[FFmpeg-devel] [PATCH] x86/dsputilenc: implement XOP version of pix_sum16

[James Almer]

SSE2: 137 cycles
XOP:   87 cycles
Signed-off-by: James Almer <jamrial at gmail.com>
---
The differences aren't as many as i originally thought after i realized 
that paddw can be used inside the loop.
The resulting macro is still a bit ugly, so commit whichever version you 
like the most.

 libavcodec/x86/dsputilenc.asm   | 29 ++++++++++++++++++++++++-----
 libavcodec/x86/dsputilenc_mmx.c |  5 +++++
 2 files changed, 29 insertions(+), 5 deletions(-)

diff --git a/libavcodec/x86/dsputilenc.asm b/libavcodec/x86/dsputilenc.asm
index 263516a..2cb70f0 100644
--- a/libavcodec/x86/dsputilenc.asm
+++ b/libavcodec/x86/dsputilenc.asm
@@ -446,13 +446,24 @@ cglobal diff_pixels, 4, 5, 5
 ; int ff_pix_sum16_mmx(uint8_t *pix, int line_size)
 ; %1 = number of xmm registers used
 ; %2 = number of loops
-%macro PIX_SUM16 2
-cglobal pix_sum16, 2, 3, %1
+; %3 = number of GPRs used
+%macro PIX_SUM16 4
+cglobal pix_sum16, 2, %3, %1
     movsxdifnidn r1, r1d
     mov          r2, %2
+%if cpuflag(xop)
+    lea          r3, [r1*3]
+%else
     pxor         m5, m5
+%endif
     pxor         m4, m4
 .loop:
+%if cpuflag(xop)
+    vphaddubq    m0, [r0]
+    vphaddubq    m1, [r0+r1]
+    vphaddubq    m2, [r0+r1*2]
+    vphaddubq    m3, [r0+r3]
+%else
     mova         m0, [r0]
 %if mmsize == 8
     mova         m1, [r0+8]
@@ -463,6 +474,7 @@ cglobal pix_sum16, 2, 3, %1
     punpcklbw    m0, m5
     punpckhbw    m3, m1, m5
     punpcklbw    m1, m5
+%endif ; cpuflag(xop)
     paddw        m1, m0
     paddw        m3, m2
     paddw        m3, m1
@@ -470,19 +482,26 @@ cglobal pix_sum16, 2, 3, %1
 %if mmsize == 8
     add          r0, r1
 %else
-    lea          r0, [r0+r1*2]
+    lea          r0, [r0+r1*%4]
 %endif
     dec r2
     jne .loop
+%if cpuflag(xop)
+    pshufd       m0, m4, q0032
+    paddd        m4, m0
+%else
     HADDW        m4, m5
+%endif
     movd        eax, m4
     RET
 %endmacro
 
 INIT_MMX mmx
-PIX_SUM16 0, 16
+PIX_SUM16 0, 16, 3, 0
 INIT_XMM sse2
-PIX_SUM16 6, 8
+PIX_SUM16 6, 8,  3, 2
+INIT_XMM xop
+PIX_SUM16 5, 4,  4, 4
 
 ; int ff_pix_norm1_mmx(uint8_t *pix, int line_size)
 ; %1 = number of xmm registers used
diff --git a/libavcodec/x86/dsputilenc_mmx.c b/libavcodec/x86/dsputilenc_mmx.c
index efe835f..4280f4b 100644
--- a/libavcodec/x86/dsputilenc_mmx.c
+++ b/libavcodec/x86/dsputilenc_mmx.c
@@ -39,6 +39,7 @@ void ff_diff_pixels_sse2(int16_t *block, const uint8_t *s1, const uint8_t *s2,
                          int stride);
 int ff_pix_sum16_mmx(uint8_t *pix, int line_size);
 int ff_pix_sum16_sse2(uint8_t *pix, int line_size);
+int ff_pix_sum16_xop(uint8_t *pix, int line_size);
 int ff_pix_norm1_mmx(uint8_t *pix, int line_size);
 int ff_pix_norm1_sse2(uint8_t *pix, int line_size);
 int ff_sum_abs_dctelem_mmx(int16_t *block);
@@ -925,5 +926,9 @@ av_cold void ff_dsputilenc_init_mmx(DSPContext *c, AVCodecContext *avctx,
 #endif
     }
 
+    if (EXTERNAL_XOP(cpu_flags)) {
+        c->pix_sum           = ff_pix_sum16_xop;
+    }
+
     ff_dsputil_init_pix_mmx(c, avctx);
 }
-- 
1.8.5.5