Implement a SIMD version of emulated_edge_mc() for x86.

From ~550 cycles (C version) to 170 (SSE/x86-64), 206 (MMX/x86-32)
and 196 (SSE2/x86-32) cycles.

2 files changed, 667 insertions, 3 deletions

diff --git a/libavcodec/x86/dsputil_mmx.c b/libavcodec/x86/dsputil_mmx.c
index 8257b3fa8a..2eb7d85f14 100644
--- a/libavcodec/x86/dsputil_mmx.c
+++ b/libavcodec/x86/dsputil_mmx.c
@@ -1664,8 +1664,80 @@ QPEL_2TAP(avg_,  8, 3dnow)
 static void just_return(void) { return; }
 #endif
 
-static void gmc_mmx(uint8_t *dst, uint8_t *src, int stride, int h, int ox, int oy,
-                    int dxx, int dxy, int dyx, int dyy, int shift, int r, int width, int height){
+#if HAVE_YASM
+typedef void emu_edge_core_func (uint8_t *buf, const uint8_t *src,
+                                 x86_reg linesize, x86_reg start_y,
+                                 x86_reg end_y, x86_reg block_h,
+                                 x86_reg start_x, x86_reg end_x,
+                                 x86_reg block_w);
+extern emu_edge_core_func ff_emu_edge_core_mmx;
+extern emu_edge_core_func ff_emu_edge_core_sse;
+
+static av_always_inline
+void emulated_edge_mc(uint8_t *buf, const uint8_t *src, int linesize,
+                      int block_w, int block_h,
+                      int src_x, int src_y, int w, int h,
+                      emu_edge_core_func *core_fn)
+{
+    int start_y, start_x, end_y, end_x, src_y_add=0;
+
+    if(src_y>= h){
+        src_y_add = h-1-src_y;
+        src_y=h-1;
+    }else if(src_y<=-block_h){
+        src_y_add = 1-block_h-src_y;
+        src_y=1-block_h;
+    }
+    if(src_x>= w){
+        src+= (w-1-src_x);
+        src_x=w-1;
+    }else if(src_x<=-block_w){
+        src+= (1-block_w-src_x);
+        src_x=1-block_w;
+    }
+
+    start_y= FFMAX(0, -src_y);
+    start_x= FFMAX(0, -src_x);
+    end_y= FFMIN(block_h, h-src_y);
+    end_x= FFMIN(block_w, w-src_x);
+    assert(start_x < end_x && block_w > 0);
+    assert(start_y < end_y && block_h > 0);
+
+    // fill in the to-be-copied part plus all above/below
+    src += (src_y_add+start_y)*linesize + start_x;
+    buf += start_x;
+    core_fn(buf, src, linesize, start_y, end_y, block_h, start_x, end_x, block_w);
+}
+
+#if ARCH_X86_32
+static av_noinline
+void emulated_edge_mc_mmx(uint8_t *buf, const uint8_t *src, int linesize,
+                          int block_w, int block_h,
+                          int src_x, int src_y, int w, int h)
+{
+    emulated_edge_mc(buf, src, linesize, block_w, block_h, src_x, src_y,
+                     w, h, &ff_emu_edge_core_mmx);
+}
+#endif
+static av_noinline
+void emulated_edge_mc_sse(uint8_t *buf, const uint8_t *src, int linesize,
+                          int block_w, int block_h,
+                          int src_x, int src_y, int w, int h)
+{
+    emulated_edge_mc(buf, src, linesize, block_w, block_h, src_x, src_y,
