src/gallium/auxiliary/gallivm/lp_bld_format_srgb.c

   1 /**************************************************************************
   2  *
   3  * Copyright 2013 VMware, Inc.
   4  * All Rights Reserved.
   5  *
   6  * Permission is hereby granted, free of charge, to any person obtaining a
   7  * copy of this software and associated documentation files (the
   8  * "Software"), to deal in the Software without restriction, including
   9  * without limitation the rights to use, copy, modify, merge, publish,
  10  * distribute, sub license, and/or sell copies of the Software, and to
  11  * permit persons to whom the Software is furnished to do so, subject to
  12  * the following conditions:
  13  *
  14  * The above copyright notice and this permission notice (including the
  15  * next paragraph) shall be included in all copies or substantial portions
  16  * of the Software.
  17  *
  18  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS
  19  * OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
  20  * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT.
  21  * IN NO EVENT SHALL VMWARE AND/OR ITS SUPPLIERS BE LIABLE FOR
  22  * ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT,
  23  * TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE
  24  * SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.
  25  *
  26  **************************************************************************/
  27
  28
  29 /**
  30  * @file
  31  * Format conversion code for srgb formats.
  32  *
  33  * Functions for converting from srgb to linear and vice versa.
  34  * From http://www.opengl.org/registry/specs/EXT/texture_sRGB.txt:
  35  *
  36  * srgb->linear:
  37  * cl = cs / 12.92,                 cs <= 0.04045
  38  * cl = ((cs + 0.055)/1.055)^2.4,   cs >  0.04045
  39  *
  40  * linear->srgb:
  41  * if (isnan(cl)) {
  42  *    Map IEEE-754 Not-a-number to zero.
  43  *    cs = 0.0;
  44  * } else if (cl > 1.0) {
  45  *    cs = 1.0;
  46  * } else if (cl < 0.0) {
  47  *    cs = 0.0;
  48  * } else if (cl < 0.0031308) {
  49  *    cs = 12.92 * cl;
  50  * } else {
  51  *    cs = 1.055 * pow(cl, 0.41666) - 0.055;
  52  * }
  53  *
  54  * This does not need to be accurate, however at least for d3d10
  55  * (http://msdn.microsoft.com/en-us/library/windows/desktop/dd607323%28v=vs.85%29.aspx):
  56  * 1) For srgb->linear, it is required that the error on the srgb side is
  57  *    not larger than 0.5f, which I interpret that if you map the value back
  58  *    to srgb from linear using the ideal conversion, it would not be off by
  59  *    more than 0.5f (that is, it would map to the same 8-bit integer value
  60  *    as it was before conversion to linear).
  61  * 2) linear->srgb is permitted 0.6f which luckily looks like quite a large
  62  *    error is allowed.
  63  * 3) Additionally, all srgb values converted to linear and back must result
  64  *    in the same value as they were originally.
  65  *
  66  * @author Roland Scheidegger <sroland@vmware.com>
  67  */
  68
  69
  70 #include "util/u_debug.h"
  71
  72 #include "lp_bld_type.h"
  73 #include "lp_bld_const.h"
  74 #include "lp_bld_arit.h"
  75 #include "lp_bld_bitarit.h"
  76 #include "lp_bld_logic.h"
  77 #include "lp_bld_format.h"
  78
  79
  80
  81 /**
  82  * Convert srgb int values to linear float values.
  83  * Several possibilities how to do this, e.g.
  84  * - table
  85  * - doing the pow() with int-to-float and float-to-int tricks
  86  *   (http://stackoverflow.com/questions/6475373/optimizations-for-pow-with-const-non-integer-exponent)
  87  * - just using standard polynomial approximation
  88  *   (3rd order polynomial is required for crappy but just sufficient accuracy)
  89  *
  90  * @param src   integer (vector) value(s) to convert
  91  *              (chan_bits bit values unpacked to 32 bit already).
