fn sgr_3x3: Deduplicate w/ generics

memorysafety · Jul 24, 2023 · 8e5428f · 8e5428f
1 parent 96d6814
commit 8e5428f
Show file tree

Hide file tree

Showing 3 changed files with 77 additions and 134 deletions.
diff --git a/src/looprestoration.rs b/src/looprestoration.rs
@@ -780,10 +780,81 @@ unsafe fn sgr_5x5_rust<BD: BitDepth>(
         while i < w {
             let v = w0 * dst[(j * 384 + i) as usize].as_::<libc::c_int>();
             *p.offset(i as isize) = iclip(
-                (*p.offset(i as isize)).as_::<libc::c_int>() + (v + ((1 as libc::c_int) << 10) >> 11),
+                (*p.offset(i as isize)).as_::<libc::c_int>()
+                    + (v + ((1 as libc::c_int) << 10) >> 11),
                 0 as libc::c_int,
                 bd.bitdepth_max().as_(),
-            ).as_();
+            )
+            .as_();
+            i += 1;
+        }
+        p = p.offset(BD::pxstride(stride as usize) as isize);
+        j += 1;
+    }
+}
+
+// TODO(randomPoison): Temporarily pub until init logic is deduplicated.
+pub(crate) unsafe extern "C" fn sgr_3x3_c<BD: BitDepth>(
+    mut p: *mut libc::c_void,
+    stride: ptrdiff_t,
+    left: *const libc::c_void,
+    mut lpf: *const libc::c_void,
+    w: libc::c_int,
+    h: libc::c_int,
+    params: *const LooprestorationParams,
+    edges: LrEdgeFlags,
+    bitdepth_max: libc::c_int,
+) {
+    sgr_3x3_rust(
+        p.cast(),
+        stride,
+        left.cast(),
+        lpf.cast(),
+        w,
+        h,
+        params,
+        edges,
+        BD::from_c(bitdepth_max),
+    )
+}
+
+unsafe fn sgr_3x3_rust<BD: BitDepth>(
+    mut p: *mut BD::Pixel,
+    stride: ptrdiff_t,
+    left: *const [BD::Pixel; 4],
+    mut lpf: *const BD::Pixel,
+    w: libc::c_int,
+    h: libc::c_int,
+    params: *const LooprestorationParams,
+    edges: LrEdgeFlags,
+    bd: BD,
+) {
+    let mut tmp: [BD::Pixel; 27300] = [0.as_(); 27300];
+    let mut dst: [BD::Coef; 24576] = [0.as_(); 24576];
+    padding::<BD>(&mut tmp, p, stride, left, lpf, w, h, edges);
+    selfguided_filter(
+        dst.as_mut_ptr(),
+        tmp.as_mut_ptr(),
+        390 as libc::c_int as ptrdiff_t,
+        w,
+        h,
+        9 as libc::c_int,
+        (*params).sgr.s1,
+        bd,
+    );
+    let w1 = (*params).sgr.w1 as libc::c_int;
+    let mut j = 0;
+    while j < h {
+        let mut i = 0;
+        while i < w {
+            let v = w1 * dst[(j * 384 + i) as usize].as_::<libc::c_int>();
+            *p.offset(i as isize) = iclip(
+                (*p.offset(i as isize)).as_::<libc::c_int>()
+                    + (v + ((1 as libc::c_int) << 10) >> 11),
+                0 as libc::c_int,
+                bd.bitdepth_max().as_(),
+            )
+            .as_();
             i += 1;
         }
         p = p.offset(BD::pxstride(stride as usize) as isize);

diff --git a/src/looprestoration_tmpl_16.rs b/src/looprestoration_tmpl_16.rs
@@ -139,7 +139,7 @@ use crate::src::looprestoration::padding;
 use crate::src::looprestoration::selfguided_filter;
 use crate::src::looprestoration::wiener_c_erased;
 use crate::src::looprestoration::Dav1dLoopRestorationDSPContext;
-use crate::src::looprestoration::sgr_5x5_c;
+use crate::src::looprestoration::{sgr_3x3_c, sgr_5x5_c};
 
 #[inline]
 unsafe extern "C" fn PXSTRIDE(x: ptrdiff_t) -> ptrdiff_t {
@@ -149,72 +149,6 @@ unsafe extern "C" fn PXSTRIDE(x: ptrdiff_t) -> ptrdiff_t {
     return x >> 1;
 }
 
