src/gallium/drivers/v3d/v3d_program.c

   1 /*
   2  * Copyright © 2014-2017 Broadcom
   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 <inttypes.h>
  25 #include "util/u_format.h"
  26 #include "util/u_math.h"
  27 #include "util/u_memory.h"
  28 #include "util/ralloc.h"
  29 #include "util/hash_table.h"
  30 #include "tgsi/tgsi_dump.h"
  31 #include "tgsi/tgsi_parse.h"
  32 #include "compiler/nir/nir.h"
  33 #include "compiler/nir/nir_builder.h"
  34 #include "nir/tgsi_to_nir.h"
  35 #include "compiler/v3d_compiler.h"
  36 #include "v3d_context.h"
  37 #include "broadcom/cle/v3d_packet_v33_pack.h"
  38 #include "mesa/state_tracker/st_glsl_types.h"
  39
  40 static gl_varying_slot
  41 v3d_get_slot_for_driver_location(nir_shader *s, uint32_t driver_location)
  42 {
  43         nir_foreach_variable(var, &s->outputs) {
  44                 if (var->data.driver_location == driver_location) {
  45                         return var->data.location;
  46                 }
  47         }
  48
  49         return -1;
  50 }
  51
  52 /**
  53  * Precomputes the TRANSFORM_FEEDBACK_OUTPUT_DATA_SPEC array for the shader.
  54  *
  55  * A shader can have 16 of these specs, and each one of them can write up to
  56  * 16 dwords.  Since we allow a total of 64 transform feedback output
  57  * components (not 16 vectors), we have to group the writes of multiple
  58  * varyings together in a single data spec.
  59  */
  60 static void
  61 v3d_set_transform_feedback_outputs(struct v3d_uncompiled_shader *so,
  62                                    const struct pipe_stream_output_info *stream_output)
  63 {
  64         if (!stream_output->num_outputs)
  65                 return;
  66
  67         struct v3d_varying_slot slots[PIPE_MAX_SO_OUTPUTS * 4];
  68         int slot_count = 0;
  69
  70         for (int buffer = 0; buffer < PIPE_MAX_SO_BUFFERS; buffer++) {
  71                 uint32_t buffer_offset = 0;
  72                 uint32_t vpm_start = slot_count;
  73
  74                 for (int i = 0; i < stream_output->num_outputs; i++) {
  75                         const struct pipe_stream_output *output =
  76                                 &stream_output->output[i];
  77
  78                         if (output->output_buffer != buffer)
  79                                 continue;
  80
  81                         /* We assume that the SO outputs appear in increasing
  82                          * order in the buffer.
  83                          */
  84                         assert(output->dst_offset >= buffer_offset);
  85
  86                         /* Pad any undefined slots in the output */
  87                         for (int j = buffer_offset; j < output->dst_offset; j++) {
  88                                 slots[slot_count] =
  89                                         v3d_slot_from_slot_and_component(VARYING_SLOT_POS, 0);
  90                                 slot_count++;
  91                                 buffer_offset++;
  92                         }
  93
  94                         /* Set the coordinate shader up to output the
  95                          * components of this varying.
  96                          */
  97                         for (int j = 0; j < output->num_components; j++) {
  98                                 gl_varying_slot slot =
  99                                         v3d_get_slot_for_driver_location(so->base.ir.nir, output->register_index);
 100
 101                                 slots[slot_count] =
 102                                         v3d_slot_from_slot_and_component(slot,
 103                                                                          output->start_component + j);
 104                                 slot_count++;
 105                                 buffer_offset++;
 106                         }
 107                 }
 108
 109                 uint32_t vpm_size = slot_count - vpm_start;
 110                 if (!vpm_size)
 111                         continue;
 112
 113                 uint32_t vpm_start_offset = vpm_start + 6;
 114
 115                 while (vpm_size) {
 116                         uint32_t write_size = MIN2(vpm_size, 1 << 4);
 117
 118                         struct V3D33_TRANSFORM_FEEDBACK_OUTPUT_DATA_SPEC unpacked = {
 119                                 /* We need the offset from the coordinate shader's VPM
 120                                  * output block, which has the [X, Y, Z, W, Xs, Ys]
 121                                  * values at the start.
