4 files changed, 313 insertions, 37 deletions

diff --git a/src/gallium/auxiliary/cso_cache/cso_cache.c b/src/gallium/auxiliary/cso_cache/cso_cache.c
index 9e77e0774d..b427b509f8 100644
--- a/src/gallium/auxiliary/cso_cache/cso_cache.c
+++ b/src/gallium/auxiliary/cso_cache/cso_cache.c
@@ -28,9 +28,22 @@
 /* Authors:  Zack Rusin <zack@tungstengraphics.com>
  */
 
+#include "pipe/p_util.h"
+
 #include "cso_cache.h"
 #include "cso_hash.h"
 
+
+struct cso_cache {
+   struct cso_hash *blend_hash;
+   struct cso_hash *depth_stencil_hash;
+   struct cso_hash *fs_hash;
+   struct cso_hash *vs_hash;
+   struct cso_hash *rasterizer_hash;
+   struct cso_hash *sampler_hash;
+   int    max_size;
+};
+
 #if 1
 static unsigned hash_key(const void *key, unsigned key_size)
 {
@@ -114,12 +127,106 @@ static int _cso_size_for_type(enum cso_cache_type type)
    return 0;
 }
 
+
+static void delete_blend_state(void *state, void *data)
+{
+   struct cso_blend *cso = (struct cso_blend *)state;
+   if (cso->delete_state && cso->data != &cso->state)
+      cso->delete_state(cso->context, cso->data);
+}
+
+static void delete_depth_stencil_state(void *state, void *data)
+{
+   struct cso_depth_stencil_alpha *cso = (struct cso_depth_stencil_alpha *)state;
+   if (cso->delete_state && cso->data != &cso->state)
+      cso->delete_state(cso->context, cso->data);
+}
+
+static void delete_sampler_state(void *state, void *data)
+{
+   struct cso_sampler *cso = (struct cso_sampler *)state;
+   if (cso->delete_state && cso->data != &cso->state)
+      cso->delete_state(cso->context, cso->data);
+}
+
+static void delete_rasterizer_state(void *state, void *data)
+{
+   struct cso_rasterizer *cso = (struct cso_rasterizer *)state;
+   if (cso->delete_state && cso->data != &cso->state)
+      cso->delete_state(cso->context, cso->data);
+}
+
+static void delete_fs_state(void *state, void *data)
+{
+   struct cso_fragment_shader *cso = (struct cso_fragment_shader *)state;
+   if (cso->delete_state && cso->data != &cso->state)
+      cso->delete_state(cso->context, cso->data);
+}
+
+static void delete_vs_state(void *state, void *data)
+{
+   struct cso_vertex_shader *cso = (struct cso_vertex_shader *)state;
+   if (cso->delete_state && cso->data != &cso->state)
+      cso->delete_state(cso->context, cso->data);
+}
+
+
+static INLINE void delete_cso(void *state, enum cso_cache_type type)
+{
+   switch (type) {
+   case CSO_BLEND: {
+      delete_blend_state(state, 0);
+   }
+      break;
+   case CSO_SAMPLER: {
+      delete_sampler_state(state, 0);
+   }
+      break;
+   case CSO_DEPTH_STENCIL_ALPHA: {
+      delete_depth_stencil_state(state, 0);
+   }
+      break;
+   case CSO_RASTERIZER: {
+      delete_rasterizer_state(state, 0);
+   }
+      break;
+   case CSO_FRAGMENT_SHADER: {
+      delete_fs_state(state, 0);
+   }
+      break;
+   case CSO_VERTEX_SHADER: {
+      delete_vs_state(state, 0);
+   }
+      break;
+   }
+   FREE(state);
+}
+
+static INLINE void sanitize_hash(struct cso_hash *hash, enum cso_cache_type type,
+                                 int max_size)
+{
+   /* if we're approach the maximum size, remove fourth of the entries
+    * otherwise every subsequent call will go through the same */
+   int max_entries = (max_size > cso_hash_size(hash)) ? max_size : cso_hash_size(hash);
+   int to_remove =  (max_size < max_entries) * max_entries/4;
+   while (to_remove) {
+      /*remove elements until we're good */
+      /*fixme: currently we pick the nodes to remove at random*/
+      struct cso_hash_iter iter = cso_hash_first_node(hash);
+      void  *cso = cso_hash_take(hash, cso_hash_iter_key(iter));
+      delete_cso(cso, type);
+      --to_remove;
+   }
+}
+
 struct cso_hash_iter
 cso_insert_state(struct cso_cache *sc,
                  unsigned hash_key, enum cso_cache_type type,
                  void *state)
 {
    struct cso_hash *hash = _cso_hash_for_type(sc, type);
+   sanitize_hash(hash, type, sc->max_size);
+
    return cso_hash_insert(hash, hash_key, state);
 }
 
