[x264-devel] x86: AVX-512 pixel_sa8d_8x8

[Henrik Gramner]

x264 | branch: master | Henrik Gramner <henrik at gramner.com> | Fri Apr 28 21:35:25 2017 +0200| [1cf7baa462ca52de7f07d6e4c795853900bb50bb] | committer: Henrik Gramner

x86: AVX-512 pixel_sa8d_8x8

> http://git.videolan.org/gitweb.cgi/x264.git/?a=commit;h=1cf7baa462ca52de7f07d6e4c795853900bb50bb
---

 common/pixel.c         |  1 +
 common/x86/pixel-a.asm | 34 +++++++++++++++++++++++++++++++++-
 common/x86/pixel.h     |  1 +
 3 files changed, 35 insertions(+), 1 deletion(-)

diff --git a/common/pixel.c b/common/pixel.c
index b68bb4c2..c33a873f 100644
--- a/common/pixel.c
+++ b/common/pixel.c
@@ -1355,6 +1355,7 @@ void x264_pixel_init( int cpu, x264_pixel_function_t *pixf )
     if( cpu&X264_CPU_AVX512 )
     {
         INIT8( satd, _avx512 );
+        pixf->sa8d[PIXEL_8x8] = x264_pixel_sa8d_8x8_avx512;
     }
 #endif //HAVE_MMX
 
diff --git a/common/x86/pixel-a.asm b/common/x86/pixel-a.asm
index 9c6ed6c8..9b3dc27b 100644
--- a/common/x86/pixel-a.asm
+++ b/common/x86/pixel-a.asm
@@ -4626,7 +4626,7 @@ cglobal intra_sad_x9_8x8, 5,7,8
     HMAXABSW2         0, 1, 2, 3
 %endmacro
 
-%macro SATD_AVX512_END 0
+%macro SATD_AVX512_END 0-1 0 ; sa8d
     paddw          m0 {k1}{z}, m1 ; zero-extend to dwords
 %if ARCH_X86_64
 %if mmsize == 64
@@ -4641,10 +4641,19 @@ cglobal intra_sad_x9_8x8, 5,7,8
     paddd        xmm0, xmm1
     movq          rax, xmm0
     rorx          rdx, rax, 32
+%if %1
+    lea           eax, [rax+rdx+1]
+    shr           eax, 1
+%else
     add           eax, edx
+%endif
 %else
     HADDD          m0, m1
     movd          eax, xm0
+%if %1
+    inc           eax
+    shr           eax, 1
+%endif
 %endif
     RET
 %endmacro
@@ -4789,6 +4798,29 @@ cglobal pixel_satd_4x4, 4,5
     SWAP      0, 1
     SATD_AVX512_END
 
+INIT_ZMM avx512
+cglobal pixel_sa8d_8x8, 4,6
+    vbroadcasti64x4 m4, [hmul_16p]
+    mov     r4d, 0x55555555
+    kmovd    k1, r4d   ; 01010101
+    kshiftlb k2, k1, 5 ; 10100000
+    kshiftlb k3, k1, 4 ; 01010000
+    lea      r4, [3*r1]
+    lea      r5, [3*r3]
+    SATD_AVX512_LOAD8 q, ym, k1, k2, k3 ; 2 0 2 0 6 4 6 4
+    DIFF_SUMSUB_SSSE3 0, 2, 1, 3, 4     ; 3 1 3 1 7 5 7 5
+    SUMSUB_BA      w, 0, 1, 2
+    SBUTTERFLY   qdq, 0, 1, 2
+    SUMSUB_BA      w, 0, 1, 2
+    shufps        m2, m0, m1, q2020
+    shufps        m1, m0, m1, q3131
+    SUMSUB_BA      w, 2, 1, 0
+    vshufi32x4    m0, m2, m1, q1010
+    vshufi32x4    m1, m2, m1, q3232
+    SUMSUB_BA      w, 0, 1, 2
+    HMAXABSW2      0, 1, 2, 3
+    SATD_AVX512_END 1
+
 %endif ; HIGH_BIT_DEPTH
 
 ;=============================================================================
diff --git a/common/x86/pixel.h b/common/x86/pixel.h
index 62b9fb42..d7753f53 100644
--- a/common/x86/pixel.h
+++ b/common/x86/pixel.h
@@ -84,6 +84,7 @@ DECL_X1( sa8d, sse4 )
 DECL_X1( sa8d, avx )
 DECL_X1( sa8d, xop )
 DECL_X1( sa8d, avx2 )
+DECL_X1( sa8d, avx512 )
 DECL_X1( sad, cache32_mmx2 );
 DECL_X1( sad, cache64_mmx2 );
 DECL_X1( sad, cache64_sse2 );