+; void pred8x8l_top_dc_8(uint8_t *src, int has_topleft, int has_topright, int stride)
+;-----------------------------------------------------------------------------
+%macro PRED8x8L_TOP_DC 0
+cglobal pred8x8l_top_dc_8, 4,4
+ sub r0, r3
+ pxor mm7, mm7
+ movq mm0, [r0-8]
+ movq mm3, [r0]
+ movq mm1, [r0+8]
+ movq mm2, mm3
+ movq mm4, mm3
+ PALIGNR mm2, mm0, 7, mm0
+ PALIGNR mm1, mm4, 1, mm4
+ test r1, r1 ; top_left
+ jz .fix_lt_2
+ test r2, r2 ; top_right
+ jz .fix_tr_1
+ jmp .body
+.fix_lt_2:
+ movq mm5, mm3
+ pxor mm5, mm2
+ psllq mm5, 56
+ psrlq mm5, 56
+ pxor mm2, mm5
+ test r2, r2 ; top_right
+ jnz .body
+.fix_tr_1:
+ movq mm5, mm3
+ pxor mm5, mm1
+ psrlq mm5, 56
+ psllq mm5, 56
+ pxor mm1, mm5
+.body:
+ PRED4x4_LOWPASS mm0, mm2, mm1, mm3, mm5
+ psadbw mm7, mm0
+ paddw mm7, [pw_4]
+ psrlw mm7, 3
+ pshufw mm7, mm7, 0
+ packuswb mm7, mm7
+%rep 3
+ movq [r0+r3*1], mm7
+ movq [r0+r3*2], mm7
+ lea r0, [r0+r3*2]
+%endrep
+ movq [r0+r3*1], mm7
+ movq [r0+r3*2], mm7
+ RET
+%endmacro
+
+INIT_MMX mmxext
+PRED8x8L_TOP_DC
+INIT_MMX ssse3
+PRED8x8L_TOP_DC
+
+;-----------------------------------------------------------------------------
+;void pred8x8l_dc_8(uint8_t *src, int has_topleft, int has_topright, int stride)
+;-----------------------------------------------------------------------------
+
+%macro PRED8x8L_DC 0
+cglobal pred8x8l_dc_8, 4,5
+ sub r0, r3
+ lea r4, [r0+r3*2]
+ movq mm0, [r0+r3*1-8]
+ punpckhbw mm0, [r0+r3*0-8]
+ movq mm1, [r4+r3*1-8]
+ punpckhbw mm1, [r0+r3*2-8]
+ mov r4, r0
+ punpckhwd mm1, mm0
+ lea r0, [r0+r3*4]
+ movq mm2, [r0+r3*1-8]
+ punpckhbw mm2, [r0+r3*0-8]
+ lea r0, [r0+r3*2]
+ movq mm3, [r0+r3*1-8]
+ punpckhbw mm3, [r0+r3*0-8]
+ punpckhwd mm3, mm2
+ punpckhdq mm3, mm1
+ lea r0, [r0+r3*2]
+ movq mm0, [r0+r3*0-8]
+ movq mm1, [r4]
+ mov r0, r4
+ movq mm4, mm3
+ movq mm2, mm3
+ PALIGNR mm4, mm0, 7, mm0
+ PALIGNR mm1, mm2, 1, mm2
+ test r1, r1
+ jnz .do_left
+.fix_lt_1:
+ movq mm5, mm3
+ pxor mm5, mm4
+ psrlq mm5, 56
+ psllq mm5, 48
+ pxor mm1, mm5
+ jmp .do_left
+.fix_lt_2:
+ movq mm5, mm3
+ pxor mm5, mm2
+ psllq mm5, 56
+ psrlq mm5, 56
+ pxor mm2, mm5
+ test r2, r2
+ jnz .body
+.fix_tr_1:
+ movq mm5, mm3
+ pxor mm5, mm1
+ psrlq mm5, 56
+ psllq mm5, 56
+ pxor mm1, mm5
+ jmp .body
+.do_left:
+ movq mm0, mm4
+ PRED4x4_LOWPASS mm2, mm1, mm4, mm3, mm5
+ movq mm4, mm0
+ movq mm7, mm2
+ PRED4x4_LOWPASS mm1, mm3, mm0, mm4, mm5
+ psllq mm1, 56
+ PALIGNR mm7, mm1, 7, mm3
+ movq mm0, [r0-8]
+ movq mm3, [r0]
+ movq mm1, [r0+8]
+ movq mm2, mm3
+ movq mm4, mm3
+ PALIGNR mm2, mm0, 7, mm0
+ PALIGNR mm1, mm4, 1, mm4
+ test r1, r1
+ jz .fix_lt_2
+ test r2, r2
+ jz .fix_tr_1
+.body:
+ lea r1, [r0+r3*2]
