SkNx_sse.h 32 KB

123456789101112131415161718192021222324252627282930313233343536373839404142434445464748495051525354555657585960616263646566676869707172737475767778798081828384858687888990919293949596979899100101102103104105106107108109110111112113114115116117118119120121122123124125126127128129130131132133134135136137138139140141142143144145146147148149150151152153154155156157158159160161162163164165166167168169170171172173174175176177178179180181182183184185186187188189190191192193194195196197198199200201202203204205206207208209210211212213214215216217218219220221222223224225226227228229230231232233234235236237238239240241242243244245246247248249250251252253254255256257258259260261262263264265266267268269270271272273274275276277278279280281282283284285286287288289290291292293294295296297298299300301302303304305306307308309310311312313314315316317318319320321322323324325326327328329330331332333334335336337338339340341342343344345346347348349350351352353354355356357358359360361362363364365366367368369370371372373374375376377378379380381382383384385386387388389390391392393394395396397398399400401402403404405406407408409410411412413414415416417418419420421422423424425426427428429430431432433434435436437438439440441442443444445446447448449450451452453454455456457458459460461462463464465466467468469470471472473474475476477478479480481482483484485486487488489490491492493494495496497498499500501502503504505506507508509510511512513514515516517518519520521522523524525526527528529530531532533534535536537538539540541542543544545546547548549550551552553554555556557558559560561562563564565566567568569570571572573574575576577578579580581582583584585586587588589590591592593594595596597598599600601602603604605606607608609610611612613614615616617618619620621622623624625626627628629630631632633634635636637638639640641642643644645646647648649650651652653654655656657658659660661662663664665666667668669670671672673674675676677678679680681682683684685686687688689690691692693694695696697698699700701702703704705706707708709710711712713714715716717718719720721722723724725726727728729730731732733734735736737738739740741742743744745746747748749750751752753754755756757758759760761762763764765766767768769770771772773774775776777778779780781782783784785786787788789790791792793794795796797798799800801802803804805806807808809810811812813814815816817818819820
  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. #ifndef SkNx_sse_DEFINED
  8. #define SkNx_sse_DEFINED
  9. #include "include/core/SkTypes.h"
  10. #if SK_CPU_SSE_LEVEL >= SK_CPU_SSE_LEVEL_SSE41
  11. #include <smmintrin.h>
  12. #elif SK_CPU_SSE_LEVEL >= SK_CPU_SSE_LEVEL_SSSE3
  13. #include <tmmintrin.h>
  14. #else
  15. #include <emmintrin.h>
  16. #endif
  17. // This file may assume <= SSE2, but must check SK_CPU_SSE_LEVEL for anything more recent.
  18. // If you do, make sure this is in a static inline function... anywhere else risks violating ODR.
  19. namespace { // NOLINT(google-build-namespaces)
  20. // Emulate _mm_floor_ps() with SSE2:
  21. // - roundtrip through integers via truncation
  22. // - subtract 1 if that's too big (possible for negative values).
  23. // This restricts the domain of our inputs to a maximum somehwere around 2^31.
  24. // Seems plenty big.
  25. AI static __m128 emulate_mm_floor_ps(__m128 v) {
  26. __m128 roundtrip = _mm_cvtepi32_ps(_mm_cvttps_epi32(v));
  27. __m128 too_big = _mm_cmpgt_ps(roundtrip, v);
  28. return _mm_sub_ps(roundtrip, _mm_and_ps(too_big, _mm_set1_ps(1.0f)));
  29. }
  30. template <>
  31. class SkNx<2, float> {
  32. public:
  33. AI SkNx(const __m128& vec) : fVec(vec) {}
  34. AI SkNx() {}
  35. AI SkNx(float val) : fVec(_mm_set1_ps(val)) {}
  36. AI static SkNx Load(const void* ptr) {
  37. return _mm_castsi128_ps(_mm_loadl_epi64((const __m128i*)ptr));
  38. }
  39. AI SkNx(float a, float b) : fVec(_mm_setr_ps(a,b,0,0)) {}
  40. AI void store(void* ptr) const { _mm_storel_pi((__m64*)ptr, fVec); }
  41. AI static void Load2(const void* ptr, SkNx* x, SkNx* y) {
  42. const float* m = (const float*)ptr;
  43. *x = SkNx{m[0], m[2]};
  44. *y = SkNx{m[1], m[3]};
  45. }
  46. AI static void Store2(void* dst, const SkNx& a, const SkNx& b) {
  47. auto vals = _mm_unpacklo_ps(a.fVec, b.fVec);
  48. _mm_storeu_ps((float*)dst, vals);
  49. }
  50. AI static void Store3(void* dst, const SkNx& a, const SkNx& b, const SkNx& c) {
  51. auto lo = _mm_setr_ps(a[0], b[0], c[0], a[1]),
