Rework:
* Jordan: Handle prog_data->total_scratch==0 in iris_upload_compute_walker
* Jordan: Resolve iris_get_scratch_space conflict with e2c5ef6cd6
* Jordan: Rebase on 4256f7ed58. broken
* Ken: Mostly fixed the rebase
* Jordan: Fix two small compilation issues
* Jordan: Rebase on Ken's ("iris: Make a pin_scratch_space() helper")
* Lionel: Fix a few bugs with scratch handles
* Jason: Tidy the patch up a bit
Reviewed-by: Jason Ekstrand <jason@jlekstrand.net>
Reviewed-by: Jordan Justen <jordan.l.justen@intel.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/11582>
Based on a patch by Rafael Antognolli.
We already had a flags parameter, but omitted it from the simple alloc
interface because most callers were passing 0. However, we'll want to
use it for selecting between device local memory and system memory, and
possibly mmap cacheability modes, in the future. At that point, many
more callers will want to specify, so I think we should include flags
in iris_bo_alloc() as well.
A few places used the iris_bo_alloc_tiled() function simply to pass
flags, so this patch converts them to use iris_bo_alloc() instead now
it does everything they want.
Reviewed-by: Tapani Pälli <tapani.palli@intel.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/11169>
Makes calling code more explicit about what is being set, and allows
take advantage of zero initialization for the ones the callsite don't
care.
Besides moving to the struct, two extra "ergonomic" changes were done:
- Add a new shader_time boolean, so shader_time_index is ignored when
unused -- this allow taking advantage of the zero initialization of
unset fields.
- Since we have a struct, provide space for the error_str pointer.
Both iris and i965 were using it, and the extra rstrdup in case of
failure shouldn't be a burden for the others.
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
Reviewed-by: Jordan Justen <jordan.l.justen@intel.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/9779>
Makes calling code more explicit about what is being set, and allows
take advantage of zero initialization for the ones the callsite don't
care.
Besides moving to the struct, two extra "ergonomic" changes were done:
- Add a new shader_time boolean, so shader_time_index is ignored when
unused -- this allow taking advantage of the zero initialization of
unset fields.
- Since we have a struct, provide space for the error_str pointer.
Both iris and i965 were using it, and the extra rstrdup in case of
failure shouldn't be a burden for the others.
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
Reviewed-by: Jordan Justen <jordan.l.justen@intel.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/9779>
When we enable u_threaded_context, the pipe->create_*_state hooks
(precompile variants) are going to be called from one thread, while
iris_update_compiled_shaders (on-the-fly variants) are going to be
called from a driver thread. BLORP shaders also happen from
clear, blit, and so on in the driver thread.
u_upload_mgr isn't thread-safe, so use an uploader for each purpose.
Reviewed-by: Zoltán Böszörményi <zboszor@gmail.com>
Reviewed-by: Ian Romanick <ian.d.romanick@intel.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/8964>
We call update_last_vue_map after updating the shaders, which compares
the new and old VUE maps. Except...updating the shaders may have
dropped the last reference to the variant that ice->shaders.last_vue_map
belonged to, leading to a classic use-after-free.
Fix this by taking a reference to the variant for the last VUE stage,
so it stays around until we're done with it.
Fixes: 1afed51445 ("iris: Store a list of shader variants in the shader itself")
Closes: https://gitlab.freedesktop.org/mesa/mesa/-/issues/4311
Reviewed-by: Lionel Landwerlin <lionel.g.landwerlin@intel.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/9143>
We've traditionally stored shader variants in a per-context hash table,
based on a key with many per-stage fields. On older hardware supported
by i965, there were potentially quite a few variants, as many features
had to be emulated in shaders, including things like texture swizzling.
However, on the modern hardware targeted by iris, our NOS dependencies
are much smaller. We almost always guess the correct state when doing
the initial precompile, and so we have maybe 1-3 variants. iris NOS
keys are also dramatically smaller (4 to 24 bytes) than i965's.
Unlike the classic world, Gallium also provides a single kind of object
for API shaders---pipe_shader_state aka iris_uncompiled_shader. We can
simply store a list of shader variants there. This makes it possible
to access shader variants across contexts, rather than compiling them
separately for each context, which better matches how the APIs work.
To look up variants, we simply walk the list and memcmp the keys.
Since the list is almost always singular (and rarely ever long),
and the keys are tiny, this should be quite low overhead.
We continue storing internally generated shaders for BLORP and
passthrough TCS in the per-context hash table, as they don't have
an associated pipe_shader_state / iris_uncompiled_shader object.
(There can also be many BLORP shaders, and the blit keys are large,
so having a hash table rather than a list makes sense there.)
Because iris_uncompiled_shaders are shared across multiple contexts,
we do require locking when accessing this list. Fortunately, this
is a per-shader lock, rather than a global one. Additionally, since
we only append variants to the list, and generate the first one at
precompile time (while only one context has the uncompiled shader),
we can assume that it is safe to access that first entry without
locking the list. This means that we only have to lock when we
have multiple variants, which is relatively uncommon.
