src/mesa/state_tracker/st_shader_cache.c

   1 /*
   2  * Copyright © 2017 Timothy Arceri
   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 <stdio.h>
  25
  26 #include "st_program.h"
  27 #include "st_shader_cache.h"
  28 #include "compiler/glsl/program.h"
  29 #include "pipe/p_shader_tokens.h"
  30 #include "program/ir_to_mesa.h"
  31 #include "util/u_memory.h"
  32
  33 static void
  34 write_stream_out_to_cache(struct blob *blob,
  35                           struct pipe_shader_state *tgsi)
  36 {
  37    blob_write_bytes(blob, &tgsi->stream_output,
  38                     sizeof(tgsi->stream_output));
  39 }
  40
  41 static void
  42 write_tgsi_to_cache(struct blob *blob, struct pipe_shader_state *tgsi,
  43                     struct st_context *st, unsigned char *sha1,
  44                     unsigned num_tokens)
  45 {
  46    blob_write_uint32(blob, num_tokens);
  47    blob_write_bytes(blob, tgsi->tokens,
  48                     num_tokens * sizeof(struct tgsi_token));
  49
  50    disk_cache_put(st->ctx->Cache, sha1, blob->data, blob->size);
  51 }
  52
  53 /**
  54  * Store tgsi and any other required state in on-disk shader cache.
  55  */
  56 void
  57 st_store_tgsi_in_disk_cache(struct st_context *st, struct gl_program *prog,
  58                             struct pipe_shader_state *out_state,
  59                             unsigned num_tokens)
  60 {
  61    if (!st->ctx->Cache)
  62       return;
  63
  64    /* Exit early when we are dealing with a ff shader with no source file to
  65     * generate a source from.
  66     */
  67    if (*prog->sh.data->sha1 == 0)
  68       return;
  69
  70    unsigned char *sha1;
  71    struct blob *blob = blob_create();
  72
  73    switch (prog->info.stage) {
  74    case MESA_SHADER_VERTEX: {
  75       struct st_vertex_program *stvp = (struct st_vertex_program *) prog;
  76       sha1 = stvp->sha1;
  77
  78       blob_write_uint32(blob, stvp->num_inputs);
  79       blob_write_bytes(blob, stvp->index_to_input,
  80                        sizeof(stvp->index_to_input));
  81       blob_write_bytes(blob, stvp->result_to_output,
  82                        sizeof(stvp->result_to_output));
  83
  84       write_stream_out_to_cache(blob, &stvp->tgsi);
  85       write_tgsi_to_cache(blob, &stvp->tgsi, st, sha1, num_tokens);
  86       break;
  87    }
  88    case MESA_SHADER_TESS_CTRL: {
  89       struct st_tessctrl_program *stcp = (struct st_tessctrl_program *) prog;
  90       sha1 = stcp->sha1;
  91
  92       write_stream_out_to_cache(blob, out_state);
  93       write_tgsi_to_cache(blob, out_state, st, sha1, num_tokens);
  94       break;
  95    }
  96    case MESA_SHADER_TESS_EVAL: {
  97       struct st_tesseval_program *step = (struct st_tesseval_program *) prog;
  98       sha1 = step->sha1;
  99
 100       write_stream_out_to_cache(blob, out_state);
 101       write_tgsi_to_cache(blob, out_state, st, sha1, num_tokens);
 102       break;
 103    }
 104    case MESA_SHADER_GEOMETRY: {
 105       struct st_geometry_program *stgp = (struct st_geometry_program *) prog;
 106       sha1 = stgp->sha1;
 107
 108       write_stream_out_to_cache(blob, out_state);
 109       write_tgsi_to_cache(blob, out_state, st, sha1, num_tokens);
 110       break;
 111    }
 112    case MESA_SHADER_FRAGMENT: {
 113       struct st_fragment_program *stfp = (struct st_fragment_program *) prog;
 114       sha1 = stfp->sha1;
 115
 116       write_tgsi_to_cache(blob, &stfp->tgsi, st, sha1, num_tokens);
 117       break;
 118    }
 119    case MESA_SHADER_COMPUTE: {
 120       struct st_compute_program *stcp = (struct st_compute_program *) prog;
 121       sha1 = stcp->sha1;
 122
