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 *state)
  46 {
  47    blob_write_bytes(blob, &state->stream_output,
  48                     sizeof(state->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_FRAGMENT:
 111    case MESA_SHADER_COMPUTE: {
 112       struct st_common_program *stcp = (struct st_common_program *) prog;
 113
 114       if (prog->info.stage == MESA_SHADER_TESS_EVAL ||
 115           prog->info.stage == MESA_SHADER_GEOMETRY)
 116          write_stream_out_to_cache(&blob, &stcp->state);
 117
 118       if (nir)
 119          write_nir_to_cache(&blob, prog);
 120       else
 121          write_tgsi_to_cache(&blob, stcp->state.tokens, prog);
 122       break;
 123    }
 124    default:
 125       unreachable("Unsupported stage");
 126    }
 127
 128    blob_finish(&blob);
 129 }
 130
 131 /**
 132  * Store TGSI or NIR and any other required state in on-disk shader cache.
 133  */
 134 void
 135 st_store_ir_in_disk_cache(struct st_context *st, struct gl_program *prog,
 136                           bool nir)
 137 {
 138    if (!st->ctx->Cache)
 139       return;
 140
 141    /* Exit early when we are dealing with a ff shader with no source file to
 142     * generate a source from.
 143     */
 144    static const char zero[sizeof(prog->sh.data->sha1)] = {0};
 145    if (memcmp(prog->sh.data->sha1, zero, sizeof(prog->sh.data->sha1)) == 0)
 146       return;
 147
 148    st_serialise_ir_program(st->ctx, prog, nir);
 149
 150    if (st->ctx->_Shader->Flags & GLSL_CACHE_INFO) {
 151       fprintf(stderr, "putting %s state tracker IR in cache\n",
 152               _mesa_shader_stage_to_string(prog->info.stage));
 153    }
 154 }
 155
 156 static void
 157 read_stream_out_from_cache(struct blob_reader *blob_reader,
 158                            struct pipe_shader_state *state)
 159 {
 160    blob_copy_bytes(blob_reader, (uint8_t *) &state->stream_output,
 161                     sizeof(state->stream_output));
 162 }
 163
 164 static void
 165 read_tgsi_from_cache(struct blob_reader *blob_reader,
 166                      const struct tgsi_token **tokens)
 167 {
 168    unsigned num_tokens  = blob_read_uint32(blob_reader);
 169    unsigned tokens_size = num_tokens * sizeof(struct tgsi_token);
 170    *tokens = (const struct tgsi_token*) MALLOC(tokens_size);
 171    blob_copy_bytes(blob_reader, (uint8_t *) *tokens, tokens_size);
 172 }
 173
 174 static void
 175 st_deserialise_ir_program(struct gl_context *ctx,
 176                           struct gl_shader_program *shProg,
 177                           struct gl_program *prog, bool nir)
 178 {
 179    struct st_context *st = st_context(ctx);
 180    size_t size = prog->driver_cache_blob_size;
 181    uint8_t *buffer = (uint8_t *) prog->driver_cache_blob;
 182    const struct nir_shader_compiler_options *options =
 183       ctx->Const.ShaderCompilerOptions[prog->info.stage].NirOptions;
 184
 185    st_set_prog_affected_state_flags(prog);
 186    _mesa_associate_uniform_storage(ctx, shProg, prog);
 187
 188    assert(prog->driver_cache_blob && prog->driver_cache_blob_size > 0);
 189
 190    struct blob_reader blob_reader;
 191    blob_reader_init(&blob_reader, buffer, size);
 192
 193    switch (prog->info.stage) {
 194    case MESA_SHADER_VERTEX: {
 195       struct st_vertex_program *stvp = (struct st_vertex_program *) prog;
 196
 197       st_release_vp_variants(st, stvp);
 198
 199       stvp->num_inputs = blob_read_uint32(&blob_reader);
 200       blob_copy_bytes(&blob_reader, (uint8_t *) stvp->index_to_input,
 201                       sizeof(stvp->index_to_input));
 202       blob_copy_bytes(&blob_reader, (uint8_t *) stvp->input_to_index,
 203                       sizeof(stvp->input_to_index));
 204       blob_copy_bytes(&blob_reader, (uint8_t *) stvp->result_to_output,
 205                       sizeof(stvp->result_to_output));
 206
 207       read_stream_out_from_cache(&blob_reader, &stvp->state);
 208
 209       if (nir) {
 210          stvp->state.type = PIPE_SHADER_IR_NIR;
 211          stvp->shader_program = shProg;
 212          stvp->state.ir.nir = nir_deserialize(NULL, options, &blob_reader);
 213          prog->nir = stvp->state.ir.nir;
 214       } else {
 215          read_tgsi_from_cache(&blob_reader, &stvp->state.tokens);
