jpeg/simd/jsimd.h

   1 /*
   2  * simd/jsimd.h
   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  */
  11
  12 /* Bitmask for supported acceleration methods */
  13
  14 #define JSIMD_NONE    0x00
  15 #define JSIMD_MMX     0x01
  16 #define JSIMD_3DNOW   0x02
  17 #define JSIMD_SSE     0x04
  18 #define JSIMD_SSE2    0x08
  19
  20 /* Short forms of external names for systems with brain-damaged linkers. */
  21
  22 #ifdef NEED_SHORT_EXTERNAL_NAMES
  23 #define jpeg_simd_cpu_support                 jSiCpuSupport
  24 #define jsimd_rgb_ycc_convert_mmx             jSRGBYCCM
  25 #define jsimd_extrgb_ycc_convert_mmx          jSEXTRGBYCCM
  26 #define jsimd_extrgbx_ycc_convert_mmx         jSEXTRGBXYCCM
  27 #define jsimd_extbgr_ycc_convert_mmx          jSEXTBGRYCCM
  28 #define jsimd_extbgrx_ycc_convert_mmx         jSEXTBGRXYCCM
  29 #define jsimd_extxbgr_ycc_convert_mmx         jSEXTXBGRYCCM
  30 #define jsimd_extxrgb_ycc_convert_mmx         jSEXTXRGBYCCM
  31 #define jsimd_ycc_rgb_convert_mmx             jSYCCRGBM
  32 #define jsimd_ycc_extrgb_convert_mmx          jSYCCEXTRGBM
  33 #define jsimd_ycc_extrgbx_convert_mmx         jSYCCEXTRGBXM
  34 #define jsimd_ycc_extbgr_convert_mmx          jSYCCEXTBGRM
  35 #define jsimd_ycc_extbgrx_convert_mmx         jSYCCEXTBGRXM
  36 #define jsimd_ycc_extxbgr_convert_mmx         jSYCCEXTXBGRM
  37 #define jsimd_ycc_extxrgb_convert_mmx         jSYCCEXTXRGBM
  38 #define jconst_rgb_ycc_convert_sse2           jSCRGBYCCS2
  39 #define jsimd_rgb_ycc_convert_sse2            jSRGBYCCS2
  40 #define jsimd_extrgb_ycc_convert_sse2         jSEXTRGBYCCS2
  41 #define jsimd_extrgbx_ycc_convert_sse2        jSEXTRGBXYCCS2
  42 #define jsimd_extbgr_ycc_convert_sse2         jSEXTBGRYCCS2
  43 #define jsimd_extbgrx_ycc_convert_sse2        jSEXTBGRXYCCS2
  44 #define jsimd_extxbgr_ycc_convert_sse2        jSEXTXBGRYCCS2
  45 #define jsimd_extxrgb_ycc_convert_sse2        jSEXTXRGBYCCS2
  46 #define jconst_ycc_rgb_convert_sse2           jSCYCCRGBS2
  47 #define jsimd_ycc_rgb_convert_sse2            jSYCCRGBS2
  48 #define jsimd_ycc_extrgb_convert_sse2         jSYCCEXTRGBS2
  49 #define jsimd_ycc_extrgbx_convert_sse2        jSYCCEXTRGBXS2
  50 #define jsimd_ycc_extbgr_convert_sse2         jSYCCEXTBGRS2
  51 #define jsimd_ycc_extbgrx_convert_sse2        jSYCCEXTBGRXS2
  52 #define jsimd_ycc_extxbgr_convert_sse2        jSYCCEXTXBGRS2
  53 #define jsimd_ycc_extxrgb_convert_sse2        jSYCCEXTXRGBS2
  54 #define jsimd_h2v2_downsample_mmx             jSDnH2V2M
  55 #define jsimd_h2v1_downsample_mmx             jSDnH2V1M
  56 #define jsimd_h2v2_downsample_sse2            jSDnH2V2S2
  57 #define jsimd_h2v1_downsample_sse2            jSDnH2V1S2
  58 #define jsimd_h2v2_upsample_mmx               jSUpH2V2M
  59 #define jsimd_h2v1_upsample_mmx               jSUpH2V1M
  60 #define jsimd_h2v2_fancy_upsample_mmx         jSFUpH2V2M
  61 #define jsimd_h2v1_fancy_upsample_mmx         jSFUpH2V1M
  62 #define jsimd_h2v2_merged_upsample_mmx        jSMUpH2V2M
  63 #define jsimd_h2v2_extrgb_merged_upsample_mmx jSMUpH2V2EXTRGBM
  64 #define jsimd_h2v2_extrgbx_merged_upsample_mmx jSMUpH2V2EXTRGBXM
  65 #define jsimd_h2v2_extbgr_merged_upsample_mmx jSMUpH2V2EXTBGRM
  66 #define jsimd_h2v2_extbgrx_merged_upsample_mmx jSMUpH2V2EXTBGRXM
  67 #define jsimd_h2v2_extxbgr_merged_upsample_mmx jSMUpH2V2EXTXBGRM
  68 #define jsimd_h2v2_extxrgb_merged_upsample_mmx jSMUpH2V2EXTXRGBM
