Skip to content

Commit

Permalink
fn {warp_affine_8x8{,t},emu_edge,resize}_{c => rust}: Deduplicate w…
Browse files Browse the repository at this point in the history 
…/ generics (#361)

This is the last of the `mc` `fn`s to deduplicate besides the
`*dsp_init*` ones, which I have to do after type erasing the `fn` ptrs.
  • Loading branch information
@kkysen
kkysen authored Aug 7, 2023
2 parents 0b9aa66 + 0e349c2 commit 97a8692
Show file tree
Hide file tree
Showing 3 changed files with 430 additions and 581 deletions.
330 changes: 330 additions & 0 deletions src/mc.rs
View file
Original file line number Diff line number Diff line change
@@ -1,9 +1,12 @@
use std::iter;

use crate::include::common::bitdepth::{AsPrimitive, BitDepth};
use crate::include::common::intops::iclip;
use crate::include::dav1d::headers::Dav1dFilterMode;
use crate::src::tables::dav1d_mc_subpel_filters;
use crate::src::tables::dav1d_mc_warp_filter;
use crate::src::tables::dav1d_obmc_masks;
use crate::src::tables::dav1d_resize_filter;

// TODO(kkysen) temporarily `pub` until `mc` callers are deduplicated
#[inline(never)]
Expand Down Expand Up @@ -896,3 +899,330 @@ pub unsafe fn w_mask_rust<BD: BitDepth>(
}
}
}

// TODO(kkysen) temporarily `pub` until `mc` callers are deduplicated
pub unsafe fn warp_affine_8x8_rust<BD: BitDepth>(
mut dst: *mut BD::Pixel,
dst_stride: libc::ptrdiff_t,
mut src: *const BD::Pixel,
src_stride: libc::ptrdiff_t,
abcd: *const i16,
mut mx: libc::c_int,
mut my: libc::c_int,
bd: BD,
) {
let intermediate_bits = bd.get_intermediate_bits();
let mut mid: [i16; 120] = [0; 120];
let mut mid_ptr: *mut i16 = mid.as_mut_ptr();
src = src.offset(-((3 * BD::pxstride(src_stride as usize)) as isize));
let mut y = 0;
while y < 15 {
let mut x = 0;
let mut tmx = mx;
while x < 8 {
let filter: *const i8 =
(dav1d_mc_warp_filter[(64 as libc::c_int + (tmx + 512 >> 10)) as usize]).as_ptr();
*mid_ptr.offset(x as isize) = (*filter.offset(0) as libc::c_int
* (*src.offset((x - 3 * 1) as isize)).as_::<libc::c_int>()
+ *filter.offset(1) as libc::c_int
* (*src.offset((x - 2 * 1) as isize)).as_::<libc::c_int>()
+ *filter.offset(2) as libc::c_int
* (*src.offset((x - 1 * 1) as isize)).as_::<libc::c_int>()
+ *filter.offset(3) as libc::c_int
* (*src.offset((x + 0 * 1) as isize)).as_::<libc::c_int>()
+ *filter.offset(4) as libc::c_int
* (*src.offset((x + 1 * 1) as isize)).as_::<libc::c_int>()
+ *filter.offset(5) as libc::c_int
* (*src.offset((x + 2 * 1) as isize)).as_::<libc::c_int>()
+ *filter.offset(6) as libc::c_int
* (*src.offset((x + 3 * 1) as isize)).as_::<libc::c_int>()
+ *filter.offset(7) as libc::c_int
* (*src.offset((x + 4 * 1) as isize)).as_::<libc::c_int>()
+ ((1 as libc::c_int) << 7 - intermediate_bits >> 1)
>> 7 - intermediate_bits) as i16;
x += 1;
tmx += *abcd.offset(0) as libc::c_int;
}
src = src.offset(BD::pxstride(src_stride as usize) as isize);
mid_ptr = mid_ptr.offset(8);
y += 1;
mx += *abcd.offset(1) as libc::c_int;
}
mid_ptr = &mut *mid.as_mut_ptr().offset((3 * 8) as isize) as *mut i16;
let mut y_0 = 0;
while y_0 < 8 {
let mut x_0 = 0;
let mut tmy = my;
while x_0 < 8 {
let filter_0: *const i8 =
(dav1d_mc_warp_filter[(64 as libc::c_int + (tmy + 512 >> 10)) as usize]).as_ptr();
*dst.offset(x_0 as isize) = bd.iclip_pixel(
*filter_0.offset(0) as libc::c_int
* *mid_ptr.offset((x_0 - 3 * 8) as isize) as libc::c_int
+ *filter_0.offset(1) as libc::c_int
* *mid_ptr.offset((x_0 - 2 * 8) as isize) as libc::c_int
+ *filter_0.offset(2) as libc::c_int
* *mid_ptr.offset((x_0 - 1 * 8) as isize) as libc::c_int
+ *filter_0.offset(3) as libc::c_int
* *mid_ptr.offset((x_0 + 0 * 8) as isize) as libc::c_int
+ *filter_0.offset(4) as libc::c_int
* *mid_ptr.offset((x_0 + 1 * 8) as isize) as libc::c_int
+ *filter_0.offset(5) as libc::c_int
* *mid_ptr.offset((x_0 + 2 * 8) as isize) as libc::c_int
+ *filter_0.offset(6) as libc::c_int
* *mid_ptr.offset((x_0 + 3 * 8) as isize) as libc::c_int
+ *filter_0.offset(7) as libc::c_int
* *mid_ptr.offset((x_0 + 4 * 8) as isize) as libc::c_int
+ ((1 as libc::c_int) << 7 + intermediate_bits >> 1)
>> 7 + intermediate_bits,
);
x_0 += 1;
tmy += *abcd.offset(2) as libc::c_int;
}
mid_ptr = mid_ptr.offset(8);
dst = dst.offset(BD::pxstride(dst_stride as usize) as isize);
y_0 += 1;
my += *abcd.offset(3) as libc::c_int;
}
}

