Chromium Code Reviews
chromiumcodereview-hr@appspot.gserviceaccount.com (chromiumcodereview-hr) | Please choose your nickname with Settings | Help | Chromium Project | Gerrit Changes | Sign out
(119)

Issues Search
    My Issues | Starred     Open | Closed | All

Side by Side Diff: src/opts/SkNx_sse.h

Issue 2133413002: try to speed-up maprect + round2i + contains (Closed) Base URL: https://skia.googlesource.com/skia.git@master
Patch Set: add dox Created 4 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
« include/private/SkNx_sse.h ('K') | « src/opts/SkNx_neon.h ('k') | tests/RectTest.cpp » ('j') | no next file with comments »
Toggle Intra-line Diffs ('i') | Expand Comments ('e') | Collapse Comments ('c') | Show Comments Hide Comments ('s')
OLDNEW
(Empty)
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 mo re recent.
14 // If you do, make sure this is in a static inline function... anywhere else ris ks violating ODR.
15
16 #define SKNX_IS_FAST
17
18 template <>
19 class SkNx<2, float> {
20 public:
21 SkNx(const __m128& vec) : fVec(vec) {}
22
23 SkNx() {}
24 SkNx(float val) : fVec(_mm_set1_ps(val)) {}
25 static SkNx Load(const void* ptr) {
26 return _mm_castsi128_ps(_mm_loadl_epi64((const __m128i*)ptr));
27 }
28 SkNx(float a, float b) : fVec(_mm_setr_ps(a,b,0,0)) {}
29
30 void store(void* ptr) const { _mm_storel_pi((__m64*)ptr, fVec); }
31
32 SkNx operator + (const SkNx& o) const { return _mm_add_ps(fVec, o.fVec); }
33 SkNx operator - (const SkNx& o) const { return _mm_sub_ps(fVec, o.fVec); }
34 SkNx operator * (const SkNx& o) const { return _mm_mul_ps(fVec, o.fVec); }
35 SkNx operator / (const SkNx& o) const { return _mm_div_ps(fVec, o.fVec); }
36
37 SkNx operator == (const SkNx& o) const { return _mm_cmpeq_ps (fVec, o.fVec); }
38 SkNx operator != (const SkNx& o) const { return _mm_cmpneq_ps(fVec, o.fVec); }
39 SkNx operator < (const SkNx& o) const { return _mm_cmplt_ps (fVec, o.fVec); }
40 SkNx operator > (const SkNx& o) const { return _mm_cmpgt_ps (fVec, o.fVec); }
41 SkNx operator <= (const SkNx& o) const { return _mm_cmple_ps (fVec, o.fVec); }
42 SkNx operator >= (const SkNx& o) const { return _mm_cmpge_ps (fVec, o.fVec); }
43
44 static SkNx Min(const SkNx& l, const SkNx& r) { return _mm_min_ps(l.fVec, r. fVec); }
45 static SkNx Max(const SkNx& l, const SkNx& r) { return _mm_max_ps(l.fVec, r. fVec); }
46
47 SkNx sqrt() const { return _mm_sqrt_ps (fVec); }
48 SkNx rsqrt() const { return _mm_rsqrt_ps(fVec); }
49 SkNx invert() const { return _mm_rcp_ps(fVec); }
50
51 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 bool allTrue() const { return 0xff == (_mm_movemask_epi8(_mm_castps_si128(fV ec)) & 0xff); }
58 bool anyTrue() const { return 0x00 != (_mm_movemask_epi8(_mm_castps_si128(fV ec)) & 0xff); }
59
60 __m128 fVec;
61 };
62
63 template <>
64 class SkNx<4, float> {
65 public:
66 SkNx(const __m128& vec) : fVec(vec) {}
67
68 SkNx() {}
69 SkNx(float val) : fVec( _mm_set1_ps(val) ) {}
