R6xx/R7xx: move shaders to r600_shader.c and fixup Xv PS

patches from Christian Koenig with some adjustments from me

1 files changed, 43 insertions, 2254 deletions

diff --git a/src/r600_exa.c b/src/r600_exa.c
index 633663c7..a44b6118 100644
--- a/src/r600_exa.c
+++ b/src/r600_exa.c
@@ -1316,7 +1316,6 @@ static Bool R600PrepareComposite(int op, PicturePtr pSrcPicture,
     uint32_t blendcntl, dst_format;
     cb_config_t cb_conf;
     shader_config_t vs_conf, ps_conf;
-    int i = 0;
     uint32_t ps[24];
 
     /* return FALSE; */
@@ -1441,221 +1440,10 @@ static Bool R600PrepareComposite(int op, PicturePtr pSrcPicture,
 	    }
 	}
 
-	/* 0 */
-	ps[i++] = CF_DWORD0(ADDR(8));
-	ps[i++] = CF_DWORD1(POP_COUNT(0),
-			    CF_CONST(0),
-			    COND(SQ_CF_COND_ACTIVE),
-			    I_COUNT(2),
-			    CALL_COUNT(0),
-			    END_OF_PROGRAM(0),
-			    VALID_PIXEL_MODE(0),
-			    CF_INST(SQ_CF_INST_TEX),
-			    WHOLE_QUAD_MODE(0),
-			    BARRIER(1));
-
-	/* 1 */
-	ps[i++] = CF_ALU_DWORD0(ADDR(3),
-				KCACHE_BANK0(0),
-				KCACHE_BANK1(0),
-				KCACHE_MODE0(SQ_CF_KCACHE_NOP));
-	ps[i++] = CF_ALU_DWORD1(KCACHE_MODE1(SQ_CF_KCACHE_NOP),
-				KCACHE_ADDR0(0),
-				KCACHE_ADDR1(0),
-				I_COUNT(4),
-				USES_WATERFALL(0),
-				CF_INST(SQ_CF_INST_ALU),
-				WHOLE_QUAD_MODE(0),
-				BARRIER(1));
-
-	/* 2 */
-	ps[i++] = CF_ALLOC_IMP_EXP_DWORD0(ARRAY_BASE(CF_PIXEL_MRT0),
-					  TYPE(SQ_EXPORT_PIXEL),
-					  RW_GPR(2),
-					  RW_REL(ABSOLUTE),
-					  INDEX_GPR(0),
-					  ELEM_SIZE(1));
-
-	ps[i++] = CF_ALLOC_IMP_EXP_DWORD1_SWIZ(SRC_SEL_X(SQ_SEL_X),
-					       SRC_SEL_Y(SQ_SEL_Y),
-					       SRC_SEL_Z(SQ_SEL_Z),
-					       SRC_SEL_W(SQ_SEL_W),
-					       R6xx_ELEM_LOOP(0),
-					       BURST_COUNT(1),
-					       END_OF_PROGRAM(1),
-					       VALID_PIXEL_MODE(0),
-					       CF_INST(SQ_CF_INST_EXPORT_DONE),
-					       WHOLE_QUAD_MODE(0),
-					       BARRIER(1));
-
-	/* 3 - alu 0 */
-	/* MUL gpr[2].x gpr[1].x gpr[0].x */
-	ps[i++] = ALU_DWORD0(SRC0_SEL(1),
-			     SRC0_REL(ABSOLUTE),
-			     SRC0_ELEM(ELEM_X),
-			     SRC0_NEG(0),
-			     SRC1_SEL(0),
-			     SRC1_REL(ABSOLUTE),
-			     SRC1_ELEM(ELEM_X),
-			     SRC1_NEG(0),
-			     INDEX_MODE(SQ_INDEX_LOOP),
-			     PRED_SEL(SQ_PRED_SEL_OFF),
-			     LAST(0));
-	ps[i++] = ALU_DWORD1_OP2(info->ChipFamily,
-				 SRC0_ABS(0),
-				 SRC1_ABS(0),
-				 UPDATE_EXECUTE_MASK(0),
-				 UPDATE_PRED(0),
-				 WRITE_MASK(1),
-				 FOG_MERGE(0),
-				 OMOD(SQ_ALU_OMOD_OFF),
-				 ALU_INST(SQ_OP2_INST_MUL),
-				 BANK_SWIZZLE(SQ_ALU_VEC_012),
-				 DST_GPR(2),
-				 DST_REL(ABSOLUTE),
-				 DST_ELEM(ELEM_X),
-				 CLAMP(1));
-	/* 4 - alu 1 */
-	/* MUL gpr[2].y gpr[1].y gpr[0].y */
-	ps[i++] = ALU_DWORD0(SRC0_SEL(1),
-			     SRC0_REL(ABSOLUTE),
-			     SRC0_ELEM(ELEM_Y),
-			     SRC0_NEG(0),
-			     SRC1_SEL(0),
-			     SRC1_REL(ABSOLUTE),
-			     SRC1_ELEM(ELEM_Y),
-			     SRC1_NEG(0),
-			     INDEX_MODE(SQ_INDEX_LOOP),
-			     PRED_SEL(SQ_PRED_SEL_OFF),
-			     LAST(0));
-	ps[i++] = ALU_DWORD1_OP2(info->ChipFamily,
-				 SRC0_ABS(0),
-				 SRC1_ABS(0),
-				 UPDATE_EXECUTE_MASK(0),
-				 UPDATE_PRED(0),
-				 WRITE_MASK(1),
-				 FOG_MERGE(0),
-				 OMOD(SQ_ALU_OMOD_OFF),
-				 ALU_INST(SQ_OP2_INST_MUL),
-				 BANK_SWIZZLE(SQ_ALU_VEC_012),
-				 DST_GPR(2),
-				 DST_REL(ABSOLUTE),
-				 DST_ELEM(ELEM_Y),
-				 CLAMP(1));
-	/* 5 - alu 2 */
-	/* MUL gpr[2].z gpr[1].z gpr[0].z */
-	ps[i++] = ALU_DWORD0(SRC0_SEL(1),
-			     SRC0_REL(ABSOLUTE),
-			     SRC0_ELEM(ELEM_Z),
-			     SRC0_NEG(0),
-			     SRC1_SEL(0),
-			     SRC1_REL(ABSOLUTE),
-			     SRC1_ELEM(ELEM_Z),
-			     SRC1_NEG(0),
-			     INDEX_MODE(SQ_INDEX_LOOP),
-			     PRED_SEL(SQ_PRED_SEL_OFF),
-			     LAST(0));
-	ps[i++] = ALU_DWORD1_OP2(info->ChipFamily,
-				 SRC0_ABS(0),
-				 SRC1_ABS(0),
-				 UPDATE_EXECUTE_MASK(0),
-				 UPDATE_PRED(0),
-				 WRITE_MASK(1),
-				 FOG_MERGE(0),
-				 OMOD(SQ_ALU_OMOD_OFF),
-				 ALU_INST(SQ_OP2_INST_MUL),
-				 BANK_SWIZZLE(SQ_ALU_VEC_012),
-				 DST_GPR(2),
-				 DST_REL(ABSOLUTE),
-				 DST_ELEM(ELEM_Z),
-				 CLAMP(1));
-	/* 6 - alu 3 */
-	/* MUL gpr[2].w gpr[1].w gpr[0].w */
-	ps[i++] = ALU_DWORD0(SRC0_SEL(1),
-			     SRC0_REL(ABSOLUTE),
-			     SRC0_ELEM(ELEM_W),
-			     SRC0_NEG(0),
-			     SRC1_SEL(0),
-			     SRC1_REL(ABSOLUTE),
-			     SRC1_ELEM(ELEM_W),
-			     SRC1_NEG(0),
-			     INDEX_MODE(SQ_INDEX_LOOP),
-			     PRED_SEL(SQ_PRED_SEL_OFF),
-			     LAST(1));
-	ps[i++] = ALU_DWORD1_OP2(info->ChipFamily,
-				 SRC0_ABS(0),
-				 SRC1_ABS(0),
-				 UPDATE_EXECUTE_MASK(0),
-				 UPDATE_PRED(0),
-				 WRITE_MASK(1),
-				 FOG_MERGE(0),
-				 OMOD(SQ_ALU_OMOD_OFF),
-				 ALU_INST(SQ_OP2_INST_MUL),
-				 BANK_SWIZZLE(SQ_ALU_VEC_012),
-				 DST_GPR(2),
-				 DST_REL(ABSOLUTE),
-				 DST_ELEM(ELEM_W),
-				 CLAMP(1));
-	/* 7 */
-	ps[i++] = 0x00000000;
-	ps[i++] = 0x00000000;
-
-	/* 8/9 - src */
-	ps[i++] = TEX_DWORD0(TEX_INST(SQ_TEX_INST_SAMPLE),
-			     BC_FRAC_MODE(0),
-			     FETCH_WHOLE_QUAD(0),
-			     RESOURCE_ID(0),
-			     SRC_GPR(0),
-			     SRC_REL(ABSOLUTE),
-			     R7xx_ALT_CONST(0));
-	ps[i++] = TEX_DWORD1(DST_GPR(0),
-			     DST_REL(ABSOLUTE),
-			     DST_SEL_X(src_r),
-			     DST_SEL_Y(src_g),
-			     DST_SEL_Z(src_b),
-			     DST_SEL_W(src_a),
-			     LOD_BIAS(0),
-			     COORD_TYPE_X(TEX_NORMALIZED),
-			     COORD_TYPE_Y(TEX_NORMALIZED),
-			     COORD_TYPE_Z(TEX_NORMALIZED),
-			     COORD_TYPE_W(TEX_NORMALIZED));
-	ps[i++] = TEX_DWORD2(OFFSET_X(0),
-			     OFFSET_Y(0),
-			     OFFSET_Z(0),
-			     SAMPLER_ID(0),
-			     SRC_SEL_X(SQ_SEL_X),
-			     SRC_SEL_Y(SQ_SEL_Y),
-			     SRC_SEL_Z(SQ_SEL_0),
-			     SRC_SEL_W(SQ_SEL_1));
-	ps[i++] = TEX_DWORD_PAD;
-	/* 10/11 - mask */
-	ps[i++] = TEX_DWORD0(TEX_INST(SQ_TEX_INST_SAMPLE),
-			     BC_FRAC_MODE(0),
-			     FETCH_WHOLE_QUAD(0),
-			     RESOURCE_ID(1),
-			     SRC_GPR(1),
-			     SRC_REL(ABSOLUTE),
-			     R7xx_ALT_CONST(0));
-	ps[i++] = TEX_DWORD1(DST_GPR(1),
-			     DST_REL(ABSOLUTE),
-			     DST_SEL_X(mask_r),
-			     DST_SEL_Y(mask_g),
-			     DST_SEL_Z(mask_b),
-			     DST_SEL_W(mask_a),
-			     LOD_BIAS(0),
-			     COORD_TYPE_X(TEX_NORMALIZED),
-			     COORD_TYPE_Y(TEX_NORMALIZED),
-			     COORD_TYPE_Z(TEX_NORMALIZED),
-			     COORD_TYPE_W(TEX_NORMALIZED));
-	ps[i++] = TEX_DWORD2(OFFSET_X(0),
-			     OFFSET_Y(0),
-			     OFFSET_Z(0),
-			     SAMPLER_ID(1),
-			     SRC_SEL_X(SQ_SEL_X),
-			     SRC_SEL_Y(SQ_SEL_Y),
-			     SRC_SEL_Z(SQ_SEL_0),
-			     SRC_SEL_W(SQ_SEL_1));
-	ps[i++] = TEX_DWORD_PAD;
+	R600_comp_mask_ps(info->ChipFamily, ps,
+			  src_a, src_r, src_g, src_b,
+			  mask_a, mask_r, mask_g, mask_b);
+
     } else {
 	int src_a, src_r, src_g, src_b;
 	/* setup pixel shader */
@@ -1675,67 +1463,9 @@ static Bool R600PrepareComposite(int op, PicturePtr pSrcPicture,
 	    src_a = SQ_SEL_W;
 	}
 
-	/* 0 */
-	ps[i++] = CF_DWORD0(ADDR(2));
-	ps[i++] = CF_DWORD1(POP_COUNT(0),
-			    CF_CONST(0),
-			    COND(SQ_CF_COND_ACTIVE),
-			    I_COUNT(1),
-			    CALL_COUNT(0),
-			    END_OF_PROGRAM(0),
-			    VALID_PIXEL_MODE(0),
-			    CF_INST(SQ_CF_INST_TEX),
-			    WHOLE_QUAD_MODE(0),
-			    BARRIER(1));
-	/* 1 */
-	ps[i++] = CF_ALLOC_IMP_EXP_DWORD0(ARRAY_BASE(CF_PIXEL_MRT0),
-					  TYPE(SQ_EXPORT_PIXEL),
-					  RW_GPR(0),
-					  RW_REL(ABSOLUTE),
-					  INDEX_GPR(0),
-					  ELEM_SIZE(1));
-
-	ps[i++] = CF_ALLOC_IMP_EXP_DWORD1_SWIZ(SRC_SEL_X(SQ_SEL_X),
-					       SRC_SEL_Y(SQ_SEL_Y),
-					       SRC_SEL_Z(SQ_SEL_Z),
-					       SRC_SEL_W(SQ_SEL_W),
-					       R6xx_ELEM_LOOP(0),
-					       BURST_COUNT(1),
-					       END_OF_PROGRAM(1),
-					       VALID_PIXEL_MODE(0),
-					       CF_INST(SQ_CF_INST_EXPORT_DONE),
-					       WHOLE_QUAD_MODE(0),
-					       BARRIER(1));
-
-
-	/* 2/3 - src */
-	ps[i++] = TEX_DWORD0(TEX_INST(SQ_TEX_INST_SAMPLE),
-			     BC_FRAC_MODE(0),
-			     FETCH_WHOLE_QUAD(0),
-			     RESOURCE_ID(0),
-			     SRC_GPR(0),
-			     SRC_REL(ABSOLUTE),
-			     R7xx_ALT_CONST(0));
-	ps[i++] = TEX_DWORD1(DST_GPR(0),
-			     DST_REL(ABSOLUTE),
-			     DST_SEL_X(src_r),
-			     DST_SEL_Y(src_g),
-			     DST_SEL_Z(src_b),
-			     DST_SEL_W(src_a),
-			     LOD_BIAS(0),
-			     COORD_TYPE_X(TEX_NORMALIZED),
-			     COORD_TYPE_Y(TEX_NORMALIZED),
-			     COORD_TYPE_Z(TEX_NORMALIZED),
-			     COORD_TYPE_W(TEX_NORMALIZED));
-	ps[i++] = TEX_DWORD2(OFFSET_X(0),
-			     OFFSET_Y(0),
-			     OFFSET_Z(0),
-			     SAMPLER_ID(0),
-			     SRC_SEL_X(SQ_SEL_X),
-			     SRC_SEL_Y(SQ_SEL_Y),
-			     SRC_SEL_Z(SQ_SEL_0),
-			     SRC_SEL_W(SQ_SEL_1));
