It's rather surprising that we've never actually hit this before.
Aparently, Ian's SPIR-V generator currently claims the Simple when you
don't do anything complex. We really shouldn't assert-fail on it.
Reviewed-by: Ian Romanick <ian.d.romanick@intel.com>
Reviewed-by: Bas Nieuwenhuizen <bas@basnieuwenhuizen.nl>
Cc: mesa-stable@lists.freedesktop.org
(cherry picked from commit 8ab9820d34)
See my LLVM patch which fixes the root cause.
Users have to apply this patch and then they have 2 choices:
- Downgrade to LLVM 5.0
- Update to LLVM git after my LLVM patch is pushed.
It won't be possible to use current and earlier development version
of LLVM 6.0.
Reviewed-by: Nicolai Hähnle <nicolai.haehnle@amd.com>
Cc: 17.3 <mesa-stable@lists.freedesktop.org>
(cherry picked from commit 3f8e3c2bd8)
For CPU access.
Signed-off-by: Leo Liu <leo.liu@amd.com>
Cc: mesa-stable@lists.freedesktop.org
Reviewed-by: Christian König <christian.koenig@amd.com>
(cherry picked from commit ea3dc75d72)
We currently have a bug where nir_lower_system_values gets called before
nir_lower_var_copies so it will miss any system value uses which come
from a copy_var intrinsic. Moving it to after brw_preprocess_nir fixes
this problem.
Reviewed-by: Lionel Landwerlin <lionel.g.landwerlin@intel.com>
Cc: mesa-stable@lists.freedesktop.org
(cherry picked from commit 279f8fb69c)
In order to implement the ballot intrinsic, we do a MOV from flag
register to some GRF. If that GRF is used in a SEL, cmod propagation
helpfully changes it into a MOV from the flag register with a cmod.
This is perfectly valid but when lower_simd_width comes along, it simply
splits into two instructions which both have conditional modifiers.
This is a problem since we're reading the flag register. This commit
makes us check whether or not flags_written() overlaps with the flag
values that we are reading via the instruction source and, if we have
any interference, will force us to emit a copy of the source.
Reviewed-by: Matt Turner <mattst88@gmail.com>
Cc: mesa-stable@lists.freedesktop.org
(cherry picked from commit fa6e74e33e)
v2: use a more generic compat function
v3: rename and formatting cleanup
Bugzilla: https://bugs.freedesktop.org/show_bug.cgi?id=103388
Signed-off-by: Jan Vesely <jan.vesely@rutgers.edu>
Reviewed-by: Francisco Jerez <currojerez@riseup.net>
CC: <mesa-stable@lists.freedesktop.org>
(cherry picked from commit a6d38f476b)
Not sure how useful this is, but it makes it more consistent.
Reviewed-by: Bas Nieuwenhuizen <bas@basnieuwenhuizen.nl>
Cc: "17.3" <mesa-stable@lists.freedesktop.org>
Signed-off-by: Dave Airlie <airlied@redhat.com>
(cherry picked from commit d8cefaa197)
This was the actual cause of GPU hangs fixed by 0fdd531457 ("radv:
Fix pipeline cache locking issues"), since multiple threads would end
up trying to create the variants for a single entry.
Now that we're locking around the whole of this function, this isn't
really necessary (we either create all or none of the variants), but
fix this anyway in case things change later.
Signed-off-by: Alex Smith <asmith@feralinteractive.com>
Reviewed-by: Bas Nieuwenhuizen <bas@basnieuwenhuizen.nl>
CC: 17.3 <mesa-stable@lists.freedesktop.org>
(cherry picked from commit fee9d05e21)
The kernel doesn't initialize the value of the INSTPM or CS_DEBUG_MODE2
registers at context initialization time. Instead, they're inherited
from whatever happened to be running on the GPU prior to first run of a
new context. So, when we started setting these, other contexts in the
system started inheriting our values. Since this controls whether
3DSTATE_CONSTANT_* takes a pointer or an offset, getting the wrong
setting is fatal for almost any process which isn't expecting this.
Unfortunately, VA-API and Beignet don't initialize this (nor does older
Mesa), so they will die horribly if we start doing this. UXA and SNA
don't use any push constants, so they are unaffected.
Until we have some kind of solution to this problem, I'm going to revert
this patch and abandon using the feature for now. It will lead to fewer
pushed UBO ranges on Broadwell+, which may lead to lower performance,
though I don't have any data on the impact.
Cc: "17.3 17.2" <mesa-stable@lists.freedesktop.org>
Bugzilla: https://bugs.freedesktop.org/show_bug.cgi?id=102774
(cherry picked from commit 013d331220)
There are two issues with the current implementation. First, it relies
on the layout(local_size_*) happening in the same shader as the main
function, and secondly it doesn't work for variable group sizes.
