These are basically just moves, so they should be safe as well.
When disabling i965's GLSL IR level scalarizer (channel expressions)
pass, I started seeing NIR code like this:
if ssa_21 {
block block_1:
/* preds: block_0 */
vec4 ssa_120 = vec4 ssa_82, ssa_83, ssa_84, ssa_30
/* succs: block_3 */
} else {
block block_2:
/* preds: block_0 */
/* succs: block_3 */
}
block block_3:
/* preds: block_1 block_2 */
vec4 ssa_33 = phi block_1: ssa_120, block_2: ssa_2
Previously, the GLSL IR scalarizer pass would break the vec4 into a
series of fmovs, which were allowed by the peephole pass. But with
the vec4 operation, they were not. We want to keep getting selects.
Normal i965 on Broadwell:
instructions in affected programs: 200 -> 176 (-12.00%)
helped: 4
With brw_fs_channel_expressions() disabled:
instructions in affected programs: 1832 -> 1646 (-10.15%)
helped: 30
Signed-off-by: Kenneth Graunke <kenneth@whitecape.org>
Reviewed-by: Jason Ekstrand <jason.ekstrand@intel.com>
Reviewed-by: Connor Abbott <cwabbott0@gmail.com>
lower_phis_to_scalar() pass recurses the instruction dependence graph to
determine if all the sources of a given instruction are scalarizable.
To prevent cycles, it temporary marks the phi instruction before recursing in,
then updates the entry with the resulting value. However, it does not consider
that the entry value may have changed after a recursion pass, hence causing
a use-after-free situation and a crash.
This patch fixes this by reloading the entry corresponding to the 'phi'
after recursing and before updating its value.
The crash can be reproduced ~20% of times with the dEQP test:
dEQP-GLES3.functional.shaders.loops.while_constant_iterations.nested_sequence_fragment
Reviewed-by: Jason Ekstrand <jason.ekstrand@intel.com>
When we compute the output swizzle we want to consider the number of
components in the add operation. So far we were using the writemask
of the multiplication for this instead, which is not correct.
Reviewed-by: Jason Ekstrand <jason.ekstrand@intel.com>
Some shaders in Civilization V and Beyond Earth do
pow(pow(x, 2.2), 0.454545)
which is converting to and from sRGB colorspace.
A more general rule that replaces pow(pow(a, b), c) with pow(a, b * c)
actually regresses two shaders in Sun Temple in which the result of the
inner pow is used twice, once by another pow and once by another
instruction. Also, since 2.2 * 0.454545 isn't exactly one, the more
general pattern would have still left us with a pow, and I'm 2.2 *
0.454545 percent sure that's not what they want.
instructions in affected programs: 934 -> 886 (-5.14%)
helped: 16
Previously, we used intrinsic->const_index[1] to represent "the number of
array elements to load" for load/store intrinsics. However, this set to 1
by every pass that ever creates a load/store intrinsic. Also, while it
might make some sense for registers, it makes no sense whatsoever in SSA.
On top of that, the i965 backend was the only backend to ever support it;
freedreno and vc4 just assert that it's always 1. Let's just delete it.
Signed-off-by: Jason Ekstrand <jason.ekstrand@intel.com>
Reviewed-by: Connor Abbott <cwabbott0@gmail.com>
Reviewed-by: Rob Clark <robclark@freedesktop.org>
This fixes bugs with special cases where we have arrays of
structures containing samplers or arrays of samplers.
I've verified that patch results in calculating same index value as
returned by _mesa_get_sampler_uniform_value for IR. Patch makes
following ES3 conformance test pass:
ES3-CTS.shaders.struct.uniform.sampler_array_fragment
v2: remove unnecessary comment (Topi)
simplify changes and the overall code (Jason)
Signed-off-by: Tapani Pälli <tapani.palli@intel.com>
Reviewed-by: Jason Ekstrand <jason.ekstrand@intel.com>
Bugzilla: https://bugs.freedesktop.org/show_bug.cgi?id=90114
Previously, this case was being handled in match_expression prior to
calling match_value. However, there is really no good reason for this
given that match_value has all of the information it needs. Also, they
weren't being handled properly in the commutative case and putting it in
match_value gives us that for free.
Reviewed-by: Connor Abbott <cwabbott0@gmail.com>
This commit switches us from the current setup of using hash sets for
use/def sets to using linked lists. Doing so should save us quite a bit of
memory because we aren't carrying around 3 hash sets per register and 2 per
SSA value. It should also save us CPU time because adding/removing things
from use/def sets is 4 pointer manipulations instead of a hash lookup.
Running shader-db 50 times with USE_NIR=0, NIR, and NIR + use/def lists:
GLSL IR Only: 586.4 +/- 1.653833
NIR with hash sets: 675.4 +/- 2.502108
NIR + use/def lists: 641.2 +/- 1.557043
I also ran a memory usage experiment with Ken's patch to delete GLSL IR and
keep NIR. This patch cuts an aditional 42.9 MiB of ralloc'd memory over
and above what we gained by deleting the GLSL IR on the same dota trace.
On the code complexity side of things, some things are now much easier and
others are a bit harder. One of the operations we perform constantly in
optimization passes is to replace one source with another. Due to the fact
that an instruction can use the same SSA value multiple times, we had to
iterate through the sources of the instruction and determine if the use we
were replacing was the only one before removing it from the set of uses.
With this patch, uses are per-source not per-instruction so we can just
remove it safely. On the other hand, trying to iterate over all of the
instructions that use a given value is more difficult. Fortunately, the
two places we do that are the ffma peephole where it doesn't matter and GCM
where we already gracefully handle duplicates visits to an instruction.
Another aspect here is that using linked lists in this way can be tricky to
get right. With sets, things were quite forgiving and the worst that
happened if you didn't properly remove a use was that it would get caught
in the validator. With linked lists, it can lead to linked list corruption
which can be harder to track. However, we do just as much validation of
the linked lists as we did of the sets so the validator should still catch
these problems. While working on this series, the vast majority of the
bugs I had to fix were caught by assertions. I don't think the lists are
going to be that much worse than the sets.
Reviewed-by: Connor Abbott <cwabbott0@gmail.com>
We were rolling our own rewrite_src variant in copy-propagation. Let's
stop doing that and use the ones in core NIR.
Reviewed-by: Connor Abbott <cwabbott0@gmail.com>
The out-of-SSA pass was one of the first passes written when getting SSA
up-and-going (for obvious reasons). As such, it came before a lot of the
nifty SSA-based helpers were introduced. This commit modernizes it so that
we're no longer doing nearly as much manual banging on use/def sets.
Reviewed-by: Connor Abbott <cwabbott0@gmail.com>
Nothing produces it, and nothing can consume it.
Signed-off-by: Ian Romanick <ian.d.romanick@intel.com>
Reviewed-by: Matt Turner <mattst88@gmail.com>
Acked-by: Jason Ekstrand <jason.ekstrand@intel.com>
All paths that produce GLSL IR for NIR lower ir_unop_log. All paths
that consume NIR will explode if they geta nir_op_flog.
Signed-off-by: Ian Romanick <ian.d.romanick@intel.com>
Reviewed-by: Matt Turner <mattst88@gmail.com>
Acked-by: Jason Ekstrand <jason.ekstrand@intel.com>
Nothing produces it, and nothing can consume it.
Signed-off-by: Ian Romanick <ian.d.romanick@intel.com>
Reviewed-by: Matt Turner <mattst88@gmail.com>
Acked-by: Jason Ekstrand <jason.ekstrand@intel.com>
All paths that produce GLSL IR for NIR lower ir_unop_exp. All paths
that consume NIR will explode if they geta nir_op_fexp.
Signed-off-by: Ian Romanick <ian.d.romanick@intel.com>
Reviewed-by: Matt Turner <mattst88@gmail.com>
Acked-by: Jason Ekstrand <jason.ekstrand@intel.com>
... and (a >= c) || (b >= c) as max(a, b) >= c.
Similar to commit 97e6c1b9.
total instructions in shared programs: 6182276 -> 6182180 (-0.00%)
instructions in affected programs: 6400 -> 6304 (-1.50%)
helped: 68
HURT: 4
Reviewed-by: Jordan Justen <jordan.l.justen@intel.com>
Reviewed-by: Glenn Kennard <glenn.kennard@gmail.com>
No changes, but does prevent some regressions in the next commit.
