add VC-1 MMX DSP functions, under MIT license.

patch by Christophe GISQUET %christophe P gisquet A free P fr%
original thread:
date: Jul 7, 2007 12:52 PM
subject: [FFmpeg-devel] [PATCH] VC-1 MMX DSP functions

Originally committed as revision 11074 to svn://svn.ffmpeg.org/ffmpeg/trunk

2 files changed, 510 insertions, 0 deletions

diff --git a/libavcodec/i386/dsputil_mmx.c b/libavcodec/i386/dsputil_mmx.c
index 7a63d2562f..1310eb4eef 100644
--- a/libavcodec/i386/dsputil_mmx.c
+++ b/libavcodec/i386/dsputil_mmx.c
@@ -2838,6 +2838,10 @@ PREFETCH(prefetch_3dnow, prefetch)
 #undef PREFETCH
 
 #include "h264dsp_mmx.c"
+#if defined(CONFIG_VC1_DECODER) || defined(CONFIG_WMV3_DECODER) \
+     || defined(CONFIG_VC1_ENCODER) || defined(CONFIG_WMV3_ENCODER)
+#include "vc1dsp_mmx.c"
+#endif
 
 /* AVS specific */
 void ff_cavsdsp_init_mmx2(DSPContext* c, AVCodecContext *avctx);
@@ -3627,6 +3631,11 @@ void dsputil_init_mmx(DSPContext* c, AVCodecContext *avctx)
             ff_cavsdsp_init_mmx2(c, avctx);
 #endif
 
+#if defined(CONFIG_VC1_DECODER) || defined(CONFIG_WMV3_DECODER) \
+     || defined(CONFIG_VC1_ENCODER) || defined(CONFIG_WMV3_ENCODER)
+            ff_vc1dsp_init_mmx(c, avctx);
+#endif
+
 #ifdef CONFIG_ENCODERS
             c->sub_hfyu_median_prediction= sub_hfyu_median_prediction_mmx2;
 #endif //CONFIG_ENCODERS
diff --git a/libavcodec/i386/vc1dsp_mmx.c b/libavcodec/i386/vc1dsp_mmx.c
new file mode 100644
index 0000000000..3032aff86f
--- /dev/null
+++ b/libavcodec/i386/vc1dsp_mmx.c
@@ -0,0 +1,501 @@
+/*
+ * VC-1 and WMV3 - DSP functions MMX-optimized
+ * Copyright (c) 2007 Christophe GISQUET <christophe.gisquet@free.fr>
+ *
+ * Permission is hereby granted, free of charge, to any person
+ * obtaining a copy of this software and associated documentation
+ * files (the "Software"), to deal in the Software without
+ * restriction, including without limitation the rights to use,
+ * copy, modify, merge, publish, distribute, sublicense, and/or sell
+ * copies of the Software, and to permit persons to whom the
+ * Software is furnished to do so, subject to the following
+ * conditions:
+ *
+ * The above copyright notice and this permission notice shall be
+ * included in all copies or substantial portions of the Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
+ * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES
+ * OF MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
+ * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT
+ * HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY,
+ * WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
+ * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR
+ * OTHER DEALINGS IN THE SOFTWARE.
