1 // Auto-generated file. Do not edit!
2 // Template: src/f32-gemm/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/gemm.h>
15
16
xnn_f32_gemminc_minmax_ukernel_4x8s4__sse(size_t mr,size_t nc,size_t kc,const float * restrict a,size_t a_stride,const float * restrict w,float * restrict c,size_t cm_stride,size_t cn_stride,const float * restrict acc,const union xnn_f32_minmax_params params[restrict XNN_MIN_ELEMENTS (1)])17 void xnn_f32_gemminc_minmax_ukernel_4x8s4__sse(
18 size_t mr,
19 size_t nc,
20 size_t kc,
21 const float*restrict a,
22 size_t a_stride,
23 const float*restrict w,
24 float*restrict c,
25 size_t cm_stride,
26 size_t cn_stride,
27 const float*restrict acc,
28 const union xnn_f32_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
29 {
30 assert(mr != 0);
31 assert(mr <= 4);
32 assert(nc != 0);
33 assert(kc != 0);
34 assert(kc % sizeof(float) == 0);
35 assert(a != NULL);
36 assert(w != NULL);
37 assert(c != NULL);
38 assert(acc != NULL);
39
40 const float* a0 = a;
41 float* c0 = c;
42 const float* a1 = (const float*) ((uintptr_t) a0 + a_stride);
43 float* c1 = (float*) ((uintptr_t) c0 + cm_stride);
44 if XNN_UNPREDICTABLE(mr < 2) {
45 a1 = a0;
46 c1 = c0;
47 }
48 const float* a2 = (const float*) ((uintptr_t) a1 + a_stride);
49 float* c2 = (float*) ((uintptr_t) c1 + cm_stride);
50 if XNN_UNPREDICTABLE(mr <= 2) {
51 a2 = a1;
52 c2 = c1;
53 }
54 const float* a3 = (const float*) ((uintptr_t) a2 + a_stride);
55 float* c3 = (float*) ((uintptr_t) c2 + cm_stride);
56 if XNN_UNPREDICTABLE(mr != 4) {
57 a3 = a2;
58 c3 = c2;
59 }
60
61 do {
62 __m128 vacc0x0123 = _mm_load_ps(acc + 0);
63 __m128 vacc0x4567 = _mm_load_ps(acc + 4);
64 __m128 vacc1x0123 = _mm_load_ps(acc + 8);
65 __m128 vacc1x4567 = _mm_load_ps(acc + 12);
66 __m128 vacc2x0123 = _mm_load_ps(acc + 16);
67 __m128 vacc2x4567 = _mm_load_ps(acc + 20);
68 __m128 vacc3x0123 = _mm_load_ps(acc + 24);
69 __m128 vacc3x4567 = _mm_load_ps(acc + 28);
70 acc += 32;
71
72 size_t k = kc;
73 while (k >= 4 * sizeof(float)) {
74 __m128 va0 = _mm_loadu_ps(a0);
75 a0 += 4;
76 __m128 va1 = _mm_loadu_ps(a1);
77 a1 += 4;
78 __m128 va2 = _mm_loadu_ps(a2);
79 a2 += 4;
80 __m128 va3 = _mm_loadu_ps(a3);
81 a3 += 4;
82
83
84 const __m128 vb0123c0 = _mm_load_ps(w + 0);
85 const __m128 vb4567c0 = _mm_load_ps(w + 4);
