rsx: Implement proper pitch compatibility lookup

- When a single row is required or is all that is available, pitch has
no meaning as the coordinate space changed to 1D

rpcs3/Emu/RSX/Common/surface_store.h

@@ -997,7 +997,7 @@ namespace rsx
 						continue;
 
 					surface = std::get<1>(tex_info).get();
-					if (surface->get_rsx_pitch() != requested_pitch)
+					if (!rsx::pitch_compatible(surface, requested_pitch, requested_height))
 						continue;
 
 					if (requested_width == 0 || requested_height == 0)
@@ -1023,7 +1023,7 @@ namespace rsx
 						continue;
 
 					surface = std::get<1>(tex_info).get();
-					if (surface->get_rsx_pitch() != requested_pitch)
+					if (!rsx::pitch_compatible(surface, requested_pitch, requested_height))
 						continue;
 
 					if (requested_width == 0 || requested_height == 0)
@@ -1059,7 +1059,7 @@ namespace rsx
 
 					auto surface = std::get<1>(tex_info).get();
 					const auto pitch = surface->get_rsx_pitch();
-					if (pitch != required_pitch)
+					if (!rsx::pitch_compatible(surface, required_pitch, required_height))
 						continue;
 
 					const auto texture_size = pitch * surface->get_surface_height();

rpcs3/Emu/RSX/Common/texture_cache.h

@@ -1040,7 +1040,7 @@ namespace rsx
 		}
 
 		template <bool check_unlocked = false>
-		std::vector<section_storage_type*> find_texture_from_range(const address_range &test_range, u16 required_pitch = 0, u32 context_mask=0xFF)
+		std::vector<section_storage_type*> find_texture_from_range(const address_range &test_range, u16 required_pitch = 0, u32 context_mask = 0xFF)
 		{
 			std::vector<section_storage_type*> results;
 
@@ -1060,7 +1060,7 @@ namespace rsx
 							continue;
 					}
 
-					if (required_pitch && tex.get_rsx_pitch() != required_pitch)
+					if (required_pitch && !rsx::pitch_compatible<false>(&tex, required_pitch, UINT16_MAX))
 					{
 						continue;
 					}
@@ -1575,8 +1575,7 @@ namespace rsx
 
 				for (u32 index = 0; index < local.size(); ++index)
 				{
-					if (local[index]->get_rsx_pitch() != pitch ||
-						local[index]->get_context() != rsx::texture_upload_context::blit_engine_dst)
+					if (local[index]->get_context() != rsx::texture_upload_context::blit_engine_dst)
 						continue;
 
 					sort_list.push_back({ local[index]->last_write_tag, 1, index });
@@ -1742,9 +1741,6 @@ namespace rsx
 				{
 					for (auto &section : local)
 					{
-						if (section->get_rsx_pitch() != pitch)
-							continue;
-
 						add_local_resource(section, current_address, slice, false);
 					}
 				}
@@ -1791,7 +1787,7 @@ namespace rsx
 			u16 tex_width, u16 tex_height, u16 tex_depth, u16 tex_pitch,
 			rsx::texture_dimension_extended extended_dimension)
 		{
-			if (texptr->get_rsx_pitch() != tex_pitch)
+			if (!rsx::pitch_compatible(texptr, tex_pitch, tex_height))
 			{
 				return false;
 			}
@@ -2008,7 +2004,7 @@ namespace rsx
 			// Check shader_read storage. In a given scene, reads from local memory far outnumber reads from the surface cache
 			const u32 lookup_mask = (is_compressed_format)? rsx::texture_upload_context::shader_read :
 				rsx::texture_upload_context::shader_read | rsx::texture_upload_context::blit_engine_dst | rsx::texture_upload_context::blit_engine_src;
-			const auto overlapping_locals = find_texture_from_range<true>(tex_range, tex_pitch, lookup_mask);
+			const auto overlapping_locals = find_texture_from_range<true>(tex_range, tex_height > 1? tex_pitch : 0, lookup_mask);
 
