VA: Use common timer and timing aspect headers

2 files changed, 168 insertions, 0 deletions

diff --git a/include/dfatool_host_dpu.ah b/include/dfatool_host_dpu.ah
new file mode 100644
index 0000000..1056a94
--- /dev/null
+++ b/include/dfatool_host_dpu.ah
@@ -0,0 +1,115 @@
+#pragma once
+
+#include <sys/time.h>
+
+aspect DfatoolHostDPUTiming {
+	struct timeval starttime;
+	struct timeval stoptime;
+	uint32_t n_ranks = 0;
+	uint32_t n_dpus = 0;
+
+	double const M_to_Mi = 1.048576; /* 2^20 / 1e6 */
+
+	advice call("% dpu_get_nr_dpus(...)") : after() {
+		n_dpus = **(tjp->arg<1>());
+	}
+
+	advice call("% dpu_get_nr_ranks(...)") : after() {
+		n_ranks = **(tjp->arg<1>());
+	}
+
+	advice call("% dpu_alloc(...)") : around() {
+		gettimeofday(&starttime, NULL);
+		tjp->proceed();
+		gettimeofday(&stoptime, NULL);
+		n_dpus = *(tjp->arg<0>());
+		printf("[::] dpu_alloc @ %s:%d | n_dpus=%u | latency_us=%f\n",
+			tjp->filename(),
+			tjp->line(),
+			n_dpus,
+			(stoptime.tv_sec - starttime.tv_sec) * 1000000.0 + (stoptime.tv_usec - starttime.tv_usec)
+		);
+	}
+
+	advice call("% dpu_alloc_ranks(...)") : around() {
+		gettimeofday(&starttime, NULL);
+		tjp->proceed();
+		gettimeofday(&stoptime, NULL);
+		n_ranks = *(tjp->arg<0>());
+		printf("[::] dpu_alloc_ranks @ %s:%d | n_ranks=%u | latency_us=%f\n",
+			tjp->filename(),
+			tjp->line(),
+			n_ranks,
+			(stoptime.tv_sec - starttime.tv_sec) * 1000000.0 + (stoptime.tv_usec - starttime.tv_usec)
+		);
+	}
+
+	advice call("% dpu_load(...)") : around() {
+		gettimeofday(&starttime, NULL);
+		tjp->proceed();
+		gettimeofday(&stoptime, NULL);
+		printf("[::] dpu_load @ %s:%d | n_dpus=%u n_ranks=%u | latency_us=%f\n",
+			tjp->filename(),
+			tjp->line(),
+			n_dpus, n_ranks,
+			(stoptime.tv_sec - starttime.tv_sec) * 1000000.0 + (stoptime.tv_usec - starttime.tv_usec)
+		);
+	}
+
+	advice call("% dpu_free(...)") : around() {
+		gettimeofday(&starttime, NULL);
+		tjp->proceed();
+		gettimeofday(&stoptime, NULL);
+		printf("[::] dpu_free @ %s:%d | n_dpus=%u n_ranks=%u | latency_us=%f\n",
+			tjp->filename(),
+			tjp->line(),
+			n_dpus, n_ranks,
+			(stoptime.tv_sec - starttime.tv_sec) * 1000000.0 + (stoptime.tv_usec - starttime.tv_usec)
+		);
+	}
+
+	advice call("% dpu_launch(...)") : around() {
+		gettimeofday(&starttime, NULL);
+		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",
+			tjp->filename(),
+			tjp->line(),
+			n_dpus, n_ranks,
+			kernel + 1,
+			input_size,
+			latency_us,
+			input_size / latency_us,
+			input_size * sizeof(T) / (latency_us * M_to_Mi)
+		);
+	}
+
+	advice call("% dpu_push_xfer(...)") : around() {
+		size_t payload_size = *(tjp->arg<4>());
+		gettimeofday(&starttime, NULL);
+		tjp->proceed();
+		gettimeofday(&stoptime, NULL);
+		double time_us = (stoptime.tv_sec - starttime.tv_sec) * 1000000.0 + (stoptime.tv_usec - starttime.tv_usec);
+		if (*(tjp->arg<1>()) == DPU_XFER_TO_DPU) {
+			printf("[::] dpu_push_to_dpu @ %s:%d | n_dpus=%u n_ranks=%u total_payload_B=%lu dpu_payload_B=%lu | latency_us=%f throughput_MiBps=%f\n",
+				tjp->filename(),
+				tjp->line(),
+				n_dpus, n_ranks,
+				payload_size * n_dpus, payload_size,
+				time_us,
+				payload_size * n_dpus / (time_us * M_to_Mi)
+			);
+		} else if (*(tjp->arg<1>()) == DPU_XFER_FROM_DPU) {
+			printf("[::] dpu_push_from_dpu @ %s:%d | n_dpus=%u n_ranks=%u total_payload_B=%lu dpu_payload_B=%lu | latency_us=%f throughput_MiBps=%f\n",
+				tjp->filename(),
+				tjp->line(),
+				n_dpus, n_ranks,
+				payload_size * n_dpus, payload_size,
+				time_us,
+				payload_size * n_dpus / (time_us * M_to_Mi)
+			);
+		}
+	}
+};
diff --git a/include/timer_base.h b/include/timer_base.h
new file mode 100644
index 0000000..fed8bec
--- /dev/null
+++ b/include/timer_base.h
@@ -0,0 +1,53 @@
+#pragma once
+
+#include <sys/time.h>
+
+#if DFATOOL_TIMING
+
+typedef struct Timer {
+
+	struct timeval startTime[N_TIMERS];
+	struct timeval stopTime[N_TIMERS];
+	double time[N_TIMERS];
+
+} Timer;
+
+#define dfatool_printf(fmt, ...) do { printf(fmt, __VA_ARGS__); } while (0)
+
+void start(Timer *timer, int i, int rep)
+{
+	if (rep == 0) {
+		timer->time[i] = 0.0;
+	}
+	gettimeofday(&timer->startTime[i], NULL);
+}
+
+void stop(Timer *timer, int i)
+{
+	gettimeofday(&timer->stopTime[i], NULL);
+	timer->time[i] +=
+	    (timer->stopTime[i].tv_sec -
+	     timer->startTime[i].tv_sec) * 1000000.0 +
+	    (timer->stopTime[i].tv_usec - timer->startTime[i].tv_usec);
+}
+
+#else
+
+#define dfatool_printf(fmt, ...) do {} while (0)
+
+typedef int Timer;
+
+void start(Timer *timer, int i, int rep)
+{
+	(void)timer;
+	(void)i;
+	(void)rep;
+}
+
+void stop(Timer *timer, int i)
+{
+	(void)timer;
+	(void)i;
+}
+
+#endif