wip (no piglit regressions on evergreen)wip-sb-tgsi

2 files changed, 66 insertions, 41 deletions

diff --git a/src/gallium/drivers/r600/sb/sb_bc_finalize.cpp b/src/gallium/drivers/r600/sb/sb_bc_finalize.cpp
index ffeb08414b..4218990a19 100644
--- a/src/gallium/drivers/r600/sb/sb_bc_finalize.cpp
+++ b/src/gallium/drivers/r600/sb/sb_bc_finalize.cpp
@@ -350,7 +350,11 @@ void bc_finalizer::finalize_alu_src(alu_group_node* g, alu_node* a) {
 
 			if (lv == literal(0))
 				src.sel = ALU_SRC_0;
-			else if (lv == literal(0.5f))
+			else if (lv == literal(0x80000000)) {
+				// XXX probably we shouldn't have -0 here in the first place?
+				src.sel = ALU_SRC_0;
+				src.neg = 1;
+			} else if (lv == literal(0.5f))
 				src.sel = ALU_SRC_0_5;
 			else if (lv == literal(-0.5f)) {
 				src.sel = ALU_SRC_0_5;
diff --git a/src/gallium/drivers/r600/sb/sb_tgsi.cpp b/src/gallium/drivers/r600/sb/sb_tgsi.cpp
index 0f0d6df745..361323dd31 100644
--- a/src/gallium/drivers/r600/sb/sb_tgsi.cpp
+++ b/src/gallium/drivers/r600/sb/sb_tgsi.cpp
@@ -46,7 +46,7 @@ const tgsi_translator::tgsi_inst_info tgsi_translator::tgsi_info_table[TGSI_OPCO
         /*   0 */ TI_DESC(ARL, 0, ti_arl, 0),
         /*   1 */ TI_DESC(MOV, ALU_OP1_MOV, ti_alu, 0),
         /*   2 */ TI_DESC(LIT, 0, ti_lit, 0),
-        /*   3 */ TI_DESC(RCP, ALU_OP1_RECIP_CLAMPED, ti_repl, 0),
+        /*   3 */ TI_DESC(RCP, ALU_OP1_RECIP_IEEE, ti_repl, 0),
         /*   4 */ TI_DESC(RSQ, ALU_OP1_RECIPSQRT_CLAMPED, ti_repl, 0),
         /*   5 */ TI_DESC(EXP, 0, ti_exp, 0),
         /*   6 */ TI_DESC(LOG, 0, ti_log, 0),
@@ -369,13 +369,15 @@ int tgsi_translator::parse_declaration() {
 		input[i].d.centroid = d->Interp.Centroid;
 		input[i].tgsi_index = d->Range.First;
 		if (tgsi_proc == TGSI_PROCESSOR_FRAGMENT) {
-			if (input[i].d.interpolate == TGSI_INTERPOLATE_LINEAR
-			        || input[i].d.interpolate == TGSI_INTERPOLATE_COLOR)
-				interp_mask |= (1 << 1);
-			else if (input[i].d.interpolate == TGSI_INTERPOLATE_PERSPECTIVE)
-				interp_mask |= (1 << 0);
-			if (input[i].d.centroid)
-				interp_mask |= (1 << 2);
+			if (input[i].d.name != TGSI_SEMANTIC_POSITION &&
+					input[i].d.name != TGSI_SEMANTIC_FACE) {
+				if (input[i].d.interpolate == TGSI_INTERPOLATE_LINEAR)
+					interp_mask |= (1 << 1);
+				else if (input[i].d.interpolate == TGSI_INTERPOLATE_PERSPECTIVE)
+					interp_mask |= (1 << 0);
+				if (input[i].d.centroid)
+					interp_mask |= (1 << 2);
+			}
 
 			input[i].d.spi_sid = spi_sid(input[i].d.name, input[i].d.sid);
 
@@ -822,7 +824,10 @@ int tgsi_translator::emit_inputs() {
 	}
 	case TARGET_PS:
 		if (ctx.is_egcm()) {
-			unsigned ij_pairs = ((interp_mask & 1) + (interp_mask >> 1))
+			if (!interp_mask)
+				interp_mask = 1;
+
+			unsigned ij_pairs = ((interp_mask & 1) + ((interp_mask >> 1) & 1))
 			        * ((interp_mask & 4) ? 2 : 1);
 
 			unsigned mask = (1u << 2 * ij_pairs) - 1;
@@ -833,7 +838,6 @@ int tgsi_translator::emit_inputs() {
 				++gpr;
 				mask >>= 4;
 			}
-
 			gpr_reserved = gpr;
 		}
 
