src/compiler/nir/nir_gather_info.c

   1 /*
   2  * Copyright © 2015 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
  24 #include "nir.h"
  25 #include "main/menums.h"
  26
  27 static void
  28 set_io_mask(nir_shader *shader, nir_variable *var, int offset, int len,
  29             bool is_output_read)
  30 {
  31    for (int i = 0; i < len; i++) {
  32       assert(var->data.location != -1);
  33
  34       int idx = var->data.location + offset + i;
  35       bool is_patch_generic = var->data.patch &&
  36                               idx != VARYING_SLOT_TESS_LEVEL_INNER &&
  37                               idx != VARYING_SLOT_TESS_LEVEL_OUTER &&
  38                               idx != VARYING_SLOT_BOUNDING_BOX0 &&
  39                               idx != VARYING_SLOT_BOUNDING_BOX1;
  40       uint64_t bitfield;
  41
  42       if (is_patch_generic) {
  43          assert(idx >= VARYING_SLOT_PATCH0 && idx < VARYING_SLOT_TESS_MAX);
  44          bitfield = BITFIELD64_BIT(idx - VARYING_SLOT_PATCH0);
  45       }
  46       else {
  47          assert(idx < VARYING_SLOT_MAX);
  48          bitfield = BITFIELD64_BIT(idx);
  49       }
  50
  51       if (var->data.mode == nir_var_shader_in) {
  52          if (is_patch_generic)
  53             shader->info.patch_inputs_read |= bitfield;
  54          else
  55             shader->info.inputs_read |= bitfield;
  56
  57          if (shader->info.stage == MESA_SHADER_FRAGMENT) {
  58             shader->info.fs.uses_sample_qualifier |= var->data.sample;
  59          }
  60       } else {
  61          assert(var->data.mode == nir_var_shader_out);
  62          if (is_output_read) {
  63             if (is_patch_generic) {
  64                shader->info.patch_outputs_read |= bitfield;
  65             } else {
  66                shader->info.outputs_read |= bitfield;
  67             }
  68          } else {
  69             if (is_patch_generic) {
  70                shader->info.patch_outputs_written |= bitfield;
  71             } else if (!var->data.read_only) {
  72                shader->info.outputs_written |= bitfield;
  73             }
  74          }
  75
  76
  77          if (var->data.fb_fetch_output)
  78             shader->info.outputs_read |= bitfield;
  79       }
  80    }
  81 }
  82
  83 /**
  84  * Mark an entire variable as used.  Caller must ensure that the variable
  85  * represents a shader input or output.
  86  */
  87 static void
  88 mark_whole_variable(nir_shader *shader, nir_variable *var, bool is_output_read)
  89 {
  90    const struct glsl_type *type = var->type;
  91
  92    if (nir_is_per_vertex_io(var, shader->info.stage)) {
  93       assert(glsl_type_is_array(type));
  94       type = glsl_get_array_element(type);
  95    }
  96
  97    const unsigned slots =
  98       var->data.compact ? DIV_ROUND_UP(glsl_get_length(type), 4)
  99                         : glsl_count_attribute_slots(type, false);
 100
 101    set_io_mask(shader, var, 0, slots, is_output_read);
 102 }
 103
 104 static unsigned
 105 get_io_offset(nir_deref_instr *deref, bool is_vertex_input)
 106 {
 107    unsigned offset = 0;
 108
 109    for (nir_deref_instr *d = deref; d; d = nir_deref_instr_parent(d)) {
 110       if (d->deref_type == nir_deref_type_array) {
 111          if (!nir_src_is_const(d->arr.index))
 112             return -1;
 113
 114          offset += glsl_count_attribute_slots(d->type, is_vertex_input) *
 115                    nir_src_as_uint(d->arr.index);
 116       }
 117       /* TODO: we can get the offset for structs here see nir_lower_io() */
 118    }
 119
 120    return offset;
 121 }
 122
 123 /**
 124  * Try to mark a portion of the given varying as used.  Caller must ensure
 125  * that the variable represents a shader input or output.
