add perf test

.gitignore

@@ -68,6 +68,7 @@ logic
 digit5x7
 pima
 save_test
+blas_perf
 .vscode
 mnist
 build/

Makefile

@@ -11,11 +11,11 @@ OBJS = ann.o tensor.o
 DEPS = ann.h tensor.h ann_config.h
 
 # use cblas
-#CFLAGS += -g -O3 -DUSE_BLAS -DCBLAS -I.
+#CFLAGS += -g -O2 -DUSE_BLAS -DCBLAS -I.
 #LFLAGS += -L. -lcblas
 
 # use openblas
-CFLAGS += -g -O3 -DUSE_BLAS -I"/opt/OpenBLAS/include"
+CFLAGS += -g -O2 -DUSE_BLAS -I"/opt/OpenBLAS/include"
 LFLAGS += -L/opt/OpenBLAS/lib/ -lopenblas
 
 #-DMKL_ILP64  -m64  -I"${MKLROOT}/include"
@@ -24,7 +24,7 @@ LFLAGS += -L/opt/OpenBLAS/lib/ -lopenblas
 #  -DMKL_ILP64  -m64  -I"${MKLROOT}/include"
 #  ${MKLROOT}/lib/libmkl_intel_ilp64.a ${MKLROOT}/lib/libmkl_tbb_thread.a ${MKLROOT}/lib/libmkl_core.a -L${TBBROOT}/lib -ltbb -lc++ -lpthread -lm -ldl
 
-all: mnist logic digit5x7 save_test save_test_binary
+all: mnist logic digit5x7 save_test save_test_binary blas_perf
 
 $(TARGET):	$(OBJS) mnist.o
 	$(CC) $(LFLAGS) -o $@ $^
@@ -41,6 +41,9 @@ save_test: $(OBJS) save_test.o
 save_test_binary: $(OBJS) save_test_binary.o
 	$(CC) $(LFLAGS) -o $@ $^
 
+blas_perf: $(OBJS) blas_perf.o
+	$(CC) $(LFLAGS) -o $@ $^
+
 %.o: %.c $(DEPS)
 	$(CC) -c $(CFLAGS) $(CPPFLAGS) $< -o $@
 

blas_perf.c

@@ -0,0 +1,133 @@
+//------------------------------------------------------
+//
+// Copyright 2023 Mark Seminatore. All rights reserved.
+//------------------------------------------------------
+#include <stdio.h>
+#include <string.h>
+#include <time.h>
+#include <cblas.h>
+
+#ifdef WIN32
+#   include <Windows.h>
+#endif
+
+#define MAX_SIZE 1024
+
+float x[MAX_SIZE], y[MAX_SIZE];
+float a[MAX_SIZE * MAX_SIZE], b[MAX_SIZE * MAX_SIZE], c[MAX_SIZE * MAX_SIZE];
+
+struct timer
+{
+#ifdef WIN32
+    LARGE_INTEGER t;
+#else
+    struct timespec t;
+#endif
+
+};
+
+//------------------------------------------------------
+//
+//------------------------------------------------------
+void timer_get_time(struct timer* t)
+{
+#ifdef WIN32
+    QueryPerformanceCounter(&t->t);
+#else
+    clock_gettime(CLOCK_PROCESS_CPUTIME_ID, &t->t);
+#endif
+
+}
+
+//------------------------------------------------------
+//
+//------------------------------------------------------
+float timer_get_delta(struct timer *t1, struct timer *t2)
+{
+    float dt;
+
+#ifdef WIN32
+    LARGE_INTEGER freq;
+    QueryPerformanceFrequency(&freq);
+
+    dt = (t2->t.QuadPart - t1->t.QuadPart) / (float)freq.QuadPart;
+
+#else
+    int seconds = (int)(t2->t.tv_sec - t1->t.tv_sec);
+    long long ns = t2->t.tv_nsec - t1->t.tv_nsec;
+    dt = (float)seconds + (float)ns / (1000000000);
+#endif
+
+    return dt;
+}
+
+//------------------------------------------------------
+//
+//------------------------------------------------------
+void test_gemm()
+{
+    struct timer t1, t2;
+    CBLAS_INDEX m, n, k;
+    float dt;
+
+    printf("Testing performance of cblas_sgemm()\n\n");
+
+    for (int i = 4; i <= MAX_SIZE; i <<= 1)
+    {
+        m = n = k = i;
+
+        timer_get_time(&t1);
+
+        cblas_sgemm(CblasRowMajor, CblasNoTrans, CblasNoTrans, m, n, k, 1.0f, a, m, b, k, 1.0f, c, k);
+
+        timer_get_time(&t2);
+
+        dt = timer_get_delta(&t1, &t2);
+
+        printf("%4d: %5.2f GFlops in %5.2fs\n", i, (float)2 * m * n * k / 1000000000 / dt, dt);
+    }
+}
+
+//------------------------------------------------------
+//
+//------------------------------------------------------
+void test_ger()
+{
+    struct timer t1, t2;
+    float dt;
+
+    printf("Testing performance of cblas_sger()\n\n");
+
+    CBLAS_INDEX m = MAX_SIZE, n = MAX_SIZE;
+
+    for (int i = 2; i <= MAX_SIZE; i <<= 1)
+    {
+        m = n = i;
+
+        timer_get_time(&t1);
+
+        cblas_sger(CblasRowMajor, m, n, 1.0f, x, 1, y, 1, a, m);
+
+        timer_get_time(&t2);
+
+        dt = timer_get_delta(&t1, &t2);
+
+        printf("%4d: %5.2f GFlops in %5.2fs\n", i, (float)2 * m * n / 1000000000 / dt, dt);
+    }
+}
+
+//------------------------------------------------------
+//
+//------------------------------------------------------
+int main(int argc, char *argv[])
+{
+	//cblas_init();
+
+    printf( "%s\n", openblas_get_config());
+    printf("    CPU uArch: %s\n", openblas_get_corename());
+    printf("Cores/Threads: %d/%d\n\n", openblas_get_num_procs(), openblas_get_num_threads());
+
+    test_gemm();
+
+	return 0;
+}

mnist.c

@@ -167,9 +167,9 @@ int main(int argc, char *argv[])
 		cblas_init();
 		if (threads != -1)
 			cblas_set_num_threads(threads);
-//		printf( "%s\n", openblas_get_config());
-//		printf("      CPU uArch: %s\n", openblas_get_corename());
-//		printf("  Cores/Threads: %d/%d\n", openblas_get_num_procs(), openblas_get_num_threads());
+		printf( "%s\n", cblas_get_config());
+		printf("      CPU uArch: %s\n", cblas_get_corename());
+		printf("  Cores/Threads: %d/%d\n", cblas_get_num_procs(), cblas_get_num_threads());
 	#else
 		if (threads != -1)
 			openblas_set_num_threads(threads);