ffmpeg/libavcodec/x86/dsputil.asm

1594 lines
40 KiB
NASM

;******************************************************************************
;* MMX optimized DSP utils
;* Copyright (c) 2008 Loren Merritt
;*
;* This file is part of FFmpeg.
;*
;* FFmpeg is free software; you can redistribute it and/or
;* modify it under the terms of the GNU Lesser General Public
;* License as published by the Free Software Foundation; either
;* version 2.1 of the License, or (at your option) any later version.
;*
;* FFmpeg is distributed in the hope that it will be useful,
;* but WITHOUT ANY WARRANTY; without even the implied warranty of
;* MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
;* Lesser General Public License for more details.
;*
;* You should have received a copy of the GNU Lesser General Public
;* License along with FFmpeg; if not, write to the Free Software
;* Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
;******************************************************************************
%include "libavutil/x86/x86util.asm"
SECTION_RODATA
pb_f: times 16 db 15
pb_zzzzzzzz77777777: times 8 db -1
pb_7: times 8 db 7
pb_zzzz3333zzzzbbbb: db -1,-1,-1,-1,3,3,3,3,-1,-1,-1,-1,11,11,11,11
pb_zz11zz55zz99zzdd: db -1,-1,1,1,-1,-1,5,5,-1,-1,9,9,-1,-1,13,13
pb_revwords: SHUFFLE_MASK_W 7, 6, 5, 4, 3, 2, 1, 0
pd_16384: times 4 dd 16384
pb_bswap32: db 3, 2, 1, 0, 7, 6, 5, 4, 11, 10, 9, 8, 15, 14, 13, 12
SECTION_TEXT
%macro SCALARPRODUCT 0
; int scalarproduct_int16(int16_t *v1, int16_t *v2, int order)
cglobal scalarproduct_int16, 3,3,3, v1, v2, order
shl orderq, 1
add v1q, orderq
add v2q, orderq
neg orderq
pxor m2, m2
.loop:
movu m0, [v1q + orderq]
movu m1, [v1q + orderq + mmsize]
pmaddwd m0, [v2q + orderq]
pmaddwd m1, [v2q + orderq + mmsize]
paddd m2, m0
paddd m2, m1
add orderq, mmsize*2
jl .loop
%if mmsize == 16
movhlps m0, m2
paddd m2, m0
pshuflw m0, m2, 0x4e
%else
pshufw m0, m2, 0x4e
%endif
paddd m2, m0
movd eax, m2
RET
; int scalarproduct_and_madd_int16(int16_t *v1, int16_t *v2, int16_t *v3, int order, int mul)
cglobal scalarproduct_and_madd_int16, 4,4,8, v1, v2, v3, order, mul
shl orderq, 1
movd m7, mulm
%if mmsize == 16
pshuflw m7, m7, 0
punpcklqdq m7, m7
%else
pshufw m7, m7, 0
%endif
pxor m6, m6
add v1q, orderq
add v2q, orderq
add v3q, orderq
neg orderq
.loop:
movu m0, [v2q + orderq]
movu m1, [v2q + orderq + mmsize]
mova m4, [v1q + orderq]
mova m5, [v1q + orderq + mmsize]
movu m2, [v3q + orderq]
movu m3, [v3q + orderq + mmsize]
pmaddwd m0, m4
pmaddwd m1, m5
pmullw m2, m7
pmullw m3, m7
paddd m6, m0
paddd m6, m1
paddw m2, m4
paddw m3, m5
mova [v1q + orderq], m2
mova [v1q + orderq + mmsize], m3
add orderq, mmsize*2
jl .loop
%if mmsize == 16
movhlps m0, m6
paddd m6, m0
pshuflw m0, m6, 0x4e
%else
pshufw m0, m6, 0x4e
%endif
paddd m6, m0
movd eax, m6
RET
%endmacro
INIT_MMX mmxext
SCALARPRODUCT
INIT_XMM sse2
SCALARPRODUCT
%macro SCALARPRODUCT_LOOP 1
align 16
.loop%1:
sub orderq, mmsize*2
%if %1
mova m1, m4
mova m4, [v2q + orderq]
mova m0, [v2q + orderq + mmsize]
palignr m1, m0, %1
palignr m0, m4, %1
mova m3, m5
mova m5, [v3q + orderq]
mova m2, [v3q + orderq + mmsize]
palignr m3, m2, %1
palignr m2, m5, %1
%else
mova m0, [v2q + orderq]
mova m1, [v2q + orderq + mmsize]
mova m2, [v3q + orderq]
mova m3, [v3q + orderq + mmsize]
%endif
%define t0 [v1q + orderq]
%define t1 [v1q + orderq + mmsize]
%if ARCH_X86_64
mova m8, t0
mova m9, t1
%define t0 m8
%define t1 m9
%endif
pmaddwd m0, t0
pmaddwd m1, t1
pmullw m2, m7
pmullw m3, m7
paddw m2, t0
paddw m3, t1
