1 | /* Copyright (C) 2011-2014 Povilas Kanapickas <povilas@radix.lt> |
2 | |
3 | Distributed under the Boost Software License, Version 1.0. |
4 | (See accompanying file LICENSE_1_0.txt or copy at |
5 | http://www.boost.org/LICENSE_1_0.txt) |
6 | */ |
7 | |
8 | #ifndef LIBSIMDPP_SIMDPP_DETAIL_INSN_I_SHIFT_L_H |
9 | #define LIBSIMDPP_SIMDPP_DETAIL_INSN_I_SHIFT_L_H |
10 | |
11 | #ifndef LIBSIMDPP_SIMD_H |
12 | #error "This file must be included through simd.h" |
13 | #endif |
14 | |
15 | #include <simdpp/types.h> |
16 | #include <simdpp/detail/not_implemented.h> |
17 | #include <simdpp/core/bit_and.h> |
18 | #include <simdpp/core/bit_andnot.h> |
19 | #include <simdpp/core/set_splat.h> |
20 | #include <simdpp/detail/insn/i_shift.h> |
21 | #include <simdpp/detail/null/math.h> |
22 | #include <simdpp/detail/vector_array_macros.h> |
23 | |
24 | namespace simdpp { |
25 | namespace SIMDPP_ARCH_NAMESPACE { |
26 | namespace detail { |
27 | namespace insn { |
28 | |
29 | |
30 | static SIMDPP_INL |
31 | uint8x16 i_shift_l(const uint8x16& a, unsigned count) |
32 | { |
33 | #if SIMDPP_USE_NULL |
34 | return detail::null::shift_l(a, count); |
35 | #elif SIMDPP_USE_AVX2 |
36 | uint16x8 mask, a16; |
37 | uint16_t mask1 = (0x00ff >> (8-count)) << 8; |
38 | |
39 | a16 = a; |
40 | mask = splat(mask1); |
41 | a16 = shift_l(a16, count); |
42 | a16 = bit_andnot(a16, mask); |
43 | return uint8x16(a16); |
44 | #elif SIMDPP_USE_SSE2 |
45 | uint16x8 mask, a16; |
46 | mask = make_ones(); |
47 | mask = shift_r(mask, 16-count); |
48 | mask = shift_l(mask, 8); |
49 | |
50 | a16 = a; |
51 | a16 = shift_l(a16, count); |
52 | a16 = bit_andnot(a16, mask); |
53 | return uint8x16(a16); |
54 | #elif SIMDPP_USE_NEON |
55 | int8x16 shift = splat(count); |
56 | return vshlq_u8(a.native(), shift.native()); |
57 | #elif SIMDPP_USE_ALTIVEC |
58 | uint8x16 shift = splat(count); |
59 | return vec_sl(a.native(), shift.native()); |
60 | #elif SIMDPP_USE_MSA |
61 | int8x16 shift = splat(count); |
62 | return (v16u8) __msa_sll_b((v16i8)a.native(), shift.native()); |
63 | #endif |
64 | } |
65 | |
66 | #if SIMDPP_USE_AVX2 |
67 | static SIMDPP_INL |
68 | uint8x32 i_shift_l(const uint8x32& a, unsigned count) |
69 | { |
70 | uint16x16 mask, a16; |
71 | uint16_t mask1 = (0x00ff >> (8-count)) << 8; |
72 | |
73 | a16 = a; |
74 | mask = splat(mask1); |
75 | a16 = shift_l(a16, count); |
76 | a16 = bit_andnot(a16, mask); |
77 | return uint8<32>(a16); |
78 | } |
79 | #endif |
80 | |
81 | #if SIMDPP_USE_AVX512BW |
82 | SIMDPP_INL uint8<64> i_shift_l(const uint8<64>& a, unsigned count) |
83 | { |
84 | uint16<32> mask, a16; |
85 | uint16_t mask1 = (0x00ff >> (8-count)) << 8; |
86 | |
87 | a16 = a; |
88 | mask = splat(mask1); |
89 | a16 = shift_l(a16, count); |
90 | a16 = bit_andnot(a16, mask); |
91 | return uint8<64>(a16); |
92 | } |
93 | #endif |
94 | |
95 | // ----------------------------------------------------------------------------- |
96 | |
97 | static SIMDPP_INL |
98 | uint16x8 i_shift_l(const uint16x8& a, unsigned count) |
99 | { |
100 | #if SIMDPP_USE_NULL |
101 | return detail::null::shift_l(a, count); |
102 | #elif SIMDPP_USE_SSE2 |
103 | return _mm_sll_epi16(a.native(), _mm_cvtsi32_si128(count)); |
104 | #elif SIMDPP_USE_NEON |
105 | int16x8 shift = splat(count); |
106 | return vshlq_u16(a.native(), shift.native()); |
107 | #elif SIMDPP_USE_ALTIVEC |
108 | uint16x8 shift = splat(count); |
109 | return vec_sl(a.native(), shift.native()); |
