media/libjpeg/simd/jcqntmmx.asm@b8a032363ba2 (annotated)
media/libjpeg/simd/jcqntmmx.asm
Thu, 22 Jan 2015 13:21:57 +0100
- author
- Michael Schloh von Bennewitz <michael@schloh.com>
- date
- Thu, 22 Jan 2015 13:21:57 +0100
- branch
- TOR_BUG_9701
- changeset 15
- b8a032363ba2
- permissions
- -rw-r--r--
Incorporate requested changes from Mozilla in review:
https://bugzilla.mozilla.org/show_bug.cgi?id=1123480#c6
| michael@0 | 1 | ; |
| michael@0 | 2 | ; jcqntmmx.asm - sample data conversion and quantization (MMX) |
| michael@0 | 3 | ; |
| michael@0 | 4 | ; Copyright 2009 Pierre Ossman <ossman@cendio.se> for Cendio AB |
| michael@0 | 5 | ; |
| michael@0 | 6 | ; Based on |
| michael@0 | 7 | ; x86 SIMD extension for IJG JPEG library |
| michael@0 | 8 | ; Copyright (C) 1999-2006, MIYASAKA Masaru. |
| michael@0 | 9 | ; For conditions of distribution and use, see copyright notice in jsimdext.inc |
| michael@0 | 10 | ; |
| michael@0 | 11 | ; This file should be assembled with NASM (Netwide Assembler), |
| michael@0 | 12 | ; can *not* be assembled with Microsoft's MASM or any compatible |
| michael@0 | 13 | ; assembler (including Borland's Turbo Assembler). |
| michael@0 | 14 | ; NASM is available from http://nasm.sourceforge.net/ or |
| michael@0 | 15 | ; http://sourceforge.net/project/showfiles.php?group_id=6208 |
| michael@0 | 16 | ; |
| michael@0 | 17 | ; [TAB8] |
| michael@0 | 18 | |
| michael@0 | 19 | %include "jsimdext.inc" |
| michael@0 | 20 | %include "jdct.inc" |
| michael@0 | 21 | |
| michael@0 | 22 | ; -------------------------------------------------------------------------- |
| michael@0 | 23 | SECTION SEG_TEXT |
| michael@0 | 24 | BITS 32 |
| michael@0 | 25 | ; |
| michael@0 | 26 | ; Load data into workspace, applying unsigned->signed conversion |
| michael@0 | 27 | ; |
| michael@0 | 28 | ; GLOBAL(void) |
| michael@0 | 29 | ; jsimd_convsamp_mmx (JSAMPARRAY sample_data, JDIMENSION start_col, |
| michael@0 | 30 | ; DCTELEM * workspace); |
| michael@0 | 31 | ; |
| michael@0 | 32 | |
| michael@0 | 33 | %define sample_data ebp+8 ; JSAMPARRAY sample_data |
| michael@0 | 34 | %define start_col ebp+12 ; JDIMENSION start_col |
| michael@0 | 35 | %define workspace ebp+16 ; DCTELEM * workspace |
| michael@0 | 36 | |
| michael@0 | 37 | align 16 |
| michael@0 | 38 | global EXTN(jsimd_convsamp_mmx) |
| michael@0 | 39 | |
| michael@0 | 40 | EXTN(jsimd_convsamp_mmx): |
| michael@0 | 41 | push ebp |
| michael@0 | 42 | mov ebp,esp |
| michael@0 | 43 | push ebx |
| michael@0 | 44 | ; push ecx ; need not be preserved |
| michael@0 | 45 | ; push edx ; need not be preserved |
| michael@0 | 46 | push esi |
| michael@0 | 47 | push edi |
