Skip to content

Commit

Permalink
fn inv_txfm_add_wht_wht_4x4_*: Deduplicate w/ generics.
Browse files Browse the repository at this point in the history
  • Loading branch information
@thedataking
thedataking committed Jan 17, 2024
1 parent cc0cff2 commit 9e19154
Show file tree
Hide file tree
Showing 3 changed files with 72 additions and 156 deletions.
68 changes: 68 additions & 0 deletions src/itx.rs
View file
Original file line number Diff line number Diff line change
Expand Up @@ -409,3 +409,71 @@ inv_txfm_fn64!(32, 64, 1);
inv_txfm_fn64!(64, 16, 2);
inv_txfm_fn64!(64, 32, 1);
inv_txfm_fn64!(64, 64, 2);

// TODO(perl): Temporarily pub until mod is deduplicated
pub(crate) unsafe extern "C" fn inv_txfm_add_wht_wht_4x4_c_erased<BD: BitDepth>(
dst: *mut DynPixel,
stride: ptrdiff_t,
coeff: *mut DynCoef,
eob: c_int,
bitdepth_max: c_int,
) {
inv_txfm_add_wht_wht_4x4_rust::<BD>(dst.cast(), stride, coeff.cast(), eob, bitdepth_max);
}

unsafe fn inv_txfm_add_wht_wht_4x4_rust<BD: BitDepth>(
mut dst: *mut BD::Pixel,
stride: ptrdiff_t,
coeff: *mut BD::Coef,
_eob: c_int,
bitdepth_max: c_int,
) {
use crate::src::itx_1d::dav1d_inv_wht4_1d_c;

let mut tmp: [i32; 16] = [0; 16];
let mut c: *mut i32 = tmp.as_mut_ptr();
let mut y = 0;
while y < 4 {
let mut x = 0;
while x < 4 {
*c.offset(x as isize) = (*coeff.offset((y + x * 4) as isize)).as_::<i32>() >> 2;
x += 1;
}
dav1d_inv_wht4_1d_c(c, 1 as c_int as ptrdiff_t);
y += 1;
c = c.offset(4);
}
memset(
coeff as *mut c_void,
0 as c_int,
::core::mem::size_of::<BD::Coef>()
.wrapping_mul(4)
.wrapping_mul(4),
);
let mut x_0 = 0;
while x_0 < 4 {
dav1d_inv_wht4_1d_c(
&mut *tmp.as_mut_ptr().offset(x_0 as isize),
4 as c_int as ptrdiff_t,
);
x_0 += 1;
}
c = tmp.as_mut_ptr();
let mut y_0 = 0;
while y_0 < 4 {
let mut x_1 = 0;
while x_1 < 4 {
let fresh1 = c;
c = c.offset(1);
*dst.offset(x_1 as isize) = iclip(
(*dst.offset(x_1 as isize)).as_::<c_int>() + *fresh1,
0 as c_int,
bitdepth_max,
)
.as_::<BD::Pixel>();
x_1 += 1;
}
y_0 += 1;
dst = dst.offset(BD::pxstride(stride as usize) as isize);
}
}
87 changes: 3 additions & 84 deletions src/itx_tmpl_16.rs
View file
Original file line number Diff line number Diff line change
@@ -1,7 +1,5 @@
use crate::include::common::bitdepth::BitDepth16;
use crate::include::common::bitdepth::DynCoef;
use crate::include::common::bitdepth::DynPixel;
use crate::include::common::intops::iclip;

use crate::src::itx::Rav1dInvTxfmDSPContext;
use crate::src::levels::ADST_ADST;
use crate::src::levels::ADST_DCT;
Expand Down Expand Up @@ -39,94 +37,15 @@ use crate::src::levels::V_ADST;
use crate::src::levels::V_DCT;
use crate::src::levels::V_FLIPADST;
use crate::src::levels::WHT_WHT;
use libc::memset;
use libc::ptrdiff_t;

use std::ffi::c_int;
use std::ffi::c_void;

