radeonsi: allow out-of-order rasterization in commutative blending cases

We do not enable this by default for additive blending, since it slightly
breaks OpenGL invariance guarantees due to non-determinism.

Still, there may be some applications can benefit from white-listing
via the radeonsi_commutative_blend_add drirc setting without any real
visible artifacts.

Reviewed-by: Marek Olšák <marek.olsak@amd.com>
Tested-by: Dieter Nützel <Dieter@nuetzel-hh.de>

src/gallium/drivers/radeonsi/driinfo_radeonsi.h

@@ -2,4 +2,5 @@
 DRI_CONF_SECTION_PERFORMANCE
     DRI_CONF_RADEONSI_ENABLE_SISCHED("false")
     DRI_CONF_RADEONSI_ASSUME_NO_Z_FIGHTS("false")
+    DRI_CONF_RADEONSI_COMMUTATIVE_BLEND_ADD("false")
 DRI_CONF_SECTION_END

src/gallium/drivers/radeonsi/si_pipe.c

@@ -1050,6 +1050,8 @@ struct pipe_screen *radeonsi_screen_create(struct radeon_winsys *ws,
 					 !(sscreen->b.debug_flags & DBG_NO_OUT_OF_ORDER);
 	sscreen->assume_no_z_fights =
 		driQueryOptionb(config->options, "radeonsi_assume_no_z_fights");
+	sscreen->commutative_blend_add =
+		driQueryOptionb(config->options, "radeonsi_commutative_blend_add");
 	sscreen->has_msaa_sample_loc_bug = (sscreen->b.family >= CHIP_POLARIS10 &&
 					    sscreen->b.family <= CHIP_POLARIS12) ||
 					   sscreen->b.family == CHIP_VEGA10 ||

src/gallium/drivers/radeonsi/si_pipe.h

@@ -96,6 +96,7 @@ struct si_screen {
 	bool				has_draw_indirect_multi;
 	bool				has_out_of_order_rast;
 	bool				assume_no_z_fights;
+	bool				commutative_blend_add;
 	bool				has_msaa_sample_loc_bug;
 	bool				dpbb_allowed;
 	bool				dfsm_allowed;

src/gallium/drivers/radeonsi/si_state.c

@@ -377,6 +377,48 @@ static uint32_t si_translate_blend_opt_factor(int blend_fact, bool is_alpha)
 	}
 }
 
+static void si_blend_check_commutativity(struct si_screen *sscreen,
+					 struct si_state_blend *blend,
+					 enum pipe_blend_func func,
+					 enum pipe_blendfactor src,
+					 enum pipe_blendfactor dst,
+					 unsigned chanmask)
+{
+	/* Src factor is allowed when it does not depend on Dst */
+	static const uint32_t src_allowed =
+		(1u << PIPE_BLENDFACTOR_ONE) |
+		(1u << PIPE_BLENDFACTOR_SRC_COLOR) |
+		(1u << PIPE_BLENDFACTOR_SRC_ALPHA) |
+		(1u << PIPE_BLENDFACTOR_SRC_ALPHA_SATURATE) |
+		(1u << PIPE_BLENDFACTOR_CONST_COLOR) |
+		(1u << PIPE_BLENDFACTOR_CONST_ALPHA) |
+		(1u << PIPE_BLENDFACTOR_SRC1_COLOR) |
+		(1u << PIPE_BLENDFACTOR_SRC1_ALPHA) |
+		(1u << PIPE_BLENDFACTOR_ZERO) |
+		(1u << PIPE_BLENDFACTOR_INV_SRC_COLOR) |
+		(1u << PIPE_BLENDFACTOR_INV_SRC_ALPHA) |
+		(1u << PIPE_BLENDFACTOR_INV_CONST_COLOR) |
+		(1u << PIPE_BLENDFACTOR_INV_CONST_ALPHA) |
+		(1u << PIPE_BLENDFACTOR_INV_SRC1_COLOR) |
+		(1u << PIPE_BLENDFACTOR_INV_SRC1_ALPHA);
+
+	if (dst == PIPE_BLENDFACTOR_ONE &&
+	    (src_allowed & (1u << src))) {
+		/* Addition is commutative, but floating point addition isn't
+		 * associative: subtle changes can be introduced via different
+		 * rounding.
+		 *
+		 * Out-of-order is also non-deterministic, which means that
+		 * this breaks OpenGL invariance requirements. So only enable
+		 * out-of-order additive blending if explicitly allowed by a
+		 * setting.
+		 */
+		if (func == PIPE_BLEND_MAX || func == PIPE_BLEND_MIN ||
+		    (func == PIPE_BLEND_ADD && sscreen->commutative_blend_add))
+			blend->commutative_4bit |= chanmask;
+	}
+}
+
 /**
  * Get rid of DST in the blend factors by commuting the operands:
  *    func(src * DST, dst * 0) ---> func(src * 0, dst * SRC)
@@ -493,6 +535,11 @@ static void *si_create_blend_state_mode(struct pipe_context *ctx,
 			continue;
 		}
 
