PrIM -- first commit

13 files changed, 1164 insertions, 0 deletions

diff --git a/SCAN-RSS/Makefile b/SCAN-RSS/Makefile
new file mode 100644
index 0000000..4930e70
--- /dev/null
+++ b/SCAN-RSS/Makefile
@@ -0,0 +1,46 @@
+DPU_DIR := dpu
+HOST_DIR := host
+BUILDDIR ?= bin
+NR_DPUS ?= 1
+NR_TASKLETS ?= 16
+BL ?= 10
+TYPE ?= INT64
+ENERGY ?= 0
+
+define conf_filename
+	${BUILDDIR}/.NR_DPUS_$(1)_NR_TASKLETS_$(2)_BL_$(3)_TYPE_$(4).conf
+endef
+CONF := $(call conf_filename,${NR_DPUS},${NR_TASKLETS},${BL},${TYPE})
+
+HOST_TARGET := ${BUILDDIR}/host_code
+DPU_TARGET := ${BUILDDIR}/dpu_code
+
+COMMON_INCLUDES := support
+HOST_SOURCES := $(wildcard ${HOST_DIR}/*.c)
+DPU_SOURCES := $(wildcard ${DPU_DIR}/*.c)
+
+.PHONY: all clean test
+
+__dirs := $(shell mkdir -p ${BUILDDIR})
+
+COMMON_FLAGS := -Wall -Wextra -g -I${COMMON_INCLUDES}
+HOST_FLAGS := ${COMMON_FLAGS} -std=c11 -O3 `dpu-pkg-config --cflags --libs dpu` -DNR_TASKLETS=${NR_TASKLETS} -DNR_DPUS=${NR_DPUS} -DBL=${BL} -D${TYPE} -DENERGY=${ENERGY}
+DPU_FLAGS := ${COMMON_FLAGS} -O2 -DNR_TASKLETS=${NR_TASKLETS} -DBL=${BL} -D${TYPE} 
+
+all: ${HOST_TARGET} ${DPU_TARGET}
+
+${CONF}:
+	$(RM) $(call conf_filename,*,*)
+	touch ${CONF}
+
+${HOST_TARGET}: ${HOST_SOURCES} ${COMMON_INCLUDES} ${CONF}
+	$(CC) -o $@ ${HOST_SOURCES} ${HOST_FLAGS}
+
+${DPU_TARGET}: ${DPU_SOURCES} ${COMMON_INCLUDES} ${CONF}
+	dpu-upmem-dpurte-clang ${DPU_FLAGS} -o $@ ${DPU_SOURCES}
+
+clean:
+	$(RM) -r $(BUILDDIR)
+
+test: all
+	./${HOST_TARGET}
diff --git a/SCAN-RSS/baselines/cpu/Makefile b/SCAN-RSS/baselines/cpu/Makefile
new file mode 100644
index 0000000..2765a1b
--- /dev/null
+++ b/SCAN-RSS/baselines/cpu/Makefile
@@ -0,0 +1,5 @@
+all:
+	g++ -O2 app_baseline.cpp -fopenmp -DTHRUST_HOST_SYSTEM=THRUST_HOST_SYSTEM_CPP -DTHRUST_DEVICE_SYSTEM=THRUST_DEVICE_SYSTEM_OMP -lgomp -I/usr/local/cuda-8.0/include -lm -o scan -D${TYPE}
+
+clean:
+	rm scan
diff --git a/SCAN-RSS/baselines/cpu/README b/SCAN-RSS/baselines/cpu/README
new file mode 100644
index 0000000..01eeccc
--- /dev/null
+++ b/SCAN-RSS/baselines/cpu/README
@@ -0,0 +1,9 @@
+Prefix sum (SCAN)
+
+Compilation instructions
+
+    TYPE=UINT64 make
+
+Execution instructions
+
+    ./scan -i 1258291200 -t 4
diff --git a/SCAN-RSS/baselines/cpu/app_baseline.cpp b/SCAN-RSS/baselines/cpu/app_baseline.cpp
new file mode 100644
index 0000000..3840eb4
--- /dev/null
+++ b/SCAN-RSS/baselines/cpu/app_baseline.cpp
@@ -0,0 +1,199 @@
+/*
+* JGL@SAFARI
+*/
+
+/**
+* CPU code with Thrust
+*/
+#include <stdio.h>
+#include <stdlib.h>
+#include <stdbool.h>
+#include <string.h>
+#include <unistd.h>
+#include <getopt.h>
+#include <assert.h>
+
+#include <iostream>
+#include <fstream>
+#include <cstdlib>
+#include <ctime>
+#include <cstdio>
+#include <math.h>
+#include <sys/time.h>
+
+#include <vector>
+
+#include <thrust/device_vector.h>
+#include <thrust/host_vector.h>
+#include <thrust/scan.h>
+#include <thrust/copy.h>
+#include <thrust/system/omp/execution_policy.h>
+#include <thrust/system/omp/vector.h>
+
+#include <omp.h>
+
+#include "../../support/common.h"
+#include "../../support/timer.h"
+
+#define ANSI_COLOR_RED     "\x1b[31m"
+#define ANSI_COLOR_GREEN   "\x1b[32m"
+#define ANSI_COLOR_RESET   "\x1b[0m"
+
+// Pointer declaration
+static T* A;
+static T* C;
+static T* C2;
+
+/**
+* @brief creates input arrays
+* @param nr_elements how many elements in input arrays
+*/
+static void read_input(T* A, unsigned int nr_elements) {
+    //srand(0);
+    printf("nr_elements\t%u\t", nr_elements);
+    for (unsigned int i = 0; i < nr_elements; i++) {
+        //A[i] = (T) (rand()) % 2;
+        A[i] = i;
+    }
+}
+
+/**
+* @brief compute output in the host
+*/
+static void scan_host(T* C, T* A, unsigned int nr_elements) {
+    C[0] = A[0];
+    for (unsigned int i = 1; i < nr_elements; i++) {
+        C[i] = C[i - 1] + A[i - 1];
+    }
+}
+
+// Params ---------------------------------------------------------------------
+typedef struct Params {
+    unsigned int   input_size;
+    int   n_warmup;
+    int   n_reps;
+    int   exp;
+    int   n_threads;
+}Params;
+
+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    -t <T>    # of threads (default=8)"
+        "\n"
+        "\nBenchmark-specific options:"
+        "\n    -i <I>    input size (default=8M elements)"
