nir_opt_algebraic is currently one of the most expensive NIR passes,
because of the many different patterns we've added over the years. Even
though patterns are already sorted by opcode, there are still way too
many patterns for common opcodes like bcsel and fadd, which means that
many patterns are tried but only a few actually match. One way to fix
this is to add a pre-pass over the code that scans it using an automaton
constructed beforehand, similar to the automatons produced by lex and
yacc for parsing source code. This automaton has to walk the SSA graph
and recognize possible pattern matches.
It turns out that the theory to do this is quite mature already, having
been developed for instruction selection as well as other non-compiler
things. I followed the presentation in the dissertation cited in the
code, "Tree algorithms: Two Taxonomies and a Toolkit," trying to keep
the naming similar. To create the automaton, we have to perform
something like the classical NFA to DFA subset construction used by lex,
but it turns out that actually computing the transition table for all
possible states would be way too expensive, with the dissertation
reporting times of almost half an hour for an example of size similar to
nir_opt_algebraic. Instead, we adopt one of the "filter" approaches
explained in the dissertation, which trade much faster table generation
and table size for a few more table lookups per instruction at runtime.
I chose the filter which resulted the fastest table generation time,
with medium table size. Right now, the table generation takes around .5
seconds, despite being implemented in pure Python, which I think is good
enough. Based on the numbers in the dissertation, the other choice might
make table compilation time 25x slower to get 4x smaller table size, but
I don't think that's worth it. As of now, we get the following binary
size before and after this patch:
text data bss dec hex filename
11979455 464720 730864 13175039 c908ff before i965_dri.so
text data bss dec hex filename
12037835 616244 791792 13445871 cd2aef after i965_dri.so
There are a number of places where I've simplified the automaton by
getting rid of details in the LHS patterns rather than complicate things
to deal with them. For example, right now the automaton doesn't
distinguish between constants with different values. This means that it
isn't as precise as it could be, but the decrease in compile time is
still worth it -- these are the compilation time numbers for a shader-db
run with my (admittedly old) database on Intel skylake:
Difference at 95.0% confidence
-42.3485 +/- 1.375
-7.20383% +/- 0.229926%
(Student's t, pooled s = 1.69843)
We can always experiment with making it more precise later.
Reviewed-by: Jason Ekstrand <jason@jlekstrand.net>
This takes the stupid simplest and most reliable approach to reducing
redundancy that I could come up with: Just use the struct declaration
as the cach key. This cuts the size of the generated C file to about
half and takes about 50 KiB off the .data section.
size before (release build):
text data bss dec hex filename
5363833 336880 13584 5714297 573179 _install/lib64/libvulkan_intel.so
size after (release build):
text data bss dec hex filename
5229017 285264 13584 5527865 545939 _install/lib64/libvulkan_intel.so
Reviewed-by: Connor Abbott <cwabbott0@gmail.com>
In compute shaders if no derivative group is defined, the derivatives
will always be zero. Specified in NV_compute_shader_derivatives.
To make the check more convenient, add a "info" local variable to the
generated code so we can refer to it in the Python rules. (Jason)
Reviewed-by: Ian Romanick <ian.d.romanick@intel.com>
Reviewed-by: Jason Ekstrand <jason@jlekstrand.net>
Consider the following search expression and NIR sequence:
('iadd', ('imul', a, b), b)
ssa_2 = imul ssa_0, ssa_1
ssa_3 = iadd ssa_2, ssa_0
The current algorithm is greedy and, the moment the imul finds a match,
it commits those variable names and returns success. In the above
example, it maps a -> ssa_0 and b -> ssa_1. When we then try to match
the iadd, it sees that ssa_0 is not b and fails to match. The iadd
match will attempt to flip itself and try again (which won't work) but
it cannot ask the imul to try a flipped match.
This commit instead counts the number of commutative ops in each
expression and assigns an index to each. It then does a loop and loops
over the full combinatorial matrix of commutative operations. In order
to keep things sane, we limit it to at most 4 commutative operations (16
combinations). There is only one optimization in opt_algebraic that
goes over this limit and it's the bitfieldReverse detection for some UE4
demo.
