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_gemm_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 union xnn_f32_minmax_params params[restrict XNN_MIN_ELEMENTS (1)])17 void xnn_f32_gemm_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 union xnn_f32_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
28 {
29 assert(mr != 0);
30 assert(mr <= 4);
31 assert(nc != 0);
32 assert(kc != 0);
33 assert(kc % sizeof(float) == 0);
34 assert(a != NULL);
35 assert(w != NULL);
36 assert(c != NULL);
37
38 const float* a0 = a;
39 float* c0 = c;
40 const float* a1 = (const float*) ((uintptr_t) a0 + a_stride);
41 float* c1 = (float*) ((uintptr_t) c0 + cm_stride);
42 if XNN_UNPREDICTABLE(mr < 2) {
43 a1 = a0;
44 c1 = c0;
45 }
46 const float* a2 = (const float*) ((uintptr_t) a1 + a_stride);
47 float* c2 = (float*) ((uintptr_t) c1 + cm_stride);
48 if XNN_UNPREDICTABLE(mr <= 2) {
49 a2 = a1;
50 c2 = c1;
51 }
52 const float* a3 = (const float*) ((uintptr_t) a2 + a_stride);
53 float* c3 = (float*) ((uintptr_t) c2 + cm_stride);
54 if XNN_UNPREDICTABLE(mr != 4) {
55 a3 = a2;
56 c3 = c2;
57 }
58
59 do {
60 __m128 vacc0x0123 = _mm_load_ps(w + 0);
61 __m128 vacc0x4567 = _mm_load_ps(w + 4);
62 __m128 vacc1x0123 = vacc0x0123;
63 __m128 vacc1x4567 = vacc0x4567;
64 __m128 vacc2x0123 = vacc0x0123;
65 __m128 vacc2x4567 = vacc0x4567;
66 __m128 vacc3x0123 = vacc0x0123;
67 __m128 vacc3x4567 = vacc0x4567;
68 w += 8;
69
70 size_t k = kc;
71 while (k >= 4 * sizeof(float)) {
72 __m128 va0 = _mm_loadu_ps(a0);
73 a0 += 4;
74 __m128 va1 = _mm_loadu_ps(a1);
75 a1 += 4;
76 __m128 va2 = _mm_loadu_ps(a2);
77 a2 += 4;
78 __m128 va3 = _mm_loadu_ps(a3);
79 a3 += 4;
80
81
82 const __m128 vb0123c0 = _mm_load_ps(w + 0);
83 const __m128 vb4567c0 = _mm_load_ps(w + 4);
84
85 vacc0x0123 = _mm_add_ps(vacc0x0123, _mm_mul_ps(va0, vb0123c0));
86 vacc1x0123 = _mm_add_ps(vacc1x0123, _mm_mul_ps(va1, vb0123c0));
87 vacc2x0123 = _mm_add_ps(vacc2x0123, _mm_mul_ps(va2, vb0123c0));
88 vacc3x0123 = _mm_add_ps(vacc3x0123, _mm_mul_ps(va3, vb0123c0));
89 vacc0x4567 = _mm_add_ps(vacc0x4567, _mm_mul_ps(va0, vb4567c0));
90 vacc1x4567 = _mm_add_ps(vacc1x4567, _mm_mul_ps(va1, vb4567c0));
91 vacc2x4567 = _mm_add_ps(vacc2x4567, _mm_mul_ps(va2, vb4567c0));
92 vacc3x4567 = _mm_add_ps(vacc3x4567, _mm_mul_ps(va3, vb4567c0));
93
94 va0 = _mm_shuffle_ps(va0, va0, _MM_SHUFFLE(0, 3, 2, 1));
95 va1 = _mm_shuffle_ps(va1, va1, _MM_SHUFFLE(0, 3, 2, 1));
96 va2 = _mm_shuffle_ps(va2, va2, _MM_SHUFFLE(0, 3, 2, 1));
97 va3 = _mm_shuffle_ps(va3, va3, _MM_SHUFFLE(0, 3, 2, 1));
98
99 const __m128 vb0123c1 = _mm_load_ps(w + 8);
100 const __m128 vb4567c1 = _mm_load_ps(w + 12);
101
