michael@0: ; michael@0: ; jimmxint.asm - accurate integer IDCT (MMX) michael@0: ; michael@0: ; Copyright 2009 Pierre Ossman for Cendio AB michael@0: ; michael@0: ; Based on michael@0: ; x86 SIMD extension for IJG JPEG library michael@0: ; Copyright (C) 1999-2006, MIYASAKA Masaru. michael@0: ; For conditions of distribution and use, see copyright notice in jsimdext.inc michael@0: ; michael@0: ; This file should be assembled with NASM (Netwide Assembler), michael@0: ; can *not* be assembled with Microsoft's MASM or any compatible michael@0: ; assembler (including Borland's Turbo Assembler). michael@0: ; NASM is available from http://nasm.sourceforge.net/ or michael@0: ; http://sourceforge.net/project/showfiles.php?group_id=6208 michael@0: ; michael@0: ; This file contains a slow-but-accurate integer implementation of the michael@0: ; inverse DCT (Discrete Cosine Transform). The following code is based michael@0: ; directly on the IJG's original jidctint.c; see the jidctint.c for michael@0: ; more details. michael@0: ; michael@0: ; [TAB8] michael@0: michael@0: %include "jsimdext.inc" michael@0: %include "jdct.inc" michael@0: michael@0: ; -------------------------------------------------------------------------- michael@0: michael@0: %define CONST_BITS 13 michael@0: %define PASS1_BITS 2 michael@0: michael@0: %define DESCALE_P1 (CONST_BITS-PASS1_BITS) michael@0: %define DESCALE_P2 (CONST_BITS+PASS1_BITS+3) michael@0: michael@0: %if CONST_BITS == 13 michael@0: F_0_298 equ 2446 ; FIX(0.298631336) michael@0: F_0_390 equ 3196 ; FIX(0.390180644) michael@0: F_0_541 equ 4433 ; FIX(0.541196100) michael@0: F_0_765 equ 6270 ; FIX(0.765366865) michael@0: F_0_899 equ 7373 ; FIX(0.899976223) michael@0: F_1_175 equ 9633 ; FIX(1.175875602) michael@0: F_1_501 equ 12299 ; FIX(1.501321110) michael@0: F_1_847 equ 15137 ; FIX(1.847759065) michael@0: F_1_961 equ 16069 ; FIX(1.961570560) michael@0: F_2_053 equ 16819 ; FIX(2.053119869) michael@0: F_2_562 equ 20995 ; FIX(2.562915447) michael@0: F_3_072 equ 25172 ; FIX(3.072711026) michael@0: %else michael@0: ; NASM cannot do compile-time arithmetic on floating-point constants. michael@0: %define DESCALE(x,n) (((x)+(1<<((n)-1)))>>(n)) michael@0: F_0_298 equ DESCALE( 320652955,30-CONST_BITS) ; FIX(0.298631336) michael@0: F_0_390 equ DESCALE( 418953276,30-CONST_BITS) ; FIX(0.390180644) michael@0: F_0_541 equ DESCALE( 581104887,30-CONST_BITS) ; FIX(0.541196100) michael@0: F_0_765 equ DESCALE( 821806413,30-CONST_BITS) ; FIX(0.765366865) michael@0: F_0_899 equ DESCALE( 966342111,30-CONST_BITS) ; FIX(0.899976223) michael@0: F_1_175 equ DESCALE(1262586813,30-CONST_BITS) ; FIX(1.175875602) michael@0: F_1_501 equ DESCALE(1612031267,30-CONST_BITS) ; FIX(1.501321110) michael@0: F_1_847 equ DESCALE(1984016188,30-CONST_BITS) ; FIX(1.847759065) michael@0: F_1_961 equ DESCALE(2106220350,30-CONST_BITS) ; FIX(1.961570560) michael@0: F_2_053 equ DESCALE(2204520673,30-CONST_BITS) ; FIX(2.053119869) michael@0: F_2_562 equ DESCALE(2751909506,30-CONST_BITS) ; FIX(2.562915447) michael@0: F_3_072 equ DESCALE(3299298341,30-CONST_BITS) ; FIX(3.072711026) michael@0: %endif michael@0: michael@0: ; -------------------------------------------------------------------------- michael@0: SECTION SEG_CONST michael@0: michael@0: alignz 16 michael@0: global EXTN(jconst_idct_islow_mmx) michael@0: michael@0: EXTN(jconst_idct_islow_mmx): michael@0: michael@0: PW_F130_F054 times 2 dw (F_0_541+F_0_765), F_0_541 michael@0: PW_F054_MF130 times 2 dw F_0_541, (F_0_541-F_1_847) michael@0: PW_MF078_F117 times 2 dw (F_1_175-F_1_961), F_1_175 michael@0: PW_F117_F078 times 2 dw F_1_175, (F_1_175-F_0_390) michael@0: PW_MF060_MF089 times 2 dw (F_0_298-F_0_899),-F_0_899 michael@0: PW_MF089_F060 times 2 dw -F_0_899, (F_1_501-F_0_899) michael@0: PW_MF050_MF256 times 2 dw (F_2_053-F_2_562),-F_2_562 michael@0: PW_MF256_F050 times 2 dw -F_2_562, (F_3_072-F_2_562) michael@0: PD_DESCALE_P1 times 2 dd 1 << (DESCALE_P1-1) michael@0: PD_DESCALE_P2 times 2 dd 1 << (DESCALE_P2-1) michael@0: PB_CENTERJSAMP times 8 db CENTERJSAMPLE michael@0: michael@0: alignz 16 michael@0: michael@0: ; -------------------------------------------------------------------------- michael@0: SECTION SEG_TEXT michael@0: BITS 32 michael@0: ; michael@0: ; Perform dequantization and inverse DCT on one block of coefficients. michael@0: ; michael@0: ; GLOBAL(void) michael@0: ; jsimd_idct_islow_mmx (void * dct_table, JCOEFPTR coef_block, michael@0: ; JSAMPARRAY output_buf, JDIMENSION output_col) michael@0: ; michael@0: michael@0: %define dct_table(b) (b)+8 ; jpeg_component_info * compptr michael@0: %define coef_block(b) (b)+12 ; JCOEFPTR coef_block michael@0: %define output_buf(b) (b)+16 ; JSAMPARRAY output_buf michael@0: %define output_col(b) (b)+20 ; JDIMENSION output_col michael@0: michael@0: %define original_ebp ebp+0 michael@0: %define wk(i) ebp-(WK_NUM-(i))*SIZEOF_MMWORD ; mmword wk[WK_NUM] michael@0: %define WK_NUM 12 michael@0: %define workspace wk(0)-DCTSIZE2*SIZEOF_JCOEF michael@0: ; JCOEF workspace[DCTSIZE2] michael@0: michael@0: align 16 michael@0: global EXTN(jsimd_idct_islow_mmx) michael@0: michael@0: EXTN(jsimd_idct_islow_mmx): michael@0: push ebp michael@0: mov eax,esp ; eax = original ebp michael@0: sub esp, byte 4 michael@0: and esp, byte (-SIZEOF_MMWORD) ; align to 64 bits michael@0: mov [esp],eax michael@0: mov ebp,esp ; ebp = aligned ebp michael@0: lea esp, [workspace] michael@0: push ebx michael@0: ; push ecx ; need not be preserved michael@0: ; push edx ; need not be preserved michael@0: push esi michael@0: push edi michael@0: michael@0: get_GOT ebx ; get GOT address michael@0: michael@0: ; ---- Pass 1: process columns from input, store into work array. michael@0: michael@0: ; mov eax, [original_ebp] michael@0: mov edx, POINTER [dct_table(eax)] ; quantptr michael@0: mov esi, JCOEFPTR [coef_block(eax)] ; inptr michael@0: lea edi, [workspace] ; JCOEF * wsptr michael@0: mov ecx, DCTSIZE/4 ; ctr michael@0: alignx 16,7 michael@0: .columnloop: michael@0: %ifndef NO_ZERO_COLUMN_TEST_ISLOW_MMX michael@0: mov eax, DWORD [DWBLOCK(1,0,esi,SIZEOF_JCOEF)] michael@0: or eax, DWORD [DWBLOCK(2,0,esi,SIZEOF_JCOEF)] michael@0: jnz short .columnDCT michael@0: michael@0: movq mm0, MMWORD [MMBLOCK(1,0,esi,SIZEOF_JCOEF)] michael@0: movq mm1, MMWORD [MMBLOCK(2,0,esi,SIZEOF_JCOEF)] michael@0: por mm0, MMWORD [MMBLOCK(3,0,esi,SIZEOF_JCOEF)] michael@0: por mm1, MMWORD [MMBLOCK(4,0,esi,SIZEOF_JCOEF)] michael@0: por mm0, MMWORD [MMBLOCK(5,0,esi,SIZEOF_JCOEF)] michael@0: por mm1, MMWORD [MMBLOCK(6,0,esi,SIZEOF_JCOEF)] michael@0: por mm0, MMWORD [MMBLOCK(7,0,esi,SIZEOF_JCOEF)] michael@0: por mm1,mm0 michael@0: packsswb mm1,mm1 michael@0: movd eax,mm1 michael@0: test eax,eax michael@0: jnz short .columnDCT michael@0: michael@0: ; -- AC terms all zero michael@0: michael@0: movq mm0, MMWORD [MMBLOCK(0,0,esi,SIZEOF_JCOEF)] michael@0: pmullw mm0, MMWORD [MMBLOCK(0,0,edx,SIZEOF_ISLOW_MULT_TYPE)] michael@0: michael@0: psllw mm0,PASS1_BITS michael@0: michael@0: movq mm2,mm0 ; mm0=in0=(00 01 02 03) michael@0: punpcklwd mm0,mm0 ; mm0=(00 00 01 01) michael@0: punpckhwd mm2,mm2 ; mm2=(02 02 03 03) michael@0: michael@0: movq mm1,mm0 michael@0: punpckldq mm0,mm0 ; mm0=(00 00 00 00) michael@0: punpckhdq mm1,mm1 ; mm1=(01 01 01 01) michael@0: movq mm3,mm2 michael@0: punpckldq mm2,mm2 ; mm2=(02 02 02 02) michael@0: punpckhdq mm3,mm3 ; mm3=(03 03 03 03) michael@0: michael@0: movq MMWORD [MMBLOCK(0,0,edi,SIZEOF_JCOEF)], mm0 michael@0: movq MMWORD [MMBLOCK(0,1,edi,SIZEOF_JCOEF)], mm0 michael@0: movq MMWORD [MMBLOCK(1,0,edi,SIZEOF_JCOEF)], mm1 michael@0: movq MMWORD [MMBLOCK(1,1,edi,SIZEOF_JCOEF)], mm1 michael@0: movq MMWORD [MMBLOCK(2,0,edi,SIZEOF_JCOEF)], mm2 michael@0: movq MMWORD [MMBLOCK(2,1,edi,SIZEOF_JCOEF)], mm2 michael@0: movq MMWORD [MMBLOCK(3,0,edi,SIZEOF_JCOEF)], mm3 michael@0: movq MMWORD [MMBLOCK(3,1,edi,SIZEOF_JCOEF)], mm3 michael@0: jmp near .nextcolumn michael@0: alignx 16,7 michael@0: %endif michael@0: .columnDCT: michael@0: michael@0: ; -- Even part michael@0: michael@0: movq mm0, MMWORD [MMBLOCK(0,0,esi,SIZEOF_JCOEF)] michael@0: movq mm1, MMWORD [MMBLOCK(2,0,esi,SIZEOF_JCOEF)] michael@0: pmullw mm0, MMWORD [MMBLOCK(0,0,edx,SIZEOF_ISLOW_MULT_TYPE)] michael@0: pmullw mm1, MMWORD [MMBLOCK(2,0,edx,SIZEOF_ISLOW_MULT_TYPE)] michael@0: movq mm2, MMWORD [MMBLOCK(4,0,esi,SIZEOF_JCOEF)] michael@0: movq mm3, MMWORD [MMBLOCK(6,0,esi,SIZEOF_JCOEF)] michael@0: pmullw mm2, MMWORD [MMBLOCK(4,0,edx,SIZEOF_ISLOW_MULT_TYPE)] michael@0: pmullw mm3, MMWORD [MMBLOCK(6,0,edx,SIZEOF_ISLOW_MULT_TYPE)] michael@0: michael@0: ; (Original) michael@0: ; z1 = (z2 + z3) * 0.541196100; michael@0: ; tmp2 = z1 + z3 * -1.847759065; michael@0: ; tmp3 = z1 + z2 * 0.765366865; michael@0: ; michael@0: ; (This implementation) michael@0: ; tmp2 = z2 * 0.541196100 + z3 * (0.541196100 - 1.847759065); michael@0: ; tmp3 = z2 * (0.541196100 + 0.765366865) + z3 * 0.541196100; michael@0: michael@0: movq mm4,mm1 ; mm1=in2=z2 michael@0: movq mm5,mm1 michael@0: punpcklwd mm4,mm3 ; mm3=in6=z3 michael@0: punpckhwd mm5,mm3 michael@0: movq mm1,mm4 michael@0: movq mm3,mm5 michael@0: pmaddwd mm4,[GOTOFF(ebx,PW_F130_F054)] ; mm4=tmp3L michael@0: pmaddwd mm5,[GOTOFF(ebx,PW_F130_F054)] ; mm5=tmp3H michael@0: pmaddwd mm1,[GOTOFF(ebx,PW_F054_MF130)] ; mm1=tmp2L michael@0: pmaddwd mm3,[GOTOFF(ebx,PW_F054_MF130)] ; mm3=tmp2H michael@0: michael@0: movq mm6,mm0 michael@0: paddw mm0,mm2 ; mm0=in0+in4 michael@0: psubw mm6,mm2 ; mm6=in0-in4 michael@0: michael@0: pxor mm7,mm7 michael@0: pxor mm2,mm2 michael@0: punpcklwd mm7,mm0 ; mm7=tmp0L michael@0: punpckhwd mm2,mm0 ; mm2=tmp0H michael@0: psrad mm7,(16-CONST_BITS) ; psrad mm7,16 & pslld mm7,CONST_BITS michael@0: psrad mm2,(16-CONST_BITS) ; psrad mm2,16 & pslld mm2,CONST_BITS michael@0: michael@0: movq mm0,mm7 michael@0: paddd mm7,mm4 ; mm7=tmp10L michael@0: psubd mm0,mm4 ; mm0=tmp13L michael@0: movq mm4,mm2 michael@0: paddd mm2,mm5 ; mm2=tmp10H michael@0: psubd mm4,mm5 ; mm4=tmp13H michael@0: michael@0: movq MMWORD [wk(0)], mm7 ; wk(0)=tmp10L michael@0: movq MMWORD [wk(1)], mm2 ; wk(1)=tmp10H michael@0: movq MMWORD [wk(2)], mm0 ; wk(2)=tmp13L michael@0: movq MMWORD [wk(3)], mm4 ; wk(3)=tmp13H michael@0: michael@0: pxor mm5,mm5 michael@0: pxor mm7,mm7 michael@0: punpcklwd mm5,mm6 ; mm5=tmp1L michael@0: punpckhwd mm7,mm6 ; mm7=tmp1H michael@0: psrad mm5,(16-CONST_BITS) ; psrad mm5,16 & pslld mm5,CONST_BITS michael@0: psrad mm7,(16-CONST_BITS) ; psrad mm7,16 & pslld mm7,CONST_BITS michael@0: michael@0: movq mm2,mm5 michael@0: paddd mm5,mm1 ; mm5=tmp11L michael@0: psubd mm2,mm1 ; mm2=tmp12L michael@0: movq mm0,mm7 michael@0: paddd mm7,mm3 ; mm7=tmp11H michael@0: psubd mm0,mm3 ; mm0=tmp12H michael@0: michael@0: movq MMWORD [wk(4)], mm5 ; wk(4)=tmp11L michael@0: movq MMWORD [wk(5)], mm7 ; wk(5)=tmp11H michael@0: movq MMWORD [wk(6)], mm2 ; wk(6)=tmp12L michael@0: movq MMWORD [wk(7)], mm0 ; wk(7)=tmp12H michael@0: michael@0: ; -- Odd part michael@0: michael@0: movq mm4, MMWORD [MMBLOCK(1,0,esi,SIZEOF_JCOEF)] michael@0: movq mm6, MMWORD [MMBLOCK(3,0,esi,SIZEOF_JCOEF)] michael@0: pmullw mm4, MMWORD [MMBLOCK(1,0,edx,SIZEOF_ISLOW_MULT_TYPE)] michael@0: pmullw mm6, MMWORD [MMBLOCK(3,0,edx,SIZEOF_ISLOW_MULT_TYPE)] michael@0: movq mm1, MMWORD [MMBLOCK(5,0,esi,SIZEOF_JCOEF)] michael@0: movq mm3, MMWORD [MMBLOCK(7,0,esi,SIZEOF_JCOEF)] michael@0: pmullw mm1, MMWORD [MMBLOCK(5,0,edx,SIZEOF_ISLOW_MULT_TYPE)] michael@0: pmullw mm3, MMWORD [MMBLOCK(7,0,edx,SIZEOF_ISLOW_MULT_TYPE)] michael@0: michael@0: movq mm5,mm6 michael@0: movq mm7,mm4 michael@0: paddw mm5,mm3 ; mm5=z3 michael@0: paddw mm7,mm1 ; mm7=z4 michael@0: michael@0: ; (Original) michael@0: ; z5 = (z3 + z4) * 1.175875602; michael@0: ; z3 = z3 * -1.961570560; z4 = z4 * -0.390180644; michael@0: ; z3 += z5; z4 += z5; michael@0: ; michael@0: ; (This implementation) michael@0: ; z3 = z3 * (1.175875602 - 1.961570560) + z4 * 1.175875602; michael@0: ; z4 = z3 * 1.175875602 + z4 * (1.175875602 - 0.390180644); michael@0: michael@0: movq mm2,mm5 michael@0: movq mm0,mm5 michael@0: punpcklwd mm2,mm7 michael@0: punpckhwd mm0,mm7 michael@0: movq mm5,mm2 michael@0: movq mm7,mm0 michael@0: pmaddwd mm2,[GOTOFF(ebx,PW_MF078_F117)] ; mm2=z3L michael@0: pmaddwd mm0,[GOTOFF(ebx,PW_MF078_F117)] ; mm0=z3H michael@0: pmaddwd mm5,[GOTOFF(ebx,PW_F117_F078)] ; mm5=z4L michael@0: pmaddwd mm7,[GOTOFF(ebx,PW_F117_F078)] ; mm7=z4H michael@0: michael@0: movq MMWORD [wk(10)], mm2 ; wk(10)=z3L michael@0: movq MMWORD [wk(11)], mm0 ; wk(11)=z3H michael@0: michael@0: ; (Original) michael@0: ; z1 = tmp0 + tmp3; z2 = tmp1 + tmp2; michael@0: ; tmp0 = tmp0 * 0.298631336; tmp1 = tmp1 * 2.053119869; michael@0: ; tmp2 = tmp2 * 3.072711026; tmp3 = tmp3 * 1.501321110; michael@0: ; z1 = z1 * -0.899976223; z2 = z2 * -2.562915447; michael@0: ; tmp0 += z1 + z3; tmp1 += z2 + z4; michael@0: ; tmp2 += z2 + z3; tmp3 += z1 + z4; michael@0: ; michael@0: ; (This implementation) michael@0: ; tmp0 = tmp0 * (0.298631336 - 0.899976223) + tmp3 * -0.899976223; michael@0: ; tmp1 = tmp1 * (2.053119869 - 2.562915447) + tmp2 * -2.562915447; michael@0: ; tmp2 = tmp1 * -2.562915447 + tmp2 * (3.072711026 - 2.562915447); michael@0: ; tmp3 = tmp0 * -0.899976223 + tmp3 * (1.501321110 - 0.899976223); michael@0: ; tmp0 += z3; tmp1 += z4; michael@0: ; tmp2 += z3; tmp3 += z4; michael@0: michael@0: movq mm2,mm3 michael@0: movq mm0,mm3 michael@0: punpcklwd mm2,mm4 michael@0: punpckhwd mm0,mm4 michael@0: movq mm3,mm2 michael@0: movq mm4,mm0 michael@0: pmaddwd mm2,[GOTOFF(ebx,PW_MF060_MF089)] ; mm2=tmp0L michael@0: pmaddwd mm0,[GOTOFF(ebx,PW_MF060_MF089)] ; mm0=tmp0H michael@0: pmaddwd mm3,[GOTOFF(ebx,PW_MF089_F060)] ; mm3=tmp3L michael@0: pmaddwd mm4,[GOTOFF(ebx,PW_MF089_F060)] ; mm4=tmp3H michael@0: michael@0: paddd mm2, MMWORD [wk(10)] ; mm2=tmp0L michael@0: paddd mm0, MMWORD [wk(11)] ; mm0=tmp0H michael@0: paddd mm3,mm5 ; mm3=tmp3L michael@0: paddd mm4,mm7 ; mm4=tmp3H michael@0: michael@0: movq MMWORD [wk(8)], mm2 ; wk(8)=tmp0L michael@0: movq MMWORD [wk(9)], mm0 ; wk(9)=tmp0H michael@0: michael@0: movq mm2,mm1 michael@0: movq mm0,mm1 michael@0: punpcklwd mm2,mm6 michael@0: punpckhwd mm0,mm6 michael@0: movq mm1,mm2 michael@0: movq mm6,mm0 michael@0: pmaddwd mm2,[GOTOFF(ebx,PW_MF050_MF256)] ; mm2=tmp1L michael@0: pmaddwd mm0,[GOTOFF(ebx,PW_MF050_MF256)] ; mm0=tmp1H michael@0: pmaddwd mm1,[GOTOFF(ebx,PW_MF256_F050)] ; mm1=tmp2L michael@0: pmaddwd mm6,[GOTOFF(ebx,PW_MF256_F050)] ; mm6=tmp2H michael@0: michael@0: paddd mm2,mm5 ; mm2=tmp1L michael@0: paddd mm0,mm7 ; mm0=tmp1H michael@0: paddd mm1, MMWORD [wk(10)] ; mm1=tmp2L michael@0: paddd mm6, MMWORD [wk(11)] ; mm6=tmp2H michael@0: michael@0: movq MMWORD [wk(10)], mm2 ; wk(10)=tmp1L michael@0: movq MMWORD [wk(11)], mm0 ; wk(11)=tmp1H michael@0: michael@0: ; -- Final output stage michael@0: michael@0: movq mm5, MMWORD [wk(0)] ; mm5=tmp10L michael@0: movq mm7, MMWORD [wk(1)] ; mm7=tmp10H michael@0: michael@0: movq mm2,mm5 michael@0: movq mm0,mm7 michael@0: paddd mm5,mm3 ; mm5=data0L michael@0: paddd mm7,mm4 ; mm7=data0H michael@0: psubd mm2,mm3 ; mm2=data7L michael@0: psubd mm0,mm4 ; mm0=data7H michael@0: michael@0: movq mm3,[GOTOFF(ebx,PD_DESCALE_P1)] ; mm3=[PD_DESCALE_P1] michael@0: michael@0: paddd mm5,mm3 michael@0: paddd mm7,mm3 michael@0: psrad mm5,DESCALE_P1 michael@0: psrad mm7,DESCALE_P1 michael@0: paddd mm2,mm3 michael@0: paddd mm0,mm3 michael@0: psrad mm2,DESCALE_P1 michael@0: psrad mm0,DESCALE_P1 michael@0: michael@0: packssdw mm5,mm7 ; mm5=data0=(00 01 02 03) michael@0: packssdw mm2,mm0 ; mm2=data7=(70 71 72 73) michael@0: michael@0: movq mm4, MMWORD [wk(4)] ; mm4=tmp11L michael@0: movq mm3, MMWORD [wk(5)] ; mm3=tmp11H michael@0: michael@0: movq mm7,mm4 michael@0: movq mm0,mm3 michael@0: paddd mm4,mm1 ; mm4=data1L michael@0: paddd mm3,mm6 ; mm3=data1H michael@0: psubd mm7,mm1 ; mm7=data6L michael@0: psubd mm0,mm6 ; mm0=data6H michael@0: michael@0: movq mm1,[GOTOFF(ebx,PD_DESCALE_P1)] ; mm1=[PD_DESCALE_P1] michael@0: michael@0: paddd mm4,mm1 michael@0: paddd mm3,mm1 michael@0: psrad mm4,DESCALE_P1 michael@0: psrad mm3,DESCALE_P1 michael@0: paddd mm7,mm1 michael@0: paddd mm0,mm1 michael@0: psrad mm7,DESCALE_P1 michael@0: psrad mm0,DESCALE_P1 michael@0: michael@0: packssdw mm4,mm3 ; mm4=data1=(10 11 12 13) michael@0: packssdw mm7,mm0 ; mm7=data6=(60 61 62 63) michael@0: michael@0: movq mm6,mm5 ; transpose coefficients(phase 1) michael@0: punpcklwd mm5,mm4 ; mm5=(00 10 01 11) michael@0: punpckhwd mm6,mm4 ; mm6=(02 12 03 13) michael@0: movq mm1,mm7 ; transpose coefficients(phase 1) michael@0: punpcklwd mm7,mm2 ; mm7=(60 70 61 71) michael@0: punpckhwd mm1,mm2 ; mm1=(62 72 63 73) michael@0: michael@0: movq mm3, MMWORD [wk(6)] ; mm3=tmp12L michael@0: movq mm0, MMWORD [wk(7)] ; mm0=tmp12H michael@0: movq mm4, MMWORD [wk(10)] ; mm4=tmp1L michael@0: movq mm2, MMWORD [wk(11)] ; mm2=tmp1H michael@0: michael@0: movq MMWORD [wk(0)], mm5 ; wk(0)=(00 10 01 11) michael@0: movq MMWORD [wk(1)], mm6 ; wk(1)=(02 12 03 13) michael@0: movq MMWORD [wk(4)], mm7 ; wk(4)=(60 70 61 71) michael@0: movq MMWORD [wk(5)], mm1 ; wk(5)=(62 72 63 73) michael@0: michael@0: movq mm5,mm3 michael@0: movq mm6,mm0 michael@0: paddd mm3,mm4 ; mm3=data2L michael@0: paddd mm0,mm2 ; mm0=data2H michael@0: psubd mm5,mm4 ; mm5=data5L michael@0: psubd mm6,mm2 ; mm6=data5H michael@0: michael@0: movq mm7,[GOTOFF(ebx,PD_DESCALE_P1)] ; mm7=[PD_DESCALE_P1] michael@0: michael@0: paddd mm3,mm7 michael@0: paddd mm0,mm7 michael@0: psrad mm3,DESCALE_P1 michael@0: psrad mm0,DESCALE_P1 michael@0: paddd mm5,mm7 michael@0: paddd mm6,mm7 michael@0: psrad mm5,DESCALE_P1 michael@0: psrad mm6,DESCALE_P1 michael@0: michael@0: packssdw mm3,mm0 ; mm3=data2=(20 21 22 23) michael@0: packssdw mm5,mm6 ; mm5=data5=(50 51 52 53) michael@0: michael@0: movq mm1, MMWORD [wk(2)] ; mm1=tmp13L michael@0: movq mm4, MMWORD [wk(3)] ; mm4=tmp13H michael@0: movq mm2, MMWORD [wk(8)] ; mm2=tmp0L michael@0: movq mm7, MMWORD [wk(9)] ; mm7=tmp0H michael@0: michael@0: movq mm0,mm1 michael@0: movq mm6,mm4 michael@0: paddd mm1,mm2 ; mm1=data3L michael@0: paddd mm4,mm7 ; mm4=data3H michael@0: psubd mm0,mm2 ; mm0=data4L michael@0: psubd mm6,mm7 ; mm6=data4H michael@0: michael@0: movq mm2,[GOTOFF(ebx,PD_DESCALE_P1)] ; mm2=[PD_DESCALE_P1] michael@0: michael@0: paddd mm1,mm2 michael@0: paddd mm4,mm2 michael@0: psrad mm1,DESCALE_P1 michael@0: psrad mm4,DESCALE_P1 michael@0: paddd mm0,mm2 michael@0: paddd mm6,mm2 michael@0: psrad mm0,DESCALE_P1 michael@0: psrad mm6,DESCALE_P1 michael@0: michael@0: packssdw mm1,mm4 ; mm1=data3=(30 31 32 33) michael@0: packssdw mm0,mm6 ; mm0=data4=(40 41 42 43) michael@0: michael@0: movq mm7, MMWORD [wk(0)] ; mm7=(00 10 01 11) michael@0: movq mm2, MMWORD [wk(1)] ; mm2=(02 12 03 13) michael@0: michael@0: movq mm4,mm3 ; transpose coefficients(phase 1) michael@0: punpcklwd mm3,mm1 ; mm3=(20 30 21 31) michael@0: punpckhwd mm4,mm1 ; mm4=(22 32 23 33) michael@0: movq mm6,mm0 ; transpose coefficients(phase 1) michael@0: punpcklwd mm0,mm5 ; mm0=(40 50 41 51) michael@0: punpckhwd mm6,mm5 ; mm6=(42 52 43 53) michael@0: michael@0: movq mm1,mm7 ; transpose coefficients(phase 2) michael@0: punpckldq mm7,mm3 ; mm7=(00 10 20 30) michael@0: punpckhdq mm1,mm3 ; mm1=(01 11 21 31) michael@0: movq mm5,mm2 ; transpose coefficients(phase 2) michael@0: punpckldq mm2,mm4 ; mm2=(02 12 22 32) michael@0: punpckhdq mm5,mm4 ; mm5=(03 13 23 33) michael@0: michael@0: movq mm3, MMWORD [wk(4)] ; mm3=(60 70 61 71) michael@0: movq mm4, MMWORD [wk(5)] ; mm4=(62 72 63 73) michael@0: michael@0: movq MMWORD [MMBLOCK(0,0,edi,SIZEOF_JCOEF)], mm7 michael@0: movq MMWORD [MMBLOCK(1,0,edi,SIZEOF_JCOEF)], mm1 michael@0: movq MMWORD [MMBLOCK(2,0,edi,SIZEOF_JCOEF)], mm2 michael@0: movq MMWORD [MMBLOCK(3,0,edi,SIZEOF_JCOEF)], mm5 michael@0: michael@0: movq mm7,mm0 ; transpose coefficients(phase 2) michael@0: punpckldq mm0,mm3 ; mm0=(40 50 60 70) michael@0: punpckhdq mm7,mm3 ; mm7=(41 51 61 71) michael@0: movq mm1,mm6 ; transpose coefficients(phase 2) michael@0: punpckldq mm6,mm4 ; mm6=(42 52 62 72) michael@0: punpckhdq mm1,mm4 ; mm1=(43 53 63 73) michael@0: michael@0: movq MMWORD [MMBLOCK(0,1,edi,SIZEOF_JCOEF)], mm0 michael@0: movq MMWORD [MMBLOCK(1,1,edi,SIZEOF_JCOEF)], mm7 michael@0: movq MMWORD [MMBLOCK(2,1,edi,SIZEOF_JCOEF)], mm6 michael@0: movq MMWORD [MMBLOCK(3,1,edi,SIZEOF_JCOEF)], mm1 michael@0: michael@0: .nextcolumn: michael@0: add esi, byte 4*SIZEOF_JCOEF ; coef_block michael@0: add edx, byte 4*SIZEOF_ISLOW_MULT_TYPE ; quantptr michael@0: add edi, byte 4*DCTSIZE*SIZEOF_JCOEF ; wsptr michael@0: dec ecx ; ctr michael@0: jnz near .columnloop michael@0: michael@0: ; ---- Pass 2: process rows from work array, store into output array. michael@0: michael@0: mov eax, [original_ebp] michael@0: lea esi, [workspace] ; JCOEF * wsptr michael@0: mov edi, JSAMPARRAY [output_buf(eax)] ; (JSAMPROW *) michael@0: mov eax, JDIMENSION [output_col(eax)] michael@0: mov ecx, DCTSIZE/4 ; ctr