| michael@0 | 48 | |
| michael@0 | 49 | pxor mm6,mm6 ; mm6=(all 0's) |
| michael@0 | 50 | pcmpeqw mm7,mm7 |
| michael@0 | 51 | psllw mm7,7 ; mm7={0xFF80 0xFF80 0xFF80 0xFF80} |
| michael@0 | 52 | |
| michael@0 | 53 | mov esi, JSAMPARRAY [sample_data] ; (JSAMPROW *) |
| michael@0 | 54 | mov eax, JDIMENSION [start_col] |
| michael@0 | 55 | mov edi, POINTER [workspace] ; (DCTELEM *) |
| michael@0 | 56 | mov ecx, DCTSIZE/4 |
| michael@0 | 57 | alignx 16,7 |
| michael@0 | 58 | .convloop: |
| michael@0 | 59 | mov ebx, JSAMPROW [esi+0*SIZEOF_JSAMPROW] ; (JSAMPLE *) |
| michael@0 | 60 | mov edx, JSAMPROW [esi+1*SIZEOF_JSAMPROW] ; (JSAMPLE *) |
| michael@0 | 61 | |
| michael@0 | 62 | movq mm0, MMWORD [ebx+eax*SIZEOF_JSAMPLE] ; mm0=(01234567) |
| michael@0 | 63 | movq mm1, MMWORD [edx+eax*SIZEOF_JSAMPLE] ; mm1=(89ABCDEF) |
| michael@0 | 64 | |
| michael@0 | 65 | mov ebx, JSAMPROW [esi+2*SIZEOF_JSAMPROW] ; (JSAMPLE *) |
| michael@0 | 66 | mov edx, JSAMPROW [esi+3*SIZEOF_JSAMPROW] ; (JSAMPLE *) |
| michael@0 | 67 | |
| michael@0 | 68 | movq mm2, MMWORD [ebx+eax*SIZEOF_JSAMPLE] ; mm2=(GHIJKLMN) |
| michael@0 | 69 | movq mm3, MMWORD [edx+eax*SIZEOF_JSAMPLE] ; mm3=(OPQRSTUV) |
| michael@0 | 70 | |
| michael@0 | 71 | movq mm4,mm0 |
| michael@0 | 72 | punpcklbw mm0,mm6 ; mm0=(0123) |
| michael@0 | 73 | punpckhbw mm4,mm6 ; mm4=(4567) |
| michael@0 | 74 | movq mm5,mm1 |
| michael@0 | 75 | punpcklbw mm1,mm6 ; mm1=(89AB) |
| michael@0 | 76 | punpckhbw mm5,mm6 ; mm5=(CDEF) |
| michael@0 | 77 | |
| michael@0 | 78 | paddw mm0,mm7 |
| michael@0 | 79 | paddw mm4,mm7 |
| michael@0 | 80 | paddw mm1,mm7 |
| michael@0 | 81 | paddw mm5,mm7 |
| michael@0 | 82 | |
| michael@0 | 83 | movq MMWORD [MMBLOCK(0,0,edi,SIZEOF_DCTELEM)], mm0 |
| michael@0 | 84 | movq MMWORD [MMBLOCK(0,1,edi,SIZEOF_DCTELEM)], mm4 |
| michael@0 | 85 | movq MMWORD [MMBLOCK(1,0,edi,SIZEOF_DCTELEM)], mm1 |
| michael@0 | 86 | movq MMWORD [MMBLOCK(1,1,edi,SIZEOF_DCTELEM)], mm5 |
| michael@0 | 87 | |
| michael@0 | 88 | movq mm0,mm2 |
| michael@0 | 89 | punpcklbw mm2,mm6 ; mm2=(GHIJ) |
| michael@0 | 90 | punpckhbw mm0,mm6 ; mm0=(KLMN) |
| michael@0 | 91 | movq mm4,mm3 |
| michael@0 | 92 | punpcklbw mm3,mm6 ; mm3=(OPQR) |
| michael@0 | 93 | punpckhbw mm4,mm6 ; mm4=(STUV) |
| michael@0 | 94 | |
| michael@0 | 95 | paddw mm2,mm7 |
| michael@0 | 96 | paddw mm0,mm7 |
| michael@0 | 97 | paddw mm3,mm7 |
| michael@0 | 98 | paddw mm4,mm7 |
| michael@0 | 99 | |
| michael@0 | 100 | movq MMWORD [MMBLOCK(2,0,edi,SIZEOF_DCTELEM)], mm2 |
| michael@0 | 101 | movq MMWORD [MMBLOCK(2,1,edi,SIZEOF_DCTELEM)], mm0 |
| michael@0 | 102 | movq MMWORD [MMBLOCK(3,0,edi,SIZEOF_DCTELEM)], mm3 |
