DRC | cdc8ac3 | 2009-06-25 20:38:31 +0000 | [diff] [blame] | 1 | ; |
| 2 | ; jcsamss2.asm - downsampling (64-bit SSE2) |
| 3 | ; |
| 4 | ; Copyright 2009 Pierre Ossman <ossman@cendio.se> for Cendio AB |
| 5 | ; Copyright 2009 D. R. Commander |
| 6 | ; |
| 7 | ; Based on |
| 8 | ; x86 SIMD extension for IJG JPEG library |
| 9 | ; Copyright (C) 1999-2006, MIYASAKA Masaru. |
| 10 | ; For conditions of distribution and use, see copyright notice in jsimdext.inc |
| 11 | ; |
| 12 | ; This file should be assembled with NASM (Netwide Assembler), |
| 13 | ; can *not* be assembled with Microsoft's MASM or any compatible |
| 14 | ; assembler (including Borland's Turbo Assembler). |
| 15 | ; NASM is available from http://nasm.sourceforge.net/ or |
| 16 | ; http://sourceforge.net/project/showfiles.php?group_id=6208 |
| 17 | ; |
| 18 | ; [TAB8] |
| 19 | |
| 20 | %include "jsimdext.inc" |
| 21 | |
| 22 | ; -------------------------------------------------------------------------- |
| 23 | SECTION SEG_TEXT |
| 24 | BITS 64 |
| 25 | ; |
| 26 | ; Downsample pixel values of a single component. |
| 27 | ; This version handles the common case of 2:1 horizontal and 1:1 vertical, |
| 28 | ; without smoothing. |
| 29 | ; |
| 30 | ; GLOBAL(void) |
| 31 | ; jsimd_h2v1_downsample_sse2 (JDIMENSION image_width, int max_v_samp_factor, |
| 32 | ; JDIMENSION v_samp_factor, JDIMENSION width_blocks, |
| 33 | ; JSAMPARRAY input_data, JSAMPARRAY output_data); |
| 34 | ; |
| 35 | |
| 36 | ; r10 = JDIMENSION image_width |
| 37 | ; r11 = int max_v_samp_factor |
| 38 | ; r12 = JDIMENSION v_samp_factor |
| 39 | ; r13 = JDIMENSION width_blocks |
| 40 | ; r14 = JSAMPARRAY input_data |
| 41 | ; r15 = JSAMPARRAY output_data |
| 42 | |
| 43 | align 16 |
| 44 | global EXTN(jsimd_h2v1_downsample_sse2) |
| 45 | |
| 46 | EXTN(jsimd_h2v1_downsample_sse2): |
| 47 | push rbp |
| 48 | mov rbp,rsp |
| 49 | collect_args |
| 50 | |
| 51 | mov rcx, r13 |
| 52 | shl rcx,3 ; imul rcx,DCTSIZE (rcx = output_cols) |
| 53 | jz near .return |
| 54 | |
| 55 | mov rdx, r10 |
| 56 | |
| 57 | ; -- expand_right_edge |
| 58 | |
| 59 | push rcx |
| 60 | shl rcx,1 ; output_cols * 2 |
| 61 | sub rcx,rdx |
| 62 | jle short .expand_end |
| 63 | |
| 64 | mov rax, r11 |
| 65 | test rax,rax |
| 66 | jle short .expand_end |
| 67 | |
| 68 | cld |
| 69 | mov rsi, r14 ; input_data |
| 70 | .expandloop: |
| 71 | push rax |
| 72 | push rcx |
