It probably doesn't matter because that buffer should have a stride of
zero. However, it still seems like a good idea just to be safe.
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/5022>
It saves addressing math, but may cause multiple loads to be done and
bcseled due to NIR not giving us good address alignment information
currently. I don't have any workloads I know of using non-const-uploaded
UBOs, so I don't have perf numbers for it
This makes us match the GLES blob's behavior, and turnip (other than being
bindful).
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/4858>
With the upcoming LDC usage in the GL driver, we don't want to be
uploading descriptors for every UBO when they aren't actually in use.
Trimming NIR's num_ubos will avoid that, and cleans up num_ubo handling
elsewhere right now.
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/4858>
The vma_heap allocator was originally designed to prefer high addresses
in order to find bugs in ANV's high address handling. However, there
are cases where you might want the allocator to prefer lower addresses
for some reason. This provides a configure bit for exactly this
purpose.
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
Reviewed-by: Eric Anholt <eric@anholt.net>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/5019>
A trace of a Hat in Time, which builds thousands of shaders
takes 339 seconds to run the second time without this patch,
and 41 seconds with it (basically there is no more loading times).
Signed-off-by: Axel Davy <davyaxel0@gmail.com>
Reviewed-by: Marek Olšák <marek.olsak@amd.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/4993>
For now this parameter doesn't do anything.
It means the implementation is allowed to use
a cache on disk.
Signed-off-by: Axel Davy <davyaxel0@gmail.com>
Reviewed-by: Marek Olšák <marek.olsak@amd.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/4993>
The HW requires a log2 width/height of the level 0 meta_* size in the
descriptors, making it pretty clear that UBWC mipmapping is all
power-of-two sized. Fixes a bunch of failures in the upcoming unit UBWC
layout unit tests.
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/4931>
Using texturator on a P3A at 1024x1024, RG8 has log2w/h of 6x7 instead of
R16I/UI's 6x8. The other blockw/h I verified other than cpp=1
(R8/R8I/R8UI didn't use UBWC) and 32 (would need a bigger type).
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/4931>
This patch adds some control flow information to the
state to keep track whether a loop contains divergent
continue or break statements to not having to
recalculate this property for every phi.
Reviewed-by: Jason Ekstrand <jason@jlekstrand.net>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/4062>
This patch splits the visit_phi() function into
three different ones according to the kind of phi
(merge-node, loop-header or loop-exit) and calls
them when visiting the cf_nodes.
This allows to revisit loops if the loop header's
phis have changed, only.
Reviewed-by: Jason Ekstrand <jason@jlekstrand.net>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/4062>
Also, rewrite the format decision code so that we correctly decide when
the linear fallback is needed, even if UBWC is disabled. As part of
that, I also moved around some of the code to handle compressed formats
to make sure that copying compressed formats with a linear staging blit
works (this is now possible since we started allowing tiled compressed
textures).
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/5007>
This is simpler than a full-blown memory reuse mechanism, but is good
enough to make sure that repeatedly doing a copy that requires the
linear staging buffer workaround won't use excessive memory or be slowed
down due to repeated allocations.
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/5007>
If one VMEM instruction uses a sampler and the other doesn't, we can't do
this optimization.
Totals from 47 (0.04% of 127638) affected shaders:
CodeSize: 271744 -> 271656 (-0.03%); split: -0.04%, +0.01%
Instrs: 52783 -> 52761 (-0.04%); split: -0.05%, +0.01%
Cycles: 5547040 -> 5546952 (-0.00%); split: -0.00%, +0.00%
VMEM: 10022 -> 9887 (-1.35%)
Signed-off-by: Rhys Perry <pendingchaos02@gmail.com>
Reviewed-by: Daniel Schürmann <daniel@schuermann.dev>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/4949>
When the LLVM version is too old or missing, SotTR applies shader
workarounds and that reduces performance by 2-5% with ACO.
SotTR workarounds are applied with LLVM 8 and older, so reporting
LLVM 9.0.1 should be fine.
Signed-off-by: Samuel Pitoiset <samuel.pitoiset@gmail.com>
Reviewed-by: Edmondo Tommasina <edmondo.tommasina@gmail.com>
Reviewed-by: Bas Nieuwenhuizen <bas@basnieuwenhuizen.nl>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/4984>
ANV and RADV have similar Python code for generating extensions
and dispatch tables.
