zoomvtools 2.0.0

#[cfg(all(target_arch = "x86_64", feature = "simd"))]
mod avx2;
#[cfg(all(target_arch = "x86_64", feature = "simd"))]
mod avx512;
mod rust;

#[cfg(test)]
mod tests;

use std::num::{NonZeroU8, NonZeroUsize};

use crate::util::Pixel;

/// Performs horizontal bilinear interpolation for sub-pixel motion estimation refinement.
///
/// This function applies bilinear interpolation horizontally to create sub-pixel samples
/// between existing pixels. Bilinear interpolation uses a simple 2-tap kernel that
/// averages adjacent horizontal pixels, providing fast and reasonably smooth interpolation
/// for motion estimation with sub-pixel accuracy.
///
/// The last column is copied directly from the source since there's no right neighbor.
///
/// # Parameters
/// - `src`: Source image buffer
/// - `dest`: Destination buffer for interpolated results
/// - `pitch`: Number of pixels per row in both buffers
/// - `width`: Width of the image in pixels
/// - `height`: Height of the image in pixels
/// - `_bits_per_sample`: Unused parameter for API consistency
#[inline]
pub fn refine_horizontal_bilinear<T: Pixel>(
    dest: &mut [T],
    src: &[T],
    pitch: NonZeroUsize,
    width: NonZeroUsize,
    height: NonZeroUsize,
    _bits_per_sample: NonZeroU8,
) {
    #[cfg(all(target_arch = "x86_64", feature = "simd"))]
    if cpudetect::x86_64::is_znver5_compatible() {
        // PERF(znver5): 5% faster than AVX2 on test machine
        // SAFETY: We check for AVX-512 first
        unsafe {
            avx512::refine_horizontal_bilinear(dest, src, pitch, width, height, _bits_per_sample);
        }
        return;
    }

    #[cfg(all(target_arch = "x86_64", feature = "simd"))]
    if cpudetect::x86_64::is_x86_64_v4_compatible() {
        cfg_select! {
            feature = "experimental" => {
                // TODO(znver4): 10% slower than scalar on test machine for 10-bit
                // SAFETY: We check for AVX-512 first
                unsafe {
                    avx512::refine_horizontal_bilinear(dest, src, pitch, width, height, _bits_per_sample);
                }
                return;
            }
            _ => {
                if size_of::<T>() == 1  {
                    // PERF(znver4): Equivalent to AVX2 on test machine
                    // SAFETY: We check for AVX-512 first
                    unsafe {
                        avx512::refine_horizontal_bilinear(dest, src, pitch, width, height, _bits_per_sample);
                    }
                    return;
                }
            }
        }
    }

    #[cfg(all(target_arch = "x86_64", feature = "simd"))]
    if cpudetect::x86_64::is_x86_64_v3_compatible() {
        cfg_select! {
            feature = "experimental" => {
                // TODO: 0-5% slower than scalar on test machine for 10-bit
                // SAFETY: We check for AVX2 first
                unsafe {
                    avx2::refine_horizontal_bilinear(dest, src, pitch, width, height, _bits_per_sample);
                }
                return;
            }
            _ => {
                if size_of::<T>() == 1  {
                    // PERF: 30-45% faster than scalar on test machine for 8-bit
                    // SAFETY: We check for AVX2 first
                    unsafe {
                        avx2::refine_horizontal_bilinear(dest, src, pitch, width, height, _bits_per_sample);
                    }
                    return;
                }
            }
        }
    }

    rust::refine_horizontal_bilinear(dest, src, pitch, width, height, _bits_per_sample);
}

