i965: Keep the CURBE BO mapped and memcpy instead of subdataing.

For the tiny bis of data we generally upload through the CURBEs, the
overhead of the kernel's pagetable trickery is actually rather high.
This improves cairo-gl gnome-terminal-vim performance by 3.8%.

3 files changed, 13 insertions, 11 deletions

diff --git a/src/mesa/drivers/dri/i965/brw_context.h b/src/mesa/drivers/dri/i965/brw_context.h
index 79818b92b75..21c4cd38a72 100644
--- a/src/mesa/drivers/dri/i965/brw_context.h
+++ b/src/mesa/drivers/dri/i965/brw_context.h
@@ -572,11 +572,6 @@ struct brw_context
 
       GLfloat *last_buf;
       GLuint last_bufsz;
-      /**
-       *  Whether we should create a new bo instead of reusing the old one
-       * (if we just dispatch the batch pointing at the old one.
-       */
-      GLboolean need_new_bo;
    } curbe;
 
    struct {
diff --git a/src/mesa/drivers/dri/i965/brw_curbe.c b/src/mesa/drivers/dri/i965/brw_curbe.c
index 3440523ea75..6cb8edb611f 100644
--- a/src/mesa/drivers/dri/i965/brw_curbe.c
+++ b/src/mesa/drivers/dri/i965/brw_curbe.c
@@ -179,6 +179,7 @@ static GLfloat fixed_plane[6][4] = {
  */
 static void prepare_constant_buffer(struct brw_context *brw)
 {
+   struct intel_context *intel = &brw->intel;
    GLcontext *ctx = &brw->intel.ctx;
    const struct brw_vertex_program *vp =
       brw_vertex_program_const(brw->vertex_program);
@@ -304,9 +305,9 @@ static void prepare_constant_buffer(struct brw_context *brw)
       brw->curbe.last_bufsz = bufsz;
 
       if (brw->curbe.curbe_bo != NULL &&
-	  (brw->curbe.need_new_bo ||
-	   brw->curbe.curbe_next_offset + bufsz > brw->curbe.curbe_bo->size))
+	  brw->curbe.curbe_next_offset + bufsz > brw->curbe.curbe_bo->size)
       {
+	 intel_bo_unmap_gtt_preferred(intel, brw->curbe.curbe_bo);
 	 dri_bo_unreference(brw->curbe.curbe_bo);
 	 brw->curbe.curbe_bo = NULL;
       }
@@ -318,7 +319,7 @@ static void prepare_constant_buffer(struct brw_context *brw)
 	 brw->curbe.curbe_bo = dri_bo_alloc(brw->intel.bufmgr, "CURBE",
 					    4096, 1 << 6);
 	 brw->curbe.curbe_next_offset = 0;
-	 brw->curbe.need_new_bo = GL_FALSE;
+	 intel_bo_map_gtt_preferred(intel, brw->curbe.curbe_bo, GL_TRUE);
       }
 
       brw->curbe.curbe_offset = brw->curbe.curbe_next_offset;
@@ -327,7 +328,9 @@ static void prepare_constant_buffer(struct brw_context *brw)
 
       /* Copy data to the buffer:
        */
-      dri_bo_subdata(brw->curbe.curbe_bo, brw->curbe.curbe_offset, bufsz, buf);
+      memcpy(brw->curbe.curbe_bo->virtual + brw->curbe.curbe_offset,
+	     buf,
+	     bufsz);
    }
 
    brw_add_validated_bo(brw, brw->curbe.curbe_bo);
diff --git a/src/mesa/drivers/dri/i965/brw_vtbl.c b/src/mesa/drivers/dri/i965/brw_vtbl.c
index e08f32c4e27..681319dedfd 100644
--- a/src/mesa/drivers/dri/i965/brw_vtbl.c
+++ b/src/mesa/drivers/dri/i965/brw_vtbl.c
@@ -139,6 +139,12 @@ static void brw_finish_batch(struct intel_context *intel)
 {
    struct brw_context *brw = brw_context(&intel->ctx);
    brw_emit_query_end(brw);
+
+   if (brw->curbe.curbe_bo) {
+      intel_bo_unmap_gtt_preferred(intel, brw->curbe.curbe_bo);
+      drm_intel_bo_unreference(brw->curbe.curbe_bo);
+      brw->curbe.curbe_bo = NULL;
+   }
 }
 
 
@@ -149,8 +155,6 @@ static void brw_new_batch( struct intel_context *intel )
 {
    struct brw_context *brw = brw_context(&intel->ctx);
 
-   brw->curbe.need_new_bo = GL_TRUE;
-
    /* 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.