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 "compiler/nir/nir.h"
  30 #include "compiler/nir/nir_serialize.h"
  31 #include "pipe/p_shader_tokens.h"
  32 #include "program/ir_to_mesa.h"
  33 #include "util/u_memory.h"
  34
  35 void
  36 st_get_program_binary_driver_sha1(struct gl_context *ctx, uint8_t *sha1)
  37 {
  38    disk_cache_compute_key(ctx->Cache, NULL, 0, sha1);
  39 }
  40
  41 static void
  42 write_stream_out_to_cache(struct blob *blob,
  43                           struct pipe_shader_state *tgsi)
  44 {
  45    blob_write_bytes(blob, &tgsi->stream_output,
  46                     sizeof(tgsi->stream_output));
  47 }
  48
  49 static void
  50 copy_blob_to_driver_cache_blob(struct blob *blob, struct gl_program *prog)
  51 {
  52    prog->driver_cache_blob = ralloc_size(NULL, blob->size);
  53    memcpy(prog->driver_cache_blob, blob->data, blob->size);
  54    prog->driver_cache_blob_size = blob->size;
  55 }
  56
  57 static void
  58 write_tgsi_to_cache(struct blob *blob, const struct tgsi_token *tokens,
  59                     struct gl_program *prog, unsigned num_tokens)
  60 {
  61    blob_write_uint32(blob, num_tokens);
  62    blob_write_bytes(blob, tokens, num_tokens * sizeof(struct tgsi_token));
  63    copy_blob_to_driver_cache_blob(blob, prog);
  64 }
  65
  66 static void
  67 write_nir_to_cache(struct blob *blob, struct gl_program *prog)
  68 {
  69    nir_serialize(blob, prog->nir);
  70    copy_blob_to_driver_cache_blob(blob, prog);
  71 }
  72
  73 static void
  74 st_serialise_ir_program(struct gl_context *ctx, struct gl_program *prog,
  75                         bool nir)
  76 {
  77    struct blob blob;
  78    blob_init(&blob);
  79
  80    switch (prog->info.stage) {
  81    case MESA_SHADER_VERTEX: {
  82       struct st_vertex_program *stvp = (struct st_vertex_program *) prog;
  83
  84       blob_write_uint32(&blob, stvp->num_inputs);
  85       blob_write_bytes(&blob, stvp->index_to_input,
  86                        sizeof(stvp->index_to_input));
  87       blob_write_bytes(&blob, stvp->result_to_output,
  88                        sizeof(stvp->result_to_output));
  89
  90       write_stream_out_to_cache(&blob, &stvp->tgsi);
  91
  92       if (nir)
  93          write_nir_to_cache(&blob, prog);
  94       else
  95          write_tgsi_to_cache(&blob, stvp->tgsi.tokens, prog,
  96                              stvp->num_tgsi_tokens);
  97       break;
  98    }
  99    case MESA_SHADER_TESS_CTRL:
 100    case MESA_SHADER_TESS_EVAL:
 101    case MESA_SHADER_GEOMETRY: {
 102       struct st_common_program *stcp = (struct st_common_program *) prog;
 103
 104       write_stream_out_to_cache(&blob, &stcp->tgsi);
 105
 106       if (nir)
 107          write_nir_to_cache(&blob, prog);
 108       else
 109          write_tgsi_to_cache(&blob, stcp->tgsi.tokens, prog,
 110                              stcp->num_tgsi_tokens);
 111       break;
 112    }
 113    case MESA_SHADER_FRAGMENT: {
 114       struct st_fragment_program *stfp = (struct st_fragment_program *) prog;
 115
 116       if (nir)
 117          write_nir_to_cache(&blob, prog);
 118       else
 119          write_tgsi_to_cache(&blob, stfp->tgsi.tokens, prog,
 120                              stfp->num_tgsi_tokens);
 121       break;
 122    }
 123    case MESA_SHADER_COMPUTE: {
 124       struct st_compute_program *stcp = (struct st_compute_program *) prog;
 125
 126       if (nir)
 127          write_nir_to_cache(&blob, prog);
 128       else
 129          write_tgsi_to_cache(&blob, stcp->tgsi.prog, prog,
 130                              stcp->num_tgsi_tokens);
 131       break;
 132    }
 133    default:
 134       unreachable("Unsupported stage");
 135    }
 136
 137    blob_finish(&blob);
 138 }
 139
 140 /**
 141  * Store tgsi and any other required state in on-disk shader cache.