+                     w, h, &ff_emu_edge_core_sse);
+}
+#endif /* HAVE_YASM */
+
+typedef void emulated_edge_mc_func (uint8_t *dst, const uint8_t *src,
+                                    int linesize, int block_w, int block_h,
+                                    int src_x, int src_y, int w, int h);
+
+static av_always_inline
+void gmc(uint8_t *dst, uint8_t *src, int stride, int h, int ox, int oy,
+         int dxx, int dxy, int dyx, int dyy, int shift, int r, int width, int height,
+         emulated_edge_mc_func *emu_edge_fn)
+{
     const int w = 8;
     const int ix = ox>>(16+shift);
     const int iy = oy>>(16+shift);
@@ -1701,7 +1773,7 @@ static void gmc_mmx(uint8_t *dst, uint8_t *src, int stride, int h, int ox, int o
     if( (unsigned)ix >= width-w ||
         (unsigned)iy >= height-h )
     {
-        ff_emulated_edge_mc(edge_buf, src, stride, w+1, h+1, ix, iy, width, height);
+        emu_edge_fn(edge_buf, src, stride, w+1, h+1, ix, iy, width, height);
         src = edge_buf;
     }
 
@@ -1782,6 +1854,30 @@ static void gmc_mmx(uint8_t *dst, uint8_t *src, int stride, int h, int ox, int o
     }
 }
 
+#if HAVE_YASM
+#if ARCH_X86_32
+static void gmc_mmx(uint8_t *dst, uint8_t *src, int stride, int h, int ox, int oy,
+                    int dxx, int dxy, int dyx, int dyy, int shift, int r, int width, int height)
+{
+    gmc(dst, src, stride, h, ox, oy, dxx, dxy, dyx, dyy, shift, r,
+        width, height, &emulated_edge_mc_mmx);
+}
+#endif
+static void gmc_sse(uint8_t *dst, uint8_t *src, int stride, int h, int ox, int oy,
+                    int dxx, int dxy, int dyx, int dyy, int shift, int r, int width, int height)
+{
+    gmc(dst, src, stride, h, ox, oy, dxx, dxy, dyx, dyy, shift, r,
+        width, height, &emulated_edge_mc_sse);
+}
+#else
+static void gmc_mmx(uint8_t *dst, uint8_t *src, int stride, int h, int ox, int oy,
+                    int dxx, int dxy, int dyx, int dyy, int shift, int r, int width, int height)
+{
+    gmc(dst, src, stride, h, ox, oy, dxx, dxy, dyx, dyy, shift, r,
+        width, height, &ff_emulated_edge_mc);
+}
+#endif
+
 #define PREFETCH(name, op) \
 static void name(void *mem, int stride, int h){\
     const uint8_t *p= mem;\
@@ -2626,7 +2722,12 @@ void dsputil_init_mmx(DSPContext* c, AVCodecContext *avctx)
         SET_HPEL_FUNCS(avg, 1, 8, mmx);
         SET_HPEL_FUNCS(avg_no_rnd, 1, 8, mmx);
 
+#if ARCH_X86_32 || !HAVE_YASM
         c->gmc= gmc_mmx;
+#endif
+#if ARCH_X86_32 && HAVE_YASM
+        c->emulated_edge_mc = emulated_edge_mc_mmx;
+#endif
 
         c->add_bytes= add_bytes_mmx;
         c->add_bytes_l2= add_bytes_l2_mmx;
@@ -2913,6 +3014,9 @@ void dsputil_init_mmx(DSPContext* c, AVCodecContext *avctx)
 #if HAVE_YASM
             c->scalarproduct_int16 = ff_scalarproduct_int16_sse2;
