yadif sse2/ssse3 optimizations

Originally committed as revision 25874 to svn://svn.ffmpeg.org/ffmpeg/trunk

libavfilter/vf_yadif.c

@@ -304,7 +304,11 @@ static av_cold int init(AVFilterContext *ctx, const char *args, void *opaque)
     if (args) sscanf(args, "%d:%d", &yadif->mode, &yadif->parity);
 
     yadif->filter_line = filter_line_c;
-    if (HAVE_MMX && cpu_flags & AV_CPU_FLAG_MMX)
+    if (HAVE_SSSE3 && cpu_flags & AV_CPU_FLAG_SSSE3)
+        yadif->filter_line = ff_yadif_filter_line_ssse3;
+    else if (HAVE_SSE && cpu_flags & AV_CPU_FLAG_SSE2)
+        yadif->filter_line = ff_yadif_filter_line_sse2;
+    else if (HAVE_MMX && cpu_flags & AV_CPU_FLAG_MMX)
         yadif->filter_line = ff_yadif_filter_line_mmx;
 
     av_log(ctx, AV_LOG_INFO, "mode:%d parity:%d\n", yadif->mode, yadif->parity);

libavfilter/x86/yadif.c

@@ -22,214 +22,24 @@
 #include "libavutil/x86_cpu.h"
 #include "libavfilter/yadif.h"
 
-#define LOAD4(mem,dst) \
-            "movd      "mem", "#dst" \n\t"\
-            "punpcklbw %%mm7, "#dst" \n\t"
-
-#define PABS(tmp,dst) \
-            "pxor     "#tmp", "#tmp" \n\t"\
-            "psubw    "#dst", "#tmp" \n\t"\
-            "pmaxsw   "#tmp", "#dst" \n\t"
-
-#define CHECK(pj,mj) \
-            "movq "#pj"(%[cur],%[mrefs]), %%mm2 \n\t" /* cur[x-refs-1+j] */\
-            "movq "#mj"(%[cur],%[prefs]), %%mm3 \n\t" /* cur[x+refs-1-j] */\
-            "movq      %%mm2, %%mm4 \n\t"\
-            "movq      %%mm2, %%mm5 \n\t"\
-            "pxor      %%mm3, %%mm4 \n\t"\
-            "pavgb     %%mm3, %%mm5 \n\t"\
-            "pand     "MANGLE(pb_1)", %%mm4 \n\t"\
-            "psubusb   %%mm4, %%mm5 \n\t"\
-            "psrlq     $8,    %%mm5 \n\t"\
-            "punpcklbw %%mm7, %%mm5 \n\t" /* (cur[x-refs+j] + cur[x+refs-j])>>1 */\
-            "movq      %%mm2, %%mm4 \n\t"\
-            "psubusb   %%mm3, %%mm2 \n\t"\
-            "psubusb   %%mm4, %%mm3 \n\t"\
-            "pmaxub    %%mm3, %%mm2 \n\t"\
-            "movq      %%mm2, %%mm3 \n\t"\
-            "movq      %%mm2, %%mm4 \n\t" /* ABS(cur[x-refs-1+j] - cur[x+refs-1-j]) */\
-            "psrlq      $8,   %%mm3 \n\t" /* ABS(cur[x-refs  +j] - cur[x+refs  -j]) */\
-            "psrlq     $16,   %%mm4 \n\t" /* ABS(cur[x-refs+1+j] - cur[x+refs+1-j]) */\
-            "punpcklbw %%mm7, %%mm2 \n\t"\
-            "punpcklbw %%mm7, %%mm3 \n\t"\
-            "punpcklbw %%mm7, %%mm4 \n\t"\
-            "paddw     %%mm3, %%mm2 \n\t"\
-            "paddw     %%mm4, %%mm2 \n\t" /* score */
-
-#define CHECK1 \
-            "movq      %%mm0, %%mm3 \n\t"\
-            "pcmpgtw   %%mm2, %%mm3 \n\t" /* if(score < spatial_score) */\
-            "pminsw    %%mm2, %%mm0 \n\t" /* spatial_score= score; */\
-            "movq      %%mm3, %%mm6 \n\t"\
-            "pand      %%mm3, %%mm5 \n\t"\
-            "pandn     %%mm1, %%mm3 \n\t"\
-            "por       %%mm5, %%mm3 \n\t"\
-            "movq      %%mm3, %%mm1 \n\t" /* spatial_pred= (cur[x-refs+j] + cur[x+refs-j])>>1; */
