third_party/qcms/src/transform-sse2.c

   1 //  qcms
   2 //  Copyright (C) 2009 Mozilla Foundation
   3 //
   4 // Permission is hereby granted, free of charge, to any person obtaining
   5 // a copy of this software and associated documentation files (the "Software"),
   6 // to deal in the Software without restriction, including without limitation
   7 // the rights to use, copy, modify, merge, publish, distribute, sublicense,
   8 // and/or sell copies of the Software, and to permit persons to whom the Software
   9 // is furnished to do so, subject to the following conditions:
  10 //
  11 // The above copyright notice and this permission notice shall be included in
  12 // all copies or substantial portions of the Software.
  13 //
  14 // THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
  15 // EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO
  16 // THE WARRANTIES OF MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
  17 // NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS BE
  18 // LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION
  19 // OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION
  20 // WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.
  21
  22 #include <emmintrin.h>
  23
  24 #include "qcmsint.h"
  25
  26 /* pre-shuffled: just load these into XMM reg instead of load-scalar/shufps sequence */
  27 #define FLOATSCALE  (float)(PRECACHE_OUTPUT_SIZE)
  28 #define CLAMPMAXVAL ( ((float) (PRECACHE_OUTPUT_SIZE - 1)) / PRECACHE_OUTPUT_SIZE )
  29 static const ALIGN float floatScaleX4[4] =
  30     { FLOATSCALE, FLOATSCALE, FLOATSCALE, FLOATSCALE};
  31 static const ALIGN float clampMaxValueX4[4] =
  32     { CLAMPMAXVAL, CLAMPMAXVAL, CLAMPMAXVAL, CLAMPMAXVAL};
  33
  34 void qcms_transform_data_rgb_out_lut_sse2(qcms_transform *transform,
  35                                           unsigned char *src,
  36                                           unsigned char *dest,
  37                                           size_t length,
  38                                           qcms_format_type output_format)
  39 {
  40     unsigned int i;
  41     float (*mat)[4] = transform->matrix;
  42     char input_back[32];
  43     /* Ensure we have a buffer that's 16 byte aligned regardless of the original
  44      * stack alignment. We can't use __attribute__((aligned(16))) or __declspec(align(32))
  45      * because they don't work on stack variables. gcc 4.4 does do the right thing
  46      * on x86 but that's too new for us right now. For more info: gcc bug #16660 */
  47     float const * input = (float*)(((uintptr_t)&input_back[16]) & ~0xf);
  48     /* share input and output locations to save having to keep the
  49      * locations in separate registers */
  50     uint32_t const * output = (uint32_t*)input;
  51
  52     /* deref *transform now to avoid it in loop */
  53     const float *igtbl_r = transform->input_gamma_table_r;
  54     const float *igtbl_g = transform->input_gamma_table_g;
  55     const float *igtbl_b = transform->input_gamma_table_b;
  56
  57     /* deref *transform now to avoid it in loop */
  58     const uint8_t *otdata_r = &transform->output_table_r->data[0];
  59     const uint8_t *otdata_g = &transform->output_table_g->data[0];
  60     const uint8_t *otdata_b = &transform->output_table_b->data[0];
  61
  62     /* input matrix values never change */
  63     const __m128 mat0  = _mm_load_ps(mat[0]);
  64     const __m128 mat1  = _mm_load_ps(mat[1]);
  65     const __m128 mat2  = _mm_load_ps(mat[2]);
  66
  67     /* these values don't change, either */