             c->scalarproduct_and_madd_int16 = ff_scalarproduct_and_madd_int16_sse2;
+
+            c->emulated_edge_mc = emulated_edge_mc_sse;
+            c->gmc= gmc_sse;
 #endif
         }
         if((mm_flags & AV_CPU_FLAG_SSSE3) && !(mm_flags & (AV_CPU_FLAG_SSE42|AV_CPU_FLAG_3DNOW)) && HAVE_YASM) // cachesplit
diff --git a/libavcodec/x86/dsputil_yasm.asm b/libavcodec/x86/dsputil_yasm.asm
index bda04727db..099f0a80df 100644
--- a/libavcodec/x86/dsputil_yasm.asm
+++ b/libavcodec/x86/dsputil_yasm.asm
@@ -421,3 +421,563 @@ cglobal scalarproduct_float_sse, 3,3,2, v1, v2, offset
     fld     dword r0m
 %endif
     RET
+
+; extern void ff_emu_edge_core(uint8_t *buf, const uint8_t *src, x86_reg linesize,
+;                              x86_reg start_y, x86_reg end_y, x86_reg block_h,
+;                              x86_reg start_x, x86_reg end_x, x86_reg block_w);
+;
+; The actual function itself is below. It basically wraps a very simple
+; w = end_x - start_x
+; if (w) {
+;   if (w > 22) {
+;     jump to the slow loop functions
+;   } else {
+;     jump to the fast loop functions
+;   }
+; }
+;
+; ... and then the same for left/right extend also. See below for loop
+; function implementations. Fast are fixed-width, slow is variable-width
+
+%macro EMU_EDGE_FUNC 1
+%ifdef ARCH_X86_64
+%define w_reg r10
+cglobal emu_edge_core_%1, 6, 7, 1
+    mov        r11, r5          ; save block_h
+%else
+%define w_reg r6
+cglobal emu_edge_core_%1, 2, 7, 0
+    mov         r4, r4m         ; end_y
+    mov         r5, r5m         ; block_h
+%endif
+
+    ; start with vertical extend (top/bottom) and body pixel copy
+    mov      w_reg, r7m
+    sub      w_reg, r6m         ; w = start_x - end_x
+    sub         r5, r4
+%ifdef ARCH_X86_64
+    sub         r4, r3
+%else
+    sub         r4, dword r3m
+%endif
+    cmp      w_reg, 22
+    jg .slow_v_extend_loop
+%ifdef ARCH_X86_32
+    mov         r2, r2m         ; linesize
+%endif
+    sal      w_reg, 7           ; w * 128
+%ifdef PIC
+    lea        rax, [.emuedge_v_extend_1 - (.emuedge_v_extend_2 - .emuedge_v_extend_1)]
+    add      w_reg, rax
+%else
+    lea      w_reg, [.emuedge_v_extend_1 - (.emuedge_v_extend_2 - .emuedge_v_extend_1)+w_reg]
+%endif
+    call     w_reg              ; fast top extend, body copy and bottom extend
+.v_extend_end:
+
+    ; horizontal extend (left/right)
+    mov      w_reg, r6m         ; start_x
+    sub         r0, w_reg
+%ifdef ARCH_X86_64
+    mov         r3, r0          ; backup of buf+block_h*linesize
+    mov         r5, r11
+%else
+    mov        r0m, r0          ; backup of buf+block_h*linesize
+    mov         r5, r5m
+%endif
+    test     w_reg, w_reg
+    jz .right_extend
+    cmp      w_reg, 22
+    jg .slow_left_extend_loop
+    mov         r1, w_reg
+    dec      w_reg
+    ; FIXME we can do a if size == 1 here if that makes any speed difference, test me