-
-#define CHECK2 /* pretend not to have checked dir=2 if dir=1 was bad.\
-                  hurts both quality and speed, but matches the C version. */\
-            "paddw    "MANGLE(pw_1)", %%mm6 \n\t"\
-            "psllw     $14,   %%mm6 \n\t"\
-            "paddsw    %%mm6, %%mm2 \n\t"\
-            "movq      %%mm0, %%mm3 \n\t"\
-            "pcmpgtw   %%mm2, %%mm3 \n\t"\
-            "pminsw    %%mm2, %%mm0 \n\t"\
-            "pand      %%mm3, %%mm5 \n\t"\
-            "pandn     %%mm1, %%mm3 \n\t"\
-            "por       %%mm5, %%mm3 \n\t"\
-            "movq      %%mm3, %%mm1 \n\t"
-
-DECLARE_ASM_CONST(16, uint64_t, pw_1) = 0x0001000100010001ULL;
-DECLARE_ASM_CONST(16, uint64_t, pb_1) = 0x0101010101010101ULL;
-
-void ff_yadif_filter_line_mmx(uint8_t *dst,
-                              uint8_t *prev, uint8_t *cur, uint8_t *next,
-                              int w, int refs, int parity, int mode)
-{
-    uint64_t tmp0, tmp1, tmp2, tmp3;
-    int x;
-
-#define FILTER\
-    for(x=0; x<w; x+=4){\
-        __asm__ volatile(\
-            "pxor      %%mm7, %%mm7 \n\t"\
-            LOAD4("(%[cur],%[mrefs])", %%mm0) /* c = cur[x-refs] */\
-            LOAD4("(%[cur],%[prefs])", %%mm1) /* e = cur[x+refs] */\
-            LOAD4("(%["prev2"])", %%mm2) /* prev2[x] */\
-            LOAD4("(%["next2"])", %%mm3) /* next2[x] */\
-            "movq      %%mm3, %%mm4 \n\t"\
-            "paddw     %%mm2, %%mm3 \n\t"\
-            "psraw     $1,    %%mm3 \n\t" /* d = (prev2[x] + next2[x])>>1 */\
-            "movq      %%mm0, %[tmp0] \n\t" /* c */\
-            "movq      %%mm3, %[tmp1] \n\t" /* d */\
-            "movq      %%mm1, %[tmp2] \n\t" /* e */\
-            "psubw     %%mm4, %%mm2 \n\t"\
-            PABS(      %%mm4, %%mm2) /* temporal_diff0 */\
-            LOAD4("(%[prev],%[mrefs])", %%mm3) /* prev[x-refs] */\
-            LOAD4("(%[prev],%[prefs])", %%mm4) /* prev[x+refs] */\
-            "psubw     %%mm0, %%mm3 \n\t"\
-            "psubw     %%mm1, %%mm4 \n\t"\
-            PABS(      %%mm5, %%mm3)\
-            PABS(      %%mm5, %%mm4)\
-            "paddw     %%mm4, %%mm3 \n\t" /* temporal_diff1 */\
-            "psrlw     $1,    %%mm2 \n\t"\
-            "psrlw     $1,    %%mm3 \n\t"\
-            "pmaxsw    %%mm3, %%mm2 \n\t"\
-            LOAD4("(%[next],%[mrefs])", %%mm3) /* next[x-refs] */\
-            LOAD4("(%[next],%[prefs])", %%mm4) /* next[x+refs] */\
-            "psubw     %%mm0, %%mm3 \n\t"\
-            "psubw     %%mm1, %%mm4 \n\t"\
-            PABS(      %%mm5, %%mm3)\
-            PABS(      %%mm5, %%mm4)\
-            "paddw     %%mm4, %%mm3 \n\t" /* temporal_diff2 */\
-            "psrlw     $1,    %%mm3 \n\t"\
-            "pmaxsw    %%mm3, %%mm2 \n\t"\
-            "movq      %%mm2, %[tmp3] \n\t" /* diff */\
-\
-            "paddw     %%mm0, %%mm1 \n\t"\
-            "paddw     %%mm0, %%mm0 \n\t"\
-            "psubw     %%mm1, %%mm0 \n\t"\
-            "psrlw     $1,    %%mm1 \n\t" /* spatial_pred */\
