SkNx_sse.h revision f44373c119290b501d4aec7385e16d12c28a1f0f
1/* 2 * Copyright 2015 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 SkNx_sse_DEFINED 9#define SkNx_sse_DEFINED 10 11#include <immintrin.h> 12 13// This file may assume <= SSE2, but must check SK_CPU_SSE_LEVEL for anything more recent. 14// If you do, make sure this is in a static inline function... anywhere else risks violating ODR. 15 16namespace { 17 18template <> 19class SkNx<2, float> { 20public: 21 AI SkNx(const __m128& vec) : fVec(vec) {} 22 23 AI SkNx() {} 24 AI SkNx(float val) : fVec(_mm_set1_ps(val)) {} 25 AI static SkNx Load(const void* ptr) { 26 return _mm_castsi128_ps(_mm_loadl_epi64((const __m128i*)ptr)); 27 } 28 AI SkNx(float a, float b) : fVec(_mm_setr_ps(a,b,0,0)) {} 29 30 AI void store(void* ptr) const { _mm_storel_pi((__m64*)ptr, fVec); } 31 32 AI SkNx operator + (const SkNx& o) const { return _mm_add_ps(fVec, o.fVec); } 33 AI SkNx operator - (const SkNx& o) const { return _mm_sub_ps(fVec, o.fVec); } 34 AI SkNx operator * (const SkNx& o) const { return _mm_mul_ps(fVec, o.fVec); } 35 AI SkNx operator / (const SkNx& o) const { return _mm_div_ps(fVec, o.fVec); } 36 37 AI SkNx operator == (const SkNx& o) const { return _mm_cmpeq_ps (fVec, o.fVec); } 38 AI SkNx operator != (const SkNx& o) const { return _mm_cmpneq_ps(fVec, o.fVec); } 39 AI SkNx operator < (const SkNx& o) const { return _mm_cmplt_ps (fVec, o.fVec); } 40 AI SkNx operator > (const SkNx& o) const { return _mm_cmpgt_ps (fVec, o.fVec); } 41 AI SkNx operator <= (const SkNx& o) const { return _mm_cmple_ps (fVec, o.fVec); } 42 AI SkNx operator >= (const SkNx& o) const { return _mm_cmpge_ps (fVec, o.fVec); } 43 44 AI static SkNx Min(const SkNx& l, const SkNx& r) { return _mm_min_ps(l.fVec, r.fVec); } 45 AI static SkNx Max(const SkNx& l, const SkNx& r) { return _mm_max_ps(l.fVec, r.fVec); } 46 47 AI SkNx sqrt() const { return _mm_sqrt_ps (fVec); } 48 AI SkNx rsqrt() const { return _mm_rsqrt_ps(fVec); } 49 AI SkNx invert() const { return _mm_rcp_ps(fVec); } 50 51 AI float operator[](int k) const { 52 SkASSERT(0 <= k && k < 2); 53 union { __m128 v; float fs[4]; } pun = {fVec}; 54 return pun.fs[k&1]; 55 } 56 57 AI bool allTrue() const { return 0xff == (_mm_movemask_epi8(_mm_castps_si128(fVec)) & 0xff); } 58 AI bool anyTrue() const { return 0x00 != (_mm_movemask_epi8(_mm_castps_si128(fVec)) & 0xff); } 59 60 __m128 fVec; 61}; 62 63template <> 64class SkNx<4, float> { 65public: 66 AI SkNx(const __m128& vec) : fVec(vec) {} 67 68 AI SkNx() {} 69 AI SkNx(float val) : fVec( _mm_set1_ps(val) ) {} 70 AI SkNx(float a, float b, float c, float d) : fVec(_mm_setr_ps(a,b,c,d)) {} 71 72 AI static SkNx Load(const void* ptr) { return _mm_loadu_ps((const float*)ptr); } 73 AI void store(void* ptr) const { _mm_storeu_ps((float*)ptr, fVec); } 74 75 AI static void Load4(const void* ptr, SkNx* r, SkNx* g, SkNx* b, SkNx* a) { 76 __m128 v0 = _mm_loadu_ps(((float*)ptr) + 0), 77 v1 = _mm_loadu_ps(((float*)ptr) + 4), 78 v2 = _mm_loadu_ps(((float*)ptr) + 8), 79 v3 = _mm_loadu_ps(((float*)ptr) + 12); 80 _MM_TRANSPOSE4_PS(v0, v1, v2, v3); 81 *r = v0; 82 *g = v1; 83 *b = v2; 84 *a = v3; 85 } 86 AI static void Store4(void* dst, const SkNx& r, const SkNx& g, const SkNx& b, const SkNx& a) { 87 __m128 v0 = r.fVec, 88 v1 = g.fVec, 89 v2 = b.fVec, 90 v3 = a.fVec; 91 _MM_TRANSPOSE4_PS(v0, v1, v2, v3); 92 _mm_storeu_ps(((float*) dst) + 0, v0); 93 _mm_storeu_ps(((float*) dst) + 4, v1); 94 _mm_storeu_ps(((float*) dst) + 8, v2); 95 _mm_storeu_ps(((float*) dst) + 12, v3); 96 } 97 98 AI SkNx operator + (const SkNx& o) const { return _mm_add_ps(fVec, o.fVec); } 99 AI SkNx operator - (const SkNx& o) const { return _mm_sub_ps(fVec, o.fVec); } 100 AI SkNx operator * (const SkNx& o) const { return _mm_mul_ps(fVec, o.fVec); } 101 AI SkNx operator / (const SkNx& o) const { return _mm_div_ps(fVec, o.fVec); } 102 103 AI SkNx operator == (const SkNx& o) const { return _mm_cmpeq_ps (fVec, o.fVec); } 104 AI SkNx operator != (const SkNx& o) const { return _mm_cmpneq_ps(fVec, o.fVec); } 105 AI SkNx operator < (const SkNx& o) const { return _mm_cmplt_ps (fVec, o.fVec); } 106 AI SkNx operator > (const SkNx& o) const { return _mm_cmpgt_ps (fVec, o.fVec); } 107 AI SkNx operator <= (const SkNx& o) const { return _mm_cmple_ps (fVec, o.fVec); } 108 AI SkNx operator >= (const SkNx& o) const { return _mm_cmpge_ps (fVec, o.fVec); } 109 110 AI static SkNx Min(const SkNx& l, const SkNx& r) { return _mm_min_ps(l.fVec, r.fVec); } 111 AI static SkNx Max(const SkNx& l, const SkNx& r) { return _mm_max_ps(l.fVec, r.fVec); } 112 113 AI SkNx abs() const { return _mm_andnot_ps(_mm_set1_ps(-0.0f), fVec); } 114 AI SkNx floor() const { 115 #if SK_CPU_SSE_LEVEL >= SK_CPU_SSE_LEVEL_SSE41 116 return _mm_floor_ps(fVec); 117 #else 118 // Emulate _mm_floor_ps() with SSE2: 119 // - roundtrip through integers via truncation 120 // - subtract 1 if that's too big (possible for negative values). 121 // This restricts the domain of our inputs to a maximum somehwere around 2^31. 122 // Seems plenty big. 123 __m128 roundtrip = _mm_cvtepi32_ps(_mm_cvttps_epi32(fVec)); 124 __m128 too_big = _mm_cmpgt_ps(roundtrip, fVec); 125 return _mm_sub_ps(roundtrip, _mm_and_ps(too_big, _mm_set1_ps(1.0f))); 126 #endif 127 } 128 129 AI SkNx sqrt() const { return _mm_sqrt_ps (fVec); } 130 AI SkNx rsqrt() const { return _mm_rsqrt_ps(fVec); } 131 AI SkNx invert() const { return _mm_rcp_ps(fVec); } 132 133 AI float operator[](int k) const { 134 SkASSERT(0 <= k && k < 4); 135 union { __m128 v; float fs[4]; } pun = {fVec}; 136 return pun.fs[k&3]; 137 } 138 139 AI bool allTrue() const { return 0xffff == _mm_movemask_epi8(_mm_castps_si128(fVec)); } 140 AI bool anyTrue() const { return 0x0000 != _mm_movemask_epi8(_mm_castps_si128(fVec)); } 141 142 AI SkNx thenElse(const SkNx& t, const SkNx& e) const { 143 #if SK_CPU_SSE_LEVEL >= SK_CPU_SSE_LEVEL_SSE41 144 return _mm_blendv_ps(e.fVec, t.fVec, fVec); 145 #else 146 return _mm_or_ps(_mm_and_ps (fVec, t.fVec), 147 _mm_andnot_ps(fVec, e.fVec)); 148 #endif 149 } 150 151 __m128 fVec; 152}; 153 154template <> 155class SkNx<4, int32_t> { 156public: 157 AI SkNx(const __m128i& vec) : fVec(vec) {} 158 159 AI SkNx() {} 160 AI SkNx(int32_t val) : fVec(_mm_set1_epi32(val)) {} 161 AI static SkNx Load(const void* ptr) { return _mm_loadu_si128((const __m128i*)ptr); } 162 AI SkNx(int32_t a, int32_t b, int32_t c, int32_t d) : fVec(_mm_setr_epi32(a,b,c,d)) {} 163 164 AI void store(void* ptr) const { _mm_storeu_si128((__m128i*)ptr, fVec); } 165 166 AI SkNx operator + (const SkNx& o) const { return _mm_add_epi32(fVec, o.fVec); } 167 AI SkNx operator - (const SkNx& o) const { return _mm_sub_epi32(fVec, o.fVec); } 168 AI SkNx operator * (const SkNx& o) const { 169 __m128i mul20 = _mm_mul_epu32(fVec, o.fVec), 170 mul31 = _mm_mul_epu32(_mm_srli_si128(fVec, 4), _mm_srli_si128(o.fVec, 4)); 171 return _mm_unpacklo_epi32(_mm_shuffle_epi32(mul20, _MM_SHUFFLE(0,0,2,0)), 172 _mm_shuffle_epi32(mul31, _MM_SHUFFLE(0,0,2,0))); 173 } 174 175 AI SkNx operator & (const SkNx& o) const { return _mm_and_si128(fVec, o.fVec); } 176 AI SkNx operator | (const SkNx& o) const { return _mm_or_si128(fVec, o.fVec); } 177 AI SkNx operator ^ (const SkNx& o) const { return _mm_xor_si128(fVec, o.fVec); } 178 179 AI SkNx operator << (int bits) const { return _mm_slli_epi32(fVec, bits); } 180 AI SkNx operator >> (int bits) const { return _mm_srai_epi32(fVec, bits); } 181 182 AI SkNx operator == (const SkNx& o) const { return _mm_cmpeq_epi32 (fVec, o.fVec); } 183 