+    sar      w_reg, 1
+    sal      w_reg, 6
+    ; r0=buf+block_h*linesize,r10(64)/r6(32)=start_x offset for funcs
+    ; r6(rax)/r3(ebx)=val,r2=linesize,r1=start_x,r5=block_h
+%ifdef PIC
+    lea        rax, [.emuedge_extend_left_2]
+    add      w_reg, rax
+%else
+    lea      w_reg, [.emuedge_extend_left_2+w_reg]
+%endif
+    call     w_reg
+
+    ; now r3(64)/r0(32)=buf,r2=linesize,r11/r5=block_h,r6/r3=val, r10/r6=end_x, r1=block_w
+.right_extend:
+%ifdef ARCH_X86_32
+    mov         r0, r0m
+    mov         r5, r5m
+%endif
+    mov      w_reg, r7m         ; end_x
+    mov         r1, r8m         ; block_w
+    mov         r4, r1
+    sub         r1, w_reg
+    jz .h_extend_end            ; if (end_x == block_w) goto h_extend_end
+    cmp         r1, 22
+    jg .slow_right_extend_loop
+    dec         r1
+    ; FIXME we can do a if size == 1 here if that makes any speed difference, test me
+    sar         r1, 1
+    sal         r1, 6
+%ifdef PIC
+    lea        rax, [.emuedge_extend_right_2]
+    add         r1, rax
+%else
+    lea         r1, [.emuedge_extend_right_2+r1]
+%endif
+    call        r1
+.h_extend_end:
+    RET
+
+%ifdef ARCH_X86_64
+%define vall  al
+%define valh  ah
+%define valw  ax
+%define valw2 r10w
+%define valw3 r3w
+%define vald eax
+%else
+%define vall  bl
+%define valh  bh
+%define valw  bx
+%define valw2 r6w
+%define valw3 valw2
+%define vald ebx
+%define stack_offset 0x14
+%endif
+
+%endmacro
+
+; macro to read/write a horizontal number of pixels (%2) to/from registers
+; on x86-64, - fills xmm0-15 for consecutive sets of 16 pixels
+;            - if (%2 & 15 == 8) fills the last 8 bytes into rax
+;            - else if (%2 & 8)  fills 8 bytes into mm0
+;            - if (%2 & 7 == 4)  fills the last 4 bytes into rax
+;            - else if (%2 & 4)  fills 4 bytes into mm0-1
+;            - if (%2 & 3 == 3)  fills 2 bytes into r10/r3, and 1 into eax
+;              (note that we're using r3 for body/bottom because it's a shorter
+;               opcode, and then the loop fits in 128 bytes)
+;            - else              fills remaining bytes into rax
+; on x86-32, - fills mm0-7 for consecutive sets of 8 pixels
+;            - if (%2 & 7 == 4)  fills 4 bytes into ebx
+;            - else if (%2 & 4)  fills 4 bytes into mm0-7
+;            - if (%2 & 3 == 3)  fills 2 bytes into r6, and 1 into ebx
+;            - else              fills remaining bytes into ebx
+; writing data out is in the same way
+%macro READ_NUM_BYTES 3
+%assign %%src_off 0 ; offset in source buffer
+%assign %%smidx   0 ; mmx register idx
+%assign %%sxidx   0 ; xmm register idx
+
+%ifnidn %3, mmx
+%rep %2/16
+    movdqu xmm %+ %%sxidx, [r1+%%src_off]
+%assign %%src_off %%src_off+16
+%assign %%sxidx   %%sxidx+1
+%endrep ; %2/16
+%endif ; !mmx
+
