Aicraft
Skip to main content

Tensor API

#include "aicraft/tensor.h"

Types​

AcTensor​

typedef struct AcTensor {
    float *data;       // Raw data pointer
    int   *shape;      // Shape array
    int   *strides;    // Stride array
    int    ndim;       // Number of dimensions
    int    size;       // Total number of elements
    bool   requires_grad;
    struct AcGradNode *grad_node;  // Autograd node
} AcTensor;

Creation​

ac_tensor_new​

AcTensor *ac_tensor_new(int *shape, int ndim);

Create a zero-initialised tensor with the given shape.

ac_tensor_rand​

AcTensor *ac_tensor_rand(int *shape, int ndim);

Create a tensor filled with uniform random values in [0, 1).

ac_tensor_ones / ac_tensor_zeros​

AcTensor *ac_tensor_ones(int *shape, int ndim);
AcTensor *ac_tensor_zeros(int *shape, int ndim);

ac_tensor_from_data​

AcTensor *ac_tensor_from_data(float *data, int *shape, int ndim);

Create a tensor from existing data (copies the data).

Operations​

ac_tensor_add / ac_tensor_mul​

AcTensor *ac_tensor_add(AcTensor *a, AcTensor *b);
AcTensor *ac_tensor_mul(AcTensor *a, AcTensor *b);

Element-wise addition and multiplication with broadcasting.

ac_tensor_matmul​

AcTensor *ac_tensor_matmul(AcTensor *a, AcTensor *b);

Matrix multiplication. Dispatches to SIMD or Vulkan backend.

ac_tensor_transpose​

AcTensor *ac_tensor_transpose(AcTensor *t);

ac_scalar​

float ac_scalar(AcTensor *t);

Extract a scalar value from a 1-element tensor.

Memory​

Tensors are allocated from the arena. Use ac_mem_checkpoint() / ac_mem_restore() for bulk deallocation.