5 files changed, 487 insertions, 0 deletions
diff --git a/Microbenchmarks/STREAM/dpu/add.c b/Microbenchmarks/STREAM/dpu/add.c
new file mode 100644
index 0000000..faf6504
--- /dev/null
+++ b/Microbenchmarks/STREAM/dpu/add.c
@@ -0,0 +1,103 @@
+/*
+* STREAM Add
+*
+*/
+#include <stdint.h>
+#include <stdio.h>
+#include <defs.h>
+#include <mram.h>
+#include <alloc.h>
+#include <perfcounter.h>
+#include <barrier.h>
+
+#include "../support/common.h"
+#include "../support/cyclecount.h"
+
+__host dpu_arguments_t DPU_INPUT_ARGUMENTS;
+__host dpu_results_t DPU_RESULTS[NR_TASKLETS];
+
+// Add
+static void add_dpu(T *bufferC, T *bufferA, T *bufferB) {
+
+    #pragma unroll
+    for (unsigned int i = 0; i < BLOCK_SIZE / sizeof(T); i++){
+        bufferC[i] = bufferA[i] + bufferB[i];
+    }
+
+}
+
+// Barrier
+BARRIER_INIT(my_barrier, NR_TASKLETS);
+
+extern int main_kernel1(void);
+
+int (*kernels[nr_kernels])(void) = {main_kernel1};
+
+int main(void) { 
+    // Kernel
+    return kernels[DPU_INPUT_ARGUMENTS.kernel](); 
+}
+
+// main_kernel1
+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
+
+        perfcounter_config(COUNT_CYCLES, true);
+    }
+    perfcounter_cycles cycles;
+    // Barrier
+    barrier_wait(&my_barrier);
+#ifndef WRAM
+    timer_start(&cycles); // START TIMER
+#endif
+
+    uint32_t input_size_dpu = DPU_INPUT_ARGUMENTS.size / sizeof(T);
+
+    dpu_results_t *result = &DPU_RESULTS[tasklet_id];
+    result->cycles = 0;
+
+    // Address of the current processing block in MRAM
+    uint32_t mram_base_addr_A = (uint32_t)(DPU_MRAM_HEAP_POINTER + (tasklet_id << BLOCK_SIZE_LOG2));
+    uint32_t mram_base_addr_B = (uint32_t)(DPU_MRAM_HEAP_POINTER + (tasklet_id << BLOCK_SIZE_LOG2) + input_size_dpu * sizeof(T));
+    uint32_t mram_base_addr_C = (uint32_t)(DPU_MRAM_HEAP_POINTER + (tasklet_id << BLOCK_SIZE_LOG2) + 2 * input_size_dpu * sizeof(T));
+
+    // 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);
+
+    for(unsigned int byte_index = 0; byte_index < input_size_dpu * sizeof(T); byte_index += BLOCK_SIZE * NR_TASKLETS){
+
+        // Load cache with current MRAM block
+        mram_read((__mram_ptr void const*)(mram_base_addr_A + byte_index), cache_A, BLOCK_SIZE);
+        mram_read((__mram_ptr void const*)(mram_base_addr_B + byte_index), cache_B, BLOCK_SIZE);
+
+#ifdef WRAM
+        // Barrier
+        barrier_wait(&my_barrier);
+        timer_start(&cycles); // START TIMER
+#endif
+
+        // Add
+        add_dpu(cache_B, cache_A, cache_B);
+
+#ifdef WRAM
+        result->cycles += timer_stop(&cycles); // STOP TIMER
+	    // Barrier
+        barrier_wait(&my_barrier);
+#endif
+
+        // Write cache to current MRAM block
+        mram_write(cache_B, (__mram_ptr void*)(mram_base_addr_C + byte_index), BLOCK_SIZE);
+
+    }
+
+#ifndef WRAM
+    result->cycles = timer_stop(&cycles); // STOP TIMER
+#endif
+    return 0;
+}
diff --git a/Microbenchmarks/STREAM/dpu/copy.c b/Microbenchmarks/STREAM/dpu/copy.c
new file mode 100644
index 0000000..ff4df42
--- /dev/null
+++ b/Microbenchmarks/STREAM/dpu/copy.c
@@ -0,0 +1,75 @@
+/*
+* STREAM Copy
+*
+*/
+#include <stdint.h>
+#include <stdio.h>
+#include <defs.h>
+#include <mram.h>
+#include <alloc.h>
+#include <perfcounter.h>
+#include <barrier.h>
+
+#include "../support/common.h"