#[cfg(feature = "asm")]
use crate::src::cpu::{rav1d_get_cpu_flags, CpuFlags};

#[cfg(feature = "asm")]
use cfg_if::cfg_if;

pub type pixel = u16;
pub type coef = i32;

#[inline]
unsafe fn PXSTRIDE(x: ptrdiff_t) -> ptrdiff_t {
if x & 1 != 0 {
unreachable!();
}
return x >> 1;
}

unsafe extern "C" fn inv_txfm_add_wht_wht_4x4_c_erased(
dst: *mut DynPixel,
stride: ptrdiff_t,
coeff: *mut DynCoef,
eob: c_int,
bitdepth_max: c_int,
) {
inv_txfm_add_wht_wht_4x4_rust(dst.cast(), stride, coeff.cast(), eob, bitdepth_max);
}

unsafe fn inv_txfm_add_wht_wht_4x4_rust(
mut dst: *mut pixel,
stride: ptrdiff_t,
coeff: *mut coef,
_eob: c_int,
bitdepth_max: c_int,
) {
use crate::src::itx_1d::dav1d_inv_wht4_1d_c;

let mut tmp: [i32; 16] = [0; 16];
let mut c: *mut i32 = tmp.as_mut_ptr();
let mut y = 0;
while y < 4 {
let mut x = 0;
while x < 4 {
*c.offset(x as isize) = *coeff.offset((y + x * 4) as isize) >> 2;
x += 1;
}
dav1d_inv_wht4_1d_c(c, 1 as c_int as ptrdiff_t);
y += 1;
c = c.offset(4);
}
memset(
coeff as *mut c_void,
0 as c_int,
::core::mem::size_of::<coef>()
.wrapping_mul(4)
.wrapping_mul(4),
);
let mut x_0 = 0;
while x_0 < 4 {
dav1d_inv_wht4_1d_c(
&mut *tmp.as_mut_ptr().offset(x_0 as isize),
4 as c_int as ptrdiff_t,
);
x_0 += 1;
}
c = tmp.as_mut_ptr();
let mut y_0 = 0;
while y_0 < 4 {
let mut x_1 = 0;
while x_1 < 4 {
let fresh1 = c;
c = c.offset(1);
*dst.offset(x_1 as isize) = iclip(
*dst.offset(x_1 as isize) as c_int + *fresh1,
0 as c_int,
bitdepth_max,
) as pixel;
x_1 += 1;
}
y_0 += 1;
dst = dst.offset(PXSTRIDE(stride) as isize);
}
}

#[cfg(all(feature = "asm", any(target_arch = "x86", target_arch = "x86_64")))]
#[inline(always)]
#[rustfmt::skip]
Expand Down Expand Up @@ -1043,7 +962,7 @@ pub unsafe fn rav1d_itx_dsp_init_16bpc(c: *mut Rav1dInvTxfmDSPContext, mut _bpc:
// TODO(legare): Temporary import until init fns are deduplicated.
use crate::src::itx::*;

(*c).itxfm_add[TX_4X4 as usize][WHT_WHT as usize] = Some(inv_txfm_add_wht_wht_4x4_c_erased);
(*c).itxfm_add[TX_4X4 as usize][WHT_WHT as usize] = Some(inv_txfm_add_wht_wht_4x4_c_erased::<BitDepth16>);
(*c).itxfm_add[TX_4X4 as usize][DCT_DCT as usize] =
Some(inv_txfm_add_dct_dct_4x4_c_erased::<BitDepth16>);
(*c).itxfm_add[TX_4X4 as usize][IDTX as usize] =
Expand Down
73 changes: 1 addition & 72 deletions src/itx_tmpl_8.rs
View file
Original file line number Diff line number Diff line change
@@ -1,7 +1,4 @@
use crate::include::common::bitdepth::BitDepth8;
use crate::include::common::bitdepth::DynCoef;
use crate::include::common::bitdepth::DynPixel;
use crate::include::common::intops::iclip_u8;
use crate::src::itx::Rav1dInvTxfmDSPContext;
use crate::src::levels::ADST_ADST;
use crate::src::levels::ADST_DCT;
Expand Down Expand Up @@ -39,82 +36,14 @@ use crate::src::levels::V_ADST;
use crate::src::levels::V_DCT;
use crate::src::levels::V_FLIPADST;
use crate::src::levels::WHT_WHT;
use libc::memset;
use libc::ptrdiff_t;
use std::ffi::c_int;
use std::ffi::c_void;

