This one is a bit more complex in that we need to handle 3-source
commutative opcodes. But it's also quite useful:
fossil-db results on Alchemist (A770):
Instrs: 151659750 -> 150164959 (-0.99%); split: -0.99%, +0.01%
Cycles: 12822686329 -> 12574996669 (-1.93%); split: -2.05%, +0.12%
Subgroup size: 7589608 -> 7589592 (-0.00%)
Send messages: 7375047 -> 7375053 (+0.00%); split: -0.00%, +0.00%
Loop count: 46313 -> 46315 (+0.00%); split: -0.01%, +0.01%
Spill count: 110184 -> 54670 (-50.38%); split: -50.79%, +0.41%
Fill count: 213724 -> 104802 (-50.96%); split: -51.43%, +0.47%
Scratch Memory Size: 9406464 -> 3375104 (-64.12%); split: -64.35%, +0.23%
Our older Shadow of the Tomb Raider fossil is particularly helped with
over a 90% reduction in scratch access (spills, fills, and scratch
size). However, benchmarking in the actual game shows no change in
performance. We're thinking the game's shaders have been updated since
our capture.
Ian noted that there was a bug here where we'd accidentally CSE two ADD3
instructions with null destinations and different src[2] that couldn't
be dead code eliminated due to conditional mods. However, this is only
a bug in the new cse_defs pass so we don't need to nominate this for
stable branches.
Reviewed-by: Ian Romanick <ian.d.romanick@intel.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/29848>
This has a number of advantages compared to the pass I wrote years ago:
- It can easily perform either Global CSE or block-local CSE, without
needing to roll any dataflow analysis, thanks to SSA def analysis.
This global CSE is able to detect and coalesce memory loads across
blocks. Although it may increase spilling a little, the reduction
in memory loads seems to more than compensate.
- Because SSA guarantees that values are never written more than once,
the new CSE pass can directly reuse an existing value. The old pass
emitted copies at the point where it discovered a value because it
had no idea whether it'd be mutated later. This led it to generate
a ton of trash for copy propagation to clean up later, and also a
nasty fragility where CSE, register coalescing, and copy propagation
could all fight one another by generating and cleaning up copies,
leading to infinite optimization loops unless we were really careful.
Generating less trash improves our CPU efficiency.
- It uses hash tables like nir_instr_set and nir_opt_cse, instead of
linearly walking lists and comparing each element. This is much more
CPU efficient.
- It doesn't use liveness analysis, which is one of the most expensive
analysis passes that we have. Def analysis is cheaper.
In addition to CSE'ing SSA values, we continue to handle flag writes,
as this is a huge source of CSE'able values. These remain block local.
However, we can simply track the last flag write, rather than creating
entire sets of instruction entries like the old pass. Much simpler.
The only real downside to this pass is that, because the backend is
currently only partially SSA, it has limited visibility and isn't able
to see all values. However, the results appear to be good enough that
the new pass can effectively replace the old pass in almost all cases.
Reviewed-by: Caio Oliveira <caio.oliveira@intel.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/28666>
We were missing the following "newer" fields:
- ex_desc
- predicate_trivial
- sdepth
- rcount
- writes_accumulator
- no_dd_clear
- no_dd_check
- check_tdr
- send_is_volatile
- send_ex_desc_scratch
- send_ex_bso
- last_rt
- keep_payload_trailing_zeroes
- has_packed_lod_ai_src
We can actually just check ex_desc and the new "bits" union to handle
most of them with fewer checks.
Reviewed-by: Caio Oliveira <caio.oliveira@intel.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/29624>
This has no changes in shader-db or fossil-db, surprisingly, but at
least CSEL will be useful shortly. Presumably the others may matter
somewhere.
Reviewed-by: Ian Romanick <ian.d.romanick@intel.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/28971>
We no longer support the old LINE+MAC lowering, and we already lower
this to MAD in NIR on Gfx11+, so the LINTERP virtual opcode always
corresponds the PLN. The only catch is that LINTERP's operands are
reversed from PLN, so we have to switch them.
Reviewed-by: Lionel Landwerlin <lionel.g.landwerlin@intel.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/28705>
This doesn't help very much now. A later commit adds a NIR optimization
pass, tentatively called nir_opt_uniform_subgroup, that converts many
kinds of subgroup operations to things involving
bitCount(ballot(true)). This commit makes a huge difference in the
results of that later commit.
No shader-db changes on any Intel platform.
Fossil-db results:
All Intel platforms had similar results. (Ice Lake shown)
Totals:
Instrs: 165558033 -> 165557519 (-0.00%)
Cycles: 15156188362 -> 15156178922 (-0.00%); split: -0.00%, +0.00%
Totals from 299 (0.05% of 656117) affected shaders:
Instrs: 88293 -> 87779 (-0.58%)
Cycles: 3709498 -> 3700058 (-0.25%); split: -0.28%, +0.03%
v2: Rebase on splitting ELK from BRW. Remove devinfo->ver >= 8 check.
