| File: | root/firefox-clang/gfx/skia/skia/modules/skcms/skcms.cc |
| Warning: | line 179, column 32 The right operand of '+' is a garbage value |
Press '?' to see keyboard shortcuts
Keyboard shortcuts:
| 1 | /* | ||||||
| 2 | * Copyright 2018 Google Inc. | ||||||
| 3 | * | ||||||
| 4 | * Use of this source code is governed by a BSD-style license that can be | ||||||
| 5 | * found in the LICENSE file. | ||||||
| 6 | */ | ||||||
| 7 | |||||||
| 8 | #include "src/skcms_public.h" // NO_G3_REWRITE | ||||||
| 9 | #include "src/skcms_internals.h" // NO_G3_REWRITE | ||||||
| 10 | #include "src/skcms_Transform.h" // NO_G3_REWRITE | ||||||
| 11 | #include <assert.h> | ||||||
| 12 | #include <float.h> | ||||||
| 13 | #include <limits.h> | ||||||
| 14 | #include <stdlib.h> | ||||||
| 15 | #include <string.h> | ||||||
| 16 | |||||||
| 17 | #if defined(__ARM_NEON) | ||||||
| 18 | #include <arm_neon.h> | ||||||
| 19 | #elif defined(__SSE__1) | ||||||
| 20 | #include <immintrin.h> | ||||||
| 21 | |||||||
| 22 | #if defined(__clang__1) | ||||||
| 23 | // That #include <immintrin.h> is usually enough, but Clang's headers | ||||||
| 24 | // "helpfully" skip including the whole kitchen sink when _MSC_VER is | ||||||
| 25 | // defined, because lots of programs on Windows would include that and | ||||||
| 26 | // it'd be a lot slower. But we want all those headers included so we | ||||||
| 27 | // can use their features after runtime checks later. | ||||||
| 28 | #include <smmintrin.h> | ||||||
| 29 | #include <avxintrin.h> | ||||||
| 30 | #include <avx2intrin.h> | ||||||
| 31 | #include <avx512fintrin.h> | ||||||
| 32 | #include <avx512dqintrin.h> | ||||||
| 33 | #endif | ||||||
| 34 | #endif | ||||||
| 35 | |||||||
| 36 | using namespace skcms_private; | ||||||
| 37 | |||||||
| 38 | static bool sAllowRuntimeCPUDetection = true; | ||||||
| 39 | |||||||
| 40 | void skcms_DisableRuntimeCPUDetection() { | ||||||
| 41 | sAllowRuntimeCPUDetection = false; | ||||||
| 42 | } | ||||||
| 43 | |||||||
| 44 | static float log2f_(float x) { | ||||||
| 45 | // The first approximation of log2(x) is its exponent 'e', minus 127. | ||||||
| 46 | int32_t bits; | ||||||
| 47 | memcpy(&bits, &x, sizeof(bits)); | ||||||
| 48 | |||||||
| 49 | float e = (float)bits * (1.0f / (1<<23)); | ||||||
| 50 | |||||||
| 51 | // If we use the mantissa too we can refine the error signficantly. | ||||||
| 52 | int32_t m_bits = (bits & 0x007fffff) | 0x3f000000; | ||||||
| 53 | float m; | ||||||
| 54 | memcpy(&m, &m_bits, sizeof(m)); | ||||||
| 55 | |||||||
| 56 | return (e - 124.225514990f | ||||||
| 57 | - 1.498030302f*m | ||||||
| 58 | - 1.725879990f/(0.3520887068f + m)); | ||||||
| 59 | } | ||||||
| 60 | static float logf_(float x) { | ||||||
| 61 | const float ln2 = 0.69314718f; | ||||||
| 62 | return ln2*log2f_(x); | ||||||
| 63 | } | ||||||
| 64 | |||||||
| 65 | static float exp2f_(float x) { | ||||||
| 66 | if (x > 128.0f) { | ||||||
| 67 | return INFINITY_; | ||||||
| 68 | } else if (x < -127.0f) { | ||||||
| 69 | return 0.0f; | ||||||
| 70 | } | ||||||
| 71 | float fract = x - floorf_(x); | ||||||
| 72 | |||||||
| 73 | float fbits = (1.0f * (1<<23)) * (x + 121.274057500f | ||||||
| 74 | - 1.490129070f*fract | ||||||
| 75 | + 27.728023300f/(4.84252568f - fract)); | ||||||
| 76 | |||||||
| 77 | // Before we cast fbits to int32_t, check for out of range values to pacify UBSAN. | ||||||
| 78 | // INT_MAX is not exactly representable as a float, so exclude it as effectively infinite. | ||||||
| 79 | // Negative values are effectively underflow - we'll end up returning a (different) negative | ||||||
| 80 | // value, which makes no sense. So clamp to zero. | ||||||
| 81 | if (fbits >= (float)INT_MAX2147483647) { | ||||||
| 82 | return INFINITY_; | ||||||
| 83 | } else if (fbits < 0) { | ||||||
| 84 | return 0; | ||||||
| 85 | } | ||||||
| 86 | |||||||
| 87 | int32_t bits = (int32_t)fbits; | ||||||
| 88 | memcpy(&x, &bits, sizeof(x)); | ||||||
| 89 | return x; | ||||||
| 90 | } | ||||||
| 91 | |||||||
| 92 | // Not static, as it's used by some test tools. | ||||||
| 93 | float powf_(float x, float y) { | ||||||
| 94 | if (x <= 0.f) { | ||||||
| 95 | return 0.f; | ||||||
| 96 | } | ||||||
| 97 | if (x == 1.f) { | ||||||
| 98 | return 1.f; | ||||||
| 99 | } | ||||||
| 100 | return exp2f_(log2f_(x) * y); | ||||||
| 101 | } | ||||||
| 102 | |||||||
| 103 | static float expf_(float x) { | ||||||
| 104 | const float log2_e = 1.4426950408889634074f; | ||||||
| 105 | return exp2f_(log2_e * x); | ||||||
| 106 | } | ||||||
| 107 | |||||||
| 108 | static float fmaxf_(float x, float y) { return x > y ? x : y; } | ||||||
| 109 | static float fminf_(float x, float y) { return x < y ? x : y; } | ||||||
| 110 | |||||||
| 111 | static bool isfinitef_(float x) { return 0 == x*0; } | ||||||
| 112 | |||||||
| 113 | static float minus_1_ulp(float x) { | ||||||
| 114 | int32_t bits; | ||||||
| 115 | memcpy(&bits, &x, sizeof(bits)); | ||||||
| 116 | bits = bits - 1; | ||||||
| 117 | memcpy(&x, &bits, sizeof(bits)); | ||||||
| 118 | return x; | ||||||
| 119 | } | ||||||
| 120 | |||||||
| 121 | // Most transfer functions we work with are sRGBish. | ||||||
| 122 | // For exotic HDR transfer functions, we encode them using a tf.g that makes no sense, | ||||||
| 123 | // and repurpose the other fields to hold the parameters of the HDR functions. | ||||||
| 124 | struct TF_PQish { float A,B,C,D,E,F; }; | ||||||
| 125 | struct TF_HLGish { float R,G,a,b,c,K_minus_1; }; | ||||||
| 126 | // We didn't originally support a scale factor K for HLG, and instead just stored 0 in | ||||||
| 127 | // the unused `f` field of skcms_TransferFunction for HLGish and HLGInvish transfer functions. | ||||||
| 128 | // By storing f=K-1, those old unusued f=0 values now mean K=1, a noop scale factor. | ||||||
| 129 | |||||||
| 130 | static float TFKind_marker(skcms_TFType kind) { | ||||||
| 131 | // We'd use different NaNs, but those aren't guaranteed to be preserved by WASM. | ||||||
| 132 | return -(float)kind; | ||||||
| 133 | } | ||||||
| 134 | |||||||
| 135 | static skcms_TFType classify(const skcms_TransferFunction& tf, TF_PQish* pq = nullptr | ||||||
| 136 | , TF_HLGish* hlg = nullptr) { | ||||||
| 137 | if (tf.g < 0) { | ||||||
| 138 | // Negative "g" is mapped to enum values; large negative are for sure invalid. | ||||||
| 139 | if (tf.g < -128) { | ||||||
| 140 | return skcms_TFType_Invalid; | ||||||
| 141 | } | ||||||
| 142 | int enum_g = -static_cast<int>(tf.g); | ||||||
| 143 | // Non-whole "g" values are invalid as well. | ||||||
| 144 | if (static_cast<float>(-enum_g) != tf.g) { | ||||||
| 145 | return skcms_TFType_Invalid; | ||||||
| 146 | } | ||||||
| 147 | // TODO: soundness checks for PQ/HLG like we do for sRGBish? | ||||||
| 148 | switch (enum_g) { | ||||||
| 149 | case skcms_TFType_PQish: | ||||||
| 150 | if (pq) { | ||||||
| 151 | memcpy(pq , &tf.a, sizeof(*pq )); | ||||||
| 152 | } | ||||||
| 153 | return skcms_TFType_PQish; | ||||||
| 154 | case skcms_TFType_HLGish: | ||||||
| 155 | if (hlg) { | ||||||
| 156 | memcpy(hlg, &tf.a, sizeof(*hlg)); | ||||||
| 157 | } | ||||||
| 158 | return skcms_TFType_HLGish; | ||||||
| 159 | case skcms_TFType_HLGinvish: | ||||||
| 160 | if (hlg) { | ||||||
| 161 | memcpy(hlg, &tf.a, sizeof(*hlg)); | ||||||
| 162 | } | ||||||
| 163 | return skcms_TFType_HLGinvish; | ||||||
| 164 | case skcms_TFType_PQ: | ||||||
| 165 | if (tf.b != 0.f || tf.c != 0.f || tf.d != 0.f || tf.e != 0.f || tf.f != 0.f) { | ||||||
| 166 | return skcms_TFType_Invalid; | ||||||
| 167 | } | ||||||
| 168 | return skcms_TFType_PQ; | ||||||
| 169 | case skcms_TFType_HLG: | ||||||
| 170 | if (tf.d != 0.f || tf.e != 0.f || tf.f != 0.f) { | ||||||
| 171 | return skcms_TFType_Invalid; | ||||||
| 172 | } | ||||||
| 173 | return skcms_TFType_HLG; | ||||||
| 174 | } | ||||||
| 175 | return skcms_TFType_Invalid; | ||||||
| 176 | } | ||||||
| 177 | |||||||
| 178 | // Basic soundness checks for sRGBish transfer functions. | ||||||
| 179 | if (isfinitef_(tf.a + tf.b + tf.c + tf.d + tf.e + tf.f + tf.g) | ||||||
| |||||||
| 180 | // a,c,d,g should be non-negative to make any sense. | ||||||
| 181 | && tf.a >= 0 | ||||||
| 182 | && tf.c >= 0 | ||||||
| 183 | && tf.d >= 0 | ||||||
| 184 | && tf.g >= 0 | ||||||
| 185 | // Raising a negative value to a fractional tf->g produces complex numbers. | ||||||
| 186 | && tf.a * tf.d + tf.b >= 0) { | ||||||
| 187 | return skcms_TFType_sRGBish; | ||||||
| 188 | } | ||||||
| 189 | |||||||
| 190 | return skcms_TFType_Invalid; | ||||||
| 191 | } | ||||||
| 192 | |||||||
| 193 | skcms_TFType skcms_TransferFunction_getType(const skcms_TransferFunction* tf) { | ||||||
| 194 | return classify(*tf); | ||||||
| 195 | } | ||||||
| 196 | bool skcms_TransferFunction_isSRGBish(const skcms_TransferFunction* tf) { | ||||||
| 197 | return classify(*tf) == skcms_TFType_sRGBish; | ||||||
| 198 | } | ||||||
| 199 | bool skcms_TransferFunction_isPQish(const skcms_TransferFunction* tf) { | ||||||
| 200 | return classify(*tf) == skcms_TFType_PQish; | ||||||
| 201 | } | ||||||
| 202 | bool skcms_TransferFunction_isHLGish(const skcms_TransferFunction* tf) { | ||||||
| 203 | return classify(*tf) == skcms_TFType_HLGish; | ||||||
| 204 | } | ||||||
| 205 | bool skcms_TransferFunction_isPQ(const skcms_TransferFunction* tf) { | ||||||
| 206 | return classify(*tf) == skcms_TFType_PQ; | ||||||
| 207 | } | ||||||
| 208 | bool skcms_TransferFunction_isHLG(const skcms_TransferFunction* tf) { | ||||||
| 209 | return classify(*tf) == skcms_TFType_HLG; | ||||||
| 210 | } | ||||||
| 211 | |||||||
| 212 | bool skcms_TransferFunction_makePQish(skcms_TransferFunction* tf, | ||||||
| 213 | float A, float B, float C, | ||||||
| 214 | float D, float E, float F) { | ||||||
| 215 | *tf = { TFKind_marker(skcms_TFType_PQish), A,B,C,D,E,F }; | ||||||
| 216 | assert(skcms_TransferFunction_isPQish(tf))(static_cast <bool> (skcms_TransferFunction_isPQish(tf) ) ? void (0) : __assert_fail ("skcms_TransferFunction_isPQish(tf)" , __builtin_FILE (), __builtin_LINE (), __extension__ __PRETTY_FUNCTION__ )); | ||||||
| 217 | return true; | ||||||
| 218 | } | ||||||
| 219 | |||||||
| 220 | bool skcms_TransferFunction_makeScaledHLGish(skcms_TransferFunction* tf, | ||||||
| 221 | float K, float R, float G, | ||||||
| 222 | float a, float b, float c) { | ||||||
| 223 | *tf = { TFKind_marker(skcms_TFType_HLGish), R,G, a,b,c, K-1.0f }; | ||||||
| 224 | assert(skcms_TransferFunction_isHLGish(tf))(static_cast <bool> (skcms_TransferFunction_isHLGish(tf )) ? void (0) : __assert_fail ("skcms_TransferFunction_isHLGish(tf)" , __builtin_FILE (), __builtin_LINE (), __extension__ __PRETTY_FUNCTION__ )); | ||||||
| 225 | return true; | ||||||
| 226 | } | ||||||
| 227 | |||||||
| 228 | void skcms_TransferFunction_makePQ( | ||||||
| 229 | skcms_TransferFunction* tf, | ||||||
| 230 | float hdr_reference_white_luminance) { | ||||||
| 231 | *tf = { TFKind_marker(skcms_TFType_PQ), | ||||||
| 232 | hdr_reference_white_luminance, | ||||||
| 233 | 0.f,0.f,0.f,0.f,0.f }; | ||||||
| 234 | assert(skcms_TransferFunction_isPQ(tf))(static_cast <bool> (skcms_TransferFunction_isPQ(tf)) ? void (0) : __assert_fail ("skcms_TransferFunction_isPQ(tf)", __builtin_FILE (), __builtin_LINE (), __extension__ __PRETTY_FUNCTION__ )); | ||||||
| 235 | } | ||||||
| 236 | |||||||
| 237 | void skcms_TransferFunction_makeHLG( | ||||||
| 238 | skcms_TransferFunction* tf, | ||||||
| 239 | float hdr_reference_white_luminance, | ||||||
| 240 | float peak_luminance, | ||||||
| 241 | float system_gamma) { | ||||||
| 242 | *tf = { TFKind_marker(skcms_TFType_HLG), | ||||||
| 243 | hdr_reference_white_luminance, | ||||||
| 244 | peak_luminance, | ||||||
| 245 | system_gamma, | ||||||
| 246 | 0.f, 0.f, 0.f }; | ||||||
| 247 | assert(skcms_TransferFunction_isHLG(tf))(static_cast <bool> (skcms_TransferFunction_isHLG(tf)) ? void (0) : __assert_fail ("skcms_TransferFunction_isHLG(tf)" , __builtin_FILE (), __builtin_LINE (), __extension__ __PRETTY_FUNCTION__ )); | ||||||
| 248 | } | ||||||
| 249 | |||||||
| 250 | float skcms_TransferFunction_eval(const skcms_TransferFunction* tf, float x) { | ||||||
| 251 | float sign = x < 0 ? -1.0f : 1.0f; | ||||||
| 252 | x *= sign; | ||||||
| 253 | |||||||
| 254 | TF_PQish pq; | ||||||
| 255 | TF_HLGish hlg; | ||||||
| 256 | switch (classify(*tf, &pq, &hlg)) { | ||||||
| 257 | case skcms_TFType_Invalid: break; | ||||||
| 258 | |||||||
| 259 | case skcms_TFType_HLG: { | ||||||
| 260 | const float a = 0.17883277f; | ||||||
| 261 | const float b = 0.28466892f; | ||||||
| 262 | const float c = 0.55991073f; | ||||||
| 263 | return sign * (x <= 0.5f ? x*x/3.f : (expf_((x-c)/a) + b) / 12.f); | ||||||
| 264 | } | ||||||
| 265 | |||||||
| 266 | case skcms_TFType_HLGish: { | ||||||
| 267 | const float K = hlg.K_minus_1 + 1.0f; | ||||||
| 268 | return K * sign * (x*hlg.R <= 1 ? powf_(x*hlg.R, hlg.G) | ||||||
| 269 | : expf_((x-hlg.c)*hlg.a) + hlg.b); | ||||||
| 270 | } | ||||||
| 271 | |||||||
| 272 | // skcms_TransferFunction_invert() inverts R, G, and a for HLGinvish so this math is fast. | ||||||
| 273 | case skcms_TFType_HLGinvish: { | ||||||
| 274 | const float K = hlg.K_minus_1 + 1.0f; | ||||||
| 275 | x /= K; | ||||||
| 276 | return sign * (x <= 1 ? hlg.R * powf_(x, hlg.G) | ||||||
| 277 | : hlg.a * logf_(x - hlg.b) + hlg.c); | ||||||
| 278 | } | ||||||
| 279 | |||||||
| 280 | case skcms_TFType_sRGBish: | ||||||
| 281 | return sign * (x < tf->d ? tf->c * x + tf->f | ||||||
| 282 | : powf_(tf->a * x + tf->b, tf->g) + tf->e); | ||||||
| 283 | |||||||
| 284 | case skcms_TFType_PQ: { | ||||||
| 285 | const float c1 = 107 / 128.f; | ||||||
| 286 | const float c2 = 2413 / 128.f; | ||||||
| 287 | const float c3 = 2392 / 128.f; | ||||||
| 288 | const float m1 = 1305 / 8192.f; | ||||||
| 289 | const float m2 = 2523 / 32.f; | ||||||
| 290 | const float p = powf_(x, 1.f / m2); | ||||||
| 291 | return powf_((p - c1) / (c2 - c3 * p), 1.f / m1); | ||||||
| 292 | } | ||||||
| 293 | |||||||
| 294 | case skcms_TFType_PQish: | ||||||
| 295 | return sign * | ||||||
| 296 | powf_((pq.A + pq.B * powf_(x, pq.C)) / (pq.D + pq.E * powf_(x, pq.C)), pq.F); | ||||||
| 297 | } | ||||||
| 298 | return 0; | ||||||
| 299 | } | ||||||
| 300 | |||||||
| 301 | |||||||
| 302 | static float eval_curve(const skcms_Curve* curve, float x) { | ||||||
| 303 | if (curve->table_entries == 0) { | ||||||
| 304 | return skcms_TransferFunction_eval(&curve->parametric, x); | ||||||
| 305 | } | ||||||
| 306 | |||||||
| 307 | float ix = fmaxf_(0, fminf_(x, 1)) * static_cast<float>(curve->table_entries - 1); | ||||||
| 308 | int lo = (int) ix , | ||||||
| 309 | hi = (int)(float)minus_1_ulp(ix + 1.0f); | ||||||
| 310 | float t = ix - (float)lo; | ||||||
| 311 | |||||||
| 312 | float l, h; | ||||||
| 313 | if (curve->table_8) { | ||||||
| 314 | l = curve->table_8[lo] * (1/255.0f); | ||||||
| 315 | h = curve->table_8[hi] * (1/255.0f); | ||||||
| 316 | } else { | ||||||
| 317 | uint16_t be_l, be_h; | ||||||
| 318 | memcpy(&be_l, curve->table_16 + 2*lo, 2); | ||||||
| 319 | memcpy(&be_h, curve->table_16 + 2*hi, 2); | ||||||
| 320 | uint16_t le_l = ((be_l << 8) | (be_l >> 8)) & 0xffff; | ||||||
| 321 | uint16_t le_h = ((be_h << 8) | (be_h >> 8)) & 0xffff; | ||||||
| 322 | l = le_l * (1/65535.0f); | ||||||
| 323 | h = le_h * (1/65535.0f); | ||||||
| 324 | } | ||||||
| 325 | return l + (h-l)*t; | ||||||
| 326 | } | ||||||
| 327 | |||||||
| 328 | float skcms_MaxRoundtripError(const skcms_Curve* curve, const skcms_TransferFunction* inv_tf) { | ||||||
| 329 | uint32_t N = curve->table_entries > 256 ? curve->table_entries : 256; | ||||||
| 330 | const float dx = 1.0f / static_cast<float>(N - 1); | ||||||
| 331 | float err = 0; | ||||||
| 332 | for (uint32_t i = 0; i < N; i++) { | ||||||
| 333 | float x = static_cast<float>(i) * dx, | ||||||
| 334 | y = eval_curve(curve, x); | ||||||
| 335 | err = fmaxf_(err, fabsf_(x - skcms_TransferFunction_eval(inv_tf, y))); | ||||||
| 336 | } | ||||||
| 337 | return err; | ||||||
| 338 | } | ||||||
| 339 | |||||||
| 340 | bool skcms_AreApproximateInverses(const skcms_Curve* curve, const skcms_TransferFunction* inv_tf) { | ||||||
| 341 | return skcms_MaxRoundtripError(curve, inv_tf) < (1/512.0f); | ||||||
| 342 | } | ||||||
| 343 | |||||||
| 344 | // Additional ICC signature values that are only used internally | ||||||
| 345 | enum { | ||||||
| 346 | // File signature | ||||||
| 347 | skcms_Signature_acsp = 0x61637370, | ||||||
| 348 | |||||||
| 349 | // Tag signatures | ||||||
| 350 | skcms_Signature_rTRC = 0x72545243, | ||||||
| 351 | skcms_Signature_gTRC = 0x67545243, | ||||||
| 352 | skcms_Signature_bTRC = 0x62545243, | ||||||
| 353 | skcms_Signature_kTRC = 0x6B545243, | ||||||
| 354 | |||||||
| 355 | skcms_Signature_rXYZ = 0x7258595A, | ||||||
| 356 | skcms_Signature_gXYZ = 0x6758595A, | ||||||
| 357 | skcms_Signature_bXYZ = 0x6258595A, | ||||||
| 358 | |||||||
| 359 | skcms_Signature_A2B0 = 0x41324230, | ||||||
| 360 | skcms_Signature_B2A0 = 0x42324130, | ||||||
| 361 | |||||||
| 362 | skcms_Signature_CHAD = 0x63686164, | ||||||
| 363 | skcms_Signature_WTPT = 0x77747074, | ||||||
| 364 | |||||||
| 365 | skcms_Signature_CICP = 0x63696370, | ||||||
| 366 | |||||||
| 367 | // Type signatures | ||||||
| 368 | skcms_Signature_curv = 0x63757276, | ||||||
| 369 | skcms_Signature_mft1 = 0x6D667431, | ||||||
| 370 | skcms_Signature_mft2 = 0x6D667432, | ||||||
| 371 | skcms_Signature_mAB = 0x6D414220, | ||||||
| 372 | skcms_Signature_mBA = 0x6D424120, | ||||||
| 373 | skcms_Signature_para = 0x70617261, | ||||||
| 374 | skcms_Signature_sf32 = 0x73663332, | ||||||
| 375 | // XYZ is also a PCS signature, so it's defined in skcms.h | ||||||
| 376 | // skcms_Signature_XYZ = 0x58595A20, | ||||||
| 377 | }; | ||||||
| 378 | |||||||
| 379 | static uint16_t read_big_u16(const uint8_t* ptr) { | ||||||
| 380 | uint16_t be; | ||||||
| 381 | memcpy(&be, ptr, sizeof(be)); | ||||||
| 382 | #if defined(_MSC_VER) | ||||||
| 383 | return _byteswap_ushort(be); | ||||||
| 384 | #else | ||||||
| 385 | return __builtin_bswap16(be); | ||||||
| 386 | #endif | ||||||
| 387 | } | ||||||
| 388 | |||||||
| 389 | static uint32_t read_big_u32(const uint8_t* ptr) { | ||||||
| 390 | uint32_t be; | ||||||
| 391 | memcpy(&be, ptr, sizeof(be)); | ||||||
| 392 | #if defined(_MSC_VER) | ||||||
| 393 | return _byteswap_ulong(be); | ||||||
| 394 | #else | ||||||
| 395 | return __builtin_bswap32(be); | ||||||
| 396 | #endif | ||||||
| 397 | } | ||||||
| 398 | |||||||
| 399 | static int32_t read_big_i32(const uint8_t* ptr) { | ||||||
| 400 | return (int32_t)read_big_u32(ptr); | ||||||
| 401 | } | ||||||
| 402 | |||||||
| 403 | static float read_big_fixed(const uint8_t* ptr) { | ||||||
| 404 | return static_cast<float>(read_big_i32(ptr)) * (1.0f / 65536.0f); | ||||||
| 405 | } | ||||||
| 406 | |||||||
| 407 | // Maps to an in-memory profile so that fields line up to the locations specified | ||||||
| 408 | // in ICC.1:2010, section 7.2 | ||||||
| 409 | typedef struct { | ||||||
| 410 | uint8_t size [ 4]; | ||||||
| 411 | uint8_t cmm_type [ 4]; | ||||||
| 412 | uint8_t version [ 4]; | ||||||
| 413 | uint8_t profile_class [ 4]; | ||||||
| 414 | uint8_t data_color_space [ 4]; | ||||||
| 415 | uint8_t pcs [ 4]; | ||||||
| 416 | uint8_t creation_date_time [12]; | ||||||
| 417 | uint8_t signature [ 4]; | ||||||
| 418 | uint8_t platform [ 4]; | ||||||
| 419 | uint8_t flags [ 4]; | ||||||
| 420 | uint8_t device_manufacturer [ 4]; | ||||||
| 421 | uint8_t device_model [ 4]; | ||||||
| 422 | uint8_t device_attributes [ 8]; | ||||||
| 423 | uint8_t rendering_intent [ 4]; | ||||||
| 424 | uint8_t illuminant_X [ 4]; | ||||||
| 425 | uint8_t illuminant_Y [ 4]; | ||||||
| 426 | uint8_t illuminant_Z [ 4]; | ||||||
| 427 | uint8_t creator [ 4]; | ||||||
| 428 | uint8_t profile_id [16]; | ||||||
| 429 | uint8_t reserved [28]; | ||||||
