src/freedreno/vulkan/tu_shader.c

   1 /*
   2  * Copyright © 2019 Google LLC
   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
  21  * DEALINGS IN THE SOFTWARE.
  22  */
  23
  24 #include "tu_private.h"
  25
  26 #include "spirv/nir_spirv.h"
  27 #include "util/mesa-sha1.h"
  28 #include "nir/nir_xfb_info.h"
  29
  30 #include "ir3/ir3_nir.h"
  31
  32 static nir_shader *
  33 tu_spirv_to_nir(struct ir3_compiler *compiler,
  34                 const uint32_t *words,
  35                 size_t word_count,
  36                 gl_shader_stage stage,
  37                 const char *entry_point_name,
  38                 const VkSpecializationInfo *spec_info)
  39 {
  40    /* TODO these are made-up */
  41    const struct spirv_to_nir_options spirv_options = {
  42       .frag_coord_is_sysval = true,
  43       .lower_ubo_ssbo_access_to_offsets = true,
  44       .caps = {
  45          .transform_feedback = compiler->gpu_id >= 600,
  46       },
  47    };
  48    const nir_shader_compiler_options *nir_options =
  49       ir3_get_compiler_options(compiler);
  50
  51    /* convert VkSpecializationInfo */
  52    struct nir_spirv_specialization *spec = NULL;
  53    uint32_t num_spec = 0;
  54    if (spec_info && spec_info->mapEntryCount) {
  55       spec = malloc(sizeof(*spec) * spec_info->mapEntryCount);
  56       if (!spec)
  57          return NULL;
  58
  59       for (uint32_t i = 0; i < spec_info->mapEntryCount; i++) {
  60          const VkSpecializationMapEntry *entry = &spec_info->pMapEntries[i];
  61          const void *data = spec_info->pData + entry->offset;
  62          assert(data + entry->size <= spec_info->pData + spec_info->dataSize);
  63          spec[i].id = entry->constantID;
  64          if (entry->size == 8)
  65             spec[i].data64 = *(const uint64_t *) data;
  66          else
  67             spec[i].data32 = *(const uint32_t *) data;
  68          spec[i].defined_on_module = false;
  69       }
  70
  71       num_spec = spec_info->mapEntryCount;
  72    }
  73
  74    nir_shader *nir =
  75       spirv_to_nir(words, word_count, spec, num_spec, stage, entry_point_name,
  76                    &spirv_options, nir_options);
  77
  78    free(spec);
  79
  80    assert(nir->info.stage == stage);
  81    nir_validate_shader(nir, "after spirv_to_nir");
  82
  83    return nir;
  84 }
  85
  86 static unsigned
  87 map_add(struct tu_descriptor_map *map, int set, int binding, int value,
  88         int array_size)
  89 {
  90    unsigned index = 0;
  91    for (unsigned i = 0; i < map->num; i++) {
  92       if (set == map->set[i] && binding == map->binding[i]) {
  93          assert(value == map->value[i]);
  94          assert(array_size == map->array_size[i]);
  95          return index;
  96       }
  97       index += map->array_size[i];
  98    }
  99
 100    assert(index == map->num_desc);
 101
 102    map->set[map->num] = set;
 103    map->binding[map->num] = binding;
 104    map->value[map->num] = value;
 105    map->array_size[map->num] = array_size;
 106    map->num++;
 107    map->num_desc += array_size;
 108
 109    return index;
 110 }
 111
 112 static void
 113 lower_tex_src_to_offset(nir_builder *b, nir_tex_instr *instr, unsigned src_idx,
 114                         struct tu_shader *shader,
 115                         const struct tu_pipeline_layout *layout)
 116 {
 117    nir_ssa_def *index = NULL;
 118    unsigned base_index = 0;
 119    unsigned array_elements = 1;
 120    nir_tex_src *src = &instr->src[src_idx];
 121    bool is_sampler = src->src_type == nir_tex_src_sampler_deref;
 122
 123    /* We compute first the offsets */
 124    nir_deref_instr *deref = nir_instr_as_deref(src->src.ssa->parent_instr);
 125    while (deref->deref_type != nir_deref_type_var) {
 126       assert(deref->parent.is_ssa);
 127       nir_deref_instr *parent =
 128          nir_instr_as_deref(deref->parent.ssa->parent_instr);
 129
 130       assert(deref->deref_type == nir_deref_type_array);