Reviewed-by: Lionel Landwerlin <lionel.g.landwerlin@intel.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/27044>
This will allow fs_builder have a reference to an fs_visitor (a
"fs_shader" really), instead of a reference to a backend_shader.
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/26323>
We haven't exposed this intrinsic as it doesn't directly correspond to
anything in SPIR-V. However, it's used internally by some NIR passes,
namely nir_opt_uniform_atomics().
We reuse most of the infrastructure in brw_find_live_channel, but with
LZD/ADD instead of FBL. A new SHADER_OPCODE_FIND_LAST_LIVE_CHANNEL is
like SHADER_OPCODE_FIND_LIVE_CHANNEL but from the other side.
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>
On Gfx4 and Gfx5, sel.l (for min) and sel.ge (for max) are implemented
using a separte cmpn and sel instruction. This lowering occurs in
fs_vistor::lower_minmax which is called very, very late... a long, long
time after the first calls to opt_cmod_propagation. As a result,
conditional modifiers can be incorrectly propagated across sel.cond on
those platforms.
No tests were affected by this change, and I find that quite shocking.
After just changing flags_written(), all of the atan tests started
failing on ILK. That required the change in cmod_propagatin (and the
addition of the prop_across_into_sel_gfx5 unit test).
Shader-db results for ILK and GM45 are below. I looked at a couple
before and after shaders... and every case that I looked at had
experienced incorrect cmod propagation. This affected a LOT of apps!
Euro Truck Simulator 2, The Talos Principle, Serious Sam 3, Sanctum 2,
Gang Beasts, and on and on... :(
I discovered this bug while working on a couple new optimization
passes. One of the passes attempts to remove condition modifiers that
are never used. The pass made no progress except on ILK and GM45.
After investigating a couple of the affected shaders, I noticed that
the code in those shaders looked wrong... investigation led to this
cause.
v2: Trivial changes in the unit tests.
v3: Fix type in comment in unit tests. Noticed by Jason and Priit.
v4: Tweak handling of BRW_OPCODE_SEL special case. Suggested by Jason.
Fixes: df1aec763e ("i965/fs: Define methods to calculate the flag subset read or written by an fs_inst.")
Reviewed-by: Jason Ekstrand <jason@jlekstrand.net>
Tested-by: Dave Airlie <airlied@redhat.com>
Iron Lake
total instructions in shared programs: 8180493 -> 8181781 (0.02%)
instructions in affected programs: 541796 -> 543084 (0.24%)
helped: 28
HURT: 1158
helped stats (abs) min: 1 max: 1 x̄: 1.00 x̃: 1
helped stats (rel) min: 0.35% max: 0.86% x̄: 0.53% x̃: 0.50%
HURT stats (abs) min: 1 max: 3 x̄: 1.14 x̃: 1
HURT stats (rel) min: 0.12% max: 4.00% x̄: 0.37% x̃: 0.23%
95% mean confidence interval for instructions value: 1.06 1.11
95% mean confidence interval for instructions %-change: 0.31% 0.38%
Instructions are HURT.
total cycles in shared programs: 239420470 -> 239421690 (<.01%)
cycles in affected programs: 2925992 -> 2927212 (0.04%)
helped: 49
HURT: 157
helped stats (abs) min: 2 max: 284 x̄: 62.69 x̃: 70
helped stats (rel) min: 0.04% max: 6.20% x̄: 1.68% x̃: 1.96%
HURT stats (abs) min: 2 max: 48 x̄: 27.34 x̃: 24
HURT stats (rel) min: 0.02% max: 2.91% x̄: 0.31% x̃: 0.20%
95% mean confidence interval for cycles value: -0.80 12.64
95% mean confidence interval for cycles %-change: -0.31% <.01%
Inconclusive result (value mean confidence interval includes 0).
GM45
total instructions in shared programs: 4985517 -> 4986207 (0.01%)
instructions in affected programs: 306935 -> 307625 (0.22%)
helped: 14
HURT: 625
helped stats (abs) min: 1 max: 1 x̄: 1.00 x̃: 1
helped stats (rel) min: 0.35% max: 0.82% x̄: 0.52% x̃: 0.49%
HURT stats (abs) min: 1 max: 3 x̄: 1.13 x̃: 1
HURT stats (rel) min: 0.12% max: 3.90% x̄: 0.34% x̃: 0.22%
95% mean confidence interval for instructions value: 1.04 1.12
95% mean confidence interval for instructions %-change: 0.29% 0.36%
Instructions are HURT.
total cycles in shared programs: 153827268 -> 153828052 (<.01%)
cycles in affected programs: 1669290 -> 1670074 (0.05%)
helped: 24
HURT: 84
helped stats (abs) min: 2 max: 232 x̄: 64.33 x̃: 67
helped stats (rel) min: 0.04% max: 4.62% x̄: 1.60% x̃: 1.94%
HURT stats (abs) min: 2 max: 48 x̄: 27.71 x̃: 24
HURT stats (rel) min: 0.02% max: 2.66% x̄: 0.34% x̃: 0.14%
95% mean confidence interval for cycles value: -1.94 16.46
95% mean confidence interval for cycles %-change: -0.29% 0.11%
Inconclusive result (value mean confidence interval includes 0).