| 430 | uint8_t tag_count [ 4]; // Technically not part of header, but required | ||||||
| 431 | } header_Layout; | ||||||
| 432 | |||||||
| 433 | typedef struct { | ||||||
| 434 | uint8_t signature [4]; | ||||||
| 435 | uint8_t offset [4]; | ||||||
| 436 | uint8_t size [4]; | ||||||
| 437 | } tag_Layout; | ||||||
| 438 | |||||||
| 439 | static const tag_Layout* get_tag_table(const skcms_ICCProfile* profile) { | ||||||
| 440 | return (const tag_Layout*)(profile->buffer + SAFE_SIZEOF(header_Layout)((uint64_t)sizeof(header_Layout))); | ||||||
| 441 | } | ||||||
| 442 | |||||||
| 443 | // s15Fixed16ArrayType is technically variable sized, holding N values. However, the only valid | ||||||
| 444 | // use of the type is for the CHAD tag that stores exactly nine values. | ||||||
| 445 | typedef struct { | ||||||
| 446 | uint8_t type [ 4]; | ||||||
| 447 | uint8_t reserved [ 4]; | ||||||
| 448 | uint8_t values [36]; | ||||||
| 449 | } sf32_Layout; | ||||||
| 450 | |||||||
| 451 | bool skcms_GetCHAD(const skcms_ICCProfile* profile, skcms_Matrix3x3* m) { | ||||||
| 452 | skcms_ICCTag tag; | ||||||
| 453 | if (!skcms_GetTagBySignature(profile, skcms_Signature_CHAD, &tag)) { | ||||||
| 454 | return false; | ||||||
| 455 | } | ||||||
| 456 | |||||||
| 457 | if (tag.type != skcms_Signature_sf32 || tag.size < SAFE_SIZEOF(sf32_Layout)((uint64_t)sizeof(sf32_Layout))) { | ||||||
| 458 | return false; | ||||||
| 459 | } | ||||||
| 460 | |||||||
| 461 | const sf32_Layout* sf32Tag = (const sf32_Layout*)tag.buf; | ||||||
| 462 | const uint8_t* values = sf32Tag->values; | ||||||
| 463 | for (int r = 0; r < 3; ++r) | ||||||
| 464 | for (int c = 0; c < 3; ++c, values += 4) { | ||||||
| 465 | m->vals[r][c] = read_big_fixed(values); | ||||||
| 466 | } | ||||||
| 467 | return true; | ||||||
| 468 | } | ||||||
| 469 | |||||||
| 470 | // XYZType is technically variable sized, holding N XYZ triples. However, the only valid uses of | ||||||
| 471 | // the type are for tags/data that store exactly one triple. | ||||||
| 472 | typedef struct { | ||||||
| 473 | uint8_t type [4]; | ||||||
| 474 | uint8_t reserved [4]; | ||||||
| 475 | uint8_t X [4]; | ||||||
| 476 | uint8_t Y [4]; | ||||||
| 477 | uint8_t Z [4]; | ||||||
| 478 | } XYZ_Layout; | ||||||
| 479 | |||||||
| 480 | static bool read_tag_xyz(const skcms_ICCTag* tag, float* x, float* y, float* z) { | ||||||
| 481 | if (tag->type != skcms_Signature_XYZ || tag->size < SAFE_SIZEOF(XYZ_Layout)((uint64_t)sizeof(XYZ_Layout))) { | ||||||
| 482 | return false; | ||||||
| 483 | } | ||||||
| 484 | |||||||
| 485 | const XYZ_Layout* xyzTag = (const XYZ_Layout*)tag->buf; | ||||||
| 486 | |||||||
| 487 | *x = read_big_fixed(xyzTag->X); | ||||||
| 488 | *y = read_big_fixed(xyzTag->Y); | ||||||
| 489 | *z = read_big_fixed(xyzTag->Z); | ||||||
| 490 | return true; | ||||||
| 491 | } | ||||||
| 492 | |||||||
| 493 | bool skcms_GetWTPT(const skcms_ICCProfile* profile, float xyz[3]) { | ||||||
| 494 | skcms_ICCTag tag; | ||||||
| 495 | return skcms_GetTagBySignature(profile, skcms_Signature_WTPT, &tag) && | ||||||
| 496 | read_tag_xyz(&tag, &xyz[0], &xyz[1], &xyz[2]); | ||||||
| 497 | } | ||||||
| 498 | |||||||
| 499 | static int data_color_space_channel_count(uint32_t data_color_space) { | ||||||
| 500 | switch (data_color_space) { | ||||||
| 501 | case skcms_Signature_CMYK: return 4; | ||||||
| 502 | case skcms_Signature_Gray: return 1; | ||||||
| 503 | case skcms_Signature_RGB: return 3; | ||||||
| 504 | case skcms_Signature_Lab: return 3; | ||||||
| 505 | case skcms_Signature_XYZ: return 3; | ||||||
| 506 | case skcms_Signature_CIELUV: return 3; | ||||||
| 507 | case skcms_Signature_YCbCr: return 3; | ||||||
| 508 | case skcms_Signature_CIEYxy: return 3; | ||||||
| 509 | case skcms_Signature_HSV: return 3; | ||||||
| 510 | case skcms_Signature_HLS: return 3; | ||||||
| 511 | case skcms_Signature_CMY: return 3; | ||||||
| 512 | case skcms_Signature_2CLR: return 2; | ||||||
| 513 | case skcms_Signature_3CLR: return 3; | ||||||
| 514 | case skcms_Signature_4CLR: return 4; | ||||||
| 515 | case skcms_Signature_5CLR: return 5; | ||||||
| 516 | case skcms_Signature_6CLR: return 6; | ||||||
| 517 | case skcms_Signature_7CLR: return 7; | ||||||
| 518 | case skcms_Signature_8CLR: return 8; | ||||||
| 519 | case skcms_Signature_9CLR: return 9; | ||||||
| 520 | case skcms_Signature_10CLR: return 10; | ||||||
| 521 | case skcms_Signature_11CLR: return 11; | ||||||
| 522 | case skcms_Signature_12CLR: return 12; | ||||||
| 523 | case skcms_Signature_13CLR: return 13; | ||||||
| 524 | case skcms_Signature_14CLR: return 14; | ||||||
| 525 | case skcms_Signature_15CLR: return 15; | ||||||
| 526 | default: return -1; | ||||||
| 527 | } | ||||||
| 528 | } | ||||||
| 529 | |||||||
| 530 | int skcms_GetInputChannelCount(const skcms_ICCProfile* profile) { | ||||||
| 531 | int a2b_count = 0; | ||||||
| 532 | if (profile->has_A2B) { | ||||||
| 533 | a2b_count = profile->A2B.input_channels != 0 | ||||||
| 534 | ? static_cast<int>(profile->A2B.input_channels) | ||||||
| 535 | : 3; | ||||||
| 536 | } | ||||||
| 537 | |||||||
| 538 | skcms_ICCTag tag; | ||||||
| 539 | int trc_count = 0; | ||||||
| 540 | if (skcms_GetTagBySignature(profile, skcms_Signature_kTRC, &tag)) { | ||||||
| 541 | trc_count = 1; | ||||||
| 542 | } else if (profile->has_trc) { | ||||||
| 543 | trc_count = 3; | ||||||
| 544 | } | ||||||
| 545 | |||||||
| 546 | int dcs_count = data_color_space_channel_count(profile->data_color_space); | ||||||
| 547 | |||||||
| 548 | if (dcs_count < 0) { | ||||||
| 549 | return -1; | ||||||
| 550 | } | ||||||
| 551 | |||||||
| 552 | if (a2b_count > 0 && a2b_count != dcs_count) { | ||||||
| 553 | return -1; | ||||||
| 554 | } | ||||||
| 555 | if (trc_count > 0 && trc_count != dcs_count) { | ||||||
| 556 | return -1; | ||||||
| 557 | } | ||||||
| 558 | |||||||
| 559 | return dcs_count; | ||||||
| 560 | } | ||||||
| 561 | |||||||
| 562 | static bool read_to_XYZD50(const skcms_ICCTag* rXYZ, const skcms_ICCTag* gXYZ, | ||||||
| 563 | const skcms_ICCTag* bXYZ, skcms_Matrix3x3* toXYZ) { | ||||||
| 564 | return read_tag_xyz(rXYZ, &toXYZ->vals[0][0], &toXYZ->vals[1][0], &toXYZ->vals[2][0]) && | ||||||
| 565 | read_tag_xyz(gXYZ, &toXYZ->vals[0][1], &toXYZ->vals[1][1], &toXYZ->vals[2][1]) && | ||||||
| 566 | read_tag_xyz(bXYZ, &toXYZ->vals[0][2], &toXYZ->vals[1][2], &toXYZ->vals[2][2]); | ||||||
| 567 | } | ||||||
| 568 | |||||||
| 569 | typedef struct { | ||||||
| 570 | uint8_t type [4]; | ||||||
| 571 | uint8_t reserved_a [4]; | ||||||
| 572 | uint8_t function_type [2]; | ||||||
| 573 | uint8_t reserved_b [2]; | ||||||
| 574 | uint8_t variable [1/*variable*/]; // 1, 3, 4, 5, or 7 s15.16, depending on function_type | ||||||
| 575 | } para_Layout; | ||||||
| 576 | |||||||
| 577 | static bool read_curve_para(const uint8_t* buf, uint32_t size, | ||||||
| 578 | skcms_Curve* curve, uint32_t* curve_size) { | ||||||
| 579 | if (size < SAFE_FIXED_SIZE(para_Layout)((uint64_t)__builtin_offsetof(para_Layout, variable))) { | ||||||
| 580 | return false; | ||||||
| 581 | } | ||||||
| 582 | |||||||
| 583 | const para_Layout* paraTag = (const para_Layout*)buf; | ||||||
| 584 | |||||||
| 585 | enum { kG = 0, kGAB = 1, kGABC = 2, kGABCD = 3, kGABCDEF = 4 }; | ||||||
| 586 | uint16_t function_type = read_big_u16(paraTag->function_type); | ||||||
| 587 | if (function_type > kGABCDEF) { | ||||||
| 588 | return false; | ||||||
| 589 | } | ||||||
| 590 | |||||||
| 591 | static const uint32_t curve_bytes[] = { 4, 12, 16, 20, 28 }; | ||||||
| 592 | if (size < SAFE_FIXED_SIZE(para_Layout)((uint64_t)__builtin_offsetof(para_Layout, variable)) + curve_bytes[function_type]) { | ||||||
| 593 | return false; | ||||||
| 594 | } | ||||||
| 595 | |||||||
| 596 | if (curve_size) { | ||||||
| 597 | *curve_size = SAFE_FIXED_SIZE(para_Layout)((uint64_t)__builtin_offsetof(para_Layout, variable)) + curve_bytes[function_type]; | ||||||
| 598 | } | ||||||
| 599 | |||||||
| 600 | curve->table_entries = 0; | ||||||
| 601 | curve->parametric.a = 1.0f; | ||||||
| 602 | curve->parametric.b = 0.0f; | ||||||
| 603 | curve->parametric.c = 0.0f; | ||||||
| 604 | curve->parametric.d = 0.0f; | ||||||
| 605 | curve->parametric.e = 0.0f; | ||||||
| 606 | curve->parametric.f = 0.0f; | ||||||
| 607 | curve->parametric.g = read_big_fixed(paraTag->variable); | ||||||
| 608 | |||||||
| 609 | switch (function_type) { | ||||||
| 610 | case kGAB: | ||||||
| 611 | curve->parametric.a = read_big_fixed(paraTag->variable + 4); | ||||||
| 612 | curve->parametric.b = read_big_fixed(paraTag->variable + 8); | ||||||
| 613 | if (curve->parametric.a == 0) { | ||||||
| 614 | return false; | ||||||
| 615 | } | ||||||
| 616 | curve->parametric.d = -curve->parametric.b / curve->parametric.a; | ||||||
| 617 | break; | ||||||
| 618 | case kGABC: | ||||||
| 619 | curve->parametric.a = read_big_fixed(paraTag->variable + 4); | ||||||
| 620 | curve->parametric.b = read_big_fixed(paraTag->variable + 8); | ||||||
| 621 | curve->parametric.e = read_big_fixed(paraTag->variable + 12); | ||||||
| 622 | if (curve->parametric.a == 0) { | ||||||
| 623 | return false; | ||||||
| 624 | } | ||||||
| 625 | curve->parametric.d = -curve->parametric.b / curve->parametric.a; | ||||||
| 626 | curve->parametric.f = curve->parametric.e; | ||||||
| 627 | break; | ||||||
| 628 | case kGABCD: | ||||||
| 629 | curve->parametric.a = read_big_fixed(paraTag->variable + 4); | ||||||
| 630 | curve->parametric.b = read_big_fixed(paraTag->variable + 8); | ||||||
| 631 | curve->parametric.c = read_big_fixed(paraTag->variable + 12); | ||||||
| 632 | curve->parametric.d = read_big_fixed(paraTag->variable + 16); | ||||||
| 633 | break; | ||||||
| 634 | case kGABCDEF: | ||||||
| 635 | curve->parametric.a = read_big_fixed(paraTag->variable + 4); | ||||||
| 636 | curve->parametric.b = read_big_fixed(paraTag->variable + 8); | ||||||
| 637 | curve->parametric.c = read_big_fixed(paraTag->variable + 12); | ||||||
| 638 | curve->parametric.d = read_big_fixed(paraTag->variable + 16); | ||||||
| 639 | curve->parametric.e = read_big_fixed(paraTag->variable + 20); | ||||||
| 640 | curve->parametric.f = read_big_fixed(paraTag->variable + 24); | ||||||
| 641 | break; | ||||||
| 642 | } | ||||||
| 643 | return skcms_TransferFunction_isSRGBish(&curve->parametric); | ||||||
| 644 | } | ||||||
| 645 | |||||||
| 646 | typedef struct { | ||||||
| 647 | uint8_t type [4]; | ||||||
| 648 | uint8_t reserved [4]; | ||||||
| 649 | uint8_t value_count [4]; | ||||||
| 650 | uint8_t variable [1/*variable*/]; // value_count, 8.8 if 1, uint16 (n*65535) if > 1 | ||||||
| 651 | } curv_Layout; | ||||||
| 652 | |||||||
| 653 | // See https://crbug.com/492744328 for how this was determined. | ||||||
| 654 | static const uint32_t kMaxTableEntries = 1 << 24; // 16,777,216 | ||||||
| 655 | |||||||
| 656 | static bool read_curve_curv(const uint8_t* buf, uint32_t size, | ||||||
| 657 | skcms_Curve* curve, uint32_t* curve_size) { | ||||||
| 658 | if (size < SAFE_FIXED_SIZE(curv_Layout)((uint64_t)__builtin_offsetof(curv_Layout, variable))) { | ||||||
| 659 | return false; | ||||||
| 660 | } | ||||||
| 661 | |||||||
| 662 | const curv_Layout* curvTag = (const curv_Layout*)buf; | ||||||
| 663 | |||||||
| 664 | uint32_t value_count = read_big_u32(curvTag->value_count); | ||||||
| 665 | if (size < SAFE_FIXED_SIZE(curv_Layout)((uint64_t)__builtin_offsetof(curv_Layout, variable)) + value_count * SAFE_SIZEOF(uint16_t)((uint64_t)sizeof(uint16_t))) { | ||||||
| 666 | return false; | ||||||
| 667 | } | ||||||
| 668 | |||||||
| 669 | if (curve_size) { | ||||||
| 670 | *curve_size = SAFE_FIXED_SIZE(curv_Layout)((uint64_t)__builtin_offsetof(curv_Layout, variable)) + value_count * SAFE_SIZEOF(uint16_t)((uint64_t)sizeof(uint16_t)); | ||||||
| 671 | } | ||||||
| 672 | |||||||
| 673 | if (value_count < 2) { | ||||||
| 674 | curve->table_entries = 0; | ||||||
| 675 | curve->parametric.a = 1.0f; | ||||||
| 676 | curve->parametric.b = 0.0f; | ||||||
| 677 | curve->parametric.c = 0.0f; | ||||||
| 678 | curve->parametric.d = 0.0f; | ||||||
| 679 | curve->parametric.e = 0.0f; | ||||||
| 680 | curve->parametric.f = 0.0f; | ||||||
| 681 | if (value_count == 0) { | ||||||
| 682 | // Empty tables are a shorthand for an identity curve | ||||||
| 683 | curve->parametric.g = 1.0f; | ||||||
| 684 | } else { | ||||||
| 685 | // Single entry tables are a shorthand for simple gamma | ||||||
| 686 | curve->parametric.g = read_big_u16(curvTag->variable) * (1.0f / 256.0f); | ||||||
| 687 | } | ||||||
| 688 | return true; | ||||||
| 689 | } | ||||||
| 690 | if (value_count > kMaxTableEntries) { | ||||||
| 691 | return false; | ||||||
| 692 | } | ||||||
| 693 | curve->table_8 = nullptr; | ||||||
| 694 | curve->table_16 = curvTag->variable; | ||||||
| 695 | curve->table_entries = value_count; | ||||||
| 696 | return true; | ||||||
| 697 | } | ||||||
| 698 | |||||||
| 699 | // Parses both curveType and parametricCurveType data. Ensures that at most 'size' bytes are read. | ||||||
| 700 | // If curve_size is not nullptr, writes the number of bytes used by the curve in (*curve_size). | ||||||
| 701 | static bool read_curve(const uint8_t* buf, uint32_t size, | ||||||
| 702 | skcms_Curve* curve, uint32_t* curve_size) { | ||||||
| 703 | if (!buf || size < 4 || !curve) { | ||||||
| 704 | return false; | ||||||
| 705 | } | ||||||
| 706 | |||||||
| 707 | uint32_t type = read_big_u32(buf); | ||||||
| 708 | if (type == skcms_Signature_para) { | ||||||
| 709 | return read_curve_para(buf, size, curve, curve_size); | ||||||
| 710 | } else if (type == skcms_Signature_curv) { | ||||||
| 711 | return read_curve_curv(buf, size, curve, curve_size); | ||||||
| 712 | } | ||||||
| 713 | |||||||
| 714 | return false; | ||||||
| 715 | } | ||||||
| 716 | |||||||
| 717 | // mft1 and mft2 share a large chunk of data | ||||||
| 718 | typedef struct { | ||||||
| 719 | uint8_t type [ 4]; | ||||||
| 720 | uint8_t reserved_a [ 4]; | ||||||
| 721 | uint8_t input_channels [ 1]; | ||||||
| 722 | uint8_t output_channels [ 1]; | ||||||
| 723 | uint8_t grid_points [ 1]; | ||||||
| 724 | uint8_t reserved_b [ 1]; | ||||||
| 725 | uint8_t matrix [36]; | ||||||
| 726 | } mft_CommonLayout; | ||||||
| 727 | |||||||
| 728 | typedef struct { | ||||||
| 729 | mft_CommonLayout common [1]; | ||||||
| 730 | |||||||
| 731 | uint8_t variable [1/*variable*/]; | ||||||
| 732 | } mft1_Layout; | ||||||
| 733 | |||||||
| 734 | typedef struct { | ||||||
| 735 | mft_CommonLayout common [1]; | ||||||
| 736 | |||||||
| 737 | uint8_t input_table_entries [2]; | ||||||
| 738 | uint8_t output_table_entries [2]; | ||||||
| 739 | uint8_t variable [1/*variable*/]; | ||||||
| 740 | } mft2_Layout; | ||||||
| 741 | |||||||
| 742 | static bool read_mft_common(const mft_CommonLayout* mftTag, skcms_A2B* a2b) { | ||||||
| 743 | // MFT matrices are applied before the first set of curves, but must be identity unless the | ||||||
| 744 | // input is PCSXYZ. We don't support PCSXYZ profiles, so we ignore this matrix. Note that the | ||||||
| 745 | // matrix in skcms_A2B is applied later in the pipe, so supporting this would require another | ||||||
| 746 | // field/flag. | ||||||
| 747 | a2b->matrix_channels = 0; | ||||||
| 748 | a2b-> input_channels = mftTag-> input_channels[0]; | ||||||
| 749 | a2b->output_channels = mftTag->output_channels[0]; | ||||||
| 750 | |||||||
| 751 | // We require exactly three (ie XYZ/Lab/RGB) output channels | ||||||
| 752 | if (a2b->output_channels != ARRAY_COUNT(a2b->output_curves)(int)(sizeof((a2b->output_curves)) / sizeof(*(a2b->output_curves )))) { | ||||||
| 753 | return false; | ||||||
| 754 | } | ||||||
| 755 | // We require at least one, and no more than four (ie CMYK) input channels | ||||||
| 756 | if (a2b->input_channels < 1 || a2b->input_channels > ARRAY_COUNT(a2b->input_curves)(int)(sizeof((a2b->input_curves)) / sizeof(*(a2b->input_curves )))) { | ||||||
| 757 | return false; | ||||||
| 758 | } | ||||||
| 759 | |||||||
| 760 | for (uint32_t i = 0; i < a2b->input_channels; ++i) { | ||||||
| 761 | a2b->grid_points[i] = mftTag->grid_points[0]; | ||||||
| 762 | } | ||||||
| 763 | // The grid only makes sense with at least two points along each axis | ||||||
| 764 | if (a2b->grid_points[0] < 2) { | ||||||
| 765 | return false; | ||||||
| 766 | } | ||||||
| 767 | return true; | ||||||
| 768 | } | ||||||
| 769 | |||||||
| 770 | // All as the A2B version above, except where noted. | ||||||
| 771 | static bool read_mft_common(const mft_CommonLayout* mftTag, skcms_B2A* b2a) { | ||||||
| 772 | // Same as A2B. | ||||||
| 773 | b2a->matrix_channels = 0; | ||||||
| 774 | b2a-> input_channels = mftTag-> input_channels[0]; | ||||||
| 775 | b2a->output_channels = mftTag->output_channels[0]; | ||||||
| 776 | |||||||
| 777 | |||||||
| 778 | // For B2A, exactly 3 input channels (XYZ) and 3 (RGB) or 4 (CMYK) output channels. | ||||||
| 779 | if (b2a->input_channels != ARRAY_COUNT(b2a->input_curves)(int)(sizeof((b2a->input_curves)) / sizeof(*(b2a->input_curves )))) { | ||||||
| 780 | return false; | ||||||
| 781 | } | ||||||
| 782 | if (b2a->output_channels < 3 || b2a->output_channels > ARRAY_COUNT(b2a->output_curves)(int)(sizeof((b2a->output_curves)) / sizeof(*(b2a->output_curves )))) { | ||||||
| 783 | return false; | ||||||
| 784 | } | ||||||
| 785 | |||||||
| 786 | // Same as A2B. | ||||||
| 787 | for (uint32_t i = 0; i < b2a->input_channels; ++i) { | ||||||
| 788 | b2a->grid_points[i] = mftTag->grid_points[0]; | ||||||
| 789 | } | ||||||
| 790 | if (b2a->grid_points[0] < 2) { | ||||||
| 791 | return false; | ||||||
| 792 | } | ||||||
| 793 | return true; | ||||||
| 794 | } | ||||||
| 795 | |||||||
| 796 | template <typename A2B_or_B2A> | ||||||
| 797 | static bool init_tables(const uint8_t* table_base, uint64_t max_tables_len, uint32_t byte_width, | ||||||
| 798 | uint32_t input_table_entries, uint32_t output_table_entries, | ||||||
| 799 | A2B_or_B2A* out) { | ||||||
| 800 | // byte_width is 1 or 2, [input|output]_table_entries are in [2, 4096], so no overflow | ||||||
| 801 | uint32_t byte_len_per_input_table = input_table_entries * byte_width; | ||||||
| 802 | uint32_t byte_len_per_output_table = output_table_entries * byte_width; | ||||||
| 803 | |||||||
| 804 | // [input|output]_channels are <= 4, so still no overflow | ||||||
| 805 | uint32_t byte_len_all_input_tables = out->input_channels * byte_len_per_input_table; | ||||||
| 806 | uint32_t byte_len_all_output_tables = out->output_channels * byte_len_per_output_table; | ||||||
| 807 | |||||||
| 808 | uint64_t grid_size = out->output_channels * byte_width; | ||||||
| 809 | for (uint32_t axis = 0; axis < out->input_channels; ++axis) { | ||||||
| 810 | grid_size *= out->grid_points[axis]; | ||||||
| 811 | } | ||||||
| 812 | |||||||
| 813 | if (max_tables_len < byte_len_all_input_tables + grid_size + byte_len_all_output_tables) { | ||||||
| 814 | return false; | ||||||
| 815 | } | ||||||
| 816 | |||||||
| 817 | for (uint32_t i = 0; i < out->input_channels; ++i) { | ||||||
| 818 | out->input_curves[i].table_entries = input_table_entries; | ||||||
| 819 | if (byte_width == 1) { | ||||||
| 820 | out->input_curves[i].table_8 = table_base + i * byte_len_per_input_table; | ||||||
| 821 | out->input_curves[i].table_16 = nullptr; | ||||||
| 822 | } else { | ||||||
| 823 | out->input_curves[i].table_8 = nullptr; | ||||||
| 824 | out->input_curves[i].table_16 = table_base + i * byte_len_per_input_table; | ||||||
| 825 | } | ||||||
| 826 | } | ||||||
| 827 | |||||||
| 828 | if (byte_width == 1) { | ||||||
| 829 | out->grid_8 = table_base + byte_len_all_input_tables; | ||||||
| 830 | out->grid_16 = nullptr; | ||||||
| 831 | } else { | ||||||
| 832 | out->grid_8 = nullptr; | ||||||
| 833 | out->grid_16 = table_base + byte_len_all_input_tables; | ||||||
| 834 | } | ||||||
| 835 | |||||||
| 836 | const uint8_t* output_table_base = table_base + byte_len_all_input_tables + grid_size; | ||||||
