some basic r500 portage

6 files changed, 742 insertions, 20 deletions

diff --git a/src/mesa/drivers/dri/r300/r300_cmdbuf.c b/src/mesa/drivers/dri/r300/r300_cmdbuf.c
index 3497738eacf..883b41a349b 100644
--- a/src/mesa/drivers/dri/r300/r300_cmdbuf.c
+++ b/src/mesa/drivers/dri/r300/r300_cmdbuf.c
@@ -281,10 +281,14 @@ void r300InitCmdBuf(r300ContextPtr r300)
 {
 	int size, mtu;
 	int has_tcl = 1;
+	int is_r500 = 0;
 
 	if (!(r300->radeon.radeonScreen->chip_flags & RADEON_CHIPSET_TCL))
 		has_tcl = 0;
 
+	if (r300->radeon.radeonScreen->chip_family >= CHIP_FAMILY_RV515)
+		is_r500 = 1;
+
 	r300->hw.max_state_size = 2 + 2;	/* reserve extra space for WAIT_IDLE and tex cache flush */
 
 	mtu = r300->radeon.glCtx->Const.MaxTextureUnits;
@@ -374,10 +378,17 @@ void r300InitCmdBuf(r300ContextPtr r300)
 	r300->hw.su_depth_scale.cmd[0] = cmdpacket0(R300_SU_DEPTH_SCALE, 2);
 	ALLOC_STATE(rc, always, R300_RC_CMDSIZE, 0);
 	r300->hw.rc.cmd[R300_RC_CMD_0] = cmdpacket0(R300_RS_COUNT, 2);
-	ALLOC_STATE(ri, always, R300_RI_CMDSIZE, 0);
-	r300->hw.ri.cmd[R300_RI_CMD_0] = cmdpacket0(R300_RS_IP_0, 8);
-	ALLOC_STATE(rr, variable, R300_RR_CMDSIZE, 0);
-	r300->hw.rr.cmd[R300_RR_CMD_0] = cmdpacket0(R300_RS_INST_0, 1);
+	if (is_r500) {
+		ALLOC_STATE(ri, always, R300_RI_CMDSIZE, 0);
+		r300->hw.ri.cmd[R300_RI_CMD_0] = cmdpacket0(R500_RS_IP_0, 16);
+		ALLOC_STATE(rr, variable, R300_RR_CMDSIZE, 0);
+		r300->hw.rr.cmd[R300_RR_CMD_0] = cmdpacket0(R500_RS_INST_0, 1);
+	} else {
+		ALLOC_STATE(ri, always, R300_RI_CMDSIZE, 0);
+		r300->hw.ri.cmd[R300_RI_CMD_0] = cmdpacket0(R300_RS_IP_0, 8);
+		ALLOC_STATE(rr, variable, R300_RR_CMDSIZE, 0);
+		r300->hw.rr.cmd[R300_RR_CMD_0] = cmdpacket0(R300_RS_ROUTE_0, 1);
+	}
 	ALLOC_STATE(sc_hyperz, always, 3, 0);
 	r300->hw.sc_hyperz.cmd[0] = cmdpacket0(R300_SC_HYPERZ, 2);
 	ALLOC_STATE(sc_screendoor, always, 2, 0);
@@ -389,14 +400,25 @@ void r300InitCmdBuf(r300ContextPtr r300)
 	r300->hw.fpt.cmd[R300_FPT_CMD_0] = cmdpacket0(R300_PFS_TEXI_0, 0);
 	ALLOC_STATE(us_out_fmt, always, 6, 0);
 	r300->hw.us_out_fmt.cmd[0] = cmdpacket0(R500_US_OUT_FMT, 5);
-	ALLOC_STATE(fpi[0], variable, R300_FPI_CMDSIZE, 0);
-	r300->hw.fpi[0].cmd[R300_FPI_CMD_0] = cmdpacket0(R300_PFS_INSTR0_0, 1);
-	ALLOC_STATE(fpi[1], variable, R300_FPI_CMDSIZE, 1);
-	r300->hw.fpi[1].cmd[R300_FPI_CMD_0] = cmdpacket0(R300_PFS_INSTR1_0, 1);
-	ALLOC_STATE(fpi[2], variable, R300_FPI_CMDSIZE, 2);
-	r300->hw.fpi[2].cmd[R300_FPI_CMD_0] = cmdpacket0(R300_PFS_INSTR2_0, 1);
-	ALLOC_STATE(fpi[3], variable, R300_FPI_CMDSIZE, 3);
-	r300->hw.fpi[3].cmd[R300_FPI_CMD_0] = cmdpacket0(R300_PFS_INSTR3_0, 1);
+
+	if (is_r500) {
+		ALLOC_STATE(r500fp, variable, R300_FPI_CMDSIZE, 0);
+		r300->hw.r500fp.cmd[R300_FPI_CMD_0] = cmdr500fp(0, 1);
+	}
+
+	if (0/*is_r500*/) {
+
+	}/* else*/
+	{
+		ALLOC_STATE(fpi[0], variable, R300_FPI_CMDSIZE, 0);
+		r300->hw.fpi[0].cmd[R300_FPI_CMD_0] = cmdpacket0(R300_PFS_INSTR0_0, 1);
+		ALLOC_STATE(fpi[1], variable, R300_FPI_CMDSIZE, 1);
+		r300->hw.fpi[1].cmd[R300_FPI_CMD_0] = cmdpacket0(R300_PFS_INSTR1_0, 1);
+		ALLOC_STATE(fpi[2], variable, R300_FPI_CMDSIZE, 2);
+		r300->hw.fpi[2].cmd[R300_FPI_CMD_0] = cmdpacket0(R300_PFS_INSTR2_0, 1);
+		ALLOC_STATE(fpi[3], variable, R300_FPI_CMDSIZE, 3);
+		r300->hw.fpi[3].cmd[R300_FPI_CMD_0] = cmdpacket0(R300_PFS_INSTR3_0, 1);
+	}
 	ALLOC_STATE(fogs, always, R300_FOGS_CMDSIZE, 0);
 	r300->hw.fogs.cmd[R300_FOGS_CMD_0] = cmdpacket0(FG_FOG_BLEND, 1);
 	ALLOC_STATE(fogc, always, R300_FOGC_CMDSIZE, 0);
@@ -456,10 +478,18 @@ void r300InitCmdBuf(r300ContextPtr r300)
 		r300->hw.vps.cmd[R300_VPS_CMD_0] =
 		    cmdvpu(R300_PVS_UPLOAD_POINTSIZE, 1);
 
