| Commit message (Collapse) | Author | Age | Files | Lines |
|
|
|
| |
Reviewed-by: Iago Toral Quiroga <itoral@igalia.com>
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
This is in preparation for dropping fs_inst::regs_read and
::regs_written in favor of more accurate alternatives expressed in
byte units. The main reason these wrappers are useful is that a
number of optimization passes implement dataflow analysis with
register granularity, so these helpers will come in handy once we've
switched register offsets and sizes to the byte representation. The
wrapper functions will also make sure that GRF misalignment (currently
neglected by most of the back-end) is taken into account correctly in
the calculation of regs_read and regs_written.
Reviewed-by: Iago Toral Quiroga <itoral@igalia.com>
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
The fs_reg::offset field in byte units introduced in this patch is a
more straightforward alternative to the current register offset
representation split between fs_reg::reg_offset and ::subreg_offset.
The split representation makes it too easy to forget about one of the
offsets while dealing with the other, which has led to multiple
back-end bugs in the past. To make the matter worse the unit
reg_offset was expressed in was rather inconsistent, for uniforms it
would be expressed in either 4B or 16B units depending on the
back-end, and for most other things it would be expressed in 32B
units.
This encodes reg_offset as a new offset field expressed consistently
in byte units. Each rvalue reference of reg_offset in existing code
like 'x = r.reg_offset' is rewritten to 'x = r.offset / reg_unit', and
each lvalue reference like 'r.reg_offset = x' is rewritten to
'r.offset = r.offset % reg_unit + x * reg_unit'.
Because the change affects a lot of places and is rather non-trivial
to verify due to the inconsistent value of reg_unit, I've tried to
avoid making any additional changes other than applying the rewrite
rule above in order to keep the patch as simple as possible, sometimes
at the cost of introducing obvious stupidity (e.g. algebraic
expressions that could be simplified given some knowledge of the
context) -- I'll clean those up later on in a second pass.
Reviewed-by: Iago Toral Quiroga <itoral@igalia.com>
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
This is required for correctness in presence of multiple 8-wide flag
writes (e.g. 8-wide instructions with a conditional mod set) which
update a different portion of the same 16-bit flag subregister. Right
now we keep track of flag dataflow with 16-bit granularity and
consider flag writes to have killed any previous definition of the
same subregister even if the write was less than 16 channels wide,
which can cause live flag register updates to be dead code-eliminated
incorrectly.
Additionally this makes sure that we handle 32-wide flag writes and
reads which may span multiple flag subregisters so the current
approach of just setting/testing a single bit from the live set
wouldn't have worked.
Reviewed-by: Jason Ekstrand <jason@jlekstrand.net>
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
When working on tessellation shaders, I created some vec4 virtual
opcodes for creating message headers through a sequence like:
mov(8) g7<1>UD 0x00000000UD { align1 WE_all 1Q compacted };
mov(1) g7.5<1>UD 0x00000100UD { align1 WE_all };
mov(1) g7<1>UD g0<0,1,0>UD { align1 WE_all compacted };
mov(1) g7.3<1>UD g8<0,1,0>UD { align1 WE_all };
This is done in the generator since the vec4 backend can't handle align1
regioning. From the visitor's point of view, this is a single opcode:
hs_set_output_urb_offsets vgrf7.0:UD, 1U, vgrf8.xxxx:UD
Normally, there's no hazard between sources and destinations - an
instruction (naturally) reads its sources, then writes the result to the
destination. However, when the virtual instruction generates multiple
hardware instructions, we can get into trouble.
In the above example, if the register allocator assigned vgrf7 and vgrf8
to the same hardware register, then we'd clobber the source with 0 in
the first instruction, and read back the wrong value in the last one.
It occured to me that this is exactly the same problem we have with
SIMD16 instructions that use W/UW or B/UB types with 0 stride. The
hardware implicitly decodes them as two SIMD8 instructions, and with
the overlapping regions, the first would clobber the second.
