src/panfrost/midgard/midgard_nir_algebraic.py

   1 #
   2 # Copyright (C) 2018 Alyssa Rosenzweig
   3 #
   4 # Copyright (C) 2016 Intel Corporation
   5 #
   6 # Permission is hereby granted, free of charge, to any person obtaining a
   7 # copy of this software and associated documentation files (the "Software"),
   8 # to deal in the Software without restriction, including without limitation
   9 # the rights to use, copy, modify, merge, publish, distribute, sublicense,
  10 # and/or sell copies of the Software, and to permit persons to whom the
  11 # Software is furnished to do so, subject to the following conditions:
  12 #
  13 # The above copyright notice and this permission notice (including the next
  14 # paragraph) shall be included in all copies or substantial portions of the
  15 # Software.
  16 #
  17 # THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
  18 # IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
  19 # FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
  20 # THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
  21 # LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
  22 # FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS
  23 # IN THE SOFTWARE.
  24
  25 import argparse
  26 import sys
  27 import math
  28
  29 a = 'a'
  30 b = 'b'
  31 c = 'c'
  32
  33 algebraic_late = [
  34     # ineg must be lowered late, but only for integers; floats will try to
  35     # have modifiers attached... hence why this has to be here rather than
  36     # a more standard lower_negate approach
  37
  38     (('ineg', a), ('isub', 0, a)),
  39
  40     # These two special-cases save space/an op than the actual csel op +
  41     # scheduler flexibility
  42
  43     (('b32csel', a, 'b@32', 0), ('iand', a, b)),
  44     (('b32csel', a, 0, 'b@32'), ('iand', ('inot', a), b)),
  45 ]
  46
  47
  48 # Midgard is able to type convert down by only one "step" per instruction; if
  49 # NIR wants more than one step, we need to break up into multiple instructions
  50
  51 converts = []
  52
  53 for op in ('u2u', 'i2i', 'f2f', 'i2f', 'u2f', 'f2i', 'f2u'):
  54     srcsz_max = 64
  55     dstsz_max = 64
  56     # 8 bit float doesn't exist
  57     srcsz_min = 8 if op[0] != 'f' else 16
  58     dstsz_min = 8 if op[2] != 'f' else 16
  59     dstsz = dstsz_min
  60     # Iterate over all possible destination and source sizes
  61     while dstsz <= dstsz_max:
  62         srcsz = srcsz_min
  63         while srcsz <= srcsz_max:
  64             # Size converter lowering is only needed if src and dst sizes are
  65             # spaced by a factor > 2.
  66             # Type converter lowering is needed as soon as src_size != dst_size
  67             if srcsz != dstsz and ((srcsz * 2 != dstsz and srcsz != dstsz * 2) or op[0] != op[2]):
  68                 cursz = srcsz
  69                 rule = a
  70                 # When converting down we first do the type conversion followed
  71                 # by one or more size conversions. When converting up, we do
  72                 # the type conversion at the end. This way we don't have to
  73                 # deal with the fact that f2f8 doesn't exists.
  74                 sizeconvop = op[0] + '2' + op[0] if srcsz < dstsz else op[2] + '2' + op[2]
  75                 if srcsz > dstsz and op[0] != op[2]:
  76                     rule = (op + str(int(cursz)), rule)
  77                 while cursz != dstsz:
  78                     cursz = cursz / 2 if dstsz < srcsz else cursz * 2
  79                     rule = (sizeconvop + str(int(cursz)), rule)
  80                 if srcsz < dstsz and op[0] != op[2]:
  81                     rule = (op + str(int(cursz)), rule)
  82                 converts += [((op + str(int(dstsz)), 'a@' + str(int(srcsz))), rule)]
  83             srcsz *= 2
  84         dstsz *= 2
  85
  86 # Midgard scales fsin/fcos arguments by pi.
  87 # Pass must be run only once, after the main loop
  88
  89 scale_trig = [
  90         (('fsin', a), ('fsin', ('fdiv', a, math.pi))),
  91         (('fcos', a), ('fcos', ('fdiv', a, math.pi))),
  92 ]
  93
  94 def main():
  95     parser = argparse.ArgumentParser()
  96     parser.add_argument('-p', '--import-path', required=True)
  97     args = parser.parse_args()
  98     sys.path.insert(0, args.import_path)
  99     run()
 100
 101
 102 def run():
 103     import nir_algebraic  # pylint: disable=import-error
 104
 105     print('#include "midgard_nir.h"')
 106
 107     print(nir_algebraic.AlgebraicPass("midgard_nir_lower_algebraic_late",
 108                                       algebraic_late + converts).render())
 109
 110     print(nir_algebraic.AlgebraicPass("midgard_nir_scale_trig",
 111                                       scale_trig).render())
 112
 113
 114 if __name__ == '__main__':
 115     main()