+ */
+
+#include "dsputil.h"
+#include "x86_cpu.h"
+
+/** Add rounder from mm7 to mm3 and pack result at destination */
+#define NORMALIZE_MMX(SHIFT)                                    \
+     "paddw     %%mm7, %%mm3           \n\t" /* +bias-r */      \
+     "paddw     %%mm7, %%mm4           \n\t" /* +bias-r */      \
+     "psraw     "SHIFT", %%mm3         \n\t"                    \
+     "psraw     "SHIFT", %%mm4         \n\t"
+
+#define TRANSFER_DO_PACK                        \
+     "packuswb  %%mm4, %%mm3           \n\t"    \
+     "movq      %%mm3, (%2)            \n\t"
+
+#define TRANSFER_DONT_PACK                      \
+     "movq      %%mm3, 0(%2)           \n\t"    \
+     "movq      %%mm4, 8(%2)           \n\t"
+
+/** @see MSPEL_FILTER13_CORE for use as UNPACK macro */
+#define DO_UNPACK(reg)  "punpcklbw %%mm0, " reg "\n\t"
+#define DONT_UNPACK(reg)
+
+/** Compute the rounder 32-r or 8-r and unpacks it to mm7 */
+#define LOAD_ROUNDER_MMX(ROUND)                 \
+     "movd      "ROUND", %%mm7         \n\t"    \
+     "punpcklwd %%mm7, %%mm7           \n\t"    \
+     "punpckldq %%mm7, %%mm7           \n\t"
+
+#define SHIFT2_LINE(OFF, R0,R1,R2,R3)           \
+    "paddw     %%mm"#R2", %%mm"#R1"    \n\t"    \
+    "movd      (%1,%4), %%mm"#R0"      \n\t"    \
+    "pmullw    %%mm6, %%mm"#R1"        \n\t"    \
+    "punpcklbw %%mm0, %%mm"#R0"        \n\t"    \
+    "movd      (%1,%3), %%mm"#R3"      \n\t"    \
+    "psubw     %%mm"#R0", %%mm"#R1"    \n\t"    \
+    "punpcklbw %%mm0, %%mm"#R3"        \n\t"    \
+    "paddw     %%mm7, %%mm"#R1"        \n\t"    \
+    "psubw     %%mm"#R3", %%mm"#R1"    \n\t"    \
+    "psraw     %5, %%mm"#R1"           \n\t"    \
+    "movq      %%mm"#R1", "#OFF"(%2)   \n\t"    \
+    "add       %3, %1                  \n\t"
+
+DECLARE_ALIGNED_16(static const uint64_t, fact_9) = 0x0009000900090009ULL;
+
+/** Sacrifying mm6 allows to pipeline loads from src */
+static void vc1_put_ver_16b_shift2_mmx(int16_t *dst,
+                                       const uint8_t *src, long int stride,
+                                       int rnd, int64_t shift)
+{
+    int  w = 3;
+
+    asm volatile(
+        LOAD_ROUNDER_MMX("%6")
+        "movq      %7, %%mm6               \n\t"
+        "1:                                \n\t"
+        "movd      (%1), %%mm2             \n\t"
+        "add       %3, %1                  \n\t"
+        "movd      (%1), %%mm3             \n\t"
+        "punpcklbw %%mm0, %%mm2            \n\t"
+        "punpcklbw %%mm0, %%mm3            \n\t"
+        SHIFT2_LINE(  0, 1, 2, 3, 4)
+        SHIFT2_LINE( 24, 2, 3, 4, 1)
+        SHIFT2_LINE( 48, 3, 4, 1, 2)
+        SHIFT2_LINE( 72, 4, 1, 2, 3)
+        SHIFT2_LINE( 96, 1, 2, 3, 4)
+        SHIFT2_LINE(120, 2, 3, 4, 1)
+        SHIFT2_LINE(144, 3, 4, 1, 2)
+        SHIFT2_LINE(168, 4, 1, 2, 3)
+        "sub       %8, %1                  \n\t"
+        "add       $8, %2                  \n\t"
+        "dec       %0                      \n\t"
+        "jnz 1b                            \n\t"
+        : "+g"(w), "+r"(src), "+r"(dst)
+        : "r"(stride), "r"(-2*stride), "m"(shift),
+          "m"(rnd), "m"(fact_9), "g"(9*stride-4)
+        : "memory"
+    );
+}
+
+/** To remove bias allowing use of MMX 16bits arithmetic */
+DECLARE_ALIGNED_16(static const uint64_t, ff_pw_128) = 0x0080008000800080ULL;
+
+/**
+ * Data is already unpacked, so some operations can directly be made from
+ * memory.