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/12191>
We're about to start using it to implement nir_jump_halt which has
nothing inherently to do with fragment shaders or discards. May as well
name it for the HW instruction it generates.
Reviewed-by: Francisco Jerez <currojerez@riseup.net>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/5071>
This involves wrapping fs_live_variables in a BRW_ANALYSIS object and
hooking it up to invalidate_analysis() so it's properly invalidated.
Seems like a lot of churn but it's fairly straightforward. The
fs_visitor invalidate_ and calculate_live_intervals() methods are no
longer necessary after this change.
Reviewed-by: Matt Turner <mattst88@gmail.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/4012>
This moves the following methods that are currently defined in
fs_visitor (even though they are side products of the liveness
analysis computation) and are already implemented in
brw_fs_live_variables.cpp:
> bool virtual_grf_interferes(int a, int b) const;
> int *virtual_grf_start;
> int *virtual_grf_end;
It makes sense for them to be part of the fs_live_variables object,
because they have the same lifetime as other liveness analysis results
and because this will allow some extra validation to happen wherever
they are accessed in order to make sure that we only ever use
up-to-date liveness analysis results.
This shortens the virtual_grf prefix in order to compensate for the
slightly increased lexical overhead from the live_intervals pointer
dereference.
Reviewed-by: Matt Turner <mattst88@gmail.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/4012>
Have fun reading through the whole back-end optimizer to verify
whether I've missed any dependency flags -- Or alternatively, just
trust that any mistake here will trigger an assertion failure during
analysis pass validation if it ever poses a problem for the
consistency of any of the analysis passes managed by the framework.
Reviewed-by: Matt Turner <mattst88@gmail.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/4012>
The invalidate_analysis() method knows what analysis passes there are
in the back-end and calls their invalidate() method to report changes
in the IR. For the moment it just calls invalidate_live_intervals()
(which will eventually be fully replaced by this function) if anything
changed.
This makes all optimization passes invalidate DEPENDENCY_EVERYTHING,
which is clearly far from ideal -- The dependency classes passed to
invalidate_analysis() will be refined in a future commit.
Reviewed-by: Matt Turner <mattst88@gmail.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/4012>
Found by inspection. This seems particularly likely to cause problems
with instructions dependent on the current execution mask like
SHADER_OPCODE_FIND_LIVE_CHANNEL or the FS_OPCODE_LOAD_LIVE_CHANNELS
instruction I'm about to introduce, but one could imagine it leading
to data corruption if CSE ever managed to combine two instructions
before and after the FS_OPCODE_PLACEHOLDER_HALT, since the one before
may not be executed for some channels.
Reviewed-by: Caio Marcelo de Oliveira Filho <caio.oliveira@intel.com>
Cc: 20.0 <mesa-stable@lists.freedesktop.org>
This reworks the current fs_inst::is_copy_payload() method into a
number of classification helpers with well-defined semantics. This
will be useful later on in order to optimize LOAD_PAYLOAD instructions
more aggressively in cases where we can determine it's safe to do so.
The closest equivalent of the present fs_inst::is_copy_payload()
method is the is_coalescing_payload() helper introduced here.
No functional nor shader-db changes.
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
The liveness calculation done by the local CSE pass in order to prune
AEB entries whose sources are no longer live is currently inaccurate,
because the live intervals are calculated once at the beginning of the
pass, so they don't take into account any of the copy instructions
inserted by the CSE pass as it makes progress. However the IP counter
used in that calculation is based on the start_ip of the basic block,
which is updated automatically whenever any instructions are inserted
into the CFG. This causes the IP counter and liveness intervals to
get out of sync in programs with multiple basic blocks, causing the
CSE pass to toss AEB entries prematurely, which can lead to missed
optimization opportunities rather non-deterministically.
On BDW this leads to the following shader-db changes:
total instructions in shared programs: 14952488 -> 14951763 (-0.00%)
instructions in affected programs: 45416 -> 44691 (-1.60%)
helped: 40
HURT: 4
total spills in shared programs: 20989 -> 20970 (-0.09%)
spills in affected programs: 103 -> 84 (-18.45%)
helped: 3
HURT: 0
total fills in shared programs: 24981 -> 24926 (-0.22%)
fills in affected programs: 127 -> 72 (-43.31%)
helped: 3
HURT: 0
In addition it avoids a number of regressions in combination with some
of the optimization changes I'm working on for SIMD32, which would
have made CSE more effective... Causing it to be less effective
elsewhere in the program astonishingly.
