src/compiler/nir/nir_lower_sincos.c

   1 /*
   2  * Copyright (c) 2014 Scott Mansell
   3  * Copyright © 2014 Broadcom
   4  * Copyright (c) 2019 Vasily Khoruzhick <anarsoul@gmail.com>
   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
  26 #include "util/u_math.h"
  27 #include "nir.h"
  28 #include "nir_builder.h"
  29
  30 static nir_ssa_def *
  31 shrink_input(nir_builder *b, nir_ssa_def *x)
  32 {
  33    nir_ssa_def *scaled_x = nir_fmul_imm(b, x, 1.0 / (M_PI * 2));
  34
  35    nir_ssa_def *xfrac = nir_ffract(b, scaled_x);
  36    /* Map [0.5, 1] to [-0.5, 0] */
  37    nir_ssa_def *xfrac_0_5_1 = nir_fadd_imm(b, xfrac, -1.0);
  38    /* Map [-1, -0.5] to [0, 0.5] */
  39    nir_ssa_def *xfrac_n1_n0_5 = nir_fadd_imm(b, xfrac, 1.0);
  40
  41    nir_ssa_def *geq_0_5 = nir_build_alu(b, nir_op_fge, xfrac, nir_imm_float(b, 0.5), NULL, NULL);
  42    nir_ssa_def *less_m0_5 = nir_build_alu(b, nir_op_flt, xfrac, nir_imm_float(b, -0.5), NULL, NULL);
  43
  44    nir_ssa_def *sel1 = nir_build_alu(b, nir_op_bcsel, geq_0_5, xfrac_0_5_1, xfrac, NULL );
  45    nir_ssa_def *sel2 = nir_build_alu(b, nir_op_bcsel, less_m0_5, xfrac_n1_n0_5, sel1, NULL);
  46
  47    return sel2;
  48 }
  49
  50 static nir_ssa_def *
  51 lower_sincos(nir_builder *b, nir_ssa_def *src, bool do_cos)
  52 {
  53    /* Fast sin/cos implementation, see
  54     * https://web.archive.org/web/20180105155939/http://forum.devmaster.net/t/fast-and-accurate-sine-cosine/9648
  55     */
  56    const float B = 8.0; //(M_PI * 2.0) * (4.0 / M_PI);
  57    const float C = -16.0; //(M_PI * 2.0) * (M_PI * 2.0) * (-4.0 / (M_PI * M_PI));
  58    const float P = 0.225;
  59
  60    if (do_cos)
  61       src = nir_fadd_imm(b, src, M_PI / 2.0);
  62
  63    nir_ssa_def *x = shrink_input(b, src);
  64
  65    nir_ssa_def *bx = nir_fmul_imm(b, x, B);
  66    nir_ssa_def *cx = nir_fmul_imm(b, x, C);
  67    nir_ssa_def *absx = nir_fabs(b, x);
  68    nir_ssa_def *cxabsx = nir_fmul(b, cx, absx);
  69
  70    /* Y1 = B * x + C * x * fabs(x) */
  71    nir_ssa_def *y1 = nir_fadd(b, bx, cxabsx);
  72
  73    /* Precision step: Y = P * (Y1 * fabs(Y1) - Y1) + Y1 */
  74    nir_ssa_def *y = nir_fabs(b, y1);
  75    y = nir_fmul(b, y, y1);
  76    y = nir_fsub(b, y, y1);
  77    y = nir_fmul_imm(b, y, P);
  78    y = nir_fadd(b, y, y1);
  79
  80    return y;
  81 }
  82
  83 static bool
  84 lower_sincos_impl(nir_function_impl *impl)
  85 {
  86    bool progress = false;
  87
  88    nir_builder b;
  89    nir_builder_init(&b, impl);
  90
  91    nir_foreach_block(block, impl) {
  92       nir_foreach_instr_safe(instr, block) {
  93          if (instr->type != nir_instr_type_alu)
  94             continue;
  95
  96          nir_alu_instr *alu_instr = nir_instr_as_alu(instr);
  97          nir_ssa_def *lower;
  98
  99          b.cursor = nir_before_instr(instr);
 100
 101          switch (alu_instr->op) {
 102          case nir_op_fsin:
 103             lower = lower_sincos(&b, nir_ssa_for_alu_src(&b, alu_instr, 0), false);
 104             break;
 105          case nir_op_fcos:
 106             lower = lower_sincos(&b, nir_ssa_for_alu_src(&b, alu_instr, 0), true);
 107             break;
 108          default:
 109             continue;
 110          }
 111
 112          nir_ssa_def_rewrite_uses(&alu_instr->dest.dest.ssa,
 113                                   nir_src_for_ssa(lower));
 114          nir_instr_remove(instr);
 115          progress = true;
 116       }
 117    }
 118
 119    if (progress) {
 120       nir_metadata_preserve(impl, nir_metadata_block_index |
 121                                   nir_metadata_dominance);
 122    }
 123
 124    return progress;
 125 }
 126
 127 bool
 128 nir_lower_sincos(nir_shader *shader)
 129 {
 130    bool progress = false;
 131
 132    nir_foreach_function(function, shader) {
 133       if (function->impl)
 134          progress |= lower_sincos_impl(function->impl);
 135    }
 136
 137    return progress;
 138 }