1; 2; jquant.asm - sample data conversion and quantization (3DNow! & MMX) 3; 4; Copyright 2009 Pierre Ossman <ossman@cendio.se> for Cendio AB 5; 6; Based on the x86 SIMD extension for IJG JPEG library 7; Copyright (C) 1999-2006, MIYASAKA Masaru. 8; For conditions of distribution and use, see copyright notice in jsimdext.inc 9; 10; This file should be assembled with NASM (Netwide Assembler), 11; can *not* be assembled with Microsoft's MASM or any compatible 12; assembler (including Borland's Turbo Assembler). 13; NASM is available from http://nasm.sourceforge.net/ or 14; http://sourceforge.net/project/showfiles.php?group_id=6208 15; 16; [TAB8] 17 18%include "jsimdext.inc" 19%include "jdct.inc" 20 21; -------------------------------------------------------------------------- 22 SECTION SEG_TEXT 23 BITS 32 24; 25; Load data into workspace, applying unsigned->signed conversion 26; 27; GLOBAL(void) 28; jsimd_convsamp_float_3dnow (JSAMPARRAY sample_data, JDIMENSION start_col, 29; FAST_FLOAT *workspace); 30; 31 32%define sample_data ebp+8 ; JSAMPARRAY sample_data 33%define start_col ebp+12 ; JDIMENSION start_col 34%define workspace ebp+16 ; FAST_FLOAT *workspace 35 36 align 16 37 global EXTN(jsimd_convsamp_float_3dnow) 38 39EXTN(jsimd_convsamp_float_3dnow): 40 push ebp 41 mov ebp,esp 42 push ebx 43; push ecx ; need not be preserved 44; push edx ; need not be preserved 45 push esi 46 push edi 47 48 pcmpeqw mm7,mm7 49 psllw mm7,7 50 packsswb mm7,mm7 ; mm7 = PB_CENTERJSAMPLE (0x808080..) 51 52 mov esi, JSAMPARRAY [sample_data] ; (JSAMPROW *) 53 mov eax, JDIMENSION [start_col] 54 mov edi, POINTER [workspace] ; (DCTELEM *) 55 mov ecx, DCTSIZE/2 56 alignx 16,7 57.convloop: 58 mov ebx, JSAMPROW [esi+0*SIZEOF_JSAMPROW] ; (JSAMPLE *) 59 mov edx, JSAMPROW [esi+1*SIZEOF_JSAMPROW] ; (JSAMPLE *) 60 61 movq mm0, MMWORD [ebx+eax*SIZEOF_JSAMPLE] 62 movq mm1, MMWORD [edx+eax*SIZEOF_JSAMPLE] 63 64 psubb mm0,mm7 ; mm0=(01234567) 65 psubb mm1,mm7 ; mm1=(89ABCDEF) 66 67 punpcklbw mm2,mm0 ; mm2=(*0*1*2*3) 68 punpckhbw mm0,mm0 ; mm0=(*4*5*6*7) 69 punpcklbw mm3,mm1 ; mm3=(*8*9*A*B) 70 punpckhbw mm1,mm1 ; mm1=(*C*D*E*F) 71 72 punpcklwd mm4,mm2 ; mm4=(***0***1) 73 punpckhwd mm2,mm2 ; mm2=(***2***3) 74 punpcklwd mm5,mm0 ; mm5=(***4***5) 75 punpckhwd mm0,mm0 ; mm0=(***6***7) 76 77 psrad mm4,(DWORD_BIT-BYTE_BIT) ; mm4=(01) 78 psrad mm2,(DWORD_BIT-BYTE_BIT) ; mm2=(23) 79 pi2fd mm4,mm4 80 pi2fd