Lowering the indirect derefs multiple times leads to very inefficient
shaders because of all the control flow inserted.
In particular on some DGC tests with mesh shaders, the tests can spin
for 1hour on an i7 and still not complete compilation.
Signed-off-by: Lionel Landwerlin <lionel.g.landwerlin@intel.com>
Reviewed-by: Caio Oliveira <caio.oliveira@intel.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/33809>
On older hardware the "use_rep_send" compile parameter was being
implicitly used to request the compilation of the SIMD16 variant of
clear pixel shaders that require it due to hardware restrictions.
However starting on Gfx12+ this flag is never set since replicated
data clears are no longer supported, but BLORP still implicitly relies
on the SIMD16 variant being generated even though there's no way for
BLORP to explicitly request it. This doesn't cause much of a problem
right now since brw_compile_fs() typically generates a SIMD16 kernel
unless the SIMD8 kernel spills or SIMD debugging flags are enabled,
but it won't work reliably on Xe3+ since we'll start using SIMD32 more
aggressively.
In order to avoid these issues use the standard required subgroup_size
parameter from shader_info to signal that the SIMD16 variant of the
shader is needed by the caller.
Reviewed-by: Lionel Landwerlin <lionel.g.landwerlin@intel.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/32664>
This pass used to unconditionally use divergence information
which forced the caller to either call divergence_analysis or
ensure that the divergence is properly reset.
Reviewed-by: Marek Olšák <marek.olsak@amd.com>
Reviewed-by: Alyssa Rosenzweig <alyssa@rosenzweig.io>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/33009>
The hope here is to replace our backend handling for loading whole
cachelines at a time from UBOs into NIR-based handling, which plays
nicely with the NIR load/store vectorizer.
Rounding down offsets to multiples of 64B allows us to globally CSE
UBO loads across basic blocks. This is really useful. However, blindly
rounding down the offset to a multiple of 64B can trigger anti-patterns
where...a single unaligned memory load could have hit all the necessary
data, but rounding it down split it into two loads.
By moving this to NIR, we gain more control of the interplay between
nir_opt_load_store_vectorize and this rebasing and CSE'ing. The backend
can then simply load between nir_def_{first,last}_component_read() and
trust that our NIR has the loads blockified appropriately.
Reviewed-by: Lionel Landwerlin <lionel.g.landwerlin@intel.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/32888>
Marek recently changed hole_size to be signed, rather than unsigned.
A negative hole_size means that the two loads overlap - and thus are
prime candidates to be combined.
My original hole_size handling was:
if hole_size > 4 * (8 - low->num_components) then don't vectorize
For non-overlapping loads, this worked: NIR's largest vector is vec16,
and if low was already a vec16, combining it with anything would exceed
that, so it'd never be considered. That meant low would always be a
vec8 or less, so (8 - low->num_components) was a positive number.
Now that we see overlapping loads, we can see a vec16 low, vec4 high,
and also a negative hole size, giving us fun comparisons like:
-16 > 4 * (8 - 16) => -16 > -32 => true, don't vectorize
Which is absolutely the wrong thing to do, because the high load's data
is entirely included within the former load's data.
The idea here was to make sure the second load would be able to pack at
least one component into the first's V8 result. But even this isn't the
best, because...even if it's simply adjacent, doing one V16 load is more
efficient than requesting two back to back V8 loads.
So, we just simplify down to a static check: if there's an entire V8 of
hole, don't vectorize. This already won't happen because the core pass
has max_hole set to 28 bytes (7 32-bit components), but that could
change based on the needs of other drivers, so let's be defensive.
fossil-db results on Alchemist:
Instrs: 161533978 -> 161295137 (-0.15%); split: -0.20%, +0.05%
Subgroup size: 8092544 -> 8092568 (+0.00%)
Send messages: 7915233 -> 7844503 (-0.89%); split: -0.94%, +0.05%
Cycle count: 16577700697 -> 16702609256 (+0.75%); split: -0.59%, +1.35%
Spill count: 72338 -> 67226 (-7.07%); split: -7.36%, +0.29%
Fill count: 134058 -> 125980 (-6.03%); split: -6.83%, +0.80%
Scratch Memory Size: 4092928 -> 3786752 (-7.48%); split: -7.53%, +0.05%
Max live registers: 33031460 -> 32945994 (-0.26%); split: -0.27%, +0.01%
Max dispatch width: 5778384 -> 5778536 (+0.00%); split: +0.26%, -0.26%
Non SSA regs after NIR: 179809505 -> 152735471 (-15.06%); split: -15.08%, +0.03%
Fixes: c21bc65ba7 ("nir/opt_load_store_vectorize: make hole_size signed to indicate overlapping loads")
Reviewed-by: Lionel Landwerlin <lionel.g.landwerlin@intel.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/32932>
A negative hole size means the loads overlap. This will be used by drivers
to handle overlapping loads in the callback easily.
