--- interpolate8x8_xmm.asm 2002/07/07 09:41:42 1.1 +++ interpolate8x8_xmm.asm 2004/08/29 10:02:38 1.9 @@ -1,60 +1,72 @@ -;/************************************************************************** +;/***************************************************************************** ; * -; * XVID MPEG-4 VIDEO CODEC -; * xmm 8x8 block-based halfpel interpolation +; * XVID MPEG-4 VIDEO CODEC +; * - mmx 8x8 block-based halfpel interpolation - ; * -; * This program is free software; you can redistribute it and/or modify -; * it under the terms of the GNU General Public License as published by -; * the Free Software Foundation; either version 2 of the License, or -; * (at your option) any later version. +; * Copyright(C) 2002 Michael Militzer +; * 2002 Pascal Massimino ; * -; * This program 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 General Public License for more details. +; * This program is free software ; you can redistribute it and/or modify +; * it under the terms of the GNU General Public License as published by +; * the Free Software Foundation ; either version 2 of the License, or +; * (at your option) any later version. ; * -; * You should have received a copy of the GNU General Public License -; * along with this program; if not, write to the Free Software -; * Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA. +; * This program 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 General Public License for more details. ; * -; *************************************************************************/ - -;/************************************************************************** -; * -; * History: +; * You should have received a copy of the GNU General Public License +; * along with this program ; if not, write to the Free Software +; * Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA 02111-1307 USA ; * -; * 04.06.2002 rewrote some funcs, mostly XMM. -Skal- -; * Heavily tuned for overlap and AGI-stalls avoidance -; * 04.02.2002 initial version (Isibaar) -; * -; *************************************************************************/ - +; ****************************************************************************/ -bits 32 +BITS 32 -%macro cglobal 1 +%macro cglobal 1 %ifdef PREFIX - global _%1 - %define %1 _%1 + %ifdef MARK_FUNCS + global _%1:function %1.endfunc-%1 + %define %1 _%1:function %1.endfunc-%1 + %else + global _%1 + %define %1 _%1 + %endif %else - global %1 + %ifdef MARK_FUNCS + global %1:function %1.endfunc-%1 + %else + global %1 + %endif %endif %endmacro -section .data +;============================================================================= +; Read only data +;============================================================================= + +%ifdef FORMAT_COFF +SECTION .rodata +%else +SECTION .rodata align=16 +%endif + +ALIGN 16 +mmx_one: + times 8 db 1 - -align 16 - -mmx_one -times 8 db 1 - -section .text +SECTION .text cglobal interpolate8x8_halfpel_h_xmm cglobal