michael@0: ; michael@0: ; jisseflt.asm - floating-point IDCT (SSE & 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 floating-point implementation of the inverse DCT michael@0: ; (Discrete Cosine Transform). The following code is based directly on michael@0: ; the IJG's original jidctflt.c; see the jidctflt.c for 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: %macro unpcklps2 2 ; %1=(0 1 2 3) / %2=(4 5 6 7) => %1=(0 1 4 5) michael@0: shufps %1,%2,0x44 michael@0: %endmacro michael@0: michael@0: %macro unpckhps2 2 ; %1=(0 1 2 3) / %2=(4 5 6 7) => %1=(2 3 6 7) michael@0: shufps %1,%2,0xEE michael@0: %endmacro michael@0: michael@0: ; -------------------------------------------------------------------------- michael@0: SECTION SEG_CONST michael@0: michael@0: alignz 16 michael@0: global EXTN(jconst_idct_float_sse) michael@0: michael@0: EXTN(jconst_idct_float_sse): michael@0: michael@0: PD_1_414 times 4 dd 1.414213562373095048801689 michael@0: PD_1_847 times 4 dd 1.847759065022573512256366 michael@0: PD_1_082 times 4 dd 1.082392200292393968799446 michael@0: PD_M2_613 times 4 dd -2.613125929752753055713286 michael@0: PD_0_125 times 4 dd 0.125 ; 1/8 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_float_sse (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 ; void * dct_table 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_XMMWORD ; xmmword wk[WK_NUM] michael@0: %define WK_NUM 2 michael@0: %define workspace wk(0)-DCTSIZE2*SIZEOF_FAST_FLOAT michael@0: ; FAST_FLOAT workspace[DCTSIZE2] michael@0: michael@0: align 16 michael@0: global EXTN(jsimd_idct_float_sse) michael@0: michael@0: EXTN(jsimd_idct_float_sse): michael@0: push ebp michael@0: mov eax,esp ; eax = original ebp michael@0: sub esp, byte 4 michael@0: and esp, byte (-SIZEOF_XMMWORD) ; align to 128 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] ; FAST_FLOAT * wsptr michael@0: mov ecx, DCTSIZE/4 ; ctr michael@0: alignx 16,7 michael@0: .columnloop: michael@0: %ifndef NO_ZERO_COLUMN_TEST_FLOAT_SSE 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 near .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: michael@0: punpckhwd mm1,mm0 ; mm1=(** 02 ** 03) michael@0: punpcklwd mm0,mm0 ; mm0=(00 00 01 01) michael@0: psrad mm1,(DWORD_BIT-WORD_BIT) ; mm1=in0H=(02 03) michael@0: psrad mm0,(DWORD_BIT-WORD_BIT) ; mm0=in0L=(00 01) michael@0: cvtpi2ps xmm3,mm1 ; xmm3=(02 03 ** **) michael@0: cvtpi2ps xmm0,mm0 ; xmm0=(00 01 ** **) michael@0: movlhps xmm0,xmm3 ; xmm0=in0=(00 01 02 03) michael@0: michael@0: mulps xmm0, XMMWORD [XMMBLOCK(0,0,edx,SIZEOF_FLOAT_MULT_TYPE)] michael@0: michael@0: movaps xmm1,xmm0 michael@0: movaps xmm2,xmm0 michael@0: movaps xmm3,xmm0 michael@0: michael@0: shufps xmm0,xmm0,0x00 ; xmm0=(00 00 00 00) michael@0: shufps xmm1,xmm1,0x55 ; xmm1=(01 01 01 01) michael@0: shufps xmm2,xmm2,0xAA ; xmm2=(02 02 02 02) michael@0: shufps xmm3,xmm3,0xFF ; xmm3=(03 03 03 03) michael@0: michael@0: movaps XMMWORD [XMMBLOCK(0,0,edi,SIZEOF_FAST_FLOAT)], xmm0 michael@0: movaps XMMWORD [XMMBLOCK(0,1,edi,SIZEOF_FAST_FLOAT)], xmm0 michael@0: movaps XMMWORD [XMMBLOCK(1,0,edi,SIZEOF_FAST_FLOAT)], xmm1 michael@0: movaps XMMWORD [XMMBLOCK(1,1,edi,SIZEOF_FAST_FLOAT)], xmm1 michael@0: movaps XMMWORD [XMMBLOCK(2,0,edi,SIZEOF_FAST_FLOAT)], xmm2 michael@0: movaps XMMWORD [XMMBLOCK(2,1,edi,SIZEOF_FAST_FLOAT)], xmm2 michael@0: movaps XMMWORD [XMMBLOCK(3,0,edi,SIZEOF_FAST_FLOAT)], xmm3 michael@0: movaps XMMWORD [XMMBLOCK(3,1,edi,SIZEOF_FAST_FLOAT)], xmm3 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: movq mm2, MMWORD [MMBLOCK(4,0,esi,SIZEOF_JCOEF)] michael@0: movq mm3, MMWORD [MMBLOCK(6,0,esi,SIZEOF_JCOEF)] michael@0: michael@0: punpckhwd mm4,mm0 ; mm4=(** 02 ** 03) michael@0: punpcklwd mm0,mm0 ; mm0=(00 00 01 01) michael@0: punpckhwd mm5,mm1 ; mm5=(** 22 ** 23) michael@0: punpcklwd mm1,mm1 ; mm1=(20 20 21 21) michael@0: michael@0: psrad mm4,(DWORD_BIT-WORD_BIT) ; mm4=in0H=(02 03) michael@0: psrad mm0,(DWORD_BIT-WORD_BIT) ; mm0=in0L=(00 01) michael@0: cvtpi2ps xmm4,mm4 ; xmm4=(02 03 ** **) michael@0: cvtpi2ps xmm0,mm0 ; xmm0=(00 01 ** **) michael@0: psrad mm5,(DWORD_BIT-WORD_BIT) ; mm5=in2H=(22 23) michael@0: psrad mm1,(DWORD_BIT-WORD_BIT) ; mm1=in2L=(20 21) michael@0: cvtpi2ps xmm5,mm5 ; xmm5=(22 23 ** **) michael@0: cvtpi2ps xmm1,mm1 ; xmm1=(20 21 ** **) michael@0: michael@0: punpckhwd mm6,mm2 ; mm6=(** 42 ** 43) michael@0: punpcklwd mm2,mm2 ; mm2=(40 40 41 41) michael@0: punpckhwd mm7,mm3 ; mm7=(** 62 ** 63) michael@0: punpcklwd mm3,mm3 ; mm3=(60 60 61 61) michael@0: michael@0: psrad mm6,(DWORD_BIT-WORD_BIT) ; mm6=in4H=(42 43) michael@0: psrad mm2,(DWORD_BIT-WORD_BIT) ; mm2=in4L=(40 41) michael@0: cvtpi2ps xmm6,mm6 ; xmm6=(42 43 ** **) michael@0: cvtpi2ps xmm2,mm2 ; xmm2=(40 41 ** **) michael@0: psrad mm7,(DWORD_BIT-WORD_BIT) ; mm7=in6H=(62 63) michael@0: psrad mm3,(DWORD_BIT-WORD_BIT) ; mm3=in6L=(60 61) michael@0: cvtpi2ps xmm7,mm7 ; xmm7=(62 63 ** **) michael@0: cvtpi2ps xmm3,mm3 ; xmm3=(60 61 ** **) michael@0: michael@0: movlhps xmm0,xmm4 ; xmm0=in0=(00 01 02 03) michael@0: movlhps xmm1,xmm5 ; xmm1=in2=(20 21 22 23) michael@0: mulps xmm0, XMMWORD [XMMBLOCK(0,0,edx,SIZEOF_FLOAT_MULT_TYPE)] michael@0: mulps xmm1, XMMWORD [XMMBLOCK(2,0,edx,SIZEOF_FLOAT_MULT_TYPE)] michael@0: michael@0: movlhps xmm2,xmm6 ; xmm2=in4=(40 41 42 43) michael@0: movlhps xmm3,xmm7 ; xmm3=in6=(60 61 62 63) michael@0: mulps xmm2, XMMWORD [XMMBLOCK(4,0,edx,SIZEOF_FLOAT_MULT_TYPE)] michael@0: mulps xmm3, XMMWORD [XMMBLOCK(6,0,edx,SIZEOF_FLOAT_MULT_TYPE)] michael@0: michael@0: movaps xmm4,xmm0 michael@0: movaps xmm5,xmm1 michael@0: subps xmm0,xmm2 ; xmm0=tmp11 michael@0: subps xmm1,xmm3 michael@0: addps xmm4,xmm2 ; xmm4=tmp10 michael@0: addps xmm5,xmm3 ; xmm5=tmp13 michael@0: michael@0: mulps xmm1,[GOTOFF(ebx,PD_1_414)] michael@0: subps xmm1,xmm5 ; xmm1=tmp12 michael@0: michael@0: movaps xmm6,xmm4 michael@0: movaps xmm7,xmm0 michael@0: subps xmm4,xmm5 ; xmm4=tmp3 michael@0: subps xmm0,xmm1 ; xmm0=tmp2 michael@0: addps xmm6,xmm5 ; xmm6=tmp0 michael@0: addps xmm7,xmm1 ; xmm7=tmp1 michael@0: michael@0: movaps XMMWORD [wk(1)], xmm4 ; tmp3 michael@0: movaps XMMWORD [wk(0)], xmm0 ; tmp2 michael@0: michael@0: ; -- Odd part michael@0: michael@0: movq mm4, MMWORD [MMBLOCK(1,0,esi,SIZEOF_JCOEF)] michael@0: movq mm0, MMWORD [MMBLOCK(3,0,esi,SIZEOF_JCOEF)] michael@0: movq mm5, MMWORD [MMBLOCK(5,0,esi,SIZEOF_JCOEF)] michael@0: movq mm1, MMWORD [MMBLOCK(7,0,esi,SIZEOF_JCOEF)] michael@0: michael@0: punpckhwd mm6,mm4 ; mm6=(** 12 ** 13) michael@0: punpcklwd mm4,mm4 ; mm4=(10 10 11 11) michael@0: punpckhwd mm2,mm0 ; mm2=(** 32 ** 33) michael@0: punpcklwd mm0,mm0 ; mm0=(30 30 31 31) michael@0: michael@0: psrad mm6,(DWORD_BIT-WORD_BIT) ; mm6=in1H=(12 13) michael@0: psrad mm4,(DWORD_BIT-WORD_BIT) ; mm4=in1L=(10 11) michael@0: cvtpi2ps xmm4,mm6 ; xmm4=(12 13 ** **) michael@0: cvtpi2ps xmm2,mm4 ; xmm2=(10 11 ** **) michael@0: psrad mm2,(DWORD_BIT-WORD_BIT) ; mm2=in3H=(32 33) michael@0: psrad mm0,(DWORD_BIT-WORD_BIT) ; mm0=in3L=(30 31) michael@0: cvtpi2ps xmm0,mm2 ; xmm0=(32 33 ** **) michael@0: cvtpi2ps xmm3,mm0 ; xmm3=(30 31 ** **) michael@0: michael@0: punpckhwd mm7,mm5 ; mm7=(** 52 ** 53) michael@0: punpcklwd mm5,mm5 ; mm5=(50 50 51 51) michael@0: punpckhwd mm3,mm1 ; mm3=(** 72 ** 73) michael@0: punpcklwd mm1,mm1 ; mm1=(70 70 71 71) michael@0: michael@0: movlhps xmm2,xmm4 ; xmm2=in1=(10 11 12 13) michael@0: movlhps