v3d_compile is now split out into a helper function that gets called a
second time if compilation fails the first time with the result
reporting the register allocation failed. The second time it is run with
the fallback scheduler to try and increase the chances of successfully
allocating the registers.
v2: Add a performance debug message when using the fallback scheduler.
Reviewed-by: Iago Toral Quiroga <itoral@igalia.com>
Reviewed-by: Alejandro Piñeiro <apinheiro@igalia.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/5953>
Instead of just having a bool status for the failure, there is now an
enum so that the compilation can report a more detailed status.
Currently this is only used to report whether the failure was due to
failed register allocation. The “failed” bool doesn’t seem to actually
have been used anywhere so this doesn’t really change a lot.
Reviewed-by: Iago Toral Quiroga <itoral@igalia.com>
Reviewed-by: Alejandro Piñeiro <apinheiro@igalia.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/5953>
The current scheduling algorithm favors parallelism a bit too
aggressively and sometimes generates shaders that fail register
allocation. This happens even if the threshold is set to zero to force
it to always use the CSR instruction choosing algorithm.
This patch adds an option to use an even more aggressive fallback that
just always picks the instruction with the shortest maximum delay in the
hope that that will generate the least register pressure. The intention
is to use this as a last resort after register allocation fails in order
to at least have a working shader.
Reviewed-by: Alejandro Piñeiro <apinheiro@igalia.com>
Acked-by: Iago Toral Quiroga <itoral@igalia.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/5953>
The input primitive ID is read from the VPM in the same memory segment
as the outputs. This means that writing the GS header to VPM location 0
needs to be done after reading the primitive ID. This patch adds a
dependency between the load_primitive_id intrinsic and the store_output
intrinsic for location 0 to stop the scheduler from reordering them.
v2: Use an enum for the dependency class number.
v3: Add "GS" to the class enum name.
Reviewed-by: Iago Toral Quiroga <itoral@igalia.com>
Reviewed-by: Alejandro Piñeiro <apinheiro@igalia.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/5953>
Adds a callback function to nir_schedule_options to give the backend a
chance to add custom dependencies between certain intrinsics. The
callback can assign a class number to the intrinsic and then set a read
or write dependency on that class.
v2: Use a linked-list of schedule nodes for the dependency classes
instead of a fixed-sized array.
Reviewed-by: Iago Toral Quiroga <itoral@igalia.com>
Reviewed-by: Alejandro Piñeiro <apinheiro@igalia.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/5953>
Instead of copying the individual members of nir_schedule_options into
the scoreboard, it now just keeps a pointer to the options. This avoids
the duplicated comments and makes it easier to add more options later.
Reviewed-by: Iago Toral Quiroga <itoral@igalia.com>
Reviewed-by: Alejandro Piñeiro <apinheiro@igalia.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/5953>
nir_schedule already has a struct for options which makes it more than
just a function declaration. Later patches intend to add more structs to
complement these options. In order to make the code easier to manage,
this moves the nir_scheduler-related parts out of nir.h to their own
header.
Reviewed-by: Iago Toral Quiroga <itoral@igalia.com>
Reviewed-by: Alejandro Piñeiro <apinheiro@igalia.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/5953>
Previously, objects of type OpTypeImage or OpTypeSampler were treated as
vtn_pointers and objects of type OpTypeSampledImage were a special-use
vtn_sampled_image struct. This commit changes that so that all of those
objects are stored in vtn_ssa_values. Each of images, samplers, and
sampled images, are stored as a scalar or vector nir_ssa_def whose
components are NIR deref values. We now use vtn_type_get_nir_type to
re-resolve those as-needed into GLSL sampler types for NIR.
This simplification has a number of benefits:
1. We can git rid of the rest of our special-cases for handling images
and samplers in function arguments. Now that they're treated as
structs at the glsl_type level, the generic paths can handle images
and samplers.
2. We can now construct composite values containing images and samplers
internally. It's unclear from the SPIR-V spec whether or not this
is allowed and it's not a pattern that GLSLang currently generates
thanks to GLSL rules. However, if we do start seeing SPIR-V that
contains such composites, we should now be able to handle it.
3. SPIR-V OpNull and OpUndef instructions can now create samplers,
images, and sampled images. The NIR generated won't likely be fully
valid but, given a NIR pass to do something sensible, it should be a
thing we can compile.
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/5278>
We're about to make the SPIR-V -> NIR path generate a bit more complex
SSA chains for certain derefs. This will ensure we don't regress anyone
when we start making vec2's of derefs.
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/5278>
There are a few cases, atomic counters being one example, where the type
used by vtn_ssa_value is not the same as the type we want NIR to use in
derefs and variables. To solve this, we add a helper which converts
between the types for us. In the next commit, we'll be adding another
major user of this: images and samplers.
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/5278>
Primarily, we check for two things:
1. That we only ever add SSA values via vtn_push_ssa_value and
vtn_copy_value.
