It isn't just load instructions that have write-after-read hazzard.
Fixes stk gaussian blur compute shaders.
Signed-off-by: Rob Clark <robdclark@gmail.com>
In some cases we can end up trying to add a write dependency on ourself,
which shouldn't trigger a flush.
Avoids an extra couple flushes per from in stk.
Signed-off-by: Rob Clark <robdclark@gmail.com>
ctx->last_fence isn't such a terribly clever idea, if batches can be
flushed out of order. Instead, each batch now holds a fence, which is
created before the batch is flushed (useful for next patch), that later
gets populated after the batch is actually flushed.
Signed-off-by: Rob Clark <robdclark@gmail.com>
In transfer_map(), when we need to flush batches that read from a
resource, we should be holding screen->lock to guard against race
conditions. Somehow deferred flush seems to make this existing
race more obvious.
Signed-off-by: Rob Clark <robdclark@gmail.com>
This is a bit more general and lets us pass additional options into the
spirv_to_nir pass beyond what capabilities we support.
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
Reviewed-by: Kristian H. Kristensen <hoegsberg@google.com>
Instead of emitting absolutely everything, just emit the few functions
that are actually referenced in some way by the entrypoint. This should
save us quite a bit of time when handed large shader modules containing
many entrypoints.
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
Reviewed-by: Kristian H. Kristensen <hoegsberg@google.com>
We have a nir_builder and it has an impl field.
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
Reviewed-by: Kristian H. Kristensen <hoegsberg@google.com>
Fixes MESA_GLSL=cache_fb with piglit
tests/spec/glsl-1.50/execution/geometry/clip-distance-vs-gs-out.shader_test
Fixes: 0610a624a1 i965/link: Serialize program to nir after linking for shader cache
Bugzilla: https://bugs.freedesktop.org/show_bug.cgi?id=103988
Signed-off-by: Jordan Justen <jordan.l.justen@intel.com>
Reviewed-by: Jason Ekstrand <jason@jlekstrand.net>
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
Since almost all BOs will be in one CL at a time, this cache will almost
always hit except for the first usage of the BO in each CL.
This didn't show up as statistically significant on the minetest trace
(n=340), but if I lop off the throttled lobe of the bimodal distribution,
it very clearly does (0.74731% +/- 0.162093%, n=269).
No significant difference in the minetest replay, but it should reduce
overhead by not requiring that we write quad indices to index buffers that
we repeatedly re-upload (and making the draw packet smaller, as well).
Over the course of the series the actual game seems to be up by 1-2 fps.
Now that there's only one user of it, it's pretty obvious how to avoid
emitting redundant ones. This should save a bunch of kernel validation
overhead.
No statistically sigificant difference on the minetest trace I was looking
at (n=169), but the maximum FPS is up by .3%
Originally there was CL code for handling various relocations back when I
had relocs for the TSDA/TA buffers. Now that the kernel handles those
entirely on its own, I can inline that code into the one place using it.
We failed to take the start into account for how many vertices to draw in
this round, so we would end up decrementing count below 0, which as an
unsigned number meant we would loop until the CLs soon ran out of space.
When I wrote the code I was thinking about how to use the previously
emitted shader state (no index bias baked into the elements) by emitting
up to 65535 and then only re-emitting with bias for the second wround, but
that doesn't work if the start is over 65535. Instead, just delay
emitting shader state until we get into the drawarrays GFXH-515 loop and
always bake the bias in when we're doing the workaround.
They are the same thing, but this is more consistent with the rest of
the project.
Signed-off-by: Dylan Baker <dylanx.c.baker@intel.com>
Reviewed-by: Eric Engestrom <eric.engestrom@imgtec.com>
v2: - Make -Dlmsensors=false work
- Simplify auto and true cases
Signed-off-by: Dylan Baker <dylanx.c.baker@intel.com>
Reviewed-by: Eric Engestrom <eric.engestrom@imgtec.com>
When we look up the DRI drawable state we need to associate an fbconfig
with the drawable. With GLX_EXT_no_config_context we can no longer infer
that from the context and must instead query the server.
Signed-off-by: Adam Jackson <ajax@redhat.com>
Reviewed-by: Ian Romanick <ian.d.romanick@intel.com>
This also fixes a bug, the error path through MakeCurrent didn't
translate the error code by the extension's error base.
Signed-off-by: Adam Jackson <ajax@redhat.com>
Reviewed-by: Ian Romanick <ian.d.romanick@intel.com>
The dummy vtable has these slots as NULL already, no need to check for
the dummy context explicitly.
Signed-off-by: Adam Jackson <ajax@redhat.com>
Reviewed-by: Emil Velikov <emil.velikov@collabora.com>
v2: Missing td tag, add Andres + Juan for 17.2.8 and 17.3.3
Signed-off-by: Emil Velikov <emil.velikov@collabora.com>
Acked-by: Nicolai Hähnle <nicolai.haehnle@amd.com> (v1)
Reviewed-by: Andres Gomez <agomez@igalia.com>
Reviewed-by: Juan A. Suarez Romero <jasuarez@igalia.com>
Aimed to work with Glide, which hasn't been a thing in over 10 years.
