Fix transpose axis and make contigous

norch/autograd/functions.py

@@ -33,7 +33,7 @@ class MatmulBackward:
 
     def backward(self, gradient):
         x, y = self.input
-        return [gradient @ y.T, x.T @ gradient]
+        return [gradient @ y.transpose(-1,-2), x.transpose(-1,-2) @ gradient]
         
 class PowBackward:
     def __init__(self, x, power):

norch/csrc/cpu.cpp

@@ -116,7 +116,14 @@ void zeros_like_tensor_cpu(Tensor* tensor, float* result_data) {
     }
 }
 
-/*void transpose_tensor_cpu(Tensor* tensor, float* result_data) {
+void transpose_1D_tensor_cpu(Tensor* tensor, float* result_data) {
+
+    for (int i = 0; i < tensor->shape[0]; i++) {
+        result_data[i] = tensor->data[i];
+    }
+}
+
+void transpose_2D_tensor_cpu(Tensor* tensor, float* result_data) {
     int rows = tensor->shape[0];
     int cols = tensor->shape[1];
 
@@ -125,33 +132,17 @@ void zeros_like_tensor_cpu(Tensor* tensor, float* result_data) {
             result_data[j * rows + i] = tensor->data[i * cols + j];
         }
     }
-}*/
+}
 
+void transpose_3D_tensor_cpu(Tensor* tensor, float* result_data) {
+    int depth = tensor->shape[0];
+    int rows = tensor->shape[1];
+    int cols = tensor->shape[2];
 
-void transpose_tensor_cpu(Tensor* tensor, float* result_data) {
-    int* shape = tensor->shape;
-    int ndim = tensor->ndim;
-    int* strides = (int*)malloc(ndim * sizeof(int));
-    int* indices = (int*)calloc(ndim, sizeof(int));
-
-    strides[ndim - 1] = 1;
-    for (int i = ndim - 2; i >= 0; i--) {
-        strides[i] = strides[i + 1] * shape[i + 1];
-    }
-
-    int idx_result;
-    for (int idx_source = 0; idx_source < tensor->size; idx_source++) {
-        idx_result = 0;
-        for (int dim = 0; dim < ndim; dim++) {
-            idx_result += indices[dim] * strides[dim];
-        }
-        result_data[idx_result] = tensor->data[idx_source];
-
-        indices[ndim - 1]++;
-        for (int dim = ndim - 1; dim > 0; dim--) {
-            if (indices[dim] == shape[dim]) {
-                indices[dim] = 0;
-                indices[dim - 1]++;
+    for (int i = 0; i < depth; i++) {
+        for (int j = 0; j < rows; j++) {
+            for (int k = 0; k < cols; k++) {
+                result_data[k * rows * depth + j * depth + i] = tensor->data[i * rows * cols + j * cols + k];
             }
         }
     }

norch/csrc/cpu.h

@@ -14,7 +14,10 @@ void pow_tensor_cpu(Tensor* tensor, float power, float* result_data);
 void scalar_mul_tensor_cpu(Tensor* tensor, float scalar, float* result_data);
 void ones_like_tensor_cpu(Tensor* tensor, float* result_data);
 void zeros_like_tensor_cpu(Tensor* tensor, float* result_data);
-void transpose_tensor_cpu(Tensor* tensor, float* result_data);
+void transpose_1D_tensor_cpu(Tensor* tensor, float* result_data);
+void transpose_2D_tensor_cpu(Tensor* tensor, float* result_data);
+void transpose_3D_tensor_cpu(Tensor* tensor, float* result_data);
+void transpose_axes_cpu(Tensor* tensor, float* result_data, int axis1, int axis2, int* new_shape);
 void assign_tensor_cpu(Tensor* tensor, float* result_data);
 
 #endif /* CPU_H */

norch/csrc/tensor.cpp

@@ -60,6 +60,15 @@ extern "C" {
         }
         printf("]\n");
 
