x86/me_cmp: port mmxext vsad functions to yasm

Also add mmxext versions of vsad8 and vsad_intra8, and sse2 versions of
vsad16 and vsad_intra16.
Since vsad8 and vsad16 are not bitexact, they are accordingly marked as
approximate.

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

2 files changed, 181 insertions, 113 deletions

diff --git a/libavcodec/x86/me_cmp.asm b/libavcodec/x86/me_cmp.asm
index b657642c41..95b99c48fd 100644
--- a/libavcodec/x86/me_cmp.asm
+++ b/libavcodec/x86/me_cmp.asm
@@ -26,6 +26,7 @@
 SECTION_RODATA
 
 cextern pb_1
+cextern pb_80
 
 SECTION .text
 
@@ -772,3 +773,163 @@ SAD_APPROX_XY2 8
 SAD_APPROX_XY2 16
 INIT_XMM sse2
 SAD_APPROX_XY2 16
+
+;--------------------------------------------------------------------
+;int ff_vsad_intra(MpegEncContext *v, uint8_t *pix1, uint8_t *pix2,
+;                  int line_size, int h);
+;--------------------------------------------------------------------
+; %1 = 8/16
+%macro VSAD_INTRA 1
+cglobal vsad_intra%1, 5, 5, 3, v, pix1, pix2, lsize, h
+    mova      m0, [pix1q]
+%if %1 == mmsize
+    mova      m2, [pix1q+lsizeq]
+    psadbw    m0, m2
+%else
+    mova      m2, [pix1q+lsizeq]
+    mova      m3, [pix1q+8]
+    mova      m4, [pix1q+lsizeq+8]
+    psadbw    m0, m2
+    psadbw    m3, m4
+    paddw     m0, m3
+%endif
+    sub       hd, 2
+
+.loop
+    lea    pix1q, [pix1q + 2*lsizeq]
+%if %1 == mmsize
+    mova      m1, [pix1q]
+    psadbw    m2, m1
+    paddw     m0, m2
+    mova      m2, [pix1q+lsizeq]
+    psadbw    m1, m2
+    paddw     m0, m1
+%else
+    mova      m1, [pix1q]
+    mova      m3, [pix1q+8]
+    psadbw    m2, m1
+    psadbw    m4, m3
+    paddw     m0, m2
+    paddw     m0, m4
+    mova      m2, [pix1q+lsizeq]
+    mova      m4, [pix1q+lsizeq+8]
+    psadbw    m1, m2
+    psadbw    m3, m4
+    paddw     m0, m1
+    paddw     m0, m3
+%endif
+    sub       hd, 2
+    jg     .loop
+
+%if mmsize == 16
+    pshufd m1, m0, 0xe
+    paddd  m0, m1
+%endif
+    movd eax, m0
+    RET
+%endmacro
+
+INIT_MMX mmxext
+VSAD_INTRA 8
+VSAD_INTRA 16
+INIT_XMM sse2
+VSAD_INTRA 16
+
+;---------------------------------------------------------------------
+;int ff_vsad_approx(MpegEncContext *v, uint8_t *pix1, uint8_t *pix2,
+;                   int line_size, int h);
+;---------------------------------------------------------------------
+; %1 = 8/16
+%macro VSAD_APPROX 1
+cglobal vsad%1_approx, 5, 5, 5, v, pix1, pix2, lsize, h
+    mova   m1, [pb_80]
+    mova   m0, [pix1q]
+%if %1 == mmsize ; vsad8_mmxext, vsad16_sse2
+    mova   m4, [pix1q+lsizeq]
+%if mmsize == 16
+    movu   m3, [pix2q]
+    movu   m2, [pix2q+lsizeq]
+    psubb  m0, m3
+    psubb  m4, m2
+%else
+    psubb  m0, [pix2q]
+    psubb  m4, [pix2q+lsizeq]
+%endif
+    pxor   m0, m1
+    pxor   m4, m1
+    psadbw m0, m4
+%else ; vsad16_mmxext
+    mova   m3, [pix1q+8]
+    psubb  m0, [pix2q]
+    psubb  m3, [pix2q+8]
+    pxor   m0, m1
+    pxor   m3, m1
+    mova   m4, [pix1q+lsizeq]
+    mova   m5, [pix1q+lsizeq+8]
+    psubb  m4, [pix2q+lsizeq]
+    psubb  m5, [pix2q+lsizeq+8]
+    pxor   m4, m1
+    pxor   m5, m1
+    psadbw m0, m4
+    psadbw m3, m5
+    paddw  m0, m3
+%endif
+    sub    hd, 2
+
+.loop
+    lea pix1q, [pix1q + 2*lsizeq]
+    lea pix2q, [pix2q + 2*lsizeq]
+    mova   m2, [pix1q]
+%if %1 == mmsize ; vsad8_mmxext, vsad16_sse2
+%if mmsize == 16
+    movu   m3, [pix2q]
+    psubb  m2, m3
+%else
+    psubb  m2, [pix2q]
+%endif
+    pxor   m2, m1
+    psadbw m4, m2
+    paddw  m0, m4
+    mova   m4, [pix1q+lsizeq]
+    movu   m3, [pix2q+lsizeq]
+    psubb  m4, m3
+    pxor   m4, m1
+    psadbw m2, m4
+    paddw  m0, m2
+%else ; vsad16_mmxext
+    mova   m3, [pix1q+8]
+    psubb  m2, [pix2q]
+    psubb  m3, [pix2q+8]
+    pxor   m2, m1
+    pxor   m3, m1
+    psadbw m4, m2
+    psadbw m5, m3
+    paddw  m0, m4
+    paddw  m0, m5
+    mova   m4, [pix1q+lsizeq]
+    mova   m5, [pix1q+lsizeq+8]
+    psubb  m4, [pix2q+lsizeq]
+    psubb  m5, [pix2q+lsizeq+8]
+    pxor   m4, m1
+    pxor   m5, m1
+    psadbw m2, m4
+    psadbw m3, m5
+    paddw  m0, m2
+    paddw  m0, m3
+%endif
+    sub    hd, 2
+    jg  .loop
+
+%if mmsize == 16
+    pshufd m1, m0, 0xe
+    paddd  m0, m1
+%endif
+    movd  eax, m0
+    RET
+%endmacro
+
+INIT_MMX mmxext
+VSAD_APPROX 8
+VSAD_APPROX 16
+INIT_XMM sse2
+VSAD_APPROX 16
diff --git a/libavcodec/x86/me_cmp_init.c b/libavcodec/x86/me_cmp_init.c
index cb47d631f9..6dc59f5aa4 100644
--- a/libavcodec/x86/me_cmp_init.c
+++ b/libavcodec/x86/me_cmp_init.c
@@ -65,6 +65,18 @@ int ff_sad16_approx_xy2_mmxext(MpegEncContext *v, uint8_t *pix1, uint8_t *pix2,
                                int stride, int h);
 int ff_sad16_approx_xy2_sse2(MpegEncContext *v, uint8_t *pix1, uint8_t *pix2,
                              int stride, int h);
+int ff_vsad_intra8_mmxext(MpegEncContext *v, uint8_t *pix1, uint8_t *pix2,
+                          int line_size, int h);
+int ff_vsad_intra16_mmxext(MpegEncContext *v, uint8_t *pix1, uint8_t *pix2,
+                           int line_size, int h);
+int ff_vsad_intra16_sse2(MpegEncContext *v, uint8_t *pix1, uint8_t *pix2,
+                         int line_size, int h);
+int ff_vsad8_approx_mmxext(MpegEncContext *v, uint8_t *pix1, uint8_t *pix2,
+                    int line_size, int h);
+int ff_vsad16_approx_mmxext(MpegEncContext *v, uint8_t *pix1, uint8_t *pix2,
+                     int line_size, int h);
+int ff_vsad16_approx_sse2(MpegEncContext *v, uint8_t *pix1, uint8_t *pix2,
+                   int line_size, int h);
 
