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    static const char zero[sizeof(prog->sh.data->sha1)] = {0};
  68    if (memcmp(prog->sh.data->sha1, zero, sizeof(prog->sh.data->sha1)) == 0)
  69       return;
  70
  71    unsigned char *sha1;
  72    struct blob *blob = blob_create();
  73
  74    switch (prog->info.stage) {
  75    case MESA_SHADER_VERTEX: {
  76       struct st_vertex_program *stvp = (struct st_vertex_program *) prog;
  77       sha1 = stvp->sha1;
  78
  79       blob_write_uint32(blob, stvp->num_inputs);
  80       blob_write_bytes(blob, stvp->index_to_input,
  81                        sizeof(stvp->index_to_input));
  82       blob_write_bytes(blob, stvp->result_to_output,
  83                        sizeof(stvp->result_to_output));
  84
  85       write_stream_out_to_cache(blob, &stvp->tgsi);
  86       write_tgsi_to_cache(blob, &stvp->tgsi, st, sha1, num_tokens);
  87       break;
  88    }
  89    case MESA_SHADER_TESS_CTRL: {
  90       struct st_tessctrl_program *stcp = (struct st_tessctrl_program *) prog;
  91       sha1 = stcp->sha1;
  92
  93       write_stream_out_to_cache(blob, out_state);
  94       write_tgsi_to_cache(blob, out_state, st, sha1, num_tokens);
  95       break;
  96    }
  97    case MESA_SHADER_TESS_EVAL: {
  98       struct st_tesseval_program *step = (struct st_tesseval_program *) prog;
  99       sha1 = step->sha1;
 100
 101       write_stream_out_to_cache(blob, out_state);
 102       write_tgsi_to_cache(blob, out_state, st, sha1, num_tokens);
 103       break;
 104    }
 105    case MESA_SHADER_GEOMETRY: {
 106       struct st_geometry_program *stgp = (struct st_geometry_program *) prog;
 107       sha1 = stgp->sha1;
 108
 109       write_stream_out_to_cache(blob, out_state);
 110       write_tgsi_to_cache(blob, out_state, st, sha1, num_tokens);
 111       break;
 112    }
 113    case MESA_SHADER_FRAGMENT: {
 114       struct st_fragment_program *stfp = (struct st_fragment_program *) prog;
 115       sha1 = stfp->sha1;
 116
 117       write_tgsi_to_cache(blob, &stfp->tgsi, st, sha1, num_tokens);
 118       break;
 119    }
 120    case MESA_SHADER_COMPUTE: {
 121       struct st_compute_program *stcp = (struct st_compute_program *) prog;
 122       sha1 = stcp->sha1;
 123
 124       write_tgsi_to_cache(blob, out_state, st, sha1, num_tokens);
 125       break;
 126    }
 127    default:
 128       unreachable("Unsupported stage");
 129    }
 130
 131    if (st->ctx->_Shader->Flags & GLSL_CACHE_INFO) {
 132       char sha1_buf[41];
 133       _mesa_sha1_format(sha1_buf, sha1);
 134       fprintf(stderr, "putting %s tgsi_tokens in cache: %s\n",
 135               _mesa_shader_stage_to_string(prog->info.stage), sha1_buf);
 136    }
 137
 138    free(blob);
 139 }
 140
 141 static void
 142 read_stream_out_from_cache(struct blob_reader *blob_reader,
 143                            struct pipe_shader_state *tgsi)
 144 {
 145    blob_copy_bytes(blob_reader, (uint8_t *) &tgsi->stream_output,
 146                     sizeof(tgsi->stream_output));
 147 }
 148
 149 static void
 150 read_tgsi_from_cache(struct blob_reader *blob_reader,
 151                      const struct tgsi_token **tokens)
 152 {
 153    uint32_t num_tokens  = blob_read_uint32(blob_reader);
 154    unsigned tokens_size = num_tokens * sizeof(struct tgsi_token);
 155    *tokens = (const struct tgsi_token*) MALLOC(tokens_size);
 156    blob_copy_bytes(blob_reader, (uint8_t *) *tokens, tokens_size);
 157 }
 158
 159 bool
 160 st_load_tgsi_from_disk_cache(struct gl_context *ctx,
 161                              struct gl_shader_program *prog)
 162 {
 163    if (!ctx->Cache)
 164       return false;
 165
 166    unsigned char *stage_sha1[MESA_SHADER_STAGES];
 167    char sha1_buf[41];
 168
 169    /* Compute and store sha1 for each stage. These will be reused by the
 170     * cache store pass if we fail to find the cached tgsi.