AI SkNx operator < (const SkNx& o) const { return _mm_cmplt_epi32 (fVec, o.fVec); } 184 AI SkNx operator > (const SkNx& o) const { return _mm_cmpgt_epi32 (fVec, o.fVec); } 185 186 AI int32_t operator[](int k) const { 187 SkASSERT(0 <= k && k < 4); 188 union { __m128i v; int32_t is[4]; } pun = {fVec}; 189 return pun.is[k&3]; 190 } 191 192 AI SkNx thenElse(const SkNx& t, const SkNx& e) const { 193 #if SK_CPU_SSE_LEVEL >= SK_CPU_SSE_LEVEL_SSE41 194 return _mm_blendv_epi8(e.fVec, t.fVec, fVec); 195 #else 196 return _mm_or_si128(_mm_and_si128 (fVec, t.fVec), 197 _mm_andnot_si128(fVec, e.fVec)); 198 #endif 199 } 200 201 __m128i fVec; 202}; 203 204template <> 205class SkNx<4, uint32_t> { 206public: 207 AI SkNx(const __m128i& vec) : fVec(vec) {} 208 209 AI SkNx() {} 210 AI SkNx(uint32_t val) : fVec(_mm_set1_epi32(val)) {} 211 AI static SkNx Load(const void* ptr) { return _mm_loadu_si128((const __m128i*)ptr); } 212 AI SkNx(uint32_t a, uint32_t b, uint32_t c, uint32_t d) : fVec(_mm_setr_epi32(a,b,c,d)) {} 213 214 AI void store(void* ptr) const { _mm_storeu_si128((__m128i*)ptr, fVec); } 215 216 AI SkNx operator + (const SkNx& o) const { return _mm_add_epi32(fVec, o.fVec); } 217 AI SkNx operator - (const SkNx& o) const { return _mm_sub_epi32(fVec, o.fVec); } 218 // Not quite sure how to best do operator * in SSE2. We probably don't use it. 219 220 AI SkNx operator & (const SkNx& o) const { return _mm_and_si128(fVec, o.fVec); } 221 AI SkNx operator | (const SkNx& o) const { return _mm_or_si128(fVec, o.fVec); } 222 AI SkNx operator ^ (const SkNx& o) const { return _mm_xor_si128(fVec, o.fVec); } 223 224 AI SkNx operator << (int bits) const { return _mm_slli_epi32(fVec, bits); } 225 AI SkNx operator >> (int bits) const { return _mm_srli_epi32(fVec, bits); } 226 227 AI SkNx operator == (const SkNx& o) const { return _mm_cmpeq_epi32 (fVec, o.fVec); } 228 // operator < and > take a little extra fiddling to make work for unsigned ints. 229 230 AI uint32_t operator[](int k) const { 231 SkASSERT(0 <= k && k < 4); 232 union { __m128i v; uint32_t us[4]; } pun = {fVec}; 233 return pun.us[k&3]; 234 } 235 236 AI SkNx thenElse(const SkNx& t, const SkNx& e) const { 237 #if SK_CPU_SSE_LEVEL >= SK_CPU_SSE_LEVEL_SSE41 238 return _mm_blendv_epi8(e.fVec, t.fVec, fVec); 239 #else 240 return _mm_or_si128(_mm_and_si128 (fVec, t.fVec), 241 _mm_andnot_si128(fVec, e.fVec)); 242 #endif 243 } 244 245 __m128i fVec; 246}; 247 248 249template <> 250class SkNx<4, uint16_t> { 251public: 252 AI SkNx(const __m128i& vec) : fVec(vec) {} 253 254 AI SkNx() {} 255 AI SkNx(uint16_t val) : fVec(_mm_set1_epi16(val)) {} 256 AI SkNx(uint16_t a, uint16_t b, uint16_t c, uint16_t d) 257 : fVec(_mm_setr_epi16(a,b,c,d,0,0,0,0)) {} 258 259 AI static SkNx Load(const void* ptr) { return _mm_loadl_epi64((const __m128i*)ptr); } 260 AI void store(void* ptr) const { _mm_storel_epi64((__m128i*)ptr, fVec); } 261 262 AI static void Load4(const void* ptr, SkNx* r, SkNx* g, SkNx* b, SkNx* a) { 263 __m128i lo = _mm_loadu_si128(((__m128i*)ptr) + 0), 264 hi = _mm_loadu_si128(((__m128i*)ptr) + 1); 265 __m128i even = _mm_unpacklo_epi16(lo, hi), // r0 r2 g0 g2 b0 b2 a0 a2 266 odd = _mm_unpackhi_epi16(lo, hi); // r1 r3 ... 267 __m128i rg = _mm_unpacklo_epi16(even, odd), // r0 r1 r2 r3 g0 g1 g2 g3 268 ba = _mm_unpackhi_epi16(even, odd); // b0 b1 ... a0 a1 ... 