110 | #elif SIMDPP_USE_MSA |
111 | int16x8 shift = splat(count); |
112 | return (v8u16) __msa_sll_h((v8i16) a.native(), shift.native()); |
113 | #endif |
114 | } |
115 | |
116 | #if SIMDPP_USE_AVX2 |
117 | static SIMDPP_INL |
118 | uint16x16 i_shift_l(const uint16x16& a, unsigned count) |
119 | { |
120 | #if SIMDPP_WORKAROUND_AVX2_SHIFT_INTRINSICS |
121 | __m256i r = a.native(); |
122 | __m128i x = _mm_cvtsi32_si128(count); |
123 | __asm("vpsllw %1, %2, %0" : "=x" (r) : "x" (x), "x" (r)); |
124 | return r; |
125 | #else |
126 | return _mm256_sll_epi16(a.native(), _mm_cvtsi32_si128(count)); |
127 | #endif |
128 | } |
129 | #endif |
130 | |
131 | #if SIMDPP_USE_AVX512BW |
132 | SIMDPP_INL uint16<32> i_shift_l(const uint16<32>& a, unsigned count) |
133 | { |
134 | return _mm512_sll_epi16(a.native(), _mm_cvtsi32_si128(count)); |
135 | } |
136 | #endif |
137 | |
138 | // ----------------------------------------------------------------------------- |
139 | |
140 | static SIMDPP_INL |
141 | uint32x4 i_shift_l(const uint32x4& a, unsigned count) |
142 | { |
143 | #if SIMDPP_USE_NULL |
144 | return detail::null::shift_l(a, count); |
145 | #elif SIMDPP_USE_SSE2 |
146 | return _mm_sll_epi32(a.native(), _mm_cvtsi32_si128(count)); |
147 | #elif SIMDPP_USE_NEON |
148 | int32x4 shift = splat(count); |
149 | return vshlq_u32(a.native(), shift.native()); |
150 | #elif SIMDPP_USE_ALTIVEC |
151 | uint32x4 shift = splat(count); |
152 | return vec_sl(a.native(), shift.native()); |
153 | #elif SIMDPP_USE_MSA |
154 | int32x4 shift = splat(count); |
155 | return (v4u32) __msa_sll_w((v4i32) a.native(), shift.native()); |
156 | #endif |
157 | } |
158 | |
159 | #if SIMDPP_USE_AVX2 |
160 | static SIMDPP_INL |
161 | uint32x8 i_shift_l(const uint32x8& a, unsigned count) |
162 | { |
163 | #if SIMDPP_WORKAROUND_AVX2_SHIFT_INTRINSICS |
164 | __m256i r = a.native(); |
165 | __m128i x = _mm_cvtsi32_si128(count); |
166 | __asm("vpslld %1, %2, %0" : "=x" (r) : "x" (x), "x" (r)); |
167 | return r; |
168 | #else |
169 | return _mm256_sll_epi32(a.native(), _mm_cvtsi32_si128(count)); |
170 | #endif |
171 | } |
172 | #endif |
173 | |
174 | #if SIMDPP_USE_AVX512F |
175 | static SIMDPP_INL |
176 | uint32<16> i_shift_l(const uint32<16>& a, unsigned count) |
177 | { |
178 | return _mm512_sll_epi32(a.native(), _mm_cvtsi32_si128(count)); |
179 | } |
180 | #endif |
181 | |
182 | // ----------------------------------------------------------------------------- |
183 | |
184 | static SIMDPP_INL |
185 | uint64x2 i_shift_l(const uint64x2& a, unsigned count) |
186 | { |
187 | #if SIMDPP_USE_SSE2 |
188 | return _mm_sll_epi64(a.native(), _mm_cvtsi32_si128(count)); |
189 | #elif SIMDPP_USE_NEON |
190 | int64x2 shift = splat(count); |
191 | return vshlq_u64(a.native(), shift.native()); |
192 | #elif SIMDPP_USE_VSX_207 |
193 | uint64x2 shift = splat(count); |
194 | return vec_sl(a.native(), shift.native()); |
195 | #elif SIMDPP_USE_MSA |
196 | int32x4 shift = splat(count); |
197 | return (v2u64) __msa_sll_d((v2i64) a.native(), (v2i64) shift.native()); |
198 | #elif SIMDPP_USE_NULL || SIMDPP_USE_ALTIVEC |
199 | return detail::null::shift_l(a, count); |
200 | #endif |
201 | } |
202 | |
203 | #if SIMDPP_USE_AVX2 |
204 | static SIMDPP_INL |
205 | uint64x4 i_shift_l(const uint64x4& a, unsigned count) |
206 | { |
207 | #if SIMDPP_WORKAROUND_AVX2_SHIFT_INTRINSICS |
208 | __m256i r = a.native(); |
209 | __m128i x = _mm_cvtsi32_si128(count); |
