i965: Convert the binding table to streamed indirect state.

This slightly reduces reduces cairo-gl firefox-talos-gfx runtime on my
Ironlake:
before:
[ # ]  backend                         test   min(s) median(s) stddev. count
[  0]       gl            firefox-talos-gfx   38.236   38.383   0.43%    5/6
after:
[  0]       gl            firefox-talos-gfx   37.799   38.203   0.39%    6/6

It turns out the cost of caching these objects and looking them up in
the cache again is greater than the cost of just computing the object
again, particularly when the overhead of having a separate BO to pin
is removed.

(Those that are paying close attention will note that this is a
reversal of the path I was moving the driver in a couple of years ago.
The major thing that has changed is that back then all state was
recomputed when we wrapped the streaming state buffer, including
recompiling our precious programs.  Now, we're uncaching just the
objects that are cheap to compute, and retaining caching of expensive
objects)

8 files changed, 132 insertions, 146 deletions

diff --git a/configure.ac b/configure.ac
index 1056fa1810..7307d89ae2 100644
--- a/configure.ac
+++ b/configure.ac
@@ -860,7 +860,7 @@ AC_SUBST([DRI_LIB_DEPS])
 
 case $DRI_DIRS in
 *i915*|*i965*)
-    PKG_CHECK_MODULES([INTEL], [libdrm_intel >= 2.4.19])
+    PKG_CHECK_MODULES([INTEL], [libdrm_intel >= 2.4.21])
     ;;
 esac
 
diff --git a/src/mesa/drivers/dri/i965/brw_context.h b/src/mesa/drivers/dri/i965/brw_context.h
index a829067383..86b86fde9a 100644
--- a/src/mesa/drivers/dri/i965/brw_context.h
+++ b/src/mesa/drivers/dri/i965/brw_context.h
@@ -131,6 +131,7 @@ struct brw_context;
 #define BRW_NEW_WM_INPUT_DIMENSIONS     0x100
 #define BRW_NEW_PSP                     0x800
 #define BRW_NEW_WM_SURFACES		0x1000
+#define BRW_NEW_BINDING_TABLE		0x2000
 #define BRW_NEW_INDICES			0x4000
 #define BRW_NEW_VERTICES		0x8000
 /**
@@ -302,7 +303,6 @@ enum brw_cache_id {
    BRW_CLIP_UNIT,
    BRW_CLIP_PROG,
    BRW_SS_SURFACE,
-   BRW_SS_SURF_BIND,
 
    BRW_MAX_CACHE
 };
@@ -377,7 +377,6 @@ struct brw_tracked_state {
 #define CACHE_NEW_CLIP_UNIT              (1<<BRW_CLIP_UNIT)
 #define CACHE_NEW_CLIP_PROG              (1<<BRW_CLIP_PROG)
 #define CACHE_NEW_SURFACE                (1<<BRW_SS_SURFACE)
-#define CACHE_NEW_SURF_BIND              (1<<BRW_SS_SURF_BIND)
 
 struct brw_cached_batch_item {
    struct header *header;
@@ -460,7 +459,7 @@ struct brw_context
        * consisting of the vertex buffers, pipelined state pointers,
        * the CURBE, the depth buffer, and a query BO.
        */
-      drm_intel_bo *validated_bos[VERT_ATTRIB_MAX + 16];
+      drm_intel_bo *validated_bos[VERT_ATTRIB_MAX + BRW_WM_MAX_SURF + 16];
       int validated_bo_count;
    } state;
 
@@ -598,6 +597,7 @@ struct brw_context
 
       /** Binding table of pointers to surf_bo entries */
       drm_intel_bo *bind_bo;
+      uint32_t bind_bo_offset;
       drm_intel_bo *surf_bo[BRW_VS_MAX_SURF];
       GLuint nr_surfaces;      
    } vs;
@@ -650,6 +650,7 @@ struct brw_context
 