 			for (auto& cached_texture : overlapping_locals)
 			{

rpcs3/Emu/RSX/GL/GLGSRender.cpp

@@ -216,7 +216,7 @@ void GLGSRender::end()
 	// Handle special memory barrier for ARGB8->D24S8 in an active DSV
 	if (ds && ds->old_contents != nullptr &&
 		ds->old_contents->get_internal_format() == gl::texture::internal_format::rgba8 &&
-		ds->get_rsx_pitch() == static_cast<gl::render_target*>(ds->old_contents)->get_rsx_pitch())
+		rsx::pitch_compatible(ds, static_cast<gl::render_target*>(ds->old_contents)))
 	{
 		gl_state.enable(GL_FALSE, GL_SCISSOR_TEST);
 

rpcs3/Emu/RSX/GL/GLRenderTargets.cpp

@@ -603,7 +603,7 @@ void gl::render_target::memory_barrier(gl::command_context& cmd, bool force_init
 	}
 
 	auto src_texture = static_cast<gl::render_target*>(old_contents);
-	if (src_texture->get_rsx_pitch() != get_rsx_pitch())
+	if (!rsx::pitch_compatible(this, src_texture))
 	{
 		LOG_TRACE(RSX, "Pitch mismatch, could not transfer inherited memory");
 		return;

rpcs3/Emu/RSX/VK/VKGSRender.cpp

@@ -1436,7 +1436,7 @@ void VKGSRender::end()
 	auto ds = std::get<1>(m_rtts.m_bound_depth_stencil);
 	if (ds && ds->old_contents &&
 		ds->old_contents->info.format == VK_FORMAT_B8G8R8A8_UNORM &&
-		ds->get_rsx_pitch() == static_cast<vk::render_target*>(ds->old_contents)->get_rsx_pitch())
+		rsx::pitch_compatible(ds, static_cast<vk::render_target*>(ds->old_contents)))
 	{
 		auto rp = vk::get_render_pass_location(VK_FORMAT_UNDEFINED, ds->info.format, 0);
 		auto render_pass = m_render_passes[rp];

rpcs3/Emu/RSX/VK/VKRenderTargets.h

@@ -100,7 +100,7 @@ namespace vk
 			}
 
 			auto src_texture = static_cast<vk::render_target*>(old_contents);
-			if (src_texture->get_rsx_pitch() != get_rsx_pitch())
+			if (!rsx::pitch_compatible(this, src_texture))
 			{
 				LOG_TRACE(RSX, "Pitch mismatch, could not transfer inherited memory");
 				return;

rpcs3/Emu/RSX/rsx_utils.h

@@ -544,6 +544,32 @@ namespace rsx
 		return std::make_tuple(u16(dst_w / scale_x), u16(dst_h / scale_y), dst_w, dst_h);
 	}
 
+	template <typename SurfaceType>
+	inline bool pitch_compatible(SurfaceType* a, SurfaceType* b)
+	{
+		if (a->get_surface_height() == 1 || b->get_surface_height() == 1)
+			return true;
+
+		return (a->get_rsx_pitch() == b->get_rsx_pitch());
+	}
+
+	template <bool __is_surface = true, typename SurfaceType>
+	inline bool pitch_compatible(SurfaceType* surface, u16 pitch_required, u16 height_required)
+	{
+		if constexpr (__is_surface)
+		{
+			if (height_required == 1 || surface->get_surface_height() == 1)
+				return true;
+		}
+		else
+		{
+			if (height_required == 1 || surface->get_height() == 1)
+				return true;
+		}
+
+		return (surface->get_rsx_pitch() == pitch_required);
+	}
+
 	/**
 	 * Remove restart index and emulate using degenerate triangles
 	 * Can be used as a workaround when restart_index doesnt work too well