+        "\n");
+}
+
+struct Params input_params(int argc, char **argv) {
+    struct Params p;
+    p.input_size    = 2 << 20;
+    p.n_warmup      = 1;
+    p.n_reps        = 3;
+    p.exp           = 0;
+    p.n_threads     = 8;
+
+    int opt;
+    while((opt = getopt(argc, argv, "hi:w:e:x:t:")) >= 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;
+        case 't': p.n_threads     = atoi(optarg); break;
+        default:
+            fprintf(stderr, "\nUnrecognized option!\n");
+            usage();
+            exit(0);
+        }
+    }
+    assert(p.n_threads > 0 && "Invalid # of threads!");
+
+    return p;
+}
+
+/**
+* @brief Main of the Host Application.
+*/
+int main(int argc, char **argv) {
+
+    struct Params p = input_params(argc, argv);
+
+    unsigned int nr_of_dpus = 1;
+    
+    unsigned int i = 0;
+    const unsigned int input_size =  p.input_size;
+    assert(input_size % (p.n_threads) == 0 && "Input size!");
+
+    // Input/output allocation
+    A = (T*)malloc(input_size * sizeof(T));
+    C = (T*)malloc(input_size * sizeof(T));
+    T *bufferA = A;
+
+    // Create an input file with arbitrary data.
+    read_input(A, input_size);
+
+    // Timer declaration
+    Timer timer;
+    float time_gpu = 0;
+
+    thrust::omp::vector<T> h_output(input_size);
+
+    // Loop over main kernel
+    for(int rep = 0; rep < p.n_warmup + p.n_reps; rep++) {
+
+        // Compute output on CPU (performance comparison and verification purposes)
+        if(rep >= p.n_warmup)
+            start(&timer, 0, rep - p.n_warmup);
+        scan_host(C, A, input_size);
+        if(rep >= p.n_warmup)
+            stop(&timer, 0);
+
+        memcpy(thrust::raw_pointer_cast(&h_output[0]), A, input_size * sizeof(T));
+
+        omp_set_num_threads(p.n_threads);
+
+        if(rep >= p.n_warmup)
+            start(&timer, 1, rep - p.n_warmup);
+        thrust::exclusive_scan(thrust::omp::par, h_output.begin(),h_output.end(),h_output.begin());
+        if(rep >= p.n_warmup)
+            stop(&timer, 1);
+    }
+
+    // Print timing results
+    printf("CPU ");
+    print(&timer, 0, p.n_reps);
+    printf("Kernel ");
+    print(&timer, 1, p.n_reps);
+
+    // Check output
+    bool status = true;
+    for (i = 0; i < input_size; i++) {
+        if(C[i] != h_output[i]){ 
+            status = false;
+            //printf("%d: %lu -- %lu\n", i, C[i], h_output[i]);
+        }
+    }
+    if (status) {
+        printf("[" ANSI_COLOR_GREEN "OK" ANSI_COLOR_RESET "] Outputs are equal\n");
+    } else {
+        printf("[" ANSI_COLOR_RED "ERROR" ANSI_COLOR_RESET "] Outputs differ!\n");
+    }
+
+    // Deallocation
+    free(A);
+    free(C);
+	
+    return 0;
+}
diff --git a/SCAN-RSS/baselines/gpu/Makefile b/SCAN-RSS/baselines/gpu/Makefile
new file mode 100644
index 0000000..1f28af7
--- /dev/null
+++ b/SCAN-RSS/baselines/gpu/Makefile
@@ -0,0 +1,5 @@
+all:
+	/usr/local/cuda/bin/nvcc app_baseline.cu -I/usr/local/cuda/include -lm -o scan -D${TYPE}
+
+clean:
+	rm scan
diff --git a/SCAN-RSS/baselines/gpu/README b/SCAN-RSS/baselines/gpu/README
new file mode 100644
index 0000000..fb58fd3
--- /dev/null
+++ b/SCAN-RSS/baselines/gpu/README
@@ -0,0 +1,9 @@
+Prefix sum (SCAN)
+
+Compilation instructions
+
+    TYPE=UINT64 make
+
+Execution instructions
+
+    ./scan -i 1258291200
diff --git a/SCAN-RSS/baselines/gpu/app_baseline.cu b/SCAN-RSS/baselines/gpu/app_baseline.cu
new file mode 100644
index 0000000..1d01620
--- /dev/null
+++ b/SCAN-RSS/baselines/gpu/app_baseline.cu
@@ -0,0 +1,214 @@
+/*
+* JGL@SAFARI
+*/
+
+/**
+* GPU code with Thrust
+*/
+#include <stdio.h>
+#include <stdlib.h>
+#include <stdbool.h>
+#include <string.h>
+#include <unistd.h>
+#include <getopt.h>
+#include <assert.h>
+
+#include <iostream>
+#include <fstream>
+#include <cstdlib>
+#include <ctime>
+#include <cstdio>
+#include <math.h>
+#include <sys/time.h>
+
+#include <vector>
+#include <thrust/device_vector.h>
+#include <thrust/host_vector.h>
+#include <thrust/scan.h>
+#include <thrust/copy.h>
+
+#include "../../support/common.h"
+#include "../../support/timer.h"
+
+#define ANSI_COLOR_RED     "\x1b[31m"
+#define ANSI_COLOR_GREEN   "\x1b[32m"
+#define ANSI_COLOR_RESET   "\x1b[0m"
+
+// Pointer declaration