-		for (i = 0; i < 6; i++) {
-		  ALLOC_STATE(vpucp[i], vpu, R300_VPUCP_CMDSIZE, 0);
-		  r300->hw.vpucp[i].cmd[R300_VPUCP_CMD_0] =
- 		    cmdvpu(R300_PVS_UPLOAD_CLIP_PLANE0+i, 1);
+		if (is_r500) {
+			for (i = 0; i < 6; i++) {
+				ALLOC_STATE(vpucp[i], vpu, R300_VPUCP_CMDSIZE, 0);
+				r300->hw.vpucp[i].cmd[R300_VPUCP_CMD_0] =
+					cmdvpu(R500_PVS_UPLOAD_CLIP_PLANE0+i, 1);
+			}
+		} else {
+			for (i = 0; i < 6; i++) {
+				ALLOC_STATE(vpucp[i], vpu, R300_VPUCP_CMDSIZE, 0);
+				r300->hw.vpucp[i].cmd[R300_VPUCP_CMD_0] =
+					cmdvpu(R300_PVS_UPLOAD_CLIP_PLANE0+i, 1);
+			}
 		}
 	}
 
diff --git a/src/mesa/drivers/dri/r300/r300_context.h b/src/mesa/drivers/dri/r300/r300_context.h
index 780d9aa5d28..ab45bf80b45 100644
--- a/src/mesa/drivers/dri/r300/r300_context.h
+++ b/src/mesa/drivers/dri/r300/r300_context.h
@@ -487,6 +487,7 @@ struct r300_hw_state {
 	struct r300_state_atom fp;	/* fragment program cntl + nodes (4600) */
 	struct r300_state_atom fpt;	/* texi - (4620) */
 	struct r300_state_atom us_out_fmt;	/* (46A4) */
+	struct r300_state_atom r500fp;	/* r500 fp instructions */
 	struct r300_state_atom fpi[4];	/* fp instructions (46C0/47C0/48C0/49C0) */
 	struct r300_state_atom fogs;	/* fog state (4BC0) */
 	struct r300_state_atom fogc;	/* fog color (4BC8) */
diff --git a/src/mesa/drivers/dri/r300/r300_emit.h b/src/mesa/drivers/dri/r300/r300_emit.h
index a6d69ec5ff8..a4f6ab997ed 100644
--- a/src/mesa/drivers/dri/r300/r300_emit.h
+++ b/src/mesa/drivers/dri/r300/r300_emit.h
@@ -74,6 +74,18 @@ static inline uint32_t cmdvpu(int addr, int count)
 	return cmd.u;
 }
 
+static inline uint32_t cmdr500fp(int addr, int count)
+{
+	drm_r300_cmd_header_t cmd;
+
+	cmd.vpu.cmd_type = R300_CMD_R500FP;
+	cmd.vpu.count = count;
+	cmd.vpu.adrhi = ((unsigned int)addr & 0xFF00) >> 8;
+	cmd.vpu.adrlo = ((unsigned int)addr & 0x00FF);
+
+	return cmd.u;
+}
+
 static inline uint32_t cmdpacket3(int packet)
 {
 	drm_r300_cmd_header_t cmd;
diff --git a/src/mesa/drivers/dri/r300/r300_reg.h b/src/mesa/drivers/dri/r300/r300_reg.h
index 2200cec6abe..d640d8b7e7d 100644
--- a/src/mesa/drivers/dri/r300/r300_reg.h
+++ b/src/mesa/drivers/dri/r300/r300_reg.h
@@ -657,7 +657,7 @@ USE OR OTHER DEALINGS IN THE SOFTWARE.
 /* This table specifies the source location and format for up to 16 texture
  * addresses (i[0]:i[15]) and four colors (c[0]:c[3])
  */
-#define R500_RS_IP_0					0x4074

+#define R500_RS_IP_0					0x4074
 #define R500_RS_IP_1					0x4078
 #define R500_RS_IP_2					0x407C
 #define R500_RS_IP_3					0x4080
@@ -1151,7 +1151,7 @@ USE OR OTHER DEALINGS IN THE SOFTWARE.
 
 
 /*  */
-#define R500_RS_INST_0					0x4320

+#define R500_RS_INST_0					0x4320
 #define R500_RS_INST_1					0x4324
 #define R500_RS_INST_2					0x4328
 #define R500_RS_INST_3					0x432c
@@ -2598,6 +2598,479 @@ enum {
 #define R300_PRIM_NUM_VERTICES_SHIFT            16
 #define R300_PRIM_NUM_VERTICES_MASK             0xffff
 
