nir_opt_large_constants balks at seeing a store_deref of a variable
where the source is a vecN operation of multiple load_consts, and thinks
that isn't a constant, so it should not bother promoting it.
Unfortunately, we were running nir_lower_load_const_to_scalar before
nir_opt_large_constants, so this prevented a ton of constant promotion.
This commit /used to help/ some shaders in shader-db. Presumably since
!16770 landed, those shaders were already helped. Currently ther are
no shader-db changes on any Intel platform.
Fossil-db results:
All Intel platforms had similar results. (Ice Lake shown)
Instructions in all programs: 141998227 -> 141421756 (-0.4%)
Instructions helped: 12515
Instructions hurt: 237
SENDs in all programs: 7437925 -> 7468033 (+0.4%)
SENDs hurt: 12806
Cycles in all programs: 9161655753 -> 9132869800 (-0.3%)
Cycles helped: 10163
Cycles hurt: 2637
Spills in all programs: 19977 -> 18678 (-6.5%)
Spills helped: 384
Spills hurt: 40
Fills in all programs: 32863 -> 31396 (-4.5%)
Fills helped: 385
Fills hurt: 42
Lost: 1
Lots of Shadow of the Tomb Raider fragment shaders and Batman Arkham
Origins vertex shaders were hurt for SENDs in this commit. A couple
Aztec Ruins compute shaders and Spaceship shaders (multiple stages)
were also hurt.
All of the shaders hurt for spills or fills were Spaceship compute
shaders. Nearly all of the shaders helped were Shadow of the Tomb
Raider fragmenet shaders. One Spaceship shader was reall, REALLY helped:
Spills helped fossils/fossil-db/Spaceship.run.9f90a2a226fcc57f.1.foz/0b507d3abe2e3c28/compute: 321 -> 13 (-96.0%)
Fills helped fossils/fossil-db/Spaceship.run.9f90a2a226fcc57f.1.foz/0b507d3abe2e3c28/compute: 279 -> 21 (-92.5%)
Overall this seems like an improvement, but we may want to actually
run these few benchmarks before landing.
Reviewed-by: Emma Anholt <emma@anholt.net>
Reviewed-by: Ian Romanick <ian.d.romanick@intel.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/16539>
NIR has two implementations of lower_idiv, keyed on the
imprecise_32bit_lowering flag. This flag is misleading: the results when
setting this flag "imprecise", they're completely wrong for some values.
If a backend has a native implementation of umul_high, the correct path
isn't that much more expensive. If it doesn't, it's substantially slower
for highp integer divison... but in practice, non-constant highp integer
division is pretty rare.
After a painful migration of the tree, this code path has no more users.
Remove it so nobody else gets the bright idea of using it again.
Closes: #6555
Signed-off-by: Alyssa Rosenzweig <alyssa@collabora.com>
Reviewed-by: Emma Anholt <emma@anholt.net>
Reviewed-by: Christian Gmeiner <christian.gmeiner@gmail.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/19303>
Setting the NIR options takes care of iris thanks to the common st/mesa
linking code, and updating brw_nir_link_shaders should handle anv.
The main effort here is updating remap_tess_levels, which needs to
handle vector stores, writemasking, and swizzling. Unfortunately,
we also need to continue handling the existing single-component
access because it's used for TES inputs, which we don't vectorize.
We could try to vectorize TES inputs too, but they're all pushed
anyway, so it wouldn't buy us much other than deleting this code.
Also, we do have opt_combine_stores, but not one for loads.
One limitation of using nir_vectorize_tess_levels is that it works
on variables, and so isn't able to combine outer/inner writes that
happen to live in the same vec4 slot (for triangle domains). That
said, it's still better than before.