+static T* A;
+static T* C;
+static T* C2;
+
+/**
+* @brief creates input arrays
+* @param nr_elements how many elements in input arrays
+*/
+static void read_input(T* A, unsigned int nr_elements) {
+    //srand(0);
+    printf("nr_elements\t%u\t", nr_elements);
+    for (unsigned int i = 0; i < nr_elements; i++) {
+        //A[i] = (T) (rand()) % 2;
+        A[i] = i;
+    }
+}
+
+/**
+* @brief compute output in the host
+*/
+static void scan_host(T* C, T* A, unsigned int nr_elements) {
+    C[0] = A[0];
+    for (unsigned int i = 1; i < nr_elements; i++) {
+        C[i] = C[i - 1] + A[i - 1];
+    }
+}
+
+// Params ---------------------------------------------------------------------
+typedef struct Params {
+    unsigned int   input_size;
+    int   n_warmup;
+    int   n_reps;
+    int   exp;
+    int   n_threads;
+}Params;
+
+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    -t <T>    # of threads (default=8)"
+        "\n"
+        "\nBenchmark-specific options:"
+        "\n    -i <I>    input size (default=640 * 3932160 elements)"
+        "\n");
+}
+
+struct Params input_params(int argc, char **argv) {
+    struct Params p;
+    p.input_size    = 1258291200;
+    p.n_warmup      = 1;
+    p.n_reps        = 3;
+    p.exp           = 0;
+    p.n_threads     = 8;
+
+    int opt;
+    while((opt = getopt(argc, argv, "hi:w:e:x:t:")) >= 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;
+        case 't': p.n_threads     = atoi(optarg); break;
+        default:
+            fprintf(stderr, "\nUnrecognized option!\n");
+            usage();
+            exit(0);
+        }
+    }
+    assert(p.n_threads > 0 && "Invalid # of threads!");
+
+    return p;
+}
+
+/**
+* @brief Main of the Host Application.
+*/
+int main(int argc, char **argv) {
+
+    cudaDeviceProp device_properties;
+    cudaGetDeviceProperties(&device_properties, 0);
+    cudaSetDevice(0);
+
+    struct Params p = input_params(argc, argv);
+
+    unsigned int nr_of_dpus = 1;
+    
+    unsigned int i = 0;
+    const unsigned int input_size = p.exp == 0 ? p.input_size * nr_of_dpus : p.input_size;
+
+    // Input/output allocation
+    A = (T*)malloc(input_size * sizeof(T));
+    C = (T*)malloc(input_size * sizeof(T));
+    C2 = (T*)malloc(input_size * sizeof(T));
+    T *bufferA = A;
+    T *bufferC = C2;
+
+    // Create an input file with arbitrary data.
+    read_input(A, input_size);
+
+    // Timer declaration
+    Timer timer;
+    float time_gpu = 0;
+
+    thrust::host_vector<T> h_output(input_size);
+
+    // Loop over main kernel
+    for(int rep = 0; rep < p.n_warmup + p.n_reps; rep++) {
+
+        // Compute output on CPU (performance comparison and verification purposes)
+        if(rep >= p.n_warmup)
+            start(&timer, 0, rep - p.n_warmup);
+        scan_host(C, A, input_size);
+        if(rep >= p.n_warmup)
+            stop(&timer, 0);
+
+
+        // Event creation
+        cudaEvent_t start, stop;
+        cudaEventCreate(&start);
+        cudaEventCreate(&stop);
+        float time1 = 0;
+
+        thrust::device_vector<T> d_input(input_size);
+        cudaMemcpy(thrust::raw_pointer_cast(&d_input[0]), A, input_size * sizeof(T), cudaMemcpyHostToDevice);
+
+        // Start timer
+        cudaEventRecord( start, 0 );
+        thrust::exclusive_scan(d_input.begin(),d_input.end(),d_input.begin());
+        // End timer
+        cudaEventRecord( stop, 0 );
+        cudaEventSynchronize( stop );
+        cudaEventElapsedTime( &time1, start, stop );
+        time_gpu += time1;
+
+        h_output = d_input;
+
+        cudaEventDestroy(start);
+        cudaEventDestroy(stop);
+    }
+
+    // Print timing results
+    printf("CPU ");
+    print(&timer, 0, p.n_reps);
+    printf("Kernel (ms):");
+    printf("%f\n", time_gpu / p.n_reps);
+
+    // Check output
+    bool status = true;
+    for (i = 0; i < input_size; i++) {
+        if(C[i] != h_output[i]){ 
+            status = false;
+            printf("%d: %lu -- %lu\n", i, C[i], h_output[i]);
+        }
+    }
+    if (status) {
+        printf("[" ANSI_COLOR_GREEN "OK" ANSI_COLOR_RESET "] Outputs are equal\n");
+    } else {
+        printf("[" ANSI_COLOR_RED "ERROR" ANSI_COLOR_RESET "] Outputs differ!\n");
+    }
+
+    // Deallocation