 122                                  */
 123                                 .first_shaded_vertex_value_to_output = vpm_start_offset,
 124                                 .number_of_consecutive_vertex_values_to_output_as_32_bit_values = write_size,
 125                                 .output_buffer_to_write_to = buffer,
 126                         };
 127
 128                         /* GFXH-1559 */
 129                         assert(unpacked.first_shaded_vertex_value_to_output != 8 ||
 130                                so->num_tf_specs != 0);
 131
 132                         assert(so->num_tf_specs != ARRAY_SIZE(so->tf_specs));
 133                         V3D33_TRANSFORM_FEEDBACK_OUTPUT_DATA_SPEC_pack(NULL,
 134                                                                        (void *)&so->tf_specs[so->num_tf_specs],
 135                                                                        &unpacked);
 136
 137                         /* If point size is being written by the shader, then
 138                          * all the VPM start offsets are shifted up by one.
 139                          * We won't know that until the variant is compiled,
 140                          * though.
 141                          */
 142                         unpacked.first_shaded_vertex_value_to_output++;
 143
 144                         /* GFXH-1559 */
 145                         assert(unpacked.first_shaded_vertex_value_to_output != 8 ||
 146                                so->num_tf_specs != 0);
 147
 148                         V3D33_TRANSFORM_FEEDBACK_OUTPUT_DATA_SPEC_pack(NULL,
 149                                                                        (void *)&so->tf_specs_psiz[so->num_tf_specs],
 150                                                                        &unpacked);
 151                         so->num_tf_specs++;
 152                         vpm_start_offset += write_size;
 153                         vpm_size -= write_size;
 154                 }
 155                 so->base.stream_output.stride[buffer] =
 156                         stream_output->stride[buffer];
 157         }
 158
 159         so->num_tf_outputs = slot_count;
 160         so->tf_outputs = ralloc_array(so->base.ir.nir, struct v3d_varying_slot,
 161                                       slot_count);
 162         memcpy(so->tf_outputs, slots, sizeof(*slots) * slot_count);
 163 }
 164
 165 static int
 166 type_size(const struct glsl_type *type)
 167 {
 168         return glsl_count_attribute_slots(type, false);
 169 }
 170
 171 static int
 172 uniforms_type_size(const struct glsl_type *type)
 173 {
 174         return st_glsl_storage_type_size(type, false);
 175 }
 176
 177 static void *
 178 v3d_shader_state_create(struct pipe_context *pctx,
 179                         const struct pipe_shader_state *cso)
 180 {
 181         struct v3d_context *v3d = v3d_context(pctx);
 182         struct v3d_uncompiled_shader *so = CALLOC_STRUCT(v3d_uncompiled_shader);
 183         if (!so)
 184                 return NULL;
 185
 186         so->program_id = v3d->next_uncompiled_program_id++;
 187
 188         nir_shader *s;
 189
 190         if (cso->type == PIPE_SHADER_IR_NIR) {
 191                 /* The backend takes ownership of the NIR shader on state
 192                  * creation.