Signed-off-by: Samuel Pitoiset <samuel.pitoiset@gmail.com>
Acked-by: Lionel Landwerlin <lionel.g.landwerlin@intel.com>
Reviewed-by: Bas Nieuwenhuizen <bas@basnieuwenhuizen.nl>
Reviewed-by: Eric Engestrom <eric@engestrom.ch>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/4987>
The GC880 on iMX6DL indicates in it's minorFeatures2 register that it
does support SEAMLESS_CUBE_MAP, however when the TE.SAMPLER_CONFIG1
VIVS_TE_SAMPLER_CONFIG1_SEAMLESS_CUBE_MAP bit is set on GC880 on iMX6DL,
the result is corrupted image. In particular, the following ~112 dEQPs
are affected and fail:
dEQP-GLES2.functional.texture.filtering.cube.*
This only happens on MX6DL GC880, MX6Q GC2000 and STM32MP1 GC400(GCnano)
do not report the minorFeatures2 SEAMLESS_CUBE_MAP bit and ignore the
TE_SAMPLER_CONFIG1 VIVS_TE_SAMPLER_CONFIG1_SEAMLESS_CUBE_MAP bit (note
that ss->seamless_cube_map is unconditionally set by mesa at times even
PIPE_CAP_SEAMLESS_CUBE_MAP_PER_TEXTURE returns 0), so there is no visible
problem and there are no failing dEQP tests on the GC2000 and GCnano.
This might imply that the minorFeatures2 SEAMLESS_CUBE_MAP has some
different meaning on GC880 or the SEAMLESS_CUBE_MAP behaves differently
on the GC880.
This patch does not set the SEAMLESS_CUBE_MAP bit on hardware which does
not indicate support for seamless cube map and on GC880, which results
in reduction in failed dEQPs: 635 to 186 on GC880, 274 to 270 on GC2000
and no change on GC400(GCnano).
Fixes: 8dd26fa2f0 ("etnaviv: support GL_ARB_seamless_cubemap_per_texture")
Reviewed-by: Christian Gmeiner <christian.gmeiner@gmail.com>
Signed-off-by: Marek Vasut <marex@denx.de>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/4865>
On a6xx we need a 0,0 based scissor in the binning pass, but can use the
blit-scissor to avoid restore/resolve of untouched pixels, and use the
conditional execution if the IB to bin to skip bins with no geometry
(due to the scissor).
Signed-off-by: Rob Clark <robdclark@chromium.org>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/5021>
Similar to what we do in postsched. It is useful for pre-RA sched to be
a bit aware of things that would cause syncs. In particular for the tex
fetches, since the vecN src/dst tends to limit postsched's ability to
re-order them.
Signed-off-by: Rob Clark <robdclark@chromium.org>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/4923>
If an instruction's only use is as an output, and it increases register
pressure, then try to avoid scheduling it until there are no other
options.
A semi-common pattern is `fragcolN.a = 1.0`, this pushes all these
immed loads to the end of the shader.
Signed-off-by: Rob Clark <robdclark@chromium.org>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/4923>
Similar to avoidance of `(ss)` syncs, it turns out to be helpful to
avoid `(sy)` syncs as well. This helps us turn an tex, (sy)alu, tex,
(sy)alu sequence into tex, tex, (sy)alu, alu, which is a big win in
gfxbench gl_fill2.
Signed-off-by: Rob Clark <robdclark@chromium.org>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/4923>
Once we schedule an instruction that will require an `(ss)` sync flag,
there is no need to delay any further instructions that consume an
SFU result (until the next SFU instruction is scheduled).
Signed-off-by: Rob Clark <robdclark@chromium.org>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/4923>
It seems for short frag shaders, too much prefetch can be detrimental.
I think what we *really* want to do is decide after pre-RA sched, when
we also know about nop's and what the actual ir3 instruction count is.
But that will require re-working how prefetch lowering works. For now
this is a super crude heuristic to attempt to approximate a good
solution.
Signed-off-by: Rob Clark <robdclark@chromium.org>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/4923>
We can no longer assume that `state->ranges[0]` is block 0. It *often*
is, but when we encounter a "real" ubo that we lower to `load_uniform`
before a block 0 `load_ubo`, it could end up another entry in the table.
Resulting in the second pass after gathering ubo ranges, not finding a
valid range. Which results in a `load_ubo` for a thing that is not
actually a ubo making it's way into ir3 frontend. Resulting in grabbing
what we think is a ubo address out of some unrelated const register, and
trying to dereference that. Which as you can imagine, fails in amusing
ways.
