+added wc-timer

examples_cuda/mandelbrot_tasks3d/Makefile

@@ -2,7 +2,7 @@
 EXAMPLE=mandelbrot_tasks3d
 CPP_SRC=mandelbrot_tasks3d.cpp mandelbrot_tasks_serial.cpp
 ISPC_SRC=mandelbrot_tasks3d.ispc
-ISPC_IA_TARGETS=avx,sse2,sse4
+ISPC_IA_TARGETS=avx
 ISPC_ARM_TARGETS=neon
 
 include ../common.mk

examples_cuda/mandelbrot_tasks3d/mandel_cu.cpp

@@ -43,6 +43,21 @@
 #include <string.h>
 #include "../timing.h"
 
+#include <sys/time.h>
+
+
+double rtc(void)
+{
+  struct timeval Tvalue;
+  double etime;
+  struct timezone dummy;
+
+  gettimeofday(&Tvalue,&dummy);
+  etime =  (double) Tvalue.tv_sec +
+    1.e-6*((double) Tvalue.tv_usec);
+  return etime;
+}
+
 #include <iostream>
 #include <cuda.h>
 #include <vector>
@@ -259,7 +274,7 @@ extern "C"
 
     fprintf(stderr, " module_name= %s \n", module_name);
     fprintf(stderr, " func_name= %s \n", func_name);
-//    fprintf(stderr, " ptx= %s \n", module);
+    //    fprintf(stderr, " ptx= %s \n", module);
     fprintf(stderr, " x0= %g  \n", *((float*)(func_args[0])));
     fprintf(stderr, " dx= %g  \n", *((float*)(func_args[1])));
     fprintf(stderr, " y0= %g  \n", *((float*)(func_args[2])));
@@ -361,15 +376,18 @@ int main(int argc, char *argv[]) {
   // time of three runs.
   //
   double minISPC = 1e30;
+#if 1
   for (int i = 0; i < 3; ++i) {
     // Clear out the buffer
     for (unsigned int i = 0; i < width * height; ++i)
       buf[i] = 0;
     reset_and_start_timer();
+    const double t0 = rtc();
     mandelbrot_ispc(x0, y0, x1, y1, width, height, maxIterations, (int*)d_buf);
-        double dt = get_elapsed_mcycles();
+    double dt = rtc() - t0; //get_elapsed_mcycles();
     minISPC = std::min(minISPC, dt);
   }
+#endif
 
   memcpyD2H(buf, d_buf, bufsize);
   deviceFree(d_buf);
@@ -388,8 +406,9 @@ int main(int argc, char *argv[]) {
     for (unsigned int i = 0; i < width * height; ++i)
       buf[i] = 0;
     reset_and_start_timer();
+    const double t0 = rtc();
     mandelbrot_serial(x0, y0, x1, y1, width, height, maxIterations, buf);
-        double dt = get_elapsed_mcycles();
+    double dt = rtc() - t0; //get_elapsed_mcycles();
     minSerial = std::min(minSerial, dt);
   }
 

examples_cuda/stencil/stencil_cu.cpp

@@ -49,6 +49,20 @@ using namespace ispc;
 #include <iostream>
 #include <cuda.h>
 #include "drvapi_error_string.h"
+#include <sys/time.h>
+
+
+double rtc(void)
+{
+  struct timeval Tvalue;
+  double etime;
+  struct timezone dummy;
+
+  gettimeofday(&Tvalue,&dummy);
+  etime =  (double) Tvalue.tv_sec +
+    1.e-6*((double) Tvalue.tv_usec);
+  return etime;
+}
 
 #define checkCudaErrors(err)  __checkCudaErrors (err, __FILE__, __LINE__)
 // These are the inline versions for all of the SDK helper functions
@@ -133,7 +147,7 @@ void memcpyH2D(CUdeviceptr d_buf, void * h_buf, const size_t size)
 }
 #define deviceLaunch(func,nbx,nby,nbz,params) \
   checkCudaErrors(cuFuncSetCacheConfig((func), CU_FUNC_CACHE_PREFER_L1)); \
-  checkCudaErrors( \
+checkCudaErrors( \
     cuLaunchKernel( \
       (func), \
       ((nbx-1)/(128/32)+1), (nby), (nbz), \
@@ -272,6 +286,7 @@ int main() {
   // the minimum time of three runs.
   //
   double minTimeISPC = 1e30;
+#if 0
   for (int i = 0; i < 3; ++i) {
     reset_and_start_timer();
     loop_stencil_ispc(0, 6, width, Nx - width, width, Ny - width,
@@ -280,6 +295,7 @@ int main() {
     double dt = get_elapsed_mcycles();
     minTimeISPC = std::min(minTimeISPC, dt);
   }
+#endif
 
   printf("[stencil ispc 1 core]:\t\t[%.3f] million cycles\n", minTimeISPC);
 