 171     */
 172    for (unsigned i = 0; i < MESA_SHADER_STAGES; i++) {
 173       if (prog->_LinkedShaders[i] == NULL)
 174          continue;
 175
 176       char *buf = ralloc_strdup(NULL, "tgsi_tokens ");
 177       _mesa_sha1_format(sha1_buf,
 178                         prog->_LinkedShaders[i]->Program->sh.data->sha1);
 179       ralloc_strcat(&buf, sha1_buf);
 180
 181       struct gl_program *glprog = prog->_LinkedShaders[i]->Program;
 182       switch (glprog->info.stage) {
 183       case MESA_SHADER_VERTEX: {
 184          struct st_vertex_program *stvp = (struct st_vertex_program *) glprog;
 185          stage_sha1[i] = stvp->sha1;
 186          ralloc_strcat(&buf, " vs");
 187          disk_cache_compute_key(ctx->Cache, buf, strlen(buf), stage_sha1[i]);
 188          break;
 189       }
 190       case MESA_SHADER_TESS_CTRL: {
 191          struct st_tessctrl_program *stcp =
 192             (struct st_tessctrl_program *) glprog;
 193          stage_sha1[i] = stcp->sha1;
 194          ralloc_strcat(&buf, " tcs");
 195          disk_cache_compute_key(ctx->Cache, buf, strlen(buf), stage_sha1[i]);
 196          break;
 197       }
 198       case MESA_SHADER_TESS_EVAL: {
 199          struct st_tesseval_program *step =
 200             (struct st_tesseval_program *) glprog;
 201          stage_sha1[i] = step->sha1;
 202          ralloc_strcat(&buf, " tes");
 203          disk_cache_compute_key(ctx->Cache, buf, strlen(buf), stage_sha1[i]);
 204          break;
 205       }
 206       case MESA_SHADER_GEOMETRY: {
 207          struct st_geometry_program *stgp =
 208             (struct st_geometry_program *) glprog;
 209          stage_sha1[i] = stgp->sha1;
 210          ralloc_strcat(&buf, " gs");
 211          disk_cache_compute_key(ctx->Cache, buf, strlen(buf), stage_sha1[i]);
 212          break;
 213       }
 214       case MESA_SHADER_FRAGMENT: {
 215          struct st_fragment_program *stfp =
 216             (struct st_fragment_program *) glprog;
 217          stage_sha1[i] = stfp->sha1;
 218          ralloc_strcat(&buf, " fs");
 219          disk_cache_compute_key(ctx->Cache, buf, strlen(buf), stage_sha1[i]);
 220          break;
 221       }
 222       case MESA_SHADER_COMPUTE: {
 223          struct st_compute_program *stcp =
 224             (struct st_compute_program *) glprog;
 225          stage_sha1[i] = stcp->sha1;
 226          ralloc_strcat(&buf, " cs");
 227          disk_cache_compute_key(ctx->Cache, buf, strlen(buf), stage_sha1[i]);
 228          break;
 229       }
 230       default:
 231          unreachable("Unsupported stage");
 232       }
 233
 234       ralloc_free(buf);
 235    }
 236
 237    /* Now that we have created the sha1 keys that will be used for writting to
 238     * the tgsi cache fallback to the regular glsl to tgsi path if we didn't
 239     * load the GLSL IR from cache. We do this as glsl to tgsi can alter things
 240     * such as gl_program_parameter_list which holds things like uniforms.
 241     */
 242    if (prog->data->LinkStatus != linking_skipped)
 243       return false;
 244
 245    struct st_context *st = st_context(ctx);
 246    uint8_t *buffer = NULL;
 247    for (unsigned i = 0; i < MESA_SHADER_STAGES; i++) {
 248       if (prog->_LinkedShaders[i] == NULL)
 249          continue;
 250
 251       unsigned char *sha1 = stage_sha1[i];
 252       size_t size;
 253       buffer = (uint8_t *) disk_cache_get(ctx->Cache, sha1, &size);
 254       if (buffer) {
 255          struct blob_reader blob_reader;
 256          blob_reader_init(&blob_reader, buffer, size);
 257
 258          struct gl_program *glprog = prog->_LinkedShaders[i]->Program;
 259          switch (glprog->info.stage) {
 260          case MESA_SHADER_VERTEX: {
 261             struct st_vertex_program *stvp =
 262                (struct st_vertex_program *) glprog;
 263
 264             st_release_vp_variants(st, stvp);
 265
 266             stvp->num_inputs = blob_read_uint32(&blob_reader);
 267             blob_copy_bytes(&blob_reader, (uint8_t *) stvp->index_to_input,
 268                             sizeof(stvp->index_to_input));
 269             blob_copy_bytes(&blob_reader, (uint8_t *) stvp->result_to_output,
 270                             sizeof(stvp->result_to_output));
 271
 272             read_stream_out_from_cache(&blob_reader, &stvp->tgsi);
 273             read_tgsi_from_cache(&blob_reader, &stvp->tgsi.tokens);
 274
 275             if (st->vp == stvp)
 276                st->dirty |= ST_NEW_VERTEX_PROGRAM(st, stvp);
 277
 278             break;
 279          }
 280          case MESA_SHADER_TESS_CTRL: {
 281             struct st_tessctrl_program *sttcp =
 282                (struct st_tessctrl_program *) glprog;
