[x265] [PATCH 2 of 3] asm: optimize nquant by PSIGND, improve 13k cycles -> 11k cycles

[Min Chen]

# HG changeset patch
# User Min Chen <chenm003 at 163.com>
# Date 1409787419 25200
# Node ID 4ca9e972f48cb4530ca7181ad7cec351568a99b3
# Parent  94bd00d1af5d8c5f6f26f97c50a727588a860714
asm: optimize nquant by PSIGND, improve 13k cycles -> 11k cycles

diff -r 94bd00d1af5d -r 4ca9e972f48c source/common/dct.cpp
--- a/source/common/dct.cpp	Wed Sep 03 16:36:44 2014 -0700
+++ b/source/common/dct.cpp	Wed Sep 03 16:36:59 2014 -0700
@@ -801,6 +801,10 @@
 {
     uint32_t numSig = 0;
 
+    X265_CHECK((numCoeff % 16) == 0, "number of quant coeff is not multiple of 4x4\n");
+    X265_CHECK((uint32_t)add < ((uint32_t)1 << qBits), "2 ^ qBits less than add\n");
+    X265_CHECK(((intptr_t)quantCoeff & 15) == 0, "quantCoeff buffer not aligned\n");
+
     for (int blockpos = 0; blockpos < numCoeff; blockpos++)
     {
         int level = coef[blockpos];
diff -r 94bd00d1af5d -r 4ca9e972f48c source/common/x86/pixel-util8.asm
--- a/source/common/x86/pixel-util8.asm	Wed Sep 03 16:36:44 2014 -0700
+++ b/source/common/x86/pixel-util8.asm	Wed Sep 03 16:36:59 2014 -0700
@@ -941,55 +941,47 @@
 ; uint32_t nquant(int32_t *coef, int32_t *quantCoeff, int32_t *qCoef, int qBits, int add, int numCoeff);
 ;-----------------------------------------------------------------------------
 INIT_XMM sse4
-cglobal nquant, 4,5,8
+cglobal nquant, 3,5,8
     movd        m6, r4m
     mov         r4d, r5m
     pxor        m7, m7          ; m7 = numZero
-    movd        m5, r3d         ; m5 = qbits
+    movd        m5, r3m         ; m5 = qbits
     pshufd      m6, m6, 0       ; m6 = add
     mov         r3d, r4d        ; r3 = numCoeff
     shr         r4d, 3
+
 .loop:
     movu        m0, [r0]        ; m0 = level
     movu        m1, [r0 + 16]   ; m1 = level
-    movu        m2, [r1]        ; m2 = qcoeff
-    movu        m3, [r1 + 16]   ; m3 = qcoeff
+
+    pabsd       m2, m0
+    pmulld      m2, [r1]        ; m4 = tmpLevel1
+    paddd       m2, m6
+    psrad       m2, m5          ; m4 = level1
+    psignd      m2, m0          ; restore sign
+
+    pabsd       m3, m1
+    pmulld      m3, [r1 + 16]   ; m4 = tmpLevel1
+    paddd       m3, m6
+    psrad       m3, m5          ; m4 = level1
+    psignd      m3, m1          ; restore sign
     add         r0, 32
     add         r1, 32
 
-    pxor        m4, m4
-    pcmpgtd     m4, m0          ; m4 = sign
-    pabsd       m0, m0
-    pmulld      m0, m2          ; m0 = tmpLevel1
-    paddd       m0, m6
-    psrad       m0, m5          ; m0 = level1
-    pxor        m0, m4
-    psubd       m0, m4
-
-    pxor        m4, m4
-    pcmpgtd     m4, m1          ; m4 = sign
-    pabsd       m1, m1
-    pmulld      m1, m3          ; m1 = tmpLevel1
-    paddd       m1, m6
-    psrad       m1, m5          ; m1 = level1
-    pxor        m1, m4
-    psubd       m1, m4
-
-    packssdw    m0, m0
-    packssdw    m1, m1
-    pmovsxwd    m0, m0
+    packssdw    m2, m3
+    pmovsxwd    m0, m2
+    movhlps     m1, m2
     pmovsxwd    m1, m1
 
-    movu        [r2], m0
+    movu        [r2     ], m0
     movu        [r2 + 16], m1
     add         r2, 32
+
+    pxor        m4, m4
+    pcmpeqw     m2, m4
+    psubw       m7, m2
+
     dec         r4d
-
-    packssdw    m0, m1
-    pxor        m4, m4
-    pcmpeqw     m0, m4
-    psubw       m7, m0
-
     jnz         .loop
 
     packuswb    m7, m7
@@ -997,10 +989,8 @@
     mov         eax, r3d
     movd        r4d, m7
     sub         eax, r4d        ; numSig
-
     RET
 
-
 ;-----------------------------------------------------------------------------
 ; void dequant_normal(const int32_t* quantCoef, int32_t* coef, int num, int scale, int shift)
 ;-----------------------------------------------------------------------------