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/7668>
There is a small gap of time where the currently bound uncompiled
shaders, and compiled shader variant, are out of sync. Specifically,
between pipe->bind_*_state() and the next draw.
Currently, shaders variants live entirely within a single context,
and when deleting an iris_uncompiled_shader, we check if any of its
variants are currently bound, and defer deleting those until the next
iris_update_compiled_shaders() hook runs and binds new shaders to
replace them. (This is due to the time gap between binding new
uncompiled shaders, and updating variants at draw time when we have
the required NOS in place.)
This works pretty well in a single context world. But as we move to
share compiled shader variants across multiple contexts, it breaks down.
When deleting a shader, we can't look at all contexts to see if its
variants are bound anywhere. We can't even quantify whether those
contexts will run a future draw any time soon, to update and unbind.
One fairly crazy solution would be to delete the variants anyway, and
leave the stale pointers to dead variants in place. This requires
removing any code that compares old and new variants. Today, we do
that sometimes for seeing if the old/new shaders toggled some feature.
Worse than that, though, we don't just have to avoid dereferences, we'd
have to avoid pointer comparisons. If we free a variant, and quickly
allocate a new variant, malloc may return the same pointer. If it's
for the same shader stage, we may get a new different program that has
the same pointer as a previously bound stale one, causing us to think
nothing had changed when we really needed to do updates. Again, this
is doable, but leaves the code fragile - we'd have to guard against
future patches adding such checks back in.
So, don't do that. Instead, do basic reference counting. When a
variant is bound in a context, up the reference. When it's unbound,
decrement it. When it hits zero, we know it's not bound anywhere and
is safe to delete, with no stale references. This ends up being
reasonably cheap anyway, since the atomic is usually uncontested.
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/7668>
Now that we're looking at shader info system values rather than
vs_prog_data, there's no reason we have to do this when updating
the shader variants. We can simply check it when binding a new
shader from the API point of view.
Reviewed-by: Anuj Phogat <anuj.phogat@gmail.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/8759>
Instead of looping over all stages and re-accessing prog_data at the
end, we can just move the URB size check into iris_update_compiled_xs,
in the new != old block, where we know the shader changed - plus where
we're disabling tessellation. We already have the prog_data handy in
these cases, with a bit less pointer chasing.
Improves performance in Piglit's drawoverhead microbenchmark
(#63: DrawArrays, no state change) by 4.07856% +/- 0.540517% (n=850).
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/8721>
Reconfiguring the URB partitioning is likely to cause shader stalls,
as the dividing line between each stage's section of memory is moving.
(Technically, 3DSTATE_URB_* are pipelined commands, but that mostly
means that the command streamer doesn't need to stall.) So it should
be beneficial to update the URB configuration less often.
If the previous URB configuration already has enough space for our
current shader's needs, we can just continue using it, assuming we
are able to allocate the maximum number of URB entries per stage.
However, if we ran out of URB space and had to limit the number of
URB entrties for a stage, and the per-entry size is larger than we
need, we should reconfigure it to try and improve concurrency.
So, we begin tracking the last URB configuration in the context,
and compare against that when updating shader variants.
Cuts 36% of the URB reconfigurations (excluding BLORP) from a
Shadow of Mordor trace, and 46% from a GFXBench Manhattan 3.0 trace.
One nice thing is that this removes the need to look at the old
prog_data when updating shaders, which should make it possible to
unbind shader variants without causing spurious URB updates.
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/8721>
According to the BSpec page for MEDIA_VFE_STATE, on Gen12 platforms
"if a fused configuration has fewer threads than the native POR
configuration, the scratch space allocation is based on the number of
threads in the base native POR configuration". However we currently
use the subslice count from devinfo->num_subslices[0], which only
includes the subslices currently enabled by the platform fusing. This
leads to scratch space underallocation and occasional hangs.
The problem is likely to affect most Gen12 GPUs with less than 96 EUs.
GFXBench5 Aztec Ruins is able to reproduce the issue fairly reliably.
Fixes: 9e5ce30da7 "intel: fix the gen 12 compute shader scratch IDs"
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/8636>
On Gen12+, we can enable additional caches in certain usage situations.
This routes that decision making to a central place in ISL, based on
surface usage flags, and updates both drivers to use it. (i965 doesn't
need to change because it doesn't support Gen12.)
We continue handling the "external" decision via an anv_mocs() wrapper
for now, since we store that flag in anv_bo, which isl doesn't know
about. (We could introduce an ISL_SURF_USAGE_EXTERNAL, but I'm not
actually sure that would be cleaner.)
This patch should not have any functional nor performance effects, as
we continue selecting the exact same MOCS values for now.
Reviewed-by: Jason Ekstrand <jason@jlekstrand.net>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/7104>
