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 "st_util.h"
  29 #include "compiler/glsl/program.h"
  30 #include "compiler/nir/nir.h"
  31 #include "compiler/nir/nir_serialize.h"
  32 #include "pipe/p_shader_tokens.h"
  33 #include "program/ir_to_mesa.h"
  34 #include "util/u_memory.h"
  35
  36 void
  37 st_get_program_binary_driver_sha1(struct gl_context *ctx, uint8_t *sha1)
  38 {
  39    disk_cache_compute_key(ctx->Cache, NULL, 0, sha1);
  40 }
  41
  42 static void
  43 write_stream_out_to_cache(struct blob *blob,
  44                           struct pipe_shader_state *tgsi)
  45 {
  46    blob_write_bytes(blob, &tgsi->stream_output,
  47                     sizeof(tgsi->stream_output));
  48 }
  49
  50 static void
  51 copy_blob_to_driver_cache_blob(struct blob *blob, struct gl_program *prog)
  52 {
  53    prog->driver_cache_blob = ralloc_size(NULL, blob->size);
  54    memcpy(prog->driver_cache_blob, blob->data, blob->size);
  55    prog->driver_cache_blob_size = blob->size;
  56 }
  57
  58 static void
  59 write_tgsi_to_cache(struct blob *blob, const struct tgsi_token *tokens,
  60                     struct gl_program *prog, unsigned num_tokens)
  61 {
  62    blob_write_uint32(blob, num_tokens);
  63    blob_write_bytes(blob, tokens, num_tokens * sizeof(struct tgsi_token));
  64    copy_blob_to_driver_cache_blob(blob, prog);
  65 }
  66
  67 static void
  68 write_nir_to_cache(struct blob *blob, struct gl_program *prog)
  69 {
  70    nir_serialize(blob, prog->nir, false);
  71    copy_blob_to_driver_cache_blob(blob, prog);
  72 }
  73
  74 static void
  75 st_serialise_ir_program(struct gl_context *ctx, struct gl_program *prog,
  76                         bool nir)
  77 {
  78    if (prog->driver_cache_blob)
  79       return;
  80
  81    struct blob blob;
  82    blob_init(&blob);
  83
  84    switch (prog->info.stage) {
  85    case MESA_SHADER_VERTEX: {
  86       struct st_vertex_program *stvp = (struct st_vertex_program *) prog;
  87
  88       blob_write_uint32(&blob, stvp->num_inputs);
  89       blob_write_bytes(&blob, stvp->index_to_input,
  90                        sizeof(stvp->index_to_input));
  91       blob_write_bytes(&blob, stvp->input_to_index,
  92                        sizeof(stvp->input_to_index));
  93       blob_write_bytes(&blob, stvp->result_to_output,
  94                        sizeof(stvp->result_to_output));
  95
  96       write_stream_out_to_cache(&blob, &stvp->tgsi);
  97
  98       if (nir)
  99          write_nir_to_cache(&blob, prog);
 100       else
 101          write_tgsi_to_cache(&blob, stvp->tgsi.tokens, prog,
 102                              stvp->num_tgsi_tokens);
 103       break;
 104    }
 105    case MESA_SHADER_TESS_CTRL:
 106    case MESA_SHADER_TESS_EVAL:
 107    case MESA_SHADER_GEOMETRY: {
 108       struct st_common_program *stcp = (struct st_common_program *) prog;
 109
 110       if (prog->info.stage == MESA_SHADER_TESS_EVAL ||
 111           prog->info.stage == MESA_SHADER_GEOMETRY)
 112          write_stream_out_to_cache(&blob, &stcp->tgsi);
 113
 114       if (nir)
 115          write_nir_to_cache(&blob, prog);
 116       else
 117          write_tgsi_to_cache(&blob, stcp->tgsi.tokens, prog,
 118                              stcp->num_tgsi_tokens);
 119       break;
 120    }
 121    case MESA_SHADER_FRAGMENT: {
 122       struct st_fragment_program *stfp = (struct st_fragment_program *) prog;
 123
 124       if (nir)
 125          write_nir_to_cache(&blob, prog);
 126       else
 127          write_tgsi_to_cache(&blob, stfp->tgsi.tokens, prog,
 128                              stfp->num_tgsi_tokens);
 129       break;
 130    }
 131    case MESA_SHADER_COMPUTE: {
 132       struct st_compute_program *stcp = (struct st_compute_program *) prog;
 133
 134       if (nir)
 135          write_nir_to_cache(&blob, prog);
 136       else
 137          write_tgsi_to_cache(&blob, stcp->tgsi.prog, prog,
 138                              stcp->num_tgsi_tokens);
 139       break;
 140    }
 141    default:
 142       unreachable("Unsupported stage");
 143    }
 144
 145    blob_finish(&blob);
 146 }
 147
 148 /**
 149  * Store tgsi and any other required state in on-disk shader cache.
