[FFmpeg-devel] [PATCH 3/4] avcodec/v210: add avx2 version of the 10-bit line encoder

[James Darnley]

Around 25% faster than the ssse3 version.
---
New patch.  Should I squash this into the previous patch before committing?
---
 libavcodec/v210enc.c          | 11 +++++++++--
 libavcodec/x86/constants.c    |  3 ++-
 libavcodec/x86/constants.h    |  2 +-
 libavcodec/x86/v210enc.asm    | 20 +++++++++++++++-----
 libavcodec/x86/v210enc_init.c |  4 ++++
 5 files changed, 31 insertions(+), 9 deletions(-)

diff --git a/libavcodec/v210enc.c b/libavcodec/v210enc.c
index ec63864..00c89dc 100644
--- a/libavcodec/v210enc.c
+++ b/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);
 
diff --git a/libavcodec/x86/constants.c b/libavcodec/x86/constants.c
index 7e6883d..11002ee 100644
--- a/libavcodec/x86/constants.c
+++ b/libavcodec/x86/constants.c
@@ -27,7 +27,8 @@ DECLARE_ALIGNED(32, const ymm_reg,  ff_pw_1)    = { 0x0001000100010001ULL, 0x000
 DECLARE_ALIGNED(32, const ymm_reg,  ff_pw_2)    = { 0x0002000200020002ULL, 0x0002000200020002ULL,
                                                     0x0002000200020002ULL, 0x0002000200020002ULL };
 DECLARE_ALIGNED(16, const xmm_reg,  ff_pw_3)    = { 0x0003000300030003ULL, 0x0003000300030003ULL };
-DECLARE_ALIGNED(16, const xmm_reg,  ff_pw_4)    = { 0x0004000400040004ULL, 0x0004000400040004ULL };
+DECLARE_ALIGNED(32, const ymm_reg,  ff_pw_4)    = { 0x0004000400040004ULL, 0x0004000400040004ULL,
+                                                    0x0004000400040004ULL, 0x0004000400040004ULL };
 DECLARE_ALIGNED(16, const xmm_reg,  ff_pw_5)    = { 0x0005000500050005ULL, 0x0005000500050005ULL };
 DECLARE_ALIGNED(16, const xmm_reg,  ff_pw_8)    = { 0x0008000800080008ULL, 0x0008000800080008ULL };
 DECLARE_ALIGNED(16, const xmm_reg,  ff_pw_9)    = { 0x0009000900090009ULL, 0x0009000900090009ULL };
diff --git a/libavcodec/x86/constants.h b/libavcodec/x86/constants.h
index 8c02864..b82aef9 100644
--- a/libavcodec/x86/constants.h
+++ b/libavcodec/x86/constants.h
@@ -28,7 +28,7 @@
 extern const ymm_reg  ff_pw_1;
 extern const ymm_reg  ff_pw_2;
 extern const xmm_reg  ff_pw_3;
-extern const xmm_reg  ff_pw_4;
+extern const ymm_reg  ff_pw_4;
 extern const xmm_reg  ff_pw_5;
 extern const xmm_reg  ff_pw_8;
 extern const xmm_reg  ff_pw_9;
diff --git a/libavcodec/x86/v210enc.asm b/libavcodec/x86/v210enc.asm
index e6776a5..d74e5ac 100644
--- a/libavcodec/x86/v210enc.asm
+++ b/libavcodec/x86/v210enc.asm
@@ -51,7 +51,7 @@ 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
@@ -61,11 +61,19 @@ 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+widthq*2+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]
@@ -79,7 +87,7 @@ 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
@@ -87,6 +95,8 @@ cglobal v210_planar_pack_10, 5, 5, 4, y, u, v, dst, width
 
 INIT_XMM ssse3
 v210_planar_pack_10
+INIT_YMM avx2
+v210_planar_pack_10
 
 %macro v210_planar_pack_8 0
 
diff --git a/libavcodec/x86/v210enc_init.c b/libavcodec/x86/v210enc_init.c
index be6698c..ee48e80 100644
--- a/libavcodec/x86/v210enc_init.c
+++ b/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)
 {
@@ -44,6 +47,7 @@ av_cold void ff_v210enc_init_x86(V210EncContext *s)
 
     if (EXTERNAL_AVX2(cpu_flags)) {
         s->pack_line_8 = ff_v210_planar_pack_8_avx2;
+        s->pack_line_10 = ff_v210_planar_pack_10_avx2;
         s->sample_factor = 2;
     }
 }
-- 
2.6.2