This reverts commit 5489033fa8.
The problem I was trying to address is that we were programming the
3DSTATE_PS::PositionXYOffsetSelect bit differently with GPL (CENTROID)
than without (NONE).
I failed to understand that this bit also impacts the thread payload
layout. GPL fragment shaders don't know ahead of time if pos_offset is
going to be used. It'll be choosen at runtime base on push constant
bits. So we need to program this bit different just to have a payload
matching the compiled shader code.
This fixes the freedoom replay with GPL FS shader in SIMD32.
Signed-off-by: Lionel Landwerlin <lionel.g.landwerlin@intel.com>
Acked-by: Ivan Briano <ivan.briano@intel.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/22938>
This can generate things like fneg! of load_const, which is silly.
Fold those away into an actual constant. Only do so on the scalar
backend because there's a comment above that the vec4 backend doesn't
want any new constants this late, and I'm inclined to believe it.
fossil-db stats show a very minor improvement:
Totals:
Instrs: 203091223 -> 203091099 (-0.00%); split: -0.00%, +0.00%
Cycles: 14410638075 -> 14410577067 (-0.00%); split: -0.00%, +0.00%
Totals from 20 (0.00% of 665070) affected shaders:
Instrs: 27067 -> 26943 (-0.46%); split: -0.47%, +0.01%
Cycles: 2687958 -> 2626950 (-2.27%); split: -2.27%, +0.00%
Reviewed-by: Lionel Landwerlin <lionel.g.landwerlin@intel.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/22881>
With the following test :
dEQP-VK.spirv_assembly.instruction.terminate_invocation.terminate.no_out_of_bounds_load
There is a :
shader_start:
... <- no control flow
g0 = some_alu
g1 = fbl
g2 = broadcast g3, g1
g4 = get_buffer_size g2
... <- no control flow
halt <- on some lanes
g5 = send <surface>, g4
eliminate_find_live_channel will remove the fbl/broadcast because it
assumes lane0 is active at get_buffer_size :
shader_start:
... <- no control flow
g0 = some_alu
g4 = get_buffer_size g0
... <- no control flow
halt <- on some lanes
g5 = send <surface>, g4
But then the instruction scheduler will move the get_buffer_size after
the halt :
shader_start:
... <- no control flow
halt <- on some lanes
g0 = some_alu
g4 = get_buffer_size g0
g5 = send <surface>, g4
get_buffer_size pulls the surface index from lane0 in g0 which could
have been turned off by the halt and we end up accessing an invalid
surface handle.
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/20765>
Comparisons which produce 32-bit boolean results (0 or 0xFFFFFFFF)
but operate on 16-bit types would first generate a CMP instruction
with W or HF types, before expanding it out. This CMP is a partial
write, which leads us to think the register may contain some prior
contents still. When placed in a loop, this causes its live range
to extend beyond its real life time.
Mark the register with UNDEF first so that we know that no prior
contents exist and need to be preserved.
This affects:
flt32, fge32, feq32, fneu32, ilt32, ult32, ige32, uge32, ieq32, ine32
On one of Cyberpunk 2077's most complex compute shaders, this reduces
the maximum live registers from 696 to 537 (22.8%). Together with the
next patch, Cyberpunk's spills and fills are cut by 10.23% and 9.19%,
respectively.
Reviewed-by: Lionel Landwerlin <lionel.g.landwerlin@intel.com>
Reviewed-by: Francisco Jerez <currojerez@riseup.net>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/22835>
This value depends on the per-sample value which can be unknown at
compile time with graphics pipeline libraries. So we need to have this
dynamic has well and pick the right value when generating the
3DSTATE_PS/3DSTATE_WM packet.
Signed-off-by: Lionel Landwerlin <lionel.g.landwerlin@intel.com>
Fixes: d8dfd153c5 ("intel/fs: Make per-sample and coarse dispatch tri-state")
Reviewed-by: Emma Anholt <emma@anholt.net>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/22728>
For MTL (verx10 == 125), float64 is supported, but int64 is not.
Therefore we need to lower cluster broadcast using 32-bit int ops.
For gfx12.5+ platforms that support int64, the register regions
used by cluster broadcast aren't supported by the 64-bit pipeline.