-	ps[i++] = TEX_DWORD_PAD;
+	R600_comp_ps(info->ChipFamily, ps,
+		     src_a, src_r, src_g, src_b);
+
     }
 
     CLEAR (cb_conf);
@@ -2246,11 +1976,10 @@ R600LoadShaders(ScrnInfoPtr pScrn, ScreenPtr pScreen)
 {
     RADEONInfoPtr info = RADEONPTR(pScrn);
     struct radeon_accel_state *accel_state = info->accel_state;
-    uint32_t *vs;
-    uint32_t *ps;
+    RADEONChipFamily ChipSet = info->ChipFamily;
+    uint32_t *shader;
     /* 512 bytes per shader for now */
     int size = 512 * 11;
-    int i;
 
     accel_state->shaders = NULL;
 
@@ -2260,1991 +1989,51 @@ R600LoadShaders(ScrnInfoPtr pScrn, ScreenPtr pScreen)
     if (accel_state->shaders == NULL)
 	return FALSE;
 
-    vs = (pointer)((char *)info->FB + accel_state->shaders->offset);
-    ps = (pointer)((char *)info->FB + accel_state->shaders->offset);
+    shader = (pointer)((char *)info->FB + accel_state->shaders->offset);
+
+    /*  solid vs --------------------------------------- */
     accel_state->solid_vs_offset = 0;
+    R600_solid_vs(ChipSet, shader + accel_state->solid_vs_offset / 4);
+
+    /*  solid ps --------------------------------------- */
     accel_state->solid_ps_offset = 512;
+    R600_solid_ps(ChipSet, shader + accel_state->solid_ps_offset / 4);
+
+    /*  copy vs --------------------------------------- */
     accel_state->copy_vs_offset = 1024;
+    R600_copy_vs(ChipSet, shader + accel_state->copy_vs_offset / 4);
+
+    /*  copy ps --------------------------------------- */
     accel_state->copy_ps_offset = 1536;
+    R600_copy_ps(ChipSet, shader + accel_state->copy_ps_offset / 4);
+
+    /*  comp vs --------------------------------------- */
     accel_state->comp_vs_offset = 2048;
+    R600_comp_vs(ChipSet, shader + accel_state->comp_vs_offset / 4);
+
+    /*  comp ps --------------------------------------- */
     accel_state->comp_ps_offset = 2560;
+    /*  not yet */
+
+    /*  comp mask vs --------------------------------------- */
     accel_state->comp_mask_vs_offset = 3072;
+    R600_comp_mask_vs(ChipSet, shader + accel_state->comp_mask_vs_offset / 4);
+
+    /*  comp mask ps --------------------------------------- */
     accel_state->comp_mask_ps_offset = 3584;
+    /*  not yet */
+
+    /*  xv vs --------------------------------------- */
     accel_state->xv_vs_offset = 4096;
-    accel_state->xv_ps_offset_packed = 4608;
-    accel_state->xv_ps_offset_planar = 5120;
+    R600_xv_vs(ChipSet, shader + accel_state->xv_vs_offset / 4);
 
-    /* solid vs --------------------------------------- */
-    i = accel_state->solid_vs_offset / 4;
-    /* 0 */
-    vs[i++] = CF_DWORD0(ADDR(4));
-    vs[i++] = CF_DWORD1(POP_COUNT(0),
-			CF_CONST(0),
-			COND(SQ_CF_COND_ACTIVE),
-			I_COUNT(1),
-			CALL_COUNT(0),
-			END_OF_PROGRAM(0),
-			VALID_PIXEL_MODE(0),
-			CF_INST(SQ_CF_INST_VTX),
-			WHOLE_QUAD_MODE(0),
-			BARRIER(1));
-    /* 1 */
-    vs[i++] = CF_ALLOC_IMP_EXP_DWORD0(ARRAY_BASE(CF_POS0),
-				      TYPE(SQ_EXPORT_POS),
-				      RW_GPR(1),
-				      RW_REL(ABSOLUTE),
-				      INDEX_GPR(0),
-				      ELEM_SIZE(0));
-    vs[i++] = CF_ALLOC_IMP_EXP_DWORD1_SWIZ(SRC_SEL_X(SQ_SEL_X),
-					   SRC_SEL_Y(SQ_SEL_Y),
-					   SRC_SEL_Z(SQ_SEL_Z),
-					   SRC_SEL_W(SQ_SEL_W),
-					   R6xx_ELEM_LOOP(0),
-					   BURST_COUNT(1),
-					   END_OF_PROGRAM(0),
-					   VALID_PIXEL_MODE(0),
-					   CF_INST(SQ_CF_INST_EXPORT_DONE),
-					   WHOLE_QUAD_MODE(0),
-					   BARRIER(1));
-    /*2 - always export a param whether it's used or not */
-    vs[i++] = CF_ALLOC_IMP_EXP_DWORD0(ARRAY_BASE(0),
-				      TYPE(SQ_EXPORT_PARAM),
-				      RW_GPR(0),
-				      RW_REL(ABSOLUTE),
-				      INDEX_GPR(0),
-				      ELEM_SIZE(0));
-    vs[i++] = CF_ALLOC_IMP_EXP_DWORD1_SWIZ(SRC_SEL_X(SQ_SEL_X),
-					   SRC_SEL_Y(SQ_SEL_Y),
-					   SRC_SEL_Z(SQ_SEL_Z),
-					   SRC_SEL_W(SQ_SEL_W),
-					   R6xx_ELEM_LOOP(0),
-					   BURST_COUNT(0),
-					   END_OF_PROGRAM(1),
-					   VALID_PIXEL_MODE(0),
-					   CF_INST(SQ_CF_INST_EXPORT_DONE),
-					   WHOLE_QUAD_MODE(0),
-		 			   BARRIER(0));
-    /* 3 - padding */
-    vs[i++] = 0x00000000;
-    vs[i++] = 0x00000000;
-    /* 4/5 */
-    vs[i++] = VTX_DWORD0(VTX_INST(SQ_VTX_INST_FETCH),
-			 FETCH_TYPE(SQ_VTX_FETCH_VERTEX_DATA),
-			 FETCH_WHOLE_QUAD(0),
-			 BUFFER_ID(0),
-			 SRC_GPR(0),
-			 SRC_REL(ABSOLUTE),
-			 SRC_SEL_X(SQ_SEL_X),
-			 MEGA_FETCH_COUNT(8));
-    vs[i++] = VTX_DWORD1_GPR(DST_GPR(1),
-			     DST_REL(0),
-			     DST_SEL_X(SQ_SEL_X),
-			     DST_SEL_Y(SQ_SEL_Y),
-			     DST_SEL_Z(SQ_SEL_0),
-			     DST_SEL_W(SQ_SEL_1),
-			     USE_CONST_FIELDS(0),
-			     DATA_FORMAT(FMT_32_32_FLOAT),
-			     NUM_FORMAT_ALL(SQ_NUM_FORMAT_NORM),
-			     FORMAT_COMP_ALL(SQ_FORMAT_COMP_SIGNED),
-			     SRF_MODE_ALL(SRF_MODE_ZERO_CLAMP_MINUS_ONE));
-    vs[i++] = VTX_DWORD2(OFFSET(0),
-			 ENDIAN_SWAP(ENDIAN_NONE),
-			 CONST_BUF_NO_STRIDE(0),
-			 MEGA_FETCH(1));
-    vs[i++] = VTX_DWORD_PAD;
-
-    /* solid ps --------------------------------------- */
-    i = accel_state->solid_ps_offset / 4;
-    /* 0 */
-    ps[i++] = CF_ALU_DWORD0(ADDR(2),
-			    KCACHE_BANK0(0),
-			    KCACHE_BANK1(0),
-			    KCACHE_MODE0(SQ_CF_KCACHE_NOP));
-    ps[i++] = CF_ALU_DWORD1(KCACHE_MODE1(SQ_CF_KCACHE_NOP),
-			    KCACHE_ADDR0(0),
-			    KCACHE_ADDR1(0),
-			    I_COUNT(4),
-			    USES_WATERFALL(0),
-			    CF_INST(SQ_CF_INST_ALU),
-			    WHOLE_QUAD_MODE(0),
-			    BARRIER(1));
-    /* 1 */
-    ps[i++] = CF_ALLOC_IMP_EXP_DWORD0(ARRAY_BASE(CF_PIXEL_MRT0),
-				      TYPE(SQ_EXPORT_PIXEL),
-				      RW_GPR(0),
-				      RW_REL(ABSOLUTE),
-				      INDEX_GPR(0),
-				      ELEM_SIZE(1));
-    ps[i++] = CF_ALLOC_IMP_EXP_DWORD1_SWIZ(SRC_SEL_X(SQ_SEL_X),
-					   SRC_SEL_Y(SQ_SEL_Y),
-					   SRC_SEL_Z(SQ_SEL_Z),
-					   SRC_SEL_W(SQ_SEL_W),
-					   R6xx_ELEM_LOOP(0),
-					   BURST_COUNT(1),
-					   END_OF_PROGRAM(1),
-					   VALID_PIXEL_MODE(0),
-					   CF_INST(SQ_CF_INST_EXPORT_DONE),
-					   WHOLE_QUAD_MODE(0),
-					   BARRIER(1));
-
-    /* 2 */
-    ps[i++] = ALU_DWORD0(SRC0_SEL(256),
-			 SRC0_REL(ABSOLUTE),
-			 SRC0_ELEM(ELEM_X),
-			 SRC0_NEG(0),
-			 SRC1_SEL(0),
-			 SRC1_REL(ABSOLUTE),
-			 SRC1_ELEM(ELEM_X),
-			 SRC1_NEG(0),
-			 INDEX_MODE(SQ_INDEX_AR_X),
-			 PRED_SEL(SQ_PRED_SEL_OFF),
-			 LAST(0));
-    ps[i++] = ALU_DWORD1_OP2(info->ChipFamily,
-			     SRC0_ABS(0),
-			     SRC1_ABS(0),
-			     UPDATE_EXECUTE_MASK(0),
-			     UPDATE_PRED(0),
-			     WRITE_MASK(1),
-			     FOG_MERGE(0),
-			     OMOD(SQ_ALU_OMOD_OFF),
-			     ALU_INST(SQ_OP2_INST_MOV),
-			     BANK_SWIZZLE(SQ_ALU_VEC_012),
-			     DST_GPR(0),
-			     DST_REL(ABSOLUTE),
-			     DST_ELEM(ELEM_X),
-			     CLAMP(1));
-    /* 3 */
-    ps[i++] = ALU_DWORD0(SRC0_SEL(256),
-			 SRC0_REL(ABSOLUTE),
-			 SRC0_ELEM(ELEM_Y),
-			 SRC0_NEG(0),
-			 SRC1_SEL(0),
-			 SRC1_REL(ABSOLUTE),
-			 SRC1_ELEM(ELEM_Y),
-			 SRC1_NEG(0),
-			 INDEX_MODE(SQ_INDEX_AR_X),
-			 PRED_SEL(SQ_PRED_SEL_OFF),
-			 LAST(0));
-    ps[i++] = ALU_DWORD1_OP2(info->ChipFamily,
-			     SRC0_ABS(0),
-			     SRC1_ABS(0),
-			     UPDATE_EXECUTE_MASK(0),
-			     UPDATE_PRED(0),
-			     WRITE_MASK(1),
-			     FOG_MERGE(0),
-			     OMOD(SQ_ALU_OMOD_OFF),
-			     ALU_INST(SQ_OP2_INST_MOV),
-			     BANK_SWIZZLE(SQ_ALU_VEC_012),
-			     DST_GPR(0),
-			     DST_REL(ABSOLUTE),
-			     DST_ELEM(ELEM_Y),
-			     CLAMP(1));
-    /* 4 */
-    ps[i++] = ALU_DWORD0(SRC0_SEL(256),
-			 SRC0_REL(ABSOLUTE),
-			 SRC0_ELEM(ELEM_Z),
-			 SRC0_NEG(0),
-			 SRC1_SEL(0),
-			 SRC1_REL(ABSOLUTE),
-			 SRC1_ELEM(ELEM_Z),
-			 SRC1_NEG(0),
-			 INDEX_MODE(SQ_INDEX_AR_X),
-			 PRED_SEL(SQ_PRED_SEL_OFF),
-			 LAST(0));
-    ps[i++] = ALU_DWORD1_OP2(info->ChipFamily,
-			     SRC0_ABS(0),
-			     SRC1_ABS(0),
-			     UPDATE_EXECUTE_MASK(0),
-			     UPDATE_PRED(0),
-			     WRITE_MASK(1),
-			     FOG_MERGE(0),
-			     OMOD(SQ_ALU_OMOD_OFF),
-			     ALU_INST(SQ_OP2_INST_MOV),
-			     BANK_SWIZZLE(SQ_ALU_VEC_012),
-			     DST_GPR(0),
-			     DST_REL(ABSOLUTE),
-			     DST_ELEM(ELEM_Z),
-			     CLAMP(1));
-    /* 5 */
-    ps[i++] = ALU_DWORD0(SRC0_SEL(256),
-			 SRC0_REL(ABSOLUTE),
-			 SRC0_ELEM(ELEM_W),
-			 SRC0_NEG(0),
-			 SRC1_SEL(0),
-			 SRC1_REL(ABSOLUTE),
-			 SRC1_ELEM(ELEM_W),
-			 SRC1_NEG(0),
-			 INDEX_MODE(SQ_INDEX_AR_X),
-			 PRED_SEL(SQ_PRED_SEL_OFF),
-			 LAST(1));
-    ps[i++] = ALU_DWORD1_OP2(info->ChipFamily,
-			     SRC0_ABS(0),
-			     SRC1_ABS(0),
-			     UPDATE_EXECUTE_MASK(0),
-			     UPDATE_PRED(0),
-			     WRITE_MASK(1),
-			     FOG_MERGE(0),
-			     OMOD(SQ_ALU_OMOD_OFF),
