<div style="line-height:1.7;color:#000000;font-size:14px;font-family:Arial"><pre>>diff -r a7c2f80c18af -r 973560d58dfb source/common/x86/intrapred8.asm
>--- a/source/common/x86/intrapred8.asm      Mon Nov 20 14:31:22 2017 +0530
>+++ b/source/common/x86/intrapred8.asm      Tue Nov 21 03:10:14 2017 +0800
>@@ -22313,11 +22313,144 @@
>     mov             [r1 + 64], r3b                  ; LeftLast
>     RET
> 
>-INIT_XMM sse4
>-cglobal intra_filter_32x32, 2,4,6
>-    mov             r2b, byte [r0 +  64]            ; topLast
>-    mov             r3b, byte [r0 + 128]            ; LeftLast
>-
>+; this function add strong intra filter
>+INIT_XMM sse4
>+cglobal intra_filter_32x32, 3,8,7
>+    xor             r3d, r3d             ; R9
>+    xor             r4d, r4d             ; R10
>+    mov             r3b, byte [r0 +  64] ; topLast
<div>>+    mov             r4b, byte [r0 + 128] ; LeftLast</div><div><br></div><div>xor+mov = movzx, the xor (clear to zero) does not spending cycle, but affect instruction decode rate</div><div><br></div>>+
>+    ; strong intra filter is diabled
>+    cmp             r2m, byte 0
>+    jz              .normal_filter32
>+    ; decide to do strong intra filter
>+    xor             r5d, r5d             ; R11
>+    xor             r6d, r6d             ; RAX
>+    xor             r7d, r7d             ; RDI
>+    mov             r5b, byte [r0]       ; topLeft
>+    mov             r6b, byte [r0 + 96]  ; leftMiddle
>+    mov             r7b, byte [r0 + 32]  ; topMiddle
>+
>+    ; threshold = 8
>+    mov             r2d, r3d             ; R8
>+    add             r2d, r5d             ; (topLast + topLeft)
>+    shl             r7d, 1               ; 2 * topMiddle
<div>>+    sub             r2d, r7d</div><div>(A+B) - 2 * C  <==> (A-C) + (B-C)</div><div><br></div>>+    mov             r7d, r2d             ; backup r2d
>+    sar             r7d, 31
>+    xor             r2d, r7d
>+    sub             r2d, r7d             ; abs(r2d)
<div>>+    cmp             r2d, 8</div><div>; how about this or instruction cdq?</div><div>; abs(x-y)</div><div>mov eax, X
sub eax, Y
sub Y, X
cmovg eax, Y</div><div><br></div><div><br></div>>+    ; bilinearAbove is false
>+    jns             .normal_filter32
>+
>+    mov             r2d, r5d
>+    add             r2d, r4d
>+    shl             r6d, 1
>+    sub             r2d, r6d
>+    mov             r6d, r2d
>+    sar             r6d, 31
>+    xor             r2d, r6d
>+    sub             r2d, r6d
>+    cmp             r2d, 8
>+    ; bilinearLeft is false
>+    jns             .normal_filter32
>+
>+    ; do strong intra filter shift = 6
>+    mov             r2d, r5d
>+    shl             r2d, 6
>+    add             r2d, 32              ; init
>+    mov             r6d, r4d
<div>>+    sub             r6w, r5w             ; deltaL size is word</div><div>partial register may stall in here</div><div><br></div>>+    mov             r7d, r3d
>+    sub             r7w, r5w             ; deltaR size is word
>+    movd            xmm0, r2d
<div>>+    vpbroadcastw    xmm0, xmm0</div><div>SSE4?</div><div><br></div>>+    mova            xmm4, xmm0
>+

</pre></div>