[x265] [PATCH] avx2: integral_init4h -> added 'INTEGRAL_FOUR_HORIZONTAL_4' macro to reduce data movement for '4' element case

[praveen at multicorewareinc.com]

# HG changeset patch
# User Praveen Tiwari <praveen at multicorewareinc.com>
# Date 1497422024 -19800
#      Wed Jun 14 12:03:44 2017 +0530
# Node ID 65e038ecbbf63a2f449ccf52358c5fbbec408b27
# Parent  28bbc349d17035a3c1fcdfbdca3b8e21ae6b669b
avx2: integral_init4h -> added 'INTEGRAL_FOUR_HORIZONTAL_4' macro to reduce data movement for '4' element case

diff -r 28bbc349d170 -r 65e038ecbbf6 source/common/x86/seaintegral.asm
--- a/source/common/x86/seaintegral.asm	Wed Jun 07 17:06:57 2017 +0530
+++ b/source/common/x86/seaintegral.asm	Wed Jun 14 12:03:44 2017 +0530
@@ -148,11 +148,6 @@
     jnz     .loop
     RET
 
-;-----------------------------------------------------------------------------
-;static void integral_init4h_c(uint32_t *sum, pixel *pix, intptr_t stride)
-;-----------------------------------------------------------------------------
-INIT_YMM avx2
-
 %macro INTEGRAL_FOUR_HORIZONTAL_16 0
     pmovzxbw       m0, [r1]
     pmovzxbw       m1, [r1 + 1]
@@ -163,6 +158,24 @@
     paddw          m0, m1
 %endmacro
 
+%macro INTEGRAL_FOUR_HORIZONTAL_4 0
+    movd       xm0, [r1]
+    movd       xm1, [r1 + 1]
+    pmovzxbw   xm0, xm0
+    pmovzxbw   xm1, xm1
+    paddw      xm0, xm1
+    movd       xm1, [r1 + 2]
+    pmovzxbw   xm1, xm1
+    paddw      xm0, xm1
+    movd       xm1, [r1 + 3]
+    pmovzxbw   xm1, xm1
+    paddw      xm0, xm1
+%endmacro
+
+;-----------------------------------------------------------------------------
+;static void integral_init4h(uint32_t *sum, pixel *pix, intptr_t stride)
+;-----------------------------------------------------------------------------
+INIT_YMM avx2
 cglobal integral4h, 3, 5, 3
     lea            r3, [4 * r2]
     sub            r0, r3
@@ -205,7 +218,7 @@
     jmp             .end
 
 .loop_4:
-    INTEGRAL_FOUR_HORIZONTAL_16
+    INTEGRAL_FOUR_HORIZONTAL_4
     pmovzxwd       xm0, xm0
     movu           xm1, [r0]
     paddd          xm0, xm1