| 837 | for (uint32_t i = 0; i < out->output_channels; ++i) { | ||||||
| 838 | out->output_curves[i].table_entries = output_table_entries; | ||||||
| 839 | if (byte_width == 1) { | ||||||
| 840 | out->output_curves[i].table_8 = output_table_base + i * byte_len_per_output_table; | ||||||
| 841 | out->output_curves[i].table_16 = nullptr; | ||||||
| 842 | } else { | ||||||
| 843 | out->output_curves[i].table_8 = nullptr; | ||||||
| 844 | out->output_curves[i].table_16 = output_table_base + i * byte_len_per_output_table; | ||||||
| 845 | } | ||||||
| 846 | } | ||||||
| 847 | |||||||
| 848 | return true; | ||||||
| 849 | } | ||||||
| 850 | |||||||
| 851 | template <typename A2B_or_B2A> | ||||||
| 852 | static bool read_tag_mft1(const skcms_ICCTag* tag, A2B_or_B2A* out) { | ||||||
| 853 | if (tag->size < SAFE_FIXED_SIZE(mft1_Layout)((uint64_t)__builtin_offsetof(mft1_Layout, variable))) { | ||||||
| 854 | return false; | ||||||
| 855 | } | ||||||
| 856 | |||||||
| 857 | const mft1_Layout* mftTag = (const mft1_Layout*)tag->buf; | ||||||
| 858 | if (!read_mft_common(mftTag->common, out)) { | ||||||
| 859 | return false; | ||||||
| 860 | } | ||||||
| 861 | |||||||
| 862 | uint32_t input_table_entries = 256; | ||||||
| 863 | uint32_t output_table_entries = 256; | ||||||
| 864 | |||||||
| 865 | return init_tables(mftTag->variable, tag->size - SAFE_FIXED_SIZE(mft1_Layout)((uint64_t)__builtin_offsetof(mft1_Layout, variable)), 1, | ||||||
| 866 | input_table_entries, output_table_entries, out); | ||||||
| 867 | } | ||||||
| 868 | |||||||
| 869 | template <typename A2B_or_B2A> | ||||||
| 870 | static bool read_tag_mft2(const skcms_ICCTag* tag, A2B_or_B2A* out) { | ||||||
| 871 | if (tag->size < SAFE_FIXED_SIZE(mft2_Layout)((uint64_t)__builtin_offsetof(mft2_Layout, variable))) { | ||||||
| 872 | return false; | ||||||
| 873 | } | ||||||
| 874 | |||||||
| 875 | const mft2_Layout* mftTag = (const mft2_Layout*)tag->buf; | ||||||
| 876 | if (!read_mft_common(mftTag->common, out)) { | ||||||
| 877 | return false; | ||||||
| 878 | } | ||||||
| 879 | |||||||
| 880 | uint32_t input_table_entries = read_big_u16(mftTag->input_table_entries); | ||||||
| 881 | uint32_t output_table_entries = read_big_u16(mftTag->output_table_entries); | ||||||
| 882 | |||||||
| 883 | // ICC spec mandates that 2 <= table_entries <= 4096 | ||||||
| 884 | if (input_table_entries < 2 || input_table_entries > 4096 || | ||||||
| 885 | output_table_entries < 2 || output_table_entries > 4096) { | ||||||
| 886 | return false; | ||||||
| 887 | } | ||||||
| 888 | |||||||
| 889 | return init_tables(mftTag->variable, tag->size - SAFE_FIXED_SIZE(mft2_Layout)((uint64_t)__builtin_offsetof(mft2_Layout, variable)), 2, | ||||||
| 890 | input_table_entries, output_table_entries, out); | ||||||
| 891 | } | ||||||
| 892 | |||||||
| 893 | static bool read_curves(const uint8_t* buf, uint32_t size, uint32_t curve_offset, | ||||||
| 894 | uint32_t num_curves, skcms_Curve* curves) { | ||||||
| 895 | for (uint32_t i = 0; i < num_curves; ++i) { | ||||||
| 896 | if (curve_offset > size) { | ||||||
| 897 | return false; | ||||||
| 898 | } | ||||||
| 899 | |||||||
| 900 | uint32_t curve_bytes; | ||||||
| 901 | if (!read_curve(buf + curve_offset, size - curve_offset, &curves[i], &curve_bytes)) { | ||||||
| 902 | return false; | ||||||
| 903 | } | ||||||
| 904 | |||||||
| 905 | if (curve_bytes > UINT32_MAX(4294967295U) - 3) { | ||||||
| 906 | return false; | ||||||
| 907 | } | ||||||
| 908 | curve_bytes = (curve_bytes + 3) & ~3U; | ||||||
| 909 | |||||||
| 910 | uint64_t new_offset_64 = (uint64_t)curve_offset + curve_bytes; | ||||||
| 911 | curve_offset = (uint32_t)new_offset_64; | ||||||
| 912 | if (new_offset_64 != curve_offset) { | ||||||
| 913 | return false; | ||||||
| 914 | } | ||||||
| 915 | } | ||||||
| 916 | |||||||
| 917 | return true; | ||||||
| 918 | } | ||||||
| 919 | |||||||
| 920 | // mAB and mBA tags use the same encoding, including color lookup tables. | ||||||
| 921 | typedef struct { | ||||||
| 922 | uint8_t type [ 4]; | ||||||
| 923 | uint8_t reserved_a [ 4]; | ||||||
| 924 | uint8_t input_channels [ 1]; | ||||||
| 925 | uint8_t output_channels [ 1]; | ||||||
| 926 | uint8_t reserved_b [ 2]; | ||||||
| 927 | uint8_t b_curve_offset [ 4]; | ||||||
| 928 | uint8_t matrix_offset [ 4]; | ||||||
| 929 | uint8_t m_curve_offset [ 4]; | ||||||
| 930 | uint8_t clut_offset [ 4]; | ||||||
| 931 | uint8_t a_curve_offset [ 4]; | ||||||
| 932 | } mAB_or_mBA_Layout; | ||||||
| 933 | |||||||
| 934 | typedef struct { | ||||||
| 935 | uint8_t grid_points [16]; | ||||||
| 936 | uint8_t grid_byte_width [ 1]; | ||||||
| 937 | uint8_t reserved [ 3]; | ||||||
| 938 | uint8_t variable [1/*variable*/]; | ||||||
| 939 | } CLUT_Layout; | ||||||
| 940 | |||||||
| 941 | static bool read_tag_mab(const skcms_ICCTag* tag, skcms_A2B* a2b, bool pcs_is_xyz, | ||||||
| 942 | const uint8_t* endOfBuffer) { | ||||||
| 943 | if (tag->size < SAFE_SIZEOF(mAB_or_mBA_Layout)((uint64_t)sizeof(mAB_or_mBA_Layout))) { | ||||||
| 944 | return false; | ||||||
| 945 | } | ||||||
| 946 | |||||||
| 947 | const mAB_or_mBA_Layout* mABTag = (const mAB_or_mBA_Layout*)tag->buf; | ||||||
| 948 | |||||||
| 949 | a2b->input_channels = mABTag->input_channels[0]; | ||||||
| 950 | a2b->output_channels = mABTag->output_channels[0]; | ||||||
| 951 | |||||||
| 952 | // We require exactly three (ie XYZ/Lab/RGB) output channels | ||||||
| 953 | if (a2b->output_channels != ARRAY_COUNT(a2b->output_curves)(int)(sizeof((a2b->output_curves)) / sizeof(*(a2b->output_curves )))) { | ||||||
| 954 | return false; | ||||||
| 955 | } | ||||||
| 956 | // We require no more than four (ie CMYK) input channels | ||||||
| 957 | if (a2b->input_channels > ARRAY_COUNT(a2b->input_curves)(int)(sizeof((a2b->input_curves)) / sizeof(*(a2b->input_curves )))) { | ||||||
| 958 | return false; | ||||||
| 959 | } | ||||||
| 960 | |||||||
| 961 | uint32_t b_curve_offset = read_big_u32(mABTag->b_curve_offset); | ||||||
| 962 | uint32_t matrix_offset = read_big_u32(mABTag->matrix_offset); | ||||||
| 963 | uint32_t m_curve_offset = read_big_u32(mABTag->m_curve_offset); | ||||||
| 964 | uint32_t clut_offset = read_big_u32(mABTag->clut_offset); | ||||||
| 965 | uint32_t a_curve_offset = read_big_u32(mABTag->a_curve_offset); | ||||||
| 966 | |||||||
| 967 | // "B" curves must be present | ||||||
| 968 | if (0 == b_curve_offset) { | ||||||
| 969 | return false; | ||||||
| 970 | } | ||||||
| 971 | |||||||
| 972 | if (!read_curves(tag->buf, tag->size, b_curve_offset, a2b->output_channels, | ||||||
| 973 | a2b->output_curves)) { | ||||||
| 974 | return false; | ||||||
| 975 | } | ||||||
| 976 | |||||||
| 977 | // "M" curves and Matrix must be used together | ||||||
| 978 | if (0 != m_curve_offset) { | ||||||
| 979 | if (0 == matrix_offset) { | ||||||
| 980 | return false; | ||||||
| 981 | } | ||||||
| 982 | a2b->matrix_channels = a2b->output_channels; | ||||||
| 983 | if (!read_curves(tag->buf, tag->size, m_curve_offset, a2b->matrix_channels, | ||||||
| 984 | a2b->matrix_curves)) { | ||||||
| 985 | return false; | ||||||
| 986 | } | ||||||
| 987 | |||||||
| 988 | // Read matrix, which is stored as a row-major 3x3, followed by the fourth column | ||||||
| 989 | if (tag->size < matrix_offset + 12 * SAFE_SIZEOF(uint32_t)((uint64_t)sizeof(uint32_t))) { | ||||||
| 990 | return false; | ||||||
| 991 | } | ||||||
| 992 | float encoding_factor = pcs_is_xyz ? (65535 / 32768.0f) : 1.0f; | ||||||
| 993 | const uint8_t* mtx_buf = tag->buf + matrix_offset; | ||||||
| 994 | a2b->matrix.vals[0][0] = encoding_factor * read_big_fixed(mtx_buf + 0); | ||||||
| 995 | a2b->matrix.vals[0][1] = encoding_factor * read_big_fixed(mtx_buf + 4); | ||||||
| 996 | a2b->matrix.vals[0][2] = encoding_factor * read_big_fixed(mtx_buf + 8); | ||||||
| 997 | a2b->matrix.vals[1][0] = encoding_factor * read_big_fixed(mtx_buf + 12); | ||||||
| 998 | a2b->matrix.vals[1][1] = encoding_factor * read_big_fixed(mtx_buf + 16); | ||||||
| 999 | a2b->matrix.vals[1][2] = encoding_factor * read_big_fixed(mtx_buf + 20); | ||||||
| 1000 | a2b->matrix.vals[2][0] = encoding_factor * read_big_fixed(mtx_buf + 24); | ||||||
| 1001 | a2b->matrix.vals[2][1] = encoding_factor * read_big_fixed(mtx_buf + 28); | ||||||
| 1002 | a2b->matrix.vals[2][2] = encoding_factor * read_big_fixed(mtx_buf + 32); | ||||||
| 1003 | a2b->matrix.vals[0][3] = encoding_factor * read_big_fixed(mtx_buf + 36); | ||||||
| 1004 | a2b->matrix.vals[1][3] = encoding_factor * read_big_fixed(mtx_buf + 40); | ||||||
| 1005 | a2b->matrix.vals[2][3] = encoding_factor * read_big_fixed(mtx_buf + 44); | ||||||
| 1006 | } else { | ||||||
| 1007 | if (0 != matrix_offset) { | ||||||
| 1008 | return false; | ||||||
| 1009 | } | ||||||
| 1010 | a2b->matrix_channels = 0; | ||||||
| 1011 | } | ||||||
| 1012 | |||||||
| 1013 | // "A" curves and CLUT must be used together | ||||||
| 1014 | if (0 != a_curve_offset) { | ||||||
| 1015 | if (0 == clut_offset) { | ||||||
| 1016 | return false; | ||||||
| 1017 | } | ||||||
| 1018 | if (!read_curves(tag->buf, tag->size, a_curve_offset, a2b->input_channels, | ||||||
| 1019 | a2b->input_curves)) { | ||||||
| 1020 | return false; | ||||||
| 1021 | } | ||||||
| 1022 | |||||||
| 1023 | if (tag->size < clut_offset + SAFE_FIXED_SIZE(CLUT_Layout)((uint64_t)__builtin_offsetof(CLUT_Layout, variable))) { | ||||||
| 1024 | return false; | ||||||
| 1025 | } | ||||||
| 1026 | const CLUT_Layout* clut = (const CLUT_Layout*)(tag->buf + clut_offset); | ||||||
| 1027 | |||||||
| 1028 | if (clut->grid_byte_width[0] == 1) { | ||||||
| 1029 | a2b->grid_8 = clut->variable; | ||||||
| 1030 | a2b->grid_16 = nullptr; | ||||||
| 1031 | } else if (clut->grid_byte_width[0] == 2) { | ||||||
| 1032 | a2b->grid_8 = nullptr; | ||||||
| 1033 | a2b->grid_16 = clut->variable; | ||||||
| 1034 | } else { | ||||||
| 1035 | return false; | ||||||
| 1036 | } | ||||||
| 1037 | |||||||
| 1038 | uint64_t grid_size = a2b->output_channels * clut->grid_byte_width[0]; // the payload | ||||||
| 1039 | for (uint32_t i = 0; i < a2b->input_channels; ++i) { | ||||||
| 1040 | a2b->grid_points[i] = clut->grid_points[i]; | ||||||
| 1041 | // The grid only makes sense with at least two points along each axis | ||||||
| 1042 | if (a2b->grid_points[i] < 2) { | ||||||
| 1043 | return false; | ||||||
| 1044 | } | ||||||
| 1045 | grid_size *= a2b->grid_points[i]; | ||||||
| 1046 | } | ||||||
| 1047 | const uint64_t table_size = clut_offset + SAFE_FIXED_SIZE(CLUT_Layout)((uint64_t)__builtin_offsetof(CLUT_Layout, variable)) + grid_size; | ||||||
| 1048 | if (table_size > tag->size) { | ||||||
| 1049 | return false; | ||||||
| 1050 | } | ||||||
| 1051 | |||||||
| 1052 | // gather_24 and gather_48 read 1 or 2 extra bytes. | ||||||
| 1053 | // We must ensure that those extra bytes are within the provided buffer limit. | ||||||
| 1054 | uint32_t slack = 0; | ||||||
| 1055 | if (a2b->output_channels == 3) { | ||||||
| 1056 | slack = clut->grid_byte_width[0] == 1 ? 1 : 2; | ||||||
| 1057 | } | ||||||
| 1058 | if (tag->buf + table_size + slack > endOfBuffer) { | ||||||
| 1059 | return false; | ||||||
| 1060 | } | ||||||
| 1061 | } else { | ||||||
| 1062 | if (0 != clut_offset) { | ||||||
| 1063 | return false; | ||||||
| 1064 | } | ||||||
| 1065 | |||||||
| 1066 | // If there is no CLUT, the number of input and output channels must match | ||||||
| 1067 | if (a2b->input_channels != a2b->output_channels) { | ||||||
| 1068 | return false; | ||||||
| 1069 | } | ||||||
| 1070 | |||||||
| 1071 | // Zero out the number of input channels to signal that we're skipping this stage | ||||||
| 1072 | a2b->input_channels = 0; | ||||||
| 1073 | } | ||||||
| 1074 | |||||||
| 1075 | return true; | ||||||
| 1076 | } | ||||||
| 1077 | |||||||
| 1078 | // Exactly the same as read_tag_mab(), except where there are comments. | ||||||
| 1079 | // TODO: refactor the two to eliminate common code? | ||||||
| 1080 | static bool read_tag_mba(const skcms_ICCTag* tag, skcms_B2A* b2a, bool pcs_is_xyz, | ||||||
| 1081 | const uint8_t* endOfBuffer) { | ||||||
| 1082 | if (tag->size < SAFE_SIZEOF(mAB_or_mBA_Layout)((uint64_t)sizeof(mAB_or_mBA_Layout))) { | ||||||
| 1083 | return false; | ||||||
| 1084 | } | ||||||
| 1085 | |||||||
| 1086 | const mAB_or_mBA_Layout* mBATag = (const mAB_or_mBA_Layout*)tag->buf; | ||||||
| 1087 | |||||||
| 1088 | b2a->input_channels = mBATag->input_channels[0]; | ||||||
| 1089 | b2a->output_channels = mBATag->output_channels[0]; | ||||||
| 1090 | |||||||
| 1091 | // Require exactly 3 inputs (XYZ) and 3 (RGB) or 4 (CMYK) outputs. | ||||||
| 1092 | if (b2a->input_channels != ARRAY_COUNT(b2a->input_curves)(int)(sizeof((b2a->input_curves)) / sizeof(*(b2a->input_curves )))) { | ||||||
| 1093 | return false; | ||||||
| 1094 | } | ||||||
| 1095 | if (b2a->output_channels < 3 || b2a->output_channels > ARRAY_COUNT(b2a->output_curves)(int)(sizeof((b2a->output_curves)) / sizeof(*(b2a->output_curves )))) { | ||||||
| 1096 | return false; | ||||||
| 1097 | } | ||||||
| 1098 | |||||||
| 1099 | uint32_t b_curve_offset = read_big_u32(mBATag->b_curve_offset); | ||||||
| 1100 | uint32_t matrix_offset = read_big_u32(mBATag->matrix_offset); | ||||||
| 1101 | uint32_t m_curve_offset = read_big_u32(mBATag->m_curve_offset); | ||||||
| 1102 | uint32_t clut_offset = read_big_u32(mBATag->clut_offset); | ||||||
| 1103 | uint32_t a_curve_offset = read_big_u32(mBATag->a_curve_offset); | ||||||
| 1104 | |||||||
| 1105 | if (0 == b_curve_offset) { | ||||||
| 1106 | return false; | ||||||
| 1107 | } | ||||||
| 1108 | |||||||
| 1109 | // "B" curves are our inputs, not outputs. | ||||||
| 1110 | if (!read_curves(tag->buf, tag->size, b_curve_offset, b2a->input_channels, | ||||||
| 1111 | b2a->input_curves)) { | ||||||
| 1112 | return false; | ||||||
| 1113 | } | ||||||
| 1114 | |||||||
| 1115 | if (0 != m_curve_offset) { | ||||||
| 1116 | if (0 == matrix_offset) { | ||||||
| 1117 | return false; | ||||||
| 1118 | } | ||||||
| 1119 | // Matrix channels is tied to input_channels (3), not output_channels. | ||||||
| 1120 | b2a->matrix_channels = b2a->input_channels; | ||||||
| 1121 | |||||||
| 1122 | if (!read_curves(tag->buf, tag->size, m_curve_offset, b2a->matrix_channels, | ||||||
| 1123 | b2a->matrix_curves)) { | ||||||
| 1124 | return false; | ||||||
| 1125 | } | ||||||
| 1126 | |||||||
| 1127 | if (tag->size < matrix_offset + 12 * SAFE_SIZEOF(uint32_t)((uint64_t)sizeof(uint32_t))) { | ||||||
| 1128 | return false; | ||||||
| 1129 | } | ||||||
| 1130 | float encoding_factor = pcs_is_xyz ? (32768 / 65535.0f) : 1.0f; // TODO: understand | ||||||
| 1131 | const uint8_t* mtx_buf = tag->buf + matrix_offset; | ||||||
| 1132 | b2a->matrix.vals[0][0] = encoding_factor * read_big_fixed(mtx_buf + 0); | ||||||
| 1133 | b2a->matrix.vals[0][1] = encoding_factor * read_big_fixed(mtx_buf + 4); | ||||||
| 1134 | b2a->matrix.vals[0][2] = encoding_factor * read_big_fixed(mtx_buf + 8); | ||||||
| 1135 | b2a->matrix.vals[1][0] = encoding_factor * read_big_fixed(mtx_buf + 12); | ||||||
| 1136 | b2a->matrix.vals[1][1] = encoding_factor * read_big_fixed(mtx_buf + 16); | ||||||
| 1137 | b2a->matrix.vals[1][2] = encoding_factor * read_big_fixed(mtx_buf + 20); | ||||||
| 1138 | b2a->matrix.vals[2][0] = encoding_factor * read_big_fixed(mtx_buf + 24); | ||||||
| 1139 | b2a->matrix.vals[2][1] = encoding_factor * read_big_fixed(mtx_buf + 28); | ||||||
| 1140 | b2a->matrix.vals[2][2] = encoding_factor * read_big_fixed(mtx_buf + 32); | ||||||
| 1141 | b2a->matrix.vals[0][3] = encoding_factor * read_big_fixed(mtx_buf + 36); | ||||||
| 1142 | b2a->matrix.vals[1][3] = encoding_factor * read_big_fixed(mtx_buf + 40); | ||||||
| 1143 | b2a->matrix.vals[2][3] = encoding_factor * read_big_fixed(mtx_buf + 44); | ||||||
| 1144 | } else { | ||||||
| 1145 | if (0 != matrix_offset) { | ||||||
| 1146 | return false; | ||||||
| 1147 | } | ||||||
| 1148 | b2a->matrix_channels = 0; | ||||||
| 1149 | } | ||||||
| 1150 | |||||||
| 1151 | if (0 != a_curve_offset) { | ||||||
| 1152 | if (0 == clut_offset) { | ||||||
| 1153 | return false; | ||||||
| 1154 | } | ||||||
| 1155 | |||||||
| 1156 | // "A" curves are our output, not input. | ||||||
| 1157 | if (!read_curves(tag->buf, tag->size, a_curve_offset, b2a->output_channels, | ||||||
| 1158 | b2a->output_curves)) { | ||||||
| 1159 | return false; | ||||||
| 1160 | } | ||||||
| 1161 | |||||||
| 1162 | if (tag->size < clut_offset + SAFE_FIXED_SIZE(CLUT_Layout)((uint64_t)__builtin_offsetof(CLUT_Layout, variable))) { | ||||||
| 1163 | return false; | ||||||
| 1164 | } | ||||||
| 1165 | const CLUT_Layout* clut = (const CLUT_Layout*)(tag->buf + clut_offset); | ||||||
| 1166 | |||||||
| 1167 | if (clut->grid_byte_width[0] == 1) { | ||||||
| 1168 | b2a->grid_8 = clut->variable; | ||||||
| 1169 | b2a->grid_16 = nullptr; | ||||||
| 1170 | } else if (clut->grid_byte_width[0] == 2) { | ||||||
| 1171 | b2a->grid_8 = nullptr; | ||||||
| 1172 | b2a->grid_16 = clut->variable; | ||||||
| 1173 | } else { | ||||||
| 1174 | return false; | ||||||
| 1175 | } | ||||||
| 1176 | |||||||
| 1177 | uint64_t grid_size = b2a->output_channels * clut->grid_byte_width[0]; | ||||||
| 1178 | for (uint32_t i = 0; i < b2a->input_channels; ++i) { | ||||||
| 1179 | b2a->grid_points[i] = clut->grid_points[i]; | ||||||
| 1180 | if (b2a->grid_points[i] < 2) { | ||||||
| 1181 | return false; | ||||||
| 1182 | } | ||||||
| 1183 | grid_size *= b2a->grid_points[i]; | ||||||
| 1184 | } | ||||||
| 1185 | if (tag->size < clut_offset + SAFE_FIXED_SIZE(CLUT_Layout)((uint64_t)__builtin_offsetof(CLUT_Layout, variable)) + grid_size) { | ||||||
| 1186 | return false; | ||||||
| 1187 | } | ||||||
| 1188 | |||||||
| 1189 | // gather_24 and gather_48 read 1 or 2 extra bytes. | ||||||
| 1190 | // We must ensure that those extra bytes are within the provided buffer limit. | ||||||
| 1191 | uint32_t slack = 0; | ||||||
| 1192 | if (b2a->output_channels == 3) { | ||||||
| 1193 | slack = clut->grid_byte_width[0] == 1 ? 1 : 2; | ||||||
| 1194 | } | ||||||
| 1195 | if (tag->buf + clut_offset + SAFE_FIXED_SIZE(CLUT_Layout)((uint64_t)__builtin_offsetof(CLUT_Layout, variable)) + grid_size + slack > endOfBuffer) { | ||||||
| 1196 | return false; | ||||||
| 1197 | } | ||||||
| 1198 | } else { | ||||||
| 1199 | if (0 != clut_offset) { | ||||||
| 1200 | return false; | ||||||
| 1201 | } | ||||||
| 1202 | |||||||
| 1203 | if (b2a->input_channels != b2a->output_channels) { | ||||||
| 1204 | return false; | ||||||
| 1205 | } | ||||||
| 1206 | |||||||
| 1207 | // Zero out *output* channels to skip this stage. | ||||||
| 1208 | b2a->output_channels = 0; | ||||||
| 1209 | } | ||||||
| 1210 | return true; | ||||||
| 1211 | } | ||||||
| 1212 | |||||||
| 1213 | // If you pass f, we'll fit a possibly-non-zero value for *f. | ||||||
| 1214 | // If you pass nullptr, we'll assume you want *f to be treated as zero. | ||||||
| 1215 | static int fit_linear(const skcms_Curve* curve, int N, float tol, | ||||||
| 1216 | float* c, float* d, float* f = nullptr) { | ||||||
| 1217 | assert(N > 1)(static_cast <bool> (N > 1) ? void (0) : __assert_fail ("N > 1", __builtin_FILE (), __builtin_LINE (), __extension__ __PRETTY_FUNCTION__)); | ||||||
| 1218 | // We iteratively fit the first points to the TF's linear piece. | ||||||
| 1219 | // We want the cx + f line to pass through the first and last points we fit exactly. | ||||||
| 1220 | // | ||||||
| 1221 | // As we walk along the points we find the minimum and maximum slope of the line before the | ||||||
| 1222 | // error would exceed our tolerance. We stop when the range [slope_min, slope_max] becomes | ||||||
| 1223 | // emtpy, when we definitely can't add any more points. | ||||||
| 1224 | // | ||||||
| 1225 | // Some points' error intervals may intersect the running interval but not lie fully | ||||||
| 1226 | // within it. So we keep track of the last point we saw that is a valid end point candidate, | ||||||