  69 #define jsimd_h2v1_merged_upsample_mmx        jSMUpH2V1M
  70 #define jsimd_h2v1_extrgb_merged_upsample_mmx jSMUpH2V1EXTRGBM
  71 #define jsimd_h2v1_extrgbx_merged_upsample_mmx jSMUpH2V1EXTRGBXM
  72 #define jsimd_h2v1_extbgr_merged_upsample_mmx jSMUpH2V1EXTBGRM
  73 #define jsimd_h2v1_extbgrx_merged_upsample_mmx jSMUpH2V1EXTBGRXM
  74 #define jsimd_h2v1_extxbgr_merged_upsample_mmx jSMUpH2V1EXTXBGRM
  75 #define jsimd_h2v1_extxrgb_merged_upsample_mmx jSMUpH2V1EXTXRGBM
  76 #define jsimd_h2v2_upsample_sse2              jSUpH2V2S2
  77 #define jsimd_h2v1_upsample_sse2              jSUpH2V1S2
  78 #define jconst_fancy_upsample_sse2            jSCFUpS2
  79 #define jsimd_h2v2_fancy_upsample_sse2        jSFUpH2V2S2
  80 #define jsimd_h2v1_fancy_upsample_sse2        jSFUpH2V1S2
  81 #define jconst_merged_upsample_sse2           jSCMUpS2
  82 #define jsimd_h2v2_merged_upsample_sse2       jSMUpH2V2S2
  83 #define jsimd_h2v2_extrgb_merged_upsample_sse2 jSMUpH2V2EXTRGBS2
  84 #define jsimd_h2v2_extrgbx_merged_upsample_sse2 jSMUpH2V2EXTRGBXS2
  85 #define jsimd_h2v2_extbgr_merged_upsample_sse2 jSMUpH2V2EXTBGRS2
  86 #define jsimd_h2v2_extbgrx_merged_upsample_sse2 jSMUpH2V2EXTBGRXS2
  87 #define jsimd_h2v2_extxbgr_merged_upsample_sse2 jSMUpH2V2EXTXBGRS2
  88 #define jsimd_h2v2_extxrgb_merged_upsample_sse2 jSMUpH2V2EXTXRGBS2
  89 #define jsimd_h2v1_merged_upsample_sse2       jSMUpH2V1S2
  90 #define jsimd_h2v1_extrgb_merged_upsample_sse2 jSMUpH2V1EXTRGBS2
  91 #define jsimd_h2v1_extrgbx_merged_upsample_sse2 jSMUpH2V1EXTRGBXS2
  92 #define jsimd_h2v1_extbgr_merged_upsample_sse2 jSMUpH2V1EXTBGRS2
  93 #define jsimd_h2v1_extbgrx_merged_upsample_sse2 jSMUpH2V1EXTBGRXS2
  94 #define jsimd_h2v1_extxbgr_merged_upsample_sse2 jSMUpH2V1EXTXBGRS2
  95 #define jsimd_h2v1_extxrgb_merged_upsample_sse2 jSMUpH2V1EXTXRGBS2
  96 #define jsimd_convsamp_mmx                    jSConvM
  97 #define jsimd_convsamp_sse2                   jSConvS2
  98 #define jsimd_convsamp_float_3dnow            jSConvF3D
  99 #define jsimd_convsamp_float_sse              jSConvFS
 100 #define jsimd_convsamp_float_sse2             jSConvFS2
 101 #define jsimd_fdct_islow_mmx                  jSFDMIS
 102 #define jsimd_fdct_ifast_mmx                  jSFDMIF
 103 #define jconst_fdct_islow_sse2                jSCFDS2IS
 104 #define jsimd_fdct_islow_sse2                 jSFDS2IS
 105 #define jconst_fdct_ifast_sse2                jSCFDS2IF
 106 #define jsimd_fdct_ifast_sse2                 jSFDS2IF
 107 #define jsimd_fdct_float_3dnow                jSFD3DF
 108 #define jconst_fdct_float_sse                 jSCFDSF
 109 #define jsimd_fdct_float_sse                  jSFDSF
 110 #define jsimd_quantize_mmx                    jSQuantM
 111 #define jsimd_quantize_sse2                   jSQuantS2
 112 #define jsimd_quantize_float_3dnow            jSQuantF3D
 113 #define jsimd_quantize_float_sse              jSQuantFS
 114 #define jsimd_quantize_float_sse2             jSQuantFS2
 115 #define jsimd_idct_2x2_mmx                    jSIDM22
 116 #define jsimd_idct_4x4_mmx                    jSIDM44
 117 #define jconst_idct_red_sse2                  jSCIDS2R
 118 #define jsimd_idct_2x2_sse2                   jSIDS222
 119 #define jsimd_idct_4x4_sse2                   jSIDS244
 120 #define jsimd_idct_islow_mmx                  jSIDMIS
 121 #define jsimd_idct_ifast_mmx                  jSIDMIF
 122 #define jconst_idct_islow_sse2                jSCIDS2IS
