R300+: pre-load vertex programs in init3D()

3 files changed, 232 insertions, 163 deletions

diff --git a/src/radeon_commonfuncs.c b/src/radeon_commonfuncs.c
index 5bc502c6..025fb564 100644
--- a/src/radeon_commonfuncs.c
+++ b/src/radeon_commonfuncs.c
@@ -144,8 +144,7 @@ static void FUNC_NAME(RADEONInit3DEngine)(ScrnInfoPtr pScrn)
 	OUT_ACCEL_REG(R300_SU_DEPTH_OFFSET, 0);
 	FINISH_ACCEL();
 
-	/* setup the VAP */
-	
+	/* setup the VAP */	
 	if (info->has_tcl) {
 	    BEGIN_ACCEL(11);
 	    OUT_ACCEL_REG(R300_VAP_CNTL_STATUS, 0);
@@ -171,6 +170,224 @@ static void FUNC_NAME(RADEONInit3DEngine)(ScrnInfoPtr pScrn)
 	}
 	FINISH_ACCEL();
 
+	/* pre-load the vertex shaders */
+	if (info->has_tcl) {
+	    /* exa mask shader program */
+	    BEGIN_ACCEL(13);
+	    OUT_ACCEL_REG(R300_VAP_PVS_VECTOR_INDX_REG, 0);
+	    /* PVS inst 0 */
+	    OUT_ACCEL_REG(R300_VAP_PVS_VECTOR_DATA_REG,
+			  (R300_PVS_DST_OPCODE(R300_VE_ADD) |
+			   R300_PVS_DST_REG_TYPE(R300_PVS_DST_REG_OUT) |
+			   R300_PVS_DST_OFFSET(0) |
+			   R300_PVS_DST_WE_X | R300_PVS_DST_WE_Y |
+			   R300_PVS_DST_WE_Z | R300_PVS_DST_WE_W));
+	    OUT_ACCEL_REG(R300_VAP_PVS_VECTOR_DATA_REG,
+			  (R300_PVS_SRC_REG_TYPE(R300_PVS_SRC_REG_INPUT) |
+			   R300_PVS_SRC_OFFSET(0) |
+			   R300_PVS_SRC_SWIZZLE_X(R300_PVS_SRC_SELECT_X) |
+			   R300_PVS_SRC_SWIZZLE_Y(R300_PVS_SRC_SELECT_Y) |
+			   R300_PVS_SRC_SWIZZLE_Z(R300_PVS_SRC_SELECT_Z) |
+			   R300_PVS_SRC_SWIZZLE_W(R300_PVS_SRC_SELECT_W)));
+	    OUT_ACCEL_REG(R300_VAP_PVS_VECTOR_DATA_REG,
+			  (R300_PVS_SRC_REG_TYPE(R300_PVS_SRC_REG_INPUT) |
+			   R300_PVS_SRC_OFFSET(0) |
+			   R300_PVS_SRC_SWIZZLE_X(R300_PVS_SRC_SELECT_FORCE_0) |
+			   R300_PVS_SRC_SWIZZLE_Y(R300_PVS_SRC_SELECT_FORCE_0) |
+			   R300_PVS_SRC_SWIZZLE_Z(R300_PVS_SRC_SELECT_FORCE_0) |
+			   R300_PVS_SRC_SWIZZLE_W(R300_PVS_SRC_SELECT_FORCE_0)));
+	    OUT_ACCEL_REG(R300_VAP_PVS_VECTOR_DATA_REG,
+			  (R300_PVS_SRC_REG_TYPE(R300_PVS_SRC_REG_INPUT) |
+			   R300_PVS_SRC_OFFSET(0) |
+			   R300_PVS_SRC_SWIZZLE_X(R300_PVS_SRC_SELECT_FORCE_0) |
+			   R300_PVS_SRC_SWIZZLE_Y(R300_PVS_SRC_SELECT_FORCE_0) |
+			   R300_PVS_SRC_SWIZZLE_Z(R300_PVS_SRC_SELECT_FORCE_0) |
+			   R300_PVS_SRC_SWIZZLE_W(R300_PVS_SRC_SELECT_FORCE_0)));
+
+	    /* PVS inst 1 */
+	    OUT_ACCEL_REG(R300_VAP_PVS_VECTOR_DATA_REG,
+			  (R300_PVS_DST_OPCODE(R300_VE_ADD) |
+			   R300_PVS_DST_REG_TYPE(R300_PVS_DST_REG_OUT) |
+			   R300_PVS_DST_OFFSET(1) |
+			   R300_PVS_DST_WE_X | R300_PVS_DST_WE_Y |
+			   R300_PVS_DST_WE_Z | R300_PVS_DST_WE_W));
+	    OUT_ACCEL_REG(R300_VAP_PVS_VECTOR_DATA_REG,
+			  (R300_PVS_SRC_REG_TYPE(R300_PVS_SRC_REG_INPUT) |