// TODO(kkysen) temporarily `pub` until `mc` callers are deduplicated
pub unsafe fn warp_affine_8x8t_rust<BD: BitDepth>(
mut tmp: *mut i16,
tmp_stride: libc::ptrdiff_t,
mut src: *const BD::Pixel,
src_stride: libc::ptrdiff_t,
abcd: *const i16,
mut mx: libc::c_int,
mut my: libc::c_int,
bd: BD,
) {
let intermediate_bits = bd.get_intermediate_bits();
let mut mid: [i16; 120] = [0; 120];
let mut mid_ptr: *mut i16 = mid.as_mut_ptr();
src = src.offset(-((3 * BD::pxstride(src_stride as usize)) as isize));
let mut y = 0;
while y < 15 {
let mut x = 0;
let mut tmx = mx;
while x < 8 {
let filter: *const i8 =
(dav1d_mc_warp_filter[(64 as libc::c_int + (tmx + 512 >> 10)) as usize]).as_ptr();
*mid_ptr.offset(x as isize) = (*filter.offset(0) as libc::c_int
* (*src.offset((x - 3 * 1) as isize)).as_::<libc::c_int>()
+ *filter.offset(1) as libc::c_int
* (*src.offset((x - 2 * 1) as isize)).as_::<libc::c_int>()
+ *filter.offset(2) as libc::c_int
* (*src.offset((x - 1 * 1) as isize)).as_::<libc::c_int>()
+ *filter.offset(3) as libc::c_int
* (*src.offset((x + 0 * 1) as isize)).as_::<libc::c_int>()
+ *filter.offset(4) as libc::c_int
* (*src.offset((x + 1 * 1) as isize)).as_::<libc::c_int>()
+ *filter.offset(5) as libc::c_int
* (*src.offset((x + 2 * 1) as isize)).as_::<libc::c_int>()
+ *filter.offset(6) as libc::c_int
* (*src.offset((x + 3 * 1) as isize)).as_::<libc::c_int>()
+ *filter.offset(7) as libc::c_int
* (*src.offset((x + 4 * 1) as isize)).as_::<libc::c_int>()
+ ((1 as libc::c_int) << 7 - intermediate_bits >> 1)
>> 7 - intermediate_bits) as i16;
x += 1;
tmx += *abcd.offset(0) as libc::c_int;
}
src = src.offset(BD::pxstride(src_stride as usize) as isize);
mid_ptr = mid_ptr.offset(8);
y += 1;
mx += *abcd.offset(1) as libc::c_int;
}
mid_ptr = &mut *mid.as_mut_ptr().offset((3 * 8) as isize) as *mut i16;
let mut y_0 = 0;
while y_0 < 8 {
let mut x_0 = 0;
let mut tmy = my;
while x_0 < 8 {
let filter_0: *const i8 =
(dav1d_mc_warp_filter[(64 as libc::c_int + (tmy + 512 >> 10)) as usize]).as_ptr();
*tmp.offset(x_0 as isize) = ((*filter_0.offset(0) as libc::c_int
* *mid_ptr.offset((x_0 - 3 * 8) as isize) as libc::c_int
+ *filter_0.offset(1) as libc::c_int
* *mid_ptr.offset((x_0 - 2 * 8) as isize) as libc::c_int
+ *filter_0.offset(2) as libc::c_int
* *mid_ptr.offset((x_0 - 1 * 8) as isize) as libc::c_int
+ *filter_0.offset(3) as libc::c_int
* *mid_ptr.offset((x_0 + 0 * 8) as isize) as libc::c_int
+ *filter_0.offset(4) as libc::c_int
* *mid_ptr.offset((x_0 + 1 * 8) as isize) as libc::c_int
+ *filter_0.offset(5) as libc::c_int
* *mid_ptr.offset((x_0 + 2 * 8) as isize) as libc::c_int
+ *filter_0.offset(6) as libc::c_int
* *mid_ptr.offset((x_0 + 3 * 8) as isize) as libc::c_int
+ *filter_0.offset(7) as libc::c_int
* *mid_ptr.offset((x_0 + 4 * 8) as isize) as libc::c_int
+ ((1 as libc::c_int) << 7 >> 1)
>> 7)
- i32::from(BD::PREP_BIAS)) as i16;
x_0 += 1;
tmy += *abcd.offset(2) as libc::c_int;
}
mid_ptr = mid_ptr.offset(8);
tmp = tmp.offset(tmp_stride as isize);
y_0 += 1;
my += *abcd.offset(3) as libc::c_int;
}
}