 131
 132       if (nir_src_is_const(deref->arr.index) && index == NULL) {
 133          /* We're still building a direct index */
 134          base_index += nir_src_as_uint(deref->arr.index) * array_elements;
 135       } else {
 136          if (index == NULL) {
 137             /* We used to be direct but not anymore */
 138             index = nir_imm_int(b, base_index);
 139             base_index = 0;
 140          }
 141
 142          index = nir_iadd(b, index,
 143                           nir_imul(b, nir_imm_int(b, array_elements),
 144                                    nir_ssa_for_src(b, deref->arr.index, 1)));
 145       }
 146
 147       array_elements *= glsl_get_length(parent->type);
 148
 149       deref = parent;
 150    }
 151
 152    if (index)
 153       index = nir_umin(b, index, nir_imm_int(b, array_elements - 1));
 154
 155    /* We have the offsets, we apply them, rewriting the source or removing
 156     * instr if needed
 157     */
 158    if (index) {
 159       nir_instr_rewrite_src(&instr->instr, &src->src,
 160                             nir_src_for_ssa(index));
 161
 162       src->src_type = is_sampler ?
 163          nir_tex_src_sampler_offset :
 164          nir_tex_src_texture_offset;
 165    } else {
 166       nir_tex_instr_remove_src(instr, src_idx);
 167    }
 168
 169    uint32_t set = deref->var->data.descriptor_set;
 170    uint32_t binding = deref->var->data.binding;
 171    struct tu_descriptor_set_layout *set_layout = layout->set[set].layout;
 172    struct tu_descriptor_set_binding_layout *binding_layout =
 173       &set_layout->binding[binding];
 174
 175    int desc_index = map_add(is_sampler ?
 176                             &shader->sampler_map : &shader->texture_map,
 177                             deref->var->data.descriptor_set,
 178                             deref->var->data.binding,
 179                             deref->var->data.index,
 180                             binding_layout->array_size) + base_index;
 181    if (is_sampler)
 182       instr->sampler_index = desc_index;
 183    else
 184       instr->texture_index = desc_index;
 185 }
 186
 187 static bool
 188 lower_sampler(nir_builder *b, nir_tex_instr *instr, struct tu_shader *shader,
 189                 const struct tu_pipeline_layout *layout)
 190 {
 191    int texture_idx =
 192       nir_tex_instr_src_index(instr, nir_tex_src_texture_deref);
 193
 194    if (texture_idx >= 0)
 195       lower_tex_src_to_offset(b, instr, texture_idx, shader, layout);
 196
 197    int sampler_idx =
 198       nir_tex_instr_src_index(instr, nir_tex_src_sampler_deref);
 199
 200    if (sampler_idx >= 0)
 201       lower_tex_src_to_offset(b, instr, sampler_idx, shader, layout);
 202
 203    if (texture_idx < 0 && sampler_idx < 0)
 204       return false;
 205
 206    return true;
 207 }
 208
 209 static void
 210 lower_load_push_constant(nir_builder *b, nir_intrinsic_instr *instr,
 211                          struct tu_shader *shader)
 212 {
 213    nir_intrinsic_instr *load =
 214       nir_intrinsic_instr_create(b->shader, nir_intrinsic_load_uniform);
 215    load->num_components = instr->num_components;
 216    uint32_t base = nir_intrinsic_base(instr);
 217    assert(base % 4 == 0);
 218    assert(base >= shader->push_consts.lo * 16);
 219    base -= shader->push_consts.lo * 16;
 220    nir_intrinsic_set_base(load, base / 4);
 221    load->src[0] =
 222       nir_src_for_ssa(nir_ushr(b, instr->src[0].ssa, nir_imm_int(b, 2)));
 223    nir_ssa_dest_init(&load->instr, &load->dest,
 224                      load->num_components, instr->dest.ssa.bit_size,
 225                      instr->dest.ssa.name);
 226    nir_builder_instr_insert(b, &load->instr);
 227    nir_ssa_def_rewrite_uses(&instr->dest.ssa, nir_src_for_ssa(&load->dest.ssa));
 228
 229    nir_instr_remove(&instr->instr);
 230 }
 231
 232 static void
 233 lower_vulkan_resource_index(nir_builder *b, nir_intrinsic_instr *instr,
 234                             struct tu_shader *shader,