+			   R300_PVS_SRC_OFFSET(1) |
+			   R300_PVS_SRC_SWIZZLE_X(R300_PVS_SRC_SELECT_X) |
+			   R300_PVS_SRC_SWIZZLE_Y(R300_PVS_SRC_SELECT_Y) |
+			   R300_PVS_SRC_SWIZZLE_Z(R300_PVS_SRC_SELECT_Z) |
+			   R300_PVS_SRC_SWIZZLE_W(R300_PVS_SRC_SELECT_W)));
+	    OUT_ACCEL_REG(R300_VAP_PVS_VECTOR_DATA_REG,
+			  (R300_PVS_SRC_REG_TYPE(R300_PVS_SRC_REG_INPUT) |
+			   R300_PVS_SRC_OFFSET(1) |
+			   R300_PVS_SRC_SWIZZLE_X(R300_PVS_SRC_SELECT_FORCE_0) |
+			   R300_PVS_SRC_SWIZZLE_Y(R300_PVS_SRC_SELECT_FORCE_0) |
+			   R300_PVS_SRC_SWIZZLE_Z(R300_PVS_SRC_SELECT_FORCE_0) |
+			   R300_PVS_SRC_SWIZZLE_W(R300_PVS_SRC_SELECT_FORCE_0)));
+	    OUT_ACCEL_REG(R300_VAP_PVS_VECTOR_DATA_REG,
+			  (R300_PVS_SRC_REG_TYPE(R300_PVS_SRC_REG_INPUT) |
+			   R300_PVS_SRC_OFFSET(1) |
+			   R300_PVS_SRC_SWIZZLE_X(R300_PVS_SRC_SELECT_FORCE_0) |
+			   R300_PVS_SRC_SWIZZLE_Y(R300_PVS_SRC_SELECT_FORCE_0) |
+			   R300_PVS_SRC_SWIZZLE_Z(R300_PVS_SRC_SELECT_FORCE_0) |
+			   R300_PVS_SRC_SWIZZLE_W(R300_PVS_SRC_SELECT_FORCE_0)));
+
+	    /* PVS inst 2 */
+	    OUT_ACCEL_REG(R300_VAP_PVS_VECTOR_DATA_REG,
+			  (R300_PVS_DST_OPCODE(R300_VE_ADD) |
+			   R300_PVS_DST_REG_TYPE(R300_PVS_DST_REG_OUT) |
+			   R300_PVS_DST_OFFSET(2) |
+			   R300_PVS_DST_WE_X | R300_PVS_DST_WE_Y |
+			   R300_PVS_DST_WE_Z | R300_PVS_DST_WE_W));
+	    OUT_ACCEL_REG(R300_VAP_PVS_VECTOR_DATA_REG,
+			  (R300_PVS_SRC_REG_TYPE(R300_PVS_SRC_REG_INPUT) |
+			   R300_PVS_SRC_OFFSET(2) |
+			   R300_PVS_SRC_SWIZZLE_X(R300_PVS_SRC_SELECT_X) |
+			   R300_PVS_SRC_SWIZZLE_Y(R300_PVS_SRC_SELECT_Y) |
+			   R300_PVS_SRC_SWIZZLE_Z(R300_PVS_SRC_SELECT_Z) |
+			   R300_PVS_SRC_SWIZZLE_W(R300_PVS_SRC_SELECT_W)));
+	    OUT_ACCEL_REG(R300_VAP_PVS_VECTOR_DATA_REG,
+			  (R300_PVS_SRC_REG_TYPE(R300_PVS_SRC_REG_INPUT) |
+			   R300_PVS_SRC_OFFSET(2) |
+			   R300_PVS_SRC_SWIZZLE_X(R300_PVS_SRC_SELECT_FORCE_0) |
+			   R300_PVS_SRC_SWIZZLE_Y(R300_PVS_SRC_SELECT_FORCE_0) |
+			   R300_PVS_SRC_SWIZZLE_Z(R300_PVS_SRC_SELECT_FORCE_0) |
+			   R300_PVS_SRC_SWIZZLE_W(R300_PVS_SRC_SELECT_FORCE_0)));
+	    OUT_ACCEL_REG(R300_VAP_PVS_VECTOR_DATA_REG,
+			  (R300_PVS_SRC_REG_TYPE(R300_PVS_SRC_REG_INPUT) |
+			   R300_PVS_SRC_OFFSET(2) |
+			   R300_PVS_SRC_SWIZZLE_X(R300_PVS_SRC_SELECT_FORCE_0) |
+			   R300_PVS_SRC_SWIZZLE_Y(R300_PVS_SRC_SELECT_FORCE_0) |
+			   R300_PVS_SRC_SWIZZLE_Z(R300_PVS_SRC_SELECT_FORCE_0) |
+			   R300_PVS_SRC_SWIZZLE_W(R300_PVS_SRC_SELECT_FORCE_0)));
+	    FINISH_ACCEL();
+
+	    BEGIN_ACCEL(9);
+	    /* exa no mask instruction */
+	    OUT_ACCEL_REG(R300_VAP_PVS_VECTOR_INDX_REG, 3);
+	    /* PVS inst 0 */
+	    OUT_ACCEL_REG(R300_VAP_PVS_VECTOR_DATA_REG,