Reviewed-by: Mel Henning <drawoc@darkrefraction.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/32699>
This is needed for implementing multiview in panvk, where the address
calculation for multiview outputs is not well-represented by lowering to
nir_intrinsic_store_output with a single offset.
The case where a variable is both per-view and per-{vertex,primitive} is
now unsupported. This would come up with drivers implementing
NV_mesh_shader or using nir_lower_multiview on geometry, tessellation,
or mesh shaders. No drivers currently do either of these. There was some
code that attempted to handle the nested per-view case by unwrapping
per-view/arrayed types twice, but it's unclear to what extent this
actually worked.
ANV and Turnip both rely on per-view outputs being assigned a unique
driver location for each view, so I've added on option to configure that
behavior rather than removing it.
Signed-off-by: Benjamin Lee <benjamin.lee@collabora.com>
Reviewed-by: Boris Brezillon <boris.brezillon@collabora.com>
Reviewed-by: Lionel Landwerlin <lionel.g.landwerlin@intel.com>
Reviewed-by: Connor Abbott <cwabbott0@gmail.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/31704>
Notably, our convergent block loads were already overfetching - we
rounded up to block sizes of 8, 16, 32, or 64(LSC-only). But we did
so in the backend, rather than NIR.
With recent changes, nir_opt_load_store_vectorizer allows holes of up
to 28 bytes (7 components at 4 bytes each). This allows us to detect
cases where we did a convergent block load for 1 component (but loaded
a whole vec8), then another load for the next vec8, and combine them
into a single V16 load. Single component loads aren't the most common,
but convergent loads of a vec2 in one group and a vec3 in another are
quite common, and it makes no sense to do V8+V8 loads instead of V16.
For non-block loads, we allow a max hole of 4 bytes. This allows the
common case of XYZ_ + XYZ_ loads (where the last component is unread)
to combine into a single larger load.
fossil-db results on Lunarlake:
Totals:
Instrs: 146692608 -> 146246432 (-0.30%); split: -0.33%, +0.02%
Subgroup size: 11100528 -> 11100512 (-0.00%)
Send messages: 7003425 -> 6862529 (-2.01%); split: -2.01%, +0.00%
Cycle count: 22396273274 -> 22523048654 (+0.57%); split: -1.08%, +1.64%
Spill count: 67671 -> 67594 (-0.11%); split: -1.59%, +1.48%
Fill count: 128999 -> 130223 (+0.95%); split: -1.73%, +2.68%
Scratch Memory Size: 5986304 -> 6042624 (+0.94%); split: -1.40%, +2.34%
Max live registers: 48898858 -> 48881655 (-0.04%); split: -0.05%, +0.01%
Non SSA regs after NIR: 172397792 -> 167577380 (-2.80%); split: -2.80%, +0.00%
Totals from 451003 (80.87% of 557667) affected shaders:
Instrs: 134111754 -> 133665578 (-0.33%); split: -0.36%, +0.03%
Subgroup size: 9039104 -> 9039088 (-0.00%)
Send messages: 6127775 -> 5986879 (-2.30%); split: -2.30%, +0.00%
Cycle count: 20306336726 -> 20433112106 (+0.62%); split: -1.19%, +1.81%
Spill count: 56230 -> 56153 (-0.14%); split: -1.92%, +1.78%
Fill count: 112920 -> 114144 (+1.08%); split: -1.97%, +3.06%
Scratch Memory Size: 3769344 -> 3825664 (+1.49%); split: -2.23%, +3.72%
Max live registers: 43750259 -> 43733056 (-0.04%); split: -0.05%, +0.01%
Non SSA regs after NIR: 158449343 -> 153628931 (-3.04%); split: -3.04%, +0.00%
In particular, sends get cut by 20.85% for Borderlands 3 DX12, 13.82%
on Cyberpunk 2077, 10.75% on Strange Brigade, and 10.20% on Red Dead
Redemption 2. Yet, spill/fills remain about the same.
fossil-db results on Alchemist are similar though not quite as good.