 126  *
 127  * If the index can't be interpreted as a constant, or some other problem
 128  * occurs, then nothing will be marked and false will be returned.
 129  */
 130 static bool
 131 try_mask_partial_io(nir_shader *shader, nir_variable *var,
 132                     nir_deref_instr *deref, bool is_output_read)
 133 {
 134    const struct glsl_type *type = var->type;
 135
 136    if (nir_is_per_vertex_io(var, shader->info.stage)) {
 137       assert(glsl_type_is_array(type));
 138       type = glsl_get_array_element(type);
 139    }
 140
 141    /* The code below only handles:
 142     *
 143     * - Indexing into matrices
 144     * - Indexing into arrays of (arrays, matrices, vectors, or scalars)
 145     *
 146     * For now, we just give up if we see varying structs and arrays of structs
 147     * here marking the entire variable as used.
 148     */
 149    if (!(glsl_type_is_matrix(type) ||
 150          (glsl_type_is_array(type) && !var->data.compact &&
 151           (glsl_type_is_numeric(glsl_without_array(type)) ||
 152            glsl_type_is_boolean(glsl_without_array(type)))))) {
 153
 154       /* If we don't know how to handle this case, give up and let the
 155        * caller mark the whole variable as used.
 156        */
 157       return false;
 158    }
 159
 160    unsigned offset = get_io_offset(deref, false);
 161    if (offset == -1)
 162       return false;
 163
 164    unsigned num_elems;
 165    unsigned elem_width = 1;
 166    unsigned mat_cols = 1;
 167    if (glsl_type_is_array(type)) {
 168       num_elems = glsl_get_aoa_size(type);
 169       if (glsl_type_is_matrix(glsl_without_array(type)))
 170          mat_cols = glsl_get_matrix_columns(glsl_without_array(type));
 171    } else {
 172       num_elems = glsl_get_matrix_columns(type);
 173    }
 174
 175    /* double element width for double types that takes two slots */
 176    if (glsl_type_is_dual_slot(glsl_without_array(type)))
 177       elem_width *= 2;
 178
 179    if (offset >= num_elems * elem_width * mat_cols) {
 180       /* Constant index outside the bounds of the matrix/array.  This could
 181        * arise as a result of constant folding of a legal GLSL program.
 182        *
 183        * Even though the spec says that indexing outside the bounds of a
 184        * matrix/array results in undefined behaviour, we don't want to pass
 185        * out-of-range values to set_io_mask() (since this could result in
 186        * slots that don't exist being marked as used), so just let the caller
 187        * mark the whole variable as used.
 188        */
 189       return false;
 190    }
 191
 192    set_io_mask(shader, var, offset, elem_width, is_output_read);
 193    return true;
 194 }
 195
 196 static void
 197 gather_intrinsic_info(nir_intrinsic_instr *instr, nir_shader *shader,
 198                       void *dead_ctx)
 199 {
 200    switch (instr->intrinsic) {
 201    case nir_intrinsic_demote:
 202    case nir_intrinsic_demote_if:
 203    case nir_intrinsic_discard:
 204    case nir_intrinsic_discard_if:
 205       assert(shader->info.stage == MESA_SHADER_FRAGMENT);
 206       shader->info.fs.uses_discard = true;
 207       break;
 208
 209    case nir_intrinsic_interp_deref_at_centroid:
 210    case nir_intrinsic_interp_deref_at_sample:
 211    case nir_intrinsic_interp_deref_at_offset:
 212    case nir_intrinsic_load_deref:
 213    case nir_intrinsic_store_deref:{
 214       nir_deref_instr *deref = nir_src_as_deref(instr->src[0]);