| michael@0 | 103 | movq MMWORD [MMBLOCK(3,1,edi,SIZEOF_DCTELEM)], mm4 |
| michael@0 | 104 | |
| michael@0 | 105 | add esi, byte 4*SIZEOF_JSAMPROW |
| michael@0 | 106 | add edi, byte 4*DCTSIZE*SIZEOF_DCTELEM |
| michael@0 | 107 | dec ecx |
| michael@0 | 108 | jnz short .convloop |
| michael@0 | 109 | |
| michael@0 | 110 | emms ; empty MMX state |
| michael@0 | 111 | |
| michael@0 | 112 | pop edi |
| michael@0 | 113 | pop esi |
| michael@0 | 114 | ; pop edx ; need not be preserved |
| michael@0 | 115 | ; pop ecx ; need not be preserved |
| michael@0 | 116 | pop ebx |
| michael@0 | 117 | pop ebp |
| michael@0 | 118 | ret |
| michael@0 | 119 | |
| michael@0 | 120 | ; -------------------------------------------------------------------------- |
| michael@0 | 121 | ; |
| michael@0 | 122 | ; Quantize/descale the coefficients, and store into coef_block |
| michael@0 | 123 | ; |
| michael@0 | 124 | ; This implementation is based on an algorithm described in |
| michael@0 | 125 | ; "How to optimize for the Pentium family of microprocessors" |
| michael@0 | 126 | ; (http://www.agner.org/assem/). |
| michael@0 | 127 | ; |
| michael@0 | 128 | ; GLOBAL(void) |
| michael@0 | 129 | ; jsimd_quantize_mmx (JCOEFPTR coef_block, DCTELEM * divisors, |
| michael@0 | 130 | ; DCTELEM * workspace); |
| michael@0 | 131 | ; |
| michael@0 | 132 | |
| michael@0 | 133 | %define RECIPROCAL(m,n,b) MMBLOCK(DCTSIZE*0+(m),(n),(b),SIZEOF_DCTELEM) |
| michael@0 | 134 | %define CORRECTION(m,n,b) MMBLOCK(DCTSIZE*1+(m),(n),(b),SIZEOF_DCTELEM) |
| michael@0 | 135 | %define SCALE(m,n,b) MMBLOCK(DCTSIZE*2+(m),(n),(b),SIZEOF_DCTELEM) |
| michael@0 | 136 | %define SHIFT(m,n,b) MMBLOCK(DCTSIZE*3+(m),(n),(b),SIZEOF_DCTELEM) |
| michael@0 | 137 | |
| michael@0 | 138 | %define coef_block ebp+8 ; JCOEFPTR coef_block |
| michael@0 | 139 | %define divisors ebp+12 ; DCTELEM * divisors |
| michael@0 | 140 | %define workspace ebp+16 ; DCTELEM * workspace |
| michael@0 | 141 | |
| michael@0 | 142 | align 16 |
| michael@0 | 143 | global EXTN(jsimd_quantize_mmx) |
| michael@0 | 144 | |
| michael@0 | 145 | EXTN(jsimd_quantize_mmx): |
| michael@0 | 146 | push ebp |
| michael@0 | 147 | mov ebp,esp |
| michael@0 | 148 | ; push ebx ; unused |
| michael@0 | 149 | ; push ecx ; unused |
| michael@0 | 150 | ; push edx ; need not be preserved |
| michael@0 | 151 | push esi |
| michael@0 | 152 | push edi |
| michael@0 | 153 | |
| michael@0 | 154 | mov esi, POINTER [workspace] |
| michael@0 | 155 | mov edx, POINTER [divisors] |
| michael@0 | 156 | mov edi, JCOEFPTR [coef_block] |
| michael@0 | 157 | mov ah, 2 |
| michael@0 | 158 | alignx 16,7 |
| michael@0 | 159 | .quantloop1: |
