makefile cuda

build/Makefile

@@ -1,50 +1,38 @@
 # Compiler
-CXX := g++
-NVCC := $(shell command -v nvcc 2>/dev/null)
+CC = g++
+NVCC = nvcc
 
 # Compiler flags
-CXXFLAGS := -Wall -Werror -fpic
-NVCCFLAGS := -Xcompiler -fPIC
+CFLAGS = -Wall -Wextra -std=c++11
+NVCCFLAGS = -std=c++11
 
-# CUDA flags
-CUDAFLAGS := -arch=sm_75
+# Directories
+SRCDIR = ../csrc
+BUILDDIR = ../build
+TARGET = ../build/libtensor.so
 
-# Source files
-SRCS := ../csrc/tensor.cpp
-CU_SRCS := $(wildcard ../csrc/*.cu)
+# Files
+SRCS = $(wildcard $(SRCDIR)/*.cpp)
+CU_SRCS = $(wildcard $(SRCDIR)/*.cu)
+OBJS = $(patsubst $(SRCDIR)/%.cpp, $(BUILDDIR)/%.o, $(SRCS))
+CU_OBJS = $(patsubst $(SRCDIR)/%.cu, $(BUILDDIR)/%.cu.o, $(CU_SRCS))
 
-# Object files
-OBJS := $(SRCS:.cpp=.o)
+# CUDA flags and libraries
+CUDAFLAGS = -arch=sm_75
+CUDALIBS = -lcudart
 
-# CUDA object files (only if nvcc is available)
-ifeq ($(NVCC),)
-CU_OBJS :=
-else
-CU_OBJS := $(CU_SRCS:%.cu=%.cu.o)
-endif
-
-# Target shared object file
-TARGET := libtensor.so
-
-# Default target
-all: $(TARGET)
-
-# Rule to build the shared object file
+# Rule to build the target
 $(TARGET): $(OBJS) $(CU_OBJS)
-	$(CXX) -shared -o $@ $^
+	$(NVCC) --shared -o $(TARGET) $(OBJS) $(CU_OBJS) $(CUDALIBS)
 
-# Rule to compile each source file into an object file
-%.o: %.cpp
-	$(CXX) $(CXXFLAGS) -c $< -o $@
+# Rule to compile C++ source files
+$(BUILDDIR)/%.o: $(SRCDIR)/%.cpp
+	$(CC) $(CFLAGS) -fPIC -c $< -o $@
 
-# Rule to compile each CUDA source file into an object file (only if nvcc is available)
-%.cu.o: %.cu
-ifneq ($(NVCC),)
-	$(NVCC) $(NVCCFLAGS) $(CUDAFLAGS) -c $< -o $@
-else
-	@echo "No CUDA compiler found, skipping CUDA compilation of $<"
-endif
+# Rule to compile CUDA source files
+$(BUILDDIR)/%.cu.o: $(SRCDIR)/%.cu
+	$(NVCC) $(NVCCFLAGS) $(CUDAFLAGS) -Xcompiler -fPIC -c $< -o $@
 
 # Clean rule
 clean:
-	$(RM) $(OBJS) $(CU_OBJS) $(TARGET)
+	rm -f $(BUILDDIR)/*.o $(TARGET)

csrc/cuda.cu

@@ -1,5 +1,7 @@
-__global__ void add_tensor_cuda(float* data1, float* data2, float* result_data, int size) {
-    int i = blockIdx.x * blockDim.x + threadIdx.x;
-    if (i < size) 
-        result_data[i] = data1[i] + data2[i];
-}
+#ifdef __CUDACC__
+    __global__ void add_tensor_cuda(float* data1, float* data2, float* result_data, int size) {
+        int i = blockIdx.x * blockDim.x + threadIdx.x;
+        if (i < size) 
+            result_data[i] = data1[i] + data2[i];
+    }
+#endif

csrc/cuda.h

@@ -1,7 +1,7 @@
 #ifndef CUDA_KERNEL_H_
 #define CUDA_KERNEL_H_
 
-#ifdef __NVCC__
+#ifdef __CUDACC__
     __global__ void add_tensor_cuda(float* data1, float* data2, float* result_data, int size);
 #endif
 

csrc/tensor.cpp

@@ -74,7 +74,8 @@ extern "C" {
     }
 
     void to_device(Tensor* tensor, char* device) {
-        #ifdef __NVCC__
+        printf("Sending tensor to device: %s\n", device);
+        #ifdef __CUDACC__
             if ((strcmp(device, "cuda") == 0) && (strcmp(tensor->device, "cpu") == 0)) {
                 
                 float* data_tmp;
@@ -84,6 +85,7 @@ extern "C" {
 
                 free(tensor_data);
                 tensor->data = data_tmp;
+                tensor->device = device;
             }
 
             else if ((strcmp(device, "cpu") == 0) && (strcmp(tensor->device, "cuda") == 0)) {
@@ -93,7 +95,10 @@ extern "C" {
                 cudaFree(tensor->data);
 
                 tensor->data = data_tmp;
+                tensor->device = device;
             }
+        #else
+            printf("Warning: CUDA is not available. Cannot perform GPU operations.\n");
         #endif
     }
 
@@ -107,7 +112,7 @@ extern "C" {
         }
 
         if (strcmp(tensor1->device, tensor2->device) != 0) {
-            fprintf(stderr, "Tensors must be on the same device\n");
+            fprintf(stderr, "Tensors must be on the same device: %s and %s\n", tensor1->device, tensor2->device);
             exit(1);
         }
 
@@ -165,7 +170,7 @@ extern "C" {
             shape[i] = tensor1->shape[i];
         }
         
-        #ifdef __NVCC__
+        #ifdef __CUDACC__
             if (strcmp(tensor1->device, "cuda") != 0) {
 
                 float* result_data;
@@ -289,7 +294,7 @@ extern "C" {
         }
 
         if (strcmp(tensor1->device, tensor2->device) != 0) {
-            fprintf(stderr, "Tensors must be on the same device\n");
+            fprintf(stderr, "Tensors must be on the same device: %s and %s\n", tensor1->device, tensor2->device);
             exit(1);
         }
 
@@ -369,7 +374,7 @@ extern "C" {
         }
 
         if (strcmp(tensor1->device, tensor2->device) != 0) {
-            fprintf(stderr, "Tensors must be on the same device\n");
+            fprintf(stderr, "Tensors must be on the same device: %s and %s\n", tensor1->device, tensor2->device);
             exit(1);
         }
 
@@ -496,5 +501,4 @@ extern "C" {
             stride *= new_shape[i];
         }
     }
-}
-
+}

norch/tensor.py

@@ -233,7 +233,7 @@ if __name__ == "__main__":
     import numpy as np
 
     
-    a = Tensor([[1, 2, 3], [1, 2, 3], [1, 2, 3]])
+    a = Tensor([[1, 2, 3], [1, 2, 3], [1, 2, 3]]).to("cuda")
     print(a ** 2)
     
     """#print(a)