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