Code for new UV rounding method to grab enough of texture.

pcsx2/GS/GSState.cpp

@@ -3842,6 +3842,7 @@ GSState::TextureMinMaxResult GSState::GetTextureMinMax(GIFRegTEX0 TEX0, GIFRegCL
 	{
 		// Optimisation aims to reduce the amount of texture loaded to only the bit which will be read
 		GSVector4 st = m_vt.m_min.t.xyxy(m_vt.m_max.t);
+		// TODO: remove this if we use the new method
 		if (linear)
 		{
 			st += GSVector4(-0.5f, 0.5f).xxyy();
@@ -3856,7 +3857,19 @@ GSState::TextureMinMaxResult GSState::GetTextureMinMax(GIFRegTEX0 TEX0, GIFRegCL
 		// draw will get scissored, adjust UVs to suit
 		const GSVector2 pos_range(std::max(m_vt.m_max.p.x - m_vt.m_min.p.x, 1.0f), std::max(m_vt.m_max.p.y - m_vt.m_min.p.y, 1.0f));
 		const GSVector2 uv_range(m_vt.m_max.t.x - m_vt.m_min.t.x, m_vt.m_max.t.y - m_vt.m_min.t.y);
+
+		GSVector4 p_min_i = m_vt.m_min.p.ceil();
+		GSVector4 p_max_i = m_vt.m_max.p.floor();
+
 		const GSVector2 grad(uv_range / pos_range);
+
+		// Linearly interpolate the UV coordinates at the mimum/maximum pixel centers in the bounding box
+		GSVector4 t_min = ((m_vt.m_max.p - p_min_i) * m_vt.m_min.t + (p_min_i - m_vt.m_min.p) * m_vt.m_max.t) / (m_vt.m_max.p - m_vt.m_min.p);
+		GSVector4 t_max = ((m_vt.m_max.p - p_max_i) * m_vt.m_min.t + (p_max_i - m_vt.m_min.p) * m_vt.m_max.t) / (m_vt.m_max.p - m_vt.m_min.p);
+		// Calculate the min/max UV coordunates for bilinear filtering
+		// Note: Should we be doing max(GSVector4.zero()) as above? Or will this be handled by the scissor?
+		t_min = (t_min - 0.5).floor().max(GSVector4.zero());
+		t_max = (t_max + 0.5).floor() + 1;
 		// Adjust texture range when sprites get scissor clipped. Since we linearly interpolate, this
 		// optimization doesn't work when perspective correction is enabled.
 		if (m_vt.m_primclass == GS_SPRITE_CLASS && PRIM->FST == 1 && m_primitive_covers_without_gaps != NoGapsType::GapsFound)
@@ -3924,14 +3937,18 @@ GSState::TextureMinMaxResult GSState::GetTextureMinMax(GIFRegTEX0 TEX0, GIFRegCL
 			}
 		}
 
-		const GSVector4i uv = GSVector4i(st.floor());
+		const GSVector4i uv = GSVector4i(t_min.xyxy(t_max));
 		uses_border = GSVector4::cast((uv < vr).blend32<0xc>(uv >= vr)).mask();
 
 		// Need to make sure we don't oversample, this can cause trouble in grabbing textures.
 		// This may be inaccurate depending on the draw, but adding 1 all the time is wrong too.
-		const int inclusive_x_req = ((m_vt.m_primclass < GS_TRIANGLE_CLASS) || (grad.x < 1.0f || (grad.x == 1.0f && m_vt.m_max.p.x != floor(m_vt.m_max.p.x)))) ? 1 : 0;
-		const int inclusive_y_req = ((m_vt.m_primclass < GS_TRIANGLE_CLASS) || (grad.y < 1.0f || (grad.y == 1.0f && m_vt.m_max.p.y != floor(m_vt.m_max.p.y)))) ? 1 : 0;
-
+		//const int inclusive_x_req = ((m_vt.m_primclass < GS_TRIANGLE_CLASS) || (grad.x < 1.0f || (grad.x == 1.0f && m_vt.m_max.p.x != floor(m_vt.m_max.p.x)))) ? 1 : 0;
+		//const int inclusive_y_req = ((m_vt.m_primclass < GS_TRIANGLE_CLASS) || (grad.y < 1.0f || (grad.y == 1.0f && m_vt.m_max.p.y != floor(m_vt.m_max.p.y)))) ? 1 : 0;
+		// Disable the current inclusive_x/y_req calculation for now
+		// TODO: Remove if we use the new method
+		const int inclusive_x_req = 0;
+		const int inclusive_y_req = 0;
+
 		// Roughly cut out the min/max of the read (Clamp)
 		switch (wms)
 		{