@@ -873,11 +877,6 @@ int tgsi_translator::emit_inputs() {
 			} else {
 				sh->add_pinned_inputs(sh->root->dst, VLK_TGSI_INPUT,
 				    in.tgsi_index, 0xF, false, in.d.gpr);
-
-				if (fragcoord_input == i) {
-					value* w = get_tgsi_value(VLK_TGSI_INPUT, i, SEL_W);
-					emit_alu(ALU_OP1_RECIP_IEEE, w, 0, asrc(w));
-				}
 			}
 
 			if (two_side) {
@@ -908,6 +907,13 @@ int tgsi_translator::emit_inputs() {
 				}
 			}
 		}
+
+		if (fragcoord_input != -1) {
+			value* w = get_tgsi_value(VLK_TGSI_INPUT, fragcoord_input, SEL_W);
+			emit_alu(ALU_OP1_RECIP_IEEE, w, 0, asrc(w));
+		}
+
+
 		break;
 	default:
 		assert(!"unexpected target");
@@ -959,10 +965,12 @@ alu_group_node* tgsi_translator::build_interp_flat(shader_io& in) {
 }
 
 inline int tgsi_translator::get_ij(shader_io& in) {
-	int ij;
-	ij = in.d.centroid ? 1 : 0;
-	if (in.d.interpolate != TGSI_INTERPOLATE_PERSPECTIVE)
-		ij += (interp_mask & 1) + ((interp_mask >> 2) & 1);
+	int ij = 0;
+	if (in.d.interpolate == TGSI_INTERPOLATE_PERSPECTIVE)
+		return in.d.centroid ? 1 : 0;
+	else if (in.d.interpolate == TGSI_INTERPOLATE_LINEAR)
+		return (interp_mask & 1) + ((interp_mask >> 2) & 1) +
+				(in.d.centroid ? 1 : 0);
 	return ij;
 }
 
@@ -992,8 +1000,6 @@ value* tgsi_translator::get_arg_value(unsigned index, unsigned chan) {
 }
 
 int tgsi_translator::ti_alu() {
-	begin_group();
-
 	switch (info->tgsi_op) {
 	case TGSI_OPCODE_SUB:
 		args.src[1].neg = !args.src[1].neg;
@@ -1004,6 +1010,7 @@ int tgsi_translator::ti_alu() {
 		break;
 	}
 
+	begin_group();
 	if (unlikely(info->flags & TIF_ALU_SWAPSRC01)) {
 		FOREACH_CHAN
 		{
@@ -1076,17 +1083,16 @@ int tgsi_translator::ti_dot() {
 }
 
 int tgsi_translator::ti_repl() {
-
 	switch (info->tgsi_op) {
 	case TGSI_OPCODE_RSQ:
 		args.src[0].abs = 1;
 		args.src[0].neg = 0;
 		break;
 	}
-	FOREACH_CHAN
-	{
-		emit_alu(info->isa_op, 0, ch);
-	}
+
+	value *t = create_temp();
+	emit_alu(info->isa_op, t, clamp, asrc(args.src[0], SEL_X));
+	ti_replicate(t);
 	return 0;
 }
 
@@ -1148,10 +1154,8 @@ alu_node* tgsi_translator::create_alu(unsigned op) {
 
 int tgsi_translator::ti_trig() {
 	value *t = prepare_trig(asrc(args.src[0], 0));
-	FOREACH_CHAN
-	{
-		emit_alu(info->isa_op, tgsi_dst(ch), clamp, t);
-	}
+	emit_alu(info->isa_op, t, clamp, asrc(t));
+	ti_replicate(t);
 	return 0;
 }
 
@@ -1160,9 +1164,9 @@ int tgsi_translator::ti_scs() {
 
 	begin_group();
 	if (write_mask & (1 << SEL_X))
-		emit_alu(ALU_OP1_COS, tgsi_dst(SEL_X), clamp, t);
+		emit_alu(ALU_OP1_COS, tgsi_dst(SEL_X), clamp, asrc(t));
 	if (write_mask & (1 << SEL_Y))
-		emit_alu(ALU_OP1_SIN, tgsi_dst(SEL_Y), clamp, t);
+		emit_alu(ALU_OP1_SIN, tgsi_dst(SEL_Y), clamp, asrc(t));
 	if (write_mask & (1 << SEL_Z))
 		emit_alu(ALU_OP1_MOV, tgsi_dst(SEL_Z), 0, asrc(0.0f));
 	if (write_mask & (1 << SEL_W))
@@ -1184,8 +1188,13 @@ value* tgsi_translator::prepare_trig(alu_src s) {
 	if (ctx.is_r600())
 		emit_alu(ALU_OP3_MULADD, t, 0, asrc(t), asrc(double_pi), asrc(neg_pi));
 	else
-//		emit_alu(ALU_OP2_ADD, t, 0, asrc(t), asrc(-0.5f));
+#if 0
+		emit_alu(ALU_OP2_ADD, t, 0, asrc(t), asrc(-0.5f));
+#else
+		// using muladd just to reduce differences from default backend for
+		// debugging
 		emit_alu(ALU_OP3_MULADD, t, 0, asrc(t), asrc(1.0f), asrc(0.5f, 0, 1));
+#endif
 