Reviewed-by: Jordan Justen <jordan.l.justen@intel.com>
Reviewed-by: Glenn Kennard <glenn.kennard@gmail.com>
Helps the same set of programs as the previous commit.
instructions in affected programs: 4490 -> 4346 (-3.21%)
helped: 8
Reviewed-by: Jordan Justen <jordan.l.justen@intel.com>
Reviewed-by: Glenn Kennard <glenn.kennard@gmail.com>
Four shaders in Unreal 4's Sun Temple are helped, and gain SIMD16
because we avoid an integer multiplication.
instructions in affected programs: 2353 -> 2245 (-4.59%)
helped: 4
GAINED: 4
Reviewed-by: Jordan Justen <jordan.l.justen@intel.com>
Reviewed-by: Glenn Kennard <glenn.kennard@gmail.com>
The nir_lower_source_mods pass does a weak form of copy propagation to
clean up all of the mov-with-negate's that get generated. However, we
weren't properly checking that the sources were SSA and so we could end up
moving a register read which is not, in general, valid.
Reviewed-by: Connor Abbott <cwabbott0@gmail.com>
The old code wasn't correctly handling the case where the new value of the
source contains an indirect.
Reviewed-by: Connor Abbott <cwabbott0@gmail.com>
Previously, this function returned the number of elements for structures
and arrays and 0 for everything else. In NIR, this is almost never what
you want because we also treat matricies as arrays so you have to
special-case constantly. This commit glsl_get_length treat matrices
as an array of columns by returning the number of columns instead of 0
This also fixes a bug in locals_to_regs caused by not checking for the
matrix case in one place.
v2: Only special-case for matrices and return a length of 0 for vectors as
we did before. This was needed to not break the TGSI-based drivers and
doesn't really affect NIR at the moment.
Reviewed-by: Connor Abbott <cwabbott0@gmail.com>
Tested-by: Rob Clark <robclark@freedesktop.org>
Suggested by Jason on a different patch after some comments /
questions by Ilia.
Signed-off-by: Ian Romanick <ian.d.romanick@intel.com>
Reviewed-by: Jason Ekstrand <jason.ekstrand@intel.com>
Reviewed-by: Connor Abbott <cwabott0@gmail.com>
We weren't comparing the right number of components when checking
swizzles. Use nir_ssa_alu_instr_num_src_components() to do the right
thing.
No piglit regressions, and no fixes either.
Reviewed-by: Ian Romanick <ian.d.romanick@intel.com>
Reviewed-by: Jason Ekstrand <jason.ekstrand@intel.com>
Signed-off-by: Connor Abbott <cwabbott0@gmail.com>
Storing this here is pretty sketchy - I don't know if any driver other
than i965 will want to use it. But this will make it a lot easier to
generate NIR code at link time. We'll probably rework it anyway.
(Ian suggested making nir_assign_var_locations_scalar_direct_first
simply modify the nir_shader's fields, rather than passing pointers
to them. If this stays long term, we should do that. But Jason and
I suspect we'll be reworking this area again in the near future.)
Signed-off-by: Kenneth Graunke <kenneth@whitecape.org>
Reviewed-by: Jason Ekstrand <jason.ekstrand@intel.com>
I guess I was looking too much at how lower_system_values worked when
writing lower_idiv.
Since ttn wasn't emitting load_var for sysvals and the only drivers
using lower_idiv were using ttn, I think nothing was broken as a result.
But might as well fix this before it becomes a problem.
Signed-off-by: Rob Clark <robclark@freedesktop.org>
Reviewed-by: Jason Ekstrand <jason.ekstrand@intel.com>
Originally you had to have one or the other. But actually I don't want
either. (Or rather I want whatever is the minimum # of instructions.)
TODO: not sure where the best place to insert a check that driver hasn't
set *both* lower_negate and lower_sub?
Signed-off-by: Rob Clark <robclark@freedesktop.org>
Now that we're not generating linker errors, we don't actually modify
this.
Signed-off-by: Kenneth Graunke <kenneth@whitecape.org>
Reviewed-by: Jason Ekstrand <jason.ekstrand@intel.com>
Reviewed-by: Ian Romanick <ian.d.romanick@intel.com>
These should never happen. Plus, NIR passes really shouldn't be
reporting linker errors - this is past link time.
Signed-off-by: Kenneth Graunke <kenneth@whitecape.org>
Reviewed-by: Jason Ekstrand <jason.ekstrand@intel.com>
Reviewed-by: Ian Romanick <ian.d.romanick@intel.com>
We don't actually need a gl_program struct. We only used it to
translate prog->Target (i.e. GL_VERTEX_PROGRAM) to the gl_shader_stage
(i.e. MESA_SHADER_VERTEX). We may as well just pass that.
Signed-off-by: Kenneth Graunke <kenneth@whitecape.org>
Reviewed-by: Jason Ekstrand <jason.ekstrand@intel.com>
Reviewed-by: Ian Romanick <ian.d.romanick@intel.com>
I want to use this in some code that doesn't currently include mtypes.h.
It seems like a better place for it anyway.
Signed-off-by: Kenneth Graunke <kenneth@whitecape.org>
Reviewed-by: Jason Ekstrand <jason.ekstrand@intel.com>
Reviewed-by: Ian Romanick <ian.d.romanick@intel.com>
This pass performs a mark and sweep pass over a nir_shader's associated
memory - anything still connected to the program will be kept, and any
dead memory we dropped on the floor will be freed.
The expectation is that this will be called when finished building and
optimizing the shader. However, it's also fine to call it earlier, and
many times, to free up memory earlier.
v2: (feedback from Jason Ekstrand)
- Skip sweeping impl->start_block, as it's already in the CF list.
- Don't sweep SSA defs (they're owned by their defining instruction)
- Don't steal phi sources (they're owned by nir_phi_instr).
- Don't steal tex->src (it's owned by the tex_inst itself)
- Don't sweep dereference chains (top-level dereferences are owned by
the instruction; sub-dereferences are owned by the parent deref).
- Don't sweep sources and destinations (SSA defs are handled as part of
the defining instruction, and registers are handled as part of
function implementations).
- Just steal instructions; don't walk them (no longer required).
v3: (feedback from Jason Ekstrand)
- Steal indirect sources from nir_src/nir_dest.
Signed-off-by: Kenneth Graunke <kenneth@whitecape.org>
Reviewed-by: Jason Ekstrand <jason.ekstrand@intel.com>
Jason pointed out that variable dereferences in NIR are really part of
their parent instruction, and should have the same lifetime.
Unlike in GLSL IR, they're not used very often - just for intrinsic
variables, call parameters & return, and indirect samplers for
texturing. Also, nir_deref_var is the top-level concept, and
nir_deref_array/nir_deref_record are child nodes.
This patch attempts to allocate nir_deref_vars out of their parent
instruction, and any sub-dereferences out of their parent deref.
It enforces these restrictions in the validator as well.
This means that freeing an instruction should free its associated
dereference chain as well. The memory sweeper pass can also happily
ignore them.
v2: Rename make_deref to evaluate_deref and make it take a nir_instr *
instead of void *. This involves adding &instr->instr everywhere.
(Requested by Jason Ekstrand.)
Signed-off-by: Kenneth Graunke <kenneth@whitecape.org>
Reviewed-by: Jason Ekstrand <jason.ekstrand@intel.com>
We can't allocate them out of the nir_ssa_def itself, because it may not
be ralloc'd (for example, nir_dest embeds a nir_ssa_def).
However, allocating them out of the instruction should work.
Signed-off-by: Kenneth Graunke <kenneth@whitecape.org>
Reviewed-by: Jason Ekstrand <jason.ekstrand@intel.com>
Phi sources are part of the phi instruction and should have the same
lifetime.
Signed-off-by: Kenneth Graunke <kenneth@whitecape.org>
Reviewed-by: Jason Ekstrand <jason.ekstrand@intel.com>
The lifetime of the params array needs to be match the nir_call_instr
itself. So, allocate it using the instruction itself as the context.
Signed-off-by: Kenneth Graunke <kenneth@whitecape.org>
Reviewed-by: Jason Ekstrand <jason.ekstrand@intel.com>
These were added in commit f2616e56, presumably in preparation for
translating ARB vp/fp into GLSL IR. That never happened, and neither did
a lowering pass that actually generated these instructions.
Reviewed-by: Jason Ekstrand <jason.ekstrand@intel.com>
Based on the algo from NV50LegalizeSSA::handleDIV() and handleMOD().
See also trans_idiv() in freedreno/ir3/ir3_compiler.c (which was an
adaptation of the nv50 code from Ilia Mirkin).
A python/numpy script which implements the same algorithm (and is
possibly useful for debugging or analysis) can be found here:
http://people.freedesktop.org/~robclark/div-lowering.py
I've tested this on i965 hacked up to insert the idiv lowering pass,
and on freedreno with NIR frontend.
