Function constexpr

compute/device/compute_algorithm.hpp:201–221

can we fallback to a sub-group level implementation?

#if FLOOR_COMPUTE_INFO_HAS_SUB_GROUPS != 0
		// can we fallback to a sub-group level implementation?
		else if constexpr (group::supports_v<group::ALGORITHM::SUB_GROUP_REDUCE, group::OP::ADD, reduced_type>) {
			constexpr const uint32_t linear_work_group_size = compute_linear_work_group_size<work_group_size>();
			
			// first pass: inclusive scan in each sub-group
			const auto sub_block_red_val = group::sub_group_reduce<group::OP::ADD>(work_item_value);
			// first sub-group item writes its result into local memory for the second pass
			if (sub_group_local_id == 0u) {
				lmem[sub_group_id_1d] = sub_block_red_val;
			}
			local_barrier();
			
			// second pass: reduction of the partial sums in each sub-group to compute the total sum, executed in the first sub-group
			reduced_type total_sum {};
			if (sub_group_id_1d == 0u) {
				// NOTE: we need to consider that the executing work-group size may be smaller than "sub_group_size * sub_group_size"
				const auto sg_in_val = (sub_group_local_id < (linear_work_group_size / sub_group_size) ? lmem[sub_group_local_id] : reduced_type(0));
				total_sum = group::sub_group_reduce<group::OP::ADD>(sg_in_val);
			}
			local_barrier();
			return total_sum;
		}
#endif
		return reduce<work_group_size>(work_item_value, lmem, plus<reduced_type> {});
	}