  52. hi = _mm_setr_ps(b[1], c[1], 0, 0);
  53. _mm_storeu_ps((float*)dst, lo);
  54. _mm_storel_pi(((__m64*)dst) + 2, hi);
  55. }
  56. AI static void Store4(void* dst, const SkNx& a, const SkNx& b, const SkNx& c, const SkNx& d) {
  57. auto lo = _mm_setr_ps(a[0], b[0], c[0], d[0]),
  58. hi = _mm_setr_ps(a[1], b[1], c[1], d[1]);
  59. _mm_storeu_ps((float*)dst, lo);
  60. _mm_storeu_ps(((float*)dst) + 4, hi);
  61. }
  62. AI SkNx operator - () const { return _mm_xor_ps(_mm_set1_ps(-0.0f), fVec); }
  63. AI SkNx operator + (const SkNx& o) const { return _mm_add_ps(fVec, o.fVec); }
  64. AI SkNx operator - (const SkNx& o) const { return _mm_sub_ps(fVec, o.fVec); }
  65. AI SkNx operator * (const SkNx& o) const { return _mm_mul_ps(fVec, o.fVec); }
  66. AI SkNx operator / (const SkNx& o) const { return _mm_div_ps(fVec, o.fVec); }
  67. AI SkNx operator == (const SkNx& o) const { return _mm_cmpeq_ps (fVec, o.fVec); }
  68. AI SkNx operator != (const SkNx& o) const { return _mm_cmpneq_ps(fVec, o.fVec); }
  69. AI SkNx operator < (const SkNx& o) const { return _mm_cmplt_ps (fVec, o.fVec); }
  70. AI SkNx operator > (const SkNx& o) const { return _mm_cmpgt_ps (fVec, o.fVec); }
  71. AI SkNx operator <= (const SkNx& o) const { return _mm_cmple_ps (fVec, o.fVec); }
  72. AI SkNx operator >= (const SkNx& o) const { return _mm_cmpge_ps (fVec, o.fVec); }
  73. AI static SkNx Min(const SkNx& l, const SkNx& r) { return _mm_min_ps(l.fVec, r.fVec); }
  74. AI static SkNx Max(const SkNx& l, const SkNx& r) { return _mm_max_ps(l.fVec, r.fVec); }
  75. AI SkNx abs() const { return _mm_andnot_ps(_mm_set1_ps(-0.0f), fVec); }
  76. AI SkNx floor() const {
  77. #if SK_CPU_SSE_LEVEL >= SK_CPU_SSE_LEVEL_SSE41
  78. return _mm_floor_ps(fVec);
  79. #else
  80. return emulate_mm_floor_ps(fVec);
  81. #endif
  82. }
  83. AI SkNx sqrt() const { return _mm_sqrt_ps (fVec); }
  84. AI SkNx rsqrt() const { return _mm_rsqrt_ps(fVec); }
  85. AI SkNx invert() const { return _mm_rcp_ps(fVec); }
  86. AI float operator[](int k) const {
  87. SkASSERT(0 <= k && k < 2);
  88. union { __m128 v; float fs[4]; } pun = {fVec};
  89. return pun.fs[k&1];
  90. }
  91. AI bool allTrue() const { return 0xff == (_mm_movemask_epi8(_mm_castps_si128(fVec)) & 0xff); }
  92. AI bool anyTrue() const { return 0x00 != (_mm_movemask_epi8(_mm_castps_si128(fVec)) & 0xff); }
  93. AI SkNx thenElse(const SkNx& t, const SkNx& e) const {
  94. #if SK_CPU_SSE_LEVEL >= SK_CPU_SSE_LEVEL_SSE41
  95. return _mm_blendv_ps(e.fVec, t.fVec, fVec);
  96. #else
  97. return _mm_or_ps(_mm_and_ps (fVec, t.fVec),
  98. _mm_andnot_ps(fVec, e.fVec));
  99. #endif
  100. }
  101. __m128 fVec;
  102. };
  103. template <>
  104. class SkNx<4, float> {
  105. public:
  106. AI SkNx(const __m128& vec) : fVec(vec) {}
  107. AI SkNx() {}
  108. AI SkNx(float val) : fVec( _mm_set1_ps(val) ) {}
  109. AI SkNx(float a, float b, float c, float d) : fVec(_mm_setr_ps(a,b,c,d)) {}
  110. AI static SkNx Load(const void* ptr) { return _mm_loadu_ps((const float*)ptr); }
  111. AI void store(void* ptr) const { _mm_storeu_ps((float*)ptr, fVec); }
  112. AI static void Load2(const void* ptr, SkNx* x, SkNx* y) {
  113. SkNx lo = SkNx::Load((const float*)ptr+0),
  114. hi = SkNx::Load((const float*)ptr+4);
  115. *x = SkNx{lo[0], lo[2], hi[0], hi[2]};
  116. *y = SkNx{lo[1], lo[3], hi[1], hi[3]};
  117. }
  118. AI static void Load4(const void* ptr, SkNx* r, SkNx* g, SkNx* b, SkNx* a) {
  119. __m128 v0 = _mm_loadu_ps(((float*)ptr) + 0),
  120. v1 = _mm_loadu_ps(((float*)ptr) + 4),
  121. v2 = _mm_loadu_ps(((float*)ptr) + 8),
  122. v3 = _mm_loadu_ps(((float*)ptr) + 12);
  123. _MM_TRANSPOSE4_PS(v0, v1, v2, v3);
  124. *r = v0;
  125. *g = v1;
  126. *b = v2;
  127. *a = v3;
  128. }
  129. AI static void Store4(void* dst, const SkNx& r, const SkNx& g, const SkNx& b, const SkNx& a) {
  130. __m128 v0 = r.fVec,
  131. v1 = g.fVec,
  132. v2 = b.fVec,
  133. v3 = a.fVec;
  134. _MM_TRANSPOSE4_PS(v0, v1, v2, v3);
  135. _mm_storeu_ps(((float*) dst) + 0, v0);
  136. _mm_storeu_ps(((float*) dst) + 4, v1);
  137. _mm_storeu_ps(((float*) dst) + 8, v2);
  138. _mm_storeu_ps(((float*) dst) + 12, v3);
  139. }
  140. AI SkNx operator - () const { return _mm_xor_ps(_mm_set1_ps(-0.0f), fVec); }
  141. AI SkNx operator + (const SkNx& o) const { return _mm_add_ps(fVec, o.fVec); }
  142. AI SkNx operator - (const SkNx& o) const { return _mm_sub_ps(fVec, o.fVec); }