xmm3,xmm0 ; xmm3=in3=(30 31 32 33) michael@0: michael@0: psrad mm7,(DWORD_BIT-WORD_BIT) ; mm7=in5H=(52 53) michael@0: psrad mm5,(DWORD_BIT-WORD_BIT) ; mm5=in5L=(50 51) michael@0: cvtpi2ps xmm4,mm7 ; xmm4=(52 53 ** **) michael@0: cvtpi2ps xmm5,mm5 ; xmm5=(50 51 ** **) michael@0: psrad mm3,(DWORD_BIT-WORD_BIT) ; mm3=in7H=(72 73) michael@0: psrad mm1,(DWORD_BIT-WORD_BIT) ; mm1=in7L=(70 71) michael@0: cvtpi2ps xmm0,mm3 ; xmm0=(72 73 ** **) michael@0: cvtpi2ps xmm1,mm1 ; xmm1=(70 71 ** **) michael@0: michael@0: mulps xmm2, XMMWORD [XMMBLOCK(1,0,edx,SIZEOF_FLOAT_MULT_TYPE)] michael@0: mulps xmm3, XMMWORD [XMMBLOCK(3,0,edx,SIZEOF_FLOAT_MULT_TYPE)] michael@0: michael@0: movlhps xmm5,xmm4 ; xmm5=in5=(50 51 52 53) michael@0: movlhps xmm1,xmm0 ; xmm1=in7=(70 71 72 73) michael@0: mulps xmm5, XMMWORD [XMMBLOCK(5,0,edx,SIZEOF_FLOAT_MULT_TYPE)] michael@0: mulps xmm1, XMMWORD [XMMBLOCK(7,0,edx,SIZEOF_FLOAT_MULT_TYPE)] michael@0: michael@0: movaps xmm4,xmm2 michael@0: movaps xmm0,xmm5 michael@0: addps xmm2,xmm1 ; xmm2=z11 michael@0: addps xmm5,xmm3 ; xmm5=z13 michael@0: subps xmm4,xmm1 ; xmm4=z12 michael@0: subps xmm0,xmm3 ; xmm0=z10 michael@0: michael@0: movaps xmm1,xmm2 michael@0: subps xmm2,xmm5 michael@0: addps xmm1,xmm5 ; xmm1=tmp7 michael@0: michael@0: mulps xmm2,[GOTOFF(ebx,PD_1_414)] ; xmm2=tmp11 michael@0: michael@0: movaps xmm3,xmm0 michael@0: addps xmm0,xmm4 michael@0: mulps xmm0,[GOTOFF(ebx,PD_1_847)] ; xmm0=z5 michael@0: mulps xmm3,[GOTOFF(ebx,PD_M2_613)] ; xmm3=(z10 * -2.613125930) michael@0: mulps xmm4,[GOTOFF(ebx,PD_1_082)] ; xmm4=(z12 * 1.082392200) michael@0: addps xmm3,xmm0 ; xmm3=tmp12 michael@0: subps xmm4,xmm0 ; xmm4=tmp10 michael@0: michael@0: ; -- Final output stage michael@0: michael@0: subps xmm3,xmm1 ; xmm3=tmp6 michael@0: movaps xmm5,xmm6 michael@0: movaps xmm0,xmm7 michael@0: addps xmm6,xmm1 ; xmm6=data0=(00 01 02 03) michael@0: addps xmm7,xmm3 ; xmm7=data1=(10 11 12 13) michael@0: subps xmm5,xmm1 ; xmm5=data7=(70 71 72 73) michael@0: subps xmm0,xmm3 ; xmm0=data6=(60 61 62 63) michael@0: subps xmm2,xmm3 ; xmm2=tmp5 michael@0: michael@0: movaps xmm1,xmm6 ; transpose coefficients(phase 1) michael@0: unpcklps xmm6,xmm7 ; xmm6=(00 10 01 11) michael@0: unpckhps xmm1,xmm7 ; xmm1=(02 12 03 13) michael@0: movaps xmm3,xmm0 ; transpose coefficients(phase 1) michael@0: unpcklps xmm0,xmm5 ; xmm0=(60 70 61 71) michael@0: unpckhps xmm3,xmm5 ; xmm3=(62 72 63 73) michael@0: michael@0: movaps xmm7, XMMWORD [wk(0)] ; xmm7=tmp2 