 #define hadamard_func(cpu)                                              \
     int ff_hadamard8_diff_ ## cpu(MpegEncContext *s, uint8_t *src1,     \
@@ -177,49 +189,6 @@ static int vsad_intra16_mmx(MpegEncContext *v, uint8_t *pix, uint8_t *dummy,
 }
 #undef SUM
 
-static int vsad_intra16_mmxext(MpegEncContext *v, uint8_t *pix, uint8_t *dummy,
-                               int line_size, int h)
-{
-    int tmp;
-
-    av_assert2((((int) pix) & 7) == 0);
-    av_assert2((line_size & 7) == 0);
-
-#define SUM(in0, in1, out0, out1)               \
-    "movq (%0), " #out0 "\n"                    \
-    "movq 8(%0), " #out1 "\n"                   \
-    "add %2, %0\n"                              \
-    "psadbw " #out0 ", " #in0 "\n"              \
-    "psadbw " #out1 ", " #in1 "\n"              \
-    "paddw " #in1 ", " #in0 "\n"                \
-    "paddw " #in0 ", %%mm6\n"
-
-    __asm__ volatile (
-        "movl %3, %%ecx\n"
-        "pxor %%mm6, %%mm6\n"
-        "pxor %%mm7, %%mm7\n"
-        "movq (%0), %%mm0\n"
-        "movq 8(%0), %%mm1\n"
-        "add %2, %0\n"
-        "jmp 2f\n"
-        "1:\n"
-
-        SUM(%%mm4, %%mm5, %%mm0, %%mm1)
-        "2:\n"
-        SUM(%%mm0, %%mm1, %%mm4, %%mm5)
-
-        "subl $2, %%ecx\n"
-        "jnz 1b\n"
-
-        "movd %%mm6, %1\n"
-        : "+r" (pix), "=r" (tmp)
-        : "r" ((x86_reg) line_size), "m" (h)
-        : "%ecx");
-
-    return tmp;
-}
-#undef SUM
-
 static int vsad16_mmx(MpegEncContext *v, uint8_t *pix1, uint8_t *pix2,
                       int line_size, int h)
 {
@@ -301,68 +270,6 @@ static int vsad16_mmx(MpegEncContext *v, uint8_t *pix1, uint8_t *pix2,
 }
 #undef SUM
 