       /** Binding table of pointers to surf_bo entries */
       drm_intel_bo *bind_bo;
+      uint32_t bind_bo_offset;
       drm_intel_bo *surf_bo[BRW_WM_MAX_SURF];
 
       drm_intel_bo *prog_bo;
diff --git a/src/mesa/drivers/dri/i965/brw_misc_state.c b/src/mesa/drivers/dri/i965/brw_misc_state.c
index 35908ee7b6..ab5d5240d0 100644
--- a/src/mesa/drivers/dri/i965/brw_misc_state.c
+++ b/src/mesa/drivers/dri/i965/brw_misc_state.c
@@ -96,12 +96,6 @@ const struct brw_tracked_state brw_drawing_rect = {
    .emit = upload_drawing_rect
 };
 
-static void prepare_binding_table_pointers(struct brw_context *brw)
-{
-   brw_add_validated_bo(brw, brw->vs.bind_bo);
-   brw_add_validated_bo(brw, brw->wm.bind_bo);
-}
-
 /**
  * Upload the binding table pointers, which point each stage's array of surface
  * state pointers.
@@ -116,23 +110,24 @@ static void upload_binding_table_pointers(struct brw_context *brw)
    BEGIN_BATCH(6);
    OUT_BATCH(CMD_BINDING_TABLE_PTRS << 16 | (6 - 2));
    if (brw->vs.bind_bo != NULL)
-      OUT_RELOC(brw->vs.bind_bo, I915_GEM_DOMAIN_SAMPLER, 0, 0); /* vs */
+      OUT_RELOC(brw->vs.bind_bo, I915_GEM_DOMAIN_SAMPLER, 0,
+		brw->vs.bind_bo_offset); /* vs */
    else
       OUT_BATCH(0);
    OUT_BATCH(0); /* gs */
    OUT_BATCH(0); /* clip */
    OUT_BATCH(0); /* sf */
-   OUT_RELOC(brw->wm.bind_bo, I915_GEM_DOMAIN_SAMPLER, 0, 0); /* wm/ps */
+   OUT_RELOC(brw->wm.bind_bo, I915_GEM_DOMAIN_SAMPLER, 0,
+	     brw->wm.bind_bo_offset); /* wm/ps */
    ADVANCE_BATCH();
 }
 
 const struct brw_tracked_state brw_binding_table_pointers = {
    .dirty = {
       .mesa = 0,
-      .brw = BRW_NEW_BATCH,
-      .cache = CACHE_NEW_SURF_BIND,
+      .brw = BRW_NEW_BATCH | BRW_NEW_BINDING_TABLE,
+      .cache = 0,
    },
-   .prepare = prepare_binding_table_pointers,
    .emit = upload_binding_table_pointers,
 };
 
@@ -154,21 +149,22 @@ static void upload_gen6_binding_table_pointers(struct brw_context *brw)
 	     GEN6_BINDING_TABLE_MODIFY_PS |
 	     (4 - 2));
    if (brw->vs.bind_bo != NULL)
-      OUT_RELOC(brw->vs.bind_bo, I915_GEM_DOMAIN_SAMPLER, 0, 0); /* vs */
+      OUT_RELOC(brw->vs.bind_bo, I915_GEM_DOMAIN_SAMPLER, 0,
+		brw->vs.bind_bo_offset); /* vs */
    else
       OUT_BATCH(0);
    OUT_BATCH(0); /* gs */
-   OUT_RELOC(brw->wm.bind_bo, I915_GEM_DOMAIN_SAMPLER, 0, 0); /* wm/ps */
+   OUT_RELOC(brw->wm.bind_bo, I915_GEM_DOMAIN_SAMPLER, 0,
+	     brw->wm.bind_bo_offset); /* wm/ps */
    ADVANCE_BATCH();
 }
 
 const struct brw_tracked_state gen6_binding_table_pointers = {
    .dirty = {
       .mesa = 0,
-      .brw = BRW_NEW_BATCH,
-      .cache = CACHE_NEW_SURF_BIND,
+      .brw = BRW_NEW_BATCH | BRW_NEW_BINDING_TABLE,
+      .cache = 0,
    },
-   .prepare = prepare_binding_table_pointers,
    .emit = upload_gen6_binding_table_pointers,
 };
 
