avcodec/x86/fpel: Port ff_put_pixels8_mmx() to SSE2

This has the advantage of not violating the ABI by using
MMX registers without issuing emms; it e.g. allows
to remove an emms_c from bink.c.

This commit uses GP registers on Unix64 (there are not
enough volatile registers to do likewise on Win64) which
reduces codesize and is faster on some CPUs.

Signed-off-by: Andreas Rheinhardt <andreas.rheinhardt@outlook.com>

libavcodec/ac3enc.c

@@ -32,7 +32,6 @@
 #include "libavutil/avassert.h"
 #include "libavutil/channel_layout.h"
 #include "libavutil/crc.h"
-#include "libavutil/emms.h"
 #include "libavutil/internal.h"
 #include "libavutil/mem.h"
 #include "libavutil/mem_internal.h"
@@ -945,8 +944,6 @@ static void ac3_process_exponents(AC3EncodeContext *s)
     compute_exp_strategy(s);
 
     encode_exponents(s);
-
-    emms_c();
 }
 
 

libavcodec/bink.c

@@ -21,7 +21,6 @@
  */
 
 #include "libavutil/attributes.h"
-#include "libavutil/emms.h"
 #include "libavutil/imgutils.h"
 #include "libavutil/mem.h"
 #include "libavutil/mem_internal.h"
@@ -1297,7 +1296,6 @@ static int decode_frame(AVCodecContext *avctx, AVFrame *frame,
         if (get_bits_count(&gb) >= bits_count)
             break;
     }
-    emms_c();
 
     if (c->version > 'b') {
         if ((ret = av_frame_replace(c->last, frame)) < 0)

libavcodec/x86/cavsdsp.c

@@ -46,13 +46,6 @@ static void cavs_idct8_add_sse2(uint8_t *dst, int16_t *block, ptrdiff_t stride)
 
 #endif /* HAVE_SSE2_EXTERNAL */
 
-static av_cold void cavsdsp_init_mmx(CAVSDSPContext *c)
-{
-#if HAVE_MMX_EXTERNAL
-    c->put_cavs_qpel_pixels_tab[1][0] = ff_put_pixels8x8_mmx;
-#endif /* HAVE_MMX_EXTERNAL */
-}
-
 #if HAVE_SSE2_EXTERNAL
 #define DEF_QPEL(OPNAME) \
     void ff_ ## OPNAME ## _cavs_qpel8_mc20_sse2(uint8_t *dst, const uint8_t *src, ptrdiff_t stride);     \
@@ -98,9 +91,6 @@ av_cold void ff_cavsdsp_init_x86(CAVSDSPContext *c)
 {
     av_unused int cpu_flags = av_get_cpu_flags();
 
-    if (X86_MMX(cpu_flags))
-        cavsdsp_init_mmx(c);
-
 #if HAVE_MMX_EXTERNAL
     if (EXTERNAL_MMXEXT(cpu_flags)) {
         c->avg_cavs_qpel_pixels_tab[1][0] = ff_avg_pixels8x8_mmxext;
@@ -113,6 +103,7 @@ av_cold void ff_cavsdsp_init_x86(CAVSDSPContext *c)
         c->put_cavs_qpel_pixels_tab[0][ 4] = put_cavs_qpel16_mc01_sse2;
         c->put_cavs_qpel_pixels_tab[0][ 8] = put_cavs_qpel16_mc02_sse2;
         c->put_cavs_qpel_pixels_tab[0][12] = put_cavs_qpel16_mc03_sse2;
+        c->put_cavs_qpel_pixels_tab[1][ 0] = ff_put_pixels8x8_sse2;
         c->put_cavs_qpel_pixels_tab[1][ 2] = ff_put_cavs_qpel8_mc20_sse2;
         c->put_cavs_qpel_pixels_tab[1][ 4] = put_cavs_qpel8_mc01_sse2;
         c->put_cavs_qpel_pixels_tab[1][ 8] = ff_put_cavs_qpel8_mc02_sse2;

libavcodec/x86/fpel.asm

@@ -27,7 +27,7 @@ SECTION .text
 
 ; void ff_put/avg_pixels(uint8_t *block, const uint8_t *pixels,
 ;                        ptrdiff_t line_size, int h)
-%macro OP_PIXELS 2
+%macro OP_PIXELS 2-3 0
 %if %2 == mmsize/2
 %define LOAD movh
 %define SAVE movh
@@ -35,14 +35,25 @@ SECTION .text
 %define LOAD movu
 %define SAVE mova
 %endif
-cglobal %1_pixels%2x%2, 3,5,4
+cglobal %1_pixels%2x%2, 3,5+4*%3,%3 ? 4 : 0
     mov         r3d, %2
     jmp         %1_pixels%2_after_prologue
 
