MIYASAKA Masaru | a2e6a9d | 2006-02-04 00:00:00 +0000 | [diff] [blame^] | 1 | ; |
| 2 | ; jcqntsse.asm - sample data conversion and quantization (SSE & MMX) |
| 3 | ; |
| 4 | ; x86 SIMD extension for IJG JPEG library |
| 5 | ; Copyright (C) 1999-2006, MIYASAKA Masaru. |
| 6 | ; For conditions of distribution and use, see copyright notice in jsimdext.inc |
| 7 | ; |
| 8 | ; This file should be assembled with NASM (Netwide Assembler), |
| 9 | ; can *not* be assembled with Microsoft's MASM or any compatible |
| 10 | ; assembler (including Borland's Turbo Assembler). |
| 11 | ; NASM is available from http://nasm.sourceforge.net/ or |
| 12 | ; http://sourceforge.net/project/showfiles.php?group_id=6208 |
| 13 | ; |
| 14 | ; Last Modified : January 12, 2005 |
| 15 | ; |
| 16 | ; [TAB8] |
| 17 | |
| 18 | %include "jsimdext.inc" |
| 19 | %include "jdct.inc" |
| 20 | |
| 21 | %ifdef DCT_FLOAT_SUPPORTED |
| 22 | %ifdef JFDCT_FLT_SSE_MMX_SUPPORTED |
| 23 | |
| 24 | ; This module is specialized to the case DCTSIZE = 8. |
| 25 | ; |
| 26 | %if DCTSIZE != 8 |
| 27 | %error "Sorry, this code only copes with 8x8 DCTs." |
| 28 | %endif |
| 29 | |
| 30 | ; -------------------------------------------------------------------------- |
| 31 | SECTION SEG_TEXT |
| 32 | BITS 32 |
| 33 | ; |
| 34 | ; Load data into workspace, applying unsigned->signed conversion |
| 35 | ; |
| 36 | ; GLOBAL(void) |
| 37 | ; jpeg_convsamp_flt_sse (JSAMPARRAY sample_data, JDIMENSION start_col, |
| 38 | ; FAST_FLOAT * workspace); |
| 39 | ; |
| 40 | |
| 41 | %define sample_data ebp+8 ; JSAMPARRAY sample_data |
| 42 | %define start_col ebp+12 ; JDIMENSION start_col |
| 43 | %define workspace ebp+16 ; FAST_FLOAT * workspace |
| 44 | |
| 45 | align 16 |
| 46 | global EXTN(jpeg_convsamp_flt_sse) |
| 47 | |
| 48 | EXTN(jpeg_convsamp_flt_sse): |
| 49 | push ebp |
| 50 | mov ebp,esp |
| 51 | push ebx |
| 52 | ; push ecx ; need not be preserved |
| 53 | ; push edx ; need not be preserved |
| 54 | push esi |
| 55 | push edi |
| 56 | |
| 57 | pcmpeqw mm7,mm7 |
| 58 | psllw mm7,7 |
| 59 | packsswb mm7,mm7 ; mm7 = PB_CENTERJSAMPLE (0x808080..) |
| 60 | |
| 61 | mov esi, JSAMPARRAY [sample_data] ; (JSAMPROW *) |
| 62 | mov eax, JDIMENSION [start_col] |
| 63 | mov edi, POINTER [workspace] ; (DCTELEM *) |
| 64 | mov ecx, DCTSIZE/2 |
| 65 | alignx 16,7 |
| 66 | .convloop: |
| 67 | mov ebx, JSAMPROW [esi+0*SIZEOF_JSAMPROW] ; (JSAMPLE *) |
| 68 | mov edx, JSAMPROW [esi+1*SIZEOF_JSAMPROW] ; (JSAMPLE *) |
| 69 | |
| 70 | movq mm0, MMWORD [ebx+eax*SIZEOF_JSAMPLE] |
