[Mesa-dev] [PATCH 2/4] st/mesa: implement a tgsi on-disk shader cache

Timothy Arceri tarceri at itsqueeze.com
Tue Feb 21 09:17:56 UTC 2017



On 21/02/17 19:57, Nicolai Hähnle wrote:
> On 20.02.2017 01:15, Timothy Arceri wrote:
>> Implements a tgsi cache for the OpenGL state tracker.
>>
>> V2: add support for compute shaders
>
> I don't think the comments I sent earlier were addressed:
> https://lists.freedesktop.org/archives/mesa-dev/2017-February/144675.html
>

Apologies I'm reading those for the first time now. Somehow missed them, 
will address.


> Thanks,
> Nicolai
>
>> ---
>>  src/mesa/state_tracker/st_glsl_to_tgsi.cpp | 222
>> +++++++++++++++++++++++++++++
>>  src/mesa/state_tracker/st_program.c        | 133 ++++++++++++++++-
>>  2 files changed, 350 insertions(+), 5 deletions(-)
>>
>> diff --git a/src/mesa/state_tracker/st_glsl_to_tgsi.cpp
>> b/src/mesa/state_tracker/st_glsl_to_tgsi.cpp
>> index 630f5af..b485776 100644
>> --- a/src/mesa/state_tracker/st_glsl_to_tgsi.cpp
>> +++ b/src/mesa/state_tracker/st_glsl_to_tgsi.cpp
>> @@ -32,6 +32,7 @@
>>
>>  #include "st_glsl_to_tgsi.h"
>>
>> +#include "compiler/glsl/blob.h"
>>  #include "compiler/glsl/glsl_parser_extras.h"
>>  #include "compiler/glsl/ir_optimization.h"
>>  #include "compiler/glsl/program.h"
>> @@ -47,6 +48,8 @@
>>  #include "pipe/p_screen.h"
>>  #include "tgsi/tgsi_ureg.h"
>>  #include "tgsi/tgsi_info.h"
>> +#include "util/disk_cache.h"
>> +#include "util/mesa-sha1.h"
>>  #include "util/u_math.h"
>>  #include "util/u_memory.h"
>>  #include "st_program.h"
>> @@ -6999,6 +7002,219 @@ has_unsupported_control_flow(exec_list *ir,
>>     return visitor.unsupported;
>>  }
>>
>> +static void
>> +read_stream_out_from_cache(struct blob_reader *blob_reader,
>> +                           struct pipe_shader_state *tgsi)
>> +{
>> +   blob_copy_bytes(blob_reader, (uint8_t *) &tgsi->stream_output,
>> +                    sizeof(tgsi->stream_output));
>> +}
>> +
>> +static void
>> +read_tgsi_from_cache(struct blob_reader *blob_reader,
>> +                     const struct tgsi_token **tokens)
>> +{
>> +   uint32_t num_tokens  = blob_read_uint32(blob_reader);
>> +   unsigned tokens_size = num_tokens * sizeof(struct tgsi_token);
>> +   *tokens = (const tgsi_token*) MALLOC(tokens_size);
>> +   blob_copy_bytes(blob_reader, (uint8_t *) *tokens, tokens_size);
>> +}
>> +
>> +static void
>> +load_tgsi_from_disk_cache(struct gl_context *ctx,
>> +                          struct gl_shader_program *prog)
>> +{
>> +   unsigned char sha1[20];
>> +   char sha1_buf[41];
>> +   struct st_context *st = st_context(ctx);
>> +
>> +   for (unsigned i = 0; i < MESA_SHADER_STAGES; i++) {
>> +      if (prog->_LinkedShaders[i] == NULL)
>> +         continue;
>> +
>> +      char *buf = ralloc_strdup(NULL, "tsgi_tokens ");
>> +      _mesa_sha1_format(sha1_buf,
>> +
>> prog->_LinkedShaders[i]->Program->sh.data->sha1);
>> +      ralloc_strcat(&buf, sha1_buf);
>> +
>> +      struct gl_program *glprog = prog->_LinkedShaders[i]->Program;
>> +      switch (glprog->info.stage) {
>> +      case MESA_SHADER_VERTEX:
>> +         ralloc_strcat(&buf, " vs");
>> +         _mesa_sha1_compute(buf, strlen(buf), sha1);
>> +         break;
>> +      case MESA_SHADER_TESS_EVAL:
>> +         ralloc_strcat(&buf, " tes");
>> +         _mesa_sha1_compute(buf, strlen(buf), sha1);
>> +         break;
>> +      case MESA_SHADER_TESS_CTRL:
>> +         ralloc_strcat(&buf, " tcs");
>> +         _mesa_sha1_compute(buf, strlen(buf), sha1);
>> +         break;
>> +      case MESA_SHADER_GEOMETRY:
>> +         ralloc_strcat(&buf, " gs");
>> +         _mesa_sha1_compute(buf, strlen(buf), sha1);
>> +         break;
>> +      case MESA_SHADER_FRAGMENT:
>> +         ralloc_strcat(&buf, " fs");
>> +         _mesa_sha1_compute(buf, strlen(buf), sha1);
>> +         break;
>> +      case MESA_SHADER_COMPUTE:
