?? skl_mb_mmx.asm
字號:
movq mm4, [ecx+%1*32+ 0] ; Dst punpckhbw mm1, mm7 punpckhbw mm3, mm7 movq mm5, [ecx+%1*32+ 8] ; Dst2 psubw mm4, mm0 psubw mm5, mm1 movq [ecx+%1*32+ 0], mm4 movq [ecx+%1*32+ 8], mm5 movq mm0, [ecx+%1*32+16] ; Dst movq mm1, [ecx+%1*32+24] ; Dst2 psubw mm0, mm2 psubw mm1, mm3 movq [ecx+%1*32+16], mm0 movq [ecx+%1*32+24], mm1%endmacroalign 16Skl_Diff_16x8_8To16_MMX: ; 89c PROLOG0 pxor mm7, mm7 DIFF 0, 0 DIFF 4, 8 lea eax,[eax+2*edx] DIFF 1, 0 DIFF 5, 8 lea eax,[eax+2*edx] DIFF 2, 0 DIFF 6, 8 lea eax,[eax+2*edx] DIFF 3, 0 DIFF 7, 8 retalign 16Skl_Diff_8x8_8To16_MMX: ; 46c PROLOG0 pxor mm7, mm7 DIFF 0, 0 lea eax,[eax+2*edx] DIFF 1, 0 lea eax,[eax+2*edx] DIFF 2, 0 lea eax,[eax+2*edx] DIFF 3, 0 ret;//////////////////////////////////////////////////////////////////////%macro DIFF2 2 movq mm0, [eax+%2] ; Src1 movq mm1, mm0 movq mm2, [eax+%2+edx] movq mm3, mm2 punpcklbw mm0, mm7 punpcklbw mm2, mm7 movq mm4, [ebx+%2] ; Src2 punpckhbw mm1, mm7 punpckhbw mm3, mm7 movq mm6, [ebx+%2+edx] ; Src2 movq mm5, mm4 punpcklbw mm4, mm7 punpckhbw mm5, mm7 psubw mm0, mm4 psubw mm1, mm5 movq mm4, mm6 movq [ecx+%1*32+ 0], mm0 punpcklbw mm6, mm7 punpckhbw mm4, mm7 movq [ecx+%1*32+ 8], mm1 psubw mm2, mm6 psubw mm3, mm4 movq [ecx+%1*32+16], mm2 movq [ecx+%1*32+24], mm3%endmacroalign 16Skl_Diff_16x8_88To16_MMX: ; 101c;SKL_RDTSC_IN push ebx mov ecx, [esp+ 4 +4] ; Dst mov eax, [esp+ 8 +4] ; Src1 mov ebx, [esp+12 +4] ; Src2 mov edx, [esp+16 +4] ; BpS pxor mm7, mm7 DIFF2 0, 0 DIFF2 4, 8 lea eax,[eax+2*edx] lea ebx,[ebx+2*edx] DIFF2 1, 0 DIFF2 5, 8 lea eax,[eax+2*edx] lea ebx,[ebx+2*edx] DIFF2 2, 0 DIFF2 6, 8 lea eax,[eax+2*edx] lea ebx,[ebx+2*edx] DIFF2 3, 0 DIFF2 7, 8 pop ebx;SKL_RDTSC_OUT retalign 16Skl_Diff_8x8_88To16_MMX: ;56 push ebx mov ecx, [esp+ 4 +4] ; Dst mov eax, [esp+ 8 +4] ; Src1 mov ebx, [esp+12 +4] ; Src2 mov edx, [esp+16 +4] ; BpS pxor mm7, mm7 DIFF2 0, 0 lea eax,[eax+2*edx] lea ebx,[ebx+2*edx] DIFF2 1, 0 lea eax,[eax+2*edx] lea ebx,[ebx+2*edx] DIFF2 2, 0 lea eax,[eax+2*edx] lea ebx,[ebx+2*edx] DIFF2 3, 0 pop ebx ret;//////////////////////////////////////////////////////////////////////;//;// 8x8 -> 16x16 upsampling (16b);//;//////////////////////////////////////////////////////////////////////%macro MUL_PACK 4 ; %1/%2: regs %3/%4: Up13/Up31 pmullw %1, %3 ; [Up13] pmullw mm4, %4 ; [Up31] pmullw %2, %3 ; [Up13] pmullw mm5, %4 ; [Up31] paddsw %1, [Cst2] paddsw %2, [Cst2] paddsw %1, mm4 paddsw %2, mm5%endmacro%macro COL03 3 ;%1/%2: regs, %3: row -output: mm4/mm5 movq %1, [edx+%3*16+0*2] ; %1 = 0|1|2|3 movq %2,[edx+%3*16+1*2] ; %2 = 1|2|3|4 movq mm5, %1 ; mm5 = 0|1|2|3 movq mm4, %1 ; mm4 = 0|1|2|3 punpckhwd mm5,%2 ; mm5 = 2|3|3|4 punpcklwd mm4,%2 ; mm4 = 0|1|1|2 punpcklwd %1,%1 ; %1 = 0|0|1|1 punpcklwd %2, mm5 ; %2 = 1|2|2|3 punpcklwd %1, mm4 ; %1 = 0|0|0|1%endmacro%macro COL47 3 ;%1-%2: regs, %3: row -output: mm4/mm5 movq mm5, [edx+%3*16+4*2] ; mm5 = 4|5|6|7 movq %1, [edx+%3*16+3*2] ; %1 = 3|4|5|6 movq %2, mm5 ; %2 = 4|5|6|7 movq mm4, mm5 ; mm4 = 4|5|6|7 punpckhwd %2, %2 ; %2 = 6|6|7|7 punpckhwd mm5, %2 ; mm5 = 6|7|7|7 movq %2, %1 ; %2 = 3|4|5|6 punpcklwd %1, mm4 ; %1 = 3|4|4|5 punpckhwd %2, mm4 ; %2 = 5|6|6|7 punpcklwd mm4, %2 ; mm4 = 4|5|5|6%endmacro%macro MIX_ROWS 4 ; %1/%2:prev %3/4:cur (preserved) mm4/mm5: output ; we need to perform: (%1,%3) -> (%1 = 3*%1+%3, mm4 = 3*%3+%1), %3 preserved. movq mm4, [Cst3] movq mm5, [Cst3] pmullw mm4, %3 pmullw mm5, %4 paddsw mm4, %1 paddsw mm5, %2 pmullw %1, [Cst3] pmullw %2, [Cst3] paddsw %1, %3 paddsw %2, %4%endmacro;////////////////////////////////////////////////////////////////////// ; Note: we can use ">>2" instead of "/4" here, since we ; are (supposed to be) averaging positive values%macro STORE_1 2 psraw %1, 2 psraw %2, 2 packuswb %1,%2 movq [ecx], %1%endmacro%macro STORE_2 2 ; pack and store (%1,%2) + (mm4,mm5) psraw %1, 4 psraw %2, 4 psraw mm4, 4 psraw mm5, 4 packuswb %1,%2 packuswb mm4, mm5 movq [ecx], %1 movq [ecx+eax], mm4 lea ecx, [ecx+2*eax]%endmacroalign 16Skl_Copy_Upsampled_8x8_16To8_MMX: ; 344c mov ecx, [esp+4] ; Dst mov edx, [esp+8] ; Src mov eax, [esp+12] ; BpS movq mm6, [Up13] movq mm7, [Up31] COL03 mm0, mm1, 0 MUL_PACK mm0,mm1, mm6, mm7 movq mm4, mm0 movq mm5, mm1 STORE_1 mm4, mm5 add ecx, eax COL03 mm2, mm3, 1 MUL_PACK mm2,mm3, mm6, mm7 MIX_ROWS mm0, mm1, mm2, mm3 STORE_2 mm0, mm1 COL03 mm0, mm1, 2 MUL_PACK mm0,mm1, mm6, mm7 MIX_ROWS mm2, mm3, mm0, mm1 STORE_2 mm2, mm3 COL03 mm2, mm3, 3 MUL_PACK mm2,mm3, mm6, mm7 MIX_ROWS mm0, mm1, mm2, mm3 STORE_2 mm0, mm1 COL03 mm0, mm1, 4 MUL_PACK mm0,mm1, mm6, mm7 MIX_ROWS mm2, mm3, mm0, mm1 STORE_2 mm2, mm3 COL03 mm2, mm3, 5 MUL_PACK mm2,mm3, mm6, mm7 MIX_ROWS mm0, mm1, mm2, mm3 STORE_2 mm0, mm1 COL03 