269 *r = rg; 270 *g = _mm_srli_si128(rg, 8); 271 *b = ba; 272 *a = _mm_srli_si128(ba, 8); 273 } 274 AI static void Store4(void* dst, const SkNx& r, const SkNx& g, const SkNx& b, const SkNx& a) { 275 __m128i rg = _mm_unpacklo_epi16(r.fVec, g.fVec); 276 __m128i ba = _mm_unpacklo_epi16(b.fVec, a.fVec); 277 __m128i lo = _mm_unpacklo_epi32(rg, ba); 278 __m128i hi = _mm_unpackhi_epi32(rg, ba); 279 _mm_storeu_si128(((__m128i*) dst) + 0, lo); 280 _mm_storeu_si128(((__m128i*) dst) + 1, hi); 281 } 282 283 AI SkNx operator + (const SkNx& o) const { return _mm_add_epi16(fVec, o.fVec); } 284 AI SkNx operator - (const SkNx& o) const { return _mm_sub_epi16(fVec, o.fVec); } 285 AI SkNx operator * (const SkNx& o) const { return _mm_mullo_epi16(fVec, o.fVec); } 286 287 AI SkNx operator << (int bits) const { return _mm_slli_epi16(fVec, bits); } 288 AI SkNx operator >> (int bits) const { return _mm_srli_epi16(fVec, bits); } 289 290 AI uint16_t operator[](int k) const { 291 SkASSERT(0 <= k && k < 4); 292 union { __m128i v; uint16_t us[8]; } pun = {fVec}; 293 return pun.us[k&3]; 294 } 295 296 __m128i fVec; 297}; 298 299template <> 300class SkNx<8, uint16_t> { 301public: 302 AI SkNx(const __m128i& vec) : fVec(vec) {} 303 304 AI SkNx() {} 305 AI SkNx(uint16_t val) : fVec(_mm_set1_epi16(val)) {} 306 AI SkNx(uint16_t a, uint16_t b, uint16_t c, uint16_t d, 307 uint16_t e, uint16_t f, uint16_t g, uint16_t h) 308 : fVec(_mm_setr_epi16(a,b,c,d,e,f,g,h)) {} 309 310 AI static SkNx Load(const void* ptr) { return _mm_loadu_si128((const __m128i*)ptr); } 311 AI void store(void* ptr) const { _mm_storeu_si128((__m128i*)ptr, fVec); } 312 313 AI static void Load4(const void* ptr, SkNx* r, SkNx* g, SkNx* b, SkNx* a) { 314 // TODO: AVX2 version 315 __m128i _01 = _mm_loadu_si128(((__m128i*)ptr) + 0), 316 _23 = _mm_loadu_si128(((__m128i*)ptr) + 1), 317 _45 = _mm_loadu_si128(((__m128i*)ptr) + 2), 318 _67 = _mm_loadu_si128(((__m128i*)ptr) + 3); 319 320 __m128i _02 = _mm_unpacklo_epi16(_01, _23), // r0 r2 g0 g2 b0 b2 a0 a2 321 _13 = _mm_unpackhi_epi16(_01, _23), // r1 r3 g1 g3 b1 b3 a1 a3 322 _46 = _mm_unpacklo_epi16(_45, _67), 323 _57 = _mm_unpackhi_epi16(_45, _67); 324 325 __m128i rg0123 = _mm_unpacklo_epi16(_02, _13), // r0 r1 r2 r3 g0 g1 g2 g3 326 ba0123 = _mm_unpackhi_epi16(_02, _13), // b0 b1 b2 b3 a0 a1 a2 a3 327 rg4567 = _mm_unpacklo_epi16(_46, _57), 328 ba4567 = _mm_unpackhi_epi16(_46, _57); 329 330 *r = _mm_unpacklo_epi64(rg0123, rg4567); 331 *g = _mm_unpackhi_epi64(rg0123, rg4567); 332 *b = _mm_unpacklo_epi64(ba0123, ba4567); 333 *a = _mm_unpackhi_epi64(ba0123, ba4567); 334 } 335 AI static void Store4(void* ptr, const SkNx& r, const SkNx& g, const SkNx& b, const SkNx& a) { 336 // TODO: AVX2 version 337 __m128i rg0123 = _mm_unpacklo_epi16(r.fVec, g.fVec), // r0 g0 r1 g1 r2 g2 r3 g3 338 rg4567 = _mm_unpackhi_epi16(r.fVec, g.fVec), // r4 g4 r5 g5 r6 g6 r7 g7 339 ba0123 = _mm_unpacklo_epi16(b.fVec, a.fVec), 340 ba4567 = _mm_unpackhi_epi16(b.fVec, a.fVec); 341 342 _mm_storeu_si128((__m128i*)ptr + 0, _mm_unpacklo_epi32(rg0123, ba0123)); 343 _mm_storeu_si128((__m128i*)ptr + 1, _mm_unpackhi_epi32(rg0123, ba0123)); 344 _mm_storeu_si128((__m128i*)ptr + 2, _mm_unpacklo_epi32(rg4567, ba4567)); 345 _mm_storeu_si128((__m128i*)ptr + 3, _mm_unpackhi_epi32(rg4567, ba4567)); 346 } 347 348 AI SkNx operator + (const SkNx& o) const { return _mm_add_epi16(fVec, o.fVec); } 349 AI SkNx operator - (const SkNx& o) const { return _mm_sub_epi16(fVec, o.fVec); } 350 AI SkNx operator * (const SkNx& o) const { return _mm_mullo_epi16(fVec, o.fVec); } 351 352 AI SkNx operator << (int bits) const { return _mm_slli_epi16(fVec, bits); } 353 AI SkNx operator >> (int bits) const { return _mm_srli_epi16(fVec, bits); } 354 355 AI static SkNx Min(const SkNx& a, const SkNx& b) { 356 // No unsigned _mm_min_epu16, so we'll shift into a space where we can use the 357 // signed version, _mm_min_epi16, then shift back. 358 const uint16_t top = 0x8000; // Keep this separate from _mm_set1_epi16 or MSVC will whine. 