 193                  */
 194                 s = cso->ir.nir;
 195
 196                 NIR_PASS_V(s, nir_lower_io, nir_var_all & ~nir_var_uniform,
 197                            type_size,
 198                            (nir_lower_io_options)0);
 199                 NIR_PASS_V(s, nir_lower_io, nir_var_uniform,
 200                            uniforms_type_size,
 201                            (nir_lower_io_options)0);
 202         } else {
 203                 assert(cso->type == PIPE_SHADER_IR_TGSI);
 204
 205                 if (V3D_DEBUG & V3D_DEBUG_TGSI) {
 206                         fprintf(stderr, "prog %d TGSI:\n",
 207                                 so->program_id);
 208                         tgsi_dump(cso->tokens, 0);
 209                         fprintf(stderr, "\n");
 210                 }
 211                 s = tgsi_to_nir(cso->tokens, &v3d_nir_options);
 212
 213                 so->was_tgsi = true;
 214         }
 215
 216         NIR_PASS_V(s, nir_opt_global_to_local);
 217         NIR_PASS_V(s, nir_lower_regs_to_ssa);
 218         NIR_PASS_V(s, nir_normalize_cubemap_coords);
 219
 220         NIR_PASS_V(s, nir_lower_load_const_to_scalar);
 221
 222         v3d_optimize_nir(s);
 223
 224         NIR_PASS_V(s, nir_remove_dead_variables, nir_var_local);
 225
 226         /* Garbage collect dead instructions */
 227         nir_sweep(s);
 228
 229         so->base.type = PIPE_SHADER_IR_NIR;
 230         so->base.ir.nir = s;
 231
 232         v3d_set_transform_feedback_outputs(so, &cso->stream_output);
 233
 234         if (V3D_DEBUG & (V3D_DEBUG_NIR |
 235                          v3d_debug_flag_for_shader_stage(s->info.stage))) {
 236                 fprintf(stderr, "%s prog %d NIR:\n",
 237                         gl_shader_stage_name(s->info.stage),
 238                         so->program_id);
 239                 nir_print_shader(s, stderr);
 240                 fprintf(stderr, "\n");
 241         }
 242
 243         return so;
 244 }
 245
 246 static struct v3d_compiled_shader *
 247 v3d_get_compiled_shader(struct v3d_context *v3d, struct v3d_key *key)
 248 {
 249         struct v3d_uncompiled_shader *shader_state = key->shader_state;
 250         nir_shader *s = shader_state->base.ir.nir;
 251
 252         struct hash_table *ht;
 253         uint32_t key_size;
 254         if (s->info.stage == MESA_SHADER_FRAGMENT) {
 255                 ht = v3d->fs_cache;
 256                 key_size = sizeof(struct v3d_fs_key);
 257         } else {
 258                 ht = v3d->vs_cache;
 259                 key_size = sizeof(struct v3d_vs_key);
 260         }
 261
 262         struct hash_entry *entry = _mesa_hash_table_search(ht, key);
 263         if (entry)
 264                 return entry->data;
 265
 266         struct v3d_compiled_shader *shader =
 267                 rzalloc(NULL, struct v3d_compiled_shader);
 268
 269         int program_id = shader_state->program_id;
 270         int variant_id =
 271                 p_atomic_inc_return(&shader_state->compiled_variant_count);
 272         uint64_t *qpu_insts;
 273         uint32_t shader_size;
 274
 275         switch (s->info.stage) {
 276         case MESA_SHADER_VERTEX:
 277                 shader->prog_data.vs = rzalloc(shader, struct v3d_vs_prog_data);
 278
 279                 qpu_insts = v3d_compile_vs(v3d->screen->compiler,
 280                                            (struct v3d_vs_key *)key,
 281                                            shader->prog_data.vs, s,
 282                                            program_id, variant_id,
 283                                            &shader_size);
 284                 break;
 285         case MESA_SHADER_FRAGMENT:
 286                 shader->prog_data.fs = rzalloc(shader, struct v3d_fs_prog_data);
 287
 288                 qpu_insts = v3d_compile_fs(v3d->screen->compiler,
 289                                            (struct v3d_fs_key *)key,
 290                                            shader->prog_data.fs, s,
 291                                            program_id, variant_id,
 292                                            &shader_size);
 293                 break;
 294         default:
 295                 unreachable("bad stage");
 296         }
 297
 298         v3d_set_shader_uniform_dirty_flags(shader);
 299
 300         shader->bo = v3d_bo_alloc(v3d->screen, shader_size, "shader");
 301         v3d_bo_map(shader->bo);
 302         memcpy(shader->bo->map, qpu_insts, shader_size);
 303
 304         free(qpu_insts);
 305
 306         struct v3d_key *dup_key;
 307         dup_key = ralloc_size(shader, key_size);
 308         memcpy(dup_key, key, key_size);
 309         _mesa_hash_table_insert(ht, dup_key, shader);
 310
 311         if (shader->prog_data.base->spill_size >
 312             v3d->prog.spill_size_per_thread) {
 313                 /* Max 4 QPUs per slice, 3 slices per core. We only do single
 314                  * core so far.  This overallocates memory on smaller cores.
