| OLD | NEW |
|---|
| (Empty) |
| 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 Softwar
e | |
| 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 <xmmintrin.h> | |
| 23 | |
| 24 #include "qcmsint.h" | |
| 25 | |
| 26 /* pre-shuffled: just load these into XMM reg instead of load-scalar/shufps sequ
ence */ | |
| 27 #define FLOATSCALE (float)(PRECACHE_OUTPUT_SIZE) | |
| 28 #define CLAMPMAXVAL ( ((float) (PRECACHE_OUTPUT_SIZE - 1)) / PRECACHE_OUTPUT_SIZ
E ) | |
| 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_sse1(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 thi
ng | |
| 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 *((__m64 *)&output[0]) = _mm_cvtps_pi32(result); | |
| 112 result = _mm_movehl_ps(result, result); | |
| 113 *((__m64 *)&output[2]) = _mm_cvtps_pi32(result) ; | |
| 114 | |
| 115 /* load for next loop while store completes */ | |
| 116 vec_r = _mm_load_ss(&igtbl_r[src[0]]); | |
| 117 vec_g = _mm_load_ss(&igtbl_g[src[1]]); | |
| 118 vec_b = _mm_load_ss(&igtbl_b[src[2]]); | |
| 119 src += 3; | |
| 120 | |
| 121 /* use calc'd indices to output RGB values */ | |
| 122 dest[r_out] = otdata_r[output[0]]; | |
| 123 dest[1] = otdata_g[output[1]]; | |
| 124 dest[b_out] = otdata_b[output[2]]; | |
| 125 dest += 3; | |
| 126 } | |
| 127 | |
| 128 /* handle final (maybe only) pixel */ | |
| 129 | |
| 130 vec_r = _mm_shuffle_ps(vec_r, vec_r, 0); | |
| 131 vec_g = _mm_shuffle_ps(vec_g, vec_g, 0); | |
| 132 vec_b = _mm_shuffle_ps(vec_b, vec_b, 0); | |
| 133 | |
| 134 vec_r = _mm_mul_ps(vec_r, mat0); | |
| 135 vec_g = _mm_mul_ps(vec_g, mat1); | |
| 136 vec_b = _mm_mul_ps(vec_b, mat2); | |
| 137 | |
| 138 vec_r = _mm_add_ps(vec_r, _mm_add_ps(vec_g, vec_b)); | |
| 139 vec_r = _mm_max_ps(min, vec_r); | |
| 140 vec_r = _mm_min_ps(max, vec_r); | |
| 141 result = _mm_mul_ps(vec_r, scale); | |
| 142 | |
| 143 *((__m64 *)&output[0]) = _mm_cvtps_pi32(result); | |
| 144 result = _mm_movehl_ps(result, result); | |
| 145 *((__m64 *)&output[2]) = _mm_cvtps_pi32(result); | |
| 146 | |
| 147 dest[r_out] = otdata_r[output[0]]; | |
| 148 dest[1] = otdata_g[output[1]]; | |
| 149 dest[b_out] = otdata_b[output[2]]; | |
| 150 | |
| 151 _mm_empty(); | |
| 152 } | |
| 153 | |
| 154 void qcms_transform_data_rgba_out_lut_sse1(qcms_transform *transform, | |
| 155 unsigned char *src, | |
| 156 unsigned char *dest, | |
| 157 size_t length, | |
| 158 qcms_format_type output_format) | |
| 159 { | |
| 160 unsigned int i; | |
| 161 float (*mat)[4] = transform->matrix; | |
| 162 char input_back[32]; | |
| 163 /* Ensure we have a buffer that's 16 byte aligned regardless of the original | |
| 164 * stack alignment. We can't use __attribute__((aligned(16))) or __declspec(
align(32)) | |
| 165 * because they don't work on stack variables. gcc 4.4 does do the right thi
ng | |
| 166 * on x86 but that's too new for us right now. For more info: gcc bug #16660
*/ | |
| 167 float const * input = (float*)(((uintptr_t)&input_back[16]) & ~0xf); | |
| 168 /* share input and output locations to save having to keep the | |
| 169 * locations in separate registers */ | |
| 170 uint32_t const * output = (uint32_t*)input; | |
| 171 | |
| 172 /* deref *transform now to avoid it in loop */ | |
| 173 const float *igtbl_r = transform->input_gamma_table_r; | |
| 174 const float *igtbl_g = transform->input_gamma_table_g; | |
| 175 const float *igtbl_b = transform->input_gamma_table_b; | |
| 176 | |
| 177 /* deref *transform now to avoid it in loop */ | |
| 178 const uint8_t *otdata_r = &transform->output_table_r->data[0]; | |
| 179 const uint8_t *otdata_g = &transform->output_table_g->data[0]; | |
| 180 const uint8_t *otdata_b = &transform->output_table_b->data[0]; | |
| 181 | |