+
+
+/*
+ * The R500 unified shader (US) registers come in banks of 512 each, one
+ * for each instruction slot in the shader.  You can't touch them directly.
+ * R500_US_VECTOR_INDEX() sets the base instruction to modify; successive
+ * writes to R500_GA_US_VECTOR_DATA autoincrement the index after the
+ * instruction is fully specified.
+ */
+#define R500_US_ALU_ALPHA_INST_0			0xa800
+#   define R500_ALPHA_OP_MAD				0
+#   define R500_ALPHA_OP_DP				1
+#   define R500_ALPHA_OP_MIN				2
+#   define R500_ALPHA_OP_MAX				3
+/* #define R500_ALPHA_OP_RESERVED			4 */
+#   define R500_ALPHA_OP_CND				5
+#   define R500_ALPHA_OP_CMP				6
+#   define R500_ALPHA_OP_FRC				7
+#   define R500_ALPHA_OP_EX2				8
+#   define R500_ALPHA_OP_LN2				9
+#   define R500_ALPHA_OP_RCP				10
+#   define R500_ALPHA_OP_RSQ				11
+#   define R500_ALPHA_OP_SIN				12
+#   define R500_ALPHA_OP_COS				13
+#   define R500_ALPHA_OP_MDH				14
+#   define R500_ALPHA_OP_MDV				15
+#   define R500_ALPHA_ADDRD(x)				(x << 4)
+#   define R500_ALPHA_ADDRD_REL				(1 << 11)
+#   define R500_ALPHA_SEL_A_SRC0			(0 << 12)
+#   define R500_ALPHA_SEL_A_SRC1			(1 << 12)
+#   define R500_ALPHA_SEL_A_SRC2			(2 << 12)
+#   define R500_ALPHA_SEL_A_SRCP			(3 << 12)
+#   define R500_ALPHA_SWIZ_A_R				(0 << 14)
+#   define R500_ALPHA_SWIZ_A_G				(1 << 14)
+#   define R500_ALPHA_SWIZ_A_B				(2 << 14)
+#   define R500_ALPHA_SWIZ_A_A				(3 << 14)
+#   define R500_ALPHA_SWIZ_A_0				(4 << 14)
+#   define R500_ALPHA_SWIZ_A_HALF			(5 << 14)
+#   define R500_ALPHA_SWIZ_A_1				(6 << 14)
+/* #define R500_ALPHA_SWIZ_A_UNUSED			(7 << 14) */
+#   define R500_ALPHA_MOD_A_NOP				(0 << 17)
+#   define R500_ALPHA_MOD_A_NEG				(1 << 17)
+#   define R500_ALPHA_MOD_A_ABS				(2 << 17)
+#   define R500_ALPHA_MOD_A_NAB				(3 << 17)
+#   define R500_ALPHA_SEL_B_SRC0			(0 << 19)
+#   define R500_ALPHA_SEL_B_SRC1			(1 << 19)
+#   define R500_ALPHA_SEL_B_SRC2			(2 << 19)
+#   define R500_ALPHA_SEL_B_SRCP			(3 << 19)
+#   define R500_ALPHA_SWIZ_B_R				(0 << 21)
+#   define R500_ALPHA_SWIZ_B_G				(1 << 21)
+#   define R500_ALPHA_SWIZ_B_B				(2 << 21)
+#   define R500_ALPHA_SWIZ_B_A				(3 << 21)
+#   define R500_ALPHA_SWIZ_B_0				(4 << 21)
+#   define R500_ALPHA_SWIZ_B_HALF			(5 << 21)
+#   define R500_ALPHA_SWIZ_B_1				(6 << 21)
+/* #define R500_ALPHA_SWIZ_B_UNUSED			(7 << 21) */
+#   define R500_ALPHA_MOD_B_NOP				(0 << 24)
+#   define R500_ALPHA_MOD_B_NEG				(1 << 24)
+#   define R500_ALPHA_MOD_B_ABS				(2 << 24)
+#   define R500_ALPHA_MOD_B_NAB				(3 << 24)
+#   define R500_ALPHA_OMOD_IDENTITY			(0 << 26)
+#   define R500_ALPHA_OMOD_MUL_2			(1 << 26)
+#   define R500_ALPHA_OMOD_MUL_4			(2 << 26)
+#   define R500_ALPHA_OMOD_MUL_8			(3 << 26)
+#   define R500_ALPHA_OMOD_DIV_2			(4 << 26)
+#   define R500_ALPHA_OMOD_DIV_4			(5 << 26)
+#   define R500_ALPHA_OMOD_DIV_8			(6 << 26)
+#   define R500_ALPHA_OMOD_DISABLE			(7 << 26)
+#   define R500_ALPHA_TARGET(x)				(x << 29)
+#   define R500_ALPHA_W_OMASK				(1 << 31)
+#define R500_US_ALU_ALPHA_ADDR_0			0x9800
+#   define R500_ALPHA_ADDR0(x)				(x << 0)
+#   define R500_ALPHA_ADDR0_CONST			(1 << 8)
+#   define R500_ALPHA_ADDR0_REL				(1 << 9)
+#   define R500_ALPHA_ADDR1(x)				(x << 10)
+#   define R500_ALPHA_ADDR1_CONST			(1 << 18)
+#   define R500_ALPHA_ADDR1_REL				(1 << 19)
+#   define R500_ALPHA_ADDR2(x)				(x << 20)
+#   define R500_ALPHA_ADDR2_CONST			(1 << 28)
+#   define R500_ALPHA_ADDR2_REL				(1 << 29)
+#   define R500_ALPHA_SRCP_OP_1_MINUS_2A0		(0 << 30)
+#   define R500_ALPHA_SRCP_OP_A1_MINUS_A0		(1 << 30)
+#   define R500_ALPHA_SRCP_OP_A1_PLUS_A0		(2 << 30)
+#   define R500_ALPHA_SRCP_OP_1_MINUS_A0		(3 << 30)
+#define R500_US_ALU_RGBA_INST_0				0xb000
+#   define R500_ALU_RGBA_OP_MAD				(0 << 0)
+#   define R500_ALU_RGBA_OP_DP3				(1 << 0)
+#   define R500_ALU_RGBA_OP_DP4				(2 << 0)
+#   define R500_ALU_RGBA_OP_D2A				(3 << 0)
+#   define R500_ALU_RGBA_OP_MIN				(4 << 0)
+#   define R500_ALU_RGBA_OP_MAX				(5 << 0)
+/* #define R500_ALU_RGBA_OP_RESERVED			(6 << 0) */
+#   define R500_ALU_RGBA_OP_CND				(7 << 0)
+#   define R500_ALU_RGBA_OP_CMP				(8 << 0)
+#   define R500_ALU_RGBA_OP_FRC				(9 << 0)
+#   define R500_ALU_RGBA_OP_SOP				(10 << 0)
+#   define R500_ALU_RGBA_OP_MDH				(11 << 0)
+#   define R500_ALU_RGBA_OP_MDV				(12 << 0)
+#   define R500_ALU_RGBA_ADDRD(x)			(x << 4)
+#   define R500_ALU_RGBA_ADDRD_REL			(1 << 11)
+#   define R500_ALU_RGBA_SEL_C_SRC0			(0 << 12)
+#   define R500_ALU_RGBA_SEL_C_SRC1			(1 << 12)
+#   define R500_ALU_RGBA_SEL_C_SRC2			(2 << 12)
+#   define R500_ALU_RGBA_SEL_C_SRCP			(3 << 12)
+#   define R500_ALU_RGBA_R_SWIZ_R			(0 << 14)
+#   define R500_ALU_RGBA_R_SWIZ_G			(1 << 14)
+#   define R500_ALU_RGBA_R_SWIZ_B			(2 << 14)
+#   define R500_ALU_RGBA_R_SWIZ_A			(3 << 14)
+#   define R500_ALU_RGBA_R_SWIZ_0			(4 << 14)
+#   define R500_ALU_RGBA_R_SWIZ_HALF			(5 << 14)
+#   define R500_ALU_RGBA_R_SWIZ_1			(6 << 14)
+/* #define R500_ALU_RGBA_R_SWIZ_UNUSED			(7 << 14) */
+#   define R500_ALU_RGBA_G_SWIZ_R			(0 << 17)
+#   define R500_ALU_RGBA_G_SWIZ_G			(1 << 17)
+#   define R500_ALU_RGBA_G_SWIZ_B			(2 << 17)
+#   define R500_ALU_RGBA_G_SWIZ_A			(3 << 17)
+#   define R500_ALU_RGBA_G_SWIZ_0			(4 << 17)
+#   define R500_ALU_RGBA_G_SWIZ_HALF			(5 << 17)
+#   define R500_ALU_RGBA_G_SWIZ_1			(6 << 17)
+/* #define R500_ALU_RGBA_G_SWIZ_UNUSED			(7 << 17) */
+#   define R500_ALU_RGBA_B_SWIZ_R			(0 << 20)
+#   define R500_ALU_RGBA_B_SWIZ_G			(1 << 20)
+#   define R500_ALU_RGBA_B_SWIZ_B			(2 << 20)
+#   define R500_ALU_RGBA_B_SWIZ_A			(3 << 20)
+#   define R500_ALU_RGBA_B_SWIZ_0			(4 << 20)
+#   define R500_ALU_RGBA_B_SWIZ_HALF			(5 << 20)
+#   define R500_ALU_RGBA_B_SWIZ_1			(6 << 20)
+/* #define R500_ALU_RGBA_B_SWIZ_UNUSED			(7 << 20) */
+#   define R500_ALU_RGBA_MOD_C_NOP			(0 << 23)
+#   define R500_ALU_RGBA_MOD_C_NEG			(1 << 23)
+#   define R500_ALU_RGBA_MOD_C_ABS			(2 << 23)
+#   define R500_ALU_RGBA_MOD_C_NAB			(3 << 23)
+#   define R500_ALU_RGBA_ALPHA_SEL_C_SRC0		(0 << 25)
+#   define R500_ALU_RGBA_ALPHA_SEL_C_SRC1		(1 << 25)
+#   define R500_ALU_RGBA_ALPHA_SEL_C_SRC2		(2 << 25)
+#   define R500_ALU_RGBA_ALPHA_SEL_C_SRCP		(3 << 25)
+#   define R500_ALU_RGBA_A_SWIZ_R			(0 << 27)
+#   define R500_ALU_RGBA_A_SWIZ_G			(1 << 27)
+#   define R500_ALU_RGBA_A_SWIZ_B			(2 << 27)
+#   define R500_ALU_RGBA_A_SWIZ_A			(3 << 27)
+#   define R500_ALU_RGBA_A_SWIZ_0			(4 << 27)
+#   define R500_ALU_RGBA_A_SWIZ_HALF			(5 << 27)
+#   define R500_ALU_RGBA_A_SWIZ_1			(6 << 27)
+/* #define R500_ALU_RGBA_A_SWIZ_UNUSED			(7 << 27) */
+#   define R500_ALU_RGBA_ALPHA_MOD_C_NOP		(0 << 30)
+#   define R500_ALU_RGBA_ALPHA_MOD_C_NEG		(1 << 30)
+#   define R500_ALU_RGBA_ALPHA_MOD_C_ABS		(2 << 30)
+#   define R500_ALU_RGBA_ALPHA_MOD_C_NAB		(3 << 30)
+#define R500_US_ALU_RGB_INST_0				0xa000
+#   define R500_ALU_RGB_SEL_A_SRC0			(0 << 0)
+#   define R500_ALU_RGB_SEL_A_SRC1			(1 << 0)
+#   define R500_ALU_RGB_SEL_A_SRC2			(2 << 0)
+#   define R500_ALU_RGB_SEL_A_SRCP			(3 << 0)
+#   define R500_ALU_RGB_R_SWIZ_A_R			(0 << 2)
+#   define R500_ALU_RGB_R_SWIZ_A_G			(1 << 2)
+#   define R500_ALU_RGB_R_SWIZ_A_B			(2 << 2)
+#   define R500_ALU_RGB_R_SWIZ_A_A			(3 << 2)
+#   define R500_ALU_RGB_R_SWIZ_A_0			(4 << 2)
+#   define R500_ALU_RGB_R_SWIZ_A_HALF			(5 << 2)
