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    case MESA_SHADER_COMPUTE: {
 109       struct st_common_program *stcp = (struct st_common_program *) prog;
 110
 111       if (prog->info.stage == MESA_SHADER_TESS_EVAL ||
 112           prog->info.stage == MESA_SHADER_GEOMETRY)
 113          write_stream_out_to_cache(&blob, &stcp->tgsi);
 114
 115       if (nir)
 116          write_nir_to_cache(&blob, prog);
 117       else
 118          write_tgsi_to_cache(&blob, stcp->tgsi.tokens, prog,
 119                              stcp->num_tgsi_tokens);
 120       break;
 121    }
 122    case MESA_SHADER_FRAGMENT: {
 123       struct st_fragment_program *stfp = (struct st_fragment_program *) prog;
 124
 125       if (nir)
 126          write_nir_to_cache(&blob, prog);
 127       else
 128          write_tgsi_to_cache(&blob, stfp->tgsi.tokens, prog,
 129                              stfp->num_tgsi_tokens);
 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                      unsigned *num_tokens)
 176 {
 177    *num_tokens  = blob_read_uint32(blob_reader);
 178    unsigned tokens_size = *num_tokens * sizeof(struct tgsi_token);
 179    *tokens = (const struct tgsi_token*) MALLOC(tokens_size);
 180    blob_copy_bytes(blob_reader, (uint8_t *) *tokens, tokens_size);
 181 }
 182
 183 static void
 184 st_deserialise_ir_program(struct gl_context *ctx,
 185                           struct gl_shader_program *shProg,
 186                           struct gl_program *prog, bool nir)
 187 {
 188    struct st_context *st = st_context(ctx);
 189    size_t size = prog->driver_cache_blob_size;
 190    uint8_t *buffer = (uint8_t *) prog->driver_cache_blob;
 191    const struct nir_shader_compiler_options *options =
 192       ctx->Const.ShaderCompilerOptions[prog->info.stage].NirOptions;
 193
 194    assert(prog->driver_cache_blob && prog->driver_cache_blob_size > 0);
 195
 196    struct blob_reader blob_reader;
 197    blob_reader_init(&blob_reader, buffer, size);
 198
 199    switch (prog->info.stage) {
 200    case MESA_SHADER_VERTEX: {
 201       struct st_vertex_program *stvp = (struct st_vertex_program *) prog;
 202
 203       st_release_vp_variants(st, stvp);
 204
 205       stvp->num_inputs = blob_read_uint32(&blob_reader);
 206       blob_copy_bytes(&blob_reader, (uint8_t *) stvp->index_to_input,
 207                       sizeof(stvp->index_to_input));
 208       blob_copy_bytes(&blob_reader, (uint8_t *) stvp->input_to_index,
 209                       sizeof(stvp->input_to_index));
 210       blob_copy_bytes(&blob_reader, (uint8_t *) stvp->result_to_output,
 211                       sizeof(stvp->result_to_output));
 212
 213       read_stream_out_from_cache(&blob_reader, &stvp->tgsi);
 214
 215       if (nir) {
 216          stvp->tgsi.type = PIPE_SHADER_IR_NIR;
 217          stvp->shader_program = shProg;
 218          stvp->tgsi.ir.nir = nir_deserialize(NULL, options, &blob_reader);
 219          prog->nir = stvp->tgsi.ir.nir;
 220       } else {
 221          read_tgsi_from_cache(&blob_reader, &stvp->tgsi.tokens,
 222                               &stvp->num_tgsi_tokens);
 223       }
 224
 225       if (st->vp == stvp)
 226          st->dirty |= ST_NEW_VERTEX_PROGRAM(st, stvp);
 227
 228       break;
 229    }
 230    case MESA_SHADER_TESS_CTRL:
 231    case MESA_SHADER_TESS_EVAL:
 232    case MESA_SHADER_GEOMETRY:
 233    case MESA_SHADER_COMPUTE: {
 234       struct st_common_program *stcp = st_common_program(prog);
 235
 236       st_release_basic_variants(st, stcp);
 237
 238       if (prog->info.stage == MESA_SHADER_TESS_EVAL ||
 239           prog->info.stage == MESA_SHADER_GEOMETRY)
 240          read_stream_out_from_cache(&blob_reader, &stcp->tgsi);
 241
 242       if (nir) {
 243          stcp->tgsi.type = PIPE_SHADER_IR_NIR;
 244          stcp->tgsi.ir.nir = nir_deserialize(NULL, options, &blob_reader);
 245          stcp->shader_program = shProg;
 246          prog->nir = stcp->tgsi.ir.nir;
 247       } else {
 248          read_tgsi_from_cache(&blob_reader, &stcp->tgsi.tokens,
 249                               &stcp->num_tgsi_tokens);
 250       }
 251
 252       if ((prog->info.stage == MESA_SHADER_TESS_CTRL && st->tcp == stcp) ||
 253           (prog->info.stage == MESA_SHADER_TESS_EVAL && st->tep == stcp) ||
 254           (prog->info.stage == MESA_SHADER_GEOMETRY && st->gp == stcp) ||
 255           (prog->info.stage == MESA_SHADER_COMPUTE && st->cp == stcp))
 256          st->dirty |= stcp->affected_states;
 257       break;
 258    }
 259    case MESA_SHADER_FRAGMENT: {
 260       struct st_fragment_program *stfp = (struct st_fragment_program *) prog;
 261
 262       st_release_fp_variants(st, stfp);
 263
 264       if (nir) {
 265          stfp->tgsi.type = PIPE_SHADER_IR_NIR;
 266          stfp->shader_program = shProg;
 267          stfp->tgsi.ir.nir = nir_deserialize(NULL, options, &blob_reader);
 268          prog->nir = stfp->tgsi.ir.nir;
 269       } else {
 270          read_tgsi_from_cache(&blob_reader, &stfp->tgsi.tokens,
 271                               &stfp->num_tgsi_tokens);
 272       }
 273
 274       if (st->fp == stfp)
 275          st->dirty |= stfp->affected_states;
 276
 277       break;
 278    }
 279    default:
 280       unreachable("Unsupported stage");
 281    }
 282
 283    /* Make sure we don't try to read more data than we wrote. This should
 284     * never happen in release builds but its useful to have this check to
 285     * catch development bugs.