 235                             const struct tu_pipeline_layout *layout)
 236 {
 237    nir_const_value *const_val = nir_src_as_const_value(instr->src[0]);
 238
 239    unsigned set = nir_intrinsic_desc_set(instr);
 240    unsigned binding = nir_intrinsic_binding(instr);
 241    struct tu_descriptor_set_layout *set_layout = layout->set[set].layout;
 242    struct tu_descriptor_set_binding_layout *binding_layout =
 243       &set_layout->binding[binding];
 244    unsigned index = 0;
 245
 246    switch (nir_intrinsic_desc_type(instr)) {
 247    case VK_DESCRIPTOR_TYPE_UNIFORM_BUFFER:
 248    case VK_DESCRIPTOR_TYPE_UNIFORM_BUFFER_DYNAMIC:
 249       if (!const_val)
 250          tu_finishme("non-constant vulkan_resource_index array index");
 251       /* skip index 0 which is used for push constants */
 252       index = map_add(&shader->ubo_map, set, binding, 0,
 253                       binding_layout->array_size) + 1;
 254       index += const_val->u32;
 255       break;
 256    case VK_DESCRIPTOR_TYPE_STORAGE_BUFFER:
 257    case VK_DESCRIPTOR_TYPE_STORAGE_BUFFER_DYNAMIC:
 258       if (!const_val)
 259          tu_finishme("non-constant vulkan_resource_index array index");
 260       index = map_add(&shader->ssbo_map, set, binding, 0,
 261                       binding_layout->array_size);
 262       index += const_val->u32;
 263       break;
 264    default:
 265       tu_finishme("unsupported desc_type for vulkan_resource_index");
 266       break;
 267    }
 268
 269    nir_ssa_def_rewrite_uses(&instr->dest.ssa,
 270                             nir_src_for_ssa(nir_imm_int(b, index)));
 271    nir_instr_remove(&instr->instr);
 272 }
 273
 274 static void
 275 lower_image_deref(nir_builder *b,
 276                   nir_intrinsic_instr *instr, struct tu_shader *shader,
 277                   const struct tu_pipeline_layout *layout)
 278 {
 279    nir_deref_instr *deref = nir_src_as_deref(instr->src[0]);
 280    nir_variable *var = nir_deref_instr_get_variable(deref);
 281
 282    uint32_t set = var->data.descriptor_set;
 283    uint32_t binding = var->data.binding;
 284    struct tu_descriptor_set_layout *set_layout = layout->set[set].layout;
 285    struct tu_descriptor_set_binding_layout *binding_layout =
 286       &set_layout->binding[binding];
 287
 288    nir_ssa_def *index = nir_imm_int(b,
 289                                     map_add(&shader->image_map,
 290                                             set, binding, var->data.index,
 291                                             binding_layout->array_size));
 292    if (deref->deref_type != nir_deref_type_var) {
 293       assert(deref->deref_type == nir_deref_type_array);
 294       index = nir_iadd(b, index, nir_ssa_for_src(b, deref->arr.index, 1));
 295    }
 296    nir_rewrite_image_intrinsic(instr, index, false);
 297 }
 298
 299 static bool
 300 lower_intrinsic(nir_builder *b, nir_intrinsic_instr *instr,
 301                 struct tu_shader *shader,
 302                 const struct tu_pipeline_layout *layout)
 303 {
 304    switch (instr->intrinsic) {
 305    case nir_intrinsic_load_layer_id:
 306       /* TODO: remove this when layered rendering is implemented */
 307       nir_ssa_def_rewrite_uses(&instr->dest.ssa,
 308                                nir_src_for_ssa(nir_imm_int(b, 0)));
 309       nir_instr_remove(&instr->instr);
 310       return true;
 311
 312    case nir_intrinsic_load_push_constant:
 313       lower_load_push_constant(b, instr, shader);
 314       return true;
 315
 316    case nir_intrinsic_vulkan_resource_index:
 317       lower_vulkan_resource_index(b, instr, shader, layout);
 318       return true;
 319
 320    case nir_intrinsic_image_deref_load:
 321    case nir_intrinsic_image_deref_store:
 322    case nir_intrinsic_image_deref_atomic_add:
 323    case nir_intrinsic_image_deref_atomic_imin:
 324    case nir_intrinsic_image_deref_atomic_umin:
 325    case nir_intrinsic_image_deref_atomic_imax:
