Home | History | Annotate | Download | only in i965
      1 #! /usr/bin/env python
      2 #
      3 # Copyright (C) 2016 Intel Corporation
      4 #
      5 # Permission is hereby granted, free of charge, to any person obtaining a
      6 # copy of this software and associated documentation files (the "Software"),
      7 # to deal in the Software without restriction, including without limitation
      8 # the rights to use, copy, modify, merge, publish, distribute, sublicense,
      9 # and/or sell copies of the Software, and to permit persons to whom the
     10 # Software is furnished to do so, subject to the following conditions:
     11 #
     12 # The above copyright notice and this permission notice (including the next
     13 # paragraph) shall be included in all copies or substantial portions of the
     14 # Software.
     15 #
     16 # THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
     17 # IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
     18 # FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
     19 # THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
     20 # LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
     21 # FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS
     22 # IN THE SOFTWARE.
     23 
     24 import nir_algebraic
     25 
     26 # Prior to Kaby Lake, The SIN and COS instructions on Intel hardware can
     27 # produce values slightly outside of the [-1.0, 1.0] range for a small set of
     28 # values.  Obviously, this can break everyone's expectations about trig
     29 # functions.  This appears to be fixed in Kaby Lake.
     30 #
     31 # According to an internal presentation, the COS instruction can produce
     32 # a value up to 1.000027 for inputs in the range (0.08296, 0.09888).  One
     33 # suggested workaround is to multiply by 0.99997, scaling down the
     34 # amplitude slightly.  Apparently this also minimizes the error function,
     35 # reducing the maximum error from 0.00006 to about 0.00003.
     36 
     37 trig_workarounds = [
     38    (('fsin', 'x'), ('fmul', ('fsin', 'x'), 0.99997)),
     39    (('fcos', 'x'), ('fmul', ('fcos', 'x'), 0.99997)),
     40 ]
     41 
     42 print '#include "brw_nir.h"'
     43 print nir_algebraic.AlgebraicPass("brw_nir_apply_trig_workarounds",
     44                                   trig_workarounds).render()
     45