+%ifdef ARCH_X86_64
+%if (%2-%%src_off) == 8
+    mov           rax, [r1+%%src_off]
+%assign %%src_off %%src_off+8
+%endif ; (%2-%%src_off) == 8
+%endif ; x86-64
+
+%rep (%2-%%src_off)/8
+    movq    mm %+ %%smidx, [r1+%%src_off]
+%assign %%src_off %%src_off+8
+%assign %%smidx   %%smidx+1
+%endrep ; (%2-%%dst_off)/8
+
+%if (%2-%%src_off) == 4
+    mov          vald, [r1+%%src_off]
+%elif (%2-%%src_off) & 4
+    movd    mm %+ %%smidx, [r1+%%src_off]
+%assign %%src_off %%src_off+4
+%endif ; (%2-%%src_off) ==/& 4
+
+%if (%2-%%src_off) == 1
+    mov          vall, [r1+%%src_off]
+%elif (%2-%%src_off) == 2
+    mov          valw, [r1+%%src_off]
+%elif (%2-%%src_off) == 3
+%ifidn %1, top
+    mov         valw2, [r1+%%src_off]
+%else ; %1 != top
+    mov         valw3, [r1+%%src_off]
+%endif ; %1 ==/!= top
+    mov          vall, [r1+%%src_off+2]
+%endif ; (%2-%%src_off) == 1/2/3
+%endmacro ; READ_NUM_BYTES
+
+%macro WRITE_NUM_BYTES 3
+%assign %%dst_off 0 ; offset in destination buffer
+%assign %%dmidx   0 ; mmx register idx
+%assign %%dxidx   0 ; xmm register idx
+
+%ifnidn %3, mmx
+%rep %2/16
+    movdqu [r0+%%dst_off], xmm %+ %%dxidx
+%assign %%dst_off %%dst_off+16
+%assign %%dxidx   %%dxidx+1
+%endrep ; %2/16
+%endif
+
+%ifdef ARCH_X86_64
+%if (%2-%%dst_off) == 8
+    mov    [r0+%%dst_off], rax
+%assign %%dst_off %%dst_off+8
+%endif ; (%2-%%dst_off) == 8
+%endif ; x86-64
+
+%rep (%2-%%dst_off)/8
+    movq   [r0+%%dst_off], mm %+ %%dmidx
+%assign %%dst_off %%dst_off+8
+%assign %%dmidx   %%dmidx+1
+%endrep ; (%2-%%dst_off)/8
+
+%if (%2-%%dst_off) == 4
+    mov    [r0+%%dst_off], vald
+%elif (%2-%%dst_off) & 4
+    movd   [r0+%%dst_off], mm %+ %%dmidx
+%assign %%dst_off %%dst_off+4
+%endif ; (%2-%%dst_off) ==/& 4
+
+%if (%2-%%dst_off) == 1
+    mov    [r0+%%dst_off], vall
+%elif (%2-%%dst_off) == 2
+    mov    [r0+%%dst_off], valw
+%elif (%2-%%dst_off) == 3
+%ifidn %1, top
+    mov    [r0+%%dst_off], valw2
+%else ; %1 != top
+    mov    [r0+%%dst_off], valw3
+%endif ; %1 ==/!= top
+    mov  [r0+%%dst_off+2], vall
+%endif ; (%2-%%dst_off) == 1/2/3
+%endmacro ; WRITE_NUM_BYTES
+
+; vertical top/bottom extend and body copy fast loops
+; these are function pointers to set-width line copy functions, i.e.
+; they read a fixed number of pixels into set registers, and write
+; those out into the destination buffer
+; r0=buf,r1=src,r2=linesize,r3(64)/r3m(32)=start_x,r4=end_y,r5=block_h
+; r6(eax/64)/r3(ebx/32)=val_reg
+%macro VERTICAL_EXTEND 1
+%assign %%n 1
+%rep 22
+ALIGN 128
+.emuedge_v_extend_ %+ %%n:
+    ; extend pixels above body
+%ifdef ARCH_X86_64
+    test           r3 , r3                   ; if (!start_y)
+    jz .emuedge_copy_body_ %+ %%n %+ _loop   ;   goto body
+%else ; ARCH_X86_32
+    cmp      dword r3m, 0
+    je .emuedge_copy_body_ %+ %%n %+ _loop
+%endif ; ARCH_X86_64/32
