OLD | NEW |
1 /* | 1 /* |
2 * Copyright 2015 Google Inc. | 2 * Copyright 2015 Google Inc. |
3 * | 3 * |
4 * Use of this source code is governed by a BSD-style license that can be | 4 * Use of this source code is governed by a BSD-style license that can be |
5 * found in the LICENSE file. | 5 * found in the LICENSE file. |
6 */ | 6 */ |
7 | 7 |
8 #ifndef SkNx_sse_DEFINED | 8 #ifndef SkNx_sse_DEFINED |
9 #define SkNx_sse_DEFINED | 9 #define SkNx_sse_DEFINED |
10 | 10 |
(...skipping 47 matching lines...) Expand 10 before | Expand all | Expand 10 after Loading... |
58 SkNx rsqrt2() const { return this->rsqrt1(); } | 58 SkNx rsqrt2() const { return this->rsqrt1(); } |
59 | 59 |
60 SkNx invert() const { return SkNx(1) / *this; } | 60 SkNx invert() const { return SkNx(1) / *this; } |
61 SkNx approxInvert() const { return _mm_rcp_ps(fVec); } | 61 SkNx approxInvert() const { return _mm_rcp_ps(fVec); } |
62 | 62 |
63 float operator[](int k) const { | 63 float operator[](int k) const { |
64 SkASSERT(0 <= k && k < 2); | 64 SkASSERT(0 <= k && k < 2); |
65 union { __m128 v; float fs[4]; } pun = {fVec}; | 65 union { __m128 v; float fs[4]; } pun = {fVec}; |
66 return pun.fs[k&1]; | 66 return pun.fs[k&1]; |
67 } | 67 } |
68 template <int k> float kth() const { return (*this)[k]; } | |
69 | 68 |
70 bool allTrue() const { return 0xff == (_mm_movemask_epi8(_mm_castps_si128(fV
ec)) & 0xff); } | 69 bool allTrue() const { return 0xff == (_mm_movemask_epi8(_mm_castps_si128(fV
ec)) & 0xff); } |
71 bool anyTrue() const { return 0x00 != (_mm_movemask_epi8(_mm_castps_si128(fV
ec)) & 0xff); } | 70 bool anyTrue() const { return 0x00 != (_mm_movemask_epi8(_mm_castps_si128(fV
ec)) & 0xff); } |
72 | 71 |
73 __m128 fVec; | 72 __m128 fVec; |
74 }; | 73 }; |
75 | 74 |
76 template <> | 75 template <> |
77 class SkNx<4, float> { | 76 class SkNx<4, float> { |
78 public: | 77 public: |
(...skipping 31 matching lines...) Expand 10 before | Expand all | Expand 10 after Loading... |
110 SkNx rsqrt2() const { return this->rsqrt1(); } | 109 SkNx rsqrt2() const { return this->rsqrt1(); } |
111 | 110 |
112 SkNx invert() const { return SkNx(1) / *this; } | 111 SkNx invert() const { return SkNx(1) / *this; } |
113 SkNx approxInvert() const { return _mm_rcp_ps(fVec); } | 112 SkNx approxInvert() const { return _mm_rcp_ps(fVec); } |
114 | 113 |
115 float operator[](int k) const { | 114 float operator[](int k) const { |
116 SkASSERT(0 <= k && k < 4); | 115 SkASSERT(0 <= k && k < 4); |
117 union { __m128 v; float fs[4]; } pun = {fVec}; | 116 union { __m128 v; float fs[4]; } pun = {fVec}; |
118 return pun.fs[k&3]; | 117 return pun.fs[k&3]; |
119 } | 118 } |
120 template <int k> float kth() const { return (*this)[k]; } | |
121 | 119 |
122 bool allTrue() const { return 0xffff == _mm_movemask_epi8(_mm_castps_si128(f
Vec)); } | 120 bool allTrue() const { return 0xffff == _mm_movemask_epi8(_mm_castps_si128(f
Vec)); } |
123 bool anyTrue() const { return 0x0000 != _mm_movemask_epi8(_mm_castps_si128(f
Vec)); } | 121 bool anyTrue() const { return 0x0000 != _mm_movemask_epi8(_mm_castps_si128(f
Vec)); } |
124 | 122 |
125 SkNx thenElse(const SkNx& t, const SkNx& e) const { | 123 SkNx thenElse(const SkNx& t, const SkNx& e) const { |
126 return _mm_or_ps(_mm_and_ps (fVec, t.fVec), | 124 return _mm_or_ps(_mm_and_ps (fVec, t.fVec), |
127 _mm_andnot_ps(fVec, e.fVec)); | 125 _mm_andnot_ps(fVec, e.fVec)); |
128 } | 126 } |
129 | 127 |
130 __m128 fVec; | 128 __m128 fVec; |
