added cpu compilation path

examples/portable/aobench/Makefile_cpu

@@ -5,4 +5,4 @@ ISPC_SRC=ao.ispc
 ISPC_IA_TARGETS=avx1-i32x8
 ISPC_ARM_TARGETS=neon
 
-include ../common.mk
+include ../common_cpu.mk

examples/portable/aobench/Makefile_gpu

@@ -1,7 +1,7 @@
 PROG=ao
 ISPC_SRC=ao.ispc
 CU_SRC=ao.cu
-CXX_SRC=ao.cpp  ao_serial.cpp
+CXX_SRC=ao.cpp 
 PTXCC_REGMAX=64
 #ISPC_FLAGS= --opt=disable-uniform-control-flow
 

examples/portable/aobench/ao.cpp

@@ -58,8 +58,6 @@
 
 #define NSUBSAMPLES        2
 
-extern void ao_serial(int w, int h, int nsubsamples, float image[]);
-
 static unsigned int test_iterations[] = {3, 7, 1};
 static unsigned int width, height;
 static unsigned char *img;
@@ -123,13 +121,8 @@ int main(int argc, char **argv)
     }
 
     // Allocate space for output images
-#if 0
     img = new unsigned char[width * height * 3];
     fimg = new float[width * height * 3];
-#else
-    ispc_malloc((void**) &img, sizeof(unsigned char)*width*height*3);
-    ispc_malloc((void**)&fimg, sizeof(        float)*width*height*3);
-#endif
 
     //
     // Run the ispc + tasks path, test_iterations times, and report the
@@ -137,7 +130,7 @@ int main(int argc, char **argv)
     //
     double minTimeISPCTasks = 1e30;
     for (unsigned int i = 0; i < test_iterations[1]; i++) {
-        ispc_memset((void *)fimg, 0, sizeof(float) * width * height * 3);
+        ispc_memset(fimg, 0, sizeof(float) * width * height * 3);
         assert(NSUBSAMPLES == 2);
 
         reset_and_start_timer();
@@ -152,29 +145,8 @@ int main(int argc, char **argv)
            minTimeISPCTasks, width, height);
     savePPM("ao-ispc-tasks.ppm", width, height); 
 
-    //
+    delete img;
-    // Run the serial path, again test_iteration times, and report the
+    delete fimg;
-    // minimum time.
-    //
-    double minTimeSerial = 1e30;
-    for (unsigned int i = 0; i < test_iterations[2]; i++) {
-        ispc_memset((void *)fimg, 0, sizeof(float) * width * height * 3);
-        reset_and_start_timer();
-        ao_serial(width, height, NSUBSAMPLES, fimg);
-        double t = get_elapsed_msec();
-        printf("@time of serial run:\t\t\t\t[%.3f] msec\n", t);
-        minTimeSerial = std::min(minTimeSerial, t);
-    }
-
-    // Report more results, save another image...
-    printf("[aobench serial]:\t\t[%.3f] msec (%d x %d image)\n", minTimeSerial, 
-           width, height);
-    printf("\t\t\t\t(%.2fx speedup from ISPC + tasks)\n", 
-           minTimeSerial / minTimeISPCTasks);
-    savePPM("ao-serial.ppm", width, height); 
-
-    ispc_free(img);
-    ispc_free(fimg);
         
