R3xx-R5xx: do EXA transforms in the vertex shader

4 files changed, 360 insertions, 98 deletions

diff --git a/src/radeon_commonfuncs.c b/src/radeon_commonfuncs.c
index ba358ab6..28bb6e56 100644
--- a/src/radeon_commonfuncs.c
+++ b/src/radeon_commonfuncs.c
@@ -220,10 +220,10 @@ static void FUNC_NAME(RADEONInit3DEngine)(ScrnInfoPtr pScrn)
 
 	/* pre-load the vertex shaders */
 	if (info->accel_state->has_tcl) {
-	    /* exa mask/Xv bicubic shader program */
-	    BEGIN_ACCEL(13);
-	    OUT_ACCEL_REG(R300_VAP_PVS_VECTOR_INDX_REG, 0);
-	    /* PVS inst 0 */
+	    BEGIN_ACCEL(37);
+	    /* exa composite shader program */
+	    OUT_ACCEL_REG(R300_VAP_PVS_VECTOR_INDX_REG, R300_PVS_VECTOR_INST_INDEX(0));
+	    /* PVS inst 0 - dst X,Y */
 	    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) |
@@ -235,8 +235,8 @@ static void FUNC_NAME(RADEONInit3DEngine)(ScrnInfoPtr pScrn)
 			   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)));
+			   R300_PVS_SRC_SWIZZLE_Z(R300_PVS_SRC_SELECT_FORCE_0) |
+			   R300_PVS_SRC_SWIZZLE_W(R300_PVS_SRC_SELECT_FORCE_1)));
 	    OUT_ACCEL_REG(R300_VAP_PVS_VECTOR_DATA_REG,
 			  (R300_PVS_SRC_REG_TYPE(R300_PVS_SRC_REG_INPUT) |
 			   R300_PVS_SRC_OFFSET(0) |
@@ -252,20 +252,26 @@ static void FUNC_NAME(RADEONInit3DEngine)(ScrnInfoPtr pScrn)
 			   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 */
+	    /* PVS inst 1 - src X */
 	    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));
+			  (R300_PVS_DST_OPCODE(R300_VE_DOT_PRODUCT) |
+			   R300_PVS_DST_REG_TYPE(R300_PVS_DST_REG_TEMPORARY) |
+			   R300_PVS_DST_OFFSET(0) |
+			   R300_PVS_DST_WE_X));
 	    OUT_ACCEL_REG(R300_VAP_PVS_VECTOR_DATA_REG,
 			  (R300_PVS_SRC_REG_TYPE(R300_PVS_SRC_REG_INPUT) |
 			   R300_PVS_SRC_OFFSET(6) |
 			   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_FORCE_1) |
+			   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_CONSTANT) |
+			   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)));
+			   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(6) |
@@ -273,6 +279,27 @@ static void FUNC_NAME(RADEONInit3DEngine)(ScrnInfoPtr pScrn)
 			   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 - src Y */
+	    OUT_ACCEL_REG(R300_VAP_PVS_VECTOR_DATA_REG,
+			  (R300_PVS_DST_OPCODE(R300_VE_DOT_PRODUCT) |
+			   R300_PVS_DST_REG_TYPE(R300_PVS_DST_REG_TEMPORARY) |
+			   R300_PVS_DST_OFFSET(0) |
+			   R300_PVS_DST_WE_Y));
+	    OUT_ACCEL_REG(R300_VAP_PVS_VECTOR_DATA_REG,
+			  (R300_PVS_SRC_REG_TYPE(R300_PVS_SRC_REG_INPUT) |
+			   R300_PVS_SRC_OFFSET(6) |
+			   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_FORCE_1) |
+			   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_CONSTANT) |
+			   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_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(6) |
@@ -281,82 +308,138 @@ static void FUNC_NAME(RADEONInit3DEngine)(ScrnInfoPtr pScrn)
 			   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 */
+	    /* PVS inst 3 - src X / w */
 	    OUT_ACCEL_REG(R300_VAP_PVS_VECTOR_DATA_REG,
-			  (R300_PVS_DST_OPCODE(R300_VE_ADD) |
+			  (R300_PVS_DST_OPCODE(R300_VE_MULTIPLY) |
 			   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));
+			   R300_PVS_DST_OFFSET(1) |
+			   R300_PVS_DST_WE_X));
 	    OUT_ACCEL_REG(R300_VAP_PVS_VECTOR_DATA_REG,
-			  (R300_PVS_SRC_REG_TYPE(R300_PVS_SRC_REG_INPUT) |
-			   R300_PVS_SRC_OFFSET(7) |
+			  (R300_PVS_SRC_REG_TYPE(R300_PVS_SRC_REG_TEMPORARY) |
+			   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)));
+			   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_CONSTANT) |
+			   R300_PVS_SRC_OFFSET(0) |
+			   R300_PVS_SRC_SWIZZLE_X(R300_PVS_SRC_SELECT_W) |
+			   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(7) |
+			   R300_PVS_SRC_OFFSET(6) |
 			   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 4 - src y / h */