-cglobal %1_pixels%2, 4,5,4
+cglobal %1_pixels%2, 4,5+4*%3,%3 ? 4 : 0
 %1_pixels%2_after_prologue:
     lea          r4, [r2*3]
 .loop:
+%if %3
+; Use GPRs on UNIX64 for put8, but not on Win64 due to a lack of volatile GPRs
+    mov         r5q, [r1]
+    mov         r6q, [r1+r2]
+    mov         r7q, [r1+r2*2]
+    mov         r8q, [r1+r4]
+    mov        [r0], r5q
+    mov     [r0+r2], r6q
+    mov   [r0+r2*2], r7q
+    mov     [r0+r4], r8q
+%else
     LOAD         m0, [r1]
     LOAD         m1, [r1+r2]
     LOAD         m2, [r1+r2*2]
@@ -57,6 +68,7 @@ cglobal %1_pixels%2, 4,5,4
     SAVE    [r0+r2], m1
     SAVE  [r0+r2*2], m2
     SAVE    [r0+r4], m3
+%endif
     sub         r3d, 4
     lea          r1, [r1+r2*4]
     lea          r0, [r0+r2*4]
@@ -64,12 +76,10 @@ cglobal %1_pixels%2, 4,5,4
     RET
 %endmacro
 
-INIT_MMX mmx
-OP_PIXELS put, 8
-
 INIT_MMX mmxext
 OP_PIXELS avg, 8
 
 INIT_XMM sse2
+OP_PIXELS put, 8, UNIX64
 OP_PIXELS put, 16
 OP_PIXELS avg, 16

libavcodec/x86/fpel.h

@@ -30,10 +30,10 @@ void ff_avg_pixels16_sse2(uint8_t *block, const uint8_t *pixels,
                           ptrdiff_t line_size, int h);
 void ff_avg_pixels16x16_sse2(uint8_t *block, const uint8_t *pixels,
                              ptrdiff_t line_size);
-void ff_put_pixels8_mmx(uint8_t *block, const uint8_t *pixels,
+void ff_put_pixels8_sse2(uint8_t *block, const uint8_t *pixels,
-                        ptrdiff_t line_size, int h);
+                         ptrdiff_t line_size, int h);
-void ff_put_pixels8x8_mmx(uint8_t *block, const uint8_t *pixels,
+void ff_put_pixels8x8_sse2(uint8_t *block, const uint8_t *pixels,
-                          ptrdiff_t line_size);
+                           ptrdiff_t line_size);
 void ff_put_pixels16_sse2(uint8_t *block, const uint8_t *pixels,
                           ptrdiff_t line_size, int h);
 void ff_put_pixels16x16_sse2(uint8_t *block, const uint8_t *pixels,

libavcodec/x86/hpeldsp_init.c

@@ -74,14 +74,6 @@ void ff_avg_pixels8_x2_mmxext(uint8_t *block, const uint8_t *pixels,
 void ff_avg_pixels8_y2_mmxext(uint8_t *block, const uint8_t *pixels,
                               ptrdiff_t line_size, int h);
 
-static void hpeldsp_init_mmx(HpelDSPContext *c, int flags)
-{
-#if HAVE_MMX_EXTERNAL
-    c->put_no_rnd_pixels_tab[1][0] =
-    c->put_pixels_tab[1][0] = ff_put_pixels8_mmx;
-#endif
-}
-
 static void hpeldsp_init_mmxext(HpelDSPContext *c, int flags)
 {
 #if HAVE_MMXEXT_EXTERNAL
@@ -115,6 +107,9 @@ static void hpeldsp_init_sse2(HpelDSPContext *c, int flags)
     c->put_no_rnd_pixels_tab[0][2] = ff_put_no_rnd_pixels16_y2_sse2;
     c->put_no_rnd_pixels_tab[0][3] = ff_put_no_rnd_pixels16_xy2_sse2;
 
+    c->put_no_rnd_pixels_tab[1][0] =
+    c->put_pixels_tab[1][0]        = ff_put_pixels8_sse2;
+
     c->avg_pixels_tab[0][0]        = ff_avg_pixels16_sse2;
     c->avg_pixels_tab[0][1]        = ff_avg_pixels16_x2_sse2;
     c->avg_pixels_tab[0][2]        = ff_avg_pixels16_y2_sse2;
@@ -143,9 +138,6 @@ av_cold void ff_hpeldsp_init_x86(HpelDSPContext *c, int flags)
 {
     int cpu_flags = av_get_cpu_flags();
 
-    if (EXTERNAL_MMX(cpu_flags))
-        hpeldsp_init_mmx(c, flags);
-
     if (EXTERNAL_MMXEXT(cpu_flags))
         hpeldsp_init_mmxext(c, flags);
 

libavcodec/x86/me_cmp.asm

@@ -23,10 +23,9 @@
 
 %include "libavutil/x86/x86util.asm"
 
-SECTION_RODATA
-
 cextern pb_1
 cextern pb_80
+cextern pw_2
 
 SECTION .text
 
@@ -468,22 +467,25 @@ HF_NOISE 16
 ;---------------------------------------------------------------------------------------
 ;int ff_sad_<opt>(MPVEncContext *v, const uint8_t *pix1, const uint8_t *pix2, ptrdiff_t stride, int h);
 ;---------------------------------------------------------------------------------------
-;%1 = 8/16
+;%1 = 8/16, %2 = a/u (whether pix1 is aligned or not)
-%macro SAD 1
+%macro SAD 1-2
-cglobal sad%1, 5, 5, 3, v, pix1, pix2, stride, h
+%ifidn %2, u
+cglobal sad%1u, 5, 5, 5, v, pix1, pix2, stride, h
+%else
+cglobal sad%1,  5, 5, 3, v, pix1, pix2, stride, h
+%endif
     movu      m2, [pix2q]
     movu      m1, [pix2q+strideq]
+%ifidn %2, u
+    movu      m0, [pix1q]
+    movu      m3, [pix1q+strideq]
+    psadbw    m2, m0
+    psadbw    m1, m3
+%else
     psadbw    m2, [pix1q]
     psadbw    m1, [pix1q+strideq]
-    paddw     m2, m1
-%if %1 != mmsize
-    movu      m0, [pix2q+8]
-    movu      m1, [pix2q+strideq+8]
-    psadbw    m0, [pix1q+8]
-    psadbw    m1, [pix1q+strideq+8]
-    paddw     m2, m0
-    paddw     m2, m1
 %endif
+    paddw     m2, m1
     sub       hd, 2
 
