# HG changeset patch
# User Shohei KOKUBO <e105744@ie.u-ryukyu.ac.jp>
# Date 1391120107 -32400
# Node ID 4eefec26e3e2a22b48a58df7a09aed471303dee3
# Parent  777bdbf6c072d87a84396726462f55fe086ea156
add file

diff -r 777bdbf6c072 -r 4eefec26e3e2 TaskManager/Cuda/CudaScheduler.h
--- a/TaskManager/Cuda/CudaScheduler.h	Fri Jan 31 06:08:13 2014 +0900
+++ b/TaskManager/Cuda/CudaScheduler.h	Fri Jan 31 07:15:07 2014 +0900
@@ -58,3 +58,5 @@
 #define CudaSchedRegister(str, filename, functionname)  \
     cuda_register_task(str, filename, functionname);
 #endif
+
+extern void cuda_register_task(int cmd, const char* filename, const char* functionname);
diff -r 777bdbf6c072 -r 4eefec26e3e2 example/Cuda/Makefile
--- a/example/Cuda/Makefile	Fri Jan 31 06:08:13 2014 +0900
+++ b/example/Cuda/Makefile	Fri Jan 31 07:15:07 2014 +0900
@@ -19,9 +19,6 @@
 .cc.o:
 	$(CC) $(CFLAGS) $(INCLUDE) -c $< -o $@
 
-.SUFFIXES: .cu .ptx
-	$(NVCC) $(NVCCFLAGS) $< $@
-
 all: $(TARGET)
 
 $(TARGET): $(OBJS) $(TASK_OBJS) $(CUDA_SRCS_TMP)
diff -r 777bdbf6c072 -r 4eefec26e3e2 example/multiply/Makefile
--- a/example/multiply/Makefile	Fri Jan 31 06:08:13 2014 +0900
+++ b/example/multiply/Makefile	Fri Jan 31 07:15:07 2014 +0900
@@ -20,6 +20,10 @@
 	@echo "Make for GPU (open cl)"
 	@$(MAKE) -f Makefile.gpu
 
+cuda: FORCE
+	@echo "Make for GPU (cuda)"
+	@$(MAKE) -f Makefile.cuda
+
 FORCE:
 
 clean:
diff -r 777bdbf6c072 -r 4eefec26e3e2 example/multiply/Makefile.cuda
--- /dev/null	Thu Jan 01 00:00:00 1970 +0000
+++ b/example/multiply/Makefile.cuda	Fri Jan 31 07:15:07 2014 +0900
@@ -0,0 +1,49 @@
+include ./Makefile.def
+
+SRCS_TMP = $(wildcard *.cc)
+SRCS_EXCLUDE =  # 除外するファイルを書く
+SRCS = $(filter-out $(SRCS_EXCLUDE),$(SRCS_TMP))
+OBJS = $(SRCS:.cc=.o)
+
+TASK_DIR  = ppe
+CUDA_TASK_DIR = cuda
+
+TASK_SRCS_TMP = $(wildcard $(TASK_DIR)/*.cc)
+TASK_SRCS_EXCLUDE = 
+TASK_SRCS = $(filter-out $(TASK_DIR)/$(TASK_SRCS_EXCLUDE),$(TASK_SRCS_TMP)) $(wildcard $(CUDA_TASK_DIR)/*.cc)
+TASK_OBJS = $(TASK_SRCS:.cc=.o)
+
+CUDA_SRCS_TMP = $(wildcard $(CUDA_TASK_DIR)/*.cu)
+CUDA_SRCS_EXCLUDE = # 除外するファイルを書く
+CUDA_SRCS = $(filter-out $(CUDA_TASK_DIR)/$(CUDA_SRCS_EXCLUDE),$(CUDA_SRCS_TMP))
+CUDA_OBJS = $(CUDA_SRCS:.cu=.ptx)
+
+CFLAGS += -D__CERIUM_GPU__
+LIBS += `sdl-config --libs` -lCudaManager -F/Library/Frameworks -framework CUDA
+
+INCLUDE += -I/Developer/NVIDIA/CUDA-5.5/include
+
+NVCC = nvcc
+NVCCFLAGS = -ptx -arch=sm_20
+
+.SUFFIXES: .cc .o
+
+.cc.o:
+	$(CC) $(CFLAGS) $(INCLUDE) -c $< -o $@
+
+all: $(TARGET) 
+
+$(TARGET): $(OBJS) $(TASK_OBJS) $(CUDA_SRCS_TMP)
+	$(NVCC) $(NVCCFLAGS) $(CUDA_SRCS_TMP)
+	$(CC) -o $@ $(OBJS) $(TASK_OBJS) $(LIBS)
+
+link:
+	$(CC) -o $(TARGET) $(OBJS) $(TASK_OBJS) $(LIBS)
+
+debug: $(TARGET)
+	sudo ppu-gdb ./$(TARGET) 
+
+clean:
+	rm -f $(TARGET) $(OBJS) $(TASK_OBJS) $(CUDA_OBJS)
+	rm -f *~ \#*
+	rm -f gpu/*~ gpu/\#*
diff -r 777bdbf6c072 -r 4eefec26e3e2 example/multiply/cuda/gpu_task_init.cc
--- /dev/null	Thu Jan 01 00:00:00 1970 +0000
+++ b/example/multiply/cuda/gpu_task_init.cc	Fri Jan 31 07:15:07 2014 +0900
@@ -0,0 +1,16 @@
+#include "Func.h"
+#include "CudaScheduler.h"
+
+/* 必ずこの位置に書いて */
+
+/**
+ * この関数は ../spe/spe-main と違って
+ * 自分で呼び出せばいい関数なので
+ * 好きな関数名でおk (SchedRegisterTask は必須)
+ */
+
+void
+gpu_task_init(void)
+{
+    CudaSchedRegister(MULTIPLY_TASK, "gpu/Multi.ptx","multi");
+}
diff -r 777bdbf6c072 -r 4eefec26e3e2 example/multiply/cuda/multiply.cu
--- /dev/null	Thu Jan 01 00:00:00 1970 +0000
+++ b/example/multiply/cuda/multiply.cu	Fri Jan 31 07:15:07 2014 +0900
@@ -0,0 +1,6 @@
+extern "C" {
+    __global__ void multi(float* A, float* B, float* C) {
+        int id = blockIdx.x * blockDim.x + threadIdx.x;
+        C[id]=A[id]*B[id];
+    }
+}
diff -r 777bdbf6c072 -r 4eefec26e3e2 example/multiply/main.cc
--- a/example/multiply/main.cc	Fri Jan 31 06:08:13 2014 +0900
+++ b/example/multiply/main.cc	Fri Jan 31 07:15:07 2014 +0900
@@ -71,7 +71,7 @@
 multi_init(TaskManager *manager)
 {
     HTask* multiply;
-    HTask* hoge;
+    //    HTask* hoge;
 
     A = new float[length];
     B = new float[length];
@@ -98,14 +98,12 @@
      */
     multiply->set_inData(0,(memaddr)A, sizeof(float)*length);
     multiply->set_inData(1,(memaddr)B, sizeof(float)*length);
-    multiply->set_inData(2,(memaddr)C,sizeof(float)*length);
     /**
      * Set of OutPut area
      *   add_outData(address of output area, size of output area);
      */
     multiply->set_outData(0,(memaddr)C, sizeof(float)*length);
     // param 0に0~length-1をsetしたtaskをlength個spawnする
-    multiply->set_param(0,(long)1);
     multiply->iterate(length); 
     
     // hoge = manager->create_task(MULTIPLY_TASK);