+			  (R300_PVS_DST_OPCODE(R300_VE_ADD) |
+			   R300_PVS_DST_REG_TYPE(R300_PVS_DST_REG_OUT) |
+			   R300_PVS_DST_OFFSET(0) |
+			   R300_PVS_DST_WE_X | R300_PVS_DST_WE_Y |
+			   R300_PVS_DST_WE_Z | R300_PVS_DST_WE_W));
+	    OUT_ACCEL_REG(R300_VAP_PVS_VECTOR_DATA_REG,
+			  (R300_PVS_SRC_REG_TYPE(R300_PVS_SRC_REG_INPUT) |
+			   R300_PVS_SRC_OFFSET(0) |
+			   R300_PVS_SRC_SWIZZLE_X(R300_PVS_SRC_SELECT_X) |
+			   R300_PVS_SRC_SWIZZLE_Y(R300_PVS_SRC_SELECT_Y) |
+			   R300_PVS_SRC_SWIZZLE_Z(R300_PVS_SRC_SELECT_Z) |
+			   R300_PVS_SRC_SWIZZLE_W(R300_PVS_SRC_SELECT_W)));
+	    OUT_ACCEL_REG(R300_VAP_PVS_VECTOR_DATA_REG,
+			  (R300_PVS_SRC_REG_TYPE(R300_PVS_SRC_REG_INPUT) |
+			   R300_PVS_SRC_OFFSET(0) |
+			   R300_PVS_SRC_SWIZZLE_X(R300_PVS_SRC_SELECT_FORCE_0) |
+			   R300_PVS_SRC_SWIZZLE_Y(R300_PVS_SRC_SELECT_FORCE_0) |
+			   R300_PVS_SRC_SWIZZLE_Z(R300_PVS_SRC_SELECT_FORCE_0) |
+			   R300_PVS_SRC_SWIZZLE_W(R300_PVS_SRC_SELECT_FORCE_0)));
+	    OUT_ACCEL_REG(R300_VAP_PVS_VECTOR_DATA_REG,
+			  (R300_PVS_SRC_REG_TYPE(R300_PVS_SRC_REG_INPUT) |
+			   R300_PVS_SRC_OFFSET(0) |
+			   R300_PVS_SRC_SWIZZLE_X(R300_PVS_SRC_SELECT_FORCE_0) |
+			   R300_PVS_SRC_SWIZZLE_Y(R300_PVS_SRC_SELECT_FORCE_0) |
+			   R300_PVS_SRC_SWIZZLE_Z(R300_PVS_SRC_SELECT_FORCE_0) |
+			   R300_PVS_SRC_SWIZZLE_W(R300_PVS_SRC_SELECT_FORCE_0)));
+
+	    /* PVS inst 1 */
+	    OUT_ACCEL_REG(R300_VAP_PVS_VECTOR_DATA_REG,
+			  (R300_PVS_DST_OPCODE(R300_VE_ADD) |
+			   R300_PVS_DST_REG_TYPE(R300_PVS_DST_REG_OUT) |
+			   R300_PVS_DST_OFFSET(1) |
+			   R300_PVS_DST_WE_X | R300_PVS_DST_WE_Y |
+			   R300_PVS_DST_WE_Z | R300_PVS_DST_WE_W));
+	    OUT_ACCEL_REG(R300_VAP_PVS_VECTOR_DATA_REG,
+			  (R300_PVS_SRC_REG_TYPE(R300_PVS_SRC_REG_INPUT) |
+			   R300_PVS_SRC_OFFSET(1) |
+			   R300_PVS_SRC_SWIZZLE_X(R300_PVS_SRC_SELECT_X) |
+			   R300_PVS_SRC_SWIZZLE_Y(R300_PVS_SRC_SELECT_Y) |
+			   R300_PVS_SRC_SWIZZLE_Z(R300_PVS_SRC_SELECT_Z) |
+			   R300_PVS_SRC_SWIZZLE_W(R300_PVS_SRC_SELECT_W)));
+	    OUT_ACCEL_REG(R300_VAP_PVS_VECTOR_DATA_REG,
+			  (R300_PVS_SRC_REG_TYPE(R300_PVS_SRC_REG_INPUT) |
+			   R300_PVS_SRC_OFFSET(1) |
+			   R300_PVS_SRC_SWIZZLE_X(R300_PVS_SRC_SELECT_FORCE_0) |
+			   R300_PVS_SRC_SWIZZLE_Y(R300_PVS_SRC_SELECT_FORCE_0) |
+			   R300_PVS_SRC_SWIZZLE_Z(R300_PVS_SRC_SELECT_FORCE_0) |
+			   R300_PVS_SRC_SWIZZLE_W(R300_PVS_SRC_SELECT_FORCE_0)));
+	    OUT_ACCEL_REG(R300_VAP_PVS_VECTOR_DATA_REG,
+			  (R300_PVS_SRC_REG_TYPE(R300_PVS_SRC_REG_INPUT) |
+			   R300_PVS_SRC_OFFSET(1) |
+			   R300_PVS_SRC_SWIZZLE_X(R300_PVS_SRC_SELECT_FORCE_0) |
+			   R300_PVS_SRC_SWIZZLE_Y(R300_PVS_SRC_SELECT_FORCE_0) |
