When the register allocator decides to spill a value, all reads of that
value are filled. This can result in cases where the same value is
filled many times in a single block. In those cases, the result of an
earlier fill may still be available when a later fill occurs.
This optimization replaces the later fill with a move from the result of
the earlier fill.
v2: Use FIXED_GRF for register overlap tests. Since this is after
register allocation, the VGRF values will not tell the whole truth.
v3: Use brw_transform_inst. Suggested by Caio. Add
brw_scratch_inst::offset instead of storing it as a source. Suggested by
Lionel.
v4: In intervening spill to the same location also invalidates the
value. 🤦
v5: Don't eliminate a fill if its destination partially overlaps the
preceeding fill destination. Fixes failures in cooperative matrix CTS.
shader-db:
Lunar Lake, Meteor Lake, and DG2 had similar results. (Lunar Lake shown)
total instructions in shared programs: 17249903 -> 17249653 (<.01%)
instructions in affected programs: 35550 -> 35300 (-0.70%)
helped: 20 / HURT: 0
total cycles in shared programs: 893092398 -> 893101836 (<.01%)
cycles in affected programs: 2501720 -> 2511158 (0.38%)
helped: 6 / HURT: 14
total fills in shared programs: 1901 -> 1776 (-6.58%)
fills in affected programs: 1757 -> 1632 (-7.11%)
helped: 20 / HURT: 0
fossil-db:
Lunar Lake, Meteor Lake, and DG2 had similar results. (Lunar Lake shown)
Totals:
Instrs: 929949528 -> 926770338 (-0.34%)
Cycle count: 105126671329 -> 104851299099 (-0.26%); split: -0.28%, +0.02%
Fill count: 6520785 -> 5021518 (-22.99%)
Totals from 54281 (2.69% of 2018922) affected shaders:
Instrs: 239616289 -> 236437099 (-1.33%)
Cycle count: 22051883404 -> 21776511174 (-1.25%); split: -1.33%, +0.08%
Fill count: 6406295 -> 4907028 (-23.40%)
Reviewed-by: Lionel Landwerlin <lionel.g.landwerlin@intel.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/37827>
When the register allocator decides to spill a value, all writes to that
value are spilled and all reads are filled. In regions where there is
not high register pressure, a spill of a value may be followed by a fill
of that same file while the spilled register is still live. This
optimization pass finds these cases, and it converts the fill to a move
from the still-live register.
The restriction that the spill and the fill must have matching NoMask
really hampers this optimization. With the restriction removed, the pass
was more than 2x helpful.
v2: Require force_writemask_all to be the same for the spill and the fill.
v3: Use FIXED_GRF for register overlap tests. Since this is after
register allocation, the VGRF values will not tell the whole truth.
v4: Use brw_transform_inst. Suggested by Caio. The allows two of the
loops to be merged. Add brw_scratch_inst::offset instead of storing it
as a source. Suggested by Lionel.
v5: Add no-fill-opt debug option to disable optimizations. Suggested by
Lionel.
v6: Move a calculation outside a loop. Suggested by Lionel.
v7: Check that spill ranges overlap instead of just checking initial
offset. Zero shaders in fossil-db were affected, but some CTS with
spill_fs were fixed (e.g.,
dEQP-VK.subgroups.arithmetic.compute.subgroupmin_uint64_t_requiredsubgroupsize).
Suggested by Lionel.
v8: Add DEBUG_NO_FILL_OPT to debug_bits in
brw_get_compiler_config_value(). Noticed by Lionel.