On MTL, dEQP-VK.subgroups.clustered.*_double* and
dEQP-VK.subgroups.clustered.*_dvec* were failing to validate the
compiled shader in debug mode, and reportedly gpu-hanging in release
mode.
With this change dEQP-VK.subgroups.clustered.*_double* passed all 48
tests and dEQP-VK.subgroups.clustered.*_dvec* passed all 140 tests on
MTL.
Rework:
* Move from generator to brw_fs_lower_regioning.cpp. (Suggested by
Francisco)
* Apply to verx10 >= 125.. (Suggested by Francisco)
Cc: 23.1 <mesa-stable>
Signed-off-by: Jordan Justen <jordan.l.justen@intel.com>
Reviewed-by: Marcin Ślusarz <marcin.slusarz@intel.com> (v1)
Reviewed-by: Francisco Jerez <currojerez@riseup.net>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/22569>
The instructions manipulation cr0 use the default mask on lane0. So if
for some reason that lane is disabled in some of the dispatchs, we can
end up not executing the instructions.
Fixes flakyness in dEQP-VK.spirv_assembly.instruction.graphics.16bit_storage.uniform_float_32_to_16.uniform_matrix_float_rtz_frag
Signed-off-by: Lionel Landwerlin <lionel.g.landwerlin@intel.com>
Cc: mesa-stable
Reviewed-by: Tapani Pälli <tapani.palli@intel.com>
Reviewed-by: Francisco Jerez <currojerez@riseup.net>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/22314>
There's no guarantee that this is a SSA value. Use the helper to handle
both SSA values and register correctly. Otherwise we read trash when we
encounter a register and make bad decisions on types, possibly leading
to our destination being UQ typed when the VGRF is only 32-bit.
Fixes compilation with -Dintel-clc=enabled since 7f6491b76d
(nir: Combine if_uses with instruction uses) but the bug is much older
than that, circa 2017. We were just getting lucky before.
Fixes: 069bf7c907 ("i965/fs: Match destination type to size for ballot")
Reviewed-by: Alyssa Rosenzweig <alyssa@collabora.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/22374>
Every nir_ssa_def is part of a chain of uses, implemented with doubly linked
lists. That means each requires 2 * 64-bit = 16 bytes per def, which is
memory intensive. Together they require 32 bytes per def. Not cool.
To cut that memory use in half, we can combine the two linked lists into a
single use list that contains both regular instruction uses and if-uses. To do
this, we augment the nir_src with a boolean "is_if", and reimplement the
abstract if-uses operations on top of that list. That boolean should fit into
the padding already in nir_src so should not actually affect memory use, and in
the future we sneak it into the bottom bit of a pointer.
However, this creates a new inefficiency: now iterating over regular uses
separate from if-uses is (nominally) more expensive. It turns out virtually
every caller of nir_foreach_if_use(_safe) also calls nir_foreach_use(_safe)
immediately before, so we rewrite most of the callers to instead call a new
single `nir_foreach_use_including_if(_safe)` which predicates the logic based on
`src->is_if`. This should mitigate the performance difference.
There's a bit of churn, but this is largely a mechanical set of changes.
Signed-off-by: Alyssa Rosenzweig <alyssa@collabora.com>
Reviewed-by: Faith Ekstrand <faith.ekstrand@collabora.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/22343>
This pass rarely makes any changes, so work a little harder to preserve
more meta data.
On my Ice Lake laptop (using a locked CPU speed and other measures to
prevent thermal throttling, etc.) using a debugoptimized build, improves
performance of Vulkan CTS "deqp-vk --deqp-case='dEQP-VK.*spir*'" by
-0.2% ± 0.1% (n = 5, pooled s = 0.431885).
v2: Add some parenthesis. Suggested by Lionel.
Reviewed-by: Lionel Landwerlin <lionel.g.landwerlin@intel.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/22299>
Two linked list management changes:
- Use the list head sentinel as the initial cursor. It is, after all, a
proper node in the list.
- Iterate the list of blocks starting with the second block instead of
skipping the first block in the loop.