+			   R300_PVS_SRC_SWIZZLE_Z(R300_PVS_SRC_SELECT_FORCE_0) |
+			   R300_PVS_SRC_SWIZZLE_W(R300_PVS_SRC_SELECT_FORCE_0)));
+	    FINISH_ACCEL();
+
+	    /* Xv shader program */
+	    BEGIN_ACCEL(9);
+	    OUT_ACCEL_REG(R300_VAP_PVS_VECTOR_INDX_REG, 5);
+
+	    OUT_ACCEL_REG(R300_VAP_PVS_VECTOR_DATA_REG,
+			  (R300_PVS_DST_OPCODE(R300_VE_ADD) |
+			   R300_PVS_DST_REG_TYPE(R300_PVS_DST_REG_OUT) |
+			   R300_PVS_DST_OFFSET(0) |
+			   R300_PVS_DST_WE_X | R300_PVS_DST_WE_Y |
+			   R300_PVS_DST_WE_Z | R300_PVS_DST_WE_W));
+	    OUT_ACCEL_REG(R300_VAP_PVS_VECTOR_DATA_REG,
+			  (R300_PVS_SRC_REG_TYPE(R300_PVS_SRC_REG_INPUT) |
+			   R300_PVS_SRC_OFFSET(0) |
+			   R300_PVS_SRC_SWIZZLE_X(R300_PVS_SRC_SELECT_X) |
+			   R300_PVS_SRC_SWIZZLE_Y(R300_PVS_SRC_SELECT_Y) |
+			   R300_PVS_SRC_SWIZZLE_Z(R300_PVS_SRC_SELECT_Z) |
+			   R300_PVS_SRC_SWIZZLE_W(R300_PVS_SRC_SELECT_W)));
+	    OUT_ACCEL_REG(R300_VAP_PVS_VECTOR_DATA_REG,
+			  (R300_PVS_SRC_REG_TYPE(R300_PVS_SRC_REG_INPUT) |
+			   R300_PVS_SRC_OFFSET(0) |
+			   R300_PVS_SRC_SWIZZLE_X(R300_PVS_SRC_SELECT_FORCE_0) |
+			   R300_PVS_SRC_SWIZZLE_Y(R300_PVS_SRC_SELECT_FORCE_0) |
+			   R300_PVS_SRC_SWIZZLE_Z(R300_PVS_SRC_SELECT_FORCE_0) |
+			   R300_PVS_SRC_SWIZZLE_W(R300_PVS_SRC_SELECT_FORCE_0)));
+	    OUT_ACCEL_REG(R300_VAP_PVS_VECTOR_DATA_REG,
+			  (R300_PVS_SRC_REG_TYPE(R300_PVS_SRC_REG_INPUT) |
+			   R300_PVS_SRC_OFFSET(0) |
+			   R300_PVS_SRC_SWIZZLE_X(R300_PVS_SRC_SELECT_FORCE_0) |
+			   R300_PVS_SRC_SWIZZLE_Y(R300_PVS_SRC_SELECT_FORCE_0) |
+			   R300_PVS_SRC_SWIZZLE_Z(R300_PVS_SRC_SELECT_FORCE_0) |
+			   R300_PVS_SRC_SWIZZLE_W(R300_PVS_SRC_SELECT_FORCE_0)));
+
+	    OUT_ACCEL_REG(R300_VAP_PVS_VECTOR_DATA_REG,
+			  (R300_PVS_DST_OPCODE(R300_VE_ADD) |
+			   R300_PVS_DST_REG_TYPE(R300_PVS_DST_REG_OUT) |
+			   R300_PVS_DST_OFFSET(1) |
+			   R300_PVS_DST_WE_X | R300_PVS_DST_WE_Y |
+			   R300_PVS_DST_WE_Z | R300_PVS_DST_WE_W));
+	    OUT_ACCEL_REG(R300_VAP_PVS_VECTOR_DATA_REG,
+			  (R300_PVS_SRC_REG_TYPE(R300_PVS_SRC_REG_INPUT) |
+			   R300_PVS_SRC_OFFSET(10) |
+			   R300_PVS_SRC_SWIZZLE_X(R300_PVS_SRC_SELECT_X) |
+			   R300_PVS_SRC_SWIZZLE_Y(R300_PVS_SRC_SELECT_Y) |
+			   R300_PVS_SRC_SWIZZLE_Z(R300_PVS_SRC_SELECT_Z) |
+			   R300_PVS_SRC_SWIZZLE_W(R300_PVS_SRC_SELECT_W)));
+	    OUT_ACCEL_REG(R300_VAP_PVS_VECTOR_DATA_REG,
+			  (R300_PVS_SRC_REG_TYPE(R300_PVS_SRC_REG_INPUT) |
+			   R300_PVS_SRC_OFFSET(10) |
+			   R300_PVS_SRC_SWIZZLE_X(R300_PVS_SRC_SELECT_FORCE_0) |
+			   R300_PVS_SRC_SWIZZLE_Y(R300_PVS_SRC_SELECT_FORCE_0) |
+			   R300_PVS_SRC_SWIZZLE_Z(R300_PVS_SRC_SELECT_FORCE_0) |
+			   R300_PVS_SRC_SWIZZLE_W(R300_PVS_SRC_SELECT_FORCE_0)));
