When using multiple binaries, we don't know the required number of VGPRs beforehand,
which means we either have to over-allocate VGPRs or avoid shared VGPRs.
As bpermute is the only instructions needing shared VGPRs, we decide for the latter.
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/22267>
This was the last missing feature for GPL. The main problem is that
the on-disk shaders cache size will increase a lot because we don't
deduplicate shaders but there is on-going work to improve that.
We also can't use the shaders cache for libraries created with the
RETAIN_LINK_TIME_OPTIMIZATION flag and module identifiers because we
don't know the SPIR-V and thus can't retain NIR shaders for linking.
Signed-off-by: Samuel Pitoiset <samuel.pitoiset@gmail.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/22264>
Even if we are able to get the assembly from the shaders cache for
graphics pipeline libraries, we still need to retain NIR shaders in
case the LTO pipelines won't be find in the cache.
Signed-off-by: Samuel Pitoiset <samuel.pitoiset@gmail.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/22264>
This is to generate a different key for a library created with
FRAGMENT_SHADER_BIT and no FS (ie. it would generate a noop FS) and
a library created with FRAGMENT_OUTPUT_INTERFACE with no CB attachments.
Otherwise, the same key would be generated and this would corrupt
the cache.
Signed-off-by: Samuel Pitoiset <samuel.pitoiset@gmail.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/22264>
With GPL, a stage can be imported from a library which means that the
binary is NULL (it's freed right after compilation) but the shader is
non-NULL. To avoid crashing, rely on non-NULL binaries because this
implies that the shader is non-NULL as well.
Signed-off-by: Samuel Pitoiset <samuel.pitoiset@gmail.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/22264>
With GPL, we don't necessarily load all stages from the same cache
entry and other stages might have been imported from libraries, so
don't overwrite them.
Signed-off-by: Samuel Pitoiset <samuel.pitoiset@gmail.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/22264>
This helps the CP (command processor) optimize mesh shader
dispatch when Y = Z = 1.
Reviewed-by: Rhys Perry <pendingchaos02@gmail.com>
Reviewed-by: Samuel Pitoiset <samuel.pitoiset@gmail.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/22222>
This just makes it possible to use the dominator
tree information during phi lowering.
No Fossil DB changes on GFX11.
Signed-off-by: Timur Kristóf <timur.kristof@gmail.com>
Reviewed-by: Daniel Schürmann <daniel@schuermann.dev>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/21493>
The branch instruction is no longer conditional when the targets are the
same, so the operand is not necessary and can be removed.
Reviewed-by: Daniel Schürmann <daniel@schuermann.dev>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/21493>
Verifying that the branch instruction reads exec is not actually
necessary because the pattern that we look for already implies that.
This prepares for the next commit which will remove the exec operand
from branches that have the same target. These branches will no
longer read exec, but they should still get the same optimization.
Reviewed-by: Daniel Schürmann <daniel@schuermann.dev>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/21493>
Don't eliminate an instruction that writes registers other than exec and scc.
It is possible that this is eg. an s_and_saveexec and the saved value is
used by a later branch.
Fixes: bc13049747
Reviewed-by: Daniel Schürmann <daniel@schuermann.dev>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/21493>
While analyzing cache loading performance, hashing the pipeline layout was
surprisingly consuming around 4% of time, sometimes close to the cost of
hashing shader modules.
Turns out we were hashing the pipeline layout on every pipeline creation.
Considering that pipeline layouts are usually deduplicated by the
application, this was amplifying the hashing cost by a big margin.
With Graphics Pipeline Library, we do need to rebuild the pipeline layout
by combining those from each library, but we can memoize the hash of the
descriptor set layout. The cost of re-hashing hashes is negligible since
each descriptor set layout can amount to 1–2KB in size.
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/22254>
It's currently used when LS store output to LDS.
The LS/HS bug fix seems does not affect this case.
But we'd better treat it as other fixed args.
Reviewed-by: Marek Olšák <marek.olsak@amd.com>
Signed-off-by: Qiang Yu <yuq825@gmail.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/22045>
AMD recommends doing this to speed up the CP when it processes
the draw ring entries. LLPC also does this.
Signed-off-by: Timur Kristóf <timur.kristof@gmail.com>
Reviewed-by: Rhys Perry <pendingchaos02@gmail.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/22211>
When writing the draw ready bit, don't write the high 24 bits
of DWORD3, because that is used by the HW for something else
according to LLPC.
Cc: mesa-stable
Signed-off-by: Timur Kristóf <timur.kristof@gmail.com>
Reviewed-by: Rhys Perry <pendingchaos02@gmail.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/22211>
If a compute pipeline is bound after a raytracing pipeline, the
computes shader slot (aka RT prolog) will be overwritten.
To fix this, move the RT prolog outside of the compute shader slot.
Fixes: d109362a3d ("radv: copy bound shaders to the cmdbuf state")
Signed-off-by: Samuel Pitoiset <samuel.pitoiset@gmail.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/22235>
Inspired by Nicolai Hähnle's commit in LLPC.
Instead of using a SALU instruction to add to the scalar
offset, rely on the buffer swizzling and use constant offset.
Fossil DB stats on GFX1100:
Totals from 47910 (35.51% of 134913) affected shaders:
CodeSize: 87927612 -> 86968136 (-1.09%)
Instrs: 17584007 -> 17440094 (-0.82%)
Latency: 97232173 -> 97126311 (-0.11%)
InvThroughput: 9904586 -> 9905288 (+0.01%); split: -0.02%, +0.02%
VClause: 544430 -> 542566 (-0.34%)
Signed-off-by: Timur Kristóf <timur.kristof@gmail.com>
Reviewed-by: Marek Olšák <marek.olsak@amd.com>
Reviewed-by: Rhys Perry <pendingchaos02@gmail.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/22227>
Sample shading can be enabled from the pipeline and from the fragment
shader, in that case we need to re-emit some states.
Signed-off-by: Samuel Pitoiset <samuel.pitoiset@gmail.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/22215>
Renames radv_declare_shader_args to declare_shader_args and runs it
twice to first gather the user SGPR count without push constants and
descriptor sets.
Reviewed-by: Samuel Pitoiset <samuel.pitoiset@gmail.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/22119>
Push constants are handled per bind point internally. Using a separate
structure in the cmdbuf state would allow us to update it easily
without relying on bound pipelines.
Signed-off-by: Samuel Pitoiset <samuel.pitoiset@gmail.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/22209>
