Heavy optimizations.

1 files changed, 58 insertions, 225 deletions

diff --git a/src/radeon_textured_videofuncs.c b/src/radeon_textured_videofuncs.c
index 14fc5ab9..59f569de 100644
--- a/src/radeon_textured_videofuncs.c
+++ b/src/radeon_textured_videofuncs.c
@@ -414,20 +414,20 @@ FUNC_NAME(RADEONDisplayTexturedVideo)(ScrnInfoPtr pScrn, RADEONPortPrivPtr pPriv
 		OUT_VIDEO_REG(R300_RS_INST_COUNT, R300_INST_COUNT_RS(1) | R300_TX_OFFSET_RS(6));
 
 		/* Pixel stack frame size. */
-		OUT_VIDEO_REG(R500_US_PIXSIZE, R500_PIX_SIZE(12));
+		OUT_VIDEO_REG(R500_US_PIXSIZE, R500_PIX_SIZE(5));
 
 		/* FP length. */
 		OUT_VIDEO_REG(R500_US_CODE_ADDR, (R500_US_CODE_START_ADDR(0) |
-							R500_US_CODE_END_ADDR(19)));
+							R500_US_CODE_END_ADDR(13)));
 		OUT_VIDEO_REG(R500_US_CODE_RANGE, (R500_US_CODE_RANGE_ADDR(0) |
-							R500_US_CODE_RANGE_SIZE(19)));
+							R500_US_CODE_RANGE_SIZE(13)));
 
 		/* Prepare for FP emission. */
 		OUT_VIDEO_REG(R500_US_CODE_OFFSET, 0);
 		OUT_VIDEO_REG(R500_GA_US_VECTOR_INDEX, 0);
 		FINISH_VIDEO();
 
-		BEGIN_VIDEO(123);
+		BEGIN_VIDEO(93);
 		/* Pixel shader.
 		 * I've gone ahead and annotated each instruction, since this
 		 * thing is MASSIVE. :3
@@ -439,8 +439,7 @@ FUNC_NAME(RADEONDisplayTexturedVideo)(ScrnInfoPtr pScrn, RADEONPortPrivPtr pPriv
 						   R500_INST_TEX_SEM_WAIT |
 						   R500_INST_RGB_WMASK_R |
 						   R500_INST_RGB_WMASK_G |
-						   R500_INST_RGB_WMASK_B |
-						   R500_INST_ALPHA_WMASK));
+						   R500_INST_RGB_WMASK_B));
 		OUT_VIDEO_REG(R500_GA_US_VECTOR_DATA, (R500_TEX_ID(1) |
 						   R500_TEX_INST_LD |
 						   R500_TEX_SEM_ACQUIRE |
@@ -464,8 +463,7 @@ FUNC_NAME(RADEONDisplayTexturedVideo)(ScrnInfoPtr pScrn, RADEONPortPrivPtr pPriv
 						   R500_INST_TEX_SEM_WAIT |
 						   R500_INST_RGB_WMASK_R |
 						   R500_INST_RGB_WMASK_G |
-						   R500_INST_RGB_WMASK_B |
-						   R500_INST_ALPHA_WMASK));
+						   R500_INST_RGB_WMASK_B));
 		OUT_VIDEO_REG(R500_GA_US_VECTOR_DATA, (R500_TEX_ID(1) |
 						   R500_TEX_INST_LD |
 						   R500_TEX_SEM_ACQUIRE |
@@ -475,7 +473,7 @@ FUNC_NAME(RADEONDisplayTexturedVideo)(ScrnInfoPtr pScrn, RADEONPortPrivPtr pPriv
 						   R500_TEX_SRC_T_SWIZ_G |
 						   R500_TEX_SRC_R_SWIZ_G |
 						   R500_TEX_SRC_Q_SWIZ_G |
-						   R500_TEX_DST_ADDR(3) |
+						   R500_TEX_DST_ADDR(1) |
 						   R500_TEX_DST_R_SWIZ_R |
 						   R500_TEX_DST_G_SWIZ_G |
 						   R500_TEX_DST_B_SWIZ_B |
@@ -500,51 +498,18 @@ FUNC_NAME(RADEONDisplayTexturedVideo)(ScrnInfoPtr pScrn, RADEONPortPrivPtr pPriv
 		OUT_VIDEO_REG(R500_GA_US_VECTOR_DATA, (R500_ALU_RGB_SEL_A_SRC0 |
 						   R500_ALU_RGB_R_SWIZ_A_R |
 						   R500_ALU_RGB_G_SWIZ_A_G |
-						   R500_ALU_RGB_B_SWIZ_A_B |
+						   R500_ALU_RGB_B_SWIZ_A_R |
 						   R500_ALU_RGB_SEL_B_SRC1 |