@@ -132,6 +239,26 @@ cso_find_state(struct cso_cache *sc,
    return cso_hash_find(hash, hash_key);
 }
 
+
+void *cso_hash_find_data_from_template( struct cso_hash *hash,
+				        unsigned hash_key, 
+				        void *templ,
+				        int size )
+{
+   struct cso_hash_iter iter = cso_hash_find(hash, hash_key);
+   while (!cso_hash_iter_is_null(iter)) {
+      void *iter_data = cso_hash_iter_data(iter);
+      if (!memcmp(iter_data, templ, size)) {
+	 /* Return the payload: 
+	  */
+         return (unsigned char *)iter_data + size;
+      }
+      iter = cso_hash_iter_next(iter);
+   }
+   return NULL;
+}
+
+
 struct cso_hash_iter cso_find_state_template(struct cso_cache *sc,
                                              unsigned hash_key, enum cso_cache_type type,
                                              void *templ)
@@ -156,8 +283,9 @@ void * cso_take_state(struct cso_cache *sc,
 
 struct cso_cache *cso_cache_create(void)
 {
-   struct cso_cache *sc = malloc(sizeof(struct cso_cache));
+   struct cso_cache *sc = MALLOC_STRUCT(cso_cache);
 
+   sc->max_size           = 4096;
    sc->blend_hash         = cso_hash_create();
    sc->sampler_hash       = cso_hash_create();
    sc->depth_stencil_hash = cso_hash_create();
@@ -168,14 +296,78 @@ struct cso_cache *cso_cache_create(void)
    return sc;
 }
 
+void cso_for_each_state(struct cso_cache *sc, enum cso_cache_type type,
+                        cso_state_callback func, void *user_data)
+{
+   struct cso_hash *hash = 0;
+   struct cso_hash_iter iter;
+
+   switch (type) {
+   case CSO_BLEND:
+      hash = sc->blend_hash;
+      break;
+   case CSO_SAMPLER:
+      hash = sc->sampler_hash;
+      break;
+   case CSO_DEPTH_STENCIL_ALPHA:
+      hash = sc->depth_stencil_hash;
+      break;
+   case CSO_RASTERIZER:
+      hash = sc->rasterizer_hash;
+      break;
+   case CSO_FRAGMENT_SHADER:
+      hash = sc->fs_hash;
+      break;
+   case CSO_VERTEX_SHADER:
+      hash = sc->vs_hash;
+      break;
+   }
+
+   iter = cso_hash_first_node(hash);
+   while (!cso_hash_iter_is_null(iter)) {
+      void *state = cso_hash_iter_data(iter);
+      if (state) {
+         func(state, user_data);
+      }
+      iter = cso_hash_iter_next(iter);
+   }
+}
+
 void cso_cache_delete(struct cso_cache *sc)
 {
    assert(sc);
+   /* delete driver data */
+   cso_for_each_state(sc, CSO_BLEND, delete_blend_state, 0);
+   cso_for_each_state(sc, CSO_DEPTH_STENCIL_ALPHA, delete_depth_stencil_state, 0);
+   cso_for_each_state(sc, CSO_FRAGMENT_SHADER, delete_fs_state, 0);
+   cso_for_each_state(sc, CSO_VERTEX_SHADER, delete_vs_state, 0);
+   cso_for_each_state(sc, CSO_RASTERIZER, delete_rasterizer_state, 0);
+   cso_for_each_state(sc, CSO_SAMPLER, delete_sampler_state, 0);
+
    cso_hash_delete(sc->blend_hash);
    cso_hash_delete(sc->sampler_hash);
    cso_hash_delete(sc->depth_stencil_hash);
    cso_hash_delete(sc->rasterizer_hash);
    cso_hash_delete(sc->fs_hash);
    cso_hash_delete(sc->vs_hash);
-   free(sc);
+   FREE(sc);
 }
+
+void cso_set_maximum_cache_size(struct cso_cache *sc, int number)
+{
+   sc->max_size = number;
+
+   sanitize_hash(sc->blend_hash, CSO_BLEND, sc->max_size);
+   sanitize_hash(sc->depth_stencil_hash, CSO_DEPTH_STENCIL_ALPHA,
+                 sc->max_size);
+   sanitize_hash(sc->fs_hash, CSO_FRAGMENT_SHADER, sc->max_size);
+   sanitize_hash(sc->vs_hash, CSO_VERTEX_SHADER, sc->max_size);
+   sanitize_hash(sc->rasterizer_hash, CSO_RASTERIZER, sc->max_size);
+   sanitize_hash(sc->sampler_hash, CSO_SAMPLER, sc->max_size);
+}
+
+int cso_maximum_cache_size(const struct cso_cache *sc)
+{
+   return sc->max_size;
+}
+
diff --git a/src/gallium/auxiliary/cso_cache/cso_cache.h b/src/gallium/auxiliary/cso_cache/cso_cache.h
index 116e2eaa2c..3b0fe100b8 100644
--- a/src/gallium/auxiliary/cso_cache/cso_cache.h
+++ b/src/gallium/auxiliary/cso_cache/cso_cache.h
@@ -1,6 +1,6 @@
 /**************************************************************************
  *
- * Copyright 2007 Tungsten Graphics, Inc., Cedar Park, Texas.
+ * Copyright 2007-2008 Tungsten Graphics, Inc., Cedar Park, Texas.
  * All Rights Reserved.
  *
  * Permission is hereby granted, free of charge, to any person obtaining a
@@ -25,9 +25,48 @@
  *
  **************************************************************************/
 