  143. AI SkNx operator * (const SkNx& o) const { return _mm_mul_ps(fVec, o.fVec); }
  144. AI SkNx operator / (const SkNx& o) const { return _mm_div_ps(fVec, o.fVec); }
  145. AI SkNx operator == (const SkNx& o) const { return _mm_cmpeq_ps (fVec, o.fVec); }
  146. AI SkNx operator != (const SkNx& o) const { return _mm_cmpneq_ps(fVec, o.fVec); }
  147. AI SkNx operator < (const SkNx& o) const { return _mm_cmplt_ps (fVec, o.fVec); }
  148. AI SkNx operator > (const SkNx& o) const { return _mm_cmpgt_ps (fVec, o.fVec); }
  149. AI SkNx operator <= (const SkNx& o) const { return _mm_cmple_ps (fVec, o.fVec); }
  150. AI SkNx operator >= (const SkNx& o) const { return _mm_cmpge_ps (fVec, o.fVec); }
  151. AI static SkNx Min(const SkNx& l, const SkNx& r) { return _mm_min_ps(l.fVec, r.fVec); }
  152. AI static SkNx Max(const SkNx& l, const SkNx& r) { return _mm_max_ps(l.fVec, r.fVec); }
  153. AI SkNx abs() const { return _mm_andnot_ps(_mm_set1_ps(-0.0f), fVec); }
  154. AI SkNx floor() const {
  155. #if SK_CPU_SSE_LEVEL >= SK_CPU_SSE_LEVEL_SSE41
  156. return _mm_floor_ps(fVec);
  157. #else
  158. return emulate_mm_floor_ps(fVec);
  159. #endif
  160. }
  161. AI SkNx sqrt() const { return _mm_sqrt_ps (fVec); }
  162. AI SkNx rsqrt() const { return _mm_rsqrt_ps(fVec); }
  163. AI SkNx invert() const { return _mm_rcp_ps(fVec); }
  164. AI float operator[](int k) const {
  165. SkASSERT(0 <= k && k < 4);
  166. union { __m128 v; float fs[4]; } pun = {fVec};
  167. return pun.fs[k&3];
  168. }
  169. AI float min() const {
  170. SkNx min = Min(*this, _mm_shuffle_ps(fVec, fVec, _MM_SHUFFLE(2,3,0,1)));
  171. min = Min(min, _mm_shuffle_ps(min.fVec, min.fVec, _MM_SHUFFLE(0,1,2,3)));
  172. return min[0];
  173. }
  174. AI float max() const {
  175. SkNx max = Max(*this, _mm_shuffle_ps(fVec, fVec, _MM_SHUFFLE(2,3,0,1)));
  176. max = Max(max, _mm_shuffle_ps(max.fVec, max.fVec, _MM_SHUFFLE(0,1,2,3)));
  177. return max[0];
  178. }
  179. AI bool allTrue() const { return 0xffff == _mm_movemask_epi8(_mm_castps_si128(fVec)); }
  180. AI bool anyTrue() const { return 0x0000 != _mm_movemask_epi8(_mm_castps_si128(fVec)); }
  181. AI SkNx thenElse(const SkNx& t, const SkNx& e) const {
  182. #if SK_CPU_SSE_LEVEL >= SK_CPU_SSE_LEVEL_SSE41
  183. return _mm_blendv_ps(e.fVec, t.fVec, fVec);
  184. #else
  185. return _mm_or_ps(_mm_and_ps (fVec, t.fVec),
  186. _mm_andnot_ps(fVec, e.fVec));
  187. #endif
  188. }
  189. __m128 fVec;
  190. };
  191. AI static __m128i mullo32(__m128i a, __m128i b) {
  192. #if SK_CPU_SSE_LEVEL >= SK_CPU_SSE_LEVEL_SSE41
  193. return _mm_mullo_epi32(a, b);
  194. #else
  195. __m128i mul20 = _mm_mul_epu32(a, b),
  196. mul31 = _mm_mul_epu32(_mm_srli_si128(a, 4), _mm_srli_si128(b, 4));
  197. return _mm_unpacklo_epi32(_mm_shuffle_epi32(mul20, _MM_SHUFFLE(0,0,2,0)),
  198. _mm_shuffle_epi32(mul31, _MM_SHUFFLE(0,0,2,0)));
  199. #endif
  200. }
  201. template <>
  202. class SkNx<4, int32_t> {
  203. public:
  204. AI SkNx(const __m128i& vec) : fVec(vec) {}
  205. AI SkNx() {}
  206. AI SkNx(int32_t val) : fVec(_mm_set1_epi32(val)) {}
  207. AI static SkNx Load(const void* ptr) { return _mm_loadu_si128((const __m128i*)ptr); }
  208. AI SkNx(int32_t a, int32_t b, int32_t c, int32_t d) : fVec(_mm_setr_epi32(a,b,c,d)) {}
  209. AI void store(void* ptr) const { _mm_storeu_si128((__m128i*)ptr, fVec); }
  210. AI SkNx operator + (const SkNx& o) const { return _mm_add_epi32(fVec, o.fVec); }
  211. AI SkNx operator - (const SkNx& o) const { return _mm_sub_epi32(fVec, o.fVec); }
  212. AI SkNx operator * (const SkNx& o) const { return mullo32(fVec, o.fVec); }
  213. AI SkNx operator & (const SkNx& o) const { return _mm_and_si128(fVec, o.fVec); }
  214. AI SkNx operator | (const SkNx& o) const { return _mm_or_si128(fVec, o.fVec); }
  215. AI SkNx operator ^ (const SkNx& o) const { return _mm_xor_si128(fVec, o.fVec); }
  216. AI SkNx operator << (int bits) const { return _mm_slli_epi32(fVec, bits); }
  217. AI SkNx operator >> (int bits) const { return _mm_srai_epi32(fVec, bits); }
  218. AI SkNx operator == (const SkNx& o) const { return _mm_cmpeq_epi32 (fVec, o.fVec); }
  219. AI SkNx operator < (const SkNx& o) const { return _mm_cmplt_epi32 (fVec, o.fVec); }
  220. AI SkNx operator > (const SkNx& o) const { return _mm_cmpgt_epi32 (fVec, o.fVec); }
  221. AI int32_t operator[](int k) const {
  222. SkASSERT(0 <= k && k < 4);
  223. union { __m128i v; int32_t is[4]; } pun = {fVec};
  224. return pun.is[k&3];
