--- mem_transfer_mmx.asm 2002/03/08 02:45:01 1.1 +++ mem_transfer_mmx.asm 2002/09/21 03:07:56 1.7 @@ -32,8 +32,10 @@ ; * ; * History: ; * +; * 04.06.2002 speed enhancement (unroll+overlap). -Skal- +; * + added transfer_8to16sub2_mmx/xmm ; * 07.01.2002 merge functions from compensate_mmx; rename functions -; * 07.11.2001 initial version; (c)2001 peter ross +; * 07.11.2001 initial version; (c)2001 peter ross ; * ; *************************************************************************/ @@ -52,6 +54,13 @@ section .text +cglobal transfer_8to16copy_mmx +cglobal transfer_16to8copy_mmx +cglobal transfer_8to16sub_mmx +cglobal transfer_8to16sub2_mmx +cglobal transfer_8to16sub2_xmm +cglobal transfer_16to8add_mmx +cglobal transfer8x8_copy_mmx ;=========================================================================== ; @@ -61,41 +70,35 @@ ; ;=========================================================================== -align 16 -cglobal transfer_8to16copy_mmx -transfer_8to16copy_mmx - - push esi - push edi - - mov edi, [esp + 8 + 4] ; dst - mov esi, [esp + 8 + 8] ; src - mov ecx, [esp + 8 + 12] ; stride - - pxor mm7, mm7 ; mm7 = zero - - mov eax, 8 - -.loop - movq mm0, [esi] - movq mm1, mm0 - punpcklbw mm0, mm7 ; mm01 = unpack([src]) - punpckhbw mm1, mm7 - - movq [edi], mm0 ; [dst] = mm01 - movq [edi + 8], mm1 - - add edi, 16 - add esi, ecx - dec eax - jnz .loop - - pop edi - pop esi - - ret +%macro COPY_8_TO_16 1 + movq mm0, [eax] + movq mm1, [eax+edx] + movq mm2, mm0 + movq mm3, mm1 + punpcklbw mm0, mm7 + movq [ecx+%1*32], mm0 + punpcklbw mm1, mm7 + movq [ecx+%1*32+16], mm1 + punpckhbw mm2, mm7 + punpckhbw mm3, mm7 + lea eax,[eax+2*edx] + movq [ecx+%1*32+8], mm2 + movq [ecx+%1*32+24], mm3 +%endmacro +align 16 +transfer_8to16copy_mmx: + mov ecx, [esp+ 4] ; Dst + mov eax, [esp+ 8] ; Src + mov edx, [esp+12] ; Stride + pxor mm7,mm7 + + COPY_8_TO_16 0 + COPY_8_TO_16 1 + COPY_8_TO_16 2 + COPY_8_TO_16 3 + ret ;=========================================================================== ; @@ -105,35 +108,32 @@ ; ;=========================================================================== -align 16 -cglobal transfer_16to8copy_mmx -transfer_16to8copy_mmx - - push esi - push edi - - mov edi, [esp + 8 + 4] ; dst - mov esi, [esp + 8 + 8] ; src - mov ecx, [esp + 8 + 12] ; stride - - mov eax, 8 - -.loop - movq mm0, [esi] - packuswb mm0, [esi + 8] ; mm0 = pack([src]) - - movq [edi], mm0 ; [dst] = mm0 - - add esi, 16 - add edi, ecx - dec eax - jnz .loop - - pop edi - pop esi +%macro COPY_16_TO_8 1 + movq mm0, [eax+%1*32] + movq mm1, [eax+%1*32+8] + packuswb mm0, mm1 + movq [ecx], mm0 + movq mm2, [eax+%1*32+16] + movq mm3, [eax+%1*32+24] + packuswb mm2, mm3 + movq [ecx+edx], mm2 +%endmacro - ret +align 16 +transfer_16to8copy_mmx: + mov ecx, [esp+ 4] ; Dst + mov eax, [esp+ 8] ; Src + mov edx, [esp+12] ; Stride + + COPY_16_TO_8 0 + lea ecx,[ecx+2*edx] + COPY_16_TO_8 1 + lea ecx,[ecx+2*edx] + COPY_16_TO_8 2 + lea ecx,[ecx+2*edx] + COPY_16_TO_8 3 + ret ;=========================================================================== ; @@ -151,213 +151,217 @@ ; * 02.12.2001 loop unrolled, code runs 10% faster now (Isibaar) ; * 30.11.2001 16 pixels are processed per iteration (Isibaar) ; * 30.11.2001 .text missing -; * 06.11.2001 inital version; (c)2001 peter ross +; * 06.11.2001 inital version; (c)2001 peter ross ; * ; *************************************************************************/ -align 16 -cglobal transfer_8to16sub_mmx -transfer_8to16sub_mmx - push esi - push edi - push ebx - - mov edi, [esp + 12 + 4] ; dct [out] - mov edx, [esp + 12 + 8] ; cur [in/out] - mov esi, [esp + 12 + 12] ; ref [in] - mov ecx, [esp + 12 + 16] ; stride [in] - - mov eax, edx ; cur -> eax - mov ebx, esi ; ref -> ebx - add eax, ecx ; cur + stride - add ebx, ecx ; ref + stride - - shl ecx, 1 - - pxor mm7, mm7 ; mm7 = zero - - movq mm0, [edx] ; mm01 = [cur] - movq mm1, mm0 - - punpcklbw mm0, mm7 - punpckhbw mm1, mm7 - - movq mm4, [eax] - movq mm5, mm4 - - punpcklbw mm4, mm7 - punpckhbw mm5, mm7 - - movq mm2, [esi] ; mm23 = [ref] - movq mm3, mm2 - - movq mm6, [ebx] - - movq [edx], mm2 ; [cur] = [ref] - movq [eax], mm6 - - punpcklbw mm2, mm7 - punpckhbw mm3, mm7 - - psubsw mm0, mm2 ; mm01 -= mm23 - - movq mm2, mm6 - - punpcklbw mm2, mm7 - punpckhbw mm6, mm7 - - psubsw mm1, mm3 - - psubsw mm4, mm2 - psubsw mm5, mm6 - - movq [edi], mm0 ; dct[] = mm01 - movq [edi + 8], mm1 - movq [edi + 16], mm4 - movq [edi + 24], mm5 - - add edx, ecx - add esi, ecx - add eax, ecx - add ebx, ecx - - movq mm0, [edx] ; mm01 = [cur] - movq mm1, mm0 - - punpcklbw mm0, mm7 - punpckhbw mm1, mm7 - - movq mm4, [eax] - movq mm5, mm4 - - punpcklbw mm4, mm7 - punpckhbw mm5, mm7 - - movq mm2, [esi] ; mm23 = [ref] - movq mm3, mm2 - - movq mm6, [ebx] - - movq [edx], mm2 ; [cur] = [ref] - movq [eax], mm6 - - punpcklbw mm2, mm7 - punpckhbw mm3, mm7 - - psubsw mm0, mm2 ; mm01 -= mm23 - - movq mm2, mm6 - - punpcklbw mm2, mm7 - punpckhbw mm6, mm7 - - psubsw mm1, mm3 - - psubsw mm4, mm2 - psubsw mm5, mm6 - - movq [edi + 32], mm0 ; dct[] = mm01 - movq [edi + 40], mm1 - movq [edi + 48], mm4 - movq [edi + 56], mm5 - - add edx, ecx - add esi, ecx - add eax, ecx - add ebx, ecx - - movq mm0, [edx] ; mm01 = [cur] - movq mm1, mm0 - - punpcklbw mm0, mm7 - punpckhbw mm1, mm7 - - movq mm4, [eax] - movq mm5, mm4 - - punpcklbw mm4, mm7 - punpckhbw mm5, mm7 - - movq mm2, [esi] ; mm23 = [ref] - movq mm3, mm2 - - movq mm6, [ebx] - - movq [edx], mm2 ; [cur] = [ref] - movq [eax], mm6 - - punpcklbw mm2, mm7 - punpckhbw mm3, mm7 - - psubsw mm0, mm2 ; mm01 -= mm23 - - movq mm2, mm6 - - punpcklbw mm2, mm7 - punpckhbw mm6, mm7 - - psubsw mm1, mm3 - - psubsw mm4, mm2 - psubsw mm5, mm6 - - movq [edi + 64], mm0 ; dct[] = mm01 - movq [edi + 72], mm1 - movq [edi + 80], mm4 - movq [edi + 88], mm5 - - add edx, ecx - add esi, ecx - add eax, ecx - add ebx, ecx - - movq mm0, [edx] ; mm01 = [cur] - movq mm1, mm0 - - punpcklbw mm0, mm7 - punpckhbw mm1, mm7 - - movq mm4, [eax] - movq mm5, mm4 - - punpcklbw mm4, mm7 - punpckhbw mm5, mm7 - - movq mm2, [esi] ; mm23 = [ref] - movq mm3, mm2 - - movq mm6, [ebx] - - movq [edx], mm2 ; [cur] = [ref] - movq [eax], mm6 - - punpcklbw mm2, mm7 - punpckhbw mm3, mm7 - - psubsw mm0, mm2 ; mm01 -= mm23 - - movq mm2, mm6 +%macro COPY_8_TO_16_SUB 1 + movq mm0, [eax] ; cur + movq mm2, [eax+edx] + movq mm1, mm0 + movq mm3, mm2 + + punpcklbw mm0, mm7 + punpcklbw mm2, mm7 + movq mm4, [ebx] ; ref + punpckhbw mm1, mm7 + punpckhbw mm3, mm7 + movq mm5, [ebx+edx] ; ref + + movq mm6, mm4 + movq [eax], mm4 + movq [eax+edx], mm5 + punpcklbw mm4, mm7 + punpckhbw mm6, mm7 + psubsw mm0, mm4 + psubsw mm1, mm6 + movq mm6, mm5 + punpcklbw mm5, mm7 + punpckhbw mm6, mm7 + psubsw mm2, mm5 + lea eax,[eax+2*edx] + psubsw mm3, mm6 + lea ebx,[ebx+2*edx] + + movq [ecx+%1*32+ 0], mm0 ; dst + movq [ecx+%1*32+ 8], mm1 + movq [ecx+%1*32+16], mm2 + movq [ecx+%1*32+24], mm3 +%endmacro - punpcklbw mm2, mm7 - punpckhbw mm6, mm7 +align 16 +transfer_8to16sub_mmx: + mov ecx, [esp + 4] ; Dst + mov eax, [esp + 8] ; Cur + push ebx + mov ebx, [esp+4+12] ; Ref + mov edx, [esp+4+16] ; Stride + pxor mm7, mm7 + + COPY_8_TO_16_SUB 0 + COPY_8_TO_16_SUB 1 + COPY_8_TO_16_SUB 2 + COPY_8_TO_16_SUB 3 - psubsw mm1, mm3 + pop ebx + ret - psubsw mm4, mm2 - psubsw mm5, mm6 +;=========================================================================== +; +; void transfer_8to16sub2_mmx(int16_t * const dct, +; uint8_t * const cur, +; const uint8_t * ref1, +; const uint8_t * ref2, +; const uint32_t stride) +; +;=========================================================================== - movq [edi + 96], mm0 ; dct[] = mm01 - movq [edi + 104], mm1 - movq [edi + 112], mm4 - movq [edi + 120], mm5 +%macro COPY_8_TO_16_SUB2_MMX 1 + movq mm0, [eax] ; cur + movq mm2, [eax+edx] + + ; mm4 <- (ref1+ref2+1) / 2 + movq mm4, [ebx] ; ref1 + movq mm1, [esi] ; ref2 + movq mm6, mm4 + movq mm3, mm1 + punpcklbw mm4, mm7 + punpcklbw mm1, mm7 + punpckhbw mm6, mm7 + punpckhbw mm3, mm7 + paddusw mm4, mm1 + paddusw mm6, mm3 + psrlw mm4,1 + psrlw mm6,1 + packuswb mm4, mm6 + + ; mm5 <- (ref1+ref2+1) / 2 + movq mm5, [ebx+edx] ; ref1 + movq mm1, [esi+edx] ; ref2 + movq mm6, mm5 + movq mm3, mm1 + punpcklbw mm5, mm7 + punpcklbw mm1, mm7 + punpckhbw mm6, mm7 + punpckhbw mm3, mm7 + paddusw mm5, mm1 + paddusw mm6, mm3 + lea esi,[esi+2*edx] + psrlw mm5,1 + psrlw mm6,1 + packuswb mm5, mm6 + + + movq mm1, mm0 + movq mm3, mm2 + punpcklbw mm0, mm7 + punpcklbw mm2, mm7 + punpckhbw mm1, mm7 + punpckhbw mm3, mm7 + + movq mm6, mm4 + punpcklbw mm4, mm7 + punpckhbw mm6, mm7 + psubsw mm0, mm4 + psubsw mm1, mm6 + movq mm6, mm5 + punpcklbw mm5, mm7 + punpckhbw mm6, mm7 + psubsw mm2, mm5 + lea eax,[eax+2*edx] + psubsw mm3, mm6 + lea ebx,[ebx+2*edx] + + movq [ecx+%1*32+ 0], mm0 ; dst + movq [ecx+%1*32+ 8], mm1 + movq [ecx+%1*32+16], mm2 + movq [ecx+%1*32+24], mm3 +%endmacro - pop ebx - pop edi - pop esi +align 16 +transfer_8to16sub2_mmx: + mov ecx, [esp + 4] ; Dst + mov eax, [esp + 8] ; Cur + push ebx + mov ebx, [esp+4+12] ; Ref1 + push esi + mov esi, [esp+8+16] ; Ref2 + mov edx, [esp+8+20] ; Stride + pxor mm7, mm7 + + COPY_8_TO_16_SUB2_MMX 0 + COPY_8_TO_16_SUB2_MMX 1 + COPY_8_TO_16_SUB2_MMX 2 + COPY_8_TO_16_SUB2_MMX 3 + + pop esi + pop ebx + ret - ret +;=========================================================================== +; +; void transfer_8to16sub2_xmm(int16_t * const dct, +; uint8_t * const cur, +; const uint8_t * ref1, +; const uint8_t * ref2, +; const uint32_t stride) +; +;=========================================================================== +%macro COPY_8_TO_16_SUB2_SSE 1 + movq mm0, [eax] ; cur + movq mm2, [eax+edx] + movq mm1, mm0 + movq mm3, mm2 + + punpcklbw mm0, mm7 + punpcklbw mm2, mm7 + movq mm4, [ebx] ; ref1 + pavgb mm4, [esi] ; ref2 + punpckhbw mm1, mm7 + punpckhbw mm3, mm7 + movq mm5, [ebx+edx] ; ref + pavgb mm5, [esi+edx] ; ref2 + + movq mm6, mm4 + punpcklbw mm4, mm7 + punpckhbw mm6, mm7 + psubsw mm0, mm4 + psubsw mm1, mm6 + lea esi,[esi+2*edx] + movq mm6, mm5 + punpcklbw mm5, mm7 + punpckhbw mm6, mm7 + psubsw mm2, mm5 + lea eax,[eax+2*edx] + psubsw mm3, mm6 + lea ebx,[ebx+2*edx] + + movq [ecx+%1*32+ 0], mm0 ; dst + movq [ecx+%1*32+ 8], mm1 + movq [ecx+%1*32+16], mm2 + movq [ecx+%1*32+24], mm3 +%endmacro +align 16 +transfer_8to16sub2_xmm: + mov ecx, [esp + 4] ; Dst + mov eax, [esp + 8] ; Cur + push ebx + mov ebx, [esp+4+12] ; Ref1 + push esi + mov esi, [esp+8+16] ; Ref2 + mov edx, [esp+8+20] ; Stride + pxor mm7, mm7 + + COPY_8_TO_16_SUB2_SSE 0 + COPY_8_TO_16_SUB2_SSE 1 + COPY_8_TO_16_SUB2_SSE 2 + COPY_8_TO_16_SUB2_SSE 3 + + pop esi + pop ebx + ret ;=========================================================================== ; @@ -367,46 +371,41 @@ ; ;=========================================================================== -align 16 -cglobal transfer_16to8add_mmx -transfer_16to8add_mmx - - push esi - push edi - - mov edi, [esp + 8 + 4] ; dst - mov esi, [esp + 8 + 8] ; src - mov ecx, [esp + 8 + 12] ; stride - - pxor mm7, mm7 - - mov eax, 8 - -.loop - movq mm0, [edi] - movq mm1, mm0 - punpcklbw mm0, mm7 ; mm23 = unpack([dst]) - punpckhbw mm1, mm7 - - movq mm2, [esi] ; mm01 = [src] - movq mm3, [esi + 8] - - paddsw mm0, mm2 ; mm01 += mm23 - paddsw mm1, mm3 - - packuswb mm0, mm1 ; [dst] = pack(mm01) - movq [edi], mm0 - - add esi, 16 - add edi, ecx - dec eax - jnz .loop - - pop edi - pop esi +%macro COPY_16_TO_8_ADD 1 + movq mm0, [ecx] + movq mm2, [ecx+edx] + movq mm1, mm0 + movq mm3, mm2 + punpcklbw mm0, mm7 + punpcklbw mm2, mm7 + punpckhbw mm1, mm7 + punpckhbw mm3, mm7 + paddsw mm0, [eax+%1*32+ 0] + paddsw mm1, [eax+%1*32+ 8] + paddsw mm2, [eax+%1*32+16] + paddsw mm3, [eax+%1*32+24] + packuswb mm0, mm1 + movq [ecx], mm0 + packuswb mm2, mm3 + movq [ecx+edx], mm2 +%endmacro - ret +align 16 +transfer_16to8add_mmx: + mov ecx, [esp+ 4] ; Dst + mov eax, [esp+ 8] ; Src + mov edx, [esp+12] ; Stride + pxor mm7, mm7 + + COPY_16_TO_8_ADD 0 + lea ecx,[ecx+2*edx] + COPY_16_TO_8_ADD 1 + lea ecx,[ecx+2*edx] + COPY_16_TO_8_ADD 2 + lea ecx,[ecx+2*edx] + COPY_16_TO_8_ADD 3 + ret ;=========================================================================== ; @@ -417,57 +416,25 @@ ; ;=========================================================================== -align 16 -cglobal transfer8x8_copy_mmx -transfer8x8_copy_mmx - push esi - push edi - - mov edi, [esp + 8 + 4] ; dst [out] - mov esi, [esp + 8 + 8] ; src [in] - mov eax, [esp + 8 + 12] ; stride [in] - - movq mm0, [esi] - movq mm1, [esi+eax] - movq [edi], mm0 - movq [edi+eax], mm1 - - add esi, eax - add edi, eax - add esi, eax - add edi, eax - - movq mm0, [esi] - movq mm1, [esi+eax] - movq [edi], mm0 - movq [edi+eax], mm1 - - add esi, eax - add edi, eax - add esi, eax - add edi, eax - - movq mm0, [esi] - movq mm1, [esi+eax] - movq [edi], mm0 - movq [edi+eax], mm1 - - add esi, eax - add edi, eax - add esi, eax - add edi, eax - - movq mm0, [esi] - movq mm1, [esi+eax] - movq [edi], mm0 - movq [edi+eax], mm1 - - add esi, eax - add edi, eax - add esi, eax - add edi, eax - - pop edi - pop esi +%macro COPY_8_TO_8 0 + movq mm0, [eax] + movq mm1, [eax+edx] + movq [ecx], mm0 + lea eax,[eax+2*edx] + movq [ecx+edx], mm1 +%endmacro - ret +align 16 +transfer8x8_copy_mmx: + mov ecx, [esp+ 4] ; Dst + mov eax, [esp+ 8] ; Src + mov edx, [esp+12] ; Stride + + COPY_8_TO_8 + lea ecx,[ecx+2*edx] + COPY_8_TO_8 + lea ecx,[ecx+2*edx] + COPY_8_TO_8 + lea ecx,[ecx+2*edx] + COPY_8_TO_8 + ret