>> +         ralloc_strcat(&buf, " cs");
>> +         _mesa_sha1_compute(buf, strlen(buf), sha1);
>> +         break;
>> +
>> +      default:
>> +         unreachable("Unsupported stage");
>> +      }
>> +
>> +      size_t size;
>> +      uint8_t *buffer = (uint8_t *) disk_cache_get(ctx->Cache, sha1,
>> &size);
>> +      if (buffer) {
>> +         struct blob_reader blob_reader;
>> +         blob_reader_init(&blob_reader, buffer, size);
>> +
>> +         switch (glprog->info.stage) {
>> +         case MESA_SHADER_VERTEX: {
>> +            struct st_vertex_program *stvp =
>> +               (struct st_vertex_program *) glprog;
>> +
>> +            st_release_vp_variants(st, stvp);
>> +
>> +            stvp->num_inputs = blob_read_uint32(&blob_reader);
>> +            blob_copy_bytes(&blob_reader, (uint8_t *)
>> stvp->index_to_input,
>> +                            sizeof(stvp->index_to_input));
>> +            blob_copy_bytes(&blob_reader, (uint8_t *)
>> stvp->result_to_output,
>> +                            sizeof(stvp->result_to_output));
>> +
>> +            read_stream_out_from_cache(&blob_reader, &stvp->tgsi);
>> +            read_tgsi_from_cache(&blob_reader, &stvp->tgsi.tokens);
>> +
>> +            if (st->vp == stvp)
>> +               st->dirty |= ST_NEW_VERTEX_PROGRAM(st, stvp);
>> +
>> +            break;
>> +         }
>> +         case MESA_SHADER_TESS_EVAL: {
>> +            struct st_tesseval_program *sttep =
>> +               (struct st_tesseval_program *) glprog;
>> +
>> +            st_release_basic_variants(st, sttep->Base.Target,
>> +                                      &sttep->variants, &sttep->tgsi);
>> +
>> +            read_stream_out_from_cache(&blob_reader, &sttep->tgsi);
>> +            read_tgsi_from_cache(&blob_reader, &sttep->tgsi.tokens);
>> +
>> +            if (st->tep == sttep)
>> +               st->dirty |= sttep->affected_states;
>> +
>> +            break;
>> +         }
>> +         case MESA_SHADER_TESS_CTRL: {
>> +            struct st_tessctrl_program *sttcp =
>> +               (struct st_tessctrl_program *) glprog;
>> +
>> +            st_release_basic_variants(st, sttcp->Base.Target,
>> +                                      &sttcp->variants, &sttcp->tgsi);
>> +
>> +            read_stream_out_from_cache(&blob_reader, &sttcp->tgsi);
>> +            read_tgsi_from_cache(&blob_reader, &sttcp->tgsi.tokens);
>> +
>> +            if (st->tcp == sttcp)
>> +               st->dirty |= sttcp->affected_states;
>> +
>> +            break;
>> +         }
>> +         case MESA_SHADER_GEOMETRY: {
>> +            struct st_geometry_program *stgp =
>> +               (struct st_geometry_program *) glprog;
>> +
>> +            st_release_basic_variants(st, stgp->Base.Target,
>> &stgp->variants,
>> +                                      &stgp->tgsi);
>> +
>> +            read_stream_out_from_cache(&blob_reader, &stgp->tgsi);
>> +            read_tgsi_from_cache(&blob_reader, &stgp->tgsi.tokens);
>> +
>> +            if (st->gp == stgp)
>> +               st->dirty |= stgp->affected_states;
>> +
>> +            break;
>> +         }
>> +         case MESA_SHADER_FRAGMENT: {
>> +            struct st_fragment_program *stfp =
>> +               (struct st_fragment_program *) glprog;
>> +
>> +            st_release_fp_variants(st, stfp);
>> +
>> +            read_tgsi_from_cache(&blob_reader, &stfp->tgsi.tokens);
>> +
>> +            if (st->fp == stfp)
>> +               st->dirty |= stfp->affected_states;
>> +
>> +            break;
>> +         }
>> +         case MESA_SHADER_COMPUTE: {
>> +            struct st_compute_program *stcp =
>> +               (struct st_compute_program *) glprog;
>> +
>> +            st_release_cp_variants(st, stcp);
>> +
>> +            read_tgsi_from_cache(&blob_reader,
>> +                                 (const tgsi_token**) &stcp->tgsi.prog);
