Elk still uses param array, so it depends on this behavior. This fixes
an assertion failure on Broadwell in gfxbench4/carchase/339.shader_test.
src/intel/compiler/elk/elk_fs_nir.cpp:148: void fs_nir_setup_uniforms(elk_fs_visitor&): Assertion `s.uniforms == s.prog_data->nr_params' failed.
Fixes: f4a0e05970 ("anv/brw/iris: get rid of param array on prog_data")
Reviewed-by: Lionel Landwerlin <lionel.g.landwerlin@intel.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/39280>
ACO got a lot better at forming VOPD instructions, and testing
feedback seems to point in a slightly positive direction for this.
gfx12 will also start requiring wave32 for dynamic VGPR allocation at
some point.
Measurements on navi31:
Cyberpunk 2077:
Difference at 95.0% confidence
1.12333 +/- 0.42876
1.88216% +/- 0.718391%
(Student's t, pooled s = 0.189165)
Black Myth Wukong benchmark:
Difference at 95.0% confidence
4 +/- 1.30862
13.9535% +/- 4.56495%
(Student's t, pooled s = 0.57735)
Portal with RTX:
66.2ms->61.5ms (~7.64% improvement)
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/39275>
Add nir_io_compact_to_higher_16 flag so that the pass knows if it can
compact 16-bit varyings into the higher 16 bits of a 32-bit varying.
Reviewed-by: Marek Olšák <marek.olsak@amd.com>
Signed-off-by: Aitor Camacho <aitor@lunarg.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/38994>
There are some situations where we don't get a binary. In those cases
it is not relevant to report the statistics.
This was detected while using shader-db, with a shader that has a xfb
vertex shader that was not writing to gl_Position. It that case
IDVS_POSITION binary was zero, so later compiling IDVS_VARYING was
skipped. Due the skip, the pan_stat structure was not initialized, so
the debug used the wrong isa, that used different measurements. This
lead to shader-db report tool failing, due having a mix-up of
measurements.
Although an alternative would be to try to ensure that the structure
is always initialized, seems just more natural to just not report on
shaders with empty binaries (as the stats would be zero for all
measurements).
Reviewed-by: Christoph Pillmayer <christoph.pillmayer@arm.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/39310>
We just... didn't do this at all??? I have no idea how this didn't blow
up before, given that plenty of apps should generate a traversal shader
that spills (and thus has a large stack size), but it did finally blow
up in function-call related work.
Cc: mesa-stable
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/29580>
G720 isn't wired up in CI, and nobody is running these regularly. So we
have missed that these CTS issues were fixed upstream.
I haven't actually verified that this works, because I don't have a G720
around at the moment. But the fixes fixed both v10 and v13 GPUs, so it
would be really, really strange if this wasn't effective on v11 as well.
Reviewed-by: Christoph Pillmayer <christoph.pillmayer@arm.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/39034>
These arguments misrenders, and gets rendered using a long hyphen
instead of double dashes. Let's fix this by using the Sphinx
option-directive, which is meant for exactly this purpose instead.
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/39286>
I'm not convinced these really should be separate opcodes at all in NIR, but
that's not what this patch is about. Here we just infer the opcodes in the
texture builder to allow simplified usage.
This lets us drop nir_txl() & nir_txb() helpers in favour of nir_tex(.lod/bias)
which is more normalized. We could also drop nir_txf_ms in favour of nir_txf but
that affects more callsites and is not obviously a win (unlike nir_txl which is
used once and nir_txb which is unused).
Signed-off-by: Alyssa Rosenzweig <alyssa.rosenzweig@intel.com>
Reviewed-by: Georg Lehmann <dadschoorse@gmail.com>
Reviewed-by: Ian Romanick <ian.d.romanick@intel.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/39271>
It is not allowed in SDS. This matters on gen8.
Possibly it wasn't enforced, or worked by luck, on older gens. But at
least for gen7 and gen8 it is documented as not allowed in SDS (so
working or not may depend on fw version, phase of the moon, etc).
Signed-off-by: Rob Clark <rob.clark@oss.qualcomm.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/39254>
The format of the packet changed with gen8. This went unnoticed because
in this particular case the contents of the packet work out the same.
But lets be correct about this.
Signed-off-by: Rob Clark <rob.clark@oss.qualcomm.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/39254>
At least on gen8 (unsure if hw or fw change) this makes things pissy and
results in writes to undefined offsets.
Non-zero chance that this was even UB on older gens, but ended up
writing somewhere harmless(ish).
Signed-off-by: Rob Clark <rob.clark@oss.qualcomm.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/39254>
vi_bindings_valid doesn't necessarily match the # of VBOs emitted,
resulting an invalid size in the CP_SET_DRAW_STATE packet. Somehow
this didn't seem to cause problems prior to Dxx (although may
potentially have been a source of flakes, depending on what random
cmds followed in memory). But caused hangs on Dxx.
See, for example, dEQP-VK.pipeline.fast_linked_library.vertex_input.misc.unused_binding
Fixes: 97da0a7734 ("tu: Rewrite to use common Vulkan dynamic state")
Signed-off-by: Rob Clark <rob.clark@oss.qualcomm.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/39254>
The Android 16 Cuttlefish build comes with a new Xe minigbm backend that
we can use on the host in the android-angle-venus-anv-adl nightly job.
Also update two outdated comments.
Signed-off-by: Valentine Burley <valentine.burley@collabora.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/39297>