70 static SkNx Load(const void* ptr) { return _mm_loadu_ps((const float*)ptr); }
71
72 SkNx(float a, float b, float c, float d) : fVec(_mm_setr_ps(a,b,c,d)) {}
73
74 void store(void* ptr) const { _mm_storeu_ps((float*)ptr, fVec); }
75
76 SkNx operator + (const SkNx& o) const { return _mm_add_ps(fVec, o.fVec); }
77 SkNx operator - (const SkNx& o) const { return _mm_sub_ps(fVec, o.fVec); }
78 SkNx operator * (const SkNx& o) const { return _mm_mul_ps(fVec, o.fVec); }
79 SkNx operator / (const SkNx& o) const { return _mm_div_ps(fVec, o.fVec); }
80
81 SkNx operator == (const SkNx& o) const { return _mm_cmpeq_ps (fVec, o.fVec); }
82 SkNx operator != (const SkNx& o) const { return _mm_cmpneq_ps(fVec, o.fVec); }
83 SkNx operator < (const SkNx& o) const { return _mm_cmplt_ps (fVec, o.fVec); }
84 SkNx operator > (const SkNx& o) const { return _mm_cmpgt_ps (fVec, o.fVec); }
85 SkNx operator <= (const SkNx& o) const { return _mm_cmple_ps (fVec, o.fVec); }
86 SkNx operator >= (const SkNx& o) const { return _mm_cmpge_ps (fVec, o.fVec); }
87
88 static SkNx Min(const SkNx& l, const SkNx& r) { return _mm_min_ps(l.fVec, r. fVec); }
89 static SkNx Max(const SkNx& l, const SkNx& r) { return _mm_max_ps(l.fVec, r. fVec); }
90
91 SkNx abs() const { return _mm_andnot_ps(_mm_set1_ps(-0.0f), fVec); }
92 SkNx floor() const {
93 #if SK_CPU_SSE_LEVEL >= SK_CPU_SSE_LEVEL_SSE41
94 return _mm_floor_ps(fVec);
95 #else
96 // Emulate _mm_floor_ps() with SSE2:
97 // - roundtrip through integers via truncation
98 // - subtract 1 if that's too big (possible for negative values).
99 // This restricts the domain of our inputs to a maximum somehwere around 2^31.
100 // Seems plenty big.
101 __m128 roundtrip = _mm_cvtepi32_ps(_mm_cvttps_epi32(fVec));
102 __m128 too_big = _mm_cmpgt_ps(roundtrip, fVec);
103 return _mm_sub_ps(roundtrip, _mm_and_ps(too_big, _mm_set1_ps(1.0f)));
104 #endif
105 }
106
107 SkNx sqrt() const { return _mm_sqrt_ps (fVec); }
108 SkNx rsqrt() const { return _mm_rsqrt_ps(fVec); }
109 SkNx invert() const { return _mm_rcp_ps(fVec); }
110
111 float operator[](int k) const {
112 SkASSERT(0 <= k && k < 4);
113 union { __m128 v; float fs[4]; } pun = {fVec};
114 return pun.fs[k&3];
115 }
116
117 bool allTrue() const { return 0xffff == _mm_movemask_epi8(_mm_castps_si128(f Vec)); }
118 bool anyTrue() const { return 0x0000 != _mm_movemask_epi8(_mm_castps_si128(f Vec)); }
119
120 SkNx thenElse(const SkNx& t, const SkNx& e) const {
121 #if SK_CPU_SSE_LEVEL >= SK_CPU_SSE_LEVEL_SSE41
122 return _mm_blendv_ps(e.fVec, t.fVec, fVec);
123 #else
124 return _mm_or_ps(_mm_and_ps (fVec, t.fVec),
125 _mm_andnot_ps(fVec, e.fVec));
126 #endif
127 }
128
129 __m128 fVec;
130 };
131
132 template <>
133 class SkNx<4, int> {
134 public:
135 SkNx(const __m128i& vec) : fVec(vec) {}
136
137 SkNx() {}
138 SkNx(int val) : fVec(_mm_set1_epi32(val)) {}
139 static SkNx Load(const void* ptr) { return _mm_loadu_si128((const __m128i*)p tr); }
140 SkNx(int a, int b, int c, int d) : fVec(_mm_setr_epi32(a,b,c,d)) {}
141