diff --git a/src/mesa/drivers/dri/i965/brw_state.h b/src/mesa/drivers/dri/i965/brw_state.h
index 364be94117..68fd7d4f80 100644
--- a/src/mesa/drivers/dri/i965/brw_state.h
+++ b/src/mesa/drivers/dri/i965/brw_state.h
@@ -81,6 +81,7 @@ const struct brw_tracked_state brw_wm_prog;
 const struct brw_tracked_state brw_wm_samplers;
 const struct brw_tracked_state brw_wm_constant_surface;
 const struct brw_tracked_state brw_wm_surfaces;
+const struct brw_tracked_state brw_wm_binding_table;
 const struct brw_tracked_state brw_wm_unit;
 
 const struct brw_tracked_state brw_psp_urb_cbs;
diff --git a/src/mesa/drivers/dri/i965/brw_state_cache.c b/src/mesa/drivers/dri/i965/brw_state_cache.c
index f1d89484fc..45f1088f4a 100644
--- a/src/mesa/drivers/dri/i965/brw_state_cache.c
+++ b/src/mesa/drivers/dri/i965/brw_state_cache.c
@@ -410,7 +410,6 @@ brw_init_surface_cache(struct brw_context *brw)
       calloc(1, cache->size * sizeof(struct brw_cache_item));
 
    brw_init_cache_id(cache, "SS_SURFACE", BRW_SS_SURFACE);
-   brw_init_cache_id(cache, "SS_SURF_BIND", BRW_SS_SURF_BIND);
 }
 
 
diff --git a/src/mesa/drivers/dri/i965/brw_state_upload.c b/src/mesa/drivers/dri/i965/brw_state_upload.c
index 08535bb59c..e345dbcf5b 100644
--- a/src/mesa/drivers/dri/i965/brw_state_upload.c
+++ b/src/mesa/drivers/dri/i965/brw_state_upload.c
@@ -69,6 +69,7 @@ static const struct brw_tracked_state *gen4_atoms[] =
    &brw_vs_surfaces,		/* must do before unit */
    &brw_wm_constant_surface,	/* must do before wm surfaces/bind bo */
    &brw_wm_surfaces,		/* must do before samplers and unit */
+   &brw_wm_binding_table,
    &brw_wm_samplers,
 
    &brw_wm_unit,
@@ -268,6 +269,8 @@ static struct dirty_bit_map brw_bits[] = {
    DEFINE_BIT(BRW_NEW_CONTEXT),
    DEFINE_BIT(BRW_NEW_WM_INPUT_DIMENSIONS),
    DEFINE_BIT(BRW_NEW_PSP),
+   DEFINE_BIT(BRW_NEW_WM_SURFACES),
+   DEFINE_BIT(BRW_NEW_BINDING_TABLE),
    DEFINE_BIT(BRW_NEW_INDICES),
    DEFINE_BIT(BRW_NEW_INDEX_BUFFER),
    DEFINE_BIT(BRW_NEW_VERTICES),
@@ -295,7 +298,6 @@ static struct dirty_bit_map cache_bits[] = {
    DEFINE_BIT(CACHE_NEW_CLIP_UNIT),
    DEFINE_BIT(CACHE_NEW_CLIP_PROG),
    DEFINE_BIT(CACHE_NEW_SURFACE),
-   DEFINE_BIT(CACHE_NEW_SURF_BIND),
    {0, 0, 0}
 };
 
diff --git a/src/mesa/drivers/dri/i965/brw_vs_surface_state.c b/src/mesa/drivers/dri/i965/brw_vs_surface_state.c
index 26164e907f..d946756af7 100644
--- a/src/mesa/drivers/dri/i965/brw_vs_surface_state.c
+++ b/src/mesa/drivers/dri/i965/brw_vs_surface_state.c
@@ -151,49 +151,29 @@ brw_update_vs_constant_surface( GLcontext *ctx,
 }
 