 315                  */
 316                 int total_spill_size =
 317                         4 * 3 * shader->prog_data.base->spill_size;
 318
 319                 v3d_bo_unreference(&v3d->prog.spill_bo);
 320                 v3d->prog.spill_bo = v3d_bo_alloc(v3d->screen,
 321                                                   total_spill_size, "spill");
 322                 v3d->prog.spill_size_per_thread =
 323                         shader->prog_data.base->spill_size;
 324         }
 325
 326         return shader;
 327 }
 328
 329 static void
 330 v3d_setup_shared_key(struct v3d_context *v3d, struct v3d_key *key,
 331                      struct v3d_texture_stateobj *texstate)
 332 {
 333         const struct v3d_device_info *devinfo = &v3d->screen->devinfo;
 334
 335         for (int i = 0; i < texstate->num_textures; i++) {
 336                 struct pipe_sampler_view *sampler = texstate->textures[i];
 337                 struct v3d_sampler_view *v3d_sampler = v3d_sampler_view(sampler);
 338                 struct pipe_sampler_state *sampler_state =
 339                         texstate->samplers[i];
 340
 341                 if (!sampler)
 342                         continue;
 343
 344                 key->tex[i].return_size =
 345                         v3d_get_tex_return_size(devinfo,
 346                                                 sampler->format,
 347                                                 sampler_state->compare_mode);
 348
 349                 /* For 16-bit, we set up the sampler to always return 2
 350                  * channels (meaning no recompiles for most statechanges),
 351                  * while for 32 we actually scale the returns with channels.
 352                  */
 353                 if (key->tex[i].return_size == 16) {
 354                         key->tex[i].return_channels = 2;
 355                 } else if (devinfo->ver > 40) {
 356                         key->tex[i].return_channels = 4;
 357                 } else {
 358                         key->tex[i].return_channels =
 359                                 v3d_get_tex_return_channels(devinfo,
 360                                                             sampler->format);
 361                 }
 362
 363                 if (key->tex[i].return_size == 32 && devinfo->ver < 40) {
 364                         memcpy(key->tex[i].swizzle,
 365                                v3d_sampler->swizzle,
 366                                sizeof(v3d_sampler->swizzle));
 367                 } else {
 368                         /* For 16-bit returns, we let the sampler state handle
 369                          * the swizzle.
 370                          */
 371                         key->tex[i].swizzle[0] = PIPE_SWIZZLE_X;
 372                         key->tex[i].swizzle[1] = PIPE_SWIZZLE_Y;
 373                         key->tex[i].swizzle[2] = PIPE_SWIZZLE_Z;
 374                         key->tex[i].swizzle[3] = PIPE_SWIZZLE_W;
 375                 }
 376
 377                 if (sampler) {
 378                         key->tex[i].compare_mode = sampler_state->compare_mode;
 379                         key->tex[i].compare_func = sampler_state->compare_func;
 380                         key->tex[i].clamp_s =
 381                                 sampler_state->wrap_s == PIPE_TEX_WRAP_CLAMP;
 382                         key->tex[i].clamp_t =
 383                                 sampler_state->wrap_t == PIPE_TEX_WRAP_CLAMP;
 384                         key->tex[i].clamp_r =
 385                                 sampler_state->wrap_r == PIPE_TEX_WRAP_CLAMP;
 386                 }
 387         }
 388
 389         key->ucp_enables = v3d->rasterizer->base.clip_plane_enable;
 390 }
 391
 392 static void
 393 v3d_update_compiled_fs(struct v3d_context *v3d, uint8_t prim_mode)
 394 {
 395         struct v3d_job *job = v3d->job;
 396         struct v3d_fs_key local_key;
 397         struct v3d_fs_key *key = &local_key;
 398
 399         if (!(v3d->dirty & (VC5_DIRTY_PRIM_MODE |
