| Jason Simmons | b86c8e4 | 2011-06-28 17:43:30 -0700 | [diff] [blame] | 1 | /* | 
|  | 2 | * Copyright (C) 2011 The Android Open Source Project | 
|  | 3 | * | 
|  | 4 | * Licensed under the Apache License, Version 2.0 (the "License"); | 
|  | 5 | * you may not use this file except in compliance with the License. | 
|  | 6 | * You may obtain a copy of the License at | 
|  | 7 | * | 
|  | 8 | *      http://www.apache.org/licenses/LICENSE-2.0 | 
|  | 9 | * | 
|  | 10 | * Unless required by applicable law or agreed to in writing, software | 
|  | 11 | * distributed under the License is distributed on an "AS IS" BASIS, | 
|  | 12 | * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. | 
|  | 13 | * See the License for the specific language governing permissions and | 
|  | 14 | * limitations under the License. | 
|  | 15 | */ | 
|  | 16 |  | 
|  | 17 | #define __STDC_LIMIT_MACROS | 
|  | 18 |  | 
|  | 19 | #include <assert.h> | 
|  | 20 | #include <stdint.h> | 
|  | 21 |  | 
|  | 22 | #include <utils/LinearTransform.h> | 
|  | 23 |  | 
|  | 24 | namespace android { | 
|  | 25 |  | 
|  | 26 | template<class T> static inline T ABS(T x) { return (x < 0) ? -x : x; } | 
|  | 27 |  | 
|  | 28 | // Static math methods involving linear transformations | 
|  | 29 | static bool scale_u64_to_u64( | 
|  | 30 | uint64_t val, | 
|  | 31 | uint32_t N, | 
|  | 32 | uint32_t D, | 
|  | 33 | uint64_t* res, | 
|  | 34 | bool round_up_not_down) { | 
|  | 35 | uint64_t tmp1, tmp2; | 
|  | 36 | uint32_t r; | 
|  | 37 |  | 
|  | 38 | assert(res); | 
|  | 39 | assert(D); | 
|  | 40 |  | 
|  | 41 | // Let U32(X) denote a uint32_t containing the upper 32 bits of a 64 bit | 
|  | 42 | // integer X. | 
|  | 43 | // Let L32(X) denote a uint32_t containing the lower 32 bits of a 64 bit | 
|  | 44 | // integer X. | 
|  | 45 | // Let X[A, B] with A <= B denote bits A through B of the integer X. | 
|  | 46 | // Let (A | B) denote the concatination of two 32 bit ints, A and B. | 
|  | 47 | // IOW X = (A | B) => U32(X) == A && L32(X) == B | 
|  | 48 | // | 
|  | 49 | // compute M = val * N (a 96 bit int) | 
|  | 50 | // --------------------------------- | 
|  | 51 | // tmp2 = U32(val) * N (a 64 bit int) | 
|  | 52 | // tmp1 = L32(val) * N (a 64 bit int) | 
|  | 53 | // which means | 
|  | 54 | // M = val * N = (tmp2 << 32) + tmp1 | 
|  | 55 | tmp2 = (val >> 32) * N; | 
|  | 56 | tmp1 = (val & UINT32_MAX) * N; | 
|  | 57 |  | 
|  | 58 | // compute M[32, 95] | 
|  | 59 | // tmp2 = tmp2 + U32(tmp1) | 
|  | 60 | //      = (U32(val) * N) + U32(L32(val) * N) | 
|  | 61 | //      = M[32, 95] | 
|  | 62 | tmp2 += tmp1 >> 32; | 
|  | 63 |  | 
|  | 64 | // if M[64, 95] >= D, then M/D has bits > 63 set and we have | 
|  | 65 | // an overflow. | 
|  | 66 | if ((tmp2 >> 32) >= D) { | 
|  | 67 | *res = UINT64_MAX; | 
|  | 68 | return false; | 
|  | 69 | } | 
|  | 70 |  | 
|  | 71 | // Divide.  Going in we know | 
|  | 72 | // tmp2 = M[32, 95] | 
|  | 73 | // U32(tmp2) < D | 
|  | 74 | r = tmp2 % D; | 
|  | 75 | tmp2 /= D; | 
|  | 76 |  | 
|  | 77 | // At this point | 
|  | 78 | // tmp1      = L32(val) * N | 
|  | 79 | // tmp2      = M[32, 95] / D | 
|  | 80 | //           = (M / D)[32, 95] | 
|  | 81 | // r         = M[32, 95] % D | 
|  | 82 | // U32(tmp2) = 0 | 
|  | 83 | // | 
|  | 84 | // compute tmp1 = (r | M[0, 31]) | 
|  | 85 | tmp1 = (tmp1 & UINT32_MAX) | ((uint64_t)r << 32); | 
