[Mesa-dev] [PATCH 11/77] i965: add initial implementation of on disk shader cache

Timothy Arceri timothy.arceri at collabora.com
Mon Oct 3 06:04:30 UTC 2016


This uses the recently-added cache.c to write out the final linked
binary for vertex and fragment shader programs.

This is based off the initial implementation done by Carl.
---
 src/mesa/drivers/dri/i965/Makefile.sources   |   1 +
 src/mesa/drivers/dri/i965/brw_shader_cache.c | 375 +++++++++++++++++++++++++++
 src/mesa/drivers/dri/i965/brw_state.h        |   7 +
 3 files changed, 383 insertions(+)
 create mode 100644 src/mesa/drivers/dri/i965/brw_shader_cache.c

diff --git a/src/mesa/drivers/dri/i965/Makefile.sources b/src/mesa/drivers/dri/i965/Makefile.sources
index df90cb4..bd2bd37 100644
--- a/src/mesa/drivers/dri/i965/Makefile.sources
+++ b/src/mesa/drivers/dri/i965/Makefile.sources
@@ -147,6 +147,7 @@ i965_FILES = \
 	brw_sf_emit.c \
 	brw_sf.h \
 	brw_sf_state.c \
+	brw_shader_cache.cpp \
 	brw_state_batch.c \
 	brw_state_cache.c \
 	brw_state_dump.c \
diff --git a/src/mesa/drivers/dri/i965/brw_shader_cache.c b/src/mesa/drivers/dri/i965/brw_shader_cache.c
new file mode 100644
index 0000000..61e51a4
--- /dev/null
+++ b/src/mesa/drivers/dri/i965/brw_shader_cache.c
@@ -0,0 +1,375 @@
+/*
+ * Copyright © 2014 Intel Corporation
+ *
+ * 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 <compiler/glsl/blob.h>
+#include <compiler/glsl/cache.h>
+#include <compiler/glsl/ir_uniform.h>
+#include <compiler/glsl/shader_cache.h>
+#include <util/macros.h>
+#include <util/mesa-sha1.h>
+#include <main/mtypes.h>
+
+#include "brw_context.h"
+#include "brw_state.h"
+#include "brw_vs.h"
+#include "brw_wm.h"
+
+static size_t
+key_size(gl_shader_stage stage)
+{
+   switch (stage) {
+   case MESA_SHADER_VERTEX:
+      return sizeof(struct brw_vs_prog_key);
+   case MESA_SHADER_TESS_CTRL:
+      return sizeof(struct brw_tcs_prog_key);
+   case MESA_SHADER_TESS_EVAL:
+      return sizeof(struct brw_tes_prog_key);
+   case MESA_SHADER_GEOMETRY:
+      return sizeof(struct brw_gs_prog_key);
+   case MESA_SHADER_FRAGMENT:
+      return sizeof(struct brw_wm_prog_key);
+   case MESA_SHADER_COMPUTE:
+      return sizeof(struct brw_cs_prog_key);
+   default:
+      unreachable("Unsupported stage!");
+   }
+}
+
+static void
+gen_shader_sha1(struct brw_context *brw, struct gl_shader_program *prog,
+                gl_shader_stage stage, void *key, unsigned char *out_sha1)
+{
+   char sha1_buf[41];
+   unsigned char sha1[20];
+   char manifest[256];
+   int offset = 0;
+
+   offset += snprintf(manifest, sizeof(manifest), "program: %s\n",
+                      _mesa_sha1_format(sha1_buf, prog->sha1));
+
+   _mesa_sha1_compute(key, key_size(stage), sha1);
+   offset += snprintf(manifest + offset, sizeof(manifest) - offset,
+                      "%s_key: %s\n",
+                      _mesa_shader_stage_to_abbrev(stage),
+                      _mesa_sha1_format(sha1_buf, sha1));
+
+   _mesa_sha1_compute(manifest, strlen(manifest), out_sha1);
+}
+
+static void
+load_program_data(struct gl_shader_program *prog, struct blob_reader *binary,
+                  struct brw_stage_prog_data *prog_data,
+                  gl_shader_stage stage, struct gl_context *ctx)
+{
+   static const gl_constant_value zero = { 0 };
+
