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 { |
| 39 unsigned int i; |
| 40 float (*mat)[4] = transform->matrix; |
| 41 char input_back[32]; |
| 42 /* Ensure we have a buffer that's 16 byte aligned regardless of the original |
| 43 * stack alignment. We can't use __attribute__((aligned(16))) or __declspec(
align(32)) |
| 44 * because they don't work on stack variables. gcc 4.4 does do the right thi
ng |
| 45 * on x86 but that's too new for us right now. For more info: gcc bug #16660
*/ |
| 46 float const * input = (float*)(((uintptr_t)&input_back[16]) & ~0xf); |
| 47 /* share input and output locations to save having to keep the |
| 48 * locations in separate registers */ |
| 49 uint32_t const * output = (uint32_t*)input; |
| 50 |
| 51 /* deref *transform now to avoid it in loop */ |
| 52 const float *igtbl_r = transform->input_gamma_table_r; |
| 53 const float *igtbl_g = transform->input_gamma_table_g; |
| 54 const float *igtbl_b = transform->input_gamma_table_b; |
| 55 |
| 56 /* deref *transform now to avoid it in loop */ |
| 57 const uint8_t *otdata_r = &transform->output_table_r->data[0]; |
| 58 const uint8_t *otdata_g = &transform->output_table_g->data[0]; |
| 59 const uint8_t *otdata_b = &transform->output_table_b->data[0]; |
| 60 |
| 61 /* input matrix values never change */ |
| 62 const __m128 mat0 = _mm_load_ps(mat[0]); |
| 63 const __m128 mat1 = _mm_load_ps(mat[1]); |
| 64 const __m128 mat2 = _mm_load_ps(mat[2]); |
| 65 |
| 66 /* these values don't change, either */ |
| 67 const __m128 max = _mm_load_ps(clampMaxValueX4); |
| 68 const __m128 min = _mm_setzero_ps(); |
| 69 const __m128 scale = _mm_load_ps(floatScaleX4); |
| 70 |
| 71 /* working variables */ |
| 72 __m128 vec_r, vec_g, vec_b, result; |
| 73 |
| 74 /* CYA */ |
| 75 if (!length) |
| 76 return; |
| 77 |
| 78 /* one pixel is handled outside of the loop */ |
| 79 length--; |
| 80 |
| 81 /* setup for transforming 1st pixel */ |
| 82 vec_r = _mm_load_ss(&igtbl_r[src[0]]); |
| 83 vec_g = _mm_load_ss(&igtbl_g[src[1]]); |
| 84 vec_b = _mm_load_ss(&igtbl_b[src[2]]); |
| 85 src += 3; |
| 86 |
| 87 /* transform all but final pixel */ |
| 88 |
| 89 for (i=0; i<length; i++) |
| 90 { |
| 91 /* position values from gamma tables */ |
| 92 vec_r = _mm_shuffle_ps(vec_r, vec_r, 0); |
| 93 vec_g = _mm_shuffle_ps(vec_g, vec_g, 0); |
| 94 vec_b = _mm_shuffle_ps(vec_b, vec_b, 0); |
| 95 |
| 96 /* gamma * matrix */ |
| 97 vec_r = _mm_mul_ps(vec_r, mat0); |
| 98 vec_g = _mm_mul_ps(vec_g, mat1); |
| 99 vec_b = _mm_mul_ps(vec_b, mat2); |
| 100 |
| 101 /* crunch, crunch, crunch */ |
| 102 vec_r = _mm_add_ps(vec_r, _mm_add_ps(vec_g, vec_b)); |
| 103 vec_r = _mm_max_ps(min, vec_r); |
| 104 vec_r = _mm_min_ps(max, vec_r); |
| 105 result = _mm_mul_ps(vec_r, scale); |
| 106 |
| 107 /* store calc'd output tables indices */ |
| 108 *((__m64 *)&output[0]) = _mm_cvtps_pi32(result); |
| 109 result = _mm_movehl_ps(result, result); |
| 110 *((__m64 *)&output[2]) = _mm_cvtps_pi32(result) ; |
| 111 |
| 112 /* load for next loop while store completes */ |
| 113 vec_r = _mm_load_ss(&igtbl_r[src[0]]); |
| 114 vec_g = _mm_load_ss(&igtbl_g[src[1]]); |
| 115 vec_b = _mm_load_ss(&igtbl_b[src[2]]); |
| 116 src += 3; |
| 117 |