+ PRED4x4_LOWPASS mm6, mm2, mm1, mm3, mm5
+ pxor mm0, mm0
+ pxor mm1, mm1
+ lea r2, [r1+r3*2]
+ psadbw mm0, mm7
+ psadbw mm1, mm6
+ paddw mm0, [pw_8]
+ paddw mm0, mm1
+ lea r4, [r2+r3*2]
+ psrlw mm0, 4
+ pshufw mm0, mm0, 0
+ packuswb mm0, mm0
+ movq [r0+r3*1], mm0
+ movq [r0+r3*2], mm0
+ movq [r1+r3*1], mm0
+ movq [r1+r3*2], mm0
+ movq [r2+r3*1], mm0
+ movq [r2+r3*2], mm0
+ movq [r4+r3*1], mm0
+ movq [r4+r3*2], mm0
+ RET
+%endmacro
+
+INIT_MMX mmxext
+PRED8x8L_DC
+INIT_MMX ssse3
+PRED8x8L_DC
+
+;-----------------------------------------------------------------------------
+; void pred8x8l_horizontal_8(uint8_t *src, int has_topleft, int has_topright, int stride)
+;-----------------------------------------------------------------------------
+
+%macro PRED8x8L_HORIZONTAL 0
+cglobal pred8x8l_horizontal_8, 4,4
+ sub r0, r3
+ lea r2, [r0+r3*2]
+ movq mm0, [r0+r3*1-8]
+ test r1, r1
+ lea r1, [r0+r3]
+ cmovnz r1, r0
+ punpckhbw mm0, [r1+r3*0-8]
+ movq mm1, [r2+r3*1-8]
+ punpckhbw mm1, [r0+r3*2-8]
+ mov r2, r0
+ punpckhwd mm1, mm0
+ lea r0, [r0+r3*4]
+ movq mm2, [r0+r3*1-8]
+ punpckhbw mm2, [r0+r3*0-8]
+ lea r0, [r0+r3*2]
+ movq mm3, [r0+r3*1-8]
+ punpckhbw mm3, [r0+r3*0-8]
+ punpckhwd mm3, mm2
+ punpckhdq mm3, mm1
+ lea r0, [r0+r3*2]
+ movq mm0, [r0+r3*0-8]
+ movq mm1, [r1+r3*0-8]
+ mov r0, r2
+ movq mm4, mm3
+ movq mm2, mm3
+ PALIGNR mm4, mm0, 7, mm0
+ PALIGNR mm1, mm2, 1, mm2
+ movq mm0, mm4
+ PRED4x4_LOWPASS mm2, mm1, mm4, mm3, mm5
+ movq mm4, mm0
+ movq mm7, mm2
+ PRED4x4_LOWPASS mm1, mm3, mm0, mm4, mm5
+ psllq mm1, 56
+ PALIGNR mm7, mm1, 7, mm3
+ movq mm3, mm7
+ lea r1, [r0+r3*2]
+ movq mm7, mm3
+ punpckhbw mm3, mm3
+ punpcklbw mm7, mm7
+ pshufw mm0, mm3, 0xff
+ pshufw mm1, mm3, 0xaa
+ lea r2, [r1+r3*2]
+ pshufw mm2, mm3, 0x55
+ pshufw mm3, mm3, 0x00
+ pshufw mm4, mm7, 0xff
+ pshufw mm5, mm7, 0xaa
+ pshufw mm6, mm7, 0x55
+ pshufw mm7, mm7, 0x00
+ movq [r0+r3*1], mm0
+ movq [r0+r3*2], mm1
+ movq [r1+r3*1], mm2
+ movq [r1+r3*2], mm3
+ movq [r2+r3*1], mm4
+ movq [r2+r3*2], mm5
+ lea r0, [r2+r3*2]
+ movq [r0+r3*1], mm6
+ movq [r0+r3*2], mm7
+ RET
+%endmacro
+
+INIT_MMX mmxext
+PRED8x8L_HORIZONTAL
+INIT_MMX ssse3
+PRED8x8L_HORIZONTAL
+
+;-----------------------------------------------------------------------------
+; void pred8x8l_vertical_8(uint8_t *src, int has_topleft, int has_topright, int stride)
+;-----------------------------------------------------------------------------
+
+%macro PRED8x8L_VERTICAL 0
+cglobal pred8x8l_vertical_8, 4,4
+ sub r0, r3
+ movq mm0, [r0-8]
+ movq mm3, [r0]
+ movq mm1, [r0+8]
+ movq mm2, mm3
+ movq mm4, mm3
+ PALIGNR mm2, mm0, 7, mm0
+ PALIGNR mm1, mm4, 1, mm4
+ test r1, r1 ; top_left
+ jz .fix_lt_2
