RED: introduce alloc/load/free overhead parameters; vary them for fgbs24a

4 files changed, 121 insertions, 61 deletions

diff --git a/RED/Makefile b/RED/Makefile
index fc8298d..f20e1f7 100644
--- a/RED/Makefile
+++ b/RED/Makefile
@@ -1,19 +1,20 @@
 NR_TASKLETS ?= 16
 BL ?= 10
 NR_DPUS ?= 1
-VERSION ?= SINGLE
-SYNC ?= HAND
 TYPE ?= INT64
 ENERGY ?= 0
 PERF ?= 0
+WITH_ALLOC_OVERHEAD ?= 0
+WITH_LOAD_OVERHEAD ?= 0
+WITH_FREE_OVERHEAD ?= 0
 
 COMMON_INCLUDES := support
 HOST_SOURCES := $(wildcard host/*.c)
 DPU_SOURCES := $(wildcard dpu/*.c)
 
 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${VERSION} -D${SYNC} -D${TYPE} -DENERGY=${ENERGY} -DPERF=${PERF}
-DPU_FLAGS := ${COMMON_FLAGS} -O2 -DNR_TASKLETS=${NR_TASKLETS} -DBL=${BL} -D${VERSION} -D${SYNC} -D${TYPE} -DPERF=${PERF}
+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} -DPERF=${PERF} -DWITH_ALLOC_OVERHEAD=${WITH_ALLOC_OVERHEAD} -DWITH_LOAD_OVERHEAD=${WITH_LOAD_OVERHEAD} -DWITH_FREE_OVERHEAD=${WITH_FREE_OVERHEAD}
+DPU_FLAGS := ${COMMON_FLAGS} -O2 -DNR_TASKLETS=${NR_TASKLETS} -DBL=${BL} -D${TYPE} -DPERF=${PERF}
 
 QUIET = @
 
diff --git a/RED/host/app.c b/RED/host/app.c
index e40fdc6..2e5d715 100644
--- a/RED/host/app.c
+++ b/RED/host/app.c
@@ -57,17 +57,31 @@ int main(int argc, char **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
+    uint32_t nr_of_ranks;
+
+    // Timer declaration
+    Timer timer;
 
     // Allocate DPUs and load binary
+#if !WITH_ALLOC_OVERHEAD
     DPU_ASSERT(dpu_alloc(NR_DPUS, NULL, &dpu_set));
+    timer.time[0] = 0; // alloc
+#endif
+#if !WITH_LOAD_OVERHEAD
     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);
+    DPU_ASSERT(dpu_get_nr_ranks(dpu_set, &nr_of_ranks));
+    assert(nr_of_dpus == NR_DPUS);
+    timer.time[1] = 0; // load
+#endif
+#if !WITH_FREE_OVERHEAD
+    timer.time[6] = 0; // free
+#endif
+
+#if ENERGY
+    struct dpu_probe_t probe;
+    DPU_ASSERT(dpu_probe_init("energy_probe", &probe));
+#endif
 
     unsigned int i = 0;
 #if PERF
@@ -75,15 +89,15 @@ int main(int argc, char **argv) {
     double cc_min = 0;
 #endif
 
-    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 = p.exp == 0 ? p.input_size * NR_DPUS : p.input_size; // Total input size (weak or strong scaling)
     const unsigned int input_size_8bytes = 
         ((input_size * sizeof(T)) % 8) != 0 ? roundup(input_size, 8) : input_size; // Input size per DPU (max.), 8-byte aligned
-    const unsigned int input_size_dpu = divceil(input_size, nr_of_dpus); // Input size per DPU (max.)
+    const unsigned int input_size_dpu = divceil(input_size, NR_DPUS); // Input size per DPU (max.)
     const unsigned int input_size_dpu_8bytes = 
         ((input_size_dpu * sizeof(T)) % 8) != 0 ? roundup(input_size_dpu, 8) : input_size_dpu; // Input size per DPU (max.), 8-byte aligned
 
     // Input/output allocation
-    A = malloc(input_size_dpu_8bytes * nr_of_dpus * sizeof(T));
+    A = malloc(input_size_dpu_8bytes * NR_DPUS * sizeof(T));
     T *bufferA = A;
     T count = 0;
     T count_host = 0;
@@ -91,34 +105,53 @@ int main(int argc, char **argv) {
     // Create an input file with arbitrary data
     read_input(A, input_size);
 
