<div style="line-height:1.7;color:#000000;font-size:14px;font-family:arial"><pre><br>At 2016-04-20 21:24:03,ramya@multicorewareinc.com wrote:
># HG changeset patch
># User Ramya Sriraman<ramya@multicorewareinc.com>
># Date 1461158053 -19800
>#      Wed Apr 20 18:44:13 2016 +0530
># Node ID 72ae446412d6e25aa3d2aa8ecb657f9815fdb635
># Parent  4f83d465d11b3baa46e6089f73b0929266d4b722
>arm: Implement quant
>
>diff -r 4f83d465d11b -r 72ae446412d6 source/common/arm/asm-primitives.cpp
>--- a/source/common/arm/asm-primitives.cpp     Wed Mar 30 17:29:13 2016 +0530
>+++ b/source/common/arm/asm-primitives.cpp     Wed Apr 20 18:44:13 2016 +0530
>@@ -820,6 +820,8 @@
>         p.chroma[X265_CSP_I444].pu[LUMA_24x32].filter_vsp = PFX(interp_4tap_vert_sp_24x32_neon);
>         p.chroma[X265_CSP_I444].pu[LUMA_48x64].filter_vsp = PFX(interp_4tap_vert_sp_48x64_neon);
> 
>+        // quant
>+        p.quant = PFX(quant_neon);
>     }
>     if (cpuMask & X265_CPU_ARMV6)
>     {
>diff -r 4f83d465d11b -r 72ae446412d6 source/common/arm/pixel-util.S
>--- a/source/common/arm/pixel-util.S   Wed Mar 30 17:29:13 2016 +0530
>+++ b/source/common/arm/pixel-util.S   Wed Apr 20 18:44:13 2016 +0530
>@@ -1962,3 +1962,63 @@
>     bx              lr
> endfunc
> 
>+function x265_quant_neon
>+    push            {r4-r9}
>+    ldr             r4, [sp, #4* 6]         //qbits
>+    ldr             r5, [sp, #4* 6 + 4]     // add
>+    ldr             r6, [sp, #4* 6 + 8]     // numcoeff
>+    mov             r7, #8
>+    sub             r7, r7 , r4             //-(qbits- 8)
>+
>+    lsr             r6, r6 ,#2
>+    mov             r8, #0
>+
>+.loop_quant:
>+
>+    vld1.s16        d0, [r0]!
>+    vmovl.s16       q1, d0                  // coef[blockpos]
>+
>+    vclt.s32        q4, q1, #0
>+    mov             r9, #1
>+    vdup.s32        q2, r9
>+    vorr.s32        q4, q4, q2              // q4= sign
>+
>+    vabs.s32        q1, q1                  // q1=level=abs(coef[blockpos])
>+    vld1.s32        {q0}, [r1]!             // quantCoeff[blockpos]
>+    vmul.i32        q0, q0, q1              // q0=tmplevel = abs(level) * quantCoeff[blockpos];
>+
>+    vdup.s32        q2, r5
r5=qbits, it is constant in the loop, the NEON have 16 of registers, so don't need load every iteration

>+    vadd.s32        q1, q0, q2             // q1= tmplevel+add
>+    vdup.s32        q2, r4
>+    vneg.s32        q2, q2
>+    vshl.s32        q1, q1, q2              // q1= level =tmplevel+add >> qbits
how about vqshrun?

>+
>+    vdup.s32        q2, r4
>+    vshl.s32        q3, q1, q2              // q3 = level << qBits
>+    vsub.s32        q8, q0, q3              // q8= tmplevel - (level << qBits)
>+    vdup.s32        q2, r7
>+    vshl.s32        q8, q8, q2              // q3= ((tmplevel - (level << qBits)) >> qBits8)
>+    vst1.s32        {q8}, [r2]!              // store deltaU
>+
>+    // numsig
>+    vclz.s32        q2, q1
>+    vshr.u32        q2, #5
>+    vadd.u32        d4, d5
>+    vpadd.u32       d4, d4
>+    vmov.32         r12, d4[0]
>+    mov             r9, #4
>+    sub             r9, r9, r12
>+    add             r8, r9
>+
>+    vmul.s32        q2, q1, q4
>+    vqmovn.s32      d0, q2
>+    vst1.s16        d0, [r3]!
>+
>+    subs            r6, #1
>+    bne             .loop_quant
>+
>+    mov             r0, r8
>+    pop             {r4-r9}
>+    bx              lr
>+endfunc

</pre></div>