gfx/qcms/transform-sse1.c

   1 #include <xmmintrin.h>
   2
   3 #include "qcmsint.h"
   4
   5 /* pre-shuffled: just load these into XMM reg instead of load-scalar/shufps sequence */
   6 #define FLOATSCALE  (float)(PRECACHE_OUTPUT_SIZE)
   7 #define CLAMPMAXVAL ( ((float) (PRECACHE_OUTPUT_SIZE - 1)) / PRECACHE_OUTPUT_SIZE )
   8 static const ALIGN float floatScaleX4[4] =
   9     { FLOATSCALE, FLOATSCALE, FLOATSCALE, FLOATSCALE};
  10 static const ALIGN float clampMaxValueX4[4] =
  11     { CLAMPMAXVAL, CLAMPMAXVAL, CLAMPMAXVAL, CLAMPMAXVAL};
  12
  13 void qcms_transform_data_rgb_out_lut_sse1(qcms_transform *transform,
  14                                           unsigned char *src,
  15                                           unsigned char *dest,
  16                                           size_t length)
  17 {
  18     unsigned int i;
  19     float (*mat)[4] = transform->matrix;
  20     char input_back[32];
  21     /* Ensure we have a buffer that's 16 byte aligned regardless of the original
  22      * stack alignment. We can't use __attribute__((aligned(16))) or __declspec(align(32))
  23      * because they don't work on stack variables. gcc 4.4 does do the right thing
  24      * on x86 but that's too new for us right now. For more info: gcc bug #16660 */
  25     float const * input = (float*)(((uintptr_t)&input_back[16]) & ~0xf);
  26     /* share input and output locations to save having to keep the
  27      * locations in separate registers */
  28     uint32_t const * output = (uint32_t*)input;
  29
  30     /* deref *transform now to avoid it in loop */
  31     const float *igtbl_r = transform->input_gamma_table_r;
  32     const float *igtbl_g = transform->input_gamma_table_g;
  33     const float *igtbl_b = transform->input_gamma_table_b;
  34
  35     /* deref *transform now to avoid it in loop */
  36     const uint8_t *otdata_r = &transform->output_table_r->data[0];
  37     const uint8_t *otdata_g = &transform->output_table_g->data[0];
  38     const uint8_t *otdata_b = &transform->output_table_b->data[0];
  39
  40     /* input matrix values never change */
  41     const __m128 mat0  = _mm_load_ps(mat[0]);
  42     const __m128 mat1  = _mm_load_ps(mat[1]);
  43     const __m128 mat2  = _mm_load_ps(mat[2]);
  44
  45     /* these values don't change, either */
  46     const __m128 max   = _mm_load_ps(clampMaxValueX4);
  47     const __m128 min   = _mm_setzero_ps();
  48     const __m128 scale = _mm_load_ps(floatScaleX4);
  49
  50     /* working variables */
  51     __m128 vec_r, vec_g, vec_b, result;
  52
  53     /* CYA */
  54     if (!length)
  55         return;
  56
  57     /* one pixel is handled outside of the loop */
  58     length--;
  59
  60     /* setup for transforming 1st pixel */
  61     vec_r = _mm_load_ss(&igtbl_r[src[0]]);
  62     vec_g = _mm_load_ss(&igtbl_g[src[1]]);
  63     vec_b = _mm_load_ss(&igtbl_b[src[2]]);
  64     src += 3;
  65
  66     /* transform all but final pixel */
  67
  68     for (i=0; i<length; i++)
  69     {
  70         /* position values from gamma tables */
  71         vec_r = _mm_shuffle_ps(vec_r, vec_r, 0);
  72         vec_g = _mm_shuffle_ps(vec_g, vec_g, 0);
  73         vec_b = _mm_shuffle_ps(vec_b, vec_b, 0);
  74
  75         /* gamma * matrix */
  76         vec_r = _mm_mul_ps(vec_r, mat0);
  77         vec_g = _mm_mul_ps(vec_g, mat1);