102 vacc0x0123 = _mm_add_ps(vacc0x0123, _mm_mul_ps(va0, vb0123c1));
103 vacc1x0123 = _mm_add_ps(vacc1x0123, _mm_mul_ps(va1, vb0123c1));
104 vacc2x0123 = _mm_add_ps(vacc2x0123, _mm_mul_ps(va2, vb0123c1));
105 vacc3x0123 = _mm_add_ps(vacc3x0123, _mm_mul_ps(va3, vb0123c1));
106 vacc0x4567 = _mm_add_ps(vacc0x4567, _mm_mul_ps(va0, vb4567c1));
107 vacc1x4567 = _mm_add_ps(vacc1x4567, _mm_mul_ps(va1, vb4567c1));
108 vacc2x4567 = _mm_add_ps(vacc2x4567, _mm_mul_ps(va2, vb4567c1));
109 vacc3x4567 = _mm_add_ps(vacc3x4567, _mm_mul_ps(va3, vb4567c1));
110
111 va0 = _mm_shuffle_ps(va0, va0, _MM_SHUFFLE(0, 3, 2, 1));
112 va1 = _mm_shuffle_ps(va1, va1, _MM_SHUFFLE(0, 3, 2, 1));
113 va2 = _mm_shuffle_ps(va2, va2, _MM_SHUFFLE(0, 3, 2, 1));
114 va3 = _mm_shuffle_ps(va3, va3, _MM_SHUFFLE(0, 3, 2, 1));
115
116 const __m128 vb0123c2 = _mm_load_ps(w + 16);
117 const __m128 vb4567c2 = _mm_load_ps(w + 20);
118
119 vacc0x0123 = _mm_add_ps(vacc0x0123, _mm_mul_ps(va0, vb0123c2));
120 vacc1x0123 = _mm_add_ps(vacc1x0123, _mm_mul_ps(va1, vb0123c2));
121 vacc2x0123 = _mm_add_ps(vacc2x0123, _mm_mul_ps(va2, vb0123c2));
122 vacc3x0123 = _mm_add_ps(vacc3x0123, _mm_mul_ps(va3, vb0123c2));
123 vacc0x4567 = _mm_add_ps(vacc0x4567, _mm_mul_ps(va0, vb4567c2));
124 vacc1x4567 = _mm_add_ps(vacc1x4567, _mm_mul_ps(va1, vb4567c2));
125 vacc2x4567 = _mm_add_ps(vacc2x4567, _mm_mul_ps(va2, vb4567c2));
126 vacc3x4567 = _mm_add_ps(vacc3x4567, _mm_mul_ps(va3, vb4567c2));
127
128 va0 = _mm_shuffle_ps(va0, va0, _MM_SHUFFLE(0, 3, 2, 1));
129 va1 = _mm_shuffle_ps(va1, va1, _MM_SHUFFLE(0, 3, 2, 1));
130 va2 = _mm_shuffle_ps(va2, va2, _MM_SHUFFLE(0, 3, 2, 1));
131 va3 = _mm_shuffle_ps(va3, va3, _MM_SHUFFLE(0, 3, 2, 1));
132
133 const __m128 vb0123c3 = _mm_load_ps(w + 24);
134 const __m128 vb4567c3 = _mm_load_ps(w + 28);
135
136 vacc0x0123 = _mm_add_ps(vacc0x0123, _mm_mul_ps(va0, vb0123c3));
137 vacc1x0123 = _mm_add_ps(vacc1x0123, _mm_mul_ps(va1, vb0123c3));
138 vacc2x0123 = _mm_add_ps(vacc2x0123, _mm_mul_ps(va2, vb0123c3));
139 vacc3x0123 = _mm_add_ps(vacc3x0123, _mm_mul_ps(va3, vb0123c3));
140 vacc0x4567 = _mm_add_ps(vacc0x4567, _mm_mul_ps(va0, vb4567c3));
141 vacc1x4567 = _mm_add_ps(vacc1x4567, _mm_mul_ps(va1, vb4567c3));
142 vacc2x4567 = _mm_add_ps(vacc2x4567, _mm_mul_ps(va2, vb4567c3));
143 vacc3x4567 = _mm_add_ps(vacc3x4567, _mm_mul_ps(va3, vb4567c3));
144
145
146 w += 32;
147 k -= 4 * sizeof(float);
148 }
149 if XNN_UNLIKELY(k != 0) {
150 __m128 va0 = _mm_loadu_ps(a0);
151 a0 = (const float*) ((uintptr_t) a0 + k);
152 __m128 va1 = _mm_loadu_ps(a1);
153 a1 = (const float*) ((uintptr_t) a1 + k);
154 __m128 va2 = _mm_loadu_ps(a2);
155 a2 = (const float*) ((uintptr_t) a2 + k);