+#   define R500_ALU_RGB_R_SWIZ_A_1			(6 << 2)
+/* #define R500_ALU_RGB_R_SWIZ_A_UNUSED			(7 << 2) */
+#   define R500_ALU_RGB_G_SWIZ_A_R			(0 << 5)
+#   define R500_ALU_RGB_G_SWIZ_A_G			(1 << 5)
+#   define R500_ALU_RGB_G_SWIZ_A_B			(2 << 5)
+#   define R500_ALU_RGB_G_SWIZ_A_A			(3 << 5)
+#   define R500_ALU_RGB_G_SWIZ_A_0			(4 << 5)
+#   define R500_ALU_RGB_G_SWIZ_A_HALF			(5 << 5)
+#   define R500_ALU_RGB_G_SWIZ_A_1			(6 << 5)
+/* #define R500_ALU_RGB_G_SWIZ_A_UNUSED			(7 << 5) */
+#   define R500_ALU_RGB_B_SWIZ_A_R			(0 << 8)
+#   define R500_ALU_RGB_B_SWIZ_A_G			(1 << 8)
+#   define R500_ALU_RGB_B_SWIZ_A_B			(2 << 8)
+#   define R500_ALU_RGB_B_SWIZ_A_A			(3 << 8)
+#   define R500_ALU_RGB_B_SWIZ_A_0			(4 << 8)
+#   define R500_ALU_RGB_B_SWIZ_A_HALF			(5 << 8)
+#   define R500_ALU_RGB_B_SWIZ_A_1			(6 << 8)
+/* #define R500_ALU_RGB_B_SWIZ_A_UNUSED			(7 << 8) */
+#   define R500_ALU_RGB_MOD_A_NOP			(0 << 11)
+#   define R500_ALU_RGB_MOD_A_NEG			(1 << 11)
+#   define R500_ALU_RGB_MOD_A_ABS			(2 << 11)
+#   define R500_ALU_RGB_MOD_A_NAB			(3 << 11)
+#   define R500_ALU_RGB_SEL_B_SRC0			(0 << 13)
+#   define R500_ALU_RGB_SEL_B_SRC1			(1 << 13)
+#   define R500_ALU_RGB_SEL_B_SRC2			(2 << 13)
+#   define R500_ALU_RGB_SEL_B_SRCP			(3 << 13)
+#   define R500_ALU_RGB_R_SWIZ_B_R			(0 << 15)
+#   define R500_ALU_RGB_R_SWIZ_B_G			(1 << 15)
+#   define R500_ALU_RGB_R_SWIZ_B_B			(2 << 15)
+#   define R500_ALU_RGB_R_SWIZ_B_A			(3 << 15)
+#   define R500_ALU_RGB_R_SWIZ_B_0			(4 << 15)
+#   define R500_ALU_RGB_R_SWIZ_B_HALF			(5 << 15)
+#   define R500_ALU_RGB_R_SWIZ_B_1			(6 << 15)
+/* #define R500_ALU_RGB_R_SWIZ_B_UNUSED			(7 << 15) */
+#   define R500_ALU_RGB_G_SWIZ_B_R			(0 << 18)
+#   define R500_ALU_RGB_G_SWIZ_B_G			(1 << 18)
+#   define R500_ALU_RGB_G_SWIZ_B_B			(2 << 18)
+#   define R500_ALU_RGB_G_SWIZ_B_A			(3 << 18)
+#   define R500_ALU_RGB_G_SWIZ_B_0			(4 << 18)
+#   define R500_ALU_RGB_G_SWIZ_B_HALF			(5 << 18)
+#   define R500_ALU_RGB_G_SWIZ_B_1			(6 << 18)
+/* #define R500_ALU_RGB_G_SWIZ_B_UNUSED			(7 << 18) */
+#   define R500_ALU_RGB_B_SWIZ_B_R			(0 << 21)
+#   define R500_ALU_RGB_B_SWIZ_B_G			(1 << 21)
+#   define R500_ALU_RGB_B_SWIZ_B_B			(2 << 21)
+#   define R500_ALU_RGB_B_SWIZ_B_A			(3 << 21)
+#   define R500_ALU_RGB_B_SWIZ_B_0			(4 << 21)
+#   define R500_ALU_RGB_B_SWIZ_B_HALF			(5 << 21)
+#   define R500_ALU_RGB_B_SWIZ_B_1			(6 << 21)
+/* #define R500_ALU_RGB_B_SWIZ_B_UNUSED			(7 << 21) */
+#   define R500_ALU_RGB_MOD_B_NOP			(0 << 24)
+#   define R500_ALU_RGB_MOD_B_NEG			(1 << 24)
+#   define R500_ALU_RGB_MOD_B_ABS			(2 << 24)
+#   define R500_ALU_RGB_MOD_B_NAB			(3 << 24)
+#   define R500_ALU_RGB_OMOD_IDENTITY			(0 << 26)
+#   define R500_ALU_RGB_OMOD_MUL_2			(1 << 26)
+#   define R500_ALU_RGB_OMOD_MUL_4			(2 << 26)
+#   define R500_ALU_RGB_OMOD_MUL_8			(3 << 26)
+#   define R500_ALU_RGB_OMOD_DIV_2			(4 << 26)
+#   define R500_ALU_RGB_OMOD_DIV_4			(5 << 26)
+#   define R500_ALU_RGB_OMOD_DIV_8			(6 << 26)
+#   define R500_ALU_RGB_OMOD_DISABLE			(7 << 26)
+#   define R500_ALU_RGB_TARGET(x)			(x << 29)
+#   define R500_ALU_RGB_WMASK				(1 << 31)
+#define R500_US_ALU_RGB_ADDR_0				0x9000
+#   define R500_RGB_ADDR0(x)				(x << 0)
+#   define R500_RGB_ADDR0_CONST				(1 << 8)
+#   define R500_RGB_ADDR0_REL				(1 << 9)
+#   define R500_RGB_ADDR1(x)				(x << 10)
+#   define R500_RGB_ADDR1_CONST				(1 << 18)
+#   define R500_RGB_ADDR1_REL				(1 << 19)
+#   define R500_RGB_ADDR2(x)				(x << 20)
+#   define R500_RGB_ADDR2_CONST				(1 << 28)
+#   define R500_RGB_ADDR2_REL				(1 << 29)
+#   define R500_RGB_SRCP_OP_1_MINUS_2RGB0		(0 << 30)
+#   define R500_RGB_SRCP_OP_RGB1_MINUS_RGB0		(1 << 30)
+#   define R500_RGB_SRCP_OP_RGB1_PLUS_RGB0		(2 << 30)
+#   define R500_RGB_SRCP_OP_1_MINUS_RGB0		(3 << 30)
+#define R500_US_CMN_INST_0				0xb800
+#   define R500_INST_TYPE_ALU				(0 << 0)
+#   define R500_INST_TYPE_OUT				(1 << 0)
+#   define R500_INST_TYPE_FC				(2 << 0)
+#   define R500_INST_TYPE_TEX				(3 << 0)
+#   define R500_INST_TEX_SEM_WAIT			(1 << 2)
+#   define R500_INST_RGB_PRED_SEL_NONE			(0 << 3)
+#   define R500_INST_RGB_PRED_SEL_RGBA			(1 << 3)
+#   define R500_INST_RGB_PRED_SEL_RRRR			(2 << 3)
+#   define R500_INST_RGB_PRED_SEL_GGGG			(3 << 3)
+#   define R500_INST_RGB_PRED_SEL_BBBB			(4 << 3)
+#   define R500_INST_RGB_PRED_SEL_AAAA			(5 << 3)
+#   define R500_INST_RGB_PRED_INV			(1 << 6)
+#   define R500_INST_WRITE_INACTIVE			(1 << 7)
+#   define R500_INST_LAST				(1 << 8)
+#   define R500_INST_NOP				(1 << 9)
+#   define R500_INST_ALU_WAIT				(1 << 10)
+#   define R500_INST_RGB_WMASK_R			(1 << 11)
+#   define R500_INST_RGB_WMASK_G			(1 << 12)
+#   define R500_INST_RGB_WMASK_B			(1 << 13)
+#   define R500_INST_ALPHA_WMASK			(1 << 14)
+#   define R500_INST_RGB_OMASK_R			(1 << 15)
+#   define R500_INST_RGB_OMASK_G			(1 << 16)
+#   define R500_INST_RGB_OMASK_B			(1 << 17)
+#   define R500_INST_ALPHA_OMASK			(1 << 18)
+#   define R500_INST_RGB_CLAMP				(1 << 19)
+#   define R500_INST_ALPHA_CLAMP			(1 << 20)
+#   define R500_INST_ALU_RESULT_SEL			(1 << 21)
+#   define R500_INST_ALPHA_PRED_INV			(1 << 22)
+#   define R500_INST_ALU_RESULT_OP_EQ			(0 << 23)
+#   define R500_INST_ALU_RESULT_OP_LT			(1 << 23)
+#   define R500_INST_ALU_RESULT_OP_GE			(2 << 23)
+#   define R500_INST_ALU_RESULT_OP_NE			(3 << 23)
+#   define R500_INST_ALPHA_PRED_SEL_NONE		(0 << 25)
+#   define R500_INST_ALPHA_PRED_SEL_RGBA		(1 << 25)
+#   define R500_INST_ALPHA_PRED_SEL_RRRR		(2 << 25)
+#   define R500_INST_ALPHA_PRED_SEL_GGGG		(3 << 25)
+#   define R500_INST_ALPHA_PRED_SEL_BBBB		(4 << 25)
+#   define R500_INST_ALPHA_PRED_SEL_AAAA		(5 << 25)
+/* XXX next four are kind of guessed */
+#   define R500_INST_STAT_WE_R				(1 << 28)
+#   define R500_INST_STAT_WE_G				(1 << 29)
+#   define R500_INST_STAT_WE_B				(1 << 30)
+#   define R500_INST_STAT_WE_A				(1 << 31)
+
+/* note that these are 8 bit lengths, despite the offsets, at least for R500 */
+#define R500_US_CODE_ADDR				0x4630
+#   define R500_US_CODE_START_ADDR(x)			(x << 0)
+#   define R500_US_CODE_END_ADDR(x)			(x << 16)
+#define R500_US_CODE_OFFSET				0x4638
+#   define R500_US_CODE_OFFSET_ADDR(x)			(x << 0)
+#define R500_US_CODE_RANGE				0x4634
+#   define R500_US_CODE_RANGE_ADDR(x)			(x << 0)
+#   define R500_US_CODE_RANGE_SIZE(x)			(x << 16)
+#define R500_US_CONFIG					0x4600
+#   define R500_ZERO_TIMES_ANYTHING_EQUALS_ZERO		(1 << 1)
+#define R500_US_FC_ADDR_0				0xa000
+#   define R500_FC_BOOL_ADDR(x)				(x << 0)
+#   define R500_FC_INT_ADDR(x)				(x << 8)
+#   define R500_FC_JUMP_ADDR(x)				(x << 16)
+#   define R500_FC_JUMP_GLOBAL				(1 << 31)
+#define R500_US_FC_BOOL_CONST				0x4620
+#   define R500_FC_KBOOL(x)				(x)
+#define R500_US_FC_CTRL					0x4624
+#   define R500_FC_TEST_EN				(1 << 30)
+#   define R500_FC_FULL_FC_EN				(1 << 31)
+#define R500_US_FC_INST_0				0x9800
+#   define R500_FC_OP_JUMP				(0 << 0)
+#   define R500_FC_OP_LOOP				(1 << 0)
+#   define R500_FC_OP_ENDLOOP				(2 << 0)
+#   define R500_FC_OP_REP				(3 << 0)
+#   define R500_FC_OP_ENDREP				(4 << 0)
+#   define R500_FC_OP_BREAKLOOP				(5 << 0)
+#   define R500_FC_OP_BREAKREP				(6 << 0)
+#   define R500_FC_OP_CONTINUE				(7 << 0)
+#   define R500_FC_B_ELSE				(1 << 4)
+#   define R500_FC_JUMP_ANY				(1 << 5)