 123       write_tgsi_to_cache(blob, out_state, st, sha1, num_tokens);
 124       break;
 125    }
 126    default:
 127       unreachable("Unsupported stage");
 128    }
 129
 130    if (st->ctx->_Shader->Flags & GLSL_CACHE_INFO) {
 131       char sha1_buf[41];
 132       _mesa_sha1_format(sha1_buf, sha1);
 133       fprintf(stderr, "putting %s tgsi_tokens in cache: %s\n",
 134               _mesa_shader_stage_to_string(prog->info.stage), sha1_buf);
 135    }
 136
 137    free(blob);
 138 }
 139
 140 static void
 141 read_stream_out_from_cache(struct blob_reader *blob_reader,
 142                            struct pipe_shader_state *tgsi)
 143 {
 144    blob_copy_bytes(blob_reader, (uint8_t *) &tgsi->stream_output,
 145                     sizeof(tgsi->stream_output));
 146 }
 147
 148 static void
 149 read_tgsi_from_cache(struct blob_reader *blob_reader,
 150                      const struct tgsi_token **tokens)
 151 {
 152    uint32_t num_tokens  = blob_read_uint32(blob_reader);
 153    unsigned tokens_size = num_tokens * sizeof(struct tgsi_token);
 154    *tokens = (const struct tgsi_token*) MALLOC(tokens_size);
 155    blob_copy_bytes(blob_reader, (uint8_t *) *tokens, tokens_size);
 156 }
 157
 158 bool
 159 st_load_tgsi_from_disk_cache(struct gl_context *ctx,
 160                              struct gl_shader_program *prog)
 161 {
 162    if (!ctx->Cache)
 163       return false;
 164
 165    unsigned char *stage_sha1[MESA_SHADER_STAGES];
 166    char sha1_buf[41];
 167
 168    /* Compute and store sha1 for each stage. These will be reused by the
 169     * cache store pass if we fail to find the cached tgsi.
 170     */
 171    for (unsigned i = 0; i < MESA_SHADER_STAGES; i++) {
 172       if (prog->_LinkedShaders[i] == NULL)
 173          continue;
 174
 175       char *buf = ralloc_strdup(NULL, "tgsi_tokens ");
 176       _mesa_sha1_format(sha1_buf,
 177                         prog->_LinkedShaders[i]->Program->sh.data->sha1);
 178       ralloc_strcat(&buf, sha1_buf);
 179
 180       struct gl_program *glprog = prog->_LinkedShaders[i]->Program;
 181       switch (glprog->info.stage) {
 182       case MESA_SHADER_VERTEX: {
 183          struct st_vertex_program *stvp = (struct st_vertex_program *) glprog;
 184          stage_sha1[i] = stvp->sha1;
 185          ralloc_strcat(&buf, " vs");
 186          disk_cache_compute_key(ctx->Cache, buf, strlen(buf), stage_sha1[i]);
 187          break;
 188       }
 189       case MESA_SHADER_TESS_CTRL: {
 190          struct st_tessctrl_program *stcp =
 191             (struct st_tessctrl_program *) glprog;
 192          stage_sha1[i] = stcp->sha1;
 193          ralloc_strcat(&buf, " tcs");
 194          disk_cache_compute_key(ctx->Cache, buf, strlen(buf), stage_sha1[i]);
 195          break;
 196       }
 197       case MESA_SHADER_TESS_EVAL: {
 198          struct st_tesseval_program *step =
 199             (struct st_tesseval_program *) glprog;
 200          stage_sha1[i] = step->sha1;
 201          ralloc_strcat(&buf, " tes");
 202          disk_cache_compute_key(ctx->Cache, buf, strlen(buf), stage_sha1[i]);
 203          break;
 204       }
 205       case MESA_SHADER_GEOMETRY: {
 206          struct st_geometry_program *stgp =
 207             (struct st_geometry_program *) glprog;
 208          stage_sha1[i] = stgp->sha1;
 209          ralloc_strcat(&buf, " gs");
 210          disk_cache_compute_key(ctx->Cache, buf, strlen(buf), stage_sha1[i]);
 211          break;
 212       }
 213       case MESA_SHADER_FRAGMENT: {
 214          struct st_fragment_program *stfp =
 215             (struct st_fragment_program *) glprog;
 216          stage_sha1[i] = stfp->sha1;
 217          ralloc_strcat(&buf, " fs");
