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
  39 static const nir_shader_compiler_options options = {
  40                 .lower_fpow = 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                 .lower_ldexp = true,
  49                 .fuse_ffma = true,
  50                 .native_integers = true,
  51                 .vertex_id_zero_based = true,
  52                 .lower_extract_byte = true,
  53                 .lower_extract_word = true,
  54                 .lower_all_io_to_temps = true,
  55                 .lower_helper_invocation = true,
  56 };
  57
  58 struct nir_shader *
  59 ir3_tgsi_to_nir(const struct tgsi_token *tokens)
  60 {
  61         return tgsi_to_nir(tokens, &options);
  62 }
  63
  64 const nir_shader_compiler_options *
  65 ir3_get_compiler_options(struct ir3_compiler *compiler)
  66 {
  67         return &options;
  68 }
  69
  70 /* for given shader key, are any steps handled in nir? */
  71 bool
  72 ir3_key_lowers_nir(const struct ir3_shader_key *key)
  73 {
  74         return key->fsaturate_s | key->fsaturate_t | key->fsaturate_r |
  75                         key->vsaturate_s | key->vsaturate_t | key->vsaturate_r |
  76                         key->ucp_enables | key->color_two_side |
  77                         key->fclamp_color | key->vclamp_color;
  78 }
  79
  80 #define OPT(nir, pass, ...) ({                             \
  81    bool this_progress = false;                             \
  82    NIR_PASS(this_progress, nir, pass, ##__VA_ARGS__);      \
  83    this_progress;                                          \
  84 })
  85
  86 #define OPT_V(nir, pass, ...) NIR_PASS_V(nir, pass, ##__VA_ARGS__)
  87
  88 static void
  89 ir3_optimize_loop(nir_shader *s)
  90 {
  91         bool progress;
  92         do {
  93                 progress = false;
  94
  95                 OPT_V(s, nir_lower_vars_to_ssa);
  96                 progress |= OPT(s, nir_opt_copy_prop_vars);
  97                 progress |= OPT(s, nir_lower_alu_to_scalar);
  98                 progress |= OPT(s, nir_lower_phis_to_scalar);
  99
 100                 progress |= OPT(s, nir_copy_prop);
 101                 progress |= OPT(s, nir_opt_dce);
 102                 progress |= OPT(s, nir_opt_cse);
 103                 static int gcm = -1;
 104                 if (gcm == -1)
 105                         gcm = env2u("GCM");
 106                 if (gcm == 1)
 107                         progress |= OPT(s, nir_opt_gcm, true);
 108                 else if (gcm == 2)
 109                         progress |= OPT(s, nir_opt_gcm, false);
 110                 progress |= OPT(s, nir_opt_peephole_select, 16);
 111                 progress |= OPT(s, nir_opt_intrinsics);
 112                 progress |= OPT(s, nir_opt_algebraic);
 113                 progress |= OPT(s, nir_opt_constant_folding);
 114                 progress |= OPT(s, nir_opt_dead_cf);
 115                 if (OPT(s, nir_opt_trivial_continues)) {
 116                         progress |= true;
 117                         /* If nir_opt_trivial_continues makes progress, then we need to clean
 118                          * things up if we want any hope of nir_opt_if or nir_opt_loop_unroll
 119                          * to make progress.
 120                          */
 121                         OPT(s, nir_copy_prop);
 122                         OPT(s, nir_opt_dce);
 123                 }
 124                 progress |= OPT(s, nir_opt_if);
 125                 progress |= OPT(s, nir_opt_remove_phis);
 126                 progress |= OPT(s, nir_opt_undef);
 127
 128         } while (progress);
 129 }
 130
 131 struct nir_shader *
 132 ir3_optimize_nir(struct ir3_shader *shader, nir_shader *s,
 133                 const struct ir3_shader_key *key)
 134 {
 135         struct nir_lower_tex_options tex_options = {
 136                         .lower_rect = 0,
 137         };
 138
 139         if (key) {
 140                 switch (shader->type) {
 141                 case SHADER_FRAGMENT:
 142                         tex_options.saturate_s = key->fsaturate_s;
 143                         tex_options.saturate_t = key->fsaturate_t;
 144                         tex_options.saturate_r = key->fsaturate_r;
 145                         break;
 146                 case SHADER_VERTEX:
 147                         tex_options.saturate_s = key->vsaturate_s;
 148                         tex_options.saturate_t = key->vsaturate_t;
 149                         tex_options.saturate_r = key->vsaturate_r;
 150                         break;
 151                 default:
 152                         /* TODO */
 153                         break;
 154                 }
 155         }
 156
 157         if (shader->compiler->gpu_id >= 400) {
 158                 /* a4xx seems to have *no* sam.p */
 159                 tex_options.lower_txp = ~0;  /* lower all txp */
 160         } else {
 161                 /* a3xx just needs to avoid sam.p for 3d tex */
 162                 tex_options.lower_txp = (1 << GLSL_SAMPLER_DIM_3D);
 163         }
 164
 165         if (fd_mesa_debug & FD_DBG_DISASM) {
 166                 debug_printf("----------------------\n");