142 void store(void* ptr) const { _mm_storeu_si128((__m128i*)ptr, fVec); }
143
144 SkNx operator + (const SkNx& o) const { return _mm_add_epi32(fVec, o.fVec); }
145 SkNx operator - (const SkNx& o) const { return _mm_sub_epi32(fVec, o.fVec); }
146 SkNx operator * (const SkNx& o) const {
147 __m128i mul20 = _mm_mul_epu32(fVec, o.fVec),
148 mul31 = _mm_mul_epu32(_mm_srli_si128(fVec, 4), _mm_srli_si128(o. fVec, 4));
149 return _mm_unpacklo_epi32(_mm_shuffle_epi32(mul20, _MM_SHUFFLE(0,0,2,0)) ,
150 _mm_shuffle_epi32(mul31, _MM_SHUFFLE(0,0,2,0)) );
151 }
152
153 SkNx operator | (const SkNx& o) const { return _mm_or_si128(fVec, o.fVec); }
154
155 SkNx operator << (int bits) const { return _mm_slli_epi32(fVec, bits); }
156 SkNx operator >> (int bits) const { return _mm_srai_epi32(fVec, bits); }
157
158 int operator[](int k) const {
159 SkASSERT(0 <= k && k < 4);
160 union { __m128i v; int is[4]; } pun = {fVec};
161 return pun.is[k&3];
162 }
163
164 __m128i fVec;
165 };
166
167 template <>
168 class SkNx<4, uint16_t> {
169 public:
170 SkNx(const __m128i& vec) : fVec(vec) {}
171
172 SkNx() {}
173 SkNx(uint16_t val) : fVec(_mm_set1_epi16(val)) {}
174 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)) {}
176
177 void store(void* ptr) const { _mm_storel_epi64((__m128i*)ptr, fVec); }
178
179 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); }
181 SkNx operator * (const SkNx& o) const { return _mm_mullo_epi16(fVec, o.fVec) ; }
182
183 SkNx operator << (int bits) const { return _mm_slli_epi16(fVec, bits); }
184 SkNx operator >> (int bits) const { return _mm_srli_epi16(fVec, bits); }
185
186 uint16_t operator[](int k) const {
187 SkASSERT(0 <= k && k < 4);
188 union { __m128i v; uint16_t us[8]; } pun = {fVec};
189 return pun.us[k&3];
190 }
191
192 __m128i fVec;
193 };
194
195 template <>
196 class SkNx<8, uint16_t> {
197 public:
198 SkNx(const __m128i& vec) : fVec(vec) {}
199
200 SkNx() {}
201 SkNx(uint16_t val) : fVec(_mm_set1_epi16(val)) {}
202 static SkNx Load(const void* ptr) { return _mm_loadu_si128((const __m128i*)p tr); }
203 SkNx(uint16_t a, uint16_t b, uint16_t c, uint16_t d,
204 uint16_t e, uint16_t f, uint16_t g, uint16_t h) : fVec(_mm_setr_epi16(a ,b,c,d,e,f,g,h)) {}
205
206 void store(void* ptr) const { _mm_storeu_si128((__m128i*)ptr, fVec); }
207
208 SkNx operator + (const SkNx& o) const { return _mm_add_epi16(fVec, o.fVec); }
209 SkNx operator - (const SkNx& o) const { return _mm_sub_epi16(fVec, o.fVec); }
210 SkNx operator * (const SkNx& o) const { return _mm_mullo_epi16(fVec, o.fVec) ; }
211
212 SkNx operator << (int bits) const { return _mm_slli_epi16(fVec, bits); }
213 SkNx operator >> (int bits) const { return _mm_srli_epi16(fVec, bits); }
214
215 static SkNx Min(const SkNx& a, const SkNx& b) {
216 // No unsigned _mm_min_epu16, so we'll shift into a space where we can u se the
217 // signed version, _mm_min_epi16, then shift back.
218 const uint16_t top = 0x8000; // Keep this separate from _mm_set1_epi16 o r MSVC will whine.