+	    OUT_ACCEL_REG(R300_VAP_PVS_VECTOR_DATA_REG,
+			  (R300_PVS_DST_OPCODE(R300_VE_MULTIPLY) |
+			   R300_PVS_DST_REG_TYPE(R300_PVS_DST_REG_OUT) |
+			   R300_PVS_DST_OFFSET(1) |
+			   R300_PVS_DST_WE_Y));
+	    OUT_ACCEL_REG(R300_VAP_PVS_VECTOR_DATA_REG,
+			  (R300_PVS_SRC_REG_TYPE(R300_PVS_SRC_REG_TEMPORARY) |
+			   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_Y) |
+			   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_CONSTANT) |
+			   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_W) |
+			   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(7) |
+			   R300_PVS_SRC_OFFSET(6) |
 			   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 */
+	    /* PVS inst 5 - mask X */
 	    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_OPCODE(R300_VE_DOT_PRODUCT) |
+			   R300_PVS_DST_REG_TYPE(R300_PVS_DST_REG_TEMPORARY) |
 			   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));
+			   R300_PVS_DST_WE_Z));
 	    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_OFFSET(7) |
+			   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_FORCE_1) |
+			   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_CONSTANT) |
+			   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)));
+			   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_OFFSET(7) |
 			   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 6 - mask Y */
+	    OUT_ACCEL_REG(R300_VAP_PVS_VECTOR_DATA_REG,
+			  (R300_PVS_DST_OPCODE(R300_VE_DOT_PRODUCT) |
+			   R300_PVS_DST_REG_TYPE(R300_PVS_DST_REG_TEMPORARY) |
+			   R300_PVS_DST_OFFSET(0) |
+			   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_OFFSET(7) |
+			   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_FORCE_1) |
+			   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_CONSTANT) |
+			   R300_PVS_SRC_OFFSET(3) |
+			   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_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(7) |
 			   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 */
+	    /* PVS inst 7 - mask X / w */
 	    OUT_ACCEL_REG(R300_VAP_PVS_VECTOR_DATA_REG,
-			  (R300_PVS_DST_OPCODE(R300_VE_ADD) |
+			  (R300_PVS_DST_OPCODE(R300_VE_MULTIPLY) |
 			   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));
+			   R300_PVS_DST_OFFSET(2) |
+			   R300_PVS_DST_WE_X));
 	    OUT_ACCEL_REG(R300_VAP_PVS_VECTOR_DATA_REG,
-			  (R300_PVS_SRC_REG_TYPE(R300_PVS_SRC_REG_INPUT) |
-			   R300_PVS_SRC_OFFSET(6) |
-			   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)));
+			  (R300_PVS_SRC_REG_TYPE(R300_PVS_SRC_REG_TEMPORARY) |
+			   R300_PVS_SRC_OFFSET(0) |
+			   R300_PVS_SRC_SWIZZLE_X(R300_PVS_SRC_SELECT_Z) |
+			   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_CONSTANT) |
+			   R300_PVS_SRC_OFFSET(2) |
+			   R300_PVS_SRC_SWIZZLE_X(R300_PVS_SRC_SELECT_W) |
+			   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(6) |
@@ -364,6 +447,27 @@ static void FUNC_NAME(RADEONInit3DEngine)(ScrnInfoPtr pScrn)
 			   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 8 - mask y / h */
+	    OUT_ACCEL_REG(R300_VAP_PVS_VECTOR_DATA_REG,
+			  (R300_PVS_DST_OPCODE(R300_VE_MULTIPLY) |
+			   R300_PVS_DST_REG_TYPE(R300_PVS_DST_REG_OUT) |
+			   R300_PVS_DST_OFFSET(2) |
+			   R300_PVS_DST_WE_Y));
+	    OUT_ACCEL_REG(R300_VAP_PVS_VECTOR_DATA_REG,
+			  (R300_PVS_SRC_REG_TYPE(R300_PVS_SRC_REG_TEMPORARY) |
+			   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_W) |
+			   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_CONSTANT) |
+			   R300_PVS_SRC_OFFSET(3) |
+			   R300_PVS_SRC_SWIZZLE_X(R300_PVS_SRC_SELECT_FORCE_0) |
+			   R300_PVS_SRC_SWIZZLE_Y(R300_PVS_SRC_SELECT_W) |
+			   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(6) |
@@ -375,7 +479,7 @@ static void FUNC_NAME(RADEONInit3DEngine)(ScrnInfoPtr pScrn)
 
