src/gallium/drivers/freedreno/ir3/ir3_nir.c

   1 /* -*- mode: C; c-file-style: "k&r"; tab-width 4; indent-tabs-mode: t; -*- */
   2
   3 /*
   4  * Copyright (C) 2015 Rob Clark <robclark@freedesktop.org>
   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 FROM,
  22  * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
  23  * SOFTWARE.
  24  *
  25  * Authors:
  26  *    Rob Clark <robclark@freedesktop.org>
  27  */
  28
  29
  30 #include "freedreno_util.h"
  31
  32 #include "ir3_nir.h"
  33 #include "ir3_compiler.h"
  34 #include "ir3_shader.h"
  35
  36 #include "nir/tgsi_to_nir.h"
  37
  38 static const nir_shader_compiler_options options = {
  39                 .lower_fpow = true,
  40                 .lower_fsat = true,
  41                 .lower_scmp = true,
  42                 .lower_flrp32 = true,
  43                 .lower_flrp64 = true,
  44                 .lower_ffract = true,
  45                 .lower_fmod32 = true,
  46                 .lower_fmod64 = true,
  47                 .lower_fdiv = true,
  48                 .fuse_ffma = true,
  49                 .native_integers = true,
  50                 .vertex_id_zero_based = true,
  51                 .lower_extract_byte = true,
  52                 .lower_extract_word = true,
  53 };
  54
  55 static const nir_shader_compiler_options options_5xx = {
  56                 .lower_fpow = true,
  57                 .lower_fsat = true,
  58                 .lower_scmp = true,
  59                 .lower_flrp32 = true,
  60                 .lower_flrp64 = true,
  61                 .lower_ffract = true,
  62                 .lower_fmod32 = true,
  63                 .lower_fmod64 = true,
  64                 .lower_fdiv = true,
  65                 .fuse_ffma = true,
  66                 .native_integers = true,
  67                 .vertex_id_zero_based = false,
  68                 .lower_extract_byte = true,
  69                 .lower_extract_word = true,
  70 };
  71
  72 struct nir_shader *
  73 ir3_tgsi_to_nir(const struct tgsi_token *tokens)
  74 {
  75         return tgsi_to_nir(tokens, &options);
  76 }
  77
  78 const nir_shader_compiler_options *
  79 ir3_get_compiler_options(struct ir3_compiler *compiler)
  80 {
  81         if (compiler->gpu_id >= 500)
  82                 return &options_5xx;
  83         return &options;
  84 }
  85
  86 /* for given shader key, are any steps handled in nir? */
  87 bool
  88 ir3_key_lowers_nir(const struct ir3_shader_key *key)
  89 {
  90         return key->fsaturate_s | key->fsaturate_t | key->fsaturate_r |
  91                         key->vsaturate_s | key->vsaturate_t | key->vsaturate_r |
  92                         key->ucp_enables | key->color_two_side |
  93                         key->fclamp_color | key->vclamp_color;
  94 }
  95
  96 #define OPT(nir, pass, ...) ({                             \
  97    bool this_progress = false;                             \
  98    NIR_PASS(this_progress, nir, pass, ##__VA_ARGS__);      \
  99    this_progress;                                          \
 100 })
 101
 102 #define OPT_V(nir, pass, ...) NIR_PASS_V(nir, pass, ##__VA_ARGS__)
 103
 104 static void
 105 ir3_optimize_loop(nir_shader *s)
 106 {
 107         bool progress;
 108         do {
 109                 progress = false;
 110
 111                 OPT_V(s, nir_lower_vars_to_ssa);
 112                 progress |= OPT(s, nir_opt_copy_prop_vars);
 113                 progress |= OPT(s, nir_lower_alu_to_scalar);
 114                 progress |= OPT(s, nir_lower_phis_to_scalar);
 115
 116                 progress |= OPT(s, nir_copy_prop);
 117                 progress |= OPT(s, nir_opt_dce);
 118                 progress |= OPT(s, nir_opt_cse);
 119                 progress |= OPT(s, ir3_nir_lower_if_else);