+ test r2, r2 ; top_right
+ jz .fix_tr_1
+ jmp .body
+.fix_lt_2:
+ movq mm5, mm3
+ pxor mm5, mm2
+ psllq mm5, 56
+ psrlq mm5, 56
+ pxor mm2, mm5
+ test r2, r2 ; top_right
+ jnz .body
+.fix_tr_1:
+ movq mm5, mm3
+ pxor mm5, mm1
+ psrlq mm5, 56
+ psllq mm5, 56
+ pxor mm1, mm5
+.body:
+ PRED4x4_LOWPASS mm0, mm2, mm1, mm3, mm5
+%rep 3
+ movq [r0+r3*1], mm0
+ movq [r0+r3*2], mm0
+ lea r0, [r0+r3*2]
+%endrep
+ movq [r0+r3*1], mm0
+ movq [r0+r3*2], mm0
+ RET
+%endmacro
+
+INIT_MMX mmxext
+PRED8x8L_VERTICAL
+INIT_MMX ssse3
+PRED8x8L_VERTICAL
+
+;-----------------------------------------------------------------------------
+;void pred8x8l_down_left_8(uint8_t *src, int has_topleft, int has_topright, int stride)
+;-----------------------------------------------------------------------------
+
+INIT_MMX mmxext
+cglobal pred8x8l_down_left_8, 4,5
+ sub r0, r3
+ movq mm0, [r0-8]
+ movq mm3, [r0]
+ movq mm1, [r0+8]
+ movq mm2, mm3
+ movq mm4, mm3
+ PALIGNR mm2, mm0, 7, mm0
+ PALIGNR mm1, mm4, 1, mm4
+ test r1, r1
+ jz .fix_lt_2
+ test r2, r2
+ jz .fix_tr_1
+ jmp .do_top
+.fix_lt_2:
+ movq mm5, mm3
+ pxor mm5, mm2
+ psllq mm5, 56
+ psrlq mm5, 56
+ pxor mm2, mm5
+ test r2, r2
+ jnz .do_top
+.fix_tr_1:
+ movq mm5, mm3
+ pxor mm5, mm1
+ psrlq mm5, 56
+ psllq mm5, 56
+ pxor mm1, mm5
+ jmp .do_top
+.fix_tr_2:
+ punpckhbw mm3, mm3
+ pshufw mm1, mm3, 0xFF
+ jmp .do_topright
+.do_top:
+ PRED4x4_LOWPASS mm4, mm2, mm1, mm3, mm5
+ movq mm7, mm4
+ test r2, r2
+ jz .fix_tr_2
+ movq mm0, [r0+8]
+ movq mm5, mm0
+ movq mm2, mm0
+ movq mm4, mm0
+ psrlq mm5, 56
+ PALIGNR mm2, mm3, 7, mm3
+ PALIGNR mm5, mm4, 1, mm4
+ PRED4x4_LOWPASS mm1, mm2, mm5, mm0, mm4
+.do_topright:
+ lea r1, [r0+r3*2]
+ movq mm6, mm1
+ psrlq mm1, 56
+ movq mm4, mm1
+ lea r2, [r1+r3*2]
+ movq mm2, mm6
+ PALIGNR mm2, mm7, 1, mm0
+ movq mm3, mm6
+ PALIGNR mm3, mm7, 7, mm0
+ PALIGNR mm4, mm6, 1, mm0
+ movq mm5, mm7
+ movq mm1, mm7
+ movq mm7, mm6
+ lea r4, [r2+r3*2]
+ psllq mm1, 8
+ PRED4x4_LOWPASS mm0, mm1, mm2, mm5, mm6
+ PRED4x4_LOWPASS mm1, mm3, mm4, mm7, mm6
+ movq [r4+r3*2], mm1
+ movq mm2, mm0
+ psllq mm1, 8
+ psrlq mm2, 56
+ psllq mm0, 8
+ por mm1, mm2
+ movq [r4+r3*1], mm1
+ movq mm2, mm0
+ psllq mm1, 8
+ psrlq mm2, 56
+ psllq mm0, 8
+ por mm1, mm2
+ movq [r2+r3*2], mm1
+ movq mm2, mm0
+ psllq mm1, 8
+ psrlq mm2, 56
+ psllq mm0, 8
+ por mm1, mm2
+ movq [r2+r3*1], mm1
+ movq mm2, mm0
+ psllq mm1, 8
+ psrlq mm2, 56
+ psllq mm0, 8
+ por mm1, mm2
+ movq [r1+r3*2], mm1
+ movq mm2, mm0
+ psllq mm1, 8
+ psrlq mm2, 56
+ psllq mm0, 8
+ por mm1, mm2
+ movq [r1+r3*1], mm1
+ movq mm2, mm0
+ psllq mm1, 8
+ psrlq mm2, 56
+ psllq mm0, 8
+ por mm1, mm2
+ movq [r0+r3*2], mm1
+ psllq mm1, 8
