i965: Unalias all GLSL source regs from the destination regs used.
We were doing it ad-hoc before, as instructions with potential
aliasing problems were identified. But thanks to swizzling basically
anything can have aliasing, so just do it generally at source reg
setup time. This is somewhat inefficient, because sometimes an
operation doesn't need unaliasing protection if the swizzling is safe,
but the unaliasing before didn't cover those cases either.
Fixes piglit glsl-fs-loop.
(cherry picked from commit 6b194dab6b)
We were patching up all the break and continues between the start of
our loop and the end of our loop, even if they were breaks/continues
for an inner loop. Avoiding patching already patched breaks/continues
fixes piglit glsl-vs-loop-nested.
(cherry picked from commit f6f547d87e)
i965: Fix up the handling of point sprite coordinate replacement.
The code was walking over the regs of pairs of attributes and checking
whether the attribute with a given reg index had point sprite enabled.
So the point sprite setup code was rarely even getting executed.
Instead, we need to determine which channels of a reg need point
sprite coordinate replacement. In addition, it was multiplying the
attribute by 1/w, when it's supposed to cover (0, 1) in each direction
regardless of w, and it wasn't filling in the Z and W components of
the texcoord as specified.
Fixes piglit point-sprite and the spriteblast demo. Bug #24431, #22245.
(cherry picked from commit bc632d0437)
i965: Add support for the CMP opcode in the GLSL path.
This would be triggered by use of sqrt() along with control flow.
Fixes piglit-fs-sqrt-branch and a bug in Yo Frankie!.
(cherry picked from commit 48dca99feb)
i965: Fix the response len of masked sampler messages for 8-wide dispatch.
The bad response length would hang the GPU with a masked sample in a
shader using control flow. For 8-wide, the response length is always
4, and masked slots are just not written to. brw_wm_glsl.c already
allocates registers in the right locations.
Fixes piglit glsl-fs-bug25902 (fd.o bug #25902).
(cherry picked from commit f6d210c284)
(cherry picked from commit dc8c035944)
radeon: Fix buffer object unmap to be called only once for dma buffers.
If flush happens inside radeonRefillCurrentMaRegion the last dma buffer would
be unmapped twice. Unmapping buffer when moving buffer to wait list fixes the
mapping error.
The llvm wrapper wasn't really an OS thing.
Use lp_bld.h for now but we eventually should rename/re-prefix all the
files/functions in the gallivm/ directory.
gallivm: add min_lod==max_lod special-case code in lp_build_lod_selector()
When min_lod==max_lod we don't need to go through all the work of
computing the lod from partial derivatives. This is hit by the mipmap
generation utility code.
LLVMBuildFPTrunc() should be used for double->float conversion, not
float->int conversion.
There should be a better way to compute floor(), ceil(), etc that doesn't
involve float->int->float conversion.
Eric Anholt