How subgroupPartitionedAddNV Works in GLSL

The subgroupPartitionedAddNV intrinsic in GLSL enables invocations within a subgroup to perform parallel addition reductions across dynamically formed subsets of execution lanes without writing to shared or global memory. Provided by the GL_NV_shader_subgroup_partitioned extension, this function solves the performance bottleneck of aggregating data across arbitrary runtime categories by using fast, register-level intra-warp communication instead of serialized atomic operations.

In standard GLSL subgroup operations like subgroupAdd, reduction operations automatically span either the entire active subgroup or a contiguous range of lanes. In real-time rendering and compute tasks—such as binning particles, clustering light sources, or aggregating material attributes—lanes in a subgroup rarely belong to the same category. Without partitioned operations, threads sharing an identical property (such as a spatial cell index or material ID) must serialize their accumulations using atomic adds in shared memory or global memory, leading to severe hardware contention.

To aggregate values over dynamically clustered lanes, the algorithm first identifies which invocations belong to the same cluster. This clustering is represented as a partition mask, typically generated using the subgroupPartitionNV(key) function. In this step, each thread passes a runtime value—such as an integer hash or identifier—into subgroupPartitionNV. The hardware evaluates the inputs across all active lanes and returns a 4-component 32-bit unsigned integer vector (uvec4) acting as a bitmask (ballot). In this ballot, bits are set for all lanes within the subgroup that provided the exact same key.

Once the partition ballot is established, each thread calls subgroupPartitionedAddNV(value, ballot). The function takes the local numerical value to be aggregated and the uvec4 ballot representing that thread's dynamic cluster.

The hardware uses the provided ballot to logically segment the subgroup into isolated execution partitions. Inside each partition:

1. The hardware filters out lanes whose corresponding bits in the caller's ballot are zero, ignoring their values completely.
2. An intra-warp reduction network executes a sum across only the lanes marked active within that ballot.
3. Depending on whether an inclusive scan or an aggregate sum is specified, each participating lane receives the computed sum of all values submitted by the threads belonging to its dynamic partition.

Because subgroupPartitionedAddNV executes directly across the GPU's register file and warp shuffle data paths, it completely bypasses the L1/L2 caches and threadgroup shared memory (TGSM). Multiple distinct partitions execute their additions simultaneously in parallel without stalling or serializing one another, yielding high throughput even when keys are randomly distributed across subgroup lanes.