[x264-devel] AVX 32-bit hpel_filter_h

[Jason Garrett-Glaser]

x264 | branch: master | Jason Garrett-Glaser <jason at x264.com> | Mon Jan 23 15:09:38 2012 -0800| [14dc11f7c52fa29576e0003c8c16857a78bf5fbf] | committer: Jason Garrett-Glaser

AVX 32-bit hpel_filter_h
Faster on Sandy Bridge.
Also add details on unsuccessful optimizations in these functions.

> http://git.videolan.org/gitweb.cgi/x264.git/?a=commit;h=14dc11f7c52fa29576e0003c8c16857a78bf5fbf
---

 common/x86/dct-64.asm |    2 +-
 common/x86/mc-a2.asm  |   12 +++++++++---
 common/x86/mc-c.c     |    2 +-
 3 files changed, 11 insertions(+), 5 deletions(-)

diff --git a/common/x86/dct-64.asm b/common/x86/dct-64.asm
index 78a2484..2a2c386 100644
--- a/common/x86/dct-64.asm
+++ b/common/x86/dct-64.asm
@@ -89,7 +89,7 @@ cextern hsub_mul
 
 %macro IDCT8_1D 11
     SUMSUB_BA %1, %6, %2, %10 ; %5=a0, %1=a2
-    
+
     psra%1   m%10, m%3, 1
     padd%1   m%10, m%3
     padd%1   m%10, m%5
diff --git a/common/x86/mc-a2.asm b/common/x86/mc-a2.asm
index 1b8788d..dd2c686 100644
--- a/common/x86/mc-a2.asm
+++ b/common/x86/mc-a2.asm
@@ -461,6 +461,7 @@ cglobal hpel_filter_c, 3,3,9
 %else
     %define tpw_32 [pw_32]
 %endif
+; This doesn't seem to be faster (with AVX) on Sandy Bridge or Bulldozer...
 %if cpuflag(misalign)
 .loop:
     movu    m4, [src-4]
@@ -559,11 +560,11 @@ cglobal hpel_filter_h_sse2, 3,3,8
     jl .loop
     REP_RET
 
-%if ARCH_X86_64 == 0
 ;-----------------------------------------------------------------------------
 ; void hpel_filter_h( uint8_t *dst, uint8_t *src, int width );
 ;-----------------------------------------------------------------------------
-cglobal hpel_filter_h_ssse3, 3,3
+%macro HPEL_H 0
+cglobal hpel_filter_h, 3,3
     add r0, r2
     add r1, r2
     neg r2
@@ -573,6 +574,9 @@ cglobal hpel_filter_h_ssse3, 3,3
     mova      m7, [pw_16]
 .loop:
     mova      m2, [src+16]
+    ; Using unaligned loads instead of palignr is marginally slower on SB and significantly
+    ; slower on Bulldozer, despite their fast load units -- even though it would let us avoid
+    ; the repeated loads of constants for pmaddubsw.
     palignr   m3, m1, m0, 14
     palignr   m4, m1, m0, 15
     palignr   m0, m2, m1, 2
@@ -596,7 +600,7 @@ cglobal hpel_filter_h_ssse3, 3,3
     add r2, 16
     jl .loop
     REP_RET
-%endif ; !ARCH_X86_64
+%endmacro
 
 INIT_MMX mmx2
 HPEL_V 0
@@ -610,9 +614,11 @@ HPEL_C
 INIT_XMM ssse3
 HPEL_C
 HPEL_V 0
+HPEL_H
 INIT_XMM avx
 HPEL_C
 HPEL_V 0
+HPEL_H
 %endif
 
 %if ARCH_X86_64
diff --git a/common/x86/mc-c.c b/common/x86/mc-c.c
index 8d3368f..dbe7d34 100644
--- a/common/x86/mc-c.c
+++ b/common/x86/mc-c.c
@@ -450,7 +450,7 @@ void x264_hpel_filter_avx( uint8_t *dsth, uint8_t *dstv, uint8_t *dstc, uint8_t
 #else
 HPEL(16, sse2, sse2, sse2, sse2)
 HPEL(16, ssse3, ssse3, ssse3, ssse3)
-HPEL(16, avx, avx, avx, ssse3)
+HPEL(16, avx, avx, avx, avx)
 #endif
 HPEL(16, sse2_misalign, sse2, sse2_misalign, sse2)
 #endif // HIGH_BIT_DEPTH