156 __m128 va3 = _mm_loadu_ps(a3);
157 a3 = (const float*) ((uintptr_t) a3 + k);
158
159
160 const __m128 vb0123c0 = _mm_load_ps(w + 0);
161 const __m128 vb4567c0 = _mm_load_ps(w + 4);
162
163 vacc0x0123 = _mm_add_ps(vacc0x0123, _mm_mul_ps(_mm_andnot_ps(_mm_cmpeq_ps(_mm_setzero_ps(), vb0123c0), va0), vb0123c0));
164 vacc1x0123 = _mm_add_ps(vacc1x0123, _mm_mul_ps(_mm_andnot_ps(_mm_cmpeq_ps(_mm_setzero_ps(), vb0123c0), va1), vb0123c0));
165 vacc2x0123 = _mm_add_ps(vacc2x0123, _mm_mul_ps(_mm_andnot_ps(_mm_cmpeq_ps(_mm_setzero_ps(), vb0123c0), va2), vb0123c0));
166 vacc3x0123 = _mm_add_ps(vacc3x0123, _mm_mul_ps(_mm_andnot_ps(_mm_cmpeq_ps(_mm_setzero_ps(), vb0123c0), va3), vb0123c0));
167 vacc0x4567 = _mm_add_ps(vacc0x4567, _mm_mul_ps(_mm_andnot_ps(_mm_cmpeq_ps(_mm_setzero_ps(), vb4567c0), va0), vb4567c0));
168 vacc1x4567 = _mm_add_ps(vacc1x4567, _mm_mul_ps(_mm_andnot_ps(_mm_cmpeq_ps(_mm_setzero_ps(), vb4567c0), va1), vb4567c0));
169 vacc2x4567 = _mm_add_ps(vacc2x4567, _mm_mul_ps(_mm_andnot_ps(_mm_cmpeq_ps(_mm_setzero_ps(), vb4567c0), va2), vb4567c0));
170 vacc3x4567 = _mm_add_ps(vacc3x4567, _mm_mul_ps(_mm_andnot_ps(_mm_cmpeq_ps(_mm_setzero_ps(), vb4567c0), va3), vb4567c0));
171
172 va0 = _mm_shuffle_ps(va0, va0, _MM_SHUFFLE(0, 3, 2, 1));
173 va1 = _mm_shuffle_ps(va1, va1, _MM_SHUFFLE(0, 3, 2, 1));
174 va2 = _mm_shuffle_ps(va2, va2, _MM_SHUFFLE(0, 3, 2, 1));
175 va3 = _mm_shuffle_ps(va3, va3, _MM_SHUFFLE(0, 3, 2, 1));
176
177 const __m128 vb0123c1 = _mm_load_ps(w + 8);
178 const __m128 vb4567c1 = _mm_load_ps(w + 12);
179
180 vacc0x0123 = _mm_add_ps(vacc0x0123, _mm_mul_ps(_mm_andnot_ps(_mm_cmpeq_ps(_mm_setzero_ps(), vb0123c1), va0), vb0123c1));
181 vacc1x0123 = _mm_add_ps(vacc1x0123, _mm_mul_ps(_mm_andnot_ps(_mm_cmpeq_ps(_mm_setzero_ps(), vb0123c1), va1), vb0123c1));
182 vacc2x0123 = _mm_add_ps(vacc2x0123, _mm_mul_ps(_mm_andnot_ps(_mm_cmpeq_ps(_mm_setzero_ps(), vb0123c1), va2), vb0123c1));
183 vacc3x0123 = _mm_add_ps(vacc3x0123, _mm_mul_ps(_mm_andnot_ps(_mm_cmpeq_ps(_mm_setzero_ps(), vb0123c1), va3), vb0123c1));
184 vacc0x4567 = _mm_add_ps(vacc0x4567, _mm_mul_ps(_mm_andnot_ps(_mm_cmpeq_ps(_mm_setzero_ps(), vb4567c1), va0), vb4567c1));
185 vacc1x4567 = _mm_add_ps(vacc1x4567, _mm_mul_ps(_mm_andnot_ps(_mm_cmpeq_ps(_mm_setzero_ps(), vb4567c1), va1), vb4567c1));
186 vacc2x4567 = _mm_add_ps(vacc2x4567, _mm_mul_ps(_mm_andnot_ps(_mm_cmpeq_ps(_mm_setzero_ps(), vb4567c1), va2), vb4567c1));
187 vacc3x4567 = _mm_add_ps(vacc3x4567, _mm_mul_ps(_mm_andnot_ps(_mm_cmpeq_ps(_mm_setzero_ps(), vb4567c1), va3), vb4567c1));
188
189 va0 = _mm_shuffle_ps(va0, va0, _MM_SHUFFLE(0, 3, 2, 1));
