libavcodec Adding ff_v210_planar_unpack AVX2

Replaced VSHUFPS with VPBLENDD to relieve port 5 bottleneck
AVX2 is 1.4x faster than AVX

3 files changed, 73 insertions, 17 deletions

diff --git a/libavcodec/v210dec.c b/libavcodec/v210dec.c
index fd8a6b0d78..bc1e1d34ff 100644
--- a/libavcodec/v210dec.c
+++ b/libavcodec/v210dec.c
@@ -123,7 +123,7 @@ static int decode_frame(AVCodecContext *avctx, void *data, int *got_frame,
         const uint32_t *src = (const uint32_t*)psrc;
         uint32_t val;
 
-        w = (avctx->width / 6) * 6;
+        w = (avctx->width / 12) * 12;
         s->unpack_frame(src, y, u, v, w);
 
         y += w;
@@ -131,6 +131,14 @@ static int decode_frame(AVCodecContext *avctx, void *data, int *got_frame,
         v += w >> 1;
         src += (w << 1) / 3;
 
+        if (w < avctx->width - 5) {
+            READ_PIXELS(u, y, v);
+            READ_PIXELS(y, u, y);
+            READ_PIXELS(v, y, u);
+            READ_PIXELS(y, v, y);
+            w += 6;
+        }
+
         if (w < avctx->width - 1) {
             READ_PIXELS(u, y, v);
 
diff --git a/libavcodec/x86/v210-init.c b/libavcodec/x86/v210-init.c
index d64dbca1a8..cb9a6cbd6a 100644
--- a/libavcodec/x86/v210-init.c
+++ b/libavcodec/x86/v210-init.c
@@ -21,9 +21,11 @@
 
 extern void ff_v210_planar_unpack_unaligned_ssse3(const uint32_t *src, uint16_t *y, uint16_t *u, uint16_t *v, int width);
 extern void ff_v210_planar_unpack_unaligned_avx(const uint32_t *src, uint16_t *y, uint16_t *u, uint16_t *v, int width);
+extern void ff_v210_planar_unpack_unaligned_avx2(const uint32_t *src, uint16_t *y, uint16_t *u, uint16_t *v, int width);
 
 extern void ff_v210_planar_unpack_aligned_ssse3(const uint32_t *src, uint16_t *y, uint16_t *u, uint16_t *v, int width);
 extern void ff_v210_planar_unpack_aligned_avx(const uint32_t *src, uint16_t *y, uint16_t *u, uint16_t *v, int width);
+extern void ff_v210_planar_unpack_aligned_avx2(const uint32_t *src, uint16_t *y, uint16_t *u, uint16_t *v, int width);
 
 av_cold void ff_v210_x86_init(V210DecContext *s)
 {
@@ -36,6 +38,9 @@ av_cold void ff_v210_x86_init(V210DecContext *s)
 
         if (HAVE_AVX_EXTERNAL && cpu_flags & AV_CPU_FLAG_AVX)
             s->unpack_frame = ff_v210_planar_unpack_aligned_avx;
+
+        if (HAVE_AVX2_EXTERNAL && cpu_flags & AV_CPU_FLAG_AVX2)
+            s->unpack_frame = ff_v210_planar_unpack_aligned_avx2;
     }
     else {
         if (cpu_flags & AV_CPU_FLAG_SSSE3)
@@ -43,6 +48,9 @@ av_cold void ff_v210_x86_init(V210DecContext *s)
 
         if (HAVE_AVX_EXTERNAL && cpu_flags & AV_CPU_FLAG_AVX)
             s->unpack_frame = ff_v210_planar_unpack_unaligned_avx;
+
+        if (HAVE_AVX2_EXTERNAL && cpu_flags & AV_CPU_FLAG_AVX2)
+            s->unpack_frame = ff_v210_planar_unpack_unaligned_avx2;
     }
 #endif
 }
diff --git a/libavcodec/x86/v210.asm b/libavcodec/x86/v210.asm
index c24c765e5b..706712313d 100644
--- a/libavcodec/x86/v210.asm
+++ b/libavcodec/x86/v210.asm
@@ -22,9 +22,14 @@
 
 %include "libavutil/x86/x86util.asm"
 
-SECTION_RODATA
+SECTION_RODATA 32
+
+; for AVX2 version only
+v210_luma_permute: dd 0,1,2,4,5,6,7,7  ; 32-byte alignment required
+v210_chroma_shuf2: db 0,1,2,3,4,5,8,9,10,11,12,13,-1,-1,-1,-1
+v210_luma_shuf_avx2: db 0,1,4,5,6,7,8,9,12,13,14,15,-1,-1,-1,-1
+v210_chroma_shuf_avx2: db 0,1,4,5,10,11,-1,-1,2,3,8,9,12,13,-1,-1
 
-v210_mask: times 4 dd 0x3ff
 v210_mult: dw 64,4,64,4,64,4,64,4
 v210_luma_shuf: db 8,9,0,1,2,3,12,13,4,5,6,7,-1,-1,-1,-1
 v210_chroma_shuf: db 0,1,8,9,6,7,-1,-1,2,3,4,5,12,13,-1,-1
@@ -34,40 +39,65 @@ SECTION .text
 %macro v210_planar_unpack 1
 