paddd m6, m0
paddd m6, m1
mova [v1q + orderq], m2
mova [v1q + orderq + mmsize], m3
jg .loop%1
%if %1
jmp .end
%endif
%endmacro
; int scalarproduct_and_madd_int16(int16_t *v1, int16_t *v2, int16_t *v3, int order, int mul)
INIT_XMM ssse3
cglobal scalarproduct_and_madd_int16, 4,5,10, v1, v2, v3, order, mul
shl orderq, 1
movd m7, mulm
pshuflw m7, m7, 0
punpcklqdq m7, m7
pxor m6, m6
mov r4d, v2d
and r4d, 15
and v2q, ~15
and v3q, ~15
mova m4, [v2q + orderq]
mova m5, [v3q + orderq]
; linear is faster than branch tree or jump table, because the branches taken are cyclic (i.e. predictable)
cmp r4d, 0
je .loop0
cmp r4d, 2
je .loop2
cmp r4d, 4
je .loop4
cmp r4d, 6
je .loop6
cmp r4d, 8
je .loop8
cmp r4d, 10
je .loop10
cmp r4d, 12
je .loop12
SCALARPRODUCT_LOOP 14
SCALARPRODUCT_LOOP 12
SCALARPRODUCT_LOOP 10
SCALARPRODUCT_LOOP 8
SCALARPRODUCT_LOOP 6
SCALARPRODUCT_LOOP 4
SCALARPRODUCT_LOOP 2
SCALARPRODUCT_LOOP 0
.end:
movhlps m0, m6
paddd m6, m0
pshuflw m0, m6, 0x4e
paddd m6, m0
movd eax, m6
RET
;-----------------------------------------------------------------------------
; void ff_apply_window_int16(int16_t *output, const int16_t *input,
; const int16_t *window, unsigned int len)
;-----------------------------------------------------------------------------
%macro REVERSE_WORDS 1-2
%if cpuflag(ssse3) && notcpuflag(atom)
pshufb %1, %2
%elif cpuflag(sse2)
pshuflw %1, %1, 0x1B
pshufhw %1, %1, 0x1B
pshufd %1, %1, 0x4E
%elif cpuflag(mmxext)
pshufw %1, %1, 0x1B
%endif
%endmacro
%macro MUL16FIXED 3
%if cpuflag(ssse3) ; dst, src, unused
; dst = ((dst * src) + (1<<14)) >> 15
pmulhrsw %1, %2
%elif cpuflag(mmxext) ; dst, src, temp
; dst = (dst * src) >> 15
; pmulhw cuts off the bottom bit, so we have to lshift by 1 and add it back
; in from the pmullw result.
mova %3, %1
pmulhw %1, %2
pmullw %3, %2
psrlw %3, 15
psllw %1, 1
por %1, %3
%endif
%endmacro
%macro APPLY_WINDOW_INT16 1 ; %1 bitexact version
%if %1
cglobal apply_window_int16, 4,5,6, output, input, window, offset, offset2
%else
cglobal apply_window_int16_round, 4,5,6, output, input, window, offset, offset2
%endif
lea offset2q, [offsetq-mmsize]
%if cpuflag(ssse3) && notcpuflag(atom)
mova m5, [pb_revwords]
ALIGN 16
%elif %1
mova m5, [pd_16384]
%endif
.loop:
%if cpuflag(ssse3)
; This version does the 16x16->16 multiplication in-place without expanding
; to 32-bit. The ssse3 version is bit-identical.
mova m0, [windowq+offset2q]
mova m1, [ inputq+offset2q]
pmulhrsw m1, m0
REVERSE_WORDS m0, m5
pmulhrsw m0, [ inputq+offsetq ]
mova [outputq+offset2q], m1
mova [outputq+offsetq ], m0
%elif %1
; This version expands 16-bit to 32-bit, multiplies by the window,
; adds 16384 for rounding, right shifts 15, then repacks back to words to
; save to the output. The window is reversed for the second half.
mova m3, [windowq+offset2q]
mova m4, [ inputq+offset2q]
pxor m0, m0
punpcklwd m0, m3
punpcklwd m1, m4
pmaddwd m0, m1
paddd m0, m5
psrad m0, 15
pxor m2, m2
punpckhwd m2, m3
punpckhwd m1, m4
pmaddwd m2, m1
paddd m2, m5
psrad m2, 15
packssdw m0, m2
mova [outputq+offset2q], m0
REVERSE_WORDS m3
mova m4, [ inputq+offsetq]
pxor m0, m0
punpcklwd m0, m3
punpcklwd m1, m4
pmaddwd m0, m1
paddd m0, m5
psrad m0, 15
pxor m2, m2
punpckhwd m2, m3
punpckhwd m1, m4
pmaddwd m2, m1
paddd m2, m5
psrad m2, 15
packssdw m0, m2
mova [outputq+offsetq], m0
%else
; This version does the 16x16->16 multiplication in-place without expanding
; to 32-bit. The mmxext and sse2 versions do not use rounding, and
; therefore are not bit-identical to the C version.