| 1227 | // and once the search is done, back up to build the line through *that* point. | ||||||
| 1228 | const float dx = 1.0f / static_cast<float>(N - 1); | ||||||
| 1229 | |||||||
| 1230 | int lin_points = 1; | ||||||
| 1231 | |||||||
| 1232 | float f_zero = 0.0f; | ||||||
| 1233 | if (f
| ||||||
| 1234 | *f = eval_curve(curve, 0); | ||||||
| 1235 | } else { | ||||||
| 1236 | f = &f_zero; | ||||||
| 1237 | } | ||||||
| 1238 | |||||||
| 1239 | |||||||
| 1240 | float slope_min = -INFINITY_; | ||||||
| 1241 | float slope_max = +INFINITY_; | ||||||
| 1242 | for (int i = 1; i
| ||||||
| 1243 | float x = static_cast<float>(i) * dx; | ||||||
| 1244 | float y = eval_curve(curve, x); | ||||||
| 1245 | |||||||
| 1246 | float slope_max_i = (y + tol - *f) / x, | ||||||
| 1247 | slope_min_i = (y - tol - *f) / x; | ||||||
| 1248 | if (slope_max_i < slope_min || slope_max < slope_min_i) { | ||||||
| 1249 | // Slope intervals would no longer overlap. | ||||||
| 1250 | break; | ||||||
| 1251 | } | ||||||
| 1252 | slope_max = fminf_(slope_max, slope_max_i); | ||||||
| 1253 | slope_min = fmaxf_(slope_min, slope_min_i); | ||||||
| 1254 | |||||||
| 1255 | float cur_slope = (y - *f) / x; | ||||||
| 1256 | if (slope_min <= cur_slope && cur_slope <= slope_max) { | ||||||
| 1257 | lin_points = i + 1; | ||||||
| 1258 | *c = cur_slope; | ||||||
| 1259 | } | ||||||
| 1260 | } | ||||||
| 1261 | |||||||
| 1262 | // Set D to the last point that met our tolerance. | ||||||
| 1263 | *d = static_cast<float>(lin_points - 1) * dx; | ||||||
| 1264 | return lin_points; | ||||||
| 1265 | } | ||||||
| 1266 | |||||||
| 1267 | // If this skcms_Curve holds an identity table, rewrite it as an identity skcms_TransferFunction. | ||||||
| 1268 | static void canonicalize_identity(skcms_Curve* curve) { | ||||||
| 1269 | if (curve->table_entries && curve->table_entries <= (uint32_t)INT_MAX2147483647) { | ||||||
| 1270 | int N = (int)curve->table_entries; | ||||||
| 1271 | |||||||
| 1272 | float c = 0.0f, d = 0.0f, f = 0.0f; | ||||||
| 1273 | if (N == fit_linear(curve, N, 1.0f/static_cast<float>(2*N), &c,&d,&f) | ||||||
| 1274 | && c == 1.0f | ||||||
| 1275 | && f == 0.0f) { | ||||||
| 1276 | curve->table_entries = 0; | ||||||
| 1277 | curve->table_8 = nullptr; | ||||||
| 1278 | curve->table_16 = nullptr; | ||||||
| 1279 | curve->parametric = skcms_TransferFunction{1,1,0,0,0,0,0}; | ||||||
| 1280 | } | ||||||
| 1281 | } | ||||||
| 1282 | } | ||||||
| 1283 | |||||||
| 1284 | static bool read_a2b(const skcms_ICCTag* tag, skcms_A2B* a2b, bool pcs_is_xyz, const uint8_t* eob) { | ||||||
| 1285 | bool ok = false; | ||||||
| 1286 | if (tag->type == skcms_Signature_mft1) { ok = read_tag_mft1(tag, a2b); } | ||||||
| 1287 | if (tag->type == skcms_Signature_mft2) { ok = read_tag_mft2(tag, a2b); } | ||||||
| 1288 | if (tag->type == skcms_Signature_mAB ) { ok = read_tag_mab(tag, a2b, pcs_is_xyz, eob); } | ||||||
| 1289 | if (!ok) { | ||||||
| 1290 | return false; | ||||||
| 1291 | } | ||||||
| 1292 | |||||||
| 1293 | if (a2b->input_channels > 0) { canonicalize_identity(a2b->input_curves + 0); } | ||||||
| 1294 | if (a2b->input_channels > 1) { canonicalize_identity(a2b->input_curves + 1); } | ||||||
| 1295 | if (a2b->input_channels > 2) { canonicalize_identity(a2b->input_curves + 2); } | ||||||
| 1296 | if (a2b->input_channels > 3) { canonicalize_identity(a2b->input_curves + 3); } | ||||||
| 1297 | |||||||
| 1298 | if (a2b->matrix_channels > 0) { canonicalize_identity(a2b->matrix_curves + 0); } | ||||||
| 1299 | if (a2b->matrix_channels > 1) { canonicalize_identity(a2b->matrix_curves + 1); } | ||||||
| 1300 | if (a2b->matrix_channels > 2) { canonicalize_identity(a2b->matrix_curves + 2); } | ||||||
| 1301 | |||||||
| 1302 | if (a2b->output_channels > 0) { canonicalize_identity(a2b->output_curves + 0); } | ||||||
| 1303 | if (a2b->output_channels > 1) { canonicalize_identity(a2b->output_curves + 1); } | ||||||
| 1304 | if (a2b->output_channels > 2) { canonicalize_identity(a2b->output_curves + 2); } | ||||||
| 1305 | |||||||
| 1306 | return true; | ||||||
| 1307 | } | ||||||
| 1308 | |||||||
| 1309 | static bool read_b2a(const skcms_ICCTag* tag, skcms_B2A* b2a, bool pcs_is_xyz, const uint8_t* eob) { | ||||||
| 1310 | bool ok = false; | ||||||
| 1311 | if (tag->type == skcms_Signature_mft1) { ok = read_tag_mft1(tag, b2a); } | ||||||
| 1312 | if (tag->type == skcms_Signature_mft2) { ok = read_tag_mft2(tag, b2a); } | ||||||
| 1313 | if (tag->type == skcms_Signature_mBA ) { ok = read_tag_mba(tag, b2a, pcs_is_xyz, eob); } | ||||||
| 1314 | if (!ok) { | ||||||
| 1315 | return false; | ||||||
| 1316 | } | ||||||
| 1317 | |||||||
| 1318 | if (b2a->input_channels > 0) { canonicalize_identity(b2a->input_curves + 0); } | ||||||
| 1319 | if (b2a->input_channels > 1) { canonicalize_identity(b2a->input_curves + 1); } | ||||||
| 1320 | if (b2a->input_channels > 2) { canonicalize_identity(b2a->input_curves + 2); } | ||||||
| 1321 | |||||||
| 1322 | if (b2a->matrix_channels > 0) { canonicalize_identity(b2a->matrix_curves + 0); } | ||||||
| 1323 | if (b2a->matrix_channels > 1) { canonicalize_identity(b2a->matrix_curves + 1); } | ||||||
| 1324 | if (b2a->matrix_channels > 2) { canonicalize_identity(b2a->matrix_curves + 2); } | ||||||
| 1325 | |||||||
| 1326 | if (b2a->output_channels > 0) { canonicalize_identity(b2a->output_curves + 0); } | ||||||
| 1327 | if (b2a->output_channels > 1) { canonicalize_identity(b2a->output_curves + 1); } | ||||||
| 1328 | if (b2a->output_channels > 2) { canonicalize_identity(b2a->output_curves + 2); } | ||||||
| 1329 | if (b2a->output_channels > 3) { canonicalize_identity(b2a->output_curves + 3); } | ||||||
| 1330 | |||||||
| 1331 | return true; | ||||||
| 1332 | } | ||||||
| 1333 | |||||||
| 1334 | typedef struct { | ||||||
| 1335 | uint8_t type [4]; | ||||||
| 1336 | uint8_t reserved [4]; | ||||||
| 1337 | uint8_t color_primaries [1]; | ||||||
| 1338 | uint8_t transfer_characteristics [1]; | ||||||
| 1339 | uint8_t matrix_coefficients [1]; | ||||||
| 1340 | uint8_t video_full_range_flag [1]; | ||||||
| 1341 | } CICP_Layout; | ||||||
| 1342 | |||||||
| 1343 | static bool read_cicp(const skcms_ICCTag* tag, skcms_CICP* cicp) { | ||||||
| 1344 | if (tag->type != skcms_Signature_CICP || tag->size < SAFE_SIZEOF(CICP_Layout)((uint64_t)sizeof(CICP_Layout))) { | ||||||
| 1345 | return false; | ||||||
| 1346 | } | ||||||
| 1347 | |||||||
| 1348 | const CICP_Layout* cicpTag = (const CICP_Layout*)tag->buf; | ||||||
| 1349 | |||||||
| 1350 | cicp->color_primaries = cicpTag->color_primaries[0]; | ||||||
| 1351 | cicp->transfer_characteristics = cicpTag->transfer_characteristics[0]; | ||||||
| 1352 | cicp->matrix_coefficients = cicpTag->matrix_coefficients[0]; | ||||||
| 1353 | cicp->video_full_range_flag = cicpTag->video_full_range_flag[0]; | ||||||
| 1354 | return true; | ||||||
| 1355 | } | ||||||
| 1356 | |||||||
| 1357 | void skcms_GetTagByIndex(const skcms_ICCProfile* profile, uint32_t idx, skcms_ICCTag* tag) { | ||||||
| 1358 | if (!profile || !profile->buffer || !tag) { return; } | ||||||
| 1359 | if (idx > profile->tag_count) { return; } | ||||||
| 1360 | const tag_Layout* tags = get_tag_table(profile); | ||||||
| 1361 | tag->signature = read_big_u32(tags[idx].signature); | ||||||
| 1362 | tag->size = read_big_u32(tags[idx].size); | ||||||
| 1363 | tag->buf = read_big_u32(tags[idx].offset) + profile->buffer; | ||||||
| 1364 | tag->type = read_big_u32(tag->buf); | ||||||
| 1365 | } | ||||||
| 1366 | |||||||
| 1367 | bool skcms_GetTagBySignature(const skcms_ICCProfile* profile, uint32_t sig, skcms_ICCTag* tag) { | ||||||
| 1368 | if (!profile || !profile->buffer || !tag) { return false; } | ||||||
| 1369 | const tag_Layout* tags = get_tag_table(profile); | ||||||
| 1370 | for (uint32_t i = 0; i < profile->tag_count; ++i) { | ||||||
| 1371 | if (read_big_u32(tags[i].signature) == sig) { | ||||||
| 1372 | tag->signature = sig; | ||||||
| 1373 | tag->size = read_big_u32(tags[i].size); | ||||||
| 1374 | tag->buf = read_big_u32(tags[i].offset) + profile->buffer; | ||||||
| 1375 | tag->type = read_big_u32(tag->buf); | ||||||
| 1376 | return true; | ||||||
| 1377 | } | ||||||
| 1378 | } | ||||||
| 1379 | return false; | ||||||
| 1380 | } | ||||||
| 1381 | |||||||
| 1382 | static bool usable_as_src(const skcms_ICCProfile* profile) { | ||||||
| 1383 | return profile->has_A2B | ||||||
| 1384 | || (profile->has_trc && profile->has_toXYZD50); | ||||||
| 1385 | } | ||||||
| 1386 | |||||||
| 1387 | bool skcms_ParseWithA2BPriority(const void* buf, size_t len, | ||||||
| 1388 | const int priority[], const int priorities, | ||||||
| 1389 | skcms_ICCProfile* profile) { | ||||||
| 1390 | static_assert(SAFE_SIZEOF(header_Layout)((uint64_t)sizeof(header_Layout)) == 132, "need to update header code"); | ||||||
| 1391 | |||||||
| 1392 | if (!profile) { | ||||||
| 1393 | return false; | ||||||
| 1394 | } | ||||||
| 1395 | memset(profile, 0, SAFE_SIZEOF(*profile)((uint64_t)sizeof(*profile))); | ||||||
| 1396 | |||||||
| 1397 | if (len < SAFE_SIZEOF(header_Layout)((uint64_t)sizeof(header_Layout))) { | ||||||
| 1398 | return false; | ||||||
| 1399 | } | ||||||
| 1400 | |||||||
| 1401 | // Byte-swap all header fields | ||||||
| 1402 | const header_Layout* header = (const header_Layout*)buf; | ||||||
| 1403 | profile->buffer = (const uint8_t*)buf; | ||||||
| 1404 | profile->size = read_big_u32(header->size); | ||||||
| 1405 | uint32_t version = read_big_u32(header->version); | ||||||
| 1406 | profile->data_color_space = read_big_u32(header->data_color_space); | ||||||
| 1407 | profile->pcs = read_big_u32(header->pcs); | ||||||
| 1408 | uint32_t signature = read_big_u32(header->signature); | ||||||
| 1409 | float illuminant_X = read_big_fixed(header->illuminant_X); | ||||||
| 1410 | float illuminant_Y = read_big_fixed(header->illuminant_Y); | ||||||
| 1411 | float illuminant_Z = read_big_fixed(header->illuminant_Z); | ||||||
| 1412 | profile->tag_count = read_big_u32(header->tag_count); | ||||||
| 1413 | |||||||
| 1414 | // Validate signature, size (smaller than buffer, large enough to hold tag table), | ||||||
| 1415 | // and major version | ||||||
| 1416 | uint64_t tag_table_size = profile->tag_count * SAFE_SIZEOF(tag_Layout)((uint64_t)sizeof(tag_Layout)); | ||||||
| 1417 | if (signature != skcms_Signature_acsp || | ||||||
| 1418 | profile->size > len || | ||||||
| 1419 | profile->size < SAFE_SIZEOF(header_Layout)((uint64_t)sizeof(header_Layout)) + tag_table_size || | ||||||
| 1420 | (version >> 24) > 4) { | ||||||
| 1421 | return false; | ||||||
| 1422 | } | ||||||
| 1423 | |||||||
| 1424 | // Validate that illuminant is D50 white | ||||||
| 1425 | if (fabsf_(illuminant_X - 0.9642f) > 0.0100f || | ||||||
| 1426 | fabsf_(illuminant_Y - 1.0000f) > 0.0100f || | ||||||
| 1427 | fabsf_(illuminant_Z - 0.8249f) > 0.0100f) { | ||||||
| 1428 | return false; | ||||||
| 1429 | } | ||||||
| 1430 | |||||||
| 1431 | // Validate that all tag entries have sane offset + size | ||||||
| 1432 | const tag_Layout* tags = get_tag_table(profile); | ||||||
| 1433 | for (uint32_t i = 0; i < profile->tag_count; ++i) { | ||||||
| 1434 | uint32_t tag_offset = read_big_u32(tags[i].offset); | ||||||
| 1435 | uint32_t tag_size = read_big_u32(tags[i].size); | ||||||
| 1436 | uint64_t tag_end = (uint64_t)tag_offset + (uint64_t)tag_size; | ||||||
| 1437 | if (tag_size < 4 || tag_end > profile->size) { | ||||||
| 1438 | return false; | ||||||
| 1439 | } | ||||||
| 1440 | } | ||||||
| 1441 | |||||||
| 1442 | if (profile->pcs != skcms_Signature_XYZ && profile->pcs != skcms_Signature_Lab) { | ||||||
| 1443 | return false; | ||||||
| 1444 | } | ||||||
| 1445 | |||||||
| 1446 | bool pcs_is_xyz = profile->pcs == skcms_Signature_XYZ; | ||||||
| 1447 | |||||||
| 1448 | // Pre-parse commonly used tags. | ||||||
| 1449 | skcms_ICCTag kTRC; | ||||||
| 1450 | if (profile->data_color_space == skcms_Signature_Gray && | ||||||
| 1451 | skcms_GetTagBySignature(profile, skcms_Signature_kTRC, &kTRC)) { | ||||||
| 1452 | if (!read_curve(kTRC.buf, kTRC.size, &profile->trc[0], nullptr)) { | ||||||
| 1453 | // Malformed tag | ||||||
| 1454 | return false; | ||||||
| 1455 | } | ||||||
| 1456 | profile->trc[1] = profile->trc[0]; | ||||||
| 1457 | profile->trc[2] = profile->trc[0]; | ||||||
| 1458 | profile->has_trc = true; | ||||||
| 1459 | |||||||
| 1460 | if (pcs_is_xyz) { | ||||||
| 1461 | profile->toXYZD50.vals[0][0] = illuminant_X; | ||||||
| 1462 | profile->toXYZD50.vals[1][1] = illuminant_Y; | ||||||
| 1463 | profile->toXYZD50.vals[2][2] = illuminant_Z; | ||||||
| 1464 | profile->has_toXYZD50 = true; | ||||||
| 1465 | } | ||||||
| 1466 | } else { | ||||||
| 1467 | skcms_ICCTag rTRC, gTRC, bTRC; | ||||||
| 1468 | if (skcms_GetTagBySignature(profile, skcms_Signature_rTRC, &rTRC) && | ||||||
| 1469 | skcms_GetTagBySignature(profile, skcms_Signature_gTRC, &gTRC) && | ||||||
| 1470 | skcms_GetTagBySignature(profile, skcms_Signature_bTRC, &bTRC)) { | ||||||
| 1471 | if (!read_curve(rTRC.buf, rTRC.size, &profile->trc[0], nullptr) || | ||||||
| 1472 | !read_curve(gTRC.buf, gTRC.size, &profile->trc[1], nullptr) || | ||||||
| 1473 | !read_curve(bTRC.buf, bTRC.size, &profile->trc[2], nullptr)) { | ||||||
| 1474 | // Malformed TRC tags | ||||||
| 1475 | return false; | ||||||
| 1476 | } | ||||||
| 1477 | profile->has_trc = true; | ||||||
| 1478 | } | ||||||
| 1479 | |||||||
| 1480 | skcms_ICCTag rXYZ, gXYZ, bXYZ; | ||||||
| 1481 | if (skcms_GetTagBySignature(profile, skcms_Signature_rXYZ, &rXYZ) && | ||||||
| 1482 | skcms_GetTagBySignature(profile, skcms_Signature_gXYZ, &gXYZ) && | ||||||
| 1483 | skcms_GetTagBySignature(profile, skcms_Signature_bXYZ, &bXYZ)) { | ||||||
| 1484 | if (!read_to_XYZD50(&rXYZ, &gXYZ, &bXYZ, &profile->toXYZD50)) { | ||||||
| 1485 | // Malformed XYZ tags | ||||||
| 1486 | return false; | ||||||
| 1487 | } | ||||||
| 1488 | profile->has_toXYZD50 = true; | ||||||
| 1489 | } | ||||||
| 1490 | } | ||||||
| 1491 | |||||||
| 1492 | const uint8_t* endOfBuffer = (const uint8_t*)buf + len; | ||||||
| 1493 | for (int i = 0; i < priorities; i++) { | ||||||
| 1494 | // enum { perceptual, relative_colormetric, saturation } | ||||||
| 1495 | if (priority[i] < 0 || priority[i] > 2) { | ||||||
| 1496 | return false; | ||||||
| 1497 | } | ||||||
| 1498 | uint32_t sig = skcms_Signature_A2B0 + static_cast<uint32_t>(priority[i]); | ||||||
| 1499 | skcms_ICCTag tag; | ||||||
| 1500 | if (skcms_GetTagBySignature(profile, sig, &tag)) { | ||||||
| 1501 | if (!read_a2b(&tag, &profile->A2B, pcs_is_xyz, endOfBuffer)) { | ||||||
| 1502 | // Malformed A2B tag | ||||||
| 1503 | return false; | ||||||
| 1504 | } | ||||||
| 1505 | profile->has_A2B = true; | ||||||
| 1506 | break; | ||||||
| 1507 | } | ||||||
| 1508 | } | ||||||
| 1509 | |||||||
| 1510 | for (int i = 0; i < priorities; i++) { | ||||||
| 1511 | // enum { perceptual, relative_colormetric, saturation } | ||||||
| 1512 | if (priority[i] < 0 || priority[i] > 2) { | ||||||
| 1513 | return false; | ||||||
| 1514 | } | ||||||
| 1515 | uint32_t sig = skcms_Signature_B2A0 + static_cast<uint32_t>(priority[i]); | ||||||
| 1516 | skcms_ICCTag tag; | ||||||
| 1517 | if (skcms_GetTagBySignature(profile, sig, &tag)) { | ||||||
| 1518 | if (!read_b2a(&tag, &profile->B2A, pcs_is_xyz, endOfBuffer)) { | ||||||
| 1519 | // Malformed B2A tag | ||||||
| 1520 | return false; | ||||||
| 1521 | } | ||||||
| 1522 | profile->has_B2A = true; | ||||||
| 1523 | break; | ||||||
| 1524 | } | ||||||
| 1525 | } | ||||||
| 1526 | |||||||
| 1527 | skcms_ICCTag cicp_tag; | ||||||
| 1528 | if (skcms_GetTagBySignature(profile, skcms_Signature_CICP, &cicp_tag)) { | ||||||
| 1529 | if (!read_cicp(&cicp_tag, &profile->CICP)) { | ||||||
| 1530 | // Malformed CICP tag | ||||||
| 1531 | return false; | ||||||
| 1532 | } | ||||||
| 1533 | profile->has_CICP = true; | ||||||
| 1534 | } | ||||||
| 1535 | |||||||
| 1536 | return usable_as_src(profile); | ||||||
| 1537 | } | ||||||
| 1538 | |||||||
| 1539 | |||||||
| 1540 | const skcms_ICCProfile* skcms_sRGB_profile() { | ||||||
| 1541 | static const skcms_ICCProfile sRGB_profile = { | ||||||
| 1542 | nullptr, // buffer, moot here | ||||||
| 1543 | |||||||
| 1544 | 0, // size, moot here | ||||||
| 1545 | skcms_Signature_RGB, // data_color_space | ||||||
| 1546 | skcms_Signature_XYZ, // pcs | ||||||
| 1547 | 0, // tag count, moot here | ||||||
| 1548 | |||||||
| 1549 | // We choose to represent sRGB with its canonical transfer function, | ||||||
| 1550 | // and with its canonical XYZD50 gamut matrix. | ||||||
| 1551 | { // the 3 trc curves | ||||||
| 1552 | {{0, {2.4f, (float)(1/1.055), (float)(0.055/1.055), (float)(1/12.92), 0.04045f, 0, 0}}}, | ||||||
| 1553 | {{0, {2.4f, (float)(1/1.055), (float)(0.055/1.055), (float)(1/12.92), 0.04045f, 0, 0}}}, | ||||||
| 1554 | {{0, {2.4f, (float)(1/1.055), (float)(0.055/1.055), (float)(1/12.92), 0.04045f, 0, 0}}}, | ||||||
| 1555 | }, | ||||||
| 1556 | |||||||
| 1557 | {{ // 3x3 toXYZD50 matrix | ||||||
| 1558 | { 0.436065674f, 0.385147095f, 0.143066406f }, | ||||||
| 1559 | { 0.222488403f, 0.716873169f, 0.060607910f }, | ||||||
| 1560 | { 0.013916016f, 0.097076416f, 0.714096069f }, | ||||||
| 1561 | }}, | ||||||
| 1562 | |||||||
| 1563 | { // an empty A2B | ||||||
| 1564 | { // input_curves | ||||||
| 1565 | {{0, {0,0, 0,0,0,0,0}}}, | ||||||
| 1566 | {{0, {0,0, 0,0,0,0,0}}}, | ||||||
| 1567 | {{0, {0,0, 0,0,0,0,0}}}, | ||||||
| 1568 | {{0, {0,0, 0,0,0,0,0}}}, | ||||||
| 1569 | }, | ||||||
| 1570 | nullptr, // grid_8 | ||||||
| 1571 | nullptr, // grid_16 | ||||||
| 1572 | 0, // input_channels | ||||||
| 1573 | {0,0,0,0}, // grid_points | ||||||
| 1574 | |||||||
| 1575 | { // matrix_curves | ||||||
| 1576 | {{0, {0,0, 0,0,0,0,0}}}, | ||||||
| 1577 | {{0, {0,0, 0,0,0,0,0}}}, | ||||||
| 1578 | {{0, {0,0, 0,0,0,0,0}}}, | ||||||
| 1579 | }, | ||||||
| 1580 | {{ // matrix (3x4) | ||||||
| 1581 | { 0,0,0,0 }, | ||||||
| 1582 | { 0,0,0,0 }, | ||||||
| 1583 | { 0,0,0,0 }, | ||||||
| 1584 | }}, | ||||||
| 1585 | 0, // matrix_channels | ||||||
| 1586 | |||||||
| 1587 | 0, // output_channels | ||||||
| 1588 | { // output_curves | ||||||
| 1589 | {{0, {0,0, 0,0,0,0,0}}}, | ||||||
| 1590 | {{0, {0,0, 0,0,0,0,0}}}, | ||||||
| 1591 | {{0, {0,0, 0,0,0,0,0}}}, | ||||||
| 1592 | }, | ||||||
| 1593 | }, | ||||||
| 1594 | |||||||
| 1595 | { // an empty B2A | ||||||
| 1596 | { // input_curves | ||||||
| 1597 | {{0, {0,0, 0,0,0,0,0}}}, | ||||||
| 1598 | {{0, {0,0, 0,0,0,0,0}}}, | ||||||
| 1599 | {{0, {0,0, 0,0,0,0,0}}}, | ||||||
| 1600 | }, | ||||||
| 1601 | 0, // input_channels | ||||||
| 1602 | |||||||
| 1603 | 0, // matrix_channels | ||||||
| 1604 | { // matrix_curves | ||||||
| 1605 | {{0, {0,0, 0,0,0,0,0}}}, | ||||||
| 1606 | {{0, {0,0, 0,0,0,0,0}}}, | ||||||
| 1607 | {{0, {0,0, 0,0,0,0,0}}}, | ||||||
| 1608 | }, | ||||||
| 1609 | {{ // matrix (3x4) | ||||||
| 1610 | { 0,0,0,0 }, | ||||||
| 1611 | { 0,0,0,0 }, | ||||||
| 1612 | { 0,0,0,0 }, | ||||||
| 1613 | }}, | ||||||
| 1614 | |||||||
| 1615 | { // output_curves | ||||||
| 1616 | {{0, {0,0, 0,0,0,0,0}}}, | ||||||
| 1617 | {{0, {0,0, 0,0,0,0,0}}}, | ||||||
| 1618 | {{0, {0,0, 0,0,0,0,0}}}, | ||||||
| 1619 | {{0, {0,0, 0,0,0,0,0}}}, | ||||||
| 1620 | }, | ||||||
| 1621 | nullptr, // grid_8 | ||||||
| 1622 | nullptr, // grid_16 | ||||||
| 1623 | {0,0,0,0}, // grid_points | ||||||
| 1624 | 0, // output_channels | ||||||
| 1625 | }, | ||||||
| 1626 | |||||||
| 1627 | { 0, 0, 0, 0 }, // an empty CICP | ||||||
| 1628 | |||||||
| 1629 | true, // has_trc | ||||||
| 1630 | true, // has_toXYZD50 | ||||||
| 1631 | false, // has_A2B | ||||||
| 1632 | false, // has B2A | ||||||
| 1633 | false, // has_CICP | ||||||
| 1634 | }; | ||||||
| 1635 | return &sRGB_profile; | ||||||
| 1636 | } | ||||||
| 1637 | |||||||
| 1638 | const skcms_ICCProfile* skcms_XYZD50_profile() { | ||||||