+#include "../support/cyclecount.h"
+
+__host dpu_arguments_t DPU_INPUT_ARGUMENTS;
+__host dpu_results_t DPU_RESULTS[NR_TASKLETS];
+
+// Barrier
+BARRIER_INIT(my_barrier, NR_TASKLETS);
+
+extern int main_kernel1(void);
+
+int (*kernels[nr_kernels])(void) = {main_kernel1};
+
+int main(void) { 
+    // Kernel
+    return kernels[DPU_INPUT_ARGUMENTS.kernel](); 
+}
+
+// main_kernel1
+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
+
+        perfcounter_config(COUNT_CYCLES, true);
+    }
+    perfcounter_cycles cycles;
+    // Barrier
+    barrier_wait(&my_barrier);
+#ifndef WRAM
+    timer_start(&cycles); // START TIMER
+#endif
+
+    uint32_t input_size_dpu = DPU_INPUT_ARGUMENTS.size / sizeof(T);
+
+    dpu_results_t *result = &DPU_RESULTS[tasklet_id];
+    result->cycles = 0;
+
+    // Address of the current processing block in MRAM
+    uint32_t mram_base_addr_A = (uint32_t)(DPU_MRAM_HEAP_POINTER + (tasklet_id << BLOCK_SIZE_LOG2));
+    uint32_t mram_base_addr_B = (uint32_t)(DPU_MRAM_HEAP_POINTER + (tasklet_id << BLOCK_SIZE_LOG2) + input_size_dpu * sizeof(T));
+
+    // Initialize a local cache to store the MRAM block
+    T *cache_A = (T *) mem_alloc(BLOCK_SIZE);
+
+    for(unsigned int byte_index = 0; byte_index < input_size_dpu * sizeof(T); byte_index += BLOCK_SIZE * NR_TASKLETS){
+
+        // Load cache with current MRAM block
+        mram_read((__mram_ptr void const*)(mram_base_addr_A + byte_index), cache_A, BLOCK_SIZE);
+
+        // Write cache to current MRAM block
+        mram_write(cache_A, (__mram_ptr void*)(mram_base_addr_B + byte_index), BLOCK_SIZE);
+
+    }
+
+#ifndef WRAM
+    result->cycles = timer_stop(&cycles); // STOP TIMER
+#endif
+    return 0;
+}
diff --git a/Microbenchmarks/STREAM/dpu/copyw.c b/Microbenchmarks/STREAM/dpu/copyw.c
new file mode 100644
index 0000000..eff7a3b
--- /dev/null
+++ b/Microbenchmarks/STREAM/dpu/copyw.c
@@ -0,0 +1,101 @@
+/*
+* STREAM Copy (WRAM)
+*
+*/
+#include <stdint.h>
+#include <stdio.h>
+#include <defs.h>
+#include <mram.h>
+#include <alloc.h>
+#include <perfcounter.h>
+#include <barrier.h>
+
+#include "../support/common.h"
+#include "../support/cyclecount.h"
+
+__host dpu_arguments_t DPU_INPUT_ARGUMENTS;
+__host dpu_results_t DPU_RESULTS[NR_TASKLETS];
+
+// Copy
+static void copyw_dpu(T *bufferB, T *bufferA) {
+
+    #pragma unroll
+    for (unsigned int i = 0; i < BLOCK_SIZE / sizeof(T); i++){
+        bufferB[i] = bufferA[i];
+    }
+
+}
+
+// Barrier
+BARRIER_INIT(my_barrier, NR_TASKLETS);
+
+extern int main_kernel1(void);
+
+int (*kernels[nr_kernels])(void) = {main_kernel1};
+
+int main(void) { 
+    // Kernel
+    return kernels[DPU_INPUT_ARGUMENTS.kernel](); 
+}
+
+// main_kernel1
+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
+
+        perfcounter_config(COUNT_CYCLES, true);
+    }
+    perfcounter_cycles cycles;
+    // Barrier
+    barrier_wait(&my_barrier);
+#ifndef WRAM
+    timer_start(&cycles); // START TIMER
+#endif
+
+    uint32_t input_size_dpu = DPU_INPUT_ARGUMENTS.size / sizeof(T);
+
+    dpu_results_t *result = &DPU_RESULTS[tasklet_id];
+    result->cycles = 0;
+
+    // Address of the current processing block in MRAM
+    uint32_t mram_base_addr_A = (uint32_t)(DPU_MRAM_HEAP_POINTER + (tasklet_id << BLOCK_SIZE_LOG2));