+	    OUT_ACCEL_REG(R300_VAP_PVS_VECTOR_DATA_REG,
+			  (R300_PVS_SRC_REG_TYPE(R300_PVS_SRC_REG_INPUT) |
+			   R300_PVS_SRC_OFFSET(10) |
+			   R300_PVS_SRC_SWIZZLE_X(R300_PVS_SRC_SELECT_FORCE_0) |
+			   R300_PVS_SRC_SWIZZLE_Y(R300_PVS_SRC_SELECT_FORCE_0) |
+			   R300_PVS_SRC_SWIZZLE_Z(R300_PVS_SRC_SELECT_FORCE_0) |
+			   R300_PVS_SRC_SWIZZLE_W(R300_PVS_SRC_SELECT_FORCE_0)));
+	    FINISH_ACCEL();
+	}
+
+
 	BEGIN_ACCEL(4);
 	OUT_ACCEL_REG(R300_US_W_FMT, 0);
 	OUT_ACCEL_REG(R300_US_OUT_FMT_1, (R300_OUT_FMT_UNUSED |
diff --git a/src/radeon_exa_render.c b/src/radeon_exa_render.c
index b806399f..2bc2c659 100644
--- a/src/radeon_exa_render.c
+++ b/src/radeon_exa_render.c
@@ -26,6 +26,7 @@
  *    Eric Anholt <anholt@FreeBSD.org>
  *    Zack Rusin <zrusin@trolltech.com>
  *    Benjamin Herrenschmidt <benh@kernel.crashing.org>
+ *    Alex Deucher <alexander.deucher@amd.com>
  *
  */
 
@@ -1168,10 +1169,10 @@ static Bool FUNC_NAME(R300PrepareComposite)(int op, PicturePtr pSrcPicture,
     }
     FINISH_ACCEL();
 
-    /* setup the vertex shader */
+    /* load the vertex shader */
     if (info->has_tcl) {
+	BEGIN_ACCEL(2);
 	if (pMask) {
-	    BEGIN_ACCEL(15);
 	    OUT_ACCEL_REG(R300_VAP_PVS_CODE_CNTL_0,
 			  ((0 << R300_PVS_FIRST_INST_SHIFT) |
 			   (2 << R300_PVS_XYZW_VALID_INST_SHIFT) |
@@ -1179,104 +1180,13 @@ static Bool FUNC_NAME(R300PrepareComposite)(int op, PicturePtr pSrcPicture,
 	    OUT_ACCEL_REG(R300_VAP_PVS_CODE_CNTL_1,
 			  (2 << R300_PVS_LAST_VTX_SRC_INST_SHIFT));
 	} else {
-	    BEGIN_ACCEL(11);
 	    OUT_ACCEL_REG(R300_VAP_PVS_CODE_CNTL_0,
-			  ((0 << R300_PVS_FIRST_INST_SHIFT) |
-			   (1 << R300_PVS_XYZW_VALID_INST_SHIFT) |
-			   (1 << R300_PVS_LAST_INST_SHIFT)));
+			  ((3 << R300_PVS_FIRST_INST_SHIFT) |
+			   (4 << R300_PVS_XYZW_VALID_INST_SHIFT) |
+			   (4 << R300_PVS_LAST_INST_SHIFT)));
 	    OUT_ACCEL_REG(R300_VAP_PVS_CODE_CNTL_1,
-			  (1 << R300_PVS_LAST_VTX_SRC_INST_SHIFT));
-	}
-	OUT_ACCEL_REG(R300_VAP_PVS_VECTOR_INDX_REG, 0);
-	/* PVS inst 0 */
-	OUT_ACCEL_REG(R300_VAP_PVS_VECTOR_DATA_REG,
-		      (R300_PVS_DST_OPCODE(R300_VE_ADD) |
-		       R300_PVS_DST_REG_TYPE(R300_PVS_DST_REG_OUT) |
-		       R300_PVS_DST_OFFSET(0) |
-		       R300_PVS_DST_WE_X | R300_PVS_DST_WE_Y |
-		       R300_PVS_DST_WE_Z | R300_PVS_DST_WE_W));
-	OUT_ACCEL_REG(R300_VAP_PVS_VECTOR_DATA_REG,
-		      (R300_PVS_SRC_REG_TYPE(R300_PVS_SRC_REG_INPUT) |
-		       R300_PVS_SRC_OFFSET(0) |
-		       R300_PVS_SRC_SWIZZLE_X(R300_PVS_SRC_SELECT_X) |
-		       R300_PVS_SRC_SWIZZLE_Y(R300_PVS_SRC_SELECT_Y) |
-		       R300_PVS_SRC_SWIZZLE_Z(R300_PVS_SRC_SELECT_Z) |
-		       R300_PVS_SRC_SWIZZLE_W(R300_PVS_SRC_SELECT_W)));