| 1639 | // Just like sRGB above, but with identity transfer functions and toXYZD50 matrix. | ||||||
| 1640 | static const skcms_ICCProfile XYZD50_profile = { | ||||||
| 1641 | nullptr, // buffer, moot here | ||||||
| 1642 | |||||||
| 1643 | 0, // size, moot here | ||||||
| 1644 | skcms_Signature_RGB, // data_color_space | ||||||
| 1645 | skcms_Signature_XYZ, // pcs | ||||||
| 1646 | 0, // tag count, moot here | ||||||
| 1647 | |||||||
| 1648 | { // the 3 trc curves | ||||||
| 1649 | {{0, {1,1, 0,0,0,0,0}}}, | ||||||
| 1650 | {{0, {1,1, 0,0,0,0,0}}}, | ||||||
| 1651 | {{0, {1,1, 0,0,0,0,0}}}, | ||||||
| 1652 | }, | ||||||
| 1653 | |||||||
| 1654 | {{ // 3x3 toXYZD50 matrix | ||||||
| 1655 | { 1,0,0 }, | ||||||
| 1656 | { 0,1,0 }, | ||||||
| 1657 | { 0,0,1 }, | ||||||
| 1658 | }}, | ||||||
| 1659 | |||||||
| 1660 | { // an empty A2B | ||||||
| 1661 | { // input_curves | ||||||
| 1662 | {{0, {0,0, 0,0,0,0,0}}}, | ||||||
| 1663 | {{0, {0,0, 0,0,0,0,0}}}, | ||||||
| 1664 | {{0, {0,0, 0,0,0,0,0}}}, | ||||||
| 1665 | {{0, {0,0, 0,0,0,0,0}}}, | ||||||
| 1666 | }, | ||||||
| 1667 | nullptr, // grid_8 | ||||||
| 1668 | nullptr, // grid_16 | ||||||
| 1669 | 0, // input_channels | ||||||
| 1670 | {0,0,0,0}, // grid_points | ||||||
| 1671 | |||||||
| 1672 | { // matrix_curves | ||||||
| 1673 | {{0, {0,0, 0,0,0,0,0}}}, | ||||||
| 1674 | {{0, {0,0, 0,0,0,0,0}}}, | ||||||
| 1675 | {{0, {0,0, 0,0,0,0,0}}}, | ||||||
| 1676 | }, | ||||||
| 1677 | {{ // matrix (3x4) | ||||||
| 1678 | { 0,0,0,0 }, | ||||||
| 1679 | { 0,0,0,0 }, | ||||||
| 1680 | { 0,0,0,0 }, | ||||||
| 1681 | }}, | ||||||
| 1682 | 0, // matrix_channels | ||||||
| 1683 | |||||||
| 1684 | 0, // output_channels | ||||||
| 1685 | { // output_curves | ||||||
| 1686 | {{0, {0,0, 0,0,0,0,0}}}, | ||||||
| 1687 | {{0, {0,0, 0,0,0,0,0}}}, | ||||||
| 1688 | {{0, {0,0, 0,0,0,0,0}}}, | ||||||
| 1689 | }, | ||||||
| 1690 | }, | ||||||
| 1691 | |||||||
| 1692 | { // an empty B2A | ||||||
| 1693 | { // input_curves | ||||||
| 1694 | {{0, {0,0, 0,0,0,0,0}}}, | ||||||
| 1695 | {{0, {0,0, 0,0,0,0,0}}}, | ||||||
| 1696 | {{0, {0,0, 0,0,0,0,0}}}, | ||||||
| 1697 | }, | ||||||
| 1698 | 0, // input_channels | ||||||
| 1699 | |||||||
| 1700 | 0, // matrix_channels | ||||||
| 1701 | { // matrix_curves | ||||||
| 1702 | {{0, {0,0, 0,0,0,0,0}}}, | ||||||
| 1703 | {{0, {0,0, 0,0,0,0,0}}}, | ||||||
| 1704 | {{0, {0,0, 0,0,0,0,0}}}, | ||||||
| 1705 | }, | ||||||
| 1706 | {{ // matrix (3x4) | ||||||
| 1707 | { 0,0,0,0 }, | ||||||
| 1708 | { 0,0,0,0 }, | ||||||
| 1709 | { 0,0,0,0 }, | ||||||
| 1710 | }}, | ||||||
| 1711 | |||||||
| 1712 | { // output_curves | ||||||
| 1713 | {{0, {0,0, 0,0,0,0,0}}}, | ||||||
| 1714 | {{0, {0,0, 0,0,0,0,0}}}, | ||||||
| 1715 | {{0, {0,0, 0,0,0,0,0}}}, | ||||||
| 1716 | {{0, {0,0, 0,0,0,0,0}}}, | ||||||
| 1717 | }, | ||||||
| 1718 | nullptr, // grid_8 | ||||||
| 1719 | nullptr, // grid_16 | ||||||
| 1720 | {0,0,0,0}, // grid_points | ||||||
| 1721 | 0, // output_channels | ||||||
| 1722 | }, | ||||||
| 1723 | |||||||
| 1724 | { 0, 0, 0, 0 }, // an empty CICP | ||||||
| 1725 | |||||||
| 1726 | true, // has_trc | ||||||
| 1727 | true, // has_toXYZD50 | ||||||
| 1728 | false, // has_A2B | ||||||
| 1729 | false, // has B2A | ||||||
| 1730 | false, // has_CICP | ||||||
| 1731 | }; | ||||||
| 1732 | |||||||
| 1733 | return &XYZD50_profile; | ||||||
| 1734 | } | ||||||
| 1735 | |||||||
| 1736 | const skcms_TransferFunction* skcms_sRGB_TransferFunction() { | ||||||
| 1737 | return &skcms_sRGB_profile()->trc[0].parametric; | ||||||
| 1738 | } | ||||||
| 1739 | |||||||
| 1740 | const skcms_TransferFunction* skcms_sRGB_Inverse_TransferFunction() { | ||||||
| 1741 | static const skcms_TransferFunction sRGB_inv = | ||||||
| 1742 | {0.416666657f, 1.137283325f, -0.0f, 12.920000076f, 0.003130805f, -0.054969788f, -0.0f}; | ||||||
| 1743 | return &sRGB_inv; | ||||||
| 1744 | } | ||||||
| 1745 | |||||||
| 1746 | const skcms_TransferFunction* skcms_Identity_TransferFunction() { | ||||||
| 1747 | static const skcms_TransferFunction identity = {1,1,0,0,0,0,0}; | ||||||
| 1748 | return &identity; | ||||||
| 1749 | } | ||||||
| 1750 | |||||||
| 1751 | const uint8_t skcms_252_random_bytes[] = { | ||||||
| 1752 | 8, 179, 128, 204, 253, 38, 134, 184, 68, 102, 32, 138, 99, 39, 169, 215, | ||||||
| 1753 | 119, 26, 3, 223, 95, 239, 52, 132, 114, 74, 81, 234, 97, 116, 244, 205, 30, | ||||||
| 1754 | 154, 173, 12, 51, 159, 122, 153, 61, 226, 236, 178, 229, 55, 181, 220, 191, | ||||||
| 1755 | 194, 160, 126, 168, 82, 131, 18, 180, 245, 163, 22, 246, 69, 235, 252, 57, | ||||||
| 1756 | 108, 14, 6, 152, 240, 255, 171, 242, 20, 227, 177, 238, 96, 85, 16, 211, | ||||||
| 1757 | 70, 200, 149, 155, 146, 127, 145, 100, 151, 109, 19, 165, 208, 195, 164, | ||||||
| 1758 | 137, 254, 182, 248, 64, 201, 45, 209, 5, 147, 207, 210, 113, 162, 83, 225, | ||||||
| 1759 | 9, 31, 15, 231, 115, 37, 58, 53, 24, 49, 197, 56, 120, 172, 48, 21, 214, | ||||||
| 1760 | 129, 111, 11, 50, 187, 196, 34, 60, 103, 71, 144, 47, 203, 77, 80, 232, | ||||||
| 1761 | 140, 222, 250, 206, 166, 247, 139, 249, 221, 72, 106, 27, 199, 117, 54, | ||||||
| 1762 | 219, 135, 118, 40, 79, 41, 251, 46, 93, 212, 92, 233, 148, 28, 121, 63, | ||||||
| 1763 | 123, 158, 105, 59, 29, 42, 143, 23, 0, 107, 176, 87, 104, 183, 156, 193, | ||||||
| 1764 | 189, 90, 188, 65, 190, 17, 198, 7, 186, 161, 1, 124, 78, 125, 170, 133, | ||||||
| 1765 | 174, 218, 67, 157, 75, 101, 89, 217, 62, 33, 141, 228, 25, 35, 91, 230, 4, | ||||||
| 1766 | 2, 13, 73, 86, 167, 237, 84, 243, 44, 185, 66, 130, 110, 150, 142, 216, 88, | ||||||
| 1767 | 112, 36, 224, 136, 202, 76, 94, 98, 175, 213 | ||||||
| 1768 | }; | ||||||
| 1769 | |||||||
| 1770 | bool skcms_ApproximatelyEqualProfiles(const skcms_ICCProfile* A, const skcms_ICCProfile* B) { | ||||||
| 1771 | // Test for exactly equal profiles first. | ||||||
| 1772 | if (A == B || 0 == memcmp(A,B, sizeof(skcms_ICCProfile))) { | ||||||
| 1773 | return true; | ||||||
| 1774 | } | ||||||
| 1775 | |||||||
| 1776 | // For now this is the essentially the same strategy we use in test_only.c | ||||||
| 1777 | // for our skcms_Transform() smoke tests: | ||||||
| 1778 | // 1) transform A to XYZD50 | ||||||
| 1779 | // 2) transform B to XYZD50 | ||||||
| 1780 | // 3) return true if they're similar enough | ||||||
| 1781 | // Our current criterion in 3) is maximum 1 bit error per XYZD50 byte. | ||||||
| 1782 | |||||||
| 1783 | // skcms_252_random_bytes are 252 of a random shuffle of all possible bytes. | ||||||
| 1784 | // 252 is evenly divisible by 3 and 4. Only 192, 10, 241, and 43 are missing. | ||||||
| 1785 | |||||||
| 1786 | // We want to allow otherwise equivalent profiles tagged as grayscale and RGB | ||||||
| 1787 | // to be treated as equal. But CMYK profiles are a totally different ballgame. | ||||||
| 1788 | const auto CMYK = skcms_Signature_CMYK; | ||||||
| 1789 | if ((A->data_color_space == CMYK) != (B->data_color_space == CMYK)) { | ||||||
| 1790 | return false; | ||||||
| 1791 | } | ||||||
| 1792 | |||||||
| 1793 | // Interpret as RGB_888 if data color space is RGB or GRAY, RGBA_8888 if CMYK. | ||||||
| 1794 | // TODO: working with RGBA_8888 either way is probably fastest. | ||||||
| 1795 | skcms_PixelFormat fmt = skcms_PixelFormat_RGB_888; | ||||||
| 1796 | size_t npixels = 84; | ||||||
| 1797 | if (A->data_color_space == skcms_Signature_CMYK) { | ||||||
| 1798 | fmt = skcms_PixelFormat_RGBA_8888; | ||||||
| 1799 | npixels = 63; | ||||||
| 1800 | } | ||||||
| 1801 | |||||||
| 1802 | // TODO: if A or B is a known profile (skcms_sRGB_profile, skcms_XYZD50_profile), | ||||||
| 1803 | // use pre-canned results and skip that skcms_Transform() call? | ||||||
| 1804 | uint8_t dstA[252], | ||||||
| 1805 | dstB[252]; | ||||||
| 1806 | if (!skcms_Transform( | ||||||
| 1807 | skcms_252_random_bytes, fmt, skcms_AlphaFormat_Unpremul, A, | ||||||
| 1808 | dstA, skcms_PixelFormat_RGB_888, skcms_AlphaFormat_Unpremul, skcms_XYZD50_profile(), | ||||||
| 1809 | npixels)) { | ||||||
| 1810 | return false; | ||||||
| 1811 | } | ||||||
| 1812 | if (!skcms_Transform( | ||||||
| 1813 | skcms_252_random_bytes, fmt, skcms_AlphaFormat_Unpremul, B, | ||||||
| 1814 | dstB, skcms_PixelFormat_RGB_888, skcms_AlphaFormat_Unpremul, skcms_XYZD50_profile(), | ||||||
| 1815 | npixels)) { | ||||||
| 1816 | return false; | ||||||
| 1817 | } | ||||||
| 1818 | |||||||
| 1819 | // TODO: make sure this final check has reasonable codegen. | ||||||
| 1820 | for (size_t i = 0; i < 252; i++) { | ||||||
| 1821 | if (abs((int)dstA[i] - (int)dstB[i]) > 1) { | ||||||
| 1822 | return false; | ||||||
| 1823 | } | ||||||
| 1824 | } | ||||||
| 1825 | return true; | ||||||
| 1826 | } | ||||||
| 1827 | |||||||
| 1828 | bool skcms_TRCs_AreApproximateInverse(const skcms_ICCProfile* profile, | ||||||
| 1829 | const skcms_TransferFunction* inv_tf) { | ||||||
| 1830 | if (!profile || !profile->has_trc) { | ||||||
| 1831 | return false; | ||||||
| 1832 | } | ||||||
| 1833 | |||||||
| 1834 | return skcms_AreApproximateInverses(&profile->trc[0], inv_tf) && | ||||||
| 1835 | skcms_AreApproximateInverses(&profile->trc[1], inv_tf) && | ||||||
| 1836 | skcms_AreApproximateInverses(&profile->trc[2], inv_tf); | ||||||
| 1837 | } | ||||||
| 1838 | |||||||
| 1839 | static bool is_zero_to_one(float x) { | ||||||
| 1840 | return 0 <= x && x <= 1; | ||||||
| 1841 | } | ||||||
| 1842 | |||||||
| 1843 | typedef struct { float vals[3]; } skcms_Vector3; | ||||||
| 1844 | |||||||
| 1845 | static skcms_Vector3 mv_mul(const skcms_Matrix3x3* m, const skcms_Vector3* v) { | ||||||
| 1846 | skcms_Vector3 dst = {{0,0,0}}; | ||||||
| 1847 | for (int row = 0; row < 3; ++row) { | ||||||
| 1848 | dst.vals[row] = m->vals[row][0] * v->vals[0] | ||||||
| 1849 | + m->vals[row][1] * v->vals[1] | ||||||
| 1850 | + m->vals[row][2] * v->vals[2]; | ||||||
| 1851 | } | ||||||
| 1852 | return dst; | ||||||
| 1853 | } | ||||||
| 1854 | |||||||
| 1855 | bool skcms_AdaptToXYZD50(float wx, float wy, | ||||||
| 1856 | skcms_Matrix3x3* toXYZD50) { | ||||||
| 1857 | if (!is_zero_to_one(wx) || !is_zero_to_one(wy) || | ||||||
| 1858 | !toXYZD50) { | ||||||
| 1859 | return false; | ||||||
| 1860 | } | ||||||
| 1861 | |||||||
| 1862 | // Assumes that Y is 1.0f. | ||||||
| 1863 | skcms_Vector3 wXYZ = { { wx / wy, 1, (1 - wx - wy) / wy } }; | ||||||
| 1864 | |||||||
| 1865 | // Now convert toXYZ matrix to toXYZD50. | ||||||
| 1866 | skcms_Vector3 wXYZD50 = { { 0.96422f, 1.0f, 0.82521f } }; | ||||||
| 1867 | |||||||
| 1868 | // Calculate the chromatic adaptation matrix. We will use the Bradford method, thus | ||||||
| 1869 | // the matrices below. The Bradford method is used by Adobe and is widely considered | ||||||
| 1870 | // to be the best. | ||||||
| 1871 | skcms_Matrix3x3 xyz_to_lms = {{ | ||||||
| 1872 | { 0.8951f, 0.2664f, -0.1614f }, | ||||||
| 1873 | { -0.7502f, 1.7135f, 0.0367f }, | ||||||
| 1874 | { 0.0389f, -0.0685f, 1.0296f }, | ||||||
| 1875 | }}; | ||||||
| 1876 | skcms_Matrix3x3 lms_to_xyz = {{ | ||||||
| 1877 | { 0.9869929f, -0.1470543f, 0.1599627f }, | ||||||
| 1878 | { 0.4323053f, 0.5183603f, 0.0492912f }, | ||||||
| 1879 | { -0.0085287f, 0.0400428f, 0.9684867f }, | ||||||
| 1880 | }}; | ||||||
| 1881 | |||||||
| 1882 | skcms_Vector3 srcCone = mv_mul(&xyz_to_lms, &wXYZ); | ||||||
| 1883 | skcms_Vector3 dstCone = mv_mul(&xyz_to_lms, &wXYZD50); | ||||||
| 1884 | |||||||
| 1885 | *toXYZD50 = {{ | ||||||
| 1886 | { dstCone.vals[0] / srcCone.vals[0], 0, 0 }, | ||||||
| 1887 | { 0, dstCone.vals[1] / srcCone.vals[1], 0 }, | ||||||
| 1888 | { 0, 0, dstCone.vals[2] / srcCone.vals[2] }, | ||||||
| 1889 | }}; | ||||||
| 1890 | *toXYZD50 = skcms_Matrix3x3_concat(toXYZD50, &xyz_to_lms); | ||||||
| 1891 | *toXYZD50 = skcms_Matrix3x3_concat(&lms_to_xyz, toXYZD50); | ||||||
| 1892 | |||||||
| 1893 | return true; | ||||||
| 1894 | } | ||||||
| 1895 | |||||||
| 1896 | bool skcms_PrimariesToXYZD50(float rx, float ry, | ||||||
| 1897 | float gx, float gy, | ||||||
| 1898 | float bx, float by, | ||||||
| 1899 | float wx, float wy, | ||||||
| 1900 | skcms_Matrix3x3* toXYZD50) { | ||||||
| 1901 | if (!is_zero_to_one(rx) || !is_zero_to_one(ry) || | ||||||
| 1902 | !is_zero_to_one(gx) || !is_zero_to_one(gy) || | ||||||
| 1903 | !is_zero_to_one(bx) || !is_zero_to_one(by) || | ||||||
| 1904 | !is_zero_to_one(wx) || !is_zero_to_one(wy) || | ||||||
| 1905 | !toXYZD50) { | ||||||
| 1906 | return false; | ||||||
| 1907 | } | ||||||
| 1908 | |||||||
| 1909 | // First, we need to convert xy values (primaries) to XYZ. | ||||||
| 1910 | skcms_Matrix3x3 primaries = {{ | ||||||
| 1911 | { rx, gx, bx }, | ||||||
| 1912 | { ry, gy, by }, | ||||||
| 1913 | { 1 - rx - ry, 1 - gx - gy, 1 - bx - by }, | ||||||
| 1914 | }}; | ||||||
| 1915 | skcms_Matrix3x3 primaries_inv; | ||||||
| 1916 | if (!skcms_Matrix3x3_invert(&primaries, &primaries_inv)) { | ||||||
| 1917 | return false; | ||||||
| 1918 | } | ||||||
| 1919 | |||||||
| 1920 | // Assumes that Y is 1.0f. | ||||||
| 1921 | skcms_Vector3 wXYZ = { { wx / wy, 1, (1 - wx - wy) / wy } }; | ||||||
| 1922 | skcms_Vector3 XYZ = mv_mul(&primaries_inv, &wXYZ); | ||||||
| 1923 | |||||||
| 1924 | skcms_Matrix3x3 toXYZ = {{ | ||||||
| 1925 | { XYZ.vals[0], 0, 0 }, | ||||||
| 1926 | { 0, XYZ.vals[1], 0 }, | ||||||
| 1927 | { 0, 0, XYZ.vals[2] }, | ||||||
| 1928 | }}; | ||||||
| 1929 | toXYZ = skcms_Matrix3x3_concat(&primaries, &toXYZ); | ||||||
| 1930 | |||||||
| 1931 | skcms_Matrix3x3 DXtoD50; | ||||||
| 1932 | if (!skcms_AdaptToXYZD50(wx, wy, &DXtoD50)) { | ||||||
| 1933 | return false; | ||||||
| 1934 | } | ||||||
| 1935 | |||||||
| 1936 | *toXYZD50 = skcms_Matrix3x3_concat(&DXtoD50, &toXYZ); | ||||||
| 1937 | return true; | ||||||
| 1938 | } | ||||||
| 1939 | |||||||
| 1940 | |||||||
| 1941 | bool skcms_Matrix3x3_invert(const skcms_Matrix3x3* src, skcms_Matrix3x3* dst) { | ||||||
| 1942 | double a00 = src->vals[0][0], | ||||||
| 1943 | a01 = src->vals[1][0], | ||||||
| 1944 | a02 = src->vals[2][0], | ||||||
| 1945 | a10 = src->vals[0][1], | ||||||
| 1946 | a11 = src->vals[1][1], | ||||||
| 1947 | a12 = src->vals[2][1], | ||||||
| 1948 | a20 = src->vals[0][2], | ||||||
| 1949 | a21 = src->vals[1][2], | ||||||
| 1950 | a22 = src->vals[2][2]; | ||||||
| 1951 | |||||||
| 1952 | double b0 = a00*a11 - a01*a10, | ||||||
| 1953 | b1 = a00*a12 - a02*a10, | ||||||
| 1954 | b2 = a01*a12 - a02*a11, | ||||||
| 1955 | b3 = a20, | ||||||
| 1956 | b4 = a21, | ||||||
| 1957 | b5 = a22; | ||||||
| 1958 | |||||||
| 1959 | double determinant = b0*b5 | ||||||
| 1960 | - b1*b4 | ||||||
| 1961 | + b2*b3; | ||||||
| 1962 | |||||||
| 1963 | if (determinant == 0) { | ||||||
| 1964 | return false; | ||||||
| 1965 | } | ||||||
| 1966 | |||||||
| 1967 | double invdet = 1.0 / determinant; | ||||||
| 1968 | if (invdet > +FLT_MAX3.40282347e+38F || invdet < -FLT_MAX3.40282347e+38F || !isfinitef_((float)invdet)) { | ||||||
| 1969 | return false; | ||||||
| 1970 | } | ||||||
| 1971 | |||||||
| 1972 | b0 *= invdet; | ||||||
| 1973 | b1 *= invdet; | ||||||
| 1974 | b2 *= invdet; | ||||||
| 1975 | b3 *= invdet; | ||||||
| 1976 | b4 *= invdet; | ||||||
| 1977 | b5 *= invdet; | ||||||
| 1978 | |||||||
| 1979 | dst->vals[0][0] = (float)( a11*b5 - a12*b4 ); | ||||||
| 1980 | dst->vals[1][0] = (float)( a02*b4 - a01*b5 ); | ||||||
| 1981 | dst->vals[2][0] = (float)( + b2 ); | ||||||
| 1982 | dst->vals[0][1] = (float)( a12*b3 - a10*b5 ); | ||||||
| 1983 | dst->vals[1][1] = (float)( a00*b5 - a02*b3 ); | ||||||
| 1984 | dst->vals[2][1] = (float)( - b1 ); | ||||||
| 1985 | dst->vals[0][2] = (float)( a10*b4 - a11*b3 ); | ||||||
| 1986 | dst->vals[1][2] = (float)( a01*b3 - a00*b4 ); | ||||||
| 1987 | dst->vals[2][2] = (float)( + b0 ); | ||||||
| 1988 | |||||||
| 1989 | for (int r = 0; r < 3; ++r) | ||||||
| 1990 | for (int c = 0; c < 3; ++c) { | ||||||
| 1991 | if (!isfinitef_(dst->vals[r][c])) { | ||||||
| 1992 | return false; | ||||||
| 1993 | } | ||||||
| 1994 | } | ||||||
| 1995 | return true; | ||||||
| 1996 | } | ||||||
| 1997 | |||||||
| 1998 | skcms_Matrix3x3 skcms_Matrix3x3_concat(const skcms_Matrix3x3* A, const skcms_Matrix3x3* B) { | ||||||
| 1999 | skcms_Matrix3x3 m = { { { 0,0,0 },{ 0,0,0 },{ 0,0,0 } } }; | ||||||
| 2000 | for (int r = 0; r < 3; r++) | ||||||
| 2001 | for (int c = 0; c < 3; c++) { | ||||||
| 2002 | m.vals[r][c] = A->vals[r][0] * B->vals[0][c] | ||||||
| 2003 | + A->vals[r][1] * B->vals[1][c] | ||||||
| 2004 | + A->vals[r][2] * B->vals[2][c]; | ||||||
| 2005 | } | ||||||
| 2006 | return m; | ||||||
| 2007 | } | ||||||
| 2008 | |||||||
| 2009 | #if defined(__clang__1) | ||||||
| 2010 | [[clang::no_sanitize("float-divide-by-zero")]] // Checked for by classify() on the way out. | ||||||
| 2011 | #endif | ||||||
| 2012 | bool skcms_TransferFunction_invert(const skcms_TransferFunction* src, skcms_TransferFunction* dst) { | ||||||
| 2013 | TF_PQish pq; | ||||||
| 2014 | TF_HLGish hlg; | ||||||
| 2015 | switch (classify(*src, &pq, &hlg)) { | ||||||
| 2016 | case skcms_TFType_Invalid: return false; | ||||||
| 2017 | case skcms_TFType_PQ: return false; | ||||||
| 2018 | case skcms_TFType_HLG: return false; | ||||||
| 2019 | case skcms_TFType_sRGBish: break; // handled below | ||||||
| 2020 | |||||||
| 2021 | case skcms_TFType_PQish: | ||||||
| 2022 | *dst = { TFKind_marker(skcms_TFType_PQish), -pq.A, pq.D, 1.0f/pq.F | ||||||
| 2023 | , pq.B, -pq.E, 1.0f/pq.C}; | ||||||
| 2024 | return true; | ||||||
| 2025 | |||||||
| 2026 | case skcms_TFType_HLGish: | ||||||
| 2027 | *dst = { TFKind_marker(skcms_TFType_HLGinvish), 1.0f/hlg.R, 1.0f/hlg.G | ||||||
| 2028 | , 1.0f/hlg.a, hlg.b, hlg.c | ||||||
| 2029 | , hlg.K_minus_1 }; | ||||||
| 2030 | return true; | ||||||
| 2031 | |||||||
| 2032 | case skcms_TFType_HLGinvish: | ||||||
| 2033 | *dst = { TFKind_marker(skcms_TFType_HLGish), 1.0f/hlg.R, 1.0f/hlg.G | ||||||
| 2034 | , 1.0f/hlg.a, hlg.b, hlg.c | ||||||
| 2035 | , hlg.K_minus_1 }; | ||||||
| 2036 | return true; | ||||||
| 2037 | } | ||||||
| 2038 | |||||||
| 2039 | assert (classify(*src) == skcms_TFType_sRGBish)(static_cast <bool> (classify(*src) == skcms_TFType_sRGBish ) ? void (0) : __assert_fail ("classify(*src) == skcms_TFType_sRGBish" , __builtin_FILE (), __builtin_LINE (), __extension__ __PRETTY_FUNCTION__ )); | ||||||
| 2040 | |||||||
| 2041 | // We're inverting this function, solving for x in terms of y. | ||||||
| 2042 | // y = (cx + f) x < d | ||||||
| 2043 | // (ax + b)^g + e x ≥ d | ||||||
| 2044 | // The inverse of this function can be expressed in the same piecewise form. | ||||||
| 2045 | skcms_TransferFunction inv = {0,0,0,0,0,0,0}; | ||||||
| 2046 | |||||||
| 2047 | // We'll start by finding the new threshold inv.d. | ||||||
| 2048 | // In principle we should be able to find that by solving for y at x=d from either side. | ||||||
| 2049 | // (If those two d values aren't the same, it's a discontinuous transfer function.) | ||||||
| 2050 | float d_l = src->c * src->d + src->f, | ||||||
| 2051 | d_r = powf_(src->a * src->d + src->b, src->g) + src->e; | ||||||
| 2052 | if (fabsf_(d_l - d_r) > 1/512.0f) { | ||||||
| 2053 | return false; | ||||||
| 2054 | } | ||||||
| 2055 | inv.d = d_l; // TODO(mtklein): better in practice to choose d_r? | ||||||
| 2056 | |||||||
| 2057 | // When d=0, the linear section collapses to a point. We leave c,d,f all zero in that case. | ||||||
| 2058 | if (inv.d > 0) { | ||||||
| 2059 | // Inverting the linear section is pretty straightfoward: | ||||||
| 2060 | // y = cx + f | ||||||
| 2061 | // y - f = cx | ||||||
| 2062 | // (1/c)y - f/c = x | ||||||
| 2063 | inv.c = 1.0f/src->c; | ||||||
| 2064 | inv.f = -src->f/src->c; | ||||||
| 2065 | } | ||||||
| 2066 | |||||||
| 2067 | // The interesting part is inverting the nonlinear section: | ||||||
| 2068 | // y = (ax + b)^g + e. | ||||||
| 2069 | // y - e = (ax + b)^g | ||||||
| 2070 | // (y - e)^1/g = ax + b | ||||||
| 2071 | // (y - e)^1/g - b = ax | ||||||