Reviewed-by: Lionel Landwerlin <lionel.g.landwerlin@intel.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/32315>
nir_opt_load_store_vectorize checks for potential address wrapping
when vectorizing two loads ("low" and "high"). It looks for cases where
"low" might have a large address, and "high" has a positive offset
which, when added together, could trigger integer wraparound. The issue
here is that if the large address of "low" was considered out-of-bounds,
adding offset could wrap around to a small address, which might actually
be in-bounds. Thus, when loaded separately, "low" will fail and trigger
robustness out-of-bound-read behavior, but "high" would read correctly.
When vectorized, the entire load would fail. This is explicitly tested
for with 32-bit SSBO addresses in the Vulkan CTS.
However, anv's 64-bit global addresses and VMA handling effectively
prevent this case. Addresses 0-4095 are a reserved page so that if
people try to use 0 as a NULL pointer, it never maps to a valid BO.
That alone guarantees that the above case where "high" gets a small
address would never be in-bounds, so we don't need to check for it.
In fact, we allocate most user allocations out of high addresses,
and have specialized allocation heaps for certain types of GPU data
structures in the lower GB of memory. For a load to wrap around and
successfully land in the right heap, it would have to load gigabytes.
Disabling this allows load vectorization and overfetching in more cases.
Reviewed-by: Lionel Landwerlin <lionel.g.landwerlin@intel.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/32315>
Moving it to intel_shader_enums.h
The plan is to make it visible to OpenCL shaders.
Signed-off-by: Lionel Landwerlin <lionel.g.landwerlin@intel.com>
Reviewed-by: Tapani Pälli <tapani.palli@intel.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/32329>
This will need to be set to true when the GLSL linker lowers IO, which
can later be unlowered by st/mesa, and then drivers can lower it again
without load_interpolated_input. Therefore, it can't be a global
immutable option.
Reviewed-by: Alyssa Rosenzweig <alyssa@rosenzweig.io>
Reviewed-by: Georg Lehmann <dadschoorse@gmail.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/32229>
At the point we were calling this, we hadn't necessarily cleaned up
derefs via nir_lower_vars_to_ssa, nor movs/vecs via copy propagation,
so it wasn't necessarily easy for this pass to see the actual usage of
the destination.
Moving this later allows us to detect f2f32(txf(...)) and avoid
converting it to a 16-bit txf (why convert with ALU instructions
when the sampler could do it for us?).
Reviewed-by: Lionel Landwerlin <lionel.g.landwerlin@intel.com>
Reviewed-by: Sushma Venkatesh Reddy <sushma.venkatesh.reddy@intel.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/31750>
This intrinsic was initially dedicated to mesh/task shaders, but the
mechanism it exposes also exists in the compute shaders on Gfx12.5+.
Signed-off-by: Lionel Landwerlin <lionel.g.landwerlin@intel.com>
Reviewed-by: Caio Oliveira <caio.oliveira@intel.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/31508>
Useful to insert debug traces a bit later in the lowering process (in
particular after load/store vectorization).
Signed-off-by: Lionel Landwerlin <lionel.g.landwerlin@intel.com>
Reviewed-by: Caio Oliveira <caio.oliveira@intel.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/31508>
If you allow an unsupported component count in the callback for loads,
nir_opt_load_store_vectorize will align num_components to the next supported
vector size, essentially overfetching.
This changes all callbacks to reject it. AMD will enable it in a later commit.
Reviewed-by: Alyssa Rosenzweig <alyssa@rosenzweig.io>
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
Reviewed-by: Rhys Perry <pendingchaos02@gmail.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/29398>
It will be used to allow merging loads with a hole between them.
Reviewed-by: Qiang Yu <yuq825@gmail.com>
Reviewed-by: Alyssa Rosenzweig <alyssa@rosenzweig.io>
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
Reviewed-by: Rhys Perry <pendingchaos02@gmail.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/29398>
This is going to handle more than atomics shortly.
Reviewed-by: Lionel Landwerlin <lionel.g.landwerlin@intel.com>
Reviewed-by: Caio Oliveira <caio.oliveira@intel.com>
Reviewed-by: Rohan Garg <rohan.garg@intel.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/30828>
Without this, the next commit tiggers assertions.
v2: Unconditionally do the lowering after brw_nir_optimize. Suggested by
Caio.
