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     (('i2i8', 'a@32'), ('i2i8', ('i2i16', a))),
  53     (('u2u8', 'a@32'), ('u2u8', ('u2u16', a))),
  54
  55     (('i2i32', 'a@8'), ('i2i32', ('i2i16', a))),
  56     (('u2u32', 'a@8'), ('u2u32', ('u2u16', a))),
  57
  58     (('f2i32', 'a@16'), ('f2i32', ('f2f32', a))),
  59     (('f2u32', 'a@16'), ('f2u32', ('f2f32', a))),
  60
  61     # Totally redundant
  62     (('~f2f16', ('f2f32', 'a@16')), a),
  63
  64     (('pack_half_2x16_split', 'a@32', 'b@32'), ('ior', ('ishl', ('i2i32', ('f2f16', b)), 16), ('i2i32', ('f2f16', a)))),
  65 ]
  66
  67 # Midgard scales fsin/fcos arguments by pi.
  68 # Pass must be run only once, after the main loop
  69
  70 scale_trig = [
  71         (('fsin', a), ('fsin', ('fdiv', a, math.pi))),
  72         (('fcos', a), ('fcos', ('fdiv', a, math.pi))),
  73 ]
  74
  75 def main():
  76     parser = argparse.ArgumentParser()
  77     parser.add_argument('-p', '--import-path', required=True)
  78     args = parser.parse_args()
  79     sys.path.insert(0, args.import_path)
  80     run()
  81
  82
  83 def run():
  84     import nir_algebraic  # pylint: disable=import-error
  85
  86     print('#include "midgard_nir.h"')
  87
  88     print(nir_algebraic.AlgebraicPass("midgard_nir_lower_algebraic_late",
  89                                       algebraic_late + converts).render())
  90
  91     print(nir_algebraic.AlgebraicPass("midgard_nir_scale_trig",
  92                                       scale_trig).render())
  93
  94
  95 if __name__ == '__main__':
  96     main()