| 2072 | // (1/a)(y - e)^1/g - b/a = x | ||||||
| 2073 | // | ||||||
| 2074 | // To make that fit our form, we need to move the (1/a) term inside the exponentiation: | ||||||
| 2075 | // let k = (1/a)^g | ||||||
| 2076 | // (1/a)( y - e)^1/g - b/a = x | ||||||
| 2077 | // (ky - ke)^1/g - b/a = x | ||||||
| 2078 | |||||||
| 2079 | float k = powf_(src->a, -src->g); // (1/a)^g == a^-g | ||||||
| 2080 | inv.g = 1.0f / src->g; | ||||||
| 2081 | inv.a = k; | ||||||
| 2082 | inv.b = -k * src->e; | ||||||
| 2083 | inv.e = -src->b / src->a; | ||||||
| 2084 | |||||||
| 2085 | // We need to enforce the same constraints here that we do when fitting a curve, | ||||||
| 2086 | // a >= 0 and ad+b >= 0. These constraints are checked by classify(), so they're true | ||||||
| 2087 | // of the source function if we're here. | ||||||
| 2088 | |||||||
| 2089 | // Just like when fitting the curve, there's really no way to rescue a < 0. | ||||||
| 2090 | if (inv.a < 0) { | ||||||
| 2091 | return false; | ||||||
| 2092 | } | ||||||
| 2093 | // On the other hand we can rescue an ad+b that's gone slightly negative here. | ||||||
| 2094 | if (inv.a * inv.d + inv.b < 0) { | ||||||
| 2095 | inv.b = -inv.a * inv.d; | ||||||
| 2096 | } | ||||||
| 2097 | |||||||
| 2098 | // That should usually make classify(inv) == sRGBish true, but there are a couple situations | ||||||
| 2099 | // where we might still fail here, like non-finite parameter values. | ||||||
| 2100 | if (classify(inv) != skcms_TFType_sRGBish) { | ||||||
| 2101 | return false; | ||||||
| 2102 | } | ||||||
| 2103 | |||||||
| 2104 | assert (inv.a >= 0)(static_cast <bool> (inv.a >= 0) ? void (0) : __assert_fail ("inv.a >= 0", __builtin_FILE (), __builtin_LINE (), __extension__ __PRETTY_FUNCTION__)); | ||||||
| 2105 | assert (inv.a * inv.d + inv.b >= 0)(static_cast <bool> (inv.a * inv.d + inv.b >= 0) ? void (0) : __assert_fail ("inv.a * inv.d + inv.b >= 0", __builtin_FILE (), __builtin_LINE (), __extension__ __PRETTY_FUNCTION__)); | ||||||
| 2106 | |||||||
| 2107 | // Now in principle we're done. | ||||||
| 2108 | // But to preserve the valuable invariant inv(src(1.0f)) == 1.0f, we'll tweak | ||||||
| 2109 | // e or f of the inverse, depending on which segment contains src(1.0f). | ||||||
| 2110 | float s = skcms_TransferFunction_eval(src, 1.0f); | ||||||
| 2111 | if (!isfinitef_(s)) { | ||||||
| 2112 | return false; | ||||||
| 2113 | } | ||||||
| 2114 | |||||||
| 2115 | float sign = s < 0 ? -1.0f : 1.0f; | ||||||
| 2116 | s *= sign; | ||||||
| 2117 | if (s < inv.d) { | ||||||
| 2118 | inv.f = 1.0f - sign * inv.c * s; | ||||||
| 2119 | } else { | ||||||
| 2120 | inv.e = 1.0f - sign * powf_(inv.a * s + inv.b, inv.g); | ||||||
| 2121 | } | ||||||
| 2122 | |||||||
| 2123 | *dst = inv; | ||||||
| 2124 | return classify(*dst) == skcms_TFType_sRGBish; | ||||||
| 2125 | } | ||||||
| 2126 | |||||||
| 2127 | // ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ // | ||||||
| 2128 | |||||||
| 2129 | // From here below we're approximating an skcms_Curve with an skcms_TransferFunction{g,a,b,c,d,e,f}: | ||||||
| 2130 | // | ||||||
| 2131 | // tf(x) = cx + f x < d | ||||||
| 2132 | // tf(x) = (ax + b)^g + e x ≥ d | ||||||
| 2133 | // | ||||||
| 2134 | // When fitting, we add the additional constraint that both pieces meet at d: | ||||||
| 2135 | // | ||||||
| 2136 | // cd + f = (ad + b)^g + e | ||||||
| 2137 | // | ||||||
| 2138 | // Solving for e and folding it through gives an alternate formulation of the non-linear piece: | ||||||
| 2139 | // | ||||||
| 2140 | // tf(x) = cx + f x < d | ||||||
| 2141 | // tf(x) = (ax + b)^g - (ad + b)^g + cd + f x ≥ d | ||||||
| 2142 | // | ||||||
| 2143 | // Our overall strategy is then: | ||||||
| 2144 | // For a couple tolerances, | ||||||
| 2145 | // - fit_linear(): fit c,d,f iteratively to as many points as our tolerance allows | ||||||
| 2146 | // - invert c,d,f | ||||||
| 2147 | // - fit_nonlinear(): fit g,a,b using Gauss-Newton given those inverted c,d,f | ||||||
| 2148 | // (and by constraint, inverted e) to the inverse of the table. | ||||||
| 2149 | // Return the parameters with least maximum error. | ||||||
| 2150 | // | ||||||
| 2151 | // To run Gauss-Newton to find g,a,b, we'll also need the gradient of the residuals | ||||||
| 2152 | // of round-trip f_inv(x), the inverse of the non-linear piece of f(x). | ||||||
| 2153 | // | ||||||
| 2154 | // let y = Table(x) | ||||||
| 2155 | // r(x) = x - f_inv(y) | ||||||
| 2156 | // | ||||||
| 2157 | // ∂r/∂g = ln(ay + b)*(ay + b)^g | ||||||
| 2158 | // - ln(ad + b)*(ad + b)^g | ||||||
| 2159 | // ∂r/∂a = yg(ay + b)^(g-1) | ||||||
| 2160 | // - dg(ad + b)^(g-1) | ||||||
| 2161 | // ∂r/∂b = g(ay + b)^(g-1) | ||||||
| 2162 | // - g(ad + b)^(g-1) | ||||||
| 2163 | |||||||
| 2164 | // Return the residual of roundtripping skcms_Curve(x) through f_inv(y) with parameters P, | ||||||
| 2165 | // and fill out the gradient of the residual into dfdP. | ||||||
| 2166 | static float rg_nonlinear(float x, | ||||||
| 2167 | const skcms_Curve* curve, | ||||||
| 2168 | const skcms_TransferFunction* tf, | ||||||
| 2169 | float dfdP[3]) { | ||||||
| 2170 | const float y = eval_curve(curve, x); | ||||||
| 2171 | |||||||
| 2172 | const float g = tf->g, a = tf->a, b = tf->b, | ||||||
| 2173 | c = tf->c, d = tf->d, f = tf->f; | ||||||
| 2174 | |||||||
| 2175 | const float Y = fmaxf_(a*y + b, 0.0f), | ||||||
| 2176 | D = a*d + b; | ||||||
| 2177 | assert (D >= 0)(static_cast <bool> (D >= 0) ? void (0) : __assert_fail ("D >= 0", __builtin_FILE (), __builtin_LINE (), __extension__ __PRETTY_FUNCTION__)); | ||||||
| 2178 | |||||||
| 2179 | // The gradient. | ||||||
| 2180 | dfdP[0] = logf_(Y)*powf_(Y, g) | ||||||
| 2181 | - logf_(D)*powf_(D, g); | ||||||
| 2182 | dfdP[1] = y*g*powf_(Y, g-1) | ||||||
| 2183 | - d*g*powf_(D, g-1); | ||||||
| 2184 | dfdP[2] = g*powf_(Y, g-1) | ||||||
| 2185 | - g*powf_(D, g-1); | ||||||
| 2186 | |||||||
| 2187 | // The residual. | ||||||
| 2188 | const float f_inv = powf_(Y, g) | ||||||
| 2189 | - powf_(D, g) | ||||||
| 2190 | + c*d + f; | ||||||
| 2191 | return x - f_inv; | ||||||
| 2192 | } | ||||||
| 2193 | |||||||
| 2194 | static bool gauss_newton_step(const skcms_Curve* curve, | ||||||
| 2195 | skcms_TransferFunction* tf, | ||||||
| 2196 | float x0, float dx, int N) { | ||||||
| 2197 | // We'll sample x from the range [x0,x1] (both inclusive) N times with even spacing. | ||||||
| 2198 | // | ||||||
| 2199 | // Let P = [ tf->g, tf->a, tf->b ] (the three terms that we're adjusting). | ||||||
| 2200 | // | ||||||
| 2201 | // We want to do P' = P + (Jf^T Jf)^-1 Jf^T r(P), | ||||||
| 2202 | // where r(P) is the residual vector | ||||||
| 2203 | // and Jf is the Jacobian matrix of f(), ∂r/∂P. | ||||||
| 2204 | // | ||||||
| 2205 | // Let's review the shape of each of these expressions: | ||||||
| 2206 | // r(P) is [N x 1], a column vector with one entry per value of x tested | ||||||
| 2207 | // Jf is [N x 3], a matrix with an entry for each (x,P) pair | ||||||
| 2208 | // Jf^T is [3 x N], the transpose of Jf | ||||||
| 2209 | // | ||||||
| 2210 | // Jf^T Jf is [3 x N] * [N x 3] == [3 x 3], a 3x3 matrix, | ||||||
| 2211 | // and so is its inverse (Jf^T Jf)^-1 | ||||||
| 2212 | // Jf^T r(P) is [3 x N] * [N x 1] == [3 x 1], a column vector with the same shape as P | ||||||
| 2213 | // | ||||||
| 2214 | // Our implementation strategy to get to the final ∆P is | ||||||
| 2215 | // 1) evaluate Jf^T Jf, call that lhs | ||||||
| 2216 | // 2) evaluate Jf^T r(P), call that rhs | ||||||
| 2217 | // 3) invert lhs | ||||||
| 2218 | // 4) multiply inverse lhs by rhs | ||||||
| 2219 | // | ||||||
| 2220 | // This is a friendly implementation strategy because we don't have to have any | ||||||
| 2221 | // buffers that scale with N, and equally nice don't have to perform any matrix | ||||||
| 2222 | // operations that are variable size. | ||||||
| 2223 | // | ||||||
| 2224 | // Other implementation strategies could trade this off, e.g. evaluating the | ||||||
| 2225 | // pseudoinverse of Jf ( (Jf^T Jf)^-1 Jf^T ) directly, then multiplying that by | ||||||
| 2226 | // the residuals. That would probably require implementing singular value | ||||||
| 2227 | // decomposition, and would create a [3 x N] matrix to be multiplied by the | ||||||
| 2228 | // [N x 1] residual vector, but on the upside I think that'd eliminate the | ||||||
| 2229 | // possibility of this gauss_newton_step() function ever failing. | ||||||
| 2230 | |||||||
| 2231 | // 0) start off with lhs and rhs safely zeroed. | ||||||
| 2232 | skcms_Matrix3x3 lhs = {{ {0,0,0}, {0,0,0}, {0,0,0} }}; | ||||||
| 2233 | skcms_Vector3 rhs = { {0,0,0} }; | ||||||
| 2234 | |||||||
| 2235 | // 1,2) evaluate lhs and evaluate rhs | ||||||
| 2236 | // We want to evaluate Jf only once, but both lhs and rhs involve Jf^T, | ||||||
| 2237 | // so we'll have to update lhs and rhs at the same time. | ||||||
| 2238 | for (int i = 0; i < N; i++) { | ||||||
| 2239 | float x = x0 + static_cast<float>(i)*dx; | ||||||
| 2240 | |||||||
| 2241 | float dfdP[3] = {0,0,0}; | ||||||
| 2242 | float resid = rg_nonlinear(x,curve,tf, dfdP); | ||||||
| 2243 | |||||||
| 2244 | for (int r = 0; r < 3; r++) { | ||||||
| 2245 | for (int c = 0; c < 3; c++) { | ||||||
| 2246 | lhs.vals[r][c] += dfdP[r] * dfdP[c]; | ||||||
| 2247 | } | ||||||
| 2248 | rhs.vals[r] += dfdP[r] * resid; | ||||||
| 2249 | } | ||||||
| 2250 | } | ||||||
| 2251 | |||||||
| 2252 | // If any of the 3 P parameters are unused, this matrix will be singular. | ||||||
| 2253 | // Detect those cases and fix them up to indentity instead, so we can invert. | ||||||
| 2254 | for (int k = 0; k < 3; k++) { | ||||||
| 2255 | if (lhs.vals[0][k]==0 && lhs.vals[1][k]==0 && lhs.vals[2][k]==0 && | ||||||
| 2256 | lhs.vals[k][0]==0 && lhs.vals[k][1]==0 && lhs.vals[k][2]==0) { | ||||||
| 2257 | lhs.vals[k][k] = 1; | ||||||
| 2258 | } | ||||||
| 2259 | } | ||||||
| 2260 | |||||||
| 2261 | // 3) invert lhs | ||||||
| 2262 | skcms_Matrix3x3 lhs_inv; | ||||||
| 2263 | if (!skcms_Matrix3x3_invert(&lhs, &lhs_inv)) { | ||||||
| 2264 | return false; | ||||||
| 2265 | } | ||||||
| 2266 | |||||||
| 2267 | // 4) multiply inverse lhs by rhs | ||||||
| 2268 | skcms_Vector3 dP = mv_mul(&lhs_inv, &rhs); | ||||||
| 2269 | tf->g += dP.vals[0]; | ||||||
| 2270 | tf->a += dP.vals[1]; | ||||||
| 2271 | tf->b += dP.vals[2]; | ||||||
| 2272 | return isfinitef_(tf->g) && isfinitef_(tf->a) && isfinitef_(tf->b); | ||||||
| 2273 | } | ||||||
| 2274 | |||||||
| 2275 | static float max_roundtrip_error_checked(const skcms_Curve* curve, | ||||||
| 2276 | const skcms_TransferFunction* tf_inv) { | ||||||
| 2277 | skcms_TransferFunction tf; | ||||||
| 2278 | if (!skcms_TransferFunction_invert(tf_inv, &tf) || skcms_TFType_sRGBish != classify(tf)) { | ||||||
| 2279 | return INFINITY_; | ||||||
| 2280 | } | ||||||
| 2281 | |||||||
| 2282 | skcms_TransferFunction tf_inv_again; | ||||||
| 2283 | if (!skcms_TransferFunction_invert(&tf, &tf_inv_again)) { | ||||||
| 2284 | return INFINITY_; | ||||||
| 2285 | } | ||||||
| 2286 | |||||||
| 2287 | return skcms_MaxRoundtripError(curve, &tf_inv_again); | ||||||
| 2288 | } | ||||||
| 2289 | |||||||
| 2290 | // Fit the points in [L,N) to the non-linear piece of tf, or return false if we can't. | ||||||
| 2291 | static bool fit_nonlinear(const skcms_Curve* curve, int L, int N, skcms_TransferFunction* tf) { | ||||||
| 2292 | // This enforces a few constraints that are not modeled in gauss_newton_step()'s optimization. | ||||||
| 2293 | auto fixup_tf = [tf]() { | ||||||
| 2294 | // a must be non-negative. That ensures the function is monotonically increasing. | ||||||
| 2295 | // We don't really know how to fix up a if it goes negative. | ||||||
| 2296 | if (tf->a < 0) { | ||||||
| 2297 | return false; | ||||||
| 2298 | } | ||||||
| 2299 | // ad+b must be non-negative. That ensures we don't end up with complex numbers in powf. | ||||||
| 2300 | // We feel just barely not uneasy enough to tweak b so ad+b is zero in this case. | ||||||
| 2301 | if (tf->a * tf->d + tf->b < 0) { | ||||||
| 2302 | tf->b = -tf->a * tf->d; | ||||||
| 2303 | } | ||||||
| 2304 | assert (tf->a >= 0 &&(static_cast <bool> (tf->a >= 0 && tf-> a * tf->d + tf->b >= 0) ? void (0) : __assert_fail ( "tf->a >= 0 && tf->a * tf->d + tf->b >= 0" , __builtin_FILE (), __builtin_LINE (), __extension__ __PRETTY_FUNCTION__ )) | ||||||
| 2305 | tf->a * tf->d + tf->b >= 0)(static_cast <bool> (tf->a >= 0 && tf-> a * tf->d + tf->b >= 0) ? void (0) : __assert_fail ( "tf->a >= 0 && tf->a * tf->d + tf->b >= 0" , __builtin_FILE (), __builtin_LINE (), __extension__ __PRETTY_FUNCTION__ )); | ||||||
| 2306 | |||||||
| 2307 | // cd+f must be ~= (ad+b)^g+e. That ensures the function is continuous. We keep e as a free | ||||||
| 2308 | // parameter so we can guarantee this. | ||||||
| 2309 | tf->e = tf->c*tf->d + tf->f | ||||||
| 2310 | - powf_(tf->a*tf->d + tf->b, tf->g); | ||||||
| 2311 | |||||||
| 2312 | return isfinitef_(tf->e); | ||||||
| 2313 | }; | ||||||
| 2314 | |||||||
| 2315 | if (!fixup_tf()) { | ||||||
| 2316 | return false; | ||||||
| 2317 | } | ||||||
| 2318 | |||||||
| 2319 | // No matter where we start, dx should always represent N even steps from 0 to 1. | ||||||
| 2320 | const float dx = 1.0f / static_cast<float>(N-1); | ||||||
| 2321 | |||||||
| 2322 | skcms_TransferFunction best_tf = *tf; | ||||||
| 2323 | float best_max_error = INFINITY_; | ||||||
| 2324 | |||||||
| 2325 | // Need this or several curves get worse... *sigh* | ||||||
| 2326 | float init_error = max_roundtrip_error_checked(curve, tf); | ||||||
| 2327 | if (init_error < best_max_error) { | ||||||
| 2328 | best_max_error = init_error; | ||||||
| 2329 | best_tf = *tf; | ||||||
| 2330 | } | ||||||
| 2331 | |||||||
| 2332 | // As far as we can tell, 1 Gauss-Newton step won't converge, and 3 steps is no better than 2. | ||||||
| 2333 | for (int j = 0; j < 8; j++) { | ||||||
| 2334 | if (!gauss_newton_step(curve, tf, static_cast<float>(L)*dx, dx, N-L) || !fixup_tf()) { | ||||||
| 2335 | *tf = best_tf; | ||||||
| 2336 | return isfinitef_(best_max_error); | ||||||
| 2337 | } | ||||||
| 2338 | |||||||
| 2339 | float max_error = max_roundtrip_error_checked(curve, tf); | ||||||
| 2340 | if (max_error < best_max_error) { | ||||||
| 2341 | best_max_error = max_error; | ||||||
| 2342 | best_tf = *tf; | ||||||
| 2343 | } | ||||||
| 2344 | } | ||||||
| 2345 | |||||||
| 2346 | *tf = best_tf; | ||||||
| 2347 | return isfinitef_(best_max_error); | ||||||
| 2348 | } | ||||||
| 2349 | |||||||
| 2350 | bool skcms_ApproximateCurve(const skcms_Curve* curve, | ||||||
| 2351 | skcms_TransferFunction* approx, | ||||||
| 2352 | float* max_error) { | ||||||
| 2353 | if (!curve
| ||||||
| 2354 | return false; | ||||||
| 2355 | } | ||||||
| 2356 | |||||||
| 2357 | if (curve->table_entries
| ||||||
| 2358 | // No point approximating an skcms_TransferFunction with an skcms_TransferFunction! | ||||||
| 2359 | return false; | ||||||
| 2360 | } | ||||||
| 2361 | |||||||
| 2362 | if (curve->table_entries == 1 || curve->table_entries > kMaxTableEntries) { | ||||||
| 2363 | // We need at least two points, and must put some reasonable cap on the maximum number. | ||||||
| 2364 | return false; | ||||||
| 2365 | } | ||||||
| 2366 | |||||||
| 2367 | int N = (int)curve->table_entries; | ||||||
| 2368 | const float dx = 1.0f / static_cast<float>(N - 1); | ||||||
| 2369 | |||||||
| 2370 | *max_error = INFINITY_; | ||||||
| 2371 | const float kTolerances[] = { 1.5f / 65535.0f, 1.0f / 512.0f }; | ||||||
| 2372 | for (int t = 0; t < ARRAY_COUNT(kTolerances)(int)(sizeof((kTolerances)) / sizeof(*(kTolerances))); t++) { | ||||||
| 2373 | skcms_TransferFunction tf, | ||||||
| 2374 | tf_inv; | ||||||
| 2375 | |||||||
| 2376 | // It's problematic to fit curves with non-zero f, so always force it to zero explicitly. | ||||||
| 2377 | tf.f = 0.0f; | ||||||
| 2378 | int L = fit_linear(curve, N, kTolerances[t], &tf.c, &tf.d); | ||||||
| 2379 | |||||||
| 2380 | if (L
| ||||||
| 2381 | // If the entire data set was linear, move the coefficients to the nonlinear portion | ||||||
| 2382 | // with G == 1. This lets use a canonical representation with d == 0. | ||||||
| 2383 | tf.g = 1; | ||||||
| 2384 | tf.a = tf.c; | ||||||
| 2385 | tf.b = tf.f; | ||||||
| 2386 | tf.c = tf.d = tf.e = tf.f = 0; | ||||||
| 2387 | } else if (L == N - 1) { | ||||||
| 2388 | // Degenerate case with only two points in the nonlinear segment. Solve directly. | ||||||
| 2389 | tf.g = 1; | ||||||
| 2390 | tf.a = (eval_curve(curve, static_cast<float>(N-1)*dx) - | ||||||
| 2391 | eval_curve(curve, static_cast<float>(N-2)*dx)) | ||||||
| 2392 | / dx; | ||||||
| 2393 | tf.b = eval_curve(curve, static_cast<float>(N-2)*dx) | ||||||
| 2394 | - tf.a * static_cast<float>(N-2)*dx; | ||||||
| 2395 | tf.e = 0; | ||||||
| 2396 | } else { | ||||||
| 2397 | // Start by guessing a gamma-only curve through the midpoint. | ||||||
| 2398 | int mid = (L + N) / 2; | ||||||
| 2399 | float mid_x = static_cast<float>(mid) / static_cast<float>(N - 1); | ||||||
| 2400 | float mid_y = eval_curve(curve, mid_x); | ||||||
| 2401 | tf.g = log2f_(mid_y) / log2f_(mid_x); | ||||||
| 2402 | tf.a = 1; | ||||||
| 2403 | tf.b = 0; | ||||||
| 2404 | tf.e = tf.c*tf.d + tf.f | ||||||
| 2405 | - powf_(tf.a*tf.d + tf.b, tf.g); | ||||||
| 2406 | |||||||
| 2407 | |||||||
| 2408 | if (!skcms_TransferFunction_invert(&tf, &tf_inv) || | ||||||
| 2409 | !fit_nonlinear(curve, L,N, &tf_inv)) { | ||||||
| 2410 | continue; | ||||||
| 2411 | } | ||||||
| 2412 | |||||||
| 2413 | // We fit tf_inv, so calculate tf to keep in sync. | ||||||
| 2414 | // fit_nonlinear() should guarantee invertibility. | ||||||
| 2415 | if (!skcms_TransferFunction_invert(&tf_inv, &tf)) { | ||||||
| 2416 | assert(false)(static_cast <bool> (false) ? void (0) : __assert_fail ( "false", __builtin_FILE (), __builtin_LINE (), __extension__ __PRETTY_FUNCTION__ )); | ||||||
| 2417 | continue; | ||||||
| 2418 | } | ||||||
| 2419 | } | ||||||
| 2420 | |||||||
| 2421 | // We'd better have a sane, sRGB-ish TF by now. | ||||||
| 2422 | // Other non-Bad TFs would be fine, but we know we've only ever tried to fit sRGBish; | ||||||
| 2423 | // anything else is just some accident of math and the way we pun tf.g as a type flag. | ||||||
| 2424 | // fit_nonlinear() should guarantee this, but the special cases may fail this test. | ||||||
| 2425 | if (skcms_TFType_sRGBish != classify(tf)) { | ||||||
| 2426 | continue; | ||||||
| 2427 | } | ||||||
| 2428 | |||||||
| 2429 | // We find our error by roundtripping the table through tf_inv. | ||||||
| 2430 | // | ||||||
| 2431 | // (The most likely use case for this approximation is to be inverted and | ||||||
| 2432 | // used as the transfer function for a destination color space.) | ||||||
| 2433 | // | ||||||
| 2434 | // We've kept tf and tf_inv in sync above, but we can't guarantee that tf is | ||||||
| 2435 | // invertible, so re-verify that here (and use the new inverse for testing). | ||||||
| 2436 | // fit_nonlinear() should guarantee this, but the special cases that don't use | ||||||
| 2437 | // it may fail this test. | ||||||
| 2438 | if (!skcms_TransferFunction_invert(&tf, &tf_inv)) { | ||||||
| 2439 | continue; | ||||||
| 2440 | } | ||||||
| 2441 | |||||||
| 2442 | float err = skcms_MaxRoundtripError(curve, &tf_inv); | ||||||
| 2443 | if (*max_error > err) { | ||||||
| 2444 | *max_error = err; | ||||||
| 2445 | *approx = tf; | ||||||
| 2446 | } | ||||||
| 2447 | } | ||||||
| 2448 | return isfinitef_(*max_error); | ||||||
| 2449 | } | ||||||
| 2450 | |||||||
| 2451 | enum class CpuType { Baseline, HSW, SKX }; | ||||||
| 2452 | |||||||
| 2453 | static CpuType cpu_type() { | ||||||
| 2454 | #if defined(SKCMS_PORTABLE) || !defined(__x86_64__1) || defined(SKCMS_FORCE_BASELINE) | ||||||
| 2455 | return CpuType::Baseline; | ||||||
| 2456 | #elif defined(SKCMS_FORCE_HSW) | ||||||
