Fixes arb_shader_image_load_store/execution/load-from-cleared-image.shader_test.
Couldn't reproduce any significant FPS regression in CPU-bound
benchmarks from the Finnish benchmarking system on neither VLV nor BSW
after 30 runs with 95% confidence level.
Bugzilla: https://bugs.freedesktop.org/show_bug.cgi?id=92849
Cc: Chris Wilson <chris@chris-wilson.co.uk>
Cc: Jason Ekstrand <jason.ekstrand@intel.com>
Cc: "11.0 11.1" <mesa-stable@lists.freedesktop.org>
Tested-by: Jordan Justen <jordan.l.justen@intel.com>
Reviewed-by: Kristian Høgsberg <krh@bitplanet.net>
Every other gen the representation of the URB size was changed and
previous ones weren't updated. I'd be willing to write a series
normalizing this to be KB on all generations if anybody else cares.
This is going to require some rather intrusive kernel changes to fix
properly, in the meantime (and forever on at least pre-v4.1 kernels)
we'll have to restore the hardware defaults at the end of every batch
in which the L3 configuration was changed to avoid interfering with
the DDX and GL clients that use an older non-L3-aware version of Mesa.
Reviewed-by: Samuel Iglesias Gonsálvez <siglesias@igalia.com>
Reviewed-by: Kristian Høgsberg <krh@bitplanet.net>
v2: Optimize look-up of the default configuration by assuming it's the
first entry of the L3 config array in order to avoid an FPS
regression in GpuTest Triangle and SynMark OglBatch2-7 on most
affected platforms.
Reviewed-by: Jordan Justen <jordan.l.justen@intel.com>
The L3 state atom calculates the target L3 partition weights when the
program bound to some shader stage is modified, and in case they are
far enough from the current partitioning it makes sure that the L3
state is re-emitted.
v2: Fix for inconsistent units the context URB size is expressed in.
Clamp URB size to 1008 KB on SKL due to FF hardware limitation.
Reviewed-by: Samuel Iglesias Gonsálvez <siglesias@igalia.com>
Acked-by: Kenneth Graunke <kenneth@whitecape.org>
Reviewed-by: Kristian Høgsberg <krh@bitplanet.net>
This calculates a rather conservative partitioning of the L3 cache
based on the shaders currently bound to the pipeline and whether they
use SLM, atomics, images or scratch space. The result is intended to
be fine-tuned later on based on other pipeline state.
Note that the L3 partitioning calculated for VLV in the non-SLM non-DC
case differs from the hardware defaults in that it doesn't include a
DC partition and has twice as much RO cache space -- This is an
intentional functional change that improves performance in several
bandwidth-bound benchmarks on VLV (5% significance): SynMark
OglTexFilterAniso by 14.18%, SynMark OglTexFilterTri by 7.15%, Unigine
Heaven by 4.91%, SynMark OglShMapPcf by 2.15%, GpuTest Fur by 1.83%,
SynMark OglDrvRes by 1.80%, SynMark OglVsTangent by 1.71%, and a few
other benchmarks from the Finnish system by less than 1%.
Reviewed-by: Samuel Iglesias Gonsálvez <siglesias@igalia.com>
Reviewed-by: Kristian Høgsberg <krh@bitplanet.net>
The input of the L3 set-up code is a vector giving the approximate
desired relative size of each partition. This implements logic to
compare the input vector against the table of validated configurations
for the device and pick the closest compatible one.
Reviewed-by: Samuel Iglesias Gonsálvez <siglesias@igalia.com>
Reviewed-by: Kristian Høgsberg <krh@bitplanet.net>
Improves performance of the arb_shader_image_load_store-atomicity
piglit test by over 25x (which isn't a real benchmark it's just heavy
on atomics -- the improvement in a microbenchmark I wrote a while ago
seemed to be even greater). The drawback is one needs to be
extra-careful not to hang the GPU (in fact the whole system). A DC
partition must have been allocated on L3, the "convert L3 cycle for DC
to UC" bit may not be set, the MOCS L3 cacheability bit must be set
for all surfaces accessed using DC atomics, and the SCRATCH1 and
ROW_CHICKEN3 bits must be kept in sync.
A fairly recent kernel is required for the command parser to allow
writes to these registers.
Reviewed-by: Samuel Iglesias Gonsálvez <siglesias@igalia.com>
Acked-by: Kenneth Graunke <kenneth@whitecape.org>
Reviewed-by: Kristian Høgsberg <krh@bitplanet.net>
It should be possible to use additional L3 configurations other than
the ones listed in the tables of validated allocations ("BSpec »
3D-Media-GPGPU Engine » L3 Cache and URB [IVB+] » L3 Cache and URB [*]
» L3 Allocation and Programming"), but it seems sensible for now to
hard-code the tables in order to stick to the hardware docs. Instead
of setting up the arbitrary L3 partitioning given as input, the
closest validated L3 configuration will be looked up in these tables
and used to program the hardware.
