* new mmx code - based upon http://aggregate.org/MAGIC

  for now it's rather sneak preview (new functions are nearly 100% faster)

Originally committed as revision 623 to svn://svn.ffmpeg.org/ffmpeg/trunk

diff --git a/libavcodec/i386/dsputil_mmx.c b/libavcodec/i386/dsputil_mmx.c
index 1702b10..4cedb1f 100644 (file)
--- a/libavcodec/i386/dsputil_mmx.c
+++ b/libavcodec/i386/dsputil_mmx.c
@@ -49,6 +49,7 @@ void ff_mmx_idct(DCTELEM *block);
 void ff_mmxext_idct(DCTELEM *block);
 
 /* pixel operations */
+static const uint64_t mm_bfe  __attribute__ ((aligned(8))) = 0xfefefefefefefefeULL;
 static const uint64_t mm_bone __attribute__ ((aligned(8))) = 0x0101010101010101ULL;
 static const uint64_t mm_wone __attribute__ ((aligned(8))) = 0x0001000100010001ULL;
 static const uint64_t mm_wtwo __attribute__ ((aligned(8))) = 0x0002000200020002ULL;
@@ -62,6 +63,7 @@ static const uint64_t mm_wtwo __attribute__ ((aligned(8))) = 0x0002000200020002U
 #define MOVQ_WONE(regd)  __asm __volatile ("movq %0, %%" #regd " \n\t" ::"m"(mm_wone))
 #define MOVQ_WTWO(regd)  __asm __volatile ("movq %0, %%" #regd " \n\t" ::"m"(mm_wtwo))
 #define MOVQ_BONE(regd)  "movq "MANGLE(mm_bone)", "#regd" \n\t"
+#define MOVQ_BFE(regd)  "movq "MANGLE(mm_bfe)", "#regd" \n\t"
 #else
 // for shared library it's better to use this way for accessing constants
 // pcmpeqd -> -1
@@ -80,8 +82,31 @@ static const uint64_t mm_wtwo __attribute__ ((aligned(8))) = 0x0002000200020002U
        "pcmpeqd " #regd ", " #regd " \n\t" \
        "psrlw $15, " #regd " \n\t"\
        "packuswb " #regd ", " #regd " \n\t"
+
+#define MOVQ_BFE(regd) \
+       "pcmpeqd " #regd ", " #regd " \n\t"\
+       "paddb " #regd ", " #regd " \n\t"
 #endif
 
+// using mm6 as temporary and for the output result
+// first argument is unmodifed and second is trashed
+// mm7 is supposed to contain 0xfefefefefefefefe
+#define PAVG_MMX_NO_RND(rega, regb) \
+    "movq " #rega ", %%mm6     \n\t"\
+    "pand " #regb ", %%mm6     \n\t"\
+    "pxor " #rega ", " #regb " \n\t"\
+    "pand %%mm7, " #regb "     \n\t"\
+    "psrlq $1, " #regb "       \n\t"\
+    "paddb " #regb ", %%mm6    \n\t"
+
+#define PAVG_MMX(rega, regb) \
+    "movq " #rega ", %%mm6     \n\t"\
+    "por  " #regb ", %%mm6     \n\t"\
+    "pxor " #rega ", " #regb " \n\t"\
+    "pand %%mm7, " #regb "             \n\t"\
+    "psrlq $1, " #regb "       \n\t"\
+    "psubb " #regb ", %%mm6    \n\t"
+
 
 /***********************************/
 /* 3Dnow specific */
@@ -291,8 +316,12 @@ static void put_pixels_mmx(UINT8 *block, const UINT8 *pixels, int line_size, int
        );
 }
 