In both cases, the simplest fix is to move the setup of these derived
values to a later time, similar to how the gl_VertexID workarounds are
done. There already exist system values defined for both of the derived
values, so we use them unconditionally, and lower them after linking is
performed.
While we're at it, we move to using gl_LocalGroupSizeARB instead of
gl_WorkGroupSize for variable group sizes.
Also the dead code elimination avoidance can be removed, since there
can be situations where gl_LocalGroupSizeARB is needed but has not been
inserted for the shader with main function. As a result, the lowering
code has to insert its own copies of the system values if needed.
Reported-by: Stephane Chevigny <stephane.chevigny@polymtl.ca>
Bugzilla: https://bugs.freedesktop.org/show_bug.cgi?id=103393
Cc: mesa-stable@lists.freedesktop.org
Signed-off-by: Ilia Mirkin <imirkin@alum.mit.edu>
Reviewed-by: Jordan Justen <jordan.l.justen@intel.com>
Reviewed-by: Samuel Pitoiset <samuel.pitoiset@gmail.com>
(cherry picked from commit 4d24a7cb97)
Vulkan CTS does not expect the value to be clamped (at least for D32),
and it makes a differences even though depth is in [0,1], due
to strict inequalities.
I couldn't find anything in the Vulkan spec about this, but the test
seemed to be copied from GL tests and the GL spec only specifies
clamping for fixed point formats. Hence I expect radeonsi to run into
this at some point as well, but given that they still have a usecase
with the Z16->Z32 promotion, I'll leave that for someone else to clean
up.
This at least fixes radv dEQP-VK.texture.shadow.* on VI.
Fixes: 0f9e32519b 'ac/nir: clamp shadow texture comparison value on VI'
Reviewed-by: Dave Airlie <airlied@redhat.com>
Since it also uses the output vector before writing to memory.
Fixes: e38685cc62 'Revert "radv: disable support for VEGA for now."'
Reviewed-by: Dave Airlie <airlied@redhat.com>
Reviewed-by: Timothy Arceri <tarceri@itsqueeze.com>
Otherwise we just need to write them to the tf ring.
this seems to improve the tessellation demo on Bonarie
~2190->~2230 fps
Reviewed-by: Bas Nieuwenhuizen <bas@basnieuwenhuizen.nl>
Signed-off-by: Dave Airlie <airlied@redhat.com>
Due to LLVM bugs. Fixes a bunch of dEQP-VK.glsl.indexing.*
tests.
Fixes: e38685cc62 'Revert "radv: disable support for VEGA for now."'
Reviewed-by: Dave Airlie <airlied@redhat.com>
When the gs_copy_shader is NULL (due to an incomplete cache), but
the main shaders are found, we still do the nir, but we shouldn't
compile the shaders again. For merged shaders we should also account
for the missing shaders.
Fixes: ce03c119ce 'radv: Add code to compile merged shaders.'
Reviewed-by: Dave Airlie <airlied@redhat.com>
With merged shaders the vertex shader may not exist. This got in
because the offending patch was written before merged shaders were
upstream, but committed after.
Fixes: 75dfab24a2 'radv: refactor indirect draws with radv_draw_info'
Reviewed-by: Samuel Pitoiset <samuel.pitoiset@gmail.com>
Otherwise for non-indexed draws we set and immediately unset
RADV_CMD_DIRTY_INDEX_BUFFER. As all the set functions should
clear their own bit, this is unnecessary.
Fixes: 341529dbee 'radv: use optimal packet order for draws'
Reviewed-by: Samuel Pitoiset <samuel.pitoiset@gmail.com>
As they were emitted after the new pipeline, the changed pipeline
detection was not working anymore.
Fixes: 341529dbee 'radv: use optimal packet order for draws'
Reviewed-by: Samuel Pitoiset <samuel.pitoiset@gmail.com>
They don't take a single wave anymore and we need the barriers.
Fixes: 6bc42855f9 'radv: enable GS on GFX9'
Reviewed-by: Dave Airlie <airlied@redhat.com>
Need to lock around the whole process of retrieving cached shaders, and
around GetPipelineCacheData.
This fixes GPU hangs observed when creating multiple pipelines in
parallel, which appeared to be due to invalid shader code being pulled
from the cache.
Signed-off-by: Alex Smith <asmith@feralinteractive.com>
Reviewed-by: Bas Nieuwenhuizen <bas@basnieuwenhuizen.nl>
v2: Warn that support is still in alpha (Jason)
Signed-off-by: Lionel Landwerlin <lionel.g.landwerlin@intel.com>
Reviewed-by: Jason Ekstrand <jason@jlekstrand.net>
This was causing Android clang version 3.8.256229 to miscompile,
presumably due to strict aliasing.