Previously, we handled that by incrementing the live range end IP by 1,
which works, but is excessive: the next instruction doesn't actually
care about that. It might also be the end of control flow. This might
keep values alive too long. What we really want is to say "my source
and destinations interfere".
This patch creates new infrastructure for doing just that, and teaches
the register allocator to add interference when there's a hazard. For
my vec4 case, we can determine this by switching on opcodes. For the
SIMD16 case, we just move the existing code there.
I audited our existing virtual opcodes that generate multiple
instructions; I believe FS_OPCODE_PACK_HALF_2x16_SPLIT needs this
treatment as well, but no others.
v2: Rebased by mattst88.
Signed-off-by: Kenneth Graunke <kenneth@whitecape.org>
Reviewed-by: Matt Turner <mattst88@gmail.com>
|
|
|
|
|
|
|
|
|
|
| |
The 2-bit hardware register file field is ARF, GRF, MRF, IMM.
Rename GRF to VGRF (virtual GRF) so that we can reuse the GRF name to
mean an assigned general purpose register.
Reviewed-by: Emil Velikov <emil.velikov@collabora.co.uk>
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
|
|
|
|
| |
Reviewed-by: Francisco Jerez <currojerez@riseup.net>
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
From Muchnick's Advanced Compiler Design and Implementation:
"To determine which variables are live at each point in a flowgraph, we
perform a backward data-flow analysis"
Previously, we were walking the blocks forwards and updating the livein and
then the liveout. However, the livein calculation depends on the liveout
and the liveout depends on the successor blocks. The net result is that it
takes one full iteration to go from liveout to livein and then another
full iteration to propagate to the predecessors. This works out to an
O(n^2) computation where n is the number of blocks. If we run things in
the other order, it's O(nl) where l is the maximum loop depth which is
practically bounded by 3.
On my HSW desktop, one particular shadertoy test gets a 20% improvement in
compile times:
N Min Max Median Avg Stddev
x 10 15.965 16.884 16.026 16.1822 0.34736846
+ 10 12.813 13.052 12.876 12.8891 0.06913666
Difference at 95.0% confidence
-3.2931 +/- 0.235316
-20.3501% +/- 1.45417%
(Student's t, pooled s = 0.250444)
Reviewed-by: Matt Turner <mattst88@gmail.com>
|
|
|
|
|
|
| |
The fs_visitor argument of fs_inst::regs_read() wasn't used at all.
Reviewed-by: Matt Turner <mattst88@gmail.com>
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
Right now virtual GRF book-keeping and allocation is performed in each
visitor class separately (among other hundred different things),
leading to duplicated logic in each visitor and preventing layering as
it forces any code that manipulates i965 IR and needs to allocate
virtual registers to depend on the specific visitor that happens to be
used to translate from GLSL IR.
v2: Use realloc()/free() to allocate VGRF book-keeping arrays (Connor).
Reviewed-by: Matt Turner <mattst88@gmail.com>
|
|
|
|
|
|
| |
Also while we're touching var_from_reg, just make it an inline function.
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
|
|
|
|
| |
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
|
|
|
|
|
|
|
| |
The next patch will be simplified because of this, and makes reading the
code a lot easier.
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
|
|
|
|
|
| |
Signed-off-by: Jason Ekstrand <jason.ekstrand@intel.com>
Reviewed-by: Matt Turner <mattst88@gmail.com>
|
|
|
|
|
| |
Signed-off-by: Jason Ekstrand <jason.ekstrand@intel.com>
Reviewed-by: Matt Turner <mattst88@gmail.com>
|
|
|
|
|
|
|
| |
Now that nothing invalidates the CFG, we can calculate_cfg() immediately
after emit_fb_writes()/emit_thread_end() and never again.
Reviewed-by: Topi Pohjolainen <topi.pohjolainen@intel.com>
|
|
|
|
|
|
| |
Everything has been converted to preserve the CFG.
