radeon: fixup r500 FP emission for new CS

3 files changed, 95 insertions, 42 deletions

diff --git a/src/mesa/drivers/dri/r300/r300_cmdbuf.c b/src/mesa/drivers/dri/r300/r300_cmdbuf.c
index ca9d36a8b6..ed94520aba 100644
--- a/src/mesa/drivers/dri/r300/r300_cmdbuf.c
+++ b/src/mesa/drivers/dri/r300/r300_cmdbuf.c
@@ -284,6 +284,45 @@ void emit_vpu(r300ContextPtr r300, struct r300_state_atom * atom)
     }
 }
 
+void emit_r500fp(r300ContextPtr r300, struct r300_state_atom * atom)
+{
+	BATCH_LOCALS(r300);
+	drm_r300_cmd_header_t cmd;
+	uint32_t addr, ndw, i, sz;
+	int type, clamp, stride;
+
+	if (!r300->radeon.radeonScreen->kernel_mm) {
+		uint32_t dwords;
+		dwords = (*atom->check) (r300, atom);
+		BEGIN_BATCH_NO_AUTOSTATE(dwords);
+		OUT_BATCH_TABLE(atom->cmd, dwords);
+		END_BATCH();
+		return;
+	}
+
+	cmd.u = atom->cmd[0];
+	sz = cmd.r500fp.count;
+	addr = ((cmd.r500fp.adrhi_flags & 1) << 8) | cmd.r500fp.adrlo;
+	type = !!(cmd.r500fp.adrhi_flags & R500FP_CONSTANT_TYPE);
+	clamp = !!(cmd.r500fp.adrhi_flags & R500FP_CONSTANT_CLAMP);
+
+	addr |= (type << 16);
+	addr |= (clamp << 17);
+
+	stride = type ? 4 : 6;
+
+	ndw = sz * stride;
+	if (ndw) {
+
+		OUT_BATCH(CP_PACKET0(R500_GA_US_VECTOR_INDEX, 0));
+		OUT_BATCH(addr);
+		OUT_BATCH(CP_PACKET0(R500_GA_US_VECTOR_DATA, ndw-1) | RADEON_ONE_REG_WR);
+		for (i = 0; i < ndw; i++) {
+			OUT_BATCH(atom->cmd[i+1]);
+		}
+	}
+}
+
 static void emit_tex_offsets(r300ContextPtr r300, struct r300_state_atom * atom)
 {
 	BATCH_LOCALS(r300);
@@ -395,7 +434,7 @@ int check_vpu(r300ContextPtr r300, struct r300_state_atom *atom)
 	return cnt ? (cnt * 4) + 1 : 0;
 }
 
-static int check_r500fp(r300ContextPtr r300, struct r300_state_atom *atom)
+int check_r500fp(r300ContextPtr r300, struct r300_state_atom *atom)
 {
 	int cnt;
 
@@ -403,11 +442,11 @@ static int check_r500fp(r300ContextPtr r300, struct r300_state_atom *atom)
 	return cnt ? (cnt * 6) + 1 : 0;
 }
 
-static int check_r500fp_const(r300ContextPtr r300, struct r300_state_atom *atom)
+int check_r500fp_const(r300ContextPtr r300, struct r300_state_atom *atom)
 {
 	int cnt;
 
-    cnt = r500fp_count(atom->cmd);
+	cnt = r500fp_count(atom->cmd);
 	return cnt ? (cnt * 4) + 1 : 0;
 }
 
@@ -570,10 +609,12 @@ void r300InitCmdBuf(r300ContextPtr r300)
 
 		ALLOC_STATE(r500fp, r500fp, R500_FPI_CMDSIZE, 0);
 		r300->hw.r500fp.cmd[R300_FPI_CMD_0] =
-            cmdr500fp(r300->radeon.radeonScreen, 0, 0, 0, 0);
+			cmdr500fp(r300->radeon.radeonScreen, 0, 0, 0, 0);
+		r300->hw.r500fp.emit = emit_r500fp;
 		ALLOC_STATE(r500fp_const, r500fp_const, R500_FPP_CMDSIZE, 0);
 		r300->hw.r500fp_const.cmd[R300_FPI_CMD_0] =
-            cmdr500fp(r300->radeon.radeonScreen, 0, 0, 1, 0);
+			cmdr500fp(r300->radeon.radeonScreen, 0, 0, 1, 0);
+		r300->hw.r500fp_const.emit = emit_r500fp;
 	} else {
 		ALLOC_STATE(fp, always, R300_FP_CMDSIZE, 0);
 		r300->hw.fp.cmd[R300_FP_CMD_0] = cmdpacket0(r300->radeon.radeonScreen, R300_US_CONFIG, 3);
diff --git a/src/mesa/drivers/dri/r300/r300_cmdbuf.h b/src/mesa/drivers/dri/r300/r300_cmdbuf.h
index 1a249c8d52..95701b49b4 100644
--- a/src/mesa/drivers/dri/r300/r300_cmdbuf.h
+++ b/src/mesa/drivers/dri/r300/r300_cmdbuf.h
@@ -132,4 +132,8 @@ void r300BeginBatch(r300ContextPtr r300,
 void emit_vpu(r300ContextPtr r300, struct r300_state_atom * atom);
 int check_vpu(r300ContextPtr r300, struct r300_state_atom *atom);
 
