diff options
| author | Birte Kristina Friesel <derf@finalrewind.org> | 2025-05-13 13:45:22 +0200 |
|---|---|---|
| committer | Birte Kristina Friesel <derf@finalrewind.org> | 2025-05-13 13:45:22 +0200 |
| commit | 05001600966baeeba4ef42ddcf1804975fe88625 (patch) | |
| tree | eff58e7f37a9bdec030e69e00ebd5b227d2cb1fc | |
| parent | 591e195f3eeb06237dedd9c5a66fcfe0b0a10889 (diff) | |
tracing: provide input_size and element_size in derived aspect
| -rw-r--r-- | VA/include/dfatool_host.ah | 14 | ||||
| -rw-r--r-- | include/dfatool_host_dpu.ah | 8 |
2 files changed, 17 insertions, 5 deletions
diff --git a/VA/include/dfatool_host.ah b/VA/include/dfatool_host.ah index 6818333..7b33cec 100644 --- a/VA/include/dfatool_host.ah +++ b/VA/include/dfatool_host.ah @@ -4,12 +4,22 @@ #include "dfatool_host_dpu.ah" aspect DfatoolHostTiming : public DfatoolHostDPUTiming { + + DfatoolHostTiming() { + element_size = sizeof(T); + } + advice call("% input_params(...)") : after() { Params* p = tjp->result(); - printf("[==] VA | n_dpus=%u n_elements=%lu e_exp=%d \n", NR_DPUS, p->input_size, p->exp); + printf("[>>] VA | n_dpus=%u n_elements=%lu e_exp=%d\n", NR_DPUS, p->input_size, p->exp); + input_size = p->input_size; } advice call("% vector_addition_host(...)") : after() { - printf("[--] VA | n_dpus=%u n_ranks=%u n_elements=%lu n_elements_per_dpu=%lu e_exp=%d \n", n_dpus, n_ranks, p.input_size, p.input_size / n_dpus, p.exp); + printf("[--] VA | n_dpus=%u n_ranks=%u n_elements=%lu e_exp=%d\n", n_dpus, n_ranks, p.input_size, p.exp); + } + + advice execution("% main(...)") : after() { + printf("[<<] VA | n_dpus=%u n_elements=%lu e_exp=%d\n", NR_DPUS, p.input_size, p.exp); } }; diff --git a/include/dfatool_host_dpu.ah b/include/dfatool_host_dpu.ah index 1056a94..560e410 100644 --- a/include/dfatool_host_dpu.ah +++ b/include/dfatool_host_dpu.ah @@ -5,6 +5,8 @@ aspect DfatoolHostDPUTiming { struct timeval starttime; struct timeval stoptime; + unsigned long input_size; + unsigned int element_size; uint32_t n_ranks = 0; uint32_t n_dpus = 0; @@ -73,16 +75,16 @@ aspect DfatoolHostDPUTiming { tjp->proceed(); gettimeofday(&stoptime, NULL); double latency_us = (stoptime.tv_sec - starttime.tv_sec) * 1000000.0 + (stoptime.tv_usec - starttime.tv_usec); - unsigned long input_size = p.input_size; - printf("[::] dpu_launch @ %s:%d | n_dpus=%u n_ranks=%u e_kernel=kernel%d n_elements=%lu | latency_us=%f throughput_Mrps=%f throughput_MiBps=%f\n", + printf("[::] dpu_launch @ %s:%d | n_dpus=%u n_ranks=%u e_kernel=kernel%d n_elements=%lu n_elements_per_dpu=%lu | latency_us=%f throughput_Mps=%f throughput_MiBps=%f\n", tjp->filename(), tjp->line(), n_dpus, n_ranks, kernel + 1, input_size, + input_size / n_dpus, latency_us, input_size / latency_us, - input_size * sizeof(T) / (latency_us * M_to_Mi) + input_size * element_size / (latency_us * M_to_Mi) ); } |