  225. }
  226. AI SkNx thenElse(const SkNx& t, const SkNx& e) const {
  227. #if SK_CPU_SSE_LEVEL >= SK_CPU_SSE_LEVEL_SSE41
  228. return _mm_blendv_epi8(e.fVec, t.fVec, fVec);
  229. #else
  230. return _mm_or_si128(_mm_and_si128 (fVec, t.fVec),
  231. _mm_andnot_si128(fVec, e.fVec));
  232. #endif
  233. }
  234. AI SkNx abs() const {
  235. #if SK_CPU_SSE_LEVEL >= SK_CPU_SSE_LEVEL_SSSE3
  236. return _mm_abs_epi32(fVec);
  237. #else
  238. SkNx mask = (*this) >> 31;
  239. return (mask ^ (*this)) - mask;
  240. #endif
  241. }
  242. AI static SkNx Min(const SkNx& x, const SkNx& y) {
  243. #if SK_CPU_SSE_LEVEL >= SK_CPU_SSE_LEVEL_SSE41
  244. return _mm_min_epi32(x.fVec, y.fVec);
  245. #else
  246. return (x < y).thenElse(x, y);
  247. #endif
  248. }
  249. AI static SkNx Max(const SkNx& x, const SkNx& y) {
  250. #if SK_CPU_SSE_LEVEL >= SK_CPU_SSE_LEVEL_SSE41
  251. return _mm_max_epi32(x.fVec, y.fVec);
  252. #else
  253. return (x > y).thenElse(x, y);
  254. #endif
  255. }
  256. __m128i fVec;
  257. };
  258. template <>
  259. class SkNx<2, uint32_t> {
  260. public:
  261. AI SkNx(const __m128i& vec) : fVec(vec) {}
  262. AI SkNx() {}
  263. AI SkNx(uint32_t val) : fVec(_mm_set1_epi32(val)) {}
  264. AI static SkNx Load(const void* ptr) { return _mm_loadl_epi64((const __m128i*)ptr); }
  265. AI SkNx(uint32_t a, uint32_t b) : fVec(_mm_setr_epi32(a,b,0,0)) {}
  266. AI void store(void* ptr) const { _mm_storel_epi64((__m128i*)ptr, fVec); }
  267. AI SkNx operator + (const SkNx& o) const { return _mm_add_epi32(fVec, o.fVec); }
  268. AI SkNx operator - (const SkNx& o) const { return _mm_sub_epi32(fVec, o.fVec); }
  269. AI SkNx operator * (const SkNx& o) const { return mullo32(fVec, o.fVec); }
  270. AI SkNx operator & (const SkNx& o) const { return _mm_and_si128(fVec, o.fVec); }
  271. AI SkNx operator | (const SkNx& o) const { return _mm_or_si128(fVec, o.fVec); }
  272. AI SkNx operator ^ (const SkNx& o) const { return _mm_xor_si128(fVec, o.fVec); }
  273. AI SkNx operator << (int bits) const { return _mm_slli_epi32(fVec, bits); }
  274. AI SkNx operator >> (int bits) const { return _mm_srli_epi32(fVec, bits); }
  275. AI SkNx operator == (const SkNx& o) const { return _mm_cmpeq_epi32 (fVec, o.fVec); }
  276. AI SkNx operator != (const SkNx& o) const { return (*this == o) ^ 0xffffffff; }
  277. // operator < and > take a little extra fiddling to make work for unsigned ints.
  278. AI uint32_t operator[](int k) const {
  279. SkASSERT(0 <= k && k < 2);
  280. union { __m128i v; uint32_t us[4]; } pun = {fVec};
  281. return pun.us[k&1];
  282. }
  283. AI SkNx thenElse(const SkNx& t, const SkNx& e) const {
  284. #if SK_CPU_SSE_LEVEL >= SK_CPU_SSE_LEVEL_SSE41
  285. return _mm_blendv_epi8(e.fVec, t.fVec, fVec);
  286. #else
  287. return _mm_or_si128(_mm_and_si128 (fVec, t.fVec),
  288. _mm_andnot_si128(fVec, e.fVec));
  289. #endif
  290. }
  291. AI bool allTrue() const { return 0xff == (_mm_movemask_epi8(fVec) & 0xff); }
  292. __m128i fVec;
  293. };
  294. template <>
  295. class SkNx<4, uint32_t> {
  296. public:
  297. AI SkNx(const __m128i& vec) : fVec(vec) {}
  298. AI SkNx() {}
  299. AI SkNx(uint32_t val) : fVec(_mm_set1_epi32(val)) {}
  300. AI static SkNx Load(const void* ptr) { return _mm_loadu_si128((const __m128i*)ptr); }
  301. AI SkNx(uint32_t a, uint32_t b, uint32_t c, uint32_t d) : fVec(_mm_setr_epi32(a,b,c,d)) {}
  302. AI void store(void* ptr) const { _mm_storeu_si128((__m128i*)ptr, fVec); }
  303. AI SkNx operator + (const SkNx& o) const { return _mm_add_epi32(fVec, o.fVec); }
  304. AI SkNx operator - (const SkNx& o) const { return _mm_sub_epi32(fVec, o.fVec); }
  305. AI SkNx operator * (const SkNx& o) const { return mullo32(fVec, o.fVec); }
  306. AI SkNx operator & (const SkNx& o) const { return _mm_and_si128(fVec, o.fVec); }
  307. AI SkNx operator | (const SkNx& o) const { return _mm_or_si128(fVec, o.fVec); }
  308. AI SkNx operator ^ (const SkNx& o) const { return _mm_xor_si128(fVec, o.fVec); }
  309. AI SkNx operator << (int bits) const { return _mm_slli_epi32(fVec, bits); }
  310. AI SkNx operator >> (int bits) const { return _mm_srli_epi32(fVec, bits); }
  311. AI SkNx operator == (const SkNx& o) const { return _mm_cmpeq_epi32 (fVec, o.fVec); }
  312. AI SkNx operator != (const SkNx& o) const { return (*this == o) ^ 0xffffffff; }
  313. // operator < and > take a little extra fiddling to make work for unsigned ints.