86
87 vacc0x0123 = _mm_add_ps(vacc0x0123, _mm_mul_ps(va0, vb0123c0));
88 vacc1x0123 = _mm_add_ps(vacc1x0123, _mm_mul_ps(va1, vb0123c0));
89 vacc2x0123 = _mm_add_ps(vacc2x0123, _mm_mul_ps(va2, vb0123c0));
90 vacc3x0123 = _mm_add_ps(vacc3x0123, _mm_mul_ps(va3, vb0123c0));
91 vacc0x4567 = _mm_add_ps(vacc0x4567, _mm_mul_ps(va0, vb4567c0));
92 vacc1x4567 = _mm_add_ps(vacc1x4567, _mm_mul_ps(va1, vb4567c0));
93 vacc2x4567 = _mm_add_ps(vacc2x4567, _mm_mul_ps(va2, vb4567c0));
94 vacc3x4567 = _mm_add_ps(vacc3x4567, _mm_mul_ps(va3, vb4567c0));
95
96 va0 = _mm_shuffle_ps(va0, va0, _MM_SHUFFLE(0, 3, 2, 1));
97 va1 = _mm_shuffle_ps(va1, va1, _MM_SHUFFLE(0, 3, 2, 1));
98 va2 = _mm_shuffle_ps(va2, va2, _MM_SHUFFLE(0, 3, 2, 1));
99 va3 = _mm_shuffle_ps(va3, va3, _MM_SHUFFLE(0, 3, 2, 1));
100
101 const __m128 vb0123c1 = _mm_load_ps(w + 8);
102 const __m128 vb4567c1 = _mm_load_ps(w + 12);
103
104 vacc0x0123 = _mm_add_ps(vacc0x0123, _mm_mul_ps(va0, vb0123c1));
105 vacc1x0123 = _mm_add_ps(vacc1x0123, _mm_mul_ps(va1, vb0123c1));
106 vacc2x0123 = _mm_add_ps(vacc2x0123, _mm_mul_ps(va2, vb0123c1));
107 vacc3x0123 = _mm_add_ps(vacc3x0123, _mm_mul_ps(va3, vb0123c1));
108 vacc0x4567 = _mm_add_ps(vacc0x4567, _mm_mul_ps(va0, vb4567c1));
109 vacc1x4567 = _mm_add_ps(vacc1x4567, _mm_mul_ps(va1, vb4567c1));
110 vacc2x4567 = _mm_add_ps(vacc2x4567, _mm_mul_ps(va2, vb4567c1));
111 vacc3x4567 = _mm_add_ps(vacc3x4567, _mm_mul_ps(va3, vb4567c1));
112
113 va0 = _mm_shuffle_ps(va0, va0, _MM_SHUFFLE(0, 3, 2, 1));
114 va1 = _mm_shuffle_ps(va1, va1, _MM_SHUFFLE(0, 3, 2, 1));
115 va2 = _mm_shuffle_ps(va2, va2, _MM_SHUFFLE(0, 3, 2, 1));
116 va3 = _mm_shuffle_ps(va3, va3, _MM_SHUFFLE(0, 3, 2, 1));
117
118 const __m128 vb0123c2 = _mm_load_ps(w + 16);
119 const __m128 vb4567c2 = _mm_load_ps(w + 20);
120
121 vacc0x0123 = _mm_add_ps(vacc0x0123, _mm_mul_ps(va0, vb0123c2));
122 vacc1x0123 = _mm_add_ps(vacc1x0123, _mm_mul_ps(va1, vb0123c2));
123 vacc2x0123 = _mm_add_ps(vacc2x0123, _mm_mul_ps(va2, vb0123c2));
124 vacc3x0123 = _mm_add_ps(vacc3x0123, _mm_mul_ps(va3, vb0123c2));
125 vacc0x4567 = _mm_add_ps(vacc0x4567, _mm_mul_ps(va0, vb4567c2));
126 vacc1x4567 = _mm_add_ps(vacc1x4567, _mm_mul_ps(va1, vb4567c2));
127 vacc2x4567 = _mm_add_ps(vacc2x4567, _mm_mul_ps(va2, vb4567c2));
128 vacc3x4567 = _mm_add_ps(vacc3x4567, _mm_mul_ps(va3, vb4567c2));