+#   define R500_FC_A_OP_NONE				(0 << 6)
+#   define R500_FC_A_OP_POP				(1 << 6)
+#   define R500_FC_A_OP_PUSH				(2 << 6)
+#   define R500_FC_JUMP_FUNC(x)				(x << 8)
+#   define R500_FC_B_POP_CNT(x)				(x << 16)
+#   define R500_FC_B_OP0_NONE				(0 << 24)
+#   define R500_FC_B_OP0_DECR				(1 << 24)
+#   define R500_FC_B_OP0_INCR				(2 << 24)
+#   define R500_FC_B_OP1_DECR				(0 << 26)
+#   define R500_FC_B_OP1_NONE				(1 << 26)
+#   define R500_FC_B_OP1_INCR				(2 << 26)
+#   define R500_FC_IGNORE_UNCOVERED			(1 << 28)
+#define R500_US_FC_INT_CONST_0				0x4c00
+#   define R500_FC_INT_CONST_KR(x)			(x << 0)
+#   define R500_FC_INT_CONST_KG(x)			(x << 8)
+#   define R500_FC_INT_CONST_KB(x)			(x << 16)
+/* _0 through _15 */
+#define R500_US_FORMAT0_0				0x4640
+#   define R500_FORMAT_TXWIDTH(x)			(x << 0)
+#   define R500_FORMAT_TXHEIGHT(x)			(x << 11)
+#   define R500_FORMAT_TXDEPTH(x)			(x << 22)
+/* _0 through _3 */
+#define R500_US_OUT_FMT_0				0x46a4
+#   define R500_OUT_FMT_C4_8				(0 << 0)
+#   define R500_OUT_FMT_C4_10				(1 << 0)
+#   define R500_OUT_FMT_C4_10_GAMMA			(2 << 0)
+#   define R500_OUT_FMT_C_16				(3 << 0)
+#   define R500_OUT_FMT_C2_16				(4 << 0)
+#   define R500_OUT_FMT_C4_16				(5 << 0)
+#   define R500_OUT_FMT_C_16_MPEG			(6 << 0)
+#   define R500_OUT_FMT_C2_16_MPEG			(7 << 0)
+#   define R500_OUT_FMT_C2_4				(8 << 0)
+#   define R500_OUT_FMT_C_3_3_2				(9 << 0)
+#   define R500_OUT_FMT_C_6_5_6				(10 << 0)
+#   define R500_OUT_FMT_C_11_11_10			(11 << 0)
+#   define R500_OUT_FMT_C_10_11_11			(12 << 0)
+#   define R500_OUT_FMT_C_2_10_10_10			(13 << 0)
+/* #define R500_OUT_FMT_RESERVED			(14 << 0) */
+#   define R500_OUT_FMT_UNUSED				(15 << 0)
+#   define R500_OUT_FMT_C_16_FP				(16 << 0)
+#   define R500_OUT_FMT_C2_16_FP			(17 << 0)
+#   define R500_OUT_FMT_C4_16_FP			(18 << 0)
+#   define R500_OUT_FMT_C_32_FP				(19 << 0)
+#   define R500_OUT_FMT_C2_32_FP			(20 << 0)
+#   define R500_OUT_FMT_C4_32_FP			(21 << 0)
+#   define R500_C0_SEL_A				(0 << 8)
+#   define R500_C0_SEL_R				(1 << 8)
+#   define R500_C0_SEL_G				(2 << 8)
+#   define R500_C0_SEL_B				(3 << 8)
+#   define R500_C1_SEL_A				(0 << 10)
+#   define R500_C1_SEL_R				(1 << 10)
+#   define R500_C1_SEL_G				(2 << 10)
+#   define R500_C1_SEL_B				(3 << 10)
+#   define R500_C2_SEL_A				(0 << 12)
+#   define R500_C2_SEL_R				(1 << 12)
+#   define R500_C2_SEL_G				(2 << 12)
+#   define R500_C2_SEL_B				(3 << 12)
+#   define R500_C3_SEL_A				(0 << 14)
+#   define R500_C3_SEL_R				(1 << 14)
+#   define R500_C3_SEL_G				(2 << 14)
+#   define R500_C3_SEL_B				(3 << 14)
+#   define R500_OUT_SIGN(x)				(x << 16)
+#   define R500_ROUND_ADJ				(1 << 20)
+#define R500_US_PIXSIZE					0x4604
+#   define R500_PIX_SIZE(x)				(x)
+#define R500_US_TEX_ADDR_0				0x9800
+#   define R500_TEX_SRC_ADDR(x)				(x << 0)
+#   define R500_TEX_SRC_ADDR_REL			(1 << 7)
+#   define R500_TEX_SRC_S_SWIZ_R			(0 << 8)
+#   define R500_TEX_SRC_S_SWIZ_G			(1 << 8)
+#   define R500_TEX_SRC_S_SWIZ_B			(2 << 8)
+#   define R500_TEX_SRC_S_SWIZ_A			(3 << 8)
+#   define R500_TEX_SRC_T_SWIZ_R			(0 << 10)
+#   define R500_TEX_SRC_T_SWIZ_G			(1 << 10)
+#   define R500_TEX_SRC_T_SWIZ_B			(2 << 10)
+#   define R500_TEX_SRC_T_SWIZ_A			(3 << 10)
+#   define R500_TEX_SRC_R_SWIZ_R			(0 << 12)
+#   define R500_TEX_SRC_R_SWIZ_G			(1 << 12)
+#   define R500_TEX_SRC_R_SWIZ_B			(2 << 12)
+#   define R500_TEX_SRC_R_SWIZ_A			(3 << 12)
+#   define R500_TEX_SRC_Q_SWIZ_R			(0 << 14)
+#   define R500_TEX_SRC_Q_SWIZ_G			(1 << 14)
+#   define R500_TEX_SRC_Q_SWIZ_B			(2 << 14)
+#   define R500_TEX_SRC_Q_SWIZ_A			(3 << 14)
+#   define R500_TEX_DST_ADDR(x)				(x << 16)
+#   define R500_TEX_DST_ADDR_REL			(1 << 23)
+#   define R500_TEX_DST_R_SWIZ_R			(0 << 24)
+#   define R500_TEX_DST_R_SWIZ_G			(1 << 24)
+#   define R500_TEX_DST_R_SWIZ_B			(2 << 24)
+#   define R500_TEX_DST_R_SWIZ_A			(3 << 24)
+#   define R500_TEX_DST_G_SWIZ_R			(0 << 26)
+#   define R500_TEX_DST_G_SWIZ_G			(1 << 26)
+#   define R500_TEX_DST_G_SWIZ_B			(2 << 26)
+#   define R500_TEX_DST_G_SWIZ_A			(3 << 26)
+#   define R500_TEX_DST_B_SWIZ_R			(0 << 28)
+#   define R500_TEX_DST_B_SWIZ_G			(1 << 28)
+#   define R500_TEX_DST_B_SWIZ_B			(2 << 28)
+#   define R500_TEX_DST_B_SWIZ_A			(3 << 28)
+#   define R500_TEX_DST_A_SWIZ_R			(0 << 30)
+#   define R500_TEX_DST_A_SWIZ_G			(1 << 30)
+#   define R500_TEX_DST_A_SWIZ_B			(2 << 30)
+#   define R500_TEX_DST_A_SWIZ_A			(3 << 30)
+#define R500_US_TEX_ADDR_DXDY_0				0xa000
+#   define R500_DX_ADDR(x)				(x << 0)
+#   define R500_DX_ADDR_REL				(1 << 7)
+#   define R500_DX_S_SWIZ_R				(0 << 8)
+#   define R500_DX_S_SWIZ_G				(1 << 8)
+#   define R500_DX_S_SWIZ_B				(2 << 8)
+#   define R500_DX_S_SWIZ_A				(3 << 8)
+#   define R500_DX_T_SWIZ_R				(0 << 10)
+#   define R500_DX_T_SWIZ_G				(1 << 10)
+#   define R500_DX_T_SWIZ_B				(2 << 10)
+#   define R500_DX_T_SWIZ_A				(3 << 10)
+#   define R500_DX_R_SWIZ_R				(0 << 12)
+#   define R500_DX_R_SWIZ_G				(1 << 12)
+#   define R500_DX_R_SWIZ_B				(2 << 12)
+#   define R500_DX_R_SWIZ_A				(3 << 12)
+#   define R500_DX_Q_SWIZ_R				(0 << 14)
+#   define R500_DX_Q_SWIZ_G				(1 << 14)
+#   define R500_DX_Q_SWIZ_B				(2 << 14)
+#   define R500_DX_Q_SWIZ_A				(3 << 14)
+#   define R500_DY_ADDR(x)				(x << 16)
+#   define R500_DY_ADDR_REL				(1 << 17)
+#   define R500_DY_S_SWIZ_R				(0 << 24)
+#   define R500_DY_S_SWIZ_G				(1 << 24)
+#   define R500_DY_S_SWIZ_B				(2 << 24)
+#   define R500_DY_S_SWIZ_A				(3 << 24)
+#   define R500_DY_T_SWIZ_R				(0 << 26)
+#   define R500_DY_T_SWIZ_G				(1 << 26)
+#   define R500_DY_T_SWIZ_B				(2 << 26)
+#   define R500_DY_T_SWIZ_A				(3 << 26)
+#   define R500_DY_R_SWIZ_R				(0 << 28)
+#   define R500_DY_R_SWIZ_G				(1 << 28)
+#   define R500_DY_R_SWIZ_B				(2 << 28)
+#   define R500_DY_R_SWIZ_A				(3 << 28)
+#   define R500_DY_Q_SWIZ_R				(0 << 30)
+#   define R500_DY_Q_SWIZ_G				(1 << 30)
+#   define R500_DY_Q_SWIZ_B				(2 << 30)
+#   define R500_DY_Q_SWIZ_A				(3 << 30)
+#define R500_US_TEX_INST_0				0x9000
+#   define R500_TEX_ID(x)				(x << 16)
+#   define R500_TEX_INST_NOP				(0 << 22)
+#   define R500_TEX_INST_LD				(1 << 22)
+#   define R500_TEX_INST_TEXKILL			(2 << 22)
+#   define R500_TEX_INST_PROJ				(3 << 22)
+#   define R500_TEX_INST_LODBIAS			(4 << 22)
+#   define R500_TEX_INST_LOD				(5 << 22)
+#   define R500_TEX_INST_DXDY				(6 << 22)
+#   define R500_TEX_SEM_ACQUIRE				(1 << 25)
+#   define R500_TEX_IGNORE_UNCOVERED			(1 << 26)
+#   define R500_TEX_UNSCALED				(1 << 27)
+#define R500_US_W_FMT					0x46b4
+#   define R500_W_FMT_W0				(0 << 0)
+#   define R500_W_FMT_W24				(1 << 0)
+#   define R500_W_FMT_W24FP				(2 << 0)
+#   define R500_W_SRC_US				(0 << 2)
+#   define R500_W_SRC_RAS				(1 << 2)
+
+
 /* Draw a primitive from vertex data in arrays loaded via 3D_LOAD_VBPNTR.
  * Two parameter dwords:
  * 0. VAP_VTX_FMT: The first parameter is not written to hardware
diff --git a/src/mesa/drivers/dri/r300/r300_state.c b/src/mesa/drivers/dri/r300/r300_state.c
index e11b5afc30c..d2fd04a550f 100644
--- a/src/mesa/drivers/dri/r300/r300_state.c
+++ b/src/mesa/drivers/dri/r300/r300_state.c
@@ -1603,6 +1603,128 @@ static void r300SetupRSUnit(GLcontext * ctx)
 		WARN_ONCE("Don't know how to satisfy InputsRead=0x%08x\n", InputsRead);
 }
 
