OLD | NEW |
1 | 1 |
2 /* | 2 /* |
3 * Copyright 2006 The Android Open Source Project | 3 * Copyright 2006 The Android Open Source Project |
4 * | 4 * |
5 * Use of this source code is governed by a BSD-style license that can be | 5 * Use of this source code is governed by a BSD-style license that can be |
6 * found in the LICENSE file. | 6 * found in the LICENSE file. |
7 */ | 7 */ |
8 | 8 |
9 | 9 |
10 #ifndef SkFloatingPoint_DEFINED | 10 #ifndef SkFloatingPoint_DEFINED |
11 #define SkFloatingPoint_DEFINED | 11 #define SkFloatingPoint_DEFINED |
12 | 12 |
13 #include "SkTypes.h" | 13 #include "SkTypes.h" |
14 #include "../private/SkOpts.h" | |
15 | 14 |
16 #include <math.h> | 15 #include <math.h> |
17 #include <float.h> | 16 #include <float.h> |
18 | 17 |
19 // For _POSIX_VERSION | 18 // For _POSIX_VERSION |
20 #if defined(__unix__) || (defined(__APPLE__) && defined(__MACH__)) | 19 #if defined(__unix__) || (defined(__APPLE__) && defined(__MACH__)) |
21 #include <unistd.h> | 20 #include <unistd.h> |
22 #endif | 21 #endif |
23 | 22 |
24 #include "SkFloatBits.h" | 23 #include "SkFloatBits.h" |
(...skipping 96 matching lines...) Expand 10 before | Expand all | Expand 10 after Loading... |
121 #define sk_double_ceil2int(x) (int)ceil(x) | 120 #define sk_double_ceil2int(x) (int)ceil(x) |
122 | 121 |
123 extern const uint32_t gIEEENotANumber; | 122 extern const uint32_t gIEEENotANumber; |
124 extern const uint32_t gIEEEInfinity; | 123 extern const uint32_t gIEEEInfinity; |
125 extern const uint32_t gIEEENegativeInfinity; | 124 extern const uint32_t gIEEENegativeInfinity; |
126 | 125 |
127 #define SK_FloatNaN (*SkTCast<const float*>(&gIEEENotANumber)) | 126 #define SK_FloatNaN (*SkTCast<const float*>(&gIEEENotANumber)) |
128 #define SK_FloatInfinity (*SkTCast<const float*>(&gIEEEInfinity)) | 127 #define SK_FloatInfinity (*SkTCast<const float*>(&gIEEEInfinity)) |
129 #define SK_FloatNegativeInfinity (*SkTCast<const float*>(&gIEEENegativeInfini
ty)) | 128 #define SK_FloatNegativeInfinity (*SkTCast<const float*>(&gIEEENegativeInfini
ty)) |
130 | 129 |
| 130 // We forward declare this to break an #include cycle. |
| 131 // (SkScalar -> SkFloatingPoint -> SkOpts.h -> SkXfermode -> SkColor -> SkScalar
) |
| 132 namespace SkOpts { extern float (*rsqrt)(float); } |
| 133 |
131 // Fast, approximate inverse square root. | 134 // Fast, approximate inverse square root. |
132 // Compare to name-brand "1.0f / sk_float_sqrt(x)". Should be around 10x faster
on SSE, 2x on NEON. | 135 // Compare to name-brand "1.0f / sk_float_sqrt(x)". Should be around 10x faster
on SSE, 2x on NEON. |
133 static inline float sk_float_rsqrt(const float x) { | 136 static inline float sk_float_rsqrt(const float x) { |
134 // We want all this inlined, so we'll inline SIMD and just take the hit when we
don't know we've got | 137 // We want all this inlined, so we'll inline SIMD and just take the hit when we
don't know we've got |
135 // it at compile time. This is going to be too fast to productively hide behind
a function pointer. | 138 // it at compile time. This is going to be too fast to productively hide behind
a function pointer. |
136 // | 139 // |
137 // We do one step of Newton's method to refine the estimates in the NEON and nul
l paths. No | 140 // We do one step of Newton's method to refine the estimates in the NEON and nul
l paths. No |
138 // refinement is faster, but very innacurate. Two steps is more accurate, but s
lower than 1/sqrt. | 141 // refinement is faster, but very innacurate. Two steps is more accurate, but s
lower than 1/sqrt. |
139 // | 142 // |
140 // Optimized constants in the null path courtesy of http://rrrola.wz.cz/inv_sqrt
.html | 143 // Optimized constants in the null path courtesy of http://rrrola.wz.cz/inv_sqrt
.html |
141 #if SK_CPU_SSE_LEVEL >= SK_CPU_SSE_LEVEL_SSE1 | 144 #if SK_CPU_SSE_LEVEL >= SK_CPU_SSE_LEVEL_SSE1 |
142 return _mm_cvtss_f32(_mm_rsqrt_ss(_mm_set_ss(x))); | 145 return _mm_cvtss_f32(_mm_rsqrt_ss(_mm_set_ss(x))); |
143 #elif defined(SK_ARM_HAS_NEON) | 146 #elif defined(SK_ARM_HAS_NEON) |
144 // Get initial estimate. | 147 // Get initial estimate. |
145 const float32x2_t xx = vdup_n_f32(x); // Clever readers will note we're doi
ng everything 2x. | 148 const float32x2_t xx = vdup_n_f32(x); // Clever readers will note we're doi
ng everything 2x. |
146 float32x2_t estimate = vrsqrte_f32(xx); | 149 float32x2_t estimate = vrsqrte_f32(xx); |
147 | 150 |
148 // One step of Newton's method to refine. | 151 // One step of Newton's method to refine. |
149 const float32x2_t estimate_sq = vmul_f32(estimate, estimate); | 152 const float32x2_t estimate_sq = vmul_f32(estimate, estimate); |
150 estimate = vmul_f32(estimate, vrsqrts_f32(xx, estimate_sq)); | 153 estimate = vmul_f32(estimate, vrsqrts_f32(xx, estimate_sq)); |
151 return vget_lane_f32(estimate, 0); // 1 will work fine too; the answer's in
both places. | 154 return vget_lane_f32(estimate, 0); // 1 will work fine too; the answer's in
both places. |
152 #else | 155 #else |
153 // Perhaps runtime-detected NEON, or a portable fallback. | 156 // Perhaps runtime-detected NEON, or a portable fallback. |
154 return SkOpts::rsqrt(x); | 157 return SkOpts::rsqrt(x); |
155 #endif | 158 #endif |
156 } | 159 } |
157 | 160 |
158 #endif | 161 #endif |
OLD | NEW |