Signed-off-by: Rob Clark <robclark@freedesktop.org>
Tested-by: Eric Anholt <eric@anholt.net> (vc4)
v2: discovered that i2b/b2i are also confused
Signed-off-by: Rob Clark <robclark@freedesktop.org>
Reviewed-by: Matt Turner <mattst88@gmail.com>
Reviewed-by: Connor Abbott <cwabbott0@gmail.com>
In freedreno these get implemented as the matching f* instruction plus a
u2f to convert the result to float 1.0/0.0. But less lines of code to
just let nir_opt_algebraic handle this for us, plus opens up some small
window for other opt passes to improve (ie. if some shader ended up with
both a flt and slt with same src args, for example).
v2: use b2f rather than u2f
Signed-off-by: Rob Clark <robclark@freedesktop.org>
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
Reviewed-by: Matt Turner <mattst88@gmail.com>
As of da5ec2a, we allocate instruction sources out of the instruction
itself. When we realloc the texture sources we need to use the right
memory context or ralloc will get angry and assert-fail
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
This commit adds a pass to L1-normalize cube-map coordinates. Some hardware
such as i965 requires that largest cube-map coordinate is +-1. We had a
pass to perform this normalization in GLSL IR but we need it in NIR for
cube maps on ARB programs to work correctly.
Reviewed-by: Jordan Justen <jordan.l.justen@intel.com>
v2 (Suggested by Eric):
- Do a vector fabs and split into components later
- Move to core NIR
Reviewed-by: Eric Anholt <eric@anholt.net>
Not much hardware wants them these days, and it might give us a chance to
do CSE or algebraic at the NIR level.
Reviewed-by: Jason Ekstrand <jason.ekstrand@intel.com>
We use nir_ssa_defs for nir_builder args, so this takes a nir_src and
makes one so it can be passed in.
Reviewed-by: Jason Ekstrand <jason.ekstrand@intel.com>
So far we'd only used nir_builder to build brand new programs. But if
we're doing modifications to instructions (like in a lowering pass), then
we want to generate new stuff before the instruction we're modifying.
Reviewed-by: Jason Ekstrand <jason.ekstrand@intel.com>
The lifetime of the sources array needs to be match the nir_tex_instr
itself. So, allocate it using the instruction itself as the context.
Signed-off-by: Kenneth Graunke <kenneth@whitecape.org>
Reviewed-by: Jason Ekstrand <jason.ekstrand@intel.com>
These sets are part of the block, and their lifetime needs to match the
block itself. So, allocate them using the block itself as the context.
Signed-off-by: Kenneth Graunke <kenneth@whitecape.org>
Reviewed-by: Jason Ekstrand <jason.ekstrand@intel.com>
The lifetime of each register's use/def/if_use sets needs to match the
register itself. So, allocate them using the register itself as the
context.
Signed-off-by: Kenneth Graunke <kenneth@whitecape.org>
Reviewed-by: Jason Ekstrand <jason.ekstrand@intel.com>
glsl_to_nir passes in the ir_function's name field; we were copying the
pointer, but not duplicating the memory.
We want to be able to free the linked GLSL IR program after translating
to NIR, so we'll need to create a copy of the function name that the NIR
shader actually owns.
Signed-off-by: Kenneth Graunke <kenneth@whitecape.org>
Reviewed-by: Jason Ekstrand <jason.ekstrand@intel.com>
We can just pass a pointer to the list of variables, and reuse the code.
Signed-off-by: Kenneth Graunke <kenneth@whitecape.org>
Reviewed-by: Jason Ekstrand <jason.ekstrand@intel.com>
No shader-db changes, probably because they're all removed by the GLSL
compiler optimization added in commit 69ad5fd4.
Reviewed-by: Eric Anholt <eric@anholt.net>
Doesn't work for analogous && cases, because of NaNs.
total instructions in shared programs: 6195712 -> 6194829 (-0.01%)
instructions in affected programs: 42000 -> 41117 (-2.10%)
helped: 403
Reviewed-by: Eric Anholt <eric@anholt.net>
total instructions in shared programs: 4422307 -> 4422363 (0.00%)
instructions in affected programs: 4230 -> 4286 (1.32%)
helped: 0
HURT: 12
While this does hurt some things, the losses are minor and it prevents the
compare-with-zero optimization from fighting with ffma which is much more
important.
Reviewed-by: Matt Turner <mattst88@gmail.com>
Previously, we couldn't generate two algebraic passes in the same file
because of multiple structure definitions. To solve this, we play the
age-old header file trick and just #define around it.
Reviewed-by: Matt Turner <mattst88@gmail.com>
Previously, NIR would just print 4 swizzle components if the swizzle was
anything other than foo.xyzw. This creates lots of noise if, for example,
you have a one-component element with a swizzle of foo.xxxx.
Reviewed-by: Kenneth Grunke <kenneth@whitecape.org>
TGSI's conditional discards take float arg and negate it, so GLSL to TGSI
generates a b2f and negates that value. Only, in NIR we want a proper
bool once again, so we compare with 0. This is a lot of pointless extra
instructions.
total instructions in shared programs: 39735 -> 39702 (-0.08%)
instructions in affected programs: 1342 -> 1309 (-2.46%)
Reviewed-by: Connor Abbott <cwabbott0@gmail.com>
Since we have patterns based on b2f, generate them if we see the b2f
equivalent using an iand. This is common when generating NIR from TGSI.
Reviewed-by: Connor Abbott <cwabbott0@gmail.com>
These are nir_cf_nodes, not ALU instructions.
Also, use unreachable() to preempt said review feedback.
v2: Do it right (thanks Ilia).
Signed-off-by: Kenneth Graunke <kenneth@whitecape.org>
Reviewed-by: Jason Ekstrand <jason.ekstrand@intel.com>
prog->nir will generate fsub opcodes, but i965 doesn't implement them.
We may as well lower them at the NIR level, since it's trivial to do.
Suggested by Connor Abbott.
Signed-off-by: Kenneth Graunke <kenneth@whitecape.org>
Reviewed-by: Connor Abbott <cwabbott0@gmail.com>
Reviewed-by: Eric Anholt <eric@anholt.net>
These will be useful for prog->nir and tgsi->nir.
v2: Don't forget to mark nir_swizzle as inline (Eric).
Signed-off-by: Kenneth Graunke <kenneth@whitecape.org>
Reviewed-by: Connor Abbott <cwabbott0@gmail.com>
Reviewed-by: Eric Anholt <eric@anholt.net>
Both prog->nir and tgsi->nir will want to use these.
Signed-off-by: Kenneth Graunke <kenneth@whitecape.org>
Reviewed-by: Connor Abbott <cwabbott0@gmail.com>
Reviewed-by: Eric Anholt <eric@anholt.net>
v2: Delete the set of indirectly accessed variables when we're done with it
v3: Rename from _packed to _scalar
Reviewed-by: Connor Abbott <cwabbott0@gmail.com>
Previously, we just assigned variable locations in nir_lower_io. Now, we
force the user to assign variable locations for us. This gives the backend
a bit more control over where variables are placed.
v2: Rename from _packed to _scalar
Reviewed-by: Connor Abbott <cwabbott0@gmail.com>
We never did a single hash table lookup in the entire NIR code base that I
found so there was no real benifit to doing it that way. I suppose that
for linking, we'll probably want to be able to lookup by name but we can
leave building that hash table to the linker. In the mean time this was
causing problems with GLSL IR -> NIR because GLSL IR doesn't guarantee us
unique names of uniforms, etc. This was causing massive rendering isues in
the unreal4 Sun Temple demo.
Reviewed-by: Connor Abbott <cwabbott0@gmail.com>
Eric's initial patch adding constant expression evaluation for
ir_unop_round_even used nearbyint. The open-coded _mesa_round_to_even
implementation came about without much explanation after a reviewer
asked whether nearbyint depended on the application not modifying the
rounding mode. Of course (as Eric commented) we rely on the application
not changing the rounding mode from its default (round-to-nearest) in
many other places, including the IROUND function used by
_mesa_round_to_even!
Worse, IROUND() is implemented using the trunc(x + 0.5) trick which
fails for x = nextafterf(0.5, 0.0).
Still worse, _mesa_round_to_even unexpectedly returns an int. I suspect
that could cause problems when rounding large integral values not
representable as an int in ir_constant_expression.cpp's
ir_unop_round_even evaluation. Its use of _mesa_round_to_even is clearly
broken for doubles (as noted during review).
The constant expression evaluation code for the packing built-in
functions also mistakenly assumed that _mesa_round_to_even returned a
float, as can be seen by the cast through a signed integer type to an
unsigned (since negative float -> unsigned conversions are undefined).
rint() and nearbyint() implement the round-half-to-even behavior we want
when the rounding mode is set to the default round-to-nearest. The only
difference between them is that nearbyint() raises the inexact
exception.
This patch implements _mesa_roundeven{f,}, a function similar to the
roundeven function added by a yet unimplemented technical specification
(ISO/IEC TS 18661-1:2014), with a small difference in behavior -- we
don't bother raising the inexact exception, which I don't think we care
about anyway.
At least recent Intel CPUs can quickly change a subset of the bits in
the x87 floating-point control register, but the exception mask bits are
not included. rint() does not need to change these bits, but nearbyint()
does (twice: save old, set new, and restore old) in order to raise the
inexact exception, which would incur some penalty.