 						   R500_ALU_RGB_R_SWIZ_B_G |
 						   R500_ALU_RGB_G_SWIZ_B_G |
-						   R500_ALU_RGB_B_SWIZ_B_G));
-		OUT_VIDEO_REG(R500_GA_US_VECTOR_DATA, (R500_ALPHA_ADDRD(4) |
-						   R500_ALPHA_OP_MAD |
-						   R500_ALPHA_SEL_A_SRC0 |
-						   R500_ALPHA_SWIZ_A_A |
-						   R500_ALPHA_SEL_B_SRC1 |
-						   R500_ALPHA_SWIZ_B_G));
-		OUT_VIDEO_REG(R500_GA_US_VECTOR_DATA, (R500_ALU_RGBA_ADDRD(4) |
-						   R500_ALU_RGBA_OP_MAD |
-						   R500_ALU_RGBA_R_SWIZ_0 |
-						   R500_ALU_RGBA_G_SWIZ_0 |
-						   R500_ALU_RGBA_B_SWIZ_0 |
-						   R500_ALU_RGBA_A_SWIZ_0));
-
-		/* MUL temp3, const0, temp0.xxxx */
-		OUT_VIDEO_REG(R500_GA_US_VECTOR_DATA, (R500_INST_TYPE_ALU |
-						   R500_INST_RGB_WMASK_R |
-						   R500_INST_RGB_WMASK_G |
-						   R500_INST_RGB_WMASK_B |
-						   R500_INST_ALPHA_WMASK));
-		OUT_VIDEO_REG(R500_GA_US_VECTOR_DATA, (R500_RGB_ADDR0(0) |
-						   R500_RGB_ADDR0_CONST |
-						   R500_RGB_ADDR1(2)));
-		OUT_VIDEO_REG(R500_GA_US_VECTOR_DATA, (R500_ALPHA_ADDR0(0) |
-						   R500_ALPHA_ADDR0_CONST |
-						   R500_ALPHA_ADDR1(2)));
-		OUT_VIDEO_REG(R500_GA_US_VECTOR_DATA, (R500_ALU_RGB_SEL_A_SRC0 |
-						   R500_ALU_RGB_R_SWIZ_A_R |
-						   R500_ALU_RGB_G_SWIZ_A_G |
-						   R500_ALU_RGB_B_SWIZ_A_B |
-						   R500_ALU_RGB_SEL_B_SRC1 |
-						   R500_ALU_RGB_R_SWIZ_B_R |
-						   R500_ALU_RGB_G_SWIZ_B_R |
 						   R500_ALU_RGB_B_SWIZ_B_R));
-		OUT_VIDEO_REG(R500_GA_US_VECTOR_DATA, (R500_ALPHA_ADDRD(5) |
+		OUT_VIDEO_REG(R500_GA_US_VECTOR_DATA, (R500_ALPHA_ADDRD(4) |
 						   R500_ALPHA_OP_MAD |
 						   R500_ALPHA_SEL_A_SRC0 |
-						   R500_ALPHA_SWIZ_A_A |
+						   R500_ALPHA_SWIZ_A_G |
 						   R500_ALPHA_SEL_B_SRC1 |
 						   R500_ALPHA_SWIZ_B_R));
-		OUT_VIDEO_REG(R500_GA_US_VECTOR_DATA, (R500_ALU_RGBA_ADDRD(5) |
+		OUT_VIDEO_REG(R500_GA_US_VECTOR_DATA, (R500_ALU_RGBA_ADDRD(4) |
 						   R500_ALU_RGBA_OP_MAD |
 						   R500_ALU_RGBA_R_SWIZ_0 |
 						   R500_ALU_RGBA_G_SWIZ_0 |
@@ -559,63 +524,27 @@ FUNC_NAME(RADEONDisplayTexturedVideo)(ScrnInfoPtr pScrn, RADEONPortPrivPtr pPriv
 						   R500_INST_ALPHA_WMASK));
 		OUT_VIDEO_REG(R500_GA_US_VECTOR_DATA, (R500_RGB_ADDR0(1) |
 						   R500_RGB_ADDR0_CONST |
-						   R500_RGB_ADDR1(3) |
+						   R500_RGB_ADDR1(1) |
 						   R500_RGB_ADDR2(4)));
 		OUT_VIDEO_REG(R500_GA_US_VECTOR_DATA, (R500_ALPHA_ADDR0(1) |
 						   R500_ALPHA_ADDR0_CONST |
-						   R500_ALPHA_ADDR1(3) |
+						   R500_ALPHA_ADDR1(1) |
 						   R500_ALPHA_ADDR2(4)));
 		OUT_VIDEO_REG(R500_GA_US_VECTOR_DATA, (R500_ALU_RGB_SEL_A_SRC0 |
 						   R500_ALU_RGB_R_SWIZ_A_R |
 						   R500_ALU_RGB_G_SWIZ_A_G |
-						   R500_ALU_RGB_B_SWIZ_A_B |
-						   R500_ALU_RGB_SEL_B_SRC1 |
-						   R500_ALU_RGB_R_SWIZ_B_G |
-						   R500_ALU_RGB_G_SWIZ_B_G |