// TODO(kkysen) temporarily `pub` until `mc` callers are deduplicated
pub unsafe fn emu_edge_rust<BD: BitDepth>(
bw: libc::intptr_t,
bh: libc::intptr_t,
iw: libc::intptr_t,
ih: libc::intptr_t,
x: libc::intptr_t,
y: libc::intptr_t,
mut dst: *mut BD::Pixel,
dst_stride: libc::ptrdiff_t,
mut r#ref: *const BD::Pixel,
ref_stride: libc::ptrdiff_t,
) {
r#ref = r#ref.offset(
iclip(y as libc::c_int, 0 as libc::c_int, ih as libc::c_int - 1) as isize
* BD::pxstride(ref_stride as usize) as isize
+ iclip(x as libc::c_int, 0 as libc::c_int, iw as libc::c_int - 1) as isize,
);
let left_ext = iclip(-x as libc::c_int, 0 as libc::c_int, bw as libc::c_int - 1);
let right_ext = iclip(
(x + bw - iw) as libc::c_int,
0 as libc::c_int,
bw as libc::c_int - 1,
);
if !(((left_ext + right_ext) as isize) < bw) {
unreachable!();
}
let top_ext = iclip(-y as libc::c_int, 0 as libc::c_int, bh as libc::c_int - 1);
let bottom_ext = iclip(
(y + bh - ih) as libc::c_int,
0 as libc::c_int,
bh as libc::c_int - 1,
);
if !(((top_ext + bottom_ext) as isize) < bh) {
unreachable!();
}
let mut blk: *mut BD::Pixel =
dst.offset((top_ext as isize * BD::pxstride(dst_stride as usize) as isize) as isize);
let center_w = (bw - left_ext as isize - right_ext as isize) as libc::c_int;
let center_h = (bh - top_ext as isize - bottom_ext as isize) as libc::c_int;
let mut y_0 = 0;
while y_0 < center_h {
BD::pixel_copy(
std::slice::from_raw_parts_mut(blk.offset(left_ext as isize), center_w as usize),
std::slice::from_raw_parts(r#ref, center_w as usize),
center_w as usize,
);
if left_ext != 0 {
BD::pixel_set(
std::slice::from_raw_parts_mut(blk, left_ext as usize),
*blk.offset(left_ext as isize),
left_ext as usize,
);
}
if right_ext != 0 {
BD::pixel_set(
std::slice::from_raw_parts_mut(
blk.offset(left_ext as isize).offset(center_w as isize),
right_ext as usize,
),
*blk.offset((left_ext + center_w - 1) as isize),
right_ext as usize,
);
}
r#ref = r#ref.offset(BD::pxstride(ref_stride as usize) as isize);
blk = blk.offset(BD::pxstride(dst_stride as usize) as isize);
y_0 += 1;
}
blk = dst.offset((top_ext as isize * BD::pxstride(dst_stride as usize) as isize) as isize);
let mut y_1 = 0;
while y_1 < top_ext {
BD::pixel_copy(
std::slice::from_raw_parts_mut(dst, bw as usize),
std::slice::from_raw_parts(blk, bw as usize),
bw as usize,
);
dst = dst.offset(BD::pxstride(dst_stride as usize) as isize);
y_1 += 1;
}
dst = dst.offset((center_h as isize * BD::pxstride(dst_stride as usize) as isize) as isize);
let mut y_2 = 0;
while y_2 < bottom_ext {
BD::pixel_copy(
std::slice::from_raw_parts_mut(dst, bw as usize),
std::slice::from_raw_parts(
dst.offset(-(BD::pxstride(dst_stride as usize) as isize)),
bw as usize,
),
bw as usize,
);
dst = dst.offset(BD::pxstride(dst_stride as usize) as isize);
y_2 += 1;
}
}