#[cfg(feature = "asm")]
use crate::src::cpu::{rav1d_get_cpu_flags, CpuFlags};

#[cfg(feature = "asm")]
use cfg_if::cfg_if;

pub type pixel = u8;
pub type coef = i16;

unsafe extern "C" fn inv_txfm_add_wht_wht_4x4_c_erased(
dst: *mut DynPixel,
stride: ptrdiff_t,
coeff: *mut DynCoef,
eob: c_int,
_bitdepth_max: c_int,
) {
inv_txfm_add_wht_wht_4x4_rust(dst.cast(), stride, coeff.cast(), eob);
}

unsafe fn inv_txfm_add_wht_wht_4x4_rust(
mut dst: *mut pixel,
stride: ptrdiff_t,
coeff: *mut coef,
_eob: c_int,
) {
use crate::src::itx_1d::dav1d_inv_wht4_1d_c;

let mut tmp: [i32; 16] = [0; 16];
let mut c: *mut i32 = tmp.as_mut_ptr();
let mut y = 0;
while y < 4 {
let mut x = 0;
while x < 4 {
*c.offset(x as isize) = *coeff.offset((y + x * 4) as isize) as c_int >> 2;
x += 1;
}
dav1d_inv_wht4_1d_c(c, 1 as c_int as ptrdiff_t);
y += 1;
c = c.offset(4);
}
memset(
coeff as *mut c_void,
0 as c_int,
::core::mem::size_of::<coef>()
.wrapping_mul(4)
.wrapping_mul(4),
);
let mut x_0 = 0;
while x_0 < 4 {
dav1d_inv_wht4_1d_c(
&mut *tmp.as_mut_ptr().offset(x_0 as isize),
4 as c_int as ptrdiff_t,
);
x_0 += 1;
}
c = tmp.as_mut_ptr();
let mut y_0 = 0;
while y_0 < 4 {
let mut x_1 = 0;
while x_1 < 4 {
let fresh1 = c;
c = c.offset(1);
*dst.offset(x_1 as isize) =
iclip_u8(*dst.offset(x_1 as isize) as c_int + *fresh1) as pixel;
x_1 += 1;
}
y_0 += 1;
dst = dst.offset(stride as isize);
}
}

#[cfg(all(feature = "asm", any(target_arch = "x86", target_arch = "x86_64")))]
#[inline(always)]
#[rustfmt::skip]
Expand Down Expand Up @@ -796,7 +725,7 @@ pub unsafe fn rav1d_itx_dsp_init_8bpc(c: *mut Rav1dInvTxfmDSPContext, mut _bpc:
// TODO(legare): Temporary import until init fns are deduplicated.
use crate::src::itx::*;

(*c).itxfm_add[TX_4X4 as usize][WHT_WHT as usize] = Some(inv_txfm_add_wht_wht_4x4_c_erased);
(*c).itxfm_add[TX_4X4 as usize][WHT_WHT as usize] = Some(inv_txfm_add_wht_wht_4x4_c_erased::<BitDepth8>);
(*c).itxfm_add[TX_4X4 as usize][DCT_DCT as usize] =
Some(inv_txfm_add_dct_dct_4x4_c_erased::<BitDepth8>);
(*c).itxfm_add[TX_4X4 as usize][IDTX as usize] =
Expand Down

0 comments on commit 9e19154

Please sign in to comment.