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_3x8s4__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_3x8s4__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 <= 3);
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
55 do {
56 __m128 vacc0x0123 = _mm_load_ps(acc + 0);
57 __m128 vacc0x4567 = _mm_load_ps(acc + 4);
58 __m128 vacc1x0123 = _mm_load_ps(acc + 8);
59 __m128 vacc1x4567 = _mm_load_ps(acc + 12);
60 __m128 vacc2x0123 = _mm_load_ps(acc + 16);
61 __m128 vacc2x4567 = _mm_load_ps(acc + 20);
62 acc += 24;
63
64 size_t k = kc;
65 while (k >= 4 * sizeof(float)) {
66 __m128 va0 = _mm_loadu_ps(a0);
67 a0 += 4;
68 __m128 va1 = _mm_loadu_ps(a1);
69 a1 += 4;
70 __m128 va2 = _mm_loadu_ps(a2);
71 a2 += 4;
72
73
74 const __m128 vb0123c0 = _mm_load_ps(w + 0);
75 const __m128 vb4567c0 = _mm_load_ps(w + 4);
76
77 vacc0x0123 = _mm_add_ps(vacc0x0123, _mm_mul_ps(va0, vb0123c0));
78 vacc1x0123 = _mm_add_ps(vacc1x0123, _mm_mul_ps(va1, vb0123c0));
79 vacc2x0123 = _mm_add_ps(vacc2x0123, _mm_mul_ps(va2, vb0123c0));
80 vacc0x4567 = _mm_add_ps(vacc0x4567, _mm_mul_ps(va0, vb4567c0));
81 vacc1x4567 = _mm_add_ps(vacc1x4567, _mm_mul_ps(va1, vb4567c0));
82 vacc2x4567 = _mm_add_ps(vacc2x4567, _mm_mul_ps(va2, vb4567c0));
83
84 va0 = _mm_shuffle_ps(va0, va0, _MM_SHUFFLE(0, 3, 2, 1));
85 va1 = _mm_shuffle_ps(va1, va1, _MM_SHUFFLE(0, 3, 2, 1));
86 va2 = _mm_shuffle_ps(va2, va2, _MM_SHUFFLE(0, 3, 2, 1));
87
88 const __m128 vb0123c1 = _mm_load_ps(w + 8);
89 const __m128 vb4567c1 = _mm_load_ps(w + 12);
90
91 vacc0x0123 = _mm_add_ps(vacc0x0123, _mm_mul_ps(va0, vb0123c1));
92 vacc1x0123 = _mm_add_ps(vacc1x0123, _mm_mul_ps(va1, vb0123c1));
93 vacc2x0123 = _mm_add_ps(vacc2x0123, _mm_mul_ps(va2, vb0123c1));
94 vacc0x4567 = _mm_add_ps(vacc0x4567, _mm_mul_ps(va0, vb4567c1));
95 vacc1x4567 = _mm_add_ps(vacc1x4567, _mm_mul_ps(va1, vb4567c1));
96 vacc2x4567 = _mm_add_ps(vacc2x4567, _mm_mul_ps(va2, vb4567c1));
97
98 va0 = _mm_shuffle_ps(va0, va0, _MM_SHUFFLE(0, 3, 2, 1));
99 va1 = _mm_shuffle_ps(va1, va1, _MM_SHUFFLE(0, 3, 2, 1));
100 va2 = _mm_shuffle_ps(va2, va2, _MM_SHUFFLE(0, 3, 2, 1));
101
102 const __m128 vb0123c2 = _mm_load_ps(w + 16);
103 const __m128 vb4567c2 = _mm_load_ps(w + 20);
104
105 vacc0x0123 = _mm_add_ps(vacc0x0123, _mm_mul_ps(va0, vb0123c2));
106 vacc1x0123 = _mm_add_ps(vacc1x0123, _mm_mul_ps(va1, vb0123c2));