+ psrlq mm0, 56
+ por mm1, mm0
+ movq [r0+r3*1], mm1
+ RET
+
+%macro PRED8x8L_DOWN_LEFT 0
+cglobal pred8x8l_down_left_8, 4,4
+ sub r0, r3
+ movq mm0, [r0-8]
+ movq mm3, [r0]
+ movq mm1, [r0+8]
+ movq mm2, mm3
+ movq mm4, mm3
+ PALIGNR mm2, mm0, 7, mm0
+ PALIGNR mm1, mm4, 1, mm4
+ test r1, r1 ; top_left
+ jz .fix_lt_2
+ test r2, r2 ; top_right
+ jz .fix_tr_1
+ jmp .do_top
+.fix_lt_2:
+ movq mm5, mm3
+ pxor mm5, mm2
+ psllq mm5, 56
+ psrlq mm5, 56
+ pxor mm2, mm5
+ test r2, r2 ; top_right
+ jnz .do_top
+.fix_tr_1:
+ movq mm5, mm3
+ pxor mm5, mm1
+ psrlq mm5, 56
+ psllq mm5, 56
+ pxor mm1, mm5
+ jmp .do_top
+.fix_tr_2:
+ punpckhbw mm3, mm3
+ pshufw mm1, mm3, 0xFF
+ jmp .do_topright
+.do_top:
+ PRED4x4_LOWPASS mm4, mm2, mm1, mm3, mm5
+ movq2dq xmm3, mm4
+ test r2, r2 ; top_right
+ jz .fix_tr_2
+ movq mm0, [r0+8]
+ movq mm5, mm0
+ movq mm2, mm0
+ movq mm4, mm0
+ psrlq mm5, 56
+ PALIGNR mm2, mm3, 7, mm3
+ PALIGNR mm5, mm4, 1, mm4
+ PRED4x4_LOWPASS mm1, mm2, mm5, mm0, mm4
+.do_topright:
+ movq2dq xmm4, mm1
+ psrlq mm1, 56
+ movq2dq xmm5, mm1
+ lea r1, [r0+r3*2]
+ pslldq xmm4, 8
+ por xmm3, xmm4
+ movdqa xmm2, xmm3
+ psrldq xmm2, 1
+ pslldq xmm5, 15
+ por xmm2, xmm5
+ lea r2, [r1+r3*2]
+ movdqa xmm1, xmm3
+ pslldq xmm1, 1
+INIT_XMM cpuname
+ PRED4x4_LOWPASS xmm0, xmm1, xmm2, xmm3, xmm4
+ psrldq xmm0, 1
+ movq [r0+r3*1], xmm0
+ psrldq xmm0, 1
+ movq [r0+r3*2], xmm0
+ psrldq xmm0, 1
+ lea r0, [r2+r3*2]
+ movq [r1+r3*1], xmm0
+ psrldq xmm0, 1
+ movq [r1+r3*2], xmm0
+ psrldq xmm0, 1
+ movq [r2+r3*1], xmm0
+ psrldq xmm0, 1
+ movq [r2+r3*2], xmm0
+ psrldq xmm0, 1
+ movq [r0+r3*1], xmm0
+ psrldq xmm0, 1
+ movq [r0+r3*2], xmm0
+ RET
+%endmacro
+
+INIT_MMX sse2
+PRED8x8L_DOWN_LEFT
+INIT_MMX ssse3
+PRED8x8L_DOWN_LEFT
+
+;-----------------------------------------------------------------------------
+;void pred8x8l_down_right_8_mmxext(uint8_t *src, int has_topleft, int has_topright, int stride)
+;-----------------------------------------------------------------------------
+
+INIT_MMX mmxext
+cglobal pred8x8l_down_right_8, 4,5
+ sub r0, r3
+ lea r4, [r0+r3*2]
+ movq mm0, [r0+r3*1-8]
+ punpckhbw mm0, [r0+r3*0-8]
+ movq mm1, [r4+r3*1-8]
+ punpckhbw mm1, [r0+r3*2-8]
+ mov r4, r0
+ punpckhwd mm1, mm0
+ lea r0, [r0+r3*4]
+ movq mm2, [r0+r3*1-8]
+ punpckhbw mm2, [r0+r3*0-8]
+ lea r0, [r0+r3*2]
+ movq mm3, [r0+r3*1-8]
+ punpckhbw mm3, [r0+r3*0-8]
+ punpckhwd mm3, mm2
+ punpckhdq mm3, mm1
+ lea r0, [r0+r3*2]
+ movq mm0, [r0+r3*0-8]
+ movq mm1, [r4]
+ mov r0, r4
+ movq mm4, mm3
+ movq mm2, mm3
+ PALIGNR mm4, mm0, 7, mm0
+ PALIGNR mm1, mm2, 1, mm2
+ test r1, r1 ; top_left
+ jz .fix_lt_1
+.do_left:
+ movq mm0, mm4
+ PRED4x4_LOWPASS mm2, mm1, mm4, mm3, mm5