Fixes: fc850080ee ("ir3: Rewrite UBO push analysis to support bindless")
Signed-off-by: Rob Clark <robdclark@chromium.org>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/4954>
Corresponds roughly to what we analyze. Note that "terminate AND
execute" is a contradiction (rather: it's equivalent to just
terminating), hence why there are only three possibilities for the
states of the flags:
.cont = continue, don't execute
.last = don't continue, don't execute
.cont.last = continue and execute
Signed-off-by: Alyssa Rosenzweig <alyssa.rosenzweig@collabora.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/5014>
This is for an optimization I plan in a following commit. I found I had
to add likely()s to avoid a perf regression from branch prediction.
On the drawoverhead 8 UBOs test, the HW can't quite keep up with the CPU,
but if I set nohw then this change is 1.32023% +/- 0.373053% (n=10).
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/4996>
This commit fixes a GS regression introduced in !4562 where
ir3's GS lowering pass was moved from common code (ir3_nir) to
freedreno-specific code (ir3_shader). For GS support in turnip, we
need to add the GS lowering pass back in, this time in tu_shader.
As for the nir_gather_info change, the GS lowering pass has always
introduced a discard_if intrinsic into the GS. Previously, we simply
ran nir_shader_gather_info before GS lowering, but now since we lower
the GS before we need to remove the assertion that only a FS can use
the discard_if intrinsic.
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/4892>
And try a bit harder to find an optimal layout. Improves on a sub-
optimal layout we arrive at in the 4 MRT pass in manhattan, picking
up a bit more than 3%.
Signed-off-by: Rob Clark <robdclark@chromium.org>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/4976>
We potentially don't know yet what the resulting scissor bounds are, so
we can't assume this when estimating number of bins per pipe for VSC
size calculations.
Signed-off-by: Rob Clark <robdclark@chromium.org>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/4976>
There is an annoying spec corner on Android. Because VkSwapchain is
implemented in the Vulkan loader on Android which may not know about
this extension, we have to handle it as a special case inside the
driver. We only have to do this on Android and only for VkSwapchainKHR.
Acked-by: Samuel Pitoiset <samuel.pitoiset@gmail.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/4882>
It seems this is actually a "minimum pitch" value. For example
TFETCH6_2_BYTE means a minimum pitch of 128 bytes for mipmap levels.
This fixes breakage with compressed formats. For example this test:
dEQP-VK.pipeline.sampler.view_type.2d.format.eac_r11_snorm_block.mipmap.linear.lod.equal_min_3_max_3
Fixes: a34b3fa198 ("freedreno/fdl: Align after dividing by block size")
Signed-off-by: Jonathan Marek <jonathan@marek.ca>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/5009>
The nir_analyze_ubo_ranges pass removes all UBO block 0 loads to reverse
what nir_lower_uniforms_to_ubo() had done, and we only upload UBO pointers
to the HW for UBO block 1-N, so let's just fix up the shader state.
Fixes an off by one in const state layout setup, and some really dodgy
register addressing trying to deal with dynamic UBO indices when the UBO
pointers happen to be at the start of the constbuf.
There's no fixes tag, though this fixes a bug from September, because it
would require the num_ubos fix in nir_lower_uniforms_to_ubo.
Reviewed-by: Kristian H. Kristensen <hoegsberg@google.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/4992>
The fixed commit was really nice in mostly fixing num_ubos to reflect the
shader after lowering, but for
dEQP-GLES31.functional.compute.basic.ubo_to_ssbo_single_invocation there
are no default uniforms and so we skipped the increment, even though we
shifted the block index up.
Fixes: 4777ee1a62 ("nir: Always create UBO variable when lowering uniforms to ubo")
Reviewed-by: Kristian H. Kristensen <hoegsberg@google.com>
Reviewed-by: Erik Faye-Lund <erik.faye-lund@collabora.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/4992>
Using non-write flags is pretty dubious -- it means the kernel tracking an
array of read-only consumers of the BO and having exclusive consumers wait
on each reader's fence. It allows multiple readers through dma-bufs to do
work in parallel, but at the cost of kernel CPU time and memory management
of the shared array. Other drivers have dropped this distinction since
dma-buf sharing is usually producer-consumer, not producer-two-consumers,
and the userspace and kernel space tracking is expensive.
For us, this lets us drop the flags passed in for relocs and tracked in
the ringbuffer reloc lists. The end result of the flags reduction work is
drawoverhead uniforms test throughput 2.37195% +/- 0.365579% (n=15)
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/4967>
It's silly to have all the reloc emitters passing around FD_RELOC_READ
when you have to have it set on all relocs (that don't include WRITE,
which implies read) for the kernel to actually track the fences on the BO.
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/4967>