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);
  48    return nir_imul(b, nir_u2u(b, group_size, bit_size),
  49                       nir_u2u(b, num_work_groups, bit_size));
  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_local_invocation_id:
  72       /* If lower_cs_local_id_from_index is true, then we derive the local
  73        * index from the local id.
  74        */
  75       if (b->shader->options->lower_cs_local_id_from_index) {
  76          /* We lower gl_LocalInvocationID from gl_LocalInvocationIndex based
  77           * on this formula:
  78           *
  79           *    gl_LocalInvocationID.x =
  80           *       gl_LocalInvocationIndex % gl_WorkGroupSize.x;
  81           *    gl_LocalInvocationID.y =
  82           *       (gl_LocalInvocationIndex / gl_WorkGroupSize.x) %
  83           *       gl_WorkGroupSize.y;
  84           *    gl_LocalInvocationID.z =
  85           *       (gl_LocalInvocationIndex /
  86           *        (gl_WorkGroupSize.x * gl_WorkGroupSize.y)) %
  87           *       gl_WorkGroupSize.z;
  88           *
  89           * However, the final % gl_WorkGroupSize.z does nothing unless we
  90           * accidentally end up with a gl_LocalInvocationIndex that is too
  91           * large so it can safely be omitted.
  92           */
  93          nir_ssa_def *local_index = nir_load_local_invocation_index(b);
  94          nir_ssa_def *local_size = nir_load_local_group_size(b);
  95
  96          /* Because no hardware supports a local workgroup size greater than
  97           * about 1K, this calculation can be done in 32-bit and can save some
  98           * 64-bit arithmetic.
  99           */
 100          nir_ssa_def *id_x, *id_y, *id_z;
 101          id_x = nir_umod(b, local_index,
 102                             nir_channel(b, local_size, 0));
 103          id_y = nir_umod(b, nir_udiv(b, local_index,
 104                                         nir_channel(b, local_size, 0)),
 105                             nir_channel(b, local_size, 1));
 106          id_z = nir_udiv(b, local_index,
 107                             nir_imul(b, nir_channel(b, local_size, 0),
 108                                         nir_channel(b, local_size, 1)));
 109          return nir_u2u(b, nir_vec3(b, id_x, id_y, id_z), bit_size);
 110       } else {
 111          return sanitize_32bit_sysval(b, intrin);
 112       }
 113
 114    case nir_intrinsic_load_local_group_size:
 115       if (b->shader->info.cs.local_size_variable) {
 116          /* If the local work group size is variable it can't be lowered at
 117           * this point.  We do, however, have to make sure that the intrinsic
 118           * is only 32-bit.
 119           */
 120          return sanitize_32bit_sysval(b, intrin);
 121       } else {
 122          /* using a 32 bit constant is safe here as no device/driver needs more
 123           * than 32 bits for the local size */
 124          nir_const_value local_size_const[3];
 125          memset(local_size_const, 0, sizeof(local_size_const));
 126          local_size_const[0].u32 = b->shader->info.cs.local_size[0];
 127          local_size_const[1].u32 = b->shader->info.cs.local_size[1];
 128          local_size_const[2].u32 = b->shader->info.cs.local_size[2];
 129          return nir_u2u(b, nir_build_imm(b, 3, 32, local_size_const), bit_size);
 130       }
 131
 132    case nir_intrinsic_load_global_invocation_id: {
 133       nir_ssa_def *group_size = nir_load_local_group_size(b);
 134       nir_ssa_def *group_id = nir_load_work_group_id(b);
 135       nir_ssa_def *local_id = nir_load_local_invocation_id(b);
 136
 137       return nir_iadd(b, nir_imul(b, nir_u2u(b, group_id, bit_size),
 138                                      nir_u2u(b, group_size, bit_size)),
 139                          nir_u2u(b, local_id, bit_size));
 140    }
 141
 142    case nir_intrinsic_load_deref: {
 143       nir_deref_instr *deref = nir_src_as_deref(intrin->src[0]);
 144       if (deref->mode != nir_var_system_value)
 145          return NULL;
 146
 147       if (deref->deref_type != nir_deref_type_var) {
 148          /* The only one system value that is an array and that is
 149           * gl_SampleMask which is always an array of one element.
 150           */
 151          assert(deref->deref_type == nir_deref_type_array);
 152          deref = nir_deref_instr_parent(deref);
 153          assert(deref->deref_type == nir_deref_type_var);
 154          assert(deref->var->data.location == SYSTEM_VALUE_SAMPLE_MASK_IN);
 155       }
 156       nir_variable *var = deref->var;
 157
 158       switch (var->data.location) {
 159       case SYSTEM_VALUE_GLOBAL_INVOCATION_INDEX: {
 160          nir_ssa_def *global_id = nir_load_global_invocation_id(b, bit_size);
 161          nir_ssa_def *global_size = build_global_group_size(b, bit_size);
 162
 163          /* index = id.x + ((id.y + (id.z * size.y)) * size.x) */
 164          nir_ssa_def *index;
 165          index = nir_imul(b, nir_channel(b, global_id, 2),
 166                              nir_channel(b, global_size, 1));
 167          index = nir_iadd(b, nir_channel(b, global_id, 1), index);
 168          index = nir_imul(b, nir_channel(b, global_size, 0), index);
 169          index = nir_iadd(b, nir_channel(b, global_id, 0), index);
 170          return index;
 171       }
 172
 173       case SYSTEM_VALUE_LOCAL_INVOCATION_INDEX: {
 174          /* If lower_cs_local_index_from_id is true, then we derive the local
 175           * index from the local id.