-    // 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++) {
 
+#if WITH_ALLOC_OVERHEAD
+        if(rep >= p.n_warmup) {
+            start(&timer, 0, 0);
+        }
+        DPU_ASSERT(dpu_alloc(NR_DPUS, NULL, &dpu_set));
+        if(rep >= p.n_warmup) {
+            stop(&timer, 0);
+        }
+#endif
+#if WITH_LOAD_OVERHEAD
+        if(rep >= p.n_warmup) {
+            start(&timer, 1, 0);
+        }
+        DPU_ASSERT(dpu_load(dpu_set, DPU_BINARY, NULL));
+        if(rep >= p.n_warmup) {
+            stop(&timer, 1);
+        }
+        DPU_ASSERT(dpu_get_nr_dpus(dpu_set, &nr_of_dpus));
+        DPU_ASSERT(dpu_get_nr_ranks(dpu_set, &nr_of_ranks));
+        assert(nr_of_dpus == NR_DPUS);
+#endif
+
         // Compute output on CPU (performance comparison and verification purposes)
         if(rep >= p.n_warmup)
-            start(&timer, 0, 0);
+            start(&timer, 2, 0);
         count_host = reduction_host(A, input_size);
         if(rep >= p.n_warmup)
-            stop(&timer, 0);
+            stop(&timer, 2);
 
         //printf("Load input data\n");
         if(rep >= p.n_warmup)
-            start(&timer, 1, 0);
+            start(&timer, 3, 0);
         count = 0;
         // Input arguments
         unsigned int kernel = 0;
         dpu_arguments_t input_arguments[NR_DPUS];
-        for(i=0; i<nr_of_dpus-1; i++) {
+        for(i=0; i<NR_DPUS-1; i++) {
             input_arguments[i].size=input_size_dpu_8bytes * sizeof(T); 
             input_arguments[i].kernel=kernel;
         }
-        input_arguments[nr_of_dpus-1].size=(input_size_8bytes - input_size_dpu_8bytes * (NR_DPUS-1)) * sizeof(T); 
-        input_arguments[nr_of_dpus-1].kernel=kernel;		
+        input_arguments[NR_DPUS-1].size=(input_size_8bytes - input_size_dpu_8bytes * (NR_DPUS-1)) * sizeof(T); 
+        input_arguments[NR_DPUS-1].kernel=kernel;
         // Copy input arrays
         i = 0;
         DPU_FOREACH(dpu_set, dpu, i) {
@@ -130,12 +163,12 @@ int main(int argc, char **argv) {
         }
         DPU_ASSERT(dpu_push_xfer(dpu_set, DPU_XFER_TO_DPU, DPU_MRAM_HEAP_POINTER_NAME, 0, input_size_dpu_8bytes * sizeof(T), DPU_XFER_DEFAULT));
         if(rep >= p.n_warmup)
-            stop(&timer, 1);
+            stop(&timer, 3);
 
         //printf("Run program on DPU(s) \n");
         // Run DPU kernel
         if(rep >= p.n_warmup) {
-            start(&timer, 2, 0);
+            start(&timer, 4, 0);
             #if ENERGY
             DPU_ASSERT(dpu_probe_start(&probe));
             #endif
@@ -143,7 +176,7 @@ int main(int argc, char **argv) {
  
         DPU_ASSERT(dpu_launch(dpu_set, DPU_SYNCHRONOUS));
         if(rep >= p.n_warmup) {
-            stop(&timer, 2);
+            stop(&timer, 4);
             #if ENERGY
             DPU_ASSERT(dpu_probe_stop(&probe));
             #endif
@@ -162,13 +195,13 @@ int main(int argc, char **argv) {
 #endif
 
         //printf("Retrieve results\n");
-        dpu_results_t results[nr_of_dpus];
-        T* results_count = malloc(nr_of_dpus * sizeof(T));
+        dpu_results_t results[NR_DPUS];
+        T* results_count = malloc(NR_DPUS * sizeof(T));
         if(rep >= p.n_warmup)
-            start(&timer, 3, 0);
+            start(&timer, 5, 0);
         i = 0;
         // PARALLEL RETRIEVE TRANSFER
-        dpu_results_t* results_retrieve[nr_of_dpus];
+        dpu_results_t* results_retrieve[NR_DPUS];
 
