GS: Accurate UV rounding for axis-aligned prims.

Author: TJnotJT

bin/resources/shaders/dx11/tfx.fx

@@ -13,6 +13,7 @@
 #define VS_IIP 0
 #define VS_TME 1
 #define VS_FST 1
+#define VS_ROUND_UV 0
 #endif
 
 #ifndef GS_IIP
@@ -79,6 +80,7 @@
 #define PS_NO_COLOR 0
 #define PS_NO_COLOR1 0
 #define PS_DATE 0
+#define PS_ROUND_UV 0
 #endif
 
 #define SW_BLEND (PS_BLEND_A || PS_BLEND_B || PS_BLEND_D)
@@ -108,6 +110,10 @@ struct VS_OUTPUT
 #else
 	nointerpolation float4 c : COLOR0;
 #endif
+
+#if VS_ROUND_UV != 0
+	nointerpolation uint4 rounduv : TEXCOORD3;
+#endif
 };
 
 struct PS_INPUT
@@ -120,6 +126,9 @@ struct PS_INPUT
 #else
 	nointerpolation float4 c : COLOR0;
 #endif
+#if PS_ROUND_UV != 0
+	nointerpolation uint4 rounduv : TEXCOORD3;
+#endif
 #if (PS_DATE >= 1 && PS_DATE <= 3) || GS_FORWARD_PRIMID
 	uint primid : SV_PrimitiveID;
 #endif
@@ -322,6 +331,37 @@ float4 clamp_wrap_uv(float4 uv)
 	return uv;
 }
 
+float4 round_uv(PS_INPUT input)
+{
+#if PS_ROUND_UV
+	// Check if we're at the prim top or left.
+	int2 topleft = int2(int2(input.p.xy) == int2(input.rounduv.xy));
+
+	// Get flags for whether to round U, V.
+	int2 round_flags = int2(input.rounduv.zw);
+
+	// Being on the top or left pixels converts round down to round up.
+	int2 round_down = int2(round_flags == PS_ROUND_UV_DOWN) & ~topleft;
+	int2 round_up = int2(round_flags == PS_ROUND_UV_UP) |
+	                (int2(round_flags == PS_ROUND_UV_DOWN) & topleft);
+
+	float2 uv = input.ti.zw; // Unnormalized UVs.
+	float2 uvi = round(input.ti.zw / 8.0f) * 8.0f; // Nearest half texel.
+	
+	// Round only if close to a half texel.
+	int2 close = int2(abs(uv - uvi) <= PS_ROUND_UV_THRESHOLD);
+	round_down &= close;
+	round_up &= close;
+
+	uv = bool2(round_down) ? uvi - PS_ROUND_UV_THRESHOLD : uv;
+	uv = bool2(round_up) ? uvi + PS_ROUND_UV_THRESHOLD : uv;
+
+	return float4(uv / 16.0f / WH.xy, uv); // Return normalized and unnormalized coords.
+#else
+	return float4(0.0f, 0.0f, 0.0f, 0.0f);
+#endif
+}
+
 float4x4 sample_4c(float4 uv, float uv_w, int2 xy)
 {
 	float4x4 c;
@@ -756,6 +796,10 @@ float4 ps_color(PS_INPUT input)
 #if PS_FST == 0
 	float2 st = input.t.xy / input.t.w;
 	float2 st_int = input.ti.zw / input.t.w;
+#elif PS_ROUND_UV != 0
+	float4 ti_rounded = round_uv(input);
+	float2 st = ti_rounded.xy;
+	float2 st_int = ti_rounded.zw;
 #else
 	float2 st = input.ti.xy;
 	float2 st_int = input.ti.zw;
@@ -1253,6 +1297,22 @@ cbuffer cb0
 	uint BaseVertex; // Only used in DX11.
 };
 