mm0, mm1, 6 MUL_PACK mm0,mm1, mm6, mm7 MIX_ROWS mm2, mm3, mm0, mm1 STORE_2 mm2, mm3 COL03 mm2, mm3, 7 MUL_PACK mm2,mm3, mm6, mm7 MIX_ROWS mm0, mm1, mm2, mm3 STORE_2 mm0, mm1 STORE_1 mm2, mm3 mov ecx, [esp+4] add ecx, 8 COL47 mm0, mm1, 0 MUL_PACK mm0,mm1, mm6, mm7 movq mm4, mm0 movq mm5, mm1 STORE_1 mm4, mm5 add ecx, eax COL47 mm2, mm3, 1 MUL_PACK mm2,mm3, mm6, mm7 MIX_ROWS mm0, mm1, mm2, mm3 STORE_2 mm0, mm1 COL47 mm0, mm1, 2 MUL_PACK mm0,mm1, mm6, mm7 MIX_ROWS mm2, mm3, mm0, mm1 STORE_2 mm2, mm3 COL47 mm2, mm3, 3 MUL_PACK mm2,mm3, mm6, mm7 MIX_ROWS mm0, mm1, mm2, mm3 STORE_2 mm0, mm1 COL47 mm0, mm1, 4 MUL_PACK mm0,mm1, mm6, mm7 MIX_ROWS mm2, mm3, mm0, mm1 STORE_2 mm2, mm3 COL47 mm2, mm3, 5 MUL_PACK mm2,mm3, mm6, mm7 MIX_ROWS mm0, mm1, mm2, mm3 STORE_2 mm0, mm1 COL47 mm0, mm1, 6 MUL_PACK mm0,mm1, mm6, mm7 MIX_ROWS mm2, mm3, mm0, mm1 STORE_2 mm2, mm3 COL47 mm2, mm3, 7 MUL_PACK mm2,mm3, mm6, mm7 MIX_ROWS mm0, mm1, mm2, mm3 STORE_2 mm0, mm1 STORE_1 mm2, mm3 ret;////////////////////////////////////////////////////////////////////// ; Note: grrr... the 'pcmpgtw' stuff are the "/4" and "/16" operators ; implemented with ">>2" and ">>4" using: ; x/4 = ( (x-(x<0))>>2 ) + (x<0) ; x/16 = ( (x-(x<0))>>4 ) + (x<0)%macro STORE_ADD_1 2 ; We substract the rounder '2' for corner pixels, ; since when 'x' is negative, (x*4 + 2)/4 is *not* ; equal to 'x'. In fact, the correct relation is: ; (x*4 + 2)/4 = x - (x<0) ; So, better revert to (x*4)/4 = x. psubsw %1, [Cst2000] psubsw %2, [Cst0002] pxor mm6, mm6 pxor mm7, mm7 pcmpgtw mm6, %1 pcmpgtw mm7, %2 paddsw %1, mm6 paddsw %2, mm7 psraw %1, 2 psraw %2, 2 psubsw %1, mm6 psubsw %2, mm7 ; mix with destination [ecx] movq mm6, [ecx] movq mm7, [ecx] punpcklbw mm6, [Cst0] punpckhbw mm7, [Cst0] paddsw %1, mm6 paddsw %2, mm7 packuswb %1,%2 movq [ecx], %1%endmacro%macro STORE_ADD_2 2 pxor mm6, mm6 pxor mm7, mm7 pcmpgtw mm6, %1 pcmpgtw mm7, %2 paddsw %1, mm6 paddsw %2, mm7 psraw %1, 4 psraw %2, 4 psubsw %1, mm6 psubsw %2, mm7 pxor mm6, mm6 pxor mm7, mm7 pcmpgtw mm6, mm4 pcmpgtw mm7, mm5 paddsw mm4, mm6 paddsw mm5, mm7 psraw mm4, 4 psraw mm5, 4 psubsw mm4, mm6 psubsw mm5, mm7 ; mix with destination movq mm6, [ecx] movq mm7, [ecx] punpcklbw mm6, [Cst0] punpckhbw mm7, [Cst0] paddsw %1, mm6 paddsw %2, mm7 movq mm6, [ecx+eax] movq mm7, [ecx+eax] punpcklbw mm6, [Cst0] punpckhbw mm7, [Cst0] paddsw