129
130 va0 = _mm_shuffle_ps(va0, va0, _MM_SHUFFLE(0, 3, 2, 1));
131 va1 = _mm_shuffle_ps(va1, va1, _MM_SHUFFLE(0, 3, 2, 1));
132 va2 = _mm_shuffle_ps(va2, va2, _MM_SHUFFLE(0, 3, 2, 1));
133 va3 = _mm_shuffle_ps(va3, va3, _MM_SHUFFLE(0, 3, 2, 1));
134
135 const __m128 vb0123c3 = _mm_load_ps(w + 24);
136 const __m128 vb4567c3 = _mm_load_ps(w + 28);
137
138 vacc0x0123 = _mm_add_ps(vacc0x0123, _mm_mul_ps(va0, vb0123c3));
139 vacc1x0123 = _mm_add_ps(vacc1x0123, _mm_mul_ps(va1, vb0123c3));
140 vacc2x0123 = _mm_add_ps(vacc2x0123, _mm_mul_ps(va2, vb0123c3));
141 vacc3x0123 = _mm_add_ps(vacc3x0123, _mm_mul_ps(va3, vb0123c3));
142 vacc0x4567 = _mm_add_ps(vacc0x4567, _mm_mul_ps(va0, vb4567c3));
143 vacc1x4567 = _mm_add_ps(vacc1x4567, _mm_mul_ps(va1, vb4567c3));
144 vacc2x4567 = _mm_add_ps(vacc2x4567, _mm_mul_ps(va2, vb4567c3));
145 vacc3x4567 = _mm_add_ps(vacc3x4567, _mm_mul_ps(va3, vb4567c3));
146
147
148 w += 32;
149 k -= 4 * sizeof(float);
150 }
151 if XNN_UNLIKELY(k != 0) {
152 __m128 va0 = _mm_loadu_ps(a0);
153 a0 = (const float*) ((uintptr_t) a0 + k);
154 __m128 va1 = _mm_loadu_ps(a1);
155 a1 = (const float*) ((uintptr_t) a1 + k);
156 __m128 va2 = _mm_loadu_ps(a2);
157 a2 = (const float*) ((uintptr_t) a2 + k);
158 __m128 va3 = _mm_loadu_ps(a3);
159 a3 = (const float*) ((uintptr_t) a3 + k);
160
161
162 const __m128 vb0123c0 = _mm_load_ps(w + 0);
163 const __m128 vb4567c0 = _mm_load_ps(w + 4);
164
165 vacc0x0123 = _mm_add_ps(vacc0x0123, _mm_mul_ps(_mm_andnot_ps(_mm_cmpeq_ps(_mm_setzero_ps(), vb0123c0), va0), vb0123c0));
166 vacc1x0123 = _mm_add_ps(vacc1x0123, _mm_mul_ps(_mm_andnot_ps(_mm_cmpeq_ps(_mm_setzero_ps(), vb0123c0), va1), vb0123c0));
167 vacc2x0123 = _mm_add_ps(vacc2x0123, _mm_mul_ps(_mm_andnot_ps(_mm_cmpeq_ps(_mm_setzero_ps(), vb0123c0), va2), vb0123c0));
168 vacc3x0123 = _mm_add_ps(vacc3x0123, _mm_mul_ps(_mm_andnot_ps(_mm_cmpeq_ps(_mm_setzero_ps(), vb0123c0), va3), vb0123c0));
169 vacc0x4567 = _mm_add_ps(vacc0x4567, _mm_mul_ps(_mm_andnot_ps(_mm_cmpeq_ps(_mm_setzero_ps(), vb4567c0), va0), vb4567c0));
170 vacc1x4567 = _mm_add_ps(vacc1x4567, _mm_mul_ps(_mm_andnot_ps(_mm_cmpeq_ps(_mm_setzero_ps(), vb4567c0), va1), vb4567c0));
171 vacc2x4567 = _mm_add_ps(vacc2x4567, _mm_mul_ps(_mm_andnot_ps(_mm_cmpeq_ps(_mm_setzero_ps(), vb4567c0), va2), vb4567c0));
