I don't know why, but we never hooked up this pass Eric wrote.
Otherwise, you can end up with stupid scalarized code such as:
vec4 ssa_7 = load_const (0.0, 0.0, 0.0, 0.0)
vec4 ssa_8 = ...
vec1 ssa_9 = feq ssa_8, ssa_7
vec1 ssa_10 = feq ssa_8.y, ssa_7.y
vec1 ssa_11 = feq ssa_8, ssa_7.z
vec1 ssa_12 = feq ssa_8.y, ssa_7.w
ssa_8.xyxy == <0, 0, 0, 0> should only take two feq instructions.
shader-db on Skylake:
total instructions in shared programs: 9121153 -> 9120749 (-0.00%)
instructions in affected programs: 32421 -> 32017 (-1.25%)
helped: 277
HURT: 69
total cycles in shared programs: 69003364 -> 69000912 (-0.00%)
cycles in affected programs: 899186 -> 896734 (-0.27%)
helped: 313
HURT: 403
This also prevents regressions when disabling channel expressions.
v2: Don't call opt_cse afterwards (requested by Matt). It should
happen in the optimization loop below anyway.
Signed-off-by: Kenneth Graunke <kenneth@whitecape.org>
Reviewed-by: Eduardo Lima Mitev <elima@igalia.com>
Reviewed-by: Matt Turner <mattst88@gmail.com>
The aim of this is to work towards removing UniformHash from the program
struct so that we don't need to hold onto it in memory and pass it around
outside the linker.
Reviewed-by: Dave Airlie <airlied@redhat.com>
There are never render target reads, so there are no scheduling hazards.
Giving the extra flexibility to the scheduler makes it possible to do
FB writes as soon as their sources are available, reducing register
pressure. It also makes it possible to do the payload setup for more
than one FB write message at a time, which could better hide latency.
shader-db results on Skylake:
total instructions in shared programs: 9110254 -> 9110211 (-0.00%)
instructions in affected programs: 2898 -> 2855 (-1.48%)
helped: 3
HURT: 0
LOST: 0
GAINED: 1
A reduction in instruction counts is surprising, but legitimate:
the three shaders helped were spilling, and reducing register
pressure allowed us to issue fewer spills/fills.
total cycles in shared programs: 69035108 -> 68928820 (-0.15%)
cycles in affected programs: 4412402 -> 4306114 (-2.41%)
helped: 4457
HURT: 213
Signed-off-by: Kenneth Graunke <kenneth@whitecape.org>
Reviewed-by: Matt Turner <mattst88@gmail.com>
Reviewed-by: Connor Abbott <cwabbott0@gmail.com>
reuse the sampler deref handling code to do the same
thing for atomics.
Acked-by: Ilia Mirkin <imirkin@alum.mit.edu>
Signed-off-by: Dave Airlie <airlied@redhat.com>
Since we fixed the glsl->tgsi conversion we no longer need
this function.
Reviewed-by: Timothy Arceri <timothy.arceri@collabora.com>
Signed-off-by: Dave Airlie <airlied@redhat.com>
The state tracker never handled this properly, and it finally
annoyed me for the second time so I decided to fix it properly.
This is inspired by the NIR sampler lowering code and I only realised
NIR seems to do its deref ordering different to GLSL at the last
minute, once I got that things got much easier.
it fixes a bunch of tests in
tests/spec/arb_gpu_shader5/execution/sampler_array_indexing/
v2: fix AoA tests when forced on.
I was right I didn't need all that code, fixing the AoA code
meant cleaning up a chunk of code I didn't like in the array
handling.
v3: start generalising the code a bit more for atomics.
v3.1: use UniformRemapTable
v4: handle uniforms differently using the param_index,
and go back to UniformStorage
fix issues identified by Timothy with deref handling.
v4.1: squash const fix and move handling 1D const out
of recursive function.
Reviewed-by: Timothy Arceri <timothy.arceri@collabora.com>
Acked-by: Ilia Mirkin <imirkin@alum.mit.edu>
Signed-off-by: Dave Airlie <airlied@redhat.com>
We have a requirement to store the index into the mesa parameterlist
for uniforms. Up until now we've overwritten var->data.location with
this info. However this then stops us accessing UniformStorage,
which is needed to do proper dereferencing.
Add a new variable to ir_variable to store this value in, and change
the two uses to use it correctly.
Reviewed-by: Timothy Arceri <timothy.arceri@collabora.com>
Signed-off-by: Dave Airlie <airlied@redhat.com>
Its previous name was somewhat misleading, this really behaves like a
RW cache flush rather than an invalidation.
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
The state cache is also L3-backed so it seems sensible to make sure
it's clean as we do for other RO caches before repartitioning the L3.
This wasn't part of my original L3 partitioning code because I was
able to reproduce hangs on Gen7 hardware when the state cache
invalidation happened asynchronously with previous 3D rendering, which
should no longer be possible after the previous change.
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
We need to split the stalling flush from the RO cache invalidation
into a different PIPE_CONTROL command to make sure that the top of the
pipe invalidation happens after any previous rendering is complete.
Otherwise it's possible for previous rendering to pollute the L3 cache
in the short window of time between RO invalidation and the completion
of the stalling flush. Fixes rendering artifacts on Unigine Heaven,
Metro Last Light Redux and Metro 2033 Redux.
