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_5x8s4__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_5x8s4__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)])
29 {
30 assert(mr != 0);
31 assert(mr <= 5);
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 const float* a4 = (const float*) ((uintptr_t) a3 + a_stride);
61 float* c4 = (float*) ((uintptr_t) c3 + cm_stride);
62 if XNN_UNPREDICTABLE(mr <= 4) {
63 a4 = a3;
64 c4 = c3;
65 }
66
67 do {
68 __m128 vacc0x0123 = _mm_load_ps(acc + 0);
69 __m128 vacc0x4567 = _mm_load_ps(acc + 4);
70 __m128 vacc1x0123 = _mm_load_ps(acc + 8);
71 __m128 vacc1x4567 = _mm_load_ps(acc + 12);
72 __m128 vacc2x0123 = _mm_load_ps(acc + 16);
73 __m128 vacc2x4567 = _mm_load_ps(acc + 20);
74 __m128 vacc3x0123 = _mm_load_ps(acc + 24);
75 __m128 vacc3x4567 = _mm_load_ps(acc + 28);
76 __m128 vacc4x0123 = _mm_load_ps(acc + 32);
77 __m128 vacc4x4567 = _mm_load_ps(acc + 36);
78 acc += 40;
79
80 size_t k = kc;
81 while (k >= 4 * sizeof(float)) {
82 __m128 va0 = _mm_loadu_ps(a0);
83 a0 += 4;
84 __m128 va1 = _mm_loadu_ps(a1);
85 a1 += 4;
86 __m128 va2 = _mm_loadu_ps(a2);
87 a2 += 4;
88 __m128 va3 = _mm_loadu_ps(a3);
89 a3 += 4;
90 __m128 va4 = _mm_loadu_ps(a4);
91 a4 += 4;
92
93
94 const __m128 vb0123c0 = _mm_load_ps(w + 0);
95 const __m128 vb4567c0 = _mm_load_ps(w + 4);
96
97 vacc0x0123 = _mm_add_ps(vacc0x0123, _mm_mul_ps(va0, vb0123c0));
98 vacc1x0123 = _mm_add_ps(vacc1x0123, _mm_mul_ps(va1, vb0123c0));
99 vacc2x0123 = _mm_add_ps(vacc2x0123, _mm_mul_ps(va2, vb0123c0));
100 vacc3x0123 = _mm_add_ps(vacc3x0123, _mm_mul_ps(va3, vb0123c0));
101 vacc4x0123 = _mm_add_ps(vacc4x0123, _mm_mul_ps(va4, vb0123c0));
102 vacc0x4567 = _mm_add_ps(vacc0x4567, _mm_mul_ps(va0, vb4567c0));
103 vacc1x4567 = _mm_add_ps(vacc1x4567, _mm_mul_ps(va1, vb4567c0));
104 vacc2x4567 = _mm_add_ps(vacc2x4567, _mm_mul_ps(va2, vb4567c0));
105 vacc3x4567 = _mm_add_ps(vacc3x4567, _mm_mul_ps(va3, vb4567c0));
106 vacc4x4567 = _mm_add_ps(vacc4x4567, _mm_mul_ps(va4, vb4567c0));
107
108 va0 = _mm_shuffle_ps(va0, va0, _MM_SHUFFLE(0, 3, 2, 1));
109 va1 = _mm_shuffle_ps(va1, va1, _MM_SHUFFLE(0, 3, 2, 1));
110 va2 = _mm_shuffle_ps(va2, va2, _MM_SHUFFLE(0, 3, 2, 1));
111 va3 = _mm_shuffle_ps(va3, va3, _MM_SHUFFLE(0, 3, 2, 1));
112 va4 = _mm_shuffle_ps(va4, va4, _MM_SHUFFLE(0, 3, 2, 1));
113