  314. AI uint32_t operator[](int k) const {
  315. SkASSERT(0 <= k && k < 4);
  316. union { __m128i v; uint32_t us[4]; } pun = {fVec};
  317. return pun.us[k&3];
  318. }
  319. AI SkNx thenElse(const SkNx& t, const SkNx& e) const {
  320. #if SK_CPU_SSE_LEVEL >= SK_CPU_SSE_LEVEL_SSE41
  321. return _mm_blendv_epi8(e.fVec, t.fVec, fVec);
  322. #else
  323. return _mm_or_si128(_mm_and_si128 (fVec, t.fVec),
  324. _mm_andnot_si128(fVec, e.fVec));
  325. #endif
  326. }
  327. AI SkNx mulHi(SkNx m) const {
  328. SkNx v20{_mm_mul_epu32(m.fVec, fVec)};
  329. SkNx v31{_mm_mul_epu32(_mm_srli_si128(m.fVec, 4), _mm_srli_si128(fVec, 4))};
  330. return SkNx{v20[1], v31[1], v20[3], v31[3]};
  331. }
  332. __m128i fVec;
  333. };
  334. template <>
  335. class SkNx<4, uint16_t> {
  336. public:
  337. AI SkNx(const __m128i& vec) : fVec(vec) {}
  338. AI SkNx() {}
  339. AI SkNx(uint16_t val) : fVec(_mm_set1_epi16(val)) {}
  340. AI SkNx(uint16_t a, uint16_t b, uint16_t c, uint16_t d)
  341. : fVec(_mm_setr_epi16(a,b,c,d,0,0,0,0)) {}
  342. AI static SkNx Load(const void* ptr) { return _mm_loadl_epi64((const __m128i*)ptr); }
  343. AI void store(void* ptr) const { _mm_storel_epi64((__m128i*)ptr, fVec); }
  344. AI static void Load4(const void* ptr, SkNx* r, SkNx* g, SkNx* b, SkNx* a) {
  345. __m128i lo = _mm_loadu_si128(((__m128i*)ptr) + 0),
  346. hi = _mm_loadu_si128(((__m128i*)ptr) + 1);
  347. __m128i even = _mm_unpacklo_epi16(lo, hi), // r0 r2 g0 g2 b0 b2 a0 a2
  348. odd = _mm_unpackhi_epi16(lo, hi); // r1 r3 ...
  349. __m128i rg = _mm_unpacklo_epi16(even, odd), // r0 r1 r2 r3 g0 g1 g2 g3
  350. ba = _mm_unpackhi_epi16(even, odd); // b0 b1 ... a0 a1 ...
  351. *r = rg;
  352. *g = _mm_srli_si128(rg, 8);
  353. *b = ba;
  354. *a = _mm_srli_si128(ba, 8);
  355. }
  356. AI static void Load3(const void* ptr, SkNx* r, SkNx* g, SkNx* b) {
  357. // The idea here is to get 4 vectors that are R G B _ _ _ _ _.
  358. // The second load is at a funny location to make sure we don't read past
  359. // the bounds of memory. This is fine, we just need to shift it a little bit.
  360. const uint8_t* ptr8 = (const uint8_t*) ptr;
  361. __m128i rgb0 = _mm_loadu_si128((const __m128i*) (ptr8 + 0));
  362. __m128i rgb1 = _mm_srli_si128(rgb0, 3*2);
  363. __m128i rgb2 = _mm_srli_si128(_mm_loadu_si128((const __m128i*) (ptr8 + 4*2)), 2*2);
  364. __m128i rgb3 = _mm_srli_si128(rgb2, 3*2);
  365. __m128i rrggbb01 = _mm_unpacklo_epi16(rgb0, rgb1);
  366. __m128i rrggbb23 = _mm_unpacklo_epi16(rgb2, rgb3);
  367. *r = _mm_unpacklo_epi32(rrggbb01, rrggbb23);
  368. *g = _mm_srli_si128(r->fVec, 4*2);
  369. *b = _mm_unpackhi_epi32(rrggbb01, rrggbb23);
  370. }
  371. AI static void Store4(void* dst, const SkNx& r, const SkNx& g, const SkNx& b, const SkNx& a) {
  372. __m128i rg = _mm_unpacklo_epi16(r.fVec, g.fVec);
  373. __m128i ba = _mm_unpacklo_epi16(b.fVec, a.fVec);
  374. __m128i lo = _mm_unpacklo_epi32(rg, ba);
  375. __m128i hi = _mm_unpackhi_epi32(rg, ba);
  376. _mm_storeu_si128(((__m128i*) dst) + 0, lo);
  377. _mm_storeu_si128(((__m128i*) dst) + 1, hi);
  378. }
  379. AI SkNx operator + (const SkNx& o) const { return _mm_add_epi16(fVec, o.fVec); }
  380. AI SkNx operator - (const SkNx& o) const { return _mm_sub_epi16(fVec, o.fVec); }
  381. AI SkNx operator * (const SkNx& o) const { return _mm_mullo_epi16(fVec, o.fVec); }
  382. AI SkNx operator & (const SkNx& o) const { return _mm_and_si128(fVec, o.fVec); }
  383. AI SkNx operator | (const SkNx& o) const { return _mm_or_si128(fVec, o.fVec); }
  384. AI SkNx operator << (int bits) const { return _mm_slli_epi16(fVec, bits); }
  385. AI SkNx operator >> (int bits) const { return _mm_srli_epi16(fVec, bits); }
  386. AI uint16_t operator[](int k) const {
  387. SkASSERT(0 <= k && k < 4);
  388. union { __m128i v; uint16_t us[8]; } pun = {fVec};
  389. return pun.us[k&3];
  390. }
  391. __m128i fVec;
  392. };
  393. template <>
  394. class SkNx<8, uint16_t> {
  395. public:
  396. AI SkNx(const __m128i& vec) : fVec(vec) {}
  397. AI SkNx() {}
  398. AI SkNx(uint16_t val) : fVec(_mm_set1_epi16(val)) {}