- /*
-  * Authors:
-  *   Zack Rusin <zack@tungstengraphics.com>
+ /**
+  * @file
+  * Constant State Object (CSO) cache.
+  *
+  * The basic idea is that the states are created via the
+  * create_state/bind_state/delete_state semantics. The driver is expected to
+  * perform as much of the Gallium state translation to whatever its internal
+  * representation is during the create call. Gallium then has a caching
+  * mechanism where it stores the created states. When the pipeline needs an
+  * actual state change, a bind call is issued. In the bind call the driver
+  * gets its already translated representation.
+  *
+  * Those semantics mean that the driver doesn't do the repeated translations
+  * of states on every frame, but only once, when a new state is actually
+  * created.
+  *
+  * Even on hardware that doesn't do any kind of state cache, it makes the
+  * driver look a lot neater, plus it avoids all the redundant state
+  * translations on every frame.
+  *
+  * Currently our constant state objects are:
+  * - alpha test
+  * - blend
+  * - depth stencil
+  * - fragment shader
+  * - rasterizer (old setup)
+  * - sampler
+  * - vertex shader
+  *
+  * Things that are not constant state objects include:
+  * - blend_color
+  * - clip_state
+  * - clear_color_state
+  * - constant_buffer
+  * - feedback_state
+  * - framebuffer_state
+  * - polygon_stipple
+  * - scissor_state
+  * - texture_state
+  * - viewport_state
+  *
+  * @author Zack Rusin <zack@tungstengraphics.com>
   */
 
 #ifndef CSO_CACHE_H
@@ -36,46 +75,57 @@
 #include "pipe/p_context.h"
 #include "pipe/p_state.h"
 
+/* cso_hash.h is necessary for cso_hash_iter, as MSVC requires structures
+ * returned by value to be fully defined */
+#include "cso_hash.h"
 
-struct cso_hash;
 
-struct cso_cache {
-   struct cso_hash *blend_hash;
-   struct cso_hash *depth_stencil_hash;
-   struct cso_hash *fs_hash;
-   struct cso_hash *vs_hash;
-   struct cso_hash *rasterizer_hash;
-   struct cso_hash *sampler_hash;
-};
+#ifdef	__cplusplus
+extern "C" {
+#endif
+
+struct cso_cache;
 
 struct cso_blend {
    struct pipe_blend_state state;
    void   *data;
+   void   (*delete_state)(void *, void  *);
+   void   *context;
 };
 