  78         vec_b = _mm_mul_ps(vec_b, mat2);
  79
  80         /* crunch, crunch, crunch */
  81         vec_r  = _mm_add_ps(vec_r, _mm_add_ps(vec_g, vec_b));
  82         vec_r  = _mm_max_ps(min, vec_r);
  83         vec_r  = _mm_min_ps(max, vec_r);
  84         result = _mm_mul_ps(vec_r, scale);
  85
  86         /* store calc'd output tables indices */
  87         *((__m64 *)&output[0]) = _mm_cvtps_pi32(result);
  88         result = _mm_movehl_ps(result, result);
  89         *((__m64 *)&output[2]) = _mm_cvtps_pi32(result) ;
  90
  91         /* load for next loop while store completes */
  92         vec_r = _mm_load_ss(&igtbl_r[src[0]]);
  93         vec_g = _mm_load_ss(&igtbl_g[src[1]]);
  94         vec_b = _mm_load_ss(&igtbl_b[src[2]]);
  95         src += 3;
  96
  97         /* use calc'd indices to output RGB values */
  98         dest[0] = otdata_r[output[0]];
  99         dest[1] = otdata_g[output[1]];
 100         dest[2] = otdata_b[output[2]];
 101         dest += 3;
 102     }
 103
 104     /* handle final (maybe only) pixel */
 105
 106     vec_r = _mm_shuffle_ps(vec_r, vec_r, 0);
 107     vec_g = _mm_shuffle_ps(vec_g, vec_g, 0);
 108     vec_b = _mm_shuffle_ps(vec_b, vec_b, 0);
 109
 110     vec_r = _mm_mul_ps(vec_r, mat0);
 111     vec_g = _mm_mul_ps(vec_g, mat1);
 112     vec_b = _mm_mul_ps(vec_b, mat2);
 113
 114     vec_r  = _mm_add_ps(vec_r, _mm_add_ps(vec_g, vec_b));
 115     vec_r  = _mm_max_ps(min, vec_r);
 116     vec_r  = _mm_min_ps(max, vec_r);
 117     result = _mm_mul_ps(vec_r, scale);
 118
 119     *((__m64 *)&output[0]) = _mm_cvtps_pi32(result);
 120     result = _mm_movehl_ps(result, result);
 121     *((__m64 *)&output[2]) = _mm_cvtps_pi32(result);
 122
 123     dest[0] = otdata_r[output[0]];
 124     dest[1] = otdata_g[output[1]];
 125     dest[2] = otdata_b[output[2]];
 126
 127     _mm_empty();
 128 }
 129
 130 void qcms_transform_data_rgba_out_lut_sse1(qcms_transform *transform,
 131                                            unsigned char *src,
 132                                            unsigned char *dest,
 133                                            size_t length)
 134 {
 135     unsigned int i;
 136     float (*mat)[4] = transform->matrix;
 137     char input_back[32];
 138     /* Ensure we have a buffer that's 16 byte aligned regardless of the original
 139      * stack alignment. We can't use __attribute__((aligned(16))) or __declspec(align(32))
 140      * because they don't work on stack variables. gcc 4.4 does do the right thing
 141      * on x86 but that's too new for us right now. For more info: gcc bug #16660 */
 142     float const * input = (float*)(((uintptr_t)&input_back[16]) & ~0xf);
 143     /* share input and output locations to save having to keep the
 144      * locations in separate registers */
 145     uint32_t const * output = (uint32_t*)input;
 146
 147     /* deref *transform now to avoid it in loop */
 148     const float *igtbl_r = transform->input_gamma_table_r;
 149     const float *igtbl_g = transform->input_gamma_table_g;
 150     const float *igtbl_b = transform->input_gamma_table_b;
 151
 152     /* deref *transform now to avoid it in loop */
 153     const uint8_t *otdata_r = &transform->output_table_r->data[0];
 154     const uint8_t *otdata_g = &transform->output_table_g->data[0];
 155     const uint8_t *otdata_b = &transform->output_table_b->data[0];
 156