Shader-db results on Kaby Lake:
total instructions in shared programs: 15310125 -> 15302469 (-0.05%)
instructions in affected programs: 1797123 -> 1789467 (-0.43%)
helped: 6751
HURT: 2264
total cycles in shared programs: 357346617 -> 357202526 (-0.04%)
cycles in affected programs: 15931005 -> 15786914 (-0.90%)
helped: 6024
HURT: 3436
total loops in shared programs: 4360 -> 4360 (0.00%)
loops in affected programs: 0 -> 0
helped: 0
HURT: 0
total spills in shared programs: 23675 -> 23666 (-0.04%)
spills in affected programs: 235 -> 226 (-3.83%)
helped: 5
HURT: 1
total fills in shared programs: 32040 -> 32032 (-0.02%)
fills in affected programs: 190 -> 182 (-4.21%)
helped: 6
HURT: 2
LOST: 18
GAINED: 5
Reviewed-by: Thomas Helland <thomashelland90@gmail.com>
NIR metadata validation verifies that the debug bit was unset (by a call
to nir_metadata_preserve) if a NIR optimization pass made progress on
the shader. With the expectation that the NIR shader consists of only a
single main function, it has been safe to call nir_metadata_preserve()
iff progress was made.
However, most optimization passes calculate progress per-function and
then return the union of those calculations. In the case that an
optimization pass makes progress only on a subset of the functions in
the shader metadata validation will detect the debug bit is still set on
any unchanged functions resulting in a failed assertion.
This patch offers a quick solution (short of a larger scale refactoring
which I do not wish to undertake as part of this series) that simply
unsets the debug bit on unchanged functions.
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
Reviewed-by: Jason Ekstrand <jason@jlekstrand.net>
The quotation marks around 1.0 cause it to be treated as a string
instead of a floating point value. The generator then treats it as an
arbitrary variable replacement, so any iand involving a ('ineg', ('b2i',
a)) matches.
v2: Remove misleading comment about sized literals (suggested by
Timothy). Add assertion that the name of a varible is entierly
alphabetic (suggested by Jason).
Signed-off-by: Ian Romanick <ian.d.romanick@intel.com>
Reviewed-by: Jason Ekstrand <jason@jlekstrand.net>
Tested-by: Timothy Arceri <tarceri@itsqueeze.com> [v1]
Reviewed-by: Timothy Arceri <tarceri@itsqueeze.com> [v1]
Fixes: 6bcd2af086 ("nir/algebraic: Add some optimizations for D3D-style Booleans")
Bugzilla: https://bugs.freedesktop.org/show_bug.cgi?id=109075
Reviewed-by: Eric Anholt <eric@anholt.net>
Reviewed-by: Bas Nieuwenhuizen <bas@basnieuwenhuizen.nl>
Tested-by: Bas Nieuwenhuizen <bas@basnieuwenhuizen.nl>
This is a squash of a bunch of individual changes:
nir/builder: Generate 32-bit bool opcodes transparently
nir/algebraic: Remap Boolean opcodes to the 32-bit variant
Use 32-bit opcodes in the NIR producers and optimizations
Generated with a little hand-editing and the following sed commands:
sed -i 's/nir_op_ball_fequal/nir_op_b32all_fequal/g' **/*.c
sed -i 's/nir_op_bany_fnequal/nir_op_b32any_fnequal/g' **/*.c
sed -i 's/nir_op_ball_iequal/nir_op_b32all_iequal/g' **/*.c
sed -i 's/nir_op_bany_inequal/nir_op_b32any_inequal/g' **/*.c
sed -i 's/nir_op_\([fiu]lt\)/nir_op_\132/g' **/*.c
sed -i 's/nir_op_\([fiu]ge\)/nir_op_\132/g' **/*.c
sed -i 's/nir_op_\([fiu]ne\)/nir_op_\132/g' **/*.c
sed -i 's/nir_op_\([fiu]eq\)/nir_op_\132/g' **/*.c
sed -i 's/nir_op_\([fi]\)ne32g/nir_op_\1neg/g' **/*.c
sed -i 's/nir_op_bcsel/nir_op_b32csel/g' **/*.c
Use 32-bit opcodes in the NIR back-ends
Generated with a little hand-editing and the following sed commands:
sed -i 's/nir_op_ball_fequal/nir_op_b32all_fequal/g' **/*.c
sed -i 's/nir_op_bany_fnequal/nir_op_b32any_fnequal/g' **/*.c
sed -i 's/nir_op_ball_iequal/nir_op_b32all_iequal/g' **/*.c
sed -i 's/nir_op_bany_inequal/nir_op_b32any_inequal/g' **/*.c
sed -i 's/nir_op_\([fiu]lt\)/nir_op_\132/g' **/*.c
sed -i 's/nir_op_\([fiu]ge\)/nir_op_\132/g' **/*.c
sed -i 's/nir_op_\([fiu]ne\)/nir_op_\132/g' **/*.c
sed -i 's/nir_op_\([fiu]eq\)/nir_op_\132/g' **/*.c
sed -i 's/nir_op_\([fi]\)ne32g/nir_op_\1neg/g' **/*.c
sed -i 's/nir_op_bcsel/nir_op_b32csel/g' **/*.c
Reviewed-by: Eric Anholt <eric@anholt.net>
Reviewed-by: Bas Nieuwenhuizen <bas@basnieuwenhuizen.nl>
Tested-by: Bas Nieuwenhuizen <bas@basnieuwenhuizen.nl>
Instead of a single i2b and b2i, we now have i2b32 and b2iN where N is
one if 8, 16, 32, or 64. This leads to having a few more opcodes but
now everything is consistent and booleans aren't a weird special case
anymore.