| 71 | movq mm1, MMWORD [edx+eax*SIZEOF_JSAMPLE] |
| 72 | |
| 73 | psubb mm0,mm7 ; mm0=(01234567) |
| 74 | psubb mm1,mm7 ; mm1=(89ABCDEF) |
| 75 | |
| 76 | punpcklbw mm2,mm0 ; mm2=(*0*1*2*3) |
| 77 | punpckhbw mm0,mm0 ; mm0=(*4*5*6*7) |
| 78 | punpcklbw mm3,mm1 ; mm3=(*8*9*A*B) |
| 79 | punpckhbw mm1,mm1 ; mm1=(*C*D*E*F) |
| 80 | |
| 81 | punpcklwd mm4,mm2 ; mm4=(***0***1) |
| 82 | punpckhwd mm2,mm2 ; mm2=(***2***3) |
| 83 | punpcklwd mm5,mm0 ; mm5=(***4***5) |
| 84 | punpckhwd mm0,mm0 ; mm0=(***6***7) |
| 85 | |
| 86 | psrad mm4,(DWORD_BIT-BYTE_BIT) ; mm4=(01) |
| 87 | psrad mm2,(DWORD_BIT-BYTE_BIT) ; mm2=(23) |
| 88 | cvtpi2ps xmm0,mm4 ; xmm0=(01**) |
| 89 | cvtpi2ps xmm1,mm2 ; xmm1=(23**) |
| 90 | psrad mm5,(DWORD_BIT-BYTE_BIT) ; mm5=(45) |
| 91 | psrad mm0,(DWORD_BIT-BYTE_BIT) ; mm0=(67) |
| 92 | cvtpi2ps xmm2,mm5 ; xmm2=(45**) |
| 93 | cvtpi2ps xmm3,mm0 ; xmm3=(67**) |
| 94 | |
| 95 | punpcklwd mm6,mm3 ; mm6=(***8***9) |
| 96 | punpckhwd mm3,mm3 ; mm3=(***A***B) |
| 97 | punpcklwd mm4,mm1 ; mm4=(***C***D) |
| 98 | punpckhwd mm1,mm1 ; mm1=(***E***F) |
| 99 | |
| 100 | psrad mm6,(DWORD_BIT-BYTE_BIT) ; mm6=(89) |
| 101 | psrad mm3,(DWORD_BIT-BYTE_BIT) ; mm3=(AB) |
| 102 | cvtpi2ps xmm4,mm6 ; xmm4=(89**) |
| 103 | cvtpi2ps xmm5,mm3 ; xmm5=(AB**) |
| 104 | psrad mm4,(DWORD_BIT-BYTE_BIT) ; mm4=(CD) |
| 105 | psrad mm1,(DWORD_BIT-BYTE_BIT) ; mm1=(EF) |
| 106 | cvtpi2ps xmm6,mm4 ; xmm6=(CD**) |
| 107 | cvtpi2ps xmm7,mm1 ; xmm7=(EF**) |
| 108 | |
| 109 | movlhps xmm0,xmm1 ; xmm0=(0123) |
| 110 | movlhps xmm2,xmm3 ; xmm2=(4567) |
| 111 | movlhps xmm4,xmm5 ; xmm4=(89AB) |
| 112 | movlhps xmm6,xmm7 ; xmm6=(CDEF) |
| 113 | |
| 114 | movaps XMMWORD [XMMBLOCK(0,0,edi,SIZEOF_FAST_FLOAT)], xmm0 |
| 115 | movaps XMMWORD [XMMBLOCK(0,1,edi,SIZEOF_FAST_FLOAT)], xmm2 |
| 116 | movaps XMMWORD [XMMBLOCK(1,0,edi,SIZEOF_FAST_FLOAT)], xmm4 |
| 117 | movaps XMMWORD [XMMBLOCK(1,1,edi,SIZEOF_FAST_FLOAT)], xmm6 |
| 118 | |
| 119 | add esi, byte 2*SIZEOF_JSAMPROW |
| 120 | add edi, byte 2*DCTSIZE*SIZEOF_FAST_FLOAT |
| 121 | dec ecx |
| 122 | jnz near .convloop |
| 123 | |
| 124 | emms ; empty MMX state |
| 125 | |
| 126 | pop edi |
| 127 | pop esi |
| 128 | ; pop edx ; need not be preserved |
| 129 | ; pop ecx ; need not be preserved |
| 130 | pop ebx |
| 131 | pop ebp |
| 132 | ret |
| 133 | |
| 134 | |
| 135 | ; -------------------------------------------------------------------------- |
| 136 | ; |
| 137 | ; Quantize/descale the coefficients, and store into coef_block |
| 138 | ; |
| 139 | ; GLOBAL(void) |
| 140 | ; jpeg_quantize_flt_sse (JCOEFPTR coef_block, FAST_FLOAT * divisors, |
