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