 120                 progress |= OPT(s, nir_opt_algebraic);
 121                 progress |= OPT(s, nir_opt_constant_folding);
 122
 123         } while (progress);
 124 }
 125
 126 struct nir_shader *
 127 ir3_optimize_nir(struct ir3_shader *shader, nir_shader *s,
 128                 const struct ir3_shader_key *key)
 129 {
 130         struct nir_lower_tex_options tex_options = {
 131                         .lower_rect = 0,
 132         };
 133
 134         if (key) {
 135                 switch (shader->type) {
 136                 case SHADER_FRAGMENT:
 137                         tex_options.saturate_s = key->fsaturate_s;
 138                         tex_options.saturate_t = key->fsaturate_t;
 139                         tex_options.saturate_r = key->fsaturate_r;
 140                         break;
 141                 case SHADER_VERTEX:
 142                         tex_options.saturate_s = key->vsaturate_s;
 143                         tex_options.saturate_t = key->vsaturate_t;
 144                         tex_options.saturate_r = key->vsaturate_r;
 145                         break;
 146                 default:
 147                         /* TODO */
 148                         break;
 149                 }
 150         }
 151
 152         if (shader->compiler->gpu_id >= 400) {
 153                 /* a4xx seems to have *no* sam.p */
 154                 tex_options.lower_txp = ~0;  /* lower all txp */
 155         } else {
 156                 /* a3xx just needs to avoid sam.p for 3d tex */
 157                 tex_options.lower_txp = (1 << GLSL_SAMPLER_DIM_3D);
 158         }
 159
 160         if (fd_mesa_debug & FD_DBG_DISASM) {
 161                 debug_printf("----------------------\n");
 162                 nir_print_shader(s, stdout);
 163                 debug_printf("----------------------\n");
 164         }
 165
 166         OPT_V(s, nir_opt_global_to_local);
 167         OPT_V(s, nir_lower_regs_to_ssa);
 168
 169         if (key) {
 170                 if (s->info.stage == MESA_SHADER_VERTEX) {
 171                         OPT_V(s, nir_lower_clip_vs, key->ucp_enables);
 172                         if (key->vclamp_color)
 173                                 OPT_V(s, nir_lower_clamp_color_outputs);
 174                 } else if (s->info.stage == MESA_SHADER_FRAGMENT) {
 175                         OPT_V(s, nir_lower_clip_fs, key->ucp_enables);
 176                         if (key->fclamp_color)
 177                                 OPT_V(s, nir_lower_clamp_color_outputs);
 178                 }
 179                 if (key->color_two_side) {
 180                         OPT_V(s, nir_lower_two_sided_color);
 181                 }
 182         } else {
 183                 /* only want to do this the first time (when key is null)
 184                  * and not again on any potential 2nd variant lowering pass:
 185                  */
 186                 OPT_V(s, ir3_nir_apply_trig_workarounds);
 187         }
 188
 189         OPT_V(s, nir_lower_tex, &tex_options);
 190         OPT_V(s, nir_lower_load_const_to_scalar);
 191
 192         ir3_optimize_loop(s);
 193
 194         /* do idiv lowering after first opt loop to give a chance for
 195          * divide by immed power-of-two to be caught first:
 196          */
 197         if (OPT(s, nir_lower_idiv))
 198                 ir3_optimize_loop(s);
 199
 200         OPT_V(s, nir_remove_dead_variables, nir_var_local);
 201
 202         if (fd_mesa_debug & FD_DBG_DISASM) {
 203                 debug_printf("----------------------\n");
 204                 nir_print_shader(s, stdout);
 205                 debug_printf("----------------------\n");
 206         }
 207
 208         nir_sweep(s);
 209
 210         return s;
 211 }