| 118 /* use calc'd indices to output RGB values */ |
| 119 dest[0] = otdata_r[output[0]]; |
| 120 dest[1] = otdata_g[output[1]]; |
| 121 dest[2] = otdata_b[output[2]]; |
| 122 dest += 3; |
| 123 } |
| 124 |
| 125 /* handle final (maybe only) pixel */ |
| 126 |
| 127 vec_r = _mm_shuffle_ps(vec_r, vec_r, 0); |
| 128 vec_g = _mm_shuffle_ps(vec_g, vec_g, 0); |
| 129 vec_b = _mm_shuffle_ps(vec_b, vec_b, 0); |
| 130 |
| 131 vec_r = _mm_mul_ps(vec_r, mat0); |
| 132 vec_g = _mm_mul_ps(vec_g, mat1); |
| 133 vec_b = _mm_mul_ps(vec_b, mat2); |
| 134 |
| 135 vec_r = _mm_add_ps(vec_r, _mm_add_ps(vec_g, vec_b)); |
| 136 vec_r = _mm_max_ps(min, vec_r); |
| 137 vec_r = _mm_min_ps(max, vec_r); |
| 138 result = _mm_mul_ps(vec_r, scale); |
| 139 |
| 140 *((__m64 *)&output[0]) = _mm_cvtps_pi32(result); |
| 141 result = _mm_movehl_ps(result, result); |
| 142 *((__m64 *)&output[2]) = _mm_cvtps_pi32(result); |
| 143 |
| 144 dest[0] = otdata_r[output[0]]; |
| 145 dest[1] = otdata_g[output[1]]; |
| 146 dest[2] = otdata_b[output[2]]; |
| 147 |
| 148 _mm_empty(); |
| 149 } |
| 150 |
| 151 void qcms_transform_data_rgba_out_lut_sse1(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 thi
ng |
| 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 __m128 mat0 = _mm_load_ps(mat[0]); |
| 180 const __m128 mat1 = _mm_load_ps(mat[1]); |
| 181 const __m128 mat2 = _mm_load_ps(mat[2]); |
| 182 |
| 183 /* these values don't change, either */ |
| 184 const __m128 max = _mm_load_ps(clampMaxValueX4); |
| 185 const __m128 min = _mm_setzero_ps(); |
| 186 const __m128 scale = _mm_load_ps(floatScaleX4); |
| 187 |
| 188 /* working variables */ |
| 189 __m128 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 = _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 *((__m64 *)&output[0]) = _mm_cvtps_pi32(result); |
| 232 result = _mm_movehl_ps(result, result); |
| 233 *((__m64 *)&output[2]) = _mm_cvtps_pi32(result); |
| 234 |
| 235 /* load gamma values for next loop while store completes */ |
| 236 vec_r = _mm_load_ss(&igtbl_r[src[0]]); |
| 237 vec_g = _mm_load_ss(&igtbl_g[src[1]]); |
| 238 vec_b = _mm_load_ss(&igtbl_b[src[2]]); |
| 239 src += 4; |
| 240 |
| 241 /* use calc'd indices to output RGB values */ |
| 242 dest[0] = otdata_r[output[0]]; |
| 243 dest[1] = otdata_g[output[1]]; |
| 244 dest[2] = otdata_b[output[2]]; |
| 245 dest += 4; |
| 246 } |
| 247 |
| 248 /* handle final (maybe only) pixel */ |
| 249 |
| 250 vec_r = _mm_shuffle_ps(vec_r, vec_r, 0); |
| 251 vec_g = _mm_shuffle_ps(vec_g, vec_g, 0); |
| 252 vec_b = _mm_shuffle_ps(vec_b, vec_b, 0); |
| 253 |
| 254 vec_r = _mm_mul_ps(vec_r, mat0); |
| 255 vec_g = _mm_mul_ps(vec_g, mat1); |
| 256 vec_b = _mm_mul_ps(vec_b, mat2); |
| 257 |
| 258 dest[3] = alpha; |
| 259 |
| 260 vec_r = _mm_add_ps(vec_r, _mm_add_ps(vec_g, vec_b)); |
| 261 vec_r = _mm_max_ps(min, vec_r); |
| 262 vec_r = _mm_min_ps(max, vec_r); |
| 263 result = _mm_mul_ps(vec_r, scale); |
| 264 |
| 265 *((__m64 *)&output[0]) = _mm_cvtps_pi32(result); |
| 266 result = _mm_movehl_ps(result, result); |
| 267 *((__m64 *)&output[2]) = _mm_cvtps_pi32(result); |
| 268 |
| 269 dest[0] = otdata_r[output[0]]; |
| 270 dest[1] = otdata_g[output[1]]; |
| 271 dest[2] = otdata_b[output[2]]; |
| 272 |
| 273 _mm_empty(); |
| 274 } |
OLD | NEW |