 align 16
@@ -492,18 +494,17 @@ align 16
     lea    pix2q, [pix2q+strideq*2]
     movu      m0, [pix2q]
     movu      m1, [pix2q+strideq]
+%ifidn %2, u
+    movu      m3, [pix1q]
+    movu      m4, [pix1q+strideq]
+    psadbw    m0, m3
+    psadbw    m1, m4
+%else
     psadbw    m0, [pix1q]
     psadbw    m1, [pix1q+strideq]
-    paddw     m2, m0
-    paddw     m2, m1
-%if %1 != mmsize
-    movu      m0, [pix2q+8]
-    movu      m1, [pix2q+strideq+8]
-    psadbw    m0, [pix1q+8]
-    psadbw    m1, [pix1q+strideq+8]
-    paddw     m2, m0
-    paddw     m2, m1
 %endif
+    paddw     m2, m0
+    paddw     m2, m1
     sub       hd, 2
     jg .loop
 %if mmsize == 16
@@ -516,9 +517,9 @@ align 16
 
 INIT_MMX mmxext
 SAD 8
-SAD 16
 INIT_XMM sse2
 SAD 16
+SAD 16, u
 
 ;------------------------------------------------------------------------------------------
 ;int ff_sad_x2_<opt>(MPVEncContext *v, const uint8_t *pix1, const uint8_t *pix2, ptrdiff_t stride, int h);
@@ -540,16 +541,6 @@ cglobal sad%1_x2, 5, 5, 5, v, pix1, pix2, stride, h
     psadbw    m0, [pix1q]
     psadbw    m2, [pix1q+strideq]
     paddw     m0, m2
-%if %1 != mmsize
-    movu      m1, [pix2q+8]
-    movu      m2, [pix2q+strideq+8]
-    pavgb     m1, [pix2q+9]
-    pavgb     m2, [pix2q+strideq+9]
-    psadbw    m1, [pix1q+8]
-    psadbw    m2, [pix1q+strideq+8]
-    paddw     m0, m1
-    paddw     m0, m2
-%endif
     sub       hd, 2
 
 align 16
@@ -571,16 +562,6 @@ align 16
     psadbw    m2, [pix1q+strideq]
     paddw     m0, m1
     paddw     m0, m2
-%if %1 != mmsize
-    movu      m1, [pix2q+8]
-    movu      m2, [pix2q+strideq+8]
-    pavgb     m1, [pix2q+9]
-    pavgb     m2, [pix2q+strideq+9]
-    psadbw    m1, [pix1q+8]
-    psadbw    m2, [pix1q+strideq+8]
-    paddw     m0, m1
-    paddw     m0, m2
-%endif
     sub       hd, 2
     jg .loop
 %if mmsize == 16
@@ -593,7 +574,6 @@ align 16
 
 INIT_MMX mmxext
 SAD_X2 8
-SAD_X2 16
 INIT_XMM sse2
 SAD_X2 16
 
@@ -612,18 +592,6 @@ cglobal sad%1_y2, 5, 5, 4, v, pix1, pix2, stride, h
     psadbw    m0, [pix1q+strideq]
     paddw     m0, m1
     mova      m1, m3
-%if %1 != mmsize
-    movu      m4, [pix2q+8]
-    movu      m5, [pix2q+strideq+8]
-    movu      m6, [pix2q+2*strideq+8]
-    pavgb     m4, m5
-    pavgb     m5, m6
-    psadbw    m4, [pix1q+8]
-    psadbw    m5, [pix1q+strideq+8]
-    paddw     m0, m4
-    paddw     m0, m5
-    mova      m4, m6
-%endif
     add    pix2q, strideq
     sub       hd, 2
 
@@ -640,17 +608,6 @@ align 16
     paddw     m0, m1
     paddw     m0, m2
     mova      m1, m3
-%if %1 != mmsize
-    movu      m5, [pix2q+8]
-    movu      m6, [pix2q+strideq+8]
-    pavgb     m4, m5
-    pavgb     m5, m6
-    psadbw    m4, [pix1q+8]
-    psadbw    m5, [pix1q+strideq+8]
-    paddw     m0, m4
-    paddw     m0, m5
-    mova      m4, m6
-%endif
     sub       hd, 2
     jg .loop
 %if mmsize == 16
@@ -663,10 +620,105 @@ align 16
 