interpolate8x8_halfpel_v_xmm cglobal interpolate8x8_halfpel_hv_xmm +cglobal interpolate8x8_halfpel_add_xmm +cglobal interpolate8x8_halfpel_h_add_xmm +cglobal interpolate8x8_halfpel_v_add_xmm +cglobal interpolate8x8_halfpel_hv_add_xmm + ;=========================================================================== ; ; void interpolate8x8_halfpel_h_xmm(uint8_t * const dst, @@ -79,29 +91,29 @@ movq mm1, [eax+edx] movq mm4, mm0 movq mm5, mm1 - movq mm2, [eax+1] + movq mm2, [eax+1] movq mm3, [eax+edx+1] pavgb mm0, mm2 pxor mm2, mm4 pavgb mm1, mm3 - lea eax,[eax+2*edx] + lea eax, [eax+2*edx] pxor mm3, mm5 pand mm2, mm7 pand mm3, mm7 psubb mm0, mm2 movq [ecx], mm0 psubb mm1, mm3 - movq [ecx+edx], mm1 + movq [ecx+edx], mm1 %endmacro -align 16 +ALIGN 16 interpolate8x8_halfpel_h_xmm: - mov eax, [esp+16]; rounding - mov ecx, [esp+ 4] ; Dst + mov eax, [esp+16] ; rounding + mov ecx, [esp+ 4] ; Dst test eax,eax - mov eax, [esp+ 8] ; Src - mov edx, [esp+12] ; stride + mov eax, [esp+ 8] ; Src + mov edx, [esp+12] ; stride jnz near .rounding1 @@ -115,7 +127,7 @@ ret .rounding1 - ; we use: (i+j)/2 = ( i+j+1 )/2 - (i^j)&1 + ; we use: (i+j)/2 = ( i+j+1 )/2 - (i^j)&1 movq mm7, [mmx_one] COPY_H_SSE_RND1 lea ecx, [ecx+2*edx] @@ -125,23 +137,24 @@ lea ecx,[ecx+2*edx] COPY_H_SSE_RND1 ret +.endfunc ;=========================================================================== ; ; void interpolate8x8_halfpel_v_xmm(uint8_t * const dst, -; const uint8_t * const src, -; const uint32_t stride, -; const uint32_t rounding); +; const uint8_t * const src, +; const uint32_t stride, +; const uint32_t rounding); ; ;=========================================================================== %macro COPY_V_SSE_RND0 0 - movq mm0, [eax] - movq mm1, [eax+edx] + movq mm0, [eax] + movq mm1, [eax+edx] pavgb mm0, mm1 pavgb mm1, [eax+2*edx] - lea eax,[eax+2*edx] - movq [ecx],mm0 + lea eax, [eax+2*edx] + movq [ecx], mm0 movq [ecx+edx],mm1 %endmacro @@ -153,28 +166,27 @@ movq mm4, mm0 movq mm5, mm1 pavgb mm0, mm1 - pxor mm4, mm1 + pxor mm4, mm1 pavgb mm1, mm2 pxor mm5, mm2 - pand mm4, mm7 ; lsb's of (i^j)... - pand mm5, mm7 ; lsb's of (i^j)... - psubb mm0, mm4 ; ...are substracted from result of pavgb + pand mm4, mm7 ; lsb's of (i^j)... + pand mm5, mm7 ; lsb's of (i^j)... + psubb mm0, mm4 ; ...are substracted from result of pavgb movq [ecx], mm0 - psubb mm1, mm5 ; ...are substracted from result of pavgb + psubb mm1, mm5 ; ...are substracted from result of pavgb movq [ecx+edx], mm1 %endmacro -align 16 +ALIGN 16 interpolate8x8_halfpel_v_xmm: mov eax, [esp+16]; rounding - mov ecx, [esp+ 4] ; Dst + mov ecx, [esp+ 4] ; Dst test eax,eax - mov eax, [esp+ 8] ; Src - mov edx, [esp+12] ; stride - - ; we process 2 line at a time + mov eax, [esp+ 8] ; Src + mov edx, [esp+12] ; stride + ; we process 2 line at a time jnz near .rounding1 COPY_V_SSE_RND0 @@ -187,9 +199,9 @@ ret .rounding1 - ; we use: (i+j)/2 = ( i+j+1 )/2 - (i^j)&1 + ; we use: (i+j)/2 = ( i+j+1 )/2 - (i^j)&1 movq mm7, [mmx_one] - movq