+static void r500SetupRSUnit(GLcontext * ctx)
+{
+	r300ContextPtr r300 = R300_CONTEXT(ctx);
+	/* I'm still unsure if these are needed */
+	GLuint interp_magic[8] = {
+		0x00,
+		R300_RS_COL_PTR(1),
+		R300_RS_COL_PTR(2),
+		R300_RS_COL_PTR(3),
+		0x00,
+		0x00,
+		0x00,
+		0x00
+	};
+	union r300_outputs_written OutputsWritten;
+	GLuint InputsRead;
+	int fp_reg, high_rr;
+	int in_texcoords, col_interp_nr;
+	int i;
+
+	if (hw_tcl_on)
+		OutputsWritten.vp_outputs = CURRENT_VERTEX_SHADER(ctx)->key.OutputsWritten;
+	else
+		RENDERINPUTS_COPY(OutputsWritten.index_bitset, r300->state.render_inputs_bitset);
+
+	if (ctx->FragmentProgram._Current)
+		InputsRead = ctx->FragmentProgram._Current->Base.InputsRead;
+	else {
+		fprintf(stderr, "No ctx->FragmentProgram._Current!!\n");
+		return;		/* This should only ever happen once.. */
+	}
+
+	R300_STATECHANGE(r300, ri);
+	R300_STATECHANGE(r300, rc);
+	R300_STATECHANGE(r300, rr);
+
+	fp_reg = in_texcoords = col_interp_nr = high_rr = 0;
+
+	r300->hw.rr.cmd[R300_RR_ROUTE_1] = 0;
+
+	if (InputsRead & FRAG_BIT_WPOS) {
+		for (i = 0; i < ctx->Const.MaxTextureUnits; i++)
+			if (!(InputsRead & (FRAG_BIT_TEX0 << i)))
+				break;
+
+		if (i == ctx->Const.MaxTextureUnits) {
+			fprintf(stderr, "\tno free texcoord found...\n");
+			_mesa_exit(-1);
+		}
+
+		InputsRead |= (FRAG_BIT_TEX0 << i);
+		InputsRead &= ~FRAG_BIT_WPOS;
+	}
+
+	for (i = 0; i < ctx->Const.MaxTextureUnits; i++) {
+		r300->hw.ri.cmd[R300_RI_INTERP_0 + i] = 0 | R300_RS_SEL_T(1) | R300_RS_SEL_R(2) | R300_RS_SEL_Q(3) | (in_texcoords << R300_RS_INTERP_SRC_SHIFT)
+		    | interp_magic[i];
+
+		r300->hw.rr.cmd[R300_RR_ROUTE_0 + fp_reg] = 0;
+		if (InputsRead & (FRAG_BIT_TEX0 << i)) {
+			//assert(r300->state.texture.tc_count != 0);
+			r300->hw.rr.cmd[R300_RR_ROUTE_0 + fp_reg] |= R300_RS_ROUTE_ENABLE | i	/* source INTERP */
+			    | (fp_reg << R300_RS_ROUTE_DEST_SHIFT);
+			high_rr = fp_reg;
+
+			/* Passing invalid data here can lock the GPU. */
+			if (R300_OUTPUTS_WRITTEN_TEST(OutputsWritten, VERT_RESULT_TEX0 + i, _TNL_ATTRIB_TEX(i))) {
+				InputsRead &= ~(FRAG_BIT_TEX0 << i);
+				fp_reg++;
+			} else {
+				WARN_ONCE("fragprog wants coords for tex%d, vp doesn't provide them!\n", i);
+			}
+		}
+		/* Need to count all coords enabled at vof */
+		if (R300_OUTPUTS_WRITTEN_TEST(OutputsWritten, VERT_RESULT_TEX0 + i, _TNL_ATTRIB_TEX(i))) {
+			in_texcoords++;
+		}
+	}
+
+	if (InputsRead & FRAG_BIT_COL0) {
+		if (R300_OUTPUTS_WRITTEN_TEST(OutputsWritten, VERT_RESULT_COL0, _TNL_ATTRIB_COLOR0)) {
+			r300->hw.rr.cmd[R300_RR_ROUTE_0] |= 0 | R300_RS_ROUTE_0_COLOR | (fp_reg++ << R300_RS_ROUTE_0_COLOR_DEST_SHIFT);
+			InputsRead &= ~FRAG_BIT_COL0;
+			col_interp_nr++;
+		} else {
+			WARN_ONCE("fragprog wants col0, vp doesn't provide it\n");
+		}
+	}
+
+	if (InputsRead & FRAG_BIT_COL1) {
+		if (R300_OUTPUTS_WRITTEN_TEST(OutputsWritten, VERT_RESULT_COL1, _TNL_ATTRIB_COLOR1)) {
+			r300->hw.rr.cmd[R300_RR_ROUTE_1] |= R300_RS_ROUTE_1_UNKNOWN11 | R300_RS_ROUTE_1_COLOR1 | (fp_reg++ << R300_RS_ROUTE_1_COLOR1_DEST_SHIFT);
+			InputsRead &= ~FRAG_BIT_COL1;
+			if (high_rr < 1)
+				high_rr = 1;
+			col_interp_nr++;
+		} else {
+			WARN_ONCE("fragprog wants col1, vp doesn't provide it\n");
+		}
+	}
+
+	/* Need at least one. This might still lock as the values are undefined... */
+	if (in_texcoords == 0 && col_interp_nr == 0) {
+		r300->hw.rr.cmd[R300_RR_ROUTE_0] |= 0 | R300_RS_ROUTE_0_COLOR | (fp_reg++ << R300_RS_ROUTE_0_COLOR_DEST_SHIFT);
+		col_interp_nr++;
+	}
+
+	r300->hw.rc.cmd[1] = 0 | ((in_texcoords << 2) << R300_IT_COUNT_SHIFT)
+	  | (col_interp_nr << R300_IC_COUNT_SHIFT)
+	  | R300_HIRES_EN;
+
+	assert(high_rr >= 0);
+	r300->hw.rr.cmd[R300_RR_CMD_0] = cmdpacket0(R300_RS_ROUTE_0, high_rr + 1);
+	r300->hw.rc.cmd[2] = 0xC0 | high_rr;
+
+	if (InputsRead)
+		WARN_ONCE("Don't know how to satisfy InputsRead=0x%08x\n", InputsRead);
+}
+
+
+
+
 #define bump_vpu_count(ptr, new_count)   do{\
 	drm_r300_cmd_header_t* _p=((drm_r300_cmd_header_t*)(ptr));\
 	int _nc=(new_count)/4; \
@@ -2163,6 +2285,81 @@ static void r300SetupPixelShader(r300ContextPtr rmesa)
 	}
 }
 
