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     # Likewise we want fsub lowered but not isub
  41     (('fsub', a, b), ('fadd', a, ('fneg', b))),
  42
  43     # These two special-cases save space/an op than the actual csel op +
  44     # scheduler flexibility
  45
  46     (('b32csel', a, 'b@32', 0), ('iand', a, b)),
  47     (('b32csel', a, 0, 'b@32'), ('iand', ('inot', a), b)),
  48 ]
  49
  50
  51 # Midgard is able to type convert down by only one "step" per instruction; if
  52 # NIR wants more than one step, we need to break up into multiple instructions
  53
  54 converts = []
  55
  56 for op in ('u2u', 'i2i', 'f2f', 'i2f', 'u2f', 'f2i', 'f2u'):
  57     srcsz_max = 64
  58     dstsz_max = 64
  59     # 8 bit float doesn't exist
  60     srcsz_min = 8 if op[0] != 'f' else 16
  61     dstsz_min = 8 if op[2] != 'f' else 16
  62     dstsz = dstsz_min
  63     # Iterate over all possible destination and source sizes
  64     while dstsz <= dstsz_max:
  65         srcsz = srcsz_min
  66         while srcsz <= srcsz_max:
  67             # Size converter lowering is only needed if src and dst sizes are
  68             # spaced by a factor > 2.
  69             # Type converter lowering is needed as soon as src_size != dst_size
  70             if srcsz != dstsz and ((srcsz * 2 != dstsz and srcsz != dstsz * 2) or op[0] != op[2]):
  71                 cursz = srcsz
  72                 rule = a
  73                 # When converting down we first do the type conversion followed
  74                 # by one or more size conversions. When converting up, we do
  75                 # the type conversion at the end. This way we don't have to
  76                 # deal with the fact that f2f8 doesn't exists.
  77                 sizeconvop = op[0] + '2' + op[0] if srcsz < dstsz else op[2] + '2' + op[2]
  78                 if srcsz > dstsz and op[0] != op[2]:
  79                     rule = (op + str(int(cursz)), rule)
  80                 while cursz != dstsz:
  81                     cursz = cursz / 2 if dstsz < srcsz else cursz * 2
  82                     rule = (sizeconvop + str(int(cursz)), rule)
  83                 if srcsz < dstsz and op[0] != op[2]:
  84                     rule = (op + str(int(cursz)), rule)
  85                 converts += [((op + str(int(dstsz)), 'a@' + str(int(srcsz))), rule)]
  86             srcsz *= 2
  87         dstsz *= 2
  88
  89 # Midgard outputs fp32 for specials. The f2f32 will be folded in later.
  90 SPECIAL = ['fexp2', 'flog2', 'fsin', 'fcos', 'frcp', 'frsq']
  91
  92 for op in SPECIAL:
  93         converts += [((op + '@16', a), ('f2f16', (op, ('f2f32', a))))]
  94
  95 # Midgard scales fsin/fcos arguments by pi.
  96 # Pass must be run only once, after the main loop
  97
  98 scale_trig = [
  99         (('fsin', a), ('fsin', ('fdiv', a, math.pi))),
 100         (('fcos', a), ('fcos', ('fdiv', a, math.pi))),
 101 ]
 102
 103 def main():
 104     parser = argparse.ArgumentParser()
 105     parser.add_argument('-p', '--import-path', required=True)
 106     args = parser.parse_args()
 107     sys.path.insert(0, args.import_path)
 108     run()
 109
 110
 111 def run():
 112     import nir_algebraic  # pylint: disable=import-error
 113
 114     print('#include "midgard_nir.h"')
 115
 116     print(nir_algebraic.AlgebraicPass("midgard_nir_lower_algebraic_late",
 117                                       algebraic_late + converts).render())
 118
 119     print(nir_algebraic.AlgebraicPass("midgard_nir_scale_trig",
 120                                       scale_trig).render())
 121
 122
 123 if __name__ == '__main__':
 124     main()