st/mesa: implement a tgsi on-disk shader cache

Implements a tgsi cache for the OpenGL state tracker.

V2: add support for compute shaders

Reviewed-by: Nicolai Hähnle <[email protected]>

5 files changed, 481 insertions, 6 deletions

diff --git a/src/mesa/Makefile.sources b/src/mesa/Makefile.sources
index ae701138f5d..17f84fd0b61 100644
--- a/src/mesa/Makefile.sources
+++ b/src/mesa/Makefile.sources
@@ -516,6 +516,8 @@ STATETRACKER_FILES = \
 	state_tracker/st_sampler_view.h \
 	state_tracker/st_scissor.c \
 	state_tracker/st_scissor.h \
+	state_tracker/st_shader_cache.c \
+	state_tracker/st_shader_cache.h \
 	state_tracker/st_texture.c \
 	state_tracker/st_texture.h \
 	state_tracker/st_tgsi_lower_yuv.c \
diff --git a/src/mesa/state_tracker/st_glsl_to_tgsi.cpp b/src/mesa/state_tracker/st_glsl_to_tgsi.cpp
index 476d18534d9..d43d821aad0 100644
--- a/src/mesa/state_tracker/st_glsl_to_tgsi.cpp
+++ b/src/mesa/state_tracker/st_glsl_to_tgsi.cpp
@@ -54,6 +54,7 @@
 #include "st_format.h"
 #include "st_glsl_types.h"
 #include "st_nir.h"
+#include "st_shader_cache.h"
 
 #include <algorithm>
 
@@ -6870,6 +6871,11 @@ 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 (st_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 3795f25a74e..4d9250b73cc 100644
--- a/src/mesa/state_tracker/st_program.c
+++ b/src/mesa/state_tracker/st_program.c
@@ -58,6 +58,7 @@
 #include "st_mesa_to_tgsi.h"
 #include "st_atifs_to_tgsi.h"
 #include "st_nir.h"
+#include "st_shader_cache.h"
 #include "cso_cache/cso_context.h"
 
 
@@ -364,7 +365,6 @@ st_release_cp_variants(struct st_context *st, struct st_compute_program *stcp)
    }
 }
 
-
 /**
  * Translate a vertex program.
  */
@@ -583,7 +583,6 @@ 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,
@@ -608,8 +607,15 @@ 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;
+      st_store_tgsi_in_disk_cache(st, &stvp->Base, NULL, num_tokens);
+   }
+
    return stvp->tgsi.tokens != NULL;
 }
 
@@ -1031,7 +1037,6 @@ 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,
@@ -1064,8 +1069,15 @@ 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;
+      st_store_tgsi_in_disk_cache(st, &stfp->Base, NULL, num_tokens);
+   }
+
    return stfp->tgsi.tokens != NULL;
 }
 
@@ -1600,13 +1612,16 @@ 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);
 
