Once we reach the intended size of the buffer (BATCH_SZ or STATE_SZ), we
try and flush. If we're not allowed to flush, we resort to growing the
buffer so that there's space for the data we need to emit.
We accidentally got the threshold wrong. The first non-wrappable call
beyond (e.g.) STATE_SZ would grow the buffer to floor(1.5 * STATE_SZ),
The next call would see we were beyond STATE_SZ and think we needed to
grow a second time - when the buffer was already large enough.
We still want to flush when we hit STATE_SZ, but for growing, we should
use the actual size of the buffer as the threshold. This way, we only
grow when actually necessary.
v2: Simplify the control flow (suggested by Jordan)
Fixes: 2dfc119f22 "i965: Grow the batch/state buffers if we need space and can't flush."
Reviewed-by: Jordan Justen <jordan.l.justen@intel.com>
(cherry picked from commit ca43616586)
The original state buffer was marked with EXEC_OBJECT_CAPTURE. When
growing it, we want to preserve that flag so we continue to capture it
in GPU hang reports.
Fixes: 2dfc119f22 "i965: Grow the batch/state buffers if we need space and can't flush."
Reviewed-by: Ian Romanick <ian.d.romanick@intel.com>
(cherry picked from commit 52d32917e1)
The intention here is make the new BO use the same alignment as the old
BO. This isn't strictly necessary, but we would have to update the
'alignment' field in the validation list when swapping it out, and we
don't bother today.
The batch and state buffers use an alignment of 4096, so this should be
equivalent - it's just clearer than cut and pasting a magic constant.
Fixes: 2dfc119f22 "i965: Grow the batch/state buffers if we need space and can't flush."
Reviewed-by: Ian Romanick <ian.d.romanick@intel.com>
Reviewed-by: Jordan Justen <jordan.l.justen@intel.com>
(cherry picked from commit 2af7085460)
STATE_BASE_ADDRESS specifies a maximum size of the dynamic state
section, beyond which data supposedly reads back as 0. On Gen8+,
we were programming it to the size of the buffer. This worked fine
until we started growing the state buffer in commit 2dfc119f22.
When the state buffer grows, the value in STATE_BASE_ADDRESS becomes
too small, and our state beyond STATE_SZ bytes would read back as 0.
To avoid having to update the value, we program it to MAX_STATE_SIZE.
We used to program the upper bound to the maximum on older hardware
anyway, so programming it too large isn't a big deal.
Bogus SURFACE_STATE can easily lead to GPU hangs and misrendering.
DiRT Rally was hitting the statebuffer growth path, and suffered from
bad texture corruption and GPU hangs (usually around the same time).
This patch fixes both issues.
Fixes: 2dfc119f22 "i965: Grow the batch/state buffers if we need space and can't flush."
Bugzilla: https://bugs.freedesktop.org/show_bug.cgi?id=103101
Tested-by: Jordan Justen <jordan.l.justen@intel.com>
Reviewed-by: Chris Wilson <chris@chris-wilson.co.uk>
Reviewed-by: Jordan Justen <jordan.l.justen@intel.com>
Reviewed-by: Jason Ekstrand <jason@jlekstrand.net>
(cherry picked from commit cfc5af588c)
When creating a context without SetPixelFormat() don't blindly take the
pixel format reported by GDI. Instead, look for our own closest pixel
format.
Minor clean-ups added by Brian Paul.
Bugzilla: https://bugs.freedesktop.org/show_bug.cgi?id=103412
Reviewed-by: Brian Paul <brianp@vmware.com>
Tested-by: Brian Paul <brianp@vmware.com>
(cherry picked from commit bf41b2b262)
I really intended to set this for all shader stages by
3835009796 but missed it for compute shaders
(because it's in a different source file...).
Reviewed-by: Dave Airlie <airlied@redhat.com>
(cherry picked from commit 71e630753e)
The docs are not very concise in what this really does, however both
Alex Deucher and Nicolai Hähnle suggested this only really affects instructions
using the CLAMP output modifier, and I've confirmed that with the newly
changed piglit isinf_and_isnan test.
So, with this bit set, if an instruction has the CLAMP modifier bit (which
clamps to [0,1]) set, then NaNs will be converted to zero, otherwise the result
will be NaN.