         DPU_FOREACH(dpu_set, dpu, i) {
             results_retrieve[i] = (dpu_results_t*)malloc(NR_TASKLETS * sizeof(dpu_results_t));
@@ -204,7 +237,7 @@ int main(int argc, char **argv) {
         }
 #endif
         if(rep >= p.n_warmup)
-            stop(&timer, 3);
+            stop(&timer, 5);
 
 #if PERF
         uint64_t max_cycles = 0;
@@ -227,23 +260,54 @@ int main(int argc, char **argv) {
         // Free memory
         free(results_count);
 
+#if WITH_ALLOC_OVERHEAD
+#if WITH_FREE_OVERHEAD
+        if(rep >= p.n_warmup) {
+            start(&timer, 6, 0);
+        }
+#endif
+        DPU_ASSERT(dpu_free(dpu_set));
+#if WITH_FREE_OVERHEAD
+        if(rep >= p.n_warmup) {
+            stop(&timer, 6);
+        }
+#endif
+#endif
+
         // Check output
         bool status = true;
         if(count != count_host) status = false;
         if (status) {
             printf("[" ANSI_COLOR_GREEN "OK" ANSI_COLOR_RESET "] Outputs are equal\n");
             if (rep >= p.n_warmup) {
-                printf("[::] RED NMC | n_dpus=%d n_tasklets=%d e_type=%s block_size_B=%d n_elements=%d "
-                    "| throughput_cpu_MBps=%f throughput_pim_MBps=%f throughput_MBps=%f",
-                    nr_of_dpus, NR_TASKLETS, XSTR(T), BLOCK_SIZE, input_size,
-                    input_size * sizeof(T) / timer.time[0],
-                    input_size * sizeof(T) / (timer.time[2] + timer.time[3]),
-                    input_size * sizeof(T) / (timer.time[1] + timer.time[2] + timer.time[3]));
-                printf(" throughput_cpu_MOpps=%f throughput_pim_MOpps=%f throughput_MOpps=%f",
-                    input_size / timer.time[0],
-                    input_size / (timer.time[2] + timer.time[3]),
-                    input_size / (timer.time[1] + timer.time[2] + timer.time[3]));
-                printall(&timer, 3);
+                printf("[::] RED UPMEM | n_dpus=%d n_tasklets=%d e_type=%s block_size_B=%d n_elements=%d",
+                    NR_DPUS, NR_TASKLETS, XSTR(T), BLOCK_SIZE, input_size);
+                printf(" b_with_alloc_overhead=%d b_with_load_overhead=%d b_with_free_overhead=%d ",
+                    WITH_ALLOC_OVERHEAD, WITH_LOAD_OVERHEAD, WITH_FREE_OVERHEAD);
+                printf("| latency_alloc_us=%f latency_load_us=%f latency_cpu_us=%f latency_write_us=%f latency_kernel_us=%f latency_read_us=%f latency_free_us=%f",
+                    timer.time[0],
+                    timer.time[1],
+                    timer.time[2],
+                    timer.time[3],
+                    timer.time[4],
+                    timer.time[5],
+                    timer.time[6]);
+                printf(" throughput_cpu_MBps=%f throughput_upmem_kernel_MBps=%f throughput_upmem_total_MBps=%f",
+                    input_size * sizeof(T) / timer.time[2],
+                    input_size * sizeof(T) / (timer.time[4]),
+                    input_size * sizeof(T) / (timer.time[0] + timer.time[1] + timer.time[3] + timer.time[4] + timer.time[5] + timer.time[6]));
+                printf(" throughput_upmem_wxr_MBps=%f throughput_upmem_lwxr_MBps=%f throughput_upmem_alwxr_MBps=%f",
+                    input_size * sizeof(T) / (timer.time[3] + timer.time[4] + timer.time[5]),
+                    input_size * sizeof(T) / (timer.time[1] + timer.time[3] + timer.time[4] + timer.time[5]),
+                    input_size * sizeof(T) / (timer.time[0] + timer.time[1] + timer.time[3] + timer.time[4] + timer.time[5]));
+                printf(" throughput_cpu_MOpps=%f throughput_upmem_kernel_MOpps=%f throughput_upmem_total_MOpps=%f",
+                    input_size / timer.time[2],
+                    input_size / (timer.time[4]),
+                    input_size / (timer.time[0] + timer.time[1] + timer.time[3] + timer.time[4] + timer.time[5] + timer.time[6]));
+                printf(" throughput_upmem_wxr_MOpps=%f throughput_upmem_lwxr_MOpps=%f throughput_upmem_alwxr_MOpps=%f\n",
+                    input_size / (timer.time[3] + timer.time[4] + timer.time[5]),
+                    input_size / (timer.time[1] + timer.time[3] + timer.time[4] + timer.time[5]),
+                    input_size / (timer.time[0] + timer.time[1] + timer.time[3] + timer.time[4] + timer.time[5]));
             }
         } else {
             printf("[" ANSI_COLOR_RED "ERROR" ANSI_COLOR_RESET "] Outputs differ!\n");
@@ -254,27 +318,17 @@ int main(int argc, char **argv) {
     printf("DPU cycles  = %g cc\n", cc / p.n_reps);
 #endif
 
