/src/Simd/src/Simd/SimdAvx512bwSynetSoftmax.cpp
Line | Count | Source (jump to first uncovered line) |
1 | | /* |
2 | | * Simd Library (http://ermig1979.github.io/Simd). |
3 | | * |
4 | | * Copyright (c) 2011-2024 Yermalayeu Ihar. |
5 | | * |
6 | | * Permission is hereby granted, free of charge, to any person obtaining a copy |
7 | | * of this software and associated documentation files (the "Software"), to deal |
8 | | * in the Software without restriction, including without limitation the rights |
9 | | * to use, copy, modify, merge, publish, distribute, sublicense, and/or sell |
10 | | * copies of the Software, and to permit persons to whom the Software is |
11 | | * furnished to do so, subject to the following conditions: |
12 | | * |
13 | | * The above copyright notice and this permission notice shall be included in |
14 | | * all copies or substantial portions of the Software. |
15 | | * |
16 | | * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR |
17 | | * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, |
18 | | * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE |
19 | | * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER |
20 | | * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, |
21 | | * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE |
22 | | * SOFTWARE. |
23 | | */ |
24 | | #include "Simd/SimdMemory.h" |
25 | | #include "Simd/SimdStore.h" |
26 | | #include "Simd/SimdBase.h" |
27 | | #include "Simd/SimdSse41.h" |
28 | | #include "Simd/SimdAvx2.h" |
29 | | #include "Simd/SimdAvx512bw.h" |
30 | | #include "Simd/SimdSynet.h" |
31 | | #include "Simd/SimdExtract.h" |
32 | | #include "Simd/SimdExp.h" |
33 | | #include "Simd/SimdPow.h" |
34 | | #include "Simd/SimdInterleave.h" |
35 | | #include "Simd/SimdDeinterleave.h" |
36 | | |
37 | | namespace Simd |
38 | | { |
39 | | #if defined(SIMD_AVX512BW_ENABLE) && defined(SIMD_SYNET_ENABLE) |
40 | | namespace Avx512bw |
41 | | { |
42 | | void SynetSoftmaxLayerForward21(const float* src, size_t outer, float* dst) |
43 | 0 | { |
44 | 0 | Exp exp; |
45 | 0 | size_t aligned = Simd::AlignLo(outer, F), tail = outer - aligned; |
46 | 0 | for (size_t o = 0; o < aligned; o += F) |
47 | 0 | { |
48 | 0 | __m512 s0 = _mm512_loadu_ps(src + 0); |
49 | 0 | __m512 s1 = _mm512_loadu_ps(src + F); |
50 | 0 | __m512 ss0 = _mm512_shuffle_ps(s0, s1, 0x88); |
51 | 0 | __m512 ss1 = _mm512_shuffle_ps(s0, s1, 0xDD); |
52 | 0 | __m512 max = _mm512_max_ps(ss0, ss1); |
53 | 0 | __m512 exp0 = exp.Exponent(_mm512_sub_ps(ss0, max)); |
54 | 0 | __m512 exp1 = exp.Exponent(_mm512_sub_ps(ss1, max)); |
55 | 0 | __m512 sum = _mm512_add_ps(exp0, exp1); |
56 | 0 | __m512 d0 = _mm512_div_ps(exp0, sum); |
57 | 0 | __m512 d1 = _mm512_div_ps(exp1, sum); |
58 | 0 | _mm512_storeu_ps(dst + 0, _mm512_unpacklo_ps(d0, d1)); |
59 | 0 | _mm512_storeu_ps(dst + F, _mm512_unpackhi_ps(d0, d1)); |
60 | 0 | src += DF; |
61 | 0 | dst += DF; |
62 | 0 | } |
63 | 0 | if (tail) |
64 | 0 | { |
65 | 0 | __mmask16 mask0 = TailMask16(tail * 2 - 0 * F); |
66 | 0 | __mmask16 mask1 = TailMask16(tail * 2 - 1 * F); |
67 | 0 | __m512 s0 = _mm512_maskz_loadu_ps(mask0, src + 0 * F); |
68 | 0 | __m512 s1 = _mm512_maskz_loadu_ps(mask1, src + 1 * F); |
69 | 0 | __m512 ss0 = _mm512_shuffle_ps(s0, s1, 0x88); |
70 | 0 | __m512 ss1 = _mm512_shuffle_ps(s0, s1, 0xDD); |
71 | 0 | __m512 max = _mm512_max_ps(ss0, ss1); |