 INIT_MMX mmxext
 SAD_Y2 8
-SAD_Y2 16
 INIT_XMM sse2
 SAD_Y2 16
 
+;------------------------------------------------------------------------------------------
+;int ff_sad_xy2_<opt>(MPVEncContext *v, const uint8_t *pix1, const uint8_t *pix2, ptrdiff_t stride, int h);
+;------------------------------------------------------------------------------------------
+
+;%1 = 8/16, %2 = aligned mov, %3 = unaligned mov
+%macro SAD_XY2 3
+cglobal sad%1_xy2, 5, 5, mmsize == 16 ? 8 + ARCH_X86_64 : 7, v, pix1, pix2, stride, h
+    mov%3     m2, [pix2q]
+    mov%3     m3, [pix2q+1]
+%if %1 == mmsize
+%if ARCH_X86_64
+    mova      m8, [pw_2]
+    %define PW_2 m8
+%else
+    %define PW_2 [pw_2]
+%endif
+%else ; %1 != mmsize
+    mova      m6, [pw_2]
+    %define PW_2 m6
+%endif
+    pxor      m1, m1
+    add    pix2q, strideq
+%if %1 != mmsize/2
+    mova      m6, m2
+    mova      m7, m3
+    punpckhbw m6, m1
+    punpckhbw m7, m1
+    paddw     m6, m7
+%endif
+    punpcklbw m2, m1
+    punpcklbw m3, m1
+    paddw     m2, m3
+    mova      m0, m1
+
+.loop:
+    mov%3     m3, [pix2q]
+    mov%3     m4, [pix2q+1]
+%if %1 != mmsize/2
+    mova      m5, m3
+    mova      m7, m4
+    punpckhbw m5, m1
+    punpckhbw m7, m1
+    paddw     m7, m5
+    paddw     m7, PW_2
+    paddw     m6, m7
+    psraw     m6, 2
+%endif
+    mov%2     m5, [pix1q]
+    punpcklbw m3, m1
+    punpcklbw m4, m1
+    paddw     m3, m4
+    paddw     m3, PW_2
+    paddw     m2, m3
+    psraw     m2, 2
+    packuswb  m2, m6
+    psadbw    m2, m5
+    paddw     m0, m2
+
+    mov%3     m2, [pix2q+strideq]
+    mov%3     m4, [pix2q+strideq+1]
+%if %1 != mmsize/2
+    mova      m5, m2
+    mova      m6, m4
+    punpckhbw m5, m1
+    punpckhbw m6, m1
+    paddw     m6, m5
+    paddw     m7, m6
+    psraw     m7, 2
+%endif
+    mov%2     m5, [pix1q+strideq]
+    punpcklbw m2, m1
+    punpcklbw m4, m1
+    paddw     m2, m4
+    paddw     m3, m2
+    psraw     m3, 2
+    packuswb  m3, m7
+    psadbw    m3, m5
+    paddw     m0, m3
+
+    sub       hd, 2
+    lea    pix1q, [pix1q+2*strideq]
+    lea    pix2q, [pix2q+2*strideq]
+    jnz    .loop
+
+%if %1 == 16
+    movhlps   m1, m0
+    paddw     m0, m1
+%endif
+    movd     eax, m0
+    RET
+%endmacro
+
+INIT_XMM sse2
+SAD_XY2  8, h, h
+SAD_XY2 16, a, u
+
 ;-------------------------------------------------------------------------------------------
 ;int ff_sad_approx_xy2_<opt>(MPVEncContext *v, const uint8_t *pix1, const uint8_t *pix2, ptrdiff_t stride, int h);
 ;-------------------------------------------------------------------------------------------
@@ -696,22 +748,6 @@ cglobal sad%1_approx_xy2, 5, 5, 7, v, pix1, pix2, stride, h
     psadbw    m0, [pix1q+strideq]
     paddw     m0, m1
     mova      m1, m3
-%if %1 != mmsize
-    movu      m5, [pix2q+8]
-    movu      m6, [pix2q+strideq+8]
-    movu      m7, [pix2q+2*strideq+8]
-    pavgb     m5, [pix2q+1+8]
-    pavgb     m6, [pix2q+strideq+1+8]
-    pavgb     m7, [pix2q+2*strideq+1+8]
-    psubusb   m6, m4
-    pavgb     m5, m6
-    pavgb     m6, m7
-    psadbw    m5, [pix1q+8]
-    psadbw    m6, [pix1q+strideq+8]
-    paddw     m0, m5
-    paddw     m0, m6
-    mova      m5, m7
-%endif
     add    pix2q, strideq
     sub       hd, 2
 
@@ -738,20 +774,6 @@ align 16
     paddw     m0, m1
     paddw     m0, m2
     mova      m1, m3
-%if %1 != mmsize
-    movu      m6, [pix2q+8]
-    movu      m7, [pix2q+strideq+8]
-    pavgb     m6, [pix2q+8+1]
-    pavgb     m7, [pix2q+strideq+8+1]
-    psubusb   m6, m4
-    pavgb     m5, m6
-    pavgb     m6, m7
-    psadbw    m5, [pix1q+8]
-    psadbw    m6, [pix1q+strideq+8]
-    paddw     m0, m5
-    paddw     m0, m6
-    mova      m5, m7
-%endif
     sub       hd, 2
     jg .loop
 %if mmsize == 16
@@ -764,7 +786,6 @@ align 16
 