-			     ALU_INST(SQ_OP2_INST_MOV),
-			     BANK_SWIZZLE(SQ_ALU_VEC_012),
-			     DST_GPR(0),
-			     DST_REL(ABSOLUTE),
-			     DST_ELEM(ELEM_W),
-			     CLAMP(1));
-
-    /* copy vs --------------------------------------- */
-    i = accel_state->copy_vs_offset / 4;
-    /* 0 */
-    vs[i++] = CF_DWORD0(ADDR(4));
-    vs[i++] = CF_DWORD1(POP_COUNT(0),
-			CF_CONST(0),
-			COND(SQ_CF_COND_ACTIVE),
-			I_COUNT(2),
-			CALL_COUNT(0),
-			END_OF_PROGRAM(0),
-			VALID_PIXEL_MODE(0),
-			CF_INST(SQ_CF_INST_VTX),
-			WHOLE_QUAD_MODE(0),
-			BARRIER(1));
-    /* 1 */
-    vs[i++] = CF_ALLOC_IMP_EXP_DWORD0(ARRAY_BASE(CF_POS0),
-				      TYPE(SQ_EXPORT_POS),
-				      RW_GPR(1),
-				      RW_REL(ABSOLUTE),
-				      INDEX_GPR(0),
-				      ELEM_SIZE(0));
-    vs[i++] = CF_ALLOC_IMP_EXP_DWORD1_SWIZ(SRC_SEL_X(SQ_SEL_X),
-					   SRC_SEL_Y(SQ_SEL_Y),
-					   SRC_SEL_Z(SQ_SEL_Z),
-					   SRC_SEL_W(SQ_SEL_W),
-					   R6xx_ELEM_LOOP(0),
-					   BURST_COUNT(0),
-					   END_OF_PROGRAM(0),
-					   VALID_PIXEL_MODE(0),
-					   CF_INST(SQ_CF_INST_EXPORT_DONE),
-					   WHOLE_QUAD_MODE(0),
-					   BARRIER(1));
-    /* 2 */
-    vs[i++] = CF_ALLOC_IMP_EXP_DWORD0(ARRAY_BASE(0),
-				      TYPE(SQ_EXPORT_PARAM),
-				      RW_GPR(0),
-				      RW_REL(ABSOLUTE),
-				      INDEX_GPR(0),
-				      ELEM_SIZE(0));
-    vs[i++] = CF_ALLOC_IMP_EXP_DWORD1_SWIZ(SRC_SEL_X(SQ_SEL_X),
-					   SRC_SEL_Y(SQ_SEL_Y),
-					   SRC_SEL_Z(SQ_SEL_Z),
-					   SRC_SEL_W(SQ_SEL_W),
-					   R6xx_ELEM_LOOP(0),
-					   BURST_COUNT(0),
-					   END_OF_PROGRAM(1),
-					   VALID_PIXEL_MODE(0),
-					   CF_INST(SQ_CF_INST_EXPORT_DONE),
-					   WHOLE_QUAD_MODE(0),
-					   BARRIER(0));
-    /* 3 */
-    vs[i++] = 0x00000000;
-    vs[i++] = 0x00000000;
-    /* 4/5 */
-    vs[i++] = VTX_DWORD0(VTX_INST(SQ_VTX_INST_FETCH),
-			 FETCH_TYPE(SQ_VTX_FETCH_VERTEX_DATA),
-			 FETCH_WHOLE_QUAD(0),
-			 BUFFER_ID(0),
-			 SRC_GPR(0),
-			 SRC_REL(ABSOLUTE),
-			 SRC_SEL_X(SQ_SEL_X),
-			 MEGA_FETCH_COUNT(16));
-    vs[i++] = VTX_DWORD1_GPR(DST_GPR(1),
-			     DST_REL(0),
-			     DST_SEL_X(SQ_SEL_X),
-			     DST_SEL_Y(SQ_SEL_Y),
-			     DST_SEL_Z(SQ_SEL_0),
-			     DST_SEL_W(SQ_SEL_1),
-			     USE_CONST_FIELDS(0),
-			     DATA_FORMAT(FMT_32_32_FLOAT),
-			     NUM_FORMAT_ALL(SQ_NUM_FORMAT_NORM),
-			     FORMAT_COMP_ALL(SQ_FORMAT_COMP_SIGNED),
-			     SRF_MODE_ALL(SRF_MODE_ZERO_CLAMP_MINUS_ONE));
-    vs[i++] = VTX_DWORD2(OFFSET(0),
-			 ENDIAN_SWAP(ENDIAN_NONE),
-			 CONST_BUF_NO_STRIDE(0),
-			 MEGA_FETCH(1));
-    vs[i++] = VTX_DWORD_PAD;
-    /* 6/7 */
-    vs[i++] = VTX_DWORD0(VTX_INST(SQ_VTX_INST_FETCH),
-			 FETCH_TYPE(SQ_VTX_FETCH_VERTEX_DATA),
-			 FETCH_WHOLE_QUAD(0),
-			 BUFFER_ID(0),
-			 SRC_GPR(0),
-			 SRC_REL(ABSOLUTE),
-			 SRC_SEL_X(SQ_SEL_X),
-			 MEGA_FETCH_COUNT(8));
-    vs[i++] = VTX_DWORD1_GPR(DST_GPR(0),
-			     DST_REL(0),
-			     DST_SEL_X(SQ_SEL_X),
-			     DST_SEL_Y(SQ_SEL_Y),
-			     DST_SEL_Z(SQ_SEL_0),
-			     DST_SEL_W(SQ_SEL_1),
-			     USE_CONST_FIELDS(0),
-			     DATA_FORMAT(FMT_32_32_FLOAT),
-			     NUM_FORMAT_ALL(SQ_NUM_FORMAT_NORM),
-			     FORMAT_COMP_ALL(SQ_FORMAT_COMP_SIGNED),
-			     SRF_MODE_ALL(SRF_MODE_ZERO_CLAMP_MINUS_ONE));
-    vs[i++] = VTX_DWORD2(OFFSET(8),
-			 ENDIAN_SWAP(ENDIAN_NONE),
-			 CONST_BUF_NO_STRIDE(0),
-			 MEGA_FETCH(0));
-    vs[i++] = VTX_DWORD_PAD;
-
-    /* copy ps --------------------------------------- */
-    i = accel_state->copy_ps_offset / 4;
-    /* CF INST 0 */
-    ps[i++] = CF_DWORD0(ADDR(2));
-    ps[i++] = CF_DWORD1(POP_COUNT(0),
-			CF_CONST(0),
-			COND(SQ_CF_COND_ACTIVE),
-			I_COUNT(1),
-			CALL_COUNT(0),
-			END_OF_PROGRAM(0),
-			VALID_PIXEL_MODE(0),
-			CF_INST(SQ_CF_INST_TEX),
-			WHOLE_QUAD_MODE(0),
-			BARRIER(1));
-    /* CF INST 1 */
-    ps[i++] = CF_ALLOC_IMP_EXP_DWORD0(ARRAY_BASE(CF_PIXEL_MRT0),
-				      TYPE(SQ_EXPORT_PIXEL),
-				      RW_GPR(0),
-				      RW_REL(ABSOLUTE),
-				      INDEX_GPR(0),
-				      ELEM_SIZE(1));
-    ps[i++] = CF_ALLOC_IMP_EXP_DWORD1_SWIZ(SRC_SEL_X(SQ_SEL_X),
-					   SRC_SEL_Y(SQ_SEL_Y),
-					   SRC_SEL_Z(SQ_SEL_Z),
-					   SRC_SEL_W(SQ_SEL_W),
-					   R6xx_ELEM_LOOP(0),
-					   BURST_COUNT(1),
-					   END_OF_PROGRAM(1),
-					   VALID_PIXEL_MODE(0),
-					   CF_INST(SQ_CF_INST_EXPORT_DONE),
-					   WHOLE_QUAD_MODE(0),
-					   BARRIER(1));
-    /* TEX INST 0 */
-    ps[i++] = TEX_DWORD0(TEX_INST(SQ_TEX_INST_SAMPLE),
-			 BC_FRAC_MODE(0),
-			 FETCH_WHOLE_QUAD(0),
-			 RESOURCE_ID(0),
-			 SRC_GPR(0),
-			 SRC_REL(ABSOLUTE),
-			 R7xx_ALT_CONST(0));
-    ps[i++] = TEX_DWORD1(DST_GPR(0),
-			 DST_REL(ABSOLUTE),
-			 DST_SEL_X(SQ_SEL_X), /* R */
-			 DST_SEL_Y(SQ_SEL_Y), /* G */
-			 DST_SEL_Z(SQ_SEL_Z), /* B */
-			 DST_SEL_W(SQ_SEL_W), /* A */
-			 LOD_BIAS(0),
-			 COORD_TYPE_X(TEX_UNNORMALIZED),
-			 COORD_TYPE_Y(TEX_UNNORMALIZED),
-			 COORD_TYPE_Z(TEX_UNNORMALIZED),
-			 COORD_TYPE_W(TEX_UNNORMALIZED));
-    ps[i++] = TEX_DWORD2(OFFSET_X(0),
-			 OFFSET_Y(0),
-			 OFFSET_Z(0),
-			 SAMPLER_ID(0),
-			 SRC_SEL_X(SQ_SEL_X),
-			 SRC_SEL_Y(SQ_SEL_Y),
-			 SRC_SEL_Z(SQ_SEL_0),
-			 SRC_SEL_W(SQ_SEL_1));
-    ps[i++] = TEX_DWORD_PAD;
-
-    /* xv vs --------------------------------------- */
-    i = accel_state->xv_vs_offset / 4;
-    /* 0 */
-    vs[i++] = CF_DWORD0(ADDR(4));
-    vs[i++] = CF_DWORD1(POP_COUNT(0),
-			CF_CONST(0),
-			COND(SQ_CF_COND_ACTIVE),
-			I_COUNT(2),
-			CALL_COUNT(0),
-			END_OF_PROGRAM(0),
-			VALID_PIXEL_MODE(0),
-			CF_INST(SQ_CF_INST_VTX),
-			WHOLE_QUAD_MODE(0),
-			BARRIER(1));
-    /* 1 */
-    vs[i++] = CF_ALLOC_IMP_EXP_DWORD0(ARRAY_BASE(CF_POS0),
-				      TYPE(SQ_EXPORT_POS),
-				      RW_GPR(1),
-				      RW_REL(ABSOLUTE),
-				      INDEX_GPR(0),
-				      ELEM_SIZE(0));
-    vs[i++] = CF_ALLOC_IMP_EXP_DWORD1_SWIZ(SRC_SEL_X(SQ_SEL_X),
-					   SRC_SEL_Y(SQ_SEL_Y),
-					   SRC_SEL_Z(SQ_SEL_Z),
-					   SRC_SEL_W(SQ_SEL_W),
-					   R6xx_ELEM_LOOP(0),
-					   BURST_COUNT(0),
-					   END_OF_PROGRAM(0),
-					   VALID_PIXEL_MODE(0),
-					   CF_INST(SQ_CF_INST_EXPORT_DONE),
-					   WHOLE_QUAD_MODE(0),
-					   BARRIER(1));
-    /* 2 */
-    vs[i++] = CF_ALLOC_IMP_EXP_DWORD0(ARRAY_BASE(0),
-				      TYPE(SQ_EXPORT_PARAM),
-				      RW_GPR(0),
-				      RW_REL(ABSOLUTE),
-				      INDEX_GPR(0),
-				      ELEM_SIZE(0));
-    vs[i++] = CF_ALLOC_IMP_EXP_DWORD1_SWIZ(SRC_SEL_X(SQ_SEL_X),
-					   SRC_SEL_Y(SQ_SEL_Y),
-					   SRC_SEL_Z(SQ_SEL_Z),
-					   SRC_SEL_W(SQ_SEL_W),
-					   R6xx_ELEM_LOOP(0),
-					   BURST_COUNT(0),
-					   END_OF_PROGRAM(1),
-					   VALID_PIXEL_MODE(0),
-					   CF_INST(SQ_CF_INST_EXPORT_DONE),
-					   WHOLE_QUAD_MODE(0),
-					   BARRIER(0));
-    /* 3 */
-    vs[i++] = 0x00000000;
-    vs[i++] = 0x00000000;
-    /* 4/5 */
-    vs[i++] = VTX_DWORD0(VTX_INST(SQ_VTX_INST_FETCH),
-			 FETCH_TYPE(SQ_VTX_FETCH_VERTEX_DATA),
-			 FETCH_WHOLE_QUAD(0),
-			 BUFFER_ID(0),
-			 SRC_GPR(0),
-			 SRC_REL(ABSOLUTE),
-			 SRC_SEL_X(SQ_SEL_X),
-			 MEGA_FETCH_COUNT(16));
-    vs[i++] = VTX_DWORD1_GPR(DST_GPR(1),
-			     DST_REL(0),
-			     DST_SEL_X(SQ_SEL_X),
-			     DST_SEL_Y(SQ_SEL_Y),
-			     DST_SEL_Z(SQ_SEL_0),
-			     DST_SEL_W(SQ_SEL_1),
-			     USE_CONST_FIELDS(0),
-			     DATA_FORMAT(FMT_32_32_FLOAT),
-			     NUM_FORMAT_ALL(SQ_NUM_FORMAT_NORM),
-			     FORMAT_COMP_ALL(SQ_FORMAT_COMP_SIGNED),
-			     SRF_MODE_ALL(SRF_MODE_ZERO_CLAMP_MINUS_ONE));
-    vs[i++] = VTX_DWORD2(OFFSET(0),
-			 ENDIAN_SWAP(ENDIAN_NONE),
-			 CONST_BUF_NO_STRIDE(0),
-			 MEGA_FETCH(1));
-    vs[i++] = VTX_DWORD_PAD;
-    /* 6/7 */
-    vs[i++] = VTX_DWORD0(VTX_INST(SQ_VTX_INST_FETCH),
-			 FETCH_TYPE(SQ_VTX_FETCH_VERTEX_DATA),
-			 FETCH_WHOLE_QUAD(0),
-			 BUFFER_ID(0),
-			 SRC_GPR(0),
-			 SRC_REL(ABSOLUTE),
-			 SRC_SEL_X(SQ_SEL_X),
-			 MEGA_FETCH_COUNT(8));
-    vs[i++] = VTX_DWORD1_GPR(DST_GPR(0),
-			     DST_REL(0),
-			     DST_SEL_X(SQ_SEL_X),
-			     DST_SEL_Y(SQ_SEL_Y),
-			     DST_SEL_Z(SQ_SEL_0),
-			     DST_SEL_W(SQ_SEL_1),
-			     USE_CONST_FIELDS(0),
-			     DATA_FORMAT(FMT_32_32_FLOAT),
-			     NUM_FORMAT_ALL(SQ_NUM_FORMAT_NORM),
-			     FORMAT_COMP_ALL(SQ_FORMAT_COMP_SIGNED),
-			     SRF_MODE_ALL(SRF_MODE_ZERO_CLAMP_MINUS_ONE));
-    vs[i++] = VTX_DWORD2(OFFSET(8),
-			 ENDIAN_SWAP(ENDIAN_NONE),
-			 CONST_BUF_NO_STRIDE(0),
-			 MEGA_FETCH(0));
-    vs[i++] = VTX_DWORD_PAD;
-
-    /* xv ps packed ---------------------------------- */
-    i = accel_state->xv_ps_offset_packed / 4;
-    /* 0 */
-    ps[i++] = CF_DWORD0(ADDR(20));
-    ps[i++] = CF_DWORD1(POP_COUNT(0),
-			CF_CONST(0),
-			COND(SQ_CF_COND_ACTIVE),