 157     /* input matrix values never change */
 158     const __m128 mat0  = _mm_load_ps(mat[0]);
 159     const __m128 mat1  = _mm_load_ps(mat[1]);
 160     const __m128 mat2  = _mm_load_ps(mat[2]);
 161
 162     /* these values don't change, either */
 163     const __m128 max   = _mm_load_ps(clampMaxValueX4);
 164     const __m128 min   = _mm_setzero_ps();
 165     const __m128 scale = _mm_load_ps(floatScaleX4);
 166
 167     /* working variables */
 168     __m128 vec_r, vec_g, vec_b, result;
 169     unsigned char alpha;
 170
 171     /* CYA */
 172     if (!length)
 173         return;
 174
 175     /* one pixel is handled outside of the loop */
 176     length--;
 177
 178     /* setup for transforming 1st pixel */
 179     vec_r = _mm_load_ss(&igtbl_r[src[0]]);
 180     vec_g = _mm_load_ss(&igtbl_g[src[1]]);
 181     vec_b = _mm_load_ss(&igtbl_b[src[2]]);
 182     alpha = src[3];
 183     src += 4;
 184
 185     /* transform all but final pixel */
 186
 187     for (i=0; i<length; i++)
 188     {
 189         /* position values from gamma tables */
 190         vec_r = _mm_shuffle_ps(vec_r, vec_r, 0);
 191         vec_g = _mm_shuffle_ps(vec_g, vec_g, 0);
 192         vec_b = _mm_shuffle_ps(vec_b, vec_b, 0);
 193
 194         /* gamma * matrix */
 195         vec_r = _mm_mul_ps(vec_r, mat0);
 196         vec_g = _mm_mul_ps(vec_g, mat1);
 197         vec_b = _mm_mul_ps(vec_b, mat2);
 198
 199         /* store alpha for this pixel; load alpha for next */
 200         dest[3] = alpha;
 201         alpha   = src[3];
 202
 203         /* crunch, crunch, crunch */
 204         vec_r  = _mm_add_ps(vec_r, _mm_add_ps(vec_g, vec_b));
 205         vec_r  = _mm_max_ps(min, vec_r);
 206         vec_r  = _mm_min_ps(max, vec_r);
 207         result = _mm_mul_ps(vec_r, scale);
 208
 209         /* store calc'd output tables indices */
 210         *((__m64 *)&output[0]) = _mm_cvtps_pi32(result);
 211         result = _mm_movehl_ps(result, result);
 212         *((__m64 *)&output[2]) = _mm_cvtps_pi32(result);
 213
 214         /* load gamma values for next loop while store completes */
 215         vec_r = _mm_load_ss(&igtbl_r[src[0]]);
 216         vec_g = _mm_load_ss(&igtbl_g[src[1]]);
 217         vec_b = _mm_load_ss(&igtbl_b[src[2]]);
 218         src += 4;
 219
 220         /* use calc'd indices to output RGB values */
 221         dest[0] = otdata_r[output[0]];
 222         dest[1] = otdata_g[output[1]];
 223         dest[2] = otdata_b[output[2]];
 224         dest += 4;
 225     }
 226
 227     /* handle final (maybe only) pixel */
 228
 229     vec_r = _mm_shuffle_ps(vec_r, vec_r, 0);
 230     vec_g = _mm_shuffle_ps(vec_g, vec_g, 0);
 231     vec_b = _mm_shuffle_ps(vec_b, vec_b, 0);
 232
 233     vec_r = _mm_mul_ps(vec_r, mat0);
 234     vec_g = _mm_mul_ps(vec_g, mat1);
 235     vec_b = _mm_mul_ps(vec_b, mat2);
 236
 237     dest[3] = alpha;
 238
 239     vec_r  = _mm_add_ps(vec_r, _mm_add_ps(vec_g, vec_b));
 240     vec_r  = _mm_max_ps(min, vec_r);
 241     vec_r  = _mm_min_ps(max, vec_r);
 242     result = _mm_mul_ps(vec_r, scale);
 243
 244     *((__m64 *)&output[0]) = _mm_cvtps_pi32(result);
 245     result = _mm_movehl_ps(result, result);
 246     *((__m64 *)&output[2]) = _mm_cvtps_pi32(result);
 247
 248     dest[0] = otdata_r[output[0]];
 249     dest[1] = otdata_g[output[1]];
 250     dest[2] = otdata_b[output[2]];
 251
 252     _mm_empty();
 253 }