190 va1 = _mm_shuffle_ps(va1, va1, _MM_SHUFFLE(0, 3, 2, 1));
191 va2 = _mm_shuffle_ps(va2, va2, _MM_SHUFFLE(0, 3, 2, 1));
192 va3 = _mm_shuffle_ps(va3, va3, _MM_SHUFFLE(0, 3, 2, 1));
193
194 const __m128 vb0123c2 = _mm_load_ps(w + 16);
195 const __m128 vb4567c2 = _mm_load_ps(w + 20);
196
197 vacc0x0123 = _mm_add_ps(vacc0x0123, _mm_mul_ps(_mm_andnot_ps(_mm_cmpeq_ps(_mm_setzero_ps(), vb0123c2), va0), vb0123c2));
198 vacc1x0123 = _mm_add_ps(vacc1x0123, _mm_mul_ps(_mm_andnot_ps(_mm_cmpeq_ps(_mm_setzero_ps(), vb0123c2), va1), vb0123c2));
199 vacc2x0123 = _mm_add_ps(vacc2x0123, _mm_mul_ps(_mm_andnot_ps(_mm_cmpeq_ps(_mm_setzero_ps(), vb0123c2), va2), vb0123c2));
200 vacc3x0123 = _mm_add_ps(vacc3x0123, _mm_mul_ps(_mm_andnot_ps(_mm_cmpeq_ps(_mm_setzero_ps(), vb0123c2), va3), vb0123c2));
201 vacc0x4567 = _mm_add_ps(vacc0x4567, _mm_mul_ps(_mm_andnot_ps(_mm_cmpeq_ps(_mm_setzero_ps(), vb4567c2), va0), vb4567c2));
202 vacc1x4567 = _mm_add_ps(vacc1x4567, _mm_mul_ps(_mm_andnot_ps(_mm_cmpeq_ps(_mm_setzero_ps(), vb4567c2), va1), vb4567c2));
203 vacc2x4567 = _mm_add_ps(vacc2x4567, _mm_mul_ps(_mm_andnot_ps(_mm_cmpeq_ps(_mm_setzero_ps(), vb4567c2), va2), vb4567c2));
204 vacc3x4567 = _mm_add_ps(vacc3x4567, _mm_mul_ps(_mm_andnot_ps(_mm_cmpeq_ps(_mm_setzero_ps(), vb4567c2), va3), vb4567c2));
205
206 va0 = _mm_shuffle_ps(va0, va0, _MM_SHUFFLE(0, 3, 2, 1));
207 va1 = _mm_shuffle_ps(va1, va1, _MM_SHUFFLE(0, 3, 2, 1));
208 va2 = _mm_shuffle_ps(va2, va2, _MM_SHUFFLE(0, 3, 2, 1));
209 va3 = _mm_shuffle_ps(va3, va3, _MM_SHUFFLE(0, 3, 2, 1));
210
211 const __m128 vb0123c3 = _mm_load_ps(w + 24);
212 const __m128 vb4567c3 = _mm_load_ps(w + 28);
213
214 vacc0x0123 = _mm_add_ps(vacc0x0123, _mm_mul_ps(_mm_andnot_ps(_mm_cmpeq_ps(_mm_setzero_ps(), vb0123c3), va0), vb0123c3));
215 vacc1x0123 = _mm_add_ps(vacc1x0123, _mm_mul_ps(_mm_andnot_ps(_mm_cmpeq_ps(_mm_setzero_ps(), vb0123c3), va1), vb0123c3));
216 vacc2x0123 = _mm_add_ps(vacc2x0123, _mm_mul_ps(_mm_andnot_ps(_mm_cmpeq_ps(_mm_setzero_ps(), vb0123c3), va2), vb0123c3));
217 vacc3x0123 = _mm_add_ps(vacc3x0123, _mm_mul_ps(_mm_andnot_ps(_mm_cmpeq_ps(_mm_setzero_ps(), vb0123c3), va3), vb0123c3));
218 vacc0x4567 = _mm_add_ps(vacc0x4567, _mm_mul_ps(_mm_andnot_ps(_mm_cmpeq_ps(_mm_setzero_ps(), vb4567c3), va0), vb4567c3));
219 vacc1x4567 = _mm_add_ps(vacc1x4567, _mm_mul_ps(_mm_andnot_ps(_mm_cmpeq_ps(_mm_setzero_ps(), vb4567c3), va1), vb4567c3));
220 vacc2x4567 = _mm_add_ps(vacc2x4567, _mm_mul_ps(_mm_andnot_ps(_mm_cmpeq_ps(_mm_setzero_ps(), vb4567c3), va2), vb4567c3));
221 vacc3x4567 = _mm_add_ps(vacc3x4567, _mm_mul_ps(_mm_andnot_ps(_mm_cmpeq_ps(_mm_setzero_ps(), vb4567c3), va3), vb4567c3));
222
223
224 w += 32;
225 }