 142  */
 143 void
 144 st_store_ir_in_disk_cache(struct st_context *st, struct gl_program *prog,
 145                           bool nir)
 146 {
 147    if (!st->ctx->Cache)
 148       return;
 149
 150    /* Exit early when we are dealing with a ff shader with no source file to
 151     * generate a source from.
 152     */
 153    static const char zero[sizeof(prog->sh.data->sha1)] = {0};
 154    if (memcmp(prog->sh.data->sha1, zero, sizeof(prog->sh.data->sha1)) == 0)
 155       return;
 156
 157    st_serialise_ir_program(st->ctx, prog, nir);
 158
 159    if (st->ctx->_Shader->Flags & GLSL_CACHE_INFO) {
 160       fprintf(stderr, "putting %s state tracker IR in cache\n",
 161               _mesa_shader_stage_to_string(prog->info.stage));
 162    }
 163 }
 164
 165 static void
 166 read_stream_out_from_cache(struct blob_reader *blob_reader,
 167                            struct pipe_shader_state *tgsi)
 168 {
 169    blob_copy_bytes(blob_reader, (uint8_t *) &tgsi->stream_output,
 170                     sizeof(tgsi->stream_output));
 171 }
 172
 173 static void
 174 read_tgsi_from_cache(struct blob_reader *blob_reader,
 175                      const struct tgsi_token **tokens,
 176                      unsigned *num_tokens)
 177 {
 178    *num_tokens  = blob_read_uint32(blob_reader);
 179    unsigned tokens_size = *num_tokens * sizeof(struct tgsi_token);
 180    *tokens = (const struct tgsi_token*) MALLOC(tokens_size);
 181    blob_copy_bytes(blob_reader, (uint8_t *) *tokens, tokens_size);
 182 }
 183
 184 static void
 185 st_deserialise_ir_program(struct gl_context *ctx,
 186                           struct gl_shader_program *shProg,
 187                           struct gl_program *prog, bool nir)
 188 {
 189    struct st_context *st = st_context(ctx);
 190    size_t size = prog->driver_cache_blob_size;
 191    uint8_t *buffer = (uint8_t *) prog->driver_cache_blob;
 192    const struct nir_shader_compiler_options *options =
 193       ctx->Const.ShaderCompilerOptions[prog->info.stage].NirOptions;
 194
 195    assert(prog->driver_cache_blob && prog->driver_cache_blob_size > 0);
 196
 197    struct blob_reader blob_reader;
 198    blob_reader_init(&blob_reader, buffer, size);
 199
 200    switch (prog->info.stage) {
 201    case MESA_SHADER_VERTEX: {
 202       struct st_vertex_program *stvp = (struct st_vertex_program *) prog;
 203
 204       st_release_vp_variants(st, stvp);
 205
 206       stvp->num_inputs = blob_read_uint32(&blob_reader);
 207       blob_copy_bytes(&blob_reader, (uint8_t *) stvp->index_to_input,
 208                       sizeof(stvp->index_to_input));
 209       blob_copy_bytes(&blob_reader, (uint8_t *) stvp->result_to_output,
 210                       sizeof(stvp->result_to_output));
 211
 212       read_stream_out_from_cache(&blob_reader, &stvp->tgsi);
 213
 214       if (nir) {
 215          stvp->tgsi.type = PIPE_SHADER_IR_NIR;
 216          stvp->shader_program = shProg;
 217          stvp->tgsi.ir.nir = nir_deserialize(NULL, options, &blob_reader);
 218          prog->nir = stvp->tgsi.ir.nir;
 219       } else {
 220          read_tgsi_from_cache(&blob_reader, &stvp->tgsi.tokens,
 221                               &stvp->num_tgsi_tokens);
 222       }
 223
 224       if (st->vp == stvp)
 225          st->dirty |= ST_NEW_VERTEX_PROGRAM(st, stvp);
 226
 227       break;
 228    }
 229    case MESA_SHADER_TESS_CTRL: {
 230       struct st_common_program *sttcp = st_common_program(prog);
 231
 232       st_release_basic_variants(st, sttcp->Base.Target,
 233                                 &sttcp->variants, &sttcp->tgsi);
 234
 235       read_stream_out_from_cache(&blob_reader, &sttcp->tgsi);
 236
 237       if (nir) {
 238          sttcp->tgsi.type = PIPE_SHADER_IR_NIR;
 239          sttcp->shader_program = shProg;