mova m0, [windowq+offset2q]
mova m1, [ inputq+offset2q]
mova m2, [ inputq+offsetq ]
MUL16FIXED m1, m0, m3
REVERSE_WORDS m0
MUL16FIXED m2, m0, m3
mova [outputq+offset2q], m1
mova [outputq+offsetq ], m2
%endif
add offsetd, mmsize
sub offset2d, mmsize
jae .loop
REP_RET
%endmacro
INIT_MMX mmxext
APPLY_WINDOW_INT16 0
INIT_XMM sse2
APPLY_WINDOW_INT16 0
INIT_MMX mmxext
APPLY_WINDOW_INT16 1
INIT_XMM sse2
APPLY_WINDOW_INT16 1
INIT_XMM ssse3
APPLY_WINDOW_INT16 1
INIT_XMM ssse3, atom
APPLY_WINDOW_INT16 1
; void add_hfyu_median_prediction_mmxext(uint8_t *dst, const uint8_t *top, const uint8_t *diff, int w, int *left, int *left_top)
INIT_MMX mmxext
cglobal add_hfyu_median_prediction, 6,6,0, dst, top, diff, w, left, left_top
movq mm0, [topq]
movq mm2, mm0
movd mm4, [left_topq]
psllq mm2, 8
movq mm1, mm0
por mm4, mm2
movd mm3, [leftq]
psubb mm0, mm4 ; t-tl
add dstq, wq
add topq, wq
add diffq, wq
neg wq
jmp .skip
.loop:
movq mm4, [topq+wq]
movq mm0, mm4
psllq mm4, 8
por mm4, mm1
movq mm1, mm0 ; t
psubb mm0, mm4 ; t-tl
.skip:
movq mm2, [diffq+wq]
%assign i 0
%rep 8
movq mm4, mm0
paddb mm4, mm3 ; t-tl+l
movq mm5, mm3
pmaxub mm3, mm1
pminub mm5, mm1
pminub mm3, mm4
pmaxub mm3, mm5 ; median
paddb mm3, mm2 ; +residual
%if i==0
movq mm7, mm3
psllq mm7, 56
%else
movq mm6, mm3
psrlq mm7, 8
psllq mm6, 56
por mm7, mm6
%endif
%if i<7
psrlq mm0, 8
psrlq mm1, 8
psrlq mm2, 8
%endif
%assign i i+1
%endrep
movq [dstq+wq], mm7
add wq, 8
jl .loop
movzx r2d, byte [dstq-1]
mov [leftq], r2d
movzx r2d, byte [topq-1]
mov [left_topq], r2d
RET
%macro ADD_HFYU_LEFT_LOOP 2 ; %1 = dst_is_aligned, %2 = src_is_aligned
add srcq, wq
add dstq, wq
neg wq
%%.loop:
%if %2
mova m1, [srcq+wq]
%else
movu m1, [srcq+wq]
%endif
mova m2, m1
psllw m1, 8
paddb m1, m2
mova m2, m1
pshufb m1, m3
paddb m1, m2
pshufb m0, m5
mova m2, m1
pshufb m1, m4
paddb m1, m2
%if mmsize == 16
mova m2, m1
pshufb m1, m6
paddb m1, m2
%endif
paddb m0, m1
%if %1
mova [dstq+wq], m0
%else
movq [dstq+wq], m0
movhps [dstq+wq+8], m0
%endif
add wq, mmsize
jl %%.loop
mov eax, mmsize-1
sub eax, wd
movd m1, eax
pshufb m0, m1
movd eax, m0
RET
%endmacro
; int add_hfyu_left_prediction(uint8_t *dst, const uint8_t *src, int w, int left)
INIT_MMX ssse3
cglobal add_hfyu_left_prediction, 3,3,7, dst, src, w, left
.skip_prologue:
mova m5, [pb_7]
mova m4, [pb_zzzz3333zzzzbbbb]
mova m3, [pb_zz11zz55zz99zzdd]
movd m0, leftm
psllq m0, 56
ADD_HFYU_LEFT_LOOP 1, 1
INIT_XMM sse4
cglobal add_hfyu_left_prediction, 3,3,7, dst, src, w, left
mova m5, [pb_f]
mova m6, [pb_zzzzzzzz77777777]
mova m4, [pb_zzzz3333zzzzbbbb]
mova m3, [pb_zz11zz55zz99zzdd]
movd m0, leftm
pslldq m0, 15
test srcq, 15
jnz .src_unaligned
test dstq, 15
jnz .dst_unaligned
ADD_HFYU_LEFT_LOOP 1, 1
.dst_unaligned:
ADD_HFYU_LEFT_LOOP 0, 1
.src_unaligned:
ADD_HFYU_LEFT_LOOP 0, 0
; float scalarproduct_float_sse(const float *v1, const float *v2, int len)
INIT_XMM sse
cglobal scalarproduct_float, 3,3,2, v1, v2, offset
neg offsetq
shl offsetq, 2
sub v1q, offsetq
sub v2q, offsetq
xorps xmm0, xmm0
.loop:
movaps xmm1, [v1q+offsetq]
mulps xmm1, [v2q+offsetq]
addps xmm0, xmm1
add offsetq, 16
js .loop
movhlps xmm1, xmm0