michael@0: movaps xmm5, XMMWORD [wk(1)] ; xmm5=tmp3 michael@0: michael@0: movaps XMMWORD [wk(0)], xmm0 ; wk(0)=(60 70 61 71) michael@0: movaps XMMWORD [wk(1)], xmm3 ; wk(1)=(62 72 63 73) michael@0: michael@0: addps xmm4,xmm2 ; xmm4=tmp4 michael@0: movaps xmm0,xmm7 michael@0: movaps xmm3,xmm5 michael@0: addps xmm7,xmm2 ; xmm7=data2=(20 21 22 23) michael@0: addps xmm5,xmm4 ; xmm5=data4=(40 41 42 43) michael@0: subps xmm0,xmm2 ; xmm0=data5=(50 51 52 53) michael@0: subps xmm3,xmm4 ; xmm3=data3=(30 31 32 33) michael@0: michael@0: movaps xmm2,xmm7 ; transpose coefficients(phase 1) michael@0: unpcklps xmm7,xmm3 ; xmm7=(20 30 21 31) michael@0: unpckhps xmm2,xmm3 ; xmm2=(22 32 23 33) michael@0: movaps xmm4,xmm5 ; transpose coefficients(phase 1) michael@0: unpcklps xmm5,xmm0 ; xmm5=(40 50 41 51) michael@0: unpckhps xmm4,xmm0 ; xmm4=(42 52 43 53) michael@0: michael@0: movaps xmm3,xmm6 ; transpose coefficients(phase 2) michael@0: unpcklps2 xmm6,xmm7 ; xmm6=(00 10 20 30) michael@0: unpckhps2 xmm3,xmm7 ; xmm3=(01 11 21 31) michael@0: movaps xmm0,xmm1 ; transpose coefficients(phase 2) michael@0: unpcklps2 xmm1,xmm2 ; xmm1=(02 12 22 32) michael@0: unpckhps2 xmm0,xmm2 ; xmm0=(03 13 23 33) michael@0: michael@0: movaps xmm7, XMMWORD [wk(0)] ; xmm7=(60 70 61 71) michael@0: movaps xmm2, XMMWORD [wk(1)] ; xmm2=(62 72 63 73) michael@0: michael@0: movaps XMMWORD [XMMBLOCK(0,0,edi,SIZEOF_FAST_FLOAT)], xmm6 michael@0: movaps XMMWORD [XMMBLOCK(1,0,edi,SIZEOF_FAST_FLOAT)], xmm3 michael@0: movaps XMMWORD [XMMBLOCK(2,0,edi,SIZEOF_FAST_FLOAT)], xmm1 michael@0: movaps XMMWORD [XMMBLOCK(3,0,edi,SIZEOF_FAST_FLOAT)], xmm0 michael@0: michael@0: movaps xmm6,xmm5 ; transpose coefficients(phase 2) michael@0: unpcklps2 xmm5,xmm7 ; xmm5=(40 50 60 70) michael@0: unpckhps2 xmm6,xmm7 ; xmm6=(41 51 61 71) michael@0: movaps xmm3,xmm4 ; transpose coefficients(phase 2) michael@0: unpcklps2 xmm4,xmm2 ; xmm4=(42 52 62 72) michael@0: unpckhps2 xmm3,xmm2 ; xmm3=(43 53 63 73) michael@0: michael@0: movaps XMMWORD [XMMBLOCK(0,1,edi,SIZEOF_FAST_FLOAT)], xmm5 michael@0: movaps XMMWORD [XMMBLOCK(1,1,edi,SIZEOF_FAST_FLOAT)], xmm6 michael@0: movaps XMMWORD [XMMBLOCK(2,1,edi,SIZEOF_FAST_FLOAT)], xmm4 michael@0: movaps XMMWORD [XMMBLOCK(3,1,edi,SIZEOF_FAST_FLOAT)], xmm3 michael@0: michael@0: .nextcolumn: michael@0: add esi, byte 4*SIZEOF_JCOEF ; coef_block michael@0: add edx, byte 4*SIZEOF_FLOAT_MULT_TYPE ; quantptr michael@0: add edi, 4*DCTSIZE*SIZEOF_FAST_FLOAT ; wsptr