-			I_COUNT(2),
-			CALL_COUNT(0),
-			END_OF_PROGRAM(0),
-			VALID_PIXEL_MODE(0),
-			CF_INST(SQ_CF_INST_TEX),
-			WHOLE_QUAD_MODE(0),
-			BARRIER(0));
-    /* 1 */
-    ps[i++] = CF_ALU_DWORD0(ADDR(3),
-			    KCACHE_BANK0(0),
-			    KCACHE_BANK1(0),
-			    KCACHE_MODE0(SQ_CF_KCACHE_NOP));
-    ps[i++] = CF_ALU_DWORD1(KCACHE_MODE1(SQ_CF_KCACHE_NOP),
-			    KCACHE_ADDR0(0),
-			    KCACHE_ADDR1(0),
-			    I_COUNT(16),
-			    USES_WATERFALL(0),
-			    CF_INST(SQ_CF_INST_ALU),
-			    WHOLE_QUAD_MODE(0),
-			    BARRIER(1));
-    /* 2 */
-    ps[i++] = CF_ALLOC_IMP_EXP_DWORD0(ARRAY_BASE(CF_PIXEL_MRT0),
-				      TYPE(SQ_EXPORT_PIXEL),
-				      RW_GPR(2),
-				      RW_REL(ABSOLUTE),
-				      INDEX_GPR(0),
-				      ELEM_SIZE(3));
-    ps[i++] = CF_ALLOC_IMP_EXP_DWORD1_SWIZ(SRC_SEL_X(SQ_SEL_X),
-					   SRC_SEL_Y(SQ_SEL_Y),
-					   SRC_SEL_Z(SQ_SEL_Z),
-					   SRC_SEL_W(SQ_SEL_W),
-					   R6xx_ELEM_LOOP(0),
-					   BURST_COUNT(1),
-					   END_OF_PROGRAM(1),
-					   VALID_PIXEL_MODE(0),
-					   CF_INST(SQ_CF_INST_EXPORT_DONE),
-					   WHOLE_QUAD_MODE(0),
-					   BARRIER(1));
-    /* Undo scaling of Y'CbCr values
-     *  Y' is scaled from 16:235
-     *  Cb/Cr are scaled from 16:240
-     */
-    /* 3 - alu 0 */
-    /* MULADD gpr[1].x gpr[1].x c[3].x c[3].y */
-    ps[i++] = ALU_DWORD0(SRC0_SEL(1),
-			 SRC0_REL(ABSOLUTE),
-			 SRC0_ELEM(ELEM_X),
-			 SRC0_NEG(0),
-			 SRC1_SEL(259),
-			 SRC1_REL(ABSOLUTE),
-			 SRC1_ELEM(ELEM_X),
-			 SRC1_NEG(0),
-			 INDEX_MODE(SQ_INDEX_LOOP),
-			 PRED_SEL(SQ_PRED_SEL_OFF),
-			 LAST(0));
-    ps[i++] = ALU_DWORD1_OP3(SRC2_SEL(259),
-			     SRC2_REL(ABSOLUTE),
-			     SRC2_ELEM(ELEM_Y),
-			     SRC2_NEG(0),
-			     ALU_INST(SQ_OP3_INST_MULADD),
-			     BANK_SWIZZLE(SQ_ALU_VEC_012),
-			     DST_GPR(1),
-			     DST_REL(ABSOLUTE),
-			     DST_ELEM(ELEM_X),
-			     CLAMP(1));
-    /* 4 - alu 1 */
-    /* MULADD gpr[1].y gpr[1].y c[3].z c[3].w */
-    ps[i++] = ALU_DWORD0(SRC0_SEL(1),
-			 SRC0_REL(ABSOLUTE),
-			 SRC0_ELEM(ELEM_Y),
-			 SRC0_NEG(0),
-			 SRC1_SEL(259),
-			 SRC1_REL(ABSOLUTE),
-			 SRC1_ELEM(ELEM_Z),
-			 SRC1_NEG(0),
-			 INDEX_MODE(SQ_INDEX_LOOP),
-			 PRED_SEL(SQ_PRED_SEL_OFF),
-			 LAST(0));
-    ps[i++] = ALU_DWORD1_OP3(SRC2_SEL(259),
-			     SRC2_REL(ABSOLUTE),
-			     SRC2_ELEM(ELEM_W),
-			     SRC2_NEG(0),
-			     ALU_INST(SQ_OP3_INST_MULADD),
-			     BANK_SWIZZLE(SQ_ALU_VEC_012),
-			     DST_GPR(1),
-			     DST_REL(ABSOLUTE),
-			     DST_ELEM(ELEM_Y),
-			     CLAMP(0));
-    /* 5 - alu 2 */
-    /* MULADD gpr[1].z gpr[1].z c[3].z c[3].w */
-    ps[i++] = ALU_DWORD0(SRC0_SEL(1),
-			 SRC0_REL(ABSOLUTE),
-			 SRC0_ELEM(ELEM_Z),
-			 SRC0_NEG(0),
-			 SRC1_SEL(259),
-			 SRC1_REL(ABSOLUTE),
-			 SRC1_ELEM(ELEM_Z),
-			 SRC1_NEG(0),
-			 INDEX_MODE(SQ_INDEX_LOOP),
-			 PRED_SEL(SQ_PRED_SEL_OFF),
-			 LAST(0));
-    ps[i++] = ALU_DWORD1_OP3(SRC2_SEL(259),
-			     SRC2_REL(ABSOLUTE),
-			     SRC2_ELEM(ELEM_W),
-			     SRC2_NEG(0),
-			     ALU_INST(SQ_OP3_INST_MULADD),
-			     BANK_SWIZZLE(SQ_ALU_VEC_012),
-			     DST_GPR(1),
-			     DST_REL(ABSOLUTE),
-			     DST_ELEM(ELEM_Z),
-			     CLAMP(0));
-    /* 6 - alu 3 */
-    /* MOV gpr[1].w 0.0 */
-    ps[i++] = ALU_DWORD0(SRC0_SEL(SQ_ALU_SRC_0),
-			 SRC0_REL(ABSOLUTE),
-			 SRC0_ELEM(ELEM_X),
-			 SRC0_NEG(0),
-			 SRC1_SEL(SQ_ALU_SRC_0),
-			 SRC1_REL(ABSOLUTE),
-			 SRC1_ELEM(ELEM_X),
-			 SRC1_NEG(0),
-			 INDEX_MODE(SQ_INDEX_LOOP),
-			 PRED_SEL(SQ_PRED_SEL_OFF),
-			 LAST(1));
-    ps[i++] = ALU_DWORD1_OP2(info->ChipFamily,
-			     SRC0_ABS(0),
-			     SRC1_ABS(0),
-			     UPDATE_EXECUTE_MASK(0),
-			     UPDATE_PRED(0),
-			     WRITE_MASK(1),
-			     FOG_MERGE(0),
-			     OMOD(SQ_ALU_OMOD_OFF),
-			     ALU_INST(SQ_OP2_INST_DOT4),
-			     BANK_SWIZZLE(SQ_ALU_VEC_102),
-			     DST_GPR(1),
-			     DST_REL(ABSOLUTE),
-			     DST_ELEM(ELEM_W),
-			     CLAMP(0));
-    /* 7 - alu 4 */
-    /* DP4 gpr[2].x gpr[1].x c[0].x */
-    ps[i++] = ALU_DWORD0(SRC0_SEL(1),
-			 SRC0_REL(ABSOLUTE),
-			 SRC0_ELEM(ELEM_X),
-			 SRC0_NEG(0),
-			 SRC1_SEL(256),
-			 SRC1_REL(ABSOLUTE),
-			 SRC1_ELEM(ELEM_X),
-			 SRC1_NEG(0),
-			 INDEX_MODE(SQ_INDEX_LOOP),
-			 PRED_SEL(SQ_PRED_SEL_OFF),
-			 LAST(0));
-    ps[i++] = ALU_DWORD1_OP2(info->ChipFamily,
-			     SRC0_ABS(0),
-			     SRC1_ABS(0),
-			     UPDATE_EXECUTE_MASK(0),
-			     UPDATE_PRED(0),
-			     WRITE_MASK(1),
-			     FOG_MERGE(0),
-			     OMOD(SQ_ALU_OMOD_OFF),
-			     ALU_INST(SQ_OP2_INST_DOT4),
-			     BANK_SWIZZLE(SQ_ALU_VEC_102),
-			     DST_GPR(2),
-			     DST_REL(ABSOLUTE),
-			     DST_ELEM(ELEM_X),
-			     CLAMP(1));
-    /* 8 - alu 5 */
-    /* DP4 gpr[2].y gpr[1].y c[0].y */
-    ps[i++] = ALU_DWORD0(SRC0_SEL(1),
-			 SRC0_REL(ABSOLUTE),
-			 SRC0_ELEM(ELEM_Y),
-			 SRC0_NEG(0),
-			 SRC1_SEL(256),
-			 SRC1_REL(ABSOLUTE),
-			 SRC1_ELEM(ELEM_Y),
-			 SRC1_NEG(0),
-			 INDEX_MODE(SQ_INDEX_LOOP),
-			 PRED_SEL(SQ_PRED_SEL_OFF),
-			 LAST(0));
-    ps[i++] = ALU_DWORD1_OP2(info->ChipFamily,
-			     SRC0_ABS(0),
-			     SRC1_ABS(0),
-			     UPDATE_EXECUTE_MASK(0),
-			     UPDATE_PRED(0),
-			     WRITE_MASK(0),
-			     FOG_MERGE(0),
-			     OMOD(SQ_ALU_OMOD_OFF),
-			     ALU_INST(SQ_OP2_INST_DOT4),
-			     BANK_SWIZZLE(SQ_ALU_VEC_102),
-			     DST_GPR(2),
-			     DST_REL(ABSOLUTE),
-			     DST_ELEM(ELEM_Y),
-			     CLAMP(1));
-    /* 9 - alu 6 */
-    /* DP4 gpr[2].z gpr[1].z c[0].z */
-    ps[i++] = ALU_DWORD0(SRC0_SEL(1),
-			 SRC0_REL(ABSOLUTE),
-			 SRC0_ELEM(ELEM_Z),
-			 SRC0_NEG(0),
-			 SRC1_SEL(256),
-			 SRC1_REL(ABSOLUTE),
-			 SRC1_ELEM(ELEM_Z),
-			 SRC1_NEG(0),
-			 INDEX_MODE(SQ_INDEX_LOOP),
-			 PRED_SEL(SQ_PRED_SEL_OFF),
-			 LAST(0));
-    ps[i++] = ALU_DWORD1_OP2(info->ChipFamily,
-			     SRC0_ABS(0),
-			     SRC1_ABS(0),
-			     UPDATE_EXECUTE_MASK(0),
-			     UPDATE_PRED(0),
-			     WRITE_MASK(0),
-			     FOG_MERGE(0),
-			     OMOD(SQ_ALU_OMOD_OFF),
-			     ALU_INST(SQ_OP2_INST_DOT4),
-			     BANK_SWIZZLE(SQ_ALU_VEC_102),
-			     DST_GPR(2),
-			     DST_REL(ABSOLUTE),
-			     DST_ELEM(ELEM_Z),
-			     CLAMP(1));
-    /* 10 - alu 7 */
-    /* DP4 gpr[2].w gpr[1].w c[0].w */
-    ps[i++] = ALU_DWORD0(SRC0_SEL(1),
-			 SRC0_REL(ABSOLUTE),
-			 SRC0_ELEM(ELEM_W),
-			 SRC0_NEG(0),
-			 SRC1_SEL(256),
-			 SRC1_REL(ABSOLUTE),
-			 SRC1_ELEM(ELEM_W),
-			 SRC1_NEG(0),
-			 INDEX_MODE(SQ_INDEX_LOOP),
-			 PRED_SEL(SQ_PRED_SEL_OFF),
-			 LAST(1));
-    ps[i++] = ALU_DWORD1_OP2(info->ChipFamily,
-			     SRC0_ABS(0),
-			     SRC1_ABS(0),
-			     UPDATE_EXECUTE_MASK(0),
-			     UPDATE_PRED(0),
-			     WRITE_MASK(0),
-			     FOG_MERGE(0),
-			     OMOD(SQ_ALU_OMOD_OFF),
-			     ALU_INST(SQ_OP2_INST_DOT4),
-			     BANK_SWIZZLE(SQ_ALU_VEC_021),
-			     DST_GPR(2),
-			     DST_REL(ABSOLUTE),
-			     DST_ELEM(ELEM_W),
-			     CLAMP(1));
-    /* 11 - alu 8 */
-    /* DP4 gpr[2].x gpr[1].x c[1].x */
-    ps[i++] = ALU_DWORD0(SRC0_SEL(1),
-			 SRC0_REL(ABSOLUTE),
-			 SRC0_ELEM(ELEM_X),
-			 SRC0_NEG(0),
-			 SRC1_SEL(257),
-			 SRC1_REL(ABSOLUTE),
-			 SRC1_ELEM(ELEM_X),
-			 SRC1_NEG(0),
-			 INDEX_MODE(SQ_INDEX_LOOP),
-			 PRED_SEL(SQ_PRED_SEL_OFF),
-			 LAST(0));
-    ps[i++] = ALU_DWORD1_OP2(info->ChipFamily,
-			     SRC0_ABS(0),
-			     SRC1_ABS(0),
-			     UPDATE_EXECUTE_MASK(0),
-			     UPDATE_PRED(0),
-			     WRITE_MASK(0),
-			     FOG_MERGE(0),
-			     OMOD(SQ_ALU_OMOD_OFF),
-			     ALU_INST(SQ_OP2_INST_DOT4),
-			     BANK_SWIZZLE(SQ_ALU_VEC_102),
-			     DST_GPR(2),
-			     DST_REL(ABSOLUTE),
-			     DST_ELEM(ELEM_X),
-			     CLAMP(1));
-    /* 12 - alu 9 */
-    /* DP4 gpr[2].y gpr[1].y c[1].y */
-    ps[i++] = ALU_DWORD0(SRC0_SEL(1),
-			 SRC0_REL(ABSOLUTE),
-			 SRC0_ELEM(ELEM_Y),
-			 SRC0_NEG(0),
-			 SRC1_SEL(257),
-			 SRC1_REL(ABSOLUTE),
-			 SRC1_ELEM(ELEM_Y),
-			 SRC1_NEG(0),
-			 INDEX_MODE(SQ_INDEX_LOOP),
-			 PRED_SEL(SQ_PRED_SEL_OFF),
-			 LAST(0));
-    ps[i++] = ALU_DWORD1_OP2(info->ChipFamily,
-			     SRC0_ABS(0),
-			     SRC1_ABS(0),
-			     UPDATE_EXECUTE_MASK(0),
-			     UPDATE_PRED(0),
-			     WRITE_MASK(1),
-			     FOG_MERGE(0),
-			     OMOD(SQ_ALU_OMOD_OFF),
-			     ALU_INST(SQ_OP2_INST_DOT4),
-			     BANK_SWIZZLE(SQ_ALU_VEC_102),
-			     DST_GPR(2),
-			     DST_REL(ABSOLUTE),
-			     DST_ELEM(ELEM_Y),
-			     CLAMP(1));
-    /* 13 - alu 10 */
-    /* DP4 gpr[2].z gpr[1].z c[1].z */
-    ps[i++] = ALU_DWORD0(SRC0_SEL(1),
-			 SRC0_REL(ABSOLUTE),
-			 SRC0_ELEM(ELEM_Z),
-			 SRC0_NEG(0),
-			 SRC1_SEL(257),
-			 SRC1_REL(ABSOLUTE),
-			 SRC1_ELEM(ELEM_Z),