-	OUT_ACCEL_REG(R300_VAP_PVS_VECTOR_DATA_REG,
-		      (R300_PVS_SRC_REG_TYPE(R300_PVS_SRC_REG_INPUT) |
-		       R300_PVS_SRC_OFFSET(0) |
-		       R300_PVS_SRC_SWIZZLE_X(R300_PVS_SRC_SELECT_FORCE_0) |
-		       R300_PVS_SRC_SWIZZLE_Y(R300_PVS_SRC_SELECT_FORCE_0) |
-		       R300_PVS_SRC_SWIZZLE_Z(R300_PVS_SRC_SELECT_FORCE_0) |
-		       R300_PVS_SRC_SWIZZLE_W(R300_PVS_SRC_SELECT_FORCE_0)));
-	OUT_ACCEL_REG(R300_VAP_PVS_VECTOR_DATA_REG,
-		      (R300_PVS_SRC_REG_TYPE(R300_PVS_SRC_REG_INPUT) |
-		       R300_PVS_SRC_OFFSET(0) |
-		       R300_PVS_SRC_SWIZZLE_X(R300_PVS_SRC_SELECT_FORCE_0) |
-		       R300_PVS_SRC_SWIZZLE_Y(R300_PVS_SRC_SELECT_FORCE_0) |
-		       R300_PVS_SRC_SWIZZLE_Z(R300_PVS_SRC_SELECT_FORCE_0) |
-		       R300_PVS_SRC_SWIZZLE_W(R300_PVS_SRC_SELECT_FORCE_0)));
-
-	/* PVS inst 1 */
-	OUT_ACCEL_REG(R300_VAP_PVS_VECTOR_DATA_REG,
-		      (R300_PVS_DST_OPCODE(R300_VE_ADD) |
-		       R300_PVS_DST_REG_TYPE(R300_PVS_DST_REG_OUT) |
-		       R300_PVS_DST_OFFSET(1) |
-		       R300_PVS_DST_WE_X | R300_PVS_DST_WE_Y |
-		       R300_PVS_DST_WE_Z | R300_PVS_DST_WE_W));
-	OUT_ACCEL_REG(R300_VAP_PVS_VECTOR_DATA_REG,
-		      (R300_PVS_SRC_REG_TYPE(R300_PVS_SRC_REG_INPUT) |
-		       R300_PVS_SRC_OFFSET(1) |
-		       R300_PVS_SRC_SWIZZLE_X(R300_PVS_SRC_SELECT_X) |
-		       R300_PVS_SRC_SWIZZLE_Y(R300_PVS_SRC_SELECT_Y) |
-		       R300_PVS_SRC_SWIZZLE_Z(R300_PVS_SRC_SELECT_Z) |
-		       R300_PVS_SRC_SWIZZLE_W(R300_PVS_SRC_SELECT_W)));
-	OUT_ACCEL_REG(R300_VAP_PVS_VECTOR_DATA_REG,
-		      (R300_PVS_SRC_REG_TYPE(R300_PVS_SRC_REG_INPUT) |
-		       R300_PVS_SRC_OFFSET(1) |
-		       R300_PVS_SRC_SWIZZLE_X(R300_PVS_SRC_SELECT_FORCE_0) |
-		       R300_PVS_SRC_SWIZZLE_Y(R300_PVS_SRC_SELECT_FORCE_0) |
-		       R300_PVS_SRC_SWIZZLE_Z(R300_PVS_SRC_SELECT_FORCE_0) |
-		       R300_PVS_SRC_SWIZZLE_W(R300_PVS_SRC_SELECT_FORCE_0)));
-	OUT_ACCEL_REG(R300_VAP_PVS_VECTOR_DATA_REG,
-		      (R300_PVS_SRC_REG_TYPE(R300_PVS_SRC_REG_INPUT) |
-		       R300_PVS_SRC_OFFSET(1) |
-		       R300_PVS_SRC_SWIZZLE_X(R300_PVS_SRC_SELECT_FORCE_0) |
-		       R300_PVS_SRC_SWIZZLE_Y(R300_PVS_SRC_SELECT_FORCE_0) |
-		       R300_PVS_SRC_SWIZZLE_Z(R300_PVS_SRC_SELECT_FORCE_0) |
-		       R300_PVS_SRC_SWIZZLE_W(R300_PVS_SRC_SELECT_FORCE_0)));
-
-	if (pMask) {
-	    /* PVS inst 2 */
-	    OUT_ACCEL_REG(R300_VAP_PVS_VECTOR_DATA_REG,
-			  (R300_PVS_DST_OPCODE(R300_VE_ADD) |
-			   R300_PVS_DST_REG_TYPE(R300_PVS_DST_REG_OUT) |
-			   R300_PVS_DST_OFFSET(2) |
-			   R300_PVS_DST_WE_X | R300_PVS_DST_WE_Y |
-			   R300_PVS_DST_WE_Z | R300_PVS_DST_WE_W));
-	    OUT_ACCEL_REG(R300_VAP_PVS_VECTOR_DATA_REG,