 400                             VC5_DIRTY_BLEND |
 401                             VC5_DIRTY_FRAMEBUFFER |
 402                             VC5_DIRTY_ZSA |
 403                             VC5_DIRTY_RASTERIZER |
 404                             VC5_DIRTY_SAMPLE_STATE |
 405                             VC5_DIRTY_FRAGTEX |
 406                             VC5_DIRTY_UNCOMPILED_FS))) {
 407                 return;
 408         }
 409
 410         memset(key, 0, sizeof(*key));
 411         v3d_setup_shared_key(v3d, &key->base, &v3d->fragtex);
 412         key->base.shader_state = v3d->prog.bind_fs;
 413         key->is_points = (prim_mode == PIPE_PRIM_POINTS);
 414         key->is_lines = (prim_mode >= PIPE_PRIM_LINES &&
 415                          prim_mode <= PIPE_PRIM_LINE_STRIP);
 416         key->clamp_color = v3d->rasterizer->base.clamp_fragment_color;
 417         if (v3d->blend->base.logicop_enable) {
 418                 key->logicop_func = v3d->blend->base.logicop_func;
 419         } else {
 420                 key->logicop_func = PIPE_LOGICOP_COPY;
 421         }
 422         if (job->msaa) {
 423                 key->msaa = v3d->rasterizer->base.multisample;
 424                 key->sample_coverage = (v3d->rasterizer->base.multisample &&
 425                                         v3d->sample_mask != (1 << VC5_MAX_SAMPLES) - 1);
 426                 key->sample_alpha_to_coverage = v3d->blend->base.alpha_to_coverage;
 427                 key->sample_alpha_to_one = v3d->blend->base.alpha_to_one;
 428         }
 429
 430         key->depth_enabled = (v3d->zsa->base.depth.enabled ||
 431                               v3d->zsa->base.stencil[0].enabled);
 432         if (v3d->zsa->base.alpha.enabled) {
 433                 key->alpha_test = true;
 434                 key->alpha_test_func = v3d->zsa->base.alpha.func;
 435         }
 436
 437         /* gl_FragColor's propagation to however many bound color buffers
 438          * there are means that the buffer count needs to be in the key.
 439          */
 440         key->nr_cbufs = v3d->framebuffer.nr_cbufs;
 441         key->swap_color_rb = v3d->swap_color_rb;
 442
 443         for (int i = 0; i < key->nr_cbufs; i++) {
 444                 struct pipe_surface *cbuf = v3d->framebuffer.cbufs[i];
 445                 if (!cbuf)
 446                         continue;
 447
 448                 const struct util_format_description *desc =
 449                         util_format_description(cbuf->format);
 450
 451                 if (desc->channel[0].type == UTIL_FORMAT_TYPE_FLOAT &&
 452                     desc->channel[0].size == 32) {
 453                         key->f32_color_rb |= 1 << i;
 454                 }
 455
 456                 if (v3d->prog.bind_fs->was_tgsi) {
 457                         if (util_format_is_pure_uint(cbuf->format))
 458                                 key->uint_color_rb |= 1 << i;
 459                         else if (util_format_is_pure_sint(cbuf->format))
 460                                 key->int_color_rb |= 1 << i;
 461                 }
 462         }
 463
 464         if (key->is_points) {
 465                 key->point_sprite_mask =
 466                         v3d->rasterizer->base.sprite_coord_enable;
 467                 key->point_coord_upper_left =
 468                         (v3d->rasterizer->base.sprite_coord_mode ==
 469                          PIPE_SPRITE_COORD_UPPER_LEFT);
 470         }
 471
 472         key->light_twoside = v3d->rasterizer->base.light_twoside;
 473         key->shade_model_flat = v3d->rasterizer->base.flatshade;
 474
 475         struct v3d_compiled_shader *old_fs = v3d->prog.fs;
 476         v3d->prog.fs = v3d_get_compiled_shader(v3d, &key->base);
 477         if (v3d->prog.fs == old_fs)
 478                 return;
 479
 480         v3d->dirty |= VC5_DIRTY_COMPILED_FS;
 481
 482         if (old_fs) {
 483                 if (v3d->prog.fs->prog_data.fs->flat_shade_flags !=
 484                     old_fs->prog_data.fs->flat_shade_flags) {