  399. AI SkNx(uint16_t a, uint16_t b, uint16_t c, uint16_t d,
  400. uint16_t e, uint16_t f, uint16_t g, uint16_t h)
  401. : fVec(_mm_setr_epi16(a,b,c,d,e,f,g,h)) {}
  402. AI static SkNx Load(const void* ptr) { return _mm_loadu_si128((const __m128i*)ptr); }
  403. AI void store(void* ptr) const { _mm_storeu_si128((__m128i*)ptr, fVec); }
  404. AI static void Load4(const void* ptr, SkNx* r, SkNx* g, SkNx* b, SkNx* a) {
  405. __m128i _01 = _mm_loadu_si128(((__m128i*)ptr) + 0),
  406. _23 = _mm_loadu_si128(((__m128i*)ptr) + 1),
  407. _45 = _mm_loadu_si128(((__m128i*)ptr) + 2),
  408. _67 = _mm_loadu_si128(((__m128i*)ptr) + 3);
  409. __m128i _02 = _mm_unpacklo_epi16(_01, _23), // r0 r2 g0 g2 b0 b2 a0 a2
  410. _13 = _mm_unpackhi_epi16(_01, _23), // r1 r3 g1 g3 b1 b3 a1 a3
  411. _46 = _mm_unpacklo_epi16(_45, _67),
  412. _57 = _mm_unpackhi_epi16(_45, _67);
  413. __m128i rg0123 = _mm_unpacklo_epi16(_02, _13), // r0 r1 r2 r3 g0 g1 g2 g3
  414. ba0123 = _mm_unpackhi_epi16(_02, _13), // b0 b1 b2 b3 a0 a1 a2 a3
  415. rg4567 = _mm_unpacklo_epi16(_46, _57),
  416. ba4567 = _mm_unpackhi_epi16(_46, _57);
  417. *r = _mm_unpacklo_epi64(rg0123, rg4567);
  418. *g = _mm_unpackhi_epi64(rg0123, rg4567);
  419. *b = _mm_unpacklo_epi64(ba0123, ba4567);
  420. *a = _mm_unpackhi_epi64(ba0123, ba4567);
  421. }
  422. AI static void Load3(const void* ptr, SkNx* r, SkNx* g, SkNx* b) {
  423. const uint8_t* ptr8 = (const uint8_t*) ptr;
  424. __m128i rgb0 = _mm_loadu_si128((const __m128i*) (ptr8 + 0*2));
  425. __m128i rgb1 = _mm_srli_si128(rgb0, 3*2);
  426. __m128i rgb2 = _mm_loadu_si128((const __m128i*) (ptr8 + 6*2));
  427. __m128i rgb3 = _mm_srli_si128(rgb2, 3*2);
  428. __m128i rgb4 = _mm_loadu_si128((const __m128i*) (ptr8 + 12*2));
  429. __m128i rgb5 = _mm_srli_si128(rgb4, 3*2);
  430. __m128i rgb6 = _mm_srli_si128(_mm_loadu_si128((const __m128i*) (ptr8 + 16*2)), 2*2);
  431. __m128i rgb7 = _mm_srli_si128(rgb6, 3*2);
  432. __m128i rgb01 = _mm_unpacklo_epi16(rgb0, rgb1);
  433. __m128i rgb23 = _mm_unpacklo_epi16(rgb2, rgb3);
  434. __m128i rgb45 = _mm_unpacklo_epi16(rgb4, rgb5);
  435. __m128i rgb67 = _mm_unpacklo_epi16(rgb6, rgb7);
  436. __m128i rg03 = _mm_unpacklo_epi32(rgb01, rgb23);
  437. __m128i bx03 = _mm_unpackhi_epi32(rgb01, rgb23);
  438. __m128i rg47 = _mm_unpacklo_epi32(rgb45, rgb67);
  439. __m128i bx47 = _mm_unpackhi_epi32(rgb45, rgb67);
  440. *r = _mm_unpacklo_epi64(rg03, rg47);
  441. *g = _mm_unpackhi_epi64(rg03, rg47);
  442. *b = _mm_unpacklo_epi64(bx03, bx47);
  443. }
  444. AI static void Store4(void* ptr, const SkNx& r, const SkNx& g, const SkNx& b, const SkNx& a) {
  445. __m128i rg0123 = _mm_unpacklo_epi16(r.fVec, g.fVec), // r0 g0 r1 g1 r2 g2 r3 g3
  446. rg4567 = _mm_unpackhi_epi16(r.fVec, g.fVec), // r4 g4 r5 g5 r6 g6 r7 g7
  447. ba0123 = _mm_unpacklo_epi16(b.fVec, a.fVec),
  448. ba4567 = _mm_unpackhi_epi16(b.fVec, a.fVec);
  449. _mm_storeu_si128((__m128i*)ptr + 0, _mm_unpacklo_epi32(rg0123, ba0123));
  450. _mm_storeu_si128((__m128i*)ptr + 1, _mm_unpackhi_epi32(rg0123, ba0123));
  451. _mm_storeu_si128((__m128i*)ptr + 2, _mm_unpacklo_epi32(rg4567, ba4567));
  452. _mm_storeu_si128((__m128i*)ptr + 3, _mm_unpackhi_epi32(rg4567, ba4567));
  453. }
  454. AI SkNx operator + (const SkNx& o) const { return _mm_add_epi16(fVec, o.fVec); }
  455. AI SkNx operator - (const SkNx& o) const { return _mm_sub_epi16(fVec, o.fVec); }
  456. AI SkNx operator * (const SkNx& o) const { return _mm_mullo_epi16(fVec, o.fVec); }
  457. AI SkNx operator & (const SkNx& o) const { return _mm_and_si128(fVec, o.fVec); }
  458. AI SkNx operator | (const SkNx& o) const { return _mm_or_si128(fVec, o.fVec); }
  459. AI SkNx operator << (int bits) const { return _mm_slli_epi16(fVec, bits); }
  460. AI SkNx operator >> (int bits) const { return _mm_srli_epi16(fVec, bits); }
  461. AI static SkNx Min(const SkNx& a, const SkNx& b) {
  462. // No unsigned _mm_min_epu16, so we'll shift into a space where we can use the
  463. // signed version, _mm_min_epi16, then shift back.
  464. const uint16_t top = 0x8000; // Keep this separate from _mm_set1_epi16 or MSVC will whine.