+float2 sign_extend_16_bit(float2 uv)
+{
+	return uv > float(0x7FFF) ? uv - float(0x10000) : uv;
+}
+
+uint4 extract_round_uv_bits(float q)
+{
+	uint qi = asuint(q);
+	return uint4(
+		(qi >> 0) & 0xFFF,  // Prim left
+		(qi >> 12) & 0xFFF, // Prim top
+		(qi >> 24) & 0xF,   // Round U flags
+		(qi >> 28) & 0xF    // Round V flags
+	);
+}
+
 VS_OUTPUT vs_main(VS_INPUT input)
 {
 	// Clamp to max depth, gs doesn't wrap
@@ -1272,7 +1332,11 @@ VS_OUTPUT vs_main(VS_INPUT input)
 
 	if(VS_TME)
 	{
-		float2 uv = input.uv - TextureOffset;
+		#if VS_ROUND_UV == 0
+			float2 uv = input.uv - TextureOffset;
+		#else
+			float2 uv = sign_extend_16_bit(input.uv) - TextureOffset; // Extend sign bit in case ST was converted to UV.
+		#endif
 		float2 st = input.st - TextureOffset;
 
 		// Integer nomalized
@@ -1291,12 +1355,20 @@ VS_OUTPUT vs_main(VS_INPUT input)
 		// Float coords
 		output.t.xy = st;
 		output.t.w = input.q;
+
+		// Get UV rounding info saved in Q.
+		#if VS_ROUND_UV
+			output.rounduv = extract_round_uv_bits(input.q);
+		#endif
 	}
 	else
 	{
 		output.t.xy = 0;
 		output.t.w = 1.0f;
 		output.ti = 0;
+		#if VS_ROUND_UV
+			output.rounduv = 0;
+		#endif
 	}
 
 	output.c = input.c;

bin/resources/shaders/opengl/tfx_fs.glsl

@@ -70,6 +70,10 @@ in SHADER
 	#else
 		flat vec4 c;
 	#endif
+
+	#if PS_ROUND_UV != 0
+		flat uvec4 rounduv;
+	#endif
 } PSin;
 
 #define TARGET_0_QUALIFIER out
@@ -252,6 +256,37 @@ vec4 clamp_wrap_uv(vec4 uv)
 	return uv_out;
 }
 
+vec4 round_uv()
+{
+#if PS_ROUND_UV
+	// Check if we're at the prim top or left.
+	ivec2 topleft = ivec2(equal(ivec2(gl_FragCoord.xy), ivec2(PSin.rounduv.xy)));
+
+	// Get flags for whether to round U, V.
+	ivec2 round_flags = ivec2(PSin.rounduv.zw);
+
+	// Being on the top or left pixels converts round down to round up.
+	ivec2 round_down = ivec2(equal(round_flags, ivec2(PS_ROUND_UV_DOWN))) & ~topleft;
+	ivec2 round_up = ivec2(equal(round_flags, ivec2(PS_ROUND_UV_UP))) |
+	                 (ivec2(equal(round_flags, ivec2(PS_ROUND_UV_DOWN))) & topleft);
+
+	vec2 uv = PSin.t_int.zw; // Unnormalized UVs.
+	vec2 uvi = round(PSin.t_int.zw / 8.0f) * 8.0f; // Nearest half texel.
+	
+	// Round only if close to a half texel.
+	ivec2 close = ivec2(lessThanEqual(abs(uv - uvi), vec2(PS_ROUND_UV_THRESHOLD)));
+	round_down &= close;
+	round_up &= close;
+
+	uv = mix(uv, uvi - vec2(PS_ROUND_UV_THRESHOLD), bvec2(round_down));
+	uv = mix(uv, uvi + vec2(PS_ROUND_UV_THRESHOLD), bvec2(round_up));
+
+	return vec4(uv / 16.0f / WH.xy, uv); // Return normalized and unnormalized coords.
+#else
+	return vec4(0.0f);
+#endif
+}
+
 mat4 sample_4c(vec4 uv)
 {
 	mat4 c;
@@ -661,6 +696,10 @@ vec4 ps_color()
 #if (PS_FST == 0)
 	vec2 st = PSin.t_float.xy / vec2(PSin.t_float.w);
 	vec2 st_int = PSin.t_int.zw / vec2(PSin.t_float.w);
+#elif PS_ROUND_UV != 0
+	vec4 t_int_rounded = round_uv();
+	vec2 st = t_int_rounded.xy;
+	vec2 st_int = t_int_rounded.zw;
 #else
 	// Note xy are normalized coordinate
 	vec2 st = PSin.t_int.xy;

bin/resources/shaders/opengl/tfx_vgs.glsl

@@ -27,10 +27,29 @@ out SHADER
 	#else
 		flat vec4 c;
 	#endif
+	#if VS_ROUND_UV != 0
+		flat uvec4 rounduv;
+	#endif
 } VSout;
 
 const float exp_min32 = exp2(-32.0f);
 