 215       if (deref->mode == nir_var_shader_in ||
 216           deref->mode == nir_var_shader_out) {
 217          nir_variable *var = nir_deref_instr_get_variable(deref);
 218          bool is_output_read = false;
 219          if (var->data.mode == nir_var_shader_out &&
 220              instr->intrinsic == nir_intrinsic_load_deref)
 221             is_output_read = true;
 222
 223          if (!try_mask_partial_io(shader, var, deref, is_output_read))
 224             mark_whole_variable(shader, var, is_output_read);
 225
 226          /* We need to track which input_reads bits correspond to a
 227           * dvec3/dvec4 input attribute */
 228          if (shader->info.stage == MESA_SHADER_VERTEX &&
 229              var->data.mode == nir_var_shader_in &&
 230              glsl_type_is_dual_slot(glsl_without_array(var->type))) {
 231             for (unsigned i = 0; i < glsl_count_attribute_slots(var->type, false); i++) {
 232                int idx = var->data.location + i;
 233                shader->info.vs.double_inputs |= BITFIELD64_BIT(idx);
 234             }
 235          }
 236       }
 237       break;
 238    }
 239
 240    case nir_intrinsic_load_draw_id:
 241    case nir_intrinsic_load_frag_coord:
 242    case nir_intrinsic_load_point_coord:
 243    case nir_intrinsic_load_front_face:
 244    case nir_intrinsic_load_vertex_id:
 245    case nir_intrinsic_load_vertex_id_zero_base:
 246    case nir_intrinsic_load_base_vertex:
 247    case nir_intrinsic_load_first_vertex:
 248    case nir_intrinsic_load_is_indexed_draw:
 249    case nir_intrinsic_load_base_instance:
 250    case nir_intrinsic_load_instance_id:
 251    case nir_intrinsic_load_sample_id:
 252    case nir_intrinsic_load_sample_pos:
 253    case nir_intrinsic_load_sample_mask_in:
 254    case nir_intrinsic_load_primitive_id:
 255    case nir_intrinsic_load_invocation_id:
 256    case nir_intrinsic_load_local_invocation_id:
 257    case nir_intrinsic_load_local_invocation_index:
 258    case nir_intrinsic_load_work_group_id:
 259    case nir_intrinsic_load_num_work_groups:
 260    case nir_intrinsic_load_tess_coord:
 261    case nir_intrinsic_load_tess_level_outer:
 262    case nir_intrinsic_load_tess_level_inner:
 263    case nir_intrinsic_load_patch_vertices_in:
 264       shader->info.system_values_read |=
 265          (1ull << nir_system_value_from_intrinsic(instr->intrinsic));
 266       break;
 267
 268    case nir_intrinsic_quad_broadcast:
 269    case nir_intrinsic_quad_swap_horizontal:
 270    case nir_intrinsic_quad_swap_vertical:
 271    case nir_intrinsic_quad_swap_diagonal:
 272       if (shader->info.stage == MESA_SHADER_FRAGMENT)
 273          shader->info.fs.needs_helper_invocations = true;
 274       break;
 275
 276    case nir_intrinsic_end_primitive:
 277    case nir_intrinsic_end_primitive_with_counter:
 278       assert(shader->info.stage == MESA_SHADER_GEOMETRY);
 279       shader->info.gs.uses_end_primitive = 1;
 280       /* fall through */
 281
 282    case nir_intrinsic_emit_vertex:
 283       if (nir_intrinsic_stream_id(instr) > 0)
 284          shader->info.gs.uses_streams = true;
 285
 286       break;
 287
 288    default:
 289       break;
 290    }
 291 }
 292
 293 static void
 294 gather_tex_info(nir_tex_instr *instr, nir_shader *shader)
 295 {
 296    if (shader->info.stage == MESA_SHADER_FRAGMENT &&
 297        nir_tex_instr_has_implicit_derivative(instr))
 298       shader->info.fs.needs_helper_invocations = true;
 299