-unsafe extern "C" fn sgr_3x3_c(
-    mut p: *mut libc::c_void,
-    stride: ptrdiff_t,
-    left: *const libc::c_void,
-    mut lpf: *const libc::c_void,
-    w: libc::c_int,
-    h: libc::c_int,
-    params: *const LooprestorationParams,
-    edges: LrEdgeFlags,
-    bitdepth_max: libc::c_int,
-) {
-    sgr_3x3_rust(
-        p.cast(),
-        stride,
-        left.cast(),
-        lpf.cast(),
-        w,
-        h,
-        params,
-        edges,
-        bitdepth_max,
-    )
-}
-
-unsafe extern "C" fn sgr_3x3_rust(
-    mut p: *mut pixel,
-    stride: ptrdiff_t,
-    left: *const [pixel; 4],
-    mut lpf: *const pixel,
-    w: libc::c_int,
-    h: libc::c_int,
-    params: *const LooprestorationParams,
-    edges: LrEdgeFlags,
-    bitdepth_max: libc::c_int,
-) {
-    let mut tmp: [pixel; 27300] = [0; 27300];
-    let mut dst: [coef; 24576] = [0; 24576];
-    padding::<BitDepth16>(&mut tmp, p, stride, left, lpf, w, h, edges);
-    selfguided_filter(
-        dst.as_mut_ptr(),
-        tmp.as_mut_ptr(),
-        390 as libc::c_int as ptrdiff_t,
-        w,
-        h,
-        9 as libc::c_int,
-        (*params).sgr.s1,
-        BitDepth16::from_c(bitdepth_max),
-    );
-    let w1 = (*params).sgr.w1 as libc::c_int;
-    let mut j = 0;
-    while j < h {
-        let mut i = 0;
-        while i < w {
-            let v = w1 * dst[(j * 384 + i) as usize];
-            *p.offset(i as isize) = iclip(
-                *p.offset(i as isize) as libc::c_int + (v + ((1 as libc::c_int) << 10) >> 11),
-                0 as libc::c_int,
-                bitdepth_max,
-            ) as pixel;
-            i += 1;
-        }
-        p = p.offset(PXSTRIDE(stride) as isize);
-        j += 1;
-    }
-}
-
 unsafe extern "C" fn sgr_mix_c(
     mut p: *mut libc::c_void,
     stride: ptrdiff_t,
@@ -816,7 +750,7 @@ pub unsafe extern "C" fn dav1d_loop_restoration_dsp_init_16bpc(
     (*c).wiener[1] = wiener_c_erased::<BitDepth16>;
     (*c).wiener[0] = (*c).wiener[1];
     (*c).sgr[0] = sgr_5x5_c::<BitDepth16>;
-    (*c).sgr[1] = sgr_3x3_c;
+    (*c).sgr[1] = sgr_3x3_c::<BitDepth16>;
     (*c).sgr[2] = sgr_mix_c;
 
     #[cfg(feature = "asm")]

diff --git a/src/looprestoration_tmpl_8.rs b/src/looprestoration_tmpl_8.rs
@@ -134,72 +134,10 @@ use crate::src::looprestoration::selfguided_filter;
 use crate::src::looprestoration::wiener_c_erased;
 use crate::src::looprestoration::Dav1dLoopRestorationDSPContext;
 use crate::src::looprestoration::LooprestorationParams;
-use crate::src::looprestoration::sgr_5x5_c;
+use crate::src::looprestoration::{sgr_3x3_c, sgr_5x5_c};
 
 use crate::include::common::intops::iclip_u8;
 
-unsafe extern "C" fn sgr_3x3_c(
-    mut p: *mut libc::c_void,
-    stride: ptrdiff_t,
-    left: *const libc::c_void,
-    mut lpf: *const libc::c_void,
-    w: libc::c_int,
-    h: libc::c_int,
-    params: *const LooprestorationParams,
-    edges: LrEdgeFlags,
-    _bitdepth_max: libc::c_int,
-) {
-    sgr_3x3_rust(
-        p.cast(),
-        stride,
-        left.cast(),
-        lpf.cast(),
-        w,
-        h,
-        params,
-        edges,
-    )
-}
-
-unsafe extern "C" fn sgr_3x3_rust(
-    mut p: *mut pixel,
-    stride: ptrdiff_t,
-    left: *const [pixel; 4],
-    mut lpf: *const pixel,
-    w: libc::c_int,
-    h: libc::c_int,
-    params: *const LooprestorationParams,
-    edges: LrEdgeFlags,
-) {
-    let mut tmp: [pixel; 27300] = [0; 27300];
-    let mut dst: [coef; 24576] = [0; 24576];
-    padding::<BitDepth8>(&mut tmp, p, stride, left, lpf, w, h, edges);
-    selfguided_filter(
-        dst.as_mut_ptr(),
-        tmp.as_mut_ptr(),
-        390 as libc::c_int as ptrdiff_t,
-        w,
-        h,
-        9 as libc::c_int,
-        (*params).sgr.s1,
-        BitDepth8::new(()),
-    );
-    let w1 = (*params).sgr.w1 as libc::c_int;
-    let mut j = 0;
-    while j < h {
-        let mut i = 0;
-        while i < w {
-            let v = w1 * dst[(j * 384 + i) as usize] as libc::c_int;
-            *p.offset(i as isize) = iclip_u8(
-                *p.offset(i as isize) as libc::c_int + (v + ((1 as libc::c_int) << 10) >> 11),
-            ) as pixel;
-            i += 1;
-        }
-        p = p.offset(stride as isize);
-        j += 1;
-    }
-}
-
 unsafe extern "C" fn sgr_mix_c(
     mut p: *mut libc::c_void,
     stride: ptrdiff_t,
@@ -771,7 +709,7 @@ pub unsafe extern "C" fn dav1d_loop_restoration_dsp_init_8bpc(
     (*c).wiener[1] = wiener_c_erased::<BitDepth8>;
     (*c).wiener[0] = (*c).wiener[1];
     (*c).sgr[0] = sgr_5x5_c::<BitDepth8>;
-    (*c).sgr[1] = sgr_3x3_c;
+    (*c).sgr[1] = sgr_3x3_c::<BitDepth8>;
     (*c).sgr[2] = sgr_mix_c;
 
     #[cfg(feature = "asm")]