226
227 const __m128 vmax = _mm_load_ps(params->sse.max);
228 vacc0x0123 = _mm_min_ps(vacc0x0123, vmax);
229 vacc1x0123 = _mm_min_ps(vacc1x0123, vmax);
230 vacc2x0123 = _mm_min_ps(vacc2x0123, vmax);
231 vacc3x0123 = _mm_min_ps(vacc3x0123, vmax);
232 vacc0x4567 = _mm_min_ps(vacc0x4567, vmax);
233 vacc1x4567 = _mm_min_ps(vacc1x4567, vmax);
234 vacc2x4567 = _mm_min_ps(vacc2x4567, vmax);
235 vacc3x4567 = _mm_min_ps(vacc3x4567, vmax);
236
237 const __m128 vmin = _mm_load_ps(params->sse.min);
238 vacc0x0123 = _mm_max_ps(vacc0x0123, vmin);
239 vacc1x0123 = _mm_max_ps(vacc1x0123, vmin);
240 vacc2x0123 = _mm_max_ps(vacc2x0123, vmin);
241 vacc3x0123 = _mm_max_ps(vacc3x0123, vmin);
242 vacc0x4567 = _mm_max_ps(vacc0x4567, vmin);
243 vacc1x4567 = _mm_max_ps(vacc1x4567, vmin);
244 vacc2x4567 = _mm_max_ps(vacc2x4567, vmin);
245 vacc3x4567 = _mm_max_ps(vacc3x4567, vmin);
246
247 if XNN_LIKELY(nc >= 8) {
248 _mm_storeu_ps(c3, vacc3x0123);
249 _mm_storeu_ps(c3 + 4, vacc3x4567);
250 c3 = (float*) ((uintptr_t) c3 + cn_stride);
251 _mm_storeu_ps(c2, vacc2x0123);
252 _mm_storeu_ps(c2 + 4, vacc2x4567);
253 c2 = (float*) ((uintptr_t) c2 + cn_stride);
254 _mm_storeu_ps(c1, vacc1x0123);
255 _mm_storeu_ps(c1 + 4, vacc1x4567);
256 c1 = (float*) ((uintptr_t) c1 + cn_stride);
257 _mm_storeu_ps(c0, vacc0x0123);
258 _mm_storeu_ps(c0 + 4, vacc0x4567);
259 c0 = (float*) ((uintptr_t) c0 + cn_stride);
260
261 a3 = (const float*) ((uintptr_t) a3 - kc);
262 a2 = (const float*) ((uintptr_t) a2 - kc);
263 a1 = (const float*) ((uintptr_t) a1 - kc);
264 a0 = (const float*) ((uintptr_t) a0 - kc);
265
266 nc -= 8;
267 } else {
268 if (nc & 4) {
269 _mm_storeu_ps(c3, vacc3x0123);
270 _mm_storeu_ps(c2, vacc2x0123);
271 _mm_storeu_ps(c1, vacc1x0123);
272 _mm_storeu_ps(c0, vacc0x0123);
273
274 vacc3x0123 = vacc3x4567;
275 vacc2x0123 = vacc2x4567;
276 vacc1x0123 = vacc1x4567;
277 vacc0x0123 = vacc0x4567;
278
279 c3 += 4;
280 c2 += 4;
281 c1 += 4;
282 c0 += 4;
283 }
284 if (nc & 2) {
285 _mm_storel_pi((__m64*) c3, vacc3x0123);
286 _mm_storel_pi((__m64*) c2, vacc2x0123);
287 _mm_storel_pi((__m64*) c1, vacc1x0123);
288 _mm_storel_pi((__m64*) c0, vacc0x0123);
289
290 vacc3x0123 = _mm_movehl_ps(vacc3x0123, vacc3x0123);
291 vacc2x0123 = _mm_movehl_ps(vacc2x0123, vacc2x0123);
292 vacc1x0123 = _mm_movehl_ps(vacc1x0123, vacc1x0123);
293 vacc0x0123 = _mm_movehl_ps(vacc0x0123, vacc0x0123);
294
295 c3 += 2;
296 c2 += 2;
297 c1 += 2;
298 c0 += 2;
299 }
300 if (nc & 1) {
301 _mm_store_ss(c3, vacc3x0123);
302 _mm_store_ss(c2, vacc2x0123);
303 _mm_store_ss(c1, vacc1x0123);
304 _mm_store_ss(c0, vacc0x0123);
305 }
306
307 nc = 0;
308 }
309 } while (nc != 0);
310 }
311