| 2457 | return CpuType::HSW; | ||||||
| 2458 | #elif defined(SKCMS_FORCE_SKX) | ||||||
| 2459 | return CpuType::SKX; | ||||||
| 2460 | #else | ||||||
| 2461 | static const CpuType type = []{ | ||||||
| 2462 | if (!sAllowRuntimeCPUDetection) { | ||||||
| 2463 | return CpuType::Baseline; | ||||||
| 2464 | } | ||||||
| 2465 | // See http://www.sandpile.org/x86/cpuid.htm | ||||||
| 2466 | |||||||
| 2467 | // First, a basic cpuid(1) lets us check prerequisites for HSW, SKX. | ||||||
| 2468 | uint32_t eax, ebx, ecx, edx; | ||||||
| 2469 | __asm__ __volatile__("cpuid" : "=a"(eax), "=b"(ebx), "=c"(ecx), "=d"(edx) | ||||||
| 2470 | : "0"(1), "2"(0)); | ||||||
| 2471 | if ((edx & (1u<<25)) && // SSE | ||||||
| 2472 | (edx & (1u<<26)) && // SSE2 | ||||||
| 2473 | (ecx & (1u<< 0)) && // SSE3 | ||||||
| 2474 | (ecx & (1u<< 9)) && // SSSE3 | ||||||
| 2475 | (ecx & (1u<<12)) && // FMA (N.B. not used, avoided even) | ||||||
| 2476 | (ecx & (1u<<19)) && // SSE4.1 | ||||||
| 2477 | (ecx & (1u<<20)) && // SSE4.2 | ||||||
| 2478 | (ecx & (1u<<26)) && // XSAVE | ||||||
| 2479 | (ecx & (1u<<27)) && // OSXSAVE | ||||||
| 2480 | (ecx & (1u<<28)) && // AVX | ||||||
| 2481 | (ecx & (1u<<29))) { // F16C | ||||||
| 2482 | |||||||
| 2483 | // Call cpuid(7) to check for AVX2 and AVX-512 bits. | ||||||
| 2484 | __asm__ __volatile__("cpuid" : "=a"(eax), "=b"(ebx), "=c"(ecx), "=d"(edx) | ||||||
| 2485 | : "0"(7), "2"(0)); | ||||||
| 2486 | // eax from xgetbv(0) will tell us whether XMM, YMM, and ZMM state is saved. | ||||||
| 2487 | uint32_t xcr0, dont_need_edx; | ||||||
| 2488 | __asm__ __volatile__("xgetbv" : "=a"(xcr0), "=d"(dont_need_edx) : "c"(0)); | ||||||
| 2489 | |||||||
| 2490 | if ((xcr0 & (1u<<1)) && // XMM register state saved? | ||||||
| 2491 | (xcr0 & (1u<<2)) && // YMM register state saved? | ||||||
| 2492 | (ebx & (1u<<5))) { // AVX2 | ||||||
| 2493 | // At this point we're at least HSW. Continue checking for SKX. | ||||||
| 2494 | if ((xcr0 & (1u<< 5)) && // Opmasks state saved? | ||||||
| 2495 | (xcr0 & (1u<< 6)) && // First 16 ZMM registers saved? | ||||||
| 2496 | (xcr0 & (1u<< 7)) && // High 16 ZMM registers saved? | ||||||
| 2497 | (ebx & (1u<<16)) && // AVX512F | ||||||
| 2498 | (ebx & (1u<<17)) && // AVX512DQ | ||||||
| 2499 | (ebx & (1u<<28)) && // AVX512CD | ||||||
| 2500 | (ebx & (1u<<30)) && // AVX512BW | ||||||
| 2501 | (ebx & (1u<<31))) { // AVX512VL | ||||||
| 2502 | return CpuType::SKX; | ||||||
| 2503 | } | ||||||
| 2504 | return CpuType::HSW; | ||||||
| 2505 | } | ||||||
| 2506 | } | ||||||
| 2507 | return CpuType::Baseline; | ||||||
| 2508 | }(); | ||||||
| 2509 | return type; | ||||||
| 2510 | #endif | ||||||
| 2511 | } | ||||||
| 2512 | |||||||
| 2513 | static bool tf_is_gamma(const skcms_TransferFunction& tf) { | ||||||
| 2514 | return tf.g > 0 && tf.a == 1 && | ||||||
| 2515 | tf.b == 0 && tf.c == 0 && tf.d == 0 && tf.e == 0 && tf.f == 0; | ||||||
| 2516 | } | ||||||
| 2517 | |||||||
| 2518 | struct OpAndArg { | ||||||
| 2519 | Op op; | ||||||
| 2520 | const void* arg; | ||||||
| 2521 | }; | ||||||
| 2522 | |||||||
| 2523 | static OpAndArg select_curve_op(const skcms_Curve* curve, int channel) { | ||||||
| 2524 | struct OpType { | ||||||
| 2525 | Op sGamma, sRGBish, PQish, HLGish, HLGinvish, table; | ||||||
| 2526 | }; | ||||||
| 2527 | static constexpr OpType kOps[] = { | ||||||
| 2528 | { Op::gamma_r, Op::tf_r, Op::pq_r, Op::hlg_r, Op::hlginv_r, Op::table_r }, | ||||||
| 2529 | { Op::gamma_g, Op::tf_g, Op::pq_g, Op::hlg_g, Op::hlginv_g, Op::table_g }, | ||||||
| 2530 | { Op::gamma_b, Op::tf_b, Op::pq_b, Op::hlg_b, Op::hlginv_b, Op::table_b }, | ||||||
| 2531 | { Op::gamma_a, Op::tf_a, Op::pq_a, Op::hlg_a, Op::hlginv_a, Op::table_a }, | ||||||
| 2532 | }; | ||||||
| 2533 | const auto& op = kOps[channel]; | ||||||
| 2534 | |||||||
| 2535 | if (curve->table_entries == 0) { | ||||||
| 2536 | const OpAndArg noop = { Op::load_a8/*doesn't matter*/, nullptr }; | ||||||
| 2537 | |||||||
| 2538 | const skcms_TransferFunction& tf = curve->parametric; | ||||||
| 2539 | |||||||
| 2540 | if (tf_is_gamma(tf)) { | ||||||
| 2541 | return tf.g != 1 ? OpAndArg{op.sGamma, &tf} | ||||||
| 2542 | : noop; | ||||||
| 2543 | } | ||||||
| 2544 | |||||||
| 2545 | switch (classify(tf)) { | ||||||
| 2546 | case skcms_TFType_Invalid: return noop; | ||||||
| 2547 | // TODO(https://issues.skia.org/issues/420956739): Consider adding | ||||||
| 2548 | // support for PQ and HLG. Generally any code that goes through this | ||||||
| 2549 | // path would also want tone mapping too. | ||||||
| 2550 | case skcms_TFType_PQ: return noop; | ||||||
| 2551 | case skcms_TFType_HLG: return noop; | ||||||
| 2552 | case skcms_TFType_sRGBish: return OpAndArg{op.sRGBish, &tf}; | ||||||
| 2553 | case skcms_TFType_PQish: return OpAndArg{op.PQish, &tf}; | ||||||
| 2554 | case skcms_TFType_HLGish: return OpAndArg{op.HLGish, &tf}; | ||||||
| 2555 | case skcms_TFType_HLGinvish: return OpAndArg{op.HLGinvish, &tf}; | ||||||
| 2556 | } | ||||||
| 2557 | } | ||||||
| 2558 | // The table_* ops make this assumption. | ||||||
| 2559 | assert(curve->table_entries <= kMaxTableEntries)(static_cast <bool> (curve->table_entries <= kMaxTableEntries ) ? void (0) : __assert_fail ("curve->table_entries <= kMaxTableEntries" , __builtin_FILE (), __builtin_LINE (), __extension__ __PRETTY_FUNCTION__ )); | ||||||
| 2560 | return OpAndArg{op.table, curve}; | ||||||
| 2561 | } | ||||||
| 2562 | |||||||
| 2563 | // Returns negative if any of the curves are malformed. | ||||||
| 2564 | static int select_curve_ops(const skcms_Curve* curves, int numChannels, OpAndArg* ops) { | ||||||
| 2565 | // We process the channels in reverse order, yielding ops in ABGR order. | ||||||
| 2566 | // (Working backwards allows us to fuse trailing B+G+R ops into a single RGB op.) | ||||||
| 2567 | int cursor = 0; | ||||||
| 2568 | for (int index = numChannels; index-- > 0; ) { | ||||||
| 2569 | if (curves[index].table_entries > kMaxTableEntries) { | ||||||
| 2570 | return -1; | ||||||
| 2571 | } | ||||||
| 2572 | ops[cursor] = select_curve_op(&curves[index], index); | ||||||
| 2573 | if (ops[cursor].arg) { | ||||||
| 2574 | ++cursor; | ||||||
| 2575 | } | ||||||
| 2576 | } | ||||||
| 2577 | |||||||
| 2578 | // Identify separate B+G+R ops and fuse them into a single RGB op. | ||||||
| 2579 | if (cursor >= 3) { | ||||||
| 2580 | struct FusableOps { | ||||||
| 2581 | Op r, g, b, rgb; | ||||||
| 2582 | }; | ||||||
| 2583 | static constexpr FusableOps kFusableOps[] = { | ||||||
| 2584 | {Op::gamma_r, Op::gamma_g, Op::gamma_b, Op::gamma_rgb}, | ||||||
| 2585 | {Op::tf_r, Op::tf_g, Op::tf_b, Op::tf_rgb}, | ||||||
| 2586 | {Op::pq_r, Op::pq_g, Op::pq_b, Op::pq_rgb}, | ||||||
| 2587 | {Op::hlg_r, Op::hlg_g, Op::hlg_b, Op::hlg_rgb}, | ||||||
| 2588 | {Op::hlginv_r, Op::hlginv_g, Op::hlginv_b, Op::hlginv_rgb}, | ||||||
| 2589 | }; | ||||||
| 2590 | |||||||
| 2591 | int posR = cursor - 1; | ||||||
| 2592 | int posG = cursor - 2; | ||||||
| 2593 | int posB = cursor - 3; | ||||||
| 2594 | for (const FusableOps& fusableOp : kFusableOps) { | ||||||
| 2595 | if (ops[posR].op == fusableOp.r && | ||||||
| 2596 | ops[posG].op == fusableOp.g && | ||||||
| 2597 | ops[posB].op == fusableOp.b && | ||||||
| 2598 | (0 == memcmp(ops[posR].arg, ops[posG].arg, sizeof(skcms_TransferFunction))) && | ||||||
| 2599 | (0 == memcmp(ops[posR].arg, ops[posB].arg, sizeof(skcms_TransferFunction)))) { | ||||||
| 2600 | // Fuse the three matching ops into one. | ||||||
| 2601 | ops[posB].op = fusableOp.rgb; | ||||||
| 2602 | cursor -= 2; | ||||||
| 2603 | break; | ||||||
| 2604 | } | ||||||
| 2605 | } | ||||||
| 2606 | } | ||||||
| 2607 | |||||||
| 2608 | return cursor; | ||||||
| 2609 | } | ||||||
| 2610 | |||||||
| 2611 | static size_t bytes_per_pixel(skcms_PixelFormat fmt) { | ||||||
| 2612 | switch (fmt >> 1) { // ignore rgb/bgr | ||||||
| 2613 | case skcms_PixelFormat_A_8 >> 1: return 1; | ||||||
| 2614 | case skcms_PixelFormat_G_8 >> 1: return 1; | ||||||
| 2615 | case skcms_PixelFormat_GA_88 >> 1: return 2; | ||||||
| 2616 | case skcms_PixelFormat_ABGR_4444 >> 1: return 2; | ||||||
| 2617 | case skcms_PixelFormat_RGB_565 >> 1: return 2; | ||||||
| 2618 | case skcms_PixelFormat_RGB_888 >> 1: return 3; | ||||||
| 2619 | case skcms_PixelFormat_RGBA_8888 >> 1: return 4; | ||||||
| 2620 | case skcms_PixelFormat_RGBA_8888_sRGB >> 1: return 4; | ||||||
| 2621 | case skcms_PixelFormat_RGBA_1010102 >> 1: return 4; | ||||||
| 2622 | case skcms_PixelFormat_RGB_101010x_XR >> 1: return 4; | ||||||
| 2623 | case skcms_PixelFormat_RGB_161616LE >> 1: return 6; | ||||||
| 2624 | case skcms_PixelFormat_RGBA_10101010_XR >> 1: return 8; | ||||||
| 2625 | case skcms_PixelFormat_RGBA_16161616LE >> 1: return 8; | ||||||
| 2626 | case skcms_PixelFormat_RGB_161616BE >> 1: return 6; | ||||||
| 2627 | case skcms_PixelFormat_RGBA_16161616BE >> 1: return 8; | ||||||
| 2628 | case skcms_PixelFormat_RGB_hhh_Norm >> 1: return 6; | ||||||
| 2629 | case skcms_PixelFormat_RGBA_hhhh_Norm >> 1: return 8; | ||||||
| 2630 | case skcms_PixelFormat_RGB_hhh >> 1: return 6; | ||||||
| 2631 | case skcms_PixelFormat_RGBA_hhhh >> 1: return 8; | ||||||
| 2632 | case skcms_PixelFormat_RGB_fff >> 1: return 12; | ||||||
| 2633 | case skcms_PixelFormat_RGBA_ffff >> 1: return 16; | ||||||
| 2634 | } | ||||||
| 2635 | assert(false)(static_cast <bool> (false) ? void (0) : __assert_fail ( "false", __builtin_FILE (), __builtin_LINE (), __extension__ __PRETTY_FUNCTION__ )); | ||||||
| 2636 | return 0; | ||||||
| 2637 | } | ||||||
| 2638 | |||||||
| 2639 | // See ITU-T H.273 Table 3 for the full list of codes. | ||||||
| 2640 | const uint8_t kTransferCicpIdPQ = 16; | ||||||
| 2641 | const uint8_t kTransferCicpIdHLG = 18; | ||||||
| 2642 | |||||||
| 2643 | static bool has_cicp_pq_trc(const skcms_ICCProfile* profile) { | ||||||
| 2644 | return profile->has_CICP | ||||||
| 2645 | && profile->CICP.transfer_characteristics == kTransferCicpIdPQ; | ||||||
| 2646 | } | ||||||
| 2647 | |||||||
| 2648 | static bool has_cicp_hlg_trc(const skcms_ICCProfile* profile) { | ||||||
| 2649 | return profile->has_CICP | ||||||
| 2650 | && profile->CICP.transfer_characteristics == kTransferCicpIdHLG; | ||||||
| 2651 | } | ||||||
| 2652 | |||||||
| 2653 | // Set tf to be the PQ transfer function, scaled such that 1.0 will map to 10,000 / 203. | ||||||
| 2654 | static void set_reference_pq_ish_trc(skcms_TransferFunction* tf) { | ||||||
| 2655 | // Initialize such that 1.0 maps to 1.0. | ||||||
| 2656 | skcms_TransferFunction_makePQish(tf, | ||||||
| 2657 | -107/128.0f, 1.0f, 32/2523.0f, 2413/128.0f, -2392/128.0f, 8192/1305.0f); | ||||||
| 2658 | |||||||
| 2659 | // Distribute scaling factor W by scaling A and B with X ^ (1/F): | ||||||
| 2660 | // ((A + Bx^C) / (D + Ex^C))^F * W = ((A + Bx^C) / (D + Ex^C) * W^(1/F))^F | ||||||
| 2661 | // See https://crbug.com/1058580#c32 for discussion. | ||||||
| 2662 | const float w = 10000.0f / 203.0f; | ||||||
| 2663 | const float ws = powf_(w, 1.0f / tf->f); | ||||||
| 2664 | tf->a = ws * tf->a; | ||||||
| 2665 | tf->b = ws * tf->b; | ||||||
| 2666 | } | ||||||
| 2667 | |||||||
| 2668 | // Set tf to be the HLG inverse OETF, scaled such that 1.0 will map to 1.0. | ||||||
| 2669 | // While this is one version of HLG, there are many others. A better version | ||||||
| 2670 | // would be to use the 1,000 nit reference version, but that will require | ||||||
| 2671 | // adding opt-optical transform support. | ||||||
| 2672 | static void set_sdr_hlg_ish_trc(skcms_TransferFunction* tf) { | ||||||
| 2673 | skcms_TransferFunction_makeHLGish(tf, | ||||||
| 2674 | 2.0f, 2.0f, 1/0.17883277f, 0.28466892f, 0.55991073f); | ||||||
| 2675 | tf->f = 1.0f / 12.0f - 1.0f; | ||||||
| 2676 | } | ||||||
| 2677 | |||||||
| 2678 | static bool prep_for_destination(const skcms_ICCProfile* profile, | ||||||
| 2679 | skcms_Matrix3x3* fromXYZD50, | ||||||
| 2680 | skcms_TransferFunction* invR, | ||||||
| 2681 | skcms_TransferFunction* invG, | ||||||
| 2682 | skcms_TransferFunction* invB, | ||||||
| 2683 | bool* dst_using_B2A, | ||||||
| 2684 | bool* dst_using_hlg_ootf) { | ||||||
| 2685 | const bool has_xyzd50 = | ||||||
| 2686 | profile->has_toXYZD50 && | ||||||
| 2687 | skcms_Matrix3x3_invert(&profile->toXYZD50, fromXYZD50); | ||||||
| 2688 | *dst_using_B2A = false; | ||||||
| 2689 | *dst_using_hlg_ootf = false; | ||||||
| 2690 | |||||||
| 2691 | // CICP-specified PQ or HLG transfer functions take precedence. | ||||||
| 2692 | // TODO: Add the ability to parse CICP primaries to not require | ||||||
| 2693 | // the XYZD50 matrix. | ||||||
| 2694 | if (has_cicp_pq_trc(profile) && has_xyzd50) { | ||||||
| 2695 | skcms_TransferFunction trc_pq; | ||||||
| 2696 | set_reference_pq_ish_trc(&trc_pq); | ||||||
| 2697 | skcms_TransferFunction_invert(&trc_pq, invR); | ||||||
| 2698 | skcms_TransferFunction_invert(&trc_pq, invG); | ||||||
| 2699 | skcms_TransferFunction_invert(&trc_pq, invB); | ||||||
| 2700 | return true; | ||||||
| 2701 | } | ||||||
| 2702 | if (has_cicp_hlg_trc(profile) && has_xyzd50) { | ||||||
| 2703 | skcms_TransferFunction trc_hlg; | ||||||
| 2704 | set_sdr_hlg_ish_trc(&trc_hlg); | ||||||
| 2705 | skcms_TransferFunction_invert(&trc_hlg, invR); | ||||||
| 2706 | skcms_TransferFunction_invert(&trc_hlg, invG); | ||||||
| 2707 | skcms_TransferFunction_invert(&trc_hlg, invB); | ||||||
| 2708 | *dst_using_hlg_ootf = true; | ||||||
| 2709 | return true; | ||||||
| 2710 | } | ||||||
| 2711 | |||||||
| 2712 | // Then prefer the B2A transformation. | ||||||
| 2713 | // skcms_Transform() supports B2A destinations. | ||||||
| 2714 | if (profile->has_B2A) { | ||||||
| 2715 | *dst_using_B2A = true; | ||||||
| 2716 | return true; | ||||||
| 2717 | } | ||||||
| 2718 | |||||||
| 2719 | // Finally use parametric transfer functions. | ||||||
| 2720 | // TODO: Reject non sRGB-ish transfer functions here. | ||||||
| 2721 | return has_xyzd50 | ||||||
| 2722 | && profile->has_trc | ||||||
| 2723 | && profile->trc[0].table_entries == 0 | ||||||
| 2724 | && profile->trc[1].table_entries == 0 | ||||||
| 2725 | && profile->trc[2].table_entries == 0 | ||||||
| 2726 | && skcms_TransferFunction_invert(&profile->trc[0].parametric, invR) | ||||||
| 2727 | && skcms_TransferFunction_invert(&profile->trc[1].parametric, invG) | ||||||
| 2728 | && skcms_TransferFunction_invert(&profile->trc[2].parametric, invB); | ||||||
| 2729 | } | ||||||
| 2730 | |||||||
| 2731 | bool skcms_Transform(const void* src, | ||||||
| 2732 | skcms_PixelFormat srcFmt, | ||||||
| 2733 | skcms_AlphaFormat srcAlpha, | ||||||
| 2734 | const skcms_ICCProfile* srcProfile, | ||||||
| 2735 | void* dst, | ||||||
| 2736 | skcms_PixelFormat dstFmt, | ||||||
| 2737 | skcms_AlphaFormat dstAlpha, | ||||||
| 2738 | const skcms_ICCProfile* dstProfile, | ||||||
| 2739 | size_t nz) { | ||||||
| 2740 | const size_t dst_bpp = bytes_per_pixel(dstFmt), | ||||||
| 2741 | src_bpp = bytes_per_pixel(srcFmt); | ||||||
| 2742 | // Let's just refuse if the request is absurdly big. | ||||||
| 2743 | if (nz * dst_bpp > INT_MAX2147483647 || nz * src_bpp > INT_MAX2147483647) { | ||||||
| 2744 | return false; | ||||||
| 2745 | } | ||||||
| 2746 | int n = (int)nz; | ||||||
| 2747 | |||||||
| 2748 | // Null profiles default to sRGB. Passing null for both is handy when doing format conversion. | ||||||
| 2749 | if (!srcProfile) { | ||||||
| 2750 | srcProfile = skcms_sRGB_profile(); | ||||||
| 2751 | } | ||||||
| 2752 | if (!dstProfile) { | ||||||
| 2753 | dstProfile = skcms_sRGB_profile(); | ||||||
| 2754 | } | ||||||
| 2755 | |||||||
| 2756 | // We can't transform in place unless the PixelFormats are the same size. | ||||||
| 2757 | if (dst == src && dst_bpp != src_bpp) { | ||||||
| 2758 | return false; | ||||||
| 2759 | } | ||||||
| 2760 | // TODO: more careful alias rejection (like, dst == src + 1)? | ||||||
| 2761 | |||||||
| 2762 | Op program[32]; | ||||||
| 2763 | const void* context[32]; | ||||||
| 2764 | |||||||
| 2765 | Op* ops = program; | ||||||
| 2766 | const void** contexts = context; | ||||||
| 2767 | |||||||
| 2768 | auto add_op = [&](Op o) { | ||||||
| 2769 | *ops++ = o; | ||||||
| 2770 | *contexts++ = nullptr; | ||||||
| 2771 | }; | ||||||
| 2772 | |||||||
| 2773 | auto add_op_ctx = [&](Op o, const void* c) { | ||||||
| 2774 | *ops++ = o; | ||||||
| 2775 | *contexts++ = c; | ||||||
| 2776 | }; | ||||||
| 2777 | |||||||
| 2778 | auto add_curve_ops = [&](const skcms_Curve* curves, int numChannels) -> bool { | ||||||
| 2779 | OpAndArg oa[4]; | ||||||
| 2780 | assert(numChannels <= ARRAY_COUNT(oa))(static_cast <bool> (numChannels <= (int)(sizeof((oa )) / sizeof(*(oa)))) ? void (0) : __assert_fail ("numChannels <= ARRAY_COUNT(oa)" , __builtin_FILE (), __builtin_LINE (), __extension__ __PRETTY_FUNCTION__ )); | ||||||
| 2781 | |||||||
| 2782 | int numOps = select_curve_ops(curves, numChannels, oa); | ||||||
| 2783 | if (numOps < 0) { | ||||||
| 2784 | return false; | ||||||
| 2785 | } | ||||||
| 2786 | |||||||
| 2787 | for (int i = 0; i < numOps; ++i) { | ||||||
| 2788 | add_op_ctx(oa[i].op, oa[i].arg); | ||||||
| 2789 | } | ||||||
| 2790 | return true; | ||||||
| 2791 | }; | ||||||
| 2792 | |||||||
| 2793 | // If the source has a TRC that is specified by CICP and not the TRC | ||||||
| 2794 | // entries, then store it here for future use. | ||||||
| 2795 | skcms_TransferFunction src_cicp_trc; | ||||||
| 2796 | |||||||
| 2797 | // These are always parametric curves of some sort. | ||||||
| 2798 | skcms_Curve dst_curves[3]; | ||||||
| 2799 | dst_curves[0].table_entries = | ||||||
| 2800 | dst_curves[1].table_entries = | ||||||
| 2801 | dst_curves[2].table_entries = 0; | ||||||
| 2802 | |||||||
| 2803 | // This will store the XYZD50 to destination gamut conversion matrix, if it is needed. | ||||||
| 2804 | skcms_Matrix3x3 dst_from_xyz; | ||||||
| 2805 | |||||||
| 2806 | // This will store the full source to destination gamut conversion matrix, if it is needed. | ||||||
| 2807 | skcms_Matrix3x3 dst_from_src; | ||||||
| 2808 | |||||||
| 2809 | switch (srcFmt >> 1) { | ||||||
| 2810 | default: return false; | ||||||
| 2811 | case skcms_PixelFormat_A_8 >> 1: add_op(Op::load_a8); break; | ||||||
| 2812 | case skcms_PixelFormat_G_8 >> 1: add_op(Op::load_g8); break; | ||||||
| 2813 | case skcms_PixelFormat_GA_88 >> 1: add_op(Op::load_ga88); break; | ||||||
| 2814 | case skcms_PixelFormat_ABGR_4444 >> 1: add_op(Op::load_4444); break; | ||||||
| 2815 | case skcms_PixelFormat_RGB_565 >> 1: add_op(Op::load_565); break; | ||||||
| 2816 | case skcms_PixelFormat_RGB_888 >> 1: add_op(Op::load_888); break; | ||||||
| 2817 | case skcms_PixelFormat_RGBA_8888 >> 1: add_op(Op::load_8888); break; | ||||||
| 2818 | case skcms_PixelFormat_RGBA_1010102 >> 1: add_op(Op::load_1010102); break; | ||||||
| 2819 | case skcms_PixelFormat_RGB_101010x_XR >> 1: add_op(Op::load_101010x_XR); break; | ||||||
| 2820 | case skcms_PixelFormat_RGBA_10101010_XR >> 1: add_op(Op::load_10101010_XR); break; | ||||||
| 2821 | case skcms_PixelFormat_RGB_161616LE >> 1: add_op(Op::load_161616LE); break; | ||||||
| 2822 | case skcms_PixelFormat_RGBA_16161616LE >> 1: add_op(Op::load_16161616LE); break; | ||||||
| 2823 | case skcms_PixelFormat_RGB_161616BE >> 1: add_op(Op::load_161616BE); break; | ||||||
| 2824 | case skcms_PixelFormat_RGBA_16161616BE >> 1: add_op(Op::load_16161616BE); break; | ||||||
| 2825 | case skcms_PixelFormat_RGB_hhh_Norm >> 1: add_op(Op::load_hhh); break; | ||||||
| 2826 | case skcms_PixelFormat_RGBA_hhhh_Norm >> 1: add_op(Op::load_hhhh); break; | ||||||
| 2827 | case skcms_PixelFormat_RGB_hhh >> 1: add_op(Op::load_hhh); break; | ||||||
| 2828 | case skcms_PixelFormat_RGBA_hhhh >> 1: add_op(Op::load_hhhh); break; | ||||||
| 2829 | case skcms_PixelFormat_RGB_fff >> 1: add_op(Op::load_fff); break; | ||||||
| 2830 | case skcms_PixelFormat_RGBA_ffff >> 1: add_op(Op::load_ffff); break; | ||||||