| 73 | |
| 74 | mov rdi, JSAMPROW [rsi] |
| 75 | add rdi,rdx |
| 76 | mov al, JSAMPLE [rdi-1] |
| 77 | |
| 78 | rep stosb |
| 79 | |
| 80 | pop rcx |
| 81 | pop rax |
| 82 | |
| 83 | add rsi, byte SIZEOF_JSAMPROW |
| 84 | dec rax |
| 85 | jg short .expandloop |
| 86 | |
| 87 | .expand_end: |
| 88 | pop rcx ; output_cols |
| 89 | |
| 90 | ; -- h2v1_downsample |
| 91 | |
| 92 | mov rax, r12 ; rowctr |
| 93 | test eax,eax |
| 94 | jle near .return |
| 95 | |
| 96 | mov rdx, 0x00010000 ; bias pattern |
| 97 | movd xmm7,edx |
| 98 | pcmpeqw xmm6,xmm6 |
| 99 | pshufd xmm7,xmm7,0x00 ; xmm7={0, 1, 0, 1, 0, 1, 0, 1} |
| 100 | psrlw xmm6,BYTE_BIT ; xmm6={0xFF 0x00 0xFF 0x00 ..} |
| 101 | |
| 102 | mov rsi, r14 ; input_data |
| 103 | mov rdi, r15 ; output_data |
| 104 | .rowloop: |
| 105 | push rcx |
| 106 | push rdi |
| 107 | push rsi |
| 108 | |
| 109 | mov rsi, JSAMPROW [rsi] ; inptr |
| 110 | mov rdi, JSAMPROW [rdi] ; outptr |
| 111 | |
| 112 | cmp rcx, byte SIZEOF_XMMWORD |
| 113 | jae short .columnloop |
| 114 | |
| 115 | .columnloop_r8: |
| 116 | movdqa xmm0, XMMWORD [rsi+0*SIZEOF_XMMWORD] |
| 117 | pxor xmm1,xmm1 |
| 118 | mov rcx, SIZEOF_XMMWORD |
| 119 | jmp short .downsample |
| 120 | |
| 121 | .columnloop: |
| 122 | movdqa xmm0, XMMWORD [rsi+0*SIZEOF_XMMWORD] |
| 123 | movdqa xmm1, XMMWORD [rsi+1*SIZEOF_XMMWORD] |
| 124 | |
| 125 | .downsample: |
| 126 | movdqa xmm2,xmm0 |
| 127 | movdqa xmm3,xmm1 |
| 128 | |
| 129 | pand xmm0,xmm6 |
| 130 | psrlw xmm2,BYTE_BIT |
| 131 | pand xmm1,xmm6 |
| 132 | psrlw xmm3,BYTE_BIT |
| 133 | |
| 134 | paddw xmm0,xmm2 |
| 135 | paddw xmm1,xmm3 |
| 136 | paddw xmm0,xmm7 |
| 137 | paddw xmm1,xmm7 |
| 138 | psrlw xmm0,1 |
| 139 | psrlw xmm1,1 |
| 140 | |
| 141 | packuswb xmm0,xmm1 |
| 142 | |
| 143 | movdqa XMMWORD [rdi+0*SIZEOF_XMMWORD], xmm0 |
| 144 | |
| 145 | sub rcx, byte SIZEOF_XMMWORD ; outcol |
| 146 | add rsi, byte 2*SIZEOF_XMMWORD ; inptr |
| 147 | add rdi, byte 1*SIZEOF_XMMWORD ; outptr |
| 148 | cmp rcx, byte SIZEOF_XMMWORD |
| 149 | jae short .columnloop |
| 150 | test rcx,rcx |
| 151 | jnz short .columnloop_r8 |
| 152 | |
| 153 | pop rsi |
| 154 | pop rdi |
| 155 | pop rcx |
| 156 | |
| 157 | add rsi, byte SIZEOF_JSAMPROW ; input_data |
| 158 | add rdi, byte SIZEOF_JSAMPROW ; output_data |
| 159 | dec rax ; rowctr |
| 160 | jg near .rowloop |
| 161 | |
| 162 | .return: |
| 163 | uncollect_args |