     return 0;
 }

examples/portable/common_cpu.mk

@@ -0,0 +1,120 @@
+
+TASK_CXX=../../tasksys.cpp ../../util/ispc_malloc.cpp
+TASK_LIB=-lpthread
+TASK_OBJ=objs/tasksys.o objs/ispc_malloc.o
+
+CXX=clang++
+CXX=icc -openmp
+CXXFLAGS+=-Iobjs/ -O2 -I../../ -I../../util 
+CXXFLAGS+=-DISPC_USE_OMP
+CC=clang
+CC=icc -openmp
+CCFLAGS+=-Iobjs/ -O2 -I../../ -I../../util
+CXXFLAGS+=-DISPC_USE_OMP
+
+LIBS=-lm $(TASK_LIB) -lstdc++
+ISPC=ispc
+ISPC_FLAGS+=-O2
+ISPC_HEADER=objs/$(ISPC_SRC:.ispc=_ispc.h)
+
+ARCH:=$(shell uname -m | sed -e s/x86_64/x86/ -e s/i686/x86/ -e s/arm.*/arm/ -e s/sa110/arm/)
+
+ifeq ($(ARCH),x86)
+  ISPC_OBJS=$(addprefix objs/, $(ISPC_SRC:.ispc=)_ispc.o)
+  COMMA=,
+  ifneq (,$(findstring $(COMMA),$(ISPC_IA_TARGETS)))
+    #$(info multi-target detected: $(ISPC_IA_TARGETS))
+    ifneq (,$(findstring sse2,$(ISPC_IA_TARGETS)))
+      ISPC_OBJS+=$(addprefix objs/, $(ISPC_SRC:.ispc=)_ispc_sse2.o)
+    endif
+    ifneq (,$(findstring sse4,$(ISPC_IA_TARGETS)))
+      ISPC_OBJS+=$(addprefix objs/, $(ISPC_SRC:.ispc=)_ispc_sse4.o)
+    endif
+    ifneq (,$(findstring avx1-,$(ISPC_IA_TARGETS)))
+      ISPC_OBJS+=$(addprefix objs/, $(ISPC_SRC:.ispc=)_ispc_avx.o)
+    endif
+    ifneq (,$(findstring avx1.1,$(ISPC_IA_TARGETS)))
+      ISPC_OBJS+=$(addprefix objs/, $(ISPC_SRC:.ispc=)_ispc_avx11.o)
+    endif
+    ifneq (,$(findstring avx2,$(ISPC_IA_TARGETS)))
+      ISPC_OBJS+=$(addprefix objs/, $(ISPC_SRC:.ispc=)_ispc_avx2.o)
+    endif
+  endif
+  ISPC_TARGETS=$(ISPC_IA_TARGETS)
+  ARCH_BIT:=$(shell getconf LONG_BIT)
+  ifeq ($(ARCH_BIT),32)
+    ISPC_FLAGS += --arch=x86
+    CXXFLAGS += -m32
+    CCFLAGS += -m32
+  else
+    ISPC_FLAGS += --arch=x86-64
+    CXXFLAGS += -m64
+    CCFLAGS += -m64
+  endif
+else ifeq ($(ARCH),arm)
+  ISPC_OBJS=$(addprefix objs/, $(ISPC_SRC:.ispc=_ispc.o))
+  ISPC_TARGETS=$(ISPC_ARM_TARGETS)
+else
+  $(error Unknown architecture $(ARCH) from uname -m)
+endif
+
+CPP_OBJS=$(addprefix objs/, $(CPP_SRC:.cpp=.o))
+CC_OBJS=$(addprefix objs/, $(CC_SRC:.c=.o))
+OBJS=$(CPP_OBJS) $(CC_OBJS) $(TASK_OBJ) $(ISPC_OBJS)
+
+default: $(EXAMPLE)
+
+all: $(EXAMPLE) $(EXAMPLE)-sse4 $(EXAMPLE)-generic16 $(EXAMPLE)-scalar
+
+.PHONY: dirs clean
+
+dirs:
+	/bin/mkdir -p objs/
+
+objs/%.cpp objs/%.o objs/%.h: dirs
+
+clean:
+	/bin/rm -rf objs *~ $(EXAMPLE) $(EXAMPLE)-sse4 $(EXAMPLE)-generic16 ref test
+
+$(EXAMPLE): $(OBJS)