 240          sttcp->tgsi.ir.nir = nir_deserialize(NULL, options, &blob_reader);
 241          prog->nir = sttcp->tgsi.ir.nir;
 242       } else {
 243          read_tgsi_from_cache(&blob_reader, &sttcp->tgsi.tokens,
 244                               &sttcp->num_tgsi_tokens);
 245       }
 246
 247       if (st->tcp == sttcp)
 248          st->dirty |= sttcp->affected_states;
 249
 250       break;
 251    }
 252    case MESA_SHADER_TESS_EVAL: {
 253       struct st_common_program *sttep = st_common_program(prog);
 254
 255       st_release_basic_variants(st, sttep->Base.Target,
 256                                 &sttep->variants, &sttep->tgsi);
 257
 258       read_stream_out_from_cache(&blob_reader, &sttep->tgsi);
 259
 260       if (nir) {
 261          sttep->tgsi.type = PIPE_SHADER_IR_NIR;
 262          sttep->shader_program = shProg;
 263          sttep->tgsi.ir.nir = nir_deserialize(NULL, options, &blob_reader);
 264          prog->nir = sttep->tgsi.ir.nir;
 265       } else {
 266          read_tgsi_from_cache(&blob_reader, &sttep->tgsi.tokens,
 267                               &sttep->num_tgsi_tokens);
 268       }
 269
 270       if (st->tep == sttep)
 271          st->dirty |= sttep->affected_states;
 272
 273       break;
 274    }
 275    case MESA_SHADER_GEOMETRY: {
 276       struct st_common_program *stgp = st_common_program(prog);
 277
 278       st_release_basic_variants(st, stgp->Base.Target, &stgp->variants,
 279                                 &stgp->tgsi);
 280
 281       read_stream_out_from_cache(&blob_reader, &stgp->tgsi);
 282
 283       if (nir) {
 284          stgp->tgsi.type = PIPE_SHADER_IR_NIR;
 285          stgp->shader_program = shProg;
 286          stgp->tgsi.ir.nir = nir_deserialize(NULL, options, &blob_reader);
 287          prog->nir = stgp->tgsi.ir.nir;
 288       } else {
 289          read_tgsi_from_cache(&blob_reader, &stgp->tgsi.tokens,
 290                               &stgp->num_tgsi_tokens);
 291       }
 292
 293       if (st->gp == stgp)
 294          st->dirty |= stgp->affected_states;
 295
 296       break;
 297    }
 298    case MESA_SHADER_FRAGMENT: {
 299       struct st_fragment_program *stfp = (struct st_fragment_program *) prog;
 300
 301       st_release_fp_variants(st, stfp);
 302
 303       if (nir) {
 304          stfp->tgsi.type = PIPE_SHADER_IR_NIR;
 305          stfp->shader_program = shProg;
 306          stfp->tgsi.ir.nir = nir_deserialize(NULL, options, &blob_reader);
 307          prog->nir = stfp->tgsi.ir.nir;
 308       } else {
 309          read_tgsi_from_cache(&blob_reader, &stfp->tgsi.tokens,
 310                               &stfp->num_tgsi_tokens);
 311       }
 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 = (struct st_compute_program *) prog;
 320
 321       st_release_cp_variants(st, stcp);
 322
 323       if (nir) {
 324          stcp->tgsi.ir_type = PIPE_SHADER_IR_NIR;
 325          stcp->shader_program = shProg;
 326          stcp->tgsi.prog = nir_deserialize(NULL, options, &blob_reader);
 327          prog->nir = (nir_shader *) stcp->tgsi.prog;
 328       } else {
 329          read_tgsi_from_cache(&blob_reader,
 330                               (const struct tgsi_token**) &stcp->tgsi.prog,
 331                               &stcp->num_tgsi_tokens);
 332       }
 333
 334       stcp->tgsi.req_local_mem = stcp->Base.info.cs.shared_size;
 335       stcp->tgsi.req_private_mem = 0;
 336       stcp->tgsi.req_input_mem = 0;
 337
 338       if (st->cp == stcp)
 339          st->dirty |= stcp->affected_states;
 340
 341       break;
 342    }
 343    default:
 344       unreachable("Unsupported stage");
 345    }
 346
 347    /* Make sure we don't try to read more data than we wrote. This should
 348     * never happen in release builds but its useful to have this check to
 349     * catch development bugs.