addps xmm0, xmm1
movss xmm1, xmm0
shufps xmm0, xmm0, 1
addss xmm0, xmm1
%if ARCH_X86_64 == 0
movss r0m, xmm0
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 0
%if ARCH_X86_64
%define w_reg r7
cglobal emu_edge_core, 6, 9, 1
mov r8, r5 ; save block_h
%else
%define w_reg r6
cglobal emu_edge_core, 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
%if ARCH_X86_64
sub r4, r3
%else
sub r4, dword r3m
%endif
cmp w_reg, 22
jg .slow_v_extend_loop
%if 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
%if ARCH_X86_64
mov r3, r0 ; backup of buf+block_h*linesize
mov r5, r8
%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,r7(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,r8/r5=block_h,r6/r3=val, r7/r6=end_x, r1=block_w
.right_extend:
%if 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
%if ARCH_X86_64
%define vall al
%define valh ah
%define valw ax
%define valw2 r7w
%define valw3 r3w
%if WIN64
%define valw4 r7w
%else ; unix64
%define valw4 r3w
%endif
%define vald eax
%else
%define vall bl
%define valh bh
%define valw bx
%define valw2 r6w
%define valw3 valw2
%define valw4 valw3
%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 r7/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 2
%assign %%src_off 0 ; offset in source buffer
%assign %%smidx 0 ; mmx register idx
%assign %%sxidx 0 ; xmm register idx
%if cpuflag(sse)
%rep %2/16
movups xmm %+ %%sxidx, [r1+%%src_off]
%assign %%src_off %%src_off+16
%assign %%sxidx %%sxidx+1
%endrep ; %2/16
%endif
%if 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]
%elifidn %1, body
mov valw3, [r1+%%src_off]
%elifidn %1, bottom
mov valw4, [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 2
%assign %%dst_off 0 ; offset in destination buffer
%assign %%dmidx 0 ; mmx register idx
%assign %%dxidx 0 ; xmm register idx
%if cpuflag(sse)
%rep %2/16
movups [r0+%%dst_off], xmm %+ %%dxidx
%assign %%dst_off %%dst_off+16
%assign %%dxidx %%dxidx+1
%endrep ; %2/16
%endif
%if 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
%elifidn %1, body
mov [r0+%%dst_off], valw3
%elifidn %1, bottom
mov [r0+%%dst_off], valw4
%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 0
%assign %%n 1
%rep 22
ALIGN 128
.emuedge_v_extend_ %+ %%n:
; extend pixels above body
%if 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 ; read bytes
.emuedge_extend_top_ %+ %%n %+ _loop: ; do {
WRITE_NUM_BYTES top, %%n ; write bytes
add r0 , r2 ; dst += linesize
%if ARCH_X86_64
dec r3d
%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 ; read bytes
WRITE_NUM_BYTES body, %%n ; write bytes
add r0 , r2 ; dst += linesize
add r1 , r2 ; src += linesize
dec r4d
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 ; read bytes
.emuedge_extend_bottom_ %+ %%n %+ _loop: ; do {
WRITE_NUM_BYTES bottom, %%n ; write bytes
add r0 , r2 ; dst += linesize
dec r5d
jnz .emuedge_extend_bottom_ %+ %%n %+ _loop ; } while (--block_h)
.emuedge_v_extend_end_ %+ %%n:
%if 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 2
mov vall, %2
mov valh, vall
%if %1 >= 8
movd mm0, vald
%if cpuflag(mmxext)
pshufw mm0, mm0, 0
%else ; mmx
punpcklwd mm0, mm0
punpckldq mm0, mm0