  465. const __m128i top_8x = _mm_set1_epi16(top);
  466. return _mm_add_epi8(top_8x, _mm_min_epi16(_mm_sub_epi8(a.fVec, top_8x),
  467. _mm_sub_epi8(b.fVec, top_8x)));
  468. }
  469. AI SkNx mulHi(const SkNx& m) const {
  470. return _mm_mulhi_epu16(fVec, m.fVec);
  471. }
  472. AI SkNx thenElse(const SkNx& t, const SkNx& e) const {
  473. return _mm_or_si128(_mm_and_si128 (fVec, t.fVec),
  474. _mm_andnot_si128(fVec, e.fVec));
  475. }
  476. AI uint16_t operator[](int k) const {
  477. SkASSERT(0 <= k && k < 8);
  478. union { __m128i v; uint16_t us[8]; } pun = {fVec};
  479. return pun.us[k&7];
  480. }
  481. __m128i fVec;
  482. };
  483. template <>
  484. class SkNx<4, uint8_t> {
  485. public:
  486. AI SkNx() {}
  487. AI SkNx(const __m128i& vec) : fVec(vec) {}
  488. AI SkNx(uint8_t a, uint8_t b, uint8_t c, uint8_t d)
  489. : fVec(_mm_setr_epi8(a,b,c,d, 0,0,0,0, 0,0,0,0, 0,0,0,0)) {}
  490. AI static SkNx Load(const void* ptr) { return _mm_cvtsi32_si128(*(const int*)ptr); }
  491. AI void store(void* ptr) const { *(int*)ptr = _mm_cvtsi128_si32(fVec); }
  492. AI uint8_t operator[](int k) const {
  493. SkASSERT(0 <= k && k < 4);
  494. union { __m128i v; uint8_t us[16]; } pun = {fVec};
  495. return pun.us[k&3];
  496. }
  497. // TODO as needed
  498. __m128i fVec;
  499. };
  500. template <>
  501. class SkNx<8, uint8_t> {
  502. public:
  503. AI SkNx(const __m128i& vec) : fVec(vec) {}
  504. AI SkNx() {}
  505. AI SkNx(uint8_t val) : fVec(_mm_set1_epi8(val)) {}
  506. AI static SkNx Load(const void* ptr) { return _mm_loadl_epi64((const __m128i*)ptr); }
  507. AI SkNx(uint8_t a, uint8_t b, uint8_t c, uint8_t d,
  508. uint8_t e, uint8_t f, uint8_t g, uint8_t h)
  509. : fVec(_mm_setr_epi8(a,b,c,d, e,f,g,h, 0,0,0,0, 0,0,0,0)) {}
  510. AI void store(void* ptr) const {_mm_storel_epi64((__m128i*)ptr, fVec);}
  511. AI SkNx saturatedAdd(const SkNx& o) const { return _mm_adds_epu8(fVec, o.fVec); }
  512. AI SkNx operator + (const SkNx& o) const { return _mm_add_epi8(fVec, o.fVec); }
  513. AI SkNx operator - (const SkNx& o) const { return _mm_sub_epi8(fVec, o.fVec); }
  514. AI static SkNx Min(const SkNx& a, const SkNx& b) { return _mm_min_epu8(a.fVec, b.fVec); }
  515. AI SkNx operator < (const SkNx& o) const {
  516. // There's no unsigned _mm_cmplt_epu8, so we flip the sign bits then use a signed compare.
  517. auto flip = _mm_set1_epi8(char(0x80));
  518. return _mm_cmplt_epi8(_mm_xor_si128(flip, fVec), _mm_xor_si128(flip, o.fVec));
  519. }
  520. AI uint8_t operator[](int k) const {
  521. SkASSERT(0 <= k && k < 16);
  522. union { __m128i v; uint8_t us[16]; } pun = {fVec};
  523. return pun.us[k&15];
  524. }
  525. AI SkNx thenElse(const SkNx& t, const SkNx& e) const {
  526. return _mm_or_si128(_mm_and_si128 (fVec, t.fVec),
  527. _mm_andnot_si128(fVec, e.fVec));
  528. }
  529. __m128i fVec;
  530. };
  531. template <>
  532. class SkNx<16, uint8_t> {
  533. public:
  534. AI SkNx(const __m128i& vec) : fVec(vec) {}
  535. AI SkNx() {}
  536. AI SkNx(uint8_t val) : fVec(_mm_set1_epi8(val)) {}
  537. AI static SkNx Load(const void* ptr) { return _mm_loadu_si128((const __m128i*)ptr); }
  538. AI SkNx(uint8_t a, uint8_t b, uint8_t c, uint8_t d,
  539. uint8_t e, uint8_t f, uint8_t g, uint8_t h,
  540. uint8_t i, uint8_t j, uint8_t k, uint8_t l,
  541. uint8_t m, uint8_t n, uint8_t o, uint8_t p)
  542. : fVec(_mm_setr_epi8(a,b,c,d, e,f,g,h, i,j,k,l, m,n,o,p)) {}
  543. AI void store(void* ptr) const { _mm_storeu_si128((__m128i*)ptr, fVec); }
  544. AI SkNx saturatedAdd(const SkNx& o) const { return _mm_adds_epu8(fVec, o.fVec); }
  545. AI SkNx operator + (const SkNx& o) const { return _mm_add_epi8(fVec, o.fVec); }
  546. AI SkNx operator - (const SkNx& o) const { return _mm_sub_epi8(fVec, o.fVec); }
  547. AI SkNx operator & (const SkNx& o) const { return _mm_and_si128(fVec, o.fVec); }
  548. AI static SkNx Min(const SkNx& a, const SkNx& b) { return _mm_min_epu8(a.fVec, b.fVec); }
  549. AI SkNx operator < (const SkNx& o) const {
  550. // There's no unsigned _mm_cmplt_epu8, so we flip the sign bits then use a signed compare.
  551. auto flip = _mm_set1_epi8(char(0x80));
  552. return _mm_cmplt_epi8(_mm_xor_si128(flip, fVec), _mm_xor_si128(flip, o.fVec));
  553. }
  554. AI uint8_t operator[](int k) const {
  555. SkASSERT(0 <= k && k < 16);
  556. union { __m128i v; uint8_t us[16]; } pun = {fVec};
  557. return pun.us[k&15];
  558. }
  559. AI SkNx thenElse(const SkNx& t, const SkNx& e) const {
  560. return _mm_or_si128(_mm_and_si128 (fVec, t.fVec),
  561. _mm_andnot_si128(fVec, e.fVec));
  562. }
  563. __m128i fVec;
  564. };
  565. template<> AI /*static*/ Sk4f SkNx_cast<float, int32_t>(const Sk4i& src) {
  566. return _mm_cvtepi32_ps(src.fVec);
  567. }
  568. template<> AI /*static*/ Sk4f SkNx_cast<float, uint32_t>(const Sk4u& src) {
  569. return SkNx_cast<float>(Sk4i::Load(&src));
  570. }
  571. template <> AI /*static*/ Sk4i SkNx_cast<int32_t, float>(const Sk4f& src) {
  572. return _mm_cvttps_epi32(src.fVec);
  573. }
  574. template<> AI /*static*/ Sk4h SkNx_cast<uint16_t, int32_t>(const Sk4i& src) {
  575. #if 0 && SK_CPU_SSE_LEVEL >= SK_CPU_SSE_LEVEL_SSE41
  576. // TODO: This seems to be causing code generation problems. Investigate?