On my Ice Lake laptop (using a locked CPU speed and other measures to
prevent thermal throttling, etc.) using a release build, improves
performance of compiling shaders from batman_arkham_city_goty.foz by
-0.24% ± 0.09% (n = 5, pooled s = 0.324106).
v2: Use nir_cursor instead of direct list manipultion. Suggested by
Lionel.
Reviewed-by: Lionel Landwerlin <lionel.g.landwerlin@intel.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/22299>
On my Ice Lake laptop (using a locked CPU speed and other measures to
prevent thermal throttling, etc.) using a release build, improves
performance of compiling shaders from batman_arkham_city_goty.foz by
-1.09% ± 0.084% (n = 5, pooled s = 0.354471)
Reduces the size of a release build by 26k.
text data bss dec hex filename
23163641 400720 231360 23795721 16b1809 before/lib64/dri/iris_dri.so
23137264 400720 231360 23769344 16ab100 after/lib64/dri/iris_dri.so
Reviewed-by: Lionel Landwerlin <lionel.g.landwerlin@intel.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/22299>
Since one of the register must always be either VGRF or FIXED_GRF, much
of regions_overlap and reg_offset can be elided.
On my Ice Lake laptop (using a locked CPU speed and other measures to
prevent thermal throttling, etc.) using a debugoptimized build, improves
performance of Vulkan CTS "deqp-vk --deqp-case='dEQP-VK.*spir*'" by
-0.29% ± 0.097% (n = 5, pooled s = 0.361697).
Using a release build, improves performance of compiling shaders from
batman_arkham_city_goty.foz by -3.3% ± 0.04% (n = 5, pooled s =
0.178312).
Reviewed-by: Lionel Landwerlin <lionel.g.landwerlin@intel.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/22299>
This function only exists in builds with assertions, so it only matters
there.
On my Ice Lake laptop (using a locked CPU speed and other measures to
prevent thermal throttling, etc.) using a debugoptimized build, improves
performance of Vulkan CTS "deqp-vk --deqp-case='dEQP-VK.*spir*'" by
-5.2% ± 0.16% (n = 5, pooled s = 0.657887).
Reviewed-by: Lionel Landwerlin <lionel.g.landwerlin@intel.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/22299>
MTL support fp64, but not int64. The fsign(double(x))*FOO optimization
would try to use a 64-bit int xor operation to conditionally toggle
the sign bit off the result.
Since this only affects high bit of the result, we can do a 32-bit
move of the low dword, and a 32-bit xor on the high dword.
Fixes dEQP-VK.spirv_assembly.instruction.compute.float_controls.fp64.input_args.modf_denorm_flush_to_zero
on MTL.
Signed-off-by: Jordan Justen <jordan.l.justen@intel.com>
Reviewed-by: Ian Romanick <ian.d.romanick@intel.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/22259>
Indeed, the parameter "mem_ctx" was not processed.
For instance, this issue is triggered with the crocus driver and
"piglit/bin/shader_runner tests/spec/arb_tessellation_shader/execution/compatibility/tes-clip-vertex-different-from-position.shader_test -auto -fbo":
SUMMARY: AddressSanitizer: 235216 byte(s) leaked in 48 allocation(s).
Fixes: 96ba0344db ("intel: Use common helpers for TCS passthrough shaders")
Signed-off-by: Patrick Lerda <patrick9876@free.fr>
Reviewed-by: Emma Anholt <emma@anholt.net>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/22173>
There are already NIR algebraic optimizations (see also ac6646129f
("nir: Move fsat outside of fmin/fmax if second arg is 0 to 1.") that
will try to remove the saturate from things like
fmax(0.5, fsat(x))
This basically reverts 40aeb558ce ("i965/fs: Allow propagation of
instructions with saturate flag to sel"). That commit message had no
shader-db information, so it's unclear whether this actually helped
anything ever.
No shader-db changes on any Intel platform.
One shader in Far Cry New Dawn was affected.
Cycles in all programs: 10933090738 -> 10933090736 (-0.0%)
Cycles helped: 1
Reviewed-by: Matt Turner <mattst88@gmail.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/22169>
Most tools looking at shader stats assume that there is only a single
resulting binary shader out of a single input. On Intel HW this is not
always the case. So having a statistic on each variant that reports
the maximum dispatch width helps showing improvement on a single
shader in terms of how large we manage to compile it.
