[x265] [PATCH] avx2: 'integral24h' asm code -> 1.40x faster than 'C' version

[jayashri at multicorewareinc.com]

# HG changeset patch
# User Jayashri Murugan <jayashri at multicorewareinc.com>
# Date 1496986500 -19800
#      Fri Jun 09 11:05:00 2017 +0530
# Node ID 0e351679c063149e0c9f7677a0c75c679918281b
# Parent  68a1e026fdd21c944c5e94a9da2de205ca63df6c
avx2: 'integral24h' asm code -> 1.40x faster than 'C' version

    integral_init24h  1.40x    1155.58         1614.98

diff -r 68a1e026fdd2 -r 0e351679c063 source/common/x86/asm-primitives.cpp
--- a/source/common/x86/asm-primitives.cpp	Fri Jun 09 10:53:23 2017 +0530
+++ b/source/common/x86/asm-primitives.cpp	Fri Jun 09 11:05:00 2017 +0530
@@ -3717,6 +3717,7 @@
         p.integral_inith[INTEGRAL_8] = PFX(integral8h_avx2);
         p.integral_inith[INTEGRAL_12] = PFX(integral12h_avx2);
         p.integral_inith[INTEGRAL_16] = PFX(integral16h_avx2);
+        p.integral_inith[INTEGRAL_24] = PFX(integral24h_avx2);
 
     }
 #endif
diff -r 68a1e026fdd2 -r 0e351679c063 source/common/x86/seaintegral.asm
--- a/source/common/x86/seaintegral.asm	Fri Jun 09 10:53:23 2017 +0530
+++ b/source/common/x86/seaintegral.asm	Fri Jun 09 11:05:00 2017 +0530
@@ -750,12 +750,146 @@
     RET
 %endif
 
+%macro INTEGRAL_TWENTYFOUR_HORIZONTAL_16 0
+    pmovzxbw       m0, [r1]
+    pmovzxbw       m1, [r1 + 1]
+    paddw          m0, m1
+    pmovzxbw       m1, [r1 + 2]
+    paddw          m0, m1
+    pmovzxbw       m1, [r1 + 3]
+    paddw          m0, m1
+    pmovzxbw       m1, [r1 + 4]
+    paddw          m0, m1
+    pmovzxbw       m1, [r1 + 5]
+    paddw          m0, m1
+    pmovzxbw       m1, [r1 + 6]
+    paddw          m0, m1
+    pmovzxbw       m1, [r1 + 7]
+    paddw          m0, m1
+    pmovzxbw       m1, [r1 + 8]
+    paddw          m0, m1
+    pmovzxbw       m1, [r1 + 9]
+    paddw          m0, m1
+    pmovzxbw       m1, [r1 + 10]
+    paddw          m0, m1
+    pmovzxbw       m1, [r1 + 11]
+    paddw          m0, m1
+    pmovzxbw       m1, [r1 + 12]
+    paddw          m0, m1
+    pmovzxbw       m1, [r1 + 13]
+    paddw          m0, m1
+    pmovzxbw       m1, [r1 + 14]
+    paddw          m0, m1
+    pmovzxbw       m1, [r1 + 15]
+    paddw          m0, m1
+    pmovzxbw       m1, [r1 + 16]
+    paddw          m0, m1
+    pmovzxbw       m1, [r1 + 17]
+    paddw          m0, m1
+    pmovzxbw       m1, [r1 + 18]
+    paddw          m0, m1
+    pmovzxbw       m1, [r1 + 19]
+    paddw          m0, m1
+    pmovzxbw       m1, [r1 + 20]
+    paddw          m0, m1
+    pmovzxbw       m1, [r1 + 21]
+    paddw          m0, m1
+    pmovzxbw       m1, [r1 + 22]
+    paddw          m0, m1
+    pmovzxbw       m1, [r1 + 23]
+    paddw          m0, m1
+%endmacro
+
+%macro INTEGRAL_TWENTYFOUR_HORIZONTAL_8 0
+    pmovzxbw       xm0, [r1]
+    pmovzxbw       xm1, [r1 + 1]
+    paddw          xm0, xm1
+    pmovzxbw       xm1, [r1 + 2]
+    paddw          xm0, xm1
+    pmovzxbw       xm1, [r1 + 3]
+    paddw          xm0, xm1
+    pmovzxbw       xm1, [r1 + 4]
+    paddw          xm0, xm1
+    pmovzxbw       xm1, [r1 + 5]
+    paddw          xm0, xm1
+    pmovzxbw       xm1, [r1 + 6]
+    paddw          xm0, xm1
+    pmovzxbw       xm1, [r1 + 7]
+    paddw          xm0, xm1
+    pmovzxbw       xm1, [r1 + 8]
+    paddw          xm0, xm1
+    pmovzxbw       xm1, [r1 + 9]
+    paddw          xm0, xm1
+    pmovzxbw       xm1, [r1 + 10]
+    paddw          xm0, xm1
+    pmovzxbw       xm1, [r1 + 11]
+    paddw          xm0, xm1
+    pmovzxbw       xm1, [r1 + 12]
+    paddw          xm0, xm1
+    pmovzxbw       xm1, [r1 + 13]
+    paddw          xm0, xm1
+    pmovzxbw       xm1, [r1 + 14]
+    paddw          xm0, xm1
+    pmovzxbw       xm1, [r1 + 15]
+    paddw          xm0, xm1
+    pmovzxbw       xm1, [r1 + 16]
+    paddw          xm0, xm1
+    pmovzxbw       xm1, [r1 + 17]
+    paddw          xm0, xm1
+    pmovzxbw       xm1, [r1 + 18]
+    paddw          xm0, xm1
+    pmovzxbw       xm1, [r1 + 19]
+    paddw          xm0, xm1
+    pmovzxbw       xm1, [r1 + 20]
+    paddw          xm0, xm1
+    pmovzxbw       xm1, [r1 + 21]
+    paddw          xm0, xm1
+    pmovzxbw       xm1, [r1 + 22]
+    paddw          xm0, xm1
+    pmovzxbw       xm1, [r1 + 23]
+    paddw          xm0, xm1
+%endmacro
+
 ;-----------------------------------------------------------------------------
 ;static void integral_init24h_c(uint32_t *sum, pixel *pix, intptr_t stride)
 ;-----------------------------------------------------------------------------
 INIT_YMM avx2
-cglobal integral24h, 3, 3, 0
- 
+cglobal integral24h, 3, 5, 3
+    lea            r3, [4 * r2]
+    sub            r0, r3
+    sub            r2, 24                      ;stride - 24
+    mov            r4, r2
+    shr            r4, 4
+
+.loop_16:
+    INTEGRAL_TWENTYFOUR_HORIZONTAL_16
+    vperm2i128     m2, m0, m0, 1
+    pmovzxwd       m2, xm2
+    pmovzxwd       m0, xm0
+    movu           m1, [r0]
+    paddd          m0, m1
+    movu           [r0 + r3], m0
+    movu           m1, [r0 + 32]
+    paddd          m2, m1
+    movu           [r0 + r3 + 32], m2
+    add            r1, 16
+    add            r0, 64
+    sub            r2, 16
+    sub            r4, 1
+    jnz            .loop_16
+    cmp            r2, 8
+    je             .loop_8
+    jmp             .end
+
+.loop_8:
+    INTEGRAL_TWENTYFOUR_HORIZONTAL_8
+    pmovzxwd       m0, xm0
+    movu           m1, [r0]
+    paddd          m0, m1
+    movu           [r0 + r3], m0
+    jmp             .end
+
+.end
     RET
 
 ;-----------------------------------------------------------------------------