1 /*
2 * Copyright 2012 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 #ifndef SkMathPriv_DEFINED
9 #define SkMathPriv_DEFINED
10
11 #include "include/core/SkMath.h"
12
13 /**
14 * Return the integer square root of value, with a bias of bitBias
15 */
16 int32_t SkSqrtBits(int32_t value, int bitBias);
17
18 /** Return the integer square root of n, treated as a SkFixed (16.16)
19 */
SkSqrt32(int32_t n)20 static inline int32_t SkSqrt32(int32_t n) { return SkSqrtBits(n, 15); }
21
22 /**
23 * Returns (value < 0 ? 0 : value) efficiently (i.e. no compares or branches)
24 */
SkClampPos(int value)25 static inline int SkClampPos(int value) {
26 return value & ~(value >> 31);
27 }
28
29 /**
30 * Stores numer/denom and numer%denom into div and mod respectively.
31 */
32 template <typename In, typename Out>
SkTDivMod(In numer,In denom,Out * div,Out * mod)33 inline void SkTDivMod(In numer, In denom, Out* div, Out* mod) {
34 #ifdef SK_CPU_ARM32
35 // If we wrote this as in the else branch, GCC won't fuse the two into one
36 // divmod call, but rather a div call followed by a divmod. Silly! This
37 // version is just as fast as calling __aeabi_[u]idivmod manually, but with
38 // prettier code.
39 //
40 // This benches as around 2x faster than the code in the else branch.
41 const In d = numer/denom;
42 *div = static_cast<Out>(d);
43 *mod = static_cast<Out>(numer-d*denom);
44 #else
45 // On x86 this will just be a single idiv.
46 *div = static_cast<Out>(numer/denom);
47 *mod = static_cast<Out>(numer%denom);
48 #endif
49 }
50
51 /** Returns -1 if n < 0, else returns 0
52 */
53 #define SkExtractSign(n) ((int32_t)(n) >> 31)
54
55 /** If sign == -1, returns -n, else sign must be 0, and returns n.
56 Typically used in conjunction with SkExtractSign().
57 */
SkApplySign(int32_t n,int32_t sign)58 static inline int32_t SkApplySign(int32_t n, int32_t sign) {
59 SkASSERT(sign == 0 || sign == -1);
60 return (n ^ sign) - sign;
61 }
62
63 /** Return x with the sign of y */
SkCopySign32(int32_t x,int32_t y)64 static inline int32_t SkCopySign32(int32_t x, int32_t y) {
65 return SkApplySign(x, SkExtractSign(x ^ y));
66 }
67
68 /** Given a positive value and a positive max, return the value
69 pinned against max.
70 Note: only works as long as max - value doesn't wrap around
71 @return max if value >= max, else value
72 */
SkClampUMax(unsigned value,unsigned max)73 static inline unsigned SkClampUMax(unsigned value, unsigned max) {
74 if (value > max) {
75 value = max;
76 }
77 return value;
78 }
79
80 // If a signed int holds min_int (e.g. 0x80000000) it is undefined what happens when
81 // we negate it (even though we *know* we're 2's complement and we'll get the same
82 // value back). So we create this helper function that casts to size_t (unsigned) first,
83 // to avoid the complaint.
sk_negate_to_size_t(int32_t value)84 static inline size_t sk_negate_to_size_t(int32_t value) {
85 #if defined(_MSC_VER)
86 #pragma warning(push)
87 #pragma warning(disable : 4146) // Thanks MSVC, we know what we're negating an unsigned
88 #endif
89 return -static_cast<size_t>(value);
90 #if defined(_MSC_VER)
91 #pragma warning(pop)
92 #endif
93 }
94
95 ///////////////////////////////////////////////////////////////////////////////
96
97 /** Return a*b/255, truncating away any fractional bits. Only valid if both
98 a and b are 0..255
99 */
SkMulDiv255Trunc(U8CPU a,U8CPU b)100 static inline U8CPU SkMulDiv255Trunc(U8CPU a, U8CPU b) {
101 SkASSERT((uint8_t)a == a);
102 SkASSERT((uint8_t)b == b);
103 unsigned prod = a*b + 1;
104 return (prod + (prod >> 8)) >> 8;
105 }
106
107 /** Return (a*b)/255, taking the ceiling of any fractional bits. Only valid if
108 both a and b are 0..255. The expected result equals (a * b + 254) / 255.
109 */
SkMulDiv255Ceiling(U8CPU a,U8CPU b)110 static inline U8CPU SkMulDiv255Ceiling(U8CPU a, U8CPU b) {
111 SkASSERT((uint8_t)a == a);
112 SkASSERT((uint8_t)b == b);
113 unsigned prod = a*b + 255;
114 return (prod + (prod >> 8)) >> 8;
115 }
116
117 /** Just the rounding step in SkDiv255Round: round(value / 255)
118 */
SkDiv255Round(unsigned prod)119 static inline unsigned SkDiv255Round(unsigned prod) {
120 prod += 128;
121 return (prod + (prod >> 8)) >> 8;
122 }
123
124 /**
125 * Swap byte order of a 4-byte value, e.g. 0xaarrggbb -> 0xbbggrraa.