-			 SRC1_NEG(0),
-			 INDEX_MODE(SQ_INDEX_LOOP),
-			 PRED_SEL(SQ_PRED_SEL_OFF),
-			 LAST(0));
-    ps[i++] = ALU_DWORD1_OP2(info->ChipFamily,
-			     SRC0_ABS(0),
-			     SRC1_ABS(0),
-			     UPDATE_EXECUTE_MASK(0),
-			     UPDATE_PRED(0),
-			     WRITE_MASK(0),
-			     FOG_MERGE(0),
-			     OMOD(SQ_ALU_OMOD_OFF),
-			     ALU_INST(SQ_OP2_INST_DOT4),
-			     BANK_SWIZZLE(SQ_ALU_VEC_102),
-			     DST_GPR(2),
-			     DST_REL(ABSOLUTE),
-			     DST_ELEM(ELEM_Z),
-			     CLAMP(1));
-    /* 14 - alu 11 */
-    /* DP4 gpr[2].w gpr[1].w c[1].w */
-    ps[i++] = ALU_DWORD0(SRC0_SEL(1),
-			 SRC0_REL(ABSOLUTE),
-			 SRC0_ELEM(ELEM_W),
-			 SRC0_NEG(0),
-			 SRC1_SEL(257),
-			 SRC1_REL(ABSOLUTE),
-			 SRC1_ELEM(ELEM_W),
-			 SRC1_NEG(0),
-			 INDEX_MODE(SQ_INDEX_LOOP),
-			 PRED_SEL(SQ_PRED_SEL_OFF),
-			 LAST(1));
-    ps[i++] = ALU_DWORD1_OP2(info->ChipFamily,
-			     SRC0_ABS(0),
-			     SRC1_ABS(0),
-			     UPDATE_EXECUTE_MASK(0),
-			     UPDATE_PRED(0),
-			     WRITE_MASK(0),
-			     FOG_MERGE(0),
-			     OMOD(SQ_ALU_OMOD_OFF),
-			     ALU_INST(SQ_OP2_INST_DOT4),
-			     BANK_SWIZZLE(SQ_ALU_VEC_021),
-			     DST_GPR(2),
-			     DST_REL(ABSOLUTE),
-			     DST_ELEM(ELEM_W),
-			     CLAMP(1));
-    /* 15 - alu 12 */
-    /* DP4 gpr[2].x gpr[1].x c[2].x */
-    ps[i++] = ALU_DWORD0(SRC0_SEL(1),
-			 SRC0_REL(ABSOLUTE),
-			 SRC0_ELEM(ELEM_X),
-			 SRC0_NEG(0),
-			 SRC1_SEL(258),
-			 SRC1_REL(ABSOLUTE),
-			 SRC1_ELEM(ELEM_X),
-			 SRC1_NEG(0),
-			 INDEX_MODE(SQ_INDEX_LOOP),
-			 PRED_SEL(SQ_PRED_SEL_OFF),
-			 LAST(0));
-    ps[i++] = ALU_DWORD1_OP2(info->ChipFamily,
-			     SRC0_ABS(0),
-			     SRC1_ABS(0),
-			     UPDATE_EXECUTE_MASK(0),
-			     UPDATE_PRED(0),
-			     WRITE_MASK(0),
-			     FOG_MERGE(0),
-			     OMOD(SQ_ALU_OMOD_OFF),
-			     ALU_INST(SQ_OP2_INST_DOT4),
-			     BANK_SWIZZLE(SQ_ALU_VEC_102),
-			     DST_GPR(2),
-			     DST_REL(ABSOLUTE),
-			     DST_ELEM(ELEM_X),
-			     CLAMP(1));
-    /* 16 - alu 13 */
-    /* DP4 gpr[2].y gpr[1].y c[2].y */
-    ps[i++] = ALU_DWORD0(SRC0_SEL(1),
-			 SRC0_REL(ABSOLUTE),
-			 SRC0_ELEM(ELEM_Y),
-			 SRC0_NEG(0),
-			 SRC1_SEL(258),
-			 SRC1_REL(ABSOLUTE),
-			 SRC1_ELEM(ELEM_Y),
-			 SRC1_NEG(0),
-			 INDEX_MODE(SQ_INDEX_LOOP),
-			 PRED_SEL(SQ_PRED_SEL_OFF),
-			 LAST(0));
-    ps[i++] = ALU_DWORD1_OP2(info->ChipFamily,
-			     SRC0_ABS(0),
-			     SRC1_ABS(0),
-			     UPDATE_EXECUTE_MASK(0),
-			     UPDATE_PRED(0),
-			     WRITE_MASK(0),
-			     FOG_MERGE(0),
-			     OMOD(SQ_ALU_OMOD_OFF),
-			     ALU_INST(SQ_OP2_INST_DOT4),
-			     BANK_SWIZZLE(SQ_ALU_VEC_102),
-			     DST_GPR(2),
-			     DST_REL(ABSOLUTE),
-			     DST_ELEM(ELEM_Y),
-			     CLAMP(1));
-    /* 17 - alu 14 */
-    /* DP4 gpr[2].z gpr[1].z c[2].z */
-    ps[i++] = ALU_DWORD0(SRC0_SEL(1),
-			 SRC0_REL(ABSOLUTE),
-			 SRC0_ELEM(ELEM_Z),
-			 SRC0_NEG(0),
-			 SRC1_SEL(258),
-			 SRC1_REL(ABSOLUTE),
-			 SRC1_ELEM(ELEM_Z),
-			 SRC1_NEG(0),
-			 INDEX_MODE(SQ_INDEX_LOOP),
-			 PRED_SEL(SQ_PRED_SEL_OFF),
-			 LAST(0));
-    ps[i++] = ALU_DWORD1_OP2(info->ChipFamily,
-			     SRC0_ABS(0),
-			     SRC1_ABS(0),
-			     UPDATE_EXECUTE_MASK(0),
-			     UPDATE_PRED(0),
-			     WRITE_MASK(1),
-			     FOG_MERGE(0),
-			     OMOD(SQ_ALU_OMOD_OFF),
-			     ALU_INST(SQ_OP2_INST_DOT4),
-			     BANK_SWIZZLE(SQ_ALU_VEC_102),
-			     DST_GPR(2),
-			     DST_REL(ABSOLUTE),
-			     DST_ELEM(ELEM_Z),
-			     CLAMP(1));
-    /* 18 - alu 15 */
-    /* DP4 gpr[2].w gpr[1].w c[2].w */
-    ps[i++] = ALU_DWORD0(SRC0_SEL(1),
-			 SRC0_REL(ABSOLUTE),
-			 SRC0_ELEM(ELEM_W),
-			 SRC0_NEG(0),
-			 SRC1_SEL(258),
-			 SRC1_REL(ABSOLUTE),
-			 SRC1_ELEM(ELEM_W),
-			 SRC1_NEG(0),
-			 INDEX_MODE(SQ_INDEX_LOOP),
-			 PRED_SEL(SQ_PRED_SEL_OFF),
-			 LAST(1));
-    ps[i++] = ALU_DWORD1_OP2(info->ChipFamily,
-			     SRC0_ABS(0),
-			     SRC1_ABS(0),
-			     UPDATE_EXECUTE_MASK(0),
-			     UPDATE_PRED(0),
-			     WRITE_MASK(0),
-			     FOG_MERGE(0),
-			     OMOD(SQ_ALU_OMOD_OFF),
-			     ALU_INST(SQ_OP2_INST_DOT4),
-			     BANK_SWIZZLE(SQ_ALU_VEC_021),
-			     DST_GPR(2),
-			     DST_REL(ABSOLUTE),
-			     DST_ELEM(ELEM_W),
-			     CLAMP(1));
-    /* 19 - alignment */
-    ps[i++] = 0x00000000;
-    ps[i++] = 0x00000000;
-    /* 20/21 - tex 0 */
-    ps[i++] = TEX_DWORD0(TEX_INST(SQ_TEX_INST_SAMPLE),
-			 BC_FRAC_MODE(0),
-			 FETCH_WHOLE_QUAD(0),
-			 RESOURCE_ID(0),
-			 SRC_GPR(0),
-			 SRC_REL(ABSOLUTE),
-			 R7xx_ALT_CONST(0));
-    ps[i++] = TEX_DWORD1(DST_GPR(1),
-			 DST_REL(ABSOLUTE),
-			 DST_SEL_X(SQ_SEL_X),    /* R */
-			 DST_SEL_Y(SQ_SEL_MASK), /* G */
-			 DST_SEL_Z(SQ_SEL_MASK), /* B */
-			 DST_SEL_W(SQ_SEL_1),    /* A */
-			 LOD_BIAS(0),
-			 COORD_TYPE_X(TEX_NORMALIZED),
-			 COORD_TYPE_Y(TEX_NORMALIZED),
-			 COORD_TYPE_Z(TEX_NORMALIZED),
-			 COORD_TYPE_W(TEX_NORMALIZED));
-    ps[i++] = TEX_DWORD2(OFFSET_X(0),
-			 OFFSET_Y(0),
-			 OFFSET_Z(0),
-			 SAMPLER_ID(0),
-			 SRC_SEL_X(SQ_SEL_X),
-			 SRC_SEL_Y(SQ_SEL_Y),
-			 SRC_SEL_Z(SQ_SEL_0),
-			 SRC_SEL_W(SQ_SEL_1));
-    ps[i++] = TEX_DWORD_PAD;
-    /* 22/23 - tex 1 */
-    ps[i++] = TEX_DWORD0(TEX_INST(SQ_TEX_INST_SAMPLE),
-			 BC_FRAC_MODE(0),
-			 FETCH_WHOLE_QUAD(0),
-			 RESOURCE_ID(1),
-			 SRC_GPR(0),
-			 SRC_REL(ABSOLUTE),
-			 R7xx_ALT_CONST(0));
-    ps[i++] = TEX_DWORD1(DST_GPR(1),
-			 DST_REL(ABSOLUTE),
-			 DST_SEL_X(SQ_SEL_MASK), /* R */
-			 DST_SEL_Y(SQ_SEL_X),    /* G */
-			 DST_SEL_Z(SQ_SEL_Y),    /* B */
-			 DST_SEL_W(SQ_SEL_MASK), /* A */
-			 LOD_BIAS(0),
-			 COORD_TYPE_X(TEX_NORMALIZED),
-			 COORD_TYPE_Y(TEX_NORMALIZED),
-			 COORD_TYPE_Z(TEX_NORMALIZED),
-			 COORD_TYPE_W(TEX_NORMALIZED));
-    ps[i++] = TEX_DWORD2(OFFSET_X(0),
-			 OFFSET_Y(0),
-			 OFFSET_Z(0),
-			 SAMPLER_ID(1),
-			 SRC_SEL_X(SQ_SEL_X),
-			 SRC_SEL_Y(SQ_SEL_Y),
-			 SRC_SEL_Z(SQ_SEL_0),
-			 SRC_SEL_W(SQ_SEL_1));
-    ps[i++] = TEX_DWORD_PAD;
-
-    /* xv ps planar ---------------------------------- */
-    i = accel_state->xv_ps_offset_planar / 4;
-    /* 0 */
-    ps[i++] = CF_DWORD0(ADDR(20));
-    ps[i++] = CF_DWORD1(POP_COUNT(0),
-			CF_CONST(0),
-			COND(SQ_CF_COND_ACTIVE),
-			I_COUNT(3),
-			CALL_COUNT(0),
-			END_OF_PROGRAM(0),
-			VALID_PIXEL_MODE(0),
-			CF_INST(SQ_CF_INST_TEX),
-			WHOLE_QUAD_MODE(0),
-			BARRIER(0));
-    /* 1 */
-    ps[i++] = CF_ALU_DWORD0(ADDR(3),
-			    KCACHE_BANK0(0),
-			    KCACHE_BANK1(0),
-			    KCACHE_MODE0(SQ_CF_KCACHE_NOP));
-    ps[i++] = CF_ALU_DWORD1(KCACHE_MODE1(SQ_CF_KCACHE_NOP),
-			    KCACHE_ADDR0(0),
-			    KCACHE_ADDR1(0),
-			    I_COUNT(16),
-			    USES_WATERFALL(0),
-			    CF_INST(SQ_CF_INST_ALU),
-			    WHOLE_QUAD_MODE(0),
-			    BARRIER(1));
-    /* 2 */
-    ps[i++] = CF_ALLOC_IMP_EXP_DWORD0(ARRAY_BASE(CF_PIXEL_MRT0),
-				      TYPE(SQ_EXPORT_PIXEL),
-				      RW_GPR(2),
-				      RW_REL(ABSOLUTE),
-				      INDEX_GPR(0),
-				      ELEM_SIZE(3));
-    ps[i++] = CF_ALLOC_IMP_EXP_DWORD1_SWIZ(SRC_SEL_X(SQ_SEL_X),
-					   SRC_SEL_Y(SQ_SEL_Y),
-					   SRC_SEL_Z(SQ_SEL_Z),
-					   SRC_SEL_W(SQ_SEL_W),
-					   R6xx_ELEM_LOOP(0),
-					   BURST_COUNT(1),
-					   END_OF_PROGRAM(1),
-					   VALID_PIXEL_MODE(0),
-					   CF_INST(SQ_CF_INST_EXPORT_DONE),
-					   WHOLE_QUAD_MODE(0),
-					   BARRIER(1));
-    /* Undo scaling of Y'CbCr values
-     *  Y' is scaled from 16:235
-     *  Cb/Cr are scaled from 16:240
-     */
-    /* 3 - alu 0 */
-    /* MULADD gpr[1].x gpr[1].x c[3].x c[3].y */
-    ps[i++] = ALU_DWORD0(SRC0_SEL(1),
-			 SRC0_REL(ABSOLUTE),
-			 SRC0_ELEM(ELEM_X),
-			 SRC0_NEG(0),
-			 SRC1_SEL(259),
-			 SRC1_REL(ABSOLUTE),
-			 SRC1_ELEM(ELEM_X),
-			 SRC1_NEG(0),
-			 INDEX_MODE(SQ_INDEX_LOOP),
-			 PRED_SEL(SQ_PRED_SEL_OFF),
-			 LAST(0));
-    ps[i++] = ALU_DWORD1_OP3(SRC2_SEL(259),
-			     SRC2_REL(ABSOLUTE),
-			     SRC2_ELEM(ELEM_Y),
-			     SRC2_NEG(0),
-			     ALU_INST(SQ_OP3_INST_MULADD),
-			     BANK_SWIZZLE(SQ_ALU_VEC_012),
-			     DST_GPR(1),
-			     DST_REL(ABSOLUTE),
-			     DST_ELEM(ELEM_X),
-			     CLAMP(1));
-    /* 4 - alu 1 */
-    /* MULADD gpr[1].y gpr[1].y c[3].z c[3].w */
-    ps[i++] = ALU_DWORD0(SRC0_SEL(1),
-			 SRC0_REL(ABSOLUTE),
-			 SRC0_ELEM(ELEM_Y),
-			 SRC0_NEG(0),
-			 SRC1_SEL(259),
-			 SRC1_REL(ABSOLUTE),
-			 SRC1_ELEM(ELEM_Z),
-			 SRC1_NEG(0),
-			 INDEX_MODE(SQ_INDEX_LOOP),
-			 PRED_SEL(SQ_PRED_SEL_OFF),
-			 LAST(0));
-    ps[i++] = ALU_DWORD1_OP3(SRC2_SEL(259),
-			     SRC2_REL(ABSOLUTE),
-			     SRC2_ELEM(ELEM_W),
-			     SRC2_NEG(0),
-			     ALU_INST(SQ_OP3_INST_MULADD),
-			     BANK_SWIZZLE(SQ_ALU_VEC_012),