Reviewed-by: Topi Pohjolainen <topi.pohjolainen@intel.com>
|
|
|
|
|
|
|
| |
Will let us avoid invalidating the CFG if the optimization pass has
removed instructions using the new basic block methods.
Reviewed-by: Topi Pohjolainen <topi.pohjolainen@intel.com>
|
|
|
|
|
| |
Use this as an opportunity to rename 'block_num' to 'num'. block->num is
clear, and block->block_num has always been redundant.
|
|
|
|
| |
Reviewed-by: Topi Pohjolainen <topi.pohjolainen@intel.com>
|
|
|
|
|
|
|
| |
We've often created the CFG immediately before, so use it when
available.
Reviewed-by: Ian Romanick <ian.d.romanick@intel.com>
|
|
|
|
| |
Reviewed-by: Ian Romanick <ian.d.romanick@intel.com>
|
|
|
|
|
|
| |
Reviewed-by: Chris Forbes <chrisf@ijw.co.nz>
Reviewed-by: Tapani Pälli <tapani.palli@intel.com>
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
|
|
|
|
|
|
|
| |
In order to remove bblock_link's inheritance of exec_node. Also makes
linked list walk code much nicer.
Acked-by: Eric Anholt <eric@anholt.net>
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
This is the only case where a fs_reg in brw_fs_visitor is used during
optimization/code generation, and it meant that optimizations had to be
careful to not move pixel_x/y's register number without updating it.
Additionally, it turns out we had a couple of other UW values that weren't
getting this treatment (like gl_SampleID), so this more general fix is
probably a good idea (though I wasn't able to replicate problems with
either pixel_[xy]'s values or gl_SampleID, even when telling the register
allocator to reuse registers immediately)
Reviewed-by: Matt Turner <mattst88@gmail.com>
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
|
|
|
|
|
|
|
|
|
|
| |
When you've got a simple solid-color shader that doesn't generate
pixel_x/y interpolation, we were deciding that the first vgrf was both the
undefined pixel_x and pixel_y, and extending its live interval to avoid
the stride problem. That tricked other optimization that tries to see if
a particular instruction is the last use of a variable.
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
|
|
|
|
|
|
|
|
|
| |
The same effect can be achieved using a combination of ::stride and
::subreg_offset. Remove the less flexible ::smear to keep the data
members of fs_reg orthogonal.
Reviewed-by: Matt Turner <mattst88@gmail.com>
Reviewed-by: Paul Berry <stereotype441@gmail.com>
|
|
|
|
|
|
|
|
|
|
|
| |
v2: Some improvements for copy propagation with non-contiguous
register strides and mismatching types.
v3: Add example of the situation that the copy propagation changes are
intended to avoid. Clarify that 'fs_reg::apply_stride()' is expected
to work with zero strides too.
Reviewed-by: Matt Turner <mattst88@gmail.com>
Reviewed-by: Paul Berry <stereotype441@gmail.com>
|
|
|
|
|
|
| |
Helped to track down a problem in a version of the next commit.
Reviewed-by: Jordan Justen <jordan.l.justen@intel.com>
|
|
|
|
|
|
|
| |
mov takes only a single source argument. Example instruction
inexplicably changed from add to mov in commit f10f5e49.
Reviewed-by: Jordan Justen <jordan.l.justen@intel.com>
|
|
|
|
|
|
|
| |
parent_mem_ctx was unused since db47074a, so remove the two wrappers
around create() and make create() the constructor.
Reviewed-by: Eric Anholt <eric@anholt.net>
|
|
|
|
|
|
|
|
|
| |
These ralloc contexts belong to a specific object and are being
deallocated manually from the class destructor. Now that we've hooked
up destructors to ralloc there's no reason for them to be children of
any other context, and doing so might to lead to double frees under
some circumstances. The class destructor has all the responsibility
of freeing class memory resources now.
|
|
|
|
| |
Reviewed-by: Eric Anholt <eric@anholt.net>
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
Looking at Lightsmark's shaders, the way we used MRFs (or in gen7's
case, GRFs) was bad in a couple of ways. One was that it prevented
compute-to-MRF for the common case of a texcoord that gets used
exactly once, but where the texcoord setup all gets emitted before the
texture calls (such as when it's a bare fragment shader input, which
gets interpolated before processing main()). Another was that it
introduced a bunch of dependencies that constrained scheduling, and
forced waits for texture operations to be done before they are
required. For example, we can now move the compute-to-MRF
interpolation for the second texture send down after the first send.
