16 #ifndef included_vector_avx512_h 17 #define included_vector_avx512_h 20 #include <x86intrin.h> 23 #define foreach_avx512_vec512i \ 24 _(i,8,64,epi8) _(i,16,32,epi16) _(i,32,16,epi32) _(i,64,8,epi64) 25 #define foreach_avx512_vec512u \ 26 _(u,8,64,epi8) _(u,16,32,epi16) _(u,32,16,epi32) _(u,64,8,epi64) 27 #define foreach_avx512_vec512f \ 28 _(f,32,8,ps) _(f,64,4,pd) 32 #define _(t, s, c, i) \ 33 static_always_inline t##s##x##c \ 34 t##s##x##c##_splat (t##s x) \ 35 { return (t##s##x##c) _mm512_set1_##i (x); } \ 37 static_always_inline t##s##x##c \ 38 t##s##x##c##_load_unaligned (void *p) \ 39 { return (t##s##x##c) _mm512_loadu_si512 (p); } \ 41 static_always_inline void \ 42 t##s##x##c##_store_unaligned (t##s##x##c v, void *p) \ 43 { _mm512_storeu_si512 ((__m512i *) p, (__m512i) v); } \ 45 static_always_inline int \ 46 t##s##x##c##_is_all_zero (t##s##x##c v) \ 47 { return (_mm512_test_epi64_mask ((__m512i) v, (__m512i) v) == 0); } \ 49 static_always_inline int \ 50 t##s##x##c##_is_equal (t##s##x##c a, t##s##x##c b) \ 51 { return t##s##x##c##_is_all_zero (a ^ b); } \ 53 static_always_inline int \ 54 t##s##x##c##_is_all_equal (t##s##x##c v, t##s x) \ 55 { return t##s##x##c##_is_equal (v, t##s##x##c##_splat (x)); } \ 57 static_always_inline u##c \ 58 t##s##x##c##_is_zero_mask (t##s##x##c v) \ 59 { return _mm512_test_##i##_mask ((__m512i) v, (__m512i) v); } \ 61 static_always_inline t##s##x##c \ 62 t##s##x##c##_interleave_lo (t##s##x##c a, t##s##x##c b) \ 63 { return (t##s##x##c) _mm512_unpacklo_##i ((__m512i) a, (__m512i) b); } \ 65 static_always_inline t##s##x##c \ 66 t##s##x##c##_interleave_hi (t##s##x##c a, t##s##x##c b) \ 67 { return (t##s##x##c) _mm512_unpackhi_##i ((__m512i) a, (__m512i) b); } \ 77 return (
u32) _mm512_movepi16_mask ((__m512i) v);
84 3, 2, 1, 0, 7, 6, 5, 4, 11, 10, 9, 8, 15, 14, 13, 12,
85 3, 2, 1, 0, 7, 6, 5, 4, 11, 10, 9, 8, 15, 14, 13, 12,
86 3, 2, 1, 0, 7, 6, 5, 4, 11, 10, 9, 8, 15, 14, 13, 12,
87 3, 2, 1, 0, 7, 6, 5, 4, 11, 10, 9, 8, 15, 14, 13, 12
89 return (u32x16) _mm512_shuffle_epi8 ((__m512i) v, (__m512i) swap);
96 1, 0, 3, 2, 5, 4, 7, 6, 9, 8, 11, 10, 13, 12, 15, 14,
97 1, 0, 3, 2, 5, 4, 7, 6, 9, 8, 11, 10, 13, 12, 15, 14,
98 1, 0, 3, 2, 5, 4, 7, 6, 9, 8, 11, 10, 13, 12, 15, 14,
99 1, 0, 3, 2, 5, 4, 7, 6, 9, 8, 11, 10, 13, 12, 15, 14
101 return (u16x32) _mm512_shuffle_epi8 ((__m512i) v, (__m512i) swap);
107 return (u32x8) _mm512_extracti64x4_epi64 ((__m512i) v, 0);
113 return (u32x8) _mm512_extracti64x4_epi64 ((__m512i) v, 1);
126 return (u32x16) _mm512_inserti64x4 ((__m512i) r, (__m256i) v, 0);
132 return (u32x16) _mm512_inserti64x4 ((__m512i) r, (__m256i) v, 1);
138 return (u64x8) _mm512_permutex2var_epi64 ((__m512i) a, (__m512i) mask,
143 #define u32x16_ternary_logic(a, b, c, d) \ 144 (u32x16) _mm512_ternarylogic_epi32 ((__m512i) a, (__m512i) b, (__m512i) c, d) 149 __m512i r[16],
a, b,
c, d, x, y;
152 __m512i pm1 = (__m512i) (u64x8) { 0, 1, 8, 9, 4, 5, 12, 13};
153 __m512i pm2 = (__m512i) (u64x8) { 2, 3, 10, 11, 6, 7, 14, 15};
154 __m512i pm3 = (__m512i) (u64x8) { 0, 1, 2, 3, 8, 9, 10, 11};