+   st_store_tgsi_in_disk_cache(st, prog, out_state, num_tokens);
+
    if ((ST_DEBUG & DEBUG_TGSI) && (ST_DEBUG & DEBUG_MESA)) {
       _mesa_print_program(prog);
       debug_printf("\n");
diff --git a/src/mesa/state_tracker/st_shader_cache.c b/src/mesa/state_tracker/st_shader_cache.c
new file mode 100644
index 00000000000..607e5b10766
--- /dev/null
+++ b/src/mesa/state_tracker/st_shader_cache.c
@@ -0,0 +1,406 @@
+/*
+ * Copyright © 2017 Timothy Arceri
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a
+ * copy of this software and associated documentation files (the "Software"),
+ * to deal in the Software without restriction, including without limitation
+ * the rights to use, copy, modify, merge, publish, distribute, sublicense,
+ * and/or sell copies of the Software, and to permit persons to whom the
+ * Software is furnished to do so, subject to the following conditions:
+ *
+ * The above copyright notice and this permission notice (including the next
+ * paragraph) shall be included in all copies or substantial portions of the
+ * Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+ * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+ * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
+ * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+ * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
+ * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER
+ * DEALINGS IN THE SOFTWARE.
+ */
+
+#include <stdio.h>
+
+#include "st_program.h"
+#include "st_shader_cache.h"
+#include "compiler/glsl/program.h"
+#include "pipe/p_shader_tokens.h"
+#include "program/ir_to_mesa.h"
+#include "util/u_memory.h"
+
+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.
+ */
+void
+st_store_tgsi_in_disk_cache(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;
+   struct blob *blob = blob_create(NULL);
+
+   switch (prog->info.stage) {
+   case MESA_SHADER_VERTEX: {
+      struct st_vertex_program *stvp = (struct st_vertex_program *) prog;
+      sha1 = stvp->sha1;
+
+      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_CTRL: {
+      struct st_tessctrl_program *stcp = (struct st_tessctrl_program *) prog;
+      sha1 = stcp->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_EVAL: {
+      struct st_tesseval_program *step = (struct st_tesseval_program *) prog;
+      sha1 = step->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: {
+      struct st_geometry_program *stgp = (struct st_geometry_program *) prog;
+      sha1 = stgp->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: {
+      struct st_fragment_program *stfp = (struct st_fragment_program *) prog;
+      sha1 = stfp->sha1;
+
+      write_tgsi_to_cache(blob, &stfp->tgsi, st, sha1, num_tokens);
+      break;
+   }
+   case MESA_SHADER_COMPUTE: {
+      struct st_compute_program *stcp = (struct st_compute_program *) prog;
+      sha1 = stcp->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) {
+      char sha1_buf[41];
+      _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);
+}
+
+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 struct tgsi_token*) MALLOC(tokens_size);
+   blob_copy_bytes(blob_reader, (uint8_t *) *tokens, tokens_size);
+}
+
+bool
+st_load_tgsi_from_disk_cache(struct gl_context *ctx,
+                             struct gl_shader_program *prog)
+{
+   if (!ctx->Cache)
+      return false;
+
+   unsigned char *stage_sha1[MESA_SHADER_STAGES];
+   char sha1_buf[41];
+
+   /* Compute and store sha1 for each stage. These will be reused by the
+    * cache store pass if we fail to find the cached tgsi.
+    */
+   for (unsigned i = 0; i < MESA_SHADER_STAGES; i++) {
+      if (prog->_LinkedShaders[i] == NULL)
+         continue;
+
+      char *buf = ralloc_strdup(NULL, "tgsi_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: {
+         struct st_vertex_program *stvp = (struct st_vertex_program *) glprog;
+         stage_sha1[i] = stvp->sha1;
+         ralloc_strcat(&buf, " vs");
+         _mesa_sha1_compute(buf, strlen(buf), stage_sha1[i]);
+         break;
+      }
+      case MESA_SHADER_TESS_CTRL: {
+         struct st_tessctrl_program *stcp =
+            (struct st_tessctrl_program *) glprog;
+         stage_sha1[i] = stcp->sha1;
+         ralloc_strcat(&buf, " tcs");
+         _mesa_sha1_compute(buf, strlen(buf), stage_sha1[i]);
+         break;
+      }
+      case MESA_SHADER_TESS_EVAL: {
+         struct st_tesseval_program *step =
+            (struct st_tesseval_program *) glprog;
+         stage_sha1[i] = step->sha1;
+         ralloc_strcat(&buf, " tes");
+         _mesa_sha1_compute(buf, strlen(buf), stage_sha1[i]);
+         break;
+      }
+      case MESA_SHADER_GEOMETRY: {
+         struct st_geometry_program *stgp =
+            (struct st_geometry_program *) glprog;
+         stage_sha1[i] = stgp->sha1;
+         ralloc_strcat(&buf, " gs");
+         _mesa_sha1_compute(buf, strlen(buf), stage_sha1[i]);
+         break;
+      }
+      case MESA_SHADER_FRAGMENT: {
+         struct st_fragment_program *stfp =
+            (struct st_fragment_program *) glprog;
+         stage_sha1[i] = stfp->sha1;
+         ralloc_strcat(&buf, " fs");
+         _mesa_sha1_compute(buf, strlen(buf), stage_sha1[i]);
+         break;
+      }
+      case MESA_SHADER_COMPUTE: {
+         struct st_compute_program *stcp =
+            (struct st_compute_program *) glprog;
+         stage_sha1[i] = stcp->sha1;
+         ralloc_strcat(&buf, " cs");
+         _mesa_sha1_compute(buf, strlen(buf), stage_sha1[i]);
+         break;
+      }
+      default:
+         unreachable("Unsupported stage");
+      }
+
+      ralloc_free(buf);
+   }
+
+   struct st_context *st = st_context(ctx);
+   for (unsigned i = 0; i < MESA_SHADER_STAGES; i++) {
+      if (prog->_LinkedShaders[i] == NULL)
+         continue;
+
+      unsigned char *sha1 = stage_sha1[i];
+      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);
+
+         struct gl_program *glprog = prog->_LinkedShaders[i]->Program;
+         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_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_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_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 struct 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);
+
+            goto fallback_recompile;
+         }
+
+         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);
+         }
+
+         st_set_prog_affected_state_flags(glprog);
+         _mesa_associate_uniform_storage(ctx, prog, glprog->Parameters);
+
+         free(buffer);
+      } else {
+         /* Failed to find a matching cached shader so fallback to recompile.
+          */
+         if (ctx->_Shader->Flags & GLSL_CACHE_INFO) {
+            fprintf(stderr, "TGSI cache item not found falling back to "
+                    "compile.\n");
+         }
+
+         goto fallback_recompile;
+      }
+   }
+
+   return true;
+
+fallback_recompile:
+
+   /* GLSL IR was compiled and linked so just fallback to the regular
+    * glsl to tgsi path.
+    */
+   if (prog->data->LinkStatus != linking_skipped)
+      return false;
+
+   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);
+
+   return true;
+}
diff --git a/src/mesa/state_tracker/st_shader_cache.h b/src/mesa/state_tracker/st_shader_cache.h
new file mode 100644
index 00000000000..f9e46158dc0
--- /dev/null
+++ b/src/mesa/state_tracker/st_shader_cache.h
@@ -0,0 +1,46 @@
+/*
+ * Copyright © 2017 Timothy Arceri
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a
+ * copy of this software and associated documentation files (the "Software"),
+ * to deal in the Software without restriction, including without limitation
+ * the rights to use, copy, modify, merge, publish, distribute, sublicense,
+ * and/or sell copies of the Software, and to permit persons to whom the
+ * Software is furnished to do so, subject to the following conditions:
+ *
+ * The above copyright notice and this permission notice (including the next
+ * paragraph) shall be included in all copies or substantial portions of the
+ * Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+ * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+ * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
+ * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+ * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
+ * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER
+ * DEALINGS IN THE SOFTWARE.
+ */
+
+#include "st_context.h"
+#include "compiler/glsl/blob.h"
+#include "main/mtypes.h"
+#include "pipe/p_state.h"
+#include "util/disk_cache.h"
+#include "util/mesa-sha1.h"
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+bool
+st_load_tgsi_from_disk_cache(struct gl_context *ctx,
+                             struct gl_shader_program *prog);
+
+void
+st_store_tgsi_in_disk_cache(struct st_context *st, struct gl_program *prog,
+                            struct pipe_shader_state *out_state,
+                            unsigned num_tokens);
+
+#ifdef __cplusplus
+}
+#endif