>> +
>> +            stcp->tgsi.req_local_mem = stcp->Base.info.cs.shared_size;
>> +            stcp->tgsi.req_private_mem = 0;
>> +            stcp->tgsi.req_input_mem = 0;
>> +
>> +            if (st->cp == stcp)
>> +                st->dirty |= stcp->affected_states;
>> +
>> +            break;
>> +         }
>> +         default:
>> +            unreachable("Unsupported stage");
>> +         }
>> +
>> +         if (blob_reader.current != blob_reader.end ||
>> blob_reader.overrun) {
>> +            /* Something very bad has gone wrong discard the item
>> from the
>> +             * cache and rebuild/link from source.
>> +             */
>> +            assert(!"Invalid TGSI shader disk cache item!");
>> +
>> +            if (ctx->_Shader->Flags & GLSL_CACHE_INFO) {
>> +               fprintf(stderr, "Error reading program from cache
>> (invalid "
>> +                       "TGSI cache item)\n");
>> +            }
>> +
>> +            disk_cache_remove(ctx->Cache, sha1);
>> +            for (unsigned i = 0; i < prog->NumShaders; i++) {
>> +               _mesa_glsl_compile_shader(ctx, prog->Shaders[i],
>> false, false,
>> +                                         true);
>> +            }
>> +
>> +            prog->data->cache_fallback = true;
>> +            _mesa_glsl_link_shader(ctx, prog);
>> +
>> +            free(buf);
>> +            return;
>> +         }
>> +
>> +         if (ctx->_Shader->Flags & GLSL_CACHE_INFO) {
>> +            _mesa_sha1_format(sha1_buf, sha1);
>> +            fprintf(stderr, "%s tgsi_tokens retrieved from cache: %s\n",
>> +                    _mesa_shader_stage_to_string(i), sha1_buf);
>> +         }
>> +
>> +         set_prog_affected_state_flags(glprog);
>> +         _mesa_associate_uniform_storage(ctx, prog, glprog->Parameters);
>> +
>> +         free(buffer);
>> +      }
>> +
>> +      ralloc_free(buf);
>> +   }
>> +}
>> +
>>  extern "C" {
>>
>>  /**
>> @@ -7010,6 +7226,12 @@ extern "C" {
>>  GLboolean
>>  st_link_shader(struct gl_context *ctx, struct gl_shader_program *prog)
>>  {
>> +   /* Return early if we are loading the shader from on-disk cache */
>> +   if (ctx->Cache && prog->data->LinkStatus == linking_skipped) {
>> +      load_tgsi_from_disk_cache(ctx, prog);
>> +      return GL_TRUE;
>> +   }
>> +
>>     struct pipe_screen *pscreen = ctx->st->pipe->screen;
>>     assert(prog->data->LinkStatus);
>>
>> diff --git a/src/mesa/state_tracker/st_program.c
>> b/src/mesa/state_tracker/st_program.c
>> index b2d1586..c85055e 100644
>> --- a/src/mesa/state_tracker/st_program.c
>> +++ b/src/mesa/state_tracker/st_program.c
>> @@ -37,7 +37,10 @@
>>  #include "program/prog_parameter.h"
>>  #include "program/prog_print.h"
>>  #include "program/programopt.h"
>> +#include "util/disk_cache.h"
>> +#include "util/mesa-sha1.h"
>>
>> +#include "compiler/glsl/blob.h"
>>  #include "compiler/nir/nir.h"
>>
>>  #include "pipe/p_context.h"
>> @@ -223,6 +226,111 @@ st_release_cp_variants(struct st_context *st,
>> struct st_compute_program *stcp)
>>     }
>>  }
>>
>> +static void
>> +write_stream_out_to_cache(struct blob *blob,
>> +                          struct pipe_shader_state *tgsi)
>> +{
>> +   blob_write_bytes(blob, &tgsi->stream_output,
>> +                    sizeof(tgsi->stream_output));
>> +}
>> +
>> +static void
>> +write_tgsi_to_cache(struct blob *blob, struct pipe_shader_state *tgsi,
>> +                    struct st_context *st, unsigned char *sha1,
>> +                    unsigned num_tokens)
>> +{
>> +   blob_write_uint32(blob, num_tokens);
>> +   blob_write_bytes(blob, tgsi->tokens,
>> +                    num_tokens * sizeof(struct tgsi_token));
>> +
>> +   disk_cache_put(st->ctx->Cache, sha1, blob->data, blob->size);
>> +}
>> +
>> +/**
>> + * Store tgsi and any other required state in on-disk shader cache.
