optimise luma copy part a bit

diff --git a/hwdec-vaapi/src/lib.rs b/hwdec-vaapi/src/lib.rs
index 1ef3e415eeb747e376dfc4b3fd1e71173105097b..40e14f445fc17e24793419ae76b82cc1d2b2eba0 100644 (file)
--- a/hwdec-vaapi/src/lib.rs
+++ b/hwdec-vaapi/src/lib.rs
@@ -283,6 +283,14 @@ pub struct VaapiH264Decoder {
     tb_den:         u32,
 }
 
+#[cfg(not(target_arch="x86_64"))]
+fn copy_luma(dst: &mut [u8], dstride: usize, src: &[u8], sstride: usize, w: usize, h: usize) {
+    for (dline, sline) in dst.chunks_mut(dstride)
+            .zip(src.chunks(sstride))
+            .take(h) {
+        dline[..w].copy_from_slice(&sline[..w]);
+    }
+}
 #[cfg(not(target_arch="x86_64"))]
 fn deint_chroma(frm: NASimpleVideoFrame<u8>, src: &[u8], sstride: usize) {
     let mut uoff = frm.offset[1];
@@ -300,6 +308,46 @@ fn deint_chroma(frm: NASimpleVideoFrame<u8>, src: &[u8], sstride: usize) {
 #[cfg(target_arch="x86_64")]
 use std::arch::asm;
 #[cfg(target_arch="x86_64")]
+fn copy_luma(dst: &mut [u8], dstride: usize, src: &[u8], sstride: usize, w: usize, h: usize) {
+    if dst.as_ptr().align_offset(32) == 0 && src.as_ptr().align_offset(32) == 0 &&
+            (w % 64) == 0 && ((dstride | sstride) % 32) == 0 {
+        unsafe {
+            asm!(
+                "2:",
+                "  mov {x}, {w}",
+                "  3:",
+                "    vmovdqa ymm0, [{src}]",
+                "    vmovdqa ymm1, [{src}+32]",
+                "    vmovdqa [{dst}], ymm0",
+                "    vmovdqa [{dst}+32], ymm1",
+                "    add {src}, 64",
+                "    add {dst}, 64",
+                "    sub {x},   64",
+                "    jnz 3b",
+                "  add {src}, {sstep}",
+                "  add {dst}, {dstep}",
+                "  dec {h}",
+                "  jnz 2b",
+                dst = inout(reg) dst.as_mut_ptr() => _,
+                src = inout(reg) src.as_ptr() => _,
+                sstep = in(reg) sstride - w,
+                dstep = in(reg) dstride - w,
+                w = in(reg) w,
+                h = in(reg) h,
+                x = out(reg) _,
+                out("ymm0") _,
+                out("ymm1") _,
+            );
+        }
+    } else {
+        for (dline, sline) in dst.chunks_mut(dstride)
+                .zip(src.chunks(sstride))
+                .take(h) {
+            dline[..w].copy_from_slice(&sline[..w]);
+        }
+    }
+}
+#[cfg(target_arch="x86_64")]
 fn deint_chroma(frm: NASimpleVideoFrame<u8>, src: &[u8], sstride: usize) {
     unsafe {
         let width = frm.width[1];
@@ -384,11 +432,7 @@ fn fill_frame(ifmt: VAImageFormat, pic: &Picture<PictureSync>, frm: &mut NABuffe
             validate!(iimg.width == (frm.width[0] as u16));
             validate!(iimg.height == (frm.height[0] as u16));
 
-            for (dline, sline) in frm.data[frm.offset[0]..].chunks_mut(frm.stride[0])
-                    .zip(imgdata[iimg.offsets[0] as usize..].chunks(iimg.pitches[0] as usize))
-                    .take(frm.height[0]) {
-                dline[..frm.width[0]].copy_from_slice(&sline[..frm.width[0]]);
-            }
+            copy_luma(&mut frm.data[frm.offset[0]..], frm.stride[0], &imgdata[iimg.offsets[0] as usize..], iimg.pitches[0] as usize, frm.width[0], frm.height[0]);
 
             deint_chroma(frm, &imgdata[iimg.offsets[1] as usize..], iimg.pitches[1] as usize);
         },