gfx/qcms/transform-altivec.c

   1 /* vim: set ts=8 sw=8 noexpandtab: */
   2 //  qcms
   3 //  Copyright (C) 2009 Mozilla Corporation
   4 //  Copyright (C) 1998-2007 Marti Maria
   5 //
   6 // Permission is hereby granted, free of charge, to any person obtaining
   7 // a copy of this software and associated documentation files (the "Software"),
   8 // to deal in the Software without restriction, including without limitation
   9 // the rights to use, copy, modify, merge, publish, distribute, sublicense,
  10 // and/or sell copies of the Software, and to permit persons to whom the Software
  11 // is furnished to do so, subject to the following conditions:
  12 //
  13 // The above copyright notice and this permission notice shall be included in
  14 // all copies or substantial portions of the Software.
  15 //
  16 // THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
  17 // EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO
  18 // THE WARRANTIES OF MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
  19 // NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS BE
  20 // LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION
  21 // OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION
  22 // WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.
  23
  24 #include <altivec.h>
  25
  26 #include "qcmsint.h"
  27
  28 #define FLOATSCALE (float)(PRECACHE_OUTPUT_SIZE)
  29 #define CLAMPMAXVAL (((float) (PRECACHE_OUTPUT_SIZE - 1)) / PRECACHE_OUTPUT_SIZE)
  30 static const ALIGN float floatScaleX4 = FLOATSCALE;
  31 static const ALIGN float clampMaxValueX4 = CLAMPMAXVAL;
  32
  33 inline vector float load_aligned_float(float *dataPtr)
  34 {
  35         vector float data = vec_lde(0, dataPtr);
  36         vector unsigned char moveToStart = vec_lvsl(0, dataPtr);
  37         return vec_perm(data, data, moveToStart);
  38 }
  39
  40 void qcms_transform_data_rgb_out_lut_altivec(qcms_transform *transform,
  41                                              unsigned char *src,
  42                                              unsigned char *dest,
  43                                              size_t length)
  44 {
  45         unsigned int i;
  46         float (*mat)[4] = transform->matrix;
  47         char input_back[32];
  48         /* Ensure we have a buffer that's 16 byte aligned regardless of the original
  49          * stack alignment. We can't use __attribute__((aligned(16))) or __declspec(align(32))
  50          * because they don't work on stack variables. gcc 4.4 does do the right thing
  51          * on x86 but that's too new for us right now. For more info: gcc bug #16660 */
  52         float const *input = (float*)(((uintptr_t)&input_back[16]) & ~0xf);
  53         /* share input and output locations to save having to keep the
  54          * locations in separate registers */
  55         uint32_t const *output = (uint32_t*)input;
  56
  57         /* deref *transform now to avoid it in loop */
  58         const float *igtbl_r = transform->input_gamma_table_r;
  59         const float *igtbl_g = transform->input_gamma_table_g;
  60         const float *igtbl_b = transform->input_gamma_table_b;
  61
  62         /* deref *transform now to avoid it in loop */
  63         const uint8_t *otdata_r = &transform->output_table_r->data[0];
  64         const uint8_t *otdata_g = &transform->output_table_g->data[0];
  65         const uint8_t *otdata_b = &transform->output_table_b->data[0];
  66
  67         /* input matrix values never change */
  68         const vector float mat0 = vec_ldl(0, (vector float*)mat[0]);
  69         const vector float mat1 = vec_ldl(0, (vector float*)mat[1]);
  70         const vector float mat2 = vec_ldl(0, (vector float*)mat[2]);
  71
  72         /* these values don't change, either */
  73         const vector float max = vec_splat(vec_lde(0, (float*)&clampMaxValueX4), 0);
  74         const vector float min = (vector float)vec_splat_u32(0);
  75         const vector float scale = vec_splat(vec_lde(0, (float*)&floatScaleX4), 0);
  76
  77         /* working variables */