359 const __m128i top_8x = _mm_set1_epi16(top); 360 return _mm_add_epi8(top_8x, _mm_min_epi16(_mm_sub_epi8(a.fVec, top_8x), 361 _mm_sub_epi8(b.fVec, top_8x))); 362 } 363 364 AI SkNx thenElse(const SkNx& t, const SkNx& e) const { 365 return _mm_or_si128(_mm_and_si128 (fVec, t.fVec), 366 _mm_andnot_si128(fVec, e.fVec)); 367 } 368 369 AI uint16_t operator[](int k) const { 370 SkASSERT(0 <= k && k < 8); 371 union { __m128i v; uint16_t us[8]; } pun = {fVec}; 372 return pun.us[k&7]; 373 } 374 375 __m128i fVec; 376}; 377 378template <> 379class SkNx<4, uint8_t> { 380public: 381 AI SkNx() {} 382 AI SkNx(const __m128i& vec) : fVec(vec) {} 383 AI SkNx(uint8_t a, uint8_t b, uint8_t c, uint8_t d) 384 : fVec(_mm_setr_epi8(a,b,c,d, 0,0,0,0, 0,0,0,0, 0,0,0,0)) {} 385 386 387 AI static SkNx Load(const void* ptr) { return _mm_cvtsi32_si128(*(const int*)ptr); } 388 AI void store(void* ptr) const { *(int*)ptr = _mm_cvtsi128_si32(fVec); } 389 390 AI uint8_t operator[](int k) const { 391 SkASSERT(0 <= k && k < 4); 392 union { __m128i v; uint8_t us[16]; } pun = {fVec}; 393 return pun.us[k&3]; 394 } 395 396 // TODO as needed 397 398 __m128i fVec; 399}; 400 401template <> 402class SkNx<16, uint8_t> { 403public: 404 AI SkNx(const __m128i& vec) : fVec(vec) {} 405 406 AI SkNx() {} 407 AI SkNx(uint8_t val) : fVec(_mm_set1_epi8(val)) {} 408 AI static SkNx Load(const void* ptr) { return _mm_loadu_si128((const __m128i*)ptr); } 409 AI SkNx(uint8_t a, uint8_t b, uint8_t c, uint8_t d, 410 uint8_t e, uint8_t f, uint8_t g, uint8_t h, 411 uint8_t i, uint8_t j, uint8_t k, uint8_t l, 412 uint8_t m, uint8_t n, uint8_t o, uint8_t p) 413 : fVec(_mm_setr_epi8(a,b,c,d, e,f,g,h, i,j,k,l, m,n,o,p)) {} 414 415 AI void store(void* ptr) const { _mm_storeu_si128((__m128i*)ptr, fVec); } 416 417 AI SkNx saturatedAdd(const SkNx& o) const { return _mm_adds_epu8(fVec, o.fVec); } 418 419 AI SkNx operator + (const SkNx& o) const { return _mm_add_epi8(fVec, o.fVec); } 420 AI SkNx operator - (const SkNx& o) const { return _mm_sub_epi8(fVec, o.fVec); } 421 422 AI static SkNx Min(const SkNx& a, const SkNx& b) { return _mm_min_epu8(a.fVec, b.fVec); } 423 AI SkNx operator < (const SkNx& o) const { 424 // There's no unsigned _mm_cmplt_epu8, so we flip the sign bits then use a signed compare. 425 auto flip = _mm_set1_epi8(char(0x80)); 426 return _mm_cmplt_epi8(_mm_xor_si128(flip, fVec), _mm_xor_si128(flip, o.fVec)); 427 } 428 429 AI uint8_t operator[](int k) const { 430 SkASSERT(0 <= k && k < 16); 431 union { __m128i v; uint8_t us[16]; } pun = {fVec}; 432 return pun.us[k&15]; 433 } 434 435 AI SkNx thenElse(const SkNx& t, const SkNx& e) const { 436 return _mm_or_si128(_mm_and_si128 (fVec, t.fVec), 437 _mm_andnot_si128(fVec, e.fVec)); 438 } 439 440 __m128i fVec; 441}; 442 443#if SK_CPU_SSE_LEVEL >= SK_CPU_SSE_LEVEL_AVX2 444 445 template <> 446 class SkNx<8, uint8_t> { 447 public: 448 AI SkNx(const __m128i& vec) : fVec(vec) {} 449 450 AI SkNx() {} 451 AI SkNx(uint8_t v) : fVec(_mm_set1_epi8(v)) {} 452 AI SkNx(uint8_t a, uint8_t b, uint8_t c, uint8_t d, 453 uint8_t e, uint8_t f, uint8_t g, uint8_t h) 454 : fVec(_mm_setr_epi8(a,b,c,d, e,f,g,h, 0,0,0,0, 0,0,0,0)) {} 455 456 457 AI static SkNx Load(const void* ptr) { return _mm_loadl_epi64((const __m128i*)ptr); } 458 AI void store(void* ptr) const { _mm_storel_epi64((__m128i*)ptr, fVec); } 459 460 AI uint8_t operator[](int k) const { 461 SkASSERT(0 <= k && k < 8); 462 union { __m128i v; uint8_t us[16]; } pun = {fVec}; 463 return pun.us[k&7]; 464 } 465 466 __m128i fVec; 467 }; 468 469 template <> 470 class SkNx<8, int32_t> { 471 public: 472 AI SkNx(const __m256i& vec) : fVec(vec) {} 473 474 AI SkNx() {} 475 AI SkNx(int32_t v) : fVec(_mm256_set1_epi32(v)) {} 476 AI SkNx(int32_t a, int32_t b, int32_t c, int32_t d, 477 int32_t e, int32_t f, int32_t g, int32_t h) 478 : fVec(_mm256_setr_epi32(a,b,c,d, e,f,g,h)) {} 479 480 AI static SkNx Load(const void* ptr) { return _mm256_loadu_si256((const __m256i*)ptr); } 481 AI void store(void* ptr) const { _mm256_storeu_si256((__m256i*)ptr, fVec); } 482 483 AI SkNx operator + (const SkNx& o) const { return _mm256_add_epi32(fVec, o.fVec); } 484 AI SkNx operator - (const SkNx& o) const { return _mm256_sub_epi32(fVec, o.fVec); } 485 AI SkNx operator * (const SkNx& o) const { return _mm256_mullo_epi32(fVec, o.fVec); } 486 487 AI SkNx operator & (const SkNx& o) const { return _mm256_and_si256(fVec, o.fVec); } 488 AI SkNx operator | (const SkNx& o) const { return _mm256_or_si256(fVec, o.fVec); } 489 AI SkNx operator ^ (const SkNx& o) const { return _mm256_xor_si256(fVec, o.fVec); } 490 491 AI SkNx operator << (int bits) const { return _mm256_slli_epi32(fVec, bits); } 492 AI SkNx operator >> (int bits) const { return _mm256_srai_epi32(fVec, bits); } 493 494 AI int32_t operator[](int k) const { 495 SkASSERT(0 <= k && k < 8); 496 union { __m256i v; int32_t is[8]; } pun = {fVec}; 497 return pun.is[k&7]; 498 } 499 500 __m256i fVec; 501 }; 502 503 template <> 504 class SkNx<8, uint32_t> { 505 public: 506 AI SkNx(const __m256i& vec) : fVec(vec) {} 507 508 AI SkNx() {} 509 AI SkNx(uint32_t v) : fVec(_mm256_set1_epi32(v)) {} 510 AI SkNx(uint32_t a, uint32_t b, uint32_t c, uint32_t d, 511 uint32_t e, uint32_t f, uint32_t g, uint32_t h) 512 : fVec(_mm256_setr_epi32(a,b,c,d, e,f,g,h)) {} 513 514 AI static SkNx Load(const void* ptr) { return _mm256_loadu_si256((const __m256i*)ptr); } 515 AI void store(void* ptr) const { _mm256_storeu_si256((__m256i*)ptr, fVec); } 516 517 AI SkNx operator + (const SkNx& o) const { return _mm256_add_epi32(fVec, o.fVec); } 518 AI SkNx operator - (const SkNx& o) const { return _mm256_sub_epi32(fVec, o.fVec); } 519 AI SkNx operator * (const SkNx& o) const { return _mm256_mullo_epi32(fVec, o.fVec); } 520 521 AI SkNx operator & (const SkNx& o) const { return _mm256_and_si256(fVec, o.fVec); } 522 AI SkNx operator | (const SkNx& o) const { return _mm256_or_si256(fVec, o.fVec); } 523 AI SkNx operator ^ (const SkNx& o) const { return _mm256_xor_si256(fVec, o.fVec); } 524 525 AI SkNx operator << (int bits) const { return _mm256_slli_epi32(fVec, bits); } 526 AI SkNx operator >> (int bits) const { return _mm256_srli_epi32(fVec, bits); } 527 528 AI uint32_t operator[](int k) const { 529 SkASSERT(0 <= k && k < 8); 530 union { __m256i v; uint32_t us[8]; } pun = {fVec}; 531 return pun.us[k&7]; 532 } 533 534 __m256i fVec; 535 }; 536 537 // _mm256_unpack{lo,hi}_pd() auto-casting to and from __m256d. 538 AI static __m256 unpacklo_pd(__m256 x, __m256 y) { 539 return _mm256_castpd_ps(_mm256_unpacklo_pd(_mm256_castps_pd(x), _mm256_castps_pd(y))); 540 } 541 AI static __m256 unpackhi_pd(__m256 x, __m256 y) { 542 return _mm256_castpd_ps(_mm256_unpackhi_pd(_mm256_castps_pd(x), _mm256_castps_pd(y))); 543 } 544 545 template <> 546 class SkNx<8, float> { 547 public: 548 AI SkNx(const __m256& vec) : fVec(vec) {} 549 550 AI SkNx() {} 551 AI SkNx(float val) : fVec(_mm256_set1_ps(val)) {} 552 AI SkNx(float a, float b, float c, float d, 553 float e, float f, float g, float h) : fVec(_mm256_setr_ps(a,b,c,d,e,f,g,h)) {} 554 555 AI static SkNx Load(const void* ptr) { return _mm256_loadu_ps((const float*)ptr); } 556 AI void store(void* ptr) const { _mm256_storeu_ps((float*)ptr, fVec); } 557 558 AI static void Store4(void* ptr, 559 const SkNx& r, const SkNx& g, const SkNx& b, const SkNx& a) { 560 __m256 rg0145 = _mm256_unpacklo_ps(r.fVec, g.fVec), // r0 g0 r1 g1 | r4 g4 r5 g5 561 rg2367 = _mm256_unpackhi_ps(r.fVec, g.fVec), // r2 ... | r6 ... 562 ba0145 = _mm256_unpacklo_ps(b.fVec, a.fVec), // b0 a0 b1 a1 | b4 a4 b5 a5 563 ba2367 = _mm256_unpackhi_ps(b.fVec, a.fVec); // b2 ... | b6 ... 