+    free(A);
+    free(C);
+    free(C2);
+	
+    return 0;
+}
diff --git a/SCAN-RSS/dpu/task.c b/SCAN-RSS/dpu/task.c
new file mode 100644
index 0000000..7a4b029
--- /dev/null
+++ b/SCAN-RSS/dpu/task.c
@@ -0,0 +1,192 @@
+/*
+* Scan with multiple tasklets (Reduce-scan-scan)
+*
+*/
+#include <stdint.h>
+#include <stdio.h>
+#include <defs.h>
+#include <mram.h>
+#include <alloc.h>
+#include <perfcounter.h>
+#include <handshake.h>
+#include <barrier.h>
+
+#include "../support/common.h"
+
+__host dpu_arguments_t DPU_INPUT_ARGUMENTS;
+__host dpu_results_t DPU_RESULTS[NR_TASKLETS];
+
+// Array for communication between adjacent tasklets
+T message[NR_TASKLETS];
+T message_partial_count;
+
+// Reduction in each tasklet
+static T reduction(T *input){
+    T output = 0;
+    #pragma unroll
+    for(unsigned int j = 0; j < REGS; j++) {
+        output += input[j];
+    }
+    return output;
+}
+// Scan in each tasklet
+static T scan(T *output, T *input){
+    output[0] = input[0];
+    #pragma unroll
+    for(unsigned int j = 1; j < REGS; j++) {
+        output[j] = output[j - 1] + input[j];
+    }
+    return output[REGS - 1];
+}
+// Handshake with adjacent tasklets
+static T handshake_sync(T l_count, unsigned int tasklet_id){
+    T p_count;
+    // Wait and read message
+    if(tasklet_id != 0){
+        handshake_wait_for(tasklet_id - 1);
+        p_count = message[tasklet_id];
+    }
+    else
+        p_count = 0;
+    // Write message and notify
+    if(tasklet_id < NR_TASKLETS - 1){
+        message[tasklet_id + 1] = p_count + l_count;
+        handshake_notify();
+    }
+    return p_count;
+}
+
+// Barrier
+BARRIER_INIT(my_barrier, NR_TASKLETS);
+
+// Add in each tasklet
+static void add(T *output, T p_count){
+    #pragma unroll
+    for(unsigned int j = 0; j < REGS; j++) {
+        output[j] += p_count;
+    }
+}
+
+extern int main_kernel1(void);
+extern int main_kernel2(void);
+
+int (*kernels[nr_kernels])(void) = {main_kernel1, main_kernel2};
+
+int main(void) { 
+    // Kernel
+    return kernels[DPU_INPUT_ARGUMENTS.kernel](); 
+}
+
+// Reduction
+int main_kernel1() {
+    unsigned int tasklet_id = me();
+#if PRINT
+    printf("tasklet_id = %u\n", tasklet_id);
+#endif
+    if (tasklet_id == 0){ // Initialize once the cycle counter
+        mem_reset(); // Reset the heap
+    }
+    // Barrier
+    barrier_wait(&my_barrier);
+
+    dpu_results_t *result = &DPU_RESULTS[tasklet_id];
+
+    uint32_t input_size_dpu_bytes = DPU_INPUT_ARGUMENTS.size; // Input size per DPU in bytes
+
+    // Address of the current processing block in MRAM
+    uint32_t base_tasklet = tasklet_id << BLOCK_SIZE_LOG2;
+    uint32_t mram_base_addr_A = (uint32_t)DPU_MRAM_HEAP_POINTER;
+
+    // Initialize a local cache to store the MRAM block
+    T *cache_A = (T *) mem_alloc(BLOCK_SIZE);
+	
+    // Local count
+    T l_count = 0;
+
+    for(unsigned int byte_index = base_tasklet; byte_index < input_size_dpu_bytes; byte_index += BLOCK_SIZE * NR_TASKLETS){
+
+        // Load cache with current MRAM block
+        mram_read((const __mram_ptr void*)(mram_base_addr_A + byte_index), cache_A, BLOCK_SIZE);
+
+        // Reduction in each tasklet
+        l_count += reduction(cache_A);
+
+    }
+
+    // Reduce local counts
+    message[tasklet_id] = l_count;
+
+    // Single-thread reduction
+    // Barrier
+    barrier_wait(&my_barrier);
+    if(tasklet_id == 0){
+        for (unsigned int each_tasklet = 1; each_tasklet < NR_TASKLETS; each_tasklet++){
+            message[0] += message[each_tasklet];
+        }
+        // Total count in this DPU
+        result->t_count = message[0];
+    }
+
+    return 0;
+}
+
+// Scan
+int main_kernel2() {
+    unsigned int tasklet_id = me();
+#if PRINT
+    printf("tasklet_id = %u\n", tasklet_id);
+#endif
+    if (tasklet_id == 0){ // Initialize once the cycle counter
+        mem_reset(); // Reset the heap
+    }
+    // Barrier
+    barrier_wait(&my_barrier);
+
+    dpu_results_t *result = &DPU_RESULTS[tasklet_id];
+
+    uint32_t input_size_dpu_bytes = DPU_INPUT_ARGUMENTS.size; // Input size per DPU in bytes
+
+    // Address of the current processing block in MRAM
+    uint32_t base_tasklet = tasklet_id << BLOCK_SIZE_LOG2;
+    uint32_t mram_base_addr_A = (uint32_t)DPU_MRAM_HEAP_POINTER;