-						   R500_ALU_RGB_B_SWIZ_B_G));
-		OUT_VIDEO_REG(R500_GA_US_VECTOR_DATA, (R500_ALPHA_ADDRD(6) |
-						   R500_ALPHA_OP_MAD |
-						   R500_ALPHA_SEL_A_SRC0 |
-						   R500_ALPHA_SWIZ_A_A |
-						   R500_ALPHA_SEL_B_SRC1 |
-						   R500_ALPHA_SWIZ_B_G));
-		OUT_VIDEO_REG(R500_GA_US_VECTOR_DATA, (R500_ALU_RGBA_ADDRD(6) |
-						   R500_ALU_RGBA_OP_MAD |
-						   R500_ALU_RGBA_SEL_C_SRC2 |
-						   R500_ALU_RGBA_R_SWIZ_R |
-						   R500_ALU_RGBA_G_SWIZ_G |
-						   R500_ALU_RGBA_B_SWIZ_B |
-						   R500_ALU_RGBA_A_SWIZ_A));
-
-		/* MAD temp5, const1, temp1.yyyy, temp3 */
-		OUT_VIDEO_REG(R500_GA_US_VECTOR_DATA, (R500_INST_TYPE_ALU |
-						   R500_INST_RGB_WMASK_R |
-						   R500_INST_RGB_WMASK_G |
-						   R500_INST_RGB_WMASK_B |
-						   R500_INST_ALPHA_WMASK));
-		OUT_VIDEO_REG(R500_GA_US_VECTOR_DATA, (R500_RGB_ADDR0(1) |
-						   R500_RGB_ADDR0_CONST |
-						   R500_RGB_ADDR1(3) |
-						   R500_RGB_ADDR2(5)));
-		OUT_VIDEO_REG(R500_GA_US_VECTOR_DATA, (R500_ALPHA_ADDR0(1) |
-						   R500_ALPHA_ADDR0_CONST |
-						   R500_ALPHA_ADDR1(3) |
-						   R500_ALPHA_ADDR2(5)));
-		OUT_VIDEO_REG(R500_GA_US_VECTOR_DATA, (R500_ALU_RGB_SEL_A_SRC0 |
-						   R500_ALU_RGB_R_SWIZ_A_R |
-						   R500_ALU_RGB_G_SWIZ_A_G |
-						   R500_ALU_RGB_B_SWIZ_A_B |
+						   R500_ALU_RGB_B_SWIZ_A_R |
 						   R500_ALU_RGB_SEL_B_SRC1 |
 						   R500_ALU_RGB_R_SWIZ_B_G |
 						   R500_ALU_RGB_G_SWIZ_B_G |
 						   R500_ALU_RGB_B_SWIZ_B_G));
-		OUT_VIDEO_REG(R500_GA_US_VECTOR_DATA, (R500_ALPHA_ADDRD(7) |
+		OUT_VIDEO_REG(R500_GA_US_VECTOR_DATA, (R500_ALPHA_ADDRD(5) |
 						   R500_ALPHA_OP_MAD |
 						   R500_ALPHA_SEL_A_SRC0 |
-						   R500_ALPHA_SWIZ_A_A |
+						   R500_ALPHA_SWIZ_A_G |
 						   R500_ALPHA_SEL_B_SRC1 |
 						   R500_ALPHA_SWIZ_B_G));
-		OUT_VIDEO_REG(R500_GA_US_VECTOR_DATA, (R500_ALU_RGBA_ADDRD(7) |
+		OUT_VIDEO_REG(R500_GA_US_VECTOR_DATA, (R500_ALU_RGBA_ADDRD(5) |
 						   R500_ALU_RGBA_OP_MAD |
 						   R500_ALU_RGBA_SEL_C_SRC2 |
 						   R500_ALU_RGBA_R_SWIZ_R |
@@ -631,16 +560,16 @@ FUNC_NAME(RADEONDisplayTexturedVideo)(ScrnInfoPtr pScrn, RADEONPortPrivPtr pPriv
 						   R500_INST_ALPHA_WMASK));
 		OUT_VIDEO_REG(R500_GA_US_VECTOR_DATA, (R500_RGB_ADDR0(1) |
 						   R500_RGB_ADDR0_CONST |
-						   R500_RGB_ADDR1(3) |
+						   R500_RGB_ADDR1(1) |
 						   R500_RGB_ADDR2(4)));
 		OUT_VIDEO_REG(R500_GA_US_VECTOR_DATA, (R500_ALPHA_ADDR0(1) |
 						   R500_ALPHA_ADDR0_CONST |
-						   R500_ALPHA_ADDR1(3) |
+						   R500_ALPHA_ADDR1(1) |
 						   R500_ALPHA_ADDR2(4)));
 		OUT_VIDEO_REG(R500_GA_US_VECTOR_DATA, (R500_ALU_RGB_SEL_A_SRC0 |
 						   R500_ALU_RGB_R_SWIZ_A_R |
 						   R500_ALU_RGB_G_SWIZ_A_G |
-						   R500_ALU_RGB_B_SWIZ_A_B |
+						   R500_ALU_RGB_B_SWIZ_A_R |
