src/freedreno/ir3/ir3_nir.c

   1 /*
   2  * Copyright (C) 2015 Rob Clark <robclark@freedesktop.org>
   3  *
   4  * Permission is hereby granted, free of charge, to any person obtaining a
   5  * copy of this software and associated documentation files (the "Software"),
   6  * to deal in the Software without restriction, including without limitation
   7  * the rights to use, copy, modify, merge, publish, distribute, sublicense,
   8  * and/or sell copies of the Software, and to permit persons to whom the
   9  * Software is furnished to do so, subject to the following conditions:
  10  *
  11  * The above copyright notice and this permission notice (including the next
  12  * paragraph) shall be included in all copies or substantial portions of the
  13  * Software.
  14  *
  15  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
  16  * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
  17  * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
  18  * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
  19  * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
  20  * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
  21  * SOFTWARE.
  22  *
  23  * Authors:
  24  *    Rob Clark <robclark@freedesktop.org>
  25  */
  26
  27
  28 #include "util/debug.h"
  29
  30 #include "ir3_nir.h"
  31 #include "ir3_compiler.h"
  32 #include "ir3_shader.h"
  33
  34 static const nir_shader_compiler_options options = {
  35                 .lower_fpow = true,
  36                 .lower_scmp = true,
  37                 .lower_flrp32 = true,
  38                 .lower_flrp64 = true,
  39                 .lower_ffract = true,
  40                 .lower_fmod32 = true,
  41                 .lower_fmod64 = true,
  42                 .lower_fdiv = true,
  43                 .lower_isign = true,
  44                 .lower_ldexp = true,
  45                 .fuse_ffma = true,
  46                 .native_integers = true,
  47                 .vertex_id_zero_based = true,
  48                 .lower_extract_byte = true,
  49                 .lower_extract_word = true,
  50                 .lower_all_io_to_temps = true,
  51                 .lower_helper_invocation = true,
  52 };
  53
  54 /* we don't want to lower vertex_id to _zero_based on newer gpus: */
  55 static const nir_shader_compiler_options options_a6xx = {
  56                 .lower_fpow = true,
  57                 .lower_scmp = true,
  58                 .lower_flrp32 = true,
  59                 .lower_flrp64 = true,
  60                 .lower_ffract = true,
  61                 .lower_fmod32 = true,
  62                 .lower_fmod64 = true,
  63                 .lower_fdiv = true,
  64                 .lower_isign = true,
  65                 .lower_ldexp = true,
  66                 .fuse_ffma = true,
  67                 .native_integers = true,
  68                 .vertex_id_zero_based = false,
  69                 .lower_extract_byte = true,
  70                 .lower_extract_word = true,
  71                 .lower_all_io_to_temps = true,
  72                 .lower_helper_invocation = true,
  73 };
  74
  75 const nir_shader_compiler_options *
  76 ir3_get_compiler_options(struct ir3_compiler *compiler)
  77 {
  78         if (compiler->gpu_id >= 600)
  79                 return &options_a6xx;
  80         return &options;
  81 }
  82
  83 /* for given shader key, are any steps handled in nir? */
  84 bool
  85 ir3_key_lowers_nir(const struct ir3_shader_key *key)
  86 {
  87         return key->fsaturate_s | key->fsaturate_t | key->fsaturate_r |
  88                         key->vsaturate_s | key->vsaturate_t | key->vsaturate_r |
  89                         key->ucp_enables | key->color_two_side |
  90                         key->fclamp_color | key->vclamp_color;
  91 }
  92
  93 #define OPT(nir, pass, ...) ({                             \
  94    bool this_progress = false;                             \
  95    NIR_PASS(this_progress, nir, pass, ##__VA_ARGS__);      \
  96    this_progress;                                          \
  97 })
  98
  99 #define OPT_V(nir, pass, ...) NIR_PASS_V(nir, pass, ##__VA_ARGS__)
 100
 101 static void
 102 ir3_optimize_loop(nir_shader *s)
 103 {
 104         bool progress;
 105         do {
 106                 progress = false;
 107
 108                 OPT_V(s, nir_lower_vars_to_ssa);