 
-/**
- * Constructs the binding table for the VS surface state.
- */
-static drm_intel_bo *
-brw_vs_get_binding_table(struct brw_context *brw)
+static void
+prepare_vs_surfaces(struct brw_context *brw)
 {
-   drm_intel_bo *bind_bo;
-
-   bind_bo = brw_search_cache(&brw->surface_cache, BRW_SS_SURF_BIND,
-			      NULL, 0,
-			      brw->vs.surf_bo, BRW_VS_MAX_SURF,
-			      NULL);
-
-   if (bind_bo == NULL) {
-      GLuint data_size = BRW_VS_MAX_SURF * sizeof(GLuint);
-      uint32_t data[BRW_VS_MAX_SURF];
-      int i;
-
-      for (i = 0; i < BRW_VS_MAX_SURF; i++)
-         if (brw->vs.surf_bo[i])
-            data[i] = brw->vs.surf_bo[i]->offset;
-         else
-            data[i] = 0;
-
-      bind_bo = brw_upload_cache( &brw->surface_cache, BRW_SS_SURF_BIND,
-				  NULL, 0,
-				  brw->vs.surf_bo, BRW_VS_MAX_SURF,
-				  data, data_size);
-
-      /* Emit binding table relocations to surface state */
-      for (i = 0; i < BRW_VS_MAX_SURF; i++) {
-	 if (brw->vs.surf_bo[i] != NULL) {
-	    /* The presumed offsets were set in the data values for
-	     * brw_upload_cache.
-	     */
-	    drm_intel_bo_emit_reloc(bind_bo, i * 4,
-				    brw->vs.surf_bo[i], 0,
-				    I915_GEM_DOMAIN_INSTRUCTION, 0);
-	 }
+   GLcontext *ctx = &brw->intel.ctx;
+   int i;
+   int nr_surfaces = 0;
+
+   brw_update_vs_constant_surface(ctx, SURF_INDEX_VERT_CONST_BUFFER);
+
+   for (i = 0; i < BRW_VS_MAX_SURF; i++) {
+      if (brw->vs.surf_bo[i] != NULL) {
+	 nr_surfaces = i + 1;
       }
    }
 
-   return bind_bo;
+   if (brw->vs.nr_surfaces != nr_surfaces) {
+      brw->state.dirty.brw |= BRW_NEW_NR_VS_SURFACES;
+      brw->vs.nr_surfaces = nr_surfaces;
+   }
+
+   for (i = 0; i < BRW_VS_MAX_SURF; i++) {
+      brw_add_validated_bo(brw, brw->vs.surf_bo[i]);
+   }
 }
 
 /**
@@ -203,43 +183,51 @@ brw_vs_get_binding_table(struct brw_context *brw)
  * to be updated, and produces BRW_NEW_NR_VS_SURFACES for the VS unit and
  * CACHE_NEW_SURF_BIND for the binding table upload.
  */
-static void prepare_vs_surfaces(struct brw_context *brw )
+static void upload_vs_surfaces(struct brw_context *brw)
 {
-   GLcontext *ctx = &brw->intel.ctx;
+   uint32_t *bind;
    int i;
-   int nr_surfaces = 0;
-
-   brw_update_vs_constant_surface(ctx, SURF_INDEX_VERT_CONST_BUFFER);
 
-   for (i = 0; i < BRW_VS_MAX_SURF; i++) {
-      if (brw->vs.surf_bo[i] != NULL) {
-	 nr_surfaces = i + 1;
+   /* BRW_NEW_NR_VS_SURFACES */
+   if (brw->vs.nr_surfaces == 0) {
+      if (brw->vs.bind_bo) {
+	 drm_intel_bo_unreference(brw->vs.bind_bo);
+	 brw->vs.bind_bo = NULL;
+	 brw->state.dirty.brw |= BRW_NEW_BINDING_TABLE;
       }
+      return;
    }
 
-   if (brw->vs.nr_surfaces != nr_surfaces) {
-      brw->state.dirty.brw |= BRW_NEW_NR_VS_SURFACES;
-      brw->vs.nr_surfaces = nr_surfaces;
-   }
-
-   /* Note that we don't end up updating the bind_bo if we don't have a
-    * surface to be pointing at.  This should be relatively harmless, as it
-    * just slightly increases our working set size.
+   /* Might want to calculate nr_surfaces first, to avoid taking up so much
+    * space for the binding table. (once we have vs samplers)
     */
-   if (brw->vs.nr_surfaces != 0) {
-      drm_intel_bo_unreference(brw->vs.bind_bo);
-      brw->vs.bind_bo = brw_vs_get_binding_table(brw);
+   bind = brw_state_batch(brw, sizeof(uint32_t) * BRW_VS_MAX_SURF,
+			  32, &brw->vs.bind_bo, &brw->vs.bind_bo_offset);
+
+   for (i = 0; i < BRW_VS_MAX_SURF; i++) {
+      /* BRW_NEW_VS_CONSTBUF */
+      if (brw->vs.surf_bo[i]) {
+	 drm_intel_bo_emit_reloc(brw->vs.bind_bo,
+				 brw->vs.bind_bo_offset + i * sizeof(uint32_t),
+				 brw->vs.surf_bo[i], 0,
+				 I915_GEM_DOMAIN_INSTRUCTION, 0);
+	 bind[i] = brw->vs.surf_bo[i]->offset;
+      } else {
+	 bind[i] = 0;
+      }
    }
+
+   brw->state.dirty.brw |= BRW_NEW_BINDING_TABLE;
 }
 