 struct cso_depth_stencil_alpha {
    struct pipe_depth_stencil_alpha_state state;
    void *data;
+   void   (*delete_state)(void *, void  *);
+   void   *context;
 };
 
 struct cso_rasterizer {
    struct pipe_rasterizer_state state;
    void *data;
+   void   (*delete_state)(void *, void  *);
+   void   *context;
 };
 
 struct cso_fragment_shader {
    struct pipe_shader_state state;
    void *data;
+   void   (*delete_state)(void *, void  *);
+   void   *context;
 };
 
 struct cso_vertex_shader {
    struct pipe_shader_state state;
    void *data;
+   void   (*delete_state)(void *, void  *);
+   void   *context;
 };
 
 struct cso_sampler {
    struct pipe_sampler_state state;
    void *data;
+   void   (*delete_state)(void *, void  *);
+   void   *context;
 };
 
 
@@ -88,6 +138,8 @@ enum cso_cache_type {
    CSO_VERTEX_SHADER
 };
 
+typedef void (*cso_state_callback)(void *, void *);
+
 unsigned cso_construct_key(void *item, int item_size);
 
 struct cso_cache *cso_cache_create(void);
@@ -101,7 +153,16 @@ struct cso_hash_iter cso_find_state(struct cso_cache *sc,
 struct cso_hash_iter cso_find_state_template(struct cso_cache *sc,
                                              unsigned hash_key, enum cso_cache_type type,
                                              void *templ);
+void cso_for_each_state(struct cso_cache *sc, enum cso_cache_type type,
+                        cso_state_callback func, void *user_data);
 void * cso_take_state(struct cso_cache *sc, unsigned hash_key,
                       enum cso_cache_type type);
 
+void cso_set_maximum_cache_size(struct cso_cache *sc, int number);
+int cso_maximum_cache_size(const struct cso_cache *sc);
+
+#ifdef	__cplusplus
+}
+#endif
+
 #endif
diff --git a/src/gallium/auxiliary/cso_cache/cso_hash.c b/src/gallium/auxiliary/cso_cache/cso_hash.c
index b40217c524..b3b4d667d2 100644
--- a/src/gallium/auxiliary/cso_cache/cso_hash.c
+++ b/src/gallium/auxiliary/cso_cache/cso_hash.c
@@ -30,12 +30,10 @@
   *   Zack Rusin <zack@tungstengraphics.com>
   */
 
-#include "cso_hash.h"
+#include "pipe/p_debug.h"
+#include "pipe/p_util.h"
 
-#include <stdio.h>
-#include <stdlib.h>
-#include <string.h>
-#include <assert.h>
+#include "cso_hash.h"
 
 #define MAX(a, b) ((a > b) ? (a) : (b))
 
@@ -98,7 +96,7 @@ struct cso_hash {
 
 static void *cso_data_allocate_node(struct cso_hash_data *hash)
 {
-   return malloc(hash->nodeSize);
+   return MALLOC(hash->nodeSize);
 }
 
 static void cso_data_free_node(struct cso_node *node)
@@ -107,10 +105,10 @@ static void cso_data_free_node(struct cso_node *node)
     * Need to cast value ptr to original cso type, then free the
     * driver-specific data hanging off of it.  For example:
    struct cso_sampler *csamp = (struct cso_sampler *) node->value;
-   free(csamp->data);
+   FREE(csamp->data);
    */
-   free(node->value);
-   free(node);
+   FREE(node->value);
+   FREE(node);
 }
 
 static struct cso_node *
@@ -134,7 +132,7 @@ static void cso_data_rehash(struct cso_hash_data *hash, int hint)
       hint = countBits(-hint);
       if (hint < MinNumBits)
          hint = MinNumBits;
-      hash->userNumBits = hint;
+      hash->userNumBits = (short)hint;
       while (primeForNumBits(hint) < (hash->size >> 1))
          ++hint;
    } else if (hint < MinNumBits) {
@@ -147,9 +145,9 @@ static void cso_data_rehash(struct cso_hash_data *hash, int hint)
       int oldNumBuckets = hash->numBuckets;
       int  i = 0;
 
-      hash->numBits = hint;
+      hash->numBits = (short)hint;
       hash->numBuckets = primeForNumBits(hint);
-      hash->buckets = malloc(sizeof(struct cso_node*) * hash->numBuckets);
+      hash->buckets = MALLOC(sizeof(struct cso_node*) * hash->numBuckets);
       for (i = 0; i < hash->numBuckets; ++i)
          hash->buckets[i] = e;
 