 109                 progress |= OPT(s, nir_opt_copy_prop_vars);
 110                 progress |= OPT(s, nir_opt_dead_write_vars);
 111                 progress |= OPT(s, nir_lower_alu_to_scalar);
 112                 progress |= OPT(s, nir_lower_phis_to_scalar);
 113
 114                 progress |= OPT(s, nir_copy_prop);
 115                 progress |= OPT(s, nir_opt_dce);
 116                 progress |= OPT(s, nir_opt_cse);
 117                 static int gcm = -1;
 118                 if (gcm == -1)
 119                         gcm = env_var_as_unsigned("GCM", 0);
 120                 if (gcm == 1)
 121                         progress |= OPT(s, nir_opt_gcm, true);
 122                 else if (gcm == 2)
 123                         progress |= OPT(s, nir_opt_gcm, false);
 124                 progress |= OPT(s, nir_opt_peephole_select, 16, true, true);
 125                 progress |= OPT(s, nir_opt_intrinsics);
 126                 progress |= OPT(s, nir_opt_algebraic);
 127                 progress |= OPT(s, nir_opt_constant_folding);
 128                 progress |= OPT(s, nir_opt_dead_cf);
 129                 if (OPT(s, nir_opt_trivial_continues)) {
 130                         progress |= true;
 131                         /* If nir_opt_trivial_continues makes progress, then we need to clean
 132                          * things up if we want any hope of nir_opt_if or nir_opt_loop_unroll
 133                          * to make progress.
 134                          */
 135                         OPT(s, nir_copy_prop);
 136                         OPT(s, nir_opt_dce);
 137                 }
 138                 progress |= OPT(s, nir_opt_if);
 139                 progress |= OPT(s, nir_opt_remove_phis);
 140                 progress |= OPT(s, nir_opt_undef);
 141
 142         } while (progress);
 143 }
 144
 145 struct nir_shader *
 146 ir3_optimize_nir(struct ir3_shader *shader, nir_shader *s,
 147                 const struct ir3_shader_key *key)
 148 {
 149         struct nir_lower_tex_options tex_options = {
 150                         .lower_rect = 0,
 151         };
 152
 153         if (key) {
 154                 switch (shader->type) {
 155                 case MESA_SHADER_FRAGMENT:
 156                         tex_options.saturate_s = key->fsaturate_s;
 157                         tex_options.saturate_t = key->fsaturate_t;
 158                         tex_options.saturate_r = key->fsaturate_r;
 159                         break;
 160                 case MESA_SHADER_VERTEX:
 161                         tex_options.saturate_s = key->vsaturate_s;
 162                         tex_options.saturate_t = key->vsaturate_t;
 163                         tex_options.saturate_r = key->vsaturate_r;
 164                         break;
 165                 default:
 166                         /* TODO */
 167                         break;
 168                 }
 169         }
 170
 171         if (shader->compiler->gpu_id >= 400) {
 172                 /* a4xx seems to have *no* sam.p */
 173                 tex_options.lower_txp = ~0;  /* lower all txp */
 174         } else {
 175                 /* a3xx just needs to avoid sam.p for 3d tex */
 176                 tex_options.lower_txp = (1 << GLSL_SAMPLER_DIM_3D);
 177         }
 178
 179         if (ir3_shader_debug & IR3_DBG_DISASM) {
 180                 debug_printf("----------------------\n");
 181                 nir_print_shader(s, stdout);
 182                 debug_printf("----------------------\n");
 183         }
 184
 185         OPT_V(s, nir_opt_global_to_local);
 186         OPT_V(s, nir_lower_regs_to_ssa);
 187
 188         if (key) {
 189                 if (s->info.stage == MESA_SHADER_VERTEX) {
 190                         OPT_V(s, nir_lower_clip_vs, key->ucp_enables, false);
 191                         if (key->vclamp_color)
 192                                 OPT_V(s, nir_lower_clamp_color_outputs);
 193                 } else if (s->info.stage == MESA_SHADER_FRAGMENT) {
 194                         OPT_V(s, nir_lower_clip_fs, key->ucp_enables);
 195                         if (key->fclamp_color)
 196                                 OPT_V(s, nir_lower_clamp_color_outputs);
 197                 }
 198                 if (key->color_two_side) {
 199                         OPT_V(s, nir_lower_two_sided_color);
 200                 }