 150  */
 151 void
 152 st_store_ir_in_disk_cache(struct st_context *st, struct gl_program *prog,
 153                           bool nir)
 154 {
 155    if (!st->ctx->Cache)
 156       return;
 157
 158    /* Exit early when we are dealing with a ff shader with no source file to
 159     * generate a source from.
 160     */
 161    static const char zero[sizeof(prog->sh.data->sha1)] = {0};
 162    if (memcmp(prog->sh.data->sha1, zero, sizeof(prog->sh.data->sha1)) == 0)
 163       return;
 164
 165    st_serialise_ir_program(st->ctx, prog, nir);
 166
 167    if (st->ctx->_Shader->Flags & GLSL_CACHE_INFO) {
 168       fprintf(stderr, "putting %s state tracker IR in cache\n",
 169               _mesa_shader_stage_to_string(prog->info.stage));
 170    }
 171 }
 172
 173 static void
 174 read_stream_out_from_cache(struct blob_reader *blob_reader,
 175                            struct pipe_shader_state *tgsi)
 176 {
 177    blob_copy_bytes(blob_reader, (uint8_t *) &tgsi->stream_output,
 178                     sizeof(tgsi->stream_output));
 179 }
 180
 181 static void
 182 read_tgsi_from_cache(struct blob_reader *blob_reader,
 183                      const struct tgsi_token **tokens,
 184                      unsigned *num_tokens)
 185 {
 186    *num_tokens  = blob_read_uint32(blob_reader);
 187    unsigned tokens_size = *num_tokens * sizeof(struct tgsi_token);
 188    *tokens = (const struct tgsi_token*) MALLOC(tokens_size);
 189    blob_copy_bytes(blob_reader, (uint8_t *) *tokens, tokens_size);
 190 }
 191
 192 static void
 193 st_deserialise_ir_program(struct gl_context *ctx,
 194                           struct gl_shader_program *shProg,
 195                           struct gl_program *prog, bool nir)
 196 {
 197    struct st_context *st = st_context(ctx);
 198    size_t size = prog->driver_cache_blob_size;
 199    uint8_t *buffer = (uint8_t *) prog->driver_cache_blob;
 200    const struct nir_shader_compiler_options *options =
 201       ctx->Const.ShaderCompilerOptions[prog->info.stage].NirOptions;
 202
 203    assert(prog->driver_cache_blob && prog->driver_cache_blob_size > 0);
 204
 205    struct blob_reader blob_reader;
 206    blob_reader_init(&blob_reader, buffer, size);
 207
 208    switch (prog->info.stage) {
 209    case MESA_SHADER_VERTEX: {
 210       struct st_vertex_program *stvp = (struct st_vertex_program *) prog;
 211
 212       st_release_vp_variants(st, stvp);
 213
 214       stvp->num_inputs = blob_read_uint32(&blob_reader);
 215       blob_copy_bytes(&blob_reader, (uint8_t *) stvp->index_to_input,
 216                       sizeof(stvp->index_to_input));
 217       blob_copy_bytes(&blob_reader, (uint8_t *) stvp->input_to_index,
 218                       sizeof(stvp->input_to_index));
 219       blob_copy_bytes(&blob_reader, (uint8_t *) stvp->result_to_output,
 220                       sizeof(stvp->result_to_output));
 221
 222       read_stream_out_from_cache(&blob_reader, &stvp->tgsi);
 223
 224       if (nir) {
 225          stvp->tgsi.type = PIPE_SHADER_IR_NIR;
 226          stvp->shader_program = shProg;
 227          stvp->tgsi.ir.nir = nir_deserialize(NULL, options, &blob_reader);
 228          prog->nir = stvp->tgsi.ir.nir;
 229       } else {
 230          read_tgsi_from_cache(&blob_reader, &stvp->tgsi.tokens,
 231                               &stvp->num_tgsi_tokens);
 232       }
 233
 234       if (st->vp == stvp)
 235          st->dirty |= ST_NEW_VERTEX_PROGRAM(st, stvp);
 236
 237       break;
 238    }
 239    case MESA_SHADER_TESS_CTRL: {
 240       struct st_common_program *sttcp = st_common_program(prog);
 241
 242       st_release_basic_variants(st, sttcp);
 243
 244       if (nir) {
 245          sttcp->tgsi.type = PIPE_SHADER_IR_NIR;
 246          sttcp->shader_program = shProg;