 	    /* Xv shader program */
 	    BEGIN_ACCEL(9);
-	    OUT_ACCEL_REG(R300_VAP_PVS_VECTOR_INDX_REG, 5);
+	    OUT_ACCEL_REG(R300_VAP_PVS_VECTOR_INDX_REG, R300_PVS_VECTOR_INST_INDEX(9));
 
 	    OUT_ACCEL_REG(R300_VAP_PVS_VECTOR_DATA_REG,
 			  (R300_PVS_DST_OPCODE(R300_VE_ADD) |
@@ -388,8 +492,8 @@ static void FUNC_NAME(RADEONInit3DEngine)(ScrnInfoPtr pScrn)
 			   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)));
+			   R300_PVS_SRC_SWIZZLE_Z(R300_PVS_SRC_SELECT_FORCE_0) |
+			   R300_PVS_SRC_SWIZZLE_W(R300_PVS_SRC_SELECT_FORCE_1)));
 	    OUT_ACCEL_REG(R300_VAP_PVS_VECTOR_DATA_REG,
 			  (R300_PVS_SRC_REG_TYPE(R300_PVS_SRC_REG_INPUT) |
 			   R300_PVS_SRC_OFFSET(0) |
@@ -409,15 +513,14 @@ static void FUNC_NAME(RADEONInit3DEngine)(ScrnInfoPtr pScrn)
 			  (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));
+			   R300_PVS_DST_WE_X | R300_PVS_DST_WE_Y));
 	    OUT_ACCEL_REG(R300_VAP_PVS_VECTOR_DATA_REG,
 			  (R300_PVS_SRC_REG_TYPE(R300_PVS_SRC_REG_INPUT) |
 			   R300_PVS_SRC_OFFSET(6) |
 			   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)));
+			   R300_PVS_SRC_SWIZZLE_Z(R300_PVS_SRC_SELECT_FORCE_0) |
+			   R300_PVS_SRC_SWIZZLE_W(R300_PVS_SRC_SELECT_FORCE_1)));
 	    OUT_ACCEL_REG(R300_VAP_PVS_VECTOR_DATA_REG,
 			  (R300_PVS_SRC_REG_TYPE(R300_PVS_SRC_REG_INPUT) |
 			   R300_PVS_SRC_OFFSET(6) |
@@ -433,6 +536,97 @@ static void FUNC_NAME(RADEONInit3DEngine)(ScrnInfoPtr pScrn)
 			   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 bicubic shader program */
+	    BEGIN_ACCEL(13);
+            OUT_ACCEL_REG(R300_VAP_PVS_VECTOR_INDX_REG, R300_PVS_VECTOR_INST_INDEX(11));
+            /* 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_FORCE_0) |
+                           R300_PVS_SRC_SWIZZLE_W(R300_PVS_SRC_SELECT_FORCE_1)));
+	    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(6) |
+                           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_FORCE_0) |
+                           R300_PVS_SRC_SWIZZLE_W(R300_PVS_SRC_SELECT_FORCE_1)));
+	    OUT_ACCEL_REG(R300_VAP_PVS_VECTOR_DATA_REG,
+                          (R300_PVS_SRC_REG_TYPE(R300_PVS_SRC_REG_INPUT) |
+                           R300_PVS_SRC_OFFSET(6) |
+                           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(6) |
+                           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(7) |
+                           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_FORCE_0) |
+                           R300_PVS_SRC_SWIZZLE_W(R300_PVS_SRC_SELECT_FORCE_1)));
+            OUT_ACCEL_REG(R300_VAP_PVS_VECTOR_DATA_REG,
+                          (R300_PVS_SRC_REG_TYPE(R300_PVS_SRC_REG_INPUT) |
+                           R300_PVS_SRC_OFFSET(7) |
+                           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(7) |
+                           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();
 	}
 