-static int vsad16_mmxext(MpegEncContext *v, uint8_t *pix1, uint8_t *pix2,
-                         int line_size, int h)
-{
-    int tmp;
-
-    av_assert2((((int) pix1) & 7) == 0);
-    av_assert2((((int) pix2) & 7) == 0);
-    av_assert2((line_size & 7) == 0);
-
-#define SUM(in0, in1, out0, out1)               \
-    "movq (%0), " #out0 "\n"                    \
-    "movq (%1), %%mm2\n"                        \
-    "movq 8(%0), " #out1 "\n"                   \
-    "movq 8(%1), %%mm3\n"                       \
-    "add %3, %0\n"                              \
-    "add %3, %1\n"                              \
-    "psubb %%mm2, " #out0 "\n"                  \
-    "psubb %%mm3, " #out1 "\n"                  \
-    "pxor %%mm7, " #out0 "\n"                   \
-    "pxor %%mm7, " #out1 "\n"                   \
-    "psadbw " #out0 ", " #in0 "\n"              \
-    "psadbw " #out1 ", " #in1 "\n"              \
-    "paddw " #in1 ", " #in0 "\n"                \
-    "paddw " #in0 ", %%mm6\n    "
-
-    __asm__ volatile (
-        "movl %4, %%ecx\n"
-        "pxor %%mm6, %%mm6\n"
-        "pcmpeqw %%mm7, %%mm7\n"
-        "psllw $15, %%mm7\n"
-        "packsswb %%mm7, %%mm7\n"
-        "movq (%0), %%mm0\n"
-        "movq (%1), %%mm2\n"
-        "movq 8(%0), %%mm1\n"
-        "movq 8(%1), %%mm3\n"
-        "add %3, %0\n"
-        "add %3, %1\n"
-        "psubb %%mm2, %%mm0\n"
-        "psubb %%mm3, %%mm1\n"
-        "pxor %%mm7, %%mm0\n"
-        "pxor %%mm7, %%mm1\n"
-        "jmp 2f\n"
-        "1:\n"
-
-        SUM(%%mm4, %%mm5, %%mm0, %%mm1)
-        "2:\n"
-        SUM(%%mm0, %%mm1, %%mm4, %%mm5)
-
-        "subl $2, %%ecx\n"
-        "jnz 1b\n"
-
-        "movd %%mm6, %2\n"
-        : "+r" (pix1), "+r" (pix2), "=r" (tmp)
-        : "r" ((x86_reg) line_size), "m" (h)
-        : "%ecx");
-
-    return tmp;
-}
-#undef SUM
-
-
-
 DECLARE_ASM_CONST(8, uint64_t, round_tab)[3] = {
     0x0000000000000000ULL,
     0x0001000100010001ULL,
@@ -667,14 +574,6 @@ av_cold void ff_me_cmp_init_x86(MECmpContext *c, AVCodecContext *avctx)
         }
     }
 
-    if (INLINE_MMXEXT(cpu_flags)) {
-        c->vsad[4] = vsad_intra16_mmxext;
-
-        if (!(avctx->flags & CODEC_FLAG_BITEXACT)) {
-            c->vsad[0] = vsad16_mmxext;
-        }
-    }
-
 #endif /* HAVE_INLINE_ASM */
 
     if (EXTERNAL_MMX(cpu_flags)) {
@@ -704,9 +603,15 @@ av_cold void ff_me_cmp_init_x86(MECmpContext *c, AVCodecContext *avctx)
         c->pix_abs[1][1] = ff_sad8_x2_mmxext;
         c->pix_abs[1][2] = ff_sad8_y2_mmxext;
 
+        c->vsad[4] = ff_vsad_intra16_mmxext;
+        c->vsad[5] = ff_vsad_intra8_mmxext;
+
         if (!(avctx->flags & CODEC_FLAG_BITEXACT)) {
             c->pix_abs[0][3] = ff_sad16_approx_xy2_mmxext;
             c->pix_abs[1][3] = ff_sad8_approx_xy2_mmxext;
+
+            c->vsad[0] = ff_vsad16_approx_mmxext;
+            c->vsad[1] = ff_vsad8_approx_mmxext;
         }
     }
 
@@ -724,8 +629,10 @@ av_cold void ff_me_cmp_init_x86(MECmpContext *c, AVCodecContext *avctx)
             c->pix_abs[0][1] = ff_sad16_x2_sse2;
             c->pix_abs[0][2] = ff_sad16_y2_sse2;
 
+            c->vsad[4]       = ff_vsad_intra16_sse2;
             if (!(avctx->flags & CODEC_FLAG_BITEXACT)) {
                 c->pix_abs[0][3] = ff_sad16_approx_xy2_sse2;
+                c->vsad[0]       = ff_vsad16_approx_sse2;
             }
         }
     }