Reviewed-by: Connor Abbott <cwabbott0@gmail.com>
All conversion opcodes require a destination size but this makes
constructing certain algebraic expressions rather cumbersome. This
commit adds support to nir_search and nir_algebraic for writing
conversion opcodes without a size. These meta-opcodes match any
conversion of that type regardless of destination size and the size gets
inferred from the sizes of the things being matched or from other
opcodes in the expression.
Reviewed-by: Connor Abbott <cwabbott0@gmail.com>
Instead of using an OrderedDict, just have a (necessarily sorted) array
of transforms and a set of opcodes.
Reviewed-by: Connor Abbott <cwabbott0@gmail.com>
Both of these things are already handled in the Value base class so we
don't need to handle them explicitly in Constant.
Reviewed-by: Connor Abbott <cwabbott0@gmail.com>
Before this commit, there were two copies of the algorithm: one in C,
that we would use to figure out what bit-size to give the replacement
expression, and one in Python, that emulated the C one and tried to
prove that the C algorithm would never fail to correctly assign
bit-sizes. That seemed pretty fragile, and likely to fall over if we
make any changes. Furthermore, the C code was really just recomputing
more-or-less the same thing as the Python code every time. Instead, we
can just store the results of the Python algorithm in the C
datastructure, and consult it to compute the bitsize of each value,
moving the "brains" entirely into Python. Since the Python algorithm no
longer has to match C, it's also a lot easier to change it to something
more closely approximating an actual type-inference algorithm. The
algorithm used is based on Hindley-Milner, although deliberately
weakened a little. It's a few more lines than the old one, judging by
the diffstat, but I think it's easier to verify that it's correct while
being as general as possible.
We could split this up into two changes, first making the C code use the
results of the Python code and then rewriting the Python algorithm, but
since the old algorithm never tracked which variable each equivalence
class, it would mean we'd have to add some non-trivial code which would
then get thrown away. I think it's better to see the final state all at
once, although I could also try splitting it up.
v2:
- Replace instances of "== None" and "!= None" with "is None" and
"is not None".
- Rename first_src to first_unsized_src
- Only merge the destination with the first unsized source, since the
sources have already been merged.
- Add a comment explaining what nir_search_value::bit_size now means.
v3:
- Fix one last instance to use "is not" instead of !=
- Don't try to be so clever when choosing which error message to print
based on whether we're in the search or replace expression.
- Fix trailing whitespace.
Reviewed-by: Jason Ekstrand <jason@jlekstrand.net>
Reviewed-by: Dylan Baker <dylan@pnwbakers.com>
This requires that we rework the interface a bit to use nir_builder but
that's a nice little modernization anyway.
Reviewed-by: Ian Romanick <ian.d.romanick@intel.com>
Reviewed-by: Eric Anholt <eric@anholt.net>
The conon_bit_class and canon_var_class variables got switched.
Fixes: 932c650e0b "nir/algebraic: Loosen a restriction on variables"
Reported-by: Ian Romanick <ian.d.romanick@intel.com>
Reviewed-by: Ian Romanick <ian.d.romanick@intel.com>
Previously, we would fail if a variable had an assigned but unknown bit
size X and we tried to assign it an actual bit size. However, this is
ok because, at the time we do the search, the variable does have an
actual bit size and it will match X because of the NIR rules.
Reviewed-by: Samuel Iglesias Gonsálvez <siglesias@igalia.com>
We rename some local variables in validate() to be more readable and
plumb the var through to get/set_var_bit_class instead of the var index.