 	/* pre-load the RS instructions */
diff --git a/src/radeon_exa_render.c b/src/radeon_exa_render.c
index 8dbbee94..89c803aa 100644
--- a/src/radeon_exa_render.c
+++ b/src/radeon_exa_render.c
@@ -1099,9 +1099,6 @@ static Bool FUNC_NAME(R300TextureSetup)(PicturePtr pPict, PixmapPtr pPix,
      */
     txformat0 |= R300_TXPITCH_EN;
 
-    info->accel_state->texW[unit] = w;
-    info->accel_state->texH[unit] = h;
-
     txfilter = (unit << R300_TX_ID_SHIFT);
 
     if (pPict->repeat) {
@@ -1160,8 +1157,61 @@ static Bool FUNC_NAME(R300TextureSetup)(PicturePtr pPict, PixmapPtr pPix,
     if (pPict->transform != 0) {
 	info->accel_state->is_transform[unit] = TRUE;
 	info->accel_state->transform[unit] = pPict->transform;
+
+	/* setup the PVS consts */
+	if (info->accel_state->has_tcl) {
+	    info->accel_state->texW[unit] = 1;
+	    info->accel_state->texH[unit] = 1;
+	    BEGIN_ACCEL(9);
+	    if (IS_R300_3D)
+		OUT_ACCEL_REG(R300_VAP_PVS_VECTOR_INDX_REG, R300_PVS_VECTOR_CONST_INDEX(unit * 2));
+	    else
+		OUT_ACCEL_REG(R300_VAP_PVS_VECTOR_INDX_REG, R500_PVS_VECTOR_CONST_INDEX(unit * 2));
+
+	    OUT_ACCEL_REG(R300_VAP_PVS_VECTOR_DATA_REG, F_TO_DW(xFixedToFloat(pPict->transform->matrix[0][0])));
+	    OUT_ACCEL_REG(R300_VAP_PVS_VECTOR_DATA_REG, F_TO_DW(xFixedToFloat(pPict->transform->matrix[0][1])));
+	    OUT_ACCEL_REG(R300_VAP_PVS_VECTOR_DATA_REG, F_TO_DW(xFixedToFloat(pPict->transform->matrix[0][2])));
+	    OUT_ACCEL_REG(R300_VAP_PVS_VECTOR_DATA_REG, F_TO_DW(1.0/w));
+
+	    OUT_ACCEL_REG(R300_VAP_PVS_VECTOR_DATA_REG, F_TO_DW(xFixedToFloat(pPict->transform->matrix[1][0])));
+	    OUT_ACCEL_REG(R300_VAP_PVS_VECTOR_DATA_REG, F_TO_DW(xFixedToFloat(pPict->transform->matrix[1][1])));
+	    OUT_ACCEL_REG(R300_VAP_PVS_VECTOR_DATA_REG, F_TO_DW(xFixedToFloat(pPict->transform->matrix[1][2])));
+	    OUT_ACCEL_REG(R300_VAP_PVS_VECTOR_DATA_REG, F_TO_DW(1.0/h));
+
+	    FINISH_ACCEL();
+	} else {
+	    info->accel_state->texW[unit] = w;
+	    info->accel_state->texH[unit] = h;
+	}
     } else {
 	info->accel_state->is_transform[unit] = FALSE;
+
+	/* setup the PVS consts */
+	if (info->accel_state->has_tcl) {
+	    info->accel_state->texW[unit] = 1;
+	    info->accel_state->texH[unit] = 1;
+
+	    BEGIN_ACCEL(9);
+	    if (IS_R300_3D)
+		OUT_ACCEL_REG(R300_VAP_PVS_VECTOR_INDX_REG, R300_PVS_VECTOR_CONST_INDEX(unit * 2));
+	    else
+		OUT_ACCEL_REG(R300_VAP_PVS_VECTOR_INDX_REG, R500_PVS_VECTOR_CONST_INDEX(unit * 2));
+
+	    OUT_ACCEL_REG(R300_VAP_PVS_VECTOR_DATA_REG, F_TO_DW(1.0));
+	    OUT_ACCEL_REG(R300_VAP_PVS_VECTOR_DATA_REG, F_TO_DW(0.0));
+	    OUT_ACCEL_REG(R300_VAP_PVS_VECTOR_DATA_REG, F_TO_DW(0.0));
+	    OUT_ACCEL_REG(R300_VAP_PVS_VECTOR_DATA_REG, F_TO_DW(1.0/w));
+
+	    OUT_ACCEL_REG(R300_VAP_PVS_VECTOR_DATA_REG, F_TO_DW(0.0));
+	    OUT_ACCEL_REG(R300_VAP_PVS_VECTOR_DATA_REG, F_TO_DW(1.0));
+	    OUT_ACCEL_REG(R300_VAP_PVS_VECTOR_DATA_REG, F_TO_DW(0.0));
+	    OUT_ACCEL_REG(R300_VAP_PVS_VECTOR_DATA_REG, F_TO_DW(1.0/h));
+
+	    FINISH_ACCEL();
+	} else {
+	    info->accel_state->texW[unit] = w;
+	    info->accel_state->texH[unit] = h;
+	}
     }
 