mm4, mm6 paddsw mm5, mm7 packuswb %1,%2 packuswb mm4, mm5 movq [ecx], %1 movq [ecx+eax], mm4 lea ecx, [ecx+2*eax]%endmacroalign 16Skl_Add_Upsampled_8x8_16To8_MMX: ; 579c mov ecx, [esp+4] ; Dst mov edx, [esp+8] ; Src mov eax, [esp+12] ; BpS COL03 mm0, mm1, 0 MUL_PACK mm0,mm1, [Up13], [Up31] movq mm4, mm0 movq mm5, mm1 STORE_ADD_1 mm4, mm5 add ecx, eax COL03 mm2, mm3, 1 MUL_PACK mm2,mm3, [Up13], [Up31] MIX_ROWS mm0, mm1, mm2, mm3 STORE_ADD_2 mm0, mm1 COL03 mm0, mm1, 2 MUL_PACK mm0,mm1, [Up13], [Up31] MIX_ROWS mm2, mm3, mm0, mm1 STORE_ADD_2 mm2, mm3 COL03 mm2, mm3, 3 MUL_PACK mm2,mm3, [Up13], [Up31] MIX_ROWS mm0, mm1, mm2, mm3 STORE_ADD_2 mm0, mm1 COL03 mm0, mm1, 4 MUL_PACK mm0,mm1, [Up13], [Up31] MIX_ROWS mm2, mm3, mm0, mm1 STORE_ADD_2 mm2, mm3 COL03 mm2, mm3, 5 MUL_PACK mm2,mm3, [Up13], [Up31] MIX_ROWS mm0, mm1, mm2, mm3 STORE_ADD_2 mm0, mm1 COL03 mm0, mm1, 6 MUL_PACK mm0,mm1, [Up13], [Up31] MIX_ROWS mm2, mm3, mm0, mm1 STORE_ADD_2 mm2, mm3 COL03 mm2, mm3, 7 MUL_PACK mm2,mm3, [Up13], [Up31] MIX_ROWS mm0, mm1, mm2, mm3 STORE_ADD_2 mm0, mm1 STORE_ADD_1 mm2, mm3 mov ecx, [esp+4] add ecx, 8 COL47 mm0, mm1, 0 MUL_PACK mm0,mm1, [Up13], [Up31] movq mm4, mm0 movq mm5, mm1 STORE_ADD_1 mm4, mm5 add ecx, eax COL47 mm2, mm3, 1 MUL_PACK mm2,mm3, [Up13], [Up31] MIX_ROWS mm0, mm1, mm2, mm3 STORE_ADD_2 mm0, mm1 COL47 mm0, mm1, 2 MUL_PACK mm0,mm1, [Up13], [Up31] MIX_ROWS mm2, mm3, mm0, mm1 STORE_ADD_2 mm2, mm3 COL47 mm2, mm3, 3 MUL_PACK mm2,mm3, [Up13], [Up31] MIX_ROWS mm0, mm1, mm2, mm3 STORE_ADD_2 mm0, mm1 COL47 mm0, mm1, 4 MUL_PACK mm0,mm1, [Up13], [Up31] MIX_ROWS mm2, mm3, mm0, mm1 STORE_ADD_2 mm2, mm3 COL47 mm2, mm3, 5 MUL_PACK mm2,mm3, [Up13], [Up31] MIX_ROWS mm0, mm1, mm2, mm3 STORE_ADD_2 mm0, mm1 COL47 mm0, mm1, 6 MUL_PACK mm0,mm1, [Up13], [Up31] MIX_ROWS mm2, mm3, mm0, mm1 STORE_ADD_2 mm2, mm3 COL47 mm2, mm3, 7 MUL_PACK mm2,mm3, [Up13], [Up31] MIX_ROWS mm0, mm1, mm2, mm3 STORE_ADD_2 mm0, mm1 STORE_ADD_1 mm2, mm3 ret;//////////////////////////////////////////////////////////////////////;//;// horizontal/vertical filtering: [x,y] -> [ (3x+y+2)>>2, (x+3y+2)>>2];// We use the trick: tmp = (x+y+2) -> [x = (tmp+2x)>>2, y = (tmp+2y)>>2];//;//////////////////////////////////////////////////////////////////////align 16Skl_HFilter_31_MMX: push esi push edi mov esi, [esp+4 +8] ; Src1 mov edi, [esp+8 +8] ; Src2 mov eax, [esp+12 +8] ; Nb_Blks lea eax,[eax*2] movq mm5, [Cst2] pxor mm7, mm7 lea esi, [esi+eax*4] lea edi, [edi+eax*4] neg eax.Loop: ;12c movd mm0, [esi+eax*4] movd mm1, [edi+eax*4] movq mm2, mm5 punpcklbw mm0, mm7 punpcklbw mm1, mm7 paddsw mm2, mm0 paddsw mm0, mm0 paddsw mm2, mm1 paddsw mm1, mm1 paddsw mm0, mm2 paddsw mm1, mm2 psraw mm0, 2 psraw mm1, 2 packuswb mm0, mm7 packuswb mm1, mm7 movd [esi+eax*4], mm0 movd [edi+eax*4], mm1 add eax,1 jl .Loop pop edi pop esi ret ; MMX is of no use here. Better use plain ASM. Moreover, ; this is for the fun of ASM coding, coz' every modern compiler can ; end up with a code that looks very much like this one...align 16Skl_VFilter_31_x86: push esi push edi push ebx push ebp mov esi, [esp+4 +16] ; Src1 mov edi, [esp+8 +16] ; Src2 mov ebp, [esp+12 +16] ; BpS mov eax, [esp+16 +16] ; Nb_Blks lea eax,[eax*8].Loop: ;7c movzx ecx, byte [esi] movzx edx, byte [edi] lea ebx, [ecx+edx+2] lea ecx,[ebx+2*ecx] lea edx,[ebx+2*edx] shr ecx,2 shr edx,2 mov [esi], cl mov [edi], dl lea esi, [esi+ebp] lea edi, [edi+ebp] dec eax jg .Loop pop ebp pop ebx pop edi pop esi ret ; this one's just a little faster than gcc's code. Very little.align 16Skl_HFilter_31_x86: push esi push edi push ebx mov esi, [esp+4 +12] ; Src1 mov edi, [esp+8 +12] ; Src2 mov eax, [esp+12 +12] ; Nb_Blks lea eax,[eax*8] lea esi, [esi+eax] lea edi, [esi+eax] neg eax.Loop: ; 6c movzx ecx, byte [esi+eax] movzx edx, byte [edi+eax] lea ebx, [ecx+edx+2] lea ecx,[ebx+2*ecx] lea edx,[ebx+2*edx] shr ecx,2 shr edx,2 mov [esi+eax], cl mov [edi+eax], dl inc eax jl .Loop pop ebx pop edi pop esi ret;//////////////////////////////////////////////////////////////////////;//;// 16b downsampling 16x16 -> 8x8;//;//////////////////////////////////////////////////////////////////////%macro HFILTER_1331 2 ;%1:src %2:dst reg. -trashes mm0/mm1/mm2 movq mm2, [Mask_ff] movq %2, [%1-1] ;-10123456 movq mm0, [%1] ; 01234567 movq mm1, [%1+1] ; 12345678 pand %2, mm2 ;-1|1|3|5 pand mm0, mm2 ; 0|2|4|6 pand mm1, mm2 ; 1|3|5|7 pand mm2, [%1+2] ; 2|4|6|8 paddusw mm0, mm1 paddusw %2, mm2 pmullw mm0, mm7 paddusw %2, mm0%endmacro%macro VFILTER_1331 4 ; %1-4: regs %1-%2: trashed paddsw %1, [Cst32] paddsw %2, %3 pmullw %2, mm7 paddsw %1, %4 paddsw %1, %2 psraw %1, 6%endmacro;//////////////////////////////////////////////////////////////////////%macro COPY_TWO_LINES_1331 