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org> [v1]
Reviewed-by: Caio Oliveira <caio.oliveira@intel.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/30251>
We were specifying align_offset = 64 and align_mul = 64, which is
invalid. nir_combined_align() asserts that align_offset < align_mul.
Our intention here is to perform cacheline-aligned (64B-aligned) block
loads, so we should set align_mul = 64 and can leave align_offset = 0.
Fixes: fbafa9cabd ("intel/nir: remove load_global_const_block_intel intrinsic")
Acked-by: Lionel Landwerlin <lionel.g.landwerlin@intel.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/30755>
load_global_constant_uniform_block_intel is equivalent in terms of
loading, then for the predicate we just do a bcsel afterward in places
where that is required.
Signed-off-by: Lionel Landwerlin <lionel.g.landwerlin@intel.com>
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/30659>
Intel GPUs have a saturate destination modifier, and
brw_fs_opt_saturate_propagation tries to replace explicit saturate
operations with this destination modifier. That pass is limited in
several ways. If the source of the explicit saturate is in a different
block or if the source of the explicit saturate is live after the
explicit saturate, brw_fs_opt_saturate_propagation will be unable to
make progress.
This optimization exists to help brw_fs_opt_saturate_propagation make
more progress. It tries to move NIR fsat instructions to the same block
that contains the definition of its source. It does this only in cases
where it will not create additional live values. It also attempts to do
this only in cases where the explicit saturate will ultimiately be
converted to a destination modifier.
v2: Fix metadata_preserve when theres no progress and use
nir_metadata_control_flow when there is progress. All suggested by
Alyssa.
v3: Fix a typo in the file header comment. Noticed by Ken. Don't
require nir_metadata_instr_index. Use nir_def_rewrite_uses_after instead
of open-coding something slightly more specific. Both suggested by Ken.
shader-db:
All Intel platforms had similar results. (Meteor Lake shown)
total instructions in shared programs: 19733645 -> 19733028 (<.01%)
instructions in affected programs: 193300 -> 192683 (-0.32%)
helped: 246
HURT: 1
helped stats (abs) min: 2 max: 48 x̄: 2.51 x̃: 2
helped stats (rel) min: 0.18% max: 0.39% x̄: 0.33% x̃: 0.34%
HURT stats (abs) min: 1 max: 1 x̄: 1.00 x̃: 1
HURT stats (rel) min: 0.31% max: 0.31% x̄: 0.31% x̃: 0.31%
95% mean confidence interval for instructions value: -2.87 -2.13
95% mean confidence interval for instructions %-change: -0.34% -0.32%
Instructions are helped.
total cycles in shared programs: 916180971 -> 916264656 (<.01%)
cycles in affected programs: 30197180 -> 30280865 (0.28%)
helped: 194
HURT: 142
helped stats (abs) min: 1 max: 21251 x̄: 872.75 x̃: 19
helped stats (rel) min: <.01% max: 23.17% x̄: 2.59% x̃: 0.23%
HURT stats (abs) min: 1 max: 28058 x̄: 1781.68 x̃: 399
HURT stats (rel) min: <.01% max: 37.21% x̄: 4.85% x̃: 1.63%
95% mean confidence interval for cycles value: -196.84 694.97
95% mean confidence interval for cycles %-change: -0.17% 1.27%
Inconclusive result (value mean confidence interval includes 0).