Reviewed-by: Samuel Iglesias Gonsálvez <siglesias@igalia.com>
Python 3 lost the long type: now everything is an int, with the right
size.
This commit makes the script compatible with Python 2 (where we check
for both int and long) and Python 3 (where we only check for int).
Signed-off-by: Mathieu Bridon <bochecha@daitauha.fr>
Reviewed-by: Eric Engestrom <eric.engestrom@intel.com>
Reviewed-by: Dylan Baker <dylan@pnwbakers.com>
Mixing the two is a long-standing recipe for errors in Python 2, so much
so that Python 3 now completely separates them.
This commit stops treating both as if they were the same, and in the
process makes the script compatible with both Python 2 and 3.
Signed-off-by: Mathieu Bridon <bochecha@daitauha.fr>
Reviewed-by: Eric Engestrom <eric.engestrom@intel.com>
Reviewed-by: Dylan Baker <dylan@pnwbakers.com>
Python 2 had two integer types: int and long. Python 3 dropped the
latter, as it made the int type automatically support bigger numbers.
As a result, Python 3 lost the 'L' suffix on integer litterals.
This probably doesn't make much difference when compiling the generated
C code, but adding it explicitly means that both Python 2 and 3 generate
the exact same C code anyway, which makes it easier to compare and check
for discrepencies when moving to Python 3.
Signed-off-by: Mathieu Bridon <bochecha@daitauha.fr>
Reviewed-by: Dylan Baker <dylan@pnwbakers.com>
The hex() builtin returns a string containing the hexa-decimal
representation of an integer.
When the argument is not an integer, then the function calls that
object's __hex__() method, if one is defined. That method is supposed to
return a string.
While that's not explicitly documented, that string is supposed to be a
valid hexa-decimal representation for a number. Python 2 doesn't enforce
this though, which is why we got away with returning things like
'NIR_TRUE' which are not numbers.
In Python 3, the hex() builtin instead calls an object's __index__()
method, which itself must return an integer. That integer is then
automatically converted to a string with its hexa-decimal representation
by the rest of the hex() function.
As a result, we really can't make this compatible with Python 3 as it
is.
The solution is to stop using the hex() builtin, and instead use a hex()
object method, which can return whatever we want, in Python 2 and 3.
Signed-off-by: Mathieu Bridon <bochecha@daitauha.fr>
Reviewed-by: Eric Engestrom <eric.engestrom@intel.com>
Reviewed-by: Dylan Baker <dylan@pnwbakers.com>
In Python 2, iterators had a .next() method.
In Python 3, instead they have a .__next__() method, which is
automatically called by the next() builtin.
In addition, it is better to use the iter() builtin to create an
iterator, rather than calling its __iter__() method.
These were also introduced in Python 2.6, so using it makes the script
compatible with Python 2 and 3.
Signed-off-by: Mathieu Bridon <bochecha@daitauha.fr>
Reviewed-by: Eric Engestrom <eric.engestrom@intel.com>
Reviewed-by: Dylan Baker <dylan@pnwbakers.com>
In Python 2, dictionaries have 2 sets of methods to iterate over their
keys and values: keys()/values()/items() and iterkeys()/itervalues()/iteritems().
The former return lists while the latter return iterators.
Python 3 dropped the method which return lists, and renamed the methods
returning iterators to keys()/values()/items().
Using those names makes the scripts compatible with both Python 2 and 3.
Signed-off-by: Mathieu Bridon <bochecha@daitauha.fr>
Reviewed-by: Eric Engestrom <eric.engestrom@intel.com>
Reviewed-by: Dylan Baker <dylan@pnwbakers.com>
In Python, dictionaries and sets are unordered, and as a result their
is no guarantee that running this script twice will produce the same
output.
Using ordered dicts and explicitly sorting items makes the build more
reproducible, and will make it possible to verify that we're not
breaking anything when we move the build scripts to Python 3.
Reviewed-by: Eric Engestrom <eric.engestrom@intel.com>
We were including it once per value, so probably around 10k times.
Let's not cause the compiler any more work than we have to.
Reviewed-by: Iago Toral Quiroga <itoral@igalia.com>
Some optimizations, like converting integer multiply/divide into left/
right shifts, have additional constraints on the search expression.
Like requiring that a variable is a constant power of two. Support
these cases by allowing a fxn name to be appended to the search var
expression (ie. "a#32(is_power_of_two)").