 218          disk_cache_compute_key(ctx->Cache, buf, strlen(buf), stage_sha1[i]);
 219          break;
 220       }
 221       case MESA_SHADER_COMPUTE: {
 222          struct st_compute_program *stcp =
 223             (struct st_compute_program *) glprog;
 224          stage_sha1[i] = stcp->sha1;
 225          ralloc_strcat(&buf, " cs");
 226          disk_cache_compute_key(ctx->Cache, buf, strlen(buf), stage_sha1[i]);
 227          break;
 228       }
 229       default:
 230          unreachable("Unsupported stage");
 231       }
 232
 233       ralloc_free(buf);
 234    }
 235
 236    /* Now that we have created the sha1 keys that will be used for writting to
 237     * the tgsi cache fallback to the regular glsl to tgsi path if we didn't
 238     * load the GLSL IR from cache. We do this as glsl to tgsi can alter things
 239     * such as gl_program_parameter_list which holds things like uniforms.
 240     */
 241    if (prog->data->LinkStatus != linking_skipped)
 242       return false;
 243
 244    struct st_context *st = st_context(ctx);
 245    uint8_t *buffer = NULL;
 246    for (unsigned i = 0; i < MESA_SHADER_STAGES; i++) {
 247       if (prog->_LinkedShaders[i] == NULL)
 248          continue;
 249
 250       unsigned char *sha1 = stage_sha1[i];
 251       size_t size;
 252       buffer = (uint8_t *) disk_cache_get(ctx->Cache, sha1, &size);
 253       if (buffer) {
 254          struct blob_reader blob_reader;
 255          blob_reader_init(&blob_reader, buffer, size);
 256
 257          struct gl_program *glprog = prog->_LinkedShaders[i]->Program;
 258          switch (glprog->info.stage) {
 259          case MESA_SHADER_VERTEX: {
 260             struct st_vertex_program *stvp =
 261                (struct st_vertex_program *) glprog;
 262
 263             st_release_vp_variants(st, stvp);
 264
 265             stvp->num_inputs = blob_read_uint32(&blob_reader);
 266             blob_copy_bytes(&blob_reader, (uint8_t *) stvp->index_to_input,
 267                             sizeof(stvp->index_to_input));
 268             blob_copy_bytes(&blob_reader, (uint8_t *) stvp->result_to_output,
 269                             sizeof(stvp->result_to_output));
 270
 271             read_stream_out_from_cache(&blob_reader, &stvp->tgsi);
 272             read_tgsi_from_cache(&blob_reader, &stvp->tgsi.tokens);
 273
 274             if (st->vp == stvp)
 275                st->dirty |= ST_NEW_VERTEX_PROGRAM(st, stvp);
 276
 277             break;
 278          }
 279          case MESA_SHADER_TESS_CTRL: {
 280             struct st_tessctrl_program *sttcp =
 281                (struct st_tessctrl_program *) glprog;
 282
 283             st_release_basic_variants(st, sttcp->Base.Target,
 284                                       &sttcp->variants, &sttcp->tgsi);
 285
 286             read_stream_out_from_cache(&blob_reader, &sttcp->tgsi);
 287             read_tgsi_from_cache(&blob_reader, &sttcp->tgsi.tokens);
 288
 289             if (st->tcp == sttcp)
 290                st->dirty |= sttcp->affected_states;
 291
 292             break;
 293          }
 294          case MESA_SHADER_TESS_EVAL: {
 295             struct st_tesseval_program *sttep =
 296                (struct st_tesseval_program *) glprog;
 297
 298             st_release_basic_variants(st, sttep->Base.Target,
 299                                       &sttep->variants, &sttep->tgsi);
 300
 301             read_stream_out_from_cache(&blob_reader, &sttep->tgsi);
 302             read_tgsi_from_cache(&blob_reader, &sttep->tgsi.tokens);
 303
 304             if (st->tep == sttep)
 305                st->dirty |= sttep->affected_states;
 306