  78         vector float vec_r, vec_g, vec_b, result;
  79
  80         /* CYA */
  81         if (!length)
  82                 return;
  83
  84         /* one pixel is handled outside of the loop */
  85         length--;
  86
  87         /* setup for transforming 1st pixel */
  88         vec_r = load_aligned_float((float*)&igtbl_r[src[0]]);
  89         vec_g = load_aligned_float((float*)&igtbl_r[src[1]]);
  90         vec_b = load_aligned_float((float*)&igtbl_r[src[2]]);
  91         src += 3;
  92
  93         /* transform all but final pixel */
  94
  95         for (i=0; i<length; i++)
  96         {
  97                 /* position values from gamma tables */
  98                 vec_r = vec_splat(vec_r, 0);
  99                 vec_g = vec_splat(vec_g, 0);
 100                 vec_b = vec_splat(vec_b, 0);
 101
 102                 /* gamma * matrix */
 103                 vec_r = vec_madd(vec_r, mat0, min);
 104                 vec_g = vec_madd(vec_g, mat1, min);
 105                 vec_b = vec_madd(vec_b, mat2, min);
 106
 107                 /* crunch, crunch, crunch */
 108                 vec_r = vec_add(vec_r, vec_add(vec_g, vec_b));
 109                 vec_r = vec_max(min, vec_r);
 110                 vec_r = vec_min(max, vec_r);
 111                 result = vec_madd(vec_r, scale, min);
 112
 113                 /* store calc'd output tables indices */
 114                 vec_st(vec_ctu(vec_round(result), 0), 0, (vector unsigned int*)output);
 115
 116                 /* load for next loop while store completes */
 117                 vec_r = load_aligned_float((float*)&igtbl_r[src[0]]);
 118                 vec_g = load_aligned_float((float*)&igtbl_r[src[1]]);
 119                 vec_b = load_aligned_float((float*)&igtbl_r[src[2]]);
 120                 src += 3;
 121
 122                 /* use calc'd indices to output RGB values */
 123                 dest[0] = otdata_r[output[0]];
 124                 dest[1] = otdata_g[output[1]];
 125                 dest[2] = otdata_b[output[2]];
 126                 dest += 3;
 127         }
 128
 129         /* handle final (maybe only) pixel */
 130
 131         vec_r = vec_splat(vec_r, 0);
 132         vec_g = vec_splat(vec_g, 0);
 133         vec_b = vec_splat(vec_b, 0);
 134
 135         vec_r = vec_madd(vec_r, mat0, min);
 136         vec_g = vec_madd(vec_g, mat1, min);
 137         vec_b = vec_madd(vec_b, mat2, min);
 138
 139         vec_r = vec_add(vec_r, vec_add(vec_g, vec_b));
 140         vec_r = vec_max(min, vec_r);
 141         vec_r = vec_min(max, vec_r);
 142         result = vec_madd(vec_r, scale, min);
 143
 144         vec_st(vec_ctu(vec_round(result),0),0,(vector unsigned int*)output);
 145
 146         dest[0] = otdata_r[output[0]];
 147         dest[1] = otdata_g[output[1]];
 148         dest[2] = otdata_b[output[2]];
 149 }
 150
 151 void qcms_transform_data_rgba_out_lut_altivec(qcms_transform *transform,
 152                                               unsigned char *src,
 153                                               unsigned char *dest,
 154                                               size_t length)
 155 {
 156         unsigned int i;
 157         float (*mat)[4] = transform->matrix;
 158         char input_back[32];
 159         /* Ensure we have a buffer that's 16 byte aligned regardless of the original
 160          * stack alignment. We can't use __attribute__((aligned(16))) or __declspec(align(32))
 161          * because they don't work on stack variables. gcc 4.4 does do the right thing
 162          * on x86 but that's too new for us right now. For more info: gcc bug #16660 */
 163         float const *input = (float*)(((uintptr_t)&input_back[16]) & ~0xf);
 164         /* share input and output locations to save having to keep the
 165          * locations in separate registers */
 166         uint32_t const *output = (uint32_t*)input;
 167
 168         /* deref *transform now to avoid it in loop */
 169         const float *igtbl_r = transform->input_gamma_table_r;
 170         const float *igtbl_g = transform->input_gamma_table_g;