+// will have to be check if it's better to have bigger
+// unrolled code also on Celerons - for now  yes
+#define LONG_UNROLL 1
 static void put_pixels_x2_mmx(UINT8 *block, const UINT8 *pixels, int line_size, int h)
 {
+#if 0
   UINT8 *p;
   const UINT8 *pix;
   p = block;
@@ -301,7 +330,7 @@ static void put_pixels_x2_mmx(UINT8 *block, const UINT8 *pixels, int line_size,
   MOVQ_WONE(mm4);
   JUMPALIGN();
   do {
-    __asm __volatile(
+      __asm __volatile(
        "movq   %1, %%mm0\n\t"
        "movq   1%1, %%mm1\n\t"
        "movq   %%mm0, %%mm2\n\t"
@@ -320,13 +349,50 @@ static void put_pixels_x2_mmx(UINT8 *block, const UINT8 *pixels, int line_size,
        "movq   %%mm0, %0\n\t"
        :"=m"(*p)
        :"m"(*pix)
-       :"memory");
+                      :"memory");
    pix += line_size; p += line_size;
   } while (--h);
+#else
+  __asm __volatile(
+       MOVQ_BFE(%%mm7)
+       "lea (%3, %3), %%eax    \n\t"
+       ".balign 8              \n\t"
+       "1:                     \n\t"
+       "movq (%1), %%mm0       \n\t"
+       "movq (%1, %3), %%mm2   \n\t"
+       "movq 1(%1), %%mm1      \n\t"
+       "movq 1(%1, %3), %%mm3  \n\t"
+       PAVG_MMX(%%mm0, %%mm1)
+       "movq %%mm6, (%2)       \n\t"
+       PAVG_MMX(%%mm2, %%mm3)
+       "movq %%mm6, (%2, %3)   \n\t"
+       "addl %%eax, %1         \n\t"
+       "addl %%eax, %2         \n\t"
+#if LONG_UNROLL
+       "movq (%1), %%mm0       \n\t"
+       "movq (%1, %3), %%mm2   \n\t"
+       "movq 1(%1), %%mm1      \n\t"
+       "movq 1(%1, %3), %%mm3  \n\t"
+       PAVG_MMX(%%mm0, %%mm1)
+       "movq %%mm6, (%2)       \n\t"
+       PAVG_MMX(%%mm2, %%mm3)
+       "movq %%mm6, (%2, %3)   \n\t"
+       "addl %%eax, %1         \n\t"
+       "addl %%eax, %2         \n\t"
+       "subl $4, %0            \n\t"
+#else
+       "subl $2, %0            \n\t"
+#endif
+       "jnz 1b                 \n\t"
+       :"+g"(h), "+S"(pixels), "+D"(block)
+       :"r"(line_size)
+       :"eax", "memory");
+#endif
 }
 
 static void put_pixels_y2_mmx(UINT8 *block, const UINT8 *pixels, int line_size, int h)
 {
+#if 0
   UINT8 *p;
   const UINT8 *pix;
   p = block;
@@ -359,6 +425,41 @@ static void put_pixels_y2_mmx(UINT8 *block, const UINT8 *pixels, int line_size,
    pix += line_size;
    p += line_size;
   } while (--h);
+#else
+  __asm __volatile(
+       MOVQ_BFE(%%mm7)
+       "lea (%3, %3), %%eax    \n\t"
+       "movq (%1), %%mm0       \n\t"
+       ".balign 8              \n\t"
+       "1:                     \n\t"
+       "movq (%1, %3), %%mm1   \n\t"
+       "movq (%1, %%eax),%%mm2 \n\t"
+       PAVG_MMX(%%mm1, %%mm0)
+       "movq %%mm6, (%2)       \n\t"
+       PAVG_MMX(%%mm2, %%mm1)
+       "movq %%mm6, (%2, %3)   \n\t"
+       "addl %%eax, %1         \n\t"
+       "addl %%eax, %2         \n\t"
+#ifdef LONG_UNROLL
+       "movq (%1, %3), %%mm1   \n\t"
+       "movq (%1, %%eax),%%mm0 \n\t"
+       PAVG_MMX(%%mm1, %%mm2)
+       "movq %%mm6, (%2)       \n\t"
+       PAVG_MMX(%%mm0, %%mm1)
+       "movq %%mm6, (%2, %3)   \n\t"
+       "addl %%eax, %1         \n\t"
+       "addl %%eax, %2         \n\t"
+       "subl $4, %0            \n\t"
+#else
+       "subl $2, %0            \n\t"
+#endif
+       "jnz 1b                 \n\t"
+       :"+g"(h), "+S"(pixels), "+D"(block)
+       :"r"(line_size)
+       :"eax", "memory");
+#endif
+
+
 }
 
 static void put_pixels_xy2_mmx(UINT8 *block, const UINT8 *pixels, int line_size, int h)