+ */
+static void vc1_put_hor_16b_shift2_mmx(uint8_t *dst, long int stride,
+                                       const int16_t *src, int rnd)
+{
+    int h = 8;
+
+    src -= 1;
+    rnd -= (-1+9+9-1)*1024; /* Add -1024 bias */
+    asm volatile(
+        LOAD_ROUNDER_MMX("%4")
+        "movq      %6, %%mm6               \n\t"
+        "movq      %5, %%mm5               \n\t"
+        "1:                                \n\t"
+        "movq      2*0+0(%1), %%mm1        \n\t"
+        "movq      2*0+8(%1), %%mm2        \n\t"
+        "movq      2*1+0(%1), %%mm3        \n\t"
+        "movq      2*1+8(%1), %%mm4        \n\t"
+        "paddw     2*3+0(%1), %%mm1        \n\t"
+        "paddw     2*3+8(%1), %%mm2        \n\t"
+        "paddw     2*2+0(%1), %%mm3        \n\t"
+        "paddw     2*2+8(%1), %%mm4        \n\t"
+        "pmullw    %%mm5, %%mm3            \n\t"
+        "pmullw    %%mm5, %%mm4            \n\t"
+        "psubw     %%mm1, %%mm3            \n\t"
+        "psubw     %%mm2, %%mm4            \n\t"
+        NORMALIZE_MMX("$7")
+        /* Remove bias */
+        "paddw     %%mm6, %%mm3            \n\t"
+        "paddw     %%mm6, %%mm4            \n\t"
+        TRANSFER_DO_PACK
+        "add       $24, %1                 \n\t"
+        "add       %3, %2                  \n\t"
+        "dec       %0                      \n\t"
+        "jnz 1b                            \n\t"
+        : "+g"(h), "+r" (src),  "+r" (dst)
+        : "g"(stride), "m"(rnd), "m"(fact_9), "m"(ff_pw_128)
+        : "memory"
+    );
+}
+
+
+/**
+ * Purely vertical or horizontal 1/2 shift interpolation.
+ * Sacrify mm6 for *9 factor.
+ */
+#define SDTEST  0
+
+static void vc1_put_shift2_mmx(uint8_t *dst, const uint8_t *src,
+                               long int stride, int rnd, long int offset)
+{
+    int h = 8;
+    START_TIMER
+    rnd = 8-rnd;
+    asm volatile(
+        LOAD_ROUNDER_MMX("%6")
+        "movq      %8, %%mm6               \n\t"
+        "1:                                \n\t"
+        "movd      0(%1   ), %%mm3         \n\t"
+        "movd      4(%1   ), %%mm4         \n\t"
+        "movd      0(%1,%3), %%mm1         \n\t"
+        "movd      4(%1,%3), %%mm2         \n\t"
+        "add       %3, %1                  \n\t"
+        "punpcklbw %%mm0, %%mm3            \n\t"
+        "punpcklbw %%mm0, %%mm4            \n\t"
+        "punpcklbw %%mm0, %%mm1            \n\t"
+        "punpcklbw %%mm0, %%mm2            \n\t"
+        "paddw     %%mm1, %%mm3            \n\t"
+        "paddw     %%mm2, %%mm4            \n\t"
+        "movd      0(%1,%3), %%mm1         \n\t"
+        "movd      4(%1,%3), %%mm2         \n\t"
+        "pmullw    %%mm6, %%mm3            \n\t" /* 0,9,9,0*/
+        "pmullw    %%mm6, %%mm4            \n\t" /* 0,9,9,0*/
+        "punpcklbw %%mm0, %%mm1            \n\t"
+        "punpcklbw %%mm0, %%mm2            \n\t"
+        "psubw     %%mm1, %%mm3            \n\t" /*-1,9,9,0*/
+        "psubw     %%mm2, %%mm4            \n\t" /*-1,9,9,0*/
+        "movd      0(%1,%3), %%mm1         \n\t"
+        "movd      4(%1,%3), %%mm2         \n\t"
+        "punpcklbw %%mm0, %%mm1            \n\t"
+        "punpcklbw %%mm0, %%mm2            \n\t"
+        "psubw     %%mm1, %%mm3            \n\t" /*-1,9,9,-1*/
+        "psubw     %%mm2, %%mm4            \n\t" /*-1,9,9,-1*/
+        NORMALIZE_MMX("$4")
+        TRANSFER_DO_PACK
+        "add       %7, %1                  \n\t"
+        "add       %5, %2                  \n\t"
+        "dec       %0                      \n\t"
+        "jnz 1b                            \n\t"
+        : "+g"(h), "+r"(src),  "+r"(dst)
+        : "r"(offset), "r"(-2*offset), "g"(stride), "m"(rnd),
+          "g"(stride-offset), "m"(fact_9)
+        : "memory"
+    );
+    STOP_TIMER("norm")
+}
+
+/**
+ * Filter coefficients made global to allow access by all 1 or 3 quarter shift
+ * interpolation functions.
+ */
+DECLARE_ALIGNED_16(static const uint64_t, fact_53) = 0x0035003500350035ULL;
+DECLARE_ALIGNED_16(static const uint64_t, fact_18) = 0x0012001200120012ULL;
+
+/**
+ * Core of the 1/4 and 3/4 shift bicubic interpolation.