564 565 __m256 _04 = unpacklo_pd(rg0145, ba0145), // r0 g0 b0 a0 | r4 g4 b4 a4 566 _15 = unpackhi_pd(rg0145, ba0145), // r1 ... | r5 ... 567 _26 = unpacklo_pd(rg2367, ba2367), // r2 ... | r6 ... 568 _37 = unpackhi_pd(rg2367, ba2367); // r3 ... | r7 ... 569 570 __m256 _01 = _mm256_permute2f128_ps(_04, _15, 16), // 16 == 010 000 == lo, lo 571 _23 = _mm256_permute2f128_ps(_26, _37, 16), 572 _45 = _mm256_permute2f128_ps(_04, _15, 25), // 25 == 011 001 == hi, hi 573 _67 = _mm256_permute2f128_ps(_26, _37, 25); 574 575 _mm256_storeu_ps((float*)ptr + 0*8, _01); 576 _mm256_storeu_ps((float*)ptr + 1*8, _23); 577 _mm256_storeu_ps((float*)ptr + 2*8, _45); 578 _mm256_storeu_ps((float*)ptr + 3*8, _67); 579 } 580 581 AI SkNx operator+(const SkNx& o) const { return _mm256_add_ps(fVec, o.fVec); } 582 AI SkNx operator-(const SkNx& o) const { return _mm256_sub_ps(fVec, o.fVec); } 583 AI SkNx operator*(const SkNx& o) const { return _mm256_mul_ps(fVec, o.fVec); } 584 AI SkNx operator/(const SkNx& o) const { return _mm256_div_ps(fVec, o.fVec); } 585 586 AI SkNx operator==(const SkNx& o) const { return _mm256_cmp_ps(fVec, o.fVec, _CMP_EQ_OQ); } 587 AI SkNx operator!=(const SkNx& o) const { return _mm256_cmp_ps(fVec, o.fVec, _CMP_NEQ_OQ); } 588 AI SkNx operator <(const SkNx& o) const { return _mm256_cmp_ps(fVec, o.fVec, _CMP_LT_OQ); } 589 AI SkNx operator >(const SkNx& o) const { return _mm256_cmp_ps(fVec, o.fVec, _CMP_GT_OQ); } 590 AI SkNx operator<=(const SkNx& o) const { return _mm256_cmp_ps(fVec, o.fVec, _CMP_LE_OQ); } 591 AI SkNx operator>=(const SkNx& o) const { return _mm256_cmp_ps(fVec, o.fVec, _CMP_GE_OQ); } 592 593 AI static SkNx Min(const SkNx& l, const SkNx& r) { return _mm256_min_ps(l.fVec, r.fVec); } 594 AI static SkNx Max(const SkNx& l, const SkNx& r) { return _mm256_max_ps(l.fVec, r.fVec); } 595 596 AI SkNx sqrt() const { return _mm256_sqrt_ps (fVec); } 597 AI SkNx rsqrt() const { return _mm256_rsqrt_ps(fVec); } 598 AI SkNx invert() const { return _mm256_rcp_ps (fVec); } 599 600 AI SkNx abs() const { return _mm256_andnot_ps(_mm256_set1_ps(-0.0f), fVec); } 601 AI SkNx floor() const { return _mm256_floor_ps(fVec); } 602 603 AI float operator[](int k) const { 604 SkASSERT(0 <= k && k < 8); 605 union { __m256 v; float fs[8]; } pun = {fVec}; 606 return pun.fs[k&7]; 607 } 608 609 AI SkNx thenElse(const SkNx& t, const SkNx& e) const { 610 return _mm256_blendv_ps(e.fVec, t.fVec, fVec); 611 } 612 613 __m256 fVec; 614 }; 615 616 AI static void SkNx_split(const Sk8f& v, Sk4f* lo, Sk4f* hi) { 617 *lo = _mm256_extractf128_ps(v.fVec, 0); 618 *hi = _mm256_extractf128_ps(v.fVec, 1); 619 } 620 621 AI static Sk8f SkNx_join(const Sk4f& lo, const Sk4f& hi) { 622 return _mm256_insertf128_ps(_mm256_castps128_ps256(lo.fVec), hi.fVec, 1); 623 } 624 625 AI static Sk8f SkNx_fma(const Sk8f& a, const Sk8f& b, const Sk8f& c) { 626 return _mm256_fmadd_ps(a.fVec, b.fVec, c.fVec); 627 } 628 629 template<> AI /*static*/ Sk8i SkNx_cast<int>(const Sk8b& src) { 630 return _mm256_cvtepu8_epi32(src.fVec); 631 } 632 633 template<> AI /*static*/ Sk8f SkNx_cast<float>(const Sk8b& src) { 634 return _mm256_cvtepi32_ps(SkNx_cast<int>(src).fVec); 635 } 636 637 template<> AI /*static*/ Sk8f SkNx_cast<float>(const Sk8i& src) { 638 return _mm256_cvtepi32_ps(src.fVec); 639 } 640 641 template<> AI /*static*/ Sk8i SkNx_cast<int>(const Sk8f& src) { 642 return _mm256_cvttps_epi32(src.fVec); 643 } 644 645 template<> AI /*static*/ Sk8i SkNx_cast<int>(const Sk8h& src) { 646 return _mm256_cvtepu16_epi32(src.fVec); 647 } 648 template<> AI /*static*/ Sk8h SkNx_cast<uint16_t>(const Sk8i& src) { 649 __m128i lo = _mm256_extractf128_si256(src.fVec, 0), 650 hi = _mm256_extractf128_si256(src.fVec, 1); 651 return _mm_packus_epi32(lo, hi); 652 } 653 template<> AI /*static*/ Sk8b SkNx_cast<uint8_t>(const Sk8i& src) { 