 						   R500_ALU_RGB_SEL_B_SRC1 |
 						   R500_ALU_RGB_R_SWIZ_B_R |
 						   R500_ALU_RGB_G_SWIZ_B_R |
@@ -648,7 +577,7 @@ FUNC_NAME(RADEONDisplayTexturedVideo)(ScrnInfoPtr pScrn, RADEONPortPrivPtr pPriv
 		OUT_VIDEO_REG(R500_GA_US_VECTOR_DATA, (R500_ALPHA_ADDRD(4) |
 						   R500_ALPHA_OP_MAD |
 						   R500_ALPHA_SEL_A_SRC0 |
-						   R500_ALPHA_SWIZ_A_A |
+						   R500_ALPHA_SWIZ_A_G |
 						   R500_ALPHA_SEL_B_SRC1 |
 						   R500_ALPHA_SWIZ_B_R));
 		OUT_VIDEO_REG(R500_GA_US_VECTOR_DATA, (R500_ALU_RGBA_ADDRD(4) |
@@ -659,42 +588,6 @@ FUNC_NAME(RADEONDisplayTexturedVideo)(ScrnInfoPtr pScrn, RADEONPortPrivPtr pPriv
 						   R500_ALU_RGBA_B_SWIZ_B |
 						   R500_ALU_RGBA_A_SWIZ_A));
 
-		/* MAD temp3, const1, temp1.xxxx, temp3 */
-		OUT_VIDEO_REG(R500_GA_US_VECTOR_DATA, (R500_INST_TYPE_ALU |
-						   R500_INST_RGB_WMASK_R |
-						   R500_INST_RGB_WMASK_G |
-						   R500_INST_RGB_WMASK_B |
-						   R500_INST_ALPHA_WMASK));
-		OUT_VIDEO_REG(R500_GA_US_VECTOR_DATA, (R500_RGB_ADDR0(1) |
-						   R500_RGB_ADDR0_CONST |
-						   R500_RGB_ADDR1(3) |
-						   R500_RGB_ADDR2(5)));
-		OUT_VIDEO_REG(R500_GA_US_VECTOR_DATA, (R500_ALPHA_ADDR0(1) |
-						   R500_ALPHA_ADDR0_CONST |
-						   R500_ALPHA_ADDR1(3) |
-						   R500_ALPHA_ADDR2(5)));
-		OUT_VIDEO_REG(R500_GA_US_VECTOR_DATA, (R500_ALU_RGB_SEL_A_SRC0 |
-						   R500_ALU_RGB_R_SWIZ_A_R |
-						   R500_ALU_RGB_G_SWIZ_A_G |
-						   R500_ALU_RGB_B_SWIZ_A_B |
-						   R500_ALU_RGB_SEL_B_SRC1 |
-						   R500_ALU_RGB_R_SWIZ_B_R |
-						   R500_ALU_RGB_G_SWIZ_B_R |
-						   R500_ALU_RGB_B_SWIZ_B_R));
-		OUT_VIDEO_REG(R500_GA_US_VECTOR_DATA, (R500_ALPHA_ADDRD(5) |
-						   R500_ALPHA_OP_MAD |
-						   R500_ALPHA_SEL_A_SRC0 |
-						   R500_ALPHA_SWIZ_A_A |
-						   R500_ALPHA_SEL_B_SRC1 |
-						   R500_ALPHA_SWIZ_B_R));
-		OUT_VIDEO_REG(R500_GA_US_VECTOR_DATA, (R500_ALU_RGBA_ADDRD(5) |
-						   R500_ALU_RGBA_OP_MAD |
-						   R500_ALU_RGBA_SEL_C_SRC2 |
-						   R500_ALU_RGBA_R_SWIZ_R |
-						   R500_ALU_RGBA_G_SWIZ_G |
-						   R500_ALU_RGBA_B_SWIZ_B |
-						   R500_ALU_RGBA_A_SWIZ_A));
-
 		/* ADD temp2, temp2, input0 */
 		OUT_VIDEO_REG(R500_GA_US_VECTOR_DATA, (R500_INST_TYPE_ALU |
 						   R500_INST_RGB_WMASK_R |
@@ -722,10 +615,10 @@ FUNC_NAME(RADEONDisplayTexturedVideo)(ScrnInfoPtr pScrn, RADEONPortPrivPtr pPriv
 						   R500_ALU_RGBA_SEL_C_SRC2 |
 						   R500_ALU_RGBA_R_SWIZ_R |
 						   R500_ALU_RGBA_G_SWIZ_G |
-						   R500_ALU_RGBA_B_SWIZ_B |
-						   R500_ALU_RGBA_A_SWIZ_A));
+						   R500_ALU_RGBA_B_SWIZ_R |
+						   R500_ALU_RGBA_A_SWIZ_G));
 
-		/* ADD temp3, temp3, input0 */
+		/* ADD temp4, temp4, input0 */
 		OUT_VIDEO_REG(R500_GA_US_VECTOR_DATA, (R500_INST_TYPE_ALU |
 						   R500_INST_RGB_WMASK_R |
