Question about atomic_work_item_fence in work_group_barrier

[misos1]

Why is used memory_order_seq_cst instead of memory_order_release/memory_order_acquire when both flags CLK_GLOBAL_MEM_FENCE and CLK_LOCAL_MEM_FENCE are present? Is not release resp. acquire enough when both address spaces are specified?

llvm-project/amd/device-libs/opencl/src/workgroup/wgbarrier.cl

Lines 25 to 26 in 6581f3d

	flags == (CLK_GLOBAL_MEM_FENCE|CLK_LOCAL_MEM_FENCE) ?
	memory_order_seq_cst : memory_order_release,

llvm-project/amd/device-libs/opencl/src/workgroup/wgbarrier.cl

Lines 32 to 33 in 6581f3d

	flags == (CLK_GLOBAL_MEM_FENCE|CLK_LOCAL_MEM_FENCE) ?
	memory_order_seq_cst : memory_order_acquire,

And then there is also CLK_IMAGE_MEM_FENCE which is treated in the same way as CLK_GLOBAL_MEM_FENCE in atomic_work_item_fence. But with it work_group_barrier will no longer use memory_order_seq_cst even when all 3 are specified (seq_cst strangely no longer needed in that case?).

llvm-project/amd/device-libs/opencl/src/misc/awif.cl

Lines 82 to 84 in 6581f3d

	// global or image is set, but not local -> fence only global memory.
	if ((flags & CLK_LOCAL_MEM_FENCE) == 0) {
	IMPL_ATOMIC_WORK_ITEM_FENCE(, "global");

[ppanchad-amd]

Hi @misos1. Internal ticket has been created to assist with your issue. Thanks!