Clean up inconsistency in enum decoration:
- Use the undecorated enums where possible.
- MAX_PROGRAM_TEXTURE_GATHER_COMPONENTS_ARB remains decorated, since it
has no undecorated equivalent in GL4.
Signed-off-by: Chris Forbes <chrisf@ijw.co.nz>
Bugzilla: https://bugs.freedesktop.org/show_bug.cgi?id=70054
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
The new surface channel select bits allow us to avoid having to
recompile the shader for this workaround.
Signed-off-by: Chris Forbes <chrisf@ijw.co.nz>
Reviewed-and-tested-by: Kenneth Graunke <kenneth@whitecape.org>
This allows us to use a different surface format for gather4, which is
required for R32G32_FLOAT to work on Gen7.
V4: - Only emit alternate surface state for shaders which will actually
use it.
- Pass a simple 'for_gather' flag rather than a function pointer.
The callee can decide what w/a to apply.
Signed-off-by: Chris Forbes <chrisf@ijw.co.nz>
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
Worst-case is that *every* texunit uses a format that needs overriding.
V4: Place the gather slots last, so shaders which don't use gather don't
get penalized by having a huge binding table.
Signed-off-by: Chris Forbes <chrisf@ijw.co.nz>
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
gather4 GREEN channel against a surface with format R32G32_FLOAT doesn't work
correctly on IVB. w/a from bspec:
- use R32G32_FLOAT_LD = 0x97 instead, for gather4 only.
- select BLUE channel to read GREEN
Signed-off-by: Chris Forbes <chrisf@ijw.co.nz>
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
V4: Only flag quirks if there are any uses of gather in the shader,
to avoid spurious recompiles just because someone happened to use
RG32F.
Signed-off-by: Chris Forbes <chrisf@ijw.co.nz>
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
Pretty much the same as the FS case. Channel select goes in the header,
V2: Less mangling.
V3: Avoid sampling at all, for degenerate swizzles.
Signed-off-by: Chris Forbes <chrisf@ijw.co.nz>
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
Lowers ir_tg4 (from textureGather and textureGatherOffset builtins) to
SHADER_OPCODE_TG4.
The usual post-sampling swizzle workaround can't work for ir_tg4,
so avoid doing that:
* For R/G/B/A swizzles use the hardware channel select (lives in the
same dword in the header as the texel offset), and then don't do
anything afterward in the shader.
* For 0/1 swizzles blast the appropriate constant over all the output
channels instead of sampling.
V2: Avoid duplicating header enabling block
V3: Avoid sampling at all, for degenerate swizzles.
Signed-off-by: Chris Forbes <chrisf@ijw.co.nz>
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
Adds the Gen7 message IDs, a new SHADER_OPCODE_TG4 pseudo-op, and
low-level support for emitting it via generate_tex().
V3: Updated for changes in master.
Signed-off-by: Chris Forbes <chrisf@ijw.co.nz>
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
When used with a cube array in VS, failed assertion in ir_validate:
Assignment count of LHS write mask channels enabled not
matching RHS vector size (3 LHS, 4 RHS).
To fix this, swizzle the RHS correctly for the writemask.
This showed up in the ARB_texture_gather tests, which exercise cube
arrays in the VS.
Signed-off-by: Chris Forbes <chrisf@ijw.co.nz>
Cc: "9.2" <mesa-stable@lists.freedesktop.org>
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
Consider only the top-left and top-right pixels to approximate DDX in a 2x2
subspan, unless the application requests a more accurate approximation via
GL_FRAGMENT_SHADER_DERIVATIVE_HINT or this optimization is disabled from the
new driconf option disable_derivative_optimization.
This results in a less accurate approximation. However, it improves the
performance of Xonotic with Ultra settings by 24.3879% +/- 0.832202% (at 95.0%
confidence) on Haswell. No noticeable image quality difference observed.
The improvement comes from faster sample_d. It seems, on Haswell, some
optimizations are introduced to allow faster sample_d when all pixels in a
subspan have the same derivative. I considered SAMPLE_STATE too, which allows
one to control the quality of sample_d on Haswell. But it gave much worse
image quality without giving better performance comparing to this change.
No piglit quick.tests regression on Haswell (tested with v1).
v2: better guess for precompile program key
Signed-off-by: Chia-I Wu <olv@lunarg.com>
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
Reviewed-by: Ian Romanick <ian.d.romanick@intel.com>
Reviewed-by: Chris Forbes <chrisf@ijw.co.nz>
We have the destination framebuffer object passed in; there's no need to
go digging around in the context.
Signed-off-by: Chris Forbes <chrisf@ijw.co.nz>
Reviewed-by: Chad Versace <chad.versace@linux.intel.com>
Using it encourages the (IMHO worrying) practice of leaving member
variables uninitialized in constructor definitions. This macro
shouldn't be necessary anymore after the last patch series fixing all
its users to initialize all member variables from the class
constructor. Remove it.
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
All member variables of glsl_to_tgsi_instruction are already being
initialized from its implicitly defined constructor, it's not
necessary to use rzalloc to allocate its memory.
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>