 167                 nir_print_shader(s, stdout);
 168                 debug_printf("----------------------\n");
 169         }
 170
 171         OPT_V(s, nir_opt_global_to_local);
 172         OPT_V(s, nir_lower_regs_to_ssa);
 173
 174         if (key) {
 175                 if (s->info.stage == MESA_SHADER_VERTEX) {
 176                         OPT_V(s, nir_lower_clip_vs, key->ucp_enables);
 177                         if (key->vclamp_color)
 178                                 OPT_V(s, nir_lower_clamp_color_outputs);
 179                 } else if (s->info.stage == MESA_SHADER_FRAGMENT) {
 180                         OPT_V(s, nir_lower_clip_fs, key->ucp_enables);
 181                         if (key->fclamp_color)
 182                                 OPT_V(s, nir_lower_clamp_color_outputs);
 183                 }
 184                 if (key->color_two_side) {
 185                         OPT_V(s, nir_lower_two_sided_color);
 186                 }
 187         } else {
 188                 /* only want to do this the first time (when key is null)
 189                  * and not again on any potential 2nd variant lowering pass:
 190                  */
 191                 OPT_V(s, ir3_nir_apply_trig_workarounds);
 192         }
 193
 194         OPT_V(s, nir_lower_tex, &tex_options);
 195         OPT_V(s, nir_lower_load_const_to_scalar);
 196         if (shader->compiler->gpu_id < 500)
 197                 OPT_V(s, ir3_nir_lower_tg4_to_tex);
 198
 199         ir3_optimize_loop(s);
 200
 201         /* do idiv lowering after first opt loop to give a chance for
 202          * divide by immed power-of-two to be caught first:
 203          */
 204         if (OPT(s, nir_lower_idiv))
 205                 ir3_optimize_loop(s);
 206
 207         OPT_V(s, nir_remove_dead_variables, nir_var_local);
 208
 209         OPT_V(s, nir_move_load_const);
 210
 211         if (fd_mesa_debug & FD_DBG_DISASM) {
 212                 debug_printf("----------------------\n");
 213                 nir_print_shader(s, stdout);
 214                 debug_printf("----------------------\n");
 215         }
 216
 217         nir_sweep(s);
 218
 219         return s;
 220 }
 221
 222 void
 223 ir3_nir_scan_driver_consts(nir_shader *shader,
 224                 struct ir3_driver_const_layout *layout)
 225 {
 226         nir_foreach_function(function, shader) {
 227                 if (!function->impl)
 228                         continue;
 229
 230                 nir_foreach_block(block, function->impl) {
 231                         nir_foreach_instr(instr, block) {
 232                                 if (instr->type != nir_instr_type_intrinsic)
 233                                         continue;
 234
 235                                 nir_intrinsic_instr *intr =
 236                                         nir_instr_as_intrinsic(instr);
 237                                 unsigned idx;
 238
 239                                 switch (intr->intrinsic) {
 240                                 case nir_intrinsic_get_buffer_size:
 241                                         idx = nir_src_as_const_value(intr->src[0])->u32[0];
 242                                         if (layout->ssbo_size.mask & (1 << idx))
 243                                                 break;
 244                                         layout->ssbo_size.mask |= (1 << idx);
 245                                         layout->ssbo_size.off[idx] =
 246                                                 layout->ssbo_size.count;
 247                                         layout->ssbo_size.count += 1; /* one const per */
 248                                         break;
 249                                 case nir_intrinsic_image_deref_atomic_add:
 250                                 case nir_intrinsic_image_deref_atomic_min:
 251                                 case nir_intrinsic_image_deref_atomic_max:
 252                                 case nir_intrinsic_image_deref_atomic_and:
 253                                 case nir_intrinsic_image_deref_atomic_or:
 254                                 case nir_intrinsic_image_deref_atomic_xor:
 255                                 case nir_intrinsic_image_deref_atomic_exchange:
 256                                 case nir_intrinsic_image_deref_atomic_comp_swap:
 257                                 case nir_intrinsic_image_deref_store:
 258                                         idx = nir_intrinsic_get_var(intr, 0)->data.driver_location;
 259                                         if (layout->image_dims.mask & (1 << idx))
 260                                                 break;
 261                                         layout->image_dims.mask |= (1 << idx);
 262                                         layout->image_dims.off[idx] =
 263                                                 layout->image_dims.count;
 264                                         layout->image_dims.count += 3; /* three const per */
 265                                         break;
 266                                 default:
 267                                         break;
 268                                 }
 269                         }
 270                 }
 271         }
 272 }