Add task-parallel version of aobench

examples/aobench/Makefile

@@ -1,8 +1,20 @@
 
-CXX=g++ -m64
-CXXFLAGS=-Iobjs/ -O3 -Wall
+ARCH = $(shell uname)
+
+TASK_CXX=../tasks_pthreads.cpp
+TASK_LIB=-lpthread
+
+ifeq ($(ARCH), Darwin)
+  TASK_CXX=../tasks_gcd.cpp
+  TASK_LIB=
+endif
+
+TASK_OBJ=$(addprefix objs/, $(subst ../,, $(TASK_CXX:.cpp=.o)))
+
+CXX=g++
+CXXFLAGS=-Iobjs/ -O3 -Wall -m64
 ISPC=ispc
-ISPCFLAGS=-O2 --fast-math --arch=x86-64
+ISPCFLAGS=-O2 --target=sse4 --arch=x86-64
 
 default: ao
 
@@ -14,12 +26,15 @@ dirs:
 clean:
 	/bin/rm -rf objs *~ ao
 
-ao: dirs objs/ao.o objs/ao_serial.o objs/ao_ispc.o
-	$(CXX) $(CXXFLAGS) -o $@ objs/ao.o objs/ao_ispc.o objs/ao_serial.o -lm -lpthread
+ao: dirs objs/ao.o objs/ao_serial.o objs/ao_ispc.o $(TASK_OBJ)
+	$(CXX) $(CXXFLAGS) -o $@ objs/ao.o objs/ao_ispc.o objs/ao_serial.o $(TASK_OBJ) -lm $(TASK_LIB)
 
 objs/%.o: %.cpp
 	$(CXX) $< $(CXXFLAGS) -c -o $@
 
+objs/%.o: ../%.cpp
+	$(CXX) $< $(CXXFLAGS) -c -o $@
+
 objs/ao.o: objs/ao_ispc.h 
 
 objs/%_ispc.h objs/%_ispc.o: %.ispc

examples/aobench/ao.cpp

@@ -173,10 +173,30 @@ int main(int argc, char **argv)
     }
 
     // Report results and save image
-    printf("[aobench ispc]:\t\t\t[%.3f] M cycles (%d x %d image)\n", minTimeISPC, 
-           width, height);
+    printf("[aobench ispc]:\t\t\t[%.3f] M cycles (%d x %d image)\n", 
+           minTimeISPC, width, height);
     savePPM("ao-ispc.ppm", width, height); 
 
+    //
+    // Run the ispc + tasks path, test_iterations times, and report the
+    // minimum time for any of them.
+    //
+    double minTimeISPCTasks = 1e30;
+    for (unsigned int i = 0; i < test_iterations; i++) {
+        memset((void *)fimg, 0, sizeof(float) * width * height * 3);
+        assert(NSUBSAMPLES == 2);
+
+        reset_and_start_timer();
+        ao_ispc_tasks(width, height, NSUBSAMPLES, fimg);
+        double t = get_elapsed_mcycles();
+        minTimeISPCTasks = std::min(minTimeISPCTasks, t);
+    }
+
+    // Report results and save image
+    printf("[aobench ispc + tasks]:\t\t[%.3f] M cycles (%d x %d image)\n", 
+           minTimeISPCTasks, width, height);
+    savePPM("ao-ispc-tasks.ppm", width, height); 
+
     //
     // Run the serial path, again test_iteration times, and report the
     // minimum time.
@@ -193,7 +213,8 @@ int main(int argc, char **argv)
     // Report more results, save another image...
     printf("[aobench serial]:\t\t[%.3f] M cycles (%d x %d image)\n", minTimeSerial, 
            width, height);
-    printf("\t\t\t\t(%.2fx speedup from ISPC)\n", minTimeSerial / minTimeISPC);
+    printf("\t\t\t\t(%.2fx speedup from ISPC, %.2fx speedup from ISPC + tasks)\n", 
+           minTimeSerial / minTimeISPC, minTimeSerial / minTimeISPCTasks);
     savePPM("ao-serial.ppm", width, height); 
         
     return 0;

examples/aobench/ao.ispc

@@ -203,8 +203,9 @@ ambient_occlusion(reference Isect isect, reference Plane plane,
 /* Compute the image for the scanlines from [y0,y1), for an overall image
    of width w and height h.
  */
-void ao_scanlines(uniform int y0, uniform int y1, uniform int w, uniform int h, 
-                  uniform int nsubsamples, reference uniform float image[]) {
+static void ao_scanlines(uniform int y0, uniform int y1, uniform int w, 
+                         uniform int h,  uniform int nsubsamples, 
+                         reference uniform float image[]) {
     static Plane plane = { { 0.0f, -0.5f, 0.0f }, { 0.f, 1.f, 0.f } };
     static Sphere spheres[3] = {
         { { -2.0f, 0.0f, -3.5f }, 0.5f },
@@ -315,3 +316,18 @@ export void ao_ispc(uniform int w, uniform int h, uniform int nsubsamples,
                     uniform float image[]) {
     ao_scanlines(0, h, w, h, nsubsamples, image);
 }
+
+
+static void task ao_task(uniform int y0, uniform int y1, uniform int width,
+                         uniform int height, uniform int nsubsamples, 
+                         uniform float image[]) {
+    ao_scanlines(y0, y1, width, height, nsubsamples, image);
+}
+
+
+export void ao_ispc_tasks(uniform int w, uniform int h, uniform int nsubsamples, 
+                          uniform float image[]) {
+    uniform int dy = 1;
+    for (uniform int y = 0; y < h; y += dy)
+        launch < ao_task(y, y+dy, w, h, nsubsamples, image) >;
+}

examples/aobench/aobench.vcxproj

@@ -1,4 +1,4 @@
-<?xml version="1.0" encoding="utf-8"?>
+<?xml version="1.0" encoding="utf-8"?>
 <Project DefaultTargets="Build" ToolsVersion="4.0" xmlns="http://schemas.microsoft.com/developer/msbuild/2003">
   <ItemGroup Label="ProjectConfigurations">
     <ProjectConfiguration Include="Debug|Win32">
@@ -21,6 +21,7 @@
   <ItemGroup>
     <ClCompile Include="ao.cpp" />
     <ClCompile Include="ao_serial.cpp" />
+    <ClCompile Include="../tasks_concrt.cpp" />
   </ItemGroup>
   <ItemGroup>
     <CustomBuild Include="ao.ispc">