 INIT_MMX mmxext
 SAD_APPROX_XY2 8
-SAD_APPROX_XY2 16
 INIT_XMM sse2
 SAD_APPROX_XY2 16
 
@@ -772,46 +793,26 @@ SAD_APPROX_XY2 16
 ;int ff_vsad_intra(MPVEncContext *v, const uint8_t *pix1, const uint8_t *pix2,
 ;                  ptrdiff_t line_size, int h);
 ;--------------------------------------------------------------------
-; %1 = 8/16
+; %1 = 8/16, %2 = a/u (whether pix1 is aligned or not)
-%macro VSAD_INTRA 1
+%macro VSAD_INTRA 2
-cglobal vsad_intra%1, 5, 5, 3, v, pix1, pix2, lsize, h
+%ifidn %2, u
-    mova      m0, [pix1q]
+cglobal vsad_intra%1u, 5, 5, 3, v, pix1, pix2, lsize, h
-%if %1 == mmsize
-    mova      m2, [pix1q+lsizeq]
-    psadbw    m0, m2
 %else
-    mova      m2, [pix1q+lsizeq]
+cglobal vsad_intra%1,  5, 5, 3, v, pix1, pix2, lsize, h
-    mova      m3, [pix1q+8]
-    mova      m4, [pix1q+lsizeq+8]
-    psadbw    m0, m2
-    psadbw    m3, m4
-    paddw     m0, m3
 %endif
+    mov%2     m0, [pix1q]
+    mov%2     m2, [pix1q+lsizeq]
+    psadbw    m0, m2
     sub       hd, 2
 
 .loop:
     lea    pix1q, [pix1q + 2*lsizeq]
-%if %1 == mmsize
+    mov%2     m1, [pix1q]
-    mova      m1, [pix1q]
     psadbw    m2, m1
     paddw     m0, m2
-    mova      m2, [pix1q+lsizeq]
+    mov%2     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
 
@@ -824,22 +825,25 @@ cglobal vsad_intra%1, 5, 5, 3, v, pix1, pix2, lsize, h
 %endmacro
 
 INIT_MMX mmxext
-VSAD_INTRA 8
+VSAD_INTRA  8, a
-VSAD_INTRA 16
 INIT_XMM sse2
-VSAD_INTRA 16
+VSAD_INTRA 16, a
+VSAD_INTRA 16, u
 
 ;---------------------------------------------------------------------
 ;int ff_vsad_approx(MPVEncContext *v, const uint8_t *pix1, const uint8_t *pix2,
 ;                   ptrdiff_t line_size, int h);
 ;---------------------------------------------------------------------
-; %1 = 8/16
+; %1 = 8/16, %2 = a/u (whether pix1 is aligned or not)
-%macro VSAD_APPROX 1
+%macro VSAD_APPROX 2
-cglobal vsad%1_approx, 5, 5, 5, v, pix1, pix2, lsize, h
+%ifidn %2, u
+cglobal vsad%1u_approx, 5, 5, 5, v, pix1, pix2, lsize, h
+%else
+cglobal vsad%1_approx,  5, 5, 5, v, pix1, pix2, lsize, h
+%endif
     mova   m1, [pb_80]
-    mova   m0, [pix1q]
+    mov%2  m0, [pix1q]
-%if %1 == mmsize ; vsad8_mmxext, vsad16_sse2
+    mov%2  m4, [pix1q+lsizeq]
-    mova   m4, [pix1q+lsizeq]
 %if mmsize == 16
     movu   m3, [pix2q]
     movu   m2, [pix2q+lsizeq]
@@ -852,29 +856,12 @@ cglobal vsad%1_approx, 5, 5, 5, v, pix1, pix2, lsize, h
     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]
+    mov%2  m2, [pix1q]
-%if %1 == mmsize ; vsad8_mmxext, vsad16_sse2
 %if mmsize == 16
     movu   m3, [pix2q]
     psubb  m2, m3
@@ -884,33 +871,12 @@ cglobal vsad%1_approx, 5, 5, 5, v, pix1, pix2, lsize, h
     pxor   m2, m1
     psadbw m4, m2
     paddw  m0, m4
-    mova   m4, [pix1q+lsizeq]
+    mov%2  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
 
@@ -923,7 +889,7 @@ cglobal vsad%1_approx, 5, 5, 5, v, pix1, pix2, lsize, h
 %endmacro
 
 INIT_MMX mmxext
-VSAD_APPROX 8
+VSAD_APPROX 8,  a
-VSAD_APPROX 16
 INIT_XMM sse2
-VSAD_APPROX 16
+VSAD_APPROX 16, a
+VSAD_APPROX 16, u