michael@0: dec ecx ; ctr michael@0: jnz near .columnloop michael@0: michael@0: ; -- Prefetch the next coefficient block michael@0: michael@0: prefetchnta [esi + (DCTSIZE2-8)*SIZEOF_JCOEF + 0*32] michael@0: prefetchnta [esi + (DCTSIZE2-8)*SIZEOF_JCOEF + 1*32] michael@0: prefetchnta [esi + (DCTSIZE2-8)*SIZEOF_JCOEF + 2*32] michael@0: prefetchnta [esi + (DCTSIZE2-8)*SIZEOF_JCOEF + 3*32] 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] ; FAST_FLOAT * 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: movaps xmm0, XMMWORD [XMMBLOCK(0,0,esi,SIZEOF_FAST_FLOAT)] michael@0: movaps xmm1, XMMWORD [XMMBLOCK(2,0,esi,SIZEOF_FAST_FLOAT)] michael@0: movaps xmm2, XMMWORD [XMMBLOCK(4,0,esi,SIZEOF_FAST_FLOAT)] michael@0: movaps xmm3, XMMWORD [XMMBLOCK(6,0,esi,SIZEOF_FAST_FLOAT)] michael@0: michael@0: movaps xmm4,xmm0 michael@0: movaps xmm5,xmm1 michael@0: subps xmm0,xmm2 ; xmm0=tmp11 michael@0: subps xmm1,xmm3 michael@0: addps xmm4,xmm2 ; xmm4=tmp10 michael@0: addps xmm5,xmm3 ; xmm5=tmp13 michael@0: michael@0: mulps xmm1,[GOTOFF(ebx,PD_1_414)] michael@0: subps xmm1,xmm5 ; xmm1=tmp12 michael@0: michael@0: movaps xmm6,xmm4 michael@0: movaps xmm7,xmm0 michael@0: subps xmm4,xmm5 ; xmm4=tmp3 michael@0: subps xmm0,xmm1 ; xmm0=tmp2 michael@0: addps xmm6,xmm5 ; xmm6=tmp0 michael@0: addps xmm7,xmm1 ; xmm7=tmp1 michael@0: michael@0: movaps XMMWORD [wk(1)], xmm4 ; tmp3 michael@0: movaps XMMWORD [wk(0)], xmm0 ; tmp2 michael@0: michael@0: ; -- Odd part michael@0: michael@0: movaps xmm2, XMMWORD [XMMBLOCK(1,0,esi,SIZEOF_FAST_FLOAT)] michael@0: movaps xmm3, XMMWORD [XMMBLOCK(3,0,esi,SIZEOF_FAST_FLOAT)] michael@0: movaps xmm5, XMMWORD [XMMBLOCK(5,0,esi,SIZEOF_FAST_FLOAT)] michael@0: movaps xmm1, XMMWORD [XMMBLOCK(7,0,esi,SIZEOF_FAST_FLOAT)] michael@0: michael@0: movaps xmm4,xmm2 michael@0: movaps xmm0,xmm5 michael@0: addps xmm2,xmm1 ; xmm2=z11 michael@0: addps xmm5,xmm3 ; xmm5=z13 michael@0: subps xmm4,xmm1 ; xmm4=z12 michael@0: subps xmm0,xmm3 ; xmm0=z10 michael@0: michael@0: movaps xmm1,xmm2 michael@0: subps xmm2,xmm5 michael@0: addps xmm1,xmm5 ; xmm1=tmp7 michael@0: michael@0: mulps xmm2,[GOTOFF(ebx,PD_1_414)] ; xmm2=tmp11 michael@0: michael@0: movaps xmm3,xmm0 michael@0: addps xmm0,xmm4 michael@0: mulps xmm0,[GOTOFF(ebx,PD_1_847)] ; xmm0=z5 michael@0: mulps xmm3,[GOTOFF(ebx,PD_M2_613)] ; xmm3=(z10 * -2.613125930) michael@0: mulps xmm4,[GOTOFF(ebx,PD_1_082)] ; xmm4=(z12 * 1.082392200) michael@0: addps xmm3,xmm0 ; xmm3=tmp12 michael@0: subps