  92  */
  93 LLVMValueRef
  94 lp_build_srgb_to_linear(struct gallivm_state *gallivm,
  95                         struct lp_type src_type,
  96                         unsigned chan_bits,
  97                         LLVMValueRef src)
  98 {
  99    struct lp_type f32_type = lp_type_float_vec(32, src_type.length * 32);
 100    struct lp_build_context f32_bld;
 101    LLVMValueRef srcf, part_lin, part_pow, is_linear, lin_const, lin_thresh;
 102    double coeffs[4] = {0.0023f,
 103                        0.0030f / 255.0f,
 104                        0.6935f / (255.0f * 255.0f),
 105                        0.3012f / (255.0f * 255.0f * 255.0f)
 106    };
 107
 108    assert(src_type.width == 32);
 109    /* Technically this would work with more bits too but would be inaccurate. */
 110    assert(chan_bits <= 8);
 111
 112    lp_build_context_init(&f32_bld, gallivm, f32_type);
 113
 114    /*
 115     * using polynomial: (src * (src * (src * 0.3012 + 0.6935) + 0.0030) + 0.0023)
 116     * ( poly =  0.3012*x^3 + 0.6935*x^2 + 0.0030*x + 0.0023)
 117     * (found with octave polyfit and some magic as I couldn't get the error
 118     * function right). Using the above mentioned error function, the values stay
 119     * within +-0.35, except for the lowest values - hence tweaking linear segment
 120     * to cover the first 16 instead of the first 11 values (the error stays
 121     * just about acceptable there too).
 122     * Hence: lin = src > 15 ? poly : src / 12.6
 123     * This function really only makes sense for vectors, should use LUT otherwise.
 124     * All in all (including float conversion) 11 instructions (with sse4.1),
 125     * 6 constants (polynomial could be done with 1 instruction less at the cost
 126     * of slightly worse dependency chain, fma should also help).
 127     */
 128    /* doing the 1/255 mul as part of the approximation */
 129    srcf = lp_build_int_to_float(&f32_bld, src);
 130    if (chan_bits != 8) {
 131       /* could adjust all the constants instead */
 132       LLVMValueRef rescale_const = lp_build_const_vec(gallivm, f32_type,
 133                                                       255.0f / ((1 << chan_bits) - 1));
 134       srcf = lp_build_mul(&f32_bld, srcf, rescale_const);
 135    }
 136    lin_const = lp_build_const_vec(gallivm, f32_type, 1.0f / (12.6f * 255.0f));
 137    part_lin = lp_build_mul(&f32_bld, srcf, lin_const);
 138
 139    part_pow = lp_build_polynomial(&f32_bld, srcf, coeffs, 4);
 140
 141    lin_thresh = lp_build_const_vec(gallivm, f32_type, 15.0f);
 142    is_linear = lp_build_compare(gallivm, f32_type, PIPE_FUNC_LEQUAL, srcf, lin_thresh);
 143    return lp_build_select(&f32_bld, is_linear, part_lin, part_pow);
 144 }
 145
 146
 147 /**
 148  * Convert linear float values to srgb int values.
 149  * Several possibilities how to do this, e.g.
 150  * - use table (based on exponent/highest order mantissa bits) and do
 151  *   linear interpolation (https://gist.github.com/rygorous/2203834)
 152  * - Chebyshev polynomial
 153  * - Approximation using reciprocals
 154  * - using int-to-float and float-to-int tricks for pow()
 155  *   (http://stackoverflow.com/questions/6475373/optimizations-for-pow-with-const-non-integer-exponent)
 156  *
 157  * @param src   float (vector) value(s) to convert.
 158  */
 159 static LLVMValueRef
 160 lp_build_linear_to_srgb(struct gallivm_state *gallivm,
 161                         struct lp_type src_type,
 162                         unsigned chan_bits,
 163                         LLVMValueRef src)
 164 {
 165    LLVMBuilderRef builder = gallivm->builder;
 166    struct lp_build_context f32_bld;
 167    LLVMValueRef lin_thresh, lin, lin_const, is_linear, tmp, pow_final;
 168
 169    lp_build_context_init(&f32_bld, gallivm, src_type);
 170
 171    src = lp_build_clamp(&f32_bld, src, f32_bld.zero, f32_bld.one);
 172
 173    if (0) {
 174       /*
 175        * using int-to-float and float-to-int trick for pow().
 176        * This is much more accurate than necessary thanks to the correction,
 177        * but it most certainly makes no sense without rsqrt available.
 178        * Bonus points if you understand how this works...
 179        * All in all (including min/max clamp, conversion) 19 instructions.
