rsx: Fix vertex input validation failure

- We switched from a allocate-per-draw-call to reusing the vertex input data. We need to make sure previous state does not affect validation results.
- Memcpy is too slow at this level of frequency, so just reset the attributes you're using before using them.
- Also changes the validation routine to only check attributes we're actually using.

rpcs3/Emu/RSX/Core/RSXVertexTypes.h

@@ -86,6 +86,7 @@ namespace rsx
 		rsx::simple_array<interleaved_range_info*> interleaved_blocks{};  // Interleaved blocks to be uploaded as-is
 		std::vector<std::pair<u8, u32>> volatile_blocks{};                // Volatile data blocks (immediate draw vertex data for example)
 		rsx::simple_array<u8> referenced_registers{};                     // Volatile register data
+		u16 attribute_mask = 0;                                           // ATTRn mask
 
 		std::array<attribute_buffer_placement, 16> attribute_placement = fill_array(attribute_buffer_placement::none);
 
@@ -108,6 +109,7 @@ namespace rsx
 		void clear()
 		{
 			m_num_used_blocks = 0;
+			attribute_mask = 0;
 			interleaved_blocks.clear();
 			volatile_blocks.clear();
 			referenced_registers.clear();
@@ -117,15 +119,20 @@ namespace rsx
 		{
 			// Criteria: At least one array stream has to be defined to feed vertex positions
 			// This stream cannot be a const register as the vertices cannot create a zero-area primitive
-
 			if (!interleaved_blocks.empty() && interleaved_blocks[0]->attribute_stride != 0)
 				return true;
 
 			if (!volatile_blocks.empty())
 				return true;
 
-			for (u8 index = 0; index < limits::vertex_count; ++index)
+			for (u16 ref_mask = attribute_mask, index = 0; ref_mask; ++index, ref_mask >>= 1)
 			{
+				if (!(ref_mask & 1))
+				{
+					// Disabled
+					continue;
+				}
+
 				switch (attribute_placement[index])
 				{
 				case attribute_buffer_placement::transient:

rpcs3/Emu/RSX/RSXThread.cpp

@@ -2210,6 +2210,7 @@ namespace rsx
 		const u32 input_mask = state.vertex_attrib_input_mask() & current_vp_metadata.referenced_inputs_mask;
 
 		result.clear();
+		result.attribute_mask = static_cast<u16>(input_mask);
 
 		if (state.current_draw_clause.command == rsx::draw_command::inlined_array)
 		{
@@ -2219,6 +2220,7 @@ namespace rsx
 			for (u8 index = 0; index < rsx::limits::vertex_count; ++index)
 			{
 				auto &vinfo = state.vertex_arrays_info[index];
+				result.attribute_placement[index] = attribute_buffer_placement::none;
 
 				if (vinfo.size() > 0)
 				{
@@ -2262,8 +2264,10 @@ namespace rsx
 				continue;
 			}
 
+			// Always reset attribute placement by default
+			result.attribute_placement[index] = attribute_buffer_placement::none;
+
 			// Check for interleaving
-			const auto &info = state.vertex_arrays_info[index];
 			if (rsx::method_registers.current_draw_clause.is_immediate_draw &&
 				rsx::method_registers.current_draw_clause.command != rsx::draw_command::indexed)
 			{
@@ -2290,6 +2294,7 @@ namespace rsx
 				continue;
 			}
 
+			const auto& info = state.vertex_arrays_info[index];
 			if (!info.size())
 			{
 				if (state.register_vertex_info[index].size > 0)