72 | 0 | __m512 exp0 = exp.Exponent(_mm512_sub_ps(ss0, max)); |
73 | 0 | __m512 exp1 = exp.Exponent(_mm512_sub_ps(ss1, max)); |
74 | 0 | __m512 sum = _mm512_add_ps(exp0, exp1); |
75 | 0 | __m512 d0 = _mm512_div_ps(exp0, sum); |
76 | 0 | __m512 d1 = _mm512_div_ps(exp1, sum); |
77 | 0 | _mm512_mask_storeu_ps(dst + 0 * F, mask0, _mm512_unpacklo_ps(d0, d1)); |
78 | 0 | _mm512_mask_storeu_ps(dst + 1 * F, mask1, _mm512_unpackhi_ps(d0, d1)); |
79 | 0 | } |
80 | 0 | } |
81 | | |
82 | | SIMD_INLINE void SynetSoftmaxLayerForward31(const Exp& exp, __m512 buf[3]) |
83 | 0 | { |
84 | 0 | __m512 max = _mm512_max_ps(buf[0], _mm512_max_ps(buf[1], buf[2])); |
85 | 0 | buf[0] = exp.Exponent(_mm512_sub_ps(buf[0], max)); |
86 | 0 | buf[1] = exp.Exponent(_mm512_sub_ps(buf[1], max)); |
87 | 0 | buf[2] = exp.Exponent(_mm512_sub_ps(buf[2], max)); |
88 | 0 | __m512 sum = _mm512_add_ps(buf[0], _mm512_add_ps(buf[1], buf[2])); |
89 | 0 | buf[0] = _mm512_div_ps(buf[0], sum); |
90 | 0 | buf[1] = _mm512_div_ps(buf[1], sum); |
91 | 0 | buf[2] = _mm512_div_ps(buf[2], sum); |
92 | 0 | } |
93 | | |
94 | | void SynetSoftmaxLayerForward31(const float* src, size_t outer, float* dst) |
95 | 0 | { |
96 | 0 | static const __m512i idx = _mm512_setr_epi32(0x00, 0x03, 0x06, 0x09, 0x0C, 0x0F, 0x12, 0x15, 0x18, 0x1B, 0x1E, 0x21, 0x24, 0x27, 0x2A, 0x2D); |
97 | 0 | Exp exp; |
98 | 0 | __m512 buf[3]; |
99 | 0 | size_t aligned = Simd::AlignLo(outer, F), tail = outer - aligned; |
100 | 0 | for (size_t o = 0; o < aligned; o += F) |
101 | 0 | { |
102 | 0 | buf[0] = _mm512_i32gather_ps(idx, src + 0, 4); |
103 | 0 | buf[1] = _mm512_i32gather_ps(idx, src + 1, 4); |
104 | 0 | buf[2] = _mm512_i32gather_ps(idx, src + 2, 4); |
105 | 0 | SynetSoftmaxLayerForward31(exp, buf); |
106 | 0 | _mm512_i32scatter_ps(dst + 0, idx, buf[0], 4); |
107 | 0 | _mm512_i32scatter_ps(dst + 1, idx, buf[1], 4); |
108 | 0 | _mm512_i32scatter_ps(dst + 2, idx, buf[2], 4); |
109 | 0 | src += 3 * F; |
110 | 0 | dst += 3 * F; |
111 | 0 | } |
112 | 0 | if (tail) |
113 | 0 | { |
114 | 0 | __mmask16 mask = TailMask16(tail); |
115 | 0 | buf[0] = _mm512_mask_i32gather_ps(_mm512_setzero_ps(), mask, idx, src + 0, 4); |
116 | 0 | buf[1] = _mm512_mask_i32gather_ps(_mm512_setzero_ps(), mask, idx, src + 1, 4); |
117 | 0 | buf[2] = _mm512_mask_i32gather_ps(_mm512_setzero_ps(), mask, idx, src + 2, 4); |
118 | 0 | SynetSoftmaxLayerForward31(exp, buf); |
119 | 0 | _mm512_mask_i32scatter_ps(dst + 0, mask, idx, buf[0], 4); |
120 | 0 | _mm512_mask_i32scatter_ps(dst + 1, mask, idx, buf[1], 4); |
121 | 0 | _mm512_mask_i32scatter_ps(dst + 2, mask, idx, buf[2], 4); |
122 | 0 | } |
123 | 0 | } |
124 | | |
125 | | SIMD_INLINE void LoadTansp16x16(const float* src, size_t srcStride, size_t cols, float* dst, __m512& max) |
126 | 0 | { |
127 | 0 | __m512 a0, a1, a2, a3, a4, a5, a6, a7, a8, a9, aA, aB, aC, aD, aE, aF, b0, b1, b2, b3, b4, b5, b6, b7, b8, b9, bA, bB, bC, bD, bE, bF; |
128 | |
|
129 | 0 | __mmask16 srcMask = __mmask16(-1) >> (16 - cols); |
130 | 0 | __m512 def = _mm512_set1_ps(-FLT_MAX); |
131 | 0 | a0 = _mm512_mask_loadu_ps(def, srcMask, src + 0x0 * srcStride); |
132 | 0 | a1 = _mm512_mask_loadu_ps(def, srcMask, src + 0x1 * srcStride); |
133 | 0 | a2 = _mm512_mask_loadu_ps(def, srcMask, src + 0x2 * srcStride); |
134 | 0 | a3 = _mm512_mask_loadu_ps(def, srcMask, src + 0x3 * srcStride); |
135 | 0 | a4 = _mm512_mask_loadu_ps(def, srcMask, src + 0x4 * srcStride); |
136 | 0 | a5 = _mm512_mask_loadu_ps(def, srcMask, src + 0x5 * srcStride); |
137 | 0 | a6 = _mm512_mask_loadu_ps(def, srcMask, src + 0x6 * srcStride); |
138 | 0 | a7 = _mm512_mask_loadu_ps(def, srcMask, src + 0x7 * srcStride); |
139 | 0 | a8 = _mm512_mask_loadu_ps(def, srcMask, src + 0x8 * srcStride); |
140 | 0 | a9 = _mm512_mask_loadu_ps(def, srcMask, src + 0x9 * srcStride); |
141 | 0 | aA = _mm512_mask_loadu_ps(def, srcMask, src + 0xA * srcStride); |
142 | 0 | aB = _mm512_mask_loadu_ps(def, srcMask, src + 0xB * srcStride); |
143 | 0 | aC = _mm512_mask_loadu_ps(def, srcMask, src + 0xC * srcStride); |
144 | 0 | aD = _mm512_mask_loadu_ps(def, srcMask, src + 0xD * srcStride); |
145 | 0 | aE = _mm512_mask_loadu_ps(def, srcMask, src + 0xE * srcStride); |
146 | 0 | aF = _mm512_mask_loadu_ps(def, srcMask, src + 0xF * srcStride); |
147 | |
|
148 | 0 | b0 = _mm512_unpacklo_ps(a0, a2); |
149 | 0 | b1 = _mm512_unpacklo_ps(a1, a3); |
150 | 0 | b2 = _mm512_unpackhi_ps(a0, a2); |
151 | 0 | b3 = _mm512_unpackhi_ps(a1, a3); |
152 | 0 | b4 = _mm512_unpacklo_ps(a4, a6); |
153 | 0 | b5 = _mm512_unpacklo_ps(a5, a7); |
154 | 0 | b6 = _mm512_unpackhi_ps(a4, a6); |
155 | 0 | b7 = _mm512_unpackhi_ps(a5, a7); |
156 | 0 | b8 = _mm512_unpacklo_ps(a8, aA); |
157 | 0 | b9 = _mm512_unpacklo_ps(a9, aB); |
158 | 0 | bA = _mm512_unpackhi_ps(a8, aA); |
159 | 0 | bB = _mm512_unpackhi_ps(a9, aB); |
160 | 0 | bC = _mm512_unpacklo_ps(aC, aE); |
161 | 0 | bD = _mm512_unpacklo_ps(aD, aF); |
162 | 0 | bE = _mm512_unpackhi_ps(aC, aE); |
163 | 0 | bF = _mm512_unpackhi_ps(aD, aF); |
164 | |
|
165 | 0 | a0 = _mm512_unpacklo_ps(b0, b1); |
166 | 0 | a1 = _mm512_unpackhi_ps(b0, b1); |
167 | 0 | a2 = _mm512_unpacklo_ps(b2, b3); |
168 | 0 | a3 = _mm512_unpackhi_ps(b2, b3); |
169 | 0 | a4 = _mm512_unpacklo_ps(b4, b5); |
170 | 0 | a5 = _mm512_unpackhi_ps(b4, b5); |
171 | 0 | a6 = _mm512_unpacklo_ps(b6, b7); |
172 | 0 | a7 = _mm512_unpackhi_ps(b6, b7); |
173 | 0 | a8 = _mm512_unpacklo_ps(b8, b9); |
174 | 0 | a9 = _mm512_unpackhi_ps(b8, b9); |
175 | 0 | aA = _mm512_unpacklo_ps(bA, bB); |
176 | 0 | aB = _mm512_unpackhi_ps(bA, bB); |
177 | 0 | aC = _mm512_unpacklo_ps(bC, bD); |
178 | 0 | aD = _mm512_unpackhi_ps(bC, bD); |
179 | 0 | aE = _mm512_unpacklo_ps(bE, bF); |
180 | 0 | aF = _mm512_unpackhi_ps(bE, bF); |
181 | |
|
182 | 0 | b0 = _mm512_shuffle_f32x4(a0, a4, 0x44); |
183 | 0 | b1 = _mm512_shuffle_f32x4(a1, a5, 0x44); |
184 | 0 | b2 = _mm512_shuffle_f32x4(a2, a6, 0x44); |
185 | 0 | b3 = _mm512_shuffle_f32x4(a3, a7, 0x44); |
186 | 0 | b4 = _mm512_shuffle_f32x4(a0, a4, 0xEE); |
187 | 0 | b5 = _mm512_shuffle_f32x4(a1, a5, 0xEE); |
188 | 0 | b6 = _mm512_shuffle_f32x4(a2, a6, 0xEE); |
189 | 0 | b7 = _mm512_shuffle_f32x4(a3, a7, 0xEE); |
190 | 0 | b8 = _mm512_shuffle_f32x4(a8, aC, 0x44); |
191 | 0 | b9 = _mm512_shuffle_f32x4(a9, aD, 0x44); |
192 | 0 | bA = _mm512_shuffle_f32x4(aA, aE, 0x44); |
193 | 0 | bB = _mm512_shuffle_f32x4(aB, aF, 0x44); |
194 | 0 | bC = _mm512_shuffle_f32x4(a8, aC, 0xEE); |
195 | 0 | bD = _mm512_shuffle_f32x4(a9, aD, 0xEE); |
196 | 0 | bE = _mm512_shuffle_f32x4(aA, aE, 0xEE); |
197 | 0 | bF = _mm512_shuffle_f32x4(aB, aF, 0xEE); |
198 | |
|
199 | 0 | a0 = _mm512_shuffle_f32x4(b0, b8, 0x88); |
200 | 0 | a1 = _mm512_shuffle_f32x4(b1, b9, 0x88); |
201 | 0 | a2 = _mm512_shuffle_f32x4(b2, bA, 0x88); |
202 | 0 | a3 = _mm512_shuffle_f32x4(b3, bB, 0x88); |
203 | 0 | a4 = _mm512_shuffle_f32x4(b0, b8, 0xDD); |
204 | 0 | a5 = _mm512_shuffle_f32x4(b1, b9, 0xDD); |
205 | 0 | a6 = _mm512_shuffle_f32x4(b2, bA, 0xDD); |
206 | 0 | a7 = _mm512_shuffle_f32x4(b3, bB, 0xDD); |
207 | 0 | a8 = _mm512_shuffle_f32x4(b4, bC, 0x88); |
208 | 0 | a9 = _mm512_shuffle_f32x4(b5, bD, 0x88); |
209 | 0 | aA = _mm512_shuffle_f32x4(b6, bE, 0x88); |
210 | 0 | aB = _mm512_shuffle_f32x4(b7, bF, 0x88); |
211 | 0 | aC = _mm512_shuffle_f32x4(b4, bC, 0xDD); |
212 | 0 | aD = _mm512_shuffle_f32x4(b5, bD, 0xDD); |
213 | 0 | aE = _mm512_shuffle_f32x4(b6, bE, 0xDD); |
214 | 0 | aF = _mm512_shuffle_f32x4(b7, bF, 0xDD); |
215 | |
|
216 | 0 | max = _mm512_max_ps(max, a0); |
217 | 0 | max = _mm512_max_ps(max, a1); |
218 | 0 | max = _mm512_max_ps(max, a2); |
219 | 0 | max = _mm512_max_ps(max, a3); |
220 | 0 | max = _mm512_max_ps(max, a4); |
221 | 0 | max = _mm512_max_ps(max, a5); |
222 | 0 | max = _mm512_max_ps(max, a6); |
223 | 0 | max = _mm512_max_ps(max, a7); |
224 | 0 | max = _mm512_max_ps(max, a8); |
225 | 0 | max = _mm512_max_ps(max, a9); |
226 | 0 | max = _mm512_max_ps(max, aA); |
227 | 0 | max = _mm512_max_ps(max, aB); |
228 | 0 | max = _mm512_max_ps(max, aC); |
229 | 0 | max = _mm512_max_ps(max, aD); |
230 | 0 | max = _mm512_max_ps(max, aE); |
231 | 0 | max = _mm512_max_ps(max, aF); |
232 | |
|
233 | 0 | _mm512_storeu_ps(dst + 0x0 * F, a0); |
234 | 0 | _mm512_storeu_ps(dst + 0x1 * F, a1); |
235 | 0 | _mm512_storeu_ps(dst + 0x2 * F, a2); |
236 | 0 | _mm512_storeu_ps(dst + 0x3 * F, a3); |
237 | 0 | _mm512_storeu_ps(dst + 0x4 * F, a4); |
238 | 0 | _mm512_storeu_ps(dst + 0x5 * F, a5); |
239 | 0 | _mm512_storeu_ps(dst + 0x6 * F, a6); |
240 | 0 | _mm512_storeu_ps(dst + 0x7 * F, a7); |
241 | 0 | _mm512_storeu_ps(dst + 0x8 * F, a8); |
242 | 0 | _mm512_storeu_ps(dst + 0x9 * F, a9); |
243 | 0 | _mm512_storeu_ps(dst + 0xA * F, aA); |
244 | 0 | _mm512_storeu_ps(dst + 0xB * F, aB); |
245 | 0 | _mm512_storeu_ps(dst + 0xC * F, aC); |
246 | 0 | _mm512_storeu_ps(dst + 0xD * F, aD); |
247 | 0 | _mm512_storeu_ps(dst + 0xE * F, aE); |
248 | 0 | _mm512_storeu_ps(dst + 0xF * F, aF); |
249 | 0 | } |
250 | | |
251 | | SIMD_INLINE void LoadTansp16x16(const float* src, size_t srcStride, size_t cols, size_t rows, float* dst, __m512& max) |
252 | 0 | { |
253 | 0 | __m512 a[16], b[16]; |
254 | |
|
255 | 0 | __mmask16 srcMask = __mmask16(-1) >> (16 - cols); |
256 | 0 | __m512 def = _mm512_set1_ps(-FLT_MAX); |
257 | 0 | for(size_t r = 0; r < rows; ++r) |
258 | 0 | a[r] = _mm512_mask_loadu_ps(def, srcMask, src + r * srcStride); |
259 | |
|
260 | 0 | for (size_t r = 0; r < rows; r += 4) |
261 | 0 | { |
262 | 0 | b[r + 0] = _mm512_unpacklo_ps(a[r + 0], a[r + 2]); |
263 | 0 | b[r + 1] = _mm512_unpacklo_ps(a[r + 1], a[r + 3]); |
264 | 0 | b[r + 2] = _mm512_unpackhi_ps(a[r + 0], a[r + 2]); |
265 | 0 | b[r + 3] = _mm512_unpackhi_ps(a[r + 1], a[r + 3]); |
266 | 0 | } |
267 | |
|
268 | 0 | for (size_t r = 0; r < rows; r += 4) |
269 | 0 | { |
270 | 0 | a[r + 0] = _mm512_unpacklo_ps(b[r + 0], b[r + 1]); |
271 | 0 | a[r + 1] = _mm512_unpackhi_ps(b[r + 0], b[r + 1]); |
272 | 0 | a[r + 2] = _mm512_unpacklo_ps(b[r + 2], b[r + 3]); |
273 | 0 | a[r + 3] = _mm512_unpackhi_ps(b[r + 2], b[r + 3]); |
274 | 0 | } |
275 | |
|
276 | 0 | for (size_t i = 0; i < 4; i += 1) |
277 | 0 | { |
278 | 0 | b[0x0 + i] = _mm512_shuffle_f32x4(a[0x0 + i], a[0x4 + i], 0x44); |
279 | 0 | b[0x4 + i] = _mm512_shuffle_f32x4(a[0x0 + i], a[0x4 + i], 0xEE); |
280 | 0 | b[0x8 + i] = _mm512_shuffle_f32x4(a[0x8 + i], a[0xC + i], 0x44); |
281 | 0 | b[0xC + i] = _mm512_shuffle_f32x4(a[0x8 + i], a[0xC + i], 0xEE); |
282 | 0 | } |
283 | |
|
284 | 0 | for (size_t i = 0; i < 4; i += 1) |
285 | 0 | { |
286 | 0 | a[0x0 + i] = _mm512_shuffle_f32x4(b[0x0 + i], b[0x8 + i], 0x88); |
287 | 0 | a[0x4 + i] = _mm512_shuffle_f32x4(b[0x0 + i], b[0x8 + i], 0xDD); |
288 | 0 | a[0x8 + i] = _mm512_shuffle_f32x4(b[0x4 + i], b[0xC + i], 0x88); |
289 | 0 | a[0xC + i] = _mm512_shuffle_f32x4(b[0x4 + i], b[0xC + i], 0xDD); |
290 | 0 | } |
291 | |
|
292 | 0 | for (size_t c = 0; c < cols; ++c) |
293 | 0 | { |
294 | 0 | max = _mm512_max_ps(max, a[c]); |
295 | 0 | _mm512_storeu_ps(dst + c * F, a[c]); |
296 | 0 | } |
297 | 0 | } |
298 | | |
299 | | SIMD_INLINE void StoreTansp16x16(const float* src, size_t cols, __m512 k, float* dst, size_t dstStride) |
300 | 0 | { |
301 | 0 | __m512 a0, a1, a2, a3, a4, a5, a6, a7, a8, a9, aA, aB, aC, aD, aE, aF, b0, b1, b2, b3, b4, b5, b6, b7, b8, b9, bA, bB, bC, bD, bE, bF; |
302 | |
|
303 | 0 | a0 = _mm512_mul_ps(k, _mm512_loadu_ps(src + 0x0 * F)); |
304 | 0 | a1 = _mm512_mul_ps(k, _mm512_loadu_ps(src + 0x1 * F)); |
305 | 0 | a2 = _mm512_mul_ps(k, _mm512_loadu_ps(src + 0x2 * F)); |
306 | 0 | a3 = _mm512_mul_ps(k, _mm512_loadu_ps(src + 0x3 * F)); |
307 | 0 | a4 = _mm512_mul_ps(k, _mm512_loadu_ps(src + 0x4 * F)); |
308 | 0 | a5 = _mm512_mul_ps(k, _mm512_loadu_ps(src + 0x5 * F)); |
309 | 0 | a6 = _mm512_mul_ps(k, _mm512_loadu_ps(src + 0x6 * F)); |
310 | 0 | a7 = _mm512_mul_ps(k, _mm512_loadu_ps(src + 0x7 * F)); |
311 | 0 | a8 = _mm512_mul_ps(k, _mm512_loadu_ps(src + 0x8 * F)); |
312 | 0 | a9 = _mm512_mul_ps(k, _mm512_loadu_ps(src + 0x9 * F)); |
313 | 0 | aA = _mm512_mul_ps(k, _mm512_loadu_ps(src + 0xA * F)); |
314 | 0 | aB = _mm512_mul_ps(k, _mm512_loadu_ps(src + 0xB * F)); |
315 | 0 | aC = _mm512_mul_ps(k, _mm512_loadu_ps(src + 0xC * F)); |
316 | 0 | aD = _mm512_mul_ps(k, _mm512_loadu_ps(src + 0xD * F)); |
317 | 0 | aE = _mm512_mul_ps(k, _mm512_loadu_ps(src + 0xE * F)); |
318 | 0 | aF = _mm512_mul_ps(k, _mm512_loadu_ps(src + 0xF * F)); |
319 | |
|
320 | 0 | b0 = _mm512_unpacklo_ps(a0, a2); |
321 | 0 | b1 = _mm512_unpacklo_ps(a1, a3); |
322 | 0 | b2 = _mm512_unpackhi_ps(a0, a2); |
323 | 0 | b3 = _mm512_unpackhi_ps(a1, a3); |
324 | 0 | b4 = _mm512_unpacklo_ps(a4, a6); |
325 | 0 | b5 = _mm512_unpacklo_ps(a5, a7); |
326 | 0 | b6 = _mm512_unpackhi_ps(a4, a6); |
327 | 0 | b7 = _mm512_unpackhi_ps(a5, a7); |
328 | 0 | b8 = _mm512_unpacklo_ps(a8, aA); |
329 | 0 | b9 = _mm512_unpacklo_ps(a9, aB); |
330 | 0 | bA = _mm512_unpackhi_ps(a8, aA); |
331 | 0 | bB = _mm512_unpackhi_ps(a9, aB); |
332 | 0 | bC = _mm512_unpacklo_ps(aC, aE); |
333 | 0 | bD = _mm512_unpacklo_ps(aD, aF); |
334 | 0 | bE = _mm512_unpackhi_ps(aC, aE); |
335 | 0 | bF = _mm512_unpackhi_ps(aD, aF); |
336 | |
|
337 | 0 | a0 = _mm512_unpacklo_ps(b0, b1); |
338 | 0 | a1 = _mm512_unpackhi_ps(b0, b1); |
339 | 0 | a2 = _mm512_unpacklo_ps(b2, b3); |
340 | 0 | a3 = _mm512_unpackhi_ps(b2, b3); |
341 | 0 | a4 = _mm512_unpacklo_ps(b4, b5); |
342 | 0 | a5 = _mm512_unpackhi_ps(b4, b5); |
343 | 0 | a6 = _mm512_unpacklo_ps(b6, b7); |
344 | 0 | a7 = _mm512_unpackhi_ps(b6, b7); |
345 | 0 | a8 = _mm512_unpacklo_ps(b8, b9); |
346 | 0 | a9 = _mm512_unpackhi_ps(b8, b9); |
347 | 0 | aA = _mm512_unpacklo_ps(bA, bB); |
348 | 0 | aB = _mm512_unpackhi_ps(bA, bB); |
349 | 0 | aC = _mm512_unpacklo_ps(bC, bD); |
350 | 0 | aD = _mm512_unpackhi_ps(bC, bD); |
351 | 0 | aE = _mm512_unpacklo_ps(bE, bF); |
352 | 0 | aF = _mm512_unpackhi_ps(bE, bF); |
353 | |
|
354 | 0 | b0 = _mm512_shuffle_f32x4(a0, a4, 0x44); |
355 | 0 | b1 = _mm512_shuffle_f32x4(a1, a5, 0x44); |
356 | 0 | b2 = _mm512_shuffle_f32x4(a2, a6, 0x44); |
357 | 0 | b3 = _mm512_shuffle_f32x4(a3, a7, 0x44); |
358 | 0 | b4 = _mm512_shuffle_f32x4(a0, a4, 0xEE); |
359 | 0 | b5 = _mm512_shuffle_f32x4(a1, a5, 0xEE); |
360 | 0 | b6 = _mm512_shuffle_f32x4(a2, a6, 0xEE); |
361 | 0 | b7 = _mm512_shuffle_f32x4(a3, a7, 0xEE); |
362 | 0 | b8 = _mm512_shuffle_f32x4(a8, aC, 0x44); |
363 | 0 | b9 = _mm512_shuffle_f32x4(a9, aD, 0x44); |
364 | 0 | bA = _mm512_shuffle_f32x4(aA, aE, 0x44); |
365 | 0 | bB = _mm512_shuffle_f32x4(aB, aF, 0x44); |
366 | 0 | bC = _mm512_shuffle_f32x4(a8, aC, 0xEE); |
367 | 0 | bD = _mm512_shuffle_f32x4(a9, aD, 0xEE); |
368 | 0 | bE = _mm512_shuffle_f32x4(aA, aE, 0xEE); |
369 | 0 | bF = _mm512_shuffle_f32x4(aB, aF, 0xEE); |
370 | |
|
371 | 0 | a0 = _mm512_shuffle_f32x4(b0, b8, 0x88); |
372 | 0 | a1 = _mm512_shuffle_f32x4(b1, b9, 0x88); |
373 | 0 | a2 = _mm512_shuffle_f32x4(b2, bA, 0x88); |
374 | 0 | a3 = _mm512_shuffle_f32x4(b3, bB, 0x88); |
375 | 0 | a4 = _mm512_shuffle_f32x4(b0, b8, 0xDD); |
376 | 0 | a5 = _mm512_shuffle_f32x4(b1, b9, 0xDD); |
377 | 0 | a6 = _mm512_shuffle_f32x4(b2, bA, 0xDD); |
378 | 0 | a7 = _mm512_shuffle_f32x4(b3, bB, 0xDD); |
379 | 0 | a8 = _mm512_shuffle_f32x4(b4, bC, 0x88); |
380 | 0 | a9 = _mm512_shuffle_f32x4(b5, bD, 0x88); |
381 | 0 | aA = _mm512_shuffle_f32x4(b6, bE, 0x88); |
382 | 0 | aB = _mm512_shuffle_f32x4(b7, bF, 0x88); |
383 | 0 | aC = _mm512_shuffle_f32x4(b4, bC, 0xDD); |
384 | 0 | aD = _mm512_shuffle_f32x4(b5, bD, 0xDD); |
385 | 0 | aE = _mm512_shuffle_f32x4(b6, bE, 0xDD); |
386 | 0 | aF = _mm512_shuffle_f32x4(b7, bF, 0xDD); |
387 | |
|
388 | 0 | __mmask16 dstMask = __mmask16(-1) >> (16 - cols); |
389 | 0 | _mm512_mask_storeu_ps(dst + 0x0 * dstStride, dstMask, a0); |
390 | 0 | _mm512_mask_storeu_ps(dst + 0x1 * dstStride, dstMask, a1); |
391 | 0 | _mm512_mask_storeu_ps(dst + 0x2 * dstStride, dstMask, a2); |
392 | 0 | _mm512_mask_storeu_ps(dst + 0x3 * dstStride, dstMask, a3); |
393 | 0 | _mm512_mask_storeu_ps(dst + 0x4 * dstStride, dstMask, a4); |
394 | 0 | _mm512_mask_storeu_ps(dst + 0x5 * dstStride, dstMask, a5); |
395 | 0 | _mm512_mask_storeu_ps(dst + 0x6 * dstStride, dstMask, a6); |
396 | 0 | _mm512_mask_storeu_ps(dst + 0x7 * dstStride, dstMask, a7); |
397 | 0 | _mm512_mask_storeu_ps(dst + 0x8 * dstStride, dstMask, a8); |
398 | 0 | _mm512_mask_storeu_ps(dst + 0x9 * dstStride, dstMask, a9); |
399 | 0 | _mm512_mask_storeu_ps(dst + 0xA * dstStride, dstMask, aA); |
400 | 0 | _mm512_mask_storeu_ps(dst + 0xB * dstStride, dstMask, aB); |
401 | 0 | _mm512_mask_storeu_ps(dst + 0xC * dstStride, dstMask, aC); |
402 | 0 | _mm512_mask_storeu_ps(dst + 0xD * dstStride, dstMask, aD); |
403 | 0 | _mm512_mask_storeu_ps(dst + 0xE * dstStride, dstMask, aE); |
404 | 0 | _mm512_mask_storeu_ps(dst + 0xF * dstStride, dstMask, aF); |
405 | 0 | } |
406 | | |
407 | | SIMD_INLINE void StoreTansp16x16(const float* src, size_t cols, size_t rows, __m512 k, float* dst, size_t dstStride) |
408 | 0 | { |
409 | 0 | __m512 a[16], b[16]; |
410 | |
|
411 | 0 | for (size_t c = 0; c < cols; ++c) |
412 | 0 | a[c] = _mm512_mul_ps(k, _mm512_loadu_ps(src + c * F)); |
413 | |
|
414 | 0 | for (size_t i = 0; i < 4; i += 1) |
415 | 0 | { |
416 | 0 | b[0x0 + i] = _mm512_shuffle_f32x4(a[0x0 + i], a[0x4 + i], 0x44); |
417 | 0 | b[0x4 + i] = _mm512_shuffle_f32x4(a[0x0 + i], a[0x4 + i], 0xEE); |
418 | 0 | b[0x8 + i] = _mm512_shuffle_f32x4(a[0x8 + i], a[0xC + i], 0x44); |
419 | 0 | b[0xC + i] = _mm512_shuffle_f32x4(a[0x8 + i], a[0xC + i], 0xEE); |
420 | 0 | } |
421 | |
|
422 | 0 | for (size_t i = 0; i < 4; i += 1) |
423 | 0 | { |
424 | 0 | a[0x0 + i] = _mm512_shuffle_f32x4(b[0x0 + i], b[0x8 + i], 0x88); |
425 | 0 | a[0x4 + i] = _mm512_shuffle_f32x4(b[0x0 + i], b[0x8 + i], 0xDD); |
426 | 0 | a[0x8 + i] = _mm512_shuffle_f32x4(b[0x4 + i], b[0xC + i], 0x88); |
427 | 0 | a[0xC + i] = _mm512_shuffle_f32x4(b[0x4 + i], b[0xC + i], 0xDD); |
428 | 0 | } |
429 | |
|
430 | 0 | for (size_t r = 0; r < rows; r += 4) |
431 | 0 | { |
432 | 0 | b[r + 0] = _mm512_unpacklo_ps(a[r + 0], a[r + 2]); |
433 | 0 | b[r + 1] = _mm512_unpacklo_ps(a[r + 1], a[r + 3]); |
434 | 0 | b[r + 2] = _mm512_unpackhi_ps(a[r + 0], a[r + 2]); |
435 | 0 | b[r + 3] = _mm512_unpackhi_ps(a[r + 1], a[r + 3]); |
436 | 0 | } |
437 | |
|
438 | 0 | for (size_t r = 0; r < rows; r += 4) |
439 | 0 | { |
440 | 0 | a[r + 0] = _mm512_unpacklo_ps(b[r + 0], b[r + 1]); |
441 | 0 | a[r + 1] = _mm512_unpackhi_ps(b[r + 0], b[r + 1]); |
442 | 0 | a[r + 2] = _mm512_unpacklo_ps(b[r + 2], b[r + 3]); |
443 | 0 | a[r + 3] = _mm512_unpackhi_ps(b[r + 2], b[r + 3]); |
444 | 0 | } |
445 | |
|
446 | 0 | __mmask16 dstMask = __mmask16(-1) >> (16 - cols); |
447 | 0 | for (size_t r = 0; r < rows; ++r) |
448 | 0 | _mm512_mask_storeu_ps(dst + r * dstStride, dstMask, a[r]); |
449 | 0 | } |
450 | | |
451 | | void SynetSoftmaxLayerForwardX1(const float* src, size_t outer, size_t count, float* dst) |
452 | 0 | { |
453 | 0 | size_t o = 0, c = 0, outerF = AlignLo(outer, F), countF = AlignLo(count, F); |
454 | 0 | Array32f buf(AlignHi(count, F) * F); |
455 | 0 | Exp exp; |
456 | 0 | for (; o < outerF; o += F) |
457 | 0 | { |
458 | 0 | __m512 _max = _mm512_set1_ps(-FLT_MAX); |
459 | 0 | for (c = 0; c < countF; c += F) |
460 | 0 | LoadTansp16x16(src + c, count, F, buf.data + c * F, _max); |
461 | 0 | if (c < count) |
462 | 0 | LoadTansp16x16(src + c, count, count - c, buf.data + c * F, _max); |
463 | 0 | __m512 _sum = _mm512_setzero_ps(); |
464 | 0 | for (c = 0; c < count; ++c) |
465 | 0 | { |
466 | 0 | __m512 _exp = exp.Exponent(_mm512_sub_ps(_mm512_loadu_ps(buf.data + c * F), _max)); |
467 | 0 | _sum = _mm512_add_ps(_sum, _exp); |
468 | 0 | _mm512_storeu_ps(buf.data + c * F, _exp); |
469 | 0 | } |
470 | 0 | __m512 _k = _mm512_div_ps(_mm512_set1_ps(1.0f), _sum); |
471 | 0 | for (c = 0; c < countF; c += F) |
472 | 0 | StoreTansp16x16(buf.data + c * F, F, _k, dst + c, count); |
473 | 0 | if (c < count) |
474 | 0 | StoreTansp16x16(buf.data + c * F, count - c, _k, dst + c, count); |
475 | 0 | src += count * F; |
476 | 0 | dst += count * F; |
477 | 0 | } |
478 | 0 | if (o < outer) |
479 | 0 | { |
480 | 0 | buf.Clear(); |
481 | 0 | __m512 _max = _mm512_set1_ps(-FLT_MAX); |
482 | 0 | for (c = 0; c < countF; c += F) |
483 | 0 | LoadTansp16x16(src + c, count, F, outer - o, buf.data + c * F, _max); |
484 | 0 | if (c < count) |
485 | 0 | LoadTansp16x16(src + c, count, count - c, outer - o, buf.data + c * F, _max); |
486 | 0 | __m512 _sum = _mm512_setzero_ps(); |
487 | 0 | for (c = 0; c < count; ++c) |
488 | 0 | { |
489 | 0 | __m512 _exp = exp.Exponent(_mm512_sub_ps(_mm512_loadu_ps(buf.data + c * F), _max)); |
490 | 0 | _sum = _mm512_add_ps(_sum, _exp); |
491 | 0 | _mm512_storeu_ps(buf.data + c * F, _exp); |
492 | 0 | } |
493 | 0 | __m512 _k = _mm512_div_ps(_mm512_set1_ps(1.0f), _sum); |
494 | 0 | for (c = 0; c < countF; c += F) |
495 | 0 | StoreTansp16x16(buf.data + c * F, F, outer - o, _k, dst + c, count); |
496 | 0 | if (c < count) |
497 | 0 | StoreTansp16x16(buf.data + c * F, count - c, outer - o, _k, dst + c, count); |
498 | 0 | } |
499 | 0 | } |
500 | | |
501 | | void SynetSoftmaxLayerForward(const float* src, size_t outer, size_t count, size_t inner, float* dst) |
502 | 0 | { |
503 | 0 | if (inner == 1) |
504 | 0 | { |
505 | 0 | if (count == 2) |
506 | 0 | SynetSoftmaxLayerForward21(src, outer, dst); |
507 | 0 | else if (count == 3) |
508 | 0 | SynetSoftmaxLayerForward31(src, outer, dst); |
509 | 0 | else |
510 | 0 | SynetSoftmaxLayerForwardX1(src, outer, count, dst); |
511 | 0 | } |
512 | 0 | else |
513 | 0 | { |
514 | 0 | Exp exp; |
515 | 0 | size_t aligned = Simd::AlignLo(inner, F); |
516 | 0 | __mmask16 tail = TailMask16(inner - aligned); |
517 | 0 | Array32f tmp(inner * 2); |
518 | 0 | const float* s; |
519 | 0 | float* max = tmp.data, * sum = tmp.data + inner, * d; |
520 | 0 | for (size_t o = 0; o < outer; ++o) |
521 | 0 | { |
522 | 0 | memcpy(max, src, inner * sizeof(float)); |
523 | 0 | s = src + inner; |
524 | 0 | for (size_t c = 1; c < count; ++c) |
525 | 0 | { |
526 | 0 | size_t i = 0; |
527 | 0 | for (; i < aligned; i += F) |
528 | 0 | _mm512_storeu_ps(max + i, _mm512_max_ps(_mm512_loadu_ps(s + i), _mm512_loadu_ps(max + i))); |
529 | 0 | if (i < inner) |
530 | 0 | _mm512_mask_storeu_ps(max + i, tail, _mm512_max_ps(_mm512_maskz_loadu_ps(tail, s + i), _mm512_maskz_loadu_ps(tail, max + i))); |
531 | 0 | s += inner; |
532 | 0 | } |
533 | |
|
534 | 0 | s = src; |
535 | 0 | d = dst; |
536 | 0 | memset(sum, 0, inner * sizeof(float)); |
537 | 0 | for (size_t c = 0; c < count; ++c) |
538 | 0 | { |
539 | 0 | size_t i = 0; |
540 | 0 | for (; i < aligned; i += F) |
541 | 0 | { |
542 | 0 | __m512 _d = exp.Exponent(_mm512_sub_ps(_mm512_loadu_ps(s + i), _mm512_loadu_ps(max + i))); |
543 | 0 | _mm512_storeu_ps(d + i, _d); |
544 | 0 | _mm512_storeu_ps(sum + i, _mm512_add_ps(_d, _mm512_loadu_ps(sum + i))); |
545 | 0 | } |
546 | 0 | if (i < inner) |
547 | 0 | { |
548 | 0 | __m512 _d = exp.Exponent(_mm512_sub_ps(_mm512_maskz_loadu_ps(tail, s + i), _mm512_maskz_loadu_ps(tail, max + i))); |
549 | 0 | _mm512_mask_storeu_ps(d + i, tail, _d); |
550 | 0 | _mm512_mask_storeu_ps(sum + i, tail, _mm512_add_ps(_d, _mm512_maskz_loadu_ps(tail, sum + i))); |
551 | 0 | } |
552 | 0 | s += inner; |
553 | 0 | d += inner; |
554 | 0 | } |
555 | |
|
556 | 0 | d = dst; |
557 | 0 | for (size_t c = 0; c < count; ++c) |
558 | 0 | { |
559 | 0 | size_t i = 0; |
560 | 0 | for (; i < aligned; i += F) |
561 | 0 | _mm512_storeu_ps(d + i, _mm512_div_ps(_mm512_loadu_ps(d + i), _mm512_loadu_ps(sum + i))); |
562 | 0 | if (i < inner) |
563 | 0 | _mm512_mask_storeu_ps(d + i, tail, _mm512_div_ps(_mm512_maskz_loadu_ps(tail, d + i), _mm512_maskz_loadu_ps(tail, sum + i))); |
564 | 0 | d += inner; |
565 | 0 | } |
566 | 0 | src += count * inner; |
567 | 0 | dst += count * inner; |
568 | 0 | } |
569 | 0 | } |
570 | 0 | } |
571 | | } |
572 | | #endif |
573 | | } |