155 __m512i pm4 = (__m512i) (u64x8) { 4, 5, 6, 7, 12, 13, 14, 15};
158 r[0] = _mm512_unpacklo_epi32 ((__m512i) m[0], (__m512i) m[1]);
159 r[1] = _mm512_unpacklo_epi32 ((__m512i) m[2], (__m512i) m[3]);
160 r[2] = _mm512_unpacklo_epi32 ((__m512i) m[4], (__m512i) m[5]);
161 r[3] = _mm512_unpacklo_epi32 ((__m512i) m[6], (__m512i) m[7]);
162 r[4] = _mm512_unpacklo_epi32 ((__m512i) m[8], (__m512i) m[9]);
163 r[5] = _mm512_unpacklo_epi32 ((__m512i) m[10], (__m512i) m[11]);
164 r[6] = _mm512_unpacklo_epi32 ((__m512i) m[12], (__m512i) m[13]);
165 r[7] = _mm512_unpacklo_epi32 ((__m512i) m[14], (__m512i) m[15]);
167 r[8] = _mm512_unpackhi_epi32 ((__m512i) m[0], (__m512i) m[1]);
168 r[9] = _mm512_unpackhi_epi32 ((__m512i) m[2], (__m512i) m[3]);
169 r[10] = _mm512_unpackhi_epi32 ((__m512i) m[4], (__m512i) m[5]);
170 r[11] = _mm512_unpackhi_epi32 ((__m512i) m[6], (__m512i) m[7]);
171 r[12] = _mm512_unpackhi_epi32 ((__m512i) m[8], (__m512i) m[9]);
172 r[13] = _mm512_unpackhi_epi32 ((__m512i) m[10], (__m512i) m[11]);
173 r[14] = _mm512_unpackhi_epi32 ((__m512i) m[12], (__m512i) m[13]);
174 r[15] = _mm512_unpackhi_epi32 ((__m512i) m[14], (__m512i) m[15]);
176 a = _mm512_unpacklo_epi64 (r[0], r[1]);
177 b = _mm512_unpacklo_epi64 (r[2], r[3]);
178 c = _mm512_unpacklo_epi64 (r[4], r[5]);
179 d = _mm512_unpacklo_epi64 (r[6], r[7]);
180 x = _mm512_permutex2var_epi64 (a, pm1, b);
181 y = _mm512_permutex2var_epi64 (c, pm1, d);
182 m[0] = (u32x16) _mm512_permutex2var_epi64 (x, pm3, y);
183 m[8] = (u32x16) _mm512_permutex2var_epi64 (x, pm4, y);
184 x = _mm512_permutex2var_epi64 (a, pm2, b);
185 y = _mm512_permutex2var_epi64 (c, pm2, d);
186 m[4] = (u32x16) _mm512_permutex2var_epi64 (x, pm3, y);
187 m[12] = (u32x16) _mm512_permutex2var_epi64 (x, pm4, y);
189 a = _mm512_unpacklo_epi64 (r[8], r[9]);
190 b = _mm512_unpacklo_epi64 (r[10], r[11]);
191 c = _mm512_unpacklo_epi64 (r[12], r[13]);
192 d = _mm512_unpacklo_epi64 (r[14], r[15]);
193 x = _mm512_permutex2var_epi64 (a, pm1, b);
194 y = _mm512_permutex2var_epi64 (c, pm1, d);
195 m[2] = (u32x16) _mm512_permutex2var_epi64 (x, pm3, y);
196 m[10] = (u32x16) _mm512_permutex2var_epi64 (x, pm4, y);
197 x = _mm512_permutex2var_epi64 (a, pm2, b);
198 y = _mm512_permutex2var_epi64 (c, pm2, d);
199 m[6] = (u32x16) _mm512_permutex2var_epi64 (x, pm3, y);
200 m[14] = (u32x16) _mm512_permutex2var_epi64 (x, pm4, y);
202 a = _mm512_unpackhi_epi64 (r[0], r[1]);
203 b = _mm512_unpackhi_epi64 (r[2], r[3]);
204 c = _mm512_unpackhi_epi64 (r[4], r[5]);
205 d = _mm512_unpackhi_epi64 (r[6], r[7]);
206 x = _mm512_permutex2var_epi64 (a, pm1, b);
207 y = _mm512_permutex2var_epi64 (c, pm1, d);
208 m[1] = (u32x16) _mm512_permutex2var_epi64 (x, pm3, y);
209 m[9] = (u32x16) _mm512_permutex2var_epi64 (x, pm4, y);
210 x = _mm512_permutex2var_epi64 (a, pm2, b);
211 y = _mm512_permutex2var_epi64 (c, pm2, d);
212 m[5] = (u32x16) _mm512_permutex2var_epi64 (x, pm3, y);
213 m[13] = (u32x16) _mm512_permutex2var_epi64 (x, pm4, y);
215 a = _mm512_unpackhi_epi64 (r[8], r[9]);
216 b = _mm512_unpackhi_epi64 (r[10], r[11]);