 						   R500_INST_RGB_WMASK_G |
@@ -742,80 +635,20 @@ FUNC_NAME(RADEONDisplayTexturedVideo)(ScrnInfoPtr pScrn, RADEONPortPrivPtr pPriv
 						   R500_ALU_RGB_R_SWIZ_B_R |
 						   R500_ALU_RGB_G_SWIZ_B_G |
 						   R500_ALU_RGB_B_SWIZ_B_B));
-		OUT_VIDEO_REG(R500_GA_US_VECTOR_DATA, (R500_ALPHA_ADDRD(5) |
-						   R500_ALPHA_OP_MAD |
-						   R500_ALPHA_SWIZ_A_1 |
-						   R500_ALPHA_SEL_B_SRC1 |
-						   R500_ALPHA_SWIZ_B_A));
-		OUT_VIDEO_REG(R500_GA_US_VECTOR_DATA, (R500_ALU_RGBA_ADDRD(5) |
-						   R500_ALU_RGBA_OP_MAD |
-						   R500_ALU_RGBA_SEL_C_SRC2 |
-						   R500_ALU_RGBA_R_SWIZ_R |
-						   R500_ALU_RGBA_G_SWIZ_G |
-						   R500_ALU_RGBA_B_SWIZ_B |
-						   R500_ALU_RGBA_A_SWIZ_A));
-
-		/* ADD temp4, temp4, input0 */
-		OUT_VIDEO_REG(R500_GA_US_VECTOR_DATA, (R500_INST_TYPE_ALU |
-						   R500_INST_RGB_WMASK_R |
-						   R500_INST_RGB_WMASK_G |
-						   R500_INST_RGB_WMASK_B |
-						   R500_INST_ALPHA_WMASK));
-		OUT_VIDEO_REG(R500_GA_US_VECTOR_DATA, (R500_RGB_ADDR1(6) |
-						   R500_RGB_ADDR2(0)));
-		OUT_VIDEO_REG(R500_GA_US_VECTOR_DATA, (R500_ALPHA_ADDR1(6) |
-						   R500_ALPHA_ADDR2(0)));
-		OUT_VIDEO_REG(R500_GA_US_VECTOR_DATA, (R500_ALU_RGB_R_SWIZ_A_1 |
-						   R500_ALU_RGB_G_SWIZ_A_1 |
-						   R500_ALU_RGB_B_SWIZ_A_1 |
-						   R500_ALU_RGB_SEL_B_SRC1 |
-						   R500_ALU_RGB_R_SWIZ_B_R |
-						   R500_ALU_RGB_G_SWIZ_B_G |
-						   R500_ALU_RGB_B_SWIZ_B_B));
-		OUT_VIDEO_REG(R500_GA_US_VECTOR_DATA, (R500_ALPHA_ADDRD(6) |
-						   R500_ALPHA_OP_MAD |
-						   R500_ALPHA_SWIZ_A_1 |
-						   R500_ALPHA_SEL_B_SRC1 |
-						   R500_ALPHA_SWIZ_B_A));
-		OUT_VIDEO_REG(R500_GA_US_VECTOR_DATA, (R500_ALU_RGBA_ADDRD(6) |
-						   R500_ALU_RGBA_OP_MAD |
-						   R500_ALU_RGBA_SEL_C_SRC2 |
-						   R500_ALU_RGBA_R_SWIZ_R |
-						   R500_ALU_RGBA_G_SWIZ_G |
-						   R500_ALU_RGBA_B_SWIZ_B |
-						   R500_ALU_RGBA_A_SWIZ_A));
-
-		/* ADD temp5, temp5, input0 */
-		OUT_VIDEO_REG(R500_GA_US_VECTOR_DATA, (R500_INST_TYPE_ALU |
-						   R500_INST_RGB_WMASK_R |
-						   R500_INST_RGB_WMASK_G |
-						   R500_INST_RGB_WMASK_B |
-						   R500_INST_ALPHA_WMASK));
-		OUT_VIDEO_REG(R500_GA_US_VECTOR_DATA, (R500_RGB_ADDR1(7) |
-						   R500_RGB_ADDR2(0)));
-		OUT_VIDEO_REG(R500_GA_US_VECTOR_DATA, (R500_ALPHA_ADDR1(7) |
-						   R500_ALPHA_ADDR2(0)));
-		OUT_VIDEO_REG(R500_GA_US_VECTOR_DATA, (R500_ALU_RGB_R_SWIZ_A_1 |
-						   R500_ALU_RGB_G_SWIZ_A_1 |
-						   R500_ALU_RGB_B_SWIZ_A_1 |
-						   R500_ALU_RGB_SEL_B_SRC1 |
-						   R500_ALU_RGB_R_SWIZ_B_R |
-						   R500_ALU_RGB_G_SWIZ_B_G |
-						   R500_ALU_RGB_B_SWIZ_B_B));
-		OUT_VIDEO_REG(R500_GA_US_VECTOR_DATA, (R500_ALPHA_ADDRD(7) |
+		OUT_VIDEO_REG(R500_GA_US_VECTOR_DATA, (R500_ALPHA_ADDRD(0) |
 						   R500_ALPHA_OP_MAD |