+static void r500SetupPixelShader(r300ContextPtr rmesa)
+{
+	GLcontext *ctx = rmesa->radeon.glCtx;
+	struct r300_fragment_program *fp = (struct r300_fragment_program *)
+	    (char *)ctx->FragmentProgram._Current;
+	int i, k;
+
+	if (!fp)		/* should only happenen once, just after context is created */
+		return;
+
+	/* emit the standard zero shader */
+	R300_STATECHANGE(rmesa, r500fp);
+	i = 1;
+	rmesa->hw.r500fp.cmd[i++] = 0x7807;
+	rmesa->hw.r500fp.cmd[i++] = R500_TEX_ID(0) | R500_TEX_INST_LD | R500_TEX_SEM_ACQUIRE | R500_TEX_IGNORE_UNCOVERED;
+	rmesa->hw.r500fp.cmd[i++] = R500_TEX_SRC_ADDR(0) |  R500_TEX_SRC_S_SWIZ_R |
+		R500_TEX_SRC_T_SWIZ_G |
+		R500_TEX_DST_ADDR(0) |
+		R500_TEX_DST_R_SWIZ_R |
+		R500_TEX_DST_G_SWIZ_G |
+		R500_TEX_DST_B_SWIZ_B |
+		R500_TEX_DST_A_SWIZ_A;
+	rmesa->hw.r500fp.cmd[i++] = R500_DX_ADDR(0) |
+		R500_DX_S_SWIZ_R |
+		R500_DX_T_SWIZ_R |
+		R500_DX_R_SWIZ_R |
+		R500_DX_Q_SWIZ_R |
+		R500_DY_ADDR(0) |
+		R500_DY_S_SWIZ_R |
+		R500_DY_T_SWIZ_R |
+		R500_DY_R_SWIZ_R |
+		R500_DY_Q_SWIZ_R;
+	rmesa->hw.r500fp.cmd[i++] = 0x0;
+	rmesa->hw.r500fp.cmd[i++] = 0x0;
+
+	rmesa->hw.r500fp.cmd[i++] = R500_INST_TYPE_OUT |
+		R500_INST_TEX_SEM_WAIT |
+		R500_INST_LAST |
+		R500_INST_RGB_OMASK_R |
+		R500_INST_RGB_OMASK_G |
+		R500_INST_RGB_OMASK_B |
+		R500_INST_ALPHA_OMASK;
+
+	rmesa->hw.r500fp.cmd[i++] = R500_RGB_ADDR0(0) |
+		R500_RGB_ADDR1(0) |
+		R500_RGB_ADDR1_CONST |
+		R500_RGB_ADDR2(0) |
+		R500_RGB_ADDR2_CONST |
+		R500_RGB_SRCP_OP_1_MINUS_2RGB0;
+	rmesa->hw.r500fp.cmd[i++] = R500_ALPHA_ADDR0(0) |
+		R500_ALPHA_ADDR1(0) |
+		R500_ALPHA_ADDR1_CONST |
+		R500_ALPHA_ADDR2(0) |
+		R500_ALPHA_ADDR2_CONST |
+		R500_ALPHA_SRCP_OP_1_MINUS_2A0;
+	rmesa->hw.r500fp.cmd[i++] = 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_SRC0 |
+		R500_ALU_RGB_R_SWIZ_B_1 |
+		R500_ALU_RGB_B_SWIZ_B_1 |
+		R500_ALU_RGB_G_SWIZ_B_1;
+	rmesa->hw.r500fp.cmd[i++] = R500_ALPHA_OP_MAD |
+		R500_ALPHA_SWIZ_A_A |
+		R500_ALPHA_SWIZ_B_1;
+	rmesa->hw.r500fp.cmd[i++] = 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;
+
+
+}
+
 void r300UpdateShaderStates(r300ContextPtr rmesa)
 {
 	GLcontext *ctx;
@@ -2170,12 +2367,19 @@ void r300UpdateShaderStates(r300ContextPtr rmesa)
 
