SkMath.h revision d77ac7792405814fac3fdf229dea348dede650f1
1 2/* 3 * Copyright 2006 The Android Open Source Project 4 * 5 * Use of this source code is governed by a BSD-style license that can be 6 * found in the LICENSE file. 7 */ 8 9 10#ifndef SkMath_DEFINED 11#define SkMath_DEFINED 12 13#include "SkTypes.h" 14 15// 64bit -> 32bit utilities 16 17/** 18 * Return true iff the 64bit value can exactly be represented in signed 32bits 19 */ 20static inline bool sk_64_isS32(int64_t value) { 21 return (int32_t)value == value; 22} 23 24/** 25 * Return the 64bit argument as signed 32bits, asserting in debug that the arg 26 * exactly fits in signed 32bits. In the release build, no checks are preformed 27 * and the return value if the arg does not fit is undefined. 28 */ 29static inline int32_t sk_64_asS32(int64_t value) { 30 SkASSERT(sk_64_isS32(value)); 31 return (int32_t)value; 32} 33 34// Handy util that can be passed two ints, and will automatically promote to 35// 64bits before the multiply, so the caller doesn't have to remember to cast 36// e.g. (int64_t)a * b; 37static inline int64_t sk_64_mul(int64_t a, int64_t b) { 38 return a * b; 39} 40 41/////////////////////////////////////////////////////////////////////////////// 42 43/** 44 * Computes numer1 * numer2 / denom in full 64 intermediate precision. 45 * It is an error for denom to be 0. There is no special handling if 46 * the result overflows 32bits. 47 */ 48static inline int32_t SkMulDiv(int32_t numer1, int32_t numer2, int32_t denom) { 49 SkASSERT(denom); 50 51 int64_t tmp = sk_64_mul(numer1, numer2) / denom; 52 return sk_64_asS32(tmp); 53} 54 55/** 56 * Return the integer square root of value, with a bias of bitBias 57 */ 58int32_t SkSqrtBits(int32_t value, int bitBias); 59 60/** Return the integer square root of n, treated as a SkFixed (16.16) 61 */ 62#define SkSqrt32(n) SkSqrtBits(n, 15) 63 64//! Returns the number of leading zero bits (0...32) 65int SkCLZ_portable(uint32_t); 66 67#ifndef SkCLZ 68 #if defined(_MSC_VER) 69 #include <intrin.h> 70 71 static inline int SkCLZ(uint32_t mask) { 72 if (mask) { 73 DWORD index; 74 _BitScanReverse(&index, mask); 75 // Suppress this bogus /analyze warning. The check for non-zero 76 // guarantees that _BitScanReverse will succeed. 77#pragma warning(suppress : 6102) // Using 'index' from failed function call 78 return index ^ 0x1F; 79 } else { 80 return 32; 81 } 82 } 83 #elif defined(SK_CPU_ARM32) || defined(__GNUC__) || defined(__clang__) 84 static inline int SkCLZ(uint32_t mask) { 85 // __builtin_clz(0) is undefined, so we have to detect that case. 86 return mask ? __builtin_clz(mask) : 32; 87 } 88 #else 89 #define SkCLZ(x) SkCLZ_portable(x) 90 #endif 91#endif 92 93/** 94 * Returns (value < 0 ? 0 : value) efficiently (i.e. no compares or branches) 95 */ 96static inline int SkClampPos(int value) { 97 return value & ~(value >> 31); 98} 99 100/** Given an integer and a positive (max) integer, return the value 101 * pinned against 0 and max, inclusive. 102 * @param value The value we want returned pinned between [0...max] 103 * @param max The positive max value 104 * @return 0 if value < 0, max if value > max, else value 105 */ 106static inline int SkClampMax(int value, int max) { 107 // ensure that max is positive 108 SkASSERT(max >= 0); 109 if (value < 0) { 110 value = 0; 111 } 112 if (value > max) { 113 value = max; 114 } 115 return value; 116} 117 118/** 119 * Returns the smallest power-of-2 that is >= the specified value. If value 120 * is already a power of 2, then it is returned unchanged. It is undefined 121 * if value is <= 0. 122 */ 123static inline int SkNextPow2(int value) { 124 SkASSERT(value > 0); 125 return 1 << (32 - SkCLZ(value - 1)); 126} 127 128/** 129 * Returns the log2 of the specified value, were that value to be rounded up 130 * to the next power of 2. It is undefined to pass 0. Examples: 131 * SkNextLog2(1) -> 0 132 * SkNextLog2(2) -> 1 133 * SkNextLog2(3) -> 2 134 * SkNextLog2(4) -> 2 135 * SkNextLog2(5) -> 3 136 */ 137static inline int SkNextLog2(uint32_t value) { 138 SkASSERT(value != 0); 139 return 32 - SkCLZ(value - 1); 140} 141 142/** 143 * Returns true if value is a power of 2. Does not explicitly check for 144 * value <= 0. 145 */ 146template <typename T> inline bool SkIsPow2(T value) { 147 return (value & (value - 1)) == 0; 148} 149 150/////////////////////////////////////////////////////////////////////////////// 151 152/** 153 * Return a*b/((1 << shift) - 1), rounding any fractional bits. 154 * Only valid if a and b are unsigned and <= 32767 and shift is > 0 and <= 8 155 */ 156static inline unsigned SkMul16ShiftRound(U16CPU a, U16CPU b, int shift) { 157 SkASSERT(a <= 32767); 158 SkASSERT(b <= 32767); 159 SkASSERT(shift > 0 && shift <= 8); 160 unsigned prod = a*b + (1 << (shift - 1)); 161 return (prod + (prod >> shift)) >> shift; 162} 163 164/** 165 * Return a*b/255, rounding any fractional bits. 166 * Only valid if a and b are unsigned and <= 32767. 167 */ 168static inline U8CPU SkMulDiv255Round(U16CPU a, U16CPU b) { 169 SkASSERT(a <= 32767); 170 SkASSERT(b <= 32767); 171 unsigned prod = a*b + 128; 172 return (prod + (prod >> 8)) >> 8; 173} 174 175/** 176 * Stores numer/denom and numer%denom into div and mod respectively. 177 */ 178template <typename In, typename Out> 179inline void SkTDivMod(In numer, In denom, Out* div, Out* mod) { 180#ifdef SK_CPU_ARM32 181 // If we wrote this as in the else branch, GCC won't fuse the two into one 182 // divmod call, but rather a div call followed by a divmod. Silly! This 183 // version is just as fast as calling __aeabi_[u]idivmod manually, but with 184 // prettier code. 185 // 186 // This benches as around 2x faster than the code in the else branch. 187 const In d = numer/denom; 188 *div = static_cast<Out>(d); 189 *mod = static_cast<Out>(numer-d*denom); 190#else 191 // On x86 this will just be a single idiv. 192 *div = static_cast<Out>(numer/denom); 193 *mod = static_cast<Out>(numer%denom); 194#endif 195} 196 197#endif 198