| 164 | pop rbp |
| 165 | ret |
| 166 | |
| 167 | ; -------------------------------------------------------------------------- |
| 168 | ; |
| 169 | ; Downsample pixel values of a single component. |
| 170 | ; This version handles the standard case of 2:1 horizontal and 2:1 vertical, |
| 171 | ; without smoothing. |
| 172 | ; |
| 173 | ; GLOBAL(void) |
| 174 | ; jsimd_h2v2_downsample_sse2 (JDIMENSION image_width, int max_v_samp_factor, |
| 175 | ; JDIMENSION v_samp_factor, JDIMENSION width_blocks, |
| 176 | ; JSAMPARRAY input_data, JSAMPARRAY output_data); |
| 177 | ; |
| 178 | |
| 179 | ; r10 = JDIMENSION image_width |
| 180 | ; r11 = int max_v_samp_factor |
| 181 | ; r12 = JDIMENSION v_samp_factor |
| 182 | ; r13 = JDIMENSION width_blocks |
| 183 | ; r14 = JSAMPARRAY input_data |
| 184 | ; r15 = JSAMPARRAY output_data |
| 185 | |
| 186 | align 16 |
| 187 | global EXTN(jsimd_h2v2_downsample_sse2) |
| 188 | |
| 189 | EXTN(jsimd_h2v2_downsample_sse2): |
| 190 | push rbp |
| 191 | mov rbp,rsp |
| 192 | collect_args |
| 193 | |
| 194 | mov rcx, r13 |
| 195 | shl rcx,3 ; imul rcx,DCTSIZE (rcx = output_cols) |
| 196 | jz near .return |
| 197 | |
| 198 | mov rdx, r10 |
| 199 | |
| 200 | ; -- expand_right_edge |
| 201 | |
| 202 | push rcx |
| 203 | shl rcx,1 ; output_cols * 2 |
| 204 | sub rcx,rdx |
| 205 | jle short .expand_end |
| 206 | |
| 207 | mov rax, r11 |
| 208 | test rax,rax |
| 209 | jle short .expand_end |
| 210 | |
| 211 | cld |
| 212 | mov rsi, r14 ; input_data |
| 213 | .expandloop: |
| 214 | push rax |
| 215 | push rcx |
| 216 | |
| 217 | mov rdi, JSAMPROW [rsi] |
| 218 | add rdi,rdx |
| 219 | mov al, JSAMPLE [rdi-1] |
| 220 | |
| 221 | rep stosb |
| 222 | |
| 223 | pop rcx |
| 224 | pop rax |
| 225 | |
| 226 | add rsi, byte SIZEOF_JSAMPROW |
| 227 | dec rax |
| 228 | jg short .expandloop |
| 229 | |
| 230 | .expand_end: |
| 231 | pop rcx ; output_cols |
| 232 | |
| 233 | ; -- h2v2_downsample |
| 234 | |
| 235 | mov rax, r12 ; rowctr |
| 236 | test rax,rax |
| 237 | jle near .return |
| 238 | |
| 239 | mov rdx, 0x00020001 ; bias pattern |
| 240 | movd xmm7,edx |
| 241 | pcmpeqw xmm6,xmm6 |
| 242 | pshufd xmm7,xmm7,0x00 ; xmm7={1, 2, 1, 2, 1, 2, 1, 2} |
| 243 | psrlw xmm6,BYTE_BIT ; xmm6={0xFF 0x00 0xFF 0x00 ..} |
| 244 | |
| 245 | mov rsi, r14 ; input_data |
| 246 | mov rdi, r15 ; output_data |
| 247 | .rowloop: |
| 248 | push rcx |
| 249 | push rdi |