The downside is that this generally prevents
remove_duplicate_mrf_writes() from doing anything, whereas previously
it avoided work for the case of sampling from the same texcoord twice.
However, I suspect that most of the win that originally justified that
code was in avoiding the WAR stall on the first send, which this patch
also avoids, rather than the small cost of the extra instruction. We
see instruction count regressions in shaders in unigine, yofrankie,
savage2, hon, and gstreamer.
Improves GLB2.7 performance by 0.633628% +/- 0.491809% (n=121/125, avg of
~66fps, outliers below 61 dropped).
Improves openarena performance by 1.01092% +/- 0.66897% (n=425).
No significant difference on Lightsmark (n=44).
v2: Squash in the fix for register unspilling for send-from-GRF, fixing a
segfault in lightsmark.
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
Acked-by: Matt Turner <mattst88@gmail.com>
|
|
|
|
|
|
|
|
|
|
| |
This will let us coalesce into texture-from-GRF arguments, which would
otherwise be prevented due to the live interval for the whole vgrf
extending across all the MOVs setting up the channels of the message
v2 (Kenneth Graunke): Rebase for renames.
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
|
|
|
|
|
|
|
|
|
|
|
| |
Now optimization passes will be able to look at the per-channel ranges.
v2: Rebase on various optimization pass changes.
v3 (Kenneth Graunke): Rename live_variables to live_intervals; split
introduction of invalidate_live_intervals() into a separate patch.
Signed-off-by: Kenneth Graunke <kenneth@whitecape.org>
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
In compute_live_intervals(), start and end are shorter names for
the virtual_grf_start and virtual_grf_end class members.
Now that the fs_live_intervals class has arrays named start and end
which are indexed by var, rather than VGRF, reusing the name is
confusing. Plus, most of the code has been factored out, so using the
long names isn't as inconvenient.
Signed-off-by: Kenneth Graunke <kenneth@whitecape.org>
Reviewed-by: Eric Anholt <eric@anholt.net>
|
|
|
|
|
|
|
|
|
|
|
| |
This is the information we'll actually use to replace the
virtual_grf_start[]/end[] arrays.
No change in shader-db.
v2 (Kenneth Graunke): Rebase; minor comment updates.
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
|
|
|
|
|
|
|
|
|
| |
These blocks are about to grow some more code, and the indentation was
getting out of hand.
v2 (Kenneth Graunke): Rebase, minor typo fixes and style changes.
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
|
|
|
|
|
|
|
|
|
|
| |
For now, this simply sets live_intervals_valid = false, but in the
future it will do something more sophisticated.
Based on a patch by Eric Anholt.
Signed-off-by: Kenneth Graunke <kenneth@whitecape.org>
Reviewed-by: Eric Anholt <eric@anholt.net>
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
This significantly improves our handling of VGRFs of size > 1.
Previously, we only marked VGRFs as def'd if the whole register was
written by a single instruction. Large VGRFs which were written
piecemeal would not be considered def'd at all, even if they were
ultimately completely written.
Without being def'd, these were then marked "live in" to the basic
block, often extending the range to preceding blocks and sometimes
even the start of the program.
The new per-component tracking gives more accurate live intervals,
which makes register coalescing more effective.
In the future, this should help with texturing from GRFs on Gen7+.
A sampler message might be represented by a 2-register VGRF which
holds the texture coordinates. If those are incoming varyings,
they'll be produced by two PLN instructions, which are piecemeal writes.