 171         const float *igtbl_b = transform->input_gamma_table_b;
 172
 173         /* deref *transform now to avoid it in loop */
 174         const uint8_t *otdata_r = &transform->output_table_r->data[0];
 175         const uint8_t *otdata_g = &transform->output_table_g->data[0];
 176         const uint8_t *otdata_b = &transform->output_table_b->data[0];
 177
 178         /* input matrix values never change */
 179         const vector float mat0 = vec_ldl(0, (vector float*)mat[0]);
 180         const vector float mat1 = vec_ldl(0, (vector float*)mat[1]);
 181         const vector float mat2 = vec_ldl(0, (vector float*)mat[2]);
 182
 183         /* these values don't change, either */
 184         const vector float max = vec_splat(vec_lde(0, (float*)&clampMaxValueX4), 0);
 185         const vector float min = (vector float)vec_splat_u32(0);
 186         const vector float scale = vec_splat(vec_lde(0, (float*)&floatScaleX4), 0);
 187
 188         /* working variables */
 189         vector float vec_r, vec_g, vec_b, result;
 190         unsigned char alpha;
 191
 192         /* CYA */
 193         if (!length)
 194                 return;
 195
 196         /* one pixel is handled outside of the loop */
 197         length--;
 198
 199         /* setup for transforming 1st pixel */
 200         vec_r = load_aligned_float((float*)&igtbl_r[src[0]]);
 201         vec_g = load_aligned_float((float*)&igtbl_r[src[1]]);
 202         vec_b = load_aligned_float((float*)&igtbl_r[src[2]]);
 203         alpha = src[3];
 204         src += 4;
 205
 206         /* transform all but final pixel */
 207
 208         for (i=0; i<length; i++)
 209         {
 210                 /* position values from gamma tables */
 211                 vec_r = vec_splat(vec_r, 0);
 212                 vec_g = vec_splat(vec_g, 0);
 213                 vec_b = vec_splat(vec_b, 0);
 214
 215                 /* gamma * matrix */
 216                 vec_r = vec_madd(vec_r, mat0, min);
 217                 vec_g = vec_madd(vec_g, mat1, min);
 218                 vec_b = vec_madd(vec_b, mat2, min);
 219
 220                 /* store alpha for this pixel; load alpha for next */
 221                 dest[3] = alpha;
 222                 alpha = src[3];
 223
 224                 /* crunch, crunch, crunch */
 225                 vec_r = vec_add(vec_r, vec_add(vec_g, vec_b));
 226                 vec_r = vec_max(min, vec_r);
 227                 vec_r = vec_min(max, vec_r);
 228                 result = vec_madd(vec_r, scale, min);
 229
 230                 /* store calc'd output tables indices */
 231                 vec_st(vec_ctu(vec_round(result), 0), 0, (vector unsigned int*)output);
 232
 233                 /* load gamma values for next loop while store completes */
 234                 vec_r = load_aligned_float((float*)&igtbl_r[src[0]]);
 235                 vec_g = load_aligned_float((float*)&igtbl_r[src[1]]);
 236                 vec_b = load_aligned_float((float*)&igtbl_r[src[2]]);
 237                 src += 4;
 238
 239                 /* use calc'd indices to output RGB values */
 240                 dest[0] = otdata_r[output[0]];
 241                 dest[1] = otdata_g[output[1]];
 242                 dest[2] = otdata_b[output[2]];
 243                 dest += 4;
 244         }
 245
 246         /* handle final (maybe only) pixel */
 247
 248         vec_r = vec_splat(vec_r, 0);
 249         vec_g = vec_splat(vec_g, 0);
 250         vec_b = vec_splat(vec_b, 0);
 251
 252         vec_r = vec_madd(vec_r, mat0, min);
 253         vec_g = vec_madd(vec_g, mat1, min);
 254         vec_b = vec_madd(vec_b, mat2, min);
 255
 256         dest[3] = alpha;
 257
 258         vec_r = vec_add(vec_r, vec_add(vec_g, vec_b));
 259         vec_r = vec_max(min, vec_r);
 260         vec_r = vec_min(max, vec_r);
 261         result = vec_madd(vec_r, scale, min);
 262
 263         vec_st(vec_ctu(vec_round(result), 0), 0, (vector unsigned int*)output);
 264
 265         dest[0] = otdata_r[output[0]];
 266         dest[1] = otdata_g[output[1]];
 267         dest[2] = otdata_b[output[2]];
 268 }
 269