%endif ; sse
%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 0
%assign %%n 2
%rep 11
ALIGN 64
.emuedge_extend_left_ %+ %%n: ; do {
sub r0, r2 ; dst -= linesize
READ_V_PIXEL %%n, [r0+r1] ; read pixels
WRITE_V_PIXEL %%n, r0 ; write pixels
dec r5
jnz .emuedge_extend_left_ %+ %%n ; } while (--block_h)
%if 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, r8/r5=block_h, r0/r6=end_x, r6/r3=val
%macro RIGHT_EXTEND 0
%assign %%n 2
%rep 11
ALIGN 64
.emuedge_extend_right_ %+ %%n: ; do {
%if ARCH_X86_64
sub r3, r2 ; dst -= linesize
READ_V_PIXEL %%n, [r3+w_reg-1] ; read pixels
WRITE_V_PIXEL %%n, r3+r4-%%n ; write pixels
dec r8
%else ; ARCH_X86_32
sub r0, r2 ; dst -= linesize
READ_V_PIXEL %%n, [r0+w_reg-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)
%if ARCH_X86_64
ret
%else ; ARCH_X86_32
rep ret
%endif ; ARCH_X86_64/32
%assign %%n %%n+2
%endrep
%if 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 2
%ifidn %1, bottom
sub r1, linesize
%endif
.%1_copy_loop:
xor cnt_reg, cnt_reg
%if notcpuflag(sse)
%define linesize r2m
V_COPY_NPX %1, mm0, movq, 8, 0xFFFFFFF8
%else ; sse
V_COPY_NPX %1, xmm0, movups, 16, 0xFFFFFFF0
%if 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 ; sse
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 0
.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
; r8(64)/r3(later-64)/r2(32)=cnt_reg,r6(64)/r3(32)=val_reg,r7(64)/r6(32)=w=end_x-start_x
%if ARCH_X86_64
push r8 ; save old value of block_h
test r3, r3
%define cnt_reg r8
jz .do_body_copy ; if (!start_y) goto do_body_copy
V_COPY_ROW top, r3
%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
%endif
.do_body_copy:
V_COPY_ROW body, r4
%if ARCH_X86_64
pop r8 ; restore old value of block_h
%define cnt_reg r3
%endif
test r5, r5
%if ARCH_X86_64
jz .v_extend_end
%else
jz .skip_bottom_extend
%endif
V_COPY_ROW bottom, r5
%if ARCH_X86_32
.skip_bottom_extend:
mov r2, r2m
%endif
jmp .v_extend_end
%endmacro
%macro SLOW_LEFT_EXTEND 0
.slow_left_extend_loop:
; r0=buf+block_h*linesize,r2=linesize,r6(64)/r3(32)=val,r5=block_h,r4=cntr,r7/r6=start_x
mov r4, 8
sub r0, linesize
READ_V_PIXEL 8, [r0+w_reg]
.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
%if ARCH_X86_32
mov r2, r2m
%endif
jmp .right_extend
%endmacro
%macro SLOW_RIGHT_EXTEND 0
.slow_right_extend_loop:
; r3(64)/r0(32)=buf+block_h*linesize,r2=linesize,r4=block_w,r8(64)/r5(32)=block_h,
; r7(64)/r6(32)=end_x,r6/r3=val,r1=cntr
%if ARCH_X86_64
%define buf_reg r3
%define bh_reg r8
%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]
.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
INIT_XMM %1
EMU_EDGE_FUNC
VERTICAL_EXTEND
LEFT_EXTEND
RIGHT_EXTEND
SLOW_V_EXTEND
SLOW_LEFT_EXTEND
SLOW_RIGHT_EXTEND
%endmacro
emu_edge sse
%if ARCH_X86_32
emu_edge mmx
%endif
;-----------------------------------------------------------------------------
; void ff_vector_clip_int32(int32_t *dst, const int32_t *src, int32_t min,
; int32_t max, unsigned int len)
;-----------------------------------------------------------------------------
; %1 = number of xmm registers used
; %2 = number of inline load/process/store loops per asm loop