+    uint32_t mram_base_addr_B = (uint32_t)(DPU_MRAM_HEAP_POINTER + input_size_dpu_bytes);
+
+    // Initialize a local cache to store the MRAM block
+    T *cache_A = (T *) mem_alloc(BLOCK_SIZE);
+    T *cache_B = (T *) mem_alloc(BLOCK_SIZE);
+	
+    // Initialize shared variable
+    if(tasklet_id == NR_TASKLETS - 1)
+        message_partial_count = DPU_INPUT_ARGUMENTS.t_count;
+    // Barrier
+    barrier_wait(&my_barrier);
+
+    for(unsigned int byte_index = base_tasklet; byte_index < input_size_dpu_bytes; byte_index += BLOCK_SIZE * NR_TASKLETS){
+
+        // Load cache with current MRAM block
+        mram_read((const __mram_ptr void*)(mram_base_addr_A + byte_index), cache_A, BLOCK_SIZE);
+
+        // Scan in each tasklet
+        T l_count = scan(cache_B, cache_A); 
+
+        // Sync with adjacent tasklets
+        T p_count = handshake_sync(l_count, tasklet_id);
+
+        // Barrier
+        barrier_wait(&my_barrier);
+
+        // Add in each tasklet
+        add(cache_B, message_partial_count + p_count);
+
+        // Write cache to current MRAM block
+        mram_write(cache_B, (__mram_ptr void*)(mram_base_addr_B + byte_index), BLOCK_SIZE);
+
+        // Total count in this DPU
+        if(tasklet_id == NR_TASKLETS - 1){
+            result->t_count = message_partial_count + p_count + l_count;
+            message_partial_count = result->t_count;
+        }
+	}
+		
+    return 0;
+}
diff --git a/SCAN-RSS/host/app.c b/SCAN-RSS/host/app.c
new file mode 100644
index 0000000..46cdd95
--- /dev/null
+++ b/SCAN-RSS/host/app.c
@@ -0,0 +1,290 @@
+/**
+* app.c
+* SCAN-RSS Host Application Source File
+*
+*/
+#include <stdio.h>
+#include <stdlib.h>
+#include <stdbool.h>
+#include <string.h>
+#include <dpu.h>
+#include <dpu_log.h>
+#include <unistd.h>
+#include <getopt.h>
+#include <assert.h>
+
+#include "../support/common.h"
+#include "../support/timer.h"
+#include "../support/params.h"
+
+// Define the DPU Binary path as DPU_BINARY here
+#ifndef DPU_BINARY
+#define DPU_BINARY "./bin/dpu_code"
+#endif
+
+#if ENERGY
+#include <dpu_probe.h>
+#endif
+
+// Pointer declaration
+static T* A;
+static T* C;
+static T* C2;
+
+// Create input arrays
+static void read_input(T* A, unsigned int nr_elements, unsigned int nr_elements_round) {
+    srand(0);
+    printf("nr_elements\t%u\t", nr_elements);
+    for (unsigned int i = 0; i < nr_elements; i++) {
+        A[i] = (T) (rand());
+    }
+    for (unsigned int i = nr_elements; i < nr_elements_round; i++) {
+        A[i] = 0;
+    }
+}
+
+// Compute output in the host
+static void scan_host(T* C, T* A, unsigned int nr_elements) {
+    C[0] = A[0];
+    for (unsigned int i = 1; i < nr_elements; i++) {
+        C[i] = C[i - 1] + A[i];
+    }
+}
+
+// Main of the Host Application
+int main(int argc, char **argv) {
+
+    struct Params p = input_params(argc, argv);
+
+    struct dpu_set_t dpu_set, dpu;
+    uint32_t nr_of_dpus;
+    
+#if ENERGY
+    struct dpu_probe_t probe;
+    DPU_ASSERT(dpu_probe_init("energy_probe", &probe));
+#endif
+
+    // Allocate DPUs and load binary
+    DPU_ASSERT(dpu_alloc(NR_DPUS, NULL, &dpu_set));
+    DPU_ASSERT(dpu_load(dpu_set, DPU_BINARY, NULL));
+    DPU_ASSERT(dpu_get_nr_dpus(dpu_set, &nr_of_dpus));
+    printf("Allocated %d DPU(s)\n", nr_of_dpus);
+
+    unsigned int i = 0;
+    T accum = 0;
+
+    const unsigned int input_size = p.exp == 0 ? p.input_size * nr_of_dpus : p.input_size; // Total input size (weak or strong scaling)
+    const unsigned int input_size_dpu_ = divceil(input_size, nr_of_dpus); // Input size per DPU (max.)
+    const unsigned int input_size_dpu_round = 
+        (input_size_dpu_ % (NR_TASKLETS * REGS) != 0) ? roundup(input_size_dpu_, (NR_TASKLETS * REGS)) : input_size_dpu_; // Input size per DPU (max.), 8-byte aligned
+
+    // Input/output allocation
+    A = malloc(input_size_dpu_round * nr_of_dpus * sizeof(T));
+    C = malloc(input_size_dpu_round * nr_of_dpus * sizeof(T));
+    C2 = malloc(input_size_dpu_round * nr_of_dpus * sizeof(T));
+    T *bufferA = A;
+    T *bufferC = C2;
+
+    // Create an input file with arbitrary data
+    read_input(A, input_size, input_size_dpu_round * nr_of_dpus);
+
+    // Timer declaration
+    Timer timer;
+