For writes, we allow the intrinsics to supply up to the full size
of the variable (vec4 for outer, vec2 for inner) even if the domain
only requires a subset of those components (i.e. triangles needs 3).
shader-db results on Icelake:
total instructions in shared programs: 19600314 -> 19597528 (-0.01%)
instructions in affected programs: 65338 -> 62552 (-4.26%)
helped: 271 / HURT: 0
helped stats (abs) min: 6 max: 24 x̄: 10.28 x̃: 12
helped stats (rel) min: 1.30% max: 18.18% x̄: 5.80% x̃: 7.59%
95% mean confidence interval for instructions value: -10.71 -9.85
95% mean confidence interval for instructions %-change: -6.17% -5.43%
Instructions are helped.
total cycles in shared programs: 851842332 -> 851808165 (<.01%)
cycles in affected programs: 618577 -> 584410 (-5.52%)
helped: 271 / HURT: 0
helped stats (abs) min: 64 max: 540 x̄: 126.08 x̃: 111
helped stats (rel) min: 2.57% max: 37.97% x̄: 6.12% x̃: 5.06%
95% mean confidence interval for cycles value: -135.35 -116.80
95% mean confidence interval for cycles %-change: -6.67% -5.57%
Cycles are helped.
total sends in shared programs: 1025238 -> 1024308 (-0.09%)
sends in affected programs: 6454 -> 5524 (-14.41%)
helped: 271 / HURT: 0
helped stats (abs) min: 2 max: 8 x̄: 3.43 x̃: 4
helped stats (rel) min: 5.71% max: 25.00% x̄: 14.98% x̃: 17.39%
95% mean confidence interval for sends value: -3.57 -3.29
95% mean confidence interval for sends %-change: -15.42% -14.54%
Sends are helped.
According to Felix DeGrood, this results in a 10% improvement in
the draw call time for certain draw calls from Strange Brigade.
v2: Fix assertions about number of components and add more of them.
Combine the quads and triangles handling as it's nearly identical.
Reviewed-by: Caio Oliveira <caio.oliveira@intel.com> [v1]
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/19061>
Setting the NIR options takes care of iris thanks to the common st/mesa
linking code, and updating brw_nir_link_shaders should handle anv.
The main effort here is updating remap_tess_levels, which needs to
handle vector stores, writemasking, and swizzling. Unfortunately,
we also need to continue handling the existing single-component
access because it's used for TES inputs, which we don't vectorize.
We could try to vectorize TES inputs too, but they're all pushed
anyway, so it wouldn't buy us much other than deleting this code.
Also, we do have opt_combine_stores, but not one for loads.
One limitation of using nir_vectorize_tess_levels is that it works
on variables, and so isn't able to combine outer/inner writes that
happen to live in the same vec4 slot (for triangle domains). That
said, it's still better than before.
For writes, we allow the intrinsics to supply up to the full size
of the variable (vec4 for outer, vec2 for inner) even if the domain
only requires a subset of those components (i.e. triangles needs 3).
shader-db results on Icelake:
total instructions in shared programs: 19605070 -> 19602284 (-0.01%)
instructions in affected programs: 65338 -> 62552 (-4.26%)
helped: 271 / HURT: 0
helped stats (abs) min: 6 max: 24 x̄: 10.28 x̃: 12
helped stats (rel) min: 1.30% max: 18.18% x̄: 5.80% x̃: 7.59%
95% mean confidence interval for instructions value: -10.71 -9.85
95% mean confidence interval for instructions %-change: -6.17% -5.43%
Instructions are helped.
total cycles in shared programs: 851854659 -> 851820320 (<.01%)
cycles in affected programs: 618749 -> 584410 (-5.55%)
helped: 271 / HURT: 0
helped stats (abs) min: 69 max: 540 x̄: 126.71 x̃: 108
helped stats (rel) min: 2.57% max: 37.97% x̄: 6.17% x̃: 5.06%
95% mean confidence interval for cycles value: -135.89 -117.54
95% mean confidence interval for cycles %-change: -6.72% -5.63%
Cycles are helped.
total sends in shared programs: 1025285 -> 1024355 (-0.09%)
sends in affected programs: 6454 -> 5524 (-14.41%)
helped: 271 / HURT: 0
helped stats (abs) min: 2 max: 8 x̄: 3.43 x̃: 4
helped stats (rel) min: 5.71% max: 25.00% x̄: 14.98% x̃: 17.39%
95% mean confidence interval for sends value: -3.57 -3.29
95% mean confidence interval for sends %-change: -15.42% -14.54%
Sends are helped.
According to Felix DeGrood, this results in a 10% improvement in
the draw call time for certain draw calls from Strange Brigade.
Reviewed-by: Caio Oliveira <caio.oliveira@intel.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/17944>
With this option enabled range of input values for fsin and fcos is
limited to [-2*pi : 2*pi] by calculating the reminder after 2*pi modulo
division. This helps to improve calculation precision for large input
arguments on Intel.