+   uint64_t parameter_values_base = blob_read_uint64(binary);
+   uint64_t uniform_data_slots_base = blob_read_uint64(binary);
+
+   uint32_t nr_params = blob_read_uint32(binary);
+   assert(nr_params == prog_data->nr_params);
+
+   prog_data->param = rzalloc_array(NULL, const gl_constant_value *,
+                                    nr_params);
+   if (ctx->_Shader->Flags & GLSL_CACHE_INFO) {
+      fprintf(stderr, "Allocating %d prog_data->params (%p)\n",
+              prog_data->nr_params, prog_data->param);
+   }
+
+   for (unsigned i = 0; i < nr_params; i++) {
+      uint64_t param = blob_read_uint64(binary);
+      ptrdiff_t p_offset, u_offset;
+      struct gl_program_parameter_list *param_list =
+         prog->_LinkedShaders[stage]->Program->Parameters;
+
+      p_offset = (param - parameter_values_base) / sizeof(gl_constant_value);
+      u_offset = (param - uniform_data_slots_base) / sizeof(gl_constant_value);
+      
+      if (p_offset >= 0 && p_offset < 4 * param_list->NumParameters) {
+         prog_data->param[i] =
+            ((gl_constant_value *) param_list->ParameterValues) + p_offset;
+      } else if (u_offset >= 0 && u_offset < prog->NumUniformDataSlots) {
+         prog_data->param[i] = prog->UniformDataSlots + u_offset;
+      } else {
+         prog_data->param[i] = &zero;
+      }
+   }
+
+   uint32_t nr_pull_params = blob_read_uint32(binary);
+   assert(nr_pull_params == prog_data->nr_pull_params);
+
+   prog_data->pull_param = rzalloc_array(NULL, const gl_constant_value *,
+                                         nr_pull_params);
+
+   for (unsigned i = 0; i < nr_pull_params; i++) {
+      /* FIXME: We need to fixup pull_params pointers here. */
+   }
+
+}
+
+#define SET_UPLOAD_PRAMS(sh, sh_caps, prog, data_base)               \
+      struct brw_##sh##_prog_data *s_prog_data = shader_prog_data;   \
+      assert(prog_data_size == sizeof *s_prog_data);                 \
+      sh##_key.program_string_id = prog->id;                         \
+      prog_data = &s_prog_data->data_base;                           \
+      brw_prog = prog;                                               \
+      cache_id = BRW_CACHE_##sh_caps##_PROG;                         \
+      out_aux = &brw->sh.prog_data;                                  \
+      key = &sh##_key;                                               \
+      max_threads = devinfo->max_##sh##_threads;                     \
+      stage_state = &brw->sh.base;                                   \
+
+static bool
+read_and_upload(struct brw_context *brw, struct program_cache *cache,
+                struct blob_reader *binary, struct gl_shader_program *prog,
+                gl_shader_stage stage)
+{
+   const struct gen_device_info *devinfo = &brw->screen->devinfo;
+
+   char sha1_buf[41];
+   unsigned char binary_sha1[20];
+   size_t size;
+   uint8_t *buffer;
+
+   struct brw_stage_prog_data *prog_data;
+   struct brw_wm_prog_key wm_key;
+   struct brw_vs_prog_key vs_key;
+   struct brw_stage_state *stage_state;
+
+   enum brw_cache_id cache_id;
+   unsigned max_threads;
+   void *out_aux;
+   void *brw_prog;
+   void *key;
+
+   brw_vs_populate_key(brw, &vs_key);
+   /* We don't care what instance of the program it is we only care if
+    * its the correct binary to load so ignore program id for on disk cache.
