Time the entire API call

Author: Arjun Ramaswami

examples/common/helper.c

@@ -1,10 +1,15 @@
 // Author: Arjun Ramaswami
 
+#define _POSIX_C_SOURCE 199309L  
 #include <stdio.h>
 #include <stdlib.h>
 #include <stdbool.h>
 #include <math.h>
 #include "helper.h"
+#include <time.h>
+#include <sys/time.h>
+#include <math.h>
+#define _USE_MATH_DEFINES
 
 /**
  * \brief  create random single precision complex floating point values  
@@ -71,14 +76,21 @@ void print_config(int N, int dim, int iter, int inv, int sp, int use_bram){
 
 /**
  * \brief  print time taken for fpga and fftw runs to a file
+ * \param  total_api_time: time taken to call iter times the host code
  * \param  timing: kernel execution and pcie transfer timing 
  * \param  N: fft size
  * \param  dim: number of dimensions of size
  * \param  iter: number of iterations of each transformation (if BATCH mode)
  * \param  inv: 1 if backward transform
  * \param  single precision floating point transformation
  */
-void display_measures(double pcie_rd, double pcie_wr, double exec_t, int N, int dim, int iter, int inv, int sp){
+void display_measures(double total_api_time, double pcie_rd, double pcie_wr, double exec_t, int N, int dim, int iter, int inv, int sp){
+
+  double avg_api_time = 0.0;
+
+  if (total_api_time != 0.0){
+    avg_api_time = total_api_time / iter;
+  }
 
   double pcie_read = pcie_rd / iter;
   double pcie_write = pcie_wr / iter;
@@ -106,4 +118,18 @@ void display_measures(double pcie_rd, double pcie_wr, double exec_t, int N, int
   printf("Kernel Execution   = %.2lfms\n", exec);
   printf("PCIe Read          = %.2lfms\n", pcie_read);
   printf("Throughput         = %.2lfGFLOPS/s | %.2lf GB/s\n", gflops, gBytes_per_sec);
+  printf("Avg API runtime    = %.2lfms\n", avg_api_time);
+}
+
+/**
+ * \brief  compute walltime in milliseconds
+ * \return time in milliseconds
+ */
+double getTimeinMilliseconds(){
+   struct timespec a;
+   if(clock_gettime(CLOCK_MONOTONIC, &a) != 0){
+     fprintf(stderr, "Error in getting wall clock time \n");
+     exit(EXIT_FAILURE);
+   }
+   return (double)(a.tv_nsec) * 1.0e-6 + (double)(a.tv_sec) * 1.0E3;
 }

examples/common/helper.h

@@ -12,6 +12,7 @@ bool fft_create_data(double2 *inp, int N);
 
 void print_config(int N, int dim, int iter, int inv, int sp, int use_bram);
 
-void display_measures(double pcie_rd, double pcie_wr, double exec, int N, int dim, int iter, int inv, int sp);
+void display_measures(double total_api_time, double pcie_rd, double pcie_wr, double exec, int N, int dim, int iter, int inv, int sp);
 
+double getTimeinMilliseconds();
 #endif // HELPER_H

examples/fft1d.c

@@ -84,7 +84,7 @@ int main(int argc, const char **argv) {
       return EXIT_FAILURE;
     }
 
-    display_measures(timing.pcie_read_t, timing.pcie_write_t, timing.exec_t, N, dim, iter, inv, sp);
+    display_measures(0.0, timing.pcie_read_t, timing.pcie_write_t, timing.exec_t, N, dim, iter, inv, sp);
   }
   else{
     fprintf(stderr, "Invalid timing measurement. Function returned prematurely\n");

examples/fft2d.c

@@ -24,6 +24,7 @@ int main(int argc, const char **argv) {
   fpga_t timing = {0.0, 0.0, 0.0, 0};
   int use_svm = 0, use_emulator = 0;
   double avg_rd = 0.0, avg_wr = 0.0, avg_exec = 0.0;
+  double temp_timer = 0.0, total_api_time = 0.0;
   bool status = true;
 
   struct argparse_option options[] = {
@@ -72,11 +73,15 @@ int main(int argc, const char **argv) {
 
       if(use_bram == 1){
         // use bram for 2d Transpose
+        temp_timer = getTimeinMilliseconds();
         timing = fftfpgaf_c2c_2d_bram(N, inp, out, inv);
+        total_api_time += getTimeinMilliseconds() - temp_timer;
       }
       else{
         // use global memory for 2d Transpose
+        temp_timer = getTimeinMilliseconds();
         timing = fftfpgaf_c2c_2d_ddr(N, inp, out, inv);
+        total_api_time += getTimeinMilliseconds() - temp_timer;
       }
 
   #ifdef USE_FFTW
@@ -108,7 +113,7 @@ int main(int argc, const char **argv) {
   fpga_final();
 
   // display performance measures
-  display_measures(avg_rd, avg_wr, avg_exec, N, dim, iter, inv, sp);
+  display_measures(total_api_time, avg_rd, avg_wr, avg_exec, N, dim, iter, inv, sp);
 
   return EXIT_SUCCESS;
 }

examples/fft3d.c

@@ -24,6 +24,7 @@ int main(int argc, const char **argv) {
   fpga_t timing = {0.0, 0.0, 0.0, 0};
   int use_svm = 0, use_emulator = 0;
   double avg_rd = 0.0, avg_wr = 0.0, avg_exec = 0.0;
+  double temp_timer = 0.0, total_api_time = 0.0;
   bool status = true;
 
   struct argparse_option options[] = {
@@ -73,11 +74,15 @@ int main(int argc, const char **argv) {
 
       if(use_bram == 1){
         // use bram for 3d Transpose
+        temp_timer = getTimeinMilliseconds();
         timing = fftfpgaf_c2c_3d_bram(N, inp, out, inv);
+        total_api_time += getTimeinMilliseconds() - temp_timer;
       }
       else{
         // use ddr for 3d Transpose
+        temp_timer = getTimeinMilliseconds();
         timing = fftfpgaf_c2c_3d_ddr(N, inp, out, inv);
+        total_api_time += getTimeinMilliseconds() - temp_timer;
       }
 
 #ifdef USE_FFTW
@@ -109,7 +114,7 @@ int main(int argc, const char **argv) {
   fpga_final();
 
   // display performance measures
-  display_measures(avg_rd, avg_wr, avg_exec, N, dim, iter, inv, sp);
+  display_measures(total_api_time, avg_rd, avg_wr, avg_exec, N, dim, iter, inv, sp);
 
   return EXIT_SUCCESS;
 }