 326    case nir_intrinsic_image_deref_atomic_umax:
 327    case nir_intrinsic_image_deref_atomic_and:
 328    case nir_intrinsic_image_deref_atomic_or:
 329    case nir_intrinsic_image_deref_atomic_xor:
 330    case nir_intrinsic_image_deref_atomic_exchange:
 331    case nir_intrinsic_image_deref_atomic_comp_swap:
 332    case nir_intrinsic_image_deref_size:
 333    case nir_intrinsic_image_deref_samples:
 334    case nir_intrinsic_image_deref_load_param_intel:
 335    case nir_intrinsic_image_deref_load_raw_intel:
 336    case nir_intrinsic_image_deref_store_raw_intel:
 337       lower_image_deref(b, instr, shader, layout);
 338       return true;
 339
 340    default:
 341       return false;
 342    }
 343 }
 344
 345 /* Figure out the range of push constants that we're actually going to push to
 346  * the shader, and tell the backend to reserve this range when pushing UBO
 347  * constants.
 348  */
 349
 350 static void
 351 gather_push_constants(nir_shader *shader, struct tu_shader *tu_shader)
 352 {
 353    uint32_t min = UINT32_MAX, max = 0;
 354    nir_foreach_function(function, shader) {
 355       if (!function->impl)
 356          continue;
 357
 358       nir_foreach_block(block, function->impl) {
 359          nir_foreach_instr_safe(instr, block) {
 360             if (instr->type != nir_instr_type_intrinsic)
 361                continue;
 362
 363             nir_intrinsic_instr *intrin = nir_instr_as_intrinsic(instr);
 364             if (intrin->intrinsic != nir_intrinsic_load_push_constant)
 365                continue;
 366
 367             uint32_t base = nir_intrinsic_base(intrin);
 368             uint32_t range = nir_intrinsic_range(intrin);
 369             min = MIN2(min, base);
 370             max = MAX2(max, base + range);
 371             break;
 372          }
 373       }
 374    }
 375
 376    if (min >= max) {
 377       tu_shader->push_consts.lo = 0;
 378       tu_shader->push_consts.count = 0;
 379       tu_shader->ir3_shader.const_state.num_reserved_user_consts = 0;
 380       return;
 381    }
 382
 383    /* CP_LOAD_STATE OFFSET and NUM_UNIT are in units of vec4 (4 dwords),
 384     * however there's an alignment requirement of 4 on OFFSET. Expand the
 385     * range and change units accordingly.
 386     */
 387    tu_shader->push_consts.lo = (min / 16) / 4 * 4;
 388    tu_shader->push_consts.count =
 389       align(max, 16) / 16 - tu_shader->push_consts.lo;
 390    tu_shader->ir3_shader.const_state.num_reserved_user_consts =
 391       align(tu_shader->push_consts.count, 4);
 392 }
 393
 394 static bool
 395 lower_impl(nir_function_impl *impl, struct tu_shader *shader,
 396             const struct tu_pipeline_layout *layout)
 397 {
 398    nir_builder b;
 399    nir_builder_init(&b, impl);
 400    bool progress = false;
 401
 402    nir_foreach_block(block, impl) {
 403       nir_foreach_instr_safe(instr, block) {
 404          b.cursor = nir_before_instr(instr);
 405          switch (instr->type) {
 406          case nir_instr_type_tex:
 407             progress |= lower_sampler(&b, nir_instr_as_tex(instr), shader, layout);
 408             break;
 409          case nir_instr_type_intrinsic:
 410             progress |= lower_intrinsic(&b, nir_instr_as_intrinsic(instr), shader, layout);
 411             break;
 412          default:
 413             break;
 414          }
 415       }
 416    }
 417
 418    return progress;
 419 }
 420
 421 static bool
 422 tu_lower_io(nir_shader *shader, struct tu_shader *tu_shader,
 423             const struct tu_pipeline_layout *layout)
 424 {
 425    bool progress = false;
 426
 427    gather_push_constants(shader, tu_shader);
 428
 429    nir_foreach_function(function, shader) {
 430       if (function->impl)
 431          progress |= lower_impl(function->impl, tu_shader, layout);
 432    }
 433
 434    /* spirv_to_nir produces num_ssbos equal to the number of SSBO-containing
 435     * variables, while ir3 wants the number of descriptors (like the gallium
 436     * path).
