radeon/r200/r300: Fix swtcl prediction to work after primitie change.

Swtcl calls flush everytime primitive changes so prediction has to made again
after flushing.

4 files changed, 86 insertions, 41 deletions

diff --git a/src/mesa/drivers/dri/r200/r200_swtcl.c b/src/mesa/drivers/dri/r200/r200_swtcl.c
index 6b7279e8db5..3d4e7011557 100644
--- a/src/mesa/drivers/dri/r200/r200_swtcl.c
+++ b/src/mesa/drivers/dri/r200/r200_swtcl.c
@@ -201,27 +201,35 @@ static void r200SetVertexFormat( GLcontext *ctx )
    }
 }
 
-
-static void r200RenderStart( GLcontext *ctx )
+static void r200_predict_emit_size( GLcontext *ctx )
 {
+   r200ContextPtr rmesa = R200_CONTEXT( ctx );
    const int vertex_array_size = 7;
    const int prim_size = 3;
-   r200ContextPtr rmesa = R200_CONTEXT( ctx );
-   r200SetVertexFormat( ctx );
-   if (RADEON_DEBUG & DEBUG_VERTS)
-      fprintf(stderr, "%s\n", __func__);
-   if (!rmesa->radeon.swtcl.primitive_counter) {
+   if (!rmesa->radeon.swtcl.emit_prediction) {
+      const int state_size = radeonCountStateEmitSize(&rmesa->radeon);
       if (rcommonEnsureCmdBufSpace(&rmesa->radeon,
-	       radeonCountStateEmitSize(&rmesa->radeon) +
+	       state_size +
 	       vertex_array_size + prim_size,
 	       __FUNCTION__))
-	 rmesa->radeon.swtcl.primitive_counter = 0;
+	 rmesa->radeon.swtcl.emit_prediction = radeonCountStateEmitSize(&rmesa->radeon);
       else
-	 rmesa->radeon.swtcl.primitive_counter = 1;
+	 rmesa->radeon.swtcl.emit_prediction = state_size;
+      rmesa->radeon.swtcl.emit_prediction += vertex_array_size + prim_size
+	 + rmesa->radeon.cmdbuf.cs->cdw;
    }
 }
 
 
+static void r200RenderStart( GLcontext *ctx )
+{
+   r200SetVertexFormat( ctx );
+   if (RADEON_DEBUG & DEBUG_VERTS)
+      fprintf(stderr, "%s\n", __func__);
+   r200_predict_emit_size( ctx );
+}
+
+
 /**
  * Set vertex state for SW TCL.  The primary purpose of this function is to
  * determine in advance whether or not the hardware can / should do the
@@ -296,8 +304,13 @@ void r200_swtcl_flush(GLcontext *ctx, uint32_t current_offset)
    r200EmitVbufPrim( rmesa,
 		     rmesa->radeon.swtcl.hw_primitive,
 		     rmesa->radeon.swtcl.numverts);
+   if ( rmesa->radeon.swtcl.emit_prediction < rmesa->radeon.cmdbuf.cs->cdw )
+      WARN_ONCE("Rendering was %d commands larger than predicted size."
+	    " We might overflow  command buffer.\n",
+	    rmesa->radeon.cmdbuf.cs->cdw - rmesa->radeon.swtcl.emit_prediction );
 
-   rmesa->radeon.swtcl.primitive_counter = 0;
+   rmesa->radeon.swtcl.emit_prediction = 0;
+   r200_predict_emit_size( ctx );
 
 }
 
@@ -905,7 +918,7 @@ void r200InitSwtcl( GLcontext *ctx )
       init_rast_tab();
       firsttime = 0;
    }
-   rmesa->radeon.swtcl.primitive_counter = 0;
+   rmesa->radeon.swtcl.emit_prediction = 0;
 