xmm4,xmm0 ; xmm4=tmp10 michael@0: michael@0: ; -- Final output stage michael@0: michael@0: subps xmm3,xmm1 ; xmm3=tmp6 michael@0: movaps xmm5,xmm6 michael@0: movaps xmm0,xmm7 michael@0: addps xmm6,xmm1 ; xmm6=data0=(00 10 20 30) michael@0: addps xmm7,xmm3 ; xmm7=data1=(01 11 21 31) michael@0: subps xmm5,xmm1 ; xmm5=data7=(07 17 27 37) michael@0: subps xmm0,xmm3 ; xmm0=data6=(06 16 26 36) michael@0: subps xmm2,xmm3 ; xmm2=tmp5 michael@0: michael@0: movaps xmm1,[GOTOFF(ebx,PD_0_125)] ; xmm1=[PD_0_125] michael@0: michael@0: mulps xmm6,xmm1 ; descale(1/8) michael@0: mulps xmm7,xmm1 ; descale(1/8) michael@0: mulps xmm5,xmm1 ; descale(1/8) michael@0: mulps xmm0,xmm1 ; descale(1/8) michael@0: michael@0: movhlps xmm3,xmm6 michael@0: movhlps xmm1,xmm7 michael@0: cvtps2pi mm0,xmm6 ; round to int32, mm0=data0L=(00 10) michael@0: cvtps2pi mm1,xmm7 ; round to int32, mm1=data1L=(01 11) michael@0: cvtps2pi mm2,xmm3 ; round to int32, mm2=data0H=(20 30) michael@0: cvtps2pi mm3,xmm1 ; round to int32, mm3=data1H=(21 31) michael@0: packssdw mm0,mm2 ; mm0=data0=(00 10 20 30) michael@0: packssdw mm1,mm3 ; mm1=data1=(01 11 21 31) michael@0: michael@0: movhlps xmm6,xmm5 michael@0: movhlps xmm7,xmm0 michael@0: cvtps2pi mm4,xmm5 ; round to int32, mm4=data7L=(07 17) michael@0: cvtps2pi mm5,xmm0 ; round to int32, mm5=data6L=(06 16) michael@0: cvtps2pi mm6,xmm6 ; round to int32, mm6=data7H=(27 37) michael@0: cvtps2pi mm7,xmm7 ; round to int32, mm7=data6H=(26 36) michael@0: packssdw mm4,mm6 ; mm4=data7=(07 17 27 37) michael@0: packssdw mm5,mm7 ; mm5=data6=(06 16 26 36) michael@0: michael@0: packsswb mm0,mm5 ; mm0=(00 10 20 30 06 16 26 36) michael@0: packsswb mm1,mm4 ; mm1=(01 11 21 31 07 17 27 37) michael@0: michael@0: movaps xmm3, XMMWORD [wk(0)] ; xmm3=tmp2 michael@0: movaps xmm1, XMMWORD [wk(1)] ; xmm1=tmp3 michael@0: michael@0: movaps xmm6,[GOTOFF(ebx,PD_0_125)] ; xmm6=[PD_0_125] michael@0: michael@0: addps xmm4,xmm2 ; xmm4=tmp4 michael@0: movaps xmm5,xmm3 michael@0: movaps xmm0,xmm1 michael@0: addps xmm3,xmm2 ; xmm3=data2=(02 12 22 32) michael@0: addps xmm1,xmm4 ; xmm1=data4=(04 14 24 34) michael@0: subps xmm5,xmm2 ; xmm5=data5=(05 15 25 35) michael@0: subps xmm0,xmm4 ; xmm0=data3=(03 13 23 33) michael@0: michael@0: mulps xmm3,xmm6 ; descale(1/8) michael@0: mulps xmm1,xmm6 ; descale(1/8) michael@0: mulps xmm5,xmm6 ; descale(1/8) michael@0: mulps xmm0,xmm6 ; descale(1/8) michael@0: michael@0: movhlps xmm7,xmm3 michael@0: movhlps xmm2,xmm1 