219 const __m128i top_8x = _mm_set1_epi16(top);
220 return _mm_add_epi8(top_8x, _mm_min_epi16(_mm_sub_epi8(a.fVec, top_8x),
221 _mm_sub_epi8(b.fVec, top_8x))) ;
222 }
223
224 SkNx thenElse(const SkNx& t, const SkNx& e) const {
225 return _mm_or_si128(_mm_and_si128 (fVec, t.fVec),
226 _mm_andnot_si128(fVec, e.fVec));
227 }
228
229 uint16_t operator[](int k) const {
230 SkASSERT(0 <= k && k < 8);
231 union { __m128i v; uint16_t us[8]; } pun = {fVec};
232 return pun.us[k&7];
233 }
234
235 __m128i fVec;
236 };
237
238 template <>
239 class SkNx<4, uint8_t> {
240 public:
241 SkNx() {}
242 SkNx(const __m128i& vec) : fVec(vec) {}
243 SkNx(uint8_t a, uint8_t b, uint8_t c, uint8_t d)
244 : fVec(_mm_setr_epi8(a,b,c,d, 0,0,0,0, 0,0,0,0, 0,0,0,0)) {}
245
246
247 static SkNx Load(const void* ptr) { return _mm_cvtsi32_si128(*(const int*)pt r); }
248 void store(void* ptr) const { *(int*)ptr = _mm_cvtsi128_si32(fVec); }
249
250 uint8_t operator[](int k) const {
251 SkASSERT(0 <= k && k < 4);
252 union { __m128i v; uint8_t us[16]; } pun = {fVec};
253 return pun.us[k&3];
254 }
255
256 // TODO as needed
257
258 __m128i fVec;
259 };
260
261 template <>
262 class SkNx<16, uint8_t> {
263 public:
264 SkNx(const __m128i& vec) : fVec(vec) {}
265
266 SkNx() {}
267 SkNx(uint8_t val) : fVec(_mm_set1_epi8(val)) {}
268 static SkNx Load(const void* ptr) { return _mm_loadu_si128((const __m128i*)p tr); }
269 SkNx(uint8_t a, uint8_t b, uint8_t c, uint8_t d,
270 uint8_t e, uint8_t f, uint8_t g, uint8_t h,
271 uint8_t i, uint8_t j, uint8_t k, uint8_t l,
272 uint8_t m, uint8_t n, uint8_t o, uint8_t p)
273 : fVec(_mm_setr_epi8(a,b,c,d, e,f,g,h, i,j,k,l, m,n,o,p)) {}
274
275 void store(void* ptr) const { _mm_storeu_si128((__m128i*)ptr, fVec); }
276
277 SkNx saturatedAdd(const SkNx& o) const { return _mm_adds_epu8(fVec, o.fVec); }
278
279 SkNx operator + (const SkNx& o) const { return _mm_add_epi8(fVec, o.fVec); }
280 SkNx operator - (const SkNx& o) const { return _mm_sub_epi8(fVec, o.fVec); }
281
282 static SkNx Min(const SkNx& a, const SkNx& b) { return _mm_min_epu8(a.fVec, b.fVec); }
283 SkNx operator < (const SkNx& o) const {
284 // There's no unsigned _mm_cmplt_epu8, so we flip the sign bits then use a signed compare.
285 auto flip = _mm_set1_epi8(char(0x80));
286 return _mm_cmplt_epi8(_mm_xor_si128(flip, fVec), _mm_xor_si128(flip, o.f Vec));
287 }
288
289 uint8_t operator[](int k) const {
290 SkASSERT(0 <= k && k < 16);
291 union { __m128i v; uint8_t us[16]; } pun = {fVec};
292 return pun.us[k&15];
293 }
294
295 SkNx thenElse(const SkNx& t, const SkNx& e) const {
296 return _mm_or_si128(_mm_and_si128 (fVec, t.fVec),
297 _mm_andnot_si128(fVec, e.fVec));
298 }
299
300 __m128i fVec;
301 };
302
303 template<> /*static*/ inline Sk4f SkNx_cast<float, int>(const Sk4i& src) {
304 return _mm_cvtepi32_ps(src.fVec);
305 }
306
307 template <> /*static*/ inline Sk4i SkNx_cast<int, float>(const Sk4f& src) {
308 return _mm_cvttps_epi32(src.fVec);
309 }
310
311 template<> /*static*/ inline Sk4h SkNx_cast<uint16_t, float>(const Sk4f& src) {
312 auto _32 = _mm_cvttps_epi32(src.fVec);
313 // Ideally we'd use _mm_packus_epi32 here. But that's SSE4.1+.