Reviewed-by: Carl Worth <cworth@cworth.org>
Shader-db results on HSW:
total instructions in shared programs: 4174156 -> 4157291 (-0.40%)
instructions in affected programs: 145397 -> 128532 (-11.60%)
helped: 383
HURT: 0
GAINED: 20
LOST: 22
There are two more tests lost than gained. However, comparing this with
GLSL IR vs. NIR results, the overall delta is reduced from 85/44
gained/lost on current master to 71/32 with this commit. Therefore, I
think it's probably a boon since we are getting "closer" to where we were
before.
Reviewed-by: Connor Abbott <cwabbott0@gmail.com>
Previously we tried to do poor-man's copy propagation as we created the
select instructions. Instead, this commit just moves the instructions from
the blocks inside the if into the block before. Copy propagation will take
care of making sure we don't have any extra mov's in there for us.
Reviewed-by: Connor Abbott <cwabbott0@gmail.com>
Previously, we stored derefs in a hash table, using the malloc'd pointer
as the key. Then, we walked through the hash table and generated code,
based on the order of the hash table's elements.
Memory addresses returned by malloc are pretty much random, which meant
that the hash was random, and the hash table's elements would be walked
in some random order. This led to successive compiles of the same
shader using different variable names and slightly different orderings
of phi-nodes. Code could not be diff'd, and the final assembly would
sometimes change slightly too.
It turns out the only point of the hash table was to avoid inserting
the same node multiple times for different dereferences. We never
actually searched the hash table! This patch uses an intrusive
linked list instead. Since exec_list uses head and tail sentinels,
checking prev or next against NULL will tell us whether the node is
already in the list.
Pair programming with Jason Ekstrand.
Signed-off-by: Jason Ekstrand <jason.ekstrand@intel.com>
Signed-off-by: Kenneth Graunke <kenneth@whitecape.org>
Reviewed-by: Connor Abbott <cwabbott0@gmail.com>
Ian and I added these around the time Connor was developing NIR. Now
that both exist, we should make them work together!
Signed-off-by: Kenneth Graunke <kenneth@whitecape.org>
Reviewed-by: Ian Romanick <ian.d.romanick@intel.com>
Reviewed-by: Jason Ekstrand <jason.ekstrand@intel.com>
Being able to see both location and driver_location can be useful when
debugging IO mistakes.
Signed-off-by: Kenneth Graunke <kenneth@whitecape.org>
Reviewed-by: Jason Ekstrand <jason.ekstrand@intel.com>
Reviewed-by: Connor Abbott <cwabbott0@gmail.com>
Vertex shaders can have shader inputs where location happens to be
VARYING_SLOT_FACE. Without predicating this on the shader stage,
we suddenly end up with load_front_face intrinsics in vertex shaders,
which is nonsensical.
Fixes spec/arb_vertex_buffer_object/pos-array when using NIR for VS.
Signed-off-by: Kenneth Graunke <kenneth@whitecape.org>
Reviewed-by: Jason Ekstrand <jason.ekstrand@intel.com>
The next commit needs to know the shader stage in glsl_to_nir().
To facilitate that, we pass the gl_shader rather than the raw exec_list
of instructions. This has both the exec_list and the stage.
Signed-off-by: Kenneth Graunke <kenneth@whitecape.org>
Reviewed-by: Jason Ekstrand <jason.ekstrand@intel.com>
glsl_to_nir, tgsi_to_nir, and prog_to_nir all want to know whether the
driver supports native integers. Presumably other passes may as well.
Adding this to nir_shader_compiler_options is an easy way to provide
that information, as it's accessible via nir_shader::options.
Signed-off-by: Kenneth Graunke <kenneth@whitecape.org>
Reviewed-by: Jason Ekstrand <jason.ekstrand@intel.com>
The code in glsl_to_nir is entirely dead, as we translate from GLSL to
NIR at link time, when there isn't a _mesa_glsl_parse_state to pass,
so every caller passes NULL.
glsl_to_nir seems like the wrong place to try and create the shader
compiler options structure anyway - tgsi_to_nir, prog_to_nir, and other
translators all would have to duplicate that code. The driver should
set this up once with whatever settings it wants, and pass it in.
Eric also added a NirOptions field to ctx->Const.ShaderCompilerOptions[]
and left a comment saying: "The memory for the options is expected to be
kept in a single static copy by the driver." This suggests the plan was
to do exactly that. That pointer was not marked const, however, and the
dead code used a mix of static structures and ralloced ones.
This patch deletes the dead code in glsl_to_nir, instead making it take
the shader compiler options as a mandatory argument. It creates an
(empty) options struct in the i965 driver, and makes NirOptions point
to that. It marks the pointer const so that we can actually do so
without generating "discards const qualifier" compiler warnings.
Signed-off-by: Kenneth Graunke <kenneth@whitecape.org>
Acked-by: Jason Ekstrand <jason.ekstrand@intel.com>
Reviewed-by: Eric Anholt <eric@anholt.net>
Nothing actually uses these, and the only caller of glsl_to_nir()
(brw_fs_nir.cpp) always passes NULL for the _mesa_glsl_parse_state
pointer, meaning they'll always be NULL and 0, respectively.
Just delete them.
Signed-off-by: Kenneth Graunke <kenneth@whitecape.org>
Reviewed-by: Jason Ekstrand <jason.ekstrand@intel.com>
v2:
- Single statement, by using memset return value as suggested by Ian
Romanick.
- No internal declaration, as suggested by Jason Ekstrand.
- Move macros to a header.
Reviewed-by: Jason Ekstrand <jason.ekstrand@intel.com>
This is to enable the code to build with -Werror=vla in the short term,
and enable the code to build with MSVC2013 soon after.
Reviewed-by: Brian Paul <brianp@vmware.com>
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
This is a conditional discard, which takes a boolean source.
Note that we don't generate ir_discard::condition today, so this
shouldn't break drivers (since none implement this intrinsic yet).
Signed-off-by: Kenneth Graunke <kenneth@whitecape.org>
Reviewed-by: Connor Abbott <cwabbott0@gmail.com>
Reviewed-by: Matt Turner <mattst88@gmail.com>
Reviewed-by: Eric Anholt <eric@anholt.net>
This adds a parent_instr field similar to the one for ssa_def. The
difference here is that the parent_instr field on a nir_register can be
NULL if the register does not have a unique definition or if that
definition does not dominate all its uses. We set this field in the
out-of-SSA pass so that backends can get SSA-like information even after
they have gone out of SSA.
Reviewed-by: Connor Abbott <cwabbott0@gmail.com>
Reviewed-by: Matt Turner <mattst88@gmail.com>
Reviewed-by: Eric Anholt <eric@anholt.net>
We were already do this for ALU operations but we haven't for non-ALU
operations. This changes that.
total NIR instructions in shared programs: 2039883 -> 2022338 (-0.86%)
NIR instructions in affected programs: 1768850 -> 1751305 (-0.99%)
helped: 14244
HURT: 124
total FS instructions in shared programs: 4083960 -> 4084036 (0.00%)
FS instructions in affected programs: 7302 -> 7378 (1.04%)
helped: 12
HURT: 51
Signed-off-by: Jason Ekstrand <jason.ekstrand@intel.com>
Reviewed-by: Connor Abbott <cwabbott0@gmail.com>
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
I initially wrote this based on the "(('fneg', ('fneg', a)), a)" above,
but we can generalize it and make it more potentially useful. In the
specific original case of a 0 for our new 'a' argument, it'll get further
algebraic optimization once the 0 is an argument to the new add.
No shader-db effects.
Reviewed-by: Matt Turner <mattst88@gmail.com>
Reviewed-by: Connor Abbott <cwabbott0@gmail.com>
We have some useful optimizations to drop things like 'ine a, 0' on a
boolean argument, but if 'a' came from logical operations on bools, it
couldn't tell. These kinds of constructs appear as a result of TGSI->NIR
quite frequently (at least with if flattening), so being a little more
aggressive in detecting booleans can pay off.
v2: Add ixor as a booleanness-preserving op (Suggestion by Connor).
vc4 results:
total instructions in shared programs: 40207 -> 39881 (-0.81%)
instructions in affected programs: 6677 -> 6351 (-4.88%)
Reviewed-by: Matt Turner <mattst88@gmail.com> (v1)
Reviewed-by: Connor Abbott <cwabbott0@gmail.com>
vc4 was already cleaning these up, but it does shave 4 NIR instructions in
shader-db.
Reviewed-by: Matt Turner <mattst88@gmail.com>
Reviewed-by: Connor Abbott <cwabbott0@gmail.com>
With -DDEBUG -UNDEBUG, this assert uses reg_state::stack_size, which
doesn't exist, breaking the build:
assert(state->states[index].index < state->states[index].stack_size);
Switch it to ifndef NDEBUG, so the field will exist if the assertion
actually generates code.