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
This reverts commit 40b3abb4d1.
It is not clear that this commit was entirely correct, and unfortunately
it was pushed by error.
CC: Jason Ekstrand <jason@jlekstrand.net>
Acked-by: Jason Ekstrand <jason@jlekstrand.net>
This function is used in two different scenarios that for 32-bit
instructions are the same, but for 16-bit instructions are not.
One scenario is that in which we are working at a SIMD8 register
level and we need to know if a register is fully defined or written.
This is useful, for example, in the context of liveness analysis or
register allocation, where we work with units of registers.
The other scenario is that in which we want to know if an instruction
is writing a full scalar component or just some subset of it. This is
useful, for example, in the context of some optimization passes
like copy propagation.
For 32-bit instructions (or larger), a SIMD8 dispatch will always write
at least a full SIMD8 register (32B) if the write is not partial. The
function is_partial_write() checks this to determine if we have a partial
write. However, when we deal with 16-bit instructions, that logic disables
some optimizations that should be safe. For example, a SIMD8 16-bit MOV will
only update half of a SIMD register, but it is still a complete write of the
variable for a SIMD8 dispatch, so we should not prevent copy propagation in
this scenario because we don't write all 32 bytes in the SIMD register
or because the write starts at offset 16B (wehere we pack components Y or
W of 16-bit vectors).
This is a problem for SIMD8 executions (VS, TCS, TES, GS) of 16-bit
instructions, which lose a number of optimizations because of this, most
important of which is copy-propagation.
This patch splits is_partial_write() into is_partial_reg_write(), which
represents the current is_partial_write(), useful for things like
liveness analysis, and is_partial_var_write(), which considers
the dispatch size to check if we are writing a full variable (rather
than a full register) to decide if the write is partial or not, which
is what we really want in many optimization passes.
Then the patch goes on and rewrites all uses of is_partial_write() to use
one or the other version. Specifically, we use is_partial_var_write()
in the following places: copy propagation, cmod propagation, common
subexpression elimination, saturate propagation and sel peephole.
Notice that the semantics of is_partial_var_write() exactly match the
current implementation of is_partial_write() for anything that is
32-bit or larger, so no changes are expected for 32-bit instructions.
Tested against ~5000 tests involving 16-bit instructions in CTS produced
the following changes in instruction counts:
Patched | Master | % |
================================================
SIMD8 | 621,900 | 706,721 | -12.00% |
================================================
SIMD16 | 93,252 | 93,252 | 0.00% |
================================================
As expected, the change only affects SIMD8 dispatches.
Reviewed-by: Topi Pohjolainen <topi.pohjolainen@intel.com>
By just assigning dst.type to src[i].type, we ensure that the offset at
the end of the loop actually offsets it by the right number of
registers. Otherwise, we'll get into a case where we copy with a Q type
and then offset with a D type and things get out of sync.
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
Previously, we tried to combine all cases where the instruction being
CSE'd writes to more than one MOV worth of registers into one case with
a bit of special casing for LOAD_PAYLOAD. This commit splits things so
that LOAD_PAYLOAD is entirely it's own case. This makes tweaking the
LOAD_PAYLOAD case simpler in the next commit.
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
The only reason it was it's own opcode was so that we could detect it
and adjust the source register based on the payload setup. Now that
we're using the ATTR file for FS inputs, there's no point in having a
magic opcode for this.
v2 (Jason Ekstrand):
- Break the bit which removes the CINTERP opcode into its own patch
Reviewed-by: Jason Ekstrand <jason@jlekstrand.net>
Reviewed-by: Matt Turner <mattst88@gmail.com>
Mostly a dummy git mv with a couple of noticable parts:
- With the earlier header cleanups, nothing in src/intel depends
files from src/mesa/drivers/dri/i965/
- Both Autoconf and Android builds are addressed. Thanks to Mauro and
Tapani for the fixups in the latter
- brw_util.[ch] is not really compiler specific, so it's moved to i965.
v2:
- move brw_eu_defines.h instead of brw_defines.h
- remove no-longer applicable includes
- add missing vulkan/ prefix in the Android build (thanks Tapani)
v3:
- don't list brw_defines.h in src/intel/Makefile.sources (Jason)
- rebase on top of the oa patches
[Emil Velikov: commit message, various small fixes througout]
Signed-off-by: Emil Velikov <emil.velikov@collabora.com>
Reviewed-by: Jason Ekstrand <jason@jlekstrand.net>
2017-03-13 11:16:34 +00:00
Renamed from src/mesa/drivers/dri/i965/brw_fs_cse.cpp (Browse further)