D3D10 would require this, glsl doesn't have modifiers (with mesa
clamp(x,0,1) would get converted to such a modifier) coupled with a
whatever-floats-your-boat specified NaN behavior, but the clamp behavior
should probably always be used (this also matches what a decomposition into
min(1.0, max(x, 0.0)) would do, if min/max also adhere to the ieee spec of
picking the non-nan result).
Some apps may in fact rely on this, as this prevents misrenderings in
This War of Mine since using ieee muls
(ce7a045fee), without having to use clamped
rcp opcode, which would also fix this bug there.
radeonsi also seems to set this bit nowadays if I see that righ (albeit the
llvm amdgpu code comment now says "Make clamp modifier on NaN input returns 0"
instead of "Do not clamp NAN to 0" since it was changed, which also looks
a bit misleading).
v2: set it in all shader stages.
Bugzilla: https://bugs.freedesktop.org/show_bug.cgi?id=103544
Reviewed-by: Dave Airlie <airlied@redhat.com>
(cherry picked from commit 3835009796)
I believe this is the safe thing to do, especially ever since the driver
actually generates NaNs for muls too.
The ISA docs are not very helpful here, however the dx10 versions will pick
a non-nan result over a NaN one (this is also the ieee754 behavior), whereas
the non-dx10 ones will pick the NaN (verified by newly changed piglit
isinf-and-isnan test).
Other "modern" drivers will most likely do the same.
This was shown to make some difference for bug 103544, albeit it is not
required to fix it.
Reviewed-by: Dave Airlie <airlied@redhat.com>
(cherry picked from commit aab0bfc648)
The dynamic index of a vector (not array!) is lowered to a sequence of
conditional assignments. However, the interpolate_at_* expressions
require that the interpolant is an l-value of a shader input.
So instead of doing conditional assignments of parts of the shader input
and then interpolating that (which is nonsensical), we interpolate the
entire shader input and then do conditional assignments of the interpolated
result.
Reviewed-by: Timothy Arceri <tarceri@itsqueeze.com>
(cherry picked from commit ca63a5ed3e)
The intended rule has been clarified in GLSL 4.60, Section 8.13.2
(Interpolation Functions):
"For all of the interpolation functions, interpolant must be an l-value
from an in declaration; this can include a variable, a block or
structure member, an array element, or some combination of these.
Component selection operators (e.g., .xy) may be used when specifying
interpolant."
For members of interface blocks, var->data.must_be_shader_input must be
determined on-the-fly after lowering interface blocks, since we don't want
to disable varying packing for an entire block just because one input in it
is used in interpolateAt*.
v2: keep setting must_be_shader_input in ast_function (Ian)
v3: follow the relaxed rule of GLSL 4.60
v4: only apply the relaxed rules to desktop GL
(the ES WG decided that the relaxed rules may apply in a future version
but not retroactively; see also
dEQP-GLES31.functional.shaders.multisample_interpolation.interpolate_at_centroid.negative.*)
Bugzilla: https://bugs.freedesktop.org/show_bug.cgi?id=101378
Reviewed-by: Ian Romanick <ian.d.romanick@intel.com> (v1)
Reviewed-by: Timothy Arceri <tarceri@itsqueeze.com>
(cherry picked from commit 4f42450b86)
We want to program the 3DSTATE_RASTER field to the gl_context value,
not the other way around.
Fixes: 13ac46557a (i965: Port Gen8+ 3DSTATE_RASTER state to genxml.)
Reviewed-by: Jason Ekstrand <jason@jlekstrand.net>
Reviewed-by: Lionel Landwerlin <lionel.g.landwerlin@intel.com>
(cherry picked from commit 760e0156df)
This fixes yet another case where DFRACEXP has only one destination. Found
by address sanitizer.
Fixes tests/spec/arb_gpu_shader_fp64/execution/built-in-functions/fs-frexp-dvec4-only-mantissa.shader_test
Fixes: 3b666aa747 ("st/glsl_to_tgsi: fix DFRACEXP with only one destination")
Acked-by: Marek Olšák <marek.olsak@amd.com>
(cherry picked from commit 7e35bdad1c)
This fixes hangs on cayman with
tests/spec/arb_tessellation_shader/execution/trivial-tess-gs_no-gs-inputs.shader_test
This has a single if/else in it, and when this peephole activated,
it would set the jump target to NULL if there was no instruction
after the final POP. This adds a NOP if we get a jump in this case,
and seems to fix the hangs, so we have a valid target for the ELSE
instruction to go to, instead of 0 (which causes infinite loops).
v2: update last_cf correctly. (I had some other patches hide this)
Cc: <mesa-stable@lists.freedesktop.org>
Signed-off-by: Dave Airlie <airlied@redhat.com>
(cherry picked from commit 579ec9c311)
Language and spelling fixups in three places.