 						   R500_ALPHA_SWIZ_A_1 |
 						   R500_ALPHA_SEL_B_SRC1 |
 						   R500_ALPHA_SWIZ_B_A));
-		OUT_VIDEO_REG(R500_GA_US_VECTOR_DATA, (R500_ALU_RGBA_ADDRD(7) |
+		OUT_VIDEO_REG(R500_GA_US_VECTOR_DATA, (R500_ALU_RGBA_ADDRD(0) |
 						   R500_ALU_RGBA_OP_MAD |
 						   R500_ALU_RGBA_SEL_C_SRC2 |
 						   R500_ALU_RGBA_R_SWIZ_R |
 						   R500_ALU_RGBA_G_SWIZ_G |
-						   R500_ALU_RGBA_B_SWIZ_B |
-						   R500_ALU_RGBA_A_SWIZ_A));
+						   R500_ALU_RGBA_B_SWIZ_R |
+						   R500_ALU_RGBA_A_SWIZ_G));
 
-		/* TEX temp2, temp2, tex0, 1D */
+		/* TEX temp3, temp3, tex0, 1D */
 		OUT_VIDEO_REG(R500_GA_US_VECTOR_DATA, (R500_INST_TYPE_TEX |
 						   R500_INST_TEX_SEM_WAIT |
 						   R500_INST_RGB_WMASK_R |
@@ -827,11 +660,11 @@ FUNC_NAME(RADEONDisplayTexturedVideo)(ScrnInfoPtr pScrn, RADEONPortPrivPtr pPriv
 						   R500_TEX_SEM_ACQUIRE |
 						   R500_TEX_IGNORE_UNCOVERED));
 		OUT_VIDEO_REG(R500_GA_US_VECTOR_DATA, (R500_TEX_SRC_ADDR(4) |
-						   R500_TEX_SRC_S_SWIZ_R |
-						   R500_TEX_SRC_T_SWIZ_G |
-						   R500_TEX_SRC_R_SWIZ_B |
-						   R500_TEX_SRC_Q_SWIZ_A |
-						   R500_TEX_DST_ADDR(4) |
+						   R500_TEX_SRC_S_SWIZ_B |
+						   R500_TEX_SRC_T_SWIZ_A |
+						   R500_TEX_SRC_R_SWIZ_R |
+						   R500_TEX_SRC_Q_SWIZ_G |
+						   R500_TEX_DST_ADDR(5) |
 						   R500_TEX_DST_R_SWIZ_R |
 						   R500_TEX_DST_G_SWIZ_G |
 						   R500_TEX_DST_B_SWIZ_B |
@@ -840,7 +673,7 @@ FUNC_NAME(RADEONDisplayTexturedVideo)(ScrnInfoPtr pScrn, RADEONPortPrivPtr pPriv
 		OUT_VIDEO_REG(R500_GA_US_VECTOR_DATA, 0x00000000);
 		OUT_VIDEO_REG(R500_GA_US_VECTOR_DATA, 0x00000000);
 
-		/* TEX temp3, temp3, tex0, 1D */
+		/* TEX temp2, temp2, tex0, 1D */
 		OUT_VIDEO_REG(R500_GA_US_VECTOR_DATA, (R500_INST_TYPE_TEX |
 						   R500_INST_TEX_SEM_WAIT |
 						   R500_INST_RGB_WMASK_R |
@@ -851,12 +684,12 @@ FUNC_NAME(RADEONDisplayTexturedVideo)(ScrnInfoPtr pScrn, RADEONPortPrivPtr pPriv
 						   R500_TEX_INST_LD |
 						   R500_TEX_SEM_ACQUIRE |
 						   R500_TEX_IGNORE_UNCOVERED));
-		OUT_VIDEO_REG(R500_GA_US_VECTOR_DATA, (R500_TEX_SRC_ADDR(5) |
+		OUT_VIDEO_REG(R500_GA_US_VECTOR_DATA, (R500_TEX_SRC_ADDR(4) |
 						   R500_TEX_SRC_S_SWIZ_R |
 						   R500_TEX_SRC_T_SWIZ_G |
 						   R500_TEX_SRC_R_SWIZ_B |
 						   R500_TEX_SRC_Q_SWIZ_A |
-						   R500_TEX_DST_ADDR(5) |
+						   R500_TEX_DST_ADDR(4) |
 						   R500_TEX_DST_R_SWIZ_R |
 						   R500_TEX_DST_G_SWIZ_G |
 						   R500_TEX_DST_B_SWIZ_B |
@@ -865,7 +698,7 @@ FUNC_NAME(RADEONDisplayTexturedVideo)(ScrnInfoPtr pScrn, RADEONPortPrivPtr pPriv
 		OUT_VIDEO_REG(R500_GA_US_VECTOR_DATA, 0x00000000);
 		OUT_VIDEO_REG(R500_GA_US_VECTOR_DATA, 0x00000000);
 
-		/* TEX temp4, temp4, tex0, 1D */
+		/* TEX temp5, temp5, tex0, 1D */
 		OUT_VIDEO_REG(R500_GA_US_VECTOR_DATA, (R500_INST_TYPE_TEX |