+    */
+   vs_key.program_string_id = 0;
+
+   switch (stage) {
+   case MESA_SHADER_VERTEX:
+      gen_shader_sha1(brw, prog, stage, &vs_key, binary_sha1);
+      break;
+   case MESA_SHADER_FRAGMENT:
+      brw_wm_populate_key(brw, &wm_key);
+      gen_shader_sha1(brw, prog, stage, &wm_key, binary_sha1);
+      break;
+   default:
+      unreachable("Unsupported stage!");
+   }
+
+   buffer = cache_get(cache, binary_sha1, &size);
+   if (buffer == NULL)
+      return false;
+
+   if (brw->ctx._Shader->Flags & GLSL_CACHE_INFO) {
+      fprintf(stderr, "attempting to populate bo cache with binary: %s\n",
+              _mesa_sha1_format(sha1_buf, binary_sha1));
+   }
+
+   blob_reader_init(binary, buffer, size);
+
+   /* Read shader program from blob. */
+   size_t program_size = blob_read_uint32(binary);
+   uint8_t *program = blob_read_bytes(binary, program_size);
+
+   /* Read shader program_data from blob. */
+   size_t prog_data_size = blob_read_uint32(binary);
+   void *shader_prog_data = blob_read_bytes(binary, prog_data_size);
+
+   switch (stage) {
+   case MESA_SHADER_VERTEX: {
+      struct brw_vertex_program *vp =
+         (struct brw_vertex_program *)brw->vertex_program;
+      SET_UPLOAD_PRAMS(vs, VS, vp, base.base)
+      break;
+   }
+   case MESA_SHADER_FRAGMENT: {
+      struct brw_fragment_program *wp =
+         (struct brw_fragment_program *)brw->fragment_program;
+      SET_UPLOAD_PRAMS(wm, FS, wp, base)
+      break;
+   }
+   default:
+      unreachable("Unsupported stage!");
+   }
+
+   load_program_data(prog, binary, prog_data, stage, &brw->ctx);
+
+   brw_alloc_stage_scratch(brw, stage_state, prog_data->total_scratch,
+                           max_threads);
+
+   brw_upload_cache(&brw->cache, cache_id, key, key_size(stage), program,
+                    program_size, prog_data, prog_data_size,
+                    &stage_state->prog_offset, out_aux, brw_prog);
+
+   free(buffer);
+
+   return true;
+}
+
+void
+upload_cached_program(struct brw_context *brw, gl_shader_stage stage)
+{
+   struct blob_reader binary;
+
+   struct program_cache *cache = brw->ctx.Cache;
+   if (cache == NULL)
+      return;
+
+   struct gl_shader_program *prog = brw->ctx.Shader.ActiveProgram;
+   if (prog == NULL)
+      return;
+
+   if (!read_and_upload(brw, cache, &binary, prog, stage))
+      goto FAIL;
+
+   if (binary.current != binary.end || binary.overrun) {
+      if (brw->ctx._Shader->Flags & GLSL_CACHE_INFO) {
+         fprintf(stderr, "Error reading program from cache (did not read "
+                 "every byte written)\n");
+      }
+      goto FAIL;
+   }
+
+   if (brw->ctx._Shader->Flags & GLSL_CACHE_INFO) {
+      fprintf(stderr, "%s: Successfully read every byte written!\n",
+              __FUNCTION__);
+   }
+   prog->program_written_to_cache = true;
+   return;
+
+FAIL:
+   /*FIXME: Fall back and compile from source here. */
+   prog->program_written_to_cache = false;
+}
+
+static bool
+write_program_data(struct brw_context *brw, struct gl_shader_program *prog,
+                   void *key, struct brw_stage_prog_data *stage_prog_data,
+                   size_t program_size, void *prog_data,
+                   size_t prog_data_size, uint32_t prog_offset,
+                   struct program_cache *cache, gl_shader_stage stage)
+{
+   unsigned char sha1[20];
+   char buf[41];
+
+   struct blob *binary = blob_create (NULL);
+   if (binary == NULL)
+      return false;
+
+   gen_shader_sha1(brw, prog, stage, key, sha1);
+
+   /* Write program to blob. */
+   blob_write_uint32(binary, program_size);
+
+   uint8_t *blob_cursor = blob_reserve_bytes(binary, program_size);
+   drm_intel_bo_get_subdata(brw->cache.bo, prog_offset, program_size,
+                            blob_cursor);
+
+   /* Write program_data to blob. */
+   blob_write_uint32(binary, prog_data_size);
+   blob_write_bytes(binary, prog_data, prog_data_size);
+
+   /* Include variable-length params from end of brw_stage_prog_data as well.