+void emit_r500fp(r300ContextPtr r300, struct r300_state_atom * atom);
+int check_r500fp(r300ContextPtr r300, struct r300_state_atom *atom);
+int check_r500fp_const(r300ContextPtr r300, struct r300_state_atom *atom);
+
 #endif				/* __R300_CMDBUF_H__ */
diff --git a/src/mesa/drivers/dri/r300/r300_ioctl.c b/src/mesa/drivers/dri/r300/r300_ioctl.c
index a86841cea4..0cf084bde5 100644
--- a/src/mesa/drivers/dri/r300/r300_ioctl.c
+++ b/src/mesa/drivers/dri/r300/r300_ioctl.c
@@ -387,6 +387,9 @@ static void r300EmitClearState(GLcontext * ctx)
 			FP_SELA(0, NO, W, FP_TMP(0), 0, 0));
 		END_BATCH();
 	} else {
+		struct r300_state_atom r500fp;
+		uint32_t _cmd[10];
+
 		R300_STATECHANGE(r300, fp);
 		R300_STATECHANGE(r300, r500fp);
 
@@ -398,44 +401,49 @@ static void r300EmitClearState(GLcontext * ctx)
 		OUT_BATCH(R500_US_CODE_START_ADDR(0) | R500_US_CODE_END_ADDR(1));
 		OUT_BATCH(R500_US_CODE_RANGE_ADDR(0) | R500_US_CODE_RANGE_SIZE(1));
 		OUT_BATCH(R500_US_CODE_OFFSET_ADDR(0));
-
-		OUT_BATCH(cmdr500fp(r300->radeon.radeonScreen, 0, 1, 0, 0));
-		OUT_BATCH(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 |
-			  R500_INST_RGB_CLAMP |
-			  R500_INST_ALPHA_CLAMP);
-		OUT_BATCH(R500_RGB_ADDR0(0) |
-			  R500_RGB_ADDR1(0) |
-			  R500_RGB_ADDR1_CONST |
-			  R500_RGB_ADDR2(0) |
-			  R500_RGB_ADDR2_CONST);
-		OUT_BATCH(R500_ALPHA_ADDR0(0) |
-			  R500_ALPHA_ADDR1(0) |
-			  R500_ALPHA_ADDR1_CONST |
-			  R500_ALPHA_ADDR2(0) |
-			  R500_ALPHA_ADDR2_CONST);
-		OUT_BATCH(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_R |
-			  R500_ALU_RGB_B_SWIZ_B_G |
-			  R500_ALU_RGB_G_SWIZ_B_B);
-		OUT_BATCH(R500_ALPHA_OP_CMP |
-			  R500_ALPHA_SWIZ_A_A |
-			  R500_ALPHA_SWIZ_B_A);
-		OUT_BATCH(R500_ALU_RGBA_OP_CMP |
-			  R500_ALU_RGBA_R_SWIZ_0 |
-			  R500_ALU_RGBA_G_SWIZ_0 |
-			  R500_ALU_RGBA_B_SWIZ_0 |
-			  R500_ALU_RGBA_A_SWIZ_0);
 		END_BATCH();
+
+		r500fp.check = check_r500fp;
+		r500fp.cmd = _cmd;
+		r500fp.cmd[0] = cmdr500fp(r300->radeon.radeonScreen, 0, 1, 0, 0);
+		r500fp.cmd[1] = 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 |
+			R500_INST_RGB_CLAMP |
+			R500_INST_ALPHA_CLAMP;
+		r500fp.cmd[2] = R500_RGB_ADDR0(0) |
+			R500_RGB_ADDR1(0) |
+			R500_RGB_ADDR1_CONST |
+			R500_RGB_ADDR2(0) |
+			R500_RGB_ADDR2_CONST;
+		r500fp.cmd[3] = R500_ALPHA_ADDR0(0) |
+			R500_ALPHA_ADDR1(0) |
+			R500_ALPHA_ADDR1_CONST |
+			R500_ALPHA_ADDR2(0) |
+			R500_ALPHA_ADDR2_CONST;
+		r500fp.cmd[4] = 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_R |
+			R500_ALU_RGB_B_SWIZ_B_G |
+			R500_ALU_RGB_G_SWIZ_B_B;
+		r500fp.cmd[5] = R500_ALPHA_OP_CMP |
+			R500_ALPHA_SWIZ_A_A |
+			R500_ALPHA_SWIZ_B_A;
+		r500fp.cmd[6] = R500_ALU_RGBA_OP_CMP |
+			R500_ALU_RGBA_R_SWIZ_0 |
+			R500_ALU_RGBA_G_SWIZ_0 |
+			R500_ALU_RGBA_B_SWIZ_0 |
+			R500_ALU_RGBA_A_SWIZ_0;
+		
+		r500fp.cmd[7] = 0;
+		emit_r500fp(r300, &r500fp);
 	}
 
 	BEGIN_BATCH(2);