-			  (R300_PVS_SRC_REG_TYPE(R300_PVS_SRC_REG_INPUT) |
-			   R300_PVS_SRC_OFFSET(2) |
-			   R300_PVS_SRC_SWIZZLE_X(R300_PVS_SRC_SELECT_X) |
-			   R300_PVS_SRC_SWIZZLE_Y(R300_PVS_SRC_SELECT_Y) |
-			   R300_PVS_SRC_SWIZZLE_Z(R300_PVS_SRC_SELECT_Z) |
-			   R300_PVS_SRC_SWIZZLE_W(R300_PVS_SRC_SELECT_W)));
-	    OUT_ACCEL_REG(R300_VAP_PVS_VECTOR_DATA_REG,
-			  (R300_PVS_SRC_REG_TYPE(R300_PVS_SRC_REG_INPUT) |
-			   R300_PVS_SRC_OFFSET(2) |
-			   R300_PVS_SRC_SWIZZLE_X(R300_PVS_SRC_SELECT_FORCE_0) |
-			   R300_PVS_SRC_SWIZZLE_Y(R300_PVS_SRC_SELECT_FORCE_0) |
-			   R300_PVS_SRC_SWIZZLE_Z(R300_PVS_SRC_SELECT_FORCE_0) |
-			   R300_PVS_SRC_SWIZZLE_W(R300_PVS_SRC_SELECT_FORCE_0)));
-	    OUT_ACCEL_REG(R300_VAP_PVS_VECTOR_DATA_REG,
-			  (R300_PVS_SRC_REG_TYPE(R300_PVS_SRC_REG_INPUT) |
-			   R300_PVS_SRC_OFFSET(2) |
-			   R300_PVS_SRC_SWIZZLE_X(R300_PVS_SRC_SELECT_FORCE_0) |
-			   R300_PVS_SRC_SWIZZLE_Y(R300_PVS_SRC_SELECT_FORCE_0) |
-			   R300_PVS_SRC_SWIZZLE_Z(R300_PVS_SRC_SELECT_FORCE_0) |
-			   R300_PVS_SRC_SWIZZLE_W(R300_PVS_SRC_SELECT_FORCE_0)));
+			  (4 << R300_PVS_LAST_VTX_SRC_INST_SHIFT));
 	}
-
 	FINISH_ACCEL();
     }
 
diff --git a/src/radeon_textured_videofuncs.c b/src/radeon_textured_videofuncs.c
index 363490a2..564f66cb 100644
--- a/src/radeon_textured_videofuncs.c
+++ b/src/radeon_textured_videofuncs.c
@@ -193,7 +193,7 @@ FUNC_NAME(RADEONDisplayTexturedVideo)(ScrnInfoPtr pScrn, RADEONPortPrivPtr pPriv
 
 	/* setup the VAP */
 	if (info->has_tcl) {
-	    BEGIN_VIDEO(16);
+	    BEGIN_VIDEO(7);
 	    OUT_VIDEO_REG(R300_VAP_PROG_STREAM_CNTL_0,
 			  ((R300_DATA_TYPE_FLOAT_2 << R300_DATA_TYPE_0_SHIFT) |
 			   (0 << R300_SKIP_DWORDS_0_SHIFT) |
@@ -244,72 +244,14 @@ FUNC_NAME(RADEONDisplayTexturedVideo)(ScrnInfoPtr pScrn, RADEONPortPrivPtr pPriv
 			    << R300_WRITE_ENA_1_SHIFT)));
 	}
 
-	/* setup vertex shader */
+	/* load the vertex shader */
 	if (info->has_tcl) {
 	    OUT_VIDEO_REG(R300_VAP_PVS_CODE_CNTL_0,
-			  ((0 << R300_PVS_FIRST_INST_SHIFT) |
-			   (1 << R300_PVS_XYZW_VALID_INST_SHIFT) |
-			   (1 << R300_PVS_LAST_INST_SHIFT)));
+			  ((5 << R300_PVS_FIRST_INST_SHIFT) |
+			   (6 << R300_PVS_XYZW_VALID_INST_SHIFT) |
+			   (6 << R300_PVS_LAST_INST_SHIFT)));
 	    OUT_VIDEO_REG(R300_VAP_PVS_CODE_CNTL_1,
-			  (1 << R300_PVS_LAST_VTX_SRC_INST_SHIFT));
-	    OUT_VIDEO_REG(R300_VAP_PVS_VECTOR_INDX_REG, 0);
-
-	    OUT_VIDEO_REG(R300_VAP_PVS_VECTOR_DATA_REG,
-			  (R300_PVS_DST_OPCODE(R300_VE_ADD) |
-			   R300_PVS_DST_REG_TYPE(R300_PVS_DST_REG_OUT) |
-			   R300_PVS_DST_OFFSET(0) |
-			   R300_PVS_DST_WE_X | R300_PVS_DST_WE_Y |
-			   R300_PVS_DST_WE_Z | R300_PVS_DST_WE_W));