shader-db:
Lunar Lake
total instructions in shared programs: 17249907 -> 17249903 (<.01%)
instructions in affected programs: 10684 -> 10680 (-0.04%)
helped: 2 / HURT: 0
total cycles in shared programs: 893092630 -> 893092398 (<.01%)
cycles in affected programs: 237320 -> 237088 (-0.10%)
helped: 2 / HURT: 0
total fills in shared programs: 1903 -> 1901 (-0.11%)
fills in affected programs: 110 -> 108 (-1.82%)
helped: 2 / HURT: 0
Meteor Lake and DG2 had similar results. (Meteor Lake shown)
total instructions in shared programs: 19968898 -> 19968778 (<.01%)
instructions in affected programs: 33020 -> 32900 (-0.36%)
helped: 10 / HURT: 0
total cycles in shared programs: 885157211 -> 884925015 (-0.03%)
cycles in affected programs: 39944544 -> 39712348 (-0.58%)
helped: 8 / HURT: 2
total fills in shared programs: 4454 -> 4394 (-1.35%)
fills in affected programs: 2678 -> 2618 (-2.24%)
helped: 10 / HURT: 0
fossil-db:
Lunar Lake
Totals:
Instrs: 930445228 -> 929949528 (-0.05%)
Cycle count: 105195579417 -> 105126671329 (-0.07%); split: -0.07%, +0.00%
Spill count: 3495279 -> 3494400 (-0.03%)
Fill count: 6767063 -> 6520785 (-3.64%)
Totals from 43844 (2.17% of 2018922) affected shaders:
Instrs: 212614840 -> 212119140 (-0.23%)
Cycle count: 19151130510 -> 19082222422 (-0.36%); split: -0.39%, +0.03%
Spill count: 2831100 -> 2830221 (-0.03%)
Fill count: 6128316 -> 5882038 (-4.02%)
Meteor Lake and DG2 had similar results. (Meteor Lake shown)
Totals:
Instrs: 1001375893 -> 1001113407 (-0.03%)
Cycle count: 92746180943 -> 92679877883 (-0.07%); split: -0.08%, +0.01%
Spill count: 3729157 -> 3728585 (-0.02%)
Fill count: 6697296 -> 6566874 (-1.95%)
Totals from 35062 (1.53% of 2284674) affected shaders:
Instrs: 179819265 -> 179556779 (-0.15%)
Cycle count: 18111194752 -> 18044891692 (-0.37%); split: -0.41%, +0.04%
Spill count: 2453752 -> 2453180 (-0.02%)
Fill count: 5279259 -> 5148837 (-2.47%)
Reviewed-by: Lionel Landwerlin <lionel.g.landwerlin@intel.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/37827>
These opcodes are emitted during register allocation instead of the
scratch reads and writes that were previously emitted. These
instructions contain additional information (i.e., the instruction
encodes the scratch offset) that enable optimizations to be added
later.
The fill and spill opcodes are lowered to scratch reads and writes
shortly after register allocation. Eventually this lower may have some
optimizations (e.g., reuse previous address calculations for successive
spills).
v2: Add brw_scratch_inst::offset instead of storing it as a
source. Suggested by Lionel.
Reviewed-by: Lionel Landwerlin <lionel.g.landwerlin@intel.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/37827>
This ensures that g0 is reserved for spilling since there is going to be
spilling.
Fixes: 8bca7e520c ("intel/brw: Only force g0's liveness to be the whole program if spilling")
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
Reviewed-by: Lionel Landwerlin <lionel.g.landwerlin@intel.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/37827>
Otherwise we get a lot of individual x/y/z stores to tesslevels when
we should really just be storing the whole thing at once.
Reviewed-by: Alyssa Rosenzweig <alyssa.rosenzweig@intel.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/38482>
The newly rewritten remap_tess_levels_legacy will have already lowered
anything it cares about to URB intrinsics. So the generic remapping
pass won't see them, as it operates on generic input/output intrinsics.
This also drops some of the callback boilerplate we needed temporarily.
Reviewed-by: Alyssa Rosenzweig <alyssa.rosenzweig@intel.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/38482>
This unifies the dynamic (SSO) and fixed (linked together) versions.
We emit piles of NIR as if we were doing the dynamic version, but
replace the tess config field access with constant values. It all
should optimize away back to something reasonable. We lower these
directly to URB read/write intrinsics.
It also rewrites the dynamic version to directly read/write the URB
rather than going through temporaries. The old version was broken
in that tessellation control shader invocations can technically use
the shared output area for cross-invocation data sharing with barriers,
although doing so using the built-in tesslevel patch outputs is very
unlikely.
Reviewed-by: Alyssa Rosenzweig <alyssa.rosenzweig@intel.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/38482>
We were using this for indirect loads of the shader input thread
payload, but there's no reason we can't use it for constant access
too. In this case we can just MOV from the ATTR file directly
without a special opcode that turns into MOV_INDIRECT later.
We also allow it to load multiple components now. This is useful
for say, returning vec4 pushed inputs. And, we allow it in more
stages than just the fragment stage.
Reviewed-by: Alyssa Rosenzweig <alyssa.rosenzweig@intel.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/38482>
We're going to change the intrinsic to a load(...) which puts "load" in
the name. Also, it's just more consistent with our usual terminology.
We also rename the corresponding backend opcode so they remain matched.
Reviewed-by: Alyssa Rosenzweig <alyssa.rosenzweig@intel.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/38482>
These are a ton more convenient. When the TCS and TES were linked
together, the legacy layouts were a hassle, but didn't impose any
significant cost. With unlinked TCS and TES, the legacy layouts
involve significant runtime code for scrambling the data, whereas
the reversed layouts are substantially less overhead.
