The docs say that we shouldn't need this workaround for gen8+, but just
removing it, causes gpu hangs. We'll revisit this, but for now, just
extend the workaround to gen9.
Signed-off-by: Damien Lespiau <damien.lespiau@intel.com>
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
Reviewed-by: Kristian Høgsberg <krh@bitplanet.net>
SKL moves the GS threadcount to dw8 from dw7, and no longer does the
divide by 2 thing.
Signed-off-by: Ben Widawsky <ben@bwidawsk.net>
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
Reviewed-by: Kristian Høgsberg <krh@bitplanet.net>
Tested-by: Kristian Høgsberg <krh@bitplanet.net>
This patch fixes this build error with G++ <= 4.6.
CXX test_vf_float_conversions.o
test_vf_float_conversions.cpp: In function ‘unsigned int f2u(float)’:
test_vf_float_conversions.cpp:63:20: error: expected primary-expression before ‘.’ token
Bugzilla: https://bugs.freedesktop.org/show_bug.cgi?id=86939
Signed-off-by: Vinson Lee <vlee@freedesktop.org>
Reviewed-by: Matt Turner <mattst88@gmail.com>
The register allocator prefers low-index registers from vc4_regs[] in the
configuration we're using, which is good because it means we prioritize
allocating the accumulators (which are faster). On the other hand, it was
causing raddr conflicts because everything beyond r0-r2 ended up in
regfile A until you got massive register pressure. By interleaving, we
end up getting more instruction pairing from getting non-conflicting
raddrs and QPU_WSes.
total instructions in shared programs: 55957 -> 52719 (-5.79%)
instructions in affected programs: 46855 -> 43617 (-6.91%)
We can avoid it by carefully ordering the packing. This is important as a
step in giving r3 to the register allocator.
total instructions in shared programs: 56087 -> 55957 (-0.23%)
instructions in affected programs: 18368 -> 18238 (-0.71%)
All uses of this require that the value be at least one, so it's
easier to report at least one than having to wrap all uses
in MAX2(max_compute_units, 1).
Reviewed-by: Marek Olšák <marek.olsak@amd.com>
Harvested GPUs have some of their render backends disabled, so
in order to prevent the hardware from trying to render things
with these disabled backends we need to correctly program
the PA_SC_RASTER_CONFIG register.
v2:
- Write RASTER_CONFIG for all SEs.
v3:
- Set GRBM_GFX_INDEX.INSTANCE_BROADCAST_WRITES bit.
- Set GRBM_GFX_INFEX.SH_BROADCAST_WRITES bit when done setting
PA_SC_RASTER_CONFIG.
- Get num_se and num_sh_per_se from kernel.
v4:
- Get correct value for num_se
- Remove loop for setting PA_SC_RASTER_CONFIG
- Only compute raster config when a backend has been disabled.
v5: Michel Dänzer
- Fix computation for chips with multiple SEs
https://bugs.freedesktop.org/show_bug.cgi?id=60879
CC: "10.4 10.3" <mesa-stable@lists.freedesktop.org>
There is a bug in the current lowering pass implementation where we lower saturate
to clamp only for vertex shaders on drivers supporting SM 3.0. The correct behavior
is to actually lower to clamp only when we don't support saturate which happens
on drivers that don't support SM 3.0
Reviewed-by: Marek Olšák <marek.olsak@amd.com>
Reviewed-by: Matt Turner <mattst88@gmail.com>
Signed-off-by: Abdiel Janulgue <abdiel.janulgue@linux.intel.com>
Fixes an infinite loop in swrast where the lowering pass unpacks saturate into
clamp but the opt_algebraic pass tries to do the opposite.
v3 (Ian):
This is a revert of commit cfa8c1cb "ir_to_mesa: lower ir_unop_saturate" on
the ir_to_mesa.cpp portion. prog_execute.c can handle saturates in vertex
shaders, so classic swrast shouldn't need this lowering pass.
Cc: "10.4" <mesa-stable@lists.freedesktop.org>
Bugzilla: https://bugs.freedesktop.org/show_bug.cgi?id=83463
Reviewed-by: Matt Turner <mattst88@gmail.com>
Signed-off-by: Abdiel Janulgue <abdiel.janulgue@linux.intel.com>
Fixes R11G11B10F rendering, and is required for SRGB format support.
Signed-off-by: Ilia Mirkin <imirkin@alum.mit.edu>
Reviewed-by: Rob Clark <robclark@freedesktop.org>
There were previously regressions regarding border colors, which the
updated swizzle logic resolves.
Signed-off-by: Ilia Mirkin <imirkin@alum.mit.edu>
Reviewed-by: Rob Clark <robclark@freedesktop.org>
This is a hack since it uses the texture information together with the
sampler, but I don't see a better way to do it. In OpenGL, there is a
1:1 correspondence.
Signed-off-by: Ilia Mirkin <imirkin@alum.mit.edu>
Reviewed-by: Rob Clark <robclark@freedesktop.org>
Expert debugging assistance provided by Chris Forbes.