210 | __asm("vpsllq %1, %2, %0" : "=x" (r) : "x" (x), "x" (r)); |
211 | return r; |
212 | #else |
213 | return _mm256_sll_epi64(a.native(), _mm_cvtsi32_si128(count)); |
214 | #endif |
215 | } |
216 | #endif |
217 | |
218 | #if SIMDPP_USE_AVX512F |
219 | static SIMDPP_INL |
220 | uint64<8> i_shift_l(const uint64<8>& a, unsigned count) |
221 | { |
222 | return _mm512_sll_epi64(a.native(), _mm_cvtsi32_si128(count)); |
223 | } |
224 | #endif |
225 | |
226 | // ----------------------------------------------------------------------------- |
227 | |
228 | template<class V> SIMDPP_INL |
229 | V i_shift_l(const V& a, unsigned count) |
230 | { |
231 | SIMDPP_VEC_ARRAY_IMPL2S(V, i_shift_l, a, count); |
232 | } |
233 | |
234 | // ----------------------------------------------------------------------------- |
235 | |
236 | template<unsigned count, unsigned N> SIMDPP_INL |
237 | uint8<N> sse_shift_l_8(const uint8<N>& a) |
238 | { |
239 | uint8_t mask1 = 0xff >> count; |
240 | uint8<N> mask = make_uint(mask1); |
241 | |
242 | uint16<N/2> a16 = (uint16<N/2>) bit_and(a, mask); |
243 | a16 = shift_l<count>(a16); |
244 | |
245 | return uint8<N>(a16); |
246 | } |
247 | |
248 | template<unsigned count> SIMDPP_INL |
249 | uint8x16 i_shift_l(const uint8x16& a) |
250 | { |
251 | static_assert(count < 8, "Shift out of bounds" ); |
252 | #if SIMDPP_USE_NULL |
253 | return i_shift_l(a, count); |
254 | #elif SIMDPP_USE_SSE2 |
255 | return sse_shift_l_8<count>(a); |
256 | #elif SIMDPP_USE_NEON |
257 | return vshlq_n_u8(a.native(), count); |
258 | #elif SIMDPP_USE_ALTIVEC |
259 | uint8x16 shift = make_uint(count); |
260 | return vec_sl(a.native(), shift.native()); |
261 | #elif SIMDPP_USE_MSA |
262 | return (v16u8) __msa_slli_b((v16i8) a.native(), count); |
263 | #endif |
264 | } |
265 | |
266 | #if SIMDPP_USE_AVX2 |
267 | template<unsigned count> SIMDPP_INL |
268 | uint8<32> i_shift_l(const uint8<32>& a) |
269 | { |
270 | static_assert(count < 8, "Shift out of bounds" ); |
271 | return sse_shift_l_8<count>(a); |
272 | } |
273 | #endif |
274 | |
275 | #if SIMDPP_USE_AVX512BW |
276 | template<unsigned count> SIMDPP_INL |
277 | uint8<64> i_shift_l(const uint8<64>& a) |
278 | { |
279 | static_assert(count < 8, "Shift out of bounds" ); |
280 | return sse_shift_l_8<count>(a); |
281 | } |
282 | #endif |
283 | |
284 | // ----------------------------------------------------------------------------- |
285 | |
286 | template<unsigned count> SIMDPP_INL |
287 | uint16x8 i_shift_l(const uint16x8& a) |
288 | { |
289 | static_assert(count < 16, "Shift out of bounds" ); |
290 | #if SIMDPP_USE_NULL |
291 | return i_shift_l(a, count); |
292 | #elif SIMDPP_USE_SSE2 |
293 | return _mm_slli_epi16(a.native(), count); |
294 | #elif SIMDPP_USE_NEON |
295 | return vshlq_n_u16(a.native(), count); |
296 | #elif SIMDPP_USE_ALTIVEC |
297 | uint16x8 shift = make_uint(count); |
298 | return vec_sl(a.native(), shift.native()); |
299 | #elif SIMDPP_USE_MSA |
300 | return (v8u16) __msa_slli_h((v8i16) a.native(), count); |
301 | #endif |
302 | } |
303 | |
304 | #if SIMDPP_USE_AVX2 |
305 | template<unsigned count> SIMDPP_INL |
306 | uint16x16 i_shift_l(const uint16x16& a) |
307 | { |
308 | static_assert(count < 16, "Shift out of bounds" ); |
309 | return _mm256_slli_epi16(a.native(), count); |
310 | } |
311 | #endif |
312 | |
313 | #if SIMDPP_USE_AVX512BW |
314 | template<unsigned count> SIMDPP_INL |
315 | uint16<32> i_shift_l(const uint16<32>& a) |
316 | { |