172 vacc3x4567 = _mm_add_ps(vacc3x4567, _mm_mul_ps(_mm_andnot_ps(_mm_cmpeq_ps(_mm_setzero_ps(), vb4567c0), va3), vb4567c0));
173
174 va0 = _mm_shuffle_ps(va0, va0, _MM_SHUFFLE(0, 3, 2, 1));
175 va1 = _mm_shuffle_ps(va1, va1, _MM_SHUFFLE(0, 3, 2, 1));
176 va2 = _mm_shuffle_ps(va2, va2, _MM_SHUFFLE(0, 3, 2, 1));
177 va3 = _mm_shuffle_ps(va3, va3, _MM_SHUFFLE(0, 3, 2, 1));
178
179 const __m128 vb0123c1 = _mm_load_ps(w + 8);
180 const __m128 vb4567c1 = _mm_load_ps(w + 12);
181
182 vacc0x0123 = _mm_add_ps(vacc0x0123, _mm_mul_ps(_mm_andnot_ps(_mm_cmpeq_ps(_mm_setzero_ps(), vb0123c1), va0), vb0123c1));
183 vacc1x0123 = _mm_add_ps(vacc1x0123, _mm_mul_ps(_mm_andnot_ps(_mm_cmpeq_ps(_mm_setzero_ps(), vb0123c1), va1), vb0123c1));
184 vacc2x0123 = _mm_add_ps(vacc2x0123, _mm_mul_ps(_mm_andnot_ps(_mm_cmpeq_ps(_mm_setzero_ps(), vb0123c1), va2), vb0123c1));
185 vacc3x0123 = _mm_add_ps(vacc3x0123, _mm_mul_ps(_mm_andnot_ps(_mm_cmpeq_ps(_mm_setzero_ps(), vb0123c1), va3), vb0123c1));
186 vacc0x4567 = _mm_add_ps(vacc0x4567, _mm_mul_ps(_mm_andnot_ps(_mm_cmpeq_ps(_mm_setzero_ps(), vb4567c1), va0), vb4567c1));
187 vacc1x4567 = _mm_add_ps(vacc1x4567, _mm_mul_ps(_mm_andnot_ps(_mm_cmpeq_ps(_mm_setzero_ps(), vb4567c1), va1), vb4567c1));
188 vacc2x4567 = _mm_add_ps(vacc2x4567, _mm_mul_ps(_mm_andnot_ps(_mm_cmpeq_ps(_mm_setzero_ps(), vb4567c1), va2), vb4567c1));
189 vacc3x4567 = _mm_add_ps(vacc3x4567, _mm_mul_ps(_mm_andnot_ps(_mm_cmpeq_ps(_mm_setzero_ps(), vb4567c1), va3), vb4567c1));
190
191 va0 = _mm_shuffle_ps(va0, va0, _MM_SHUFFLE(0, 3, 2, 1));
192 va1 = _mm_shuffle_ps(va1, va1, _MM_SHUFFLE(0, 3, 2, 1));
193 va2 = _mm_shuffle_ps(va2, va2, _MM_SHUFFLE(0, 3, 2, 1));
194 va3 = _mm_shuffle_ps(va3, va3, _MM_SHUFFLE(0, 3, 2, 1));
195
196 const __m128 vb0123c2 = _mm_load_ps(w + 16);
197 const __m128 vb4567c2 = _mm_load_ps(w + 20);
198
199 vacc0x0123 = _mm_add_ps(vacc0x0123, _mm_mul_ps(_mm_andnot_ps(_mm_cmpeq_ps(_mm_setzero_ps(), vb0123c2), va0), vb0123c2));
200 vacc1x0123 = _mm_add_ps(vacc1x0123, _mm_mul_ps(_mm_andnot_ps(_mm_cmpeq_ps(_mm_setzero_ps(), vb0123c2), va1), vb0123c2));
201 vacc2x0123 = _mm_add_ps(vacc2x0123, _mm_mul_ps(_mm_andnot_ps(_mm_cmpeq_ps(_mm_setzero_ps(), vb0123c2), va2), vb0123c2));
202 vacc3x0123 = _mm_add_ps(vacc3x0123, _mm_mul_ps(_mm_andnot_ps(_mm_cmpeq_ps(_mm_setzero_ps(), vb0123c2), va3), vb0123c2));
