hwdec-vaapi: add SSE branch for copy_luma()

diff --git a/hwdec-vaapi/src/lib.rs b/hwdec-vaapi/src/lib.rs
index 378289b53d2edf7e920ed8b5a4867913a7d2f539..85749d4fe347b107b09e5fc1ab146449749b8740 100644 (file)
--- a/hwdec-vaapi/src/lib.rs
+++ b/hwdec-vaapi/src/lib.rs
@@ -347,6 +347,42 @@ fn copy_luma(dst: &mut [u8], dstride: usize, src: &[u8], sstride: usize, w: usiz
                 out("ymm1") _,
             );
         }
+    } else if dst.as_ptr().align_offset(16) == 0 && src.as_ptr().align_offset(16) == 0 &&
+            (w % 64) == 0 && ((dstride | sstride) % 16) == 0 {
+        unsafe {
+            asm!(
+                "2:",
+                "  mov {x}, {w}",
+                "  3:",
+                "    movdqa xmm0, [{src}]",
+                "    movdqa xmm1, [{src}+16]",
+                "    movdqa xmm2, [{src}+32]",
+                "    movdqa xmm3, [{src}+48]",
+                "    movdqa [{dst}], xmm0",
+                "    movdqa [{dst}+16], xmm1",
+                "    movdqa [{dst}+32], xmm2",
+                "    movdqa [{dst}+48], xmm3",
+                "    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 = inout(reg) h => _,
+                x = out(reg) _,
+                out("xmm0") _,
+                out("xmm1") _,
+                out("xmm2") _,
+                out("xmm3") _,
+            );
+        }
     } else {
         let copy_len = dstride.min(w);
         for (dline, sline) in dst.chunks_mut(dstride)