     return TRUE;
@@ -1310,9 +1360,10 @@ static Bool FUNC_NAME(R300PrepareComposite)(int op, PicturePtr pSrcPicture,
     /* setup the VAP */
     if (info->accel_state->has_tcl) {
 	if (pMask)
-	    BEGIN_ACCEL(8);
+	    BEGIN_ACCEL(10);
 	else
-	    BEGIN_ACCEL(7);
+	    BEGIN_ACCEL(9);
+	OUT_ACCEL_REG(R300_VAP_PVS_STATE_FLUSH_REG, 0);
     } else {
 	if (pMask)
 	    BEGIN_ACCEL(6);
@@ -1363,22 +1414,28 @@ static Bool FUNC_NAME(R300PrepareComposite)(int op, PicturePtr pSrcPicture,
 
     /* load the vertex shader
      * We pre-load vertex programs in RADEONInit3DEngine():
-     * - exa no mask
-     * - exa mask
+     * - exa
      * - Xv
+     * - Xv bicubic
      * Here we select the offset of the vertex program we want to use
      */
     if (info->accel_state->has_tcl) {
 	if (pMask) {
+	    /* consts used by vertex shaders */
+	    OUT_ACCEL_REG(R300_VAP_PVS_CONST_CNTL, (R300_PVS_CONST_BASE_OFFSET(0) |
+						    R300_PVS_MAX_CONST_ADDR(3)));
 	    OUT_ACCEL_REG(R300_VAP_PVS_CODE_CNTL_0,
 			  ((0 << R300_PVS_FIRST_INST_SHIFT) |
-			   (2 << R300_PVS_XYZW_VALID_INST_SHIFT) |
-			   (2 << R300_PVS_LAST_INST_SHIFT)));
+			   (8 << R300_PVS_XYZW_VALID_INST_SHIFT) |
+			   (8 << R300_PVS_LAST_INST_SHIFT)));
 	    OUT_ACCEL_REG(R300_VAP_PVS_CODE_CNTL_1,
-			  (2 << R300_PVS_LAST_VTX_SRC_INST_SHIFT));
+			  (8 << R300_PVS_LAST_VTX_SRC_INST_SHIFT));
 	} else {
+	    /* consts used by vertex shaders */
+	    OUT_ACCEL_REG(R300_VAP_PVS_CONST_CNTL, (R300_PVS_CONST_BASE_OFFSET(0) |
+						    R300_PVS_MAX_CONST_ADDR(3)));
 	    OUT_ACCEL_REG(R300_VAP_PVS_CODE_CNTL_0,
-			  ((3 << R300_PVS_FIRST_INST_SHIFT) |
+			  ((0 << 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,
@@ -2054,10 +2111,12 @@ static void FUNC_NAME(RadeonCompositeTile)(ScrnInfoPtr pScrn,
     srcBottomRight.y = IntToxFixed(srcY + h);
 
     if (info->accel_state->is_transform[0]) {
-	transformPoint(info->accel_state->transform[0], &srcTopLeft);
-	transformPoint(info->accel_state->transform[0], &srcTopRight);
-	transformPoint(info->accel_state->transform[0], &srcBottomLeft);
-	transformPoint(info->accel_state->transform[0], &srcBottomRight);
+	if ((info->ChipFamily < CHIP_FAMILY_R300) || !info->accel_state->has_tcl) {
+	    transformPoint(info->accel_state->transform[0], &srcTopLeft);
+	    transformPoint(info->accel_state->transform[0], &srcTopRight);
+	    transformPoint(info->accel_state->transform[0], &srcBottomLeft);
+	    transformPoint(info->accel_state->transform[0], &srcBottomRight);
+	}
     }
 