libavcodec/x86/me_cmp_init.c

@@ -24,8 +24,6 @@
 
 #include "libavutil/attributes.h"
 #include "libavutil/cpu.h"
-#include "libavutil/mem_internal.h"
-#include "libavutil/x86/asm.h"
 #include "libavutil/x86/cpu.h"
 #include "libavcodec/me_cmp.h"
 #include "libavcodec/mpegvideoenc.h"
@@ -42,40 +40,38 @@ int ff_hf_noise8_mmx(const uint8_t *pix1, ptrdiff_t stride, int h);
 int ff_hf_noise16_mmx(const uint8_t *pix1, ptrdiff_t stride, int h);
 int ff_sad8_mmxext(MPVEncContext *v, const uint8_t *pix1, const uint8_t *pix2,
                    ptrdiff_t stride, int h);
-int ff_sad16_mmxext(MPVEncContext *v, const uint8_t *pix1, const uint8_t *pix2,
-                    ptrdiff_t stride, int h);
 int ff_sad16_sse2(MPVEncContext *v, const uint8_t *pix1, const uint8_t *pix2,
                   ptrdiff_t stride, int h);
+int ff_sad16u_sse2(MPVEncContext *v, const uint8_t *pix1, const uint8_t *pix2,
+                   ptrdiff_t stride, int h);
 int ff_sad8_x2_mmxext(MPVEncContext *v, const uint8_t *pix1, const uint8_t *pix2,
                       ptrdiff_t stride, int h);
-int ff_sad16_x2_mmxext(MPVEncContext *v, const uint8_t *pix1, const uint8_t *pix2,
-                       ptrdiff_t stride, int h);
 int ff_sad16_x2_sse2(MPVEncContext *v, const uint8_t *pix1, const uint8_t *pix2,
                      ptrdiff_t stride, int h);
 int ff_sad8_y2_mmxext(MPVEncContext *v, const uint8_t *pix1, const uint8_t *pix2,
                       ptrdiff_t stride, int h);
-int ff_sad16_y2_mmxext(MPVEncContext *v, const uint8_t *pix1, const uint8_t *pix2,
-                       ptrdiff_t stride, int h);
 int ff_sad16_y2_sse2(MPVEncContext *v, const uint8_t *pix1, const uint8_t *pix2,
                      ptrdiff_t stride, int h);
 int ff_sad8_approx_xy2_mmxext(MPVEncContext *v, const uint8_t *pix1, const uint8_t *pix2,
                               ptrdiff_t stride, int h);
-int ff_sad16_approx_xy2_mmxext(MPVEncContext *v, const uint8_t *pix1, const uint8_t *pix2,
+int ff_sad8_xy2_sse2(MPVEncContext *v, const uint8_t *pix1, const uint8_t *pix2,
-                               ptrdiff_t stride, int h);
+                     ptrdiff_t stride, int h);
 int ff_sad16_approx_xy2_sse2(MPVEncContext *v, const uint8_t *pix1, const uint8_t *pix2,
                              ptrdiff_t stride, int h);
+int ff_sad16_xy2_sse2(MPVEncContext *v, const uint8_t *pix1, const uint8_t *pix2,
+                      ptrdiff_t stride, int h);
 int ff_vsad_intra8_mmxext(MPVEncContext *v, const uint8_t *pix1, const uint8_t *pix2,
                           ptrdiff_t stride, int h);
-int ff_vsad_intra16_mmxext(MPVEncContext *v, const uint8_t *pix1, const uint8_t *pix2,
-                           ptrdiff_t stride, int h);
 int ff_vsad_intra16_sse2(MPVEncContext *v, const uint8_t *pix1, const uint8_t *pix2,
                          ptrdiff_t stride, int h);
+int ff_vsad_intra16u_sse2(MPVEncContext *v, const uint8_t *pix1, const uint8_t *pix2,
+                          ptrdiff_t stride, int h);
 int ff_vsad8_approx_mmxext(MPVEncContext *v, const uint8_t *pix1, const uint8_t *pix2,
                     ptrdiff_t stride, int h);
-int ff_vsad16_approx_mmxext(MPVEncContext *v, const uint8_t *pix1, const uint8_t *pix2,
-                     ptrdiff_t stride, int h);
 int ff_vsad16_approx_sse2(MPVEncContext *v, const uint8_t *pix1, const uint8_t *pix2,
                    ptrdiff_t stride, int h);
+int ff_vsad16u_approx_sse2(MPVEncContext *v, const uint8_t *pix1, const uint8_t *pix2,
+                           ptrdiff_t stride, int h);
 
 #define hadamard_func(cpu)                                                       \
     int ff_hadamard8_diff_ ## cpu(MPVEncContext *s, const uint8_t *src1,         \
@@ -121,127 +117,10 @@ static int nsse8_mmx(MPVEncContext *c, const uint8_t *pix1, const uint8_t *pix2,
 