Cc: "17.2" "17.3" <mesa-stable@lists.freedesktop.org>
Signed-off-by: Ben Crocker <bcrocker@redhat.com>
Reviewed-by: Eric Engestrom <eric.engestrom@imgtec.com>
[Eric: move two fixes from the other patch to this one.]
Signed-off-by: Eric Engestrom <eric.engestrom@imgtec.com>
(cherry picked from commit b43daf7bf6)
Signed-off-by: Kai Wasserbäch <kai@dev.carbon-project.org>
Reviewed-by: Eric Engestrom <eric.engestrom@imgtec.com>
(cherry picked from commit d25123e23a)
When floating point textures are created on OpenGL ES 2.0, driver
is free to choose used internal format. Mesa makes this decision in
adjust_for_oes_float_texture. Error checking for glTexImage2D properly
checks that sized formats are not used. We use same error checking
path for glTexSubImage2D (since there is lot of overlap), however since
those checks include internalFormat checks, we need to pass original
internalFormat passed by the client. Patch adds oes_float_internal_format
that does reverse adjust_for_oes_float_texture to get that format.
Fixes following test failure:
ES2-CTS.gtf.GL2ExtensionTests.texture_float.texture_float
(when running test with MESA_GLES_VERSION_OVERRIDE=2.0)
Signed-off-by: Tapani Pälli <tapani.palli@intel.com>
Bugzilla: https://bugs.freedesktop.org/show_bug.cgi?id=103227
Cc: "17.3" <mesa-stable@lists.freedesktop.org>
Reviewed-by: Matt Turner <mattst88@gmail.com>
(cherry picked from commit 1e508e10d9)
The header can be included from C++, hence contents should have
appropriate notation.
Cc: mesa-stable@lists.freedesktop.org
Cc: Dylan Baker <dylan@pnwbakers.com>
Signed-off-by: Emil Velikov <emil.velikov@collabora.com>
Reviewed-by: Eric Engestrom <eric.engestrom@imgtec.com>
(cherry picked from commit c7616ac069)
>From GLSL 4.5 spec, section "7.1 Built-In Language Variables", page 130 of
the PDF states:
"If multiple shaders using members of a built-in block belonging to
the same interface are linked together in the same program, they must
all redeclare the built-in block in the same way, as described in
section 4.3.9 “Interface Blocks” for interface-block matching, or a
link-time error will result."
Fixes:
* GL45-CTS.CommonBugs.CommonBug_PerVertexValidation
v2 (Neil Roberts):
Explicitly look for gl_PerVertex in the symbol tables instead of
waiting to find a variable in the interface.
Bugzilla: https://bugs.freedesktop.org/show_bug.cgi?id=102677
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
Signed-off-by: Eduardo Lima Mitev <elima@igalia.com>
Signed-off-by: Neil Roberts <nroberts@igalia.com>
(cherry picked from commit f9de7f5596)
This effectively factorizes a couple of similar routines.
v2 (Neil Roberts): Non-trivial rebase on master
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
Signed-off-by: Eduardo Lima Mitev <elima@igalia.com>
Signed-off-by: Neil Roberts <nroberts@igalia.com>
(cherry picked from commit f5fe99ac85)
Some symbols gathered in the symbols table during parsing are needed
later for the compile and link stages, so they are moved along the
process. Currently, only functions and non-temporary variables are
copied between symbol tables. However, the built-in gl_PerVertex
interface blocks are also needed during the linking stage (the last
step), to match re-declared blocks of inter-stage shaders.
This patch adds a new utility function that will factorize current code
that copies functions and variables between two symbol tables, and in
addition will copy explicitly declared gl_PerVertex blocks too.
The function will be used in a subsequent patch.
v2 (Neil Roberts):
Allow the src symbol table to be NULL and explicitly copy the
gl_PerVertex symbols in case they are not referenced in the exec_list.