| 141 | ; FAST_FLOAT * workspace); |
| 142 | ; |
| 143 | |
| 144 | %define coef_block ebp+8 ; JCOEFPTR coef_block |
| 145 | %define divisors ebp+12 ; FAST_FLOAT * divisors |
| 146 | %define workspace ebp+16 ; FAST_FLOAT * workspace |
| 147 | |
| 148 | align 16 |
| 149 | global EXTN(jpeg_quantize_flt_sse) |
| 150 | |
| 151 | EXTN(jpeg_quantize_flt_sse): |
| 152 | push ebp |
| 153 | mov ebp,esp |
| 154 | ; push ebx ; unused |
| 155 | ; push ecx ; unused |
| 156 | ; push edx ; need not be preserved |
| 157 | push esi |
| 158 | push edi |
| 159 | |
| 160 | mov esi, POINTER [workspace] |
| 161 | mov edx, POINTER [divisors] |
| 162 | mov edi, JCOEFPTR [coef_block] |
| 163 | mov eax, DCTSIZE2/16 |
| 164 | alignx 16,7 |
| 165 | .quantloop: |
| 166 | movaps xmm0, XMMWORD [XMMBLOCK(0,0,esi,SIZEOF_FAST_FLOAT)] |
| 167 | movaps xmm1, XMMWORD [XMMBLOCK(0,1,esi,SIZEOF_FAST_FLOAT)] |
| 168 | mulps xmm0, XMMWORD [XMMBLOCK(0,0,edx,SIZEOF_FAST_FLOAT)] |
| 169 | mulps xmm1, XMMWORD [XMMBLOCK(0,1,edx,SIZEOF_FAST_FLOAT)] |
| 170 | movaps xmm2, XMMWORD [XMMBLOCK(1,0,esi,SIZEOF_FAST_FLOAT)] |
| 171 | movaps xmm3, XMMWORD [XMMBLOCK(1,1,esi,SIZEOF_FAST_FLOAT)] |
| 172 | mulps xmm2, XMMWORD [XMMBLOCK(1,0,edx,SIZEOF_FAST_FLOAT)] |
| 173 | mulps xmm3, XMMWORD [XMMBLOCK(1,1,edx,SIZEOF_FAST_FLOAT)] |
| 174 | |
| 175 | movhlps xmm4,xmm0 |
| 176 | movhlps xmm5,xmm1 |
| 177 | |
| 178 | cvtps2pi mm0,xmm0 |
| 179 | cvtps2pi mm1,xmm1 |
| 180 | cvtps2pi mm4,xmm4 |
| 181 | cvtps2pi mm5,xmm5 |
| 182 | |
| 183 | movhlps xmm6,xmm2 |
| 184 | movhlps xmm7,xmm3 |
| 185 | |
| 186 | cvtps2pi mm2,xmm2 |
| 187 | cvtps2pi mm3,xmm3 |
| 188 | cvtps2pi mm6,xmm6 |
| 189 | cvtps2pi mm7,xmm7 |
| 190 | |
| 191 | packssdw mm0,mm4 |
| 192 | packssdw mm1,mm5 |
| 193 | packssdw mm2,mm6 |
| 194 | packssdw mm3,mm7 |
| 195 | |
| 196 | movq MMWORD [MMBLOCK(0,0,edi,SIZEOF_JCOEF)], mm0 |
| 197 | movq MMWORD [MMBLOCK(0,1,edi,SIZEOF_JCOEF)], mm1 |
| 198 | movq MMWORD [MMBLOCK(1,0,edi,SIZEOF_JCOEF)], mm2 |
| 199 | movq MMWORD [MMBLOCK(1,1,edi,SIZEOF_JCOEF)], mm3 |
| 200 | |
| 201 | add esi, byte 16*SIZEOF_FAST_FLOAT |
| 202 | add edx, byte 16*SIZEOF_FAST_FLOAT |
| 203 | add edi, byte 16*SIZEOF_JCOEF |
| 204 | dec eax |
| 205 | jnz short .quantloop |
| 206 | |
| 207 | emms ; empty MMX state |
| 208 | |
| 209 | pop edi |
| 210 | pop esi |
| 211 | ; pop edx ; need not be preserved |
| 212 | ; pop ecx ; unused |
| 213 | ; pop ebx ; unused |
| 214 | pop ebp |
| 215 | ret |
| 216 | |
| 217 | %endif ; JFDCT_FLT_SSE_MMX_SUPPORTED |
| 218 | %endif ; DCT_FLOAT_SUPPORTED |