-	    OUT_VIDEO_REG(R300_VAP_PVS_VECTOR_DATA_REG,
-			  (R300_PVS_SRC_REG_TYPE(R300_PVS_SRC_REG_INPUT) |
-			   R300_PVS_SRC_OFFSET(0) |
-			   R300_PVS_SRC_SWIZZLE_X(R300_PVS_SRC_SELECT_X) |
-			   R300_PVS_SRC_SWIZZLE_Y(R300_PVS_SRC_SELECT_Y) |
-			   R300_PVS_SRC_SWIZZLE_Z(R300_PVS_SRC_SELECT_Z) |
-			   R300_PVS_SRC_SWIZZLE_W(R300_PVS_SRC_SELECT_W)));
-	    OUT_VIDEO_REG(R300_VAP_PVS_VECTOR_DATA_REG,
-			  (R300_PVS_SRC_REG_TYPE(R300_PVS_SRC_REG_INPUT) |
-			   R300_PVS_SRC_OFFSET(0) |
-			   R300_PVS_SRC_SWIZZLE_X(R300_PVS_SRC_SELECT_FORCE_0) |
-			   R300_PVS_SRC_SWIZZLE_Y(R300_PVS_SRC_SELECT_FORCE_0) |
-			   R300_PVS_SRC_SWIZZLE_Z(R300_PVS_SRC_SELECT_FORCE_0) |
-			   R300_PVS_SRC_SWIZZLE_W(R300_PVS_SRC_SELECT_FORCE_0)));
-	    OUT_VIDEO_REG(R300_VAP_PVS_VECTOR_DATA_REG,
-			  (R300_PVS_SRC_REG_TYPE(R300_PVS_SRC_REG_INPUT) |
-			   R300_PVS_SRC_OFFSET(0) |
-			   R300_PVS_SRC_SWIZZLE_X(R300_PVS_SRC_SELECT_FORCE_0) |
-			   R300_PVS_SRC_SWIZZLE_Y(R300_PVS_SRC_SELECT_FORCE_0) |
-			   R300_PVS_SRC_SWIZZLE_Z(R300_PVS_SRC_SELECT_FORCE_0) |
-			   R300_PVS_SRC_SWIZZLE_W(R300_PVS_SRC_SELECT_FORCE_0)));
-
-	    OUT_VIDEO_REG(R300_VAP_PVS_VECTOR_DATA_REG,
-			  (R300_PVS_DST_OPCODE(R300_VE_ADD) |
-			   R300_PVS_DST_REG_TYPE(R300_PVS_DST_REG_OUT) |
-			   R300_PVS_DST_OFFSET(1) |
-			   R300_PVS_DST_WE_X | R300_PVS_DST_WE_Y |
-			   R300_PVS_DST_WE_Z | R300_PVS_DST_WE_W));
-	    OUT_VIDEO_REG(R300_VAP_PVS_VECTOR_DATA_REG,
-			  (R300_PVS_SRC_REG_TYPE(R300_PVS_SRC_REG_INPUT) |
-			   R300_PVS_SRC_OFFSET(10) |
-			   R300_PVS_SRC_SWIZZLE_X(R300_PVS_SRC_SELECT_X) |
-			   R300_PVS_SRC_SWIZZLE_Y(R300_PVS_SRC_SELECT_Y) |
-			   R300_PVS_SRC_SWIZZLE_Z(R300_PVS_SRC_SELECT_Z) |
-			   R300_PVS_SRC_SWIZZLE_W(R300_PVS_SRC_SELECT_W)));
-	    OUT_VIDEO_REG(R300_VAP_PVS_VECTOR_DATA_REG,
-			  (R300_PVS_SRC_REG_TYPE(R300_PVS_SRC_REG_INPUT) |
-			   R300_PVS_SRC_OFFSET(10) |
-			   R300_PVS_SRC_SWIZZLE_X(R300_PVS_SRC_SELECT_FORCE_0) |
-			   R300_PVS_SRC_SWIZZLE_Y(R300_PVS_SRC_SELECT_FORCE_0) |
-			   R300_PVS_SRC_SWIZZLE_Z(R300_PVS_SRC_SELECT_FORCE_0) |
-			   R300_PVS_SRC_SWIZZLE_W(R300_PVS_SRC_SELECT_FORCE_0)));
-	    OUT_VIDEO_REG(R300_VAP_PVS_VECTOR_DATA_REG,
-			  (R300_PVS_SRC_REG_TYPE(R300_PVS_SRC_REG_INPUT) |
-			   R300_PVS_SRC_OFFSET(10) |
-			   R300_PVS_SRC_SWIZZLE_X(R300_PVS_SRC_SELECT_FORCE_0) |
-			   R300_PVS_SRC_SWIZZLE_Y(R300_PVS_SRC_SELECT_FORCE_0) |
-			   R300_PVS_SRC_SWIZZLE_Z(R300_PVS_SRC_SELECT_FORCE_0) |
-			   R300_PVS_SRC_SWIZZLE_W(R300_PVS_SRC_SELECT_FORCE_0)));
-
+			  (6 << R300_PVS_LAST_VTX_SRC_INST_SHIFT));
 	}
 
 	OUT_VIDEO_REG(R300_VAP_OUT_VTX_FMT_0, R300_VTX_POS_PRESENT);