 123 #define jsimd_idct_islow_sse2                 jSIDS2IS
 124 #define jconst_idct_ifast_sse2                jSCIDS2IF
 125 #define jsimd_idct_ifast_sse2                 jSIDS2IF
 126 #define jsimd_idct_float_3dnow                jSID3DF
 127 #define jconst_fdct_float_sse                 jSCIDSF
 128 #define jsimd_idct_float_sse                  jSIDSF
 129 #define jconst_fdct_float_sse2                jSCIDS2F
 130 #define jsimd_idct_float_sse2                 jSIDS2F
 131 #endif /* NEED_SHORT_EXTERNAL_NAMES */
 132
 133 /* SIMD Ext: retrieve SIMD/CPU information */
 134 EXTERN(unsigned int) jpeg_simd_cpu_support JPP((void));
 135
 136 /* SIMD Color Space Conversion */
 137 EXTERN(void) jsimd_rgb_ycc_convert_mmx
 138         JPP((JDIMENSION img_width,
 139              JSAMPARRAY input_buf, JSAMPIMAGE output_buf,
 140              JDIMENSION output_row, int num_rows));
 141 EXTERN(void) jsimd_extrgb_ycc_convert_mmx
 142         JPP((JDIMENSION img_width,
 143              JSAMPARRAY input_buf, JSAMPIMAGE output_buf,
 144              JDIMENSION output_row, int num_rows));
 145 EXTERN(void) jsimd_extrgbx_ycc_convert_mmx
 146         JPP((JDIMENSION img_width,
 147              JSAMPARRAY input_buf, JSAMPIMAGE output_buf,
 148              JDIMENSION output_row, int num_rows));
 149 EXTERN(void) jsimd_extbgr_ycc_convert_mmx
 150         JPP((JDIMENSION img_width,
 151              JSAMPARRAY input_buf, JSAMPIMAGE output_buf,
 152              JDIMENSION output_row, int num_rows));
 153 EXTERN(void) jsimd_extbgrx_ycc_convert_mmx
 154         JPP((JDIMENSION img_width,
 155              JSAMPARRAY input_buf, JSAMPIMAGE output_buf,
 156              JDIMENSION output_row, int num_rows));
 157 EXTERN(void) jsimd_extxbgr_ycc_convert_mmx
 158         JPP((JDIMENSION img_width,
 159              JSAMPARRAY input_buf, JSAMPIMAGE output_buf,
 160              JDIMENSION output_row, int num_rows));
 161 EXTERN(void) jsimd_extxrgb_ycc_convert_mmx
 162         JPP((JDIMENSION img_width,
 163              JSAMPARRAY input_buf, JSAMPIMAGE output_buf,
 164              JDIMENSION output_row, int num_rows));
 165
 166 EXTERN(void) jsimd_ycc_rgb_convert_mmx
 167         JPP((JDIMENSION out_width,
 168              JSAMPIMAGE input_buf, JDIMENSION input_row,
 169              JSAMPARRAY output_buf, int num_rows));
 170 EXTERN(void) jsimd_ycc_extrgb_convert_mmx
 171         JPP((JDIMENSION out_width,
 172              JSAMPIMAGE input_buf, JDIMENSION input_row,
 173              JSAMPARRAY output_buf, int num_rows));
 174 EXTERN(void) jsimd_ycc_extrgbx_convert_mmx
 175         JPP((JDIMENSION out_width,
 176              JSAMPIMAGE input_buf, JDIMENSION input_row,
 177              JSAMPARRAY output_buf, int num_rows));
 178 EXTERN(void) jsimd_ycc_extbgr_convert_mmx
 179         JPP((JDIMENSION out_width,
 180              JSAMPIMAGE input_buf, JDIMENSION input_row,
 181              JSAMPARRAY output_buf, int num_rows));
 182 EXTERN(void) jsimd_ycc_extbgrx_convert_mmx
 183         JPP((JDIMENSION out_width,
 184              JSAMPIMAGE input_buf, JDIMENSION input_row,
 185              JSAMPARRAY output_buf, int num_rows));
 186 EXTERN(void) jsimd_ycc_extxbgr_convert_mmx
 187         JPP((JDIMENSION out_width,
 188              JSAMPIMAGE input_buf, JDIMENSION input_row,
 189              JSAMPARRAY output_buf, int num_rows));
 190 EXTERN(void) jsimd_ycc_extxrgb_convert_mmx
 191         JPP((JDIMENSION out_width,
