diff --git a/src/looprestoration.rs b/src/looprestoration.rs
index 54b402136..a10bf78a2 100644
--- a/src/looprestoration.rs
+++ b/src/looprestoration.rs
@@ -725,11 +725,11 @@ pub(crate) unsafe extern "C" fn selfguided_filter<BD: BitDepth>(
 }
 
 // TODO(randomPoison): Temporarily pub until init logic is deduplicated.
-pub(crate) unsafe extern "C" fn sgr_5x5_c<BD: BitDepth>(
-    mut p: *mut libc::c_void,
+pub(crate) unsafe extern "C" fn sgr_5x5_c_erased<BD: BitDepth>(
+    p: *mut libc::c_void,
     stride: ptrdiff_t,
     left: *const libc::c_void,
-    mut lpf: *const libc::c_void,
+    lpf: *const libc::c_void,
     w: libc::c_int,
     h: libc::c_int,
     params: *const LooprestorationParams,
@@ -753,7 +753,7 @@ unsafe fn sgr_5x5_rust<BD: BitDepth>(
     mut p: *mut BD::Pixel,
     stride: ptrdiff_t,
     left: *const [BD::Pixel; 4],
-    mut lpf: *const BD::Pixel,
+    lpf: *const BD::Pixel,
     w: libc::c_int,
     h: libc::c_int,
     params: *const LooprestorationParams,
@@ -779,13 +779,10 @@ unsafe fn sgr_5x5_rust<BD: BitDepth>(
         let mut i = 0;
         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) = bd.iclip_pixel(
                 (*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);
@@ -794,11 +791,11 @@ unsafe fn sgr_5x5_rust<BD: BitDepth>(
 }
 
 // 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,
+pub(crate) unsafe extern "C" fn sgr_3x3_c_erased<BD: BitDepth>(
+    p: *mut libc::c_void,
     stride: ptrdiff_t,
     left: *const libc::c_void,
-    mut lpf: *const libc::c_void,
+    lpf: *const libc::c_void,
     w: libc::c_int,
     h: libc::c_int,
     params: *const LooprestorationParams,
@@ -822,7 +819,7 @@ 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,
+    lpf: *const BD::Pixel,
     w: libc::c_int,
     h: libc::c_int,
     params: *const LooprestorationParams,
@@ -848,13 +845,10 @@ unsafe fn sgr_3x3_rust<BD: BitDepth>(
         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) = bd.iclip_pixel(
                 (*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);
@@ -863,11 +857,11 @@ unsafe fn sgr_3x3_rust<BD: BitDepth>(
 }
 
 // TODO(randomPoison): Temporarily pub until init logic is deduplicated.
-pub(crate) unsafe extern "C" fn sgr_mix_c<BD: BitDepth>(
-    mut p: *mut libc::c_void,
+pub(crate) unsafe extern "C" fn sgr_mix_c_erased<BD: BitDepth>(
+    p: *mut libc::c_void,
     stride: ptrdiff_t,
     left: *const libc::c_void,
-    mut lpf: *const libc::c_void,
+    lpf: *const libc::c_void,
     w: libc::c_int,
     h: libc::c_int,
     params: *const LooprestorationParams,
@@ -891,7 +885,7 @@ unsafe fn sgr_mix_rust<BD: BitDepth>(
     mut p: *mut BD::Pixel,
     stride: ptrdiff_t,
     left: *const [BD::Pixel; 4],
-    mut lpf: *const BD::Pixel,
+    lpf: *const BD::Pixel,
     w: libc::c_int,
     h: libc::c_int,
     params: *const LooprestorationParams,
@@ -930,13 +924,10 @@ unsafe fn sgr_mix_rust<BD: BitDepth>(
         while i < w {
             let v = w0 * dst0[(j * 384 + i) as usize].as_::<libc::c_int>()
                 + w1 * dst1[(j * 384 + i) as usize].as_::<libc::c_int>();
-            *p.offset(i as isize) = iclip(
+            *p.offset(i as isize) = bd.iclip_pixel(
                 (*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
index d049e0933..9d32c351a 100644
--- a/src/looprestoration_tmpl_16.rs
+++ b/src/looprestoration_tmpl_16.rs
@@ -131,17 +131,18 @@ pub type pixel = uint16_t;
 pub type coef = int32_t;
 pub type const_left_pixel_row = *const [pixel; 4];
 