 ; v210_planar_unpack(const uint32_t *src, uint16_t *y, uint16_t *u, uint16_t *v, int width)
-cglobal v210_planar_unpack_%1, 5, 5, 7
+cglobal v210_planar_unpack_%1, 5, 5, 8
     movsxdifnidn r4, r4d
     lea    r1, [r1+2*r4]
     add    r2, r4
     add    r3, r4
     neg    r4
 
-    mova   m3, [v210_mult]
-    mova   m4, [v210_mask]
-    mova   m5, [v210_luma_shuf]
-    mova   m6, [v210_chroma_shuf]
+    VBROADCASTI128   m3, [v210_mult]
+    VBROADCASTI128   m5, [v210_chroma_shuf]
+
+%if cpuflag(avx2)
+    VBROADCASTI128   m4, [v210_luma_shuf_avx2]
+    VBROADCASTI128   m5, [v210_chroma_shuf_avx2]
+    mova             m6, [v210_luma_permute]
+    VBROADCASTI128   m7, [v210_chroma_shuf2]
+%else
+    VBROADCASTI128   m4, [v210_luma_shuf]
+    VBROADCASTI128   m5, [v210_chroma_shuf]
+%endif
+
 .loop:
 %ifidn %1, unaligned
-    movu   m0, [r0]
+    movu   m0, [r0]    ; yB v5 yA  u5 y9 v4  y8 u4 y7  v3 y6 u3  y5 v2 y4  u2 y3 v1  y2 u1 y1  v0 y0 u0
 %else
     mova   m0, [r0]
 %endif
 
     pmullw m1, m0, m3
-    psrld  m0, 10
-    psrlw  m1, 6  ; u0 v0 y1 y2 v1 u2 y4 y5
-    pand   m0, m4 ; y0 __ u1 __ y3 __ v2 __
+    pslld  m0, 12
+    psrlw  m1, 6                       ; yB yA u5 v4 y8 y7 v3 u3 y5 y4 u2 v1 y2 y1 v0 u0
+    psrld  m0, 22                      ; 00 v5 00 y9 00 u4 00 y6 00 v2 00 y3 00 u1 00 y0
+
+%if cpuflag(avx2)
+    vpblendd m2, m1, m0, 0x55          ; yB yA 00 y9 y8 y7 00 y6 y5 y4 00 y3 y2 y1 00 y0
+    pshufb m2, m4                      ; 00 00 yB yA y9 y8 y7 y6 00 00 y5 y4 y3 y2 y1 y0
+    vpermd m2, m6, m2                  ; 00 00 00 00 yB yA y9 y8 y7 y6 y5 y4 y3 y2 y1 y0
+    movu   [r1+2*r4], m2
 
-    shufps m2, m1, m0, 0x8d ; y1 y2 y4 y5 y0 __ y3 __
-    pshufb m2, m5 ; y0 y1 y2 y3 y4 y5 __ __
+    vpblendd m1, m0, 0xaa              ; 00 v5 u5 v4 00 u4 v3 u3 00 v2 u2 v1 00 u1 v0 u0
+    pshufb m1, m5                      ; 00 v5 v4 v3 00 u5 u4 u3 00 v2 v1 v0 00 u2 u1 u0
+    vpermq m1, m1, 0xd8                ; 00 v5 v4 v3 00 v2 v1 v0 00 u5 u4 u3 00 u2 u1 u0
+    pshufb m1, m7                      ; 00 00 v5 v4 v3 v2 v1 v0 00 00 u5 u4 u3 u2 u1 u0
+
+    movu   [r2+r4], xm1
+    vextracti128 [r3+r4], m1, 1
+%else
+    shufps m2, m1, m0, 0x8d            ; 00 y9 00 y6 yB yA y8 y7 00 y3 00 y0 y5 y4 y2 y1
+    pshufb m2, m4                      ; 00 00 yB yA y9 y8 y7 y6 00 00 y5 y4 y3 y2 y1 y0
     movu   [r1+2*r4], m2
 
-    shufps m1, m0, 0xd8 ; u0 v0 v1 u2 u1 __ v2 __
-    pshufb m1, m6 ; u0 u1 u2 __ v0 v1 v2 __
+    shufps m1, m0, 0xd8                ; 00 v5 00 u4 u5 v4 v3 u3 00 v2 00 u1 u2 v1 v0 u0
+    pshufb m1, m5                      ; 00 v5 v4 v3 00 u5 u4 u3 00 v2 v1 v0 00 u2 u1 u0
+
     movq   [r2+r4], m1
     movhps [r3+r4], m1
+%endif
 
     add r0, mmsize
-    add r4, 6
+    add r4, (mmsize*3)/8
     jl  .loop
 
     REP_RET
@@ -81,6 +111,11 @@ INIT_XMM avx
 v210_planar_unpack unaligned
 %endif
 
+%if HAVE_AVX2_EXTERNAL
+INIT_YMM avx2
+v210_planar_unpack unaligned
+%endif
+
 INIT_XMM ssse3
 v210_planar_unpack aligned
 
@@ -88,3 +123,8 @@ v210_planar_unpack aligned
 INIT_XMM avx
 v210_planar_unpack aligned
 %endif
+
+%if HAVE_AVX2_EXTERNAL
+INIT_YMM avx2
+v210_planar_unpack aligned
+%endif