Fixes: 14dc281c13 ("vc4: Enforce one-uniform-per-instruction after optimization.")
This is based heavily on Daniel Stone's work for the same, rebased on
master and with a number of TODO's fixed.
This does not implement glvnd (which is coming in a later patch)
Meson builds egl slightly differently than autotools, namely it doesn't
build an intermediate shared library. It doesn't do this because meson
doesn't have problems with the name of the library being dynamically
generated, so the glvnd and non-glvnd code can follow the same path.
v2: - Don't reuse variable (Eric E.)
Signed-off-by: Dylan Baker <dylanx.c.baker@intel.com>
Tested-by: Eric Engestrom <eric.engestrom@imgtec.com>
Reviewed-by: Daniel Stone <daniels@collabora.com>
These files are needed by both vulkan wayland-wsi and by egl
wayland-wsi, since the XML file is in src/egl/wayland/wayland-drm and we
can include this directory in such a way that it will be loaded before
egl and vulkan this allows us to avoid multiple calls to the same
generator.
Signed-off-by: Dylan Baker <dylanx.c.baker@intel.com>
Reviewed-and-Tested-by: Eric Engestrom <eric@engestrom.ch>
Previously this failed to change with_glx to disabled from auto if
platform_x11 was unset or if no opengl apis were being built.
v2: - swap conditional positions
Signed-off-by: Dylan Baker <dylanx.c.baker@intel.com>
Reviewed-and-Tested-by: Eric Engestrom <eric@engestrom.ch>
fixes: b603725703 ("configure.ac: Bump libdrm_amdgpu version to 2.4.85.")
Signed-off-by: Dylan Baker <dylanx.c.baker@intel.com>
Reviewed-by: Eric Anholt <eric@anholt.net>
Reviewed-by: Bas Nieuwenhuizen <bas@basnieuwenhuizen.nl>
The pass only looks at var load/store intrinsics, not input load/store
intrinsics, so assert that we don't see the other type.
v2: Adjust comment indentation.
Reviewed-by: Timothy Arceri <tarceri@itsqueeze.com>
Implicit sync kicks in when a buffer is used by two different amdgpu
contexts simultaneously. Jobs that use explicit synchronization
mechanisms end up needlessly waiting to be scheduled for long periods
of time in order to achieve serialized execution.
This patch disables implicit synchronization for all radv allocations
except for wsi bos. The only systems that require implicit
synchronization are DRI2/3 and PRIME.
v2: mark wsi bos as RADV_MEM_IMPLICIT_SYNC
v3: Add drm version check (Bas)
Signed-off-by: Andres Rodriguez <andresx7@gmail.com>
Reviewed-by: Bas Nieuwenhuizen <bas@basnieuwenhuizen.nl>
This allows us to pass extra parameters to the memory allocation
operation that are not defined in the vulkan spec. This is useful for
internal usage.
Signed-off-by: Andres Rodriguez <andresx7@gmail.com>
Reviewed-by: Bas Nieuwenhuizen <bas@basnieuwenhuizen.nl>
This patch helps lower high priority compute latency. Found by
bisecting a perf regression on computeparticles with high priority
compute queues enabled.
Reverting this micro-optimization doesn't seem to have any negative
effect on performance on Dota2 or ssao.
Signed-off-by: Andres Rodriguez <andresx7@gmail.com>
Reviewed-by: Bas Nieuwenhuizen <bas@basnieuwenhuizen.nl>
This extension allows the caller to change a queue's system wide
priority. This is useful for applications with specific
latency constraints.
Signed-off-by: Andres Rodriguez <andresx7@gmail.com>
Reviewed-by: Bas Nieuwenhuizen <bas@basnieuwenhuizen.nl>
This is part of a cooperative scheduling approach used by radv. All
drivers in the stack must opt-in to resource arbitration, otherwise GL
based apps will be able to ignore system priorities.
We always hardcode the field to its maximum value, instead of attempting
to calculate an approximate usage. In testing, there were no benefits to
using anything other than the maximum.
Signed-off-by: Andres Rodriguez <andresx7@gmail.com>
Reviewed-by: Bas Nieuwenhuizen <bas@basnieuwenhuizen.nl>
When WAVE_LIMIT is set, a submission will opt-in for SPI based resource
scheduling. Because this mechanism is cooperative, we must ensure that
all submissions have this field set, otherwise they will bypass resource
arbitration.
We always hardcode the field to its maximum value, instead of attempting
to calculate an approximate usage. In testing, there were no benefits to
using anything other than the maximum.
Signed-off-by: Andres Rodriguez <andresx7@gmail.com>
Reviewed-by: Bas Nieuwenhuizen <bas@basnieuwenhuizen.nl>