(...skipping 21 matching lines...) Expand all Loading... |
152 } | 150 } |
153 | 151 |
154 SkNx operator << (int bits) const { return _mm_slli_epi32(fVec, bits); } | 152 SkNx operator << (int bits) const { return _mm_slli_epi32(fVec, bits); } |
155 SkNx operator >> (int bits) const { return _mm_srai_epi32(fVec, bits); } | 153 SkNx operator >> (int bits) const { return _mm_srai_epi32(fVec, bits); } |
156 | 154 |
157 int operator[](int k) const { | 155 int operator[](int k) const { |
158 SkASSERT(0 <= k && k < 4); | 156 SkASSERT(0 <= k && k < 4); |
159 union { __m128i v; int is[4]; } pun = {fVec}; | 157 union { __m128i v; int is[4]; } pun = {fVec}; |
160 return pun.is[k&3]; | 158 return pun.is[k&3]; |
161 } | 159 } |
162 template <int k> int kth() const { return (*this)[k]; } | |
163 | 160 |
164 __m128i fVec; | 161 __m128i fVec; |
165 }; | 162 }; |
166 | 163 |
167 template <> | 164 template <> |
168 class SkNx<4, uint16_t> { | 165 class SkNx<4, uint16_t> { |
169 public: | 166 public: |
170 SkNx(const __m128i& vec) : fVec(vec) {} | 167 SkNx(const __m128i& vec) : fVec(vec) {} |
171 | 168 |
172 SkNx() {} | 169 SkNx() {} |
173 SkNx(uint16_t val) : fVec(_mm_set1_epi16(val)) {} | 170 SkNx(uint16_t val) : fVec(_mm_set1_epi16(val)) {} |
174 static SkNx Load(const void* ptr) { return _mm_loadl_epi64((const __m128i*)p
tr); } | 171 static SkNx Load(const void* ptr) { return _mm_loadl_epi64((const __m128i*)p
tr); } |
175 SkNx(uint16_t a, uint16_t b, uint16_t c, uint16_t d) : fVec(_mm_setr_epi16(a
,b,c,d,0,0,0,0)) {} | 172 SkNx(uint16_t a, uint16_t b, uint16_t c, uint16_t d) : fVec(_mm_setr_epi16(a
,b,c,d,0,0,0,0)) {} |
176 | 173 |
177 void store(void* ptr) const { _mm_storel_epi64((__m128i*)ptr, fVec); } | 174 void store(void* ptr) const { _mm_storel_epi64((__m128i*)ptr, fVec); } |
178 | 175 |
179 SkNx operator + (const SkNx& o) const { return _mm_add_epi16(fVec, o.fVec);
} | 176 SkNx operator + (const SkNx& o) const { return _mm_add_epi16(fVec, o.fVec);
} |
180 SkNx operator - (const SkNx& o) const { return _mm_sub_epi16(fVec, o.fVec);
} | 177 SkNx operator - (const SkNx& o) const { return _mm_sub_epi16(fVec, o.fVec);
} |
181 SkNx operator * (const SkNx& o) const { return _mm_mullo_epi16(fVec, o.fVec)
; } | 178 SkNx operator * (const SkNx& o) const { return _mm_mullo_epi16(fVec, o.fVec)
; } |
182 | 179 |
183 SkNx operator << (int bits) const { return _mm_slli_epi16(fVec, bits); } | 180 SkNx operator << (int bits) const { return _mm_slli_epi16(fVec, bits); } |
184 SkNx operator >> (int bits) const { return _mm_srli_epi16(fVec, bits); } | 181 SkNx operator >> (int bits) const { return _mm_srli_epi16(fVec, bits); } |
185 | 182 |
186 uint16_t operator[](int k) const { | 183 uint16_t operator[](int k) const { |
187 SkASSERT(0 <= k && k < 4); | 184 SkASSERT(0 <= k && k < 4); |
188 union { __m128i v; uint16_t us[8]; } pun = {fVec}; | 185 union { __m128i v; uint16_t us[8]; } pun = {fVec}; |
189 return pun.us[k&3]; | 186 return pun.us[k&3]; |
190 } | 187 } |
191 template <int k> uint16_t kth() const { return (*this)[k]; } | |
192 | 188 |
193 __m128i fVec; | 189 __m128i fVec; |
194 }; | 190 }; |
195 | 191 |
196 template <> | 192 template <> |
197 class SkNx<8, uint16_t> { | 193 class SkNx<8, uint16_t> { |
198 public: | 194 public: |
199 SkNx(const __m128i& vec) : fVec(vec) {} | 195 SkNx(const __m128i& vec) : fVec(vec) {} |
200 | 196 |
201 SkNx() {} | 197 SkNx() {} |
(...skipping 23 matching lines...) Expand all Loading... |