+vec2 sign_extend_16_bit(vec2 uv)
+{
+	return mix(uv, uv - vec2(0x10000), greaterThan(uv, vec2(0x7FFF)));
+}
+
+uvec4 extract_round_uv_bits(float q)
+{
+	uint qi = floatBitsToUint(q);
+	return uvec4(
+		(qi >> 0) & 0xFFF,  // Prim left
+		(qi >> 12) & 0xFFF, // Prim top
+		(qi >> 24) & 0xF,   // Round U flags
+		(qi >> 28) & 0xF    // Round V flags
+	);
+}
+
 #if VS_EXPAND == 0
 
 layout(location = 0) in vec2  i_st;
@@ -43,7 +62,11 @@ layout(location = 7) in vec4  i_f;
 
 void texture_coord()
 {
-	vec2 uv = vec2(i_uv) - TextureOffset;
+	#if VS_ROUND_UV == 0
+		vec2 uv = vec2(i_uv) - TextureOffset;
+	#else
+		vec2 uv = sign_extend_16_bit(vec2(i_uv)) - TextureOffset; // Extend sign bit in case ST was converted to UV.
+	#endif
 	vec2 st = i_st - TextureOffset;
 
 	// Float coordinate
@@ -59,6 +82,11 @@ void texture_coord()
 	// Some games uses float coordinate for post-processing effect
 	VSout.t_int.zw = st / TextureScale;
 #endif
+
+// Get UV rounding info saved in Q.
+#if VS_ROUND_UV
+	VSout.rounduv = extract_round_uv_bits(i_q);
+#endif
 }
 
 void vs_main()
@@ -78,7 +106,7 @@ void vs_main()
 	texture_coord();
 
 	VSout.c = i_c;
-	VSout.t_float.z = i_f.x; // pack for with texture
+	VSout.t_float.z = i_f.x; // pack fog with texture
 
 	#if VS_POINT_SIZE
 		gl_PointSize = PointSize.x;
@@ -108,6 +136,9 @@ struct ProcessedVertex
 	vec4 t_float;
 	vec4 t_int;
 	vec4 c;
+#if VS_ROUND_UV
+	uvec4 rounduv;
+#endif
 };
 
 ProcessedVertex load_vertex(uint index)
@@ -131,7 +162,11 @@ ProcessedVertex load_vertex(uint index)
 	vtx.p.z = float(z) * exp_min32;
 	vtx.p.w = 1.0f;
 
-	vec2 uv = vec2(i_uv) - TextureOffset;
+	#if VS_ROUND_UV == 0
+		vec2 uv = vec2(i_uv) - TextureOffset;
+	#else
+		vec2 uv = sign_extend_16_bit(vec2(i_uv)) - TextureOffset; // Extend sign bit in case ST was converted to UV.
+	#endif
 	vec2 st = i_st - TextureOffset;
 
 	vtx.t_float.xy = st;
@@ -144,6 +179,11 @@ ProcessedVertex load_vertex(uint index)
 	vtx.t_int.zw = st / TextureScale;
 #endif
 
+// Get UV rounding info saved in Q.
+#if VS_ROUND_UV
+	vtx.rounduv = extract_round_uv_bits(i_q);
+#endif
+
 	vtx.c = i_c;
 	vtx.t_float.z = i_f.x;
 
@@ -210,6 +250,9 @@ void main()
 	VSout.t_float = vtx.t_float;
 	VSout.t_int = vtx.t_int;
 	VSout.c = vtx.c;
+#if VS_ROUND_UV
+	VSout.rounduv = vtx.rounduv;
+#endif
 }
 
 #endif // VS_EXPAND