+ movq mm4, mm0
+ movq mm7, mm2
+ movq mm6, mm2
+ PRED4x4_LOWPASS mm1, mm3, mm0, mm4, mm5
+ psllq mm1, 56
+ PALIGNR mm7, mm1, 7, mm3
+ movq mm0, [r0-8]
+ movq mm3, [r0]
+ movq mm1, [r0+8]
+ movq mm2, mm3
+ movq mm4, mm3
+ PALIGNR mm2, mm0, 7, mm0
+ PALIGNR mm1, mm4, 1, mm4
+ test r1, r1 ; top_left
+ jz .fix_lt_2
+ test r2, r2 ; top_right
+ jz .fix_tr_1
+.do_top:
+ PRED4x4_LOWPASS mm4, mm2, mm1, mm3, mm5
+ movq mm5, mm4
+ jmp .body
+.fix_lt_1:
+ movq mm5, mm3
+ pxor mm5, mm4
+ psrlq mm5, 56
+ psllq mm5, 48
+ pxor mm1, mm5
+ jmp .do_left
+.fix_lt_2:
+ movq mm5, mm3
+ pxor mm5, mm2
+ psllq mm5, 56
+ psrlq mm5, 56
+ pxor mm2, mm5
+ test r2, r2 ; top_right
+ jnz .do_top
+.fix_tr_1:
+ movq mm5, mm3
+ pxor mm5, mm1
+ psrlq mm5, 56
+ psllq mm5, 56
+ pxor mm1, mm5
+ jmp .do_top
+.body:
+ lea r1, [r0+r3*2]
+ movq mm1, mm7
+ movq mm7, mm5
+ movq mm5, mm6
+ movq mm2, mm7
+ lea r2, [r1+r3*2]
+ PALIGNR mm2, mm6, 1, mm0
+ movq mm3, mm7
+ PALIGNR mm3, mm6, 7, mm0
+ movq mm4, mm7
+ lea r4, [r2+r3*2]
+ psrlq mm4, 8
+ PRED4x4_LOWPASS mm0, mm1, mm2, mm5, mm6
+ PRED4x4_LOWPASS mm1, mm3, mm4, mm7, mm6
+ movq [r4+r3*2], mm0
+ movq mm2, mm1
+ psrlq mm0, 8
+ psllq mm2, 56
+ psrlq mm1, 8
+ por mm0, mm2
+ movq [r4+r3*1], mm0
+ movq mm2, mm1
+ psrlq mm0, 8
+ psllq mm2, 56
+ psrlq mm1, 8
+ por mm0, mm2
+ movq [r2+r3*2], mm0
+ movq mm2, mm1
+ psrlq mm0, 8
+ psllq mm2, 56
+ psrlq mm1, 8
+ por mm0, mm2
+ movq [r2+r3*1], mm0
+ movq mm2, mm1
+ psrlq mm0, 8
+ psllq mm2, 56
+ psrlq mm1, 8
+ por mm0, mm2
+ movq [r1+r3*2], mm0
+ movq mm2, mm1
+ psrlq mm0, 8
+ psllq mm2, 56
+ psrlq mm1, 8
+ por mm0, mm2
+ movq [r1+r3*1], mm0
+ movq mm2, mm1
+ psrlq mm0, 8
+ psllq mm2, 56
+ psrlq mm1, 8
+ por mm0, mm2
+ movq [r0+r3*2], mm0
+ psrlq mm0, 8
+ psllq mm1, 56
+ por mm0, mm1
+ movq [r0+r3*1], mm0
+ RET
+
+%macro PRED8x8L_DOWN_RIGHT 0
+cglobal pred8x8l_down_right_8, 4,5
+ sub r0, r3
+ lea r4, [r0+r3*2]
+ movq mm0, [r0+r3*1-8]
+ punpckhbw mm0, [r0+r3*0-8]
+ movq mm1, [r4+r3*1-8]
+ punpckhbw mm1, [r0+r3*2-8]
+ mov r4, r0
+ punpckhwd mm1, mm0
+ lea r0, [r0+r3*4]
+ movq mm2, [r0+r3*1-8]
+ punpckhbw mm2, [r0+r3*0-8]
+ lea r0, [r0+r3*2]
+ movq mm3, [r0+r3*1-8]
+ punpckhbw mm3, [r0+r3*0-8]
+ punpckhwd mm3, mm2
+ punpckhdq mm3, mm1
+ lea r0, [r0+r3*2]
+ movq mm0, [r0+r3*0-8]
+ movq mm1, [r4]
+ mov r0, r4
+ movq mm4, mm3
+ movq mm2, mm3
+ PALIGNR mm4, mm0, 7, mm0
+ PALIGNR mm1, mm2, 1, mm2
+ test r1, r1
+ jz .fix_lt_1
+ jmp .do_left
+.fix_lt_1:
+ movq mm5, mm3
+ pxor mm5, mm4
+ psrlq mm5, 56
+ psllq mm5, 48
+ pxor mm1, mm5
+ jmp .do_left
+.fix_lt_2:
+ movq mm5, mm3
+ pxor mm5, mm2
+ psllq mm5, 56
+ psrlq mm5, 56
+ pxor mm2, mm5
+ test r2, r2
+ jnz .do_top
+.fix_tr_1:
+ movq mm5, mm3
+ pxor mm5, mm1
+ psrlq mm5, 56
+ psllq mm5, 56
+ pxor mm1, mm5
+ jmp .do_top
+.do_left:
+ movq mm0, mm4
+ PRED4x4_LOWPASS mm2, mm1, mm4, mm3, mm5
+ movq mm4, mm0
+ movq mm7, mm2
+ movq2dq xmm3, mm2
+ PRED4x4_LOWPASS mm1, mm3, mm0, mm4, mm5
+ psllq mm1, 56
+ PALIGNR mm7, mm1, 7, mm3
+ movq2dq xmm1, mm7
+ movq mm0, [r0-8]
+ movq mm3, [r0]
+ movq mm1, [r0+8]
+ movq mm2, mm3
+ movq mm4, mm3
+ PALIGNR mm2, mm0, 7, mm0
+ PALIGNR mm1, mm4, 1, mm4
+ test r1, r1
+ jz .fix_lt_2
+ test r2, r2
+ jz .fix_tr_1
+.do_top:
+ PRED4x4_LOWPASS mm4, mm2, mm1, mm3, mm5
+ movq2dq xmm4, mm4
+ lea r1, [r0+r3*2]
+ movdqa xmm0, xmm3
+ pslldq xmm4, 8
+ por xmm3, xmm4
+ lea r2, [r1+r3*2]
+ pslldq xmm4, 1
+ por xmm1, xmm4
+ psrldq xmm0, 7
+ pslldq xmm0, 15
+ psrldq xmm0, 7
+ por xmm1, xmm0
+ lea r0, [r2+r3*2]
+ movdqa xmm2, xmm3
+ psrldq xmm2, 1
+INIT_XMM cpuname
+ PRED4x4_LOWPASS xmm0, xmm1, xmm2, xmm3, xmm4
+ movdqa xmm1, xmm0
+ psrldq xmm1, 1
+ movq [r0+r3*2], xmm0
+ movq [r0+r3*1], xmm1
+ psrldq xmm0, 2
+ psrldq xmm1, 2
+ movq [r2+r3*2], xmm0
+ movq [r2+r3*1], xmm1
+ psrldq xmm0, 2
+ psrldq xmm1, 2
+ movq [r1+r3*2], xmm0
+ movq [r1+r3*1], xmm1
+ psrldq xmm0, 2
+ psrldq xmm1, 2
+ movq [r4+r3*2], xmm0
+ movq [r4+r3*1], xmm1
+ RET
+%endmacro
+
+INIT_MMX sse2
+PRED8x8L_DOWN_RIGHT
+INIT_MMX ssse3
+PRED8x8L_DOWN_RIGHT
+
+;-----------------------------------------------------------------------------
+; void pred8x8l_vertical_right_8(uint8_t *src, int has_topleft, int has_topright, int stride)
+;-----------------------------------------------------------------------------
+
+INIT_MMX mmxext
+cglobal pred8x8l_vertical_right_8, 4,5
+ sub r0, r3
+ lea r4, [r0+r3*2]
+ movq mm0, [r0+r3*1-8]
+ punpckhbw mm0, [r0+r3*0-8]
+ movq mm1, [r4+r3*1-8]
+ punpckhbw mm1, [r0+r3*2-8]
+ mov r4, r0
+ punpckhwd mm1, mm0
+ lea r0, [r0+r3*4]
+ movq mm2, [r0+r3*1-8]
+ punpckhbw mm2, [r0+r3*0-8]
+ lea r0, [r0+r3*2]
+ movq mm3, [r0+r3*1-8]
+ punpckhbw mm3, [r0+r3*0-8]
+ punpckhwd mm3, mm2
+ punpckhdq mm3, mm1
+ lea r0, [r0+r3*2]
+ movq mm0, [r0+r3*0-8]
+ movq mm1, [r4]
+ mov r0, r4
+ movq mm4, mm3
+ movq mm2, mm3
+ PALIGNR mm4, mm0, 7, mm0
+ PALIGNR mm1, mm2, 1, mm2
+ test r1, r1
+ jz .fix_lt_1
+ jmp .do_left
+.fix_lt_1:
+ movq mm5, mm3
+ pxor mm5, mm4
+ psrlq mm5, 56
+ psllq mm5, 48
+ pxor mm1, mm5
+ jmp .do_left
+.fix_lt_2:
+ movq mm5, mm3
+ pxor mm5, mm2
+ psllq mm5, 56
+ psrlq mm5, 56
+ pxor mm2, mm5
+ test r2, r2
+ jnz .do_top
+.fix_tr_1:
+ movq mm5, mm3
+ pxor mm5, mm1
+ psrlq mm5, 56
+ psllq mm5, 56
+ pxor mm1, mm5
+ jmp .do_top
+.do_left:
+ movq mm0, mm4
+ PRED4x4_LOWPASS mm2, mm1, mm4, mm3, mm5
+ movq mm7, mm2
+ movq mm0, [r0-8]
+ movq mm3, [r0]
