<div dir="ltr">Thanks for the improvements min. Modified patch below.<br><br># HG changeset patch<br># User Ramya Sriraman<<a href="mailto:ramya@multicorewareinc.com">ramya@multicorewareinc.com</a>><br># Date 1461158053 -19800<br>#      Wed Apr 20 18:44:13 2016 +0530<br># Node ID e21b86fb24567ad92be78eaadd4278fd34a7d161<br># Parent  4f83d465d11b3baa46e6089f73b0929266d4b722<br>arm: Implement quant<br><br>diff -r 4f83d465d11b -r e21b86fb2456 source/common/arm/asm-primitives.cpp<br>--- a/source/common/arm/asm-primitives.cpp    Wed Mar 30 17:29:13 2016 +0530<br>+++ b/source/common/arm/asm-primitives.cpp    Wed Apr 20 18:44:13 2016 +0530<br>@@ -820,6 +820,8 @@<br>         p.chroma[X265_CSP_I444].pu[LUMA_24x32].filter_vsp = PFX(interp_4tap_vert_sp_24x32_neon);<br>         p.chroma[X265_CSP_I444].pu[LUMA_48x64].filter_vsp = PFX(interp_4tap_vert_sp_48x64_neon);<br> <br>+        // quant<br>+        p.quant = PFX(quant_neon);<br>     }<br>     if (cpuMask & X265_CPU_ARMV6)<br>     {<br>diff -r 4f83d465d11b -r e21b86fb2456 source/common/arm/pixel-util.S<br>--- a/source/common/arm/pixel-util.S    Wed Mar 30 17:29:13 2016 +0530<br>+++ b/source/common/arm/pixel-util.S    Wed Apr 20 18:44:13 2016 +0530<br>@@ -1962,3 +1962,63 @@<br>     bx              lr<br> endfunc<br> <br>+function x265_quant_neon<br>+    push            {r4-r8}<br>+    ldr             r4, [sp, #4* 5]         //qbits<br>+    vdup.s32        q8, r4<br>+    ldr             r5, [sp, #4* 5 + 4]     // add<br>+    vdup.s32        q9, r5<br>+    ldr             r6, [sp, #4* 5 + 8]     // numcoeff<br>+    mov             r5, #8<br>+    sub             r5, r4, r5<br>+    vdup.s32        q10, r5<br>+    vneg.s32        q10, q10                // -(qbits- 8)<br>+    mov             r5, #1<br>+    vdup.u32        q11, r5<br>+<br>+    lsr             r6, r6 ,#2<br>+    eor             r7, r7<br>+<br>+.loop_quant:<br>+<br>+    vld1.s16        d0, [r0]!<br>+    vmovl.s16       q1, d0                  // coef[blockpos]<br>+<br>+    vclt.s32        q4, q1, #0<br>+    vorr.s32        q4, q4, q11             // q4= sign<br>+<br>+    vabs.s32        q1, q1                  // q1=level=abs(coef[blockpos])<br>+    vld1.s32        {q0}, [r1]!             // quantCoeff[blockpos]<br>+    vmul.i32        q0, q0, q1              // q0=tmplevel = abs(level) * quantCoeff[blockpos];<br>+<br>+    vadd.s32        q1, q0, q9              // q1= tmplevel+add<br>+    vneg.s32        q12, q8<br>+    vshl.s32        q1, q1, q12             // q1= level =(tmplevel+add) >> qbits<br>+<br>+    vshl.s32        q3, q1, q8              // q3 = level << qBits<br>+    vsub.s32        q13, q0, q3             // q8= tmplevel - (level << qBits)<br>+    vshl.s32        q13, q13, q10           // q3= ((tmplevel - (level << qBits)) >> qBits8)<br>+    vst1.s32        {q13}, [r2]!            // store deltaU<br>+<br>+    // numsig<br>+    vclz.s32        q2, q1<br>+    vshr.u32        q2, #5<br>+    vadd.u32        d4, d5<br>+    vpadd.u32       d4, d4<br>+    vmov.32         r12, d4[0]<br>+    mov             r8, #4<br>+    sub             r8, r8, r12<br>+    add             r7, r8<br>+<br>+    vmul.s32        q2, q1, q4<br>+    vqmovn.s32      d0, q2<br>+    vst1.s16        d0, [r3]!<br>+<br>+    subs            r6, #1<br>+    bne             .loop_quant<br>+<br>+    mov             r0, r7<br>+    pop             {r4-r8}<br>+    bx              lr<br>+endfunc<br>+<br>diff -r 4f83d465d11b -r e21b86fb2456 source/common/arm/pixel-util.h<br>--- a/source/common/arm/pixel-util.h    Wed Mar 30 17:29:13 2016 +0530<br>+++ b/source/common/arm/pixel-util.h    Wed Apr 20 18:44:13 2016 +0530<br>@@ -78,4 +78,6 @@<br> int x265_pixel_sa8d_32x32_neon(const pixel* pix1, intptr_t i_pix1, const pixel* pix2, intptr_t i_pix2);<br> int x265_pixel_sa8d_32x64_neon(const pixel* pix1, intptr_t i_pix1, const pixel* pix2, intptr_t i_pix2);<br> int x265_pixel_sa8d_64x64_neon(const pixel* pix1, intptr_t i_pix1, const pixel* pix2, intptr_t i_pix2);<br>+<br>+uint32_t x265_quant_neon(const int16_t* coef, const int32_t* quantCoeff, int32_t* deltaU, int16_t* qCoef, int qBits, int add, int numCoeff);<br> #endif // ifndef X265_PIXEL_UTIL_ARM_H<br><br></div><div class="gmail_extra"><br clear="all"><div><div class="gmail_signature"><div dir="ltr"><div><div dir="ltr"><div><div><span style="color:rgb(56,118,29)"><br></span></div><div><span style="color:rgb(56,118,29)">Thank you<br></span></div><span style="color:rgb(56,118,29)">Regards<br></span></div><span style="color:rgb(56,118,29)">Ramya</span><br></div></div></div></div></div>
<br><div class="gmail_quote">On Wed, Apr 20, 2016 at 10:05 PM, chen <span dir="ltr"><<a href="mailto:chenm003@163.com" target="_blank">chenm003@163.com</a>></span> wrote:<br><blockquote class="gmail_quote" style="margin:0 0 0 .8ex;border-left:1px #ccc solid;padding-left:1ex"><div style="line-height:1.7;color:#000000;font-size:14px;font-family:arial"><pre><div><div class="h5"><br>At 2016-04-20 21:24:03,<a href="mailto:ramya@multicorewareinc.com" target="_blank">ramya@multicorewareinc.com</a> wrote:
># HG changeset patch
># User Ramya Sriraman<<a href="mailto:ramya@multicorewareinc.com" target="_blank">ramya@multicorewareinc.com</a>>
># 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) * </div></div>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><br>_______________________________________________<br>
x265-devel mailing list<br>
<a href="mailto:x265-devel@videolan.org">x265-devel@videolan.org</a><br>
<a href="https://mailman.videolan.org/listinfo/x265-devel" rel="noreferrer" target="_blank">https://mailman.videolan.org/listinfo/x265-devel</a><br>
<br></blockquote></div><br></div>