Module kernel 

Raw-pointer kernel functions that operate directly on f64 slices.

These bypass the Value::Tensor wrapper, operating on contiguous &[f64] data. The interpreter resolves tensor pointers once at call entry, then dispatches to these zero-overhead kernels.

All functions here are safe Rust — they accept slices, not raw pointers. The “raw pointer” concept means: the interpreter does one to_vec() or buffer.borrow() at entry, then passes the contiguous slice through.

Functions

conv1d_circular

1D convolution on a sliding window of a circular buffer.

conv1d_dispatched

1D convolution using dispatched summation.

conv1d_raw

1D convolution with stride=1, no padding (“valid” mode).

conv2d_dispatched

2D convolution using dispatched summation strategy.

conv2d_raw

2D convolution — NCHW layout, valid mode (no padding), configurable stride.

gelu_raw

Approximate GELU: x * 0.5 * (1 + tanh(sqrt(2/pi) * (x + 0.044715 * x^3)))

layer_norm_dispatched

Layer normalization using dispatched summation for mean/variance.

layer_norm_raw

Layer normalization over the last dimension.

linear_dispatched

Linear projection using dispatched summation.

linear_raw

Linear projection: Y[outer, out_f] = X[outer, in_f] @ W^T[out_f, in_f] + bias[out_f]

matmul_dispatched

Matrix multiply using dispatched summation strategy.

matmul_raw

Matrix multiply: C[m,n] = A[m,k] × B[k,n] with Kahan-summed dots.

maxpool1d_raw

Max-pooling over 1D signal, stride = pool_size.

maxpool2d_raw

2D max-pooling — NCHW layout, stride = pool_size (non-overlapping).

relu_raw

ReLU: max(0, x) element-wise.

softmax_raw

Softmax over the last dimension of a contiguous buffer.