 192              JSAMPIMAGE input_buf, JDIMENSION input_row,
 193              JSAMPARRAY output_buf, int num_rows));
 194
 195 extern const int jconst_rgb_ycc_convert_sse2[];
 196 EXTERN(void) jsimd_rgb_ycc_convert_sse2
 197         JPP((JDIMENSION img_width,
 198              JSAMPARRAY input_buf, JSAMPIMAGE output_buf,
 199              JDIMENSION output_row, int num_rows));
 200 EXTERN(void) jsimd_extrgb_ycc_convert_sse2
 201         JPP((JDIMENSION img_width,
 202              JSAMPARRAY input_buf, JSAMPIMAGE output_buf,
 203              JDIMENSION output_row, int num_rows));
 204 EXTERN(void) jsimd_extrgbx_ycc_convert_sse2
 205         JPP((JDIMENSION img_width,
 206              JSAMPARRAY input_buf, JSAMPIMAGE output_buf,
 207              JDIMENSION output_row, int num_rows));
 208 EXTERN(void) jsimd_extbgr_ycc_convert_sse2
 209         JPP((JDIMENSION img_width,
 210              JSAMPARRAY input_buf, JSAMPIMAGE output_buf,
 211              JDIMENSION output_row, int num_rows));
 212 EXTERN(void) jsimd_extbgrx_ycc_convert_sse2
 213         JPP((JDIMENSION img_width,
 214              JSAMPARRAY input_buf, JSAMPIMAGE output_buf,
 215              JDIMENSION output_row, int num_rows));
 216 EXTERN(void) jsimd_extxbgr_ycc_convert_sse2
 217         JPP((JDIMENSION img_width,
 218              JSAMPARRAY input_buf, JSAMPIMAGE output_buf,
 219              JDIMENSION output_row, int num_rows));
 220 EXTERN(void) jsimd_extxrgb_ycc_convert_sse2
 221         JPP((JDIMENSION img_width,
 222              JSAMPARRAY input_buf, JSAMPIMAGE output_buf,
 223              JDIMENSION output_row, int num_rows));
 224
 225 extern const int jconst_ycc_rgb_convert_sse2[];
 226 EXTERN(void) jsimd_ycc_rgb_convert_sse2
 227         JPP((JDIMENSION out_width,
 228              JSAMPIMAGE input_buf, JDIMENSION input_row,
 229              JSAMPARRAY output_buf, int num_rows));
 230 EXTERN(void) jsimd_ycc_extrgb_convert_sse2
 231         JPP((JDIMENSION out_width,
 232              JSAMPIMAGE input_buf, JDIMENSION input_row,
 233              JSAMPARRAY output_buf, int num_rows));
 234 EXTERN(void) jsimd_ycc_extrgbx_convert_sse2
 235         JPP((JDIMENSION out_width,
 236              JSAMPIMAGE input_buf, JDIMENSION input_row,
 237              JSAMPARRAY output_buf, int num_rows));
 238 EXTERN(void) jsimd_ycc_extbgr_convert_sse2
 239         JPP((JDIMENSION out_width,
 240              JSAMPIMAGE input_buf, JDIMENSION input_row,
 241              JSAMPARRAY output_buf, int num_rows));
 242 EXTERN(void) jsimd_ycc_extbgrx_convert_sse2
 243         JPP((JDIMENSION out_width,
 244              JSAMPIMAGE input_buf, JDIMENSION input_row,
 245              JSAMPARRAY output_buf, int num_rows));
 246 EXTERN(void) jsimd_ycc_extxbgr_convert_sse2
 247         JPP((JDIMENSION out_width,
 248              JSAMPIMAGE input_buf, JDIMENSION input_row,
 249              JSAMPARRAY output_buf, int num_rows));
 250 EXTERN(void) jsimd_ycc_extxrgb_convert_sse2
 251         JPP((JDIMENSION out_width,
 252              JSAMPIMAGE input_buf, JDIMENSION input_row,
 253              JSAMPARRAY output_buf, int num_rows));
 254
 255 /* SIMD Downsample */
 256 EXTERN(void) jsimd_h2v2_downsample_mmx
 257         JPP((JDIMENSION image_width, int max_v_samp_factor,
 258              JDIMENSION v_samp_factor, JDIMENSION width_blocks,
 259              JSAMPARRAY input_data, JSAMPARRAY output_data));