     if (info->accel_state->has_mask) {
@@ -2071,10 +2130,12 @@ static void FUNC_NAME(RadeonCompositeTile)(ScrnInfoPtr pScrn,
 	maskBottomRight.y = IntToxFixed(maskY + h);
 
 	if (info->accel_state->is_transform[1]) {
-	    transformPoint(info->accel_state->transform[1], &maskTopLeft);
-	    transformPoint(info->accel_state->transform[1], &maskTopRight);
-	    transformPoint(info->accel_state->transform[1], &maskBottomLeft);
-	    transformPoint(info->accel_state->transform[1], &maskBottomRight);
+	    if ((info->ChipFamily < CHIP_FAMILY_R300) || !info->accel_state->has_tcl) {
+		transformPoint(info->accel_state->transform[1], &maskTopLeft);
+		transformPoint(info->accel_state->transform[1], &maskTopRight);
+		transformPoint(info->accel_state->transform[1], &maskBottomLeft);
+		transformPoint(info->accel_state->transform[1], &maskBottomRight);
+	    }
 	}
 
 	vtx_count = 6;
diff --git a/src/radeon_reg.h b/src/radeon_reg.h
index 248cb427..9261b39d 100644
--- a/src/radeon_reg.h
+++ b/src/radeon_reg.h
@@ -4284,6 +4284,12 @@
 #define R300_VAP_PVS_CODE_CNTL_1			0x22D8
 #       define R300_PVS_LAST_VTX_SRC_INST_SHIFT         0
 #define R300_VAP_PVS_VECTOR_INDX_REG		        0x2200
+#       define R300_PVS_CODE_START                      0
+#       define R300_PVS_CONST_START                     512
+#       define R500_PVS_CONST_START                     1024
+#       define R300_PVS_VECTOR_INST_INDEX(x)            ((x) + R300_PVS_CODE_START)
+#       define R300_PVS_VECTOR_CONST_INDEX(x)           ((x) + R300_PVS_CONST_START)
+#       define R500_PVS_VECTOR_CONST_INDEX(x)           ((x) + R500_PVS_CONST_START)
 #define R300_VAP_PVS_VECTOR_DATA_REG		        0x2204
 /* PVS instructions */
 /* Opcode and dst instruction */
@@ -4402,6 +4408,10 @@
 #define R300_PVS_SRC_ADDR_SEL(x)                        ((x) << 29)
 #define R300_PVS_SRC_ADDR_MODE_1                        (1 << 31)
 
+#define R300_VAP_PVS_CONST_CNTL                         0x22d4
+#       define R300_PVS_CONST_BASE_OFFSET(x)            ((x) << 0)
+#       define R300_PVS_MAX_CONST_ADDR(x)               ((x) << 16)
+
 #define R300_VAP_PVS_FLOW_CNTL_OPC		        0x22dc
 #define R300_VAP_OUT_VTX_FMT_0			        0x2090
 #       define R300_VTX_POS_PRESENT                     (1 << 0)
@@ -5441,9 +5451,6 @@
 #   define R500_W_SRC_US				(0 << 2)
 #   define R500_W_SRC_RAS				(1 << 2)
 
-#define R500_GA_US_VECTOR_INDEX 0x4250
-#define R500_GA_US_VECTOR_DATA 0x4254
-
 #define R500_RS_INST_0					0x4320
 #define R500_RS_INST_1					0x4324
 #   define R500_RS_INST_TEX_ID_SHIFT			0
diff --git a/src/radeon_textured_videofuncs.c b/src/radeon_textured_videofuncs.c
index bbc5caf9..8ead2a40 100644
--- a/src/radeon_textured_videofuncs.c
+++ b/src/radeon_textured_videofuncs.c
@@ -1213,26 +1213,26 @@ FUNC_NAME(R300DisplayTexturedVideo)(ScrnInfoPtr pScrn, RADEONPortPrivPtr pPriv)
 