 300    switch (instr->op) {
 301    case nir_texop_tg4:
 302       shader->info.uses_texture_gather = true;
 303       break;
 304    default:
 305       break;
 306    }
 307 }
 308
 309 static void
 310 gather_alu_info(nir_alu_instr *instr, nir_shader *shader)
 311 {
 312    switch (instr->op) {
 313    case nir_op_fddx:
 314    case nir_op_fddy:
 315       shader->info.uses_fddx_fddy = true;
 316       /* Fall through */
 317    case nir_op_fddx_fine:
 318    case nir_op_fddy_fine:
 319    case nir_op_fddx_coarse:
 320    case nir_op_fddy_coarse:
 321       if (shader->info.stage == MESA_SHADER_FRAGMENT)
 322          shader->info.fs.needs_helper_invocations = true;
 323       break;
 324    default:
 325       break;
 326    }
 327
 328    shader->info.uses_64bit |= instr->dest.dest.ssa.bit_size == 64;
 329    unsigned num_srcs = nir_op_infos[instr->op].num_inputs;
 330    for (unsigned i = 0; i < num_srcs; i++) {
 331       shader->info.uses_64bit |= nir_src_bit_size(instr->src[i].src) == 64;
 332    }
 333 }
 334
 335 static void
 336 gather_info_block(nir_block *block, nir_shader *shader, void *dead_ctx)
 337 {
 338    nir_foreach_instr(instr, block) {
 339       switch (instr->type) {
 340       case nir_instr_type_alu:
 341          gather_alu_info(nir_instr_as_alu(instr), shader);
 342          break;
 343       case nir_instr_type_intrinsic:
 344          gather_intrinsic_info(nir_instr_as_intrinsic(instr), shader, dead_ctx);
 345          break;
 346       case nir_instr_type_tex:
 347          gather_tex_info(nir_instr_as_tex(instr), shader);
 348          break;
 349       case nir_instr_type_call:
 350          assert(!"nir_shader_gather_info only works if functions are inlined");
 351          break;
 352       default:
 353          break;
 354       }
 355    }
 356 }
 357
 358 void
 359 nir_shader_gather_info(nir_shader *shader, nir_function_impl *entrypoint)
 360 {
 361    shader->info.num_textures = 0;
 362    shader->info.num_images = 0;
 363    shader->info.last_msaa_image = -1;
 364    nir_foreach_variable(var, &shader->uniforms) {
 365       /* Bindless textures and images don't use non-bindless slots. */
 366       if (var->data.bindless)
 367          continue;
 368
 369       shader->info.num_textures += glsl_type_get_sampler_count(var->type);
 370       shader->info.num_images += glsl_type_get_image_count(var->type);
 371
 372       /* Assuming image slots don't have holes (e.g. OpenGL) */
 373       if (glsl_type_is_image(var->type) &&
 374           glsl_get_sampler_dim(var->type) == GLSL_SAMPLER_DIM_MS)
 375          shader->info.last_msaa_image = shader->info.num_images - 1;
 376    }
 377
 378    shader->info.inputs_read = 0;
 379    shader->info.outputs_written = 0;
 380    shader->info.outputs_read = 0;
 381    shader->info.patch_outputs_read = 0;
 382    shader->info.patch_inputs_read = 0;
 383    shader->info.patch_outputs_written = 0;
 384    shader->info.system_values_read = 0;
 385    if (shader->info.stage == MESA_SHADER_VERTEX) {
 386       shader->info.vs.double_inputs = 0;
 387    }
 388    if (shader->info.stage == MESA_SHADER_FRAGMENT) {
 389       shader->info.fs.uses_sample_qualifier = false;
 390       shader->info.fs.uses_discard = false;
 391       shader->info.fs.needs_helper_invocations = false;
 392    }
 393
 394    void *dead_ctx = ralloc_context(NULL);
 395    nir_foreach_block(block, entrypoint) {
 396       gather_info_block(block, shader, dead_ctx);
 397    }
 398    ralloc_free(dead_ctx);
 399 }