-            PABS(      %%mm2, %%mm0)      /* ABS(c-e) */\
-\
-            "movq -1(%[cur],%[mrefs]), %%mm2 \n\t" /* cur[x-refs-1] */\
-            "movq -1(%[cur],%[prefs]), %%mm3 \n\t" /* cur[x+refs-1] */\
-            "movq      %%mm2, %%mm4 \n\t"\
-            "psubusb   %%mm3, %%mm2 \n\t"\
-            "psubusb   %%mm4, %%mm3 \n\t"\
-            "pmaxub    %%mm3, %%mm2 \n\t"\
-            "pshufw $9,%%mm2, %%mm3 \n\t"\
-            "punpcklbw %%mm7, %%mm2 \n\t" /* ABS(cur[x-refs-1] - cur[x+refs-1]) */\
-            "punpcklbw %%mm7, %%mm3 \n\t" /* ABS(cur[x-refs+1] - cur[x+refs+1]) */\
-            "paddw     %%mm2, %%mm0 \n\t"\
-            "paddw     %%mm3, %%mm0 \n\t"\
-            "psubw    "MANGLE(pw_1)", %%mm0 \n\t" /* spatial_score */\
-\
-            CHECK(-2,0)\
-            CHECK1\
-            CHECK(-3,1)\
-            CHECK2\
-            CHECK(0,-2)\
-            CHECK1\
-            CHECK(1,-3)\
-            CHECK2\
-\
-            /* if(p->mode<2) ... */\
-            "movq    %[tmp3], %%mm6 \n\t" /* diff */\
-            "cmpl      $2, %[mode] \n\t"\
-            "jge       1f \n\t"\
-            LOAD4("(%["prev2"],%[mrefs],2)", %%mm2) /* prev2[x-2*refs] */\
-            LOAD4("(%["next2"],%[mrefs],2)", %%mm4) /* next2[x-2*refs] */\
-            LOAD4("(%["prev2"],%[prefs],2)", %%mm3) /* prev2[x+2*refs] */\
-            LOAD4("(%["next2"],%[prefs],2)", %%mm5) /* next2[x+2*refs] */\
-            "paddw     %%mm4, %%mm2 \n\t"\
-            "paddw     %%mm5, %%mm3 \n\t"\
-            "psrlw     $1,    %%mm2 \n\t" /* b */\
-            "psrlw     $1,    %%mm3 \n\t" /* f */\
-            "movq    %[tmp0], %%mm4 \n\t" /* c */\
-            "movq    %[tmp1], %%mm5 \n\t" /* d */\
-            "movq    %[tmp2], %%mm7 \n\t" /* e */\
-            "psubw     %%mm4, %%mm2 \n\t" /* b-c */\
-            "psubw     %%mm7, %%mm3 \n\t" /* f-e */\
-            "movq      %%mm5, %%mm0 \n\t"\
-            "psubw     %%mm4, %%mm5 \n\t" /* d-c */\
-            "psubw     %%mm7, %%mm0 \n\t" /* d-e */\
-            "movq      %%mm2, %%mm4 \n\t"\
-            "pminsw    %%mm3, %%mm2 \n\t"\
-            "pmaxsw    %%mm4, %%mm3 \n\t"\
-            "pmaxsw    %%mm5, %%mm2 \n\t"\
-            "pminsw    %%mm5, %%mm3 \n\t"\
-            "pmaxsw    %%mm0, %%mm2 \n\t" /* max */\
-            "pminsw    %%mm0, %%mm3 \n\t" /* min */\
-            "pxor      %%mm4, %%mm4 \n\t"\
-            "pmaxsw    %%mm3, %%mm6 \n\t"\
-            "psubw     %%mm2, %%mm4 \n\t" /* -max */\
-            "pmaxsw    %%mm4, %%mm6 \n\t" /* diff= MAX3(diff, min, -max); */\
-            "1: \n\t"\
-\
-            "movq    %[tmp1], %%mm2 \n\t" /* d */\
-            "movq      %%mm2, %%mm3 \n\t"\
-            "psubw     %%mm6, %%mm2 \n\t" /* d-diff */\
-            "paddw     %%mm6, %%mm3 \n\t" /* d+diff */\
-            "pmaxsw    %%mm2, %%mm1 \n\t"\
-            "pminsw    %%mm3, %%mm1 \n\t" /* d = clip(spatial_pred, d-diff, d+diff); */\