 						   R500_INST_TEX_SEM_WAIT |
 						   R500_INST_RGB_WMASK_R |
@@ -876,12 +709,12 @@ FUNC_NAME(RADEONDisplayTexturedVideo)(ScrnInfoPtr pScrn, RADEONPortPrivPtr pPriv
 						   R500_TEX_INST_LD |
 						   R500_TEX_SEM_ACQUIRE |
 						   R500_TEX_IGNORE_UNCOVERED));
-		OUT_VIDEO_REG(R500_GA_US_VECTOR_DATA, (R500_TEX_SRC_ADDR(6) |
-						   R500_TEX_SRC_S_SWIZ_R |
-						   R500_TEX_SRC_T_SWIZ_G |
+		OUT_VIDEO_REG(R500_GA_US_VECTOR_DATA, (R500_TEX_SRC_ADDR(0) |
+						   R500_TEX_SRC_S_SWIZ_B |
+						   R500_TEX_SRC_T_SWIZ_A |
 						   R500_TEX_SRC_R_SWIZ_B |
 						   R500_TEX_SRC_Q_SWIZ_A |
-						   R500_TEX_DST_ADDR(6) |
+						   R500_TEX_DST_ADDR(3) |
 						   R500_TEX_DST_R_SWIZ_R |
 						   R500_TEX_DST_G_SWIZ_G |
 						   R500_TEX_DST_B_SWIZ_B |
@@ -890,7 +723,7 @@ FUNC_NAME(RADEONDisplayTexturedVideo)(ScrnInfoPtr pScrn, RADEONPortPrivPtr pPriv
 		OUT_VIDEO_REG(R500_GA_US_VECTOR_DATA, 0x00000000);
 		OUT_VIDEO_REG(R500_GA_US_VECTOR_DATA, 0x00000000);
 
-		/* TEX temp5, temp5, tex0, 1D */
+		/* TEX temp4, temp4, tex0, 1D */
 		OUT_VIDEO_REG(R500_GA_US_VECTOR_DATA, (R500_INST_TYPE_TEX |
 						   R500_INST_TEX_SEM_WAIT |
 						   R500_INST_RGB_WMASK_R |
@@ -901,12 +734,12 @@ FUNC_NAME(RADEONDisplayTexturedVideo)(ScrnInfoPtr pScrn, RADEONPortPrivPtr pPriv
 						   R500_TEX_INST_LD |
 						   R500_TEX_SEM_ACQUIRE |
 						   R500_TEX_IGNORE_UNCOVERED));
-		OUT_VIDEO_REG(R500_GA_US_VECTOR_DATA, (R500_TEX_SRC_ADDR(7) |
+		OUT_VIDEO_REG(R500_GA_US_VECTOR_DATA, (R500_TEX_SRC_ADDR(0) |
 						   R500_TEX_SRC_S_SWIZ_R |
 						   R500_TEX_SRC_T_SWIZ_G |
 						   R500_TEX_SRC_R_SWIZ_B |
 						   R500_TEX_SRC_Q_SWIZ_A |
-						   R500_TEX_DST_ADDR(7) |
+						   R500_TEX_DST_ADDR(0) |
 						   R500_TEX_DST_R_SWIZ_R |
 						   R500_TEX_DST_G_SWIZ_G |
 						   R500_TEX_DST_B_SWIZ_B |
@@ -924,14 +757,14 @@ FUNC_NAME(RADEONDisplayTexturedVideo)(ScrnInfoPtr pScrn, RADEONPortPrivPtr pPriv
 						   R500_INST_RGB_WMASK_G |
 						   R500_INST_RGB_WMASK_B |
 						   R500_INST_ALPHA_WMASK));
-		OUT_VIDEO_REG(R500_GA_US_VECTOR_DATA, (R500_RGB_ADDR0(7) |
+		OUT_VIDEO_REG(R500_GA_US_VECTOR_DATA, (R500_RGB_ADDR0(3) |
 						   R500_RGB_SRCP_OP_RGB1_MINUS_RGB0 |
 						   R500_RGB_ADDR1(5) |
-						   R500_RGB_ADDR2(3)));
-		OUT_VIDEO_REG(R500_GA_US_VECTOR_DATA, (R500_ALPHA_ADDR0(7) |
+						   R500_RGB_ADDR2(1)));
+		OUT_VIDEO_REG(R500_GA_US_VECTOR_DATA, (R500_ALPHA_ADDR0(3) |
 						   R500_ALPHA_SRCP_OP_A1_MINUS_A0 |
 						   R500_ALPHA_ADDR1(5) |
-						   R500_ALPHA_ADDR1(3)));
+						   R500_ALPHA_ADDR1(1)));
 		OUT_VIDEO_REG(R500_GA_US_VECTOR_DATA, (R500_ALU_RGB_SEL_A_SRC2 |
 						   R500_ALU_RGB_R_SWIZ_A_B |
 						   R500_ALU_RGB_G_SWIZ_A_B |