-			     DST_GPR(1),
-			     DST_REL(ABSOLUTE),
-			     DST_ELEM(ELEM_Y),
-			     CLAMP(0));
-    /* 5 - alu 2 */
-    /* MULADD gpr[1].z gpr[1].z c[3].z c[3].w */
-    ps[i++] = ALU_DWORD0(SRC0_SEL(1),
-			 SRC0_REL(ABSOLUTE),
-			 SRC0_ELEM(ELEM_Z),
-			 SRC0_NEG(0),
-			 SRC1_SEL(259),
-			 SRC1_REL(ABSOLUTE),
-			 SRC1_ELEM(ELEM_Z),
-			 SRC1_NEG(0),
-			 INDEX_MODE(SQ_INDEX_LOOP),
-			 PRED_SEL(SQ_PRED_SEL_OFF),
-			 LAST(0));
-    ps[i++] = ALU_DWORD1_OP3(SRC2_SEL(259),
-			     SRC2_REL(ABSOLUTE),
-			     SRC2_ELEM(ELEM_W),
-			     SRC2_NEG(0),
-			     ALU_INST(SQ_OP3_INST_MULADD),
-			     BANK_SWIZZLE(SQ_ALU_VEC_012),
-			     DST_GPR(1),
-			     DST_REL(ABSOLUTE),
-			     DST_ELEM(ELEM_Z),
-			     CLAMP(0));
-    /* 6 - alu 3 */
-    /* MOV gpr[1].w 0.0 */
-    ps[i++] = ALU_DWORD0(SRC0_SEL(SQ_ALU_SRC_0),
-			 SRC0_REL(ABSOLUTE),
-			 SRC0_ELEM(ELEM_X),
-			 SRC0_NEG(0),
-			 SRC1_SEL(SQ_ALU_SRC_0),
-			 SRC1_REL(ABSOLUTE),
-			 SRC1_ELEM(ELEM_X),
-			 SRC1_NEG(0),
-			 INDEX_MODE(SQ_INDEX_LOOP),
-			 PRED_SEL(SQ_PRED_SEL_OFF),
-			 LAST(1));
-    ps[i++] = ALU_DWORD1_OP2(info->ChipFamily,
-			     SRC0_ABS(0),
-			     SRC1_ABS(0),
-			     UPDATE_EXECUTE_MASK(0),
-			     UPDATE_PRED(0),
-			     WRITE_MASK(1),
-			     FOG_MERGE(0),
-			     OMOD(SQ_ALU_OMOD_OFF),
-			     ALU_INST(SQ_OP2_INST_DOT4),
-			     BANK_SWIZZLE(SQ_ALU_VEC_102),
-			     DST_GPR(1),
-			     DST_REL(ABSOLUTE),
-			     DST_ELEM(ELEM_W),
-			     CLAMP(0));
-    /* 7 - alu 4 */
-    /* DP4 gpr[2].x gpr[1].x c[0].x */
-    ps[i++] = ALU_DWORD0(SRC0_SEL(1),
-			 SRC0_REL(ABSOLUTE),
-			 SRC0_ELEM(ELEM_X),
-			 SRC0_NEG(0),
-			 SRC1_SEL(256),
-			 SRC1_REL(ABSOLUTE),
-			 SRC1_ELEM(ELEM_X),
-			 SRC1_NEG(0),
-			 INDEX_MODE(SQ_INDEX_LOOP),
-			 PRED_SEL(SQ_PRED_SEL_OFF),
-			 LAST(0));
-    ps[i++] = ALU_DWORD1_OP2(info->ChipFamily,
-			     SRC0_ABS(0),
-			     SRC1_ABS(0),
-			     UPDATE_EXECUTE_MASK(0),
-			     UPDATE_PRED(0),
-			     WRITE_MASK(1),
-			     FOG_MERGE(0),
-			     OMOD(SQ_ALU_OMOD_OFF),
-			     ALU_INST(SQ_OP2_INST_DOT4),
-			     BANK_SWIZZLE(SQ_ALU_VEC_102),
-			     DST_GPR(2),
-			     DST_REL(ABSOLUTE),
-			     DST_ELEM(ELEM_X),
-			     CLAMP(1));
-    /* 8 - alu 5 */
-    /* DP4 gpr[2].y gpr[1].y c[0].y */
-    ps[i++] = ALU_DWORD0(SRC0_SEL(1),
-			 SRC0_REL(ABSOLUTE),
-			 SRC0_ELEM(ELEM_Y),
-			 SRC0_NEG(0),
-			 SRC1_SEL(256),
-			 SRC1_REL(ABSOLUTE),
-			 SRC1_ELEM(ELEM_Y),
-			 SRC1_NEG(0),
-			 INDEX_MODE(SQ_INDEX_LOOP),
-			 PRED_SEL(SQ_PRED_SEL_OFF),
-			 LAST(0));
-    ps[i++] = ALU_DWORD1_OP2(info->ChipFamily,
-			     SRC0_ABS(0),
-			     SRC1_ABS(0),
-			     UPDATE_EXECUTE_MASK(0),
-			     UPDATE_PRED(0),
-			     WRITE_MASK(0),
-			     FOG_MERGE(0),
-			     OMOD(SQ_ALU_OMOD_OFF),
-			     ALU_INST(SQ_OP2_INST_DOT4),
-			     BANK_SWIZZLE(SQ_ALU_VEC_102),
-			     DST_GPR(2),
-			     DST_REL(ABSOLUTE),
-			     DST_ELEM(ELEM_Y),
-			     CLAMP(1));
-    /* 9 - alu 6 */
-    /* DP4 gpr[2].z gpr[1].z c[0].z */
-    ps[i++] = ALU_DWORD0(SRC0_SEL(1),
-			 SRC0_REL(ABSOLUTE),
-			 SRC0_ELEM(ELEM_Z),
-			 SRC0_NEG(0),
-			 SRC1_SEL(256),
-			 SRC1_REL(ABSOLUTE),
-			 SRC1_ELEM(ELEM_Z),
-			 SRC1_NEG(0),
-			 INDEX_MODE(SQ_INDEX_LOOP),
-			 PRED_SEL(SQ_PRED_SEL_OFF),
-			 LAST(0));
-    ps[i++] = ALU_DWORD1_OP2(info->ChipFamily,
-			     SRC0_ABS(0),
-			     SRC1_ABS(0),
-			     UPDATE_EXECUTE_MASK(0),
-			     UPDATE_PRED(0),
-			     WRITE_MASK(0),
-			     FOG_MERGE(0),
-			     OMOD(SQ_ALU_OMOD_OFF),
-			     ALU_INST(SQ_OP2_INST_DOT4),
-			     BANK_SWIZZLE(SQ_ALU_VEC_102),
-			     DST_GPR(2),
-			     DST_REL(ABSOLUTE),
-			     DST_ELEM(ELEM_Z),
-			     CLAMP(1));
-    /* 10 - alu 7 */
-    /* DP4 gpr[2].w gpr[1].w c[0].w */
-    ps[i++] = ALU_DWORD0(SRC0_SEL(1),
-			 SRC0_REL(ABSOLUTE),
-			 SRC0_ELEM(ELEM_W),
-			 SRC0_NEG(0),
-			 SRC1_SEL(256),
-			 SRC1_REL(ABSOLUTE),
-			 SRC1_ELEM(ELEM_W),
-			 SRC1_NEG(0),
-			 INDEX_MODE(SQ_INDEX_LOOP),
-			 PRED_SEL(SQ_PRED_SEL_OFF),
-			 LAST(1));
-    ps[i++] = ALU_DWORD1_OP2(info->ChipFamily,
-			     SRC0_ABS(0),
-			     SRC1_ABS(0),
-			     UPDATE_EXECUTE_MASK(0),
-			     UPDATE_PRED(0),
-			     WRITE_MASK(0),
-			     FOG_MERGE(0),
-			     OMOD(SQ_ALU_OMOD_OFF),
-			     ALU_INST(SQ_OP2_INST_DOT4),
-			     BANK_SWIZZLE(SQ_ALU_VEC_021),
-			     DST_GPR(2),
-			     DST_REL(ABSOLUTE),
-			     DST_ELEM(ELEM_W),
-			     CLAMP(1));
-    /* 11 - alu 8 */
-    /* DP4 gpr[2].x gpr[1].x c[1].x */
-    ps[i++] = ALU_DWORD0(SRC0_SEL(1),
-			 SRC0_REL(ABSOLUTE),
-			 SRC0_ELEM(ELEM_X),
-			 SRC0_NEG(0),
-			 SRC1_SEL(257),
-			 SRC1_REL(ABSOLUTE),
-			 SRC1_ELEM(ELEM_X),
-			 SRC1_NEG(0),
-			 INDEX_MODE(SQ_INDEX_LOOP),
-			 PRED_SEL(SQ_PRED_SEL_OFF),
-			 LAST(0));
-    ps[i++] = ALU_DWORD1_OP2(info->ChipFamily,
-			     SRC0_ABS(0),
-			     SRC1_ABS(0),
-			     UPDATE_EXECUTE_MASK(0),
-			     UPDATE_PRED(0),
-			     WRITE_MASK(0),
-			     FOG_MERGE(0),
-			     OMOD(SQ_ALU_OMOD_OFF),
-			     ALU_INST(SQ_OP2_INST_DOT4),
-			     BANK_SWIZZLE(SQ_ALU_VEC_102),
-			     DST_GPR(2),
-			     DST_REL(ABSOLUTE),
-			     DST_ELEM(ELEM_X),
-			     CLAMP(1));
-    /* 12 - alu 9 */
-    /* DP4 gpr[2].y gpr[1].y c[1].y */
-    ps[i++] = ALU_DWORD0(SRC0_SEL(1),
-			 SRC0_REL(ABSOLUTE),
-			 SRC0_ELEM(ELEM_Y),
-			 SRC0_NEG(0),
-			 SRC1_SEL(257),
-			 SRC1_REL(ABSOLUTE),
-			 SRC1_ELEM(ELEM_Y),
-			 SRC1_NEG(0),
-			 INDEX_MODE(SQ_INDEX_LOOP),
-			 PRED_SEL(SQ_PRED_SEL_OFF),
-			 LAST(0));
-    ps[i++] = ALU_DWORD1_OP2(info->ChipFamily,
-			     SRC0_ABS(0),
-			     SRC1_ABS(0),
-			     UPDATE_EXECUTE_MASK(0),
-			     UPDATE_PRED(0),
-			     WRITE_MASK(1),
-			     FOG_MERGE(0),
-			     OMOD(SQ_ALU_OMOD_OFF),
-			     ALU_INST(SQ_OP2_INST_DOT4),
-			     BANK_SWIZZLE(SQ_ALU_VEC_102),
-			     DST_GPR(2),
-			     DST_REL(ABSOLUTE),
-			     DST_ELEM(ELEM_Y),
-			     CLAMP(1));
-    /* 13 - alu 10 */
-    /* DP4 gpr[2].z gpr[1].z c[1].z */
-    ps[i++] = ALU_DWORD0(SRC0_SEL(1),
-			 SRC0_REL(ABSOLUTE),
-			 SRC0_ELEM(ELEM_Z),
-			 SRC0_NEG(0),
-			 SRC1_SEL(257),
-			 SRC1_REL(ABSOLUTE),
-			 SRC1_ELEM(ELEM_Z),
-			 SRC1_NEG(0),
-			 INDEX_MODE(SQ_INDEX_LOOP),
-			 PRED_SEL(SQ_PRED_SEL_OFF),
-			 LAST(0));
-    ps[i++] = ALU_DWORD1_OP2(info->ChipFamily,
-			     SRC0_ABS(0),
-			     SRC1_ABS(0),
-			     UPDATE_EXECUTE_MASK(0),
-			     UPDATE_PRED(0),
-			     WRITE_MASK(0),
-			     FOG_MERGE(0),
-			     OMOD(SQ_ALU_OMOD_OFF),
-			     ALU_INST(SQ_OP2_INST_DOT4),
-			     BANK_SWIZZLE(SQ_ALU_VEC_102),
-			     DST_GPR(2),
-			     DST_REL(ABSOLUTE),
-			     DST_ELEM(ELEM_Z),
-			     CLAMP(1));
-    /* 14 - alu 11 */
-    /* DP4 gpr[2].w gpr[1].w c[1].w */
-    ps[i++] = ALU_DWORD0(SRC0_SEL(1),
-			 SRC0_REL(ABSOLUTE),
-			 SRC0_ELEM(ELEM_W),
-			 SRC0_NEG(0),
-			 SRC1_SEL(257),
-			 SRC1_REL(ABSOLUTE),
-			 SRC1_ELEM(ELEM_W),
-			 SRC1_NEG(0),
-			 INDEX_MODE(SQ_INDEX_LOOP),
-			 PRED_SEL(SQ_PRED_SEL_OFF),
-			 LAST(1));
-    ps[i++] = ALU_DWORD1_OP2(info->ChipFamily,
-			     SRC0_ABS(0),
-			     SRC1_ABS(0),
-			     UPDATE_EXECUTE_MASK(0),
-			     UPDATE_PRED(0),
-			     WRITE_MASK(0),
-			     FOG_MERGE(0),
-			     OMOD(SQ_ALU_OMOD_OFF),
-			     ALU_INST(SQ_OP2_INST_DOT4),
-			     BANK_SWIZZLE(SQ_ALU_VEC_021),
-			     DST_GPR(2),
-			     DST_REL(ABSOLUTE),
-			     DST_ELEM(ELEM_W),
-			     CLAMP(1));
-    /* 15 - alu 12 */
-    /* DP4 gpr[2].x gpr[1].x c[2].x */
-    ps[i++] = ALU_DWORD0(SRC0_SEL(1),
-			 SRC0_REL(ABSOLUTE),
-			 SRC0_ELEM(ELEM_X),
-			 SRC0_NEG(0),
-			 SRC1_SEL(258),
-			 SRC1_REL(ABSOLUTE),
-			 SRC1_ELEM(ELEM_X),
-			 SRC1_NEG(0),
-			 INDEX_MODE(SQ_INDEX_LOOP),
-			 PRED_SEL(SQ_PRED_SEL_OFF),
-			 LAST(0));
-    ps[i++] = ALU_DWORD1_OP2(info->ChipFamily,
-			     SRC0_ABS(0),
-			     SRC1_ABS(0),
-			     UPDATE_EXECUTE_MASK(0),
-			     UPDATE_PRED(0),
-			     WRITE_MASK(0),
-			     FOG_MERGE(0),
-			     OMOD(SQ_ALU_OMOD_OFF),
-			     ALU_INST(SQ_OP2_INST_DOT4),
-			     BANK_SWIZZLE(SQ_ALU_VEC_102),
-			     DST_GPR(2),
-			     DST_REL(ABSOLUTE),
-			     DST_ELEM(ELEM_X),
-			     CLAMP(1));
-    /* 16 - alu 13 */
-    /* DP4 gpr[2].y gpr[1].y c[2].y */
-    ps[i++] = ALU_DWORD0(SRC0_SEL(1),
-			 SRC0_REL(ABSOLUTE),
-			 SRC0_ELEM(ELEM_Y),
-			 SRC0_NEG(0),
-			 SRC1_SEL(258),
-			 SRC1_REL(ABSOLUTE),
-			 SRC1_ELEM(ELEM_Y),
-			 SRC1_NEG(0),
-			 INDEX_MODE(SQ_INDEX_LOOP),
-			 PRED_SEL(SQ_PRED_SEL_OFF),
-			 LAST(0));
-    ps[i++] = ALU_DWORD1_OP2(info->ChipFamily,
-			     SRC0_ABS(0),
-			     SRC1_ABS(0),
-			     UPDATE_EXECUTE_MASK(0),