  577. return _mm_packus_epi32(src.fVec);
  578. #elif SK_CPU_SSE_LEVEL >= SK_CPU_SSE_LEVEL_SSSE3
  579. // With SSSE3, we can just shuffle the low 2 bytes from each lane right into place.
  580. const int _ = ~0;
  581. return _mm_shuffle_epi8(src.fVec, _mm_setr_epi8(0,1, 4,5, 8,9, 12,13, _,_,_,_,_,_,_,_));
  582. #else
  583. // With SSE2, we have to sign extend our input, making _mm_packs_epi32 do the pack we want.
  584. __m128i x = _mm_srai_epi32(_mm_slli_epi32(src.fVec, 16), 16);
  585. return _mm_packs_epi32(x,x);
  586. #endif
  587. }
  588. template<> AI /*static*/ Sk4h SkNx_cast<uint16_t, float>(const Sk4f& src) {
  589. return SkNx_cast<uint16_t>(SkNx_cast<int32_t>(src));
  590. }
  591. template<> AI /*static*/ Sk4b SkNx_cast<uint8_t, float>(const Sk4f& src) {
  592. auto _32 = _mm_cvttps_epi32(src.fVec);
  593. #if SK_CPU_SSE_LEVEL >= SK_CPU_SSE_LEVEL_SSSE3
  594. const int _ = ~0;
  595. return _mm_shuffle_epi8(_32, _mm_setr_epi8(0,4,8,12, _,_,_,_, _,_,_,_, _,_,_,_));
  596. #else
  597. auto _16 = _mm_packus_epi16(_32, _32);
  598. return _mm_packus_epi16(_16, _16);
  599. #endif
  600. }
  601. template<> AI /*static*/ Sk4u SkNx_cast<uint32_t, uint8_t>(const Sk4b& src) {
  602. #if SK_CPU_SSE_LEVEL >= SK_CPU_SSE_LEVEL_SSSE3
  603. const int _ = ~0;
  604. return _mm_shuffle_epi8(src.fVec, _mm_setr_epi8(0,_,_,_, 1,_,_,_, 2,_,_,_, 3,_,_,_));
  605. #else
  606. auto _16 = _mm_unpacklo_epi8(src.fVec, _mm_setzero_si128());
  607. return _mm_unpacklo_epi16(_16, _mm_setzero_si128());
  608. #endif
  609. }
  610. template<> AI /*static*/ Sk4i SkNx_cast<int32_t, uint8_t>(const Sk4b& src) {
  611. return SkNx_cast<uint32_t>(src).fVec;
  612. }
  613. template<> AI /*static*/ Sk4f SkNx_cast<float, uint8_t>(const Sk4b& src) {
  614. return _mm_cvtepi32_ps(SkNx_cast<int32_t>(src).fVec);
  615. }
  616. template<> AI /*static*/ Sk4f SkNx_cast<float, uint16_t>(const Sk4h& src) {
  617. auto _32 = _mm_unpacklo_epi16(src.fVec, _mm_setzero_si128());
  618. return _mm_cvtepi32_ps(_32);
  619. }
  620. template<> AI /*static*/ Sk8b SkNx_cast<uint8_t, int32_t>(const Sk8i& src) {
  621. Sk4i lo, hi;
  622. SkNx_split(src, &lo, &hi);
  623. auto t = _mm_packs_epi32(lo.fVec, hi.fVec);
  624. return _mm_packus_epi16(t, t);
  625. }
  626. template<> AI /*static*/ Sk16b SkNx_cast<uint8_t, float>(const Sk16f& src) {
  627. Sk8f ab, cd;
  628. SkNx_split(src, &ab, &cd);
  629. Sk4f a,b,c,d;
  630. SkNx_split(ab, &a, &b);
  631. SkNx_split(cd, &c, &d);
  632. return _mm_packus_epi16(_mm_packus_epi16(_mm_cvttps_epi32(a.fVec),
  633. _mm_cvttps_epi32(b.fVec)),
  634. _mm_packus_epi16(_mm_cvttps_epi32(c.fVec),
  635. _mm_cvttps_epi32(d.fVec)));
  636. }
  637. template<> AI /*static*/ Sk4h SkNx_cast<uint16_t, uint8_t>(const Sk4b& src) {
  638. return _mm_unpacklo_epi8(src.fVec, _mm_setzero_si128());
  639. }
  640. template<> AI /*static*/ Sk8h SkNx_cast<uint16_t, uint8_t>(const Sk8b& src) {
  641. return _mm_unpacklo_epi8(src.fVec, _mm_setzero_si128());
  642. }
  643. template<> AI /*static*/ Sk4b SkNx_cast<uint8_t, uint16_t>(const Sk4h& src) {
  644. return _mm_packus_epi16(src.fVec, src.fVec);
  645. }
  646. template<> AI /*static*/ Sk8b SkNx_cast<uint8_t, uint16_t>(const Sk8h& src) {
  647. return _mm_packus_epi16(src.fVec, src.fVec);
  648. }
  649. template<> AI /*static*/ Sk4i SkNx_cast<int32_t, uint16_t>(const Sk4h& src) {
  650. return _mm_unpacklo_epi16(src.fVec, _mm_setzero_si128());
  651. }
  652. template<> AI /*static*/ Sk4b SkNx_cast<uint8_t, int32_t>(const Sk4i& src) {
  653. return _mm_packus_epi16(_mm_packus_epi16(src.fVec, src.fVec), src.fVec);
  654. }
  655. template<> AI /*static*/ Sk4b SkNx_cast<uint8_t, uint32_t>(const Sk4u& src) {
  656. return _mm_packus_epi16(_mm_packus_epi16(src.fVec, src.fVec), src.fVec);
  657. }
  658. template<> AI /*static*/ Sk4i SkNx_cast<int32_t, uint32_t>(const Sk4u& src) {
  659. return src.fVec;
  660. }
  661. AI static Sk4i Sk4f_round(const Sk4f& x) {
  662. return _mm_cvtps_epi32(x.fVec);
  663. }
  664. } // namespace
  665. #endif//SkNx_sse_DEFINED