203 vacc0x4567 = _mm_add_ps(vacc0x4567, _mm_mul_ps(_mm_andnot_ps(_mm_cmpeq_ps(_mm_setzero_ps(), vb4567c2), va0), vb4567c2));
204 vacc1x4567 = _mm_add_ps(vacc1x4567, _mm_mul_ps(_mm_andnot_ps(_mm_cmpeq_ps(_mm_setzero_ps(), vb4567c2), va1), vb4567c2));
205 vacc2x4567 = _mm_add_ps(vacc2x4567, _mm_mul_ps(_mm_andnot_ps(_mm_cmpeq_ps(_mm_setzero_ps(), vb4567c2), va2), vb4567c2));
206 vacc3x4567 = _mm_add_ps(vacc3x4567, _mm_mul_ps(_mm_andnot_ps(_mm_cmpeq_ps(_mm_setzero_ps(), vb4567c2), va3), vb4567c2));
207
208 va0 = _mm_shuffle_ps(va0, va0, _MM_SHUFFLE(0, 3, 2, 1));
209 va1 = _mm_shuffle_ps(va1, va1, _MM_SHUFFLE(0, 3, 2, 1));
210 va2 = _mm_shuffle_ps(va2, va2, _MM_SHUFFLE(0, 3, 2, 1));
211 va3 = _mm_shuffle_ps(va3, va3, _MM_SHUFFLE(0, 3, 2, 1));
212
213 const __m128 vb0123c3 = _mm_load_ps(w + 24);
214 const __m128 vb4567c3 = _mm_load_ps(w + 28);
215
216 vacc0x0123 = _mm_add_ps(vacc0x0123, _mm_mul_ps(_mm_andnot_ps(_mm_cmpeq_ps(_mm_setzero_ps(), vb0123c3), va0), vb0123c3));
217 vacc1x0123 = _mm_add_ps(vacc1x0123, _mm_mul_ps(_mm_andnot_ps(_mm_cmpeq_ps(_mm_setzero_ps(), vb0123c3), va1), vb0123c3));
218 vacc2x0123 = _mm_add_ps(vacc2x0123, _mm_mul_ps(_mm_andnot_ps(_mm_cmpeq_ps(_mm_setzero_ps(), vb0123c3), va2), vb0123c3));
219 vacc3x0123 = _mm_add_ps(vacc3x0123, _mm_mul_ps(_mm_andnot_ps(_mm_cmpeq_ps(_mm_setzero_ps(), vb0123c3), va3), vb0123c3));
220 vacc0x4567 = _mm_add_ps(vacc0x4567, _mm_mul_ps(_mm_andnot_ps(_mm_cmpeq_ps(_mm_setzero_ps(), vb4567c3), va0), vb4567c3));
221 vacc1x4567 = _mm_add_ps(vacc1x4567, _mm_mul_ps(_mm_andnot_ps(_mm_cmpeq_ps(_mm_setzero_ps(), vb4567c3), va1), vb4567c3));
222 vacc2x4567 = _mm_add_ps(vacc2x4567, _mm_mul_ps(_mm_andnot_ps(_mm_cmpeq_ps(_mm_setzero_ps(), vb4567c3), va2), vb4567c3));
223 vacc3x4567 = _mm_add_ps(vacc3x4567, _mm_mul_ps(_mm_andnot_ps(_mm_cmpeq_ps(_mm_setzero_ps(), vb4567c3), va3), vb4567c3));
224
225
226 w += 32;
227 }
228
229 const __m128 vmax = _mm_load_ps(params->sse.max);
230 vacc0x0123 = _mm_min_ps(vacc0x0123, vmax);
231 vacc1x0123 = _mm_min_ps(vacc1x0123, vmax);
232 vacc2x0123 = _mm_min_ps(vacc2x0123, vmax);
233 vacc3x0123 = _mm_min_ps(vacc3x0123, vmax);
234 vacc0x4567 = _mm_min_ps(vacc0x4567, vmax);
235 vacc1x4567 = _mm_min_ps(vacc1x4567, vmax);