 180        */
 181
 182       float exp_f = 2.0f / 3.0f;
 183       /* some compilers can't do exp2f, so this is exp2f(127.0f/exp_f - 127.0f) */
 184       float exp2f_c = 1.30438178253e+19f;
 185       float coeff_f = 0.62996f;
 186       LLVMValueRef pow_approx, coeff, x2, exponent, pow_1, pow_2;
 187       struct lp_type int_type = lp_int_type(src_type);
 188
 189       /*
 190        * First calculate approx x^8/12
 191        */
 192       exponent = lp_build_const_vec(gallivm, src_type, exp_f);
 193       coeff = lp_build_const_vec(gallivm, src_type,
 194                                  exp2f_c * powf(coeff_f, 1.0f / exp_f));
 195
 196       /* premultiply src */
 197       tmp = lp_build_mul(&f32_bld, coeff, src);
 198       /* "log2" */
 199       tmp = LLVMBuildBitCast(builder, tmp, lp_build_vec_type(gallivm, int_type), "");
 200       tmp = lp_build_int_to_float(&f32_bld, tmp);
 201       /* multiply for pow */
 202       tmp = lp_build_mul(&f32_bld, tmp, exponent);
 203       /* "exp2" */
 204       pow_approx = lp_build_itrunc(&f32_bld, tmp);
 205       pow_approx = LLVMBuildBitCast(builder, pow_approx,
 206                                     lp_build_vec_type(gallivm, src_type), "");
 207
 208       /*
 209        * Since that pow was inaccurate (like 3 bits, though each sqrt step would
 210        * give another bit), compensate the error (which is why we chose another
 211        * exponent in the first place).
 212        */
 213       /* x * x^(8/12) = x^(20/12) */
 214       pow_1 = lp_build_mul(&f32_bld, pow_approx, src);
 215
 216       /* x * x * x^(-4/12) = x^(20/12) */
 217       /* Should avoid using rsqrt if it's not available, but
 218        * using x * x^(4/12) * x^(4/12) instead will change error weight */
 219       tmp = lp_build_fast_rsqrt(&f32_bld, pow_approx);
 220       x2 = lp_build_mul(&f32_bld, src, src);
 221       pow_2 = lp_build_mul(&f32_bld, x2, tmp);
 222
 223       /* average the values so the errors cancel out, compensate bias,
 224        * we also squeeze the 1.055 mul of the srgb conversion plus the 255.0 mul
 225        * for conversion to int in here */
 226       tmp = lp_build_add(&f32_bld, pow_1, pow_2);
 227       coeff = lp_build_const_vec(gallivm, src_type,
 228                                  1.0f / (3.0f * coeff_f) * 0.999852f *
 229                                  powf(1.055f * 255.0f, 4.0f));
 230       pow_final = lp_build_mul(&f32_bld, tmp, coeff);
 231
 232       /* x^(5/12) = rsqrt(rsqrt(x^20/12)) */
 233       if (lp_build_fast_rsqrt_available(src_type)) {
 234          pow_final = lp_build_fast_rsqrt(&f32_bld,
 235                         lp_build_fast_rsqrt(&f32_bld, pow_final));
 236       }
 237       else {
 238          pow_final = lp_build_sqrt(&f32_bld, lp_build_sqrt(&f32_bld, pow_final));
 239       }
 240       pow_final = lp_build_add(&f32_bld, pow_final,
 241                                lp_build_const_vec(gallivm, src_type, -0.055f * 255.0f));
 242    }
 243
 244    else {
 245       /*
 246        * using "rational polynomial" approximation here.
 247        * Essentially y = a*x^0.375 + b*x^0.5 + c, with also
 248        * factoring in the 255.0 mul and the scaling mul.
 249        * (a is closer to actual value so has higher weight than b.)
 250        * Note: the constants are magic values. They were found empirically,
 251        * possibly could be improved but good enough (be VERY careful with
 252        * error metric if you'd want to tweak them, they also MUST fit with
 253        * the crappy polynomial above for srgb->linear since it is required
 254        * that each srgb value maps back to the same value).
 255        * This function has an error of max +-0.17. Not sure this is actually
 256        * enough, we require +-0.6 but that may include the +-0.5 from integer
 257        * conversion. Seems to pass all relevant tests though...
 258        * For the approximated srgb->linear values the error is naturally larger
 259        * (+-0.42) but still accurate enough (required +-0.5 essentially).
 260        * All in all (including min/max clamp, conversion) 15 instructions.
 261        * FMA would help (minus 2 instructions).
 262        */
 263
 264       LLVMValueRef x05, x0375, a_const, b_const, c_const, tmp2;
 265
 266       if (lp_build_fast_rsqrt_available(src_type)) {
 267          tmp = lp_build_fast_rsqrt(&f32_bld, src);
 268          x05 = lp_build_mul(&f32_bld, src, tmp);
 269       }
 270       else {
 271          /*
 272           * I don't really expect this to be practical without rsqrt
 273           * but there's no reason for triple punishment so at least
 274           * save the otherwise resulting division and unnecessary mul...