+    uint32_t mram_base_addr_B = (uint32_t)(DPU_MRAM_HEAP_POINTER + (tasklet_id << BLOCK_SIZE_LOG2) + input_size_dpu * sizeof(T));
+
+    // 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);
+
+    for(unsigned int byte_index = 0; byte_index < input_size_dpu * sizeof(T); byte_index += BLOCK_SIZE * NR_TASKLETS){
+
+        // Load cache with current MRAM block
+        mram_read((__mram_ptr void const*)(mram_base_addr_A + byte_index), cache_A, BLOCK_SIZE);
+
+#ifdef WRAM
+	    // Barrier
+	    barrier_wait(&my_barrier);
+        timer_start(&cycles); // START TIMER
+#endif
+
+        // Copy
+        copyw_dpu(cache_B, cache_A);
+
+#ifdef WRAM
+        result->cycles += timer_stop(&cycles); // STOP TIMER
+	    // Barrier
+	    barrier_wait(&my_barrier);
+#endif
+
+        // Write cache to current MRAM block
+        mram_write(cache_B, (__mram_ptr void*)(mram_base_addr_B + byte_index), BLOCK_SIZE);
+
+    }
+
+#ifndef WRAM
+    result->cycles = timer_stop(&cycles); // STOP TIMER
+#endif
+    return 0;
+}
diff --git a/Microbenchmarks/STREAM/dpu/scale.c b/Microbenchmarks/STREAM/dpu/scale.c
new file mode 100644
index 0000000..4247aac
--- /dev/null
+++ b/Microbenchmarks/STREAM/dpu/scale.c
@@ -0,0 +1,103 @@
+/*
+* STREAM Scale
+*
+*/
+#include <stdint.h>
+#include <stdio.h>
+#include <defs.h>
+#include <mram.h>
+#include <alloc.h>
+#include <perfcounter.h>
+#include <barrier.h>
+
+#include "../support/common.h"
+#include "../support/cyclecount.h"
+
+__host dpu_arguments_t DPU_INPUT_ARGUMENTS;
+__host dpu_results_t DPU_RESULTS[NR_TASKLETS];
+
+// Scale
+static void scale_dpu(T *bufferB, T *bufferA, T scalar) {
+
+    #pragma unroll
+    for (unsigned int i = 0; i < BLOCK_SIZE / sizeof(T); i++){
+        bufferB[i] = scalar * bufferA[i];
+    }
+
+}
+
+// Barrier
+BARRIER_INIT(my_barrier, NR_TASKLETS);
+
+extern int main_kernel1(void);
+
+int (*kernels[nr_kernels])(void) = {main_kernel1};
+
+int main(void) { 
+    // Kernel
+    return kernels[DPU_INPUT_ARGUMENTS.kernel](); 
+}
+
+// main_kernel1
+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
+
+        perfcounter_config(COUNT_CYCLES, true);
+    }
+    perfcounter_cycles cycles;
+    // Barrier
+    barrier_wait(&my_barrier);
+#ifndef WRAM
+    timer_start(&cycles); // START TIMER
+#endif
+
+    uint32_t input_size_dpu = DPU_INPUT_ARGUMENTS.size / sizeof(T);
+
+    T scalar = (T)input_size_dpu; // Simply use this number as a scalar
+
+    dpu_results_t *result = &DPU_RESULTS[tasklet_id];
+    result->cycles = 0;
+
+    // Address of the current processing block in MRAM
+    uint32_t mram_base_addr_A = (uint32_t)(DPU_MRAM_HEAP_POINTER + (tasklet_id << BLOCK_SIZE_LOG2));
+    uint32_t mram_base_addr_B = (uint32_t)(DPU_MRAM_HEAP_POINTER + (tasklet_id << BLOCK_SIZE_LOG2) + input_size_dpu * sizeof(T));
+
+    // 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);
+
+    for(unsigned int byte_index = 0; byte_index < input_size_dpu * sizeof(T); byte_index += BLOCK_SIZE * NR_TASKLETS){
+
+        // Load cache with current MRAM block
+        mram_read((__mram_ptr void const*)(mram_base_addr_A + byte_index), cache_A, BLOCK_SIZE);
+
+#ifdef WRAM
+	    // Barrier
+	    barrier_wait(&my_barrier);
+        timer_start(&cycles); // START TIMER
+#endif
+
+        // Scale
+        scale_dpu(cache_B, cache_A, scalar);
+
+#ifdef WRAM