Reviewed-by: Alyssa Rosenzweig <alyssa.rosenzweig@intel.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/38482>
st/nir lowers this for iris, and brw_link_shaders lowers this for anv,
but for unlinked tessellation control / evaluation shaders, the lowering
was not happening for TCS.
Just do it unconditionally when lowering TCS outputs and TES inputs.
This lets the remapping code just assume vectors all the time, rather
than getting single component stores with nir_intrinsic_component set
(which came from nir_lower_io lowering compact arrays).
This also requires changes to the dynamic unlinked TCS/TES lowering to
temporaries, which needs to use vectors rather than arrays with this
change. That code is going away in future patches anyway, but this
keeps it going for now to avoid interim breakage.
Reviewed-by: Alyssa Rosenzweig <alyssa.rosenzweig@intel.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/38482>
We now call remap_tess_levels before remap_non_header_patch_urb_offsets.
The latter already excludes tess levels anyway, so the order shouldn't
matter.
This paves the way for remap_tess_levels to skip handling some header
values in certain cases, because with reversed layouts, many of them no
longer need any special handling and we can just let the generic pass
handle them.
Reviewed-by: Alyssa Rosenzweig <alyssa.rosenzweig@intel.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/38482>
Just have a single remap_tess_levels that does either the
statically-known-primitive or the dynamic (unlinked) mode.
Reviewed-by: Alyssa Rosenzweig <alyssa.rosenzweig@intel.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/38482>
Our current legacy patch header layout handling doesn't actually care
which is which slot, and remaps everything to its correct spot anyway.
For using the newer "reversed" patch header layouts, it will be more
convenient to have outer as slot 0, and inner as slot 1, as that just
works with no special remapping needed for both quads and triangles
(but unfortunately isolines are still a pain).
Reviewed-by: Alyssa Rosenzweig <alyssa.rosenzweig@intel.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/38482>
These work the same regardless of stage.
v2 (Ken): Rebase, move from mesh to all stages, add reorderable load
variant, allow channel masks to be non-constant even on Xe2.
Signed-off-by: Lionel Landwerlin <lionel.g.landwerlin@intel.com>
Signed-off-by: Kenneth Graunke <kenneth@whitecape.org>
Reviewed-by: Alyssa Rosenzweig <alyssa.rosenzweig@intel.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/38482>
Be consistent with lowering that happens after, so that it gets a full
vector register and can stride into it.
Signed-off-by: Lionel Landwerlin <lionel.g.landwerlin@intel.com>
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
Reviewed-by: Alyssa Rosenzweig <alyssa.rosenzweig@intel.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/38482>
Allows us to better determine if we need Z/W payload delivery.
Signed-off-by: Lionel Landwerlin <lionel.g.landwerlin@intel.com>
Reviewed-by: Alyssa Rosenzweig <alyssa.rosenzweig@intel.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/36392>
anv sets device->uses_ex_bso on verx10 >= 125 and then sets the
compiler->extended_bindless_surface_offset to that.
iris was not setting anything. However, LSC_ADDR_SURFTYPE_SS used for
scratch on Gfx12.5 is bindless, and Xe2 uses ExBSO for all UGM access,
so we need to be setting this.
Just set it in the compiler so both drivers have it set.
Fixes piglit arb_tessellation_shader-tes-gs-max-output -small -scan 1 50
on iris.
Fixes: 80c89909f3 ("brw: fixup immediate bindless surface handling")
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/38645>
In 80c89909f3 ("brw: fixup immediate bindless surface handling") I
forgot that we have a special usage for the only _SS surface (the
scratch surface).
Because it's only delivered in the 31:10 bits of R0 and because we
want to minimize the amount of shader instructions for scratch
messages, the surface offset in shifted right by the driver to align
things properly for the 31:6 extended descriptor format.
This is unfortunately incompatible with the full 32bit format of
ExBSO. So this surface type currently cannot be considered bindless.
We might revisit later if we start using _SS surfaces for other
things.
Signed-off-by: Lionel Landwerlin <lionel.g.landwerlin@intel.com>
Fixes: 80c89909f3 ("brw: fixup immediate bindless surface handling")
Reviewed-by: Rohan Garg <rohan.garg@intel.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/38618>
Fixes this test on Xe2+:
INTEL_DEBUG=no32 ./deqp-vk -n dEQP-VK.spirv_assembly.instruction.maint9_vectorization.bit_field_u_extract.result_v16i-base_v16i-offset_s64u-count_s16i
Generate invalid code for that platform:
and(16) g37<1>UW g65<16,4,4>UW 0x000fUW { align1 1H I@5 };
ERROR: Invalid register region for source 0. See special restrictions section.
Several helpers like has_subdword_integer_region_restriction() do not
see the final type of the source, so compute it early.