+ *
+ * @param UNPACK  Macro unpacking arguments from 8 to 16bits (can be empty).
+ * @param MOVQ    "movd 1" or "movq 2", if data read is already unpacked.
+ * @param A1      Address of 1st tap (beware of unpacked/packed).
+ * @param A2      Address of 2nd tap
+ * @param A3      Address of 3rd tap
+ * @param A4      Address of 4th tap
+ */
+#define MSPEL_FILTER13_CORE(UNPACK, MOVQ, A1, A2, A3, A4, POS)  \
+     MOVQ "*0+"A1", %%mm1       \n\t"                           \
+     MOVQ "*4+"A1", %%mm2       \n\t"                           \
+     UNPACK("%%mm1")                                            \
+     UNPACK("%%mm2")                                            \
+     "pmullw    "POS", %%mm1    \n\t"                           \
+     "pmullw    "POS", %%mm2    \n\t"                           \
+     MOVQ "*0+"A2", %%mm3       \n\t"                           \
+     MOVQ "*4+"A2", %%mm4       \n\t"                           \
+     UNPACK("%%mm3")                                            \
+     UNPACK("%%mm4")                                            \
+     "pmullw    %%mm6, %%mm3    \n\t" /* *18 */                 \
+     "pmullw    %%mm6, %%mm4    \n\t" /* *18 */                 \
+     "psubw     %%mm1, %%mm3    \n\t" /* 18,-3 */               \
+     "psubw     %%mm2, %%mm4    \n\t" /* 18,-3 */               \
+     MOVQ "*0+"A4", %%mm1       \n\t"                           \
+     MOVQ "*4+"A4", %%mm2       \n\t"                           \
+     UNPACK("%%mm1")                                            \
+     UNPACK("%%mm2")                                            \
+     "psllw     $2, %%mm1       \n\t" /* 4* */                  \
+     "psllw     $2, %%mm2       \n\t" /* 4* */                  \
+     "psubw     %%mm1, %%mm3    \n\t" /* -4,18,-3 */            \
+     "psubw     %%mm2, %%mm4    \n\t" /* -4,18,-3 */            \
+     MOVQ "*0+"A3", %%mm1       \n\t"                           \
+     MOVQ "*4+"A3", %%mm2       \n\t"                           \
+     UNPACK("%%mm1")                                            \
+     UNPACK("%%mm2")                                            \
+     "pmullw    %%mm5, %%mm1    \n\t" /* *53 */                 \
+     "pmullw    %%mm5, %%mm2    \n\t" /* *53 */                 \
+     "paddw     %%mm1, %%mm3    \n\t" /* 4,53,18,-3 */          \
+     "paddw     %%mm2, %%mm4    \n\t" /* 4,53,18,-3 */
+
+/**
+ * Macro to build the vertical 16bits version of vc1_put_shift[13].
+ * Here, offset=src_stride. Parameters passed A1 to A4 must use
+ * %3 (src_stride) and %4 (3*src_stride).
+ *
+ * @param  NAME   Either 1 or 3
+ * @see MSPEL_FILTER13_CORE for information on A1->A4
+ */
+#define MSPEL_FILTER13_VER_16B(NAME, A1, A2, A3, A4)                    \
+static void                                                             \
+vc1_put_ver_16b_ ## NAME ## _mmx(int16_t *dst, const uint8_t *src,      \
+                                 long int src_stride,                   \
+                                 int rnd, int64_t shift)                \
+{                                                                       \
+    int h = 8;                                                          \
+    src -= src_stride;                                                  \
+    asm volatile(                                                       \
+        LOAD_ROUNDER_MMX("%5")                                          \
+        "movq      %7, %%mm5       \n\t"                                \
+        "movq      %8, %%mm6       \n\t"                                \
+        ASMALIGN(3)                                                     \
+        "1:                        \n\t"                                \
+        MSPEL_FILTER13_CORE(DO_UNPACK, "movd  1", A1, A2, A3, A4, "%9") \
+        NORMALIZE_MMX("%6")                                             \
+        TRANSFER_DONT_PACK                                              \
+        /* Last 3 (in fact 4) bytes on the line */                      \
+        "movd      8+"A1", %%mm1   \n\t"                                \
+        DO_UNPACK("%%mm1")                                              \
+        "movq      %%mm1, %%mm3    \n\t"                                \
+        "paddw     %%mm1, %%mm1    \n\t"                                \
+        "paddw     %%mm3, %%mm1    \n\t" /* 3* */                       \