No reduction in shader-db instruction counts. However, code which
prints the live interval ranges does show that some VGRFs now have
smaller (and more correct) live intervals.
v2: Rebase on current send-from-GRF code requiring adding extra use[]s.
v3: Rebase on live intervals fix to include defs in the end of the
interval.
v4 (Kenneth Graunke): Rebase; split off a few preparatory patches;
add lots of comments; minor style changes; rewrite commit message.
v5 (Eric Anholt): whitespace nit.
Written-by: Eric Anholt <eric@anholt.net> [v1-3]
Signed-off-by: Kenneth Graunke <kenneth@whitecape.org> [v4]
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
Reviewed-by: Eric Anholt <eric@anholt.net> (v4)
|
|
|
|
|
|
|
|
|
| |
num_vars was shorthand for the number of virtual GRFs. num_vgrfs is a
bit clearer. Plus, the next patch will introduce "vars" which are
distinct from vgrfs.
Signed-off-by: Kenneth Graunke <kenneth@whitecape.org>
Reviewed-by: Eric Anholt <eric@anholt.net>
|
|
|
|
|
|
|
|
| |
This has no functional effect, but should make subsequent changes a
little simpler.
Signed-off-by: Kenneth Graunke <kenneth@whitecape.org>
Reviewed-by: Eric Anholt <eric@anholt.net>
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
We were expanding the live range too far, breaking register_coalesce_2()
and compute_to_mrf() on 16-wide shaders. Turning it back on improves
GLB2.7 performance by 0.239355% +/- 0.0850649% (n=398). shader-db stats
are:
total instructions in shared programs: 1627211 -> 1609262 (-1.10%)
instructions in affected programs: 450351 -> 432402 (-3.99%)
While 33 new 16-wide shaders are gained, 70 are lost. Despite that,
tropics (the app that lost the most 16-wide) shows a .41% +/- .16%
(n=7/8, first-run outlier removed) performance improvement on my HSW.
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
Previously, we would sometimes not consider a write to a register to
extend the end of the interval, nor would we consider a read before a
write to extend the start. This made for a bunch of complicated logic
related to how to treat the results when dead code might be present.
Instead, just extend the interval and fix dead code elimination to know
how to remove it.
Interestingly, this actually results in a tiny bit more optimization:
total instructions in shared programs: 1391220 -> 1390799 (-0.03%)
instructions in affected programs: 14037 -> 13616 (-3.00%)
v2: Fix a theoretical problem with the simd16 workaround if dst == src,
where we would revert the bump of the live range.
Reviewed-by: Ian Romanick <ian.d.romanick@intel.com> (v1)
|
|
|
|
|
|
|
|
| |
These checks were all over, and every time I wrote one I had to try to
decide again what the cases were for partial updates.
v2: Fix inadvertent reladdr check removal.
Reviewed-by: Matt Turner <mattst88@gmail.com>
|
|
|
|
| |
Reviewed-by: Matt Turner <mattst88@gmail.com>
|
|
|
|
|
|
|
|
|
| |
For sampler messages, it depends on the target gen, and on gen4
SIMD16-sampler-on-SIMD8-execution we were returning 4 instead of 8 like we
should.
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
NOTE: This is a candidate for the 9.1 branch.
|
|
|
|
|
|
|
|
|
|
|
|
| |
We can execute way fewer instructions by doing our boolean manipulation
on an "int" of bits at a time, while also reducing our working set size.
Reduces compile time of L4D2's slowest shader from 4s to 1.1s
(-72.4% +/- 0.2%, n=10)
v2: Remove redundant masking (noted by Ken)
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
We were allowing a compressed instruction to write a register that
contained the last use of a uniform pull constant (either UBO load or push
constant spillover), so it would get half its values smashed.
Since we need to see the actual instruction to decide this, move the
pre-gen6 pixel_x/y logic here, which should improve the performance of
register allocation since virtual_grf_interferes() is called more than
once per instruction.
NOTE: This is a candidate for the stable branches.
Bugzilla: https://bugs.freedesktop.org/show_bug.cgi?id=61317
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
|