| 182 /* input matrix values never change */ | |
| 183 const __m128 mat0 = _mm_load_ps(mat[0]); | |
| 184 const __m128 mat1 = _mm_load_ps(mat[1]); | |
| 185 const __m128 mat2 = _mm_load_ps(mat[2]); | |
| 186 | |
| 187 /* these values don't change, either */ | |
| 188 const __m128 max = _mm_load_ps(clampMaxValueX4); | |
| 189 const __m128 min = _mm_setzero_ps(); | |
| 190 const __m128 scale = _mm_load_ps(floatScaleX4); | |
| 191 | |
| 192 /* working variables */ | |
| 193 __m128 vec_r, vec_g, vec_b, result; | |
| 194 const int r_out = output_format.r; | |
| 195 const int b_out = output_format.b; | |
| 196 unsigned char alpha; | |
| 197 | |
| 198 /* CYA */ | |
| 199 if (!length) | |
| 200 return; | |
| 201 | |
| 202 /* one pixel is handled outside of the loop */ | |
| 203 length--; | |
| 204 | |
| 205 /* setup for transforming 1st pixel */ | |
| 206 vec_r = _mm_load_ss(&igtbl_r[src[0]]); | |
| 207 vec_g = _mm_load_ss(&igtbl_g[src[1]]); | |
| 208 vec_b = _mm_load_ss(&igtbl_b[src[2]]); | |
| 209 alpha = src[3]; | |
| 210 src += 4; | |
| 211 | |
| 212 /* transform all but final pixel */ | |
| 213 | |
| 214 for (i=0; i<length; i++) | |
| 215 { | |
| 216 /* position values from gamma tables */ | |
| 217 vec_r = _mm_shuffle_ps(vec_r, vec_r, 0); | |
| 218 vec_g = _mm_shuffle_ps(vec_g, vec_g, 0); | |
| 219 vec_b = _mm_shuffle_ps(vec_b, vec_b, 0); | |
| 220 | |
| 221 /* gamma * matrix */ | |
| 222 vec_r = _mm_mul_ps(vec_r, mat0); | |
| 223 vec_g = _mm_mul_ps(vec_g, mat1); | |
| 224 vec_b = _mm_mul_ps(vec_b, mat2); | |
| 225 | |
| 226 /* store alpha for this pixel; load alpha for next */ | |
| 227 dest[3] = alpha; | |
| 228 alpha = src[3]; | |
| 229 | |
| 230 /* crunch, crunch, crunch */ | |
| 231 vec_r = _mm_add_ps(vec_r, _mm_add_ps(vec_g, vec_b)); | |
| 232 vec_r = _mm_max_ps(min, vec_r); | |
| 233 vec_r = _mm_min_ps(max, vec_r); | |
| 234 result = _mm_mul_ps(vec_r, scale); | |
| 235 | |
| 236 /* store calc'd output tables indices */ | |
| 237 *((__m64 *)&output[0]) = _mm_cvtps_pi32(result); | |
| 238 result = _mm_movehl_ps(result, result); | |
| 239 *((__m64 *)&output[2]) = _mm_cvtps_pi32(result); | |
| 240 | |
| 241 /* load gamma values for next loop while store completes */ | |
| 242 vec_r = _mm_load_ss(&igtbl_r[src[0]]); | |
| 243 vec_g = _mm_load_ss(&igtbl_g[src[1]]); | |
| 244 vec_b = _mm_load_ss(&igtbl_b[src[2]]); | |
| 245 src += 4; | |
| 246 | |
| 247 /* use calc'd indices to output RGB values */ | |
| 248 dest[r_out] = otdata_r[output[0]]; | |
| 249 dest[1] = otdata_g[output[1]]; | |
| 250 dest[b_out] = otdata_b[output[2]]; | |
| 251 dest += 4; | |
| 252 } | |
| 253 | |
| 254 /* handle final (maybe only) pixel */ | |
| 255 | |
| 256 vec_r = _mm_shuffle_ps(vec_r, vec_r, 0); | |
| 257 vec_g = _mm_shuffle_ps(vec_g, vec_g, 0); | |
| 258 vec_b = _mm_shuffle_ps(vec_b, vec_b, 0); | |
| 259 | |
| 260 vec_r = _mm_mul_ps(vec_r, mat0); | |
| 261 vec_g = _mm_mul_ps(vec_g, mat1); | |
| 262 vec_b = _mm_mul_ps(vec_b, mat2); | |
| 263 | |
| 264 dest[3] = alpha; | |
| 265 | |
| 266 vec_r = _mm_add_ps(vec_r, _mm_add_ps(vec_g, vec_b)); | |
| 267 vec_r = _mm_max_ps(min, vec_r); | |
| 268 vec_r = _mm_min_ps(max, vec_r); | |
| 269 result = _mm_mul_ps(vec_r, scale); | |
| 270 | |
| 271 *((__m64 *)&output[0]) = _mm_cvtps_pi32(result); | |
| 272 result = _mm_movehl_ps(result, result); | |
| 273 *((__m64 *)&output[2]) = _mm_cvtps_pi32(result); | |
| 274 | |
| 275 dest[r_out] = otdata_r[output[0]]; | |
| 276 dest[1] = otdata_g[output[1]]; | |
| 277 dest[b_out] = otdata_b[output[2]]; | |
| 278 | |
| 279 _mm_empty(); | |
| 280 } | |
| OLD | NEW |
|---|
Chromium Code Reviews
Issue 1438093002:
Remove //third_party/qcms. (Closed)
Base URL: https://github.com/domokit/mojo.git@master