 216       }
 217       break;
 218    }
 219    case MESA_SHADER_TESS_CTRL:
 220    case MESA_SHADER_TESS_EVAL:
 221    case MESA_SHADER_GEOMETRY:
 222    case MESA_SHADER_FRAGMENT:
 223    case MESA_SHADER_COMPUTE: {
 224       struct st_common_program *stcp = st_common_program(prog);
 225
 226       if (prog->info.stage == MESA_SHADER_FRAGMENT)
 227          st_release_fp_variants(st, stcp);
 228       else
 229          st_release_common_variants(st, stcp);
 230
 231       if (prog->info.stage == MESA_SHADER_TESS_EVAL ||
 232           prog->info.stage == MESA_SHADER_GEOMETRY)
 233          read_stream_out_from_cache(&blob_reader, &stcp->state);
 234
 235       if (nir) {
 236          stcp->state.type = PIPE_SHADER_IR_NIR;
 237          stcp->state.ir.nir = nir_deserialize(NULL, options, &blob_reader);
 238          stcp->shader_program = shProg;
 239          prog->nir = stcp->state.ir.nir;
 240       } else {
 241          read_tgsi_from_cache(&blob_reader, &stcp->state.tokens);
 242       }
 243       break;
 244    }
 245    default:
 246       unreachable("Unsupported stage");
 247    }
 248
 249    /* Make sure we don't try to read more data than we wrote. This should
 250     * never happen in release builds but its useful to have this check to
 251     * catch development bugs.
 252     */
 253    if (blob_reader.current != blob_reader.end || blob_reader.overrun) {
 254       assert(!"Invalid TGSI shader disk cache item!");
 255
 256       if (ctx->_Shader->Flags & GLSL_CACHE_INFO) {
 257          fprintf(stderr, "Error reading program from cache (invalid "
 258                  "TGSI cache item)\n");
 259       }
 260    }
 261
 262    st_finalize_program(st, prog);
 263 }
 264
 265 bool
 266 st_load_ir_from_disk_cache(struct gl_context *ctx,
 267                            struct gl_shader_program *prog,
 268                            bool nir)
 269 {
 270    if (!ctx->Cache)
 271       return false;
 272
 273    /* If we didn't load the GLSL metadata from cache then we could not have
 274     * loaded TGSI or NIR either.
 275     */
 276    if (prog->data->LinkStatus != LINKING_SKIPPED)
 277       return false;
 278
 279    for (unsigned i = 0; i < MESA_SHADER_STAGES; i++) {
 280       if (prog->_LinkedShaders[i] == NULL)
 281          continue;
 282
 283       struct gl_program *glprog = prog->_LinkedShaders[i]->Program;
 284       st_deserialise_ir_program(ctx, prog, glprog, nir);
 285
 286       /* We don't need the cached blob anymore so free it */
 287       ralloc_free(glprog->driver_cache_blob);
 288       glprog->driver_cache_blob = NULL;
 289       glprog->driver_cache_blob_size = 0;
 290
 291       if (ctx->_Shader->Flags & GLSL_CACHE_INFO) {
 292          fprintf(stderr, "%s state tracker IR retrieved from cache\n",
 293                  _mesa_shader_stage_to_string(i));
 294       }
 295    }
 296
 297    return true;
 298 }
 299
 300 void
 301 st_serialise_tgsi_program(struct gl_context *ctx, struct gl_program *prog)
 302 {
 303    st_serialise_ir_program(ctx, prog, false);
 304 }
 305
 306 void
 307 st_serialise_tgsi_program_binary(struct gl_context *ctx,
 308                                  struct gl_shader_program *shProg,
 309                                  struct gl_program *prog)
 310 {
 311    st_serialise_ir_program(ctx, prog, false);
 312 }
 313
 314 void
 315 st_deserialise_tgsi_program(struct gl_context *ctx,
 316                             struct gl_shader_program *shProg,
 317                             struct gl_program *prog)
 318 {
 319    st_deserialise_ir_program(ctx, shProg, prog, false);
 320 }
 321
 322 void
 323 st_serialise_nir_program(struct gl_context *ctx, struct gl_program *prog)
 324 {
 325    st_serialise_ir_program(ctx, prog, true);
 326 }
 327
 328 void
 329 st_serialise_nir_program_binary(struct gl_context *ctx,
 330                                 struct gl_shader_program *shProg,
 331                                 struct gl_program *prog)
 332 {
 333    st_serialise_ir_program(ctx, prog, true);
 334 }
 335
 336 void
 337 st_deserialise_nir_program(struct gl_context *ctx,
 338                            struct gl_shader_program *shProg,
 339                            struct gl_program *prog)
 340 {
 341    st_deserialise_ir_program(ctx, shProg, prog, true);
 342 }