 const struct brw_tracked_state brw_vs_surfaces = {
    .dirty = {
       .mesa = 0,
-      .brw = (BRW_NEW_VS_CONSTBUF),
+      .brw = (BRW_NEW_VS_CONSTBUF |
+	      BRW_NEW_NR_VS_SURFACES |
+	      BRW_NEW_BATCH),
       .cache = 0
    },
    .prepare = prepare_vs_surfaces,
+   .emit = upload_vs_surfaces,
 };
-
-
-
diff --git a/src/mesa/drivers/dri/i965/brw_wm_surface_state.c b/src/mesa/drivers/dri/i965/brw_wm_surface_state.c
index 2b216fddbb..ba6a6258f5 100644
--- a/src/mesa/drivers/dri/i965/brw_wm_surface_state.c
+++ b/src/mesa/drivers/dri/i965/brw_wm_surface_state.c
@@ -639,57 +639,10 @@ brw_update_renderbuffer_surface(struct brw_context *brw,
    }
 }
 
-
-/**
- * Constructs the binding table for the WM surface state, which maps unit
- * numbers to surface state objects.
- */
-static drm_intel_bo *
-brw_wm_get_binding_table(struct brw_context *brw)
-{
-   drm_intel_bo *bind_bo;
-
-   assert(brw->wm.nr_surfaces <= BRW_WM_MAX_SURF);
-
-   bind_bo = brw_search_cache(&brw->surface_cache, BRW_SS_SURF_BIND,
-			      NULL, 0,
-			      brw->wm.surf_bo, brw->wm.nr_surfaces,
-			      NULL);
-
-   if (bind_bo == NULL) {
-      GLuint data_size = brw->wm.nr_surfaces * sizeof(GLuint);
-      uint32_t data[BRW_WM_MAX_SURF];
-      int i;
-
-      for (i = 0; i < brw->wm.nr_surfaces; i++)
-         if (brw->wm.surf_bo[i])
-            data[i] = brw->wm.surf_bo[i]->offset;
-         else
-            data[i] = 0;
-
-      bind_bo = brw_upload_cache( &brw->surface_cache, BRW_SS_SURF_BIND,
-				  NULL, 0,
-				  brw->wm.surf_bo, brw->wm.nr_surfaces,
-				  data, data_size);
-
-      /* Emit binding table relocations to surface state */
-      for (i = 0; i < BRW_WM_MAX_SURF; i++) {
-	 if (brw->wm.surf_bo[i] != NULL) {
-	    drm_intel_bo_emit_reloc(bind_bo, i * sizeof(GLuint),
-				    brw->wm.surf_bo[i], 0,
-				    I915_GEM_DOMAIN_INSTRUCTION, 0);
-	 }
-      }
-   }
-
-   return bind_bo;
-}
-
 static void prepare_wm_surfaces(struct brw_context *brw )
 {
    GLcontext *ctx = &brw->intel.ctx;
    GLuint i;
-   int old_nr_surfaces;
 
    /* _NEW_BUFFERS | _NEW_COLOR */
    /* Update surfaces for drawing buffers */
@@ -703,32 +656,21 @@ static void prepare_wm_surfaces(struct brw_context *brw )
       brw_update_renderbuffer_surface(brw, NULL, 0);
    }
 