mm2,mm2 81 psrad mm5,(DWORD_BIT-BYTE_BIT) ; mm5=(45) 82 psrad mm0,(DWORD_BIT-BYTE_BIT) ; mm0=(67) 83 pi2fd mm5,mm5 84 pi2fd mm0,mm0 85 86 movq MMWORD [MMBLOCK(0,0,edi,SIZEOF_FAST_FLOAT)], mm4 87 movq MMWORD [MMBLOCK(0,1,edi,SIZEOF_FAST_FLOAT)], mm2 88 movq MMWORD [MMBLOCK(0,2,edi,SIZEOF_FAST_FLOAT)], mm5 89 movq MMWORD [MMBLOCK(0,3,edi,SIZEOF_FAST_FLOAT)], mm0 90 91 punpcklwd mm6,mm3 ; mm6=(***8***9) 92 punpckhwd mm3,mm3 ; mm3=(***A***B) 93 punpcklwd mm4,mm1 ; mm4=(***C***D) 94 punpckhwd mm1,mm1 ; mm1=(***E***F) 95 96 psrad mm6,(DWORD_BIT-BYTE_BIT) ; mm6=(89) 97 psrad mm3,(DWORD_BIT-BYTE_BIT) ; mm3=(AB) 98 pi2fd mm6,mm6 99 pi2fd mm3,mm3 100 psrad mm4,(DWORD_BIT-BYTE_BIT) ; mm4=(CD) 101 psrad mm1,(DWORD_BIT-BYTE_BIT) ; mm1=(EF) 102 pi2fd mm4,mm4 103 pi2fd mm1,mm1 104 105 movq MMWORD [MMBLOCK(1,0,edi,SIZEOF_FAST_FLOAT)], mm6 106 movq MMWORD [MMBLOCK(1,1,edi,SIZEOF_FAST_FLOAT)], mm3 107 movq MMWORD [MMBLOCK(1,2,edi,SIZEOF_FAST_FLOAT)], mm4 108 movq MMWORD [MMBLOCK(1,3,edi,SIZEOF_FAST_FLOAT)], mm1 109 110 add esi, byte 2*SIZEOF_JSAMPROW 111 add edi, byte 2*DCTSIZE*SIZEOF_FAST_FLOAT 112 dec ecx 113 jnz near .convloop 114 115 femms ; empty MMX/3DNow! state 116 117 pop edi 118 pop esi 119; pop edx ; need not be preserved 120; pop ecx ; need not be preserved 121 pop ebx 122 pop ebp 123 ret 124 125 126; -------------------------------------------------------------------------- 127; 128; Quantize/descale the coefficients, and store into coef_block 129; 130; GLOBAL(void) 131; jsimd_quantize_float_3dnow (JCOEFPTR coef_block, FAST_FLOAT *divisors, 132; FAST_FLOAT *workspace); 133; 134 135%define coef_block ebp+8 ; JCOEFPTR coef_block 136%define divisors ebp+12 ; FAST_FLOAT *divisors 137%define workspace ebp+16 ; FAST_FLOAT *workspace 138 139 align 16 140 global EXTN(jsimd_quantize_float_3dnow) 141 142EXTN(jsimd_quantize_float_3dnow): 143 push ebp 144 mov ebp,esp 145; push ebx ; unused 146; push ecx ; unused 147; push edx ; need not be preserved 148 push esi 149 push edi 150 151 mov eax, 0x4B400000 ; (float)0x00C00000 (rndint_magic) 152 movd mm7,eax 153 punpckldq mm7,mm7 ; mm7={12582912.0F 12582912.0F} 154 155 mov esi, POINTER [workspace] 156 mov edx, POINTER [divisors] 157 mov edi, JCOEFPTR [coef_block] 158 mov eax, DCTSIZE2/16 159 alignx 16,7 160.quantloop: 161 movq mm0, MMWORD [MMBLOCK(0,0,esi,SIZEOF_FAST_FLOAT)] 162 movq mm1, MMWORD [MMBLOCK(0,1,esi,SIZEOF_FAST_FLOAT)] 