i965: Reuse existing program data when a new compiled program matches.

It's common in applications just before the advent of
EXT_separate_shader_objects to have multiple linked shaders with the
same VS or FS.  While we aren't detecting those at the Mesa level, we
can detect when our compiled output happens to match an existing
compiled program.

This patch was created after noting the incredible amount of compiled
program data generated by Heroes of Newerth.  It reduces the program
data in use at the start menu (replayed by apitrace) from 828kb to
632kb, and reduces CACHE_NEW_WM_PROG state flagging by 3/4.  It
doesn't impact our rate of hardware state changes yet, because things
depending on CACHE_NEW_WM_PROG also depend on BRW_NEW_FRAGMENT_PROGRAM
which is still being flagged.

Reviewed-by: Ian Romanick <ian.d.romanick@intel.com>

diff --git a/src/mesa/drivers/dri/i965/brw_context.h b/src/mesa/drivers/dri/i965/brw_context.h
index 16b71f6b1c986144e992615661a3695ccf3e094a..559dc6886524af5247e3198026bc3f8d64ba5551 100644 (file)
--- a/src/mesa/drivers/dri/i965/brw_context.h
+++ b/src/mesa/drivers/dri/i965/brw_context.h
@@ -365,6 +365,7 @@ struct brw_cache_item {
    /** 32-bit hash of the key data */
    GLuint hash;
    GLuint key_size;            /* for variable-sized keys */
+   GLuint aux_size;
    const void *key;
 
    uint32_t offset;

diff --git a/src/mesa/drivers/dri/i965/brw_state_cache.c b/src/mesa/drivers/dri/i965/brw_state_cache.c
index d13711b19b7825ea360d2e6a3a7a355e5103174c..3988625ea9189466bc75c86d7f3e31fafe5e9a4b 100644 (file)
--- a/src/mesa/drivers/dri/i965/brw_state_cache.c
+++ b/src/mesa/drivers/dri/i965/brw_state_cache.c
@@ -187,6 +187,77 @@ brw_cache_new_bo(struct brw_cache *cache, uint32_t new_size)
    brw->state.dirty.brw |= BRW_NEW_PROGRAM_CACHE;
 }
 
+/**
+ * Attempts to find an item in the cache with identical data and aux
+ * data to use
+ */
+static bool
+brw_try_upload_using_copy(struct brw_cache *cache,
+                         struct brw_cache_item *result_item,
+                         const void *data,
+                         const void *aux)
+{
+   int i;
+   struct brw_cache_item *item;
+
+   for (i = 0; i < cache->size; i++) {
+      for (item = cache->items[i]; item; item = item->next) {
+        const void *item_aux = item->key + item->key_size;
+        int ret;
+
+        if (item->cache_id != result_item->cache_id ||
+            item->size != result_item->size ||
+            item->aux_size != result_item->aux_size) {
+           continue;
+        }
+
+        if (memcmp(item_aux, aux, item->aux_size) != 0) {
+           continue;
+        }
+
+        drm_intel_bo_map(cache->bo, false);
+        ret = memcmp(cache->bo->virtual + item->offset, data, item->size);
+        drm_intel_bo_unmap(cache->bo);
+        if (ret)
+           continue;
+
+        result_item->offset = item->offset;
+
+        return true;
+      }
+   }
+
+   return false;
+}
+
+static void
+brw_upload_item_data(struct brw_cache *cache,
+                    struct brw_cache_item *item,
+                    const void *data)
+{
+   /* Allocate space in the cache BO for our new program. */
+   if (cache->next_offset + item->size > cache->bo->size) {
+      uint32_t new_size = cache->bo->size * 2;
+
+      while (cache->next_offset + item->size > new_size)
+        new_size *= 2;
+
+      brw_cache_new_bo(cache, new_size);
+   }
+
+   /* If we would block on writing to an in-use program BO, just
+    * recreate it.
+    */
+   if (cache->bo_used_by_gpu) {
+      brw_cache_new_bo(cache, cache->bo->size);
+   }
+
+   item->offset = cache->next_offset;
+
+   /* Programs are always 64-byte aligned, so set up the next one now */
+   cache->next_offset = ALIGN(item->offset + item->size, 64);
+}
+
 void
 brw_upload_cache(struct brw_cache *cache,
                 enum brw_cache_id cache_id,
@@ -204,34 +275,24 @@ brw_upload_cache(struct brw_cache *cache,
    void *tmp;
 
    item->cache_id = cache_id;
+   item->size = data_size;
    item->key = key;
    item->key_size = key_size;
+   item->aux_size = aux_size;
    hash = hash_key(item);
    item->hash = hash;
 
-   /* Allocate space in the cache BO for our new program. */
-   if (cache->next_offset + data_size > cache->bo->size) {
-      uint32_t new_size = cache->bo->size * 2;
-
-      while (cache->next_offset + data_size > new_size)
-        new_size *= 2;
-
-      brw_cache_new_bo(cache, new_size);
-   }
-
-   /* If we would block on writing to an in-use program BO, just
-    * recreate it.
+   /* If we can find a matching prog/prog_data combo in the cache
+    * already, then reuse the existing stuff.  This will mean not
+    * flagging CACHE_NEW_* when transitioning between the two
+    * equivalent hash keys.  This is notably useful for programs
+    * generating shaders at runtime, where multiple shaders may
+    * compile to the thing in our backend.
     */
-   if (cache->bo_used_by_gpu) {
-      brw_cache_new_bo(cache, cache->bo->size);
+   if (!brw_try_upload_using_copy(cache, item, data, aux)) {
+      brw_upload_item_data(cache, item, data);
    }
 
-   item->offset = cache->next_offset;
-   item->size = data_size;
-
-   /* Programs are always 64-byte aligned, so set up the next one now */
-   cache->next_offset = ALIGN(item->offset + data_size, 64);
-
    /* Set up the memory containing the key and aux_data */
    tmp = malloc(key_size + aux_size);