4 files changed, 137 insertions, 13 deletions

diff --git a/VA/support/common.h b/VA/support/common.h
index cee09e2..6ce6e23 100755
--- a/VA/support/common.h
+++ b/VA/support/common.h
@@ -1,14 +1,20 @@
 #ifndef _COMMON_H_
 #define _COMMON_H_
 
+enum kernels {
+	kernel1 = 0,
+	nr_kernels = 1,
+};
+
 // Structures used by both the host and the dpu to communicate information
 typedef struct {
+	/*
+	 * Size per DPU cannot exceed 32 bit, as each DPU only has 64 MiB of memory
+	 * (i.e., only needs 26 bit for addressing).
+	 */
 	uint32_t size;
 	uint32_t transfer_size;
-	enum kernels {
-		kernel1 = 0,
-		nr_kernels = 1,
-	} kernel;
+	enum kernels kernel;
 } dpu_arguments_t;
 
 // Transfer size between MRAM and WRAM
diff --git a/VA/support/dfatool_host.ah b/VA/support/dfatool_host.ah
new file mode 100644
index 0000000..f207ea2
--- /dev/null
+++ b/VA/support/dfatool_host.ah
@@ -0,0 +1,106 @@
+#pragma once
+
+#include <sys/time.h>
+
+aspect DfatoolHostTiming {
+	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("% input_params(...)") : after() {
+		Params* p = tjp->result();
+		printf("[--] n_dpus=%u n_elements=%lu e_exp=%d \n", NR_DPUS, p->input_size, p->exp);
+	}
+
+	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 | n_dpus=%u | latency_us=%f\n",
+			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 | n_ranks=%u | latency_us=%f\n",
+			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 | n_dpus=%u n_ranks=%u | latency_us=%f\n",
+			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 | n_dpus=%u n_ranks=%u | latency_us=%f\n",
+			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 | n_dpus=%u n_ranks=%u e_kernel=kernel%d n_elements=%lu | latency_us=%f throughput_Mrps=%f throughput_MiBps=%f\n",
+			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 | n_dpus=%u n_ranks=%u total_payload_B=%lu dpu_payload_B=%lu | latency_us=%f throughput_MiBps=%f\n",
+				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 | n_dpus=%u n_ranks=%u total_payload_B=%lu dpu_payload_B=%lu | latency_us=%f throughput_MiBps=%f\n",
+				n_dpus, n_ranks,
+				payload_size * n_dpus, payload_size,
+				time_us,
+				payload_size * n_dpus / (time_us * M_to_Mi)
+			);
+		}
+	}
+};
diff --git a/VA/support/params.h b/VA/support/params.h
index 47c10ef..4f41adc 100644
--- a/VA/support/params.h
+++ b/VA/support/params.h
@@ -4,7 +4,7 @@
 #include "common.h"
 
 typedef struct Params {
-	unsigned int input_size;
+	unsigned long int input_size;
 	int n_warmup;
 	int n_reps;
 	int exp;
@@ -41,7 +41,7 @@ struct Params input_params(int argc, char **argv)
 			exit(0);
 			break;
 		case 'i':
-			p.input_size = atoi(optarg);
+			p.input_size = atol(optarg);
 			break;
 		case 'w':
 			p.n_warmup = atoi(optarg);
diff --git a/VA/support/timer.h b/VA/support/timer.h
index df68334..c0208d3 100755
--- a/VA/support/timer.h
+++ b/VA/support/timer.h
@@ -35,6 +35,8 @@
 
 #include <sys/time.h>
 
+#if DFATOOL_TIMING
+
 typedef struct Timer {
 
 	struct timeval startTime[7];
@@ -43,6 +45,8 @@ typedef struct Timer {
 
 } Timer;
 
+#define dfatool_printf(fmt, ...) do { printf(fmt, __VA_ARGS__); } while (0)
+
 void start(Timer *timer, int i, int rep)
 {
 	if (rep == 0) {
@@ -60,15 +64,23 @@ void stop(Timer *timer, int i)
 	    (timer->stopTime[i].tv_usec - timer->startTime[i].tv_usec);
 }
 
-void print(Timer *timer, int i, int REP)
+#else
+
+#define dfatool_printf(fmt, ...) do {} while (0)
+
+typedef int Timer;
+
+void start(Timer *timer, int i, int rep)
 {
-	printf("Time (ms): %f\t", timer->time[i] / (1000 * REP));
+	(void)timer;
+	(void)i;
+	(void)rep;
 }
 
-void printall(Timer *timer, int maxt)
+void stop(Timer *timer, int i)
 {
-	for (int i = 0; i <= maxt; i++) {
-		printf(" timer%d_us=%f", i, timer->time[i]);
-	}
-	printf("\n");
+	(void)timer;
+	(void)i;
 }
+
+#endif