Signed-off-by: Kenneth Graunke <kenneth@whitecape.org>
Reviewed-by: Matt Turner <mattst88@gmail.com>
One less new directory necessary for gallium code that wants to interact
with NIR.
Reviewed-by: Connor Abbott <cwabbott0@gmail.com>
Reviewed-by: Jose Fonseca <jfonseca@vmware.com>
This lets us be slightly more efficient by not walking the CFG extra times.
Also, it may make it easier to ensure that GVN happens on only unpinned
instructions.
Reviewed-by: Reviewed-by: Connor Abbott <cwabbott0@gmail.com>
v2 Jason Ekstrand <jason.ekstrand@intel.com>:
- Use nir_dominance_lca for computing least common anscestors
- Use the block index for comparing dominance tree depths
- Pin things that do partial derivatives
Reviewed-by: Reviewed-by: Connor Abbott <cwabbott0@gmail.com>
Right now, the nir_instr_prev function function blindly looks up the
previous element in the exec list and casts it to an instruction even if
it's the tail sentinel. The next commit will change this to return null if
it's the first instruction. Making this change first avoids getting a
segfault between commits. The only reason we never noticed is that, thanks
to the way things are laid out in nir_block, the casted instruction's type
was never parallal_copy.
Reviewed-by: Connor Abbott <cwabbott0@gmail.com>
Previously, if you remved a CF node that still had instructions in it, none
of the use/def information from those instructions would get cleaned up.
Also, we weren't removing if statements from the if_uses of the
corresponding register or SSA def. This commit fixes both of these
problems
Reviewed-by: Connor Abbott <cwabbott0@gmail.com>
This is mostly thanks to Connor. The idea is to do a depth-first search
that computes pre and post indices for all the blocks. We can then figure
out if one block dominates another in constant time by two simple
comparison operations.
Reviewed-by: Connor Abbott <cwabbott0@gmail.com>
Being able to find the least common anscestor in the dominance tree is a
useful thing that we may want to do in other passes. In particular, we
need it for GCM.
v2: Handle NULL inputs by returning the other block
Reviewed-by: Connor Abbott <cwabbott0@gmail.com>
To silence compiler warning about unhandled switch case.
v2: move GLSL_TYPE_DOUBLE to the "not reached" section, per Ilia.
Reviewed-by: Ilia Mirkin <imirkin@alum.mit.edu>
v2: Rebase on the nir_opcodes.h python code generation support.
v3: Use SSA values, and set an appropriate writemask on dot products.
v4: Make the arguments be SSA references as well. This lets you stack up
expressions in the arguments of other expressions, at the cost of
having to insert a fmov/imov if you want to swizzle. Also, add
the generated file to NIR_GENERATED_FILES.
v5: Use more pythonish style for iterating the list.
v6: Infer the size of the dest from the size of the srcs, and auto-swizzle
a single small src out to the appropriate size.
v7: Add little helpers for initializing the struct, add a typedef for the
struct like other nir types have.
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org> (v6)
Reviewed-by: Connor Abbott <cwabbott0@gmail.com> (v7)
This could be done in a separate pass like we do in GLSL IR, but it seems
to me like having the definitions of the transformations in the two
directions next to each other makes a lot of sense.
v2: Reorder the comment about the transformation.
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
Mesa has a shader compiler struct flagging whether GLSL IR's opt_algebraic
and other passes should try and generate certain types of opcodes or
patterns. Extend that to NIR by defining our own struct, which is
automatically generated from the Mesa struct in glsl_to_nir and provided
directly by the driver in TGSI-to-NIR.
v2: Split out the previous two prep patches.
v3: Rebase to master (no TGSI->NIR present)
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org> (v2)
This will be used so that we can customize the transforms for the target
GPU, so we don't un-lower expressions that had already been lowered (or
introduce new lowering transformations that not all GPUs want)
v2: Drop the complication of having the condition->index dictionary, since
we don't actually expect there to be many different conditions (change
by Kenneth).
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
This will be used to give the optimization passes a chance to customize
behavior for the particular target device.
v2: Rebase to master (no TGSI->NIR present)
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org> (v1)
GLSL IR labels gl_FrontFacing as an input variable and not a system value.
This commit makes NIR silently translate gl_FrontFacing to a system value
so that it properly gets translated into a load_system_value intrinsic.
Reviewed-by: Connor Abbott <cwabbott0@gmail.com>
Reviewed-by: Matt Turner <mattst88@gmail.com>
And unfortunately other shaders do the same thing but with >=/<= which
we can't apply this optimization to because of NaNs.
instructions in affected programs: 23309 -> 22938 (-1.59%)
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
No change on shader-db on i965.
v2: Reword the comment due to feedback from Erik Faye-Lund
Reviewed-by: Connor Abbott <cwabbott0@gmail.com> (v1)
Reviewed-by: Jason Ekstrand <jason.ekstrand@intel.com> (v1)
We want the size of a float per component, not the size of a whole vec4.
NIR instructions on i965:
total instructions in shared programs: 1261937 -> 1261929 (-0.00%)
instructions in affected programs: 114 -> 106 (-7.02%)
Looking at one of these examples (tesseract), it's from vec4 load_consts
for a MRT solid fill, which do get CSEed now that we don't memcmp off the
end of the const value and into the SSA def. For the 1-component loads
that are common in i965, we were only memcmping off into the rest of the
usually zero-filled const_value.
Reviewed-by: Connor Abbott <cwabbott0@gmail.com>
Printing instructions doesn't modify them, so we can mark the parameter
const.
Signed-off-by: Kenneth Graunke <kenneth@whitecape.org>
Reviewed-by: Jason Ekstrand <jason.ekstrand@intel.com>
We've probably never seen this ridiculous pattern in the wild, so it
didn't matter.
Reviewed-by: Connor Abbott <cwabbott0@gmail.com>
Reviewed-by: Jason Ekstrand <jason.ekstrand@intel.com>
v2 Jason Ekstrand <jason.ekstrand@intel.com>:
- Add better comments
- Use nir_ssa_dest_init and nir_src_for_ssa more places
- Fix some void * casts
v3 Jason Ekstrand <jason.ekstrand@intel.com>:
- Rework the way we determine whether or not to sccalarize a phi node to
make the recursion non-bogus
- Treat load_const instructions as scalarizable
v4 Jason Ekstrand <jason.ekstrand@intel.com>:
- Allow uniform and input loads to be scalarizable
v5 Jason Ekstrand <jason.ekstrand@intel.com>:
- Also consider loads of inputs (varying, uniform, or ubo) to be
scalarizable. We were already doing this for load_var on uniforms and
inputs.
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
This allows you to match on an unknown value but only if it is of a given
type. 90% of the uses of this are for matching only booleans, but adding
the generality of arbitrary types is no more complex.
nir_algebraic.py doesn't handle this yet but that's ok because the C
language will ensure that the default type on all variables is void.
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
There are some algebraic transformations that we want to do but only if
certain things are constants. For instance, we may want to replace
a * (b + c) with (a * b) + (a * c) as long as a and either b or c is constant.
While this generates more instructions, some of it will get constant
folded.
nir_algebraic.py doesn't handle this yet, but that's ok because the C
language will make sure that false is the default for now.
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
We end up with these from TGSI-to-NIR because the pass generating the
comparisons doesn't know if the arg is actually a bool input or not. vc4
results:
total instructions in shared programs: 41801 -> 41508 (-0.70%)
instructions in affected programs: 4253 -> 3960 (-6.89%)
Reviewed-by: Matt Turner <mattst88@gmail.com>
This will be used by tgsi_to_nir, which needs to get vec4 types for
declaring shader input/output variables.
v2: Add a missing space.
Reviewed-by: Matt Turner <mattst88@gmail.com> (v2)
Reviewed-by: Jason Ekstrand <jason.ekstrand@intel.com>
It now emits vector MOVs instead of a series of individual MOVs, which
should be useful to any vector backends. This pushes the problem of
src/dest aliasing of channels on a scalar chip to the backend, but if
there are any vector operations in your shader then you needed to be
handling this already.
Fixes fs-swap-problem with my scalarizing patches.
v2: Rename to insert_mov(), and add a comment about what it does.
v3: Rewrite the comment.
Reviewed-by: Connor Abbott <cwabbott0@gmail.com> (v3)
This reverts commit d7d340fb2f.
We have an isnormal() implementation available, the only problem was that
we had the wrong return type (fixed in a later patch).
