From 82f1307ee4a5a34978694f132076e86a11b26967 Mon Sep 17 00:00:00 2001
From: Kostya Shishkov <kostya.shishkov@gmail.com>
Date: Wed, 31 Aug 2022 16:14:47 +0200
Subject: [PATCH] h264/dsp: split chroma_interp() by width

---
 nihav-itu/src/codecs/h264/dsp/mc/debug.rs   | 14 +++++++++++++-
 nihav-itu/src/codecs/h264/dsp/mc/mod.rs     |  8 +++++---
 nihav-itu/src/codecs/h264/dsp/mc/release.rs | 14 +++++++++++++-
 3 files changed, 31 insertions(+), 5 deletions(-)

diff --git a/nihav-itu/src/codecs/h264/dsp/mc/debug.rs b/nihav-itu/src/codecs/h264/dsp/mc/debug.rs
index 34d59da..e180800 100644
--- a/nihav-itu/src/codecs/h264/dsp/mc/debug.rs
+++ b/nihav-itu/src/codecs/h264/dsp/mc/debug.rs
@@ -164,7 +164,7 @@ fn h264_mc33(dst: &mut [u8], dstride: usize, src: &[u8], sstride: usize, w: usiz
 }
 
 
-pub fn chroma_interp(dst: &mut [u8], dstride: usize, src: &[u8], sstride: usize, dx: u16, dy: u16, w: usize, h: usize) {
+fn chroma_interp(dst: &mut [u8], dstride: usize, src: &[u8], sstride: usize, dx: u16, dy: u16, w: usize, h: usize) {
     let a0 = 8 - dx;
     let a1 = dx;
     let b0 = 8 - dy;
@@ -202,6 +202,18 @@ pub fn chroma_interp(dst: &mut [u8], dstride: usize, src: &[u8], sstride: usize,
     }
 }
 
+pub fn chroma_interp_8(dst: &mut [u8], dstride: usize, src: &[u8], sstride: usize, dx: u16, dy: u16, h: usize) {
+    chroma_interp(dst, dstride, src, sstride, dx, dy, 8, h);
+}
+
+pub fn chroma_interp_4(dst: &mut [u8], dstride: usize, src: &[u8], sstride: usize, dx: u16, dy: u16, h: usize) {
+    chroma_interp(dst, dstride, src, sstride, dx, dy, 4, h);
+}
+
+pub fn chroma_interp_2(dst: &mut [u8], dstride: usize, src: &[u8], sstride: usize, dx: u16, dy: u16, h: usize) {
+    chroma_interp(dst, dstride, src, sstride, dx, dy, 2, h);
+}
+
 macro_rules! luma_mc {
     ($orig:ident, $func4:ident, $func8:ident, $func16:ident) => {
         fn $func4(dst: &mut [u8], dstride: usize, src: &[u8], sstride: usize, h: usize) {
diff --git a/nihav-itu/src/codecs/h264/dsp/mc/mod.rs b/nihav-itu/src/codecs/h264/dsp/mc/mod.rs
index e2036cc..55b124e 100644
--- a/nihav-itu/src/codecs/h264/dsp/mc/mod.rs
+++ b/nihav-itu/src/codecs/h264/dsp/mc/mod.rs
@@ -23,6 +23,7 @@ pub struct H264MC {
     avg_buf:    NAVideoBufferRef<u8>,
     pub put_block_weighted:     [fn (dst: &mut [u8], stride: usize, src: &[u8], h: usize, wparams: [i8; 3]); 4],
     pub put_block_weighted2:    [fn (dst: &mut [u8], stride: usize, src0: &[u8], src1: &[u8], h: usize, wparams: [i8; 5]); 4],
+    pub chroma_interp:          [fn (dst: &mut [u8], dstride: usize, src: &[u8], sstride: usize, dx: u16, dy: u16, h: usize); 3],
 }
 