@@ -963,14 +796,14 @@ FUNC_NAME(RADEONDisplayTexturedVideo)(ScrnInfoPtr pScrn, RADEONPortPrivPtr pPriv
 						   R500_INST_RGB_WMASK_G |
 						   R500_INST_RGB_WMASK_B |
 						   R500_INST_ALPHA_WMASK));
-		OUT_VIDEO_REG(R500_GA_US_VECTOR_DATA, (R500_RGB_ADDR0(6) |
+		OUT_VIDEO_REG(R500_GA_US_VECTOR_DATA, (R500_RGB_ADDR0(0) |
 						   R500_RGB_SRCP_OP_RGB1_MINUS_RGB0 |
 						   R500_RGB_ADDR1(4) |
-						   R500_RGB_ADDR2(3)));
-		OUT_VIDEO_REG(R500_GA_US_VECTOR_DATA, (R500_ALPHA_ADDR0(6) |
+						   R500_RGB_ADDR2(1)));
+		OUT_VIDEO_REG(R500_GA_US_VECTOR_DATA, (R500_ALPHA_ADDR0(0) |
 						   R500_ALPHA_SRCP_OP_A1_MINUS_A0 |
 						   R500_ALPHA_ADDR1(4) |
-						   R500_ALPHA_ADDR1(3)));
+						   R500_ALPHA_ADDR1(1)));
 		OUT_VIDEO_REG(R500_GA_US_VECTOR_DATA, (R500_ALU_RGB_SEL_A_SRC2 |
 						   R500_ALU_RGB_R_SWIZ_A_B |
 						   R500_ALU_RGB_G_SWIZ_A_B |
@@ -979,13 +812,13 @@ FUNC_NAME(RADEONDisplayTexturedVideo)(ScrnInfoPtr pScrn, RADEONPortPrivPtr pPriv
 						   R500_ALU_RGB_R_SWIZ_B_R |
 						   R500_ALU_RGB_G_SWIZ_B_G |
 						   R500_ALU_RGB_B_SWIZ_B_B));
-		OUT_VIDEO_REG(R500_GA_US_VECTOR_DATA, (R500_ALPHA_ADDRD(4) |
+		OUT_VIDEO_REG(R500_GA_US_VECTOR_DATA, (R500_ALPHA_ADDRD(0) |
 						   R500_ALPHA_OP_MAD |
 						   R500_ALPHA_SEL_A_SRC2 |
 						   R500_ALPHA_SWIZ_A_B |
 						   R500_ALPHA_SEL_B_SRCP |
 						   R500_ALPHA_SWIZ_B_A));
-		OUT_VIDEO_REG(R500_GA_US_VECTOR_DATA, (R500_ALU_RGBA_ADDRD(4) |
+		OUT_VIDEO_REG(R500_GA_US_VECTOR_DATA, (R500_ALU_RGBA_ADDRD(0) |
 						   R500_ALU_RGBA_OP_MAD |
 						   R500_ALU_RGBA_SEL_C_SRC0 |
 						   R500_ALU_RGBA_R_SWIZ_R |
@@ -1008,11 +841,11 @@ FUNC_NAME(RADEONDisplayTexturedVideo)(ScrnInfoPtr pScrn, RADEONPortPrivPtr pPriv
 						   R500_INST_RGB_OMASK_G |
 						   R500_INST_RGB_OMASK_B |
 						   R500_INST_ALPHA_OMASK));
-		OUT_VIDEO_REG(R500_GA_US_VECTOR_DATA, (R500_RGB_ADDR0(4) |
+		OUT_VIDEO_REG(R500_GA_US_VECTOR_DATA, (R500_RGB_ADDR0(0) |
 						   R500_RGB_SRCP_OP_RGB1_MINUS_RGB0 |
 						   R500_RGB_ADDR1(5) |
 						   R500_RGB_ADDR2(2)));
-		OUT_VIDEO_REG(R500_GA_US_VECTOR_DATA, (R500_ALPHA_ADDR0(4) |
+		OUT_VIDEO_REG(R500_GA_US_VECTOR_DATA, (R500_ALPHA_ADDR0(0) |
 						   R500_ALPHA_SRCP_OP_A1_MINUS_A0 |
 						   R500_ALPHA_ADDR1(5) |
 						   R500_ALPHA_ADDR1(2)));
@@ -1046,7 +879,7 @@ FUNC_NAME(RADEONDisplayTexturedVideo)(ScrnInfoPtr pScrn, RADEONPortPrivPtr pPriv
 		OUT_VIDEO_REG_F(R500_GA_US_VECTOR_DATA, 0x0);
 		OUT_VIDEO_REG_F(R500_GA_US_VECTOR_DATA, 0x0);
 		OUT_VIDEO_REG_F(R500_GA_US_VECTOR_DATA, 0x0);
-		/* const1 = {0, 1 / -texture[0].height, 0, 0) */
+		/* const0 = {1 / texture[0].height, 0, 0, 0} */
 		OUT_VIDEO_REG_F(R500_GA_US_VECTOR_DATA, 0x0);
 		OUT_VIDEO_REG_F(R500_GA_US_VECTOR_DATA, (1.0/(float)pPriv->h));
 		OUT_VIDEO_REG_F(R500_GA_US_VECTOR_DATA, 0x0);