Maybe new_src could be used in more cases. Being conservative for now.
Signed-off-by: Lionel Landwerlin <lionel.g.landwerlin@intel.com>
Cc: mesa-stable
Reviewed-by: Francisco Jerez <currojerez@riseup.net>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/38548>
This is strictly a GL thing. iris can manage it in its own program keys
without polluting the compiler with stuff nobody else cares about.
We can also drop a lot of padding that was introduced in commit
a18835a9ca which doesn't appear to be
necessary.
Reviewed-by: Alyssa Rosenzweig <alyssa.rosenzweig@intel.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/38556>
Rather than adding another boolean to optionally lower PLS vars, pass
the types we want to lowers through a nir_variable_mode bitmask.
Signed-off-by: Boris Brezillon <boris.brezillon@collabora.com>
Reviewed-by: Marek Olšák <marek.olsak@amd.com>
Reviewed-by: Lionel Landwerlin <lionel.g.landwerlin@intel.com>
Reviewed-by: Eric R. Smith <eric.smith@collabora.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/37110>
send.ugm (1|M0) r125 r0 null:0 0x0 0x0200651F {$9} // wr:1+0, rd:0; fence invalid flush type scoped to tile
When destination of Send(s) is not null, the response length must not be 0.
Should only affect DG2 products.
Signed-off-by: Lionel Landwerlin <lionel.g.landwerlin@intel.com>
Cc: mesa-stable
Reviewed-by: Alyssa Rosenzweig <alyssa.rosenzweig@intel.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/38478>
We need to make sure the data part returned by sampler messages is
always aligned to a physical register. Just like the residency data
lives in a single physical register after the data.
Lowering a vec3 16bits per components led to a half a physical
register allocation which then confused the descriptor lowering
(expecting physical register units).
Signed-off-by: Lionel Landwerlin <lionel.g.landwerlin@intel.com>
Fixes: 295734bf88 ("intel/fs: fix residency handling on Xe2")
Closes: https://gitlab.freedesktop.org/mesa/mesa/-/issues/12794
Reviewed-by: Alyssa Rosenzweig <alyssa.rosenzweig@intel.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/34008>
This is unused at the moment but the backend incorrectly assumes
immediate handles are for the binding table (therefore not bindless).
Some new CTS tests are using an immediate bindless handle which is
broken.
Signed-off-by: Lionel Landwerlin <lionel.g.landwerlin@intel.com>
Acked-by: Alyssa Rosenzweig <alyssa.rosenzweig@intel.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/38359>
This is done by grep ALIGN( to align(
docs,*.xml,blake3 is excluded
Signed-off-by: Yonggang Luo <luoyonggang@gmail.com>
Reviewed-by: Timur Kristóf <timur.kristof@gmail.com>
Acked-by: Alyssa Rosenzweig <alyssa.rosenzweig@intel.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/38365>
We add a bunch of new helpers to avoid the need to touch >parent_instr,
including the full set of:
* nir_def_is_*
* nir_def_as_*_or_null
* nir_def_as_* [assumes the right instr type]
* nir_src_is_*
* nir_src_as_*
* nir_scalar_is_*
* nir_scalar_as_*
Plus nir_def_instr() where there's no more suitable helper.
Also an existing helper is renamed to unify all the names, while we're
churning the tree:
* nir_src_as_alu_instr -> nir_src_as_alu
..and then we port the tree to use the helpers as much as possible, using
nir_def_instr() where that does not work.
Acked-by: Marek Olšák <maraeo@gmail.com>
---
To eliminate nir_def::parent_instr we need to churn the tree anyway, so I'm
taking this opportunity to clean up a lot of NIR patterns.
Co-authored-by: Konstantin Seurer <konstantin.seurer@gmail.com>
Signed-off-by: Alyssa Rosenzweig <alyssa.rosenzweig@intel.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/38313>
Xe2 uses byte offsets rather than OWord offsets. We've been storing the
per-slot offsets in bytes on Xe2 for a while, but kept the global offset
immediate in OWords for some reason, choosing to lower it during logical
send lowering.
This patch makes both offsets (global immediate, per-slot) in the same
units, so they could be added together if necessary without scaling.
Reviewed-by: Lionel Landwerlin <lionel.g.landwerlin@intel.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/38343>
Much clearer, especially since we're dealing with at least four
different kinds of intrinsics. These helpers were introduced years ago,
but probably didn't exist when we first wrote this code.
Reviewed-by: Faith Ekstrand <faith.ekstrand@collabora.com>
Reviewed-by: Lionel Landwerlin <lionel.g.landwerlin@intel.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/38343>