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
Signed-off-by: Eduardo Lima Mitev <elima@igalia.com>
Signed-off-by: Neil Roberts <nroberts@igalia.com>
(cherry picked from commit 4c62a270a9)
The cache-test test program attempts to create a collision (using key_a
and key_a_collide) by making the first two bytes identical. The idea is
fine -- the shader cache wants to use the first four characters of a
SHA1 hex digest as the index.
The following program
unsigned char array[4] = {1, 2, 3, 4};
int *ptr = (int *)array;
for (int i = 0; i < 4; i++) {
printf("%02x", array[i]);
}
printf("\n");
printf("%08x\n", *ptr);
prints
01020304
04030201
on little endian, and
01020304
01020304
on big endian.
On big endian platforms reading the character array back as an int (as
is done in disk_cache.c) does not yield the same results as reading the
byte array.
To get the first four characters of the SHA1 hex digest when we mask
with CACHE_INDEX_KEY_MASK, we need to byte swap the int on big endian
platforms.
Bugzilla: https://bugs.freedesktop.org/103668
Bugzilla: https://bugs.gentoo.org/637060
Bugzilla: https://bugs.gentoo.org/636326
Fixes: 87ab26b2ab ("glsl: Add initial functions to implement an
on-disk cache")
Reviewed-by: Emil Velikov <emil.velikov@collabora.com>
(cherry picked from commit c690a7a8cd)
The code defines a macro blk0(i) based on the preprocessor condition
BYTE_ORDER == LITTLE_ENDIAN. If true, blk0(i) is defined as a byte swap
operation. Unfortunately, if the preprocessor macros used in the test
are no defined, then the comparison becomes 0 == 0 and it evaluates as
true.
Fixes: d1efa09d34 ("util: import sha1 implementation from OpenBSD")
Reviewed-by: Emil Velikov <emil.velikov@collabora.com>
(cherry picked from commit 532674303a)
MADs don't take immediate sources, but we allow them in the IR since it
simplifies a lot of things. I neglected to consider that case.
Fixes: 4009a9ead4 ("i965/fs: Allow saturate propagation to propagate
negations into MADs.")
Bugzilla: https://bugs.freedesktop.org/show_bug.cgi?id=103616
Reported-and-Tested-by: Ruslan Kabatsayev <b7.10110111@gmail.com>
Reviewed-by: Ian Romanick <ian.d.romanick@intel.com>
(cherry picked from commit a05af1f7b8)
We need to account for SGPR locations in merged shaders.
This case is exercised by KHR-GL45.enhanced_layouts.vertex_attrib_locations
Fixes: 79c2e7388c ("radeonsi/gfx9: use SPI_SHADER_USER_DATA_COMMON")
Reviewed-by: Marek Olšák <marek.olsak@amd.com>
(cherry picked from commit df5ebe0c26)
generate_array_index fails to check whether the target of a subroutine
call exists in the AST, potentially passing around null ir_rvalue
pointers eventuating in abort/segfault.
Fixes: fd01840c0b ("glsl: add AoA support to subroutines")
Reviewed-by: Timothy Arceri <tarceri@itsqueeze.com>
Bugzilla: https://bugs.freedesktop.org/show_bug.cgi?id=100438
(cherry picked from commit f09c2cefdd)
So far on pre-cayman chipsets the CF instructions CF_OP_LOOP_END,
CF_OP_CALL_FS, CF_OP_POP, and CF_OP_GDS an extra CF_NOP instruction
was added to add the EOP flag, even though this is not actually
needed, because all these instrutions support the EOP flag.
This patch removes the fixup code, adds setting the EOP flag for the
according instructions as well as others like CF_OP_TEX and CF_OP_VTX,
and adds writing out EOP for this type of instruction in the disassembler.
This also fixes a bug where shaders were created that didn't actually have
the EOP flag set in the last CF instruction, which might have resulted
in GPU lockups.