217 c = _mm512_unpackhi_epi64 (r[12], r[13]);
218 d = _mm512_unpackhi_epi64 (r[14], r[15]);
219 x = _mm512_permutex2var_epi64 (a, pm1, b);
220 y = _mm512_permutex2var_epi64 (c, pm1, d);
221 m[3] = (u32x16) _mm512_permutex2var_epi64 (x, pm3, y);
222 m[11] = (u32x16) _mm512_permutex2var_epi64 (x, pm4, y);
223 x = _mm512_permutex2var_epi64 (a, pm2, b);
224 y = _mm512_permutex2var_epi64 (c, pm2, d);
225 m[7] = (u32x16) _mm512_permutex2var_epi64 (x, pm3, y);
226 m[15] = (u32x16) _mm512_permutex2var_epi64 (x, pm4, y);
237 __m512i pm1 = (__m512i) (u64x8) { 0, 1, 8, 9, 4, 5, 12, 13};
238 __m512i pm2 = (__m512i) (u64x8) { 2, 3, 10, 11, 6, 7, 14, 15};
239 __m512i pm3 = (__m512i) (u64x8) { 0, 1, 2, 3, 8, 9, 10, 11};
240 __m512i pm4 = (__m512i) (u64x8) { 4, 5, 6, 7, 12, 13, 14, 15};
243 r[0] = _mm512_unpacklo_epi64 ((__m512i) m[0], (__m512i) m[1]);
244 r[1] = _mm512_unpacklo_epi64 ((__m512i) m[2], (__m512i) m[3]);
245 r[2] = _mm512_unpacklo_epi64 ((__m512i) m[4], (__m512i) m[5]);
246 r[3] = _mm512_unpacklo_epi64 ((__m512i) m[6], (__m512i) m[7]);
247 r[4] = _mm512_unpackhi_epi64 ((__m512i) m[0], (__m512i) m[1]);
248 r[5] = _mm512_unpackhi_epi64 ((__m512i) m[2], (__m512i) m[3]);
249 r[6] = _mm512_unpackhi_epi64 ((__m512i) m[4], (__m512i) m[5]);
250 r[7] = _mm512_unpackhi_epi64 ((__m512i) m[6], (__m512i) m[7]);
252 x = _mm512_permutex2var_epi64 (r[0], pm1, r[1]);
253 y = _mm512_permutex2var_epi64 (r[2], pm1, r[3]);
254 m[0] = (u64x8) _mm512_permutex2var_epi64 (x, pm3, y);
255 m[4] = (u64x8) _mm512_permutex2var_epi64 (x, pm4, y);
256 x = _mm512_permutex2var_epi64 (r[0], pm2, r[1]);
257 y = _mm512_permutex2var_epi64 (r[2], pm2, r[3]);
258 m[2] = (u64x8) _mm512_permutex2var_epi64 (x, pm3, y);
259 m[6] = (u64x8) _mm512_permutex2var_epi64 (x, pm4, y);
261 x = _mm512_permutex2var_epi64 (r[4], pm1, r[5]);
262 y = _mm512_permutex2var_epi64 (r[6], pm1, r[7]);
263 m[1] = (u64x8) _mm512_permutex2var_epi64 (x, pm3, y);
264 m[5] = (u64x8) _mm512_permutex2var_epi64 (x, pm4, y);
265 x = _mm512_permutex2var_epi64 (r[4], pm2, r[5]);
266 y = _mm512_permutex2var_epi64 (r[6], pm2, r[7]);
267 m[3] = (u64x8) _mm512_permutex2var_epi64 (x, pm3, y);
268 m[7] = (u64x8) _mm512_permutex2var_epi64 (x, pm4, y);
static_always_inline u64x8 u64x8_permute(u64x8 a, u64x8 b, u64x8 mask)
static_always_inline void u64x8_transpose(u64x8 m[8])
static_always_inline u32 u32x8_min_scalar(u32x8 v)
static_always_inline void u32x16_transpose(u32x16 m[16])
static_always_inline u32x8 u32x16_extract_hi(u32x16 v)
static_always_inline u32x16 u32x16_insert_lo(u32x16 r, u32x8 v)
#define static_always_inline
#define foreach_avx512_vec512i
static_always_inline u16x32 u16x32_byte_swap(u16x32 v)
static_always_inline u32x16 u32x16_byte_swap(u32x16 v)
static_always_inline u32x8 u32x16_extract_lo(u32x16 v)
static_always_inline u32 u32x16_min_scalar(u32x16 v)
static_always_inline u32x8 u32x8_min(u32x8 a, u32x8 b)
foreach_avx512_vec512i foreach_avx512_vec512u static_always_inline u32 u16x32_msb_mask(u16x32 v)
static_always_inline u32x16 u32x16_insert_hi(u32x16 r, u32x8 v)
#define foreach_avx512_vec512u