v210: Add avx2 version of the 10-bit line encoder

Around 25% faster than the ssse3 version.
Signed-off-by: Luca Barbato <lu_zero@gentoo.org>

libavcodec/v210enc.c

@@ -135,13 +135,20 @@ static int encode_frame(AVCodecContext *avctx, AVPacket *pkt,
         const uint16_t *v = (const uint16_t *)pic->data[2];
         for (h = 0; h < avctx->height; h++) {
             uint32_t val;
-            w = (avctx->width / 6) * 6;
+            w = (avctx->width / (6 * s->sample_factor)) * 6 * s->sample_factor;
             s->pack_line_10(y, u, v, dst, w);
 
             y += w;
             u += w >> 1;
             v += w >> 1;
-            dst += (w / 6) * 16;
+            dst += (w / (6 * s->sample_factor)) * 16 * s->sample_factor;
+
+            for (; w < avctx->width - 5; w += 6) {
+                WRITE_PIXELS(u, y, v);
+                WRITE_PIXELS(y, u, y);
+                WRITE_PIXELS(v, y, u);
+                WRITE_PIXELS(y, v, y);
+            }
             if (w < avctx->width - 1) {
                 WRITE_PIXELS(u, y, v);
 

libavcodec/x86/v210enc.asm

@@ -23,8 +23,8 @@
 
 SECTION_RODATA 32
 
-v210_enc_min_10: times 16 dw 0x4
-v210_enc_max_10: times 16 dw 0x3fb
+v210_enc_min_10: times 32 dw 0x4
+v210_enc_max_10: times 32 dw 0x3fb
 
 v210_enc_luma_mult_10: times 2 dw 4,1,16,4,1,16,0,0
 v210_enc_luma_shuf_10: times 2 db -1,0,1,-1,2,3,4,5,-1,6,7,-1,8,9,10,11
@@ -47,7 +47,7 @@ SECTION .text
 %macro v210_planar_pack_10 0
 
 ; v210_planar_pack_10(const uint16_t *y, const uint16_t *u, const uint16_t *v, uint8_t *dst, ptrdiff_t width)
-cglobal v210_planar_pack_10, 5, 5, 4, y, u, v, dst, width
+cglobal v210_planar_pack_10, 5, 5, 4+cpuflag(avx2), y, u, v, dst, width
     lea     r0, [yq+2*widthq]
     add     uq, widthq
     add     vq, widthq
@@ -57,11 +57,19 @@ cglobal v210_planar_pack_10, 5, 5, 4, y, u, v, dst, width
     mova    m3, [v210_enc_max_10]
 
 .loop
-    movu    m0, [yq+2*widthq]
+    movu        xm0, [yq+2*widthq]
+%if cpuflag(avx2)
+    vinserti128 m0, m0, [yq+2*widthq+12], 1
+%endif
     CLIPW   m0, m2, m3
 
-    movq    m1, [uq+widthq]
-    movhps  m1, [vq+widthq]
+    movq    xm1, [uq+widthq]
+    movhps  xm1, [vq+widthq]
+%if cpuflag(avx2)
+    movq         xm4, [uq+widthq+6]
+    movhps       xm4, [vq+widthq+6]
+    vinserti128  m1, m1, xm4, 1
+%endif
     CLIPW   m1, m2, m3
 
     pmullw  m0, [v210_enc_luma_mult_10]
@@ -75,7 +83,7 @@ cglobal v210_planar_pack_10, 5, 5, 4, y, u, v, dst, width
     movu    [dstq], m0
 
     add     dstq, mmsize
-    add     widthq, 6
+    add     widthq, (mmsize*3)/8
     jl .loop
 
     RET
@@ -85,6 +93,10 @@ cglobal v210_planar_pack_10, 5, 5, 4, y, u, v, dst, width
 INIT_XMM ssse3
 v210_planar_pack_10
 %endif
+%if HAVE_AVX2_EXTERNAL
+INIT_YMM avx2
+v210_planar_pack_10
+%endif
 
 %macro v210_planar_pack_8 0
 

libavcodec/x86/v210enc_init.c

@@ -29,6 +29,9 @@ void ff_v210_planar_pack_8_avx2(const uint8_t *y, const uint8_t *u,
 void ff_v210_planar_pack_10_ssse3(const uint16_t *y, const uint16_t *u,
                                   const uint16_t *v, uint8_t *dst,
                                   ptrdiff_t width);
+void ff_v210_planar_pack_10_avx2(const uint16_t *y, const uint16_t *u,
+                                 const uint16_t *v, uint8_t *dst,
+                                 ptrdiff_t width);
 
 av_cold void ff_v210enc_init_x86(V210EncContext *s)
 {
@@ -45,5 +48,6 @@ av_cold void ff_v210enc_init_x86(V210EncContext *s)
     if (EXTERNAL_AVX2(cpu_flags)) {
         s->sample_factor = 2;
         s->pack_line_8   = ff_v210_planar_pack_8_avx2;
+        s->pack_line_10  = ff_v210_planar_pack_10_avx2;
     }
 }