Block Hash Kernel

void trt_edgellm::rt::launchFnv1aHashChunksKernel(

uint8_t const *deviceData,

size_t totalSize,

int32_t numChunks,

int32_t numCtas,

uint64_t *partialsOut,

cudaStream_t stream

)

Launch a multi-CTA parallel FNV-1a Phase 1 kernel on device data.

The kernel splits the payload into numChunks equal chunks distributed across numCtas CTAs. Each thread hashes its chunk independently and writes the 128-bit partial digest to partialsOut as two consecutive uint64_t (hi, lo) per chunk. The caller must synchronize stream then perform the Phase 2 reduction on CPU.