-v2: Add limit_trig_input_range option to prog_key to update shader
cache (Lionel)
Signed-off-by: Vadym Shovkoplias <vadym.shovkoplias@globallogic.com>
Reviewed-by: Lionel Landwerlin <lionel.g.landwerlin@intel.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/16388>
This controls the whole lowering of "make tex ops with implicit
derivatives on non-implicit-derivative stages be tex ops with an explicit
lod of 0 instead", but it's really hard to describe that in a git commit
summary.
All existing callers get it added except:
- nir_to_tgsi which didn't want it.
- nouveau, which didn't want it (fixes regressions in shadowcube and
shadow2darray with NIR, since the shading languages don't expose txl of
those sampler types and thus it's not supported in HW)
- optional lowering passes in mesa/st (lower_rect, YUV lowering, etc)
Reviewed-by: Marek Olšák <marek.olsak@amd.com>
Reviewed-by: Jason Ekstrand <jason.ekstrand@collabora.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/16156>
DG2 can only do sample_d and sample_d_c on 1D and 2D surfaces. Cube
maps and 3D surfaces were already handled, but 1D array and 2D array
surfaces were not.
Fixes the following Vulkan CTS failures on DG2:
dEQP-VK.glsl.texture_functions.texturegradclamp.isampler1darray_fragment
dEQP-VK.glsl.texture_functions.texturegradclamp.isampler2darray_fragment
dEQP-VK.glsl.texture_functions.texturegradclamp.sampler1darray_fixed_fragment
dEQP-VK.glsl.texture_functions.texturegradclamp.sampler1darray_float_fragment
dEQP-VK.glsl.texture_functions.texturegradclamp.sampler2darray_fixed_fragment
dEQP-VK.glsl.texture_functions.texturegradclamp.sampler2darray_float_fragment
dEQP-VK.glsl.texture_functions.texturegradclamp.usampler1darray_fragment
dEQP-VK.glsl.texture_functions.texturegradclamp.usampler2darray_fragment
The Fixes: tag below is a bit misleading. This commit adds another
lowering, similar to the one in the Fixes: commit, that probably should
have been added at the same time. I just want to make sure this commit
gets applied everywhere that commit was also applied.
Fixes: 635ed58e52 ("intel/compiler: Lower txd for 3D samplers on XeHP.")
Reviewed-by: Jason Ekstrand <jason.ekstrand@collabora.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/15681>
In general, an atomic intrinsic may perform separate atomics for every
enabled SIMD channel, as each channel may operate on different memory.
However, an extremely common case is for all channels to access the same
memory location. In this case, we can simply perform a reduction/scan
across the subgroup, and perform one atomic for the whole subgroup,
rather than one per channel. For example, if an intrinsic says to take
the minimum value of the existing memory and the value in each channel,
we can do a thread-local minimum of all enabled channels, then do a
single atomic to take the minimum of that and the existing memory.
Our hardware doesn't optimize the case where multiple channels ask for
atomics on the same memory location; it assumes the compiler will do so.
nir_opt_uniform_atomics() uses divergence analysis to detect this case,
adds the necessary subgroup operations, and moves the atomic inside a
conditional that disables all but a single invocation. It even detects
cases where the shader code already performs this kind of optimization,
and avoids doing it a second time.
This may not be the optimal solution for us. In the backend, we could
detect this case and emit send(1) instructions with NoMask, rather than
generating if...send(16)...endif, and a lot of unnecessary ALU ops. But
it's simple to do, reuses the same path as ACO, and still provides most
of the benefit by cutting up to 16x atomics down to a single atomic,
which is more merciful to the memory bus.
Improves performance of Shadow of the Tomb Raider by 5.5% on XeHP.
Improves performance of a customer-internal benchmark on XeHP at
3840x2160 and low settings by approximately 30%.
Reviewed-by: Ian Romanick <ian.d.romanick@intel.com>
Reviewed-by: Caio Oliveira <caio.oliveira@intel.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/15484>
This option only applies to relative shuffles (up/down/xor), and in a
moment we're going to add an option to lower normal shuffles, so rename
it.
While we're here, rename lower_shuffle() to lower_to_shuffle() for
similar reasons.