 275           */
 276          x05 = lp_build_sqrt(&f32_bld, src);
 277       }
 278
 279       tmp = lp_build_mul(&f32_bld, x05, src);
 280       if (lp_build_fast_rsqrt_available(src_type)) {
 281          x0375 = lp_build_fast_rsqrt(&f32_bld, lp_build_fast_rsqrt(&f32_bld, tmp));
 282       }
 283       else {
 284          x0375 = lp_build_sqrt(&f32_bld, lp_build_sqrt(&f32_bld, tmp));
 285       }
 286
 287       a_const = lp_build_const_vec(gallivm, src_type, 0.675f * 1.0622 * 255.0f);
 288       b_const = lp_build_const_vec(gallivm, src_type, 0.325f * 1.0622 * 255.0f);
 289       c_const = lp_build_const_vec(gallivm, src_type, -0.0620f * 255.0f);
 290
 291       tmp = lp_build_mul(&f32_bld, a_const, x0375);
 292       tmp2 = lp_build_mad(&f32_bld, b_const, x05, c_const);
 293       pow_final = lp_build_add(&f32_bld, tmp, tmp2);
 294    }
 295
 296    /* linear part is easy */
 297    lin_const = lp_build_const_vec(gallivm, src_type, 12.92f * 255.0f);
 298    lin = lp_build_mul(&f32_bld, src, lin_const);
 299
 300    lin_thresh = lp_build_const_vec(gallivm, src_type, 0.0031308f);
 301    is_linear = lp_build_compare(gallivm, src_type, PIPE_FUNC_LEQUAL, src, lin_thresh);
 302    tmp = lp_build_select(&f32_bld, is_linear, lin, pow_final);
 303
 304    if (chan_bits != 8) {
 305       /* could adjust all the constants instead */
 306       LLVMValueRef rescale_const = lp_build_const_vec(gallivm, src_type,
 307                                                       ((1 << chan_bits) - 1) / 255.0f);
 308       tmp = lp_build_mul(&f32_bld, tmp, rescale_const);
 309    }
 310
 311    f32_bld.type.sign = 0;
 312    return lp_build_iround(&f32_bld, tmp);
 313 }
 314
 315
 316 /**
 317  * Convert linear float soa values to packed srgb AoS values.
 318  * This only handles packed formats which are 4x8bit in size
 319  * (rgba and rgbx plus swizzles), and 16bit 565-style formats
 320  * with no alpha. (In the latter case the return values won't be
 321  * fully packed, it will look like r5g6b5x16r5g6b5x16...)
 322  *
 323  * @param src   float SoA (vector) values to convert.
 324  */
 325 LLVMValueRef
 326 lp_build_float_to_srgb_packed(struct gallivm_state *gallivm,
 327                               const struct util_format_description *dst_fmt,
 328                               struct lp_type src_type,
 329                               LLVMValueRef *src)
 330 {
 331    LLVMBuilderRef builder = gallivm->builder;
 332    unsigned chan;
 333    struct lp_build_context f32_bld;
 334    struct lp_type int32_type = lp_int_type(src_type);
 335    LLVMValueRef tmpsrgb[4], alpha, dst;
 336
 337    lp_build_context_init(&f32_bld, gallivm, src_type);
 338
 339    /* rgb is subject to linear->srgb conversion, alpha is not */
 340    for (chan = 0; chan < 3; chan++) {
 341       unsigned chan_bits = dst_fmt->channel[dst_fmt->swizzle[chan]].size;
 342       tmpsrgb[chan] = lp_build_linear_to_srgb(gallivm, src_type, chan_bits, src[chan]);
 343    }
 344    /*
 345     * can't use lp_build_conv since we want to keep values as 32bit
 346     * here so we can interleave with rgb to go from SoA->AoS.
 347     */
 348    alpha = lp_build_clamp_zero_one_nanzero(&f32_bld, src[3]);
 349    alpha = lp_build_mul(&f32_bld, alpha,
 350                         lp_build_const_vec(gallivm, src_type, 255.0f));
 351    tmpsrgb[3] = lp_build_iround(&f32_bld, alpha);
 352
 353    dst = lp_build_zero(gallivm, int32_type);
 354    for (chan = 0; chan < dst_fmt->nr_channels; chan++) {
 355       if (dst_fmt->swizzle[chan] <= PIPE_SWIZZLE_W) {
 356          unsigned ls;
 357          LLVMValueRef shifted, shift_val;
 358          ls = dst_fmt->channel[dst_fmt->swizzle[chan]].shift;
 359          shift_val = lp_build_const_int_vec(gallivm, int32_type, ls);
 360          shifted = LLVMBuildShl(builder, tmpsrgb[chan], shift_val, "");
 361          dst = LLVMBuildOr(builder, dst, shifted, "");
 362       }
 363    }
 364    return dst;
 365 }