 	r300UpdateTextureState(ctx);
 
-	r300SetupPixelShader(rmesa);
+	if (rmesa->radeon.radeonScreen->chip_family >= CHIP_FAMILY_RV515)
+		r500SetupPixelShader(rmesa);
+	else
+		r300SetupPixelShader(rmesa);
 	r300SetupTextures(ctx);
 
 	if ((rmesa->radeon.radeonScreen->chip_flags & RADEON_CHIPSET_TCL))
 		r300SetupVertexProgram(rmesa);
-	r300SetupRSUnit(ctx);
+
+	if (rmesa->radeon.radeonScreen->chip_family >= CHIP_FAMILY_RV515)
+		r500SetupRSUnit(ctx);
+	else
+		r300SetupRSUnit(ctx);
 }
 
 /**
diff --git a/src/mesa/drivers/dri/radeon/radeon_chipset.h b/src/mesa/drivers/dri/radeon/radeon_chipset.h
index 6ad441bdd04..9e375474a03 100644
--- a/src/mesa/drivers/dri/radeon/radeon_chipset.h
+++ b/src/mesa/drivers/dri/radeon/radeon_chipset.h
@@ -169,6 +169,8 @@ enum {
    CHIP_FAMILY_RV410,
    CHIP_FAMILY_RS400,
    CHIP_FAMILY_RS690,
+   CHIP_FAMILY_RV515,
+   CHIP_FAMILY_R520,
    CHIP_FAMILY_LAST
 };