114 const __m128 vb0123c1 = _mm_load_ps(w + 8);
115 const __m128 vb4567c1 = _mm_load_ps(w + 12);
116
117 vacc0x0123 = _mm_add_ps(vacc0x0123, _mm_mul_ps(va0, vb0123c1));
118 vacc1x0123 = _mm_add_ps(vacc1x0123, _mm_mul_ps(va1, vb0123c1));
119 vacc2x0123 = _mm_add_ps(vacc2x0123, _mm_mul_ps(va2, vb0123c1));
120 vacc3x0123 = _mm_add_ps(vacc3x0123, _mm_mul_ps(va3, vb0123c1));
121 vacc4x0123 = _mm_add_ps(vacc4x0123, _mm_mul_ps(va4, vb0123c1));
122 vacc0x4567 = _mm_add_ps(vacc0x4567, _mm_mul_ps(va0, vb4567c1));
123 vacc1x4567 = _mm_add_ps(vacc1x4567, _mm_mul_ps(va1, vb4567c1));
124 vacc2x4567 = _mm_add_ps(vacc2x4567, _mm_mul_ps(va2, vb4567c1));
125 vacc3x4567 = _mm_add_ps(vacc3x4567, _mm_mul_ps(va3, vb4567c1));
126 vacc4x4567 = _mm_add_ps(vacc4x4567, _mm_mul_ps(va4, vb4567c1));
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 va4 = _mm_shuffle_ps(va4, va4, _MM_SHUFFLE(0, 3, 2, 1));
133
134 const __m128 vb0123c2 = _mm_load_ps(w + 16);
135 const __m128 vb4567c2 = _mm_load_ps(w + 20);
136
137 vacc0x0123 = _mm_add_ps(vacc0x0123, _mm_mul_ps(va0, vb0123c2));
138 vacc1x0123 = _mm_add_ps(vacc1x0123, _mm_mul_ps(va1, vb0123c2));
139 vacc2x0123 = _mm_add_ps(vacc2x0123, _mm_mul_ps(va2, vb0123c2));
140 vacc3x0123 = _mm_add_ps(vacc3x0123, _mm_mul_ps(va3, vb0123c2));
141 vacc4x0123 = _mm_add_ps(vacc4x0123, _mm_mul_ps(va4, vb0123c2));
142 vacc0x4567 = _mm_add_ps(vacc0x4567, _mm_mul_ps(va0, vb4567c2));
143 vacc1x4567 = _mm_add_ps(vacc1x4567, _mm_mul_ps(va1, vb4567c2));
144 vacc2x4567 = _mm_add_ps(vacc2x4567, _mm_mul_ps(va2, vb4567c2));
145 vacc3x4567 = _mm_add_ps(vacc3x4567, _mm_mul_ps(va3, vb4567c2));
146 vacc4x4567 = _mm_add_ps(vacc4x4567, _mm_mul_ps(va4, vb4567c2));
147
148 va0 = _mm_shuffle_ps(va0, va0, _MM_SHUFFLE(0, 3, 2, 1));
149 va1 = _mm_shuffle_ps(va1, va1, _MM_SHUFFLE(0, 3, 2, 1));
150 va2 = _mm_shuffle_ps(va2, va2, _MM_SHUFFLE(0, 3, 2, 1));
151 va3 = _mm_shuffle_ps(va3, va3, _MM_SHUFFLE(0, 3, 2, 1));
152 va4 = _mm_shuffle_ps(va4, va4, _MM_SHUFFLE(0, 3, 2, 1));
153
154 const __m128 vb0123c3 = _mm_load_ps(w + 24);
155 const __m128 vb4567c3 = _mm_load_ps(w + 28);
156
157 vacc0x0123 = _mm_add_ps(vacc0x0123, _mm_mul_ps(va0, vb0123c3));
158 vacc1x0123 = _mm_add_ps(vacc1x0123, _mm_mul_ps(va1, vb0123c3));
159 vacc2x0123 = _mm_add_ps(vacc2x0123, _mm_mul_ps(va2, vb0123c3));