+        "movd      8+"A2", %%mm3   \n\t"                                \
+        DO_UNPACK("%%mm3")                                              \
+        "pmullw    %%mm6, %%mm3    \n\t" /* *18 */                      \
+        "psubw     %%mm1, %%mm3    \n\t" /*18,-3 */                     \
+        "movd      8+"A3", %%mm1   \n\t"                                \
+        DO_UNPACK("%%mm1")                                              \
+        "pmullw    %%mm5, %%mm1    \n\t" /* *53 */                      \
+        "paddw     %%mm1, %%mm3    \n\t" /*53,18,-3 */                  \
+        "movd      8+"A4", %%mm1   \n\t"                                \
+        DO_UNPACK("%%mm1")                                              \
+        "psllw     $2, %%mm1       \n\t" /* 4* */                       \
+        "psubw     %%mm1, %%mm3    \n\t"                                \
+        "paddw     %%mm7, %%mm3    \n\t"                                \
+        "psraw     %6, %%mm3       \n\t"                                \
+        "movq      %%mm3, 16(%2)   \n\t"                                \
+        "add       %3, %1          \n\t"                                \
+        "add       $24, %2         \n\t"                                \
+        "dec       %0              \n\t"                                \
+        "jnz 1b                    \n\t"                                \
+        : "+g"(h), "+r" (src),  "+r" (dst)                              \
+        : "r"(src_stride), "r"(3*src_stride),                           \
+          "m"(rnd), "m"(shift),                                         \
+          "m"(fact_53), "m"(fact_18), "m"(ff_pw_3)                      \
+        : "memory"                                                      \
+    );                                                                  \
+}
+
+/**
+ * Macro to build the horizontal 16bits version of vc1_put_shift[13].
+ * Here, offset=16bits, so parameters passed A1 to A4 should be simple.
+ *
+ * @param  NAME   Either 1 or 3
+ * @see MSPEL_FILTER13_CORE for information on A1->A4
+ */
+#define MSPEL_FILTER13_HOR_16B(NAME, A1, A2, A3, A4)                    \
+static void                                                             \
+vc1_put_hor_16b_ ## NAME ## _mmx(uint8_t *dst, long int stride,         \
+                                 const int16_t *src, int rnd)           \
+{                                                                       \
+    int h = 8;                                                          \
+    src -= 1;                                                           \
+    rnd -= (-4+58+13-3)*256; /* Add -256 bias */                        \
+    asm volatile(                                                       \
+        LOAD_ROUNDER_MMX("%4")                                          \
+        "movq      %6, %%mm6       \n\t"                                \
+        "movq      %5, %%mm5       \n\t"                                \
+        ASMALIGN(3)                                                     \
+        "1:                        \n\t"                                \
+        MSPEL_FILTER13_CORE(DONT_UNPACK, "movq 2", A1, A2, A3, A4, "%8")\
+        NORMALIZE_MMX("$7")                                             \
+        /* Remove bias */                                               \
+        "paddw     %7, %%mm3       \n\t"                                \
+        "paddw     %7, %%mm4       \n\t"                                \
+        TRANSFER_DO_PACK                                                \
+        "add       $24, %1         \n\t"                                \
+        "add       %3, %2          \n\t"                                \
+        "dec       %0              \n\t"                                \
+        "jnz 1b                    \n\t"                                \
+        : "+g"(h), "+r" (src),  "+r" (dst)                              \
+        : "g"(stride), "m"(rnd), "m"(fact_53), "m"(fact_18),            \
+          "m"(ff_pw_128), "m"(ff_pw_3)                                  \
+        : "memory"                                                      \
+    );                                                                  \
+}
+
+/**
+ * Macro to build the 8bits, any direction, version of vc1_put_shift[13].
+ * Here, offset=src_stride. Parameters passed A1 to A4 must use
+ * %3 (offset) and %4 (3*offset).
