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    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    struct st_context *st = st_context(ctx);
 229    uint8_t *buffer = NULL;
 230    for (unsigned i = 0; i < MESA_SHADER_STAGES; i++) {
 231       if (prog->_LinkedShaders[i] == NULL)
 232          continue;
 233
 234       unsigned char *sha1 = stage_sha1[i];
 235       size_t size;
 236       buffer = (uint8_t *) disk_cache_get(ctx->Cache, sha1, &size);
 237       if (buffer) {
 238          struct blob_reader blob_reader;
 239          blob_reader_init(&blob_reader, buffer, size);
 240
 241          struct gl_program *glprog = prog->_LinkedShaders[i]->Program;
 242          switch (glprog->info.stage) {
 243          case MESA_SHADER_VERTEX: {
 244             struct st_vertex_program *stvp =
 245                (struct st_vertex_program *) glprog;
 246
 247             st_release_vp_variants(st, stvp);
 248
 249             stvp->num_inputs = blob_read_uint32(&blob_reader);
 250             blob_copy_bytes(&blob_reader, (uint8_t *) stvp->index_to_input,
 251                             sizeof(stvp->index_to_input));
 252             blob_copy_bytes(&blob_reader, (uint8_t *) stvp->result_to_output,
 253                             sizeof(stvp->result_to_output));
 254
 255             read_stream_out_from_cache(&blob_reader, &stvp->tgsi);
 256             read_tgsi_from_cache(&blob_reader, &stvp->tgsi.tokens);
 257
 258             if (st->vp == stvp)
 259                st->dirty |= ST_NEW_VERTEX_PROGRAM(st, stvp);
 260
 261             break;
 262          }
 263          case MESA_SHADER_TESS_CTRL: {
 264             struct st_common_program *sttcp = st_common_program(glprog);
 265
 266             st_release_basic_variants(st, sttcp->Base.Target,
 267                                       &sttcp->variants, &sttcp->tgsi);
 268
 269             read_stream_out_from_cache(&blob_reader, &sttcp->tgsi);
 270             read_tgsi_from_cache(&blob_reader, &sttcp->tgsi.tokens);
 271
 272             if (st->tcp == sttcp)
 273                st->dirty |= sttcp->affected_states;
 274
 275             break;
 276          }
 277          case MESA_SHADER_TESS_EVAL: {
 278             struct st_common_program *sttep = st_common_program(glprog);
 279
 280             st_release_basic_variants(st, sttep->Base.Target,
 281                                       &sttep->variants, &sttep->tgsi);
 282
 283             read_stream_out_from_cache(&blob_reader, &sttep->tgsi);
 284             read_tgsi_from_cache(&blob_reader, &sttep->tgsi.tokens);
 285
 286             if (st->tep == sttep)
 287                st->dirty |= sttep->affected_states;
 288
 289             break;
 290          }
 291          case MESA_SHADER_GEOMETRY: {
 292             struct st_common_program *stgp = st_common_program(glprog);
 293
 294             st_release_basic_variants(st, stgp->Base.Target, &stgp->variants,
 295                                       &stgp->tgsi);
 296
 297             read_stream_out_from_cache(&blob_reader, &stgp->tgsi);
 298             read_tgsi_from_cache(&blob_reader, &stgp->tgsi.tokens);
 299
 300             if (st->gp == stgp)
 301                st->dirty |= stgp->affected_states;
 302
 303             break;
 304          }
 305          case MESA_SHADER_FRAGMENT: {
 306             struct st_fragment_program *stfp =
 307                (struct st_fragment_program *) glprog;
 308
 309             st_release_fp_variants(st, stfp);
 310
 311             read_tgsi_from_cache(&blob_reader, &stfp->tgsi.tokens);
 312
 313             if (st->fp == stfp)
 314                st->dirty |= stfp->affected_states;
 315
 316             break;
 317          }
 318          case MESA_SHADER_COMPUTE: {
 319             struct st_compute_program *stcp =
 320                (struct st_compute_program *) glprog;
 321
 322             st_release_cp_variants(st, stcp);
 323
 324             read_tgsi_from_cache(&blob_reader,
 325                                  (const struct tgsi_token**) &stcp->tgsi.prog);
 326
 327             stcp->tgsi.req_local_mem = stcp->Base.info.cs.shared_size;
 328             stcp->tgsi.req_private_mem = 0;
 329             stcp->tgsi.req_input_mem = 0;
 330
 331             if (st->cp == stcp)
 332                 st->dirty |= stcp->affected_states;
 333
 334             break;
 335          }
 336          default:
 337             unreachable("Unsupported stage");
 338          }
 339
 340          if (blob_reader.current != blob_reader.end || blob_reader.overrun) {
 341             /* Something very bad has gone wrong discard the item from the
 342              * cache and rebuild/link from source.
 343              */
 344             assert(!"Invalid TGSI shader disk cache item!");
 345
 346             if (ctx->_Shader->Flags & GLSL_CACHE_INFO) {
 347                fprintf(stderr, "Error reading program from cache (invalid "
 348                        "TGSI cache item)\n");
 349             }
 350
 351             disk_cache_remove(ctx->Cache, sha1);
 352
 353             goto fallback_recompile;
 354          }
 355
 356          if (ctx->_Shader->Flags & GLSL_CACHE_INFO) {
 357             _mesa_sha1_format(sha1_buf, sha1);
 358             fprintf(stderr, "%s tgsi_tokens retrieved from cache: %s\n",
 359                     _mesa_shader_stage_to_string(i), sha1_buf);
 360          }
 361
 362          st_set_prog_affected_state_flags(glprog);
 363          _mesa_associate_uniform_storage(ctx, prog, glprog->Parameters,
 364                                          false);
 365
 366          free(buffer);
 367       } else {
 368          /* Failed to find a matching cached shader so fallback to recompile.
 369           */
 370          if (ctx->_Shader->Flags & GLSL_CACHE_INFO) {
 371             fprintf(stderr, "TGSI cache item not found falling back to "
 372                     "compile.\n");
 373          }
 374
 375          goto fallback_recompile;
 376       }
 377    }
 378
 379    return true;
 380
 381 fallback_recompile:
 382    free(buffer);
 383
 384    for (unsigned i = 0; i < prog->NumShaders; i++) {
 385       _mesa_glsl_compile_shader(ctx, prog->Shaders[i], false, false, true);
 386    }
 387
 388    prog->data->cache_fallback = true;
 389    _mesa_glsl_link_shader(ctx, prog);
 390
 391    return true;
 392 }