There are no drivers that implement it, so annotate it as obsolete
v2: Move the extension to OLD/
Signed-off-by: Emil Velikov <emil.velikov@collabora.com>
Reviewed-by: Brian Paul <brianp@vmware.com> (v1)
Reviewed-by: Adam Jackson <ajax@redhat.com> (v1)
Reviewed-by: Ian Romanick <ian.d.romanick@intel.com>
The implementation is a simple 'return EGL_FALSE'. Stop pretending and
simply remove it.
Note: the removal of XMesa API is fine, since there hasn't been any
users for it in years.
Signed-off-by: Emil Velikov <emil.velikov@collabora.com>
Reviewed-by: Brian Paul <brianp@vmware.com>
Reviewed-by: Adam Jackson <ajax@redhat.com>
Reviewed-by: Ian Romanick <ian.d.romanick@intel.com>
No Mesa driver has implemented the extension in ages. Seemingly non Mesa
drivers don't implement it either.
As mentioned by Ian, the extension is effectively superseded by
ARB_vertex_buffer_object.
v2: Move the extension to OLD/
Signed-off-by: Emil Velikov <emil.velikov@collabora.com>
Reviewed-by: Brian Paul <brianp@vmware.com> (v1)
Reviewed-by: Adam Jackson <ajax@redhat.com> (v1)
Reviewed-by: Ian Romanick <ian.d.romanick@intel.com>
The extension was never implemented and seemingly never will.
The DRI based libGL dropped support for it over 10 years ago.
Signed-off-by: Emil Velikov <emil.velikov@collabora.com>
Reviewed-by: Brian Paul <brianp@vmware.com>
Reviewed-by: Adam Jackson <ajax@redhat.com>
Reviewed-by: Ian Romanick <ian.d.romanick@intel.com>
The extension was never implemented and seemingly never will.
The DRI based libGL dropped support for it over 10 years ago.
Signed-off-by: Emil Velikov <emil.velikov@collabora.com>
Reviewed-by: Brian Paul <brianp@vmware.com>
Reviewed-by: Adam Jackson <ajax@redhat.com>
Reviewed-by: Ian Romanick <ian.d.romanick@intel.com>
I believe the workaround describes that the MI_LOAD_REGISTER_IMM should
come right after the 3DSTATE_SAMPLE_PATTERN.
This fixes GPU hangs in the i965 initial state batchbuffer when running
some Piglit tests with always_flush_batch=true.
Signed-off-by: Rafael Antognolli <rafael.antognolli@intel.com>
Cc: Kenneth Graunke <kenneth@whitecape.org>
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
On CNL, we see multiple multisample failures on piglit tests. By
emitting this extra state, though not documented in the bspec, those
failures seem to go away.
This workaround could be removed if we ever find out a better solution,
but it should be good enough for now.
Signed-off-by: Rafael Antognolli <rafael.antognolli@intel.com>
Cc: Kenneth Graunke <kenneth@whitecape.org>
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
Reported-by: Marc Dietrich <marvin24@gmx.de>
Signed-off-by: Dylan Baker <dylanx.c.baker@intel.com>
Reviewed-by: Emil Velikov <emil.velikov@collabora.com>
Reviewed-by: Eric Engestrom <eric.engestrom@imgtec.com>
In this condition dri2_dpy->driver_name string always equals
NULL, so call to free() is useless
Signed-off-by: Vadym Shovkoplias <vadym.shovkoplias@globallogic.com>
Reviewed-by: Eric Engestrom <eric.engestrom@imgtec.com>
deviceName string is declared, assigned and freed but actually
never used in dri3_create_screen() function.
Fixes: 2d94601582 ("Add DRI3+Present loader")
Signed-off-by: Vadym Shovkoplias <vadym.shovkoplias@globallogic.com>
Reviewed-by: Eric Engestrom <eric.engestrom@imgtec.com>
"vkAllocateCommandBuffers can be used to create multiple command
buffers. If the creation of any of those command buffers fails, the
implementation must destroy all successfully created command buffer
objects from this command, set all entries of the pCommandBuffers
array to NULL and return the error."
This has been suggested by gabriel@system.is.
Signed-off-by: Samuel Pitoiset <samuel.pitoiset@gmail.com>
Reviewed-by: Bas Nieuwenhuizen <bas@basnieuwenhuizen.nl>
It's really annoying and this pollutes the output especially
when a bunch of non-meta shaders are compiled.
Signed-off-by: Samuel Pitoiset <samuel.pitoiset@gmail.com>
Reviewed-by: Bas Nieuwenhuizen <bas@basnieuwenhuizen.nl>
This just builds on the image support. Evergreen only has ssbo
for fragment and compute no other stages.
v2: handle images and ssbo in the same shader properly (Ilia)
v3: fix RESQ on buffers,
fix missing atom emit
fix first element offset
use R32 format
write separate buffer rat store path.
(from running deqp gles3.1 tests)
Signed-off-by: Dave Airlie <airlied@redhat.com>
On cayman it appears the cmp component is now in Z.
Fixes:
arb_shader_image_load_store-dead-fragments on cayman.
Signed-off-by: Dave Airlie <airlied@redhat.com>