+ *
+ * @param  NAME   Either 1 or 3
+ * @see MSPEL_FILTER13_CORE for information on A1->A4
+ */
+#define MSPEL_FILTER13_8B(NAME, A1, A2, A3, A4)                         \
+static void                                                             \
+vc1_put_## NAME ## _mmx(uint8_t *dst, const uint8_t *src,               \
+                        long int stride, int rnd, long int offset)      \
+{                                                                       \
+    int h = 8;                                                          \
+    src -= offset;                                                      \
+    rnd = 32-rnd;                                                       \
+    asm volatile (                                                      \
+        LOAD_ROUNDER_MMX("%6")                                          \
+        "movq      %7, %%mm5       \n\t"                                \
+        "movq      %8, %%mm6       \n\t"                                \
+        ASMALIGN(3)                                                     \
+        "1:                        \n\t"                                \
+        MSPEL_FILTER13_CORE(DO_UNPACK, "movd   1", A1, A2, A3, A4, "%9")\
+        NORMALIZE_MMX("$6")                                             \
+        TRANSFER_DO_PACK                                                \
+        "add       %5, %1          \n\t"                                \
+        "add       %5, %2          \n\t"                                \
+        "dec       %0              \n\t"                                \
+        "jnz 1b                    \n\t"                                \
+        : "+g"(h), "+r" (src),  "+r" (dst)                              \
+        : "r"(offset), "r"(3*offset), "g"(stride), "m"(rnd),            \
+          "m"(fact_53), "m"(fact_18), "m"(ff_pw_3)                      \
+        : "memory"                                                      \
+    );                                                                  \
+}
+
+/** 1/4 shift bicubic interpolation */
+MSPEL_FILTER13_8B     (shift1, "0(%1,%4  )", "0(%1,%3,2)", "0(%1,%3  )", "0(%1     )")
+MSPEL_FILTER13_VER_16B(shift1, "0(%1,%4  )", "0(%1,%3,2)", "0(%1,%3  )", "0(%1     )")
+MSPEL_FILTER13_HOR_16B(shift1, "2*3(%1)", "2*2(%1)", "2*1(%1)", "2*0(%1)")
+
+/** 3/4 shift bicubic interpolation */
+MSPEL_FILTER13_8B     (shift3, "0(%1     )", "0(%1,%3  )", "0(%1,%3,2)", "0(%1,%4  )")
+MSPEL_FILTER13_VER_16B(shift3, "0(%1     )", "0(%1,%3  )", "0(%1,%3,2)", "0(%1,%4  )")
+MSPEL_FILTER13_HOR_16B(shift3, "2*0(%1)", "2*1(%1)", "2*2(%1)", "2*3(%1)")
+
+typedef void (*vc1_mspel_mc_filter_ver_16bits)(int16_t *dst, const uint8_t *src, long int src_stride, int rnd, int64_t shift);
+typedef void (*vc1_mspel_mc_filter_hor_16bits)(uint8_t *dst, long int dst_stride, const int16_t *src, int rnd);
+typedef void (*vc1_mspel_mc_filter_8bits)(uint8_t *dst, const uint8_t *src, long int stride, int rnd, long int offset);
+
+/**
+ * Interpolates fractional pel values by applying proper vertical then
+ * horizontal filter.
+ *
+ * @param  dst     Destination buffer for interpolated pels.
+ * @param  src     Source buffer.
+ * @param  stride  Stride for both src and dst buffers.
+ * @param  hmode   Horizontal filter (expressed in quarter pixels shift).
+ * @param  hmode   Vertical filter.
+ * @param  rnd     Rounding bias.