-            "packuswb  %%mm1, %%mm1 \n\t"\
-\
-            :[tmp0]"=m"(tmp0),\
-             [tmp1]"=m"(tmp1),\
-             [tmp2]"=m"(tmp2),\
-             [tmp3]"=m"(tmp3)\
-            :[prev] "r"(prev),\
-             [cur]  "r"(cur),\
-             [next] "r"(next),\
-             [prefs]"r"((x86_reg)refs),\
-             [mrefs]"r"((x86_reg)-refs),\
-             [mode] "g"(mode)\
-        );\
-        __asm__ volatile("movd %%mm1, %0" :"=m"(*dst));\
-        dst += 4;\
-        prev+= 4;\
-        cur += 4;\
-        next+= 4;\
-    }
-
-    if (parity) {
-#define prev2 "prev"
-#define next2 "cur"
-        FILTER
-#undef prev2
-#undef next2
-    } else {
-#define prev2 "cur"
-#define next2 "next"
-        FILTER
-#undef prev2
-#undef next2
-    }
-}
-#undef LOAD4
-#undef PABS
-#undef CHECK
-#undef CHECK1
-#undef CHECK2
-#undef FILTER
+#if HAVE_SSSE3
+#define COMPILE_TEMPLATE_SSE 1
+#define COMPILE_TEMPLATE_SSSE3 1
+#undef RENAME
+#define RENAME(a) a ## _ssse3
+#include "yadif_template.c"
+#undef COMPILE_TEMPLATE_SSSE3
+#endif
+
+#if HAVE_SSE
+#undef RENAME
+#define RENAME(a) a ## _sse2
+#include "yadif_template.c"
+#undef COMPILE_TEMPLATE_SSE
+#endif
+
+#if HAVE_MMX
+#undef RENAME
+#define RENAME(a) a ## _mmx
+#include "yadif_template.c"
+#endif

libavfilter/x86/yadif_template.c

@@ -0,0 +1,268 @@
+/*
+ * Copyright (C) 2006 Michael Niedermayer <michaelni@gmx.at>
+ *
+ * This file is part of FFmpeg.
+ *
+ * FFmpeg is free software; you can redistribute it and/or modify
+ * it under the terms of the GNU General Public License as published by
+ * the Free Software Foundation; either version 2 of the License, or
+ * (at your option) any later version.
+ *
+ * FFmpeg is distributed in the hope that it will be useful,
+ * but WITHOUT ANY WARRANTY; without even the implied warranty of
+ * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+ * GNU General Public License for more details.
+ *
+ * You should have received a copy of the GNU General Public License along
+ * with FFmpeg; if not, write to the Free Software Foundation, Inc.,
+ * 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
+ */
+
+#ifdef COMPILE_TEMPLATE_SSE
+#define MM "%%xmm"
+#define MOV  "movq"
+#define MOVQ "movdqa"
+#define MOVQU "movdqu"
+#define STEP 8
+#define LOAD(mem,dst) \
+            MOV"       "mem", "dst" \n\t"\
+            "punpcklbw "MM"7, "dst" \n\t"
+#define PSRL1(reg) "psrldq $1, "reg" \n\t"
+#define PSRL2(reg) "psrldq $2, "reg" \n\t"
+#define PSHUF(src,dst) "movdqa "dst", "src" \n\t"\
+                       "psrldq $2, "src"     \n\t"
+#else
+#define MM "%%mm"
+#define MOV  "movd"
+#define MOVQ "movq"
+#define MOVQU "movq"
+#define STEP 4
+#define LOAD(mem,dst) \
+            MOV"       "mem", "dst" \n\t"\
+            "punpcklbw "MM"7, "dst" \n\t"
+#define PSRL1(reg) "psrlq $8, "reg" \n\t"
+#define PSRL2(reg) "psrlq $16, "reg" \n\t"
+#define PSHUF(src,dst) "pshufw $9, "dst", "src" \n\t"
+#endif
+
+#ifdef COMPILE_TEMPLATE_SSSE3
+#define PABS(tmp,dst) \