 	return t;
 }
@@ -1228,11 +1237,11 @@ int tgsi_translator::ti_log() {
 
 	begin_group();
 	if (write_mask & (1 << SEL_X))
-		emit_alu(ALU_OP1_MOV, tgsi_dst(SEL_X), clamp, t2);
+		emit_alu(ALU_OP1_MOV, tgsi_dst(SEL_X), clamp, asrc(t2));
 	if (write_mask & (1 << SEL_Y))
 		emit_alu(ALU_OP2_MUL, tgsi_dst(SEL_Y), clamp, s, asrc(t3));
 	if (write_mask & (1 << SEL_Z))
-		emit_alu(ALU_OP1_MOV, tgsi_dst(SEL_Z), clamp, t);
+		emit_alu(ALU_OP1_MOV, tgsi_dst(SEL_Z), clamp, asrc(t));
 	if (write_mask & (1 << SEL_W))
 		emit_alu(ALU_OP1_MOV, tgsi_dst(SEL_W), 0, asrc(1.0f));
 	end_group();
@@ -1278,7 +1287,7 @@ int tgsi_translator::ti_pow() {
 	value* t = create_temp();
 
 	emit_alu(ALU_OP1_LOG_IEEE, t, 0, asrc(args.src[0], 0));
-	emit_alu(ALU_OP2_MUL, t, 0, asrc(args.src[1], 0), t);
+	emit_alu(ALU_OP2_MUL, t, 0, asrc(args.src[1], 0), asrc(t));
 	emit_alu(ALU_OP1_EXP_IEEE, t, clamp, asrc(t));
 	ti_replicate(t);
 	return 0;
@@ -1433,10 +1442,12 @@ int tgsi_translator::ti_f2iu() {
 	{
 		emit_alu(ALU_OP1_TRUNC, t[ch], 0, asrc(args.src[0], ch));
 	}
+	begin_group();
 	FOREACH_CHAN
 	{
 		emit_alu(info->isa_op, tgsi_dst(ch), 0, t[ch]);
 	}
+	end_group();
 	return 0;
 }
 
@@ -1500,6 +1511,9 @@ int tgsi_translator::ti_divmod() {
 	value *t2z = create_temp();
 	value *t3x = create_temp();
 
+	vvec dst;
+	create_temps(dst, 4);
+
 	FOREACH_CHAN
 	{
 		if (signed_op) {
@@ -1602,20 +1616,26 @@ int tgsi_translator::ti_divmod() {
 			if (mod) {
 				/* sign of the remainder is the same as the sign of src0 */
 				/* tmp0.x = src0>=0 ? tmp0.z : tmp0.x */
-				emit_alu(ALU_OP3_CNDGE_INT, tgsi_dst(ch), 0, asrc(t2x),
+				emit_alu(ALU_OP3_CNDGE_INT, dst[ch], 0, asrc(t2x),
 				    asrc(t0z), asrc(t0x));
 			} else {
 				/* fix the quotient sign (same as the sign of src0*src1) */
 				/* tmp0.x = tmp2.z>=0 ? tmp0.z : tmp0.x */
-				emit_alu(ALU_OP3_CNDGE_INT, tgsi_dst(ch), 0, asrc(t2z),
+				emit_alu(ALU_OP3_CNDGE_INT, dst[ch], 0, asrc(t2z),
 				    asrc(t0z), asrc(t0x));
 			}
 		} else { // unsigned
 			/* 19. dst = tmp1.y==0 ? tmp1.w : tmp0.z */
-			emit_alu(ALU_OP3_CNDE_INT, tgsi_dst(ch), 0, asrc(t1y), asrc(t1w),
+			emit_alu(ALU_OP3_CNDE_INT, dst[ch], 0, asrc(t1y), asrc(t1w),
 			    asrc(t0z));
 		}
 	}
+	begin_group();
+	FOREACH_CHAN
+	{
+		emit_alu(ALU_OP1_MOV, tgsi_dst(ch), 0, asrc(dst[ch]));
+	}
+	end_group();
 	return 0;
 }
 
@@ -1990,6 +2010,7 @@ int tgsi_translator::ti_tex() {
 		}
 
 		/* for cube forms of lod and bias we need to route things */
+		// XXX just copy target value pointer to src?
 		if (tgsi_op == TGSI_OPCODE_TXB || tgsi_op == TGSI_OPCODE_TXL)
 			emit_alu(ALU_OP1_MOV, tv[SEL_Z], 0, asrc(args.src[0], SEL_W));
 		else if (tgsi_op == TGSI_OPCODE_TXB2 || tgsi_op == TGSI_OPCODE_TXL2)
@@ -2178,7 +2199,7 @@ int tgsi_translator::ti_if() {
 	alu_node *a = build_alu(info->isa_op, sh->get_special_value(SV_EXEC_MASK),
 	    0, asrc(args.src[0], 0), asrc(0.0f));
 
-	a->dst.insert(a->dst.begin(), 2, NULL);
+	a->dst.insert(a->dst.begin(), 2, (value*)NULL);
 	emit_node(a);
 
 	region_node *r = sh->create_region();