     /* load the vertex shader
      * We pre-load vertex programs in RADEONInit3DEngine():
-     * - exa mask/Xv bicubic
-     * - exa no mask
+     * - exa
      * - Xv
+     * - Xv bicubic
      * Here we select the offset of the vertex program we want to use
      */
     if (info->accel_state->has_tcl) {
 	if (pPriv->bicubic_enabled) {
 	    OUT_ACCEL_REG(R300_VAP_PVS_CODE_CNTL_0,
-			  ((0 << R300_PVS_FIRST_INST_SHIFT) |
-			   (2 << R300_PVS_XYZW_VALID_INST_SHIFT) |
-			   (2 << R300_PVS_LAST_INST_SHIFT)));
+			  ((11 << R300_PVS_FIRST_INST_SHIFT) |
+			   (13 << R300_PVS_XYZW_VALID_INST_SHIFT) |
+			   (13 << R300_PVS_LAST_INST_SHIFT)));
 	    OUT_ACCEL_REG(R300_VAP_PVS_CODE_CNTL_1,
-			  (2 << R300_PVS_LAST_VTX_SRC_INST_SHIFT));
+			  (13 << R300_PVS_LAST_VTX_SRC_INST_SHIFT));
 	} else {
 	    OUT_ACCEL_REG(R300_VAP_PVS_CODE_CNTL_0,
-			  ((5 << R300_PVS_FIRST_INST_SHIFT) |
-			   (6 << R300_PVS_XYZW_VALID_INST_SHIFT) |
-			   (6 << R300_PVS_LAST_INST_SHIFT)));
+			  ((9 << R300_PVS_FIRST_INST_SHIFT) |
+			   (10 << R300_PVS_XYZW_VALID_INST_SHIFT) |
+			   (10 << R300_PVS_LAST_INST_SHIFT)));
 	    OUT_ACCEL_REG(R300_VAP_PVS_CODE_CNTL_1,
-			  (6 << R300_PVS_LAST_VTX_SRC_INST_SHIFT));
+			  (10 << R300_PVS_LAST_VTX_SRC_INST_SHIFT));
 	}
     }
 
@@ -2643,26 +2643,26 @@ FUNC_NAME(R500DisplayTexturedVideo)(ScrnInfoPtr pScrn, RADEONPortPrivPtr pPriv)
 
     /* load the vertex shader
      * We pre-load vertex programs in RADEONInit3DEngine():
-     * - exa mask/Xv bicubic
-     * - exa no mask
+     * - exa
      * - Xv
+     * - Xv bicubic
      * Here we select the offset of the vertex program we want to use
      */
     if (info->accel_state->has_tcl) {
 	if (pPriv->bicubic_enabled) {
 	    OUT_ACCEL_REG(R300_VAP_PVS_CODE_CNTL_0,
-			  ((0 << R300_PVS_FIRST_INST_SHIFT) |
-			   (2 << R300_PVS_XYZW_VALID_INST_SHIFT) |
-			   (2 << R300_PVS_LAST_INST_SHIFT)));
+			  ((11 << R300_PVS_FIRST_INST_SHIFT) |
+			   (13 << R300_PVS_XYZW_VALID_INST_SHIFT) |
+			   (13 << R300_PVS_LAST_INST_SHIFT)));
 	    OUT_ACCEL_REG(R300_VAP_PVS_CODE_CNTL_1,
-			  (2 << R300_PVS_LAST_VTX_SRC_INST_SHIFT));
+			  (13 << R300_PVS_LAST_VTX_SRC_INST_SHIFT));
 	} else {
 	    OUT_ACCEL_REG(R300_VAP_PVS_CODE_CNTL_0,
-			  ((5 << R300_PVS_FIRST_INST_SHIFT) |
-			   (6 << R300_PVS_XYZW_VALID_INST_SHIFT) |
-			   (6 << R300_PVS_LAST_INST_SHIFT)));
+			  ((9 << R300_PVS_FIRST_INST_SHIFT) |
+			   (10 << R300_PVS_XYZW_VALID_INST_SHIFT) |
+			   (10 << R300_PVS_LAST_INST_SHIFT)));
 	    OUT_ACCEL_REG(R300_VAP_PVS_CODE_CNTL_1,
-			  (6 << R300_PVS_LAST_VTX_SRC_INST_SHIFT));
+			  (10 << R300_PVS_LAST_VTX_SRC_INST_SHIFT));
 	}
     }