GEMV: add MEMCPY variant

2 files changed, 89 insertions, 7 deletions

diff --git a/GEMV/baselines/cpu/Makefile b/GEMV/baselines/cpu/Makefile
index 382fa96..016d561 100644
--- a/GEMV/baselines/cpu/Makefile
+++ b/GEMV/baselines/cpu/Makefile
@@ -1,4 +1,5 @@
 NUMA ?= 0
+NUMA_MEMCPY ?= 0
 FLAGS =
 TYPE ?= double
 
@@ -10,7 +11,7 @@ endif
 all: gemv
 
 gemv: gemv_openmp.c
-	gcc -Wall -Wextra -pedantic -march=native -O2 -o gemv -fopenmp -DNUMA=${NUMA} -DT=${TYPE} -DTYPE_${TYPE} gemv_openmp.c ${FLAGS}
+	gcc -ggdb -Wall -Wextra -pedantic -march=native -O2 -o gemv -fopenmp -DNUMA=${NUMA} -DNUMA_MEMCPY=${NUMA_MEMCPY} -DT=${TYPE} -DTYPE_${TYPE} gemv_openmp.c ${FLAGS}
 
 gemv_O0: gemv_openmp.c
 	gcc -o gemv_O0 -fopenmp gemv_openmp.c
diff --git a/GEMV/baselines/cpu/gemv_openmp.c b/GEMV/baselines/cpu/gemv_openmp.c
index 798f7de..3da1504 100644
--- a/GEMV/baselines/cpu/gemv_openmp.c
+++ b/GEMV/baselines/cpu/gemv_openmp.c
@@ -21,6 +21,12 @@ int numa_node_out = -1;
 int numa_node_cpu = -1;
 #endif
 
+#if NUMA_MEMCPY
+struct bitmask* bitmask_cpu;
+int numa_node_local = -1;
+int numa_node_in_is_local = 0;
+#endif
+
 #define XSTR(x) STR(x)
 #define STR(x) #x
 
@@ -40,13 +46,18 @@ int main(int argc, char *argv[])
 
     T **A, *b, *x;
 
+    T **A_local, *x_local;
+
 #if NUMA
     bitmask_in    = numa_parse_nodestring(argv[1]);
     bitmask_out   = numa_parse_nodestring(argv[2]);
     numa_node_cpu = atoi(argv[3]);
+#if NUMA_MEMCPY
+    bitmask_cpu   = numa_parse_nodestring(argv[4]);
+#endif // NUMA_MEMCPY
 #else
     (void) argv;
-#endif
+#endif // NUMA
 
 #if NUMA
     if (bitmask_out) {
@@ -79,11 +90,21 @@ int main(int argc, char *argv[])
     make_hilbert_mat(rows,cols, &A);
 
 #if NUMA
+#if NUMA_MEMCPY
+    if (bitmask_cpu) {
+        numa_set_membind(bitmask_cpu);
+        numa_free_nodemask(bitmask_cpu);
+    }
+#else
     struct bitmask *bitmask_all = numa_allocate_nodemask();
     numa_bitmask_setall(bitmask_all);
     numa_set_membind(bitmask_all);
     numa_free_nodemask(bitmask_all);
-#endif
+#endif // NUMA_MEMCPY
+#endif // NUMA
+
+    A_local = A;
+    x_local = x;
 
 #if NUMA
     mp_pages[0] = A;
@@ -116,6 +137,10 @@ int main(int argc, char *argv[])
     }
 #endif
 
+#if NUMA_MEMCPY
+    numa_node_in_is_local = ((numa_node_cpu == numa_node_in) || (numa_node_cpu + 8 == numa_node_in)) * 1;
+#endif
+
     Timer timer;
     for (int i = 0; i < 100; i++) {
 
@@ -132,14 +157,70 @@ int main(int argc, char *argv[])
         }
         }
 
+#if NUMA_MEMCPY
+        start(&timer, 1, 0);
+        if (!numa_node_in_is_local) {
+            x_local = (T*) numa_alloc(sizeof(T)*cols);
+            allocate_dense(rows, cols, &A_local);
+        }
+        stop(&timer, 1);
+
+        start(&timer, 2, 0);
+        if (!numa_node_in_is_local) {
+            //for (size_t i=0; i < rows; i++ ) {
+             //   memcpy(A_local[i], A[i], cols * sizeof(T));
+            //}
+            memcpy(*A_local, *A, rows * cols * sizeof(T));
+            memcpy(x_local, x, cols * sizeof(T));
+        } else {
+            A_local = A;
+            x_local = x;
+        }
+        stop(&timer, 2);
+
+        mp_pages[0] = A_local;
+        if (move_pages(0, 1, mp_pages, NULL, mp_status, 0) == -1) {
+            perror("move_pages(A_local)");
+        }
+        else if (mp_status[0] < 0) {
+            printf("move_pages error: %d", mp_status[0]);
+        }
+        else {
+            numa_node_local = mp_status[0];
+        }
+#endif
+
         unsigned int nr_threads = 0;
 #pragma omp parallel
 #pragma omp atomic
         nr_threads++;
 
         start(&timer, 0, 0);
-        gemv(A, x, rows, cols, &b);
+        gemv(A_local, x_local, rows, cols, &b);
         stop(&timer, 0);
+
+#if NUMA_MEMCPY
+        start(&timer, 3, 0);
+        if (!numa_node_in_is_local) {
+            numa_free(x_local, sizeof(T) * rows);
+            numa_free(*A_local, sizeof(T) * rows * cols);
+            numa_free(A_local, sizeof(void*) * rows);
+        }
+        stop(&timer, 3);
+#endif
+
+#if NUMA_MEMCPY
+        printf("[::] GEMV-CPU-MEMCPY | n_threads=%d e_type=%s n_elements=%ld"
+            " numa_node_in=%d numa_node_out=%d numa_node_cpu=%d numa_node_local=%d numa_distance_in_cpu=%d numa_distance_cpu_out=%d"
+            " | throughput_MBps=%f throughput_MOpps=%f",
+            nr_threads, XSTR(T), rows * cols,
+            numa_node_in, numa_node_out, numa_node_cpu, numa_node_local, numa_distance(numa_node_in, numa_node_cpu), numa_distance(numa_node_cpu, numa_node_out),
+            rows * cols * sizeof(T) / timer.time[0],
+            rows * cols / timer.time[0]);
+        printf(" latency_kernel_us=%f latency_alloc_us=%f latency_memcpy_us=%f latency_free_us=%f latency_total_us=%f\n",
+            timer.time[0], timer.time[1], timer.time[2], timer.time[3],
+            timer.time[0] + timer.time[1] + timer.time[2] + timer.time[3]);
+#else
         printf("[::] GEMV-CPU | n_threads=%d e_type=%s n_elements=%ld"
 #if NUMA
             " numa_node_in=%d numa_node_out=%d numa_node_cpu=%d numa_distance_in_cpu=%d numa_distance_cpu_out=%d"
@@ -150,9 +231,9 @@ int main(int argc, char *argv[])
             numa_node_in, numa_node_out, numa_node_cpu, numa_distance(numa_node_in, numa_node_cpu), numa_distance(numa_node_cpu, numa_node_out),
 #endif
             rows * cols * sizeof(T) / timer.time[0]);
-        printf(" throughput_MOpps=%f",
-            rows * cols / timer.time[0]);
-        printall(&timer, 0);
+        printf(" throughput_MOpps=%f latency_us=%f\n",
+            rows * cols / timer.time[0], timer.time[0]);
+#endif
     }