 260 EXTERN(void) jsimd_h2v1_downsample_mmx
 261         JPP((JDIMENSION image_width, int max_v_samp_factor,
 262              JDIMENSION v_samp_factor, JDIMENSION width_blocks,
 263              JSAMPARRAY input_data, JSAMPARRAY output_data));
 264
 265 EXTERN(void) jsimd_h2v2_downsample_sse2
 266         JPP((JDIMENSION image_width, int max_v_samp_factor,
 267              JDIMENSION v_samp_factor, JDIMENSION width_blocks,
 268              JSAMPARRAY input_data, JSAMPARRAY output_data));
 269 EXTERN(void) jsimd_h2v1_downsample_sse2
 270         JPP((JDIMENSION image_width, int max_v_samp_factor,
 271              JDIMENSION v_samp_factor, JDIMENSION width_blocks,
 272              JSAMPARRAY input_data, JSAMPARRAY output_data));
 273
 274 /* SIMD Upsample */
 275 EXTERN(void) jsimd_h2v2_upsample_mmx
 276         JPP((int max_v_samp_factor, JDIMENSION output_width,
 277              JSAMPARRAY input_data, JSAMPARRAY * output_data_ptr));
 278 EXTERN(void) jsimd_h2v1_upsample_mmx
 279         JPP((int max_v_samp_factor, JDIMENSION output_width,
 280              JSAMPARRAY input_data, JSAMPARRAY * output_data_ptr));
 281
 282 EXTERN(void) jsimd_h2v2_fancy_upsample_mmx
 283         JPP((int max_v_samp_factor, JDIMENSION downsampled_width,
 284              JSAMPARRAY input_data, JSAMPARRAY * output_data_ptr));
 285 EXTERN(void) jsimd_h2v1_fancy_upsample_mmx
 286         JPP((int max_v_samp_factor, JDIMENSION downsampled_width,
 287              JSAMPARRAY input_data, JSAMPARRAY * output_data_ptr));
 288
 289 EXTERN(void) jsimd_h2v2_merged_upsample_mmx
 290         JPP((JDIMENSION output_width, JSAMPIMAGE input_buf,
 291              JDIMENSION in_row_group_ctr, JSAMPARRAY output_buf));
 292 EXTERN(void) jsimd_h2v2_extrgb_merged_upsample_mmx
 293         JPP((JDIMENSION output_width, JSAMPIMAGE input_buf,
 294              JDIMENSION in_row_group_ctr, JSAMPARRAY output_buf));
 295 EXTERN(void) jsimd_h2v2_extrgbx_merged_upsample_mmx
 296         JPP((JDIMENSION output_width, JSAMPIMAGE input_buf,
 297              JDIMENSION in_row_group_ctr, JSAMPARRAY output_buf));
 298 EXTERN(void) jsimd_h2v2_extbgr_merged_upsample_mmx
 299         JPP((JDIMENSION output_width, JSAMPIMAGE input_buf,
 300              JDIMENSION in_row_group_ctr, JSAMPARRAY output_buf));
 301 EXTERN(void) jsimd_h2v2_extbgrx_merged_upsample_mmx
 302         JPP((JDIMENSION output_width, JSAMPIMAGE input_buf,
 303              JDIMENSION in_row_group_ctr, JSAMPARRAY output_buf));
 304 EXTERN(void) jsimd_h2v2_extxbgr_merged_upsample_mmx
 305         JPP((JDIMENSION output_width, JSAMPIMAGE input_buf,
 306              JDIMENSION in_row_group_ctr, JSAMPARRAY output_buf));
 307 EXTERN(void) jsimd_h2v2_extxrgb_merged_upsample_mmx
 308         JPP((JDIMENSION output_width, JSAMPIMAGE input_buf,
 309              JDIMENSION in_row_group_ctr, JSAMPARRAY output_buf));
 310 EXTERN(void) jsimd_h2v1_merged_upsample_mmx
 311         JPP((JDIMENSION output_width, JSAMPIMAGE input_buf,
 312              JDIMENSION in_row_group_ctr, JSAMPARRAY output_buf));
 313 EXTERN(void) jsimd_h2v1_extrgb_merged_upsample_mmx
 314         JPP((JDIMENSION output_width, JSAMPIMAGE input_buf,
 315              JDIMENSION in_row_group_ctr, JSAMPARRAY output_buf));
 316 EXTERN(void) jsimd_h2v1_extrgbx_merged_upsample_mmx
 317         JPP((JDIMENSION output_width, JSAMPIMAGE input_buf,
 318              JDIMENSION in_row_group_ctr, JSAMPARRAY output_buf));