  68     const __m128 max   = _mm_load_ps(clampMaxValueX4);
  69     const __m128 min   = _mm_setzero_ps();
  70     const __m128 scale = _mm_load_ps(floatScaleX4);
  71
  72     /* working variables */
  73     __m128 vec_r, vec_g, vec_b, result;
  74     const int r_out = output_format.r;
  75     const int b_out = output_format.b;
  76
  77     /* CYA */
  78     if (!length)
  79         return;
  80
  81     /* one pixel is handled outside of the loop */
  82     length--;
  83
  84     /* setup for transforming 1st pixel */
  85     vec_r = _mm_load_ss(&igtbl_r[src[0]]);
  86     vec_g = _mm_load_ss(&igtbl_g[src[1]]);
  87     vec_b = _mm_load_ss(&igtbl_b[src[2]]);
  88     src += 3;
  89
  90     /* transform all but final pixel */
  91
  92     for (i=0; i<length; i++)
  93     {
  94         /* position values from gamma tables */
  95         vec_r = _mm_shuffle_ps(vec_r, vec_r, 0);
  96         vec_g = _mm_shuffle_ps(vec_g, vec_g, 0);
  97         vec_b = _mm_shuffle_ps(vec_b, vec_b, 0);
  98
  99         /* gamma * matrix */
 100         vec_r = _mm_mul_ps(vec_r, mat0);
 101         vec_g = _mm_mul_ps(vec_g, mat1);
 102         vec_b = _mm_mul_ps(vec_b, mat2);
 103
 104         /* crunch, crunch, crunch */
 105         vec_r  = _mm_add_ps(vec_r, _mm_add_ps(vec_g, vec_b));
 106         vec_r  = _mm_max_ps(min, vec_r);
 107         vec_r  = _mm_min_ps(max, vec_r);
 108         result = _mm_mul_ps(vec_r, scale);
 109
 110         /* store calc'd output tables indices */
 111         _mm_store_si128((__m128i*)output, _mm_cvtps_epi32(result));
 112
 113         /* load for next loop while store completes */
 114         vec_r = _mm_load_ss(&igtbl_r[src[0]]);
 115         vec_g = _mm_load_ss(&igtbl_g[src[1]]);
 116         vec_b = _mm_load_ss(&igtbl_b[src[2]]);
 117         src += 3;
 118
 119         /* use calc'd indices to output RGB values */
 120         dest[r_out] = otdata_r[output[0]];
 121         dest[1]     = otdata_g[output[1]];
 122         dest[b_out] = otdata_b[output[2]];
 123         dest += 3;
 124     }
 125
 126     /* handle final (maybe only) pixel */
 127
 128     vec_r = _mm_shuffle_ps(vec_r, vec_r, 0);
 129     vec_g = _mm_shuffle_ps(vec_g, vec_g, 0);
 130     vec_b = _mm_shuffle_ps(vec_b, vec_b, 0);
 131
 132     vec_r = _mm_mul_ps(vec_r, mat0);
 133     vec_g = _mm_mul_ps(vec_g, mat1);
 134     vec_b = _mm_mul_ps(vec_b, mat2);
 135
 136     vec_r  = _mm_add_ps(vec_r, _mm_add_ps(vec_g, vec_b));
 137     vec_r  = _mm_max_ps(min, vec_r);
 138     vec_r  = _mm_min_ps(max, vec_r);
 139     result = _mm_mul_ps(vec_r, scale);
 140
 141     _mm_store_si128((__m128i*)output, _mm_cvtps_epi32(result));
 142
 143     dest[r_out] = otdata_r[output[0]];
 144     dest[1]     = otdata_g[output[1]];
 145     dest[b_out] = otdata_b[output[2]];
 146 }
 147
 148 void qcms_transform_data_rgba_out_lut_sse2(qcms_transform *transform,
 149                                            unsigned char *src,
 150                                            unsigned char *dest,
 151                                            size_t length,
 152                                            qcms_format_type output_format)
 153 {
 154     unsigned int i;
 155     float (*mat)[4] = transform->matrix;
 156     char input_back[32];
 157     /* Ensure we have a buffer that's 16 byte aligned regardless of the original
 158      * stack alignment. We can't use __attribute__((aligned(16))) or __declspec(align(32))
 159      * because they don't work on stack variables. gcc 4.4 does do the right thing
 160      * on x86 but that's too new for us right now. For more info: gcc bug #16660 */
