1 files changed, 51 insertions, 8 deletions

diff --git a/src/mesa/drivers/dri/r600/r700_render.c b/src/mesa/drivers/dri/r600/r700_render.c
index cdf3220a7f..d64e921bda 100644
--- a/src/mesa/drivers/dri/r600/r700_render.c
+++ b/src/mesa/drivers/dri/r600/r700_render.c
@@ -53,6 +53,8 @@
 #include "r700_fragprog.h"
 #include "r700_state.h"
 
+#include "radeon_common_context.h"
+
 void r700WaitForIdle(context_t *context);
 void r700WaitForIdleClean(context_t *context);
 GLboolean r700SendTextureState(context_t *context);
@@ -298,22 +300,52 @@ static void r700RunRenderPrimitive(GLcontext * ctx, int start, int end, int prim
 
 }
 
+/* start 3d, idle, cb/db flush */
+#define PRE_EMIT_STATE_BUFSZ 10 + 5 + 14
+
+static GLuint r700PredictRenderSize(GLcontext* ctx)
+{
+    context_t *context = R700_CONTEXT(ctx);
+    TNLcontext *tnl = TNL_CONTEXT(ctx);
+    struct r700_vertex_program *vpc
+        = (struct r700_vertex_program *)ctx->VertexProgram._Current;
+    struct vertex_buffer *vb = &tnl->vb;
+    GLboolean flushed;
+    GLuint dwords, i;
+    GLuint state_size;
+    /* pre calculate aos count so state prediction works */
+    context->radeon.tcl.aos_count = _mesa_bitcount(vpc->mesa_program.Base.InputsRead);
+
+    dwords = PRE_EMIT_STATE_BUFSZ;
+    for (i = 0; i < vb->PrimitiveCount; i++)
+        dwords += vb->Primitive[i].count + 10;
+    state_size = radeonCountStateEmitSize(&context->radeon);
+    flushed = rcommonEnsureCmdBufSpace(&context->radeon,
+            dwords + state_size, __FUNCTION__);
+
+    if (flushed)
+        dwords += radeonCountStateEmitSize(&context->radeon);
+    else
+        dwords += state_size;
+
+    if (RADEON_DEBUG & DEBUG_PRIMS)
+        fprintf(stderr, "%s: total prediction size is %d.\n", __FUNCTION__, dwords);
+    return dwords;
+}
+
 static GLboolean r700RunRender(GLcontext * ctx,
 			       struct tnl_pipeline_stage *stage)
 {
     context_t *context = R700_CONTEXT(ctx);
     radeonContextPtr radeon = &context->radeon;
-    unsigned int i, ind_count = 0, id = 0;
+    unsigned int i, id = 0;
     TNLcontext *tnl = TNL_CONTEXT(ctx);
     struct vertex_buffer *vb = &tnl->vb;
     struct radeon_renderbuffer *rrb;
 
-    for (i = 0; i < vb->PrimitiveCount; i++)
-	    ind_count += vb->Primitive[i].count + 10;
-
-    /* just an estimate, need to properly calculate this */
-    rcommonEnsureCmdBufSpace(&context->radeon,
-			     radeon->hw.max_state_size + ind_count, __FUNCTION__);
+    if (RADEON_DEBUG & DEBUG_PRIMS)
+        fprintf(stderr, "%s: cs begin at %d\n",
+                __func__, context->radeon.cmdbuf.cs->cdw);
 
     /* always emit CB base to prevent
      * lock ups on some chips.
@@ -327,6 +359,9 @@ static GLboolean r700RunRender(GLcontext * ctx,
     r700SetupVertexProgram(ctx);
     r700SetupFragmentProgram(ctx);
     r600UpdateTextureState(ctx);
+
+    GLuint emit_end = r700PredictRenderSize(ctx) 
+        + context->radeon.cmdbuf.cs->cdw;
     r700SetupStreams(ctx);
 
     radeonEmitState(radeon);
@@ -336,7 +371,7 @@ static GLboolean r700RunRender(GLcontext * ctx,
         GLuint prim = _tnl_translate_prim(&vb->Primitive[i]);
         GLuint start = vb->Primitive[i].start;
         GLuint end = vb->Primitive[i].start + vb->Primitive[i].count;
-	r700RunRenderPrimitive(ctx, start, end, prim);
+        r700RunRenderPrimitive(ctx, start, end, prim);
     }
 
     /* Flush render op cached for last several quads. */
@@ -354,6 +389,14 @@ static GLboolean r700RunRender(GLcontext * ctx,
 
     radeonReleaseArrays(ctx, ~0);
 
+    if (RADEON_DEBUG & DEBUG_PRIMS)
+        fprintf(stderr, "%s: cs end at %d\n",
+                __func__, context->radeon.cmdbuf.cs->cdw);
+
+    if ( emit_end < context->radeon.cmdbuf.cs->cdw )
+       WARN_ONCE("Rendering was %d commands larger than predicted size."
+	       " We might overflow  command buffer.\n", context->radeon.cmdbuf.cs->cdw - emit_end);
+
     return GL_FALSE;
 }