Chromium Code Reviews
chromiumcodereview-hr@appspot.gserviceaccount.com (chromiumcodereview-hr) | Please choose your nickname with Settings | Help | Chromium Project | Gerrit Changes | Sign out
(680)

Side by Side Diff: include/core/SkFloatingPoint.h

Issue 1248503004: Have Windows use _mm_rsqrt_ss too. (Closed) Base URL: https://skia.googlesource.com/skia.git@master
Patch Set: Created 5 years, 5 months ago
Use n/p to move between diff chunks; N/P to move between comments. Draft comments are only viewable by you.
Jump to:
View unified diff | Download patch
« no previous file with comments | « no previous file | no next file » | no next file with comments »
Toggle Intra-line Diffs ('i') | Expand Comments ('e') | Collapse Comments ('c') | Show Comments Hide Comments ('s')
OLDNEW
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
(...skipping 125 matching lines...) Expand 10 before | Expand all | Expand 10 after
136 #define sk_double_ceil2int(x) (int)ceil(x) 136 #define sk_double_ceil2int(x) (int)ceil(x)
137 137
138 extern const uint32_t gIEEENotANumber; 138 extern const uint32_t gIEEENotANumber;
139 extern const uint32_t gIEEEInfinity; 139 extern const uint32_t gIEEEInfinity;
140 extern const uint32_t gIEEENegativeInfinity; 140 extern const uint32_t gIEEENegativeInfinity;
141 141
142 #define SK_FloatNaN (*SkTCast<const float*>(&gIEEENotANumber)) 142 #define SK_FloatNaN (*SkTCast<const float*>(&gIEEENotANumber))
143 #define SK_FloatInfinity (*SkTCast<const float*>(&gIEEEInfinity)) 143 #define SK_FloatInfinity (*SkTCast<const float*>(&gIEEEInfinity))
144 #define SK_FloatNegativeInfinity (*SkTCast<const float*>(&gIEEENegativeInfini ty)) 144 #define SK_FloatNegativeInfinity (*SkTCast<const float*>(&gIEEENegativeInfini ty))
145 145
146 #if defined(__SSE__)
147 #include <xmmintrin.h>
148 #elif defined(SK_ARM_HAS_NEON)
149 #include <arm_neon.h>
150 #endif
151
152 // Fast, approximate inverse square root. 146 // Fast, approximate inverse square root.
153 // Compare to name-brand "1.0f / sk_float_sqrt(x)". Should be around 10x faster on SSE, 2x on NEON. 147 // Compare to name-brand "1.0f / sk_float_sqrt(x)". Should be around 10x faster on SSE, 2x on NEON.
154 static inline float sk_float_rsqrt(const float x) { 148 static inline float sk_float_rsqrt(const float x) {
155 // We want all this inlined, so we'll inline SIMD and just take the hit when we don't know we've got 149 // We want all this inlined, so we'll inline SIMD and just take the hit when we don't know we've got
156 // it at compile time. This is going to be too fast to productively hide behind a function pointer. 150 // it at compile time. This is going to be too fast to productively hide behind a function pointer.
157 // 151 //
158 // We do one step of Newton's method to refine the estimates in the NEON and nul l paths. No 152 // We do one step of Newton's method to refine the estimates in the NEON and nul l paths. No
159 // refinement is faster, but very innacurate. Two steps is more accurate, but s lower than 1/sqrt. 153 // refinement is faster, but very innacurate. Two steps is more accurate, but s lower than 1/sqrt.
160 #if defined(__SSE__) 154 #if SK_CPU_SSE_LEVEL >= SK_CPU_SSE_LEVEL_SSE1
161 float result; 155 return _mm_cvtss_f32(_mm_rsqrt_ss(_mm_set_ss(x)));
162 _mm_store_ss(&result, _mm_rsqrt_ss(_mm_set_ss(x)));
163 return result;
164 #elif defined(SK_ARM_HAS_NEON) 156 #elif defined(SK_ARM_HAS_NEON)
165 // Get initial estimate. 157 // Get initial estimate.
166 const float32x2_t xx = vdup_n_f32(x); // Clever readers will note we're doi ng everything 2x. 158 const float32x2_t xx = vdup_n_f32(x); // Clever readers will note we're doi ng everything 2x.
167 float32x2_t estimate = vrsqrte_f32(xx); 159 float32x2_t estimate = vrsqrte_f32(xx);
168 160
169 // One step of Newton's method to refine. 161 // One step of Newton's method to refine.
170 const float32x2_t estimate_sq = vmul_f32(estimate, estimate); 162 const float32x2_t estimate_sq = vmul_f32(estimate, estimate);
171 estimate = vmul_f32(estimate, vrsqrts_f32(xx, estimate_sq)); 163 estimate = vmul_f32(estimate, vrsqrts_f32(xx, estimate_sq));
172 return vget_lane_f32(estimate, 0); // 1 will work fine too; the answer's in both places. 164 return vget_lane_f32(estimate, 0); // 1 will work fine too; the answer's in both places.
173 #else 165 #else
174 // Get initial estimate. 166 // Get initial estimate.
175 int i = *SkTCast<int*>(&x); 167 int i = *SkTCast<int*>(&x);
176 i = 0x5f3759df - (i>>1); 168 i = 0x5f3759df - (i>>1);
177 float estimate = *SkTCast<float*>(&i); 169 float estimate = *SkTCast<float*>(&i);
178 170
179 // One step of Newton's method to refine. 171 // One step of Newton's method to refine.
180 const float estimate_sq = estimate*estimate; 172 const float estimate_sq = estimate*estimate;
181 estimate *= (1.5f-0.5f*x*estimate_sq); 173 estimate *= (1.5f-0.5f*x*estimate_sq);
182 return estimate; 174 return estimate;
183 #endif 175 #endif
184 } 176 }
185 177
186 #endif 178 #endif
OLDNEW
« no previous file with comments | « no previous file | no next file » | no next file with comments »

Powered by Google App Engine
This is Rietveld 408576698