+		si_blend_check_commutativity(sctx->screen, blend,
+					     eqRGB, srcRGB, dstRGB, 0x7 << (4 * i));
+		si_blend_check_commutativity(sctx->screen, blend,
+					     eqA, srcA, dstA, 0x8 << (4 * i));
+
 		/* Blending optimizations for RB+.
 		 * These transformations don't change the behavior.
 		 *
@@ -636,6 +683,7 @@ static void si_bind_blend_state(struct pipe_context *ctx, void *state)
 	    (!old_blend ||
 	     (old_blend->blend_enable_4bit != blend->blend_enable_4bit ||
 	      old_blend->cb_target_enabled_4bit != blend->cb_target_enabled_4bit ||
+	      old_blend->commutative_4bit != blend->commutative_4bit ||
 	      old_blend->logicop_enable != blend->logicop_enable)))
 		si_mark_atom_dirty(sctx, &sctx->msaa_config);
 }
@@ -3208,12 +3256,23 @@ static bool si_out_of_order_rasterization(struct si_context *sctx)
 	if (!colormask)
 		return true;
 
-	bool blend_enabled = (colormask & blend->blend_enable_4bit) != 0;
+	unsigned blendmask = colormask & blend->blend_enable_4bit;
 
-	if (blend_enabled)
+	if (blendmask) {
-		return false; /* TODO */
+		/* Only commutative blending. */
+		if (blendmask & ~blend->commutative_4bit)
+			return false;
 
-	return dsa_order_invariant.pass_last;
+		if (!dsa_order_invariant.pass_set)
+			return false;
+	}
+
+	if (colormask & ~blendmask) {
+		if (!dsa_order_invariant.pass_last)
+			return false;
+	}
+
+	return true;
 }
 
 static void si_emit_msaa_config(struct si_context *sctx, struct r600_atom *atom)

src/gallium/drivers/radeonsi/si_state.h

@@ -55,6 +55,7 @@ struct si_state_blend {
 	unsigned		cb_target_enabled_4bit;
 	unsigned		blend_enable_4bit;
 	unsigned		need_src_alpha_4bit;
+	unsigned		commutative_4bit;
 	bool			alpha_to_coverage:1;
 	bool			alpha_to_one:1;
 	bool			dual_src_blend:1;

src/util/xmlpool/t_options.h

@@ -443,3 +443,8 @@ DRI_CONF_OPT_END
 DRI_CONF_OPT_BEGIN_B(radeonsi_assume_no_z_fights, def) \
         DRI_CONF_DESC(en,gettext("Assume no Z fights (enables aggressive out-of-order rasterization to improve performance; may cause rendering errors)")) \
 DRI_CONF_OPT_END
+
+#define DRI_CONF_RADEONSI_COMMUTATIVE_BLEND_ADD(def) \
+DRI_CONF_OPT_BEGIN_B(radeonsi_commutative_blend_add, def) \
+        DRI_CONF_DESC(en,gettext("Commutative additive blending optimizations (may cause rendering errors)")) \
+DRI_CONF_OPT_END