 485                         v3d->dirty |= VC5_DIRTY_FLAT_SHADE_FLAGS;
 486                 }
 487
 488                 if (v3d->prog.fs->prog_data.fs->noperspective_flags !=
 489                     old_fs->prog_data.fs->noperspective_flags) {
 490                         v3d->dirty |= VC5_DIRTY_NOPERSPECTIVE_FLAGS;
 491                 }
 492
 493                 if (v3d->prog.fs->prog_data.fs->centroid_flags !=
 494                     old_fs->prog_data.fs->centroid_flags) {
 495                         v3d->dirty |= VC5_DIRTY_CENTROID_FLAGS;
 496                 }
 497         }
 498
 499         if (old_fs && memcmp(v3d->prog.fs->prog_data.fs->input_slots,
 500                              old_fs->prog_data.fs->input_slots,
 501                              sizeof(v3d->prog.fs->prog_data.fs->input_slots))) {
 502                 v3d->dirty |= VC5_DIRTY_FS_INPUTS;
 503         }
 504 }
 505
 506 static void
 507 v3d_update_compiled_vs(struct v3d_context *v3d, uint8_t prim_mode)
 508 {
 509         struct v3d_vs_key local_key;
 510         struct v3d_vs_key *key = &local_key;
 511
 512         if (!(v3d->dirty & (VC5_DIRTY_PRIM_MODE |
 513                             VC5_DIRTY_RASTERIZER |
 514                             VC5_DIRTY_VERTTEX |
 515                             VC5_DIRTY_VTXSTATE |
 516                             VC5_DIRTY_UNCOMPILED_VS |
 517                             VC5_DIRTY_FS_INPUTS))) {
 518                 return;
 519         }
 520
 521         memset(key, 0, sizeof(*key));
 522         v3d_setup_shared_key(v3d, &key->base, &v3d->verttex);
 523         key->base.shader_state = v3d->prog.bind_vs;
 524         key->num_fs_inputs = v3d->prog.fs->prog_data.fs->base.num_inputs;
 525         STATIC_ASSERT(sizeof(key->fs_inputs) ==
 526                       sizeof(v3d->prog.fs->prog_data.fs->input_slots));
 527         memcpy(key->fs_inputs, v3d->prog.fs->prog_data.fs->input_slots,
 528                sizeof(key->fs_inputs));
 529         key->clamp_color = v3d->rasterizer->base.clamp_vertex_color;
 530
 531         key->per_vertex_point_size =
 532                 (prim_mode == PIPE_PRIM_POINTS &&
 533                  v3d->rasterizer->base.point_size_per_vertex);
 534
 535         struct v3d_compiled_shader *vs =
 536                 v3d_get_compiled_shader(v3d, &key->base);
 537         if (vs != v3d->prog.vs) {
 538                 v3d->prog.vs = vs;
 539                 v3d->dirty |= VC5_DIRTY_COMPILED_VS;
 540         }
 541
 542         key->is_coord = true;
 543         /* Coord shaders only output varyings used by transform feedback. */
 544         struct v3d_uncompiled_shader *shader_state = key->base.shader_state;
 545         memcpy(key->fs_inputs, shader_state->tf_outputs,
 546                sizeof(*key->fs_inputs) * shader_state->num_tf_outputs);
 547         if (shader_state->num_tf_outputs < key->num_fs_inputs) {
 548                 memset(&key->fs_inputs[shader_state->num_tf_outputs],
 549                        0,
 550                        sizeof(*key->fs_inputs) * (key->num_fs_inputs -
 551                                                   shader_state->num_tf_outputs));
 552         }
 553         key->num_fs_inputs = shader_state->num_tf_outputs;
 554
 555         struct v3d_compiled_shader *cs =
 556                 v3d_get_compiled_shader(v3d, &key->base);
 557         if (cs != v3d->prog.cs) {
 558                 v3d->prog.cs = cs;
 559                 v3d->dirty |= VC5_DIRTY_COMPILED_CS;
 560         }
 561 }
 562
 563 void
 564 v3d_update_compiled_shaders(struct v3d_context *v3d, uint8_t prim_mode)
 565 {
 566         v3d_update_compiled_fs(v3d, prim_mode);
 567         v3d_update_compiled_vs(v3d, prim_mode);
 568 }
 569
 570 static uint32_t
 571 fs_cache_hash(const void *key)
 572 {
 573         return _mesa_hash_data(key, sizeof(struct v3d_fs_key));
 574 }
 575
 576 static uint32_t
 577 vs_cache_hash(const void *key)
 578 {
 579         return _mesa_hash_data(key, sizeof(struct v3d_vs_key));