+            "pabsw     "dst", "dst" \n\t"
+#else
+#define PABS(tmp,dst) \
+            "pxor     "tmp", "tmp" \n\t"\
+            "psubw    "dst", "tmp" \n\t"\
+            "pmaxsw   "tmp", "dst" \n\t"
+#endif
+
+#define CHECK(pj,mj) \
+            MOVQU" "#pj"(%[cur],%[mrefs]), "MM"2 \n\t" /* cur[x-refs-1+j] */\
+            MOVQU" "#mj"(%[cur],%[prefs]), "MM"3 \n\t" /* cur[x+refs-1-j] */\
+            MOVQ"      "MM"2, "MM"4 \n\t"\
+            MOVQ"      "MM"2, "MM"5 \n\t"\
+            "pxor      "MM"3, "MM"4 \n\t"\
+            "pavgb     "MM"3, "MM"5 \n\t"\
+            "pand     "MANGLE(ff_pb_1)", "MM"4 \n\t"\
+            "psubusb   "MM"4, "MM"5 \n\t"\
+            PSRL1(MM"5")                 \
+            "punpcklbw "MM"7, "MM"5 \n\t" /* (cur[x-refs+j] + cur[x+refs-j])>>1 */\
+            MOVQ"      "MM"2, "MM"4 \n\t"\
+            "psubusb   "MM"3, "MM"2 \n\t"\
+            "psubusb   "MM"4, "MM"3 \n\t"\
+            "pmaxub    "MM"3, "MM"2 \n\t"\
+            MOVQ"      "MM"2, "MM"3 \n\t"\
+            MOVQ"      "MM"2, "MM"4 \n\t" /* ABS(cur[x-refs-1+j] - cur[x+refs-1-j]) */\
+            PSRL1(MM"3")                  /* ABS(cur[x-refs  +j] - cur[x+refs  -j]) */\
+            PSRL2(MM"4")                  /* ABS(cur[x-refs+1+j] - cur[x+refs+1-j]) */\
+            "punpcklbw "MM"7, "MM"2 \n\t"\
+            "punpcklbw "MM"7, "MM"3 \n\t"\
+            "punpcklbw "MM"7, "MM"4 \n\t"\
+            "paddw     "MM"3, "MM"2 \n\t"\
+            "paddw     "MM"4, "MM"2 \n\t" /* score */
+
+#define CHECK1 \
+            MOVQ"      "MM"0, "MM"3 \n\t"\
+            "pcmpgtw   "MM"2, "MM"3 \n\t" /* if(score < spatial_score) */\
+            "pminsw    "MM"2, "MM"0 \n\t" /* spatial_score= score; */\
+            MOVQ"      "MM"3, "MM"6 \n\t"\
+            "pand      "MM"3, "MM"5 \n\t"\
+            "pandn     "MM"1, "MM"3 \n\t"\
+            "por       "MM"5, "MM"3 \n\t"\
+            MOVQ"      "MM"3, "MM"1 \n\t" /* spatial_pred= (cur[x-refs+j] + cur[x+refs-j])>>1; */
+
+#define CHECK2 /* pretend not to have checked dir=2 if dir=1 was bad.\
+                  hurts both quality and speed, but matches the C version. */\
+            "paddw    "MANGLE(ff_pw_1)", "MM"6 \n\t"\
+            "psllw     $14,   "MM"6 \n\t"\
+            "paddsw    "MM"6, "MM"2 \n\t"\
+            MOVQ"      "MM"0, "MM"3 \n\t"\
+            "pcmpgtw   "MM"2, "MM"3 \n\t"\
+            "pminsw    "MM"2, "MM"0 \n\t"\
+            "pand      "MM"3, "MM"5 \n\t"\
+            "pandn     "MM"1, "MM"3 \n\t"\
+            "por       "MM"5, "MM"3 \n\t"\
+            MOVQ"      "MM"3, "MM"1 \n\t"
+
+void RENAME(ff_yadif_filter_line)(uint8_t *dst,
+                                  uint8_t *prev, uint8_t *cur, uint8_t *next,
+                                  int w, int refs, int parity, int mode)
+{
+    DECLARE_ALIGNED(16, uint8_t, tmp0[16]);
+    DECLARE_ALIGNED(16, uint8_t, tmp1[16]);
+    DECLARE_ALIGNED(16, uint8_t, tmp2[16]);