+ movq mm1, [r0+8]
+ movq mm2, mm3
+ movq mm4, mm3
+ PALIGNR mm2, mm0, 7, mm0
+ PALIGNR mm1, mm4, 1, mm4
+ test r1, r1
+ jz .fix_lt_2
+ test r2, r2
+ jz .fix_tr_1
+.do_top:
+ PRED4x4_LOWPASS mm6, mm2, mm1, mm3, mm5
+ lea r1, [r0+r3*2]
+ movq mm2, mm6
+ movq mm3, mm6
+ PALIGNR mm3, mm7, 7, mm0
+ PALIGNR mm6, mm7, 6, mm1
+ movq mm4, mm3
+ pavgb mm3, mm2
+ lea r2, [r1+r3*2]
+ PRED4x4_LOWPASS mm0, mm6, mm2, mm4, mm5
+ movq [r0+r3*1], mm3
+ movq [r0+r3*2], mm0
+ movq mm5, mm0
+ movq mm6, mm3
+ movq mm1, mm7
+ movq mm2, mm1
+ psllq mm2, 8
+ movq mm3, mm1
+ psllq mm3, 16
+ lea r4, [r2+r3*2]
+ PRED4x4_LOWPASS mm0, mm1, mm3, mm2, mm4
+ PALIGNR mm6, mm0, 7, mm2
+ movq [r1+r3*1], mm6
+ psllq mm0, 8
+ PALIGNR mm5, mm0, 7, mm1
+ movq [r1+r3*2], mm5
+ psllq mm0, 8
+ PALIGNR mm6, mm0, 7, mm2
+ movq [r2+r3*1], mm6
+ psllq mm0, 8
+ PALIGNR mm5, mm0, 7, mm1
+ movq [r2+r3*2], mm5
+ psllq mm0, 8
+ PALIGNR mm6, mm0, 7, mm2
+ movq [r4+r3*1], mm6
+ psllq mm0, 8
+ PALIGNR mm5, mm0, 7, mm1
+ movq [r4+r3*2], mm5
+ RET
+
+%macro PRED8x8L_VERTICAL_RIGHT 0
+cglobal pred8x8l_vertical_right_8, 4,5,7
+ ; manually spill XMM registers for Win64 because
+ ; the code here is initialized with INIT_MMX
+ WIN64_SPILL_XMM 7
+ sub r0, r3
+ lea r4, [r0+r3*2]
+ movq mm0, [r0+r3*1-8]
+ punpckhbw mm0, [r0+r3*0-8]
+ movq mm1, [r4+r3*1-8]
+ punpckhbw mm1, [r0+r3*2-8]
+ mov r4, r0
+ punpckhwd mm1, mm0
+ lea r0, [r0+r3*4]
+ movq mm2, [r0+r3*1-8]
+ punpckhbw mm2, [r0+r3*0-8]
+ lea r0, [r0+r3*2]
+ movq mm3, [r0+r3*1-8]
+ punpckhbw mm3, [r0+r3*0-8]
+ punpckhwd mm3, mm2
+ punpckhdq mm3, mm1
+ lea r0, [r0+r3*2]
+ movq mm0, [r0+r3*0-8]
+ movq mm1, [r4]
+ mov r0, r4
+ movq mm4, mm3
+ movq mm2, mm3
+ PALIGNR mm4, mm0, 7, mm0
+ PALIGNR mm1, mm2, 1, mm2
+ test r1, r1
+ jnz .do_left
+.fix_lt_1:
+ movq mm5, mm3
+ pxor mm5, mm4
+ psrlq mm5, 56
+ psllq mm5, 48
+ pxor mm1, mm5
+ jmp .do_left
+.fix_lt_2:
+ movq mm5, mm3
+ pxor mm5, mm2
+ psllq mm5, 56
+ psrlq mm5, 56
+ pxor mm2, mm5
+ test r2, r2
+ jnz .do_top
+.fix_tr_1:
+ movq mm5, mm3
+ pxor mm5, mm1
+ psrlq mm5, 56
+ psllq mm5, 56
+ pxor mm1, mm5
+ jmp .do_top
+.do_left:
+ movq mm0, mm4
+ PRED4x4_LOWPASS mm2, mm1, mm4, mm3, mm5
+ movq2dq xmm0, mm2
+ movq mm0, [r0-8]
+ movq mm3, [r0]
+ movq mm1, [r0+8]
+ movq mm2, mm3
+ movq mm4, mm3
+ PALIGNR mm2, mm0, 7, mm0
+ PALIGNR mm1, mm4, 1, mm4
+ test r1, r1
+ jz .fix_lt_2
+ test r2, r2
+ jz .fix_tr_1
+.do_top:
+ PRED4x4_LOWPASS mm6, mm2, mm1, mm3, mm5
+ lea r1, [r0+r3*2]
+ movq2dq xmm4, mm6
+ pslldq xmm4, 8
+ por xmm0, xmm4
+ movdqa xmm6, [pw_ff00]
+ movdqa xmm1, xmm0
+ lea r2, [r1+r3*2]
+ movdqa xmm2, xmm0
+ movdqa xmm3, xmm0
+ pslldq xmm0, 1
+ pslldq xmm1, 2
+ pavgb xmm2, xmm0