>> + */
>> +static void
>> +cache_tgsi(struct st_context *st, struct gl_program *prog,
>> +           struct pipe_shader_state *out_state, unsigned num_tokens)
>> +{
>> +   if (!st->ctx->Cache)
>> +      return;
>> +
>> +   /* Exit early when we are dealing with a ff shader with no source
>> file to
>> +    * generate a source from.
>> +    */
>> +   if (*prog->sh.data->sha1 == 0)
>> +      return;
>> +
>> +   unsigned char sha1[20];
>> +   char sha1_buf[41];
>> +   char *buf = ralloc_strdup(NULL, "tsgi_tokens ");
>> +   struct blob *blob = blob_create(NULL);
>> +
>> +   _mesa_sha1_format(sha1_buf, prog->sh.data->sha1);
>> +   ralloc_strcat(&buf, sha1_buf);
>> +
>> +   switch (prog->info.stage) {
>> +   case MESA_SHADER_VERTEX: {
>> +      ralloc_strcat(&buf, " vs");
>> +     _mesa_sha1_compute(buf, strlen(buf), sha1);
>> +
>> +      struct st_vertex_program *stvp = (struct st_vertex_program *)
>> prog;
>> +
>> +      blob_write_uint32(blob, stvp->num_inputs);
>> +      blob_write_bytes(blob, stvp->index_to_input,
>> +                       sizeof(stvp->index_to_input));
>> +      blob_write_bytes(blob, stvp->result_to_output,
>> +                       sizeof(stvp->result_to_output));
>> +
>> +      write_stream_out_to_cache(blob, &stvp->tgsi);
>> +      write_tgsi_to_cache(blob, &stvp->tgsi, st, sha1, num_tokens);
>> +      break;
>> +   }
>> +   case MESA_SHADER_TESS_EVAL:
>> +      ralloc_strcat(&buf, " tes");
>> +      _mesa_sha1_compute(buf, strlen(buf), sha1);
>> +      write_stream_out_to_cache(blob, out_state);
>> +      write_tgsi_to_cache(blob, out_state, st, sha1, num_tokens);
>> +      break;
>> +   case MESA_SHADER_TESS_CTRL:
>> +      ralloc_strcat(&buf, " tcs");
>> +      _mesa_sha1_compute(buf, strlen(buf), sha1);
>> +      write_stream_out_to_cache(blob, out_state);
>> +      write_tgsi_to_cache(blob, out_state, st, sha1, num_tokens);
>> +      break;
>> +   case MESA_SHADER_GEOMETRY:
>> +      ralloc_strcat(&buf, " gs");
>> +      _mesa_sha1_compute(buf, strlen(buf), sha1);
>> +      write_stream_out_to_cache(blob, out_state);
>> +      write_tgsi_to_cache(blob, out_state, st, sha1, num_tokens);
>> +      break;
>> +   case MESA_SHADER_FRAGMENT: {
>> +      ralloc_strcat(&buf, " fs");
>> +      _mesa_sha1_compute(buf, strlen(buf), sha1);
>> +
>> +      struct st_fragment_program *stfp = (struct st_fragment_program
>> *) prog;
>> +      write_tgsi_to_cache(blob, &stfp->tgsi, st, sha1, num_tokens);
>> +      break;
>> +   }
>> +   case MESA_SHADER_COMPUTE:
>> +      ralloc_strcat(&buf, " cs");
>> +      _mesa_sha1_compute(buf, strlen(buf), sha1);
>> +      write_tgsi_to_cache(blob, out_state, st, sha1, num_tokens);
>> +      break;
>> +   default:
>> +      unreachable("Unsupported stage");
>> +   }
>> +
>> +   if (st->ctx->_Shader->Flags & GLSL_CACHE_INFO) {
>> +      _mesa_sha1_format(sha1_buf, sha1);
>> +      fprintf(stderr, "putting %s tgsi_tokens in cache: %s\n",
>> +              _mesa_shader_stage_to_string(prog->info.stage), sha1_buf);
>> +   }
>> +
>> +   ralloc_free(blob);
>> +   ralloc_free(buf);
>> +}
>>
>>  /**
>>   * Translate a vertex program.