Reviewed-by: Danylo Piliaiev <dpiliaiev@igalia.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/14412>
Based on Rafael's:
* "nir/lower_tex: Add option to lower offset for tg4 too."
* "intel/compiler: Lower offsets for tg4 on gen9+."
* "WIP: Do not lower basic offsets."
* "WIP: intel/compiler: Enable lowering offsets restriction."
But, with these changes:
* Fixed range checking to be signed 4 bits
* Converted to filter
* Apply only to gfx12.5+
* Use nir_src_is_const / nir_src_comp_as_int (s-b Jason)
Signed-off-by: Jordan Justen <jordan.l.justen@intel.com>
Reviewed-by: Jason Ekstrand <jason@jlekstrand.net>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/14142>
Instead of checking for MESA_SHADER_COMPUTE (and KERNEL). Where
appropriate, also use gl_shader_stage_is_compute().
This allows most of the workgroup-related lowering to be applied to
Task and Mesh shaders. These will be added later and "inherit" from
cs_prog_data structure.
Reviewed-by: Lionel Landwerlin <lionel.g.landwerlin@intel.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/13629>
In the vec4 compiler, 8-bit types should never exist.
In the scalar compiler, 8-bit types should only ever be able to exist on
Gfx ver 8 and 9.
Some instructions are handled in non-obvious ways.
Hopefully this will save the next person some time.
Reviewed-by: Lionel Landwerlin <lionel.g.landwerlin@intel.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/9025>
This is where it should be rather than having to pass it into the
optimisation pass every time.
It also allows us to call the loop analysis pass without having to
duplicate these options which we will do later in this series.
Reviewed-by: Ian Romanick <ian.d.romanick@intel.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/12064>
Only fragment and some compute shaders support implicit derivatives.
They're totally meaningless without helper invocations and some
understanding of the dispatch pattern. We've got code to lower
nir_texop_tex in these shader stages to use an explicit derivative of 0
but it was pretty badly broken:
1. It only handled nir_texop_tex, not nir_texop_txb or nir_texop_lod.
2. It didn't take min_lod into account
3. It was conflated with adding a missing LOD parameter to opcodes
which expect one such as nir_texop_txf. While not really a bug,
this does make it way harder to reason about the code.
4. Unless you set a flag (which most drivers don't), it left the
opcode nir_texop_tex instead of nir_texop_txl which it should have
been.
This reworks it to go through roughly the same path as other LOD
lowering only with a constant lod of 0 instead of calling out to
nir_texop_lod. We also get rid of the lower_tex_without_implicit_lod
flag because most drivers set it and those that don't are probably
subtly broken. If someone really wants to get nir_texop_tex in their
vertex shaders, they can write a new patch to add the flag back in.
Fixes: e382890e25 "nir: set default lod to texture opcodes that..."
Fixes: d5ac5d6e83 "nir: Add option to lower tex to txl when..."
Reviewed-by: Connor Abbott <cwabbott0@gmail.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/11775>
Qualcomm has a mode with a subgroup size of 128, so just emitting larger
integer operations and then lowering them later isn't an option. This
makes the pass able to handle the lowering itself, so that we don't have
to go down to 64-thread wavefronts when ballots are used.
(The GLSL and legacy SPIR-V extensions only support a maximum of 64
threads, but I guess we'll cross that bridge when we come to it...)
Reviewed-by: Jason Ekstrand <jason@jlekstrand.net>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/6752>
965 and the mesa st disagree on how vertex elements are ordered when
edgeflags are involved. 965 wants them in gl_vert_attrib order,
but gallium supplies the edgeflag as the last vertex element regardless.
This adds a flag which is enabled for gen4/5 to denote that the
edgeflag is at the end. When we reap 965 later we can resolve this
better.
Reviewed-by: Jason Ekstrand <jason@jlekstrand.net>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/11146>
We don't want to have to deal with vector phis in freedreno, because
vectors are always split/unsplit around vectorized instructions anyways,
and the stated reason for not scalarising them (it hurting coalescing)
won't apply to us because we won't be using nir_from_ssa. Add this
option so that we don't have to do the equivalent thing while
translating from NIR.
Reviewed-by: Rob Clark <robdclark@gmail.com>
Reviewed-by: Daniel Schürmann <daniel@schuermann.dev>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/10809>
