[x264-devel] x86: AVX2 high bit-depth predict_16x16_h

[Henrik Gramner]

x264 | branch: master | Henrik Gramner <henrik at gramner.com> | Tue Apr 16 23:27:04 2013 +0200| [989cbf04b0eb43fb8de2dd330edc733ab5399150] | committer: Jason Garrett-Glaser

x86: AVX2 high bit-depth predict_16x16_h

> http://git.videolan.org/gitweb.cgi/x264.git/?a=commit;h=989cbf04b0eb43fb8de2dd330edc733ab5399150
---

 common/common.h          |    2 +-
 common/x86/predict-a.asm |   43 ++++++++++---------------------------------
 common/x86/predict-c.c   |    3 +++
 common/x86/predict.h     |    1 +
 tools/checkasm.c         |    2 +-
 5 files changed, 16 insertions(+), 35 deletions(-)

diff --git a/common/common.h b/common/common.h
index ecd743b..53a6ff0 100644
--- a/common/common.h
+++ b/common/common.h
@@ -753,7 +753,7 @@ struct x264_t
 #define FENC_STRIDE 16
 #define FDEC_STRIDE 32
             ALIGNED_16( pixel fenc_buf[48*FENC_STRIDE] );
-            ALIGNED_16( pixel fdec_buf[52*FDEC_STRIDE] );
+            ALIGNED_N( pixel fdec_buf[52*FDEC_STRIDE] );
 
             /* i4x4 and i8x8 backup data, for skipping the encode stage when possible */
             ALIGNED_16( pixel i4x4_fdec_buf[16*16] );
diff --git a/common/x86/predict-a.asm b/common/x86/predict-a.asm
index 10f1d9f..f255ef1 100644
--- a/common/x86/predict-a.asm
+++ b/common/x86/predict-a.asm
@@ -2031,46 +2031,23 @@ cglobal predict_16x16_v, 1,1
 ;-----------------------------------------------------------------------------
 %macro PREDICT_16x16_H 0
 cglobal predict_16x16_h, 1,2
-    mov r1, 12*FDEC_STRIDEB
-%if HIGH_BIT_DEPTH
-.vloop:
-%assign Y 0
-%rep 4
-    movd        m0, [r0+r1+Y*FDEC_STRIDEB-2*SIZEOF_PIXEL]
-    SPLATW      m0, m0, 1
-    mova [r0+r1+Y*FDEC_STRIDEB+ 0], m0
-    mova [r0+r1+Y*FDEC_STRIDEB+16], m0
-%if mmsize==8
-    mova [r0+r1+Y*FDEC_STRIDEB+ 8], m0
-    mova [r0+r1+Y*FDEC_STRIDEB+24], m0
-%endif
-%assign Y Y+1
-%endrep
-
-%else ; !HIGH_BIT_DEPTH
-%if cpuflag(ssse3)
-    mova   m1, [pb_3]
-%endif
-.vloop:
-%assign Y 0
-%rep 4
-    SPLATB_LOAD m0, r0+r1+FDEC_STRIDE*Y-1, m1
-    mova [r0+r1+FDEC_STRIDE*Y], m0
-%if mmsize==8
-    mova [r0+r1+FDEC_STRIDE*Y+8], m0
+%if cpuflag(ssse3) && notcpuflag(avx2)
+    mova  m2, [pb_3]
 %endif
-%assign Y Y+1
-%endrep
-%endif ; HIGH_BIT_DEPTH
-    sub r1, 4*FDEC_STRIDEB
-    jge .vloop
+    mov  r1d, 4
+.loop:
+    PRED_H_4ROWS 16, 1
+    dec  r1d
+    jg .loop
     RET
 %endmacro
 
 INIT_MMX mmx2
 PREDICT_16x16_H
-INIT_XMM sse2
 %if HIGH_BIT_DEPTH
+INIT_XMM sse2
+PREDICT_16x16_H
+INIT_YMM avx2
 PREDICT_16x16_H
 %else
 ;no SSE2 for 8-bit, it's slower than MMX on all systems that don't support SSSE3
diff --git a/common/x86/predict-c.c b/common/x86/predict-c.c
index 2c31ded..871e8af 100644
--- a/common/x86/predict-c.c
+++ b/common/x86/predict-c.c
@@ -341,6 +341,9 @@ void x264_predict_16x16_init_mmx( int cpu, x264_predict_t pf[7] )
 #if HAVE_X86_INLINE_ASM
     pf[I_PRED_16x16_P]       = x264_predict_16x16_p_sse2;
 #endif
+    if( !(cpu&X264_CPU_AVX2) )
+        return;
+    pf[I_PRED_16x16_H]       = x264_predict_16x16_h_avx2;
 #else
 #if !ARCH_X86_64
     pf[I_PRED_16x16_P]       = x264_predict_16x16_p_mmx2;
diff --git a/common/x86/predict.h b/common/x86/predict.h
index d3e415d..8f10721 100644
--- a/common/x86/predict.h
+++ b/common/x86/predict.h
@@ -38,6 +38,7 @@ void x264_predict_16x16_v_sse ( pixel *src );
 void x264_predict_16x16_h_mmx2( pixel *src );
 void x264_predict_16x16_h_sse2( uint16_t *src );
 void x264_predict_16x16_h_ssse3( uint8_t *src );
+void x264_predict_16x16_h_avx2( uint16_t *src );
 void x264_predict_16x16_dc_mmx2( pixel *src );
 void x264_predict_16x16_dc_sse2( pixel *src );
 void x264_predict_16x16_dc_core_mmx2( pixel *src, int i_dc_left );
diff --git a/tools/checkasm.c b/tools/checkasm.c
index f871d58..33c2378 100644
--- a/tools/checkasm.c
+++ b/tools/checkasm.c
@@ -2147,7 +2147,7 @@ static int check_intra( int cpu_ref, int cpu_new )
     int ret = 0, ok = 1, used_asm = 0;
     ALIGNED_ARRAY_32( pixel, edge,[36] );
     ALIGNED_ARRAY_32( pixel, edge2,[36] );
-    ALIGNED_16( pixel fdec[FDEC_STRIDE*20] );
+    ALIGNED_ARRAY_32( pixel, fdec,[FDEC_STRIDE*20] );
     struct
     {
         x264_predict_t      predict_16x16[4+3];