The included tables should work for Gen7-9. Note that the quantities
are specified in ways rather than in KB, this is because the L3
control registers expect the value in ways, and because by doing that
we can re-use a single table for all GT variants of the same
generation (and in the case of IVB/HSW and CHV/SKL across different
generations) which generally have different L3 way sizes but allow the
same combinations of way allocations.
v2: Use slice count from the devinfo structure instead of the gt
number to implement get_l3_way_size().
Reviewed-by: Samuel Iglesias Gonsálvez <siglesias@igalia.com>
Acked-by: Kenneth Graunke <kenneth@whitecape.org>
Reviewed-by: Kristian Høgsberg <krh@bitplanet.net>
According to the hardware docs a DC flush is sufficient to make
CS_STALL happy, there's no need to add STALL_AT_SCOREBOARD whenever
it's present.
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
Reviewed-by: Kristian Høgsberg <krh@bitplanet.net>
This will make sure that we recalculate the URB layout anytime the URB
size is modified by the L3 partitioning code.
Reviewed-by: Jordan Justen <jordan.l.justen@intel.com>
Reviewed-by: Kristian Høgsberg <krh@bitplanet.net>
This stores the result of can_do_pipelined_register_writes() in the
context struct so we can find out later whether LRI can be used to
program the L3 configuration.
v2:
* Split change of gen check in can_do_pipelined_register_writes (jljusten)
Reviewed-by: Jordan Justen <jordan.l.justen@intel.com>
Reviewed-by: Kristian Høgsberg <krh@bitplanet.net>
Allow for pipelined register writes for gen < 7.
v2:
* Split from another patch and adjust comment (jljusten)
Reviewed-by: Jordan Justen <jordan.l.justen@intel.com>
Reviewed-by: Kristian Høgsberg <krh@bitplanet.net>
This adds the double code to the geometry shader input handling.
Reviewed-by: Michel Dänzer <michel.daenzer@amd.com>
Cc: "11.0 11.1" <mesa-stable@lists.freedesktop.org>
Signed-off-by: Dave Airlie <airlied@redhat.com>
This handles loading doubles from LDS properly.
Reviewed-by: Michel Dänzer <michel.daenzer@amd.com>
Cc: "11.0 11.1" <mesa-stable@lists.fedoraproject.org>
Signed-off-by: Dave Airlie <airlied@redhat.com>
This fixes:
glsl-1.50/execution/geometry/dynamic_input_array_index.shader_test
my profanity.
We need to load the AR register with the value from the index reg
Cc: "11.0 11.1" <mesa-stable@lists.freedesktop.org>
Signed-off-by: Dave Airlie <airlied@redhat.com>
This fixes:
gs-input-array-vec4-index-rd
The others run out of gprs unfortunately.
Cc: "11.0 11.1" <mesa-stable@lists.freedesktop.org>
Signed-off-by: Dave Airlie <airlied@redhat.com>
This fixes:
arb_transform_feedback3-ext_interleaved_two_bufs_gs
arb_transform_feedback3-ext_interleaved_two_bufs_gs_max
transform-feedback-builtins
If we are only emitting one ring, then emit all output
buffers on it.
Cc: "11.0 11.1" <mesa-stable@lists.freedesktop.org>
Signed-off-by: Dave Airlie <airlied@redhat.com>
[airlied: update relnotes]
Reviewed-by: Marek Olšák <marek.olsak@amd.com>
Signed-off-by: Edward O'Callaghan <eocallaghan@alterapraxis.com>
Signed-off-by: Dave Airlie <airlied@redhat.com>
See: `commit e82c527f1fc2f8ddc64954ecd06b0de3cea92e93`
which is where a block in src maps to a pixel in dst and vice versa.
e.g. DXT1 <-> R32G32_UINT
DXT5 <-> R32G32B32A32_UINT
Reviewed-by: Marek Olšák <marek.olsak@amd.com>
Signed-off-by: Edward O'Callaghan <eocallaghan@alterapraxis.com>
Signed-off-by: Dave Airlie <airlied@redhat.com>
The only effect here is a space savings - 822 programs in shader-db
affected with the following overall change:
total bytes used in shared programs : 44154976 -> 44139880 (-0.03%)
Fixes: 641eda0c (nv50/ir: r63 is only 0 if we are using less than 63 registers)
Signed-off-by: Ilia Mirkin <imirkin@alum.mit.edu>
Cc: "11.0 11.1" <mesa-stable@lists.freedesktop.org>
This allows us to use the short encoding, and potentially fold
immediates in later on.
total instructions in shared programs : 6379731 -> 6367861 (-0.19%)
total gprs used in shared programs : 728502 -> 728683 (0.02%)
total local used in shared programs : 9904 -> 9904 (0.00%)
total bytes used in shared programs : 44661008 -> 44154976 (-1.13%)
local gpr inst bytes
helped 0 51 7267 20306
hurt 0 232 125 274
Signed-off-by: Ilia Mirkin <imirkin@alum.mit.edu>
Operations that take immediates can only encode registers up to 64. This
fixes a shader in a "Powered by Unity" intro.
