src/compiler/nir/nir_lower_system_values.c

   1 /*
   2  * Copyright © 2014 Intel Corporation
   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
  20  * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS
  21  * IN THE SOFTWARE.
  22  *
  23  * Authors:
  24  *    Connor Abbott (cwabbott0@gmail.com)
  25  *
  26  */
  27
  28 #include "nir.h"
  29 #include "nir_builder.h"
  30
  31 static nir_ssa_def *
  32 sanitize_32bit_sysval(nir_builder *b, nir_intrinsic_instr *intrin)
  33 {
  34    assert(intrin->dest.is_ssa);
  35    const unsigned bit_size = intrin->dest.ssa.bit_size;
  36    if (bit_size == 32)
  37       return NULL;
  38
  39    intrin->dest.ssa.bit_size = 32;
  40    return nir_u2u(b, &intrin->dest.ssa, bit_size);
  41 }
  42
  43 static nir_ssa_def*
  44 build_global_group_size(nir_builder *b, unsigned bit_size)
  45 {
  46    nir_ssa_def *group_size = nir_load_local_group_size(b);
  47    nir_ssa_def *num_work_groups = nir_load_num_work_groups(b, bit_size);
  48    return nir_imul(b, nir_u2u(b, group_size, bit_size),
  49                       num_work_groups);
  50 }
  51
  52 static bool
  53 lower_system_value_filter(const nir_instr *instr, const void *_state)
  54 {
  55    return instr->type == nir_instr_type_intrinsic;
  56 }
  57
  58 static nir_ssa_def *
  59 lower_system_value_instr(nir_builder *b, nir_instr *instr, void *_state)
  60 {
  61    nir_intrinsic_instr *intrin = nir_instr_as_intrinsic(instr);
  62
  63    /* All the intrinsics we care about are loads */
  64    if (!nir_intrinsic_infos[intrin->intrinsic].has_dest)
  65       return NULL;
  66
  67    assert(intrin->dest.is_ssa);
  68    const unsigned bit_size = intrin->dest.ssa.bit_size;
  69
  70    switch (intrin->intrinsic) {
  71    case nir_intrinsic_load_vertex_id:
  72       if (b->shader->options->vertex_id_zero_based) {
  73          return nir_iadd(b, nir_load_vertex_id_zero_base(b),
  74                             nir_load_first_vertex(b));
  75       } else {
  76          return NULL;
  77       }
  78
  79    case nir_intrinsic_load_base_vertex:
  80       /**
  81        * From the OpenGL 4.6 (11.1.3.9 Shader Inputs) specification:
  82        *
  83        * "gl_BaseVertex holds the integer value passed to the baseVertex
  84        * parameter to the command that resulted in the current shader
  85        * invocation. In the case where the command has no baseVertex
  86        * parameter, the value of gl_BaseVertex is zero."
  87        */
  88       if (b->shader->options->lower_base_vertex) {
  89          return nir_iand(b, nir_load_is_indexed_draw(b),
  90                             nir_load_first_vertex(b));
  91       } else {
  92          return NULL;
  93       }
  94
  95    case nir_intrinsic_load_local_invocation_id:
  96       /* If lower_cs_local_id_from_index is true, then we derive the local
  97        * index from the local id.
  98        */
  99       if (b->shader->options->lower_cs_local_id_from_index) {
 100          /* We lower gl_LocalInvocationID from gl_LocalInvocationIndex based
 101           * on this formula:
 102           *
 103           *    gl_LocalInvocationID.x =
 104           *       gl_LocalInvocationIndex % gl_WorkGroupSize.x;
 105           *    gl_LocalInvocationID.y =
 106           *       (gl_LocalInvocationIndex / gl_WorkGroupSize.x) %
 107           *       gl_WorkGroupSize.y;
 108           *    gl_LocalInvocationID.z =
 109           *       (gl_LocalInvocationIndex /
 110           *        (gl_WorkGroupSize.x * gl_WorkGroupSize.y)) %
 111           *       gl_WorkGroupSize.z;
 112           *
 113           * However, the final % gl_WorkGroupSize.z does nothing unless we
 114           * accidentally end up with a gl_LocalInvocationIndex that is too
 115           * large so it can safely be omitted.
 116           */
 117          nir_ssa_def *local_index = nir_load_local_invocation_index(b);
 118          nir_ssa_def *local_size = nir_load_local_group_size(b);
 119
 120          /* Because no hardware supports a local workgroup size greater than
 121           * about 1K, this calculation can be done in 32-bit and can save some
 122           * 64-bit arithmetic.