225 SkNx thenElse(const SkNx& t, const SkNx& e) const { | 221 SkNx thenElse(const SkNx& t, const SkNx& e) const { |
226 return _mm_or_si128(_mm_and_si128 (fVec, t.fVec), | 222 return _mm_or_si128(_mm_and_si128 (fVec, t.fVec), |
227 _mm_andnot_si128(fVec, e.fVec)); | 223 _mm_andnot_si128(fVec, e.fVec)); |
228 } | 224 } |
229 | 225 |
230 uint16_t operator[](int k) const { | 226 uint16_t operator[](int k) const { |
231 SkASSERT(0 <= k && k < 8); | 227 SkASSERT(0 <= k && k < 8); |
232 union { __m128i v; uint16_t us[8]; } pun = {fVec}; | 228 union { __m128i v; uint16_t us[8]; } pun = {fVec}; |
233 return pun.us[k&7]; | 229 return pun.us[k&7]; |
234 } | 230 } |
235 template <int k> uint16_t kth() const { return (*this)[k]; } | |
236 | 231 |
237 __m128i fVec; | 232 __m128i fVec; |
238 }; | 233 }; |
239 | 234 |
240 template <> | 235 template <> |
241 class SkNx<4, uint8_t> { | 236 class SkNx<4, uint8_t> { |
242 public: | 237 public: |
243 SkNx(const __m128i& vec) : fVec(vec) {} | 238 SkNx(const __m128i& vec) : fVec(vec) {} |
244 | 239 |
245 SkNx() {} | 240 SkNx() {} |
(...skipping 31 matching lines...) Expand 10 before | Expand all | Expand 10 after Loading... |
277 // There's no unsigned _mm_cmplt_epu8, so we flip the sign bits then use
a signed compare. | 272 // There's no unsigned _mm_cmplt_epu8, so we flip the sign bits then use
a signed compare. |
278 auto flip = _mm_set1_epi8(char(0x80)); | 273 auto flip = _mm_set1_epi8(char(0x80)); |
279 return _mm_cmplt_epi8(_mm_xor_si128(flip, fVec), _mm_xor_si128(flip, o.f
Vec)); | 274 return _mm_cmplt_epi8(_mm_xor_si128(flip, fVec), _mm_xor_si128(flip, o.f
Vec)); |
280 } | 275 } |
281 | 276 |
282 uint8_t operator[](int k) const { | 277 uint8_t operator[](int k) const { |
283 SkASSERT(0 <= k && k < 16); | 278 SkASSERT(0 <= k && k < 16); |
284 union { __m128i v; uint8_t us[16]; } pun = {fVec}; | 279 union { __m128i v; uint8_t us[16]; } pun = {fVec}; |
285 return pun.us[k&15]; | 280 return pun.us[k&15]; |
286 } | 281 } |
287 template <int k> uint8_t kth() const { return (*this)[k]; } | |
288 | 282 |
289 SkNx thenElse(const SkNx& t, const SkNx& e) const { | 283 SkNx thenElse(const SkNx& t, const SkNx& e) const { |
290 return _mm_or_si128(_mm_and_si128 (fVec, t.fVec), | 284 return _mm_or_si128(_mm_and_si128 (fVec, t.fVec), |
291 _mm_andnot_si128(fVec, e.fVec)); | 285 _mm_andnot_si128(fVec, e.fVec)); |
292 } | 286 } |
293 | 287 |
294 __m128i fVec; | 288 __m128i fVec; |
295 }; | 289 }; |
296 | 290 |
297 template<> /*static*/ inline Sk4f SkNx_cast<float, int>(const Sk4i& src) { | 291 template<> /*static*/ inline Sk4f SkNx_cast<float, int>(const Sk4i& src) { |
(...skipping 56 matching lines...) Expand 10 before | Expand all | Expand 10 after Loading... |
354 | 348 |
355 template<> /*static*/ inline Sk4h SkNx_cast<uint16_t, uint8_t>(const Sk4b& src)
{ | 349 template<> /*static*/ inline Sk4h SkNx_cast<uint16_t, uint8_t>(const Sk4b& src)
{ |
356 return _mm_unpacklo_epi8(src.fVec, _mm_setzero_si128()); | 350 return _mm_unpacklo_epi8(src.fVec, _mm_setzero_si128()); |
357 } | 351 } |
358 | 352 |
359 template<> /*static*/ inline Sk4b SkNx_cast<uint8_t, uint16_t>(const Sk4h& src)
{ | 353 template<> /*static*/ inline Sk4b SkNx_cast<uint8_t, uint16_t>(const Sk4h& src)
{ |
360 return _mm_packus_epi16(src.fVec, src.fVec); | 354 return _mm_packus_epi16(src.fVec, src.fVec); |
361 } | 355 } |
362 | 356 |
363 #endif//SkNx_sse_DEFINED | 357 #endif//SkNx_sse_DEFINED |
OLD | NEW |