Bugzilla: https://bugs.freedesktop.org/show_bug.cgi?id=88806
Acked-by: Matt Turner <mattst88@gmail.com>
Before, we were only copying the first N channels, where N is the size
of the SSA destination, which is fine for per-component instructions,
but non-per-component instructions like fdot3 can have more source
components than destination components. Fix this using the helper
function introduced in the last patch.
v2: use new helper name
Reviewed-by: Jason Ekstrand <jason.ekstrand@intel.com>
Signed-off-by: Connor Abbott <cwabbott0@gmail.com>
Unlike with non-SSA ALU instructions, where if they're per-component
you have to look at the writemask to know which source channels are
being used, SSA ALU instructions always have all the possible channels
enabled so we can just look at the number of components in the SSA
definition for per-component instructions to say how many source
components are being used.
v2: use new name nir_ssa_alu_instr_src_components()
Reviewed-by: Jason Ekstrand <jason.ekstrand@intel.com>
Signed-off-by: Connor Abbott <cwabbott0@gmail.com>
Previously, we called the abs() function in math.h. However, this involves
unnecessarily going through double. This commit changes it to use integers
directly with a ternary.
Reviewed-by: Connor Abbott <cwabbott0@gmail.com>
Reviewed-by: Eric Anholt <eric@anholt.net>
Previously, these functions were explicitly writing to dst.x and dst.y.
However they both return only one component so writing to dst.y is invalid.
Also, since they only return one component, we don't need the explicit
assignment in the expression and can simplify it use an implicit
assignment.
Reviewed-by: Connor Abbott <cwabbott0@gmail.com>
This avoids the overhead of copying structures and better matches the newly
added nir_alu_src_copy and nir_alu_dest_copy.
Reviewed-by: Eric Anholt <eric@anholt.net>
Reviewed-by: Connor Abbott <cwabbott0@gmail.com>
Add a required field to the Opcode class, const_expr, that contains an
expression or statement that computes the result of the opcode given known
constant inputs. Then take those const_expr's and expand them into a function
that takes an opcode and an array of constant inputs and spits out the constant
result. This means that when adding opcodes, there's one less place to update,
and almost all the opcodes are self-documenting since the information on how to
compute the result is right next to the definition.
The helper functions in nir_constant_expressions.c were taken from
ir_constant_expressions.cpp.
v3 Jason Ekstrand <jason.ekstrand@iastate.edu>
- Use mako to generate one function per opcode instead of doing piles of
string splicing
v4 Jason Ekstrand <jason.ekstrand@iastate.edu>
- More comments and better indentation in the mako
- Add a description of the constant expression language in nir_opcodes.py
- Added nir_constant_expressions.py to EXTRA_DIST in Makefile.am
Signed-off-by: Jason Ekstrand <jason.ekstrand@intel.com>
Reviewed-by: Connor Abbott <cwabbott0@gmail.com>
Before, we used a system where a file, nir_opcodes.h, defined some macros that
were included to generate the enum values and the nir_op_infos structure. This
worked pretty well, but for development the error messages were never very
useful, Python tools couldn't understand the opcode list, and it was difficult
to use nir_opcodes.h to do other things like autogenerate a builder API. Now, we
store opcode information in nir_opcodes.py, and we have nir_opcodes_c.py to
generate the old nir_opcodes.c and nir_opcodes_h.py to generate nir_opcodes.h,
which contains all the enum names and gets included into nir.h like before. In
addition to solving the above problems, using Python and Mako to generate
everything means that it's much easier to add keep information centralized as we
add new things like constant propagation that require per-opcode information.
v2:
- make Opcode derive from object (Dylan)
- don't use assert like it's a function (Dylan)
- style fixes for fnoise, use xrange (Dylan)
- use iterkeys() in nir_opcodes_h.py (Dylan)
- use pydoc-style comments (Jason)
- don't make fmin/fmax commutative and associative yet (Jason)
Signed-off-by: Connor Abbott <cwabbott0@gmail.com>
Reviewed-by: Jason Ekstrand <jason.ekstrand@intel.com>
v3 Jason Ekstrand <jason.ekstrand@intel.com>
- Alphabetize source file lists
- Generate nir_opcodes.h in the builddir instead of the source dir
- Include $(builddir)/src/glsl/nir in the i965 build
- Rework nir_opcodes.h generation so it generates a complete header file
instead of one that has to be embedded inside an enum declaration
It's nice to have this present in your default cases so you can see what
instruction is triggering an abort.
v2: Just pass a NULL state, now that it won't crash when you do.
Reviewed-by: Jason Ekstrand <jason.ekstrand@intel.com>
This is the equivalent of brw_fs_channel_expressions.cpp, which I wanted
for vc4.
v2: Use the nir_src_for_ssa() helper, and another instance of
nir_alu_src_copy().
v3: Drop the non-SSA support. All intended callers will have SSA-only ALU
ops.
v4: Use insert_before, drop stale bcsel/fcsel comment, drop now-unused
unsupported() function, drop lower_context struct.
v5: Completely rename the pass to nir_lower_alu_to_scalar(), add an assert
about weird input_sizes[].
Reviewed-by: Jason Ekstrand <jason.ekstrand@iastate.edu>
There aren't many users yet, but I wanted to do this from my scalarizing
pass.
v2: Constify the src arguments.
Reviewed-by: Jason Ekstrand <jason.ekstrand@intel.com>
Most of these exist in the GLSL IR algebraic pass already. However,
SSA allows us to find more instances of the patterns.
total NIR instructions in shared programs: 2015593 -> 2011430 (-0.21%)
NIR instructions in affected programs: 124189 -> 120026 (-3.35%)
helped: 604
total i965 instructions in shared programs: 6025505 -> 6018717 (-0.11%)
i965 instructions in affected programs: 261295 -> 254507 (-2.60%)
helped: 1295
HURT: 3
Signed-off-by: Kenneth Graunke <kenneth@whitecape.org>
Reviewed-by: Jason Ekstrand <jason.ekstrand@intel.com>
Reviewed-by: Matt Turner <mattst88@gmail.com>
The first batch removes bonus fnot/inot operations, possibly allowing
other optimizations to better recognize patterns.
The next batch replaces a fadd and constant 0.0 with an fneg - negation
is usually free on GPUs, while addition is not.
total NIR instructions in shared programs: 2020814 -> 2015593 (-0.26%)
NIR instructions in affected programs: 411143 -> 405922 (-1.27%)
helped: 2233
HURT: 214
A few shaders are hurt by a few instructions due to moving neg such
that it has a constant operand, which is then folded, resulting in two
distinct load_consts for x and -x. We can always clean that up later.
total i965 instructions in shared programs: 6035392 -> 6025505 (-0.16%)
i965 instructions in affected programs: 784980 -> 775093 (-1.26%)
helped: 4508
HURT: 2
Signed-off-by: Kenneth Graunke <kenneth@whitecape.org>
Reviewed-by: Jason Ekstrand <jason.ekstrand@intel.com>
Reviewed-by: Matt Turner <mattst88@gmail.com>
The GLSL IR optimization pass contained these; we may as well include
them too.
v2: Fix a >> 0 and a << 0 optimizations (caught by Matt).
No change in the number of NIR instructions on a shader-db run.
total i965 instructions in shared programs: 6035397 -> 6035392 (-0.00%)
i965 instructions in affected programs: 542 -> 537 (-0.92%)
helped: 2 (in glamor)
Signed-off-by: Kenneth Graunke <kenneth@whitecape.org>
Reviewed-by: Jason Ekstrand <jason.ekstrand@intel.com>
Reviewed-by: Matt Turner <mattst88@gmail.com>
Matt and I noticed a bunch of "val <- ior a a" operations in a shader,
so we decided to add an algebraic optimization for that. While there,
I decided to add a bunch more of them.
v2: Delete bogus fand/for optimizations (caught by Jason).
total NIR instructions in shared programs: 2023511 -> 2020814 (-0.13%)
NIR instructions in affected programs: 149634 -> 146937 (-1.80%)
helped: 1032
total i965 instructions in shared programs: 6035392 -> 6035397 (0.00%)
i965 instructions in affected programs: 537 -> 542 (0.93%)
HURT: 2
Signed-off-by: Kenneth Graunke <kenneth@whitecape.org>
Reviewed-by: Jason Ekstrand <jason.ekstrand@intel.com>
Reviewed-by: Matt Turner <mattst88@gmail.com>
Matt and I noticed that one of the shaders hurt by INTEL_USE_NIR=1 had
load_input and load_uniform intrinsics repeated several times, with the
same parameters, but each one generating a distinct SSA value. This
made ALU operations on those values appear distinct as well.
Generating distinct SSA values is silly - these are read only variables.
CSE'ing them makes everything use a single SSA value, which then allows
other operations to be CSE'd away as well.