236 vacc2x4567 = _mm_min_ps(vacc2x4567, vmax);
237 vacc3x4567 = _mm_min_ps(vacc3x4567, vmax);
238
239 const __m128 vmin = _mm_load_ps(params->sse.min);
240 vacc0x0123 = _mm_max_ps(vacc0x0123, vmin);
241 vacc1x0123 = _mm_max_ps(vacc1x0123, vmin);
242 vacc2x0123 = _mm_max_ps(vacc2x0123, vmin);
243 vacc3x0123 = _mm_max_ps(vacc3x0123, vmin);
244 vacc0x4567 = _mm_max_ps(vacc0x4567, vmin);
245 vacc1x4567 = _mm_max_ps(vacc1x4567, vmin);
246 vacc2x4567 = _mm_max_ps(vacc2x4567, vmin);
247 vacc3x4567 = _mm_max_ps(vacc3x4567, vmin);
248
249 if XNN_LIKELY(nc >= 8) {
250 _mm_storeu_ps(c3, vacc3x0123);
251 _mm_storeu_ps(c3 + 4, vacc3x4567);
252 c3 = (float*) ((uintptr_t) c3 + cn_stride);
253 _mm_storeu_ps(c2, vacc2x0123);
254 _mm_storeu_ps(c2 + 4, vacc2x4567);
255 c2 = (float*) ((uintptr_t) c2 + cn_stride);
256 _mm_storeu_ps(c1, vacc1x0123);
257 _mm_storeu_ps(c1 + 4, vacc1x4567);
258 c1 = (float*) ((uintptr_t) c1 + cn_stride);
259 _mm_storeu_ps(c0, vacc0x0123);
260 _mm_storeu_ps(c0 + 4, vacc0x4567);
261 c0 = (float*) ((uintptr_t) c0 + cn_stride);
262
263 a3 = (const float*) ((uintptr_t) a3 - kc);
264 a2 = (const float*) ((uintptr_t) a2 - kc);
265 a1 = (const float*) ((uintptr_t) a1 - kc);
266 a0 = (const float*) ((uintptr_t) a0 - kc);
267
268 nc -= 8;
269 } else {
270 if (nc & 4) {
271 _mm_storeu_ps(c3, vacc3x0123);
272 _mm_storeu_ps(c2, vacc2x0123);
273 _mm_storeu_ps(c1, vacc1x0123);
274 _mm_storeu_ps(c0, vacc0x0123);
275
276 vacc3x0123 = vacc3x4567;
277 vacc2x0123 = vacc2x4567;
278 vacc1x0123 = vacc1x4567;
279 vacc0x0123 = vacc0x4567;
280
281 c3 += 4;
282 c2 += 4;
283 c1 += 4;
284 c0 += 4;
285 }
286 if (nc & 2) {
287 _mm_storel_pi((__m64*) c3, vacc3x0123);
288 _mm_storel_pi((__m64*) c2, vacc2x0123);
289 _mm_storel_pi((__m64*) c1, vacc1x0123);
290 _mm_storel_pi((__m64*) c0, vacc0x0123);
291
292 vacc3x0123 = _mm_movehl_ps(vacc3x0123, vacc3x0123);
293 vacc2x0123 = _mm_movehl_ps(vacc2x0123, vacc2x0123);
294 vacc1x0123 = _mm_movehl_ps(vacc1x0123, vacc1x0123);
295 vacc0x0123 = _mm_movehl_ps(vacc0x0123, vacc0x0123);
296
297 c3 += 2;
298 c2 += 2;
299 c1 += 2;
300 c0 += 2;
301 }
302 if (nc & 1) {
303 _mm_store_ss(c3, vacc3x0123);
304 _mm_store_ss(c2, vacc2x0123);
305 _mm_store_ss(c1, vacc1x0123);
306 _mm_store_ss(c0, vacc0x0123);
307 }
308
309 nc = 0;
310 }
311 } while (nc != 0);
312 }
313