Bugzilla: https://bugs.freedesktop.org/show_bug.cgi?id=93540
Bugzilla: https://bugs.freedesktop.org/show_bug.cgi?id=93599
Tested-by: Darius Spitznagel <d.spitznagel@goodbytez.de>
Tested-by: Martin Peres <martin.peres@linux.intel.com>
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
The SEL instruction with predication mode NONE emitted when the atomic
operation doesn't need to be predicated is a no-op and might rely on
undocumented hardware behaviour. Noticed by chance while looking at
the assembly output.
Reviewed-by: Matt Turner <mattst88@gmail.com>
Parameterize build_asin() on the fit coefficients so the
implementation can be shared while still using different polynomials
for asin and acos. Also switch back to implementing acos in terms of
asin -- The improvement obtained from cancelling out the pi/2 terms
was negligible compared to the approximation error.
vtn_handle_type() creates a signed type regardless of the value of the
signedness flag, which usually doesn't make much of a difference
except when the type is used as base sampled type of an image type,
what will cause the base type of the NIR image variable to be
inconsistent with its format and cause an assertion failure in the
back-end (most likely only reproducible on Gen7), and may change the
semantics of the image intrinsic subtly (e.g. UMIN may become IMIN).
New functions for examining instructions, declarations, etc.
Reviewed-by: Marek Olšák <marek.olsak@amd.com>
Reviewed-by: Edward O'Callaghan <eocallaghan@alterapraxis.com>
The errors.c file had grown quite large so split off this extension
code into its own file. This involved making a handful of functions
non-static.
Acked-by: Timothy Arceri <timothy.arceri@collabora.com>
This fixes a crash with bin/arb_clear_texture-base-formats and
probably some other tests which use clear_texture().
Signed-off-by: Samuel Pitoiset <samuel.pitoiset@gmail.com>
Reviewed-by: Ilia Mirkin <imirkin@alum.mit.edu>
There is no need to allocate memory when unwrapping the indirect buf.
Signed-off-by: Samuel Pitoiset <samuel.pitoiset@gmail.com>
Reviewed-by: Ilia Mirkin <imirkin@alum.mit.edu>
The builtin data can get released with a glReleaseShaderCompiler call.
We're careful everywhere to clone everything that comes out of builtins
except here, where we accidentally return the signature belonging to the
builtin version, rather than the locally-cloned one.
Signed-off-by: Ilia Mirkin <imirkin@alum.mit.edu>
Reviewed-by: Timothy Arceri <timothy.arceri@collabora.com>
Tested-by: Rob Herring <robh@kernel.org>
Cc: mesa-stable@lists.freedesktop.org
The builtin function shader is part of the builtin state, released
when glReleaseShaderCompiler is called. We must ensure that the
builtins have been (re)initialized before attempting to link with the
builtin shader.
Signed-off-by: Ilia Mirkin <imirkin@alum.mit.edu>
Reviewed-by: Timothy Arceri <timothy.arceri@collabora.com>
Tested-by: Rob Herring <robh@kernel.org>
Cc: mesa-stable@lists.freedesktop.org
All interface blocks will have been lowered by this point so just
use an assert. Returning false would have caused all sorts of
problems if they were not lowered yet and there is an assert to
catch this later anyway.
We also update the tests to reflect this change.
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
We use the simple batch helper to submit a batch at driver startup time
which holds all the state that never changes. We don't have a whole lot
and once we enable tesselation there'll be even less. Even so, it's a
simple mechanism and reduces our steady state batch sizes a bit.
The gen7 pipeline has a useful helper function for this, let's use it in
gen8_pipeline.c too. The gen7 function has an off-by-one bug though: we
have to compute log2(size / 1024) - 1, but we divide by 2048 instead so
as to avoid the case where size is less than 1024 and we'd return -1.
Fixes compute since 7dd31b81fe
gallium/radeon: support PIPE_CAP_SURFACE_REINTERPRET_BLOCKS
Signed-off-by: Jan Vesely <jan.vesely@rutgers.edu>
Signed-off-by: Marek Olšák <marek.olsak@amd.com>
The vec4 backend, at the end, does this:
if (inst->is_3src()) {
for (int i = 0; i < 3; i++) {
if (inst->src[i].vstride == BRW_VERTICAL_STRIDE_0)
assert(brw_is_single_value_swizzle(inst->src[i].swizzle));
So make sure that we use the same conditions when trying to
copy-propagate. UNIFORMs will be converted to vstride 0 in
convert_to_hw_regs, but so will ATTRs when interleaved (as will happen
in a GS with multiple attributes). Since the vstride is not set at
copy-prop time, infer it by inspecting dispatch_mode and reject ATTRs if
they have non-scalar swizzles and are interleaved.
Fixes assertion errors in dolphin-generated geometry shaders (or
misrendering on opt builds) on Sandybridge or on IVB/HSW with
INTEL_DEBUG=nodualobj.
Co-authored-by: Ilia Mirkin <imirkin@alum.mit.edu>
Reviewed-by: Ilia Mirkin <imirkin@alum.mit.edu>
Bugzilla: https://bugs.freedesktop.org/show_bug.cgi?id=93418
Cc: "11.0 11.1" <mesa-stable@lists.freedesktop.org>