 437     */
 438    shader->info.num_ssbos = tu_shader->ssbo_map.num_desc;
 439
 440    return progress;
 441 }
 442
 443 static void
 444 tu_gather_xfb_info(nir_shader *nir, struct tu_shader *shader)
 445 {
 446    struct ir3_stream_output_info *info = &shader->ir3_shader.stream_output;
 447    nir_xfb_info *xfb = nir_gather_xfb_info(nir, NULL);
 448
 449    if (!xfb)
 450       return;
 451
 452    /* creating a map from VARYING_SLOT_* enums to consecutive index */
 453    uint8_t num_outputs = 0;
 454    uint64_t outputs_written = 0;
 455    for (int i = 0; i < xfb->output_count; i++)
 456       outputs_written |= BITFIELD64_BIT(xfb->outputs[i].location);
 457
 458    uint8_t output_map[VARYING_SLOT_TESS_MAX];
 459    memset(output_map, 0, sizeof(output_map));
 460
 461    for (unsigned attr = 0; attr < VARYING_SLOT_MAX; attr++) {
 462       if (outputs_written & BITFIELD64_BIT(attr))
 463          output_map[attr] = num_outputs++;
 464    }
 465
 466    assert(xfb->output_count < IR3_MAX_SO_OUTPUTS);
 467    info->num_outputs = xfb->output_count;
 468
 469    for (int i = 0; i < IR3_MAX_SO_BUFFERS; i++)
 470       info->stride[i] = xfb->buffers[i].stride / 4;
 471
 472    for (int i = 0; i < xfb->output_count; i++) {
 473       info->output[i].register_index = output_map[xfb->outputs[i].location];
 474       info->output[i].start_component = xfb->outputs[i].component_offset;
 475       info->output[i].num_components =
 476                            util_bitcount(xfb->outputs[i].component_mask);
 477       info->output[i].output_buffer  = xfb->outputs[i].buffer;
 478       info->output[i].dst_offset = xfb->outputs[i].offset / 4;
 479       info->output[i].stream = xfb->buffer_to_stream[xfb->outputs[i].buffer];
 480    }
 481
 482    ralloc_free(xfb);
 483 }
 484
 485 struct tu_shader *
 486 tu_shader_create(struct tu_device *dev,
 487                  gl_shader_stage stage,
 488                  const VkPipelineShaderStageCreateInfo *stage_info,
 489                  struct tu_pipeline_layout *layout,
 490                  const VkAllocationCallbacks *alloc)
 491 {
 492    const struct tu_shader_module *module =
 493       tu_shader_module_from_handle(stage_info->module);
 494    struct tu_shader *shader;
 495
 496    const uint32_t max_variant_count = (stage == MESA_SHADER_VERTEX) ? 2 : 1;
 497    shader = vk_zalloc2(
 498       &dev->alloc, alloc,
 499       sizeof(*shader) + sizeof(struct ir3_shader_variant) * max_variant_count,
 500       8, VK_SYSTEM_ALLOCATION_SCOPE_COMMAND);
 501    if (!shader)
 502       return NULL;
 503
 504    /* translate SPIR-V to NIR */
 505    assert(module->code_size % 4 == 0);
 506    nir_shader *nir = tu_spirv_to_nir(
 507       dev->compiler, (const uint32_t *) module->code, module->code_size / 4,
 508       stage, stage_info->pName, stage_info->pSpecializationInfo);
 509    if (!nir) {
 510       vk_free2(&dev->alloc, alloc, shader);
 511       return NULL;
 512    }
 513
 514    if (unlikely(dev->physical_device->instance->debug_flags & TU_DEBUG_NIR)) {
 515       fprintf(stderr, "translated nir:\n");
 516       nir_print_shader(nir, stderr);
 517    }
 518
 519    /* multi step inlining procedure */
 520    NIR_PASS_V(nir, nir_lower_variable_initializers, nir_var_function_temp);
 521    NIR_PASS_V(nir, nir_lower_returns);
 522    NIR_PASS_V(nir, nir_inline_functions);
 523    NIR_PASS_V(nir, nir_opt_deref);
 524    foreach_list_typed_safe(nir_function, func, node, &nir->functions) {
 525       if (!func->is_entrypoint)
 526          exec_node_remove(&func->node);
 527    }
 528    assert(exec_list_length(&nir->functions) == 1);
 529    NIR_PASS_V(nir, nir_lower_variable_initializers, ~nir_var_function_temp);
 530
 531    /* Split member structs.  We do this before lower_io_to_temporaries so that
 532     * it doesn't lower system values to temporaries by accident.
