From a439fb0b8b38fc265d79140a2fa9592232744161 Mon Sep 17 00:00:00 2001
From: Kostya Shishkov <kostya.shiskov@gmail.com>
Date: Sat, 21 Oct 2023 09:54:50 +0200
Subject: [PATCH] hwdec-vaapi: optimise chroma deinterleaving

---
 hwdec-vaapi/src/lib.rs | 96 +++++++++++++++++++++++++++++++++++++-----
 1 file changed, 86 insertions(+), 10 deletions(-)

diff --git a/hwdec-vaapi/src/lib.rs b/hwdec-vaapi/src/lib.rs
index ca843c2..7b580d5 100644
--- a/hwdec-vaapi/src/lib.rs
+++ b/hwdec-vaapi/src/lib.rs
@@ -283,6 +283,91 @@ pub struct VaapiH264Decoder {
     tb_den:         u32,
 }
 
+#[cfg(not(target_arch="x86_64"))]
+fn deint_chroma(frm: NASimpleVideoFrame<u8>, src: &[u8], sstride: usize) {
+    let mut uoff = frm.offset[1];
+    let mut voff = frm.offset[2];
+    for cline in src.chunks(sstride).take(frm.height[1]) {
+        for (x, pair) in cline.chunks_exact(2).take(frm.width[1]).enumerate() {
+            frm.data[uoff + x] = pair[0];
+            frm.data[voff + x] = pair[1];
+        }
+        uoff += frm.stride[1];
+        voff += frm.stride[2];
+    }
+}
+
+#[cfg(target_arch="x86_64")]
+use std::arch::asm;
+#[cfg(target_arch="x86_64")]
+fn deint_chroma(frm: NASimpleVideoFrame<u8>, src: &[u8], sstride: usize) {
+    unsafe {
+        let width = frm.width[1];
+        let height = frm.height[1];
+        let dst = frm.data.as_mut_ptr();
+        let udst = dst.add(frm.offset[1]);
+        let vdst = dst.add(frm.offset[2]);
+        let dstep = frm.stride[1] - width;
+        let sstep = sstride - width * 2;
+        asm!(
+            "2:",
+            "  mov {tmp}, {width}",
+            "  test {width}, 8",
+            "  jz 3f",
+            "  movaps   xmm0,     [{src}]",
+            "  movaps   xmm1,     xmm0",
+            "  psllw    xmm0,     8",
+            "  psrlw    xmm1,     8",
+            "  psrlw    xmm0,     8",
+            "  packuswb xmm1,     xmm1",
+            "  packuswb xmm0,     xmm0",
+            "  movq     [{vdst}], xmm1",
+            "  movq     [{udst}], xmm0",
+            "  add      {src},    16",
+            "  add      {vdst},   8",
+            "  add      {udst},   8",
+            "  sub      {tmp},    8",
+            "  3:",
+            "    movaps   xmm0,     [{src}]",
+            "    movaps   xmm1,     [{src} + 16]",
+            "    movaps   xmm2,     xmm0",
+            "    movaps   xmm3,     xmm1",
+            "    psllw    xmm0,     8",
+            "    psllw    xmm1,     8",
+            "    psrlw    xmm2,     8",
+            "    psrlw    xmm3,     8",
+            "    psrlw    xmm0,     8",
+            "    psrlw    xmm1,     8",
+            "    packuswb xmm2,     xmm3",
+            "    packuswb xmm0,     xmm1",
+            "    movups   [{vdst}], xmm2",
+            "    movups   [{udst}], xmm0",
+            "    add      {src},    32",
+            "    add      {vdst},   16",
+            "    add      {udst},   16",
+            "    sub      {tmp},    16",
+            "    jnz      3b",
+            "  add {udst}, {dstep}",
+            "  add {vdst}, {dstep}",
+            "  add {src},  {sstep}",
+            "  dec {height}",
+            "  jnz 2b",
+            src = inout(reg) src.as_ptr() => _,
+            udst = inout(reg) udst => _,
+            vdst = inout(reg) vdst => _,
+            width = in(reg) width,
+            height = inout(reg) height => _,
+            dstep = in(reg) dstep,
+            sstep = in(reg) sstep,
+            tmp = out(reg) _,
+            out("xmm0") _,
+            out("xmm1") _,
+            out("xmm2") _,
+            out("xmm3") _,
+        );
+    }
+}
+
 fn fill_frame(ifmt: VAImageFormat, pic: &Picture<PictureSync>, frm: &mut NABufferType) -> DecoderResult<()> {
     let mut vbuf = frm.get_vbuf().unwrap();
     let (w, h) = pic.surface_size();
@@ -305,16 +390,7 @@ fn fill_frame(ifmt: VAImageFormat, pic: &Picture<PictureSync>, frm: &mut NABuffe
                 dline[..frm.width[0]].copy_from_slice(&sline[..frm.width[0]]);
             }
 
-            let mut uoff = frm.offset[1];
-            let mut voff = frm.offset[2];
-            for cline in imgdata[iimg.offsets[1] as usize..].chunks(iimg.pitches[1] as usize).take(frm.height[1]) {
-                for (x, pair) in cline.chunks_exact(2).take(frm.width[1]).enumerate() {
-                    frm.data[uoff + x] = pair[0];
-                    frm.data[voff + x] = pair[1];
-                }
-                uoff += frm.stride[1];
-                voff += frm.stride[2];
-            }
+            deint_chroma(frm, &imgdata[iimg.offsets[1] as usize..], iimg.pitches[1] as usize);
         },
         _ => unimplemented!(),
     };
-- 
2.39.5