 176           */
 177          if (!b->shader->options->lower_cs_local_index_from_id)
 178             break;
 179
 180          /* From the GLSL man page for gl_LocalInvocationIndex:
 181           *
 182           *    "The value of gl_LocalInvocationIndex is equal to
 183           *    gl_LocalInvocationID.z * gl_WorkGroupSize.x *
 184           *    gl_WorkGroupSize.y + gl_LocalInvocationID.y *
 185           *    gl_WorkGroupSize.x + gl_LocalInvocationID.x"
 186           */
 187          nir_ssa_def *local_id = nir_load_local_invocation_id(b);
 188
 189          nir_ssa_def *size_x =
 190             nir_imm_int(b, b->shader->info.cs.local_size[0]);
 191          nir_ssa_def *size_y =
 192             nir_imm_int(b, b->shader->info.cs.local_size[1]);
 193
 194          /* Because no hardware supports a local workgroup size greater than
 195           * about 1K, this calculation can be done in 32-bit and can save some
 196           * 64-bit arithmetic.
 197           */
 198          nir_ssa_def *index;
 199          index = nir_imul(b, nir_channel(b, local_id, 2),
 200                              nir_imul(b, size_x, size_y));
 201          index = nir_iadd(b, index,
 202                              nir_imul(b, nir_channel(b, local_id, 1), size_x));
 203          index = nir_iadd(b, index, nir_channel(b, local_id, 0));
 204          index = nir_u2u(b, index, bit_size);
 205          return index;
 206       }
 207
 208       case SYSTEM_VALUE_VERTEX_ID:
 209          if (b->shader->options->vertex_id_zero_based) {
 210             return nir_iadd(b, nir_load_vertex_id_zero_base(b),
 211                                nir_load_first_vertex(b));
 212          }
 213          break;
 214
 215       case SYSTEM_VALUE_BASE_VERTEX:
 216          /**
 217           * From the OpenGL 4.6 (11.1.3.9 Shader Inputs) specification:
 218           *
 219           * "gl_BaseVertex holds the integer value passed to the baseVertex
 220           * parameter to the command that resulted in the current shader
 221           * invocation. In the case where the command has no baseVertex
 222           * parameter, the value of gl_BaseVertex is zero."
 223           */
 224          if (b->shader->options->lower_base_vertex) {
 225             return nir_iand(b, nir_load_is_indexed_draw(b),
 226                                nir_load_first_vertex(b));
 227          }
 228          break;
 229
 230       case SYSTEM_VALUE_HELPER_INVOCATION:
 231          if (b->shader->options->lower_helper_invocation) {
 232             nir_ssa_def *tmp;
 233
 234             tmp = nir_ishl(b,
 235                            nir_imm_int(b, 1),
 236                            nir_load_sample_id_no_per_sample(b));
 237
 238             tmp = nir_iand(b,
 239                            nir_load_sample_mask_in(b),
 240                            tmp);
 241
 242             return nir_inot(b, nir_i2b(b, tmp));
 243          }
 244          break;
 245
 246       case SYSTEM_VALUE_INSTANCE_INDEX:
 247          return nir_iadd(b, nir_load_instance_id(b),
 248                             nir_load_base_instance(b));
 249
 250       case SYSTEM_VALUE_SUBGROUP_EQ_MASK:
 251       case SYSTEM_VALUE_SUBGROUP_GE_MASK:
 252       case SYSTEM_VALUE_SUBGROUP_GT_MASK:
 253       case SYSTEM_VALUE_SUBGROUP_LE_MASK:
 254       case SYSTEM_VALUE_SUBGROUP_LT_MASK: {
 255          nir_intrinsic_op op =
 256             nir_intrinsic_from_system_value(var->data.location);
 257          nir_intrinsic_instr *load = nir_intrinsic_instr_create(b->shader, op);
 258          nir_ssa_dest_init_for_type(&load->instr, &load->dest,
 259                                     var->type, NULL);
 260          load->num_components = load->dest.ssa.num_components;
 261          nir_builder_instr_insert(b, &load->instr);
 262          return &load->dest.ssa;
 263       }
 264
 265       case SYSTEM_VALUE_DEVICE_INDEX:
 266          if (b->shader->options->lower_device_index_to_zero)
 267             return nir_imm_int(b, 0);
 268          break;
 269
 270       case SYSTEM_VALUE_GLOBAL_GROUP_SIZE:
 271          return build_global_group_size(b, bit_size);
 272
 273       default:
 274          break;
 275       }
 276
 277       nir_intrinsic_op sysval_op =
 278          nir_intrinsic_from_system_value(var->data.location);
 279       return nir_load_system_value(b, sysval_op, 0,
 280                                       intrin->dest.ssa.bit_size);
 281    }
 282
 283    default:
 284       return NULL;
 285    }
 286 }
 287
 288 bool
 289 nir_lower_system_values(nir_shader *shader)
 290 {
 291    bool progress = nir_shader_lower_instructions(shader,
 292                                                  lower_system_value_filter,
 293                                                  lower_system_value_instr,
 294                                                  NULL);
 295
 296    /* We're going to delete the variables so we need to clean up all those
 297     * derefs we left lying around.
 298     */
 299    if (progress)
 300       nir_remove_dead_derefs(shader);
 301
 302    exec_list_make_empty(&shader->system_values);
 303
 304    return progress;
 305 }