 307             break;
 308          }
 309          case MESA_SHADER_GEOMETRY: {
 310             struct st_geometry_program *stgp =
 311                (struct st_geometry_program *) glprog;
 312
 313             st_release_basic_variants(st, stgp->Base.Target, &stgp->variants,
 314                                       &stgp->tgsi);
 315
 316             read_stream_out_from_cache(&blob_reader, &stgp->tgsi);
 317             read_tgsi_from_cache(&blob_reader, &stgp->tgsi.tokens);
 318
 319             if (st->gp == stgp)
 320                st->dirty |= stgp->affected_states;
 321
 322             break;
 323          }
 324          case MESA_SHADER_FRAGMENT: {
 325             struct st_fragment_program *stfp =
 326                (struct st_fragment_program *) glprog;
 327
 328             st_release_fp_variants(st, stfp);
 329
 330             read_tgsi_from_cache(&blob_reader, &stfp->tgsi.tokens);
 331
 332             if (st->fp == stfp)
 333                st->dirty |= stfp->affected_states;
 334
 335             break;
 336          }
 337          case MESA_SHADER_COMPUTE: {
 338             struct st_compute_program *stcp =
 339                (struct st_compute_program *) glprog;
 340
 341             st_release_cp_variants(st, stcp);
 342
 343             read_tgsi_from_cache(&blob_reader,
 344                                  (const struct tgsi_token**) &stcp->tgsi.prog);
 345
 346             stcp->tgsi.req_local_mem = stcp->Base.info.cs.shared_size;
 347             stcp->tgsi.req_private_mem = 0;
 348             stcp->tgsi.req_input_mem = 0;
 349
 350             if (st->cp == stcp)
 351                 st->dirty |= stcp->affected_states;
 352
 353             break;
 354          }
 355          default:
 356             unreachable("Unsupported stage");
 357          }
 358
 359          if (blob_reader.current != blob_reader.end || blob_reader.overrun) {
 360             /* Something very bad has gone wrong discard the item from the
 361              * cache and rebuild/link from source.
 362              */
 363             assert(!"Invalid TGSI shader disk cache item!");
 364
 365             if (ctx->_Shader->Flags & GLSL_CACHE_INFO) {
 366                fprintf(stderr, "Error reading program from cache (invalid "
 367                        "TGSI cache item)\n");
 368             }
 369
 370             disk_cache_remove(ctx->Cache, sha1);
 371
 372             goto fallback_recompile;
 373          }
 374
 375          if (ctx->_Shader->Flags & GLSL_CACHE_INFO) {
 376             _mesa_sha1_format(sha1_buf, sha1);
 377             fprintf(stderr, "%s tgsi_tokens retrieved from cache: %s\n",
 378                     _mesa_shader_stage_to_string(i), sha1_buf);
 379          }
 380
 381          st_set_prog_affected_state_flags(glprog);
 382          _mesa_associate_uniform_storage(ctx, prog, glprog->Parameters,
 383                                          false);
 384
 385          free(buffer);
 386       } else {
 387          /* Failed to find a matching cached shader so fallback to recompile.
 388           */
 389          if (ctx->_Shader->Flags & GLSL_CACHE_INFO) {
 390             fprintf(stderr, "TGSI cache item not found falling back to "
 391                     "compile.\n");
 392          }
 393
 394          goto fallback_recompile;
 395       }
 396    }
 397
 398    return true;
 399
 400 fallback_recompile:
 401    free(buffer);
 402
 403    for (unsigned i = 0; i < prog->NumShaders; i++) {
 404       _mesa_glsl_compile_shader(ctx, prog->Shaders[i], false, false, true);
 405    }
 406
 407    prog->data->cache_fallback = true;
 408    _mesa_glsl_link_shader(ctx, prog);
 409
 410    return true;
 411 }