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