@@ -173,7 +171,7 @@ static void cso_data_rehash(struct cso_hash_data *hash, int hint)
             firstNode = afterLastNode;
          }
       }
-      free(oldBuckets);
+      FREE(oldBuckets);
    }
 }
 
@@ -235,13 +233,13 @@ struct cso_hash_iter cso_hash_insert(struct cso_hash *hash,
 
 struct cso_hash * cso_hash_create(void)
 {
-   struct cso_hash *hash = malloc(sizeof(struct cso_hash));
-   hash->data.d = malloc(sizeof(struct cso_hash_data));
+   struct cso_hash *hash = MALLOC_STRUCT(cso_hash);
+   hash->data.d = MALLOC_STRUCT(cso_hash_data);
    hash->data.d->fakeNext = 0;
    hash->data.d->buckets = 0;
    hash->data.d->size = 0;
    hash->data.d->nodeSize = sizeof(struct cso_node);
-   hash->data.d->userNumBits = MinNumBits;
+   hash->data.d->userNumBits = (short)MinNumBits;
    hash->data.d->numBits = 0;
    hash->data.d->numBuckets = 0;
 
@@ -261,9 +259,9 @@ void cso_hash_delete(struct cso_hash *hash)
          cur = next;
       }
    }
-   free(hash->data.d->buckets);
-   free(hash->data.d);
-   free(hash);
+   FREE(hash->data.d->buckets);
+   FREE(hash->data.d);
+   FREE(hash);
 }
 
 struct cso_hash_iter cso_hash_find(struct cso_hash *hash,
@@ -301,7 +299,7 @@ static struct cso_node *cso_hash_data_next(struct cso_node *node)
 
    a.next = node->next;
    if (!a.next) {
-      fprintf(stderr, "iterating beyond the last element\n");
+      debug_printf("iterating beyond the last element\n");
       return 0;
    }
    if (a.next->next)
@@ -352,7 +350,7 @@ static struct cso_node *cso_hash_data_prev(struct cso_node *node)
       --bucket;
       --start;
    }
-   fprintf(stderr, "iterating backward beyond first element\n");
+   debug_printf("iterating backward beyond first element\n");
    return a.e;
 }
 
@@ -397,3 +395,8 @@ struct cso_hash_iter cso_hash_first_node(struct cso_hash *hash)
    struct cso_hash_iter iter = {hash, cso_data_first_node(hash->data.d)};
    return iter;
 }
+
+int cso_hash_size(struct cso_hash *hash)
+{
+   return hash->data.d->size;
+}
diff --git a/src/gallium/auxiliary/cso_cache/cso_hash.h b/src/gallium/auxiliary/cso_cache/cso_hash.h
index b4aa111860..d5bca9d591 100644
--- a/src/gallium/auxiliary/cso_cache/cso_hash.h
+++ b/src/gallium/auxiliary/cso_cache/cso_hash.h
@@ -33,6 +33,11 @@
 #ifndef CSO_HASH_H
 #define CSO_HASH_H
 
+
+#ifdef	__cplusplus
+extern "C" {
+#endif
+
 struct cso_hash;
 struct cso_node;
 
@@ -42,7 +47,9 @@ struct cso_hash_iter {
 };
 
 struct cso_hash *cso_hash_create(void);
-void              cso_hash_delete(struct cso_hash *hash);
+void             cso_hash_delete(struct cso_hash *hash);
+
+int              cso_hash_size(struct cso_hash *hash);
 
 struct cso_hash_iter cso_hash_insert(struct cso_hash *hash, unsigned key,
                                      void *data);
@@ -59,4 +66,17 @@ void     *cso_hash_iter_data(struct cso_hash_iter iter);
 struct cso_hash_iter cso_hash_iter_next(struct cso_hash_iter iter);
 struct cso_hash_iter cso_hash_iter_prev(struct cso_hash_iter iter);
 
+
+/* KW: a convenience routine: 
+ */
+void *cso_hash_find_data_from_template( struct cso_hash *hash,
+				        unsigned hash_key, 
+				        void *templ,
+				        int size );
+
+
+#ifdef	__cplusplus
+}
+#endif
+
 #endif