4 files changed, 153 insertions, 0 deletions
diff --git a/VA/include/common.h b/VA/include/common.h
new file mode 100644
index 0000000..6ce6e23
--- /dev/null
+++ b/VA/include/common.h
@@ -0,0 +1,68 @@
+#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 kernel;
+} dpu_arguments_t;
+
+// Transfer size between MRAM and WRAM
+#ifdef BL
+#define BLOCK_SIZE_LOG2 BL
+#define BLOCK_SIZE (1 << BLOCK_SIZE_LOG2)
+#else
+#define BLOCK_SIZE_LOG2 8
+#define BLOCK_SIZE (1 << BLOCK_SIZE_LOG2)
+#define BL BLOCK_SIZE_LOG2
+#endif
+
+// Data type
+#ifdef UINT32
+#define T uint32_t
+#define DIV 2			// Shift right to divide by sizeof(T)
+#elif UINT64
+#define T uint64_t
+#define DIV 3			// Shift right to divide by sizeof(T)
+#elif INT32
+#define T int32_t
+#define DIV 2			// Shift right to divide by sizeof(T)
+#elif INT64
+#define T int64_t
+#define DIV 3			// Shift right to divide by sizeof(T)
+#elif FLOAT
+#define T float
+#define DIV 2			// Shift right to divide by sizeof(T)
+#elif DOUBLE
+#define T double
+#define DIV 3			// Shift right to divide by sizeof(T)
+#elif CHAR
+#define T char
+#define DIV 0			// Shift right to divide by sizeof(T)
+#elif SHORT
+#define T short
+#define DIV 1			// Shift right to divide by sizeof(T)
+#endif
+
+#ifndef ENERGY
+#define ENERGY 0
+#endif
+#define PRINT 0
+
+#define ANSI_COLOR_RED     "\x1b[31m"
+#define ANSI_COLOR_GREEN   "\x1b[32m"
+#define ANSI_COLOR_RESET   "\x1b[0m"
+
+#define divceil(n, m) (((n)-1) / (m) + 1)
+#define roundup(n, m) ((n / m) * m + m)
+#endif
diff --git a/VA/include/dfatool_host.ah b/VA/include/dfatool_host.ah
new file mode 100644
index 0000000..6818333
--- /dev/null
+++ b/VA/include/dfatool_host.ah
@@ -0,0 +1,15 @@
+#pragma once
+
+#include <sys/time.h>
+#include "dfatool_host_dpu.ah"
+
+aspect DfatoolHostTiming : public DfatoolHostDPUTiming {
+	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);
+	}
+
+	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);
+	}
+};
diff --git a/VA/include/params.h b/VA/include/params.h
new file mode 100644
index 0000000..4f41adc
--- /dev/null
+++ b/VA/include/params.h
@@ -0,0 +1,65 @@
+#ifndef _PARAMS_H_
+#define _PARAMS_H_
+
+#include "common.h"
+
+typedef struct Params {
+	unsigned long int input_size;
+	int n_warmup;
+	int n_reps;
+	int exp;
+} Params;
+
+static void usage()
+{
+	fprintf(stderr,
+		"\nUsage:  ./program [options]"
+		"\n"
+		"\nGeneral options:"
+		"\n    -h        help"
+		"\n    -w <W>    # of untimed warmup iterations (default=1)"
+		"\n    -e <E>    # of timed repetition iterations (default=3)"
+		"\n    -x <X>    Weak (0) or strong (1) scaling (default=0)"
+		"\n"
+		"\nBenchmark-specific options:"
+		"\n    -i <I>    input size (default=2621440 elements)" "\n");
+}
+
+struct Params input_params(int argc, char **argv)
+{
+	struct Params p;
+	p.input_size = 2621440;
+	p.n_warmup = 1;
+	p.n_reps = 3;
+	p.exp = 0;
+
+	int opt;
+	while ((opt = getopt(argc, argv, "hi:w:e:x:")) >= 0) {
+		switch (opt) {
+		case 'h':
+			usage();
+			exit(0);
+			break;
+		case 'i':
+			p.input_size = atol(optarg);
+			break;
+		case 'w':
+			p.n_warmup = atoi(optarg);
+			break;
+		case 'e':
+			p.n_reps = atoi(optarg);
+			break;
+		case 'x':
+			p.exp = atoi(optarg);
+			break;
+		default:
+			fprintf(stderr, "\nUnrecognized option!\n");
+			usage();
+			exit(0);
+		}
+	}
+	assert(NR_DPUS > 0 && "Invalid # of dpus!");
+
+	return p;
+}
+#endif
diff --git a/VA/include/timer.h b/VA/include/timer.h
new file mode 100644
index 0000000..7b80823
--- /dev/null
+++ b/VA/include/timer.h
@@ -0,0 +1,5 @@
+#pragma once
+
+#define N_TIMERS 7
+#include "../../include/timer_base.h"
+#undef N_TIMERS