+    printf("NR_TASKLETS\t%d\tBL\t%d\n", NR_TASKLETS, BL);
+
+    // Loop over main kernel
+    for(int rep = 0; rep < p.n_warmup + p.n_reps; rep++) {
+
+        // Compute output on CPU (performance comparison and verification purposes)
+        if(rep >= p.n_warmup)
+            start(&timer, 0, rep - p.n_warmup);
+        scan_host(C, A, input_size);
+        if(rep >= p.n_warmup)
+            stop(&timer, 0);
+
+        printf("Load input data\n");
+        if(rep >= p.n_warmup)
+            start(&timer, 1, rep - p.n_warmup);
+        // Input arguments
+        const unsigned int input_size_dpu = input_size_dpu_round;
+        unsigned int kernel = 0;
+        dpu_arguments_t input_arguments = {input_size_dpu * sizeof(T), kernel, 0};
+        // Copy input arrays
+        i = 0;
+        DPU_FOREACH(dpu_set, dpu, i) {
+            DPU_ASSERT(dpu_prepare_xfer(dpu, &input_arguments));
+        }
+        DPU_ASSERT(dpu_push_xfer(dpu_set, DPU_XFER_TO_DPU, "DPU_INPUT_ARGUMENTS", 0, sizeof(input_arguments), DPU_XFER_DEFAULT));
+        DPU_FOREACH(dpu_set, dpu, i) {
+            DPU_ASSERT(dpu_prepare_xfer(dpu, bufferA + input_size_dpu * i));
+        }
+        DPU_ASSERT(dpu_push_xfer(dpu_set, DPU_XFER_TO_DPU, DPU_MRAM_HEAP_POINTER_NAME, 0, input_size_dpu * sizeof(T), DPU_XFER_DEFAULT));
+        if(rep >= p.n_warmup)
+            stop(&timer, 1);
+
+        printf("Run program on DPU(s) \n");
+        // Run DPU kernel
+        if(rep >= p.n_warmup) {
+            start(&timer, 2, rep - p.n_warmup);
+            #if ENERGY
+            DPU_ASSERT(dpu_probe_start(&probe));
+            #endif
+        }
+ 
+        DPU_ASSERT(dpu_launch(dpu_set, DPU_SYNCHRONOUS));
+        if(rep >= p.n_warmup) {
+            stop(&timer, 2);
+            #if ENERGY
+            DPU_ASSERT(dpu_probe_stop(&probe));
+            #endif
+        }
+
+#if PRINT
+        {
+            unsigned int each_dpu = 0;
+            printf("Display DPU Logs\n");
+            DPU_FOREACH (dpu_set, dpu) {
+                printf("DPU#%d:\n", each_dpu);
+                DPU_ASSERT(dpulog_read_for_dpu(dpu.dpu, stdout));
+                each_dpu++;
+            }
+        }
+#endif
+
+        printf("Retrieve results\n");
+        dpu_results_t results[nr_of_dpus];
+        T* results_scan = malloc(nr_of_dpus * sizeof(T));
+        i = 0;
+        accum = 0;
+		
+        if(rep >= p.n_warmup)
+            start(&timer, 3, rep - p.n_warmup);
+        // PARALLEL RETRIEVE TRANSFER
+        dpu_results_t* results_retrieve[nr_of_dpus];
+
+        DPU_FOREACH(dpu_set, dpu, i) {
+            results_retrieve[i] = (dpu_results_t*)malloc(NR_TASKLETS * sizeof(dpu_results_t));
+            DPU_ASSERT(dpu_prepare_xfer(dpu, results_retrieve[i]));
+        }
+        DPU_ASSERT(dpu_push_xfer(dpu_set, DPU_XFER_FROM_DPU, "DPU_RESULTS", 0, NR_TASKLETS * sizeof(dpu_results_t), DPU_XFER_DEFAULT));
+
+        DPU_FOREACH(dpu_set, dpu, i) {
+            // Retrieve tasklet timings
+            for (unsigned int each_tasklet = 0; each_tasklet < NR_TASKLETS; each_tasklet++) {
+                if(each_tasklet == 0)
+                    results[i].t_count = results_retrieve[i][each_tasklet].t_count;
+            }
+            free(results_retrieve[i]);
+            // Sequential scan
+            T temp = results[i].t_count;
+            results_scan[i] = accum;
+            accum += temp;
+#if PRINT
+            printf("i=%d -- %lu,  %lu, %lu\n", i, results_scan[i], accum, temp);
+#endif
+        }
+
+        // Arguments for scan kernel (2nd kernel)
+        kernel = 1;
+        dpu_arguments_t input_arguments_2[NR_DPUS];
+        for(i=0; i<nr_of_dpus; i++) {
+            input_arguments_2[i].size=input_size_dpu * sizeof(T); 
+            input_arguments_2[i].kernel=kernel;
+            input_arguments_2[i].t_count=results_scan[i];
+        }
+        DPU_FOREACH(dpu_set, dpu, i) {
+            DPU_ASSERT(dpu_prepare_xfer(dpu, &input_arguments_2[i]));
+        }
+        DPU_ASSERT(dpu_push_xfer(dpu_set, DPU_XFER_TO_DPU, "DPU_INPUT_ARGUMENTS", 0, sizeof(input_arguments_2[0]), DPU_XFER_DEFAULT));
+        if(rep >= p.n_warmup)
+            stop(&timer, 3);
+
+        printf("Run program on DPU(s) \n");
+        // Run DPU kernel
