use crate::{utils::broadcast::pad_shape, Tensor, TensorNode};
use maidenx_core::error::{Error, Result};
impl Tensor {
pub fn broadcast(&self, shape: &[usize]) -> Result<Self> {
if self.shape() == shape {
return Ok(self.clone());
}
let old_shape = self.shape();
let rank = shape.len();
if old_shape.is_empty() {
return self.broadcast_scalar_to(shape);
}
if shape.is_empty() {
if !old_shape.is_empty() {
return Err(Error::InvalidShape {
message: format!("Cannot broadcast non-scalar shape {:?} to scalar ()", old_shape),
});
} else {
return Ok(self.clone());
}
}
let padded_old = pad_shape(old_shape, rank);
for i in 0..rank {
let dim_old = padded_old[i];
let dim_new = shape[i];
if dim_old != 1 && dim_old != dim_new {
return Err(Error::InvalidShape {
message: format!(
"Cannot broadcast dimension {} -> {} (shape {:?} -> {:?})",
dim_old, dim_new, old_shape, shape
),
});
}
}
let mut result = Self::empty_with_spec(shape, self.device(), self.dtype())?;
let mut temp_buf = vec![0u8; self.size() * self.dtype().size_in_bytes()];
unsafe {
self.buffer()
.copy_to_host(temp_buf.as_mut_ptr() as *mut std::ffi::c_void, temp_buf.len(), self.offset(), 0)?;
}
let mut result_buf = vec![0u8; result.size() * result.dtype().size_in_bytes()];
let src_strides = self.strides();
let padded_src_strides = pad_shape(src_strides, rank);
let padded_src_shape = padded_old;
let mut effective_strides = padded_src_strides.clone();
for i in 0..rank {
if padded_src_shape[i] == 1 {
effective_strides[i] = 0;
}
}
let mut pos = vec![0; rank];
for i in 0..result.size() {
let mut remainder = i;
for d in (0..rank).rev() {
pos[d] = remainder % shape[d];
remainder /= shape[d];
}
let mut src_idx = 0;
for d in 0..rank {
src_idx += pos[d] * effective_strides[d];
}
let src_offset = src_idx * self.dtype().size_in_bytes();
let dst_offset = i * self.dtype().size_in_bytes();
result_buf[dst_offset..dst_offset + self.dtype().size_in_bytes()]
.copy_from_slice(&temp_buf[src_offset..src_offset + self.dtype().size_in_bytes()]);
}
unsafe {
result.with_buffer_mut(|buf| {
buf.copy_from_host(result_buf.as_ptr() as *const std::ffi::c_void, result_buf.len(), 0, 0)?;
Ok(())
})?;
}
if self.requires_grad() {
result.with_grad()?;
let orig_shape = self.shape().to_vec();
let backward_fn =
Box::new(move |_inputs: &[Tensor], grad_out: &Tensor| -> Result<Vec<Tensor>> { Ok(vec![grad_out.sum_to_shape(&orig_shape)?]) });
let node = TensorNode::new("broadcast".to_string(), vec![self.clone()], Some(backward_fn));
result.node = Some(node);
}
Ok(result)
}
pub fn broadcast_like(&self, other: &Self) -> Result<Self> {
self.broadcast(other.shape())
}
pub fn broadcast_left(&self, batch_dims: &[usize]) -> Result<Self> {
let mut new_shape = batch_dims.as_ref().to_vec();
new_shape.extend(self.shape());
self.broadcast(&new_shape)
}
fn broadcast_scalar_to(&self, shape: &[usize]) -> Result<Self> {
let mut result = Self::empty_with_spec(shape, self.device(), self.dtype())?;
let mut scalar_buf = vec![0u8; self.dtype().size_in_bytes()];
unsafe {
self.buffer()
.copy_to_host(scalar_buf.as_mut_ptr() as *mut std::ffi::c_void, scalar_buf.len(), self.offset(), 0)?;
let mut result_buf = vec![0u8; result.size() * result.dtype().size_in_bytes()];
for i in 0..result.size() {
let offset = i * result.dtype().size_in_bytes();
result_buf[offset..offset + scalar_buf.len()].copy_from_slice(&scalar_buf);
}
result.with_buffer_mut(|buf| {
buf.copy_from_host(result_buf.as_ptr() as *const std::ffi::c_void, result_buf.len(), 0, 0)?;
Ok(())
})?;
}
Ok(result)
}
}