 161     float const * input = (float*)(((uintptr_t)&input_back[16]) & ~0xf);
 162     /* share input and output locations to save having to keep the
 163      * locations in separate registers */
 164     uint32_t const * output = (uint32_t*)input;
 165
 166     /* deref *transform now to avoid it in loop */
 167     const float *igtbl_r = transform->input_gamma_table_r;
 168     const float *igtbl_g = transform->input_gamma_table_g;
 169     const float *igtbl_b = transform->input_gamma_table_b;
 170
 171     /* deref *transform now to avoid it in loop */
 172     const uint8_t *otdata_r = &transform->output_table_r->data[0];
 173     const uint8_t *otdata_g = &transform->output_table_g->data[0];
 174     const uint8_t *otdata_b = &transform->output_table_b->data[0];
 175
 176     /* input matrix values never change */
 177     const __m128 mat0  = _mm_load_ps(mat[0]);
 178     const __m128 mat1  = _mm_load_ps(mat[1]);
 179     const __m128 mat2  = _mm_load_ps(mat[2]);
 180
 181     /* these values don't change, either */
 182     const __m128 max   = _mm_load_ps(clampMaxValueX4);
 183     const __m128 min   = _mm_setzero_ps();
 184     const __m128 scale = _mm_load_ps(floatScaleX4);
 185
 186     /* working variables */
 187     __m128 vec_r, vec_g, vec_b, result;
 188     const int r_out = output_format.r;
 189     const int b_out = output_format.b;
 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 = _mm_load_ss(&igtbl_r[src[0]]);
 201     vec_g = _mm_load_ss(&igtbl_g[src[1]]);
 202     vec_b = _mm_load_ss(&igtbl_b[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 = _mm_shuffle_ps(vec_r, vec_r, 0);
 212         vec_g = _mm_shuffle_ps(vec_g, vec_g, 0);
 213         vec_b = _mm_shuffle_ps(vec_b, vec_b, 0);
 214
 215         /* gamma * matrix */
 216         vec_r = _mm_mul_ps(vec_r, mat0);
 217         vec_g = _mm_mul_ps(vec_g, mat1);
 218         vec_b = _mm_mul_ps(vec_b, mat2);
 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  = _mm_add_ps(vec_r, _mm_add_ps(vec_g, vec_b));
 226         vec_r  = _mm_max_ps(min, vec_r);
 227         vec_r  = _mm_min_ps(max, vec_r);
 228         result = _mm_mul_ps(vec_r, scale);
 229
 230         /* store calc'd output tables indices */
 231         _mm_store_si128((__m128i*)output, _mm_cvtps_epi32(result));
 232
 233         /* load gamma values for next loop while store completes */
 234         vec_r = _mm_load_ss(&igtbl_r[src[0]]);
 235         vec_g = _mm_load_ss(&igtbl_g[src[1]]);
 236         vec_b = _mm_load_ss(&igtbl_b[src[2]]);
 237         src += 4;
 238
 239         /* use calc'd indices to output RGB values */
 240         dest[r_out] = otdata_r[output[0]];
 241         dest[1]     = otdata_g[output[1]];
 242         dest[b_out] = otdata_b[output[2]];
 243         dest += 4;
 244     }
 245
 246     /* handle final (maybe only) pixel */
 247
 248     vec_r = _mm_shuffle_ps(vec_r, vec_r, 0);
 249     vec_g = _mm_shuffle_ps(vec_g, vec_g, 0);
 250     vec_b = _mm_shuffle_ps(vec_b, vec_b, 0);
 251
 252     vec_r = _mm_mul_ps(vec_r, mat0);
 253     vec_g = _mm_mul_ps(vec_g, mat1);
 254     vec_b = _mm_mul_ps(vec_b, mat2);
 255
 256     dest[3] = alpha;
 257
 258     vec_r  = _mm_add_ps(vec_r, _mm_add_ps(vec_g, vec_b));
 259     vec_r  = _mm_max_ps(min, vec_r);
 260     vec_r  = _mm_min_ps(max, vec_r);
 261     result = _mm_mul_ps(vec_r, scale);
 262
 263     _mm_store_si128((__m128i*)output, _mm_cvtps_epi32(result));
 264
 265     dest[r_out] = otdata_r[output[0]];
 266     dest[1]     = otdata_g[output[1]];
 267     dest[b_out] = otdata_b[output[2]];
 268 }