+        printf("Strides: [");
+        for (int i = 0; i < ndim; i++) {
+            printf("%d", tensor->strides[i]);
+            if (i < ndim - 1) {
+                printf(", ");
+            }
+        }
+        printf("]\n");
+
         /*printf("Data:\n[");
         for (int i = 0; i < stride; i++) {
             printf("%.2f", tensor->data[i]);
@@ -715,8 +724,112 @@ extern "C" {
                 fprintf(stderr, "Memory allocation failed\n");
                 exit(1);
             }
-            transpose_tensor_cpu(tensor, result_data);
+            switch (ndim) {
+                case 1:
+                    transpose_1D_tensor_cpu(tensor, result_data);
+                    break;
+                case 2:
+                    transpose_2D_tensor_cpu(tensor, result_data);
+                    break;
+                case 3:
+                    transpose_3D_tensor_cpu(tensor, result_data);
+                    break;
+                default:
+                    fprintf(stderr, "Transpose only supports tensors up to 3 dimensions.\n");
+                    exit(-1);
+            }
             return create_tensor(result_data, shape, ndim, device);
         }
     }
-}
+
+    Tensor* transpose_axes_tensor(Tensor* tensor, int axis1, int axis2) {
+        char* device = (char*)malloc(strlen(tensor->device) + 1);
+        if (device != NULL) {
+            strcpy(device, tensor->device);
+        } else {
+            fprintf(stderr, "Memory allocation failed\n");
+            exit(-1);
+        }
+
+        int ndim = tensor->ndim;
+        int* shape = (int*)malloc(ndim * sizeof(int));
+        if (shape == NULL) {
+            fprintf(stderr, "Memory allocation failed\n");
+            exit(-1);
+        }
+
+        for (int i = 0; i < ndim; i++) {
+            shape[i] = tensor->shape[i];
+        }
+
+        shape[axis1] = tensor->shape[axis2];
+        shape[axis2] = tensor->shape[axis1];
+
+        int size = tensor->size;
+
+        if (strcmp(tensor->device, "cuda") == 0) {
+
+            float* result_data;
+            cudaMalloc((void **)&result_data, size * sizeof(float));
+            //transpose_axes_cuda(tensor, result_data);
+            return create_tensor(result_data, shape, ndim, device);
+        } 
+        else {
+            float* result_data = (float*)malloc(size * sizeof(float));
+            if (result_data == NULL) {
+                fprintf(stderr, "Memory allocation failed\n");
+                exit(1);
+            }
+            //transpose_axes_cpu(tensor, result_data, axis1, axis2, shape);
+            assign_tensor_cpu(tensor, result_data);
+            
+            Tensor* new_tensor = create_tensor(result_data, shape, ndim, device);
+            for (int i = 0; i < ndim; i++) {
+                new_tensor->strides[i] = tensor->strides[i];
+            }
+            new_tensor->strides[axis1] = tensor->strides[axis2];
+            new_tensor->strides[axis2] = tensor->strides[axis1];
+            make_contiguous(new_tensor);
+            return new_tensor;
+        }
+    }
+
+    void make_contiguous(Tensor* tensor) {
+        float* new_data = (float*)malloc(tensor->size * sizeof(float));
+        if (new_data == NULL) {
+            // Handle memory allocation failure
+            return;
+        }
+
+        int* new_strides = (int*)malloc(tensor->ndim * sizeof(int));
+        if (new_strides == NULL) {
+            free(new_data);
+            // Handle memory allocation failure
+            return;
+        }
+
+        // Calculate new strides assuming C-contiguous order
+        int stride = 1;
+        for (int i = tensor->ndim - 1; i >= 0; i--) {
+            new_strides[i] = stride;
+            stride *= tensor->shape[i];
+        }
+
+        // Rearrange data
+        for (int i = 0; i < tensor->size; i++) {
+            int index = 0;
+            int offset = i;
+            for (int j = 0; j < tensor->ndim; j++) {
+                index += (offset / new_strides[j]) * tensor->strides[j];
+                offset %= new_strides[j];
+            }
+            new_data[i] = tensor->data[index];
+        }
+
+        // Free old data and update tensor properties
+        free(tensor->data);
+        free(tensor->strides);
+        tensor->data = new_data;
+        tensor->strides = new_strides;
+        }
+    }

norch/csrc/tensor.h

@@ -26,6 +26,9 @@ extern "C" {
     void to_device(Tensor* tensor, char* device);
     Tensor* ones_like_tensor(Tensor* tensor);
     Tensor* zeros_like_tensor(Tensor* tensor);
+    Tensor* transpose_tensor(Tensor* tensor);
+    Tensor* transpose_axes_tensor(Tensor* tensor, int axis1, int axis2);
+    void make_contiguous(Tensor* tensor);
 }
 