 283
 284             st_release_basic_variants(st, sttcp->Base.Target,
 285                                       &sttcp->variants, &sttcp->tgsi);
 286
 287             read_stream_out_from_cache(&blob_reader, &sttcp->tgsi);
 288             read_tgsi_from_cache(&blob_reader, &sttcp->tgsi.tokens);
 289
 290             if (st->tcp == sttcp)
 291                st->dirty |= sttcp->affected_states;
 292
 293             break;
 294          }
 295          case MESA_SHADER_TESS_EVAL: {
 296             struct st_tesseval_program *sttep =
 297                (struct st_tesseval_program *) glprog;
 298
 299             st_release_basic_variants(st, sttep->Base.Target,
 300                                       &sttep->variants, &sttep->tgsi);
 301
 302             read_stream_out_from_cache(&blob_reader, &sttep->tgsi);
 303             read_tgsi_from_cache(&blob_reader, &sttep->tgsi.tokens);
 304
 305             if (st->tep == sttep)
 306                st->dirty |= sttep->affected_states;
 307
 308             break;
 309          }
 310          case MESA_SHADER_GEOMETRY: {
 311             struct st_geometry_program *stgp =
 312                (struct st_geometry_program *) glprog;
 313
 314             st_release_basic_variants(st, stgp->Base.Target, &stgp->variants,
 315                                       &stgp->tgsi);
 316
 317             read_stream_out_from_cache(&blob_reader, &stgp->tgsi);
 318             read_tgsi_from_cache(&blob_reader, &stgp->tgsi.tokens);
 319
 320             if (st->gp == stgp)
 321                st->dirty |= stgp->affected_states;
 322
 323             break;
 324          }
 325          case MESA_SHADER_FRAGMENT: {
 326             struct st_fragment_program *stfp =
 327                (struct st_fragment_program *) glprog;
 328
 329             st_release_fp_variants(st, stfp);
 330
 331             read_tgsi_from_cache(&blob_reader, &stfp->tgsi.tokens);
 332
 333             if (st->fp == stfp)
 334                st->dirty |= stfp->affected_states;
 335
 336             break;
 337          }
 338          case MESA_SHADER_COMPUTE: {
 339             struct st_compute_program *stcp =
 340                (struct st_compute_program *) glprog;
 341
 342             st_release_cp_variants(st, stcp);
 343
 344             read_tgsi_from_cache(&blob_reader,
 345                                  (const struct tgsi_token**) &stcp->tgsi.prog);
 346
 347             stcp->tgsi.req_local_mem = stcp->Base.info.cs.shared_size;
 348             stcp->tgsi.req_private_mem = 0;
 349             stcp->tgsi.req_input_mem = 0;
 350
 351             if (st->cp == stcp)
 352                 st->dirty |= stcp->affected_states;
 353
 354             break;
 355          }
 356          default:
 357             unreachable("Unsupported stage");
 358          }
 359
 360          if (blob_reader.current != blob_reader.end || blob_reader.overrun) {
 361             /* Something very bad has gone wrong discard the item from the
 362              * cache and rebuild/link from source.
 363              */
 364             assert(!"Invalid TGSI shader disk cache item!");
 365
 366             if (ctx->_Shader->Flags & GLSL_CACHE_INFO) {
 367                fprintf(stderr, "Error reading program from cache (invalid "
 368                        "TGSI cache item)\n");
 369             }
 370
 371             disk_cache_remove(ctx->Cache, sha1);
 372
 373             goto fallback_recompile;
 374          }
 375
 376          if (ctx->_Shader->Flags & GLSL_CACHE_INFO) {
 377             _mesa_sha1_format(sha1_buf, sha1);
 378             fprintf(stderr, "%s tgsi_tokens retrieved from cache: %s\n",
 379                     _mesa_shader_stage_to_string(i), sha1_buf);
 380          }
 381
 382          st_set_prog_affected_state_flags(glprog);
 383          _mesa_associate_uniform_storage(ctx, prog, glprog->Parameters,
 384                                          false);
 385
 386          free(buffer);
 387       } else {
 388          /* Failed to find a matching cached shader so fallback to recompile.
 389           */
 390          if (ctx->_Shader->Flags & GLSL_CACHE_INFO) {
 391             fprintf(stderr, "TGSI cache item not found falling back to "
 392                     "compile.\n");
 393          }
 394
 395          goto fallback_recompile;
 396       }
 397    }
 398
 399    return true;
 400
 401 fallback_recompile:
 402    free(buffer);
 403
 404    for (unsigned i = 0; i < prog->NumShaders; i++) {
 405       _mesa_glsl_compile_shader(ctx, prog->Shaders[i], false, false, true);
 406    }
 407
 408    prog->data->cache_fallback = true;
 409    _mesa_glsl_link_shader(ctx, prog);
 410
 411    return true;
 412 }