 319 EXTERN(void) jsimd_h2v1_extbgr_merged_upsample_mmx
 320         JPP((JDIMENSION output_width, JSAMPIMAGE input_buf,
 321              JDIMENSION in_row_group_ctr, JSAMPARRAY output_buf));
 322 EXTERN(void) jsimd_h2v1_extbgrx_merged_upsample_mmx
 323         JPP((JDIMENSION output_width, JSAMPIMAGE input_buf,
 324              JDIMENSION in_row_group_ctr, JSAMPARRAY output_buf));
 325 EXTERN(void) jsimd_h2v1_extxbgr_merged_upsample_mmx
 326         JPP((JDIMENSION output_width, JSAMPIMAGE input_buf,
 327              JDIMENSION in_row_group_ctr, JSAMPARRAY output_buf));
 328 EXTERN(void) jsimd_h2v1_extxrgb_merged_upsample_mmx
 329         JPP((JDIMENSION output_width, JSAMPIMAGE input_buf,
 330              JDIMENSION in_row_group_ctr, JSAMPARRAY output_buf));
 331
 332 EXTERN(void) jsimd_h2v2_upsample_sse2
 333         JPP((int max_v_samp_factor, JDIMENSION output_width,
 334              JSAMPARRAY input_data, JSAMPARRAY * output_data_ptr));
 335 EXTERN(void) jsimd_h2v1_upsample_sse2
 336         JPP((int max_v_samp_factor, JDIMENSION output_width,
 337              JSAMPARRAY input_data, JSAMPARRAY * output_data_ptr));
 338
 339 extern const int jconst_fancy_upsample_sse2[];
 340 EXTERN(void) jsimd_h2v2_fancy_upsample_sse2
 341         JPP((int max_v_samp_factor, JDIMENSION downsampled_width,
 342              JSAMPARRAY input_data, JSAMPARRAY * output_data_ptr));
 343 EXTERN(void) jsimd_h2v1_fancy_upsample_sse2
 344         JPP((int max_v_samp_factor, JDIMENSION downsampled_width,
 345              JSAMPARRAY input_data, JSAMPARRAY * output_data_ptr));
 346
 347 extern const int jconst_merged_upsample_sse2[];
 348 EXTERN(void) jsimd_h2v2_merged_upsample_sse2
 349         JPP((JDIMENSION output_width, JSAMPIMAGE input_buf,
 350              JDIMENSION in_row_group_ctr, JSAMPARRAY output_buf));
 351 EXTERN(void) jsimd_h2v2_extrgb_merged_upsample_sse2
 352         JPP((JDIMENSION output_width, JSAMPIMAGE input_buf,
 353              JDIMENSION in_row_group_ctr, JSAMPARRAY output_buf));
 354 EXTERN(void) jsimd_h2v2_extrgbx_merged_upsample_sse2
 355         JPP((JDIMENSION output_width, JSAMPIMAGE input_buf,
 356              JDIMENSION in_row_group_ctr, JSAMPARRAY output_buf));
 357 EXTERN(void) jsimd_h2v2_extbgr_merged_upsample_sse2
 358         JPP((JDIMENSION output_width, JSAMPIMAGE input_buf,
 359              JDIMENSION in_row_group_ctr, JSAMPARRAY output_buf));
 360 EXTERN(void) jsimd_h2v2_extbgrx_merged_upsample_sse2
 361         JPP((JDIMENSION output_width, JSAMPIMAGE input_buf,
 362              JDIMENSION in_row_group_ctr, JSAMPARRAY output_buf));
 363 EXTERN(void) jsimd_h2v2_extxbgr_merged_upsample_sse2
 364         JPP((JDIMENSION output_width, JSAMPIMAGE input_buf,
 365              JDIMENSION in_row_group_ctr, JSAMPARRAY output_buf));
 366 EXTERN(void) jsimd_h2v2_extxrgb_merged_upsample_sse2
 367         JPP((JDIMENSION output_width, JSAMPIMAGE input_buf,
 368              JDIMENSION in_row_group_ctr, JSAMPARRAY output_buf));
 369 EXTERN(void) jsimd_h2v1_merged_upsample_sse2
 370         JPP((JDIMENSION output_width, JSAMPIMAGE input_buf,
 371              JDIMENSION in_row_group_ctr, JSAMPARRAY output_buf));
 372 EXTERN(void) jsimd_h2v1_extrgb_merged_upsample_sse2
 373         JPP((JDIMENSION output_width, JSAMPIMAGE input_buf,
 374              JDIMENSION in_row_group_ctr, JSAMPARRAY output_buf));
 375 EXTERN(void) jsimd_h2v1_extrgbx_merged_upsample_sse2