+        result->cycles += timer_stop(&cycles); // STOP TIMER
+	    // Barrier
+	    barrier_wait(&my_barrier);
+#endif
+
+        // Write cache to current MRAM block
+        mram_write(cache_B, (__mram_ptr void*)(mram_base_addr_B + byte_index), BLOCK_SIZE);
+
+    }
+
+#ifndef WRAM
+    result->cycles = timer_stop(&cycles); // STOP TIMER
+#endif
+    return 0;
+}
diff --git a/Microbenchmarks/STREAM/dpu/triad.c b/Microbenchmarks/STREAM/dpu/triad.c
new file mode 100644
index 0000000..e81cb59
--- /dev/null
+++ b/Microbenchmarks/STREAM/dpu/triad.c
@@ -0,0 +1,105 @@
+/*
+* STREAM Triad
+*
+*/
+#include <stdint.h>
+#include <stdio.h>
+#include <defs.h>
+#include <mram.h>
+#include <alloc.h>
+#include <perfcounter.h>
+#include <barrier.h>
+
+#include "../support/common.h"
+#include "../support/cyclecount.h"
+
+__host dpu_arguments_t DPU_INPUT_ARGUMENTS;
+__host dpu_results_t DPU_RESULTS[NR_TASKLETS];
+
+// Triad
+static void triad_dpu(T *bufferC, T *bufferA, T *bufferB, T scalar) {
+
+    #pragma unroll
+    for (unsigned int i = 0; i < BLOCK_SIZE / sizeof(T); i++){
+        bufferC[i] = bufferA[i] + scalar * bufferB[i];
+    }
+
+}
+
+// Barrier
+BARRIER_INIT(my_barrier, NR_TASKLETS);
+
+extern int main_kernel1(void);
+
+int (*kernels[nr_kernels])(void) = {main_kernel1};
+
+int main(void) { 
+    // Kernel
+    return kernels[DPU_INPUT_ARGUMENTS.kernel](); 
+}
+
+// main_kernel1
+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
+
+        perfcounter_config(COUNT_CYCLES, true);
+    }
+    perfcounter_cycles cycles;
+    // Barrier
+    barrier_wait(&my_barrier);
+#ifndef WRAM
+    timer_start(&cycles); // START TIMER
+#endif
+
+    uint32_t input_size_dpu = DPU_INPUT_ARGUMENTS.size / sizeof(T);
+
+    T scalar = (T)input_size_dpu; // Simply use this number as a scalar
+
+    dpu_results_t *result = &DPU_RESULTS[tasklet_id];
+    result->cycles = 0;
+
+    // Address of the current processing block in MRAM
+    uint32_t mram_base_addr_A = (uint32_t)(DPU_MRAM_HEAP_POINTER + (tasklet_id << BLOCK_SIZE_LOG2));
+    uint32_t mram_base_addr_B = (uint32_t)(DPU_MRAM_HEAP_POINTER + (tasklet_id << BLOCK_SIZE_LOG2) + input_size_dpu * sizeof(T));
+    uint32_t mram_base_addr_C = (uint32_t)(DPU_MRAM_HEAP_POINTER + (tasklet_id << BLOCK_SIZE_LOG2) + 2 * input_size_dpu * sizeof(T));
+
+    // 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);
+
+    for(unsigned int byte_index = 0; byte_index < input_size_dpu * sizeof(T); byte_index += BLOCK_SIZE * NR_TASKLETS){
+
+        // Load cache with current MRAM block
+        mram_read((__mram_ptr void const*)(mram_base_addr_A + byte_index), cache_A, BLOCK_SIZE);
+        mram_read((__mram_ptr void const*)(mram_base_addr_B + byte_index), cache_B, BLOCK_SIZE);
+
+#ifdef WRAM
+	    // Barrier
+	    barrier_wait(&my_barrier);
+        timer_start(&cycles); // START TIMER
+#endif
+
+        // Triad
+        triad_dpu(cache_B, cache_A, cache_B, scalar);
+
+#ifdef WRAM
+        result->cycles += timer_stop(&cycles); // STOP TIMER
+	    // Barrier
+	    barrier_wait(&my_barrier);
+#endif
+
+        // Write cache to current MRAM block
+        mram_write(cache_B, (__mram_ptr void*)(mram_base_addr_C + byte_index), BLOCK_SIZE);
+
+    }
+
+#ifndef WRAM
+    result->cycles = timer_stop(&cycles); // STOP TIMER
+#endif
+    return 0;
+}