 #endif /* TENSOR_H */

norch/tensor.py

@@ -383,6 +383,28 @@ class Tensor:
 
         return result_data
     
+    def transpose(self, axis1, axis2):
+        if axis1 < 0:
+            axis1 = self.ndim + axis1
+        if axis2 < 0:
+            axis2 = self.ndim + axis2
+
+        Tensor._C.transpose_axes_tensor.argtypes = [ctypes.POINTER(CTensor), ctypes.c_int, ctypes.c_int]
+        Tensor._C.transpose_axes_tensor.restype = ctypes.POINTER(CTensor)
+
+        result_tensor_ptr = Tensor._C.transpose_axes_tensor(self.tensor, axis1, axis2)
+
+        result_data = Tensor()
+        result_data.tensor = result_tensor_ptr
+        result_data.shape = self.shape.copy()
+        result_data.shape[axis1] = self.shape[axis2]
+        result_data.shape[axis2] = self.shape[axis1]
+        result_data.ndim = self.ndim
+        result_data.device = self.device
+
+        return result_data
+
+    
     @property
     def T(self):
         Tensor._C.transpose_tensor.argtypes = [ctypes.POINTER(CTensor)]

test.py

@@ -28,22 +28,72 @@ if __name__ == "__main__":
         [[7.890, 8.901], [9.012, 1.234], [2.345, 3.456]]
     ], requires_grad=True)
 
-    b = norch.Tensor([
+    b = norch.Tensor([[
         [1.234, 2.123, 1.5],
         [5.678, 6.789, 1.293],
         [3.635, 4.456, 1.0202],
         [7.890, 8.901, 1.91],
-    ])
+    ],[
+        [1.234, 2.123, 1.5],
+        [5.678, 6.789, 1.293],
+        [3.635, 4.456, 1.0202],
+        [7.890, 8.901, 1.91],
+    ],[
+        [1.234, 2.123, 1.5],
+        [5.678, 6.789, 1.293],
+        [3.635, 4.456, 1.0202],
+        [7.890, 8.901, 1.91],
+    ],[
+        [1.234, 2.123, 1.5],
+        [5.678, 6.789, 1.293],
+        [3.635, 4.456, 1.0202],
+        [7.890, 8.901, 1.91],
+    ],[
+        [1.234, 2.123, 1.5],
+        [5.678, 6.789, 1.293],
+        [3.635, 4.456, 1.0202],
+        [7.890, 8.901, 1.91],
+    ]])
 
+    #print(a.T)
+    #print(a.T.shape)
+    #[5, 3, 2] [4, 3] [5, 4, 2]
+    #print(a.shape, b.shape)
     #b = norch.Tensor([
     #    [1.234, 2.123, 1.5]])
+    #result = b @ a
 
-    #print(a.shape)
-    c = a.reshape([2,3,5])
-    print(b @ c)
+
+    #### testar transpose axes!!!! make it contiguous
+
+    tensor1 = norch.Tensor([[[1, 2], [3, 4], [5, 6]],
+                        [[7, 8], [9, 10], [11, 12]],
+                        [[13, 14], [15, 16], [17, 18]],
+                        [[19, 20], [21, 22], [23, 24]],
+                        [[25, 26], [27, 28], [29, 30]]])
+
+    # Reshape tensor1 to 2x3x5
+    reshaped_tensor = tensor1.transpose(1, 0)
+
+    # Create a 5x4 tensor
+    tensor2 = norch.Tensor([[1, 2, 3],
+                            [5, 6, 7],
+                            [9, 10, 11],
+                            [13, 14, 15],
+                            [17, 18, 19]])
+
+    tensor2 = tensor2.transpose(1,0)
+    
+
+    # Multiply reshaped_tensor by tensor2
+    result = tensor2 @ reshaped_tensor
+    print(result)
+
+    #print(a.shape, b.shape, result.shape)
     #c = result.sum()
     #c.backward()
     #print(a.grad)
+    #print(a.transpose(2,1))
     
     #a = norch.Tensor([[1, 2, 3], [1, 2, 3], [1, 2, 3]])#.to("cuda")
     #b = Tensor([[1, 2, 3], [1, 2, 3], [1, 2, 3]])#.to("cuda")