src/compiler/nir/nir_lower_int_to_float.c

   1 /*
   2  * Copyright © 2018 Intel Corporation
   3  * Copyright © 2019 Vasily Khoruzhick <anarsoul@gmail.com>
   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
  25 #include "nir.h"
  26 #include "nir_builder.h"
  27
  28 static bool
  29 assert_ssa_def_is_not_int(nir_ssa_def *def, void *arg)
  30 {
  31    ASSERTED BITSET_WORD *int_types = arg;
  32    assert(!BITSET_TEST(int_types, def->index));
  33    return true;
  34 }
  35
  36 static bool
  37 lower_alu_instr(nir_builder *b, nir_alu_instr *alu)
  38 {
  39    const nir_op_info *info = &nir_op_infos[alu->op];
  40
  41    bool is_bool_only = alu->dest.dest.ssa.bit_size == 1;
  42    for (unsigned i = 0; i < info->num_inputs; i++) {
  43       if (alu->src[i].src.ssa->bit_size != 1)
  44          is_bool_only = false;
  45    }
  46
  47    if (is_bool_only) {
  48       /* avoid lowering integers ops are used for booleans (ieq,ine,etc) */
  49       return false;
  50    }
  51
  52    /* Replacement SSA value */
  53    nir_ssa_def *rep = NULL;
  54    switch (alu->op) {
  55    case nir_op_mov:
  56    case nir_op_vec2:
  57    case nir_op_vec3:
  58    case nir_op_vec4:
  59    case nir_op_bcsel:
  60       /* These we expect to have integers but the opcode doesn't change */
  61       break;
  62
  63    case nir_op_b2i32: alu->op = nir_op_b2f32; break;
  64    case nir_op_i2f32: alu->op = nir_op_mov; break;
  65    case nir_op_u2f32: alu->op = nir_op_mov; break;
  66    case nir_op_f2i32: alu->op = nir_op_ftrunc; break;
  67    case nir_op_f2u32: alu->op = nir_op_ffloor; break;
  68    case nir_op_i2b1: alu->op = nir_op_f2b1; break;
  69
  70    case nir_op_ilt: alu->op = nir_op_flt; break;
  71    case nir_op_ige: alu->op = nir_op_fge; break;
  72    case nir_op_ieq: alu->op = nir_op_feq; break;
  73    case nir_op_ine: alu->op = nir_op_fneu; break;
  74    case nir_op_ult: alu->op = nir_op_flt; break;
  75    case nir_op_uge: alu->op = nir_op_fge; break;
  76
  77    case nir_op_iadd: alu->op = nir_op_fadd; break;
  78    case nir_op_isub: alu->op = nir_op_fsub; break;
  79    case nir_op_imul: alu->op = nir_op_fmul; break;
  80    case nir_op_idiv:
  81       rep = nir_ftrunc(b, nir_fdiv(b,
  82                                    nir_ssa_for_alu_src(b, alu, 0),
  83                                    nir_ssa_for_alu_src(b, alu, 1)));
  84       break;
  85    case nir_op_iabs: alu->op = nir_op_fabs; break;
  86    case nir_op_ineg: alu->op = nir_op_fneg; break;
  87    case nir_op_imax: alu->op = nir_op_fmax; break;
  88    case nir_op_imin: alu->op = nir_op_fmin; break;
  89    case nir_op_umax: alu->op = nir_op_fmax; break;
  90    case nir_op_umin: alu->op = nir_op_fmin; break;
  91
  92    case nir_op_ball_iequal2:  alu->op = nir_op_ball_fequal2; break;
  93    case nir_op_ball_iequal3:  alu->op = nir_op_ball_fequal3; break;
  94    case nir_op_ball_iequal4:  alu->op = nir_op_ball_fequal4; break;
  95    case nir_op_bany_inequal2: alu->op = nir_op_bany_fnequal2; break;
  96    case nir_op_bany_inequal3: alu->op = nir_op_bany_fnequal3; break;
  97    case nir_op_bany_inequal4: alu->op = nir_op_bany_fnequal4; break;
  98
  99    default:
 100       assert(nir_alu_type_get_base_type(info->output_type) != nir_type_int &&
 101              nir_alu_type_get_base_type(info->output_type) != nir_type_uint);
 102       for (unsigned i = 0; i < info->num_inputs; i++) {
 103          assert(nir_alu_type_get_base_type(info->input_types[i]) != nir_type_int &&
 104                 nir_alu_type_get_base_type(info->input_types[i]) != nir_type_uint);
 105       }
 106       return false;
 107    }
 108
 109    if (rep) {
 110       /* We've emitted a replacement instruction */
 111       nir_ssa_def_rewrite_uses(&alu->dest.dest.ssa, nir_src_for_ssa(rep));
 112       nir_instr_remove(&alu->instr);
 113    }
 114
 115    return true;
 116 }
 117
 118 static bool
 119 nir_lower_int_to_float_impl(nir_function_impl *impl)
 120 {
 121    bool progress = false;
 122    BITSET_WORD *float_types = NULL, *int_types = NULL;
 123
 124    nir_builder b;
 125    nir_builder_init(&b, impl);
 126
 127    nir_index_ssa_defs(impl);
 128    float_types = calloc(BITSET_WORDS(impl->ssa_alloc),
 129                         sizeof(BITSET_WORD));
 130    int_types = calloc(BITSET_WORDS(impl->ssa_alloc),
 131                       sizeof(BITSET_WORD));
 132    nir_gather_ssa_types(impl, float_types, int_types);
 133
 134    nir_foreach_block(block, impl) {
 135       nir_foreach_instr_safe(instr, block) {
 136          switch (instr->type) {
 137          case nir_instr_type_alu:
 138             progress |= lower_alu_instr(&b, nir_instr_as_alu(instr));
 139             break;
 140
 141          case nir_instr_type_load_const: {
 142             nir_load_const_instr *load = nir_instr_as_load_const(instr);
 143             if (load->def.bit_size != 1 && BITSET_TEST(int_types, load->def.index)) {
 144                for (unsigned i = 0; i < load->def.num_components; i++)
 145                   load->value[i].f32 = load->value[i].i32;
 146             }
 147             break;
 148          }
 149
 150          case nir_instr_type_intrinsic:
 151          case nir_instr_type_ssa_undef:
 152          case nir_instr_type_phi:
 153          case nir_instr_type_tex:
 154             break;
 155
 156          default:
 157             nir_foreach_ssa_def(instr, assert_ssa_def_is_not_int, (void *)int_types);
 158             break;
 159          }
 160       }
 161    }
 162
 163    if (progress) {
 164       nir_metadata_preserve(impl, nir_metadata_block_index |
 165                                   nir_metadata_dominance);
 166    }
 167
 168    free(float_types);
 169    free(int_types);
 170
 171    return progress;
 172 }
 173
 174 bool
 175 nir_lower_int_to_float(nir_shader *shader)
 176 {
 177    bool progress = false;
 178
 179    nir_foreach_function(function, shader) {
 180       if (function->impl && nir_lower_int_to_float_impl(function->impl))
 181          progress = true;
 182    }
 183
 184    return progress;
 185 }