160 vacc3x0123 = _mm_add_ps(vacc3x0123, _mm_mul_ps(va3, vb0123c3));
161 vacc4x0123 = _mm_add_ps(vacc4x0123, _mm_mul_ps(va4, 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 vacc4x4567 = _mm_add_ps(vacc4x4567, _mm_mul_ps(va4, vb4567c3));
167
168
169 w += 32;
170 k -= 4 * sizeof(float);
171 }
172 if XNN_UNLIKELY(k != 0) {
173 do {
174 const __m128 va0 = _mm_load1_ps(a0);
175 a0 += 1;
176 const __m128 va1 = _mm_load1_ps(a1);
177 a1 += 1;
178 const __m128 va2 = _mm_load1_ps(a2);
179 a2 += 1;
180 const __m128 va3 = _mm_load1_ps(a3);
181 a3 += 1;
182 const __m128 va4 = _mm_load1_ps(a4);
183 a4 += 1;
184
185 const __m128 vb0123 = _mm_load_ps(w);
186 const __m128 vb4567 = _mm_load_ps(w + 4);
187 w += 8;
188
189 vacc0x0123 = _mm_add_ps(vacc0x0123, _mm_mul_ps(va0, vb0123));
190 vacc1x0123 = _mm_add_ps(vacc1x0123, _mm_mul_ps(va1, vb0123));
191 vacc2x0123 = _mm_add_ps(vacc2x0123, _mm_mul_ps(va2, vb0123));
192 vacc3x0123 = _mm_add_ps(vacc3x0123, _mm_mul_ps(va3, vb0123));
193 vacc4x0123 = _mm_add_ps(vacc4x0123, _mm_mul_ps(va4, vb0123));
194 vacc0x4567 = _mm_add_ps(vacc0x4567, _mm_mul_ps(va0, vb4567));
195 vacc1x4567 = _mm_add_ps(vacc1x4567, _mm_mul_ps(va1, vb4567));
196 vacc2x4567 = _mm_add_ps(vacc2x4567, _mm_mul_ps(va2, vb4567));
197 vacc3x4567 = _mm_add_ps(vacc3x4567, _mm_mul_ps(va3, vb4567));
198 vacc4x4567 = _mm_add_ps(vacc4x4567, _mm_mul_ps(va4, vb4567));
199
200 k -= sizeof(float);
201 } while (k != 0);
202 }
203
204 const __m128 vmax = _mm_load_ps(params->sse.max);
205 vacc0x0123 = _mm_min_ps(vacc0x0123, vmax);
206 vacc1x0123 = _mm_min_ps(vacc1x0123, vmax);
207 vacc2x0123 = _mm_min_ps(vacc2x0123, vmax);
208 vacc3x0123 = _mm_min_ps(vacc3x0123, vmax);
209 vacc4x0123 = _mm_min_ps(vacc4x0123, vmax);
210 vacc0x4567 = _mm_min_ps(vacc0x4567, vmax);
211 vacc1x4567 = _mm_min_ps(vacc1x4567, vmax);
212 vacc2x4567 = _mm_min_ps(vacc2x4567, vmax);
213 vacc3x4567 = _mm_min_ps(vacc3x4567, vmax);
214 vacc4x4567 = _mm_min_ps(vacc4x4567, vmax);
215
216 const __m128 vmin = _mm_load_ps(params->sse.min);
217 vacc0x0123 = _mm_max_ps(vacc0x0123, vmin);
218 vacc1x0123 = _mm_max_ps(vacc1x0123, vmin);
219 vacc2x0123 = _mm_max_ps(vacc2x0123, vmin);
220 vacc3x0123 = _mm_max_ps(vacc3x0123, vmin);
221 vacc4x0123 = _mm_max_ps(vacc4x0123, vmin);
222 vacc0x4567 = _mm_max_ps(vacc0x4567, vmin);
223 vacc1x4567 = _mm_max_ps(vacc1x4567, vmin);