>> @@ -442,7 +550,6 @@ st_translate_vertex_program(struct st_context *st,
>>                                        &stvp->tgsi.stream_output);
>>
>>        free_glsl_to_tgsi_visitor(stvp->glsl_to_tgsi);
>> -      stvp->glsl_to_tgsi = NULL;
>>     } else
>>        error = st_translate_mesa_program(st->ctx,
>>                                          PIPE_SHADER_VERTEX,
>> @@ -467,8 +574,15 @@ st_translate_vertex_program(struct st_context *st,
>>        return false;
>>     }
>>
>> -   stvp->tgsi.tokens = ureg_get_tokens(ureg, NULL);
>> +   unsigned num_tokens;
>> +   stvp->tgsi.tokens = ureg_get_tokens(ureg, &num_tokens);
>>     ureg_destroy(ureg);
>> +
>> +   if (stvp->glsl_to_tgsi) {
>> +      stvp->glsl_to_tgsi = NULL;
>> +      cache_tgsi(st, &stvp->Base, NULL, num_tokens);
>> +   }
>> +
>>     return stvp->tgsi.tokens != NULL;
>>  }
>>
>> @@ -890,7 +1004,6 @@ st_translate_fragment_program(struct st_context *st,
>>                             fs_output_semantic_index);
>>
>>        free_glsl_to_tgsi_visitor(stfp->glsl_to_tgsi);
>> -      stfp->glsl_to_tgsi = NULL;
>>     } else if (stfp->ati_fs)
>>        st_translate_atifs_program(ureg,
>>                                   stfp->ati_fs,
>> @@ -923,8 +1036,15 @@ st_translate_fragment_program(struct st_context
>> *st,
>>                                  fs_output_semantic_name,
>>                                  fs_output_semantic_index);
>>
>> -   stfp->tgsi.tokens = ureg_get_tokens(ureg, NULL);
>> +   unsigned num_tokens;
>> +   stfp->tgsi.tokens = ureg_get_tokens(ureg, &num_tokens);
>>     ureg_destroy(ureg);
>> +
>> +   if (stfp->glsl_to_tgsi) {
>> +      stfp->glsl_to_tgsi = NULL;
>> +      cache_tgsi(st, &stfp->Base, NULL, num_tokens);
>> +   }
>> +
>>     return stfp->tgsi.tokens != NULL;
>>  }
>>
>> @@ -1459,13 +1579,16 @@ st_translate_program_common(struct st_context
>> *st,
>>                          output_semantic_name,
>>                          output_semantic_index);
>>
>> -   out_state->tokens = ureg_get_tokens(ureg, NULL);
>> +   unsigned num_tokens;
>> +   out_state->tokens = ureg_get_tokens(ureg, &num_tokens);
>>     ureg_destroy(ureg);
>>
>>     st_translate_stream_output_info(glsl_to_tgsi,
>>                                     outputMapping,
>>                                     &out_state->stream_output);
>>
>> +   cache_tgsi(st, prog, out_state, num_tokens);
>> +
>>     if ((ST_DEBUG & DEBUG_TGSI) && (ST_DEBUG & DEBUG_MESA)) {
>>        _mesa_print_program(prog);
>>        debug_printf("\n");
>>
>


More information about the mesa-dev mailing list