1 ; %1: dst HFILTER_1331 edx , mm5 HFILTER_1331 edx+eax, mm6 lea edx, [edx+2*eax] VFILTER_1331 mm3,mm4,mm5, mm6 movq [%1], mm3 HFILTER_1331 edx , mm3 HFILTER_1331 edx+eax, mm4 lea edx, [edx+2*eax] VFILTER_1331 mm5,mm6,mm3,mm4 movq [%1+16], mm5%endmacroalign 16Skl_Filter_18x18_To_8x8_MMX: ; 283c (~4.4c per output pixel) mov ecx, [esp+4] ; Dst mov edx, [esp+8] ; Src mov eax, [esp+12] ; BpS movq mm7, [Cst3] sub edx, eax ; mm3/mm4/mm5/mm6 is used as a 4-samples delay line. ; process columns 0-3 HFILTER_1331 edx , mm3 ; pre-load mm3/mm4 HFILTER_1331 edx+eax, mm4 lea edx, [edx+2*eax] COPY_TWO_LINES_1331 ecx + 0*16 COPY_TWO_LINES_1331 ecx + 2*16 COPY_TWO_LINES_1331 ecx + 4*16 COPY_TWO_LINES_1331 ecx + 6*16 ; process columns 4-7 mov edx, [esp+8] sub edx, eax add edx, 8 HFILTER_1331 edx , mm3 ; pre-load mm3/mm4 HFILTER_1331 edx+eax, mm4 lea edx, [edx+2*eax] COPY_TWO_LINES_1331 ecx + 0*16 +8 COPY_TWO_LINES_1331 ecx + 2*16 +8 COPY_TWO_LINES_1331 ecx + 4*16 +8 COPY_TWO_LINES_1331 ecx + 6*16 +8 ret;//////////////////////////////////////////////////////////////////////%macro DIFF_TWO_LINES_1331 1 ; %1: dst HFILTER_1331 edx , mm5 HFILTER_1331 edx+eax, mm6 lea edx, [edx+2*eax] movq mm2, [%1] VFILTER_1331 mm3,mm4,mm5, mm6 psubsw mm2, mm3 movq [%1], mm2 HFILTER_1331 edx , mm3 HFILTER_1331 edx+eax, mm4 lea edx, [edx+2*eax] movq mm2, [%1+16] VFILTER_1331 mm5,mm6,mm3,mm4 psubsw mm2, mm5 movq [%1+16], mm2%endmacroalign 16Skl_Filter_Diff_18x18_To_8x8_MMX: ; 302c;SKL_RDTSC_IN mov ecx, [esp+4] ; Dst mov edx, [esp+8] ; Src mov eax, [esp+12] ; BpS movq mm7, [Cst3] sub edx, eax ; mm3/mm4/mm5/mm6 is used as a 4-samples delay line. ; process columns 0-3 HFILTER_1331 edx , mm3 ; pre-load mm3/mm4 HFILTER_1331 edx+eax, mm4 lea edx, [edx+2*eax] DIFF_TWO_LINES_1331 ecx + 0*16 DIFF_TWO_LINES_1331 ecx + 2*16 DIFF_TWO_LINES_1331 ecx + 4*16 DIFF_TWO_LINES_1331 ecx + 6*16 ; process columns 4-7 mov edx, [esp+8] sub edx, eax add edx, 8 HFILTER_1331 edx , mm3 ; pre-load mm3/mm4 HFILTER_1331 edx+eax, mm4 lea edx, [edx+2*eax] DIFF_TWO_LINES_1331 ecx + 0*16 +8 DIFF_TWO_LINES_1331 ecx + 2*16 +8 DIFF_TWO_LINES_1331 ecx + 4*16 +8 DIFF_TWO_LINES_1331 ecx + 6*16 +8;SKL_RDTSC_OUT ret;//////////////////////////////////////////////////////////////////////?? 快捷鍵說明
復制代碼
Ctrl + C
搜索代碼
Ctrl + F
全屏模式
F11
切換主題
Ctrl + Shift + D
顯示快捷鍵
?
增大字號
Ctrl + =
減小字號
Ctrl + -