michael@0: cvtps2pi mm2,xmm3 ; round to int32, mm2=data2L=(02 12) michael@0: cvtps2pi mm3,xmm1 ; round to int32, mm3=data4L=(04 14) michael@0: cvtps2pi mm6,xmm7 ; round to int32, mm6=data2H=(22 32) michael@0: cvtps2pi mm7,xmm2 ; round to int32, mm7=data4H=(24 34) michael@0: packssdw mm2,mm6 ; mm2=data2=(02 12 22 32) michael@0: packssdw mm3,mm7 ; mm3=data4=(04 14 24 34) michael@0: michael@0: movhlps xmm4,xmm5 michael@0: movhlps xmm6,xmm0 michael@0: cvtps2pi mm5,xmm5 ; round to int32, mm5=data5L=(05 15) michael@0: cvtps2pi mm4,xmm0 ; round to int32, mm4=data3L=(03 13) michael@0: cvtps2pi mm6,xmm4 ; round to int32, mm6=data5H=(25 35) michael@0: cvtps2pi mm7,xmm6 ; round to int32, mm7=data3H=(23 33) michael@0: packssdw mm5,mm6 ; mm5=data5=(05 15 25 35) michael@0: packssdw mm4,mm7 ; mm4=data3=(03 13 23 33) michael@0: michael@0: movq mm6,[GOTOFF(ebx,PB_CENTERJSAMP)] ; mm6=[PB_CENTERJSAMP] michael@0: michael@0: packsswb mm2,mm3 ; mm2=(02 12 22 32 04 14 24 34) michael@0: packsswb mm4,mm5 ; mm4=(03 13 23 33 05 15 25 35) michael@0: michael@0: paddb mm0,mm6 michael@0: paddb mm1,mm6 michael@0: paddb mm2,mm6 michael@0: paddb mm4,mm6 michael@0: michael@0: movq mm7,mm0 ; transpose coefficients(phase 1) michael@0: punpcklbw mm0,mm1 ; mm0=(00 01 10 11 20 21 30 31) michael@0: punpckhbw mm7,mm1 ; mm7=(06 07 16 17 26 27 36 37) michael@0: movq mm3,mm2 ; transpose coefficients(phase 1) michael@0: punpcklbw mm2,mm4 ; mm2=(02 03 12 13 22 23 32 33) michael@0: punpckhbw mm3,mm4 ; mm3=(04 05 14 15 24 25 34 35) michael@0: michael@0: movq mm5,mm0 ; transpose coefficients(phase 2) michael@0: punpcklwd mm0,mm2 ; mm0=(00 01 02 03 10 11 12 13) michael@0: punpckhwd mm5,mm2 ; mm5=(20 21 22 23 30 31 32 33) michael@0: movq mm6,mm3 ; transpose coefficients(phase 2) michael@0: punpcklwd mm3,mm7 ; mm3=(04 05 06 07 14 15 16 17) michael@0: punpckhwd mm6,mm7 ; mm6=(24 25 26 27 34 35 36 37) michael@0: michael@0: movq mm1,mm0 ; transpose coefficients(phase 3) michael@0: punpckldq mm0,mm3 ; mm0=(00 01 02 03 04 05 06 07) michael@0: punpckhdq mm1,mm3 ; mm1=(10 11 12 13 14 15 16 17) michael@0: movq mm4,mm5 ; transpose coefficients(phase 3) michael@0: punpckldq mm5,mm6 ; mm5=(20 21 22 23 24 25 26 27) michael@0: punpckhdq mm4,mm6 ; 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], mm1 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], mm5 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_FAST_FLOAT ; 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