+    DECLARE_ALIGNED(16, uint8_t, tmp3[16]);
+    int x;
+
+#define FILTER\
+    for(x=0; x<w; x+=STEP){\
+        __asm__ volatile(\
+            "pxor      "MM"7, "MM"7 \n\t"\
+            LOAD("(%[cur],%[mrefs])", MM"0") /* c = cur[x-refs] */\
+            LOAD("(%[cur],%[prefs])", MM"1") /* e = cur[x+refs] */\
+            LOAD("(%["prev2"])", MM"2") /* prev2[x] */\
+            LOAD("(%["next2"])", MM"3") /* next2[x] */\
+            MOVQ"      "MM"3, "MM"4 \n\t"\
+            "paddw     "MM"2, "MM"3 \n\t"\
+            "psraw     $1,    "MM"3 \n\t" /* d = (prev2[x] + next2[x])>>1 */\
+            MOVQ"      "MM"0, %[tmp0] \n\t" /* c */\
+            MOVQ"      "MM"3, %[tmp1] \n\t" /* d */\
+            MOVQ"      "MM"1, %[tmp2] \n\t" /* e */\
+            "psubw     "MM"4, "MM"2 \n\t"\
+            PABS(      MM"4", MM"2") /* temporal_diff0 */\
+            LOAD("(%[prev],%[mrefs])", MM"3") /* prev[x-refs] */\
+            LOAD("(%[prev],%[prefs])", MM"4") /* prev[x+refs] */\
+            "psubw     "MM"0, "MM"3 \n\t"\
+            "psubw     "MM"1, "MM"4 \n\t"\
+            PABS(      MM"5", MM"3")\
+            PABS(      MM"5", MM"4")\
+            "paddw     "MM"4, "MM"3 \n\t" /* temporal_diff1 */\
+            "psrlw     $1,    "MM"2 \n\t"\
+            "psrlw     $1,    "MM"3 \n\t"\
+            "pmaxsw    "MM"3, "MM"2 \n\t"\
+            LOAD("(%[next],%[mrefs])", MM"3") /* next[x-refs] */\
+            LOAD("(%[next],%[prefs])", MM"4") /* next[x+refs] */\
+            "psubw     "MM"0, "MM"3 \n\t"\
+            "psubw     "MM"1, "MM"4 \n\t"\
+            PABS(      MM"5", MM"3")\
+            PABS(      MM"5", MM"4")\
+            "paddw     "MM"4, "MM"3 \n\t" /* temporal_diff2 */\
+            "psrlw     $1,    "MM"3 \n\t"\
+            "pmaxsw    "MM"3, "MM"2 \n\t"\
+            MOVQ"      "MM"2, %[tmp3] \n\t" /* diff */\
+\
+            "paddw     "MM"0, "MM"1 \n\t"\
+            "paddw     "MM"0, "MM"0 \n\t"\
+            "psubw     "MM"1, "MM"0 \n\t"\
+            "psrlw     $1,    "MM"1 \n\t" /* spatial_pred */\
+            PABS(      MM"2", MM"0")      /* ABS(c-e) */\
+\
+            MOVQU" -1(%[cur],%[mrefs]), "MM"2 \n\t" /* cur[x-refs-1] */\
+            MOVQU" -1(%[cur],%[prefs]), "MM"3 \n\t" /* cur[x+refs-1] */\
+            MOVQ"      "MM"2, "MM"4 \n\t"\
+            "psubusb   "MM"3, "MM"2 \n\t"\
+            "psubusb   "MM"4, "MM"3 \n\t"\
+            "pmaxub    "MM"3, "MM"2 \n\t"\
+            PSHUF(MM"3", MM"2") \
+            "punpcklbw "MM"7, "MM"2 \n\t" /* ABS(cur[x-refs-1] - cur[x+refs-1]) */\
+            "punpcklbw "MM"7, "MM"3 \n\t" /* ABS(cur[x-refs+1] - cur[x+refs+1]) */\
+            "paddw     "MM"2, "MM"0 \n\t"\
+            "paddw     "MM"3, "MM"0 \n\t"\
+            "psubw    "MANGLE(ff_pw_1)", "MM"0 \n\t" /* spatial_score */\
+\
+            CHECK(-2,0)\
+            CHECK1\
+            CHECK(-3,1)\
+            CHECK2\
+            CHECK(0,-2)\
+            CHECK1\
+            CHECK(1,-3)\
+            CHECK2\