+    READ_NUM_BYTES  top,    %%n, %1          ; read bytes
+.emuedge_extend_top_ %+ %%n %+ _loop:        ; do {
+    WRITE_NUM_BYTES top,    %%n, %1          ;   write bytes
+    add            r0 , r2                   ;   dst += linesize
+%ifdef ARCH_X86_64
+    dec            r3
+%else ; ARCH_X86_32
+    dec      dword r3m
+%endif ; ARCH_X86_64/32
+    jnz .emuedge_extend_top_ %+ %%n %+ _loop ; } while (--start_y)
+
+    ; copy body pixels
+.emuedge_copy_body_ %+ %%n %+ _loop:         ; do {
+    READ_NUM_BYTES  body,   %%n, %1          ;   read bytes
+    WRITE_NUM_BYTES body,   %%n, %1          ;   write bytes
+    add            r0 , r2                   ;   dst += linesize
+    add            r1 , r2                   ;   src += linesize
+    dec            r4
+    jnz .emuedge_copy_body_ %+ %%n %+ _loop  ; } while (--end_y)
+
+    ; copy bottom pixels
+    test           r5 , r5                   ; if (!block_h)
+    jz .emuedge_v_extend_end_ %+ %%n         ;   goto end
+    sub            r1 , r2                   ; src -= linesize
+    READ_NUM_BYTES  bottom, %%n, %1          ; read bytes
+.emuedge_extend_bottom_ %+ %%n %+ _loop:     ; do {
+    WRITE_NUM_BYTES bottom, %%n, %1          ;   write bytes
+    add            r0 , r2                   ;   dst += linesize
+    dec            r5
+    jnz .emuedge_extend_bottom_ %+ %%n %+ _loop ; } while (--block_h)
+
+.emuedge_v_extend_end_ %+ %%n:
+%ifdef ARCH_X86_64
+    ret
+%else ; ARCH_X86_32
+    rep ret
+%endif ; ARCH_X86_64/32
+%assign %%n %%n+1
+%endrep
+%endmacro VERTICAL_EXTEND
+
+; left/right (horizontal) fast extend functions
+; these are essentially identical to the vertical extend ones above,
+; just left/right separated because number of pixels to extend is
+; obviously not the same on both sides.
+; for reading, pixels are placed in eax (x86-64) or ebx (x86-64) in the
+; lowest two bytes of the register (so val*0x0101), and are splatted
+; into each byte of mm0 as well if n_pixels >= 8
+
+%macro READ_V_PIXEL 3
+    mov        vall, %2
+    mov        valh, vall
+%if %1 >= 8
+    movd        mm0, vald
+%ifidn %3, mmx
+    punpcklwd   mm0, mm0
+    punpckldq   mm0, mm0
+%else ; !mmx
+    pshufw      mm0, mm0, 0
+%endif ; mmx
+%endif ; %1 >= 8
+%endmacro
+
+%macro WRITE_V_PIXEL 2
+%assign %%dst_off 0
+%rep %1/8
+    movq [%2+%%dst_off], mm0
+%assign %%dst_off %%dst_off+8
+%endrep
+%if %1 & 4
+%if %1 >= 8
+    movd [%2+%%dst_off], mm0
+%else ; %1 < 8
+    mov  [%2+%%dst_off]  , valw
+    mov  [%2+%%dst_off+2], valw
+%endif ; %1 >=/< 8
+%assign %%dst_off %%dst_off+4
+%endif ; %1 & 4
+%if %1&2
+    mov  [%2+%%dst_off], valw
+%endif ; %1 & 2
+%endmacro
+
+; r0=buf+block_h*linesize, r1=start_x, r2=linesize, r5=block_h, r6/r3=val
+%macro LEFT_EXTEND 1
+%assign %%n 2
+%rep 11
+ALIGN 64
+.emuedge_extend_left_ %+ %%n:          ; do {
