diff --git a/rpcs3/Emu/Cell/lv2/sys_rsx.cpp b/rpcs3/Emu/Cell/lv2/sys_rsx.cpp
index 707524af8d..1e85acafcf 100644
--- a/rpcs3/Emu/Cell/lv2/sys_rsx.cpp
+++ b/rpcs3/Emu/Cell/lv2/sys_rsx.cpp
@@ -260,6 +260,7 @@ error_code sys_rsx_context_allocate(vm::ptr<u32> context_id, vm::ptr<u64> lpar_d
 	render->current_display_buffer = 0;
 	render->label_addr = vm::cast(*lpar_reports, HERE);
 	render->device_addr = rsx_cfg->device_addr;
+	render->dma_address = rsx_cfg->dma_address;
 	render->local_mem_size = rsx_cfg->memory_size;
 	render->init(vm::cast(*lpar_dma_control, HERE));
 
diff --git a/rpcs3/Emu/RSX/RSXThread.h b/rpcs3/Emu/RSX/RSXThread.h
index 595601b66a..c6ee64a2f0 100644
--- a/rpcs3/Emu/RSX/RSXThread.h
+++ b/rpcs3/Emu/RSX/RSXThread.h
@@ -618,6 +618,7 @@ namespace rsx
 
 	public:
 		RsxDmaControl* ctrl = nullptr;
+		u32 dma_address{0};
 		rsx_iomap_table iomap_table;
 		u32 restore_point = 0;
 		atomic_t<u32> external_interrupt_lock{ 0 };
diff --git a/rpcs3/Emu/RSX/rsx_methods.cpp b/rpcs3/Emu/RSX/rsx_methods.cpp
index 8bfdbe2d23..31ac692104 100644
--- a/rpcs3/Emu/RSX/rsx_methods.cpp
+++ b/rpcs3/Emu/RSX/rsx_methods.cpp
@@ -45,7 +45,9 @@ namespace rsx
 		void set_reference(thread* rsx, u32 _reg, u32 arg)
 		{
 			rsx->sync();
-			rsx->ctrl->ref.exchange(arg);
+
+			// Write ref+get atomically (get will be written again with the same value at command end)
+			vm::_ref<atomic_be_t<u64>>(rsx->dma_address + ::offset32(&RsxDmaControl::get)).store(u64{rsx->fifo_ctrl->get_pos()} << 32 | arg);
 		}
 
 		void semaphore_acquire(thread* rsx, u32 /*_reg*/, u32 arg)