 201         } else {
 202                 /* only want to do this the first time (when key is null)
 203                  * and not again on any potential 2nd variant lowering pass:
 204                  */
 205                 OPT_V(s, ir3_nir_apply_trig_workarounds);
 206         }
 207
 208         OPT_V(s, nir_lower_tex, &tex_options);
 209         OPT_V(s, nir_lower_load_const_to_scalar);
 210         if (shader->compiler->gpu_id < 500)
 211                 OPT_V(s, ir3_nir_lower_tg4_to_tex);
 212
 213         ir3_optimize_loop(s);
 214
 215         /* do idiv lowering after first opt loop to give a chance for
 216          * divide by immed power-of-two to be caught first:
 217          */
 218         if (OPT(s, nir_lower_idiv))
 219                 ir3_optimize_loop(s);
 220
 221         OPT_V(s, nir_remove_dead_variables, nir_var_function_temp);
 222
 223         OPT_V(s, nir_move_load_const);
 224
 225         if (ir3_shader_debug & IR3_DBG_DISASM) {
 226                 debug_printf("----------------------\n");
 227                 nir_print_shader(s, stdout);
 228                 debug_printf("----------------------\n");
 229         }
 230
 231         nir_sweep(s);
 232
 233         return s;
 234 }
 235
 236 void
 237 ir3_nir_scan_driver_consts(nir_shader *shader,
 238                 struct ir3_driver_const_layout *layout)
 239 {
 240         nir_foreach_function(function, shader) {
 241                 if (!function->impl)
 242                         continue;
 243
 244                 nir_foreach_block(block, function->impl) {
 245                         nir_foreach_instr(instr, block) {
 246                                 if (instr->type != nir_instr_type_intrinsic)
 247                                         continue;
 248
 249                                 nir_intrinsic_instr *intr =
 250                                         nir_instr_as_intrinsic(instr);
 251                                 unsigned idx;
 252
 253                                 switch (intr->intrinsic) {
 254                                 case nir_intrinsic_get_buffer_size:
 255                                         idx = nir_src_as_const_value(intr->src[0])->u32[0];
 256                                         if (layout->ssbo_size.mask & (1 << idx))
 257                                                 break;
 258                                         layout->ssbo_size.mask |= (1 << idx);
 259                                         layout->ssbo_size.off[idx] =
 260                                                 layout->ssbo_size.count;
 261                                         layout->ssbo_size.count += 1; /* one const per */
 262                                         break;
 263                                 case nir_intrinsic_image_deref_atomic_add:
 264                                 case nir_intrinsic_image_deref_atomic_min:
 265                                 case nir_intrinsic_image_deref_atomic_max:
 266                                 case nir_intrinsic_image_deref_atomic_and:
 267                                 case nir_intrinsic_image_deref_atomic_or:
 268                                 case nir_intrinsic_image_deref_atomic_xor:
 269                                 case nir_intrinsic_image_deref_atomic_exchange:
 270                                 case nir_intrinsic_image_deref_atomic_comp_swap:
 271                                 case nir_intrinsic_image_deref_store:
 272                                 case nir_intrinsic_image_deref_size:
 273                                         idx = nir_intrinsic_get_var(intr, 0)->data.driver_location;
 274                                         if (layout->image_dims.mask & (1 << idx))
 275                                                 break;
 276                                         layout->image_dims.mask |= (1 << idx);
 277                                         layout->image_dims.off[idx] =
 278                                                 layout->image_dims.count;
 279                                         layout->image_dims.count += 3; /* three const per */
 280                                         break;
 281                                 default:
 282                                         break;
 283                                 }
 284                         }
 285                 }
 286         }
 287 }