5 files changed, 184 insertions, 0 deletions
diff --git a/RED/include/common.h b/RED/include/common.h
new file mode 100644
index 0000000..6cc1ae2
--- /dev/null
+++ b/RED/include/common.h
@@ -0,0 +1,75 @@
+#ifndef _COMMON_H_
+#define _COMMON_H_
+
+// 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
+
+enum kernels {
+	kernel1 = 0,
+	nr_kernels = 1,
+};
+
+// Structures used by both the host and the dpu to communicate information
+typedef struct {
+	uint32_t size;
+	enum kernels kernel;
+	T t_count;
+} dpu_arguments_t;
+
+typedef struct {
+	uint64_t cycles;
+	T t_count;
+} dpu_results_t;
+
+#ifndef PERF
+#define PERF 0 // Use perfcounters?
+#endif
+#ifndef PERF_SYNC
+#define PERF_SYNC 0 // Use perfcounters to time sync primitives?
+#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/RED/include/cyclecount.h b/RED/include/cyclecount.h
new file mode 100644
index 0000000..c4247b5
--- /dev/null
+++ b/RED/include/cyclecount.h
@@ -0,0 +1,19 @@
+#include <perfcounter.h>
+
+// Timer
+typedef struct perfcounter_cycles{
+    perfcounter_t start;
+    perfcounter_t end;
+    perfcounter_t end2;
+
+}perfcounter_cycles;
+
+void timer_start(perfcounter_cycles *cycles){
+    cycles->start = perfcounter_get(); // START TIMER
+}
+
+uint64_t timer_stop(perfcounter_cycles *cycles){
+    cycles->end = perfcounter_get(); // STOP TIMER
+    cycles->end2 = perfcounter_get(); // STOP TIMER
+    return(((uint64_t)((uint32_t)(((cycles->end >> 4) - (cycles->start >> 4)) - ((cycles->end2 >> 4) - (cycles->end >> 4))))) << 4);
+}
diff --git a/RED/include/dfatool_host.ah b/RED/include/dfatool_host.ah
new file mode 100644
index 0000000..88dfbd8
--- /dev/null
+++ b/RED/include/dfatool_host.ah
@@ -0,0 +1,29 @@
+#pragma once
+
+#include <sys/time.h>
+#include "dfatool_host_dpu.ah"
+
+aspect DfatoolHostTiming : public DfatoolHostDPUTiming {
+	unsigned long n_elements;
+	unsigned int element_size;
+
+	virtual int getKernel() { return 1; }
+
+	DfatoolHostTiming() {
+		element_size = sizeof(T);
+	}
+
+	advice call("% input_params(...)") : after() {
+		Params* p = tjp->result();
+		n_elements = p->input_size;
+		printf("[>>] RED | n_dpus=%u n_elements=%lu\n", NR_DPUS, n_elements);
+	}
+
+	advice call("% reduction_host(...)") : after() {
+		printf("[--] RED | n_dpus=%u n_elements=%lu\n", n_dpus, n_elements);
+	}
+
+	advice execution("% main(...)") : after() {
+		printf("[<<] RED | n_dpus=%u n_elements=%lu\n", NR_DPUS, n_elements);
+	}
+};
diff --git a/RED/include/params.h b/RED/include/params.h
new file mode 100644
index 0000000..ee90908
--- /dev/null
+++ b/RED/include/params.h
@@ -0,0 +1,56 @@
+#ifndef _PARAMS_H_
+#define _PARAMS_H_
+
+#include "common.h"
+
+typedef struct Params {
+    unsigned 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=1)"
+        "\n"
+        "\nBenchmark-specific options:"
+        "\n    -i <I>    input size (default=6553600 elements)"
+        "\n");
+}
+
+struct Params input_params(int argc, char **argv) {
+    struct Params p;
+    p.input_size    = 6553600;
+    p.n_warmup      = 1;
+    p.n_reps        = 3;
+    p.exp           = 1;
+
+    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    = atoi(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/RED/include/timer.h b/RED/include/timer.h
new file mode 100644
index 0000000..7b80823
--- /dev/null
+++ b/RED/include/timer.h
@@ -0,0 +1,5 @@
+#pragma once
+
+#define N_TIMERS 7
+#include "../../include/timer_base.h"
+#undef N_TIMERS