tracking for sRGB was based on piles of sRGBEnabled checks that I'm not at all sure added up to correct code. When trying to better test the CCS_E modifier, I patched waffle to start using modifiers whenever the GBM back-end was in-use. When I ran piglit with these waffle patches and my old CCS_E series, our pass rate was under 50%. I think part of that was due to bugs with sRGB and part of it was due to not having a plan for falling back to CCS_D once the CCS_E modifier gets used. This series is that plan. The first 5 patches are a couple of bugfixes and the removal of a couple of bogus restrictions. In particular, we were disabling CCS_E on all renderbuffers for no good reason. Patch 3 fixes bugs exposed by patch 2 related to glBlitFramebuffers with both color and depth bits specified. The next 5 add a partial resolve pass for MCS and hook it up so that we can handle clear colors with texture views correctly. The last 22 patches rework things so that we can properly fall back to CCS_D whenever we can't render with CCS_E. This requires adding a seventh value to the isl_aux_state enum to describe a "partial clear" state which is the state a CCS_E image is in when it's been fast-cleared and then rendered to using CCS_D. Tracking this additional state allows us to turn on CCS_E even when we have an sRGB visual and then just silently resolve if we ever need to render with sRGB encode enabled. If they just turn on sRGB encode and leave it on, then they don't get a resolve because our tracking code knows that you can do CCS_D rendering on a CCS_E surface that is in the CLEAR state and the end result is the PARTIAL_CLEAR state. We need to land this series before we flip on the CCS_E modifier and I'd like to land it in time for the 17.2 release if we can. Happy Reviewing! Jason Ekstrand (32): i965/surface_state: Use the minified depth for number of image layers i965/blorp: Use the renderbuffer format for clears i965/blorp: Do flushes around depth resolves i965/miptree: Stop setting FOR_SCANOUT for renderbuffers i965/miptree: Remove some unneeded restrictions intel/blorp: Add a partial resolve pass for MCS i965/miptree: Make layer_range_length return locical layers i965/miptree: Tighten up finish_mcs_write i965/miptree: Add support for partially resolving MCS i965/miptree: Partially resolve MCS for texture views i965/miptree: Add a helper for getting the aux usage for texturing i965/miptree: Rework prepare/finish_render to be in terms of aux_usage i965/blorp: Do prepare/finsh manually i965/blorp: Use texture/render_aux_usage for blits i965/blorp: Be more accurate about aux usage in blorp_copy i965/blorp: Use render_aux_usage for color clears i965/blorp: Use prepare/finish_depth for depth clears i965/miptree: Refactor some things to use mt->aux_usage i965/miptree: Take an aux_usage in prepare/finish intel/isl: Add an aux state for "partial clear" i965/miptree: Use ISL_AUX_STATE_PARTIAL_CLEAR for CCS_D i965/miptree: Allow for accessing a CCS_E image as CCS_D i965/miptree: Use miptree range helpers in has_color_unresolved i965/miptree: Take an isl_format in prepare_texture i965/surface_state: Take an isl_aux_usage in emit_surface_state i965/surface_state: Get the aux usage from the miptree code intel/isl/format: Dedent the template in gen_format_layout.py intel/isl/format: Add an srgb_to_linear helper i965: Weaken the texture view rules for formats slightly intel/blorp: Allow blorp_copy on sRGB formats intel/isl: Add a helper for determining if a color is 0/1 i965: Enable regular fast-clears (CCS_D) on gen9+ src/intel/blorp/blorp.h | 6 + src/intel/blorp/blorp_blit.c | 18 +- src/intel/blorp/blorp_clear.c | 105 ++++- src/intel/blorp/blorp_nir_builder.h | 102 +++++ src/intel/blorp/blorp_priv.h | 1 + src/intel/isl/gen_format_layout.py | 161 +++++--- src/intel/isl/isl.c | 27 ++ src/intel/isl/isl.h | 99 +++-- src/mesa/drivers/dri/i965/brw_blorp.c | 349 ++++++++++------ src/mesa/drivers/dri/i965/brw_blorp.h | 5 + src/mesa/drivers/dri/i965/brw_clear.c | 3 +- src/mesa/drivers/dri/i965/brw_draw.c | 14 +- src/mesa/drivers/dri/i965/brw_meta_util.c | 11 - src/mesa/drivers/dri/i965/brw_wm_surface_state.c | 160 +++---- src/mesa/drivers/dri/i965/intel_mipmap_tree.c | 504 +++++++++++++++-------- src/mesa/drivers/dri/i965/intel_mipmap_tree.h | 20 +- 16 files changed, 1058 insertions(+), 527 deletions(-) create mode 100644 src/intel/blorp/blorp_nir_builder.h -- 2.5.0.400.gff86faf _______________________________________________ mesa-dev mailing list mesa-dev@lists.freedesktop.org https://lists.freedesktop.org/mailman/listinfo/mesa-dev