317 | static_assert(count < 16, "Shift out of bounds" ); |
318 | return _mm512_slli_epi16(a.native(), count); |
319 | } |
320 | #endif |
321 | |
322 | // ----------------------------------------------------------------------------- |
323 | |
324 | template<unsigned count> SIMDPP_INL |
325 | uint32x4 i_shift_l(const uint32x4& a) |
326 | { |
327 | static_assert(count < 32, "Shift out of bounds" ); |
328 | #if SIMDPP_USE_NULL |
329 | return i_shift_l(a, count); |
330 | #elif SIMDPP_USE_SSE2 |
331 | return _mm_slli_epi32(a.native(), count); |
332 | #elif SIMDPP_USE_NEON |
333 | return vshlq_n_u32(a.native(), count); |
334 | #elif SIMDPP_USE_ALTIVEC |
335 | uint32x4 shift = make_uint(count); |
336 | return vec_sl(a.native(), shift.native()); |
337 | #elif SIMDPP_USE_MSA |
338 | return (v4u32) __msa_slli_w((v4i32) a.native(), count); |
339 | #endif |
340 | } |
341 | |
342 | #if SIMDPP_USE_AVX2 |
343 | template<unsigned count> SIMDPP_INL |
344 | uint32x8 i_shift_l(const uint32x8& a) |
345 | { |
346 | static_assert(count < 32, "Shift out of bounds" ); |
347 | return _mm256_slli_epi32(a.native(), count); |
348 | } |
349 | #endif |
350 | |
351 | #if SIMDPP_USE_AVX512F |
352 | template<unsigned count> SIMDPP_INL |
353 | uint32<16> i_shift_l(const uint32<16>& a) |
354 | { |
355 | static_assert(count < 32, "Shift out of bounds" ); |
356 | return _mm512_slli_epi32(a.native(), count); |
357 | } |
358 | #endif |
359 | |
360 | // ----------------------------------------------------------------------------- |
361 | |
362 | template<unsigned count> SIMDPP_INL |
363 | uint64x2 i_shift_l(const uint64x2& a) |
364 | { |
365 | static_assert(count < 64, "Shift out of bounds" ); |
366 | #if SIMDPP_USE_NULL || SIMDPP_USE_ALTIVEC |
367 | return i_shift_l(a, count); |
368 | #elif SIMDPP_USE_SSE2 |
369 | return _mm_slli_epi64(a.native(), count); |
370 | #elif SIMDPP_USE_NEON |
371 | return vshlq_n_u64(a.native(), count); |
372 | #elif SIMDPP_USE_MSA |
373 | return (v2u64) __msa_slli_d((v2i64) a.native(), count); |
374 | #else |
375 | return SIMDPP_NOT_IMPLEMENTED1(a); |
376 | #endif |
377 | } |
378 | |
379 | #if SIMDPP_USE_AVX2 |
380 | template<unsigned count> SIMDPP_INL |
381 | uint64x4 i_shift_l(const uint64x4& a) |
382 | { |
383 | static_assert(count < 64, "Shift out of bounds" ); |
384 | return _mm256_slli_epi64(a.native(), count); |
385 | } |
386 | #endif |
387 | |
388 | #if SIMDPP_USE_AVX512F |
389 | template<unsigned count> SIMDPP_INL |
390 | uint64<8> i_shift_l(const uint64<8>& a) |
391 | { |
392 | static_assert(count < 64, "Shift out of bounds" ); |
393 | return _mm512_slli_epi64(a.native(), count); |
394 | } |
395 | #endif |
396 | |
397 | // ----------------------------------------------------------------------------- |
398 | |
399 | template<unsigned count, class V> SIMDPP_INL |
400 | V i_shift_l(const V& a) |
401 | { |
402 | SIMDPP_VEC_ARRAY_IMPL1(V, i_shift_l<count>, a); |
403 | } |
404 | |
405 | template<bool no_shift> |
406 | struct i_shift_l_wrapper { |
407 | template<unsigned count, class V> |
408 | static SIMDPP_INL V run(const V& arg) { return i_shift_l<count>(arg); } |
409 | }; |
410 | template<> |
411 | struct i_shift_l_wrapper<true> { |
412 | template<unsigned count, class V> |
413 | static SIMDPP_INL V run(const V& arg) { return arg; } |
414 | }; |
415 | |
416 | } // namespace insn |
417 | } // namespace detail |
418 | } // namespace SIMDPP_ARCH_NAMESPACE |
419 | } // namespace simdpp |
420 | |
421 | #endif |
422 | |
423 | |