 247          sttcp->tgsi.ir.nir = nir_deserialize(NULL, options, &blob_reader);
 248          prog->nir = sttcp->tgsi.ir.nir;
 249       } else {
 250          read_tgsi_from_cache(&blob_reader, &sttcp->tgsi.tokens,
 251                               &sttcp->num_tgsi_tokens);
 252       }
 253
 254       if (st->tcp == sttcp)
 255          st->dirty |= sttcp->affected_states;
 256
 257       break;
 258    }
 259    case MESA_SHADER_TESS_EVAL: {
 260       struct st_common_program *sttep = st_common_program(prog);
 261
 262       st_release_basic_variants(st, sttep);
 263       read_stream_out_from_cache(&blob_reader, &sttep->tgsi);
 264
 265       if (nir) {
 266          sttep->tgsi.type = PIPE_SHADER_IR_NIR;
 267          sttep->shader_program = shProg;
 268          sttep->tgsi.ir.nir = nir_deserialize(NULL, options, &blob_reader);
 269          prog->nir = sttep->tgsi.ir.nir;
 270       } else {
 271          read_tgsi_from_cache(&blob_reader, &sttep->tgsi.tokens,
 272                               &sttep->num_tgsi_tokens);
 273       }
 274
 275       if (st->tep == sttep)
 276          st->dirty |= sttep->affected_states;
 277
 278       break;
 279    }
 280    case MESA_SHADER_GEOMETRY: {
 281       struct st_common_program *stgp = st_common_program(prog);
 282
 283       st_release_basic_variants(st, stgp);
 284       read_stream_out_from_cache(&blob_reader, &stgp->tgsi);
 285
 286       if (nir) {
 287          stgp->tgsi.type = PIPE_SHADER_IR_NIR;
 288          stgp->shader_program = shProg;
 289          stgp->tgsi.ir.nir = nir_deserialize(NULL, options, &blob_reader);
 290          prog->nir = stgp->tgsi.ir.nir;
 291       } else {
 292          read_tgsi_from_cache(&blob_reader, &stgp->tgsi.tokens,
 293                               &stgp->num_tgsi_tokens);
 294       }
 295
 296       if (st->gp == stgp)
 297          st->dirty |= stgp->affected_states;
 298
 299       break;
 300    }
 301    case MESA_SHADER_FRAGMENT: {
 302       struct st_fragment_program *stfp = (struct st_fragment_program *) prog;
 303
 304       st_release_fp_variants(st, stfp);
 305
 306       if (nir) {
 307          stfp->tgsi.type = PIPE_SHADER_IR_NIR;
 308          stfp->shader_program = shProg;
 309          stfp->tgsi.ir.nir = nir_deserialize(NULL, options, &blob_reader);
 310          prog->nir = stfp->tgsi.ir.nir;
 311       } else {
 312          read_tgsi_from_cache(&blob_reader, &stfp->tgsi.tokens,
 313                               &stfp->num_tgsi_tokens);
 314       }
 315
 316       if (st->fp == stfp)
 317          st->dirty |= stfp->affected_states;
 318
 319       break;
 320    }
 321    case MESA_SHADER_COMPUTE: {
 322       struct st_compute_program *stcp = (struct st_compute_program *) prog;
 323
 324       st_release_cp_variants(st, stcp);
 325
 326       if (nir) {
 327          stcp->tgsi.ir_type = PIPE_SHADER_IR_NIR;
 328          stcp->shader_program = shProg;
 329          stcp->tgsi.prog = nir_deserialize(NULL, options, &blob_reader);
 330          prog->nir = (nir_shader *) stcp->tgsi.prog;
 331       } else {
 332          read_tgsi_from_cache(&blob_reader,
 333                               (const struct tgsi_token**) &stcp->tgsi.prog,
 334                               &stcp->num_tgsi_tokens);
 335       }
 336
 337       stcp->tgsi.req_local_mem = stcp->Base.info.cs.shared_size;
 338       stcp->tgsi.req_private_mem = 0;
 339       stcp->tgsi.req_input_mem = 0;
 340
 341       if (st->cp == stcp)
 342          st->dirty |= stcp->affected_states;
 343
 344       break;
 345    }
 346    default:
 347       unreachable("Unsupported stage");
 348    }
 349
 350    /* Make sure we don't try to read more data than we wrote. This should
 351     * never happen in release builds but its useful to have this check to
 352     * catch development bugs.