| 250 | push rsi |
| 251 | |
| 252 | mov rdx, JSAMPROW [rsi+0*SIZEOF_JSAMPROW] ; inptr0 |
| 253 | mov rsi, JSAMPROW [rsi+1*SIZEOF_JSAMPROW] ; inptr1 |
| 254 | mov rdi, JSAMPROW [rdi] ; outptr |
| 255 | |
| 256 | cmp rcx, byte SIZEOF_XMMWORD |
| 257 | jae short .columnloop |
| 258 | |
| 259 | .columnloop_r8: |
| 260 | movdqa xmm0, XMMWORD [rdx+0*SIZEOF_XMMWORD] |
| 261 | movdqa xmm1, XMMWORD [rsi+0*SIZEOF_XMMWORD] |
| 262 | pxor xmm2,xmm2 |
| 263 | pxor xmm3,xmm3 |
| 264 | mov rcx, SIZEOF_XMMWORD |
| 265 | jmp short .downsample |
| 266 | |
| 267 | .columnloop: |
| 268 | movdqa xmm0, XMMWORD [rdx+0*SIZEOF_XMMWORD] |
| 269 | movdqa xmm1, XMMWORD [rsi+0*SIZEOF_XMMWORD] |
| 270 | movdqa xmm2, XMMWORD [rdx+1*SIZEOF_XMMWORD] |
| 271 | movdqa xmm3, XMMWORD [rsi+1*SIZEOF_XMMWORD] |
| 272 | |
| 273 | .downsample: |
| 274 | movdqa xmm4,xmm0 |
| 275 | movdqa xmm5,xmm1 |
| 276 | pand xmm0,xmm6 |
| 277 | psrlw xmm4,BYTE_BIT |
| 278 | pand xmm1,xmm6 |
| 279 | psrlw xmm5,BYTE_BIT |
| 280 | paddw xmm0,xmm4 |
| 281 | paddw xmm1,xmm5 |
| 282 | |
| 283 | movdqa xmm4,xmm2 |
| 284 | movdqa xmm5,xmm3 |
| 285 | pand xmm2,xmm6 |
| 286 | psrlw xmm4,BYTE_BIT |
| 287 | pand xmm3,xmm6 |
| 288 | psrlw xmm5,BYTE_BIT |
| 289 | paddw xmm2,xmm4 |
| 290 | paddw xmm3,xmm5 |
| 291 | |
| 292 | paddw xmm0,xmm1 |
| 293 | paddw xmm2,xmm3 |
| 294 | paddw xmm0,xmm7 |
| 295 | paddw xmm2,xmm7 |
| 296 | psrlw xmm0,2 |
| 297 | psrlw xmm2,2 |
| 298 | |
| 299 | packuswb xmm0,xmm2 |
| 300 | |
| 301 | movdqa XMMWORD [rdi+0*SIZEOF_XMMWORD], xmm0 |
| 302 | |
| 303 | sub rcx, byte SIZEOF_XMMWORD ; outcol |
| 304 | add rdx, byte 2*SIZEOF_XMMWORD ; inptr0 |
| 305 | add rsi, byte 2*SIZEOF_XMMWORD ; inptr1 |
| 306 | add rdi, byte 1*SIZEOF_XMMWORD ; outptr |
| 307 | cmp rcx, byte SIZEOF_XMMWORD |
| 308 | jae near .columnloop |
| 309 | test rcx,rcx |
| 310 | jnz near .columnloop_r8 |
| 311 | |
| 312 | pop rsi |
| 313 | pop rdi |
| 314 | pop rcx |
| 315 | |
| 316 | add rsi, byte 2*SIZEOF_JSAMPROW ; input_data |
| 317 | add rdi, byte 1*SIZEOF_JSAMPROW ; output_data |
| 318 | dec rax ; rowctr |
| 319 | jg near .rowloop |
| 320 | |
| 321 | .return: |
| 322 | uncollect_args |
| 323 | pop rbp |
| 324 | ret |
DRC | 132b5fd | 2009-10-08 09:04:56 +0000 | [diff] [blame^] | 325 | |
| 326 | ; For some reason, the OS X linker does not honor the request to align the |
| 327 | ; segment unless we do this. |
| 328 | align 16 |