+        if(rep >= p.n_warmup) {
+            start(&timer, 4, rep - p.n_warmup);
+            #if ENERGY
+            DPU_ASSERT(dpu_probe_start(&probe));
+            #endif
+        }
+        DPU_ASSERT(dpu_launch(dpu_set, DPU_SYNCHRONOUS));
+        if(rep >= p.n_warmup) {
+            stop(&timer, 4);
+            #if ENERGY
+            DPU_ASSERT(dpu_probe_stop(&probe));
+            #endif
+        }
+#if PRINT
+        {
+            unsigned int each_dpu = 0;
+            printf("Display DPU Logs\n");
+            DPU_FOREACH (dpu_set, dpu) {
+                printf("DPU#%d:\n", each_dpu);
+                DPU_ASSERT(dpulog_read_for_dpu(dpu.dpu, stdout));
+                each_dpu++;
+            }
+        }
+#endif
+
+        printf("Retrieve results\n");
+        if(rep >= p.n_warmup)
+            start(&timer, 5, rep - p.n_warmup);
+        i = 0;
+        // PARALLEL RETRIEVE TRANSFER
+        DPU_FOREACH(dpu_set, dpu, i) {
+            DPU_ASSERT(dpu_prepare_xfer(dpu, bufferC + input_size_dpu * i));
+        }
+        DPU_ASSERT(dpu_push_xfer(dpu_set, DPU_XFER_FROM_DPU, DPU_MRAM_HEAP_POINTER_NAME, input_size_dpu * sizeof(T), input_size_dpu * sizeof(T), DPU_XFER_DEFAULT));
+        if(rep >= p.n_warmup)
+            stop(&timer, 5);
+
+        // Free memory
+        free(results_scan);
+    }
+
+    // Print timing results
+    printf("CPU ");
+    print(&timer, 0, p.n_reps);
+    printf("CPU-DPU ");
+    print(&timer, 1, p.n_reps);
+    printf("DPU Kernel Reduction ");
+    print(&timer, 2, p.n_reps);
+    printf("Inter-DPU (Scan) ");
+    print(&timer, 3, p.n_reps);
+    printf("DPU Kernel Scan ");
+    print(&timer, 4, p.n_reps);
+    printf("DPU-CPU ");
+    print(&timer, 5, p.n_reps);
+
+    #if ENERGY
+    double energy;
+    DPU_ASSERT(dpu_probe_get(&probe, DPU_ENERGY, DPU_AVERAGE, &energy));
+    printf("DPU Energy (J): %f\t", energy);
+    #endif	
+
+
+    // Check output
+    bool status = true;
+    for (i = 0; i < input_size; i++) {
+        if(C[i] != bufferC[i]){ 
+            status = false;
+#if PRINT
+            printf("%d: %lu -- %lu\n", i, C[i], bufferC[i]);
+#endif
+        }
+    }
+    if (status) {
+        printf("[" ANSI_COLOR_GREEN "OK" ANSI_COLOR_RESET "] Outputs are equal\n");
+    } else {
+        printf("[" ANSI_COLOR_RED "ERROR" ANSI_COLOR_RESET "] Outputs differ!\n");
+    }
+
+    // Deallocation
+    free(A);
+    free(C);
+    free(C2);
+    DPU_ASSERT(dpu_free(dpu_set));
+	
+    return status ? 0 : -1;
+}
diff --git a/SCAN-RSS/run.sh b/SCAN-RSS/run.sh
new file mode 100755
index 0000000..1c39f7c
--- /dev/null
+++ b/SCAN-RSS/run.sh
@@ -0,0 +1,11 @@
+#!/bin/bash
+
+for i in  2048 4096 8192 16384 65536 262144 1048576 3932160
+do
+		    NR_DPUS=1 NR_TASKLETS=16 BL=10 VERSION=SINGLE make all
+			wait
+            ./bin/host_code -w 10 -e 100 -i ${i} > profile/out${i}_tl16_bl10_dpu11
+			wait
+			make clean
+			wait
+done
diff --git a/SCAN-RSS/support/common.h b/SCAN-RSS/support/common.h
new file mode 100755
index 0000000..be19a8c
--- /dev/null
+++ b/SCAN-RSS/support/common.h
@@ -0,0 +1,69 @@
+#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
+
+#define REGS (BLOCK_SIZE >> DIV)
+
+// Structures used by both the host and the dpu to communicate information 
+typedef struct {
+    uint32_t size;
+	enum kernels {
+	    kernel1 = 0,
+	    kernel2 = 1,
+	    nr_kernels = 2,
+	} kernel;
+    T t_count;
+} dpu_arguments_t;
+
+typedef struct {
+    T t_count;
+} dpu_results_t;
+
+#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/SCAN-RSS/support/params.h b/SCAN-RSS/support/params.h
new file mode 100644
index 0000000..bb86211
--- /dev/null
+++ b/SCAN-RSS/support/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=0)"
+        "\n"
+        "\nBenchmark-specific options:"
+        "\n    -i <I>    input size (default=3932160 elements)"
+        "\n");
+}
+
+struct Params input_params(int argc, char **argv) {
+    struct Params p;
+    p.input_size    = 3932160;
+    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    = 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/SCAN-RSS/support/timer.h b/SCAN-RSS/support/timer.h
new file mode 100755
index 0000000..b53d95f
--- /dev/null
+++ b/SCAN-RSS/support/timer.h
@@ -0,0 +1,59 @@
+/*