michael@0: alignx 16,7 michael@0: .rowloop: michael@0: michael@0: ; -- Even part michael@0: michael@0: movq mm0, MMWORD [MMBLOCK(0,0,esi,SIZEOF_JCOEF)] michael@0: movq mm1, MMWORD [MMBLOCK(2,0,esi,SIZEOF_JCOEF)] michael@0: movq mm2, MMWORD [MMBLOCK(4,0,esi,SIZEOF_JCOEF)] michael@0: movq mm3, MMWORD [MMBLOCK(6,0,esi,SIZEOF_JCOEF)] michael@0: michael@0: ; (Original) michael@0: ; z1 = (z2 + z3) * 0.541196100; michael@0: ; tmp2 = z1 + z3 * -1.847759065; michael@0: ; tmp3 = z1 + z2 * 0.765366865; michael@0: ; michael@0: ; (This implementation) michael@0: ; tmp2 = z2 * 0.541196100 + z3 * (0.541196100 - 1.847759065); michael@0: ; tmp3 = z2 * (0.541196100 + 0.765366865) + z3 * 0.541196100; michael@0: michael@0: movq mm4,mm1 ; mm1=in2=z2 michael@0: movq mm5,mm1 michael@0: punpcklwd mm4,mm3 ; mm3=in6=z3 michael@0: punpckhwd mm5,mm3 michael@0: movq mm1,mm4 michael@0: movq mm3,mm5 michael@0: pmaddwd mm4,[GOTOFF(ebx,PW_F130_F054)] ; mm4=tmp3L michael@0: pmaddwd mm5,[GOTOFF(ebx,PW_F130_F054)] ; mm5=tmp3H michael@0: pmaddwd mm1,[GOTOFF(ebx,PW_F054_MF130)] ; mm1=tmp2L michael@0: pmaddwd mm3,[GOTOFF(ebx,PW_F054_MF130)] ; mm3=tmp2H michael@0: michael@0: movq mm6,mm0 michael@0: paddw mm0,mm2 ; mm0=in0+in4 michael@0: psubw mm6,mm2 ; mm6=in0-in4 michael@0: michael@0: pxor mm7,mm7 michael@0: pxor mm2,mm2 michael@0: punpcklwd mm7,mm0 ; mm7=tmp0L michael@0: punpckhwd mm2,mm0 ; mm2=tmp0H michael@0: psrad mm7,(16-CONST_BITS) ; psrad mm7,16 & pslld mm7,CONST_BITS michael@0: psrad mm2,(16-CONST_BITS) ; psrad mm2,16 & pslld mm2,CONST_BITS michael@0: michael@0: movq mm0,mm7 michael@0: paddd mm7,mm4 ; mm7=tmp10L michael@0: psubd mm0,mm4 ; mm0=tmp13L michael@0: movq mm4,mm2 michael@0: paddd mm2,mm5 ; mm2=tmp10H michael@0: psubd mm4,mm5 ; mm4=tmp13H michael@0: michael@0: movq MMWORD [wk(0)], mm7 ; wk(0)=tmp10L michael@0: movq MMWORD [wk(1)], mm2 ; wk(1)=tmp10H michael@0: movq MMWORD [wk(2)], mm0 ; wk(2)=tmp13L michael@0: movq MMWORD [wk(3)], mm4 ; wk(3)=tmp13H michael@0: michael@0: pxor mm5,mm5 michael@0: pxor mm7,mm7 michael@0: punpcklwd mm5,mm6 ; mm5=tmp1L michael@0: punpckhwd mm7,mm6 ; mm7=tmp1H michael@0: psrad mm5,(16-CONST_BITS) ; psrad mm5,16 & pslld mm5,CONST_BITS michael@0: psrad mm7,(16-CONST_BITS) ; psrad mm7,16 & pslld mm7,CONST_BITS michael@0: michael@0: movq mm2,mm5 michael@0: paddd mm5,mm1 ; mm5=tmp11L michael@0: psubd mm2,mm1 ; mm2=tmp12L michael@0: movq mm0,mm7 michael@0: paddd mm7,mm3 ; mm7=tmp11H michael@0: psubd mm0,mm3 ; mm0=tmp12H michael@0: michael@0: movq MMWORD [wk(4)], mm5 ; wk(4)=tmp11L michael@0: movq MMWORD [wk(5)], mm7 ; wk(5)=tmp11H michael@0: movq MMWORD [wk(6)], mm2 ; wk(6)=tmp12L michael@0: movq MMWORD [wk(7)], mm0 ; wk(7)=tmp12H michael@0: michael@0: ; -- Odd part michael@0: michael@0: movq mm4, MMWORD [MMBLOCK(1,0,esi,SIZEOF_JCOEF)] michael@0: movq mm6, MMWORD [MMBLOCK(3,0,esi,SIZEOF_JCOEF)] michael@0: movq mm1, MMWORD [MMBLOCK(5,0,esi,SIZEOF_JCOEF)] michael@0: movq mm3, MMWORD [MMBLOCK(7,0,esi,SIZEOF_JCOEF)] michael@0: michael@0: movq mm5,mm6 michael@0: movq mm7,mm4 michael@0: paddw mm5,mm3 ; mm5=z3 michael@0: paddw mm7,mm1 ; mm7=z4 michael@0: michael@0: ; (Original) michael@0: ; z5 = (z3 + z4) * 1.175875602; michael@0: ; z3 = z3 * -1.961570560; z4 = z4 * -0.390180644; michael@0: ; z3 += z5; z4 += z5; michael@0: ; michael@0: ; (This implementation) michael@0: ; z3 = z3 * (1.175875602 - 1.961570560) + z4 * 1.175875602; michael@0: ; z4 = z3 * 1.175875602 + z4 * (1.175875602 - 0.390180644); michael@0: michael@0: movq mm2,mm5 michael@0: movq mm0,mm5 michael@0: punpcklwd mm2,mm7 michael@0: punpckhwd mm0,mm7 michael@0: movq mm5,mm2 michael@0: movq mm7,mm0 michael@0: pmaddwd mm2,[GOTOFF(ebx,PW_MF078_F117)] ; mm2=z3L michael@0: pmaddwd mm0,[GOTOFF(ebx,PW_MF078_F117)] ; mm0=z3H michael@0: pmaddwd mm5,[GOTOFF(ebx,PW_F117_F078)] ; mm5=z4L michael@0: pmaddwd mm7,[GOTOFF(ebx,PW_F117_F078)] ; mm7=z4H michael@0: michael@0: movq MMWORD [wk(10)], mm2 ; wk(10)=z3L michael@0: movq MMWORD [wk(11)], mm0 ; wk(11)=z3H michael@0: michael@0: ; (Original) michael@0: ; z1 = tmp0 + tmp3; z2 = tmp1 + tmp2; michael@0: ; tmp0 = tmp0 * 0.298631336; tmp1 = tmp1 * 2.053119869; michael@0: ; tmp2 = tmp2 * 3.072711026; tmp3 = tmp3 * 