2. That the type of the SSA value matches the SPIR-V destination type.
Reviewed-by: Caio Marcelo de Oliveira Filho <caio.oliveira@intel.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/5278>
Previously, we created our vtn_ssa_value in _vtn_variable_load_store
manually as we did the recursive load/store. Instead, we now create the
SSA value before calling into the recursive function. This is a tiny
bit less efficient but it removes a case of hand-rolling vtn_ssa_value
creation. For symmetry, we make _vtn_block_load_store assume the value
is already created. Finally, we remove a trivial hand-rolled case in
vtn_composite_extract.
Reviewed-by: Caio Marcelo de Oliveira Filho <caio.oliveira@intel.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/5278>
For three different functions which create vtn_ssa_values, we had three
completely different implementations. This unifies them all to roughly
the same algorithm. While we're at it, we take advantage of the
nir_build_imm helper to avoid some extra code in vtn_const_ssa_value.
Reviewed-by: Caio Marcelo de Oliveira Filho <caio.oliveira@intel.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/5278>
The w++ is to handle a differences between the KHR extension and Vulkan
1.1 feature where the Vulkan 1.1 instructions take an scope parameter.
While incrementing w technically works, it's really subtle and very easy
to miss when reading the code.
Reviewed-by: Caio Marcelo de Oliveira Filho <caio.oliveira@intel.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/5278>
For some reason, we were doing a signed shift vectors and an unsigned
shift for scalars. We then plug it into i2b so it should make no
difference whatsoever. The fact that we're doing different things for
vectors vs. scalars is bonkers. Let's simplify the code a bit.
Reviewed-by: Caio Marcelo de Oliveira Filho <caio.oliveira@intel.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/5278>
The original implementation of SPV_EXT_descriptor_indexing was extremely
paranoid about the NonUniform qualifier, trying to fetch it from every
possible location and propagate it through access chains etc. However,
the Vulkan spec is quite nice to us on this and has very strict rules
for where the NonUniform decoration has to be placed. For image and
texture operations, we can search for the decoration on the spot when we
process the image or texture op. For pointers, we continue putting it
on the pointer but we don't bother trying to do anything silly like
propagate it through casts.
Reviewed-by: Caio Marcelo de Oliveira Filho <caio.oliveira@intel.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/5278>
For cross-process timelines we have to have a thread to wait
till the requested points become available.
The functions actually dealing with timeline semaphores stubbed out, to
implement in the next patch. As such the thread code shouldn't trigger
yet.
The core idea is that we still use the refcount mechanism that we use with
emulated timelines, though the native timeline syncobj don't participate
in the refcounting. This way we keep the ordering of submission in a queue
as each submission is also blocked by its predecessor.
Where we change behavior is when the number of blockers reaches 0. In the
new code we check if we need to wait for the timeline semaphores to
be available and if so we won't execute the submission immediately but
pass it to the submission thread.
Reviewed-by: Samuel Pitoiset <samuel.pitoiset@gmail.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/5600>
To facilitate cross-process timeline semaphores we have to deal with
the fact that the syncobj signal operation might be submitted a
small finite time after the wait operation.
For that we start using DRM_SYNCOBJ_WAIT_FLAGS_WAIT_FOR_SUBMIT during
the wait operation so we properly wait instead of returning an error.
Furthermore, to make this effective for syncobjs that get reused we
actually have to reset them after the wait. Otherwise the wait before
submit would get the previous fence instead of waiting for the new
thing to submit.
The obvious choice, resetting the syncobj after the CS submission
has 2 issues though:
1) If the same semaphore is used for wait and signal we can't reset it.
This is solvable by only resetting the semaphores that are not in the
signal list.
2) The submitted work might be complete before we get to resetting the
syncobj. If there is a cycle of submissions that signals it again and
finishes before we get to the reset we are screwed.
Solution:
Copy the fence into a new syncobj and reset the old syncobj before
submission. Yes I know it is more syscalls :( At least I reduced the
alloc/free overhead by keeping a cache of temporary syncobjs.
This also introduces a syncobj_reset_count as we don't want to reset
syncobjs that are part of an emulated timeline semaphore. (yes, if
the kernel supports timeline syncobjs we should use those instead,
but those still need to be implemented and if we depend on them in
this patch ordering dependencies get hard ...)
Reviewed-by: Samuel Pitoiset <samuel.pitoiset@gmail.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/5600>
8e1b75b3 introduced umax/umin in order to lower iand/ior for (n)eq zero.
That breaks the lower_int_to_float pass, because umax and umin weren't
handled there.
Tested with lima. The other users of nir_lower_int_to_float
(etnaviv, freedreno) should also have that issue.
Reviewed-by: Vasily Khoruzhick <anarsoul@gmail.com>
Reviewed-by: Jason Ekstrand <jason@jlekstrand.net>
Signed-off-by: Andreas Baierl <ichgeh@imkreisrum.de>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/6043>
