[965] Improve performance by allocating CURBE buffers a page at a time.

Since each one is only 64b, and kernel allocations are a page anyway, this
lets us reduce buffer allocation by packing many CURBEs into one buffer, for
each batchbuffer submitted.  Improves openarena performance by around 10%.

3 files changed, 33 insertions, 9 deletions

diff --git a/src/mesa/drivers/dri/i965/brw_context.h b/src/mesa/drivers/dri/i965/brw_context.h
index 9ddd41d3f39..7a2073d7c16 100644
--- a/src/mesa/drivers/dri/i965/brw_context.h
+++ b/src/mesa/drivers/dri/i965/brw_context.h
@@ -538,6 +538,10 @@ struct brw_context
       struct brw_tracked_state tracked_state;
 
       dri_bo *curbe_bo;
+      /** Offset within curbe_bo of space for current curbe entry */
+      GLuint curbe_offset;
+      /** Offset within curbe_bo of space for next curbe entry */
+      GLuint curbe_next_offset;
 
       GLfloat *last_buf;
       GLuint last_bufsz;
diff --git a/src/mesa/drivers/dri/i965/brw_curbe.c b/src/mesa/drivers/dri/i965/brw_curbe.c
index 2e39ec48055..f41f659b331 100644
--- a/src/mesa/drivers/dri/i965/brw_curbe.c
+++ b/src/mesa/drivers/dri/i965/brw_curbe.c
@@ -286,7 +286,8 @@ static void upload_constant_buffer(struct brw_context *brw)
 		   brw->curbe.last_buf ? memcmp(buf, brw->curbe.last_buf, bufsz) : -1);
    }
 
-   if (brw->curbe.last_buf &&
+   if (brw->curbe.curbe_bo != NULL &&
+       brw->curbe.last_buf &&
        bufsz == brw->curbe.last_bufsz &&
        memcmp(buf, brw->curbe.last_buf, bufsz) == 0) {
       free(buf);
@@ -297,16 +298,32 @@ static void upload_constant_buffer(struct brw_context *brw)
       brw->curbe.last_buf = buf;
       brw->curbe.last_bufsz = bufsz;
 
-      dri_bo_unreference(brw->curbe.curbe_bo);
-      brw->curbe.curbe_bo = dri_bo_alloc(brw->intel.bufmgr, "CURBE",
-					 bufsz, 1 << 6,
-					 DRM_BO_FLAG_MEM_LOCAL |
-					 DRM_BO_FLAG_CACHED |
-					 DRM_BO_FLAG_CACHED_MAPPED);
+      if (brw->curbe.curbe_bo != NULL &&
+	  brw->curbe.curbe_next_offset + bufsz > brw->curbe.curbe_bo->size)
+      {
+	 dri_bo_unreference(brw->curbe.curbe_bo);
+	 brw->curbe.curbe_bo = NULL;
+      }
+
+      if (brw->curbe.curbe_bo == NULL) {
+	 /* Allocate a single page for CURBE entries for this batchbuffer.
+	  * They're generally around 64b.
+	  */
+	 brw->curbe.curbe_bo = dri_bo_alloc(brw->intel.bufmgr, "CURBE",
+					    4096, 1 << 6,
+					    DRM_BO_FLAG_MEM_LOCAL |
+					    DRM_BO_FLAG_CACHED |
+					    DRM_BO_FLAG_CACHED_MAPPED);
+	 brw->curbe.curbe_next_offset = 0;
+      }
+
+      brw->curbe.curbe_offset = brw->curbe.curbe_next_offset;
+      brw->curbe.curbe_next_offset += bufsz;
+      brw->curbe.curbe_next_offset = ALIGN(brw->curbe.curbe_next_offset, 64);
 
       /* Copy data to the buffer:
        */
-      dri_bo_subdata(brw->curbe.curbe_bo, 0, bufsz, buf);
+      dri_bo_subdata(brw->curbe.curbe_bo, brw->curbe.curbe_offset, bufsz, buf);
    }
 
    /* Because this provokes an action (ie copy the constants into the
@@ -325,7 +342,7 @@ static void upload_constant_buffer(struct brw_context *brw)
    BEGIN_BATCH(2, IGNORE_CLIPRECTS);
    OUT_BATCH((CMD_CONST_BUFFER << 16) | (1 << 8) | (2 - 2));
    OUT_RELOC(brw->curbe.curbe_bo, DRM_BO_FLAG_MEM_TT | DRM_BO_FLAG_READ,
-	     (sz - 1));
+	     (sz - 1) + brw->curbe.curbe_offset);
    ADVANCE_BATCH();
 }
 
diff --git a/src/mesa/drivers/dri/i965/brw_vtbl.c b/src/mesa/drivers/dri/i965/brw_vtbl.c
index e9fed4dae10..126e6558395 100644
--- a/src/mesa/drivers/dri/i965/brw_vtbl.c
+++ b/src/mesa/drivers/dri/i965/brw_vtbl.c
@@ -94,6 +94,9 @@ static void brw_new_batch( struct intel_context *intel )
    /* Check that we didn't just wrap our batchbuffer at a bad time. */
    assert(!brw->no_batch_wrap);
 
+   dri_bo_unreference(brw->curbe.curbe_bo);
+   brw->curbe.curbe_bo = NULL;
+
    /* Mark all context state as needing to be re-emitted.
     * This is probably not as severe as on 915, since almost all of our state
     * is just in referenced buffers.