[airlied: cleaned up a little]
Signed-off-by: Gert Wollny <gw.fossdev@gmail.com>
Cc: <mesa-stable@lists.freedesktop.org>
Signed-off-by: Dave Airlie <airlied@redhat.com>
(cherry picked from commit 1d076aafbc)
Almost all of our BO export paths were already properly marked the BO as
external and added it to the handle table. Most export use-cases go
through a prime fd or flink where we have a brw_bo export helper that
does the right thing. The one missing one happens when you call
queryImage and ask for __DRI_IMAGE_ATTRIB_HANDLE. We just grabbed the
gem handle out of the BO (because it's really easy to do that) and
handed it off to the client; what could go wrong? As it turns out, this
path is used by basically every compositor that wants to turn around and
call drmModeAddFB2 on it so it can hand it off to display. The result,
as of 4b1e70cc57, is that we no longer set
MOCS_PTE on those surfaces and the kernel's attempts to disable caching
fail and we scanout gets corruption.
Bugzilla: https://bugs.freedesktop.org/show_bug.cgi?id=103759
Fixes: 4b1e70cc57
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
Cc: mesa-stable@lists.freedesktop.org
(cherry picked from commit 0a6a137eb2)
Otherwise, if the image is not bound to the start of the buffer, we're
going to be reading and writing its fast clear state in the wrong spot.
Reviewed-by: Lionel Landwerlin <lionel.g.landwerlin@intel.com>
Cc: mesa-stable@lists.freedesktop.org
(cherry picked from commit a07f7b2619)
Number of dwords in MI_FLUSH_DW changed from 4 to 5 in gen8+.
Signed-off-by: Anuj Phogat <anuj.phogat@gmail.com>
Cc: <mesa-stable@lists.freedesktop.org>
(cherry picked from commit 1dc45d75bb)
[Emil Velikov: trivial conflicts]
Signed-off-by: Emil Velikov <emil.velikov@collabora.com>
Conflicts:
src/mesa/drivers/dri/i965/intel_blit.c
This build system is rather incomplete in the 17.3 branch, with multiple
bugs and user facing changes already addressed in master.
It's not shipped in the tarball and we don't want to receive bug reports
about 17.3, 18.0 is the release that I hope to have the meson build in
shape for.
Simply error() out, if anyone tries to use it.
Signed-off-by: Emil Velikov <emil.velikov@collabora.com>
This reverts commit e8c9e65185.
With the actual bug fixed (by commit 6ac2d16901), this is not
necessary. I'm doubtful of its correctness in any case.
(cherry picked from commit a31d038208)
Fixes the following tests on CHV, BXT, and GLK:
KHR-GL46.shader_ballot_tests.ShaderBallotFunctionBallot
dEQP-VK.spirv_assembly.instruction.compute.uconvert.uint32_to_int64
Bugzilla: https://bugs.freedesktop.org/show_bug.cgi?id=103115
(cherry picked from commit cfcfa0b9cd)
The MOV instruction can extract bytes to words/double words, and
words/double words to quadwords, but not byte to quadwords.
For unsigned byte to quadword, we can read them as words and AND off the
high byte and extract to quadword in one instruction. For signed bytes,
we need to first sign extend to word and the sign extend that word to a
quadword.
Fixes the following test on CHV, BXT, and GLK:
KHR-GL46.shader_ballot_tests.ShaderBallotBitmasks
Bugzilla: https://bugs.freedesktop.org/show_bug.cgi?id=103628
Reviewed-by: Jason Ekstrand <jason@jlekstrand.net>
(cherry picked from commit 6ac2d16901)
When queryImage doesn't support __DRI_IMAGE_ATTRIB_FOURCC wayland clients
will die with a NULL derefence in wl_proxy_add_listener.
Attempt to provide a simple fallback to keep ancient systems working.
Fixes: 6595c69951 ("egl/wayland: Remove more surface specifics from
create_wl_buffer")
Bugzilla: https://bugs.freedesktop.org/show_bug.cgi?id=103519
Signed-off-by: Derek Foreman <derekf@osg.samsung.com>
Reviewed-by: Emil Velikov <emil.velikov@collabora.com>
Acked-by: Daniel Stone <daniels@collabora.com>
Reviewed-by: Eric Engestrom <eric.engestrom@imgtec.com>
(cherry picked from commit 0db36caa19)
Squashed with:
egl: fix var type
queryImage() takes an `int*`; compiler is warning about the
signed<->unsigned pointer mismatch.