fossil-db:
Meteor Lake, DG2, and Tiger Lake had similar results. (Meteor Lake shown)
Totals:
Instrs: 151512021 -> 151511351 (-0.00%); split: -0.00%, +0.00%
Cycle count: 17209013596 -> 17209840995 (+0.00%); split: -0.02%, +0.02%
Max live registers: 32013312 -> 32013549 (+0.00%)
Max dispatch width: 5512304 -> 5512136 (-0.00%)
Totals from 774 (0.12% of 630172) affected shaders:
Instrs: 1559285 -> 1558615 (-0.04%); split: -0.05%, +0.01%
Cycle count: 1312656268 -> 1313483667 (+0.06%); split: -0.24%, +0.30%
Max live registers: 82195 -> 82432 (+0.29%)
Max dispatch width: 6664 -> 6496 (-2.52%)
Ice Lake
Totals:
Instrs: 151416791 -> 151416137 (-0.00%); split: -0.00%, +0.00%
Cycle count: 15162468885 -> 15163298824 (+0.01%); split: -0.00%, +0.01%
Max live registers: 32471367 -> 32471603 (+0.00%)
Max dispatch width: 5623752 -> 5623712 (-0.00%)
Totals from 733 (0.12% of 635598) affected shaders:
Instrs: 877965 -> 877311 (-0.07%); split: -0.09%, +0.01%
Cycle count: 190763628 -> 191593567 (+0.44%); split: -0.21%, +0.64%
Max live registers: 72067 -> 72303 (+0.33%)
Max dispatch width: 6216 -> 6176 (-0.64%)
Skylake
Totals:
Instrs: 140794845 -> 140794075 (-0.00%); split: -0.00%, +0.00%
Cycle count: 14665159301 -> 14665320514 (+0.00%); split: -0.00%, +0.01%
Max live registers: 31783341 -> 31783662 (+0.00%); split: -0.00%, +0.00%
Totals from 659 (0.11% of 625670) affected shaders:
Instrs: 829061 -> 828291 (-0.09%); split: -0.09%, +0.00%
Cycle count: 185478478 -> 185639691 (+0.09%); split: -0.33%, +0.41%
Max live registers: 67491 -> 67812 (+0.48%); split: -0.01%, +0.48%
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/29774>
Gfx8 required us to emulate image load store with untyped messages,
whereas Gfx9 just has typed message support for everything. brw no
longer supports Gfx8, so all of this code is effectively dead.
Reviewed-by: Caio Oliveira <caio.oliveira@intel.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/30576>
on agx (and mali), we predicate atomics on "if (!helper)", so doing so again in
this pass is redundant. and would cause a problem since we'd then have to lower
the "is helper inv?" flag late. so just skip the extra lowering code.
Signed-off-by: Alyssa Rosenzweig <alyssa@rosenzweig.io>
Acked-by: Rhys Perry <pendingchaos02@gmail.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/30488>
UBO loads with a non-indirect buffer index should be safe to perform
speculatively. With a direct offset, we may sometimes turn them into
push constants, at which point it's just reading a register with no
cost at all. Otherwise, we access them via messages that use surface
state, and automatically perform bounds checking. So we shouldn't have
any issues with reading out of bounds and page faulting, for example.
This allows nir_opt_peephole_sel() to operate on load_ubo intrinsics,
so we can turn simple if's with loads on both sides to bcsels. In some
cases this can collapse a surprising amount of control flow, allowing
other optimizations to work better.
The i965 OpenGL driver used load_uniform intrinsics, which are allowed
in NIR's peephole select pass. But iris uses the Gallium NIR pass that
translates uniforms to loads from UBO 0, so we haven't been able to take
advantage of NIR's peephole select pass there. The backend pass was
still able to handle this to some extent, however.
fossil-db results on Alchemist:
Totals:
Instrs: 150656329 -> 150645307 (-0.01%); split: -0.01%, +0.00%
Cycles: 12635230179 -> 12633696811 (-0.01%); split: -0.02%, +0.00%
Send messages: 7416330 -> 7416261 (-0.00%)
Spill count: 52471 -> 52473 (+0.00%)
Fill count: 100818 -> 100803 (-0.01%); split: -0.02%, +0.00%
Scratch Memory Size: 3197952 -> 3198976 (+0.03%)
Totals from 1848 (0.29% of 630003) affected shaders:
Instrs: 1412300 -> 1401278 (-0.78%); split: -0.80%, +0.02%
Cycles: 1809789567 -> 1808256199 (-0.08%); split: -0.11%, +0.03%
Send messages: 59829 -> 59760 (-0.12%)
Spill count: 3870 -> 3872 (+0.05%)
Fill count: 9693 -> 9678 (-0.15%); split: -0.18%, +0.02%
Scratch Memory Size: 174080 -> 175104 (+0.59%)
Reviewed-by: Caio Oliveira <caio.oliveira@intel.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/30498>
API requires samplers to return 32-bit even though hardware can handle
16-bit floating point, so we detect that case and make more efficient
use of memory BW. This is helping improve performance of encode and
decode tokens during LLM by at least 5% across multiple platforms.
Thank you Kenneth Graunke for suggesting and guiding me throughout
this implementation.
Signed-off-by: Sushma Venkatesh Reddy <sushma.venkatesh.reddy@intel.com>
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/30447>
Floating-point offsets work fine in combination with the
floating-point arithmetic we're about to lower these intrinsics into,
and they require less instructions than converting to fixed-point and
then back. No reason to take the precision/range hit nor the extra
instructions.
Reviewed-by: Caio Oliveira <caio.oliveira@intel.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/29847>