 376         JPP((JDIMENSION output_width, JSAMPIMAGE input_buf,
 377              JDIMENSION in_row_group_ctr, JSAMPARRAY output_buf));
 378 EXTERN(void) jsimd_h2v1_extbgr_merged_upsample_sse2
 379         JPP((JDIMENSION output_width, JSAMPIMAGE input_buf,
 380              JDIMENSION in_row_group_ctr, JSAMPARRAY output_buf));
 381 EXTERN(void) jsimd_h2v1_extbgrx_merged_upsample_sse2
 382         JPP((JDIMENSION output_width, JSAMPIMAGE input_buf,
 383              JDIMENSION in_row_group_ctr, JSAMPARRAY output_buf));
 384 EXTERN(void) jsimd_h2v1_extxbgr_merged_upsample_sse2
 385         JPP((JDIMENSION output_width, JSAMPIMAGE input_buf,
 386              JDIMENSION in_row_group_ctr, JSAMPARRAY output_buf));
 387 EXTERN(void) jsimd_h2v1_extxrgb_merged_upsample_sse2
 388         JPP((JDIMENSION output_width, JSAMPIMAGE input_buf,
 389              JDIMENSION in_row_group_ctr, JSAMPARRAY output_buf));
 390
 391 /* SIMD Sample Conversion */
 392 EXTERN(void) jsimd_convsamp_mmx JPP((JSAMPARRAY sample_data,
 393                                      JDIMENSION start_col,
 394                                      DCTELEM * workspace));
 395
 396 EXTERN(void) jsimd_convsamp_sse2 JPP((JSAMPARRAY sample_data,
 397                                       JDIMENSION start_col,
 398                                       DCTELEM * workspace));
 399
 400 EXTERN(void) jsimd_convsamp_float_3dnow JPP((JSAMPARRAY sample_data,
 401                                              JDIMENSION start_col,
 402                                              FAST_FLOAT * workspace));
 403
 404 EXTERN(void) jsimd_convsamp_float_sse JPP((JSAMPARRAY sample_data,
 405                                            JDIMENSION start_col,
 406                                            FAST_FLOAT * workspace));
 407
 408 EXTERN(void) jsimd_convsamp_float_sse2 JPP((JSAMPARRAY sample_data,
 409                                             JDIMENSION start_col,
 410                                             FAST_FLOAT * workspace));
 411
 412 /* SIMD Forward DCT */
 413 EXTERN(void) jsimd_fdct_islow_mmx JPP((DCTELEM * data));
 414 EXTERN(void) jsimd_fdct_ifast_mmx JPP((DCTELEM * data));
 415
 416 extern const int jconst_fdct_ifast_sse2[];
 417 EXTERN(void) jsimd_fdct_islow_sse2 JPP((DCTELEM * data));
 418 extern const int jconst_fdct_islow_sse2[];
 419 EXTERN(void) jsimd_fdct_ifast_sse2 JPP((DCTELEM * data));
 420
 421 EXTERN(void) jsimd_fdct_float_3dnow JPP((FAST_FLOAT * data));
 422
 423 extern const int jconst_fdct_float_sse[];
 424 EXTERN(void) jsimd_fdct_float_sse JPP((FAST_FLOAT * data));
 425
 426 /* SIMD Quantization */
 427 EXTERN(void) jsimd_quantize_mmx JPP((JCOEFPTR coef_block,
 428                                      DCTELEM * divisors,
 429                                      DCTELEM * workspace));
 430
 431 EXTERN(void) jsimd_quantize_sse2 JPP((JCOEFPTR coef_block,
 432                                       DCTELEM * divisors,
 433                                       DCTELEM * workspace));
 434
 435 EXTERN(void) jsimd_quantize_float_3dnow JPP((JCOEFPTR coef_block,
 436                                              FAST_FLOAT * divisors,
 437                                              FAST_FLOAT * workspace));
 438
 439 EXTERN(void) jsimd_quantize_float_sse JPP((JCOEFPTR coef_block,
 440                                            FAST_FLOAT * divisors,
 441                                            FAST_FLOAT * workspace));
 442