// TODO(kkysen) temporarily `pub` until `mc` callers are deduplicated
pub unsafe fn resize_rust<BD: BitDepth>(
mut dst: *mut BD::Pixel,
dst_stride: libc::ptrdiff_t,
mut src: *const BD::Pixel,
src_stride: libc::ptrdiff_t,
dst_w: libc::c_int,
mut h: libc::c_int,
src_w: libc::c_int,
dx: libc::c_int,
mx0: libc::c_int,
bd: BD,
) {
loop {
let mut mx = mx0;
let mut src_x = -(1 as libc::c_int);
let mut x = 0;
while x < dst_w {
let F: *const i8 = (dav1d_resize_filter[(mx >> 8) as usize]).as_ptr();
*dst.offset(x as isize) = bd.iclip_pixel(
-(*F.offset(0) as libc::c_int
* (*src.offset(iclip(src_x - 3, 0 as libc::c_int, src_w - 1) as isize))
.as_::<libc::c_int>()
+ *F.offset(1) as libc::c_int
* (*src.offset(iclip(src_x - 2, 0 as libc::c_int, src_w - 1) as isize))
.as_::<libc::c_int>()
+ *F.offset(2) as libc::c_int
* (*src.offset(iclip(src_x - 1, 0 as libc::c_int, src_w - 1) as isize))
.as_::<libc::c_int>()
+ *F.offset(3) as libc::c_int
* (*src.offset(iclip(src_x + 0, 0 as libc::c_int, src_w - 1) as isize))
.as_::<libc::c_int>()
+ *F.offset(4) as libc::c_int
* (*src.offset(iclip(src_x + 1, 0 as libc::c_int, src_w - 1) as isize))
.as_::<libc::c_int>()
+ *F.offset(5) as libc::c_int
* (*src.offset(iclip(src_x + 2, 0 as libc::c_int, src_w - 1) as isize))
.as_::<libc::c_int>()
+ *F.offset(6) as libc::c_int
* (*src.offset(iclip(src_x + 3, 0 as libc::c_int, src_w - 1) as isize))
.as_::<libc::c_int>()
+ *F.offset(7) as libc::c_int
* (*src.offset(iclip(src_x + 4, 0 as libc::c_int, src_w - 1) as isize))
.as_::<libc::c_int>())
+ 64
>> 7,
);
mx += dx;
src_x += mx >> 14;
mx &= 0x3fff as libc::c_int;
x += 1;
}
dst = dst.offset(BD::pxstride(dst_stride as usize) as isize);
src = src.offset(BD::pxstride(src_stride as usize) as isize);
h -= 1;
if !(h != 0) {
break;
}
}
}
Loading

0 comments on commit 97a8692

Please sign in to comment.