314 #if SK_CPU_SSE_LEVEL >= SK_CPU_SSE_LEVEL_SSSE3
315 // With SSSE3, we can just shuffle the low 2 bytes from each lane right into place.
316 const int _ = ~0;
317 return _mm_shuffle_epi8(_32, _mm_setr_epi8(0,1, 4,5, 8,9, 12,13, _,_,_,_,_,_ ,_,_));
318 #else
319 // With SSE2, we have to emulate _mm_packus_epi32 with _mm_packs_epi32:
320 _32 = _mm_sub_epi32(_32, _mm_set1_epi32((int)0x00008000));
321 return _mm_add_epi16(_mm_packs_epi32(_32, _32), _mm_set1_epi16((short)0x8000 ));
322 #endif
323 }
324
325 template<> /*static*/ inline Sk4b SkNx_cast<uint8_t, float>(const Sk4f& src) {
326 auto _32 = _mm_cvttps_epi32(src.fVec);
327 #if SK_CPU_SSE_LEVEL >= SK_CPU_SSE_LEVEL_SSSE3
328 const int _ = ~0;
329 return _mm_shuffle_epi8(_32, _mm_setr_epi8(0,4,8,12, _,_,_,_, _,_,_,_, _,_,_ ,_));
330 #else
331 auto _16 = _mm_packus_epi16(_32, _32);
332 return _mm_packus_epi16(_16, _16);
333 #endif
334 }
335
336 template<> /*static*/ inline Sk4f SkNx_cast<float, uint8_t>(const Sk4b& src) {
337 #if SK_CPU_SSE_LEVEL >= SK_CPU_SSE_LEVEL_SSSE3
338 const int _ = ~0;
339 auto _32 = _mm_shuffle_epi8(src.fVec, _mm_setr_epi8(0,_,_,_, 1,_,_,_, 2,_,_, _, 3,_,_,_));
340 #else
341 auto _16 = _mm_unpacklo_epi8(src.fVec, _mm_setzero_si128()),
342 _32 = _mm_unpacklo_epi16(_16, _mm_setzero_si128());
343 #endif
344 return _mm_cvtepi32_ps(_32);
345 }
346
347 template<> /*static*/ inline Sk4f SkNx_cast<float, uint16_t>(const Sk4h& src) {
348 auto _32 = _mm_unpacklo_epi16(src.fVec, _mm_setzero_si128());
349 return _mm_cvtepi32_ps(_32);
350 }
351
352 template<> /*static*/ inline Sk16b SkNx_cast<uint8_t, float>(const Sk16f& src) {
353 Sk8f ab, cd;
354 SkNx_split(src, &ab, &cd);
355
356 Sk4f a,b,c,d;
357 SkNx_split(ab, &a, &b);
358 SkNx_split(cd, &c, &d);
359
360 return _mm_packus_epi16(_mm_packus_epi16(_mm_cvttps_epi32(a.fVec),
361 _mm_cvttps_epi32(b.fVec)),
362 _mm_packus_epi16(_mm_cvttps_epi32(c.fVec),
363 _mm_cvttps_epi32(d.fVec)));
364 }
365
366 template<> /*static*/ inline Sk4h SkNx_cast<uint16_t, uint8_t>(const Sk4b& src) {
367 return _mm_unpacklo_epi8(src.fVec, _mm_setzero_si128());
368 }
369
370 template<> /*static*/ inline Sk4b SkNx_cast<uint8_t, uint16_t>(const Sk4h& src) {
371 return _mm_packus_epi16(src.fVec, src.fVec);
372 }
373
374 #endif//SkNx_sse_DEFINED
OLDNEW
« include/private/SkNx_sse.h ('K') | « src/opts/SkNx_neon.h ('k') | tests/RectTest.cpp » ('j') | no next file with comments »

Powered by Google App Engine
This is Rietveld 408576698