107 vacc2x0123 = _mm_add_ps(vacc2x0123, _mm_mul_ps(va2, vb0123c2));
108 vacc0x4567 = _mm_add_ps(vacc0x4567, _mm_mul_ps(va0, vb4567c2));
109 vacc1x4567 = _mm_add_ps(vacc1x4567, _mm_mul_ps(va1, vb4567c2));
110 vacc2x4567 = _mm_add_ps(vacc2x4567, _mm_mul_ps(va2, vb4567c2));
111
112 va0 = _mm_shuffle_ps(va0, va0, _MM_SHUFFLE(0, 3, 2, 1));
113 va1 = _mm_shuffle_ps(va1, va1, _MM_SHUFFLE(0, 3, 2, 1));
114 va2 = _mm_shuffle_ps(va2, va2, _MM_SHUFFLE(0, 3, 2, 1));
115
116 const __m128 vb0123c3 = _mm_load_ps(w + 24);
117 const __m128 vb4567c3 = _mm_load_ps(w + 28);
118
119 vacc0x0123 = _mm_add_ps(vacc0x0123, _mm_mul_ps(va0, vb0123c3));
120 vacc1x0123 = _mm_add_ps(vacc1x0123, _mm_mul_ps(va1, vb0123c3));
121 vacc2x0123 = _mm_add_ps(vacc2x0123, _mm_mul_ps(va2, vb0123c3));
122 vacc0x4567 = _mm_add_ps(vacc0x4567, _mm_mul_ps(va0, vb4567c3));
123 vacc1x4567 = _mm_add_ps(vacc1x4567, _mm_mul_ps(va1, vb4567c3));
124 vacc2x4567 = _mm_add_ps(vacc2x4567, _mm_mul_ps(va2, vb4567c3));
125
126
127 w += 32;
128 k -= 4 * sizeof(float);
129 }
130 if XNN_UNLIKELY(k != 0) {
131 __m128 va0 = _mm_loadu_ps(a0);
132 a0 = (const float*) ((uintptr_t) a0 + k);
133 __m128 va1 = _mm_loadu_ps(a1);
134 a1 = (const float*) ((uintptr_t) a1 + k);
135 __m128 va2 = _mm_loadu_ps(a2);
136 a2 = (const float*) ((uintptr_t) a2 + k);
137
138
139 const __m128 vb0123c0 = _mm_load_ps(w + 0);
140 const __m128 vb4567c0 = _mm_load_ps(w + 4);
141
142 vacc0x0123 = _mm_add_ps(vacc0x0123, _mm_mul_ps(_mm_andnot_ps(_mm_cmpeq_ps(_mm_setzero_ps(), vb0123c0), va0), vb0123c0));
143 vacc1x0123 = _mm_add_ps(vacc1x0123, _mm_mul_ps(_mm_andnot_ps(_mm_cmpeq_ps(_mm_setzero_ps(), vb0123c0), va1), vb0123c0));
144 vacc2x0123 = _mm_add_ps(vacc2x0123, _mm_mul_ps(_mm_andnot_ps(_mm_cmpeq_ps(_mm_setzero_ps(), vb0123c0), va2), vb0123c0));
145 vacc0x4567 = _mm_add_ps(vacc0x4567, _mm_mul_ps(_mm_andnot_ps(_mm_cmpeq_ps(_mm_setzero_ps(), vb4567c0), va0), vb4567c0));
146 vacc1x4567 = _mm_add_ps(vacc1x4567, _mm_mul_ps(_mm_andnot_ps(_mm_cmpeq_ps(_mm_setzero_ps(), vb4567c0), va1), vb4567c0));
147 vacc2x4567 = _mm_add_ps(vacc2x4567, _mm_mul_ps(_mm_andnot_ps(_mm_cmpeq_ps(_mm_setzero_ps(), vb4567c0), va2), vb4567c0));
148
149 va0 = _mm_shuffle_ps(va0, va0, _MM_SHUFFLE(0, 3, 2, 1));
150 va1 = _mm_shuffle_ps(va1, va1, _MM_SHUFFLE(0, 3, 2, 1));