+INIT_XMM cpuname
+ PRED4x4_LOWPASS xmm4, xmm3, xmm1, xmm0, xmm5
+ pandn xmm6, xmm4
+ movdqa xmm5, xmm4
+ psrlw xmm4, 8
+ packuswb xmm6, xmm4
+ movhlps xmm4, xmm6
+ movhps [r0+r3*2], xmm5
+ movhps [r0+r3*1], xmm2
+ psrldq xmm5, 4
+ movss xmm5, xmm6
+ psrldq xmm2, 4
+ movss xmm2, xmm4
+ lea r0, [r2+r3*2]
+ psrldq xmm5, 1
+ psrldq xmm2, 1
+ movq [r0+r3*2], xmm5
+ movq [r0+r3*1], xmm2
+ psrldq xmm5, 1
+ psrldq xmm2, 1
+ movq [r2+r3*2], xmm5
+ movq [r2+r3*1], xmm2
+ psrldq xmm5, 1
+ psrldq xmm2, 1
+ movq [r1+r3*2], xmm5
+ movq [r1+r3*1], xmm2
+ RET
+%endmacro
+
+INIT_MMX sse2
+PRED8x8L_VERTICAL_RIGHT
+INIT_MMX ssse3
+PRED8x8L_VERTICAL_RIGHT
+
+;-----------------------------------------------------------------------------
+;void pred8x8l_vertical_left_8(uint8_t *src, int has_topleft, int has_topright, int stride)
+;-----------------------------------------------------------------------------
+
+%macro PRED8x8L_VERTICAL_LEFT 0
+cglobal pred8x8l_vertical_left_8, 4,4
+ sub r0, r3
+ movq mm0, [r0-8]
+ movq mm3, [r0]
+ movq mm1, [r0+8]
+ movq mm2, mm3
+ movq mm4, mm3
+ PALIGNR mm2, mm0, 7, mm0
+ PALIGNR mm1, mm4, 1, mm4
+ test r1, r1
+ jz .fix_lt_2
+ test r2, r2
+ jz .fix_tr_1
+ jmp .do_top
+.fix_lt_2:
+ movq mm5, mm3
+ pxor mm5, mm2
+ psllq mm5, 56
+ psrlq mm5, 56
+ pxor mm2, mm5
+ test r2, r2
+ jnz .do_top
+.fix_tr_1:
+ movq mm5, mm3
+ pxor mm5, mm1
+ psrlq mm5, 56
+ psllq mm5, 56
+ pxor mm1, mm5
+ jmp .do_top
+.fix_tr_2:
+ punpckhbw mm3, mm3
+ pshufw mm1, mm3, 0xFF
+ jmp .do_topright
+.do_top:
+ PRED4x4_LOWPASS mm4, mm2, mm1, mm3, mm5
+ movq2dq xmm4, mm4
+ test r2, r2
+ jz .fix_tr_2
+ movq mm0, [r0+8]
+ movq mm5, mm0
+ movq mm2, mm0
+ movq mm4, mm0
+ psrlq mm5, 56
+ PALIGNR mm2, mm3, 7, mm3
+ PALIGNR mm5, mm4, 1, mm4
+ PRED4x4_LOWPASS mm1, mm2, mm5, mm0, mm4
+.do_topright:
+ movq2dq xmm3, mm1
+ lea r1, [r0+r3*2]
+ pslldq xmm3, 8
+ por xmm4, xmm3
+ movdqa xmm2, xmm4
+ movdqa xmm1, xmm4
+ movdqa xmm3, xmm4
+ psrldq xmm2, 1
+ pslldq xmm1, 1
+ pavgb xmm3, xmm2
+ lea r2, [r1+r3*2]
+INIT_XMM cpuname
+ PRED4x4_LOWPASS xmm0, xmm1, xmm2, xmm4, xmm5
+ psrldq xmm0, 1
+ movq [r0+r3*1], xmm3
+ movq [r0+r3*2], xmm0
+ lea r0, [r2+r3*2]
+ psrldq xmm3, 1
+ psrldq xmm0, 1
+ movq [r1+r3*1], xmm3
+ movq [r1+r3*2], xmm0
+ psrldq xmm3, 1
+ psrldq xmm0, 1
+ movq [r2+r3*1], xmm3
+ movq [r2+r3*2], xmm0
+ psrldq xmm3, 1
+ psrldq xmm0, 1
+ movq [r0+r3*1], xmm3
+ movq [r0+r3*2], xmm0
+ RET
+%endmacro
+
+INIT_MMX sse2
+PRED8x8L_VERTICAL_LEFT
+INIT_MMX ssse3
+PRED8x8L_VERTICAL_LEFT
+
+;-----------------------------------------------------------------------------
+; void pred8x8l_horizontal_up_8(uint8_t *src, int has_topleft, int has_topright, int stride)