@@ -296,10 +312,11 @@ int main() {
   double minTimeISPCTasks = 1e30;
   for (int i = 0; i < 3; ++i) {
     reset_and_start_timer();
+    const double t0 = rtc();
     loop_stencil_ispc_tasks(0, 6, width, Nx - width, width, Ny - width,
         width, Nz - width, Nx, Ny, Nz, (double*)d_coeff, (double*)d_vsq,
         (double*)d_Aispc0, (double*)d_Aispc1);
-    double dt = get_elapsed_mcycles();
+    double dt = rtc() - t0; //get_elapsed_mcycles();
     minTimeISPCTasks = std::min(minTimeISPCTasks, dt);
   }
   memcpyD2H(Aispc[1], d_Aispc1, bufsize);

examples_cuda/volume_rendering/Makefile

@@ -1,7 +1,7 @@
 
 EXAMPLE=volume
 CPP_SRC=volume.cpp volume_serial.cpp
-ISPC_SRC=volume.ispc
+ISPC_SRC=volume1.ispc
 ISPC_IA_TARGETS=avx
 
 include ../common.mk

examples_cuda/volume_rendering/camera.dat

@@ -1,4 +1,4 @@
-896 1184
+1792 2368
 
 0.000155 0.000000 0.000000 -0.069927
 0.000000 -0.000155 0.000000 0.093236

examples_cuda/volume_rendering/volume.cpp

@@ -44,6 +44,21 @@
 #include "volume_ispc.h"
 using namespace ispc;
 
+#include <sys/time.h>
+
+
+static inline double rtc(void)
+{
+  struct timeval Tvalue;
+  double etime;
+  struct timezone dummy;
+
+  gettimeofday(&Tvalue,&dummy);
+  etime =  (double) Tvalue.tv_sec +
+    1.e-6*((double) Tvalue.tv_usec);
+  return etime;
+}
+
 extern void volume_serial(float density[], int nVoxels[3], 
     const float raster2camera[4][4],
     const float camera2world[4][4], 
@@ -70,8 +85,8 @@ writePPM(float *buf, int width, int height, const char *fn) {
 
 
 /* Load image and viewing parameters from a camera data file.
-   FIXME: we should add support to be able to specify viewing parameters
+FIXME: we should add support to be able to specify viewing parameters
-   in the program here directly. */
+in the program here directly. */
 static void
 loadCamera(const char *fn, int *width, int *height, float raster2camera[4][4],
     float camera2world[4][4]) {
@@ -156,6 +171,7 @@ int main(int argc, char *argv[]) {
   // time of three runs.
   //
   double minISPC = 1e30;
+#if 0
   for (int i = 0; i < 3; ++i) {
     reset_and_start_timer();
     volume_ispc(density, n, raster2camera, camera2world,
@@ -163,6 +179,7 @@ int main(int argc, char *argv[]) {
     double dt = get_elapsed_mcycles();
     minISPC = std::min(minISPC, dt);
   }
+#endif
 
   printf("[volume ispc 1 core]:\t\t[%.3f] million cycles\n", minISPC);
   writePPM(image, width, height, "volume-ispc-1core.ppm");
@@ -178,9 +195,10 @@ int main(int argc, char *argv[]) {
   double minISPCtasks = 1e30;
   for (int i = 0; i < 3; ++i) {
     reset_and_start_timer();
+    const double t0 = rtc();
     volume_ispc_tasks(density, n, raster2camera, camera2world,
         width, height, image);
-        double dt = get_elapsed_mcycles();
+    double dt = rtc() - t0; //get_elapsed_mcycles();
     minISPCtasks = std::min(minISPCtasks, dt);
   }
 

examples_cuda/volume_rendering/volume_cu.cpp

@@ -44,6 +44,19 @@
 #include "volume_ispc.h"
 using namespace ispc;
 
+#include <sys/time.h>
+static inline double rtc(void)
+{
+  struct timeval Tvalue;
+  double etime;
+  struct timezone dummy;
+
+  gettimeofday(&Tvalue,&dummy);
+  etime =  (double) Tvalue.tv_sec +
+    1.e-6*((double) Tvalue.tv_usec);
+  return etime;
+}
+
 #include <cassert>
 #include <iostream>
 #include <cuda.h>
@@ -414,6 +427,7 @@ int main(int argc, char *argv[]) {
     double minISPCtasks = 1e30;
     for (int i = 0; i < 3; ++i) {
         reset_and_start_timer();
+        const double t0 = rtc();
         volume_ispc_tasks(
               (float*)d_density, 
               (int*)d_n, 
@@ -421,7 +435,7 @@ int main(int argc, char *argv[]) {
               (float(*)[4])d_camera2world,
               width, height, 
               (float*)d_image);
-        double dt = get_elapsed_mcycles();
+        double dt = rtc() - t0; //get_elapsed_mcycles();
         minISPCtasks = std::min(minISPCtasks, dt);
     }