; %3 = process 4*mmsize (%3=0) or 8*mmsize (%3=1) bytes per loop
; %4 = CLIPD function takes min/max as float instead of int (CLIPD_SSE2)
; %5 = suffix
%macro VECTOR_CLIP_INT32 4-5
cglobal vector_clip_int32%5, 5,5,%1, dst, src, min, max, len
%if %4
cvtsi2ss m4, minm
cvtsi2ss m5, maxm
%else
movd m4, minm
movd m5, maxm
%endif
SPLATD m4
SPLATD m5
.loop:
%assign %%i 1
%rep %2
mova m0, [srcq+mmsize*0*%%i]
mova m1, [srcq+mmsize*1*%%i]
mova m2, [srcq+mmsize*2*%%i]
mova m3, [srcq+mmsize*3*%%i]
%if %3
mova m7, [srcq+mmsize*4*%%i]
mova m8, [srcq+mmsize*5*%%i]
mova m9, [srcq+mmsize*6*%%i]
mova m10, [srcq+mmsize*7*%%i]
%endif
CLIPD m0, m4, m5, m6
CLIPD m1, m4, m5, m6
CLIPD m2, m4, m5, m6
CLIPD m3, m4, m5, m6
%if %3
CLIPD m7, m4, m5, m6
CLIPD m8, m4, m5, m6
CLIPD m9, m4, m5, m6
CLIPD m10, m4, m5, m6
%endif
mova [dstq+mmsize*0*%%i], m0
mova [dstq+mmsize*1*%%i], m1
mova [dstq+mmsize*2*%%i], m2
mova [dstq+mmsize*3*%%i], m3
%if %3
mova [dstq+mmsize*4*%%i], m7
mova [dstq+mmsize*5*%%i], m8
mova [dstq+mmsize*6*%%i], m9
mova [dstq+mmsize*7*%%i], m10
%endif
%assign %%i %%i+1
%endrep
add srcq, mmsize*4*(%2+%3)
add dstq, mmsize*4*(%2+%3)
sub lend, mmsize*(%2+%3)
jg .loop
REP_RET
%endmacro
INIT_MMX mmx
%define CLIPD CLIPD_MMX
VECTOR_CLIP_INT32 0, 1, 0, 0
INIT_XMM sse2
VECTOR_CLIP_INT32 6, 1, 0, 0, _int
%define CLIPD CLIPD_SSE2
VECTOR_CLIP_INT32 6, 2, 0, 1
INIT_XMM sse4
%define CLIPD CLIPD_SSE41
%ifdef m8
VECTOR_CLIP_INT32 11, 1, 1, 0
%else
VECTOR_CLIP_INT32 6, 1, 0, 0
%endif
;-----------------------------------------------------------------------------
; void vector_fmul_reverse(float *dst, const float *src0, const float *src1,
; int len)
;-----------------------------------------------------------------------------
%macro VECTOR_FMUL_REVERSE 0
cglobal vector_fmul_reverse, 4,4,2, dst, src0, src1, len
lea lenq, [lend*4 - 2*mmsize]
ALIGN 16
.loop:
%if cpuflag(avx)
vmovaps xmm0, [src1q + 16]
vinsertf128 m0, m0, [src1q], 1
vshufps m0, m0, m0, q0123
vmovaps xmm1, [src1q + mmsize + 16]
vinsertf128 m1, m1, [src1q + mmsize], 1
vshufps m1, m1, m1, q0123
%else
mova m0, [src1q]
mova m1, [src1q + mmsize]
shufps m0, m0, q0123
shufps m1, m1, q0123
%endif
mulps m0, m0, [src0q + lenq + mmsize]
mulps m1, m1, [src0q + lenq]
mova [dstq + lenq + mmsize], m0
mova [dstq + lenq], m1
add src1q, 2*mmsize
sub lenq, 2*mmsize
jge .loop
REP_RET
%endmacro
INIT_XMM sse
VECTOR_FMUL_REVERSE
%if HAVE_AVX_EXTERNAL
INIT_YMM avx
VECTOR_FMUL_REVERSE
%endif
;-----------------------------------------------------------------------------
; vector_fmul_add(float *dst, const float *src0, const float *src1,
; const float *src2, int len)
;-----------------------------------------------------------------------------
%macro VECTOR_FMUL_ADD 0
cglobal vector_fmul_add, 5,5,2, dst, src0, src1, src2, len
lea lenq, [lend*4 - 2*mmsize]
ALIGN 16
.loop:
mova m0, [src0q + lenq]
mova m1, [src0q + lenq + mmsize]
mulps m0, m0, [src1q + lenq]
mulps m1, m1, [src1q + lenq + mmsize]
addps m0, m0, [src2q + lenq]
addps m1, m1, [src2q + lenq + mmsize]
mova [dstq + lenq], m0
mova [dstq + lenq + mmsize], m1
sub lenq, 2*mmsize
jge .loop
REP_RET
%endmacro