|  | 86 |  | 
|  | 87 | // Divide again.  Keep the remainder around in order to round properly. | 
|  | 88 | r = tmp1 % D; | 
|  | 89 | tmp1 /= D; | 
|  | 90 |  | 
|  | 91 | // At this point | 
|  | 92 | // tmp2      = (M / D)[32, 95] | 
|  | 93 | // tmp1      = (M / D)[ 0, 31] | 
|  | 94 | // r         =  M % D | 
|  | 95 | // U32(tmp1) = 0 | 
|  | 96 | // U32(tmp2) = 0 | 
|  | 97 |  | 
|  | 98 | // Pack the result and deal with the round-up case (As well as the | 
|  | 99 | // remote possiblility over overflow in such a case). | 
|  | 100 | *res = (tmp2 << 32) | tmp1; | 
|  | 101 | if (r && round_up_not_down) { | 
|  | 102 | ++(*res); | 
|  | 103 | if (!(*res)) { | 
|  | 104 | *res = UINT64_MAX; | 
|  | 105 | return false; | 
|  | 106 | } | 
|  | 107 | } | 
|  | 108 |  | 
|  | 109 | return true; | 
|  | 110 | } | 
|  | 111 |  | 
|  | 112 | static bool linear_transform_s64_to_s64( | 
|  | 113 | int64_t  val, | 
|  | 114 | int64_t  basis1, | 
|  | 115 | int32_t  N, | 
|  | 116 | uint32_t D, | 
| John Grossman | 885a2fe | 2012-06-27 15:34:43 -0700 | [diff] [blame] | 117 | bool     invert_frac, | 
| Jason Simmons | b86c8e4 | 2011-06-28 17:43:30 -0700 | [diff] [blame] | 118 | int64_t  basis2, | 
|  | 119 | int64_t* out) { | 
|  | 120 | uint64_t scaled, res; | 
|  | 121 | uint64_t abs_val; | 
|  | 122 | bool is_neg; | 
|  | 123 |  | 
|  | 124 | if (!out) | 
|  | 125 | return false; | 
|  | 126 |  | 
|  | 127 | // Compute abs(val - basis_64). Keep track of whether or not this delta | 
|  | 128 | // will be negative after the scale opertaion. | 
|  | 129 | if (val < basis1) { | 
|  | 130 | is_neg = true; | 
|  | 131 | abs_val = basis1 - val; | 
|  | 132 | } else { | 
|  | 133 | is_neg = false; | 
|  | 134 | abs_val = val - basis1; | 
|  | 135 | } | 
|  | 136 |  | 
|  | 137 | if (N < 0) | 
|  | 138 | is_neg = !is_neg; | 
|  | 139 |  | 
|  | 140 | if (!scale_u64_to_u64(abs_val, | 
| John Grossman | 885a2fe | 2012-06-27 15:34:43 -0700 | [diff] [blame] | 141 | invert_frac ? D : ABS(N), | 
|  | 142 | invert_frac ? ABS(N) : D, | 
| Jason Simmons | b86c8e4 | 2011-06-28 17:43:30 -0700 | [diff] [blame] | 143 | &scaled, | 
|  | 144 | is_neg)) | 
|  | 145 | return false; // overflow/undeflow | 
|  | 146 |  | 
|  | 147 | // if scaled is >= 0x8000<etc>, then we are going to overflow or | 
|  | 148 | // underflow unless ABS(basis2) is large enough to pull us back into the | 
|  | 149 | // non-overflow/underflow region. | 
|  | 150 | if (scaled & INT64_MIN) { | 
|  | 151 | if (is_neg && (basis2 < 0)) | 
|  | 152 | return false; // certain underflow | 
|  | 153 |  | 
|  | 154 | if (!is_neg && (basis2 >= 0)) | 
|  | 155 | return false; // certain overflow | 
|  | 156 |  | 
|  | 157 | if (ABS(basis2) <= static_cast<int64_t>(scaled & INT64_MAX)) | 
|  | 158 | return false; // not enough | 
|  | 159 |  | 
|  | 160 | // Looks like we are OK | 
|  | 161 | *out = (is_neg ? (-scaled) : scaled) + basis2; | 
|  | 162 | } else { | 
|  | 163 | // Scaled fits within signed bounds, so we just need to check for | 
|  | 164 | // over/underflow for two signed integers.  Basically, if both scaled | 
|  | 165 | // and basis2 have the same sign bit, and the result has a different | 
|  | 166 | // sign bit, then we have under/overflow.  An easy way to compute this | 
|  | 167 | // is | 
|  | 168 | // (scaled_signbit XNOR basis_signbit) && | 