+\
+            /* if(p->mode<2) ... */\
+            MOVQ"    %[tmp3], "MM"6 \n\t" /* diff */\
+            "cmpl      $2, %[mode] \n\t"\
+            "jge       1f \n\t"\
+            LOAD("(%["prev2"],%[mrefs],2)", MM"2") /* prev2[x-2*refs] */\
+            LOAD("(%["next2"],%[mrefs],2)", MM"4") /* next2[x-2*refs] */\
+            LOAD("(%["prev2"],%[prefs],2)", MM"3") /* prev2[x+2*refs] */\
+            LOAD("(%["next2"],%[prefs],2)", MM"5") /* next2[x+2*refs] */\
+            "paddw     "MM"4, "MM"2 \n\t"\
+            "paddw     "MM"5, "MM"3 \n\t"\
+            "psrlw     $1,    "MM"2 \n\t" /* b */\
+            "psrlw     $1,    "MM"3 \n\t" /* f */\
+            MOVQ"    %[tmp0], "MM"4 \n\t" /* c */\
+            MOVQ"    %[tmp1], "MM"5 \n\t" /* d */\
+            MOVQ"    %[tmp2], "MM"7 \n\t" /* e */\
+            "psubw     "MM"4, "MM"2 \n\t" /* b-c */\
+            "psubw     "MM"7, "MM"3 \n\t" /* f-e */\
+            MOVQ"      "MM"5, "MM"0 \n\t"\
+            "psubw     "MM"4, "MM"5 \n\t" /* d-c */\
+            "psubw     "MM"7, "MM"0 \n\t" /* d-e */\
+            MOVQ"      "MM"2, "MM"4 \n\t"\
+            "pminsw    "MM"3, "MM"2 \n\t"\
+            "pmaxsw    "MM"4, "MM"3 \n\t"\
+            "pmaxsw    "MM"5, "MM"2 \n\t"\
+            "pminsw    "MM"5, "MM"3 \n\t"\
+            "pmaxsw    "MM"0, "MM"2 \n\t" /* max */\
+            "pminsw    "MM"0, "MM"3 \n\t" /* min */\
+            "pxor      "MM"4, "MM"4 \n\t"\
+            "pmaxsw    "MM"3, "MM"6 \n\t"\
+            "psubw     "MM"2, "MM"4 \n\t" /* -max */\
+            "pmaxsw    "MM"4, "MM"6 \n\t" /* diff= MAX3(diff, min, -max); */\
+            "1: \n\t"\
+\
+            MOVQ"    %[tmp1], "MM"2 \n\t" /* d */\
+            MOVQ"      "MM"2, "MM"3 \n\t"\
+            "psubw     "MM"6, "MM"2 \n\t" /* d-diff */\
+            "paddw     "MM"6, "MM"3 \n\t" /* d+diff */\
+            "pmaxsw    "MM"2, "MM"1 \n\t"\
+            "pminsw    "MM"3, "MM"1 \n\t" /* d = clip(spatial_pred, d-diff, d+diff); */\
+            "packuswb  "MM"1, "MM"1 \n\t"\
+\
+            :[tmp0]"=m"(tmp0),\
+             [tmp1]"=m"(tmp1),\
+             [tmp2]"=m"(tmp2),\
+             [tmp3]"=m"(tmp3)\
+            :[prev] "r"(prev),\
+             [cur]  "r"(cur),\
+             [next] "r"(next),\
+             [prefs]"r"((x86_reg)refs),\
+             [mrefs]"r"((x86_reg)-refs),\
+             [mode] "g"(mode)\
+        );\
+        __asm__ volatile(MOV" "MM"1, %0" :"=m"(*dst));\
+        dst += STEP;\
+        prev+= STEP;\
+        cur += STEP;\
+        next+= STEP;\
+    }
+
+    if (parity) {
+#define prev2 "prev"
+#define next2 "cur"
+        FILTER
+#undef prev2
+#undef next2
+    } else {
+#define prev2 "cur"
+#define next2 "next"
+        FILTER
+#undef prev2
+#undef next2
+    }
+}
+#undef STEP
+#undef MM
+#undef MOV
+#undef MOVQ
+#undef MOVQU
+#undef PSHUF
+#undef PSRL1
+#undef PSRL2
+#undef LOAD
+#undef PABS
+#undef CHECK
+#undef CHECK1
+#undef CHECK2
+#undef FILTER
+