+ * Copyright (c) 2016 University of Cordoba and University of Illinois

+ * All rights reserved.

+ *

+ * Developed by:    IMPACT Research Group

+ *                  University of Cordoba and University of Illinois

+ *                  http://impact.crhc.illinois.edu/

+ *

+ * Permission is hereby granted, free of charge, to any person obtaining a copy

+ * of this software and associated documentation files (the "Software"), to deal

+ * with the Software without restriction, including without limitation the 

+ * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or

+ * sell copies of the Software, and to permit persons to whom the Software is

+ * furnished to do so, subject to the following conditions:

+ *

+ *      > Redistributions of source code must retain the above copyright notice,

+ *        this list of conditions and the following disclaimers.

+ *      > Redistributions in binary form must reproduce the above copyright

+ *        notice, this list of conditions and the following disclaimers in the

+ *        documentation and/or other materials provided with the distribution.

+ *      > Neither the names of IMPACT Research Group, University of Cordoba, 

+ *        University of Illinois nor the names of its contributors may be used 

+ *        to endorse or promote products derived from this Software without 

+ *        specific prior written permission.

+ *

+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR

+ * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,

+ * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE 

+ * CONTRIBUTORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER

+ * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,

+ * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS WITH

+ * THE SOFTWARE.

+ *

+ */

+

+#include <sys/time.h>

+

+typedef struct Timer{

+

+    struct timeval startTime[7];

+    struct timeval stopTime[7];

+    double         time[7];

+

+}Timer;

+

+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);

+}

+

+void print(Timer *timer, int i, int REP) { printf("Time (ms): %f\t", timer->time[i] / (1000 * REP)); }