mm2, [eax] ; loop invariant + movq mm2, [eax] ; loop invariant add eax, edx COPY_V_SSE_RND1 @@ -200,13 +212,14 @@ lea ecx,[ecx+2*edx] COPY_V_SSE_RND1 ret +.endfunc ;=========================================================================== ; ; void interpolate8x8_halfpel_hv_xmm(uint8_t * const dst, -; const uint8_t * const src, -; const uint32_t stride, -; const uint32_t rounding); +; const uint8_t * const src, +; const uint32_t stride, +; const uint32_t rounding); ; ; ;=========================================================================== @@ -214,7 +227,7 @@ ; The trick is to correct the result of 'pavgb' with some combination of the ; lsb's of the 4 input values i,j,k,l, and their intermediate 'pavgb' (s and t). ; The boolean relations are: -; (i+j+k+l+3)/4 = (s+t+1)/2 - (ij&kl)&st +; (i+j+k+l+3)/4 = (s+t+1)/2 - (ij&kl)&st ; (i+j+k+l+2)/4 = (s+t+1)/2 - (ij|kl)&st ; (i+j+k+l+1)/4 = (s+t+1)/2 - (ij&kl)|st ; (i+j+k+l+0)/4 = (s+t+1)/2 - (ij|kl)|st @@ -223,91 +236,91 @@ ; Moreover, we process 2 lines at a times, for better overlapping (~15% faster). %macro COPY_HV_SSE_RND0 0 - lea eax,[eax+edx] + lea eax, [eax+edx] - movq mm0, [eax] - movq mm1, [eax+1] + movq mm0, [eax] + movq mm1, [eax+1] - movq mm6, mm0 - pavgb mm0, mm1 ; mm0=(j+k+1)/2. preserved for next step - lea eax,[eax+edx] - pxor mm1, mm6 ; mm1=(j^k). preserved for next step - - por mm3, mm1 ; ij |= jk - movq mm6, mm2 - pxor mm6, mm0 ; mm6 = s^t - pand mm3, mm6 ; (ij|jk) &= st - pavgb mm2, mm0 ; mm2 = (s+t+1)/2 - pand mm3, mm7 ; mask lsb - psubb mm2, mm3 ; apply. - - movq [ecx], mm2 - - movq mm2, [eax] - movq mm3, [eax+1] - movq mm6, mm2 - pavgb mm2, mm3 ; preserved for next iteration - lea ecx,[ecx+edx] - pxor mm3, mm6 ; preserved for next iteration - - por mm1, mm3 - movq mm6, mm0 - pxor mm6, mm2 - pand mm1, mm6 - pavgb mm0, mm2 + movq mm6, mm0 + pavgb mm0, mm1 ; mm0=(j+k+1)/2. preserved for next step + lea eax, [eax+edx] + pxor mm1, mm6 ; mm1=(j^k). preserved for next step + + por mm3, mm1 ; ij |= jk + movq mm6, mm2 + pxor mm6, mm0 ; mm6 = s^t + pand mm3, mm6 ; (ij|jk) &= st + pavgb mm2, mm0 ; mm2 = (s+t+1)/2 + pand mm3, mm7 ; mask lsb + psubb mm2, mm3 ; apply. - pand mm1, mm7 - psubb mm0, mm1 + movq [ecx], mm2 - movq [ecx], mm0 + movq mm2, [eax] + movq mm3, [eax+1] + movq mm6, mm2 + pavgb mm2, mm3 ; preserved for next iteration + lea ecx,[ecx+edx] + pxor mm3, mm6 ; preserved for next iteration + + por mm1, mm3 + movq mm6, mm0 + pxor mm6, mm2 + pand mm1, mm6 + pavgb mm0, mm2 + + pand mm1, mm7 + psubb mm0, mm1 + + movq [ecx], mm0 %endmacro %macro COPY_HV_SSE_RND1 0 - lea eax,[eax+edx] + lea eax, [eax+edx] - movq mm0, [eax] - movq mm1, [eax+1] + movq mm0, [eax] + movq mm1, [eax+1] - movq mm6, mm0 - pavgb mm0, mm1 ; mm0=(j+k+1)/2. preserved for next step - lea eax,[eax+edx] - pxor mm1, mm6 ; mm1=(j^k). preserved for next step - - pand mm3, mm1 - movq mm6, mm2 - pxor mm6, mm0 - por mm3, mm6 - pavgb mm2, mm0 - pand mm3, mm7 - psubb mm2, mm3 - - movq [ecx], mm2 - - movq mm2, [eax] - movq mm3, [eax+1] - movq mm6, mm2 - pavgb mm2, mm3 ; preserved for next iteration - lea ecx,[ecx+edx] - pxor mm3, mm6 ; preserved for next iteration - - pand mm1, mm3 - movq mm6, mm0 - pxor mm6, mm2 - por mm1, mm6 - pavgb mm0, mm2 - pand mm1, mm7 - psubb mm0, mm1 + movq mm6, mm0 + pavgb mm0, mm1 ; mm0=(j+k+1)/2. preserved for next step + lea eax, [eax+edx] + pxor mm1, mm6 ; mm1=(j^k). preserved for next step + + pand mm3, mm1 + movq mm6, mm2 + pxor mm6, mm0 + por mm3, mm6 + pavgb mm2, mm0 + pand mm3, mm7 + psubb mm2, mm3 + + movq [ecx], mm2 + + movq mm2, [eax] + movq mm3, [eax+1] + movq mm6, mm2 + pavgb mm2, mm3 ; preserved for next iteration + lea ecx,[ecx+edx] + pxor mm3, mm6 ; preserved for next iteration + + pand mm1, mm3 + movq mm6, mm0 + pxor mm6, mm2 + por mm1, mm6 + pavgb mm0, mm2 + pand mm1, mm7 + psubb mm0, mm1 - movq [ecx], mm0 + movq [ecx], mm0 %endmacro -align 16 +ALIGN 16 interpolate8x8_halfpel_hv_xmm: - mov eax, [esp+16] ; rounding - mov ecx, [esp+ 4] ; Dst - test eax,eax - mov eax, [esp+ 8] ; Src - mov edx, [esp+12] ; stride + mov eax, [esp+16] ; rounding + mov ecx, [esp+ 4] ; Dst + test eax, eax + mov eax, [esp+ 8] ; Src + mov edx, [esp+12] ; stride movq mm7, [mmx_one] @@ -316,7 +329,7 @@ movq mm3, [eax+1] movq mm6, mm2 pavgb mm2, mm3 - pxor mm3, mm6 ; mm2/mm3 ready + pxor mm3, mm6 ; mm2/mm3 ready jnz near .rounding1 @@ -337,4 +350,366 @@ COPY_HV_SSE_RND1 add ecx, edx COPY_HV_SSE_RND1 - ret + ret +.endfunc + +;=========================================================================== +; +; The next functions combine both source halfpel interpolation step and the +; averaging (with rouding) step to avoid wasting memory bandwidth computing +; intermediate halfpel images and then averaging them. +; +;=========================================================================== + +%macro PROLOG0 0 + mov ecx, [esp+ 4] ; Dst + mov eax, [esp+ 8] ; Src + mov edx, [esp+12] ; BpS +%endmacro +%macro PROLOG1 0 + PROLOG0 + test dword [esp+16], 1; Rounding? +%endmacro +%macro EPILOG 0 + ret +%endmacro + +;=========================================================================== +; +; void interpolate8x8_halfpel_add_xmm(uint8_t * const dst, +; const uint8_t * const src, +; const uint32_t stride, +; const uint32_t rounding); +; +; +;=========================================================================== + +%macro ADD_FF 2 + movq mm0, [eax+%1] + movq mm1, [eax+%2] +;;--- +;; movq mm2, mm0 +;; movq mm3, mm1 +;;--- + pavgb mm0, [ecx+%1] + pavgb mm1, [ecx+%2] +;;-- +;; por mm2, [ecx+%1] +;; por mm3, [ecx+%2] +;; pand mm2, [mmx_one] +;; pand mm3, [mmx_one] +;; psubsb mm0, mm2 +;; psubsb mm1, mm3 +;;-- + movq [ecx+%1], mm0 + movq [ecx+%2], mm1 +%endmacro + +ALIGN 16 +interpolate8x8_halfpel_add_xmm: ; 23c + PROLOG1 + ADD_FF 0, edx + lea eax,[eax+2*edx] + lea ecx,[ecx+2*edx] + ADD_FF 0, edx + lea eax,[eax+2*edx] + lea ecx,[ecx+2*edx] + ADD_FF 0, edx + lea eax,[eax+2*edx] + lea ecx,[ecx+2*edx] + ADD_FF 