 580 }
 581
 582 static bool
 583 fs_cache_compare(const void *key1, const void *key2)
 584 {
 585         return memcmp(key1, key2, sizeof(struct v3d_fs_key)) == 0;
 586 }
 587
 588 static bool
 589 vs_cache_compare(const void *key1, const void *key2)
 590 {
 591         return memcmp(key1, key2, sizeof(struct v3d_vs_key)) == 0;
 592 }
 593
 594 static void
 595 delete_from_cache_if_matches(struct hash_table *ht,
 596                              struct v3d_compiled_shader **last_compile,
 597                              struct hash_entry *entry,
 598                              struct v3d_uncompiled_shader *so)
 599 {
 600         const struct v3d_key *key = entry->key;
 601
 602         if (key->shader_state == so) {
 603                 struct v3d_compiled_shader *shader = entry->data;
 604                 _mesa_hash_table_remove(ht, entry);
 605                 v3d_bo_unreference(&shader->bo);
 606
 607                 if (shader == *last_compile)
 608                         *last_compile = NULL;
 609
 610                 ralloc_free(shader);
 611         }
 612 }
 613
 614 static void
 615 v3d_shader_state_delete(struct pipe_context *pctx, void *hwcso)
 616 {
 617         struct v3d_context *v3d = v3d_context(pctx);
 618         struct v3d_uncompiled_shader *so = hwcso;
 619
 620         struct hash_entry *entry;
 621         hash_table_foreach(v3d->fs_cache, entry) {
 622                 delete_from_cache_if_matches(v3d->fs_cache, &v3d->prog.fs,
 623                                              entry, so);
 624         }
 625         hash_table_foreach(v3d->vs_cache, entry) {
 626                 delete_from_cache_if_matches(v3d->vs_cache, &v3d->prog.vs,
 627                                              entry, so);
 628         }
 629
 630         ralloc_free(so->base.ir.nir);
 631         free(so);
 632 }
 633
 634 static void
 635 v3d_fp_state_bind(struct pipe_context *pctx, void *hwcso)
 636 {
 637         struct v3d_context *v3d = v3d_context(pctx);
 638         v3d->prog.bind_fs = hwcso;
 639         v3d->dirty |= VC5_DIRTY_UNCOMPILED_FS;
 640 }
 641
 642 static void
 643 v3d_vp_state_bind(struct pipe_context *pctx, void *hwcso)
 644 {
 645         struct v3d_context *v3d = v3d_context(pctx);
 646         v3d->prog.bind_vs = hwcso;
 647         v3d->dirty |= VC5_DIRTY_UNCOMPILED_VS;
 648 }
 649
 650 void
 651 v3d_program_init(struct pipe_context *pctx)
 652 {
 653         struct v3d_context *v3d = v3d_context(pctx);
 654
 655         pctx->create_vs_state = v3d_shader_state_create;
 656         pctx->delete_vs_state = v3d_shader_state_delete;
 657
 658         pctx->create_fs_state = v3d_shader_state_create;
 659         pctx->delete_fs_state = v3d_shader_state_delete;
 660
 661         pctx->bind_fs_state = v3d_fp_state_bind;
 662         pctx->bind_vs_state = v3d_vp_state_bind;
 663
 664         v3d->fs_cache = _mesa_hash_table_create(pctx, fs_cache_hash,
 665                                                 fs_cache_compare);
 666         v3d->vs_cache = _mesa_hash_table_create(pctx, vs_cache_hash,
 667                                                 vs_cache_compare);
 668 }
 669
 670 void
 671 v3d_program_fini(struct pipe_context *pctx)
 672 {
 673         struct v3d_context *v3d = v3d_context(pctx);
 674
 675         struct hash_entry *entry;
 676         hash_table_foreach(v3d->fs_cache, entry) {
 677                 struct v3d_compiled_shader *shader = entry->data;
 678                 v3d_bo_unreference(&shader->bo);
 679                 ralloc_free(shader);
 680                 _mesa_hash_table_remove(v3d->fs_cache, entry);
 681         }
 682
 683         hash_table_foreach(v3d->vs_cache, entry) {
 684                 struct v3d_compiled_shader *shader = entry->data;
 685                 v3d_bo_unreference(&shader->bo);
 686                 ralloc_free(shader);
 687                 _mesa_hash_table_remove(v3d->vs_cache, entry);
 688         }
 689
 690         v3d_bo_unreference(&v3d->prog.spill_bo);
 691 }