GS/HW: Fix Z Tex in RT regions + read back sources for SW if needed

pcsx2/GS/Renderers/HW/GSRendererHW.cpp

@@ -504,12 +504,12 @@ void GSRendererHW::ConvertSpriteTextureShuffle(u32& process_rg, u32& process_ba,
 
 				const int width_diff = static_cast<int>(m_env.CTXT[m_env.PRIM.CTXT].TEX0.TBW) - static_cast<int>((m_cached_ctx.FRAME.FBW + 1) >> 1);
 				// We can check the future for a clue as this can be more accurate, be careful of different draws like channel shuffles or single page draws.
-				if (m_env.CTXT[m_env.PRIM.CTXT].TEX0.TBP0 == m_cached_ctx.FRAME.Block() && GSLocalMemory::m_psm[m_env.CTXT[m_env.PRIM.CTXT].TEX0.PSM].bpp == 32 && width_diff >= 0)
+				if (m_env.CTXT[m_env.PRIM.CTXT].TEX0.TBP0 == m_cached_ctx.FRAME.Block() && GSLocalMemory::m_psm[m_env.CTXT[m_env.PRIM.CTXT].TEX0.PSM].bpp > 16 && GSLocalMemory::m_psm[m_env.CTXT[m_env.PRIM.CTXT].TEX0.PSM].bpp == 32 && width_diff >= 0)
 				{
 					// width_diff will be zero is both are BW == 1, so be careful of that.
 					const bool same_width = width_diff > 0 || (m_cached_ctx.FRAME.FBW == 1 && width_diff == 0);
 					// Draw is double width and the draw is twice the width of the next draws texture.
-					if (!same_width && max_tex_draw_width >= (m_cached_ctx.FRAME.FBW * 64))
+					if ((!same_width && max_tex_draw_width >= (m_cached_ctx.FRAME.FBW * 64)) || (single_direction_doubled && (m_vt.m_max.p.x >= (rt->m_valid.z * 2))))
 					{
 						half_bottom_uv = false;
 						half_bottom_vert = false;
@@ -7383,6 +7383,7 @@ bool GSRendererHW::CanUseSwPrimRender(bool no_rt, bool no_ds, bool draw_sprite_t
 			if ((req_color && !src_target->m_valid_rgb) || (req_alpha && (!src_target->m_valid_alpha_low || !src_target->m_valid_alpha_high)))
 				return true;
 
+			bool req_readback = false;
 			// If the EE has written over our sample area, we're fine to do this on the CPU, despite the target.
 			if (!src_target->m_dirty.empty())
 			{
@@ -7408,11 +7409,22 @@ bool GSRendererHW::CanUseSwPrimRender(bool no_rt, bool no_ds, bool draw_sprite_t
 					}
 				}
 			}
+			else
+			{
+				// If the target isn't dirty we might have valid data, so let's check their areas overlap, if so we need to read it back for SW.
+				GSVector4i src_rect = GSVector4i(m_vt.m_min.t.x, m_vt.m_min.t.y, m_vt.m_max.t.x, m_vt.m_max.t.x);
+				GSVector4i area = g_texture_cache->TranslateAlignedRectByPage(src_target, m_cached_ctx.TEX0.TBP0, m_cached_ctx.TEX0.PSM, m_cached_ctx.TEX0.TBW, src_rect, false);
+				req_readback = !area.rintersect(src_target->m_drawn_since_read).eq(GSVector4i::zero());
+			}
 			// Make sure it actually makes sense to use this target as a source, given the formats, and it wouldn't just sample as garbage.
 			// We can't rely exclusively on the dirty rect check above, because sometimes the targets are from older frames and too large.
 			if (!GSUtil::HasSameSwizzleBits(m_cached_ctx.TEX0.PSM, src_target->m_TEX0.PSM) &&
 				(!src_target->m_32_bits_fmt || GSLocalMemory::m_psm[m_cached_ctx.TEX0.PSM].bpp != 16))
-					return true;
+			{
+				if (req_readback)
+					g_texture_cache->Read(src_target, src_target->m_drawn_since_read);
+				return true;
+			}
 
 			return false;
 		}

pcsx2/GS/Renderers/HW/GSTextureCache.cpp

@@ -1160,8 +1160,8 @@ GSTextureCache::Source* GSTextureCache::LookupDepthSource(const bool is_depth, c
 		if (inside_target)
 		{
 			// Need to set it up as a region target.
-			src->m_region.SetX(block_boundary_rect.x, block_boundary_rect.z);
-			src->m_region.SetY(block_boundary_rect.y, block_boundary_rect.w);
+			src->m_region.SetX(block_boundary_rect.x, src->m_from_target->m_valid.z);
+			src->m_region.SetY(block_boundary_rect.y, src->m_from_target->m_valid.w);
 		}
 
 		if (GSRendererHW::GetInstance()->IsTBPFrameOrZ(dst->m_TEX0.TBP0))