x86/dsputilenc: implement SSE2 versions of pix_{sum16, norm1}

Signed-off-by: James Almer <jamrial@gmail.com>
Signed-off-by: Michael Niedermayer <michaelni@gmx.at>

3 files changed, 72 insertions, 40 deletions

diff --git a/libavcodec/x86/dsputilenc.asm b/libavcodec/x86/dsputilenc.asm
index 6269532a0e..0ff0e605b5 100644
--- a/libavcodec/x86/dsputilenc.asm
+++ b/libavcodec/x86/dsputilenc.asm
@@ -23,6 +23,10 @@
 
 %include "libavutil/x86/x86util.asm"
 
+SECTION_RODATA
+
+cextern pw_1
+
 SECTION .text
 
 %macro DIFF_PIXELS_1 4
@@ -439,73 +443,92 @@ cglobal diff_pixels, 4, 5, 5
     jne .loop
     RET
 
-INIT_MMX mmx
 ; int ff_pix_sum16_mmx(uint8_t *pix, int line_size)
-cglobal pix_sum16, 2, 3
+; %1 = number of xmm registers used
+; %2 = number of loops
+%macro PIX_SUM16 2
+cglobal pix_sum16, 2, 3, %1
     movsxdifnidn r1, r1d
-    mov          r2, r1
-    neg          r2
-    shl          r2, 4
-    sub          r0, r2
-    pxor         m7, m7
-    pxor         m6, m6
+    mov          r2, %2
+    pxor         m5, m5
+    pxor         m4, m4
 .loop:
-    mova         m0, [r0+r2+0]
-    mova         m1, [r0+r2+0]
-    mova         m2, [r0+r2+8]
-    mova         m3, [r0+r2+8]
-    punpcklbw    m0, m7
-    punpckhbw    m1, m7
-    punpcklbw    m2, m7
-    punpckhbw    m3, m7
+    mova         m0, [r0]
+%if mmsize == 8
+    mova         m1, [r0+8]
+%else
+    mova         m1, [r0+r1]
+%endif
+    punpckhbw    m2, m0, m5
+    punpcklbw    m0, m5
+    punpckhbw    m3, m1, m5
+    punpcklbw    m1, m5
     paddw        m1, m0
     paddw        m3, m2
     paddw        m3, m1
-    paddw        m6, m3
-    add          r2, r1
-    js .loop
-    mova         m5, m6
-    psrlq        m6, 32
-    paddw        m6, m5
-    mova         m5, m6
-    psrlq        m6, 16
-    paddw        m6, m5
-    movd        eax, m6
+    paddw        m4, m3
+%if mmsize == 8
+    add          r0, r1
+%else
+    lea          r0, [r0+r1*2]
+%endif
+    dec r2
+    jne .loop
+    HADDW        m4, m5
+    movd        eax, m4
     and         eax, 0xffff
     RET
+%endmacro
 
 INIT_MMX mmx
+PIX_SUM16 0, 16
+INIT_XMM sse2
+PIX_SUM16 6, 8
+
 ; int ff_pix_norm1_mmx(uint8_t *pix, int line_size)
-cglobal pix_norm1, 2, 4
+; %1 = number of xmm registers used
+; %2 = number of loops
+%macro PIX_NORM1 2
+cglobal pix_norm1, 2, 3, %1
     movsxdifnidn r1, r1d
-    mov          r2, 16
+    mov          r2, %2
     pxor         m0, m0
-    pxor         m7, m7
+    pxor         m5, m5
 .loop:
     mova         m2, [r0+0]
+%if mmsize == 8
     mova         m3, [r0+8]
-    mova         m1, m2
-    punpckhbw    m1, m0
+%else
+    mova         m3, [r0+r1]
+%endif
+    punpckhbw    m1, m2, m0
     punpcklbw    m2, m0
-    mova         m4, m3
-    punpckhbw    m3, m0
-    punpcklbw    m4, m0
+    punpckhbw    m4, m3, m0
+    punpcklbw    m3, m0
     pmaddwd      m1, m1
     pmaddwd      m2, m2
     pmaddwd      m3, m3
     pmaddwd      m4, m4
     paddd        m2, m1
     paddd        m4, m3
-    paddd        m7, m2
+    paddd        m5, m2
+    paddd        m5, m4
+%if mmsize == 8
     add          r0, r1
-    paddd        m7, m4
+%else
+    lea          r0, [r0+r1*2]
+%endif
     dec r2
     jne .loop
-    mova         m1, m7
-    psrlq        m7, 32
-    paddd        m1, m7
-    movd        eax, m1
+    HADDD        m5, m1
+    movd        eax, m5
     RET
+%endmacro
+
+INIT_MMX mmx
+PIX_NORM1 0, 16
+INIT_XMM sse2
+PIX_NORM1 6, 8
 
 ;-----------------------------------------------
 ;int ff_sum_abs_dctelem(int16_t *block)
diff --git a/libavcodec/x86/dsputilenc_mmx.c b/libavcodec/x86/dsputilenc_mmx.c
index 5aae14791b..efe835fac9 100644
--- a/libavcodec/x86/dsputilenc_mmx.c
+++ b/libavcodec/x86/dsputilenc_mmx.c
@@ -38,7 +38,9 @@ void ff_diff_pixels_mmx(int16_t *block, const uint8_t *s1, const uint8_t *s2,
 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_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);
 int ff_sum_abs_dctelem_mmxext(int16_t *block);
 int ff_sum_abs_dctelem_sse2(int16_t *block);
@@ -906,6 +908,8 @@ av_cold void ff_dsputilenc_init_mmx(DSPContext *c, AVCodecContext *avctx,
         c->sse[0] = ff_sse16_sse2;
         c->sum_abs_dctelem   = ff_sum_abs_dctelem_sse2;
         c->diff_pixels = ff_diff_pixels_sse2;
+        c->pix_sum     = ff_pix_sum16_sse2;
+        c->pix_norm1   = ff_pix_norm1_sse2;
 
 #if HAVE_ALIGNED_STACK
         c->hadamard8_diff[0] = ff_hadamard8_diff16_sse2;
diff --git a/libavutil/x86/x86util.asm b/libavutil/x86/x86util.asm
index 67d7905132..807e87e60e 100644
--- a/libavutil/x86/x86util.asm
+++ b/libavutil/x86/x86util.asm
@@ -288,7 +288,12 @@
     paddd   %1, %2
 %endif
 %if notcpuflag(xop) || sizeof%1 != 16
+%if cpuflag(mmxext)
     PSHUFLW %2, %1, q0032
+%else ; mmx
+    mova    %2, %1
+    psrlq   %2, 32
+%endif
     paddd   %1, %2
 %endif
 %undef %1