nvshmem_add_perftest(shmem_atomic_bw.cu)
nvshmem_add_perftest(shmem_atomic_latency.cu)
nvshmem_add_perftest(shmem_atomic_ping_pong_latency.cu)
nvshmem_add_perftest(shmem_g_bw.cu)
nvshmem_add_perftest(shmem_g_latency.cu)
nvshmem_add_perftest(shmem_get_bw.cu)
nvshmem_add_perftest(shmem_get_latency.cu)
nvshmem_add_perftest(shmem_nbi_issue_rate.cu)
nvshmem_add_perftest(shmem_region_nbi_rate.cu)
nvshmem_add_perftest(shmem_p_bw.cu)
nvshmem_add_perftest(shmem_p_latency.cu)
nvshmem_add_perftest(shmem_p_ping_pong_latency.cu)
nvshmem_add_perftest(shmem_put_atomic_ping_pong_latency.cu)
nvshmem_add_perftest(shmem_put_bw.cu)
nvshmem_add_perftest(shmem_put_tma_bw.cu)
nvshmem_add_perftest(shmem_flush_bench.cu)
nvshmem_add_perftest(shmem_put_latency.cu)
nvshmem_add_perftest(shmem_put_ping_pong_latency.cu)
nvshmem_add_perftest(shmem_put_signal_ping_pong_latency.cu)
nvshmem_add_perftest(shmem_putmem_signal_ping_pong_latency.cu)
nvshmem_add_perftest(shmem_signal_ping_pong_latency.cu)
nvshmem_add_perftest(shmem_st_bw.cu)
if(NVSHMEM_CFT_HANDLES_SUPPORT)
    nvshmem_add_perftest(shmem_putmem_signal_counted_ping_pong_latency.cu)
endif()

if(CUDAToolkit_VERSION VERSION_GREATER_EQUAL "13.4")
    nvshmem_add_perftest(shmem_get_bw_locality_domains.cu)
    nvshmem_add_perftest(shmem_put_bw_locality_domains.cu)
endif()

if(NVSHMEM_BUILD_BITCODE_LIBRARY)
    nvshmem_add_cubin_perftest(shmem_atomic_bw.cu)
    nvshmem_add_cubin_perftest(shmem_atomic_latency.cu)
    nvshmem_add_cubin_perftest(shmem_atomic_ping_pong_latency.cu)
    nvshmem_add_cubin_perftest(shmem_g_latency.cu)
    nvshmem_add_cubin_perftest(shmem_get_latency.cu)
    nvshmem_add_cubin_perftest(shmem_p_latency.cu)
    nvshmem_add_cubin_perftest(shmem_p_ping_pong_latency.cu)
    nvshmem_add_cubin_perftest(shmem_put_atomic_ping_pong_latency.cu)
    nvshmem_add_cubin_perftest(shmem_put_latency.cu)
    nvshmem_add_cubin_perftest(shmem_put_ping_pong_latency.cu)
    nvshmem_add_cubin_perftest(shmem_put_signal_ping_pong_latency.cu)
    nvshmem_add_cubin_perftest(shmem_signal_ping_pong_latency.cu)
endif()

if(NVSHMEM_BUILD_LTOIR_LIBRARY)
    nvshmem_add_ltoir_cubin_perftest(shmem_atomic_bw.cu)
    nvshmem_add_ltoir_cubin_perftest(shmem_atomic_latency.cu)
    nvshmem_add_ltoir_cubin_perftest(shmem_atomic_ping_pong_latency.cu)
    nvshmem_add_ltoir_cubin_perftest(shmem_g_latency.cu)
    nvshmem_add_ltoir_cubin_perftest(shmem_get_latency.cu)
    nvshmem_add_ltoir_cubin_perftest(shmem_p_latency.cu)
    nvshmem_add_ltoir_cubin_perftest(shmem_p_ping_pong_latency.cu)
    nvshmem_add_ltoir_cubin_perftest(shmem_put_atomic_ping_pong_latency.cu)
    nvshmem_add_ltoir_cubin_perftest(shmem_put_latency.cu)
    nvshmem_add_ltoir_cubin_perftest(shmem_put_ping_pong_latency.cu)
    nvshmem_add_ltoir_cubin_perftest(shmem_put_signal_ping_pong_latency.cu)
    nvshmem_add_ltoir_cubin_perftest(shmem_signal_ping_pong_latency.cu)
endif()