Fixes: 0db36caa19 "egl/wayland: Add a fallback when fourcc
query isn't supported"
Signed-off-by: Eric Engestrom <eric.engestrom@imgtec.com>
Reviewed-by: Frank Binns <frank.binns@imgtec.com>
Reviewed-by: Tapani Pälli <tapani.palli@intel.com>
Reviewed-by: Emil Velikov <emil.velikov@collabora.com>
Reviewed-by: Derek Foreman <derekf@osg.samsung.com>
(cherry picked from commit ca95d7ad4e)
Otherwise we leak it.
Fixes: eaa56eab6d "radv: initial support for shared semaphores (v2)"
Reviewed-by: Samuel Pitoiset <samuel.pitoiset@gmail.com>
(cherry picked from commit 7c25578863)
Otherwise we can leak the old syncobj.
Fixes: eaa56eab6d "radv: initial support for shared semaphores (v2)"
Reviewed-by: Dave Airlie <airlied@redhat.com>
Reviewed-by: Samuel Pitoiset <samuel.pitoiset@gmail.com>
(cherry picked from commit 917d3b43f2)
It turned out that with recent changes that call into dri3 from glFinish(),
it appears like different thread end up waiting for X events simultaneously,
causing deadlocks since they steal events from eachoter and update the dri3
counters behind eachothers backs.
This patch intends to improve on that. It allows at most one thread at a
time to wait on events for a single drawable. If another thread intends to
do the same, it's put to sleep until the first thread finishes waiting, and
then it rechecks counters and optionally retries the waiting. Threads that
poll for X events never pulls X events off the event queue if there are
other threads waiting for events on that drawable. Counters in the
dri3 drawable structure are protected by a mutex. Finally, the mutex we
introduce is never held while waiting for the X server to avoid
unnecessary stalls.
This does not make dri3 drawables completely thread-safe but at least it's a
first step.
Bugzilla: https://bugs.freedesktop.org/show_bug.cgi?id=102358
Fixes: d5ba75f888 "st/dri2 Plumb the flush_swapbuffer functionality through to dri3"
Signed-off-by: Thomas Hellstrom <thellstrom@vmware.com>
Acked-by: Nicolai Hähnle <nicolai.haehnle@amd.com>
(cherry picked from commit 54a58b2856)
We renamed "Function Enable" to "Enable", which broke our detection
of whether shaders are enabled or not. So, we'd see a bunch of HS/DS
packets with program offsets of 0, and think that was a valid TCS/TES.
Fixes: c032cae9ff (genxml: Rename "Function Enable" to "Enable".)
Reviewed-by: Lionel Landwerlin <lionel.g.landwerlin@intel.com>
(cherry picked from commit 9a0465b3a3)
We want to emit invariant state at the start of a render batch. In the
past, this more or less happened: a new batch flagged BRW_NEW_CONTEXT
(because we don't have hardware contexts), which triggered the
brw_invariant_state atom. So, it would be emitted before any 3D
drawing. (Technically, there might be some BLT commands in the batch
because Gen4-5 have a single combined render/BLT ring, but that should
be harmless).
With the advent of BLORP, this broke. The first item in a batch might
be a BLORP operation, which bypasses the normal draw upload path. So,
we need to ensure invariant state happens first. To do that, we just
upload it when creating a new batch. On Gen6+ we'd need to worry about
whether it's a RENDER or BLT batch, but because we have a combined ring,
this approach should work fine on Gen4-5.
Seems to fix GPU hangs when playing hardware accelerated video with
mpv -hwdec=vaapi on Ironlake.
Cc: mesa-stable@lists.freedesktop.org
Bugzilla: https://bugs.freedesktop.org/show_bug.cgi?id=103529
Reviewed-by: Jason Ekstrand <jason@jlekstrand.net>
(cherry picked from commit 8f91aa35a5)
...and provide a better citation for the existing one.
v2:
- Apply the workaround to Gen8 too, as intended (caught by Topi).
- Restructure to add bits instead of an extra flush (based on a similar
patch by Rafael Antognolli).
Cc: mesa-stable@lists.freedesktop.org
Reviewed-by: Rafael Antognolli <rafael.antognolli@intel.com>
(cherry picked from commit 8d48671492)
Speed up simd16 frontend (default) on avx/avx2 platforms;
fixes performance regression caused by switch to simdlib.
Reviewed-by: Bruce Cherniak <bruce.cherniak@intel.com>
Cc: mesa-stable@lists.freedesktop.org
(cherry picked from commit d8489517a5)