-			     UPDATE_PRED(0),
-			     WRITE_MASK(0),
-			     FOG_MERGE(0),
-			     OMOD(SQ_ALU_OMOD_OFF),
-			     ALU_INST(SQ_OP2_INST_DOT4),
-			     BANK_SWIZZLE(SQ_ALU_VEC_102),
-			     DST_GPR(2),
-			     DST_REL(ABSOLUTE),
-			     DST_ELEM(ELEM_Y),
-			     CLAMP(1));
-    /* 17 - alu 14 */
-    /* DP4 gpr[2].z gpr[1].z c[2].z */
-    ps[i++] = ALU_DWORD0(SRC0_SEL(1),
-			 SRC0_REL(ABSOLUTE),
-			 SRC0_ELEM(ELEM_Z),
-			 SRC0_NEG(0),
-			 SRC1_SEL(258),
-			 SRC1_REL(ABSOLUTE),
-			 SRC1_ELEM(ELEM_Z),
-			 SRC1_NEG(0),
-			 INDEX_MODE(SQ_INDEX_LOOP),
-			 PRED_SEL(SQ_PRED_SEL_OFF),
-			 LAST(0));
-    ps[i++] = ALU_DWORD1_OP2(info->ChipFamily,
-			     SRC0_ABS(0),
-			     SRC1_ABS(0),
-			     UPDATE_EXECUTE_MASK(0),
-			     UPDATE_PRED(0),
-			     WRITE_MASK(1),
-			     FOG_MERGE(0),
-			     OMOD(SQ_ALU_OMOD_OFF),
-			     ALU_INST(SQ_OP2_INST_DOT4),
-			     BANK_SWIZZLE(SQ_ALU_VEC_102),
-			     DST_GPR(2),
-			     DST_REL(ABSOLUTE),
-			     DST_ELEM(ELEM_Z),
-			     CLAMP(1));
-    /* 18 - alu 15 */
-    /* DP4 gpr[2].w gpr[1].w c[2].w */
-    ps[i++] = ALU_DWORD0(SRC0_SEL(1),
-			 SRC0_REL(ABSOLUTE),
-			 SRC0_ELEM(ELEM_W),
-			 SRC0_NEG(0),
-			 SRC1_SEL(258),
-			 SRC1_REL(ABSOLUTE),
-			 SRC1_ELEM(ELEM_W),
-			 SRC1_NEG(0),
-			 INDEX_MODE(SQ_INDEX_LOOP),
-			 PRED_SEL(SQ_PRED_SEL_OFF),
-			 LAST(1));
-    ps[i++] = ALU_DWORD1_OP2(info->ChipFamily,
-			     SRC0_ABS(0),
-			     SRC1_ABS(0),
-			     UPDATE_EXECUTE_MASK(0),
-			     UPDATE_PRED(0),
-			     WRITE_MASK(0),
-			     FOG_MERGE(0),
-			     OMOD(SQ_ALU_OMOD_OFF),
-			     ALU_INST(SQ_OP2_INST_DOT4),
-			     BANK_SWIZZLE(SQ_ALU_VEC_021),
-			     DST_GPR(2),
-			     DST_REL(ABSOLUTE),
-			     DST_ELEM(ELEM_W),
-			     CLAMP(1));
-    /* 19 - alignment */
-    ps[i++] = 0x00000000;
-    ps[i++] = 0x00000000;
-    /* 20/21 - tex 0 */
-    ps[i++] = TEX_DWORD0(TEX_INST(SQ_TEX_INST_SAMPLE),
-			 BC_FRAC_MODE(0),
-			 FETCH_WHOLE_QUAD(0),
-			 RESOURCE_ID(0),
-			 SRC_GPR(0),
-			 SRC_REL(ABSOLUTE),
-			 R7xx_ALT_CONST(0));
-    ps[i++] = TEX_DWORD1(DST_GPR(1),
-			 DST_REL(ABSOLUTE),
-			 DST_SEL_X(SQ_SEL_X),    /* R */
-			 DST_SEL_Y(SQ_SEL_MASK), /* G */
-			 DST_SEL_Z(SQ_SEL_MASK), /* B */
-			 DST_SEL_W(SQ_SEL_1),    /* A */
-			 LOD_BIAS(0),
-			 COORD_TYPE_X(TEX_NORMALIZED),
-			 COORD_TYPE_Y(TEX_NORMALIZED),
-			 COORD_TYPE_Z(TEX_NORMALIZED),
-			 COORD_TYPE_W(TEX_NORMALIZED));
-    ps[i++] = TEX_DWORD2(OFFSET_X(0),
-			 OFFSET_Y(0),
-			 OFFSET_Z(0),
-			 SAMPLER_ID(0),
-			 SRC_SEL_X(SQ_SEL_X),
-			 SRC_SEL_Y(SQ_SEL_Y),
-			 SRC_SEL_Z(SQ_SEL_0),
-			 SRC_SEL_W(SQ_SEL_1));
-    ps[i++] = TEX_DWORD_PAD;
-    /* 22/23 - tex 1 */
-    ps[i++] = TEX_DWORD0(TEX_INST(SQ_TEX_INST_SAMPLE),
-			 BC_FRAC_MODE(0),
-			 FETCH_WHOLE_QUAD(0),
-			 RESOURCE_ID(1),
-			 SRC_GPR(0),
-			 SRC_REL(ABSOLUTE),
-			 R7xx_ALT_CONST(0));
-    ps[i++] = TEX_DWORD1(DST_GPR(1),
-			 DST_REL(ABSOLUTE),
-			 DST_SEL_X(SQ_SEL_MASK), /* R */
-			 DST_SEL_Y(SQ_SEL_MASK), /* G */
-			 DST_SEL_Z(SQ_SEL_X),    /* B */
-			 DST_SEL_W(SQ_SEL_MASK), /* A */
-			 LOD_BIAS(0),
-			 COORD_TYPE_X(TEX_NORMALIZED),
-			 COORD_TYPE_Y(TEX_NORMALIZED),
-			 COORD_TYPE_Z(TEX_NORMALIZED),
-			 COORD_TYPE_W(TEX_NORMALIZED));
-    ps[i++] = TEX_DWORD2(OFFSET_X(0),
-			 OFFSET_Y(0),
-			 OFFSET_Z(0),
-			 SAMPLER_ID(1),
-			 SRC_SEL_X(SQ_SEL_X),
-			 SRC_SEL_Y(SQ_SEL_Y),
-			 SRC_SEL_Z(SQ_SEL_0),
-			 SRC_SEL_W(SQ_SEL_1));
-    ps[i++] = TEX_DWORD_PAD;
-    /* 24/25 - tex 2 */
-    ps[i++] = TEX_DWORD0(TEX_INST(SQ_TEX_INST_SAMPLE),
-			 BC_FRAC_MODE(0),
-			 FETCH_WHOLE_QUAD(0),
-			 RESOURCE_ID(2),
-			 SRC_GPR(0),
-			 SRC_REL(ABSOLUTE),
-			 R7xx_ALT_CONST(0));
-    ps[i++] = TEX_DWORD1(DST_GPR(1),
-			 DST_REL(ABSOLUTE),
-			 DST_SEL_X(SQ_SEL_MASK), /* R */
-			 DST_SEL_Y(SQ_SEL_X),    /* G */
-			 DST_SEL_Z(SQ_SEL_MASK), /* B */
-			 DST_SEL_W(SQ_SEL_MASK), /* A */
-			 LOD_BIAS(0),
-			 COORD_TYPE_X(TEX_NORMALIZED),
-			 COORD_TYPE_Y(TEX_NORMALIZED),
-			 COORD_TYPE_Z(TEX_NORMALIZED),
-			 COORD_TYPE_W(TEX_NORMALIZED));
-    ps[i++] = TEX_DWORD2(OFFSET_X(0),
-			 OFFSET_Y(0),
-			 OFFSET_Z(0),
-			 SAMPLER_ID(2),
-			 SRC_SEL_X(SQ_SEL_X),
-			 SRC_SEL_Y(SQ_SEL_Y),
-			 SRC_SEL_Z(SQ_SEL_0),
-			 SRC_SEL_W(SQ_SEL_1));
-    ps[i++] = TEX_DWORD_PAD;
-
-    /* comp mask vs --------------------------------------- */
-    i = accel_state->comp_mask_vs_offset / 4;
-    /* 0 */
-    vs[i++] = CF_DWORD0(ADDR(4));
-    vs[i++] = CF_DWORD1(POP_COUNT(0),
-			CF_CONST(0),
-			COND(SQ_CF_COND_ACTIVE),
-			I_COUNT(3),
-			CALL_COUNT(0),
-			END_OF_PROGRAM(0),
-			VALID_PIXEL_MODE(0),
-			CF_INST(SQ_CF_INST_VTX),
-			WHOLE_QUAD_MODE(0),
-			BARRIER(1));
-    /* 1 - dst */
-    vs[i++] = CF_ALLOC_IMP_EXP_DWORD0(ARRAY_BASE(CF_POS0),
-				      TYPE(SQ_EXPORT_POS),
-				      RW_GPR(2),
-				      RW_REL(ABSOLUTE),
-				      INDEX_GPR(0),
-				      ELEM_SIZE(0));
-    vs[i++] = CF_ALLOC_IMP_EXP_DWORD1_SWIZ(SRC_SEL_X(SQ_SEL_X),
-					   SRC_SEL_Y(SQ_SEL_Y),
-					   SRC_SEL_Z(SQ_SEL_Z),
-					   SRC_SEL_W(SQ_SEL_W),
-					   R6xx_ELEM_LOOP(0),
-					   BURST_COUNT(1),
-					   END_OF_PROGRAM(0),
-					   VALID_PIXEL_MODE(0),
-					   CF_INST(SQ_CF_INST_EXPORT_DONE),
-					   WHOLE_QUAD_MODE(0),
-					   BARRIER(1));
-    /* 2 - src */
-    vs[i++] = CF_ALLOC_IMP_EXP_DWORD0(ARRAY_BASE(0),
-				      TYPE(SQ_EXPORT_PARAM),
-				      RW_GPR(1),
-				      RW_REL(ABSOLUTE),
-				      INDEX_GPR(0),
-				      ELEM_SIZE(0));
-    vs[i++] = CF_ALLOC_IMP_EXP_DWORD1_SWIZ(SRC_SEL_X(SQ_SEL_X),
-					   SRC_SEL_Y(SQ_SEL_Y),
-					   SRC_SEL_Z(SQ_SEL_Z),
-					   SRC_SEL_W(SQ_SEL_W),
-					   R6xx_ELEM_LOOP(0),
-					   BURST_COUNT(1),
-					   END_OF_PROGRAM(0),
-					   VALID_PIXEL_MODE(0),
-					   CF_INST(SQ_CF_INST_EXPORT),
-					   WHOLE_QUAD_MODE(0),
-					   BARRIER(0));
-    /* 3 - mask */
-    vs[i++] = CF_ALLOC_IMP_EXP_DWORD0(ARRAY_BASE(1),
-				      TYPE(SQ_EXPORT_PARAM),
-				      RW_GPR(0),
-				      RW_REL(ABSOLUTE),
-				      INDEX_GPR(0),
-				      ELEM_SIZE(0));
-    vs[i++] = CF_ALLOC_IMP_EXP_DWORD1_SWIZ(SRC_SEL_X(SQ_SEL_X),
-					   SRC_SEL_Y(SQ_SEL_Y),
-					   SRC_SEL_Z(SQ_SEL_Z),
-					   SRC_SEL_W(SQ_SEL_W),
-					   R6xx_ELEM_LOOP(0),
-					   BURST_COUNT(1),
-					   END_OF_PROGRAM(1),
-					   VALID_PIXEL_MODE(0),
-					   CF_INST(SQ_CF_INST_EXPORT_DONE),
-					   WHOLE_QUAD_MODE(0),
-					   BARRIER(0));
-    /* 4/5 - dst */
-    vs[i++] = VTX_DWORD0(VTX_INST(SQ_VTX_INST_FETCH),
-			 FETCH_TYPE(SQ_VTX_FETCH_VERTEX_DATA),
-			 FETCH_WHOLE_QUAD(0),
-			 BUFFER_ID(0),
-			 SRC_GPR(0),
-			 SRC_REL(ABSOLUTE),
-			 SRC_SEL_X(SQ_SEL_X),
-			 MEGA_FETCH_COUNT(24));
-    vs[i++] = VTX_DWORD1_GPR(DST_GPR(2),
-			     DST_REL(0),
-			     DST_SEL_X(SQ_SEL_X),
-			     DST_SEL_Y(SQ_SEL_Y),
-			     DST_SEL_Z(SQ_SEL_0),
-			     DST_SEL_W(SQ_SEL_1),
-			     USE_CONST_FIELDS(0),
-			     DATA_FORMAT(FMT_32_32_FLOAT),
-			     NUM_FORMAT_ALL(SQ_NUM_FORMAT_NORM),
-			     FORMAT_COMP_ALL(SQ_FORMAT_COMP_SIGNED),
-			     SRF_MODE_ALL(SRF_MODE_ZERO_CLAMP_MINUS_ONE));
-    vs[i++] = VTX_DWORD2(OFFSET(0),
-			 ENDIAN_SWAP(ENDIAN_NONE),
-			 CONST_BUF_NO_STRIDE(0),
-			 MEGA_FETCH(1));
-    vs[i++] = VTX_DWORD_PAD;
-    /* 6/7 - src */
-    vs[i++] = VTX_DWORD0(VTX_INST(SQ_VTX_INST_FETCH),
-			 FETCH_TYPE(SQ_VTX_FETCH_VERTEX_DATA),
-			 FETCH_WHOLE_QUAD(0),
-			 BUFFER_ID(0),
-			 SRC_GPR(0),
-			 SRC_REL(ABSOLUTE),
-			 SRC_SEL_X(SQ_SEL_X),
-			 MEGA_FETCH_COUNT(8));
-    vs[i++] = VTX_DWORD1_GPR(DST_GPR(1),
-			     DST_REL(0),
-			     DST_SEL_X(SQ_SEL_X),
-			     DST_SEL_Y(SQ_SEL_Y),
-			     DST_SEL_Z(SQ_SEL_0),
-			     DST_SEL_W(SQ_SEL_1),
-			     USE_CONST_FIELDS(0),
-			     DATA_FORMAT(FMT_32_32_FLOAT),
-			     NUM_FORMAT_ALL(SQ_NUM_FORMAT_NORM),
-			     FORMAT_COMP_ALL(SQ_FORMAT_COMP_SIGNED),
-			     SRF_MODE_ALL(SRF_MODE_ZERO_CLAMP_MINUS_ONE));
-    vs[i++] = VTX_DWORD2(OFFSET(8),
-			 ENDIAN_SWAP(ENDIAN_NONE),
-			 CONST_BUF_NO_STRIDE(0),
-			 MEGA_FETCH(0));
-    vs[i++] = VTX_DWORD_PAD;
-    /* 8/9 - mask */
-    vs[i++] = VTX_DWORD0(VTX_INST(SQ_VTX_INST_FETCH),
-			 FETCH_TYPE(SQ_VTX_FETCH_VERTEX_DATA),
-			 FETCH_WHOLE_QUAD(0),
-			 BUFFER_ID(0),
-			 SRC_GPR(0),
-			 SRC_REL(ABSOLUTE),
-			 SRC_SEL_X(SQ_SEL_X),
-			 MEGA_FETCH_COUNT(8));
-    vs[i++] = VTX_DWORD1_GPR(DST_GPR(0),
-			     DST_REL(0),
-			     DST_SEL_X(SQ_SEL_X),
-			     DST_SEL_Y(SQ_SEL_Y),
-			     DST_SEL_Z(SQ_SEL_0),
-			     DST_SEL_W(SQ_SEL_1),
-			     USE_CONST_FIELDS(0),
-			     DATA_FORMAT(FMT_32_32_FLOAT),
-			     NUM_FORMAT_ALL(SQ_NUM_FORMAT_NORM),