|  | 169 | // (scaled_signbit XOR res_signbit) | 
|  | 170 | // == | 
|  | 171 | // (scaled_signbit XOR basis_signbit XOR 1) && | 
|  | 172 | // (scaled_signbit XOR res_signbit) | 
|  | 173 |  | 
|  | 174 | if (is_neg) | 
|  | 175 | scaled = -scaled; | 
|  | 176 | res = scaled + basis2; | 
|  | 177 |  | 
|  | 178 | if ((scaled ^ basis2 ^ INT64_MIN) & (scaled ^ res) & INT64_MIN) | 
|  | 179 | return false; | 
|  | 180 |  | 
|  | 181 | *out = res; | 
|  | 182 | } | 
|  | 183 |  | 
|  | 184 | return true; | 
|  | 185 | } | 
|  | 186 |  | 
|  | 187 | bool LinearTransform::doForwardTransform(int64_t a_in, int64_t* b_out) const { | 
|  | 188 | if (0 == a_to_b_denom) | 
|  | 189 | return false; | 
|  | 190 |  | 
|  | 191 | return linear_transform_s64_to_s64(a_in, | 
|  | 192 | a_zero, | 
|  | 193 | a_to_b_numer, | 
|  | 194 | a_to_b_denom, | 
| John Grossman | 885a2fe | 2012-06-27 15:34:43 -0700 | [diff] [blame] | 195 | false, | 
| Jason Simmons | b86c8e4 | 2011-06-28 17:43:30 -0700 | [diff] [blame] | 196 | b_zero, | 
|  | 197 | b_out); | 
|  | 198 | } | 
|  | 199 |  | 
|  | 200 | bool LinearTransform::doReverseTransform(int64_t b_in, int64_t* a_out) const { | 
|  | 201 | if (0 == a_to_b_numer) | 
|  | 202 | return false; | 
|  | 203 |  | 
|  | 204 | return linear_transform_s64_to_s64(b_in, | 
|  | 205 | b_zero, | 
| Jason Simmons | b86c8e4 | 2011-06-28 17:43:30 -0700 | [diff] [blame] | 206 | a_to_b_numer, | 
| John Grossman | 885a2fe | 2012-06-27 15:34:43 -0700 | [diff] [blame] | 207 | a_to_b_denom, | 
|  | 208 | true, | 
| Jason Simmons | b86c8e4 | 2011-06-28 17:43:30 -0700 | [diff] [blame] | 209 | a_zero, | 
|  | 210 | a_out); | 
|  | 211 | } | 
|  | 212 |  | 
|  | 213 | template <class T> void LinearTransform::reduce(T* N, T* D) { | 
|  | 214 | T a, b; | 
|  | 215 | if (!N || !D || !(*D)) { | 
|  | 216 | assert(false); | 
|  | 217 | return; | 
|  | 218 | } | 
|  | 219 |  | 
|  | 220 | a = *N; | 
|  | 221 | b = *D; | 
|  | 222 |  | 
|  | 223 | if (a == 0) { | 
|  | 224 | *D = 1; | 
|  | 225 | return; | 
|  | 226 | } | 
|  | 227 |  | 
|  | 228 | // This implements Euclid's method to find GCD. | 
|  | 229 | if (a < b) { | 
|  | 230 | T tmp = a; | 
|  | 231 | a = b; | 
|  | 232 | b = tmp; | 
|  | 233 | } | 
|  | 234 |  | 
|  | 235 | while (1) { | 
|  | 236 | // a is now the greater of the two. | 
|  | 237 | const T remainder = a % b; | 
|  | 238 | if (remainder == 0) { | 
|  | 239 | *N /= b; | 
|  | 240 | *D /= b; | 
|  | 241 | return; | 
|  | 242 | } | 
|  | 243 | // by swapping remainder and b, we are guaranteeing that a is | 
|  | 244 | // still the greater of the two upon entrance to the loop. | 
|  | 245 | a = b; | 
|  | 246 | b = remainder; | 
|  | 247 | } | 
|  | 248 | }; | 
|  | 249 |  | 
|  | 250 | template void LinearTransform::reduce<uint64_t>(uint64_t* N, uint64_t* D); | 
|  | 251 | template void LinearTransform::reduce<uint32_t>(uint32_t* N, uint32_t* D); | 
|  | 252 |  | 
|  | 253 | void LinearTransform::reduce(int32_t* N, uint32_t* D) { | 
|  | 254 | if (N && D && *D) { | 
|  | 255 | if (*N < 0) { | 
|  | 256 | *N = -(*N); | 
|  | 257 | reduce(reinterpret_cast<uint32_t*>(N), D); | 
|  | 258 | *N = -(*N); | 
|  | 259 | } else { | 
|  | 260 | reduce(reinterpret_cast<uint32_t*>(N), D); | 
|  | 261 | } | 
|  | 262 | } | 
|  | 263 | } | 
|  | 264 |  | 
|  | 265 | }  // namespace android |