Generalizing a bit, it seems like we should be able to safely CSE any
intrinsics that can be eliminated and reordered. I didn't implement
support for variables for the time being.
v2: Assert that info->num_variables == 0 (requested by Jason).
total NIR instructions in shared programs: 2435936 -> 2023511 (-16.93%)
NIR instructions in affected programs: 2413496 -> 2001071 (-17.09%)
helped: 16872
total i965 instructions in shared programs: 6028987 -> 6008427 (-0.34%)
i965 instructions in affected programs: 640654 -> 620094 (-3.21%)
helped: 2071
HURT: 585
GAINED: 14
LOST: 25
Signed-off-by: Kenneth Graunke <kenneth@whitecape.org>
Reviewed-by: Jason Ekstrand <jason.ekstrand@intel.com>
This should not be a change in behavior, as all current cases that
potentially answer "yes" require SSA.
The next patch will introduce another case that requires SSA.
Signed-off-by: Kenneth Graunke <kenneth@whitecape.org>
Reviewed-by: Jason Ekstrand <jason.ekstrand@intel.com>
Reviewed-by: Matt Turner <mattst88@gmail.com>
brw_fs_nir has only seen scalar bools so far, thanks to vector splitting,
and the ralloc of in glsl_to_nir.cpp will *usually* get you a 0-filled
chunk of memory, so reading too large of a value will usually get you the
right bool value. But once we start doing vector bools in a few commits,
we end up getting bad values.
Reviewed-by: Jason Ekstrand <jason.ekstrand@intel.com>
Almost all instructions we nir_ssa_def_init() for are nir_dests, and you
have to keep from forgetting to set is_ssa when you do. Just provide the
simpler helper, instead.
Reviewed-by: Jason Ekstrand <jason.ekstrand@intel.com>
Designated initializers with anonymous unions don't work in MSVC or
GCC < 4.6. With a couple of constructor methods, we don't need them any
more and the code is actually cleaner.
Bugzilla: https://bugs.freedesktop.org/show_bug.cgi?id=88467
Reviewed-by: Connor Abbot <cwabbott0@gmail.com>
Fix build error on Mac OS X.
CC nir_to_ssa.lo
nir_to_ssa.c:29:10: fatal error: 'malloc.h' file not found
^
Bugzilla: https://bugs.freedesktop.org/show_bug.cgi?id=88478
Signed-off-by: Vinson Lee <vlee@freedesktop.org>
This is a rework of the liveness algorithm using a worklist as suggested by
Connor. Doing so reduces the number of times we walk over the instructions
because we don't have to do an entire pointless walk over the instructions
just to figure out it's time to stop. Also, the stuff after the last loop
in the funciton will only ever get visited once.
Reviewed-by: Connor Abbott <cwabbott0@gmail.com>
A worklist is a common concept in optimizations. This adds a structure
that we can reuse for many different types of optimizations.
Reviewed-by: Connor Abbott <cwabbott0@gmail.com>
Previously, the set API required the user to do all of the hashing of keys
as it passed them in. Since the hashing function is intrinsically tied to
the comparison function, it makes sense for the hash set to know about
it. Also, it makes for a somewhat clumsy API as the user is constantly
calling hashing functions many of which have long names. This is
especially bad when the standard call looks something like
_mesa_set_add(ht, _mesa_pointer_hash(key), key);
In the above case, there is no reason why the hash set shouldn't do the
hashing for you. We leave the option for you to do your own hashing if
it's more efficient, but it's no longer needed. Also, if you do do your
own hashing, the hash set will assert that your hash matches what it
expects out of the hashing function. This should make it harder to mess up
your hashing.
This is analygous to 94303a0750 where we did this for hash_table
Signed-off-by: Jason Ekstrand <jason.ekstrand@intel.com>
Reviewed-by: Matt Turner <mattst88@gmail.com>
Reviewed-by: Eric Anholt <eric@anholt.net>
We already have search_pre_hashed. This makes the APIs match better.
Reviewed-by: Matt Turner <mattst88@gmail.com>
Reviewed-by: Eric Anholt <eric@anholt.net>
Without the break, it was possible that an instruction would match multiple
expressions. If this happened, you could end up trying to replace it
multiple times and get a segfault. This makes it so that, after a
successful replacement, it moves on to the next instruction.
Reviewed-by: Connor Abbott <cwabbott0@gmail.com>
This refactor allows you to more easily get the deref node associated with
a given variable. We then use that new functionality in the
deref_may_be_aliased function instead of creating a 1-element deref chain.
Reviewed-by: Connor Abbott <cwabbott0@gmail.com>
The original name wasn't particularly descriptive. This one indicates that
it actually gives you SSA values as opposed to the old pass which lowered
variables to registers.
Reviewed-by: Connor Abbott <cwabbott0@gmail.com>
This solves a number of problems. First is the ability to change the
number of sources that a texture instruction has. Second, it solves the
delema that may occur if a texture instruction has more than 4 sources.
Reviewed-by: Connor Abbott <cwabbott0@gmail.com>
Additional description was added to a variety of places. Also, we no
longer use the term "leaf" to describe fully-qualified direct derefs.
Instead, we simply use the term "direct" or spell it out completely.
Reviewed-by: Connor Abbott <cwabbott0@gmail.com>
This should be much better for debugging as GDB will pick up on the fact
that it's an enum and actually tell you what you're looking at instead of
giving you some arbitrary hex value you have to go look up.
Reviewed-by: Connor Abbott <cwabbott0@gmail.com>
This should make debugging a lot easier as GDB handles static inlines much
better than macros. Also, static inlines are typesafe.
Reviewed-By: Glenn Kennard <glenn.kennard@gmail.com>
Reviewed-by: Connor Abbott <cwabbott0@gmail.com>
This was a left-over relic of GLSL IR that we aren't using for anything.
If we ever want that value again, we can add it back, but NIR constant
folding should be just as good as GLSL IR's if not better pretty soon, so
I'm not worried about it.
Reviewed-by: Connor Abbott <cwabbott0@gmail.com>
Previously, our variable renaming algorithm, while similar to the one in
the Cytron paper, was not the same. While I'm pretty sure it was correct,
it will be easier for readers of the code in the variable renaming pass if
it follows more closely. This commit removes the automatic stack popping
we were doing and replaces it with explicit popping like Cytron does.
Reviewed-by: Connor Abbott <cwabbott0@gmail.com>
This commit seeks to make the lower_variables pass much more clear by
adding a pile of comments and re-arranging a few things. There are no
functional or algorithmic changes.
Reviewed-by: Connor Abbott <cwabbott0@gmail.com>
parallel_copy_copy was a silly name. Also, things were getting long and
annoying, so I added a foreach macro. For historical reasons, several of
the original iterations over parallel copy entries in from_ssa used the
_safe variants of the loop. However, all of these no longer ever remove an
entry so it's ok to make them all use the normal iterator.
Reviewed-by: Connor Abbott <cwabbott0@gmail.com>
Previously, we were doing a lazy creation of the parallel copy
instructions. This is confusing, hard to get right, and involves some
extra state tracking of the copies. This commit adds an extra walk over
the basic blocks to add the block-end parallel copies up front. This
should be much less confusing and, consequently, easier to get right. This
commit also adds more comments about parallel copies to help explain what
all is going on.
As a consequence of these changes, we can now remove the at_end parameter
from nir_parallel_copy_instr.
Reviewed-by: Connor Abbott <cwabbott0@gmail.com>
Originally, this field was intended for determining if the given
instruction acted per-component or if it had mismatching source and
destination sizes that would have to be interpreted specially. However, we
can easily derive this from output_size == 0, so it's not really that
useful. Also, the values we were setting in nir_opcodes.h for this field
were completely bogus and it was never used.
Reviewed-by: Connor Abbott <cwabbott0@gmail.com>
Prior to this commit, we had a big switch statement for this. Now it's
baked into the opcode metadata so we can just use that.
Reviewed-by: Connor Abbott <cwabbott0@gmail.com>
This commit adds some algebraic properties to the metadata of each opcode
in NIR. In particular, you now know, just from the metadata, if a given
opcode is commutative or associative. This will be useful for algebraic
transformation passes that want to be able to match a + b as well as b + a
in one go.
v2: Make algebraic properties all caps. This was more consistent with the
intrinsics flags and seems better for flags in general.
Also, the enums are now declared with (1 << n) rather then hex values.
v3: fmin and fmax technically aren't commutative or associative. Things
get funny when one of the arguments is a NaN.
Reviewed-by: Connor Abbott <cwabbott0@gmail.com>
As it was, we weren't ever using load_const in a non-SSA way. This allows
us to substantially simplify the load_const instruction. If we ever need a
non-SSA constant load, we can do a load_const and an imov.
Reviewed-by: Connor Abbott <cwabbott0@gmail.com>
Previously, lower_atomics was non-SSA only. We assert-failed if the
destination of an atomic operation intrinsic was an SSA def and we used
temporary registers for computing offsets. This commit changes both of
these behaviors. We now use SSA values for computing offsets (so we can
optimize them) and we handle SSA destinations. We also move the pass to
run before we go out of SSA on i965 as it now generates SSA values.