/// Performs vertical bilinear interpolation for sub-pixel motion estimation refinement.
///
/// This function applies bilinear interpolation vertically to create sub-pixel samples
/// between existing pixels. Bilinear interpolation uses a simple 2-tap kernel that
/// averages adjacent vertical pixels, providing fast and reasonably smooth interpolation
/// for motion estimation with sub-pixel accuracy.
///
/// The last row is copied directly from the source since there's no bottom neighbor.
///
/// # Parameters
/// - `src`: Source image buffer
/// - `dest`: Destination buffer for interpolated results
/// - `pitch`: Number of pixels per row in both buffers
/// - `width`: Width of the image in pixels
/// - `height`: Height of the image in pixels
/// - `_bits_per_sample`: Unused parameter for API consistency
#[inline]
pub fn refine_vertical_bilinear<T: Pixel>(
    dest: &mut [T],
    src: &[T],
    pitch: NonZeroUsize,
    width: NonZeroUsize,
    height: NonZeroUsize,
    _bits_per_sample: NonZeroU8,
) {
    #[cfg(all(target_arch = "x86_64", feature = "simd"))]
    if cpudetect::x86_64::is_x86_64_v4_compatible() {
        cfg_select! {
            feature = "experimental" => {
                // TODO: 40% slower than scalar on test machine for 8-bit
                // TODO: 15% slower than scalar on test machine for 10-bit
                // SAFETY: We check for AVX-512 first
                unsafe {
                    avx512::refine_vertical_bilinear(dest, src, pitch, width, height, _bits_per_sample);
                }
                return;
            }
            _ => {}
        }
    }

    #[cfg(all(target_arch = "x86_64", feature = "simd"))]
    if cpudetect::x86_64::is_x86_64_v3_compatible() {
        cfg_select! {
            feature = "experimental" => {
                // TODO: 5% slower than scalar on test machine for 10-bit
                // SAFETY: We check for AVX2 first
                unsafe {
                    avx2::refine_vertical_bilinear(dest, src, pitch, width, height, _bits_per_sample);
                }
                return;
            }
            _ => {
                if size_of::<T>() == 1  {
                    // PERF: Equivalent to scalar on test machine for 8-bit
                    // SAFETY: We check for AVX2 first
                    unsafe {
                        avx2::refine_vertical_bilinear(dest, src, pitch, width, height, _bits_per_sample);
                    }
                    return;
                }
            }
        }
    }

    rust::refine_vertical_bilinear(dest, src, pitch, width, height, _bits_per_sample);
}

/// Performs diagonal bilinear interpolation for sub-pixel motion estimation refinement.
///
/// This function applies bilinear interpolation diagonally to create sub-pixel samples
/// at quarter-pixel positions. It averages 2x2 blocks of pixels to create interpolated
/// values at diagonal positions, which is essential for sub-pixel motion estimation
/// that requires samples at positions like (0.5, 0.5).
///
/// Edge pixels and the last row/column use simplified interpolation due to missing neighbors.
///
/// # Parameters
/// - `src`: Source image buffer
/// - `dest`: Destination buffer for interpolated results
/// - `pitch`: Number of pixels per row in both buffers
/// - `width`: Width of the image in pixels
/// - `height`: Height of the image in pixels
/// - `_bits_per_sample`: Unused parameter for API consistency
#[inline]
pub fn refine_diagonal_bilinear<T: Pixel>(
    dest: &mut [T],
    src: &[T],
    pitch: NonZeroUsize,
    width: NonZeroUsize,
    height: NonZeroUsize,
    _bits_per_sample: NonZeroU8,
) {
    #[cfg(all(target_arch = "x86_64", feature = "simd"))]
    if cpudetect::x86_64::is_x86_64_v4_compatible() {
        cfg_select! {
            feature = "experimental" => {
                // TODO: 40% slower than scalar for 8-bit
                // SAFETY: We check for AVX-512 first
                unsafe {
                    avx512::refine_diagonal_bilinear(dest, src, pitch, width, height, _bits_per_sample);
                }
                return;
            }
             _ => {
                if size_of::<T>() == 2  {
                    // PERF: 0-5% faster than AVX2 on test machine for 10-bit
                    // SAFETY: We check for AVX-512 first
                    unsafe {
                        avx512::refine_diagonal_bilinear(dest, src, pitch, width, height, _bits_per_sample);
                    }
                    return;
                }
            }
        }
    }

    #[cfg(all(target_arch = "x86_64", feature = "simd"))]
    if cpudetect::x86_64::is_x86_64_v3_compatible() {
        cfg_select! {
            feature = "experimental" => {
                // TODO: 50-90% slower than scalar for 8-bit
                // TODO: 20-40% slower than scalar for 10-bit
                // SAFETY: We check for AVX2 first
                unsafe {
                    avx2::refine_diagonal_bilinear(dest, src, pitch, width, height, _bits_per_sample);
                }
                return;
            }
            _ => {}
        }
    }

    rust::refine_diagonal_bilinear(dest, src, pitch, width, height, _bits_per_sample);
}