-#[cfg(feature = "asm")]
-cfg_if! {
-    if #[cfg(any(target_arch = "arm", target_arch = "aarch64"))] {
-        use crate::src::looprestoration::LrEdgeFlags;
-        use crate::src::looprestoration::LooprestorationParams;
-    }
-}
-
+#[cfg(all(feature = "asm", any(target_arch = "arm", target_arch = "aarch64")))]
+#[rustfmt::skip]
+use crate::{
+    src::looprestoration::LrEdgeFlags,
+    src::looprestoration::LooprestorationParams,
+};
+
+use crate::src::looprestoration::sgr_3x3_c_erased;
+use crate::src::looprestoration::sgr_5x5_c_erased;
+use crate::src::looprestoration::sgr_mix_c_erased;
 use crate::src::looprestoration::wiener_c_erased;
 use crate::src::looprestoration::Dav1dLoopRestorationDSPContext;
-use crate::src::looprestoration::{sgr_3x3_c, sgr_5x5_c, sgr_mix_c};
 
 #[inline]
 unsafe extern "C" fn PXSTRIDE(x: ptrdiff_t) -> ptrdiff_t {
@@ -673,9 +674,9 @@ 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::<BitDepth16>;
-    (*c).sgr[2] = sgr_mix_c::<BitDepth16>;
+    (*c).sgr[0] = sgr_5x5_c_erased::<BitDepth16>;
+    (*c).sgr[1] = sgr_3x3_c_erased::<BitDepth16>;
+    (*c).sgr[2] = sgr_mix_c_erased::<BitDepth16>;
 
     #[cfg(feature = "asm")]
     cfg_if! {
diff --git a/src/looprestoration_tmpl_8.rs b/src/looprestoration_tmpl_8.rs
index 42b30be7f..24a029004 100644
--- a/src/looprestoration_tmpl_8.rs
+++ b/src/looprestoration_tmpl_8.rs
@@ -126,18 +126,19 @@ extern "C" {
 pub type pixel = uint8_t;
 pub type coef = int16_t;
 pub type const_left_pixel_row = *const [pixel; 4];
+use crate::src::looprestoration::sgr_3x3_c_erased;
+use crate::src::looprestoration::sgr_5x5_c_erased;
+use crate::src::looprestoration::sgr_mix_c_erased;
 use crate::src::looprestoration::wiener_c_erased;
 use crate::src::looprestoration::Dav1dLoopRestorationDSPContext;
-use crate::src::looprestoration::{sgr_3x3_c, sgr_5x5_c, sgr_mix_c};
 
-#[cfg(feature = "asm")]
-cfg_if! {
-    if #[cfg(any(target_arch = "arm", target_arch = "aarch64"))] {
-        use crate::include::stddef::*;
-        use crate::src::looprestoration::LrEdgeFlags;
-        use crate::src::looprestoration::LooprestorationParams;
-    }
-}
+#[cfg(all(feature = "asm", any(target_arch = "arm", target_arch = "aarch64")))]
+#[rustfmt::skip]
+use crate::{
+    include::stddef::ptrdiff_t,
+    src::looprestoration::LrEdgeFlags,
+    src::looprestoration::LooprestorationParams,
+};
 
 #[cfg(all(feature = "asm", any(target_arch = "x86", target_arch = "x86_64")))]
 #[inline(always)]
@@ -634,9 +635,9 @@ 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::<BitDepth8>;
-    (*c).sgr[2] = sgr_mix_c::<BitDepth8>;
+    (*c).sgr[0] = sgr_5x5_c_erased::<BitDepth8>;
+    (*c).sgr[1] = sgr_3x3_c_erased::<BitDepth8>;
+    (*c).sgr[2] = sgr_mix_c_erased::<BitDepth8>;
 
     #[cfg(feature = "asm")]
     cfg_if! {