 533     */
 534    NIR_PASS_V(nir, nir_split_var_copies);
 535    NIR_PASS_V(nir, nir_split_per_member_structs);
 536
 537    NIR_PASS_V(nir, nir_remove_dead_variables,
 538               nir_var_shader_in | nir_var_shader_out | nir_var_system_value | nir_var_mem_shared);
 539
 540    /* Gather information for transform feedback.
 541     * This should be called after nir_split_per_member_structs.
 542     * Also needs to be called after nir_remove_dead_variables with varyings,
 543     * so that we could align stream outputs correctly.
 544     */
 545    if (nir->info.stage == MESA_SHADER_VERTEX ||
 546          nir->info.stage == MESA_SHADER_TESS_EVAL ||
 547          nir->info.stage == MESA_SHADER_GEOMETRY)
 548       tu_gather_xfb_info(nir, shader);
 549
 550    NIR_PASS_V(nir, nir_propagate_invariant);
 551
 552    NIR_PASS_V(nir, nir_lower_io_to_temporaries, nir_shader_get_entrypoint(nir), true, true);
 553
 554    NIR_PASS_V(nir, nir_lower_global_vars_to_local);
 555    NIR_PASS_V(nir, nir_split_var_copies);
 556    NIR_PASS_V(nir, nir_lower_var_copies);
 557
 558    NIR_PASS_V(nir, nir_opt_copy_prop_vars);
 559    NIR_PASS_V(nir, nir_opt_combine_stores, nir_var_all);
 560
 561    /* ir3 doesn't support indirect input/output */
 562    NIR_PASS_V(nir, nir_lower_indirect_derefs, nir_var_shader_in | nir_var_shader_out);
 563
 564    NIR_PASS_V(nir, nir_lower_io_arrays_to_elements_no_indirects, false);
 565
 566    nir_assign_io_var_locations(&nir->inputs, &nir->num_inputs, stage);
 567    nir_assign_io_var_locations(&nir->outputs, &nir->num_outputs, stage);
 568
 569    NIR_PASS_V(nir, nir_lower_system_values);
 570    NIR_PASS_V(nir, nir_lower_frexp);
 571
 572    if (stage == MESA_SHADER_FRAGMENT)
 573       NIR_PASS_V(nir, nir_lower_input_attachments, true);
 574
 575    NIR_PASS_V(nir, tu_lower_io, shader, layout);
 576
 577    NIR_PASS_V(nir, nir_lower_io, nir_var_all, ir3_glsl_type_size, 0);
 578
 579    if (stage == MESA_SHADER_FRAGMENT) {
 580       /* NOTE: lower load_barycentric_at_sample first, since it
 581        * produces load_barycentric_at_offset:
 582        */
 583       NIR_PASS_V(nir, ir3_nir_lower_load_barycentric_at_sample);
 584       NIR_PASS_V(nir, ir3_nir_lower_load_barycentric_at_offset);
 585
 586       NIR_PASS_V(nir, ir3_nir_move_varying_inputs);
 587    }
 588
 589    nir_shader_gather_info(nir, nir_shader_get_entrypoint(nir));
 590
 591    /* num_uniforms only used by ir3 for size of ubo 0 (push constants) */
 592    nir->num_uniforms = MAX_PUSH_CONSTANTS_SIZE / 16;
 593
 594    shader->ir3_shader.compiler = dev->compiler;
 595    shader->ir3_shader.type = stage;
 596    shader->ir3_shader.nir = nir;
 597
 598    return shader;
 599 }
 600
 601 void
 602 tu_shader_destroy(struct tu_device *dev,
 603                   struct tu_shader *shader,
 604                   const VkAllocationCallbacks *alloc)
 605 {
 606    if (shader->ir3_shader.nir)
 607       ralloc_free(shader->ir3_shader.nir);
 608
 609    for (uint32_t i = 0; i < 1 + shader->has_binning_pass; i++) {
 610       if (shader->variants[i].ir)
 611          ir3_destroy(shader->variants[i].ir);
 612    }
 613
 614    if (shader->ir3_shader.const_state.immediates)