163 pfmul mm0, MMWORD [MMBLOCK(0,0,edx,SIZEOF_FAST_FLOAT)] 164 pfmul mm1, MMWORD [MMBLOCK(0,1,edx,SIZEOF_FAST_FLOAT)] 165 movq mm2, MMWORD [MMBLOCK(0,2,esi,SIZEOF_FAST_FLOAT)] 166 movq mm3, MMWORD [MMBLOCK(0,3,esi,SIZEOF_FAST_FLOAT)] 167 pfmul mm2, MMWORD [MMBLOCK(0,2,edx,SIZEOF_FAST_FLOAT)] 168 pfmul mm3, MMWORD [MMBLOCK(0,3,edx,SIZEOF_FAST_FLOAT)] 169 170 pfadd mm0,mm7 ; mm0=(00 ** 01 **) 171 pfadd mm1,mm7 ; mm1=(02 ** 03 **) 172 pfadd mm2,mm7 ; mm0=(04 ** 05 **) 173 pfadd mm3,mm7 ; mm1=(06 ** 07 **) 174 175 movq mm4,mm0 176 punpcklwd mm0,mm1 ; mm0=(00 02 ** **) 177 punpckhwd mm4,mm1 ; mm4=(01 03 ** **) 178 movq mm5,mm2 179 punpcklwd mm2,mm3 ; mm2=(04 06 ** **) 180 punpckhwd mm5,mm3 ; mm5=(05 07 ** **) 181 182 punpcklwd mm0,mm4 ; mm0=(00 01 02 03) 183 punpcklwd mm2,mm5 ; mm2=(04 05 06 07) 184 185 movq mm6, MMWORD [MMBLOCK(1,0,esi,SIZEOF_FAST_FLOAT)] 186 movq mm1, MMWORD [MMBLOCK(1,1,esi,SIZEOF_FAST_FLOAT)] 187 pfmul mm6, MMWORD [MMBLOCK(1,0,edx,SIZEOF_FAST_FLOAT)] 188 pfmul mm1, MMWORD [MMBLOCK(1,1,edx,SIZEOF_FAST_FLOAT)] 189 movq mm3, MMWORD [MMBLOCK(1,2,esi,SIZEOF_FAST_FLOAT)] 190 movq mm4, MMWORD [MMBLOCK(1,3,esi,SIZEOF_FAST_FLOAT)] 191 pfmul mm3, MMWORD [MMBLOCK(1,2,edx,SIZEOF_FAST_FLOAT)] 192 pfmul mm4, MMWORD [MMBLOCK(1,3,edx,SIZEOF_FAST_FLOAT)] 193 194 pfadd mm6,mm7 ; mm0=(10 ** 11 **) 195 pfadd mm1,mm7 ; mm4=(12 ** 13 **) 196 pfadd mm3,mm7 ; mm0=(14 ** 15 **) 197 pfadd mm4,mm7 ; mm4=(16 ** 17 **) 198 199 movq mm5,mm6 200 punpcklwd mm6,mm1 ; mm6=(10 12 ** **) 201 punpckhwd mm5,mm1 ; mm5=(11 13 ** **) 202 movq mm1,mm3 203 punpcklwd mm3,mm4 ; mm3=(14 16 ** **) 204 punpckhwd mm1,mm4 ; mm1=(15 17 ** **) 205 206 punpcklwd mm6,mm5 ; mm6=(10 11 12 13) 207 punpcklwd mm3,mm1 ; mm3=(14 15 16 17) 208 209 movq MMWORD [MMBLOCK(0,0,edi,SIZEOF_JCOEF)], mm0 210 movq MMWORD [MMBLOCK(0,1,edi,SIZEOF_JCOEF)], mm2 211 movq MMWORD [MMBLOCK(1,0,edi,SIZEOF_JCOEF)], mm6 212 movq MMWORD [MMBLOCK(1,1,edi,SIZEOF_JCOEF)], mm3 213 214 add esi, byte 16*SIZEOF_FAST_FLOAT 215 add edx, byte 16*SIZEOF_FAST_FLOAT 216 add edi, byte 16*SIZEOF_JCOEF 217 dec eax 218 jnz near .quantloop 219 220 femms ; empty MMX/3DNow! state 221 222 pop edi 223 pop esi 224; pop edx ; need not be preserved 225; pop ecx ; unused 226; pop ebx ; unused 227 pop ebp 228 ret 229 230; For some reason, the OS X linker does not honor the request to align the 231; segment unless we do this. 232 align 16 233