 #endif /* HAVE_X86ASM */
 
-#if HAVE_INLINE_ASM
-
-DECLARE_ASM_CONST(8, uint64_t, round_tab)[3] = {
-    0x0000000000000000ULL,
-    0x0001000100010001ULL,
-    0x0002000200020002ULL,
-};
-
-static inline void sad8_4_mmx(const uint8_t *blk1, const uint8_t *blk2,
-                              ptrdiff_t stride, int h)
-{
-    x86_reg len = -stride * h;
-    __asm__ volatile (
-        "movq  (%1, %%"FF_REG_a"), %%mm0\n\t"
-        "movq 1(%1, %%"FF_REG_a"), %%mm2\n\t"
-        "movq %%mm0, %%mm1              \n\t"
-        "movq %%mm2, %%mm3              \n\t"
-        "punpcklbw %%mm7, %%mm0         \n\t"
-        "punpckhbw %%mm7, %%mm1         \n\t"
-        "punpcklbw %%mm7, %%mm2         \n\t"
-        "punpckhbw %%mm7, %%mm3         \n\t"
-        "paddw %%mm2, %%mm0             \n\t"
-        "paddw %%mm3, %%mm1             \n\t"
-        ".p2align 4                     \n\t"
-        "1:                             \n\t"
-        "movq  (%2, %%"FF_REG_a"), %%mm2\n\t"
-        "movq 1(%2, %%"FF_REG_a"), %%mm4\n\t"
-        "movq %%mm2, %%mm3              \n\t"
-        "movq %%mm4, %%mm5              \n\t"
-        "punpcklbw %%mm7, %%mm2         \n\t"
-        "punpckhbw %%mm7, %%mm3         \n\t"
-        "punpcklbw %%mm7, %%mm4         \n\t"
-        "punpckhbw %%mm7, %%mm5         \n\t"
-        "paddw %%mm4, %%mm2             \n\t"
-        "paddw %%mm5, %%mm3             \n\t"
-        "movq %5, %%mm5                 \n\t"
-        "paddw %%mm2, %%mm0             \n\t"
-        "paddw %%mm3, %%mm1             \n\t"
-        "paddw %%mm5, %%mm0             \n\t"
-        "paddw %%mm5, %%mm1             \n\t"
-        "movq (%3, %%"FF_REG_a"), %%mm4 \n\t"
-        "movq (%3, %%"FF_REG_a"), %%mm5 \n\t"
-        "psrlw $2, %%mm0                \n\t"
-        "psrlw $2, %%mm1                \n\t"
-        "packuswb %%mm1, %%mm0          \n\t"
-        "psubusb %%mm0, %%mm4           \n\t"
-        "psubusb %%mm5, %%mm0           \n\t"
-        "por %%mm4, %%mm0               \n\t"
-        "movq %%mm0, %%mm4              \n\t"
-        "punpcklbw %%mm7, %%mm0         \n\t"
-        "punpckhbw %%mm7, %%mm4         \n\t"
-        "paddw %%mm0, %%mm6             \n\t"
-        "paddw %%mm4, %%mm6             \n\t"
-        "movq  %%mm2, %%mm0             \n\t"
-        "movq  %%mm3, %%mm1             \n\t"
-        "add %4, %%"FF_REG_a"           \n\t"
-        " js 1b                         \n\t"
-        : "+a" (len)
-        : "r" (blk1 - len), "r" (blk1 - len + stride), "r" (blk2 - len),
-          "r" (stride), "m" (round_tab[2]));
-}
-
-static inline int sum_mmx(void)
-{
-    int ret;
-    __asm__ volatile (
-        "movq %%mm6, %%mm0              \n\t"
-        "psrlq $32, %%mm6               \n\t"
-        "paddw %%mm0, %%mm6             \n\t"
-        "movq %%mm6, %%mm0              \n\t"
-        "psrlq $16, %%mm6               \n\t"
-        "paddw %%mm0, %%mm6             \n\t"
-        "movd %%mm6, %0                 \n\t"
-        : "=r" (ret));
-    return ret & 0xFFFF;
-}
-
-#define PIX_SADXY(suf)                                                  \
-static int sad8_xy2_ ## suf(MPVEncContext *v, const uint8_t *blk2,      \
-                            const uint8_t *blk1, ptrdiff_t stride, int h) \
-{                                                                       \
-    __asm__ volatile (                                                  \
-        "pxor %%mm7, %%mm7     \n\t"                                    \
-        "pxor %%mm6, %%mm6     \n\t"                                    \
-        ::);                                                            \
-                                                                        \
-    sad8_4_ ## suf(blk1, blk2, stride, h);                              \
-                                                                        \
-    return sum_ ## suf();                                               \
-}                                                                       \
-                                                                        \
-static int sad16_xy2_ ## suf(MPVEncContext *v, const uint8_t *blk2,     \
-                             const uint8_t *blk1, ptrdiff_t stride, int h) \
-{                                                                       \
-    __asm__ volatile (                                                  \
-        "pxor %%mm7, %%mm7     \n\t"                                    \
-        "pxor %%mm6, %%mm6     \n\t"                                    \
-        ::);                                                            \
-                                                                        \
-    sad8_4_ ## suf(blk1,     blk2,     stride, h);                      \
-    sad8_4_ ## suf(blk1 + 8, blk2 + 8, stride, h);                      \
-                                                                        \
-    return sum_ ## suf();                                               \
-}                                                                       \
-
-PIX_SADXY(mmx)
-
-#endif /* HAVE_INLINE_ASM */
-
 av_cold void ff_me_cmp_init_x86(MECmpContext *c, AVCodecContext *avctx)
 {
     int cpu_flags = av_get_cpu_flags();
 
-#if HAVE_INLINE_ASM
-    if (INLINE_MMX(cpu_flags)) {
-        c->pix_abs[0][3] = sad16_xy2_mmx;
-        c->pix_abs[1][3] = sad8_xy2_mmx;
-    }
-
-#endif /* HAVE_INLINE_ASM */
-
     if (EXTERNAL_MMX(cpu_flags)) {
         c->sse[1]            = ff_sse8_mmx;
 #if HAVE_X86ASM
@@ -256,24 +135,17 @@ av_cold void ff_me_cmp_init_x86(MECmpContext *c, AVCodecContext *avctx)
         c->hadamard8_diff[1] = ff_hadamard8_diff_mmxext;
 #endif
 
-        c->sad[0] = ff_sad16_mmxext;
         c->sad[1] = ff_sad8_mmxext;
 
-        c->pix_abs[0][0] = ff_sad16_mmxext;
-        c->pix_abs[0][1] = ff_sad16_x2_mmxext;
-        c->pix_abs[0][2] = ff_sad16_y2_mmxext;
         c->pix_abs[1][0] = ff_sad8_mmxext;
         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 & AV_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;
         }
     }
@@ -282,21 +154,36 @@ av_cold void ff_me_cmp_init_x86(MECmpContext *c, AVCodecContext *avctx)
         c->sse[0] = ff_sse16_sse2;
         c->sum_abs_dctelem   = ff_sum_abs_dctelem_sse2;
 