 443 EXTERN(void) jsimd_quantize_float_sse2 JPP((JCOEFPTR coef_block,
 444                                             FAST_FLOAT * divisors,
 445                                             FAST_FLOAT * workspace));
 446
 447 /* SIMD Reduced Inverse DCT */
 448 EXTERN(void) jsimd_idct_2x2_mmx JPP((void * dct_table,
 449                                      JCOEFPTR coef_block,
 450                                      JSAMPARRAY output_buf,
 451                                      JDIMENSION output_col));
 452 EXTERN(void) jsimd_idct_4x4_mmx JPP((void * dct_table,
 453                                      JCOEFPTR coef_block,
 454                                      JSAMPARRAY output_buf,
 455                                      JDIMENSION output_col));
 456
 457 extern const int jconst_idct_red_sse2[];
 458 EXTERN(void) jsimd_idct_2x2_sse2 JPP((void * dct_table,
 459                                       JCOEFPTR coef_block,
 460                                       JSAMPARRAY output_buf,
 461                                       JDIMENSION output_col));
 462 EXTERN(void) jsimd_idct_4x4_sse2 JPP((void * dct_table,
 463                                       JCOEFPTR coef_block,
 464                                       JSAMPARRAY output_buf,
 465                                       JDIMENSION output_col));
 466
 467 /* SIMD Inverse DCT */
 468 EXTERN(void) jsimd_idct_islow_mmx JPP((void * dct_table,
 469                                        JCOEFPTR coef_block,
 470                                        JSAMPARRAY output_buf,
 471                                        JDIMENSION output_col));
 472 EXTERN(void) jsimd_idct_ifast_mmx JPP((void * dct_table,
 473                                        JCOEFPTR coef_block,
 474                                        JSAMPARRAY output_buf,
 475                                        JDIMENSION output_col));
 476
 477 extern const int jconst_idct_islow_sse2[];
 478 EXTERN(void) jsimd_idct_islow_sse2 JPP((void * dct_table,
 479                                         JCOEFPTR coef_block,
 480                                         JSAMPARRAY output_buf,
 481                                         JDIMENSION output_col));
 482 extern const int jconst_idct_ifast_sse2[];
 483 EXTERN(void) jsimd_idct_ifast_sse2 JPP((void * dct_table,
 484                                         JCOEFPTR coef_block,
 485                                         JSAMPARRAY output_buf,
 486                                         JDIMENSION output_col));
 487
 488 EXTERN(void) jsimd_idct_float_3dnow JPP((void * dct_table,
 489                                          JCOEFPTR coef_block,
 490                                          JSAMPARRAY output_buf,
 491                                          JDIMENSION output_col));
 492
 493 extern const int jconst_idct_float_sse[];
 494 EXTERN(void) jsimd_idct_float_sse JPP((void * dct_table,
 495                                        JCOEFPTR coef_block,
 496                                        JSAMPARRAY output_buf,
 497                                        JDIMENSION output_col));
 498
 499 extern const int jconst_idct_float_sse2[];
 500 EXTERN(void) jsimd_idct_float_sse2 JPP((void * dct_table,
 501                                         JCOEFPTR coef_block,
 502                                         JSAMPARRAY output_buf,
 503                                         JDIMENSION output_col));
 504