 impl H264MC {
@@ -31,6 +32,7 @@ impl H264MC {
             avg_buf,
             put_block_weighted:     [put_blk_w_2, put_blk_w_4, put_blk_w_8, put_blk_w_16],
             put_block_weighted2:    [put_blk_w2_2, put_blk_w2_4, put_blk_w2_8, put_blk_w2_16],
+            chroma_interp:          [chroma_interp_2, chroma_interp_4, chroma_interp_8],
         };
         obj.register_simd();
         obj
@@ -85,7 +87,7 @@ impl H264MC {
             };
         for chroma in 1..3 {
             let off = frm.offset[chroma] + xpos / 2 + (ypos / 2) * frm.stride[chroma];
-            chroma_interp(&mut frm.data[off..], frm.stride[chroma], csrc[chroma - 1], cstride[chroma - 1], dx, dy, cbw, cbh);
+            (self.chroma_interp[wmode])(&mut frm.data[off..], frm.stride[chroma], csrc[chroma - 1], cstride[chroma - 1], dx, dy, cbh);
         }
     }
 
@@ -144,8 +146,8 @@ impl H264MC {
                  &src[svoff + (src_x as usize) + (src_y as usize) * svstride..]],
                  [sustride, svstride])
             };
-        chroma_interp(udst, 16, csrc[0], cstride[0], dx, dy, cbw, cbh);
-        chroma_interp(vdst, 16, csrc[1], cstride[1], dx, dy, cbw, cbh);
+        (self.chroma_interp[wmode])(udst, 16, csrc[0], cstride[0], dx, dy, cbh);
+        (self.chroma_interp[wmode])(vdst, 16, csrc[1], cstride[1], dx, dy, cbh);
     }
 
     pub fn avg(&mut self, dst: &mut [u8], dstride: usize, bw: usize, bh: usize, comp: usize) {
diff --git a/nihav-itu/src/codecs/h264/dsp/mc/release.rs b/nihav-itu/src/codecs/h264/dsp/mc/release.rs
index 0a521d0..e4e7e16 100644
--- a/nihav-itu/src/codecs/h264/dsp/mc/release.rs
+++ b/nihav-itu/src/codecs/h264/dsp/mc/release.rs
@@ -204,7 +204,7 @@ fn h264_mc33(dst: &mut [u8], dstride: usize, src: &[u8], sstride: usize, w: usiz
 }
 
 
-pub fn chroma_interp(dst: &mut [u8], dstride: usize, src: &[u8], sstride: usize, dx: u16, dy: u16, w: usize, h: usize) {
+fn chroma_interp(dst: &mut [u8], dstride: usize, src: &[u8], sstride: usize, dx: u16, dy: u16, w: usize, h: usize) {
     let a0 = 8 - dx;
     let a1 = dx;
     let b0 = 8 - dy;
@@ -274,6 +274,18 @@ pub fn chroma_interp(dst: &mut [u8], dstride: usize, src: &[u8], sstride: usize,
     }
 }
 
+pub fn chroma_interp_8(dst: &mut [u8], dstride: usize, src: &[u8], sstride: usize, dx: u16, dy: u16, h: usize) {
+    chroma_interp(dst, dstride, src, sstride, dx, dy, 8, h);
+}
+
+pub fn chroma_interp_4(dst: &mut [u8], dstride: usize, src: &[u8], sstride: usize, dx: u16, dy: u16, h: usize) {
+    chroma_interp(dst, dstride, src, sstride, dx, dy, 4, h);
+}
+
+pub fn chroma_interp_2(dst: &mut [u8], dstride: usize, src: &[u8], sstride: usize, dx: u16, dy: u16, h: usize) {
+    chroma_interp(dst, dstride, src, sstride, dx, dy, 2, h);
+}
+
 macro_rules! luma_mc {
     ($orig:ident, $func4:ident, $func8:ident, $func16:ident) => {
         fn $func4(dst: &mut [u8], dstride: usize, src: &[u8], sstride: usize, h: usize) {
-- 
2.30.2