+    sub         r0, r2                 ;   dst -= linesize
+    READ_V_PIXEL  %%n, [r0+r1], %1     ;   read pixels
+    WRITE_V_PIXEL %%n, r0              ;   write pixels
+    dec         r5
+    jnz .emuedge_extend_left_ %+ %%n   ; } while (--block_h)
+%ifdef ARCH_X86_64
+    ret
+%else ; ARCH_X86_32
+    rep ret
+%endif ; ARCH_X86_64/32
+%assign %%n %%n+2
+%endrep
+%endmacro ; LEFT_EXTEND
+
+; r3/r0=buf+block_h*linesize, r2=linesize, r11/r5=block_h, r0/r6=end_x, r6/r3=val
+%macro RIGHT_EXTEND 1
+%assign %%n 2
+%rep 11
+ALIGN 64
+.emuedge_extend_right_ %+ %%n:          ; do {
+%ifdef ARCH_X86_64
+    sub        r3, r2                   ;   dst -= linesize
+    READ_V_PIXEL  %%n, [r3+w_reg-1], %1 ;   read pixels
+    WRITE_V_PIXEL %%n, r3+r4-%%n        ;   write pixels
+    dec       r11
+%else ; ARCH_X86_32
+    sub        r0, r2                   ;   dst -= linesize
+    READ_V_PIXEL  %%n, [r0+w_reg-1], %1 ;   read pixels
+    WRITE_V_PIXEL %%n, r0+r4-%%n        ;   write pixels
+    dec     r5
+%endif ; ARCH_X86_64/32
+    jnz .emuedge_extend_right_ %+ %%n   ; } while (--block_h)
+%ifdef ARCH_X86_64
+    ret
+%else ; ARCH_X86_32
+    rep ret
+%endif ; ARCH_X86_64/32
+%assign %%n %%n+2
+%endrep
+
+%ifdef ARCH_X86_32
+%define stack_offset 0x10
+%endif
+%endmacro ; RIGHT_EXTEND
+
+; below follow the "slow" copy/extend functions, these act on a non-fixed
+; width specified in a register, and run a loop to copy the full amount
+; of bytes. They are optimized for copying of large amounts of pixels per
+; line, so they unconditionally splat data into mm registers to copy 8
+; bytes per loop iteration. It could be considered to use xmm for x86-64
+; also, but I haven't optimized this as much (i.e. FIXME)
+%macro V_COPY_NPX 4-5
+%if %0 == 4
+    test     w_reg, %4
+    jz .%1_skip_%4_px
+%else ; %0 == 5
+.%1_%4_px_loop:
+%endif
+    %3          %2, [r1+cnt_reg]
+    %3 [r0+cnt_reg], %2
+    add    cnt_reg, %4
+%if %0 == 5
+    sub      w_reg, %4
+    test     w_reg, %5
+    jnz .%1_%4_px_loop
+%endif
+.%1_skip_%4_px:
+%endmacro
+
+%macro V_COPY_ROW 3
+%ifidn %1, bottom
+    sub         r1, linesize
+%endif
+.%1_copy_loop:
+    xor    cnt_reg, cnt_reg
+%ifidn %3, mmx
+%define linesize r2m
+    V_COPY_NPX %1,  mm0, movq,    8, 0xFFFFFFF8
+%else ; !mmx
+    V_COPY_NPX %1, xmm0, movdqu, 16, 0xFFFFFFF0
+%ifdef ARCH_X86_64
+%define linesize r2
+    V_COPY_NPX %1, rax , mov,     8
+%else ; ARCH_X86_32
+%define linesize r2m
+    V_COPY_NPX %1,  mm0, movq,    8
+%endif ; ARCH_X86_64/32
+%endif ; mmx
+    V_COPY_NPX %1, vald, mov,     4
+    V_COPY_NPX %1, valw, mov,     2
+    V_COPY_NPX %1, vall, mov,     1
+    mov      w_reg, cnt_reg
+%ifidn %1, body
+    add         r1, linesize
+%endif
+    add         r0, linesize
+    dec         %2
+    jnz .%1_copy_loop
+%endmacro
+
+%macro SLOW_V_EXTEND 1