| 2831 | |||||||
| 2832 | case skcms_PixelFormat_RGBA_8888_sRGB >> 1: | ||||||
| 2833 | add_op(Op::load_8888); | ||||||
| 2834 | add_op_ctx(Op::tf_rgb, skcms_sRGB_TransferFunction()); | ||||||
| 2835 | break; | ||||||
| 2836 | } | ||||||
| 2837 | if (srcFmt == skcms_PixelFormat_RGB_hhh_Norm || | ||||||
| 2838 | srcFmt == skcms_PixelFormat_RGBA_hhhh_Norm) { | ||||||
| 2839 | add_op(Op::clamp); | ||||||
| 2840 | } | ||||||
| 2841 | if (srcFmt & 1) { | ||||||
| 2842 | add_op(Op::swap_rb); | ||||||
| 2843 | } | ||||||
| 2844 | skcms_ICCProfile gray_dst_profile; | ||||||
| 2845 | switch (dstFmt >> 1) { | ||||||
| 2846 | case skcms_PixelFormat_G_8: | ||||||
| 2847 | case skcms_PixelFormat_GA_88: | ||||||
| 2848 | // When transforming to gray, stop at XYZ (by setting toXYZ to identity), then transform | ||||||
| 2849 | // luminance (Y) by the destination transfer function. | ||||||
| 2850 | gray_dst_profile = *dstProfile; | ||||||
| 2851 | skcms_SetXYZD50(&gray_dst_profile, &skcms_XYZD50_profile()->toXYZD50); | ||||||
| 2852 | dstProfile = &gray_dst_profile; | ||||||
| 2853 | break; | ||||||
| 2854 | default: | ||||||
| 2855 | break; | ||||||
| 2856 | } | ||||||
| 2857 | |||||||
| 2858 | if (srcProfile->data_color_space == skcms_Signature_CMYK) { | ||||||
| 2859 | // Photoshop creates CMYK images as inverse CMYK. | ||||||
| 2860 | // These happen to be the only ones we've _ever_ seen. | ||||||
| 2861 | add_op(Op::invert); | ||||||
| 2862 | // With CMYK, ignore the alpha type, to avoid changing K or conflating CMY with K. | ||||||
| 2863 | srcAlpha = skcms_AlphaFormat_Unpremul; | ||||||
| 2864 | } | ||||||
| 2865 | |||||||
| 2866 | if (srcAlpha == skcms_AlphaFormat_Opaque) { | ||||||
| 2867 | add_op(Op::force_opaque); | ||||||
| 2868 | } else if (srcAlpha == skcms_AlphaFormat_PremulAsEncoded) { | ||||||
| 2869 | add_op(Op::unpremul); | ||||||
| 2870 | } | ||||||
| 2871 | |||||||
| 2872 | if (dstProfile != srcProfile) { | ||||||
| 2873 | |||||||
| 2874 | // Track whether or not the A2B or B2A transforms are used. the CICP | ||||||
| 2875 | // values take precedence over A2B and B2A. | ||||||
| 2876 | bool src_using_A2B = false; | ||||||
| 2877 | bool src_using_hlg_ootf = false; | ||||||
| 2878 | bool dst_using_B2A = false; | ||||||
| 2879 | bool dst_using_hlg_ootf = false; | ||||||
| 2880 | |||||||
| 2881 | if (!prep_for_destination(dstProfile, | ||||||
| 2882 | &dst_from_xyz, | ||||||
| 2883 | &dst_curves[0].parametric, | ||||||
| 2884 | &dst_curves[1].parametric, | ||||||
| 2885 | &dst_curves[2].parametric, | ||||||
| 2886 | &dst_using_B2A, | ||||||
| 2887 | &dst_using_hlg_ootf)) { | ||||||
| 2888 | return false; | ||||||
| 2889 | } | ||||||
| 2890 | |||||||
| 2891 | if (has_cicp_pq_trc(srcProfile) && srcProfile->has_toXYZD50) { | ||||||
| 2892 | set_reference_pq_ish_trc(&src_cicp_trc); | ||||||
| 2893 | add_op_ctx(Op::pq_rgb, &src_cicp_trc); | ||||||
| 2894 | } else if (has_cicp_hlg_trc(srcProfile) && srcProfile->has_toXYZD50) { | ||||||
| 2895 | src_using_hlg_ootf = true; | ||||||
| 2896 | set_sdr_hlg_ish_trc(&src_cicp_trc); | ||||||
| 2897 | add_op_ctx(Op::hlg_rgb, &src_cicp_trc); | ||||||
| 2898 | } else if (srcProfile->has_A2B) { | ||||||
| 2899 | src_using_A2B = true; | ||||||
| 2900 | if (srcProfile->A2B.input_channels) { | ||||||
| 2901 | if (!add_curve_ops(srcProfile->A2B.input_curves, | ||||||
| 2902 | (int)srcProfile->A2B.input_channels)) { | ||||||
| 2903 | return false; | ||||||
| 2904 | } | ||||||
| 2905 | add_op(Op::clamp); | ||||||
| 2906 | add_op_ctx(Op::clut_A2B, &srcProfile->A2B); | ||||||
| 2907 | } | ||||||
| 2908 | |||||||
| 2909 | if (srcProfile->A2B.matrix_channels == 3) { | ||||||
| 2910 | if (!add_curve_ops(srcProfile->A2B.matrix_curves, /*numChannels=*/3)) { | ||||||
| 2911 | return false; | ||||||
| 2912 | } | ||||||
| 2913 | |||||||
| 2914 | static const skcms_Matrix3x4 I = {{ | ||||||
| 2915 | {1,0,0,0}, | ||||||
| 2916 | {0,1,0,0}, | ||||||
| 2917 | {0,0,1,0}, | ||||||
| 2918 | }}; | ||||||
| 2919 | if (0 != memcmp(&I, &srcProfile->A2B.matrix, sizeof(I))) { | ||||||
| 2920 | add_op_ctx(Op::matrix_3x4, &srcProfile->A2B.matrix); | ||||||
| 2921 | } | ||||||
| 2922 | } | ||||||
| 2923 | |||||||
| 2924 | if (srcProfile->A2B.output_channels == 3) { | ||||||
| 2925 | if (!add_curve_ops(srcProfile->A2B.output_curves, /*numChannels=*/3)) { | ||||||
| 2926 | return false; | ||||||
| 2927 | } | ||||||
| 2928 | } | ||||||
| 2929 | |||||||
| 2930 | if (srcProfile->pcs == skcms_Signature_Lab) { | ||||||
| 2931 | add_op(Op::lab_to_xyz); | ||||||
| 2932 | } | ||||||
| 2933 | |||||||
| 2934 | } else if (srcProfile->has_trc && srcProfile->has_toXYZD50) { | ||||||
| 2935 | if (!add_curve_ops(srcProfile->trc, /*numChannels=*/3)) { | ||||||
| 2936 | return false; | ||||||
| 2937 | } | ||||||
| 2938 | } else { | ||||||
| 2939 | return false; | ||||||
| 2940 | } | ||||||
| 2941 | |||||||
| 2942 | // A2B sources are in XYZD50 by now, but TRC sources are still in their original gamut. | ||||||
| 2943 | assert (srcProfile->has_A2B || srcProfile->has_toXYZD50)(static_cast <bool> (srcProfile->has_A2B || srcProfile ->has_toXYZD50) ? void (0) : __assert_fail ("srcProfile->has_A2B || srcProfile->has_toXYZD50" , __builtin_FILE (), __builtin_LINE (), __extension__ __PRETTY_FUNCTION__ )); | ||||||
| 2944 | |||||||
| 2945 | if (dst_using_B2A) { | ||||||
| 2946 | // B2A needs its input in XYZD50, so transform TRC sources now. | ||||||
| 2947 | if (!src_using_A2B) { | ||||||
| 2948 | add_op_ctx(Op::matrix_3x3, &srcProfile->toXYZD50); | ||||||
| 2949 | // Apply the HLG OOTF in XYZD50 space, if needed. | ||||||
| 2950 | if (src_using_hlg_ootf) { | ||||||
| 2951 | add_op(Op::hlg_ootf_scale); | ||||||
| 2952 | } | ||||||
| 2953 | } | ||||||
| 2954 | |||||||
| 2955 | if (dstProfile->pcs == skcms_Signature_Lab) { | ||||||
| 2956 | add_op(Op::xyz_to_lab); | ||||||
| 2957 | } | ||||||
| 2958 | |||||||
| 2959 | if (dstProfile->B2A.input_channels == 3) { | ||||||
| 2960 | if (!add_curve_ops(dstProfile->B2A.input_curves, /*numChannels=*/3)) { | ||||||
| 2961 | return false; | ||||||
| 2962 | } | ||||||
| 2963 | } | ||||||
| 2964 | |||||||
| 2965 | if (dstProfile->B2A.matrix_channels == 3) { | ||||||
| 2966 | static const skcms_Matrix3x4 I = {{ | ||||||
| 2967 | {1,0,0,0}, | ||||||
| 2968 | {0,1,0,0}, | ||||||
| 2969 | {0,0,1,0}, | ||||||
| 2970 | }}; | ||||||
| 2971 | if (0 != memcmp(&I, &dstProfile->B2A.matrix, sizeof(I))) { | ||||||
| 2972 | add_op_ctx(Op::matrix_3x4, &dstProfile->B2A.matrix); | ||||||
| 2973 | } | ||||||
| 2974 | |||||||
| 2975 | if (!add_curve_ops(dstProfile->B2A.matrix_curves, /*numChannels=*/3)) { | ||||||
| 2976 | return false; | ||||||
| 2977 | } | ||||||
| 2978 | } | ||||||
| 2979 | |||||||
| 2980 | if (dstProfile->B2A.output_channels) { | ||||||
| 2981 | add_op(Op::clamp); | ||||||
| 2982 | add_op_ctx(Op::clut_B2A, &dstProfile->B2A); | ||||||
| 2983 | |||||||
| 2984 | if (!add_curve_ops(dstProfile->B2A.output_curves, | ||||||
| 2985 | (int)dstProfile->B2A.output_channels)) { | ||||||
| 2986 | return false; | ||||||
| 2987 | } | ||||||
| 2988 | } | ||||||
| 2989 | } else { | ||||||
| 2990 | // This is a TRC destination. | ||||||
| 2991 | |||||||
| 2992 | // Transform to the destination gamut. | ||||||
| 2993 | if (src_using_hlg_ootf != dst_using_hlg_ootf) { | ||||||
| 2994 | // If just the src or the dst has an HLG OOTF then we will apply the OOTF in XYZD50 | ||||||
| 2995 | // space. If both the src and dst has an HLG OOTF then they will cancel. | ||||||
| 2996 | if (!src_using_A2B) { | ||||||
| 2997 | add_op_ctx(Op::matrix_3x3, &srcProfile->toXYZD50); | ||||||
| 2998 | } | ||||||
| 2999 | if (src_using_hlg_ootf) { | ||||||
| 3000 | add_op(Op::hlg_ootf_scale); | ||||||
| 3001 | } | ||||||
| 3002 | if (dst_using_hlg_ootf) { | ||||||
| 3003 | add_op(Op::hlginv_ootf_scale); | ||||||
| 3004 | } | ||||||
| 3005 | add_op_ctx(Op::matrix_3x3, &dst_from_xyz); | ||||||
| 3006 | } else if (src_using_A2B) { | ||||||
| 3007 | // If the source is A2B then we are already in XYZD50. Just apply the xyz->dst | ||||||
| 3008 | // matrix. | ||||||
| 3009 | add_op_ctx(Op::matrix_3x3, &dst_from_xyz); | ||||||
| 3010 | } else { | ||||||
| 3011 | const skcms_Matrix3x3* to_xyz = &srcProfile->toXYZD50; | ||||||
| 3012 | // There's a chance the source and destination gamuts are identical, | ||||||
| 3013 | // in which case we can skip the gamut transform. | ||||||
| 3014 | if (0 != memcmp(&dstProfile->toXYZD50, to_xyz, sizeof(skcms_Matrix3x3))) { | ||||||
| 3015 | // Concat the entire gamut transform into dst_from_src. | ||||||
| 3016 | dst_from_src = skcms_Matrix3x3_concat(&dst_from_xyz, to_xyz); | ||||||
| 3017 | add_op_ctx(Op::matrix_3x3, &dst_from_src); | ||||||
| 3018 | } | ||||||
| 3019 | } | ||||||
| 3020 | |||||||
| 3021 | // Encode back to dst RGB using its parametric transfer functions. | ||||||
| 3022 | OpAndArg oa[3]; | ||||||
| 3023 | int numOps = select_curve_ops(dst_curves, /*numChannels=*/3, oa); | ||||||
| 3024 | // All dst_curves should be parametric, not table-based, so select_curve_ops should | ||||||
| 3025 | // not fail. | ||||||
| 3026 | assert(numOps >= 0)(static_cast <bool> (numOps >= 0) ? void (0) : __assert_fail ("numOps >= 0", __builtin_FILE (), __builtin_LINE (), __extension__ __PRETTY_FUNCTION__)); | ||||||
| 3027 | for (int index = 0; index < numOps; ++index) { | ||||||
| 3028 | assert(oa[index].op != Op::table_r &&(static_cast <bool> (oa[index].op != Op::table_r && oa[index].op != Op::table_g && oa[index].op != Op::table_b && oa[index].op != Op::table_a) ? void (0) : __assert_fail ("oa[index].op != Op::table_r && oa[index].op != Op::table_g && oa[index].op != Op::table_b && oa[index].op != Op::table_a" , __builtin_FILE (), __builtin_LINE (), __extension__ __PRETTY_FUNCTION__ )) | ||||||
| 3029 | oa[index].op != Op::table_g &&(static_cast <bool> (oa[index].op != Op::table_r && oa[index].op != Op::table_g && oa[index].op != Op::table_b && oa[index].op != Op::table_a) ? void (0) : __assert_fail ("oa[index].op != Op::table_r && oa[index].op != Op::table_g && oa[index].op != Op::table_b && oa[index].op != Op::table_a" , __builtin_FILE (), __builtin_LINE (), __extension__ __PRETTY_FUNCTION__ )) | ||||||
| 3030 | oa[index].op != Op::table_b &&(static_cast <bool> (oa[index].op != Op::table_r && oa[index].op != Op::table_g && oa[index].op != Op::table_b && oa[index].op != Op::table_a) ? void (0) : __assert_fail ("oa[index].op != Op::table_r && oa[index].op != Op::table_g && oa[index].op != Op::table_b && oa[index].op != Op::table_a" , __builtin_FILE (), __builtin_LINE (), __extension__ __PRETTY_FUNCTION__ )) | ||||||
| 3031 | oa[index].op != Op::table_a)(static_cast <bool> (oa[index].op != Op::table_r && oa[index].op != Op::table_g && oa[index].op != Op::table_b && oa[index].op != Op::table_a) ? void (0) : __assert_fail ("oa[index].op != Op::table_r && oa[index].op != Op::table_g && oa[index].op != Op::table_b && oa[index].op != Op::table_a" , __builtin_FILE (), __builtin_LINE (), __extension__ __PRETTY_FUNCTION__ )); | ||||||
| 3032 | add_op_ctx(oa[index].op, oa[index].arg); | ||||||
| 3033 | } | ||||||
| 3034 | } | ||||||
| 3035 | } | ||||||
| 3036 | |||||||
| 3037 | // Clamp here before premul to make sure we're clamping to normalized values _and_ gamut, | ||||||
| 3038 | // not just to values that fit in [0,1]. | ||||||
| 3039 | // | ||||||
| 3040 | // E.g. r = 1.1, a = 0.5 would fit fine in fixed point after premul (ra=0.55,a=0.5), | ||||||
| 3041 | // but would be carrying r > 1, which is really unexpected for downstream consumers. | ||||||
| 3042 | if (dstFmt < skcms_PixelFormat_RGB_hhh) { | ||||||
| 3043 | add_op(Op::clamp); | ||||||
| 3044 | } | ||||||
| 3045 | |||||||
| 3046 | if (dstProfile->data_color_space == skcms_Signature_CMYK) { | ||||||
| 3047 | // Photoshop creates CMYK images as inverse CMYK. | ||||||
| 3048 | // These happen to be the only ones we've _ever_ seen. | ||||||
| 3049 | add_op(Op::invert); | ||||||
| 3050 | |||||||
| 3051 | // CMYK has no alpha channel, so make sure dstAlpha is a no-op. | ||||||
| 3052 | dstAlpha = skcms_AlphaFormat_Unpremul; | ||||||
| 3053 | } | ||||||
| 3054 | |||||||
| 3055 | if (dstAlpha == skcms_AlphaFormat_Opaque) { | ||||||
| 3056 | add_op(Op::force_opaque); | ||||||
| 3057 | } else if (dstAlpha == skcms_AlphaFormat_PremulAsEncoded) { | ||||||
| 3058 | add_op(Op::premul); | ||||||
| 3059 | } | ||||||
| 3060 | if (dstFmt & 1) { | ||||||
| 3061 | add_op(Op::swap_rb); | ||||||
| 3062 | } | ||||||
| 3063 | switch (dstFmt >> 1) { | ||||||
| 3064 | default: return false; | ||||||
| 3065 | case skcms_PixelFormat_A_8 >> 1: add_op(Op::store_a8); break; | ||||||
| 3066 | case skcms_PixelFormat_G_8 >> 1: add_op(Op::store_g8); break; | ||||||
| 3067 | case skcms_PixelFormat_GA_88 >> 1: add_op(Op::store_ga88); break; | ||||||
| 3068 | case skcms_PixelFormat_ABGR_4444 >> 1: add_op(Op::store_4444); break; | ||||||
| 3069 | case skcms_PixelFormat_RGB_565 >> 1: add_op(Op::store_565); break; | ||||||
| 3070 | case skcms_PixelFormat_RGB_888 >> 1: add_op(Op::store_888); break; | ||||||
| 3071 | case skcms_PixelFormat_RGBA_8888 >> 1: add_op(Op::store_8888); break; | ||||||
| 3072 | case skcms_PixelFormat_RGBA_1010102 >> 1: add_op(Op::store_1010102); break; | ||||||
| 3073 | case skcms_PixelFormat_RGB_161616LE >> 1: add_op(Op::store_161616LE); break; | ||||||
| 3074 | case skcms_PixelFormat_RGBA_16161616LE >> 1: add_op(Op::store_16161616LE); break; | ||||||
| 3075 | case skcms_PixelFormat_RGB_161616BE >> 1: add_op(Op::store_161616BE); break; | ||||||
| 3076 | case skcms_PixelFormat_RGBA_16161616BE >> 1: add_op(Op::store_16161616BE); break; | ||||||
| 3077 | case skcms_PixelFormat_RGB_hhh_Norm >> 1: add_op(Op::store_hhh); break; | ||||||
| 3078 | case skcms_PixelFormat_RGBA_hhhh_Norm >> 1: add_op(Op::store_hhhh); break; | ||||||
| 3079 | case skcms_PixelFormat_RGB_101010x_XR >> 1: add_op(Op::store_101010x_XR); break; | ||||||
| 3080 | case skcms_PixelFormat_RGBA_10101010_XR >> 1: add_op(Op::store_10101010_XR); break; | ||||||
| 3081 | case skcms_PixelFormat_RGB_hhh >> 1: add_op(Op::store_hhh); break; | ||||||
| 3082 | case skcms_PixelFormat_RGBA_hhhh >> 1: add_op(Op::store_hhhh); break; | ||||||
| 3083 | case skcms_PixelFormat_RGB_fff >> 1: add_op(Op::store_fff); break; | ||||||
| 3084 | case skcms_PixelFormat_RGBA_ffff >> 1: add_op(Op::store_ffff); break; | ||||||
| 3085 | |||||||
| 3086 | case skcms_PixelFormat_RGBA_8888_sRGB >> 1: | ||||||
| 3087 | add_op_ctx(Op::tf_rgb, skcms_sRGB_Inverse_TransferFunction()); | ||||||
| 3088 | add_op(Op::store_8888); | ||||||
| 3089 | break; | ||||||
| 3090 | } | ||||||
| 3091 | |||||||
| 3092 | assert(ops <= program + ARRAY_COUNT(program))(static_cast <bool> (ops <= program + (int)(sizeof(( program)) / sizeof(*(program)))) ? void (0) : __assert_fail ( "ops <= program + ARRAY_COUNT(program)", __builtin_FILE () , __builtin_LINE (), __extension__ __PRETTY_FUNCTION__)); | ||||||
| 3093 | assert(contexts <= context + ARRAY_COUNT(context))(static_cast <bool> (contexts <= context + (int)(sizeof ((context)) / sizeof(*(context)))) ? void (0) : __assert_fail ("contexts <= context + ARRAY_COUNT(context)", __builtin_FILE (), __builtin_LINE (), __extension__ __PRETTY_FUNCTION__)); | ||||||
| 3094 | |||||||
| 3095 | auto run = baseline::run_program; | ||||||
| 3096 | switch (cpu_type()) { | ||||||
| 3097 | case CpuType::SKX: | ||||||
| 3098 | #if !defined(SKCMS_DISABLE_SKX1) | ||||||
| 3099 | run = skx::run_program; | ||||||
| 3100 | break; | ||||||
| 3101 | #endif | ||||||
| 3102 | |||||||
| 3103 | case CpuType::HSW: | ||||||
| 3104 | #if !defined(SKCMS_DISABLE_HSW) | ||||||
| 3105 | run = hsw::run_program; | ||||||
| 3106 | break; | ||||||
| 3107 | #endif | ||||||
| 3108 | |||||||
| 3109 | case CpuType::Baseline: | ||||||
| 3110 | break; | ||||||
| 3111 | } | ||||||
| 3112 | |||||||
| 3113 | run(program, context, ops - program, (const char*)src, (char*)dst, n, src_bpp,dst_bpp); | ||||||
| 3114 | return true; | ||||||
| 3115 | } | ||||||
| 3116 | |||||||
| 3117 | static void assert_usable_as_destination(const skcms_ICCProfile* profile) { | ||||||
| 3118 | #if defined(NDEBUG) | ||||||
| 3119 | (void)profile; | ||||||
| 3120 | #else | ||||||
| 3121 | skcms_Matrix3x3 fromXYZD50; | ||||||
| 3122 | skcms_TransferFunction invR, invG, invB; | ||||||
| 3123 | bool useB2A = false; | ||||||
| 3124 | bool useHlgOotf = false; | ||||||
| 3125 | assert(prep_for_destination(profile, &fromXYZD50, &invR, &invG, &invB, &useB2A, &useHlgOotf))(static_cast <bool> (prep_for_destination(profile, & fromXYZD50, &invR, &invG, &invB, &useB2A, & useHlgOotf)) ? void (0) : __assert_fail ("prep_for_destination(profile, &fromXYZD50, &invR, &invG, &invB, &useB2A, &useHlgOotf)" , __builtin_FILE (), __builtin_LINE (), __extension__ __PRETTY_FUNCTION__ )); | ||||||
| 3126 | #endif | ||||||
| 3127 | } | ||||||
| 3128 | |||||||
| 3129 | bool skcms_MakeUsableAsDestination(skcms_ICCProfile* profile) { | ||||||
| 3130 | if (!profile->has_B2A
| ||||||
| 3131 | skcms_Matrix3x3 fromXYZD50; | ||||||
| 3132 | if (!profile->has_trc || !profile->has_toXYZD50 | ||||||
| 3133 | || !skcms_Matrix3x3_invert(&profile->toXYZD50, &fromXYZD50)) { | ||||||
| 3134 | return false; | ||||||
| 3135 | } | ||||||
| 3136 | |||||||
| 3137 | skcms_TransferFunction tf[3]; | ||||||
| 3138 | for (int i = 0; i < 3; i++) { | ||||||
| 3139 | skcms_TransferFunction inv; | ||||||
| 3140 | if (profile->trc[i].table_entries == 0 | ||||||
| 3141 | && skcms_TransferFunction_invert(&profile->trc[i].parametric, &inv)) { | ||||||
| 3142 | tf[i] = profile->trc[i].parametric; | ||||||
| 3143 | continue; | ||||||
| 3144 | } | ||||||
| 3145 | |||||||
| 3146 | float max_error; | ||||||
| 3147 | // Parametric curves from skcms_ApproximateCurve() are guaranteed to be invertible. | ||||||
| 3148 | if (!skcms_ApproximateCurve(&profile->trc[i], &tf[i], &max_error)) { | ||||||
| 3149 | return false; | ||||||
| 3150 | } | ||||||
| 3151 | } | ||||||
| 3152 | |||||||
| 3153 | for (int i = 0; i < 3; ++i) { | ||||||
| 3154 | profile->trc[i].table_entries = 0; | ||||||
| 3155 | profile->trc[i].parametric = tf[i]; | ||||||
| 3156 | } | ||||||
| 3157 | } | ||||||
| 3158 | assert_usable_as_destination(profile); | ||||||
| 3159 | return true; | ||||||
| 3160 | } | ||||||
| 3161 | |||||||
| 3162 | bool skcms_MakeUsableAsDestinationWithSingleCurve(skcms_ICCProfile* profile) { | ||||||
| 3163 | // Call skcms_MakeUsableAsDestination() with B2A disabled; | ||||||
| 3164 | // on success that'll return a TRC/XYZ profile with three skcms_TransferFunctions. | ||||||
| 3165 | skcms_ICCProfile result = *profile; | ||||||
| 3166 | result.has_B2A = false; | ||||||
| 3167 | if (!skcms_MakeUsableAsDestination(&result)) { | ||||||
| |||||||
| 3168 | return false; | ||||||
| 3169 | } | ||||||
| 3170 | |||||||
| 3171 | // Of the three, pick the transfer function that best fits the other two. | ||||||
| 3172 | int best_tf = 0; | ||||||
| 3173 | float min_max_error = INFINITY_; | ||||||
| 3174 | for (int i = 0; i < 3; i++) { | ||||||
| 3175 | skcms_TransferFunction inv; | ||||||
| 3176 | if (!skcms_TransferFunction_invert(&result.trc[i].parametric, &inv)) { | ||||||
| 3177 | return false; | ||||||
| 3178 | } | ||||||
| 3179 | |||||||
| 3180 | float err = 0; | ||||||
| 3181 | for (int j = 0; j < 3; ++j) { | ||||||
| 3182 | err = fmaxf_(err, skcms_MaxRoundtripError(&profile->trc[j], &inv)); | ||||||
| 3183 | } | ||||||
| 3184 | if (min_max_error > err) { | ||||||
| 3185 | min_max_error = err; | ||||||
| 3186 | best_tf = i; | ||||||
| 3187 | } | ||||||
| 3188 | } | ||||||
| 3189 | |||||||
| 3190 | for (int i = 0; i < 3; i++) { | ||||||
| 3191 | result.trc[i].parametric = result.trc[best_tf].parametric; | ||||||
| 3192 | } | ||||||
| 3193 | |||||||
| 3194 | *profile = result; | ||||||
| 3195 | assert_usable_as_destination(profile); | ||||||
| 3196 | return true; | ||||||
| 3197 | } |