151 va2 = _mm_shuffle_ps(va2, va2, _MM_SHUFFLE(0, 3, 2, 1));
152
153 const __m128 vb0123c1 = _mm_load_ps(w + 8);
154 const __m128 vb4567c1 = _mm_load_ps(w + 12);
155
156 vacc0x0123 = _mm_add_ps(vacc0x0123, _mm_mul_ps(_mm_andnot_ps(_mm_cmpeq_ps(_mm_setzero_ps(), vb0123c1), va0), vb0123c1));
157 vacc1x0123 = _mm_add_ps(vacc1x0123, _mm_mul_ps(_mm_andnot_ps(_mm_cmpeq_ps(_mm_setzero_ps(), vb0123c1), va1), vb0123c1));
158 vacc2x0123 = _mm_add_ps(vacc2x0123, _mm_mul_ps(_mm_andnot_ps(_mm_cmpeq_ps(_mm_setzero_ps(), vb0123c1), va2), vb0123c1));
159 vacc0x4567 = _mm_add_ps(vacc0x4567, _mm_mul_ps(_mm_andnot_ps(_mm_cmpeq_ps(_mm_setzero_ps(), vb4567c1), va0), vb4567c1));
160 vacc1x4567 = _mm_add_ps(vacc1x4567, _mm_mul_ps(_mm_andnot_ps(_mm_cmpeq_ps(_mm_setzero_ps(), vb4567c1), va1), vb4567c1));
161 vacc2x4567 = _mm_add_ps(vacc2x4567, _mm_mul_ps(_mm_andnot_ps(_mm_cmpeq_ps(_mm_setzero_ps(), vb4567c1), va2), vb4567c1));
162
163 va0 = _mm_shuffle_ps(va0, va0, _MM_SHUFFLE(0, 3, 2, 1));
164 va1 = _mm_shuffle_ps(va1, va1, _MM_SHUFFLE(0, 3, 2, 1));
165 va2 = _mm_shuffle_ps(va2, va2, _MM_SHUFFLE(0, 3, 2, 1));
166
167 const __m128 vb0123c2 = _mm_load_ps(w + 16);
168 const __m128 vb4567c2 = _mm_load_ps(w + 20);
169
170 vacc0x0123 = _mm_add_ps(vacc0x0123, _mm_mul_ps(_mm_andnot_ps(_mm_cmpeq_ps(_mm_setzero_ps(), vb0123c2), va0), vb0123c2));
171 vacc1x0123 = _mm_add_ps(vacc1x0123, _mm_mul_ps(_mm_andnot_ps(_mm_cmpeq_ps(_mm_setzero_ps(), vb0123c2), va1), vb0123c2));
172 vacc2x0123 = _mm_add_ps(vacc2x0123, _mm_mul_ps(_mm_andnot_ps(_mm_cmpeq_ps(_mm_setzero_ps(), vb0123c2), va2), vb0123c2));
173 vacc0x4567 = _mm_add_ps(vacc0x4567, _mm_mul_ps(_mm_andnot_ps(_mm_cmpeq_ps(_mm_setzero_ps(), vb4567c2), va0), vb4567c2));
174 vacc1x4567 = _mm_add_ps(vacc1x4567, _mm_mul_ps(_mm_andnot_ps(_mm_cmpeq_ps(_mm_setzero_ps(), vb4567c2), va1), vb4567c2));
175 vacc2x4567 = _mm_add_ps(vacc2x4567, _mm_mul_ps(_mm_andnot_ps(_mm_cmpeq_ps(_mm_setzero_ps(), vb4567c2), va2), vb4567c2));
176
177 va0 = _mm_shuffle_ps(va0, va0, _MM_SHUFFLE(0, 3, 2, 1));
178 va1 = _mm_shuffle_ps(va1, va1, _MM_SHUFFLE(0, 3, 2, 1));
179 va2 = _mm_shuffle_ps(va2, va2, _MM_SHUFFLE(0, 3, 2, 1));
180
181 const __m128 vb0123c3 = _mm_load_ps(w + 24);
182 const __m128 vb4567c3 = _mm_load_ps(w + 28);
183
184 vacc0x0123 = _mm_add_ps(vacc0x0123, _mm_mul_ps(_mm_andnot_ps(_mm_cmpeq_ps(_mm_setzero_ps(), vb0123c3), va0), vb0123c3));