| michael@0 | 160 | mov al, DCTSIZE2/8/2 |
| michael@0 | 161 | alignx 16,7 |
| michael@0 | 162 | .quantloop2: |
| michael@0 | 163 | movq mm2, MMWORD [MMBLOCK(0,0,esi,SIZEOF_DCTELEM)] |
| michael@0 | 164 | movq mm3, MMWORD [MMBLOCK(0,1,esi,SIZEOF_DCTELEM)] |
| michael@0 | 165 | |
| michael@0 | 166 | movq mm0,mm2 |
| michael@0 | 167 | movq mm1,mm3 |
| michael@0 | 168 | |
| michael@0 | 169 | psraw mm2,(WORD_BIT-1) ; -1 if value < 0, 0 otherwise |
| michael@0 | 170 | psraw mm3,(WORD_BIT-1) |
| michael@0 | 171 | |
| michael@0 | 172 | pxor mm0,mm2 ; val = -val |
| michael@0 | 173 | pxor mm1,mm3 |
| michael@0 | 174 | psubw mm0,mm2 |
| michael@0 | 175 | psubw mm1,mm3 |
| michael@0 | 176 | |
| michael@0 | 177 | ; |
| michael@0 | 178 | ; MMX is an annoyingly crappy instruction set. It has two |
| michael@0 | 179 | ; misfeatures that are causing problems here: |
| michael@0 | 180 | ; |
| michael@0 | 181 | ; - All multiplications are signed. |
| michael@0 | 182 | ; |
| michael@0 | 183 | ; - The second operand for the shifts is not treated as packed. |
| michael@0 | 184 | ; |
| michael@0 | 185 | ; |
| michael@0 | 186 | ; We work around the first problem by implementing this algorithm: |
| michael@0 | 187 | ; |
| michael@0 | 188 | ; unsigned long unsigned_multiply(unsigned short x, unsigned short y) |
| michael@0 | 189 | ; { |
| michael@0 | 190 | ; enum { SHORT_BIT = 16 }; |
| michael@0 | 191 | ; signed short sx = (signed short) x; |
| michael@0 | 192 | ; signed short sy = (signed short) y; |
| michael@0 | 193 | ; signed long sz; |
| michael@0 | 194 | ; |
| michael@0 | 195 | ; sz = (long) sx * (long) sy; /* signed multiply */ |
| michael@0 | 196 | ; |
| michael@0 | 197 | ; if (sx < 0) sz += (long) sy << SHORT_BIT; |
| michael@0 | 198 | ; if (sy < 0) sz += (long) sx << SHORT_BIT; |
| michael@0 | 199 | ; |
| michael@0 | 200 | ; return (unsigned long) sz; |
| michael@0 | 201 | ; } |
| michael@0 | 202 | ; |
| michael@0 | 203 | ; (note that a negative sx adds _sy_ and vice versa) |
| michael@0 | 204 | ; |
| michael@0 | 205 | ; For the second problem, we replace the shift by a multiplication. |
| michael@0 | 206 | ; Unfortunately that means we have to deal with the signed issue again. |
| michael@0 | 207 | ; |
| michael@0 | 208 | |
| michael@0 | 209 | paddw mm0, MMWORD [CORRECTION(0,0,edx)] ; correction + roundfactor |
| michael@0 | 210 | paddw mm1, MMWORD [CORRECTION(0,1,edx)] |
| michael@0 | 211 | |
| michael@0 | 212 | movq mm4,mm0 ; store current value for later |
| michael@0 | 213 | movq mm5,mm1 |
| michael@0 | 214 | pmulhw mm0, MMWORD [RECIPROCAL(0,0,edx)] ; reciprocal |
| michael@0 | 215 | pmulhw mm1, MMWORD [RECIPROCAL(0,1,edx)] |