-   old_nr_surfaces = brw->wm.nr_surfaces;
-   brw->wm.nr_surfaces = BRW_MAX_DRAW_BUFFERS;
-
-   if (brw->wm.surf_bo[SURF_INDEX_FRAG_CONST_BUFFER] != NULL)
-       brw->wm.nr_surfaces = SURF_INDEX_FRAG_CONST_BUFFER + 1;
-
    /* Update surfaces for textures */
    for (i = 0; i < BRW_MAX_TEX_UNIT; i++) {
       const struct gl_texture_unit *texUnit = &ctx->Texture.Unit[i];
       const GLuint surf = SURF_INDEX_TEXTURE(i);
 
-      /* _NEW_TEXTURE, BRW_NEW_TEXDATA */
+      /* _NEW_TEXTURE */
       if (texUnit->_ReallyEnabled) {
 	 brw_update_texture_surface(ctx, i);
-	 brw->wm.nr_surfaces = surf + 1;
       } else {
          drm_intel_bo_unreference(brw->wm.surf_bo[surf]);
          brw->wm.surf_bo[surf] = NULL;
       }
    }
 
-   drm_intel_bo_unreference(brw->wm.bind_bo);
-   brw->wm.bind_bo = brw_wm_get_binding_table(brw);
-
-   if (brw->wm.nr_surfaces != old_nr_surfaces)
-      brw->state.dirty.brw |= BRW_NEW_NR_WM_SURFACES;
+   brw->state.dirty.brw |= BRW_NEW_WM_SURFACES;
 }
 
 const struct brw_tracked_state brw_wm_surfaces = {
@@ -736,12 +678,69 @@ const struct brw_tracked_state brw_wm_surfaces = {
       .mesa = (_NEW_COLOR |
                _NEW_TEXTURE |
                _NEW_BUFFERS),
-      .brw = (BRW_NEW_CONTEXT |
-	      BRW_NEW_WM_SURFACES),
+      .brw = (BRW_NEW_CONTEXT),
       .cache = 0
    },
    .prepare = prepare_wm_surfaces,
 };
 
+static void
+brw_wm_prepare_binding_table(struct brw_context *brw)
+{
+   int i;
 
+   for (i = 0; i < BRW_WM_MAX_SURF; i++) {
+      if (brw->wm.surf_bo[i]) {
+	 brw_add_validated_bo(brw, brw->wm.surf_bo[i]);
+      }
+   }
+}
 
+/**
+ * Constructs the binding table for the WM surface state, which maps unit
+ * numbers to surface state objects.
+ */
+static void
+brw_wm_upload_binding_table(struct brw_context *brw)
+{
+   uint32_t *bind;
+   int i, nr_surfaces = 0;
+
+   /* Might want to calculate nr_surfaces first, to avoid taking up so much
+    * space for the binding table.
+    */
+   bind = brw_state_batch(brw, sizeof(uint32_t) * BRW_WM_MAX_SURF,
+			  32, &brw->wm.bind_bo, &brw->wm.bind_bo_offset);
+
+   for (i = 0; i < BRW_WM_MAX_SURF; i++) {
+      /* BRW_NEW_WM_SURFACES */
+      if (brw->wm.surf_bo[i]) {
+	 drm_intel_bo_emit_reloc(brw->wm.bind_bo,
+				 brw->wm.bind_bo_offset + i * sizeof(uint32_t),
+				 brw->wm.surf_bo[i], 0,
+				 I915_GEM_DOMAIN_INSTRUCTION, 0);
+	 bind[i] = brw->wm.surf_bo[i]->offset;
+	 nr_surfaces = i + 1;
+      } else {
+	 bind[i] = 0;
+      }
+   }
+
+   if (brw->wm.nr_surfaces != nr_surfaces) {
+      brw->wm.nr_surfaces = nr_surfaces;
+      brw->state.dirty.brw |= BRW_NEW_NR_WM_SURFACES;
+   }
+
+   brw->state.dirty.brw |= BRW_NEW_BINDING_TABLE;
+}
+
+const struct brw_tracked_state brw_wm_binding_table = {
+   .dirty = {
+      .mesa = 0,
+      .brw = (BRW_NEW_BATCH |
+	      BRW_NEW_WM_SURFACES),
+      .cache = 0
+   },
+   .prepare = brw_wm_prepare_binding_table,
+   .emit = brw_wm_upload_binding_table,
+};