 123           */
 124          nir_ssa_def *id_x, *id_y, *id_z;
 125          id_x = nir_umod(b, local_index,
 126                             nir_channel(b, local_size, 0));
 127          id_y = nir_umod(b, nir_udiv(b, local_index,
 128                                         nir_channel(b, local_size, 0)),
 129                             nir_channel(b, local_size, 1));
 130          id_z = nir_udiv(b, local_index,
 131                             nir_imul(b, nir_channel(b, local_size, 0),
 132                                         nir_channel(b, local_size, 1)));
 133          return nir_u2u(b, nir_vec3(b, id_x, id_y, id_z), bit_size);
 134       } else {
 135          return sanitize_32bit_sysval(b, intrin);
 136       }
 137
 138    case nir_intrinsic_load_local_invocation_index:
 139       /* If lower_cs_local_index_from_id is true, then we derive the local
 140        * index from the local id.
 141        */
 142       if (b->shader->options->lower_cs_local_index_from_id) {
 143          /* From the GLSL man page for gl_LocalInvocationIndex:
 144           *
 145           *    "The value of gl_LocalInvocationIndex is equal to
 146           *    gl_LocalInvocationID.z * gl_WorkGroupSize.x *
 147           *    gl_WorkGroupSize.y + gl_LocalInvocationID.y *
 148           *    gl_WorkGroupSize.x + gl_LocalInvocationID.x"
 149           */
 150          nir_ssa_def *local_id = nir_load_local_invocation_id(b);
 151
 152          nir_ssa_def *size_x =
 153             nir_imm_int(b, b->shader->info.cs.local_size[0]);
 154          nir_ssa_def *size_y =
 155             nir_imm_int(b, b->shader->info.cs.local_size[1]);
 156
 157          /* Because no hardware supports a local workgroup size greater than
 158           * about 1K, this calculation can be done in 32-bit and can save some
 159           * 64-bit arithmetic.
 160           */
 161          nir_ssa_def *index;
 162          index = nir_imul(b, nir_channel(b, local_id, 2),
 163                              nir_imul(b, size_x, size_y));
 164          index = nir_iadd(b, index,
 165                              nir_imul(b, nir_channel(b, local_id, 1), size_x));
 166          index = nir_iadd(b, index, nir_channel(b, local_id, 0));
 167          return nir_u2u(b, index, bit_size);
 168       } else {
 169          return sanitize_32bit_sysval(b, intrin);
 170       }
 171
 172    case nir_intrinsic_load_local_group_size:
 173       if (b->shader->info.cs.local_size_variable) {
 174          /* If the local work group size is variable it can't be lowered at
 175           * this point.  We do, however, have to make sure that the intrinsic
 176           * is only 32-bit.
 177           */
 178          return sanitize_32bit_sysval(b, intrin);
 179       } else {
 180          /* using a 32 bit constant is safe here as no device/driver needs more
 181           * than 32 bits for the local size */
 182          nir_const_value local_size_const[3];
 183          memset(local_size_const, 0, sizeof(local_size_const));
 184          local_size_const[0].u32 = b->shader->info.cs.local_size[0];
 185          local_size_const[1].u32 = b->shader->info.cs.local_size[1];
 186          local_size_const[2].u32 = b->shader->info.cs.local_size[2];
 187          return nir_u2u(b, nir_build_imm(b, 3, 32, local_size_const), bit_size);
 188       }
 189
 190    case nir_intrinsic_load_global_invocation_id: {
 191       nir_ssa_def *group_size = nir_load_local_group_size(b);
 192       nir_ssa_def *group_id = nir_load_work_group_id(b, bit_size);
 193       nir_ssa_def *local_id = nir_load_local_invocation_id(b);
 194
 195       return nir_iadd(b, nir_imul(b, group_id,
 196                                      nir_u2u(b, group_size, bit_size)),
 197                          nir_u2u(b, local_id, bit_size));
 198    }
 199
 200    case nir_intrinsic_load_global_invocation_index: {
 201       nir_ssa_def *global_id = nir_load_global_invocation_id(b, bit_size);
 202       nir_ssa_def *global_size = build_global_group_size(b, bit_size);
 203
 204       /* index = id.x + ((id.y + (id.z * size.y)) * size.x) */
 205       nir_ssa_def *index;
 206       index = nir_imul(b, nir_channel(b, global_id, 2),
 207                           nir_channel(b, global_size, 1));
 208       index = nir_iadd(b, nir_channel(b, global_id, 1), index);
 209       index = nir_imul(b, nir_channel(b, global_size, 0), index);
 210       index = nir_iadd(b, nir_channel(b, global_id, 0), index);
 211       return index;
 212    }
 213
 214    case nir_intrinsic_load_helper_invocation:
 215       if (b->shader->options->lower_helper_invocation) {
 216          nir_ssa_def *tmp;
 217          tmp = nir_ishl(b, nir_imm_int(b, 1),
 218                            nir_load_sample_id_no_per_sample(b));
 219          tmp = nir_iand(b, nir_load_sample_mask_in(b), tmp);
 220          return nir_inot(b, nir_i2b(b, tmp));
 221       } else {
 222          return NULL;
 223       }
 224
 225    case nir_intrinsic_load_deref: {
 226       nir_deref_instr *deref = nir_src_as_deref(intrin->src[0]);
 227       if (deref->mode != nir_var_system_value)
 228          return NULL;
 229
 230       if (deref->deref_type != nir_deref_type_var) {
 231          /* The only one system value that is an array and that is
 232           * gl_SampleMask which is always an array of one element.