+	$(CXX) $(CXXFLAGS) -o $@ $^ $(LIBS)
+
+objs/%.o: %.cpp dirs $(ISPC_HEADER)
+	$(CXX) $< $(CXXFLAGS) -c -o $@
+
+objs/%.o: %.c dirs $(ISPC_HEADER)
+	$(CC) $< $(CCFLAGS) -c -o $@
+
+objs/%.o: ../../%.cpp dirs
+	$(CXX) $< $(CXXFLAGS) -c -o $@
+objs/%.o: ../../util/%.cpp dirs
+	$(CXX) $< $(CXXFLAGS) -c -o $@
+
+objs/$(EXAMPLE).o: objs/$(EXAMPLE)_ispc.h dirs
+
+objs/%_ispc.h objs/%_ispc.o objs/%_ispc_sse2.o objs/%_ispc_sse4.o objs/%_ispc_avx.o objs/%_ispc_avx11.o objs/%_ispc_avx2.o: %.ispc dirs
+	$(ISPC) $(ISPC_FLAGS) --target=$(ISPC_TARGETS) $< -o objs/$*_ispc.o -h objs/$*_ispc.h
+
+objs/$(ISPC_SRC:.ispc=)_sse4.cpp: $(ISPC_SRC)
+	$(ISPC) $(ISPC_FLAGS) $< -o $@ --target=generic-4 --emit-c++ --c++-include-file=sse4.h
+
+objs/$(ISPC_SRC:.ispc=)_sse4.o: objs/$(ISPC_SRC:.ispc=)_sse4.cpp
+	$(CXX) -I../../intrinsics -msse4.2 $< $(CXXFLAGS) -c -o $@
+
+$(EXAMPLE)-sse4: $(CPP_OBJS) objs/$(ISPC_SRC:.ispc=)_sse4.o
+	$(CXX) $(CXXFLAGS) -o $@ $^ $(LIBS)
+
+objs/$(ISPC_SRC:.ispc=)_generic16.cpp: $(ISPC_SRC)
+	$(ISPC) $(ISPC_FLAGS) $< -o $@ --target=generic-16 --emit-c++ --c++-include-file=generic-16.h
+
+objs/$(ISPC_SRC:.ispc=)_generic16.o: objs/$(ISPC_SRC:.ispc=)_generic16.cpp
+	$(CXX) -I../../intrinsics $< $(CXXFLAGS) -c -o $@
+
+$(EXAMPLE)-generic16: $(CPP_OBJS) objs/$(ISPC_SRC:.ispc=)_generic16.o
+	$(CXX) $(CXXFLAGS) -o $@ $^ $(LIBS)
+
+objs/$(ISPC_SRC:.ispc=)_scalar.o: $(ISPC_SRC)
+	$(ISPC) $(ISPC_FLAGS) $< -o $@ --target=generic-1
+
+$(EXAMPLE)-scalar: $(CPP_OBJS) objs/$(ISPC_SRC:.ispc=)_scalar.o
+	$(CXX) $(CXXFLAGS) -o $@ $^ $(LIBS)

examples/tasksys.cpp

@@ -960,17 +960,22 @@ InitTaskSystem() {
 
 inline void
 TaskGroup::Launch(int baseIndex, int count) {
-#pragma omp parallel for
+#pragma omp parallel
-    for(int i = 0; i < count; i++) {
+  {
+    const int threadIndex = omp_get_thread_num();
+    const int threadCount = omp_get_num_threads();
+
+#pragma omp for schedule(runtime)
+    for(int i = 0; i < count; i++) 
+    {
         TaskInfo *ti = GetTaskInfo(baseIndex + i);
 
         // Actually run the task. 
-        int threadIndex = omp_get_thread_num();
-        int threadCount = omp_get_num_threads();
         ti->func(ti->data, threadIndex, threadCount, ti->taskIndex, ti->taskCount(),
             ti->taskIndex0(), ti->taskIndex1(), ti->taskIndex2(),
             ti->taskCount0(), ti->taskCount1(), ti->taskCount2());
     }
+  }
 }
 
 inline void