 615            free(shader->ir3_shader.const_state.immediates);
 616    if (shader->binary)
 617       free(shader->binary);
 618    if (shader->binning_binary)
 619       free(shader->binning_binary);
 620
 621    vk_free2(&dev->alloc, alloc, shader);
 622 }
 623
 624 void
 625 tu_shader_compile_options_init(
 626    struct tu_shader_compile_options *options,
 627    const VkGraphicsPipelineCreateInfo *pipeline_info)
 628 {
 629    bool has_gs = false;
 630    if (pipeline_info) {
 631       for (uint32_t i = 0; i < pipeline_info->stageCount; i++) {
 632          if (pipeline_info->pStages[i].stage == VK_SHADER_STAGE_GEOMETRY_BIT) {
 633             has_gs = true;
 634             break;
 635          }
 636       }
 637    }
 638
 639    *options = (struct tu_shader_compile_options) {
 640       /* TODO: Populate the remaining fields of ir3_shader_key. */
 641       .key = {
 642          .has_gs = has_gs,
 643       },
 644       /* TODO: VK_PIPELINE_CREATE_DISABLE_OPTIMIZATION_BIT
 645        * some optimizations need to happen otherwise shader might not compile
 646        */
 647       .optimize = true,
 648       .include_binning_pass = true,
 649    };
 650 }
 651
 652 static uint32_t *
 653 tu_compile_shader_variant(struct ir3_shader *shader,
 654                           const struct ir3_shader_key *key,
 655                           struct ir3_shader_variant *nonbinning,
 656                           struct ir3_shader_variant *variant)
 657 {
 658    variant->shader = shader;
 659    variant->type = shader->type;
 660    variant->key = *key;
 661    variant->binning_pass = !!nonbinning;
 662    variant->nonbinning = nonbinning;
 663
 664    int ret = ir3_compile_shader_nir(shader->compiler, variant);
 665    if (ret)
 666       return NULL;
 667
 668    /* when assemble fails, we rely on tu_shader_destroy to clean up the
 669     * variant
 670     */
 671    return ir3_shader_assemble(variant, shader->compiler->gpu_id);
 672 }
 673
 674 VkResult
 675 tu_shader_compile(struct tu_device *dev,
 676                   struct tu_shader *shader,
 677                   const struct tu_shader *next_stage,
 678                   const struct tu_shader_compile_options *options,
 679                   const VkAllocationCallbacks *alloc)
 680 {
 681    if (options->optimize) {
 682       /* ignore the key for the first pass of optimization */
 683       ir3_optimize_nir(&shader->ir3_shader, shader->ir3_shader.nir, NULL);
 684
 685       if (unlikely(dev->physical_device->instance->debug_flags &
 686                    TU_DEBUG_NIR)) {
 687          fprintf(stderr, "optimized nir:\n");
 688          nir_print_shader(shader->ir3_shader.nir, stderr);
 689       }
 690    }
 691
 692    shader->binary = tu_compile_shader_variant(
 693       &shader->ir3_shader, &options->key, NULL, &shader->variants[0]);
 694    if (!shader->binary)
 695       return VK_ERROR_OUT_OF_HOST_MEMORY;
 696
 697    if (shader_debug_enabled(shader->ir3_shader.type)) {
 698       fprintf(stdout, "Native code for unnamed %s shader %s:\n",
 699               ir3_shader_stage(&shader->variants[0]), shader->ir3_shader.nir->info.name);
 700        if (shader->ir3_shader.type == MESA_SHADER_FRAGMENT)