| michael@0 | 216 | paddw mm0,mm4 ; reciprocal is always negative (MSB=1), |
| michael@0 | 217 | paddw mm1,mm5 ; so we always need to add the initial value |
| michael@0 | 218 | ; (input value is never negative as we |
| michael@0 | 219 | ; inverted it at the start of this routine) |
| michael@0 | 220 | |
| michael@0 | 221 | ; here it gets a bit tricky as both scale |
| michael@0 | 222 | ; and mm0/mm1 can be negative |
| michael@0 | 223 | movq mm6, MMWORD [SCALE(0,0,edx)] ; scale |
| michael@0 | 224 | movq mm7, MMWORD [SCALE(0,1,edx)] |
| michael@0 | 225 | movq mm4,mm0 |
| michael@0 | 226 | movq mm5,mm1 |
| michael@0 | 227 | pmulhw mm0,mm6 |
| michael@0 | 228 | pmulhw mm1,mm7 |
| michael@0 | 229 | |
| michael@0 | 230 | psraw mm6,(WORD_BIT-1) ; determine if scale is negative |
| michael@0 | 231 | psraw mm7,(WORD_BIT-1) |
| michael@0 | 232 | |
| michael@0 | 233 | pand mm6,mm4 ; and add input if it is |
| michael@0 | 234 | pand mm7,mm5 |
| michael@0 | 235 | paddw mm0,mm6 |
| michael@0 | 236 | paddw mm1,mm7 |
| michael@0 | 237 | |
| michael@0 | 238 | psraw mm4,(WORD_BIT-1) ; then check if negative input |
| michael@0 | 239 | psraw mm5,(WORD_BIT-1) |
| michael@0 | 240 | |
| michael@0 | 241 | pand mm4, MMWORD [SCALE(0,0,edx)] ; and add scale if it is |
| michael@0 | 242 | pand mm5, MMWORD [SCALE(0,1,edx)] |
| michael@0 | 243 | paddw mm0,mm4 |
| michael@0 | 244 | paddw mm1,mm5 |
| michael@0 | 245 | |
| michael@0 | 246 | pxor mm0,mm2 ; val = -val |
| michael@0 | 247 | pxor mm1,mm3 |
| michael@0 | 248 | psubw mm0,mm2 |
| michael@0 | 249 | psubw mm1,mm3 |
| michael@0 | 250 | |
| michael@0 | 251 | movq MMWORD [MMBLOCK(0,0,edi,SIZEOF_DCTELEM)], mm0 |
| michael@0 | 252 | movq MMWORD [MMBLOCK(0,1,edi,SIZEOF_DCTELEM)], mm1 |
| michael@0 | 253 | |
| michael@0 | 254 | add esi, byte 8*SIZEOF_DCTELEM |
| michael@0 | 255 | add edx, byte 8*SIZEOF_DCTELEM |
| michael@0 | 256 | add edi, byte 8*SIZEOF_JCOEF |
| michael@0 | 257 | dec al |
| michael@0 | 258 | jnz near .quantloop2 |
| michael@0 | 259 | dec ah |
| michael@0 | 260 | jnz near .quantloop1 ; to avoid branch misprediction |
| michael@0 | 261 | |
| michael@0 | 262 | emms ; empty MMX state |
| michael@0 | 263 | |
| michael@0 | 264 | pop edi |
| michael@0 | 265 | pop esi |
| michael@0 | 266 | ; pop edx ; need not be preserved |
| michael@0 | 267 | ; pop ecx ; unused |
| michael@0 | 268 | ; pop ebx ; unused |
| michael@0 | 269 | pop ebp |
| michael@0 | 270 | ret |
| michael@0 | 271 | |
| michael@0 | 272 | ; For some reason, the OS X linker does not honor the request to align the |
| michael@0 | 273 | ; segment unless we do this. |
| michael@0 | 274 | align 16 |