 353     */
 354    if (blob_reader.current != blob_reader.end || blob_reader.overrun) {
 355       assert(!"Invalid TGSI shader disk cache item!");
 356
 357       if (ctx->_Shader->Flags & GLSL_CACHE_INFO) {
 358          fprintf(stderr, "Error reading program from cache (invalid "
 359                  "TGSI cache item)\n");
 360       }
 361    }
 362
 363    st_set_prog_affected_state_flags(prog);
 364    _mesa_associate_uniform_storage(ctx, shProg, prog);
 365
 366    /* Create Gallium shaders now instead of on demand. */
 367    if (ST_DEBUG & DEBUG_PRECOMPILE ||
 368        st->shader_has_one_variant[prog->info.stage])
 369       st_precompile_shader_variant(st, prog);
 370 }
 371
 372 bool
 373 st_load_ir_from_disk_cache(struct gl_context *ctx,
 374                            struct gl_shader_program *prog,
 375                            bool nir)
 376 {
 377    if (!ctx->Cache)
 378       return false;
 379
 380    /* If we didn't load the GLSL metadata from cache then we could not have
 381     * loaded the tgsi either.
 382     */
 383    if (prog->data->LinkStatus != LINKING_SKIPPED)
 384       return false;
 385
 386    for (unsigned i = 0; i < MESA_SHADER_STAGES; i++) {
 387       if (prog->_LinkedShaders[i] == NULL)
 388          continue;
 389
 390       struct gl_program *glprog = prog->_LinkedShaders[i]->Program;
 391       st_deserialise_ir_program(ctx, prog, glprog, nir);
 392
 393       /* We don't need the cached blob anymore so free it */
 394       ralloc_free(glprog->driver_cache_blob);
 395       glprog->driver_cache_blob = NULL;
 396       glprog->driver_cache_blob_size = 0;
 397
 398       if (ctx->_Shader->Flags & GLSL_CACHE_INFO) {
 399          fprintf(stderr, "%s state tracker IR retrieved from cache\n",
 400                  _mesa_shader_stage_to_string(i));
 401       }
 402    }
 403
 404    return true;
 405 }
 406
 407 void
 408 st_serialise_tgsi_program(struct gl_context *ctx, struct gl_program *prog)
 409 {
 410    st_serialise_ir_program(ctx, prog, false);
 411 }
 412
 413 void
 414 st_serialise_tgsi_program_binary(struct gl_context *ctx,
 415                                  struct gl_shader_program *shProg,
 416                                  struct gl_program *prog)
 417 {
 418    st_serialise_ir_program(ctx, prog, false);
 419 }
 420
 421 void
 422 st_deserialise_tgsi_program(struct gl_context *ctx,
 423                             struct gl_shader_program *shProg,
 424                             struct gl_program *prog)
 425 {
 426    st_deserialise_ir_program(ctx, shProg, prog, false);
 427 }
 428
 429 void
 430 st_serialise_nir_program(struct gl_context *ctx, struct gl_program *prog)
 431 {
 432    st_serialise_ir_program(ctx, prog, true);
 433 }
 434
 435 void
 436 st_serialise_nir_program_binary(struct gl_context *ctx,
 437                                 struct gl_shader_program *shProg,
 438                                 struct gl_program *prog)
 439 {
 440    st_serialise_ir_program(ctx, prog, true);
 441 }
 442
 443 void
 444 st_deserialise_nir_program(struct gl_context *ctx,
 445                            struct gl_shader_program *shProg,
 446                            struct gl_program *prog)
 447 {
 448    st_deserialise_ir_program(ctx, shProg, prog, true);
 449 }