 233           */
 234          assert(deref->deref_type == nir_deref_type_array);
 235          deref = nir_deref_instr_parent(deref);
 236          assert(deref->deref_type == nir_deref_type_var);
 237          assert(deref->var->data.location == SYSTEM_VALUE_SAMPLE_MASK_IN);
 238       }
 239       nir_variable *var = deref->var;
 240
 241       switch (var->data.location) {
 242       case SYSTEM_VALUE_INSTANCE_INDEX:
 243          return nir_iadd(b, nir_load_instance_id(b),
 244                             nir_load_base_instance(b));
 245
 246       case SYSTEM_VALUE_SUBGROUP_EQ_MASK:
 247       case SYSTEM_VALUE_SUBGROUP_GE_MASK:
 248       case SYSTEM_VALUE_SUBGROUP_GT_MASK:
 249       case SYSTEM_VALUE_SUBGROUP_LE_MASK:
 250       case SYSTEM_VALUE_SUBGROUP_LT_MASK: {
 251          nir_intrinsic_op op =
 252             nir_intrinsic_from_system_value(var->data.location);
 253          nir_intrinsic_instr *load = nir_intrinsic_instr_create(b->shader, op);
 254          nir_ssa_dest_init_for_type(&load->instr, &load->dest,
 255                                     var->type, NULL);
 256          load->num_components = load->dest.ssa.num_components;
 257          nir_builder_instr_insert(b, &load->instr);
 258          return &load->dest.ssa;
 259       }
 260
 261       case SYSTEM_VALUE_DEVICE_INDEX:
 262          if (b->shader->options->lower_device_index_to_zero)
 263             return nir_imm_int(b, 0);
 264          break;
 265
 266       case SYSTEM_VALUE_GLOBAL_GROUP_SIZE:
 267          return build_global_group_size(b, bit_size);
 268
 269       case SYSTEM_VALUE_BARYCENTRIC_LINEAR_PIXEL:
 270          return nir_load_barycentric(b, nir_intrinsic_load_barycentric_pixel,
 271                                      INTERP_MODE_NOPERSPECTIVE);
 272
 273       case SYSTEM_VALUE_BARYCENTRIC_LINEAR_CENTROID:
 274          return nir_load_barycentric(b, nir_intrinsic_load_barycentric_centroid,
 275                                      INTERP_MODE_NOPERSPECTIVE);
 276
 277       case SYSTEM_VALUE_BARYCENTRIC_LINEAR_SAMPLE:
 278          return nir_load_barycentric(b, nir_intrinsic_load_barycentric_sample,
 279                                      INTERP_MODE_NOPERSPECTIVE);
 280
 281       case SYSTEM_VALUE_BARYCENTRIC_PERSP_PIXEL:
 282          return nir_load_barycentric(b, nir_intrinsic_load_barycentric_pixel,
 283                                      INTERP_MODE_SMOOTH);
 284
 285       case SYSTEM_VALUE_BARYCENTRIC_PERSP_CENTROID:
 286          return nir_load_barycentric(b, nir_intrinsic_load_barycentric_centroid,
 287                                      INTERP_MODE_SMOOTH);
 288
 289       case SYSTEM_VALUE_BARYCENTRIC_PERSP_SAMPLE:
 290          return nir_load_barycentric(b, nir_intrinsic_load_barycentric_sample,
 291                                      INTERP_MODE_SMOOTH);
 292
 293       case SYSTEM_VALUE_BARYCENTRIC_PULL_MODEL:
 294          return nir_load_barycentric(b, nir_intrinsic_load_barycentric_model,
 295                                      INTERP_MODE_NONE);
 296
 297       default:
 298          break;
 299       }
 300
 301       nir_intrinsic_op sysval_op =
 302          nir_intrinsic_from_system_value(var->data.location);
 303       return nir_load_system_value(b, sysval_op, 0,
 304                                       intrin->dest.ssa.num_components,
 305                                       intrin->dest.ssa.bit_size);
 306    }
 307
 308    default:
 309       return NULL;
 310    }
 311 }
 312
 313 bool
 314 nir_lower_system_values(nir_shader *shader)
 315 {
 316    bool progress = nir_shader_lower_instructions(shader,
 317                                                  lower_system_value_filter,
 318                                                  lower_system_value_instr,
 319                                                  NULL);
 320
 321    /* We're going to delete the variables so we need to clean up all those
 322     * derefs we left lying around.
 323     */
 324    if (progress)
 325       nir_remove_dead_derefs(shader);
 326
 327    nir_foreach_variable_with_modes_safe(var, shader, nir_var_system_value)
 328       exec_node_remove(&var->node);
 329
 330    return progress;
 331 }