For shaders that can be compiled in multiple SIMD width (like fragment
shaders), this will report the maximum dispatch width in the
statistics of each variants.
Signed-off-by: Lionel Landwerlin <lionel.g.landwerlin@intel.com>
Reviewed-by: Sagar Ghuge <sagar.ghuge@intel.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/22014>
This fixes the behavior of register coalesce in cases where the source
of a copy is written elsewhere in the program by a force_writemask_all
instruction, which could cause the overwrite to be executed for an
inactive channel under non-uniform control flow, causing
can_coalesce_vars() to give incorrect results. This has been reported
in cases like:
> while (true) {
> x = imageSize(img);
> if (non_uniform_condition()) {
> y = x;
> break;
> }
> }
> use(y);
Currently the register coalesce pass would coalesce x and y in the
example above, which is invalid since in the example above imageSize()
is implemented as a force_writemask_all SEND message, whose result is
broadcast to all channels, so when a given channel executes 'y = x'
and breaks out of the loop, another divergent channel can execute a
subsequent iteration of the loop overwriting 'x' with a different
value, hence coalescing y and x into the same register changes the
behavior of the program.
Note that this is a regression introduced by commit a4b36cd3dd. In
order to avoid the problem without reverting that patch, we prevent
register coalesce if there is an overwrite of the source with
force_writemask_all behavior inconsistent with the copy and this
occurs anywhere in the intersection of the live ranges of source and
destination, even if it occurs lexically before the copy, since it
might be physically executed after the copy under divergent loop
control flow.
Fixes: a4b36cd3dd ("intel/fs: Coalesce when the src live range is contained in the dst")
Reported-by: Lionel Landwerlin <lionel.g.landwerlin@intel.com>
Reviewed-by: Lionel Landwerlin <lionel.g.landwerlin@intel.com>
Acked-by: Matt Turner <mattst88@gmail.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/21351>
This fixes the behavior of copy propagation in cases where either the
source or destination of an ACP is overwritten elsewhere in the
program by a force_writemask_all instruction, which could cause the
overwrite to be executed for an inactive channel under non-uniform
control flow, causing the current per-channel dataflow propagation to
give incorrect results. This has been reported in cases like:
> while (true) {
> x = imageSize(img);
> if (non_uniform_condition()) {
> y = x;
> break;
> }
> }
> use(y);
Currently the copy propagation pass would propagate copy 'y = x' into
'use(y)', which is invalid since in the example above imageSize() is
implemented as a force_writemask_all SEND message, whose result is
broadcast to all channels, so when a given channel executes 'y = x'
and breaks out of the loop, another divergent channel can execute a
subsequent iteration of the loop overwriting 'x' with a different
value, hence replacing 'y' with 'x' at 'use(y)' changes the behavior
of the program.
This patch extends the global dataflow analysis algorithm to determine
whether there is any control flow path from a given copy to an
overwrite of its source or destination which has force_writemask_all
behavior inconsistent with the copy, and in such case prevents copy
propagation for that ACP entry at any point of the program which can
be reached from the overwrite, even if the copy is statically
re-executed along all such control flow paths (as in the example
above), since the execution of the overwrite for a given channel i may
corrupt other channels j!=i inactive for the subsequently re-executed
copy.
Note that a simpler solution has been attempted which fully shuts down
copy propagation if such a force_writemask_all ACP overwrite is
present /anywhere/ in the program regardless of its location in the
control flow graph, however that led to large shader-db regressions in
some programs from shader-db (like a CS from Car Chase which would
emit 53% more instructions). With this solution the only handful of
shaders that suffer instruction count regressions seem to be getting
misoptimized right now (e.g. some compute shaders from Deus Ex
Mankind). This solution doesn't seem to affect the run-time of
shader-db significantly, it's less than 1% higher with the fix
applied.
Reported-by: Lionel Landwerlin <lionel.g.landwerlin@intel.com>
Reviewed-by: Lionel Landwerlin <lionel.g.landwerlin@intel.com>
Acked-by: Matt Turner <mattst88@gmail.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/21351>