    tnl->Driver.Render.Start = r200RenderStart;
    tnl->Driver.Render.Finish = r200RenderFinish;
diff --git a/src/mesa/drivers/dri/r300/r300_swtcl.c b/src/mesa/drivers/dri/r300/r300_swtcl.c
index 18af51e666a..ca17f306d14 100644
--- a/src/mesa/drivers/dri/r300/r300_swtcl.c
+++ b/src/mesa/drivers/dri/r300/r300_swtcl.c
@@ -490,6 +490,27 @@ static void r300ChooseRenderState( GLcontext *ctx )
 		rmesa->radeon.swtcl.RenderIndex = index;
 	}
 }
+static void r300_predict_emit_size( GLcontext *ctx )
+{
+	r300ContextPtr rmesa = R300_CONTEXT( ctx );
+	if (!rmesa->radeon.swtcl.emit_prediction) {
+		const int vertex_size = 7;
+		const int prim_size = 3;
+		const int cache_flush_size = 4;
+		const int state_size = radeonCountStateEmitSize(&rmesa->radeon);
+
+		if (rcommonEnsureCmdBufSpace(&rmesa->radeon,
+					state_size +
+					+ vertex_size + prim_size,
+					__FUNCTION__))
+			rmesa->radeon.swtcl.emit_prediction = radeonCountStateEmitSize(&rmesa->radeon);
+		else
+			rmesa->radeon.swtcl.emit_prediction = state_size;
+
+		rmesa->radeon.swtcl.emit_prediction += rmesa->radeon.cmdbuf.cs->cdw
+			+ vertex_size + prim_size + cache_flush_size * 2;
+	}
+}
 
 
 void r300RenderStart(GLcontext *ctx)
@@ -508,20 +529,7 @@ void r300RenderStart(GLcontext *ctx)
 
 	r300UpdateShaderStates(rmesa);
 
-	const int vertex_size = 7;
-	const int prim_size = 3;
-
-	if (!rmesa->radeon.swtcl.primitive_counter) {
-		if (rcommonEnsureCmdBufSpace(&rmesa->radeon,
-					radeonCountStateEmitSize(&rmesa->radeon) +
-					+ vertex_size + prim_size,
-					__FUNCTION__))
-			rmesa->radeon.swtcl.primitive_counter = 0;
-		else
-			rmesa->radeon.swtcl.primitive_counter = 1;
-	}
-
-	r300EmitCacheFlush(rmesa);
+	r300_predict_emit_size( ctx );
 
 	/* investigate if we can put back flush optimisation if needed */
 	if (rmesa->radeon.dma.flush != NULL) {
@@ -577,7 +585,7 @@ void r300InitSwtcl(GLcontext *ctx)
 		init_rast_tab();
 		firsttime = 0;
 	}
-	rmesa->radeon.swtcl.primitive_counter = 0;
+	rmesa->radeon.swtcl.emit_prediction = 0;
 
 	tnl->Driver.Render.Start = r300RenderStart;
 	tnl->Driver.Render.Finish = r300RenderFinish;
@@ -644,6 +652,8 @@ void r300_swtcl_flush(GLcontext *ctx, uint32_t current_offset)
 		fprintf(stderr, "%s\n", __func__);
 	r300ContextPtr rmesa = R300_CONTEXT(ctx);
 
+	r300EmitCacheFlush(rmesa);
+
 	radeonEmitState(&rmesa->radeon);
     r300_emit_scissor(ctx);
 	r300EmitVertexAOS(rmesa,
@@ -655,6 +665,11 @@ void r300_swtcl_flush(GLcontext *ctx, uint32_t current_offset)
 		   rmesa->radeon.swtcl.hw_primitive,
 		   rmesa->radeon.swtcl.numverts);
 	r300EmitCacheFlush(rmesa);
-	rmesa->radeon.swtcl.primitive_counter = 0;
+	if ( rmesa->radeon.swtcl.emit_prediction < rmesa->radeon.cmdbuf.cs->cdw )
+		WARN_ONCE("Rendering was %d commands larger than predicted size."
+			" We might overflow  command buffer.\n",
+			rmesa->radeon.cmdbuf.cs->cdw - rmesa->radeon.swtcl.emit_prediction );
+	rmesa->radeon.swtcl.emit_prediction = 0;
+	r300_predict_emit_size( ctx );
 	COMMIT_BATCH();
 }
diff --git a/src/mesa/drivers/dri/radeon/radeon_common_context.h b/src/mesa/drivers/dri/radeon/radeon_common_context.h
index 07a0c4c0554..427eb946ffa 100644
--- a/src/mesa/drivers/dri/radeon/radeon_common_context.h
+++ b/src/mesa/drivers/dri/radeon/radeon_common_context.h
@@ -342,7 +342,7 @@ struct radeon_swtcl_info {
 	struct tnl_attr_map vertex_attrs[VERT_ATTRIB_MAX];
 	GLuint vertex_attr_count;
 