Signed-off-by: Ilia Mirkin <imirkin@alum.mit.edu>
We already semi-did this but the list of uses as unsorted, so it was
unreliable. Sort the uses by bb and serial, and don't unspill for each
instruction in a sequence. (And also don't unspill multiple times for a
single instruction that uses the value in question multiple times.)
This causes a minor reduction in generated instructions for shader-db
(as few programs spill) but more importantly it brings determinism to
each run's output.
On SM10:
total instructions in shared programs : 6387945 -> 6379359 (-0.13%)
total gprs used in shared programs : 728544 -> 728544 (0.00%)
total local used in shared programs : 9904 -> 9904 (0.00%)
local gpr inst bytes
helped 0 0 322 322
hurt 0 0 0 0
Signed-off-by: Ilia Mirkin <imirkin@alum.mit.edu>
This fixes the fetching of fp64 inputs to the geometry shader,
this fixes the recently posted piglit's
arb_gpu_shader_fp64/execution/gs-fs-vs-double-array.shader_test
arb_vertex_attrib_64bit/execution/gs-fs-vs-attrib-double-array.shader_test
Reviewed-by: Roland Scheidegger <sroland@vmware.com>
Signed-off-by: Dave Airlie <airlied@redhat.com>
Reviewed-by: Jason Ekstrand <jason.ekstrand@intel.com> [v1]
Reviewed-by: Eric Anholt <eric@anholt.net> [v1]
v2: Move new rule to Boolean simplification section
Add a a@bool != true simplification
Suggested-by: Neil Roberts <neil@linux.intel.com>
... and allow the "binding" qualifier in ES 3.1 as well.
GLSL ES 3.1 incorporates only a few features from the extension
ARB_shading_language_420pack: the relaxed qualifier ordering
requirements and the binding qualifier.
Cc: "11.1" <mesa-stable@lists.freedesktop.org>
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
These features would not have been enabled with #version 420 otherwise.
Cc: "11.1" <mesa-stable@lists.freedesktop.org>
Reviewed-by: Ilia Mirkin <imirkin@alum.mit.edu>
This interaction was missed in the addition of ARB_image_load_store.
Cc: "11.0 11.1" <mesa-stable@lists.freedesktop.org>
Bugzilla: https://bugs.freedesktop.org/show_bug.cgi?id=93266
Reviewed-by: Ilia Mirkin <imirkin@alum.mit.edu>
Reviewed-by: Ian Romanick <ian.d.romanick@intel.com>
Unforunately the Appveyor -> SourceForge connection seems a bit
unreliable, causing frequent build failures while downloading
winflexbison (approx once every 2 days).
Fetching winflexbison archive into Appveyor's cache should eliminate
these.
Fetching Python modules from PyPI doesn't seem to be a problem, so they
are left alone for now, though they could eventually get the same
treatment.
We still have several failures in the newly enabled tests in simulation:
sRGB downsampling is done as if it was just linear, stencil blits are not
supported on MSAA either, and derivatives are still not supported
(breaking some MSAA simulation shaders). So, other than sRGB downsampling
quality, things seem to be in good shape.
This is the core of ARB_texture_multisample. Most of the piglit tests for
GL_ARB_texture_multisample require GL 3.0, but exposing support for this
lets us use the gallium blitter for multisample resolves. We can
sometimes multisample resolve using just the RCL, but that requires that
the blit is 1:1, unflipped, and aligned to tile boundaries.
This includes GL_SAMPLE_COVERAGE, GL_SAMPLE_ALPHA_TO_ONE, and
GL_SAMPLE_ALPHA_TO_COVAGE.
I haven't implemented a dithering function yet, and gallium doesn't give
me a good chance to do so for GL_SAMPLE_COVERAGE.
I only stumbled on this while experimenting due to reading about HW-2905.
I don't know if the EZ disable in the Z-clear is actually necessary, but
go with it for now.
The recent unaligned fix successfully prevented RCL blits that weren't
aligned inside of the surface, but we also want to be able to do RCL blits
for the whole surface when the width or height of the surface aren't
aligned (we don't care what renders inside of the padding).
Emil Velikov
Francisco Jerez
Dave Airlie
Edward O'Callaghan
Ilia Mirkin
Matt Turner
Jose Fonseca
Eric Anholt