-    // Print timing results
-/*
-    printf("CPU ");
-    print(&timer, 0, p.n_reps);
-    printf("CPU-DPU ");
-    print(&timer, 1, p.n_reps);
-    printf("DPU Kernel ");
-    print(&timer, 2, p.n_reps);
-    printf("Inter-DPU ");
-    print(&timer, 3, 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	
+    #endif
 
     // Deallocation
     free(A);
+#if !WITH_ALLOC_OVERHEAD
     DPU_ASSERT(dpu_free(dpu_set));
-	
+#endif
+
     return 0;
 }
diff --git a/RED/run-fgbs24a.sh b/RED/run-fgbs24a.sh
index becee22..962bb46 100755
--- a/RED/run-fgbs24a.sh
+++ b/RED/run-fgbs24a.sh
@@ -4,6 +4,8 @@ set -e
 
 mkdir -p $(hostname)
 
+ts=$(date +%Y%m%d)
+
 # BL: use 2^(BL) B blocks for MRAM <-> WRAM transfers on PIM module
 # T: data type
 # -w: number of un-timed warmup iterations
@@ -16,13 +18,16 @@ echo "Started at $(date)"
 echo "Revision $(git describe --always)"
 
 # >2048 are not part of upstream
-for nr_dpus in 2543 2304 2048; do
+for nr_dpus in 2304 2048 2543; do
 	for nr_tasklets in 16; do
 		echo
-		if make -B NR_DPUS=${nr_dpus} NR_TASKLETS=${nr_tasklets} BL=10 VERSION=SINGLE; then
+		if make -B NR_DPUS=${nr_dpus} NR_TASKLETS=${nr_tasklets} BL=10; then
+			timeout --foreground -k 1m 30m bin/host_code -w 0 -e 100 -i 419430400 -x 1 || true
+		fi
+		if make -B NR_DPUS=${nr_dpus} NR_TASKLETS=${nr_tasklets} BL=10 WITH_ALLOC_OVERHEAD=1 WITH_LOAD_OVERHEAD=1 WITH_FREE_OVERHEAD=1; then
 			timeout --foreground -k 1m 30m bin/host_code -w 0 -e 100 -i 419430400 -x 1 || true
 		fi
 	done
 done
 echo "Completed at $(date)"
-) | tee "$(hostname)/fgbs24a.txt"
+) | tee "$(hostname)/${ts}-fgbs24a.txt"
diff --git a/RED/support/timer.h b/RED/support/timer.h
index 5c00213..4d597b9 100755
--- a/RED/support/timer.h
+++ b/RED/support/timer.h
@@ -37,9 +37,9 @@
 

 typedef struct Timer{

 

-    struct timeval startTime[4];

-    struct timeval stopTime[4];

-    double         time[4];

+    struct timeval startTime[7];

+    struct timeval stopTime[7];

+    double         time[7];

 

 }Timer;