126 */
127 #if defined(_MSC_VER)
128 #include <stdlib.h>
SkBSwap32(uint32_t v)129 static inline uint32_t SkBSwap32(uint32_t v) { return _byteswap_ulong(v); }
130 #else
SkBSwap32(uint32_t v)131 static inline uint32_t SkBSwap32(uint32_t v) { return __builtin_bswap32(v); }
132 #endif
133
134 //! Returns the number of leading zero bits (0...32)
135 int SkCLZ_portable(uint32_t);
136
137 #ifndef SkCLZ
138 #if defined(SK_BUILD_FOR_WIN)
139 #include <intrin.h>
140
SkCLZ(uint32_t mask)141 static inline int SkCLZ(uint32_t mask) {
142 if (mask) {
143 unsigned long index;
144 _BitScanReverse(&index, mask);
145 // Suppress this bogus /analyze warning. The check for non-zero
146 // guarantees that _BitScanReverse will succeed.
147 #pragma warning(suppress : 6102) // Using 'index' from failed function call
148 return index ^ 0x1F;
149 } else {
150 return 32;
151 }
152 }
153 #elif defined(SK_CPU_ARM32) || defined(__GNUC__) || defined(__clang__)
SkCLZ(uint32_t mask)154 static inline int SkCLZ(uint32_t mask) {
155 // __builtin_clz(0) is undefined, so we have to detect that case.
156 return mask ? __builtin_clz(mask) : 32;
157 }
158 #else
159 #define SkCLZ(x) SkCLZ_portable(x)
160 #endif
161 #endif
162
163 /**
164 * Returns the smallest power-of-2 that is >= the specified value. If value
165 * is already a power of 2, then it is returned unchanged. It is undefined
166 * if value is <= 0.
167 */
SkNextPow2(int value)168 static inline int SkNextPow2(int value) {
169 SkASSERT(value > 0);
170 return 1 << (32 - SkCLZ(value - 1));
171 }
172
173 /**
174 * Returns the largest power-of-2 that is <= the specified value. If value
175 * is already a power of 2, then it is returned unchanged. It is undefined
176 * if value is <= 0.
177 */
SkPrevPow2(int value)178 static inline int SkPrevPow2(int value) {
179 SkASSERT(value > 0);
180 return 1 << (32 - SkCLZ(value >> 1));
181 }
182
183 /**
184 * Returns the log2 of the specified value, were that value to be rounded up
185 * to the next power of 2. It is undefined to pass 0. Examples:
186 * SkNextLog2(1) -> 0
187 * SkNextLog2(2) -> 1
188 * SkNextLog2(3) -> 2
189 * SkNextLog2(4) -> 2
190 * SkNextLog2(5) -> 3
191 */
SkNextLog2(uint32_t value)192 static inline int SkNextLog2(uint32_t value) {
193 SkASSERT(value != 0);
194 return 32 - SkCLZ(value - 1);
195 }
196
197 /**
198 * Returns the log2 of the specified value, were that value to be rounded down
199 * to the previous power of 2. It is undefined to pass 0. Examples:
200 * SkPrevLog2(1) -> 0
201 * SkPrevLog2(2) -> 1
202 * SkPrevLog2(3) -> 1
203 * SkPrevLog2(4) -> 2
204 * SkPrevLog2(5) -> 2
205 */
SkPrevLog2(uint32_t value)206 static inline int SkPrevLog2(uint32_t value) {
207 SkASSERT(value != 0);
208 return 32 - SkCLZ(value >> 1);
209 }
210
211 ///////////////////////////////////////////////////////////////////////////////
212
213 /**
214 * Return the smallest power-of-2 >= n.
215 */
GrNextPow2(uint32_t n)216 static inline uint32_t GrNextPow2(uint32_t n) {
217 return n ? (1 << (32 - SkCLZ(n - 1))) : 1;
218 }
219
220 /**
221 * Returns the next power of 2 >= n or n if the next power of 2 can't be represented by size_t.
222 */
GrNextSizePow2(size_t n)223 static inline size_t GrNextSizePow2(size_t n) {
224 constexpr int kNumSizeTBits = 8 * sizeof(size_t);
225 constexpr size_t kHighBitSet = size_t(1) << (kNumSizeTBits - 1);
226
227 if (!n) {
228 return 1;
229 } else if (n >= kHighBitSet) {
230 return n;
231 }
232
233 n--;
234 uint32_t shift = 1;
235 while (shift < kNumSizeTBits) {
236 n |= n >> shift;
237 shift <<= 1;
238 }
239 return n + 1;
240 }
241
242 // conservative check. will return false for very large values that "could" fit
SkFitsInFixed(T x)243 template <typename T> static inline bool SkFitsInFixed(T x) {
244 return SkTAbs(x) <= 32767.0f;
245 }
246
247 #endif
248