del tensor free memory

README.md

@@ -29,7 +29,7 @@ x1 = norch.Tensor([[1, 2],
                   [3, 4]], requires_grad=True).to("cuda")
 
 x2 = norch.Tensor([[4, 3], 
-                  [2, 1]], requires_grad=True).to("cuda)
+                  [2, 1]], requires_grad=True).to("cuda")
 
 x3 = x1 @ x2
 result = x3.sum()

norch/csrc/cuda.cu

@@ -42,6 +42,10 @@ __host__ void cuda_to_cpu(Tensor* tensor) {
     strcpy(tensor->device, device_str); 
 }
 
+__host__ void free_cuda(float* data) {
+    cudaFree(data);
+}
+
 __global__ void add_tensor_cuda_kernel(float* data1, float* data2, float* result_data, int size) {
     
     int i = blockIdx.x * blockDim.x + threadIdx.x;

norch/csrc/cuda.h

@@ -4,6 +4,7 @@
 
     __host__ void cpu_to_cuda(Tensor* tensor, int device_id);
     __host__ void cuda_to_cpu(Tensor* tensor);
+    __host__ void free_cuda(float* data);
     
     __global__ void add_tensor_cuda_kernel(float* data1, float* data2, float* result_data, int size);
     __host__ void add_tensor_cuda(Tensor* tensor1, Tensor* tensor2, float* result_data);

norch/csrc/tensor.cpp

@@ -16,28 +16,27 @@ extern "C" {
             fprintf(stderr, "Memory allocation failed\n");
             exit(1);
         }
-        tensor->data = data;
-        tensor->shape = shape;
         tensor->ndim = ndim;
 
-        tensor->device = (char*)malloc(strlen(device) + 1);
-        if (device != NULL) {
-            strcpy(tensor->device, device);
-        } else {
-            fprintf(stderr, "Memory allocation failed\n");
-            exit(-1);
-        }
-
         tensor->size = 1;
         for (int i = 0; i < ndim; i++) {
             tensor->size *= shape[i];
         }
 
+        tensor->device = strdup(device);
         tensor->strides = (int*)malloc(ndim * sizeof(int));
-        if (tensor->strides == NULL) {
+        tensor->shape = (int*)malloc(ndim * sizeof(int));
+        tensor->data = (float*)malloc(tensor->size * sizeof(float));
+
+        if (tensor->device == NULL || tensor->strides == NULL || tensor->shape == NULL || tensor->data == NULL) {
             fprintf(stderr, "Memory allocation failed\n");
             exit(1);
         }
+
+        memcpy(tensor->shape, shape, tensor->ndim * sizeof(int));
+        strcpy(tensor->device, device);
+        memcpy(tensor->data, data, tensor->size * sizeof(float));
+
         int stride = 1;
         for (int i = ndim - 1; i >= 0; i--) {
             tensor->strides[i] = stride;
@@ -47,6 +46,43 @@ extern "C" {
         return tensor;
     }
 
+    void delete_tensor(Tensor* tensor) {
+        if (tensor != NULL) {
+            
+            if (tensor->strides != NULL) {
+                free(tensor->strides);
+                tensor->strides = NULL;
+            }
+            
+            if (tensor->shape != NULL) {
+                free(tensor->shape);
+                tensor->shape = NULL;
+            }
+            
+            if (tensor->data != NULL) {
+                if (strcmp(tensor->device, "cpu") == 0) {
+                    free(tensor->data);
+                } else {
+                    free_cuda(tensor->data);
+                }
+                tensor->data = NULL;
+            }
+
+            if (tensor->device != NULL) {
+                free(tensor->device);
+                tensor->device = NULL;
+            }
+
+            if (tensor->strides != NULL) {
+                free(tensor->strides);
+                tensor->strides = NULL;
+            }
+            
+            free(tensor);
+        }
+    }
+
+
     float get_item(Tensor* tensor, int* indices) {
         int index = 0;
         for (int i = 0; i < tensor->ndim; i++) {

norch/csrc/tensor.h

@@ -5,8 +5,6 @@ typedef struct {
     float* data;
     int* strides;
     int* shape;
-    int* strides_cuda;
-    int* shape_cuda;
     int ndim;
     int size;
     char* device;
@@ -14,6 +12,7 @@ typedef struct {
 
 extern "C" {
     Tensor* create_tensor(float* data, int* shape, int ndim, char* device);
+    void delete_tensor(Tensor* tensor);
     float get_item(Tensor* tensor, int* indices);
     Tensor* add_tensor(Tensor* tensor1, Tensor* tensor2);
     Tensor* sum_tensor(Tensor* tensor, int axis, bool keepdims);

norch/optim/optimizers/sgd.py

@@ -26,3 +26,4 @@ class SGD(Optimizer):
 
             parameter.detach()
             velocity.detach()
+            del parameter

norch/tensor.py

@@ -81,6 +81,12 @@ class Tensor:
         flat_data, shape = flatten_recursively(nested_list)
         return flat_data, shape
     
+    def __del__(self):
+        if self.tensor is not None:
+            Tensor._C.delete_tensor.argtypes = [ctypes.POINTER(CTensor)]
+            Tensor._C.delete_tensor.restype = None
+            Tensor._C.delete_tensor(self.tensor)
+    
     def __setattr__(self, name, value):
         if name == 'grad':
             for hook in self.hooks: