1 // Auto-generated file. Do not edit!
2 // Template: src/f32-igemm/sse-shuffle.c.in
3 // Generator: tools/xngen
4 //
5 // Copyright 2019 Google LLC
6 //
7 // This source code is licensed under the BSD-style license found in the
8 // LICENSE file in the root directory of this source tree.
9
10 #include <assert.h>
11
12 #include <xmmintrin.h>
13
14 #include <xnnpack/igemm.h>
15
16
xnn_f32_igemm_minmax_ukernel_4x8s4__sse(size_t mr,size_t nc,size_t kc,size_t ks,const float ** restrict a,const float * restrict w,float * restrict c,size_t cm_stride,size_t cn_stride,size_t a_offset,const float * zero,const union xnn_f32_minmax_params params[restrict XNN_MIN_ELEMENTS (1)])17 void xnn_f32_igemm_minmax_ukernel_4x8s4__sse(
18 size_t mr,
19 size_t nc,
20 size_t kc,
21 size_t ks,
22 const float**restrict a,
23 const float*restrict w,
24 float*restrict c,
25 size_t cm_stride,
26 size_t cn_stride,
27 size_t a_offset,
28 const float* zero,
29 const union xnn_f32_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
30 {
31 assert(mr != 0);
32 assert(mr <= 4);
33 assert(nc != 0);
34 assert(kc != 0);
35 assert(kc % sizeof(float) == 0);
36 assert(ks != 0);
37 assert(ks % (4 * sizeof(void*)) == 0);
38 assert(a_offset % sizeof(float) == 0);
39 assert(a != NULL);
40 assert(w != NULL);
41 assert(c != NULL);
42
43 float* c0 = c;
44 float* c1 = (float*) ((uintptr_t) c0 + cm_stride);
45 if XNN_UNPREDICTABLE(mr < 2) {
46 c1 = c0;
47 }
48 float* c2 = (float*) ((uintptr_t) c1 + cm_stride);
49 if XNN_UNPREDICTABLE(mr <= 2) {
50 c2 = c1;
51 }
52 float* c3 = (float*) ((uintptr_t) c2 + cm_stride);
53 if XNN_UNPREDICTABLE(mr != 4) {
54 c3 = c2;
55 }
56
57 do {
58 __m128 vacc0x0123 = _mm_load_ps(w);
59 __m128 vacc0x4567 = _mm_load_ps(w + 4);
60 __m128 vacc1x0123 = vacc0x0123;
61 __m128 vacc1x4567 = vacc0x4567;
62 __m128 vacc2x0123 = vacc0x0123;
63 __m128 vacc2x4567 = vacc0x4567;
64 __m128 vacc3x0123 = vacc0x0123;
65 __m128 vacc3x4567 = vacc0x4567;
66 w += 8;
67
68 size_t p = ks;
69 do {
70 const float* restrict a0 = a[0];
71 assert(a0 != NULL);
72 if XNN_UNPREDICTABLE(a0 != zero) {
73 a0 = (const float*) ((uintptr_t) a0 + a_offset);
74 }
75 const float* restrict a1 = a[1];
76 assert(a1 != NULL);
77 if XNN_UNPREDICTABLE(a1 != zero) {
78 a1 = (const float*) ((uintptr_t) a1 + a_offset);
79 }
80 const float* restrict a2 = a[2];
81 assert(a2 != NULL);
82 if XNN_UNPREDICTABLE(a2 != zero) {
83 a2 = (const float*) ((uintptr_t) a2 + a_offset);
84 }
85 const float* restrict a3 = a[3];
86 assert(a3 != NULL);
87 if XNN_UNPREDICTABLE(a3 != zero) {
88 a3 = (const float*) ((uintptr_t) a3 + a_offset);
89 }
90 a += 4;
91
92 size_t k = kc;
93 while (k >= 4 * sizeof(float)) {
94 __m128 va0 = _mm_loadu_ps(a0);
95 a0 += 4;
96 __m128 va1 = _mm_loadu_ps(a1);
97 a1 += 4;
98 __m128 va2 = _mm_loadu_ps(a2);
99 a2 += 4;
100 __m128 va3 = _mm_loadu_ps(a3);
101 a3 += 4;
102
103
104 const __m128 vb0123c0 = _mm_load_ps(w + 0);
105 const __m128 vb4567c0 = _mm_load_ps(w + 4);
106
107 vacc0x0123 = _mm_add_ps(vacc0x0123, _mm_mul_ps(va0, vb0123c0));
108 vacc1x0123 = _mm_add_ps(vacc1x0123, _mm_mul_ps(va1, vb0123c0));
109 vacc2x0123 = _mm_add_ps(vacc2x0123, _mm_mul_ps(va2, vb0123c0));
110 vacc3x0123 = _mm_add_ps(vacc3x0123, _mm_mul_ps(va3, vb0123c0));
111 vacc0x4567 = _mm_add_ps(vacc0x4567, _mm_mul_ps(va0, vb4567c0));
112 vacc1x4567 = _mm_add_ps(vacc1x4567, _mm_mul_ps(va1, vb4567c0));
113 vacc2x4567 = _mm_add_ps(vacc2x4567, _mm_mul_ps(va2, vb4567c0));
114 vacc3x4567 = _mm_add_ps(vacc3x4567, _mm_mul_ps(va3, vb4567c0));
115
116 va0 = _mm_shuffle_ps(va0, va0, _MM_SHUFFLE(0, 3, 2, 1));
117 va1 = _mm_shuffle_ps(va1, va1, _MM_SHUFFLE(0, 3, 2, 1));
118 va2 = _mm_shuffle_ps(va2, va2, _MM_SHUFFLE(0, 3, 2, 1));
119 va3 = _mm_shuffle_ps(va3, va3, _MM_SHUFFLE(0, 3, 2, 1));
120
121 const __m128 vb0123c1 = _mm_load_ps(w + 8);
122 const __m128 vb4567c1 = _mm_load_ps(w + 12);
123
124 vacc0x0123 = _mm_add_ps(vacc0x0123, _mm_mul_ps(va0, vb0123c1));
125 vacc1x0123 = _mm_add_ps(vacc1x0123, _mm_mul_ps(va1, vb0123c1));
126 vacc2x0123 = _mm_add_ps(vacc2x0123, _mm_mul_ps(va2, vb0123c1));
127 vacc3x0123 = _mm_add_ps(vacc3x0123, _mm_mul_ps(va3, vb0123c1));
128 vacc0x4567 = _mm_add_ps(vacc0x4567, _mm_mul_ps(va0, vb4567c1));
129 vacc1x4567 = _mm_add_ps(vacc1x4567, _mm_mul_ps(va1, vb4567c1));
130 vacc2x4567 = _mm_add_ps(vacc2x4567, _mm_mul_ps(va2, vb4567c1));
131 vacc3x4567 = _mm_add_ps(vacc3x4567, _mm_mul_ps(va3, vb4567c1));
132
133 va0 = _mm_shuffle_ps(va0, va0, _MM_SHUFFLE(0, 3, 2, 1));
134 va1 = _mm_shuffle_ps(va1, va1, _MM_SHUFFLE(0, 3, 2, 1));
135 va2 = _mm_shuffle_ps(va2, va2, _MM_SHUFFLE(0, 3, 2, 1));
136 va3 = _mm_shuffle_ps(va3, va3, _MM_SHUFFLE(0, 3, 2, 1));
137
138 const __m128 vb0123c2 = _mm_load_ps(w + 16);
139 const __m128 vb4567c2 = _mm_load_ps(w + 20);
140
141 vacc0x0123 = _mm_add_ps(vacc0x0123, _mm_mul_ps(va0, vb0123c2));
142 vacc1x0123 = _mm_add_ps(vacc1x0123, _mm_mul_ps(va1, vb0123c2));
143 vacc2x0123 = _mm_add_ps(vacc2x0123, _mm_mul_ps(va2, vb0123c2));
144 vacc3x0123 = _mm_add_ps(vacc3x0123, _mm_mul_ps(va3, vb0123c2));
145 vacc0x4567 = _mm_add_ps(vacc0x4567, _mm_mul_ps(va0, vb4567c2));
146 vacc1x4567 = _mm_add_ps(vacc1x4567, _mm_mul_ps(va1, vb4567c2));
147 vacc2x4567 = _mm_add_ps(vacc2x4567, _mm_mul_ps(va2, vb4567c2));
148 vacc3x4567 = _mm_add_ps(vacc3x4567, _mm_mul_ps(va3, vb4567c2));
149
150 va0 = _mm_shuffle_ps(va0, va0, _MM_SHUFFLE(0, 3, 2, 1));
151 va1 = _mm_shuffle_ps(va1, va1, _MM_SHUFFLE(0, 3, 2, 1));
152 va2 = _mm_shuffle_ps(va2, va2, _MM_SHUFFLE(0, 3, 2, 1));
153 va3 = _mm_shuffle_ps(va3, va3, _MM_SHUFFLE(0, 3, 2, 1));
154
155 const __m128 vb0123c3 = _mm_load_ps(w + 24);
156 const __m128 vb4567c3 = _mm_load_ps(w + 28);
157
158 vacc0x0123 = _mm_add_ps(vacc0x0123, _mm_mul_ps(va0, vb0123c3));
159 vacc1x0123 = _mm_add_ps(vacc1x0123, _mm_mul_ps(va1, vb0123c3));
160 vacc2x0123 = _mm_add_ps(vacc2x0123, _mm_mul_ps(va2, vb0123c3));
161 vacc3x0123 = _mm_add_ps(vacc3x0123, _mm_mul_ps(va3, vb0123c3));
162 vacc0x4567 = _mm_add_ps(vacc0x4567, _mm_mul_ps(va0, vb4567c3));
163 vacc1x4567 = _mm_add_ps(vacc1x4567, _mm_mul_ps(va1, vb4567c3));
164 vacc2x4567 = _mm_add_ps(vacc2x4567, _mm_mul_ps(va2, vb4567c3));
165 vacc3x4567 = _mm_add_ps(vacc3x4567, _mm_mul_ps(va3, vb4567c3));
166
167
168 w += 32;
169 k -= 4 * sizeof(float);
170 }
171 if XNN_UNLIKELY(k != 0) {
172 __m128 va0 = _mm_loadu_ps(a0);
173 a0 = (const float*) ((uintptr_t) a0 + k);
174 __m128 va1 = _mm_loadu_ps(a1);
175 a1 = (const float*) ((uintptr_t) a1 + k);
176 __m128 va2 = _mm_loadu_ps(a2);
177 a2 = (const float*) ((uintptr_t) a2 + k);
178 __m128 va3 = _mm_loadu_ps(a3);
179 a3 = (const float*) ((uintptr_t) a3 + k);
180
181
182 const __m128 vb0123c0 = _mm_load_ps(w + 0);
183 const __m128 vb4567c0 = _mm_load_ps(w + 4);
184
185 vacc0x0123 = _mm_add_ps(vacc0x0123, _mm_mul_ps(_mm_andnot_ps(_mm_cmpeq_ps(_mm_setzero_ps(), vb0123c0), va0), vb0123c0));
186 vacc1x0123 = _mm_add_ps(vacc1x0123, _mm_mul_ps(_mm_andnot_ps(_mm_cmpeq_ps(_mm_setzero_ps(), vb0123c0), va1), vb0123c0));
187 vacc2x0123 = _mm_add_ps(vacc2x0123, _mm_mul_ps(_mm_andnot_ps(_mm_cmpeq_ps(_mm_setzero_ps(), vb0123c0), va2), vb0123c0));
188 vacc3x0123 = _mm_add_ps(vacc3x0123, _mm_mul_ps(_mm_andnot_ps(_mm_cmpeq_ps(_mm_setzero_ps(), vb0123c0), va3), vb0123c0));
189 vacc0x4567 = _mm_add_ps(vacc0x4567, _mm_mul_ps(_mm_andnot_ps(_mm_cmpeq_ps(_mm_setzero_ps(), vb4567c0), va0), vb4567c0));
190 vacc1x4567 = _mm_add_ps(vacc1x4567, _mm_mul_ps(_mm_andnot_ps(_mm_cmpeq_ps(_mm_setzero_ps(), vb4567c0), va1), vb4567c0));
191 vacc2x4567 = _mm_add_ps(vacc2x4567, _mm_mul_ps(_mm_andnot_ps(_mm_cmpeq_ps(_mm_setzero_ps(), vb4567c0), va2), vb4567c0));
192 vacc3x4567 = _mm_add_ps(vacc3x4567, _mm_mul_ps(_mm_andnot_ps(_mm_cmpeq_ps(_mm_setzero_ps(), vb4567c0), va3), vb4567c0));
193
194 va0 = _mm_shuffle_ps(va0, va0, _MM_SHUFFLE(0, 3, 2, 1));
195 va1 = _mm_shuffle_ps(va1, va1, _MM_SHUFFLE(0, 3, 2, 1));
196 va2 = _mm_shuffle_ps(va2, va2, _MM_SHUFFLE(0, 3, 2, 1));
197 va3 = _mm_shuffle_ps(va3, va3, _MM_SHUFFLE(0, 3, 2, 1));
198
199 const __m128 vb0123c1 = _mm_load_ps(w + 8);
200 const __m128 vb4567c1 = _mm_load_ps(w + 12);
201
202 vacc0x0123 = _mm_add_ps(vacc0x0123, _mm_mul_ps(_mm_andnot_ps(_mm_cmpeq_ps(_mm_setzero_ps(), vb0123c1), va0), vb0123c1));
203 vacc1x0123 = _mm_add_ps(vacc1x0123, _mm_mul_ps(_mm_andnot_ps(_mm_cmpeq_ps(_mm_setzero_ps(), vb0123c1), va1), vb0123c1));
204 vacc2x0123 = _mm_add_ps(vacc2x0123, _mm_mul_ps(_mm_andnot_ps(_mm_cmpeq_ps(_mm_setzero_ps(), vb0123c1), va2), vb0123c1));
205 vacc3x0123 = _mm_add_ps(vacc3x0123, _mm_mul_ps(_mm_andnot_ps(_mm_cmpeq_ps(_mm_setzero_ps(), vb0123c1), va3), vb0123c1));
206 vacc0x4567 = _mm_add_ps(vacc0x4567, _mm_mul_ps(_mm_andnot_ps(_mm_cmpeq_ps(_mm_setzero_ps(), vb4567c1), va0), vb4567c1));
207 vacc1x4567 = _mm_add_ps(vacc1x4567, _mm_mul_ps(_mm_andnot_ps(_mm_cmpeq_ps(_mm_setzero_ps(), vb4567c1), va1), vb4567c1));
208 vacc2x4567 = _mm_add_ps(vacc2x4567, _mm_mul_ps(_mm_andnot_ps(_mm_cmpeq_ps(_mm_setzero_ps(), vb4567c1), va2), vb4567c1));
209 vacc3x4567 = _mm_add_ps(vacc3x4567, _mm_mul_ps(_mm_andnot_ps(_mm_cmpeq_ps(_mm_setzero_ps(), vb4567c1), va3), vb4567c1));
210
211 va0 = _mm_shuffle_ps(va0, va0, _MM_SHUFFLE(0, 3, 2, 1));
212 va1 = _mm_shuffle_ps(va1, va1, _MM_SHUFFLE(0, 3, 2, 1));
213 va2 = _mm_shuffle_ps(va2, va2, _MM_SHUFFLE(0, 3, 2, 1));
214 va3 = _mm_shuffle_ps(va3, va3, _MM_SHUFFLE(0, 3, 2, 1));
215
216 const __m128 vb0123c2 = _mm_load_ps(w + 16);
217 const __m128 vb4567c2 = _mm_load_ps(w + 20);
218
219 vacc0x0123 = _mm_add_ps(vacc0x0123, _mm_mul_ps(_mm_andnot_ps(_mm_cmpeq_ps(_mm_setzero_ps(), vb0123c2), va0), vb0123c2));
220 vacc1x0123 = _mm_add_ps(vacc1x0123, _mm_mul_ps(_mm_andnot_ps(_mm_cmpeq_ps(_mm_setzero_ps(), vb0123c2), va1), vb0123c2));
221 vacc2x0123 = _mm_add_ps(vacc2x0123, _mm_mul_ps(_mm_andnot_ps(_mm_cmpeq_ps(_mm_setzero_ps(), vb0123c2), va2), vb0123c2));
222 vacc3x0123 = _mm_add_ps(vacc3x0123, _mm_mul_ps(_mm_andnot_ps(_mm_cmpeq_ps(_mm_setzero_ps(), vb0123c2), va3), vb0123c2));
223 vacc0x4567 = _mm_add_ps(vacc0x4567, _mm_mul_ps(_mm_andnot_ps(_mm_cmpeq_ps(_mm_setzero_ps(), vb4567c2), va0), vb4567c2));
224 vacc1x4567 = _mm_add_ps(vacc1x4567, _mm_mul_ps(_mm_andnot_ps(_mm_cmpeq_ps(_mm_setzero_ps(), vb4567c2), va1), vb4567c2));
225 vacc2x4567 = _mm_add_ps(vacc2x4567, _mm_mul_ps(_mm_andnot_ps(_mm_cmpeq_ps(_mm_setzero_ps(), vb4567c2), va2), vb4567c2));
226 vacc3x4567 = _mm_add_ps(vacc3x4567, _mm_mul_ps(_mm_andnot_ps(_mm_cmpeq_ps(_mm_setzero_ps(), vb4567c2), va3), vb4567c2));
227
228 va0 = _mm_shuffle_ps(va0, va0, _MM_SHUFFLE(0, 3, 2, 1));
229 va1 = _mm_shuffle_ps(va1, va1, _MM_SHUFFLE(0, 3, 2, 1));
230 va2 = _mm_shuffle_ps(va2, va2, _MM_SHUFFLE(0, 3, 2, 1));
231 va3 = _mm_shuffle_ps(va3, va3, _MM_SHUFFLE(0, 3, 2, 1));
232
233 const __m128 vb0123c3 = _mm_load_ps(w + 24);
234 const __m128 vb4567c3 = _mm_load_ps(w + 28);
235
236 vacc0x0123 = _mm_add_ps(vacc0x0123, _mm_mul_ps(_mm_andnot_ps(_mm_cmpeq_ps(_mm_setzero_ps(), vb0123c3), va0), vb0123c3));
237 vacc1x0123 = _mm_add_ps(vacc1x0123, _mm_mul_ps(_mm_andnot_ps(_mm_cmpeq_ps(_mm_setzero_ps(), vb0123c3), va1), vb0123c3));
238 vacc2x0123 = _mm_add_ps(vacc2x0123, _mm_mul_ps(_mm_andnot_ps(_mm_cmpeq_ps(_mm_setzero_ps(), vb0123c3), va2), vb0123c3));
239 vacc3x0123 = _mm_add_ps(vacc3x0123, _mm_mul_ps(_mm_andnot_ps(_mm_cmpeq_ps(_mm_setzero_ps(), vb0123c3), va3), vb0123c3));
240 vacc0x4567 = _mm_add_ps(vacc0x4567, _mm_mul_ps(_mm_andnot_ps(_mm_cmpeq_ps(_mm_setzero_ps(), vb4567c3), va0), vb4567c3));
241 vacc1x4567 = _mm_add_ps(vacc1x4567, _mm_mul_ps(_mm_andnot_ps(_mm_cmpeq_ps(_mm_setzero_ps(), vb4567c3), va1), vb4567c3));
242 vacc2x4567 = _mm_add_ps(vacc2x4567, _mm_mul_ps(_mm_andnot_ps(_mm_cmpeq_ps(_mm_setzero_ps(), vb4567c3), va2), vb4567c3));
243 vacc3x4567 = _mm_add_ps(vacc3x4567, _mm_mul_ps(_mm_andnot_ps(_mm_cmpeq_ps(_mm_setzero_ps(), vb4567c3), va3), vb4567c3));
244
245
246 w += 32;
247 }
248 p -= 4 * sizeof(void*);
249 } while (p != 0);
250
251 const __m128 vmax = _mm_load_ps(params->sse.max);
252 vacc0x0123 = _mm_min_ps(vacc0x0123, vmax);
253 vacc1x0123 = _mm_min_ps(vacc1x0123, vmax);
254 vacc2x0123 = _mm_min_ps(vacc2x0123, vmax);
255 vacc3x0123 = _mm_min_ps(vacc3x0123, vmax);
256 vacc0x4567 = _mm_min_ps(vacc0x4567, vmax);
257 vacc1x4567 = _mm_min_ps(vacc1x4567, vmax);
258 vacc2x4567 = _mm_min_ps(vacc2x4567, vmax);
259 vacc3x4567 = _mm_min_ps(vacc3x4567, vmax);
260
261 const __m128 vmin = _mm_load_ps(params->sse.min);
262 vacc0x0123 = _mm_max_ps(vacc0x0123, vmin);
263 vacc1x0123 = _mm_max_ps(vacc1x0123, vmin);
264 vacc2x0123 = _mm_max_ps(vacc2x0123, vmin);
265 vacc3x0123 = _mm_max_ps(vacc3x0123, vmin);
266 vacc0x4567 = _mm_max_ps(vacc0x4567, vmin);
267 vacc1x4567 = _mm_max_ps(vacc1x4567, vmin);
268 vacc2x4567 = _mm_max_ps(vacc2x4567, vmin);
269 vacc3x4567 = _mm_max_ps(vacc3x4567, vmin);
270
271 if XNN_LIKELY(nc >= 8) {
272 _mm_storeu_ps(c3, vacc3x0123);
273 _mm_storeu_ps(c3 + 4, vacc3x4567);
274 c3 = (float*) ((uintptr_t) c3 + cn_stride);
275 _mm_storeu_ps(c2, vacc2x0123);
276 _mm_storeu_ps(c2 + 4, vacc2x4567);
277 c2 = (float*) ((uintptr_t) c2 + cn_stride);
278 _mm_storeu_ps(c1, vacc1x0123);
279 _mm_storeu_ps(c1 + 4, vacc1x4567);
280 c1 = (float*) ((uintptr_t) c1 + cn_stride);
281 _mm_storeu_ps(c0, vacc0x0123);
282 _mm_storeu_ps(c0 + 4, vacc0x4567);
283 c0 = (float*) ((uintptr_t) c0 + cn_stride);
284
285 a = (const float**restrict) ((uintptr_t) a - ks);
286 nc -= 8;
287 } else {
288 if (nc & 4) {
289 _mm_storeu_ps(c3, vacc3x0123);
290 _mm_storeu_ps(c2, vacc2x0123);
291 _mm_storeu_ps(c1, vacc1x0123);
292 _mm_storeu_ps(c0, vacc0x0123);
293
294 vacc3x0123 = vacc3x4567;
295 vacc2x0123 = vacc2x4567;
296 vacc1x0123 = vacc1x4567;
297 vacc0x0123 = vacc0x4567;
298
299 c3 += 4;
300 c2 += 4;
301 c1 += 4;
302 c0 += 4;
303 }
304 if (nc & 2) {
305 _mm_storel_pi((__m64*) c3, vacc3x0123);
306 _mm_storel_pi((__m64*) c2, vacc2x0123);
307 _mm_storel_pi((__m64*) c1, vacc1x0123);
308 _mm_storel_pi((__m64*) c0, vacc0x0123);
309
310 vacc3x0123 = _mm_movehl_ps(vacc3x0123, vacc3x0123);
311 vacc2x0123 = _mm_movehl_ps(vacc2x0123, vacc2x0123);
312 vacc1x0123 = _mm_movehl_ps(vacc1x0123, vacc1x0123);
313 vacc0x0123 = _mm_movehl_ps(vacc0x0123, vacc0x0123);
314
315 c3 += 2;
316 c2 += 2;
317 c1 += 2;
318 c0 += 2;
319 }
320 if (nc & 1) {
321 _mm_store_ss(c3, vacc3x0123);
322 _mm_store_ss(c2, vacc2x0123);
323 _mm_store_ss(c1, vacc1x0123);
324 _mm_store_ss(c0, vacc0x0123);
325 }
326
327 nc = 0;
328 }
329 } while (nc != 0);
330 }
331