Signed-off-by: Ilia Mirkin <imirkin@alum.mit.edu>
Reviewed-by: Rob Clark <robclark@freedesktop.org>
This was an oversight in the original patch. When PolygonMode is
used, then front faces, back faces, or both may be rendered as
points and are affected by point sprite state.
Note that SNB/IVB can't actually be fully conformant here, for
a legacy context -- we don't have separate sets of pointsprite
enables for front and back faces. Haswell ignores pointsprite
state correctly in hardware for non-point rasterization, so can
do this correctly, but it doesn't seem worth it.
Signed-off-by: Chris Forbes <chrisf@ijw.co.nz>
Cc: "10.4" <mesa-stable@lists.freedesktop.org>
Bugzilla: https://bugs.freedesktop.org/show_bug.cgi?id=86764
Reviewed-by: Matt Turner <mattst88@gmail.com>
Dead code elimination was eating the Y offset.
Fixes the piglit test:
spec/ARB_gpu_shader5/arb_gpu_shader5-interpolateAtOffset-nonconst
Signed-off-by: Chris Forbes <chrisf@ijw.co.nz>
Reviewed-by: Jason Ekstrand <jason.ekstrand@intel.com>
The original idea was to optimize away the condition by integrating it directly
into the CMP instruction. However, with native integers this requires an extra
I2F instruction. It is also fishy because the negation used didn't really honor
ieee754 float comparison rules, not to mention the CMP instruction itself
(being pretty much a legacy instruction) doesn't really have defined special
float value behavior in any case.
So, use UCMP and adjust the code trying to optimize the condition away
accordingly (I have absolutely no idea if such conditions are actually hit
or would be translated away somewhere else already).
v2: cosmetic changes
No piglit regressions on llvmpipe.
Reviewed-by: Jose Fonseca <jfonseca@vmware.com>
Reviewed-by: Brian Paul <brianp@vmware.com>
Multiple scenes per context are meant to be used so a new scene can be built
while another one is processed in rasterization. However, quite surprisingly,
this does not actually work (and according to git log, possibly never did,
though maybe it did at some point further back (5 years+) but was buggy)
because we always wait immediately on the rasterizer to finish the scene when
contexts (and hence setup/scene) is flushed. This means when we try to get
an empty scene later, any old one is already empty again.
Thus using multiple scenes is just a waste of memory (not too bad, since the
additional scenes are guaranteed to be empty, which means their size ought to
be one data block (64kB) plus the size of some structs), without actually
really doing anything. (There is also quite some code for the whole concept of
multiple scenes which doesn't really do much in practice, but keep it hoping
the wait-on-scene-flush can be fixed some day.)
Reviewed-by: Jose Fonseca <jfonseca@vmware.com>
The prim assembler may change the prim type when injecting prim ids now,
which isn't reflected by what's stored in emit.
This looks brittle and potentially dangerous (it is not obvious if such prim
type changes are really supported by pt emit, the prim type is actually also
set in prepare which would then be different).
This fixes piglit primitive-id-no-gs-first-vertex.shader_test.
Reviewed-by: Jose Fonseca <jfonseca@vmware.com>
The decomposition done in the prim assembler will turn tri fans into tris,
but this wasn't reflected in the output prim type. Meaning with a tri fan
with 6 verts input, the output was a tri fan with 12 vertices instead of a
tri list with 12 vertices (not as bad as it sounds, since the additional tris
created would all be degenerate since they'd all have two times vertex zero
but still bogus).
This is because the prim assembler is used if either the input topology is
something with adjacency, or if prim id needs to be injected, and for the
latter case topologies without adjacency can be converted to basic ones.
Unfortunately decomposition here for inserting prim ids is necessary, at
least for the indexed case where we can't just insert the prim id at the
right place depending on provoking vertex.
Reviewed-by: Jose Fonseca <jfonseca@vmware.com>
The default macros when the adjacency macros aren't defined will already
exactly do that (that is, drop the adjacent vertices and call the non-adjacent
macro).
Reviewed-by: Jose Fonseca <jfonseca@vmwarec.com>
Safe from causing optimization loops, since we don't constant propagate
VF arguments.
(for this and the previous patch):
total instructions in shared programs: 4289075 -> 4271932 (-0.40%)
instructions in affected programs: 1616779 -> 1599636 (-1.06%)
Reviewed-by: Ian Romanick <ian.d.romanick@intel.com>
The LINE instruction performs a multiply-add instruction (a * b + c)
where b and c are scalar arguments. It reads b and c from offsets in
src0 such that you can load them (it they're representable) as a
vector-float immediate with a single instruction.
Hurts some programs, but that'll all get better once we CSE the
vector-float MOVs in the next patch.
Bugzilla: https://bugs.freedesktop.org/show_bug.cgi?id=77544
Reviewed-by: Ian Romanick <ian.d.romanick@intel.com>
The PRMs say that
<src0> region must be a replicated scalar
(with HorzStride = VertStride = 0).
but apparently that doesn't actually apply to all generations. I did
notice when implementing the optimization later in this series that G45
and ILK needed this regioning.
Reviewed-by: Ian Romanick <ian.d.romanick@intel.com>