 286     */
 287    if (blob_reader.current != blob_reader.end || blob_reader.overrun) {
 288       assert(!"Invalid TGSI shader disk cache item!");
 289
 290       if (ctx->_Shader->Flags & GLSL_CACHE_INFO) {
 291          fprintf(stderr, "Error reading program from cache (invalid "
 292                  "TGSI cache item)\n");
 293       }
 294    }
 295
 296    st_set_prog_affected_state_flags(prog);
 297    _mesa_associate_uniform_storage(ctx, shProg, prog);
 298
 299    /* Create Gallium shaders now instead of on demand. */
 300    if (ST_DEBUG & DEBUG_PRECOMPILE ||
 301        st->shader_has_one_variant[prog->info.stage])
 302       st_precompile_shader_variant(st, prog);
 303 }
 304
 305 bool
 306 st_load_ir_from_disk_cache(struct gl_context *ctx,
 307                            struct gl_shader_program *prog,
 308                            bool nir)
 309 {
 310    if (!ctx->Cache)
 311       return false;
 312
 313    /* If we didn't load the GLSL metadata from cache then we could not have
 314     * loaded the tgsi either.
 315     */
 316    if (prog->data->LinkStatus != LINKING_SKIPPED)
 317       return false;
 318
 319    for (unsigned i = 0; i < MESA_SHADER_STAGES; i++) {
 320       if (prog->_LinkedShaders[i] == NULL)
 321          continue;
 322
 323       struct gl_program *glprog = prog->_LinkedShaders[i]->Program;
 324       st_deserialise_ir_program(ctx, prog, glprog, nir);
 325
 326       /* We don't need the cached blob anymore so free it */
 327       ralloc_free(glprog->driver_cache_blob);
 328       glprog->driver_cache_blob = NULL;
 329       glprog->driver_cache_blob_size = 0;
 330
 331       if (ctx->_Shader->Flags & GLSL_CACHE_INFO) {
 332          fprintf(stderr, "%s state tracker IR retrieved from cache\n",
 333                  _mesa_shader_stage_to_string(i));
 334       }
 335    }
 336
 337    return true;
 338 }
 339
 340 void
 341 st_serialise_tgsi_program(struct gl_context *ctx, struct gl_program *prog)
 342 {
 343    st_serialise_ir_program(ctx, prog, false);
 344 }
 345
 346 void
 347 st_serialise_tgsi_program_binary(struct gl_context *ctx,
 348                                  struct gl_shader_program *shProg,
 349                                  struct gl_program *prog)
 350 {
 351    st_serialise_ir_program(ctx, prog, false);
 352 }
 353
 354 void
 355 st_deserialise_tgsi_program(struct gl_context *ctx,
 356                             struct gl_shader_program *shProg,
 357                             struct gl_program *prog)
 358 {
 359    st_deserialise_ir_program(ctx, shProg, prog, false);
 360 }
 361
 362 void
 363 st_serialise_nir_program(struct gl_context *ctx, struct gl_program *prog)
 364 {
 365    st_serialise_ir_program(ctx, prog, true);
 366 }
 367
 368 void
 369 st_serialise_nir_program_binary(struct gl_context *ctx,
 370                                 struct gl_shader_program *shProg,
 371                                 struct gl_program *prog)
 372 {
 373    st_serialise_ir_program(ctx, prog, true);
 374 }
 375
 376 void
 377 st_deserialise_nir_program(struct gl_context *ctx,
 378                            struct gl_shader_program *shProg,
 379                            struct gl_program *prog)
 380 {
 381    st_deserialise_ir_program(ctx, shProg, prog, true);
 382 }