224 vacc2x4567 = _mm_max_ps(vacc2x4567, vmin);
225 vacc3x4567 = _mm_max_ps(vacc3x4567, vmin);
226 vacc4x4567 = _mm_max_ps(vacc4x4567, vmin);
227
228 if XNN_LIKELY(nc >= 8) {
229 _mm_storeu_ps(c4, vacc4x0123);
230 _mm_storeu_ps(c4 + 4, vacc4x4567);
231 c4 = (float*) ((uintptr_t) c4 + cn_stride);
232 _mm_storeu_ps(c3, vacc3x0123);
233 _mm_storeu_ps(c3 + 4, vacc3x4567);
234 c3 = (float*) ((uintptr_t) c3 + cn_stride);
235 _mm_storeu_ps(c2, vacc2x0123);
236 _mm_storeu_ps(c2 + 4, vacc2x4567);
237 c2 = (float*) ((uintptr_t) c2 + cn_stride);
238 _mm_storeu_ps(c1, vacc1x0123);
239 _mm_storeu_ps(c1 + 4, vacc1x4567);
240 c1 = (float*) ((uintptr_t) c1 + cn_stride);
241 _mm_storeu_ps(c0, vacc0x0123);
242 _mm_storeu_ps(c0 + 4, vacc0x4567);
243 c0 = (float*) ((uintptr_t) c0 + cn_stride);
244
245 a4 = (const float*) ((uintptr_t) a4 - kc);
246 a3 = (const float*) ((uintptr_t) a3 - kc);
247 a2 = (const float*) ((uintptr_t) a2 - kc);
248 a1 = (const float*) ((uintptr_t) a1 - kc);
249 a0 = (const float*) ((uintptr_t) a0 - kc);
250
251 nc -= 8;
252 } else {
253 if (nc & 4) {
254 _mm_storeu_ps(c4, vacc4x0123);
255 _mm_storeu_ps(c3, vacc3x0123);
256 _mm_storeu_ps(c2, vacc2x0123);
257 _mm_storeu_ps(c1, vacc1x0123);
258 _mm_storeu_ps(c0, vacc0x0123);
259
260 vacc4x0123 = vacc4x4567;
261 vacc3x0123 = vacc3x4567;
262 vacc2x0123 = vacc2x4567;
263 vacc1x0123 = vacc1x4567;
264 vacc0x0123 = vacc0x4567;
265
266 c4 += 4;
267 c3 += 4;
268 c2 += 4;
269 c1 += 4;
270 c0 += 4;
271 }
272 if (nc & 2) {
273 _mm_storel_pi((__m64*) c4, vacc4x0123);
274 _mm_storel_pi((__m64*) c3, vacc3x0123);
275 _mm_storel_pi((__m64*) c2, vacc2x0123);
276 _mm_storel_pi((__m64*) c1, vacc1x0123);
277 _mm_storel_pi((__m64*) c0, vacc0x0123);
278
279 vacc4x0123 = _mm_movehl_ps(vacc4x0123, vacc4x0123);
280 vacc3x0123 = _mm_movehl_ps(vacc3x0123, vacc3x0123);
281 vacc2x0123 = _mm_movehl_ps(vacc2x0123, vacc2x0123);
282 vacc1x0123 = _mm_movehl_ps(vacc1x0123, vacc1x0123);
283 vacc0x0123 = _mm_movehl_ps(vacc0x0123, vacc0x0123);
284
285 c4 += 2;
286 c3 += 2;
287 c2 += 2;
288 c1 += 2;
289 c0 += 2;
290 }
291 if (nc & 1) {
292 _mm_store_ss(c4, vacc4x0123);
293 _mm_store_ss(c3, vacc3x0123);
294 _mm_store_ss(c2, vacc2x0123);
295 _mm_store_ss(c1, vacc1x0123);
296 _mm_store_ss(c0, vacc0x0123);
297 }
298
299 nc = 0;
300 }
301 } while (nc != 0);
302 }
303