 350     */
 351    if (blob_reader.current != blob_reader.end || blob_reader.overrun) {
 352       assert(!"Invalid TGSI shader disk cache item!");
 353
 354       if (ctx->_Shader->Flags & GLSL_CACHE_INFO) {
 355          fprintf(stderr, "Error reading program from cache (invalid "
 356                  "TGSI cache item)\n");
 357       }
 358    }
 359
 360    st_set_prog_affected_state_flags(prog);
 361    _mesa_associate_uniform_storage(ctx, shProg, prog, false);
 362
 363    /* Create Gallium shaders now instead of on demand. */
 364    if (ST_DEBUG & DEBUG_PRECOMPILE ||
 365        st->shader_has_one_variant[prog->info.stage])
 366       st_precompile_shader_variant(st, prog);
 367 }
 368
 369 bool
 370 st_load_ir_from_disk_cache(struct gl_context *ctx,
 371                            struct gl_shader_program *prog,
 372                            bool nir)
 373 {
 374    if (!ctx->Cache)
 375       return false;
 376
 377    /* If we didn't load the GLSL metadata from cache then we could not have
 378     * loaded the tgsi either.
 379     */
 380    if (prog->data->LinkStatus != LINKING_SKIPPED)
 381       return false;
 382
 383    for (unsigned i = 0; i < MESA_SHADER_STAGES; i++) {
 384       if (prog->_LinkedShaders[i] == NULL)
 385          continue;
 386
 387       struct gl_program *glprog = prog->_LinkedShaders[i]->Program;
 388       st_deserialise_ir_program(ctx, prog, glprog, nir);
 389
 390       /* We don't need the cached blob anymore so free it */
 391       ralloc_free(glprog->driver_cache_blob);
 392       glprog->driver_cache_blob = NULL;
 393       glprog->driver_cache_blob_size = 0;
 394
 395       if (ctx->_Shader->Flags & GLSL_CACHE_INFO) {
 396          fprintf(stderr, "%s state tracker IR retrieved from cache\n",
 397                  _mesa_shader_stage_to_string(i));
 398       }
 399    }
 400
 401    return true;
 402 }
 403
 404 void
 405 st_serialise_tgsi_program(struct gl_context *ctx, struct gl_program *prog)
 406 {
 407    st_serialise_ir_program(ctx, prog, false);
 408 }
 409
 410 void
 411 st_deserialise_tgsi_program(struct gl_context *ctx,
 412                             struct gl_shader_program *shProg,
 413                             struct gl_program *prog)
 414 {
 415    st_deserialise_ir_program(ctx, shProg, prog, false);
 416 }
 417
 418 void
 419 st_serialise_nir_program(struct gl_context *ctx, struct gl_program *prog)
 420 {
 421    st_serialise_ir_program(ctx, prog, true);
 422 }
 423
 424 void
 425 st_deserialise_nir_program(struct gl_context *ctx,
 426                            struct gl_shader_program *shProg,
 427                            struct gl_program *prog)
 428 {
 429    st_deserialise_ir_program(ctx, shProg, prog, true);
 430 }