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