-	GLuint primitive_counter;
+	GLuint emit_prediction;
 };
 
 #define RADEON_MAX_AOS_ARRAYS		16
diff --git a/src/mesa/drivers/dri/radeon/radeon_swtcl.c b/src/mesa/drivers/dri/radeon/radeon_swtcl.c
index c9e399f9b61..7358e22255b 100644
--- a/src/mesa/drivers/dri/radeon/radeon_swtcl.c
+++ b/src/mesa/drivers/dri/radeon/radeon_swtcl.c
@@ -220,6 +220,27 @@ static void radeonSetVertexFormat( GLcontext *ctx )
    }
 }
 
+static void radeon_predict_emit_size( GLcontext* ctx )
+{
+    r100ContextPtr rmesa = R100_CONTEXT( ctx );
+
+    if (!rmesa->radeon.swtcl.emit_prediction) {
+        const int state_size = radeonCountStateEmitSize( &rmesa->radeon );
+        const int scissor_size = 8;
+        const int prims_size = 8;
+        const int vertex_size = 7;
+
+        if (rcommonEnsureCmdBufSpace(&rmesa->radeon,
+                    state_size +
+                    (scissor_size + prims_size + vertex_size),
+                    __FUNCTION__))
+            rmesa->radeon.swtcl.emit_prediction = radeonCountStateEmitSize( &rmesa->radeon );
+        else
+            rmesa->radeon.swtcl.emit_prediction = state_size;
+        rmesa->radeon.swtcl.emit_prediction += scissor_size + prims_size + vertex_size
+            + rmesa->radeon.cmdbuf.cs->cdw;
+    }
+}
 
 static void radeonRenderStart( GLcontext *ctx )
 {
@@ -230,16 +251,7 @@ static void radeonRenderStart( GLcontext *ctx )
     if (rmesa->radeon.dma.flush != 0 &&
             rmesa->radeon.dma.flush != rcommon_flush_last_swtcl_prim)
         rmesa->radeon.dma.flush( ctx );
-
-    if (!rmesa->radeon.swtcl.primitive_counter) {
-        if (rcommonEnsureCmdBufSpace(&rmesa->radeon,
-                    radeonCountStateEmitSize( &rmesa->radeon ) +
-                    (8 + 8 + 7), /* scissor + primis + VertexAOS */
-                    __FUNCTION__))
-            rmesa->radeon.swtcl.primitive_counter = 0;
-        else
-            rmesa->radeon.swtcl.primitive_counter = 1;
-    }
+    radeon_predict_emit_size( ctx );
 }
 
 
@@ -307,9 +319,14 @@ void r100_swtcl_flush(GLcontext *ctx, uint32_t current_offset)
 		       rmesa->swtcl.vertex_format,
 		       rmesa->radeon.swtcl.hw_primitive,
 		       rmesa->radeon.swtcl.numverts);
+   if ( rmesa->radeon.swtcl.emit_prediction < rmesa->radeon.cmdbuf.cs->cdw )
+     WARN_ONCE("Rendering was %d commands larger than predicted size."
+	 " We might overflow  command buffer.\n",
+	 rmesa->radeon.cmdbuf.cs->cdw - rmesa->radeon.swtcl.emit_prediction );
 
+   radeon_predict_emit_size( ctx );
 
-   rmesa->radeon.swtcl.primitive_counter = 0;
+   rmesa->radeon.swtcl.emit_prediction = 0;
 
 }
 
@@ -814,8 +831,8 @@ void radeonInitSwtcl( GLcontext *ctx )
    if (firsttime) {
       init_rast_tab();
       firsttime = 0;
-      rmesa->radeon.swtcl.primitive_counter = 0;
    }
+   rmesa->radeon.swtcl.emit_prediction = 0;
 
    tnl->Driver.Render.Start = radeonRenderStart;
    tnl->Driver.Render.Finish = radeonRenderFinish;