Reviewed-by: Connor Abbott <cwabbott0@gmail.com>
Before, we were using foreach_dest and switching on whether the destination
was an SSA value. This works, except not all destinations are SSA values
so we have to special-case ssa_undef instructions. Now that we have a
foreach_ssa_def function, we can iterate over all of the register
destinations in one pass and iterate over the SSA destinations in a second.
This way, if we add other ssa-only instructions, we won't have to worry
about adding them to the special case we have for ssa_undef.
Reviewed-by: Connor Abbott <cwabbott0@gmail.com>
There are some functions whose destinations are SSA-only and so aren't a
nir_dest. This provides a function that is capable of iterating over the
SSA definitions defined by those functions. If you want registers, you
should use the old iterator.
v2: Kenneth Graunke <kenneth@whitecape.org>:
- Fix nir_foreach_ssa_def's return value.
Reviewed-by: Connor Abbott <cwabbott0@gmail.com>
Previously, we were just iterating over the program "in order" which
kind-of approximates a DFS, but not really. In particular, we got the
following case wrong:
loop {
a = 3;
if (foo) {
a = 5;
} else {
break;
}
use(a);
}
where use(a) would get 3 instead of 5 because of premature popping of the
SSA def stack. Now, since we do an actaul DFS, we should evaluate use(a)
immediately after a = 5 and we should be ok.
Reviewed-by: Connor Abbott <cwabbott0@gmail.com>
We stopped generating predicates in glsl_to_nir some time ago. Right now,
it's all dead untested code that I'm not convinced always worked in the
first place. If we decide we want them back, we can revert this patch.
Reviewed-by: Connor Abbott <cwabbott0@gmail.com>
Previously, the condition was a scalar that applied to all components
simultaneously. As of this commit, the condition is a vector and each
component is switched seperately.
Reviewed-by: Connor Abbott <cwabbott0@gmail.com>
nir_metadata_dirty was a terrible name because the parameter it takes is
the metadata to be preserved. This is really confusing because it looks
like it's doing the opposite of what it is actually doing. Now it's named
sensibly.
Reviewed-by: Connor Abbott <cwabbott0@gmail.com>
v2 Jason Ekstrand <jason.ekstrand@intel.com>:
- Use the nir_tex_src_sampler_offset source type instead of the
sampler_indirect thing that I cooked up before.
Reviewed-by: Chris Forbes <chrisf@ijw.co.nz>
In particular, we rename nir_tex_src_sampler_index to _sampler_offset and
add a sampler_array_size field to nir_tex_instr. This way we can pass the
size of sampler arrays through to backends even after removing the variable
information and, with it, the type.
Reviewed-by: Connor Abbott <cwabbott0@gmail.com>
In GLSL-to-NIR we were just setting the base index to 0 whenever there was
an indirect so having it expressed as a sum makes no sense. Also, while a
base offset may make sense for the memory location (first element in the
array, etc.) it makes less sense for the actual uniform buffer index. This
may change later, but it seems to make more sense for now.
Reviewed-by: Connor Abbott <cwabbott0@gmail.com>
This commit renames nir_instr_as_texture to nir_instr_as_tex and renames
nir_instr_type_texture to nir_instr_type_tex to be consistent with
nir_tex_instr.
Reviewed-by: Connor Abbott <cwabbott0@gmail.com>
This pass uses the previously built algebraic transformations framework and
should act as an example for anyone else wanting to make an algebraic
transformation pass for NIR.
Reviewed-by: Matt Turner <mattst88@gmail.com>
Reviewed-by: Connor Abbott <cwabbott0@gmail.com>
This commit builds on the nir_search.h infastructure by adding a bit of
python code that makes it stupid easy to write an algebraic transformation
pass. The nir_algebraic.py file contains four python classes that
correspond directly to the datastructures in nir_search.c and allow you to
easily generate the C code to represent them. Given a list of
search-and-replace operations, it can then generate a function that applies
those transformations to a shader.
The transformations can be specified manually, or they can be specified
using nested tuples. The nested tuples make a neat little language for
specifying expression trees and search-and-replace operations in a very
readable and easy-to-edit fasion.
The generated code is also fairly efficient. Insteady of blindly calling
nir_replace_instr with every single transformation and on every single
instruction, it uses a switch statement on the instruction opcode to do a
first-order culling and only calls nir_replace_instr if the opcode is known
to match the first opcode in the search expression.
Reviewed-by: Connor Abbott <cwabbott0@gmail.com>
This framework provides a simple way to do simple search-and-replace
operations on NIR code. The nir_search.h header provides four simple data
structures for representing expressions: nir_value and four subtypes:
nir_variable, nir_constant, and nir_expression. An expression tree can
then be represented by nesting these data structures as needed. The
nir_replace_instr function takes an instruction, an expression, and a
value; if the instruction matches the expression, it is replaced with a new
chain of instructions to generate the given replacement value. The
framework keeps track of swizzles on sources and automatically generates
the currect swizzles for the replacement value.
Reviewed-by: Connor Abbott <cwabbott0@gmail.com>
Previously, the casting operations were macros. While this is usually
fine, the casting macro used the input parameter twice leading to strange
behavior when you passed the result of another function into it. Since we
know the source and destination types explicitly, we don't loose anything
by making it a function.
Also, this gives us a nice little macro for creating cast function that
will hopefully prevent mistyping.
Reviewed-by: Connor Abbott <cwabbott0@gmail.com>
We used to have the number of components built into the intrinsic. This
meant that all of our load/store intrinsics had vec1, vec2, vec3, and vec4
variants. This lead to piles of switch statements to generate the correct
intrinsic names, and introspection to figure out the number of components.
We can make things much nicer by allowing "vectorized" intrinsics.
Reviewed-by: Connor Abbott <cwabbott0@gmail.com>
With this commit, the GLSL IR -> NIR pass generates NIR in more-or-less SSA
form. It's SSA in the sense that it doesn't have any registers, but it
isn't really useful SSA because it still has a pile of load/store
intrinsics that we will need to get rid of.
Reviewed-by: Connor Abbott <cwabbott0@gmail.com>
This pass analizes all of the load/store operations and, when a variable is
never aliased (potentially used by an indirect operation), it is lowered
directly to an SSA value. This pass translates to SSA directly and does
not require any fixup by the original to-SSA pass.
Reviewed-by: Connor Abbott <cwabbott0@gmail.com>
Instead, we give SSA definitions a temporary index of 0xFFFFFFFF if the
instruction does not have a block and a proper index when it actually gets
added to the list.
Reviewed-by: Connor Abbott <cwabbott0@gmail.com>
Previously, we used a string name. It was nice for translating out of GLSL
IR (which also does that) but cumbersome the rest of the time.
Reviewed-by: Connor Abbott <cwabbott0@gmail.com>
Backends want to be able to do special things with constant values such as
put them into immediates or make decisions based on whether or not a value
is constant. Before, constants always got lowered to a load_const into a
register and then a register use. Now we leave constants as SSA values so
backends can special-case them if they want. Since handling constant SSA
values is trivial, this shouldn't be a problem for backends.
Reviewed-by: Connor Abbott <cwabbott0@gmail.com>
This pass is still fairly basic. It only handles ALU operations, constant
loads, and phi nodes. No texture ops or intrinsics yet.
Reviewed-by: Connor Abbott <cwabbott0@gmail.com>
The unlink_blocks function moves successors around to make sure that, if
there is a remaining successor, it is in the first successors slot and not
the second. To fix this, we simply get both successors up front.
Reviewed-by: Connor Abbott <cwabbott0@gmail.com>
We also make the return types match GLSL. The GLSL spec specifies that
findMSB and findLSB return a signed integer. Previously, nir had them
return unsigned. This updates nir's behavior to match what GLSL expects.
We also update the nir-to-fs generator to take the new instructions. While
we're at it, we fix the case where the input to findMSB is zero.
Reviewed-by: Connor Abbott <cwabbott0@gmail.com>
These indices should now be reasonably stable/consistent. Redoing the
indices in the print functions makes it harder to debug problems.
Reviewed-by: Connor Abbott <cwabbott0@gmail.com>
This commit rewrites the out-of-SSA pass to not be nearly as naieve. It's
based on "Revisiting Out-of-SSA Translation for Correctness, Code Quality,
and Efficiency" by Boissinot et. al. It should be fairly close to
state-of-the art.
Reviewed-by: Connor Abbott <cwabbott0@gmail.com>
Since we don't actually have an "if" instruction, this is a very common
pattern when iterating over instructions. This adds a helper function for
it to make things a little less painful.
Reviewed-by: Connor Abbott <cwabbott0@gmail.com>
This pass is kind of stupidly implemented but it should be enough to get us
up and going. We probably want something better that doesn't generate all
of the redundant moves eventually. However, the i965 backend should be
able to handle the movs, so I'm not too worried about it in the short term.