INIT_XMM sse
VECTOR_FMUL_ADD
%if HAVE_AVX_EXTERNAL
INIT_YMM avx
VECTOR_FMUL_ADD
%endif
;-----------------------------------------------------------------------------
; void ff_butterflies_float_interleave(float *dst, const float *src0,
; const float *src1, int len);
;-----------------------------------------------------------------------------
%macro BUTTERFLIES_FLOAT_INTERLEAVE 0
cglobal butterflies_float_interleave, 4,4,3, dst, src0, src1, len
%if ARCH_X86_64
movsxd lenq, lend
%endif
test lenq, lenq
jz .end
shl lenq, 2
lea src0q, [src0q + lenq]
lea src1q, [src1q + lenq]
lea dstq, [ dstq + 2*lenq]
neg lenq
.loop:
mova m0, [src0q + lenq]
mova m1, [src1q + lenq]
subps m2, m0, m1
addps m0, m0, m1
unpcklps m1, m0, m2
unpckhps m0, m0, m2
%if cpuflag(avx)
vextractf128 [dstq + 2*lenq ], m1, 0
vextractf128 [dstq + 2*lenq + 16], m0, 0
vextractf128 [dstq + 2*lenq + 32], m1, 1
vextractf128 [dstq + 2*lenq + 48], m0, 1
%else
mova [dstq + 2*lenq ], m1
mova [dstq + 2*lenq + mmsize], m0
%endif
add lenq, mmsize
jl .loop
.end:
REP_RET
%endmacro
INIT_XMM sse
BUTTERFLIES_FLOAT_INTERLEAVE
%if HAVE_AVX_EXTERNAL
INIT_YMM avx
BUTTERFLIES_FLOAT_INTERLEAVE
%endif
; %1 = aligned/unaligned
%macro BSWAP_LOOPS 1
mov r3, r2
sar r2, 3
jz .left4_%1
.loop8_%1:
mov%1 m0, [r1 + 0]
mov%1 m1, [r1 + 16]
%if cpuflag(ssse3)
pshufb m0, m2
pshufb m1, m2
mova [r0 + 0], m0
mova [r0 + 16], m1
%else
pshuflw m0, m0, 10110001b
pshuflw m1, m1, 10110001b
pshufhw m0, m0, 10110001b
pshufhw m1, m1, 10110001b
mova m2, m0
mova m3, m1
psllw m0, 8
psllw m1, 8
psrlw m2, 8
psrlw m3, 8
por m2, m0
por m3, m1
mova [r0 + 0], m2
mova [r0 + 16], m3
%endif
add r0, 32
add r1, 32
dec r2
jnz .loop8_%1
.left4_%1:
mov r2, r3
and r3, 4
jz .left
mov%1 m0, [r1]
%if cpuflag(ssse3)
pshufb m0, m2
mova [r0], m0
%else
pshuflw m0, m0, 10110001b
pshufhw m0, m0, 10110001b
mova m2, m0
psllw m0, 8
psrlw m2, 8
por m2, m0
mova [r0], m2
%endif
add r1, 16
add r0, 16
%endmacro
; void bswap_buf(uint32_t *dst, const uint32_t *src, int w);
%macro BSWAP32_BUF 0
%if cpuflag(ssse3)
cglobal bswap32_buf, 3,4,3
mov r3, r1
mova m2, [pb_bswap32]
%else
cglobal bswap32_buf, 3,4,5
mov r3, r1
%endif
and r3, 15
jz .start_align
BSWAP_LOOPS u
jmp .left
.start_align:
BSWAP_LOOPS a
.left:
%if cpuflag(ssse3)
mov r3, r2
and r2, 2
jz .left1
movq m0, [r1]
pshufb m0, m2
movq [r0], m0
add r1, 8
add r0, 8
.left1:
and r3, 1
jz .end
mov r2d, [r1]
bswap r2d
mov [r0], r2d
%else
and r2, 3
jz .end
.loop2:
mov r3d, [r1]
bswap r3d
mov [r0], r3d
add r1, 4
add r0, 4
dec r2
jnz .loop2
%endif
.end:
RET
%endmacro
INIT_XMM sse2
BSWAP32_BUF
INIT_XMM ssse3
BSWAP32_BUF
%macro op_avgh 3
movh %3, %2
pavgb %1, %3
movh %2, %1
%endmacro
%macro op_avg 2
pavgb %1, %2
mova %2, %1
%endmacro
%macro op_puth 2-3
movh %2, %1
%endmacro
%macro op_put 2
mova %2, %1
%endmacro
; void pixels4_l2_mmxext(uint8_t *dst, uint8_t *src1, uint8_t *src2, int dstStride, int src1Stride, int h)
%macro PIXELS4_L2 1
%define OP op_%1h
cglobal %1_pixels4_l2, 6,6
movsxdifnidn r3, r3d
movsxdifnidn r4, r4d
test r5d, 1
je .loop
movd m0, [r1]
movd m1, [r2]
add r1, r4
add r2, 4
pavgb m0, m1
OP m0, [r0], m3
add r0, r3
dec r5d
.loop:
mova m0, [r1]
mova m1, [r1+r4]
lea r1, [r1+2*r4]
pavgb m0, [r2]
pavgb m1, [r2+4]
OP m0, [r0], m3
OP m1, [r0+r3], m3
lea r0, [r0+2*r3]
mova m0, [r1]
mova m1, [r1+r4]
lea r1, [r1+2*r4]
pavgb m0, [r2+8]
pavgb m1, [r2+12]
OP m0, [r0], m3
OP m1, [r0+r3], m3
lea r0, [r0+2*r3]
add r2, 16
sub r5d, 4
jne .loop
REP_RET
%endmacro
INIT_MMX mmxext
PIXELS4_L2 put
PIXELS4_L2 avg
; void pixels8_l2_mmxext(uint8_t *dst, uint8_t *src1, uint8_t *src2, int dstStride, int src1Stride, int h)
%macro PIXELS8_L2 1
%define OP op_%1
cglobal %1_pixels8_l2, 6,6
movsxdifnidn r3, r3d
movsxdifnidn r4, r4d
test r5d, 1
je .loop
mova m0, [r1]
mova m1, [r2]
add r1, r4
add r2, 8
pavgb m0, m1
OP m0, [r0]
add r0, r3
dec r5d
.loop:
mova m0, [r1]
mova m1, [r1+r4]
lea r1, [r1+2*r4]
pavgb m0, [r2]
pavgb m1, [r2+8]
OP m0, [r0]
OP m1, [r0+r3]
lea r0, [r0+2*r3]
mova m0, [r1]
mova m1, [r1+r4]
lea r1, [r1+2*r4]
pavgb m0, [r2+16]
pavgb m1, [r2+24]
OP m0, [r0]
OP m1, [r0+r3]
lea r0, [r0+2*r3]
add r2, 32
sub r5d, 4
jne .loop
REP_RET
%endmacro
INIT_MMX mmxext
PIXELS8_L2 put
PIXELS8_L2 avg
; void pixels16_l2_mmxext(uint8_t *dst, uint8_t *src1, uint8_t *src2, int dstStride, int src1Stride, int h)
%macro PIXELS16_L2 1
%define OP op_%1
cglobal %1_pixels16_l2, 6,6
movsxdifnidn r3, r3d
movsxdifnidn r4, r4d
test r5d, 1
je .loop
mova m0, [r1]
mova m1, [r1+8]
pavgb m0, [r2]
pavgb m1, [r2+8]
add r1, r4
add r2, 16
OP m0, [r0]
OP m1, [r0+8]
add r0, r3
dec r5d
.loop:
mova m0, [r1]
mova m1, [r1+8]
add r1, r4
pavgb m0, [r2]
pavgb m1, [r2+8]
OP m0, [r0]
OP m1, [r0+8]
add r0, r3
mova m0, [r1]
mova m1, [r1+8]
add r1, r4
pavgb m0, [r2+16]
pavgb m1, [r2+24]
OP m0, [r0]
OP m1, [r0+8]
add r0, r3
add r2, 32
sub r5d, 2
jne .loop
REP_RET
%endmacro
INIT_MMX mmxext
PIXELS16_L2 put
PIXELS16_L2 avg
INIT_MMX mmxext
; void pixels(uint8_t *block, const uint8_t *pixels, int line_size, int h)
%macro PIXELS48 2
%if %2 == 4
%define OP movh
%else
%define OP mova
%endif
cglobal %1_pixels%2, 4,5
movsxdifnidn r2, r2d
lea r4, [r2*3]
.loop:
OP m0, [r1]
OP m1, [r1+r2]
OP m2, [r1+r2*2]
OP m3, [r1+r4]
lea r1, [r1+r2*4]
%ifidn %1, avg
pavgb m0, [r0]
pavgb m1, [r0+r2]
pavgb m2, [r0+r2*2]
pavgb m3, [r0+r4]
%endif
OP [r0], m0
OP [r0+r2], m1
OP [r0+r2*2], m2
OP [r0+r4], m3
sub r3d, 4
lea r0, [r0+r2*4]
jne .loop
RET
%endmacro
PIXELS48 put, 4
PIXELS48 avg, 4
PIXELS48 put, 8
PIXELS48 avg, 8
INIT_XMM sse2
; void put_pixels16_sse2(uint8_t *block, const uint8_t *pixels, int line_size, int h)
cglobal put_pixels16, 4,5,4
movsxdifnidn r2, r2d
lea r4, [r2*3]
.loop:
movu m0, [r1]
movu m1, [r1+r2]
movu m2, [r1+r2*2]
movu m3, [r1+r4]
lea r1, [r1+r2*4]
mova [r0], m0
mova [r0+r2], m1
mova [r0+r2*2], m2
mova [r0+r4], m3
sub r3d, 4
lea r0, [r0+r2*4]
jnz .loop
REP_RET
; void avg_pixels16_sse2(uint8_t *block, const uint8_t *pixels, int line_size, int h)
cglobal avg_pixels16, 4,5,4
movsxdifnidn r2, r2d
lea r4, [r2*3]
.loop:
movu m0, [r1]
movu m1, [r1+r2]
movu m2, [r1+r2*2]
movu m3, [r1+r4]
lea r1, [r1+r2*4]
pavgb m0, [r0]
pavgb m1, [r0+r2]
pavgb m2, [r0+r2*2]
pavgb m3, [r0+r4]
mova [r0], m0
mova [r0+r2], m1
mova [r0+r2*2], m2
mova [r0+r4], m3
sub r3d, 4
lea r0, [r0+r2*4]
jnz .loop
REP_RET