+    *
+    * Before writing either of the params or pull_params arrays, we first
+    * write out the addresses of the ParameterValues and UniformDataSlots
+    * storage. The pointers within params will be pointers to within one of
+    * these blocks of storage. So we can use the addresses of this storage
+    * together with the pointer values to correctly construct pointers to the
+    * actual storage when the program data is loaded from the cache.
+    */
+   blob_write_uint64(binary,
+                     ptr_to_uint64_t(prog->_LinkedShaders[stage]->
+                        Program->Parameters->ParameterValues));
+
+   blob_write_uint64(binary, ptr_to_uint64_t(prog->UniformDataSlots));
+
+   blob_write_uint32(binary, stage_prog_data->nr_params);
+
+   for (unsigned i = 0; i < stage_prog_data->nr_params; i++) {
+      blob_write_uint64(binary,
+                        ptr_to_uint64_t((void *) stage_prog_data->param[i]));
+   }
+
+   blob_write_uint32(binary, stage_prog_data->nr_pull_params);
+
+   if (brw->ctx._Shader->Flags & GLSL_CACHE_INFO) {
+      fprintf(stderr, "putting binary in cache: %s\n",
+              _mesa_sha1_format(buf, sha1));
+   }
+
+   cache_put(cache, sha1, binary->data, binary->size);
+   ralloc_free (binary);
+
+   return true;
+}
+
+void
+write_cached_program(struct brw_context *brw)
+{
+   struct program_cache *cache = brw->ctx.Cache;
+   if (cache == NULL)
+      return;
+
+   struct gl_shader_program *prog = brw->ctx.Shader.ActiveProgram;
+   if (prog == NULL)
+      return;
+
+   if (prog->program_written_to_cache)
+      return;
+
+   if (prog->_LinkedShaders[MESA_SHADER_VERTEX]) {
+      struct brw_vs_prog_key vs_key;
+      brw_vs_populate_key(brw, &vs_key);
+      vs_key.program_string_id = 0;
+
+      if (!write_program_data(brw, prog, &vs_key,
+                              &brw->vs.prog_data->base.base,
+                              brw->vs.prog_data->program_size,
+                              brw->vs.prog_data, sizeof *brw->vs.prog_data,
+                              brw->vs.base.prog_offset, cache,
+                              MESA_SHADER_VERTEX)) {
+         return;
+      }
+   }
+
+   if (prog->_LinkedShaders[MESA_SHADER_FRAGMENT]) {
+      struct brw_wm_prog_key wm_key;
+      brw_wm_populate_key(brw, &wm_key);
+      wm_key.program_string_id = 0;
+
+      if (!write_program_data(brw, prog, &wm_key, &brw->wm.prog_data->base,
+                              brw->wm.prog_data->program_size,
+                              brw->wm.prog_data, sizeof *brw->wm.prog_data,
+                              brw->wm.base.prog_offset, cache,
+                              MESA_SHADER_FRAGMENT)) {
+         return;
+      }
+   }
+
+   prog->program_written_to_cache = true;
+}
diff --git a/src/mesa/drivers/dri/i965/brw_state.h b/src/mesa/drivers/dri/i965/brw_state.h
index aba9508..2a11a55 100644
--- a/src/mesa/drivers/dri/i965/brw_state.h
+++ b/src/mesa/drivers/dri/i965/brw_state.h
@@ -196,6 +196,13 @@ void brw_upload_state_base_address(struct brw_context *brw);
 void gen8_write_pma_stall_bits(struct brw_context *brw,
                                uint32_t pma_stall_bits);
 
+/* brw_shader_cache.h */
+void
+upload_cached_program(struct brw_context *brw, gl_shader_stage stage);
+
+void
+write_cached_program(struct brw_context *brw);
+
 /***********************************************************************
  * brw_state.c
  */
-- 
2.7.4



More information about the mesa-dev mailing list