0, edx + EPILOG +.endfunc + +;=========================================================================== +; +; void interpolate8x8_halfpel_h_add_xmm(uint8_t * const dst, +; const uint8_t * const src, +; const uint32_t stride, +; const uint32_t rounding); +; +; +;=========================================================================== + + +%macro ADD_FH_RND0 2 + movq mm0, [eax+%1] + movq mm1, [eax+%2] + pavgb mm0, [eax+%1+1] + pavgb mm1, [eax+%2+1] + pavgb mm0, [ecx+%1] + pavgb mm1, [ecx+%2] + movq [ecx+%1],mm0 + movq [ecx+%2],mm1 +%endmacro + +%macro ADD_FH_RND1 2 + movq mm0, [eax+%1] + movq mm1, [eax+%2] + movq mm4, mm0 + movq mm5, mm1 + movq mm2, [eax+%1+1] + movq mm3, [eax+%2+1] + pavgb mm0, mm2 + ; lea ?? + pxor mm2, mm4 + pavgb mm1, mm3 + pxor mm3, mm5 + pand mm2, [mmx_one] + pand mm3, [mmx_one] + psubb mm0, mm2 + psubb mm1, mm3 + pavgb mm0, [ecx+%1] + pavgb mm1, [ecx+%2] + movq [ecx+%1],mm0 + movq [ecx+%2],mm1 +%endmacro + +ALIGN 16 +interpolate8x8_halfpel_h_add_xmm: ; 32c + PROLOG1 + jnz near .Loop1 + ADD_FH_RND0 0, edx + lea eax,[eax+2*edx] + lea ecx,[ecx+2*edx] + ADD_FH_RND0 0, edx + lea eax,[eax+2*edx] + lea ecx,[ecx+2*edx] + ADD_FH_RND0 0, edx + lea eax,[eax+2*edx] + lea ecx,[ecx+2*edx] + ADD_FH_RND0 0, edx + EPILOG + +.Loop1 + ; we use: (i+j)/2 = ( i+j+1 )/2 - (i^j)&1 + ; movq mm7, [mmx_one] + ADD_FH_RND1 0, edx + lea eax,[eax+2*edx] + lea ecx,[ecx+2*edx] + ADD_FH_RND1 0, edx + lea eax,[eax+2*edx] + lea ecx,[ecx+2*edx] + ADD_FH_RND1 0, edx + lea eax,[eax+2*edx] + lea ecx,[ecx+2*edx] + ADD_FH_RND1 0, edx + EPILOG +.endfunc + + +;=========================================================================== +; +; void interpolate8x8_halfpel_v_add_xmm(uint8_t * const dst, +; const uint8_t * const src, +; const uint32_t stride, +; const uint32_t rounding); +; +; +;=========================================================================== + +%macro ADD_8_HF_RND0 0 + movq mm0, [eax] + movq mm1, [eax+edx] + pavgb mm0, mm1 + pavgb mm1, [eax+2*edx] + lea eax,[eax+2*edx] + pavgb mm0, [ecx] + pavgb mm1, [ecx+edx] + movq [ecx],mm0 + movq [ecx+edx],mm1 +%endmacro + +%macro ADD_8_HF_RND1 0 + movq mm1, [eax+edx] + movq mm2, [eax+2*edx] + lea eax,[eax+2*edx] + movq mm4, mm0 + movq mm5, mm1 + pavgb mm0, mm1 + pxor mm4, mm1 + pavgb mm1, mm2 + pxor mm5, mm2 + pand mm4, mm7 ; lsb's of (i^j)... + pand mm5, mm7 ; lsb's of (i^j)... + psubb mm0, mm4 ; ...are substracted from result of pavgb + pavgb mm0, [ecx] + movq [ecx], mm0 + psubb mm1, mm5 ; ...are substracted from result of pavgb + pavgb mm1, [ecx+edx] + movq [ecx+edx], mm1 +%endmacro + +ALIGN 16 +interpolate8x8_halfpel_v_add_xmm: + PROLOG1 + + jnz near .Loop1 + pxor mm7, mm7 ; this is a NOP + + ADD_8_HF_RND0 + lea ecx,[ecx+2*edx] + ADD_8_HF_RND0 + lea ecx,[ecx+2*edx] + ADD_8_HF_RND0 + lea ecx,[ecx+2*edx] + ADD_8_HF_RND0 + EPILOG + +.Loop1 + movq mm0, [eax] ; loop invariant + movq mm7, [mmx_one] + + ADD_8_HF_RND1 + movq mm0, mm2 + lea ecx,[ecx+2*edx] + ADD_8_HF_RND1 + movq mm0, mm2 + lea ecx,[ecx+2*edx] + ADD_8_HF_RND1 + movq mm0, mm2 + lea ecx,[ecx+2*edx] + ADD_8_HF_RND1 + EPILOG +.endfunc + +; The trick is to correct the result of 'pavgb' with some combination of the +; lsb's of the 4 input values i,j,k,l, and their intermediate 'pavgb' (s and t). +; The boolean relations are: +; (i+j+k+l+3)/4 = (s+t+1)/2 - (ij&kl)&st +; (i+j+k+l+2)/4 = (s+t+1)/2 - (ij|kl)&st +; (i+j+k+l+1)/4 = (s+t+1)/2 - (ij&kl)|st +; (i+j+k+l+0)/4 = (s+t+1)/2 - (ij|kl)|st +; with s=(i+j+1)/2, t=(k+l+1)/2, ij = i^j, kl = k^l, st = s^t. + +; Moreover, we process 2 lines at a times, for better overlapping (~15% faster). + +;=========================================================================== +; +; void interpolate8x8_halfpel_hv_add_xmm(uint8_t * const dst, +; const uint8_t * const src, +; const uint32_t stride, +; const uint32_t rounding); +; +; +;=========================================================================== + +%macro ADD_HH_RND0 0 + lea eax,[eax+edx] + + movq mm0, [eax] + movq mm1, [eax+1] + + movq mm6, mm0 + pavgb mm0, mm1 ; mm0=(j+k+1)/2. preserved for next step + lea eax,[eax+edx] + pxor mm1, mm6 ; mm1=(j^k). preserved for next step + + por mm3, mm1 ; ij |= jk + movq mm6, mm2 + pxor mm6, mm0 ; mm6 = s^t + pand mm3, mm6 ; (ij|jk) &= st + pavgb mm2, mm0 ; mm2 = (s+t+1)/2 + pand mm3, mm7 ; mask lsb + psubb mm2, mm3 ; apply. + + pavgb mm2, [ecx] + movq [ecx], mm2 + + movq mm2, [eax] + movq mm3, [eax+1] + movq mm6, mm2 + pavgb mm2, mm3 ; preserved for next iteration + lea ecx,[ecx+edx] + pxor mm3, mm6 ; preserved for next iteration + + por mm1, mm3 + movq mm6, mm0 + pxor mm6, mm2 + pand mm1, mm6 + pavgb mm0, mm2 + + pand mm1, mm7 + psubb mm0, mm1 + + pavgb mm0, [ecx] + movq [ecx], mm0 +%endmacro + +%macro ADD_HH_RND1 0 + lea eax,[eax+edx] + + movq mm0, [eax] + movq mm1, [eax+1] + + movq mm6, mm0 + pavgb mm0, mm1 ; mm0=(j+k+1)/2. preserved for next step + lea eax,[eax+edx] + pxor mm1, mm6 ; mm1=(j^k). preserved for next step + + pand mm3, mm1 + movq mm6, mm2 + pxor mm6, mm0 + por mm3, mm6 + pavgb mm2, mm0 + pand mm3, mm7 + psubb mm2, mm3 + + pavgb mm2, [ecx] + movq [ecx], mm2 + + movq mm2, [eax] + movq mm3, [eax+1] + movq mm6, mm2 + pavgb mm2, mm3 ; preserved for next iteration + lea ecx,[ecx+edx] + pxor mm3, mm6 ; preserved for next iteration + + pand mm1, mm3 + movq mm6, mm0 + pxor mm6, mm2 + por mm1, mm6 + pavgb mm0, mm2 + pand mm1, mm7 + psubb mm0, mm1 + + pavgb mm0, [ecx] + movq [ecx], mm0 +%endmacro + +ALIGN 16 +interpolate8x8_halfpel_hv_add_xmm: + PROLOG1 + + movq mm7, [mmx_one] + + ; loop invariants: mm2=(i+j+1)/2 and mm3= i^j + movq mm2, [eax] + movq mm3, [eax+1] + movq mm6, mm2 + pavgb mm2, mm3 + pxor mm3, mm6 ; mm2/mm3 ready + + jnz near .Loop1 + + ADD_HH_RND0 + add ecx, edx + ADD_HH_RND0 + add ecx, edx + ADD_HH_RND0 + add ecx, edx + ADD_HH_RND0 + EPILOG + +.Loop1 + ADD_HH_RND1 + add ecx, edx + ADD_HH_RND1 + add ecx, edx + ADD_HH_RND1 + add ecx, edx + ADD_HH_RND1 + + EPILOG +.endfunc +