[x265] [PATCH] avx2: 'integral4h' HIGH_BIT_DEPTH asm code -> 5.37x faster than 'C' version

[jayashri at multicorewareinc.com]

# HG changeset patch
# User Jayashri Murugan <jayashri at multicorewareinc.com>
# Date 1497421166 -19800
#      Wed Jun 14 11:49:26 2017 +0530
# Node ID 0e11a209ad118de136971184fcdf026fa275f9f2
# Parent  65e038ecbbf63a2f449ccf52358c5fbbec408b27
avx2: 'integral4h' HIGH_BIT_DEPTH asm code -> 5.37x faster than 'C' version

    integral_init4h  5.37x    457.68          2456.91

diff -r 65e038ecbbf6 -r 0e11a209ad11 source/common/x86/asm-primitives.cpp
--- a/source/common/x86/asm-primitives.cpp	Wed Jun 14 12:03:44 2017 +0530
+++ b/source/common/x86/asm-primitives.cpp	Wed Jun 14 11:49:26 2017 +0530
@@ -2164,6 +2164,7 @@
         p.integral_initv[INTEGRAL_16] = PFX(integral16v_avx2);
         p.integral_initv[INTEGRAL_24] = PFX(integral24v_avx2);
         p.integral_initv[INTEGRAL_32] = PFX(integral32v_avx2);
+        p.integral_inith[INTEGRAL_4] = PFX(integral4h_avx2);
 
         /* TODO: This kernel needs to be modified to work with HIGH_BIT_DEPTH only 
         p.planeClipAndMax = PFX(planeClipAndMax_avx2); */
diff -r 65e038ecbbf6 -r 0e11a209ad11 source/common/x86/seaintegral.asm
--- a/source/common/x86/seaintegral.asm	Wed Jun 14 12:03:44 2017 +0530
+++ b/source/common/x86/seaintegral.asm	Wed Jun 14 11:49:26 2017 +0530
@@ -172,10 +172,55 @@
     paddw      xm0, xm1
 %endmacro
 
+%macro INTEGRAL_FOUR_HORIZONTAL_8_HBD 0
+    pmovzxwd       m0, [r1]
+    pmovzxwd       m1, [r1 + 2]
+    paddd          m0, m1
+    pmovzxwd       m1, [r1 + 4]
+    paddd          m0, m1
+    pmovzxwd       m1, [r1 + 6]
+    paddd          m0, m1
+%endmacro
+
+%macro INTEGRAL_FOUR_HORIZONTAL_4_HBD 0
+    pmovzxwd       xm0, [r1]
+    pmovzxwd       xm1, [r1 + 2]
+    paddd          xm0, xm1
+    pmovzxwd       xm1, [r1 + 4]
+    paddd          xm0, xm1
+    pmovzxwd       xm1, [r1 + 6]
+    paddd          xm0, xm1
+%endmacro
+
 ;-----------------------------------------------------------------------------
 ;static void integral_init4h(uint32_t *sum, pixel *pix, intptr_t stride)
 ;-----------------------------------------------------------------------------
 INIT_YMM avx2
+%if HIGH_BIT_DEPTH
+cglobal integral4h, 3, 5, 3
+    lea            r3, [4 * r2]
+    sub            r0, r3
+    sub            r2, 4                      ;stride - 4
+    mov            r4, r2
+    shr            r4, 3
+
+.loop_8:
+    INTEGRAL_FOUR_HORIZONTAL_8_HBD
+    movu           m1, [r0]
+    paddd          m0, m1
+    movu           [r0 + r3], m0 
+    add            r1, 16
+    add            r0, 32
+    sub            r2, 8
+    sub            r4, 1
+    jnz            .loop_8
+    INTEGRAL_FOUR_HORIZONTAL_4_HBD
+    movu           xm1, [r0]
+    paddd          xm0, xm1
+    movu           [r0 + r3], xm0
+    RET
+
+%else
 cglobal integral4h, 3, 5, 3
     lea            r3, [4 * r2]
     sub            r0, r3
@@ -227,6 +272,7 @@
 
 .end
     RET
+%endif
 
 ;-----------------------------------------------------------------------------
 ;static void integral_init8h_c(uint32_t *sum, pixel *pix, intptr_t stride)