 701                fprintf(stdout, "SIMD0\n");
 702        ir3_shader_disasm(&shader->variants[0], shader->binary, stdout);
 703    }
 704
 705    /* compile another variant for the binning pass */
 706    if (options->include_binning_pass &&
 707        shader->ir3_shader.type == MESA_SHADER_VERTEX) {
 708       shader->binning_binary = tu_compile_shader_variant(
 709          &shader->ir3_shader, &options->key, &shader->variants[0],
 710          &shader->variants[1]);
 711       if (!shader->binning_binary)
 712          return VK_ERROR_OUT_OF_HOST_MEMORY;
 713
 714       shader->has_binning_pass = true;
 715
 716       if (shader_debug_enabled(MESA_SHADER_VERTEX)) {
 717          fprintf(stdout, "Native code for unnamed binning shader %s:\n",
 718                  shader->ir3_shader.nir->info.name);
 719           ir3_shader_disasm(&shader->variants[1], shader->binary, stdout);
 720       }
 721    }
 722
 723    if (unlikely(dev->physical_device->instance->debug_flags & TU_DEBUG_IR3)) {
 724       fprintf(stderr, "disassembled ir3:\n");
 725       fprintf(stderr, "shader: %s\n",
 726               gl_shader_stage_name(shader->ir3_shader.type));
 727       ir3_shader_disasm(&shader->variants[0], shader->binary, stderr);
 728
 729       if (shader->has_binning_pass) {
 730          fprintf(stderr, "disassembled ir3:\n");
 731          fprintf(stderr, "shader: %s (binning)\n",
 732                  gl_shader_stage_name(shader->ir3_shader.type));
 733          ir3_shader_disasm(&shader->variants[1], shader->binning_binary,
 734                            stderr);
 735       }
 736    }
 737
 738    return VK_SUCCESS;
 739 }
 740
 741 VkResult
 742 tu_CreateShaderModule(VkDevice _device,
 743                       const VkShaderModuleCreateInfo *pCreateInfo,
 744                       const VkAllocationCallbacks *pAllocator,
 745                       VkShaderModule *pShaderModule)
 746 {
 747    TU_FROM_HANDLE(tu_device, device, _device);
 748    struct tu_shader_module *module;
 749
 750    assert(pCreateInfo->sType == VK_STRUCTURE_TYPE_SHADER_MODULE_CREATE_INFO);
 751    assert(pCreateInfo->flags == 0);
 752    assert(pCreateInfo->codeSize % 4 == 0);
 753
 754    module = vk_alloc2(&device->alloc, pAllocator,
 755                       sizeof(*module) + pCreateInfo->codeSize, 8,
 756                       VK_SYSTEM_ALLOCATION_SCOPE_OBJECT);
 757    if (module == NULL)
 758       return vk_error(device->instance, VK_ERROR_OUT_OF_HOST_MEMORY);
 759
 760    module->code_size = pCreateInfo->codeSize;
 761    memcpy(module->code, pCreateInfo->pCode, pCreateInfo->codeSize);
 762
 763    _mesa_sha1_compute(module->code, module->code_size, module->sha1);
 764
 765    *pShaderModule = tu_shader_module_to_handle(module);
 766
 767    return VK_SUCCESS;
 768 }
 769
 770 void
 771 tu_DestroyShaderModule(VkDevice _device,
 772                        VkShaderModule _module,
 773                        const VkAllocationCallbacks *pAllocator)
 774 {
 775    TU_FROM_HANDLE(tu_device, device, _device);
 776    TU_FROM_HANDLE(tu_shader_module, module, _module);
 777
 778    if (!module)
 779       return;
 780
 781    vk_free2(&device->alloc, pAllocator, module);
 782 }