+.slow_v_extend_loop:
+; r0=buf,r1=src,r2(64)/r2m(32)=linesize,r3(64)/r3m(32)=start_x,r4=end_y,r5=block_h
+; r11(64)/r3(later-64)/r2(32)=cnt_reg,r6(64)/r3(32)=val_reg,r10(64)/r6(32)=w=end_x-start_x
+%ifdef ARCH_X86_64
+    push       r11              ; save old value of block_h
+    test        r3, r3
+%define cnt_reg r11
+    jz .do_body_copy            ; if (!start_y) goto do_body_copy
+    V_COPY_ROW top, r3, %1
+%else
+    cmp  dword r3m, 0
+%define cnt_reg r2
+    je .do_body_copy            ; if (!start_y) goto do_body_copy
+    V_COPY_ROW top, dword r3m, %1
+%endif
+
+.do_body_copy:
+    V_COPY_ROW body, r4, %1
+
+%ifdef ARCH_X86_64
+    pop        r11              ; restore old value of block_h
+%define cnt_reg r3
+%endif
+    test        r5, r5
+%ifdef ARCH_X86_64
+    jz .v_extend_end
+%else
+    jz .skip_bottom_extend
+%endif
+    V_COPY_ROW bottom, r5, %1
+%ifdef ARCH_X86_32
+.skip_bottom_extend:
+    mov         r2, r2m
+%endif
+    jmp .v_extend_end
+%endmacro
+
+%macro SLOW_LEFT_EXTEND 1
+.slow_left_extend_loop:
+; r0=buf+block_h*linesize,r2=linesize,r6(64)/r3(32)=val,r5=block_h,r4=cntr,r10/r6=start_x
+    mov         r4, 8
+    sub         r0, linesize
+    READ_V_PIXEL 8, [r0+w_reg], %1
+.left_extend_8px_loop:
+    movq [r0+r4-8], mm0
+    add         r4, 8
+    cmp         r4, w_reg
+    jle .left_extend_8px_loop
+    sub         r4, 8
+    cmp         r4, w_reg
+    jge .left_extend_loop_end
+.left_extend_2px_loop:
+    mov    [r0+r4], valw
+    add         r4, 2
+    cmp         r4, w_reg
+    jl .left_extend_2px_loop
+.left_extend_loop_end:
+    dec         r5
+    jnz .slow_left_extend_loop
+%ifdef ARCH_X86_32
+    mov         r2, r2m
+%endif
+    jmp .right_extend
+%endmacro
+
+%macro SLOW_RIGHT_EXTEND 1
+.slow_right_extend_loop:
+; r3(64)/r0(32)=buf+block_h*linesize,r2=linesize,r4=block_w,r11(64)/r5(32)=block_h,
+; r10(64)/r6(32)=end_x,r6/r3=val,r1=cntr
+%ifdef ARCH_X86_64
+%define buf_reg r3
+%define bh_reg r11
+%else
+%define buf_reg r0
+%define bh_reg r5
+%endif
+    lea         r1, [r4-8]
+    sub    buf_reg, linesize
+    READ_V_PIXEL 8, [buf_reg+w_reg-1], %1
+.right_extend_8px_loop:
+    movq [buf_reg+r1], mm0
+    sub         r1, 8
+    cmp         r1, w_reg
+    jge .right_extend_8px_loop
+    add         r1, 8
+    cmp         r1, w_reg
+    je .right_extend_loop_end
+.right_extend_2px_loop:
+    sub         r1, 2
+    mov [buf_reg+r1], valw
+    cmp         r1, w_reg
+    jg .right_extend_2px_loop
+.right_extend_loop_end:
+    dec         bh_reg
+    jnz .slow_right_extend_loop
+    jmp .h_extend_end
+%endmacro
+
+%macro emu_edge 1
+EMU_EDGE_FUNC     %1
+VERTICAL_EXTEND   %1
+LEFT_EXTEND       %1
+RIGHT_EXTEND      %1
+SLOW_V_EXTEND     %1
+SLOW_LEFT_EXTEND  %1
+SLOW_RIGHT_EXTEND %1
+%endmacro
+
+emu_edge sse
+%ifdef ARCH_X86_32
+emu_edge mmx
+%endif