-			     FORMAT_COMP_ALL(SQ_FORMAT_COMP_SIGNED),
-			     SRF_MODE_ALL(SRF_MODE_ZERO_CLAMP_MINUS_ONE));
-    vs[i++] = VTX_DWORD2(OFFSET(16),
-			 ENDIAN_SWAP(ENDIAN_NONE),
-			 CONST_BUF_NO_STRIDE(0),
-			 MEGA_FETCH(0));
-    vs[i++] = VTX_DWORD_PAD;
-
-    /* comp mask vs --------------------------------------- */
-    i = accel_state->comp_mask_vs_offset / 4;
-    /* 0 */
-    vs[i++] = CF_DWORD0(ADDR(4));
-    vs[i++] = CF_DWORD1(POP_COUNT(0),
-			CF_CONST(0),
-			COND(SQ_CF_COND_ACTIVE),
-			I_COUNT(3),
-			CALL_COUNT(0),
-			END_OF_PROGRAM(0),
-			VALID_PIXEL_MODE(0),
-			CF_INST(SQ_CF_INST_VTX),
-			WHOLE_QUAD_MODE(0),
-			BARRIER(1));
-    /* 1 - dst */
-    vs[i++] = CF_ALLOC_IMP_EXP_DWORD0(ARRAY_BASE(CF_POS0),
-				      TYPE(SQ_EXPORT_POS),
-				      RW_GPR(2),
-				      RW_REL(ABSOLUTE),
-				      INDEX_GPR(0),
-				      ELEM_SIZE(0));
-    vs[i++] = CF_ALLOC_IMP_EXP_DWORD1_SWIZ(SRC_SEL_X(SQ_SEL_X),
-					   SRC_SEL_Y(SQ_SEL_Y),
-					   SRC_SEL_Z(SQ_SEL_Z),
-					   SRC_SEL_W(SQ_SEL_W),
-					   R6xx_ELEM_LOOP(0),
-					   BURST_COUNT(1),
-					   END_OF_PROGRAM(0),
-					   VALID_PIXEL_MODE(0),
-					   CF_INST(SQ_CF_INST_EXPORT_DONE),
-					   WHOLE_QUAD_MODE(0),
-					   BARRIER(1));
-    /* 2 - src */
-    vs[i++] = CF_ALLOC_IMP_EXP_DWORD0(ARRAY_BASE(0),
-				      TYPE(SQ_EXPORT_PARAM),
-				      RW_GPR(1),
-				      RW_REL(ABSOLUTE),
-				      INDEX_GPR(0),
-				      ELEM_SIZE(0));
-    vs[i++] = CF_ALLOC_IMP_EXP_DWORD1_SWIZ(SRC_SEL_X(SQ_SEL_X),
-					   SRC_SEL_Y(SQ_SEL_Y),
-					   SRC_SEL_Z(SQ_SEL_Z),
-					   SRC_SEL_W(SQ_SEL_W),
-					   R6xx_ELEM_LOOP(0),
-					   BURST_COUNT(1),
-					   END_OF_PROGRAM(0),
-					   VALID_PIXEL_MODE(0),
-					   CF_INST(SQ_CF_INST_EXPORT),
-					   WHOLE_QUAD_MODE(0),
-					   BARRIER(0));
-    /* 3 - mask */
-    vs[i++] = CF_ALLOC_IMP_EXP_DWORD0(ARRAY_BASE(1),
-				      TYPE(SQ_EXPORT_PARAM),
-				      RW_GPR(0),
-				      RW_REL(ABSOLUTE),
-				      INDEX_GPR(0),
-				      ELEM_SIZE(0));
-    vs[i++] = CF_ALLOC_IMP_EXP_DWORD1_SWIZ(SRC_SEL_X(SQ_SEL_X),
-					   SRC_SEL_Y(SQ_SEL_Y),
-					   SRC_SEL_Z(SQ_SEL_Z),
-					   SRC_SEL_W(SQ_SEL_W),
-					   R6xx_ELEM_LOOP(0),
-					   BURST_COUNT(1),
-					   END_OF_PROGRAM(1),
-					   VALID_PIXEL_MODE(0),
-					   CF_INST(SQ_CF_INST_EXPORT_DONE),
-					   WHOLE_QUAD_MODE(0),
-					   BARRIER(0));
-    /* 4/5 - dst */
-    vs[i++] = VTX_DWORD0(VTX_INST(SQ_VTX_INST_FETCH),
-			 FETCH_TYPE(SQ_VTX_FETCH_VERTEX_DATA),
-			 FETCH_WHOLE_QUAD(0),
-			 BUFFER_ID(0),
-			 SRC_GPR(0),
-			 SRC_REL(ABSOLUTE),
-			 SRC_SEL_X(SQ_SEL_X),
-			 MEGA_FETCH_COUNT(24));
-    vs[i++] = VTX_DWORD1_GPR(DST_GPR(2),
-			     DST_REL(0),
-			     DST_SEL_X(SQ_SEL_X),
-			     DST_SEL_Y(SQ_SEL_Y),
-			     DST_SEL_Z(SQ_SEL_0),
-			     DST_SEL_W(SQ_SEL_1),
-			     USE_CONST_FIELDS(0),
-			     DATA_FORMAT(FMT_32_32_FLOAT),
-			     NUM_FORMAT_ALL(SQ_NUM_FORMAT_NORM),
-			     FORMAT_COMP_ALL(SQ_FORMAT_COMP_SIGNED),
-			     SRF_MODE_ALL(SRF_MODE_ZERO_CLAMP_MINUS_ONE));
-    vs[i++] = VTX_DWORD2(OFFSET(0),
-			 ENDIAN_SWAP(ENDIAN_NONE),
-			 CONST_BUF_NO_STRIDE(0),
-			 MEGA_FETCH(1));
-    vs[i++] = VTX_DWORD_PAD;
-    /* 6/7 - src */
-    vs[i++] = VTX_DWORD0(VTX_INST(SQ_VTX_INST_FETCH),
-			 FETCH_TYPE(SQ_VTX_FETCH_VERTEX_DATA),
-			 FETCH_WHOLE_QUAD(0),
-			 BUFFER_ID(0),
-			 SRC_GPR(0),
-			 SRC_REL(ABSOLUTE),
-			 SRC_SEL_X(SQ_SEL_X),
-			 MEGA_FETCH_COUNT(8));
-    vs[i++] = VTX_DWORD1_GPR(DST_GPR(1),
-			     DST_REL(0),
-			     DST_SEL_X(SQ_SEL_X),
-			     DST_SEL_Y(SQ_SEL_Y),
-			     DST_SEL_Z(SQ_SEL_0),
-			     DST_SEL_W(SQ_SEL_1),
-			     USE_CONST_FIELDS(0),
-			     DATA_FORMAT(FMT_32_32_FLOAT),
-			     NUM_FORMAT_ALL(SQ_NUM_FORMAT_NORM),
-			     FORMAT_COMP_ALL(SQ_FORMAT_COMP_SIGNED),
-			     SRF_MODE_ALL(SRF_MODE_ZERO_CLAMP_MINUS_ONE));
-    vs[i++] = VTX_DWORD2(OFFSET(8),
-			 ENDIAN_SWAP(ENDIAN_NONE),
-			 CONST_BUF_NO_STRIDE(0),
-			 MEGA_FETCH(0));
-    vs[i++] = VTX_DWORD_PAD;
-    /* 8/9 - mask */
-    vs[i++] = VTX_DWORD0(VTX_INST(SQ_VTX_INST_FETCH),
-			 FETCH_TYPE(SQ_VTX_FETCH_VERTEX_DATA),
-			 FETCH_WHOLE_QUAD(0),
-			 BUFFER_ID(0),
-			 SRC_GPR(0),
-			 SRC_REL(ABSOLUTE),
-			 SRC_SEL_X(SQ_SEL_X),
-			 MEGA_FETCH_COUNT(8));
-    vs[i++] = VTX_DWORD1_GPR(DST_GPR(0),
-			     DST_REL(0),
-			     DST_SEL_X(SQ_SEL_X),
-			     DST_SEL_Y(SQ_SEL_Y),
-			     DST_SEL_Z(SQ_SEL_0),
-			     DST_SEL_W(SQ_SEL_1),
-			     USE_CONST_FIELDS(0),
-			     DATA_FORMAT(FMT_32_32_FLOAT),
-			     NUM_FORMAT_ALL(SQ_NUM_FORMAT_NORM),
-			     FORMAT_COMP_ALL(SQ_FORMAT_COMP_SIGNED),
-			     SRF_MODE_ALL(SRF_MODE_ZERO_CLAMP_MINUS_ONE));
-    vs[i++] = VTX_DWORD2(OFFSET(16),
-			 ENDIAN_SWAP(ENDIAN_NONE),
-			 CONST_BUF_NO_STRIDE(0),
-			 MEGA_FETCH(0));
-    vs[i++] = VTX_DWORD_PAD;
-
-    // comp mask ps --------------------------------------- */
-    /* not yet */
-
-    /* comp vs --------------------------------------- */
-    i = accel_state->comp_vs_offset / 4;
-    /* 0 */
-    vs[i++] = CF_DWORD0(ADDR(4));
-    vs[i++] = CF_DWORD1(POP_COUNT(0),
-			CF_CONST(0),
-			COND(SQ_CF_COND_ACTIVE),
-			I_COUNT(2),
-			CALL_COUNT(0),
-			END_OF_PROGRAM(0),
-			VALID_PIXEL_MODE(0),
-			CF_INST(SQ_CF_INST_VTX),
-			WHOLE_QUAD_MODE(0),
-			BARRIER(1));
-    /* 1 - dst */
-    vs[i++] = CF_ALLOC_IMP_EXP_DWORD0(ARRAY_BASE(CF_POS0),
-				      TYPE(SQ_EXPORT_POS),
-				      RW_GPR(1),
-				      RW_REL(ABSOLUTE),
-				      INDEX_GPR(0),
-				      ELEM_SIZE(0));
-    vs[i++] = CF_ALLOC_IMP_EXP_DWORD1_SWIZ(SRC_SEL_X(SQ_SEL_X),
-					   SRC_SEL_Y(SQ_SEL_Y),
-					   SRC_SEL_Z(SQ_SEL_Z),
-					   SRC_SEL_W(SQ_SEL_W),
-					   R6xx_ELEM_LOOP(0),
-					   BURST_COUNT(0),
-					   END_OF_PROGRAM(0),
-					   VALID_PIXEL_MODE(0),
-					   CF_INST(SQ_CF_INST_EXPORT_DONE),
-					   WHOLE_QUAD_MODE(0),
-					   BARRIER(1));
-    /* 2 - src */
-    vs[i++] = CF_ALLOC_IMP_EXP_DWORD0(ARRAY_BASE(0),
-				      TYPE(SQ_EXPORT_PARAM),
-				      RW_GPR(0),
-				      RW_REL(ABSOLUTE),
-				      INDEX_GPR(0),
-				      ELEM_SIZE(0));
-    vs[i++] = CF_ALLOC_IMP_EXP_DWORD1_SWIZ(SRC_SEL_X(SQ_SEL_X),
-					   SRC_SEL_Y(SQ_SEL_Y),
-					   SRC_SEL_Z(SQ_SEL_Z),
-					   SRC_SEL_W(SQ_SEL_W),
-					   R6xx_ELEM_LOOP(0),
-					   BURST_COUNT(0),
-					   END_OF_PROGRAM(1),
-					   VALID_PIXEL_MODE(0),
-					   CF_INST(SQ_CF_INST_EXPORT_DONE),
-					   WHOLE_QUAD_MODE(0),
-					   BARRIER(0));
-    /* 3 */
-    vs[i++] = 0x00000000;
-    vs[i++] = 0x00000000;
-    /* 4/5 - dst */
-    vs[i++] = VTX_DWORD0(VTX_INST(SQ_VTX_INST_FETCH),
-			 FETCH_TYPE(SQ_VTX_FETCH_VERTEX_DATA),
-			 FETCH_WHOLE_QUAD(0),
-			 BUFFER_ID(0),
-			 SRC_GPR(0),
-			 SRC_REL(ABSOLUTE),
-			 SRC_SEL_X(SQ_SEL_X),
-			 MEGA_FETCH_COUNT(16));
-    vs[i++] = VTX_DWORD1_GPR(DST_GPR(1),
-			     DST_REL(0),
-			     DST_SEL_X(SQ_SEL_X),
-			     DST_SEL_Y(SQ_SEL_Y),
-			     DST_SEL_Z(SQ_SEL_0),
-			     DST_SEL_W(SQ_SEL_1),
-			     USE_CONST_FIELDS(0),
-			     DATA_FORMAT(FMT_32_32_FLOAT),
-			     NUM_FORMAT_ALL(SQ_NUM_FORMAT_NORM),
-			     FORMAT_COMP_ALL(SQ_FORMAT_COMP_SIGNED),
-			     SRF_MODE_ALL(SRF_MODE_ZERO_CLAMP_MINUS_ONE));
-    vs[i++] = VTX_DWORD2(OFFSET(0),
-			 ENDIAN_SWAP(ENDIAN_NONE),
-			 CONST_BUF_NO_STRIDE(0),
-			 MEGA_FETCH(1));
-    vs[i++] = VTX_DWORD_PAD;
-    /* 6/7 - src */
-    vs[i++] = VTX_DWORD0(VTX_INST(SQ_VTX_INST_FETCH),
-			 FETCH_TYPE(SQ_VTX_FETCH_VERTEX_DATA),
-			 FETCH_WHOLE_QUAD(0),
-			 BUFFER_ID(0),
-			 SRC_GPR(0),
-			 SRC_REL(ABSOLUTE),
-			 SRC_SEL_X(SQ_SEL_X),
-			 MEGA_FETCH_COUNT(8));
-    vs[i++] = VTX_DWORD1_GPR(DST_GPR(0),
-			     DST_REL(0),
-			     DST_SEL_X(SQ_SEL_X),
-			     DST_SEL_Y(SQ_SEL_Y),
-			     DST_SEL_Z(SQ_SEL_0),
-			     DST_SEL_W(SQ_SEL_1),
-			     USE_CONST_FIELDS(0),
-			     DATA_FORMAT(FMT_32_32_FLOAT),
-			     NUM_FORMAT_ALL(SQ_NUM_FORMAT_NORM),
-			     FORMAT_COMP_ALL(SQ_FORMAT_COMP_SIGNED),
-			     SRF_MODE_ALL(SRF_MODE_ZERO_CLAMP_MINUS_ONE));
-    vs[i++] = VTX_DWORD2(OFFSET(8),
-			 ENDIAN_SWAP(ENDIAN_NONE),
-			 CONST_BUF_NO_STRIDE(0),
-			 MEGA_FETCH(0));
-    vs[i++] = VTX_DWORD_PAD;
-
-    /* comp ps --------------------------------------- */
-    /* not yet */
+    /*  xv ps packed --------------------------------------- */
+    accel_state->xv_ps_offset_packed = 4608;
+    R600_xv_ps_packet(ChipSet, shader + accel_state->xv_ps_offset_packed / 4);
 
+    /*  xv ps planar ---------------------------------- */
+    accel_state->xv_ps_offset_planar = 5120;
+    R600_xv_ps_planar(ChipSet, shader + accel_state->xv_ps_offset_planar / 4);
 
     return TRUE;
 }