1.501321110; michael@0: ; z1 = z1 * -0.899976223; z2 = z2 * -2.562915447; michael@0: ; tmp0 += z1 + z3; tmp1 += z2 + z4; michael@0: ; tmp2 += z2 + z3; tmp3 += z1 + z4; michael@0: ; michael@0: ; (This implementation) michael@0: ; tmp0 = tmp0 * (0.298631336 - 0.899976223) + tmp3 * -0.899976223; michael@0: ; tmp1 = tmp1 * (2.053119869 - 2.562915447) + tmp2 * -2.562915447; michael@0: ; tmp2 = tmp1 * -2.562915447 + tmp2 * (3.072711026 - 2.562915447); michael@0: ; tmp3 = tmp0 * -0.899976223 + tmp3 * (1.501321110 - 0.899976223); michael@0: ; tmp0 += z3; tmp1 += z4; michael@0: ; tmp2 += z3; tmp3 += z4; michael@0: michael@0: movq mm2,mm3 michael@0: movq mm0,mm3 michael@0: punpcklwd mm2,mm4 michael@0: punpckhwd mm0,mm4 michael@0: movq mm3,mm2 michael@0: movq mm4,mm0 michael@0: pmaddwd mm2,[GOTOFF(ebx,PW_MF060_MF089)] ; mm2=tmp0L michael@0: pmaddwd mm0,[GOTOFF(ebx,PW_MF060_MF089)] ; mm0=tmp0H michael@0: pmaddwd mm3,[GOTOFF(ebx,PW_MF089_F060)] ; mm3=tmp3L michael@0: pmaddwd mm4,[GOTOFF(ebx,PW_MF089_F060)] ; mm4=tmp3H michael@0: michael@0: paddd mm2, MMWORD [wk(10)] ; mm2=tmp0L michael@0: paddd mm0, MMWORD [wk(11)] ; mm0=tmp0H michael@0: paddd mm3,mm5 ; mm3=tmp3L michael@0: paddd mm4,mm7 ; mm4=tmp3H michael@0: michael@0: movq MMWORD [wk(8)], mm2 ; wk(8)=tmp0L michael@0: movq MMWORD [wk(9)], mm0 ; wk(9)=tmp0H michael@0: michael@0: movq mm2,mm1 michael@0: movq mm0,mm1 michael@0: punpcklwd mm2,mm6 michael@0: punpckhwd mm0,mm6 michael@0: movq mm1,mm2 michael@0: movq mm6,mm0 michael@0: pmaddwd mm2,[GOTOFF(ebx,PW_MF050_MF256)] ; mm2=tmp1L michael@0: pmaddwd mm0,[GOTOFF(ebx,PW_MF050_MF256)] ; mm0=tmp1H michael@0: pmaddwd mm1,[GOTOFF(ebx,PW_MF256_F050)] ; mm1=tmp2L michael@0: pmaddwd mm6,[GOTOFF(ebx,PW_MF256_F050)] ; mm6=tmp2H michael@0: michael@0: paddd mm2,mm5 ; mm2=tmp1L michael@0: paddd mm0,mm7 ; mm0=tmp1H michael@0: paddd mm1, MMWORD [wk(10)] ; mm1=tmp2L michael@0: paddd mm6, MMWORD [wk(11)] ; mm6=tmp2H michael@0: michael@0: movq MMWORD [wk(10)], mm2 ; wk(10)=tmp1L michael@0: movq MMWORD [wk(11)], mm0 ; wk(11)=tmp1H michael@0: michael@0: ; -- Final output stage michael@0: michael@0: movq mm5, MMWORD [wk(0)] ; mm5=tmp10L michael@0: movq mm7, MMWORD [wk(1)] ; mm7=tmp10H michael@0: michael@0: movq mm2,mm5 michael@0: movq mm0,mm7 michael@0: paddd mm5,mm3 ; mm5=data0L michael@0: paddd mm7,mm4 ; mm7=data0H michael@0: psubd mm2,mm3 ; mm2=data7L michael@0: psubd mm0,mm4 ; mm0=data7H michael@0: michael@0: movq mm3,[GOTOFF(ebx,PD_DESCALE_P2)] ; mm3=[PD_DESCALE_P2] michael@0: michael@0: paddd mm5,mm3 michael@0: paddd mm7,mm3 michael@0: psrad mm5,DESCALE_P2 michael@0: psrad mm7,DESCALE_P2 michael@0: paddd mm2,mm3 michael@0: paddd mm0,mm3 michael@0: psrad mm2,DESCALE_P2 michael@0: psrad mm0,DESCALE_P2 michael@0: michael@0: packssdw mm5,mm7 ; mm5=data0=(00 10 20 30) michael@0: packssdw mm2,mm0 ; mm2=data7=(07 17 27 37) michael@0: michael@0: movq mm4, MMWORD [wk(4)] ; mm4=tmp11L michael@0: movq mm3, MMWORD [wk(5)] ; mm3=tmp11H michael@0: michael@0: movq mm7,mm4 michael@0: movq mm0,mm3 michael@0: paddd mm4,mm1 ; mm4=data1L michael@0: paddd mm3,mm6 ; mm3=data1H michael@0: psubd mm7,mm1 ; mm7=data6L michael@0: psubd mm0,mm6 ; mm0=data6H michael@0: michael@0: movq mm1,[GOTOFF(ebx,PD_DESCALE_P2)] ; mm1=[PD_DESCALE_P2] michael@0: michael@0: paddd mm4,mm1 michael@0: paddd mm3,mm1 michael@0: psrad mm4,DESCALE_P2 michael@0: psrad mm3,DESCALE_P2 michael@0: paddd mm7,mm1 michael@0: paddd mm0,mm1 michael@0: psrad mm7,DESCALE_P2 michael@0: psrad mm0,DESCALE_P2 michael@0: michael@0: packssdw mm4,mm3 ; mm4=data1=(01 11 21 31) michael@0: packssdw mm7,mm0 ; mm7=data6=(06 16 26 36) michael@0: michael@0: packsswb mm5,mm7 ; mm5=(00 10 20 30 06 16 26 36) michael@0: packsswb mm4,mm2 ; mm4=(01 11 21 31 07 17 27 37) michael@0: michael@0: movq mm6, MMWORD [wk(6)] ; mm6=tmp12L michael@0: movq mm1, MMWORD [wk(7)] ; mm1=tmp12H michael@0: movq mm3, MMWORD [wk(10)] ; mm3=tmp1L michael@0: movq mm0, MMWORD [wk(11)] ; mm0=tmp1H michael@0: michael@0: movq MMWORD [wk(0)], mm5 ; wk(0)=(00 10 20 30 06 16 26 36) michael@0: movq MMWORD [wk(1)], mm4 ; wk(1)=(01 11 21 31 07 17 27 37) michael@0: michael@0: movq mm7,mm6 michael@0: movq mm2,mm1 michael@0: paddd mm6,mm3 ; mm6=data2L michael@0: paddd mm1,mm0 ; mm1=data2H michael@0: psubd mm7,mm3 ; mm7=data5L michael@0: psubd mm2,mm0 ; mm2=data5H michael@0: michael@0: movq mm5,[GOTOFF(ebx,PD_DESCALE_P2)] ; mm5=[PD_DESCALE_P2] michael@0: michael@0: paddd mm6,mm5 michael@0: paddd mm1,mm5 michael@0: psrad mm6,DESCALE_P2 michael@0: psrad mm1,DESCALE_P2 michael@0: paddd mm7,mm5 michael@0: paddd mm2,mm5 michael@0: psrad mm7,DESCALE_P2 michael@0: psrad mm2,DESCALE_P2 michael@0: michael@0: packssdw mm6,mm1 ; mm6=data2=(02 12 22 32) michael@0: packssdw mm7,mm2 ; mm7=data5=(05 15 25 35) michael@0: michael@0: movq mm4, MMWORD [wk(2)] ; mm4=tmp13L michael@0: movq mm3, MMWORD [wk(3)] ; mm3=tmp13H michael@0: movq mm0, MMWORD [wk(8)] ; mm0=tmp0L michael@0: movq mm5, MMWORD [wk(9)] ; mm5=tmp0H michael@0: michael@0: movq mm1,mm4 michael@0: movq mm2,mm3 michael@0: paddd mm4,mm0 ; mm4=data3L michael@0: paddd mm3,mm5 ; mm3=data3H michael@0: psubd mm1,mm0 ; mm1=data4L michael@0: psubd mm2,mm5 ; mm2=data4H michael@0: michael@0: movq mm0,[GOTOFF(ebx,PD_DESCALE_P2)] ; mm0=[PD_DESCALE_P2] michael@0: michael@0: paddd mm4,mm0 michael@0: paddd mm3,mm0 michael@0: psrad mm4,DESCALE_P2 michael@0: psrad mm3,DESCALE_P2 michael@0: paddd mm1,mm0 michael@0: paddd mm2,mm0 michael@0: psrad mm1,DESCALE_P2 michael@0: psrad mm2,DESCALE_P2 michael@0: michael@0: movq mm5,[GOTOFF(ebx,PB_CENTERJSAMP)] ; mm5=[PB_CENTERJSAMP] michael@0: michael@0: packssdw mm4,mm3 ; mm4=data3=(03 13 23 33) michael@0: packssdw mm1,mm2 ; mm1=data4=(04 14 24 34) michael@0: michael@0: movq mm0, MMWORD [wk(0)] ; mm0=(00 10 20 30 06 16 26 36) michael@0: movq mm3, MMWORD [wk(1)] ; mm3=(01 11 21 31 07 17 27 37) michael@0: michael@0: packsswb mm6,mm1 ; mm6=(02 12 22 32 04 14 24 34) michael@0: packsswb mm4,mm7 ; mm4=(03 13 23 33 05 15 25 35) michael@0: michael@0: paddb mm0,mm5 michael@0: paddb mm3,mm5 michael@0: paddb mm6,mm5 michael@0: paddb mm4,mm5 michael@0: michael@0: movq mm2,mm0 ; transpose coefficients(phase 1) michael@0: punpcklbw mm0,mm3 ; mm0=(00 01 10 11 20 21 30 31) michael@0: punpckhbw mm2,mm3 ; mm2=(06 07 16 17 26 27 36 37) michael@0: movq mm1,mm6 ; transpose coefficients(phase 1) michael@0: punpcklbw mm6,mm4 ; mm6=(02 03 12 13 22 23 32 33) michael@0: punpckhbw mm1,mm4 ; mm1=(04 05 14 15 24 25 34 35) michael@0: michael@0: movq mm7,mm0 ; transpose coefficients(phase 2) michael@0: punpcklwd mm0,mm6 ; mm0=(00 01 02 03 10 11 12 13) michael@0: punpckhwd mm7,mm6 ; mm7=(20 21 22 23 30 31 32 33) michael@0: movq mm5,mm1 ; transpose coefficients(phase 2) michael@0: punpcklwd mm1,mm2 ; mm1=(04 05 06 07 14 15 16 17) michael@0: punpckhwd mm5,mm2 ; mm5=(24 25 26 27 34 35 36 37) michael@0: michael@0: movq mm3,mm0 ; transpose coefficients(phase 3) michael@0: punpckldq mm0,mm1 ; mm0=(00 01 02 03 04 05 06 07) michael@0: punpckhdq mm3,mm1 ; mm3=(10 11 12 13 14 15 16 17) michael@0: movq mm4,mm7 ; transpose coefficients(phase 3) michael@0: punpckldq mm7,mm5 ; mm7=(20 21 22 23 24 25 26 27) michael@0: punpckhdq mm4,mm5 ; mm4=(30 31 32 33 34 35 36 37) michael@0: michael@0: pushpic ebx ; save GOT address michael@0: michael@0: mov edx, JSAMPROW [edi+0*SIZEOF_JSAMPROW] michael@0: mov ebx, JSAMPROW [edi+1*SIZEOF_JSAMPROW] michael@0: movq MMWORD [edx+eax*SIZEOF_JSAMPLE], mm0 michael@0: movq MMWORD [ebx+eax*SIZEOF_JSAMPLE], mm3 michael@0: mov edx, JSAMPROW [edi+2*SIZEOF_JSAMPROW] michael@0: mov ebx, JSAMPROW [edi+3*SIZEOF_JSAMPROW] michael@0: movq MMWORD [edx+eax*SIZEOF_JSAMPLE], mm7 michael@0: movq MMWORD [ebx+eax*SIZEOF_JSAMPLE], mm4 michael@0: michael@0: poppic ebx ; restore GOT address michael@0: michael@0: add esi, byte 4*SIZEOF_JCOEF ; wsptr michael@0: add edi, byte 4*SIZEOF_JSAMPROW michael@0: dec ecx ; ctr michael@0: jnz near .rowloop michael@0: michael@0: emms ; empty MMX state michael@0: michael@0: pop edi michael@0: pop esi michael@0: ; pop edx ; need not be preserved michael@0: ; pop ecx ; need not be preserved michael@0: pop ebx michael@0: mov esp,ebp ; esp <- aligned ebp michael@0: pop esp ; esp <- original ebp michael@0: pop ebp michael@0: ret michael@0: michael@0: ; For some reason, the OS X linker does not honor the request to align the michael@0: ; segment unless we do this. michael@0: align 16