Signed-off-by: Rob Clark <robclark@freedesktop.org>
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
Reviewed-by: Jason Ekstrand <jason@jlekstrand.net>
This matches the "foreach x in container" pattern found in many other
programming languages. Generated by the following regular expression:
s/nir_foreach_function(\([^,]*\),\s*\([^,]*\))/nir_foreach_function(\2, \1)/
Reviewed-by: Ian Romanick <ian.d.romanick@intel.com>
This matches the "foreach x in container" pattern found in many other
programming languages. Generated by the following regular expression:
s/nir_foreach_instr(\([^,]*\),\s*\([^,]*\))/nir_foreach_instr(\2, \1)/
and similar expressions for nir_foreach_instr_safe etc.
Reviewed-by: Ian Romanick <ian.d.romanick@intel.com>
This commit adds a validator that ensures that all expressions passed
through nir_algebraic are 100% non-ambiguous as far as bit-sizes are
concerned. This way it's a compile-time error rather than a hard-to-trace
C exception some time later.
Reviewed-by: Samuel Iglesias Gonsálvez <siglesias@igalia.com>
This is consistent with the rename done for the rest of NIR. Currently,
"bool" is the only type specifier used in nir_opt_algebraic.py so this is
really a no-op.
Reviewed-by: Samuel Iglesias Gonsálvez <siglesias@igalia.com>
Reviewed-by: Iago Toral Quiroga <itoral@igalia.com>
Previously, if an exception was encountered anywhere, nir_algebraic would
just die in a fire with no indication whatsoever as to where the actual bug
is. This commit makes it print out the particular search-and-replace
expression that is causing problems along with the exception. Also, it
will now report all of the errors it finds and then exit at the end like a
standard C compiler would do.
Reviewed-by: Dylan Baker <dylan@pnwbakers.com>
Reviewed-by: Iago Toral Quiroga <itoral@igalia.com>
Some passes may not refer to options->..., at which point the compiler
will warn about an unused variable. Just cast to void unconditionally
to shut it up.
Signed-off-by: Kenneth Graunke <kenneth@whitecape.org>
Reviewed-by: Eduardo Lima Mitev <elima@igalia.com>
Reviewed-by: Jason Ekstrand <jason@jlekstrand.net>
Reviewed-by: Matt Turner <mattst88@gmail.com>
When we replace an expresion we have to compute bitsize information for the
replacement. We do this in two passes to validate that bitsize information
is consistent and correct: first we propagate bitsize from child nodes to
parent, then we do it the other way around, starting from the original's
instruction destination bitsize.
v2 (Iago):
- Always use nir_type_bool32 instead of nir_type_bool when generating
algebraic optimizations. Before we used nir_type_bool32 with constants
and nir_type_bool with variables.
- Fix bool comparisons in nir_search.c to account for bitsized types.
v3 (Sam):
- Unpack the double constant value as unsigned long long (8 bytes) in
nir_algrebraic.py.
v4 (Sam):
- Use helpers to get type size and base type from nir_alu_type.
Signed-off-by: Iago Toral Quiroga <itoral@igalia.com>
Signed-off-by: Samuel Iglesias Gonsálvez <siglesias@igalia.com>
Reviewed-by: Jason Ekstrand <jason@jlekstrand.net>
Reviewed-by: Samuel Iglesias Gonsálvez <siglesias@igalia.com>
Reviewed-by: Iago Toral Quiroga <itoral@igalia.com>
The next patch adds an algebraic rule that uses the constant 0xff00ff00.
Without this change, the build fails with
return hex(struct.unpack('I', struct.pack('i', self.value))[0])
struct.error: 'i' format requires -2147483648 <= number <= 2147483647
The hex() function handles integers of any size, and assigning a
negative value to an unsigned does what we want in C. The pack/unpack is
unnecessary (and as we see, buggy).
Reviewed-by: Dylan Baker <baker.dylan.c@gmail.com>
Walking the SSA definitions in order means that we consider the smallest
algebraic optimizations before larger optimizations. So if a smaller
rule is part of a larger rule, the smaller one will happen first,
preventing the larger one from happening.
instructions in affected programs: 32721 -> 32611 (-0.34%)
helped: 106
In programs whose nir_optimize loop count changes (129 of them):
before: 1164 optimization loops
after: 1071 optimization loops
Of the 129 affected, 16 programs' optimization loop counts increased.
Prevents regressions and annoyances in the next commits.
Reviewed-by: Eduardo Lima Mitev <elima@igalia.com>
Reviewed-by: Jason Ekstrand <jason.ekstrand@intel.com>