+ */
+static void vc1_mspel_mc(uint8_t *dst, const uint8_t *src, int stride,
+                         int hmode, int vmode, int rnd)
+{
+    static const vc1_mspel_mc_filter_ver_16bits vc1_put_shift_ver_16bits[] =
+         { NULL, vc1_put_ver_16b_shift1_mmx, vc1_put_ver_16b_shift2_mmx, vc1_put_ver_16b_shift3_mmx };
+    static const vc1_mspel_mc_filter_hor_16bits vc1_put_shift_hor_16bits[] =
+         { NULL, vc1_put_hor_16b_shift1_mmx, vc1_put_hor_16b_shift2_mmx, vc1_put_hor_16b_shift3_mmx };
+    static const vc1_mspel_mc_filter_8bits vc1_put_shift_8bits[] =
+         { NULL, vc1_put_shift1_mmx, vc1_put_shift2_mmx, vc1_put_shift3_mmx };
+
+    asm volatile(
+        "pxor %%mm0, %%mm0         \n\t"
+        ::: "memory"
+    );
+
+    if (vmode) { /* Vertical filter to apply */
+        if (hmode) { /* Horizontal filter to apply, output to tmp */
+            static const int shift_value[] = { 0, 5, 1, 5 };
+            int              shift = (shift_value[hmode]+shift_value[vmode])>>1;
+            int              r;
+            DECLARE_ALIGNED_16(int16_t, tmp[12*8]);
+
+            r = (1<<(shift-1)) + rnd-1;
+            vc1_put_shift_ver_16bits[vmode](tmp, src-1, stride, r, shift);
+
+            vc1_put_shift_hor_16bits[hmode](dst, stride, tmp+1, 64-rnd);
+            return;
+        }
+        else { /* No horizontal filter, output 8 lines to dst */
+            vc1_put_shift_8bits[vmode](dst, src, stride, 1-rnd, stride);
+            return;
+        }
+    }
+
+    /* Horizontal mode with no vertical mode */
+    vc1_put_shift_8bits[hmode](dst, src, stride, rnd, 1);
+}
+
+static void put_vc1_mspel_mc00_mmx(uint8_t *dst, const uint8_t *src, int stride, int rnd) {
+    put_pixels8_mmx(dst, src, stride, 8);
+}
+
+/** Macro to ease bicubic filter interpolation functions declarations */
+#define DECLARE_FUNCTION(a, b)                                          \
+static void put_vc1_mspel_mc ## a ## b ## _mmx(uint8_t *dst, const uint8_t *src, int stride, int rnd) { \
+     vc1_mspel_mc(dst, src, stride, a, b, rnd);                         \
+}
+
+DECLARE_FUNCTION(0, 1)
+DECLARE_FUNCTION(0, 2)
+DECLARE_FUNCTION(0, 3)
+
+DECLARE_FUNCTION(1, 0)
+DECLARE_FUNCTION(1, 1)
+DECLARE_FUNCTION(1, 2)
+DECLARE_FUNCTION(1, 3)
+
+DECLARE_FUNCTION(2, 0)
+DECLARE_FUNCTION(2, 1)
+DECLARE_FUNCTION(2, 2)
+DECLARE_FUNCTION(2, 3)
+
+DECLARE_FUNCTION(3, 0)
+DECLARE_FUNCTION(3, 1)
+DECLARE_FUNCTION(3, 2)
+DECLARE_FUNCTION(3, 3)
+
+void ff_vc1dsp_init_mmx(DSPContext* dsp, AVCodecContext *avctx) {
+    dsp->put_vc1_mspel_pixels_tab[ 0] = put_vc1_mspel_mc00_mmx;
+    dsp->put_vc1_mspel_pixels_tab[ 4] = put_vc1_mspel_mc01_mmx;
+    dsp->put_vc1_mspel_pixels_tab[ 8] = put_vc1_mspel_mc02_mmx;
+    dsp->put_vc1_mspel_pixels_tab[12] = put_vc1_mspel_mc03_mmx;
+
+    dsp->put_vc1_mspel_pixels_tab[ 1] = put_vc1_mspel_mc10_mmx;
+    dsp->put_vc1_mspel_pixels_tab[ 5] = put_vc1_mspel_mc11_mmx;
+    dsp->put_vc1_mspel_pixels_tab[ 9] = put_vc1_mspel_mc12_mmx;
+    dsp->put_vc1_mspel_pixels_tab[13] = put_vc1_mspel_mc13_mmx;
+
+    dsp->put_vc1_mspel_pixels_tab[ 2] = put_vc1_mspel_mc20_mmx;
+    dsp->put_vc1_mspel_pixels_tab[ 6] = put_vc1_mspel_mc21_mmx;
+    dsp->put_vc1_mspel_pixels_tab[10] = put_vc1_mspel_mc22_mmx;
+    dsp->put_vc1_mspel_pixels_tab[14] = put_vc1_mspel_mc23_mmx;
+
+    dsp->put_vc1_mspel_pixels_tab[ 3] = put_vc1_mspel_mc30_mmx;
+    dsp->put_vc1_mspel_pixels_tab[ 7] = put_vc1_mspel_mc31_mmx;
+    dsp->put_vc1_mspel_pixels_tab[11] = put_vc1_mspel_mc32_mmx;
+    dsp->put_vc1_mspel_pixels_tab[15] = put_vc1_mspel_mc33_mmx;
+}