+        c->pix_abs[0][0] = ff_sad16_sse2;
+        c->pix_abs[0][1] = ff_sad16_x2_sse2;
+        c->pix_abs[0][2] = ff_sad16_y2_sse2;
+        c->pix_abs[0][3] = ff_sad16_xy2_sse2;
+
 #if HAVE_ALIGNED_STACK
         c->hadamard8_diff[0] = ff_hadamard8_diff16_sse2;
         c->hadamard8_diff[1] = ff_hadamard8_diff_sse2;
 #endif
-        if (!(cpu_flags & AV_CPU_FLAG_SSE2SLOW) && avctx->codec_id != AV_CODEC_ID_SNOW) {
+        if (avctx->codec_id != AV_CODEC_ID_SNOW) {
             c->sad[0]        = ff_sad16_sse2;
-            c->pix_abs[0][0] = ff_sad16_sse2;
-            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 & AV_CODEC_FLAG_BITEXACT)) {
-                c->pix_abs[0][3] = ff_sad16_approx_xy2_sse2;
                 c->vsad[0]       = ff_vsad16_approx_sse2;
             }
+        } else {
+            // Snow does not abide by the alignment requirements
+            // of blk1, so we use special versions without them for it.
+            c->sad[0]        = ff_sad16u_sse2;
+
+            c->vsad[4]       = ff_vsad_intra16u_sse2;
+            if (!(avctx->flags & AV_CODEC_FLAG_BITEXACT)) {
+                c->vsad[0]       = ff_vsad16u_approx_sse2;
+            }
+        }
+        if (avctx->flags & AV_CODEC_FLAG_BITEXACT) {
+            c->pix_abs[1][3] = ff_sad8_xy2_sse2;
+        } else {
+            c->pix_abs[0][3] = ff_sad16_approx_xy2_sse2;
         }
     }
 

libavcodec/x86/qpeldsp_init.c

@@ -521,8 +521,6 @@ av_cold void ff_qpeldsp_init_x86(QpelDSPContext *c)
         SET_QPEL_FUNCS(avg_qpel,        1,  8, mmxext, );
 
         SET_QPEL_FUNCS(put_qpel,        0, 16, mmxext, );
-        c->put_no_rnd_qpel_pixels_tab[1][0] =
-        c->put_qpel_pixels_tab[1][0] = ff_put_pixels8x8_mmx;
         SET_QPEL_FUNCS(put_qpel,        1,  8, mmxext, );
         SET_QPEL_FUNCS(put_no_rnd_qpel, 0, 16, mmxext, );
         SET_QPEL_FUNCS(put_no_rnd_qpel, 1,  8, mmxext, );
@@ -532,6 +530,8 @@ av_cold void ff_qpeldsp_init_x86(QpelDSPContext *c)
     if (EXTERNAL_SSE2(cpu_flags)) {
         c->put_no_rnd_qpel_pixels_tab[0][0] =
         c->put_qpel_pixels_tab[0][0] = ff_put_pixels16x16_sse2;
+        c->put_no_rnd_qpel_pixels_tab[1][0] =
+        c->put_qpel_pixels_tab[1][0] = ff_put_pixels8x8_sse2;
         c->avg_qpel_pixels_tab[0][0] = ff_avg_pixels16x16_sse2;
     }
 #endif

libavcodec/x86/vc1dsp_init.c

@@ -73,7 +73,7 @@ static void vc1_h_loop_filter16_sse4(uint8_t *src, ptrdiff_t stride, int pq)
         ff_ ## OP ## pixels ## DEPTH ## INSN(dst, src, stride, DEPTH);     \
     }
 
-DECLARE_FUNCTION(put_,  8, _mmx)
+DECLARE_FUNCTION(put_,  8, _sse2)
 DECLARE_FUNCTION(avg_,  8, _mmxext)
 DECLARE_FUNCTION(put_, 16, _sse2)
 DECLARE_FUNCTION(avg_, 16, _sse2)
@@ -125,7 +125,6 @@ av_cold void ff_vc1dsp_init_x86(VC1DSPContext *dsp)
     if (EXTERNAL_MMX(cpu_flags)) {
         dsp->put_no_rnd_vc1_chroma_pixels_tab[0] = ff_put_vc1_chroma_mc8_nornd_mmx;
 
-        dsp->put_vc1_mspel_pixels_tab[1][0]      = put_vc1_mspel_mc00_8_mmx;
     }
     if (EXTERNAL_MMXEXT(cpu_flags)) {
         ASSIGN_LF4(mmxext);
@@ -142,6 +141,7 @@ av_cold void ff_vc1dsp_init_x86(VC1DSPContext *dsp)
         ASSIGN_LF816(sse2);
 
         dsp->put_vc1_mspel_pixels_tab[0][0]      = put_vc1_mspel_mc00_16_sse2;
+        dsp->put_vc1_mspel_pixels_tab[1][0]      = put_vc1_mspel_mc00_8_sse2;
         dsp->avg_vc1_mspel_pixels_tab[0][0]      = avg_vc1_mspel_mc00_16_sse2;
     }
     if (EXTERNAL_SSSE3(cpu_flags)) {