use super::{CpuBackend, TensorInfo, linear};
use crate::{Result, invalid};
fn integer(parameters: &[u8; 64], index: usize) -> i32 {
i32::from_le_bytes(std::array::from_fn(|offset| parameters[index * 4 + offset]))
}
impl CpuBackend {
pub(super) fn backward_image_values(
&self,
info: &TensorInfo,
gradient: &[f32],
) -> Result<Vec<f32>> {
let gradient_info = self.info(
*info
.sources
.first()
.ok_or_else(|| invalid("CPU image gradient missing"))?,
)?;
if info.kind != 0 || gradient_info.kind != 0 || gradient.len() != gradient_info.elements() {
return Err(invalid("CPU image gradients require binary32 storage"));
}
match info.operation {
53 => self.im2col_backward_values(info, gradient_info.shape, gradient),
62 => self.pool_backward_values(info, gradient_info.shape, gradient),
_ => Err(invalid("CPU image backward operation differs")),
}
}
#[allow(
clippy::too_many_lines,
reason = "Pinned im2col backward retains its pixel and kernel loop ordering."
)]
fn im2col_backward_values(
&self,
info: &TensorInfo,
gradient_shape: [usize; 4],
gradient: &[f32],
) -> Result<Vec<f32>> {
let kernel = self.info(
*info
.sources
.get(1)
.ok_or_else(|| invalid("CPU im2col backward kernel missing"))?,
)?;
if !matches!(kernel.kind, 0 | 1) {
return Err(invalid("CPU im2col backward kernel type differs"));
}
let parameters = &info.parameters;
let two_dimensional = integer(parameters, 6) == 1;
let shape = info.shape;
let batches = shape[if two_dimensional { 3 } else { 2 }];
let channels = shape[if two_dimensional { 2 } else { 1 }];
let height = if two_dimensional { shape[1] } else { 1 };
let kernel_height = if two_dimensional { kernel.shape[1] } else { 1 };
let kernel_width = kernel.shape[0];
let stride = [integer(parameters, 0), integer(parameters, 1)];
let padding = [integer(parameters, 2), integer(parameters, 3)];
let dilation = [integer(parameters, 4), integer(parameters, 5)];
let gradient_width = channels
.checked_mul(kernel_height)
.and_then(|count| count.checked_mul(kernel_width))
.ok_or_else(|| invalid("CPU im2col backward channel extent overflow"))?;
if stride[0] <= 0
|| dilation[0] <= 0
|| two_dimensional && (stride[1] <= 0 || dilation[1] <= 0)
|| gradient_shape[0] != gradient_width
|| two_dimensional && gradient_shape[3] != batches
|| !two_dimensional
&& (shape[3] != 1 || gradient_shape[2] != batches || gradient_shape[3] != 1)
{
return Err(invalid("CPU im2col backward geometry differs"));
}
let output_height = if two_dimensional {
gradient_shape[2]
} else {
1
};
let output_width = gradient_shape[1];
let mut output = vec![0.0_f32; info.elements()];
for batch in 0..batches {
for channel in 0..channels {
for y in 0..height {
for x in 0..shape[0] {
let mut sum = 0.0_f32;
for kernel_y in 0..kernel_height {
for kernel_x in 0..kernel_width {
let numerator_x = x as i64 + i64::from(padding[0])
- kernel_x as i64 * i64::from(dilation[0]);
if numerator_x % i64::from(stride[0]) != 0 {
continue;
}
let source_x = numerator_x / i64::from(stride[0]);
let source_y = if two_dimensional {
let numerator_y = y as i64 + i64::from(padding[1])
- kernel_y as i64 * i64::from(dilation[1]);
if numerator_y % i64::from(stride[1]) != 0 {
continue;
}
numerator_y / i64::from(stride[1])
} else {
0
};
if source_x < 0
|| source_x >= output_width as i64
|| source_y < 0
|| source_y >= output_height as i64
{
continue;
}
let kernel_index =
(channel * kernel_height + kernel_y) * kernel_width + kernel_x;
let source = if two_dimensional {
[kernel_index, source_x as usize, source_y as usize, batch]
} else {
[kernel_index, source_x as usize, batch, 0]
};
sum += gradient[linear(source, gradient_shape)];
}
}
let destination = if two_dimensional {
[x, y, channel, batch]
} else {
[x, channel, batch, 0]
};
output[linear(destination, shape)] = sum;
}
}
}
}
Ok(output)
}
#[allow(
clippy::too_many_lines,
reason = "Pinned pooling backward retains window visitation and first-maximum selection."
)]
fn pool_backward_values(
&self,
info: &TensorInfo,
gradient_shape: [usize; 4],
gradient: &[f32],
) -> Result<Vec<f32>> {
let forward = *info
.sources
.get(1)
.ok_or_else(|| invalid("CPU pooling backward forward input missing"))?;
let forward_info = self.info(forward)?;
let parameters = &info.parameters;
let average = match integer(parameters, 0) {
0 => false,
1 => true,
_ => return Err(invalid("CPU pooling backward operation differs")),
};
let kernel = [integer(parameters, 1), integer(parameters, 2)];
let stride = [integer(parameters, 3), integer(parameters, 4)];
let padding = [integer(parameters, 5), integer(parameters, 6)];
let shape = info.shape;
if forward_info.kind != 0
|| forward_info.shape != shape
|| gradient_shape[2..] != shape[2..]
|| kernel.iter().chain(&stride).any(|&value| value <= 0)
{
return Err(invalid("CPU pooling backward geometry differs"));
}
let kernel_area = kernel[0]
.checked_mul(kernel[1])
.ok_or_else(|| invalid("CPU pooling backward kernel area overflow"))?;
let forward = if average {
Vec::new()
} else {
self.values(forward)?
};
let mut output = vec![0.0_f32; info.elements()];
for batch in 0..shape[3] {
for channel in 0..shape[2] {
for output_y in 0..gradient_shape[1] {
for output_x in 0..gradient_shape[0] {
let gradient =
gradient[linear([output_x, output_y, channel, batch], gradient_shape)];
let origin_x =
output_x as i64 * i64::from(stride[0]) - i64::from(padding[0]);
let origin_y =
output_y as i64 * i64::from(stride[1]) - i64::from(padding[1]);
let contribution = gradient / kernel_area as f32;
let mut maximum = -f32::MAX;
let mut maximum_index = None;
for kernel_y in 0..kernel[1] {
let y = origin_y + i64::from(kernel_y);
if y < 0 || y >= shape[1] as i64 {
continue;
}
for kernel_x in 0..kernel[0] {
let x = origin_x + i64::from(kernel_x);
if x < 0 || x >= shape[0] as i64 {
continue;
}
let index = linear([x as usize, y as usize, channel, batch], shape);
if average {
output[index] += contribution;
} else {
let value = forward[index];
if value <= maximum {
continue;
}
maximum = value;
maximum_index = Some(index);
}
}
}
if !average && let Some(index) = maximum_index {
output[index] += gradient;
}
}
}
}
}
Ok(output)
}
}