654 auto _16 = SkNx_cast<uint16_t>(src); 655 return _mm_packus_epi16(_16.fVec, _16.fVec); 656 } 657 658#endif 659 660template<> AI /*static*/ Sk4f SkNx_cast<float, int32_t>(const Sk4i& src) { 661 return _mm_cvtepi32_ps(src.fVec); 662} 663template<> AI /*static*/ Sk4f SkNx_cast<float, uint32_t>(const Sk4u& src) { 664 return SkNx_cast<float>(Sk4i::Load(&src)); 665} 666 667template <> AI /*static*/ Sk4i SkNx_cast<int32_t, float>(const Sk4f& src) { 668 return _mm_cvttps_epi32(src.fVec); 669} 670 671template<> AI /*static*/ Sk4h SkNx_cast<uint16_t, int32_t>(const Sk4i& src) { 672#if 0 && SK_CPU_SSE_LEVEL >= SK_CPU_SSE_LEVEL_SSE41 673 // TODO: This seems to be causing code generation problems. Investigate? 674 return _mm_packus_epi32(src.fVec); 675#elif SK_CPU_SSE_LEVEL >= SK_CPU_SSE_LEVEL_SSSE3 676 // With SSSE3, we can just shuffle the low 2 bytes from each lane right into place. 677 const int _ = ~0; 678 return _mm_shuffle_epi8(src.fVec, _mm_setr_epi8(0,1, 4,5, 8,9, 12,13, _,_,_,_,_,_,_,_)); 679#else 680 // With SSE2, we have to sign extend our input, making _mm_packs_epi32 do the pack we want. 681 __m128i x = _mm_srai_epi32(_mm_slli_epi32(src.fVec, 16), 16); 682 return _mm_packs_epi32(x,x); 683#endif 684} 685 686template<> AI /*static*/ Sk4h SkNx_cast<uint16_t, float>(const Sk4f& src) { 687 return SkNx_cast<uint16_t>(SkNx_cast<int32_t>(src)); 688} 689 690template<> AI /*static*/ Sk4b SkNx_cast<uint8_t, float>(const Sk4f& src) { 691 auto _32 = _mm_cvttps_epi32(src.fVec); 692#if SK_CPU_SSE_LEVEL >= SK_CPU_SSE_LEVEL_SSSE3 693 const int _ = ~0; 694 return _mm_shuffle_epi8(_32, _mm_setr_epi8(0,4,8,12, _,_,_,_, _,_,_,_, _,_,_,_)); 695#else 696 auto _16 = _mm_packus_epi16(_32, _32); 697 return _mm_packus_epi16(_16, _16); 698#endif 699} 700 701template<> AI /*static*/ Sk4i SkNx_cast<int32_t, uint8_t>(const Sk4b& src) { 702#if SK_CPU_SSE_LEVEL >= SK_CPU_SSE_LEVEL_SSSE3 703 const int _ = ~0; 704 return _mm_shuffle_epi8(src.fVec, _mm_setr_epi8(0,_,_,_, 1,_,_,_, 2,_,_,_, 3,_,_,_)); 705#else 706 auto _16 = _mm_unpacklo_epi8(src.fVec, _mm_setzero_si128()); 707 return _mm_unpacklo_epi16(_16, _mm_setzero_si128()); 708#endif 709} 710 711template<> AI /*static*/ Sk4f SkNx_cast<float, uint8_t>(const Sk4b& src) { 712 return _mm_cvtepi32_ps(SkNx_cast<int32_t>(src).fVec); 713} 714 715template<> AI /*static*/ Sk4f SkNx_cast<float, uint16_t>(const Sk4h& src) { 716 auto _32 = _mm_unpacklo_epi16(src.fVec, _mm_setzero_si128()); 717 return _mm_cvtepi32_ps(_32); 718} 719 720template<> AI /*static*/ Sk16b SkNx_cast<uint8_t, float>(const Sk16f& src) { 721 Sk8f ab, cd; 722 SkNx_split(src, &ab, &cd); 723 724 Sk4f a,b,c,d; 725 SkNx_split(ab, &a, &b); 726 SkNx_split(cd, &c, &d); 727 728 return _mm_packus_epi16(_mm_packus_epi16(_mm_cvttps_epi32(a.fVec), 729 _mm_cvttps_epi32(b.fVec)), 730 _mm_packus_epi16(_mm_cvttps_epi32(c.fVec), 731 _mm_cvttps_epi32(d.fVec))); 732} 733 734template<> AI /*static*/ Sk4h SkNx_cast<uint16_t, uint8_t>(const Sk4b& src) { 735 return _mm_unpacklo_epi8(src.fVec, _mm_setzero_si128()); 736} 737 738template<> AI /*static*/ Sk4b SkNx_cast<uint8_t, uint16_t>(const Sk4h& src) { 739 return _mm_packus_epi16(src.fVec, src.fVec); 740} 741 742template<> AI /*static*/ Sk4i SkNx_cast<int32_t, uint16_t>(const Sk4h& src) { 743 return _mm_unpacklo_epi16(src.fVec, _mm_setzero_si128()); 744} 745 746template<> AI /*static*/ Sk4b SkNx_cast<uint8_t, int32_t>(const Sk4i& src) { 747 return _mm_packus_epi16(_mm_packus_epi16(src.fVec, src.fVec), src.fVec); 748} 749 750template<> AI /*static*/ Sk4i SkNx_cast<int32_t, uint32_t>(const Sk4u& src) { 751 return src.fVec; 752} 753 754AI static Sk4i Sk4f_round(const Sk4f& x) { 755 return _mm_cvtps_epi32(x.fVec); 756} 757 758} // namespace 759 760#endif//SkNx_sse_DEFINED 761