185 vacc1x0123 = _mm_add_ps(vacc1x0123, _mm_mul_ps(_mm_andnot_ps(_mm_cmpeq_ps(_mm_setzero_ps(), vb0123c3), va1), vb0123c3));
186 vacc2x0123 = _mm_add_ps(vacc2x0123, _mm_mul_ps(_mm_andnot_ps(_mm_cmpeq_ps(_mm_setzero_ps(), vb0123c3), va2), vb0123c3));
187 vacc0x4567 = _mm_add_ps(vacc0x4567, _mm_mul_ps(_mm_andnot_ps(_mm_cmpeq_ps(_mm_setzero_ps(), vb4567c3), va0), vb4567c3));
188 vacc1x4567 = _mm_add_ps(vacc1x4567, _mm_mul_ps(_mm_andnot_ps(_mm_cmpeq_ps(_mm_setzero_ps(), vb4567c3), va1), vb4567c3));
189 vacc2x4567 = _mm_add_ps(vacc2x4567, _mm_mul_ps(_mm_andnot_ps(_mm_cmpeq_ps(_mm_setzero_ps(), vb4567c3), va2), vb4567c3));
190
191
192 w += 32;
193 }
194
195 const __m128 vmax = _mm_load_ps(params->sse.max);
196 vacc0x0123 = _mm_min_ps(vacc0x0123, vmax);
197 vacc1x0123 = _mm_min_ps(vacc1x0123, vmax);
198 vacc2x0123 = _mm_min_ps(vacc2x0123, vmax);
199 vacc0x4567 = _mm_min_ps(vacc0x4567, vmax);
200 vacc1x4567 = _mm_min_ps(vacc1x4567, vmax);
201 vacc2x4567 = _mm_min_ps(vacc2x4567, vmax);
202
203 const __m128 vmin = _mm_load_ps(params->sse.min);
204 vacc0x0123 = _mm_max_ps(vacc0x0123, vmin);
205 vacc1x0123 = _mm_max_ps(vacc1x0123, vmin);
206 vacc2x0123 = _mm_max_ps(vacc2x0123, vmin);
207 vacc0x4567 = _mm_max_ps(vacc0x4567, vmin);
208 vacc1x4567 = _mm_max_ps(vacc1x4567, vmin);
209 vacc2x4567 = _mm_max_ps(vacc2x4567, vmin);
210
211 if XNN_LIKELY(nc >= 8) {
212 _mm_storeu_ps(c2, vacc2x0123);
213 _mm_storeu_ps(c2 + 4, vacc2x4567);
214 c2 = (float*) ((uintptr_t) c2 + cn_stride);
215 _mm_storeu_ps(c1, vacc1x0123);
216 _mm_storeu_ps(c1 + 4, vacc1x4567);
217 c1 = (float*) ((uintptr_t) c1 + cn_stride);
218 _mm_storeu_ps(c0, vacc0x0123);
219 _mm_storeu_ps(c0 + 4, vacc0x4567);
220 c0 = (float*) ((uintptr_t) c0 + cn_stride);
221
222 a2 = (const float*) ((uintptr_t) a2 - kc);
223 a1 = (const float*) ((uintptr_t) a1 - kc);
224 a0 = (const float*) ((uintptr_t) a0 - kc);
225
226 nc -= 8;
227 } else {
228 if (nc & 4) {
229 _mm_storeu_ps(c2, vacc2x0123);
230 _mm_storeu_ps(c1, vacc1x0123);
231 _mm_storeu_ps(c0, vacc0x0123);
232
233 vacc2x0123 = vacc2x4567;
234 vacc1x0123 = vacc1x4567;
235 vacc0x0123 = vacc0x4567;
236
237 c2 += 4;
238 c1 += 4;
239 c0 += 4;
240 }
241 if (nc & 2) {
242 _mm_storel_pi((__m64*) c2, vacc2x0123);
243 _mm_storel_pi((__m64*) c1, vacc1x0123);
244 _mm_storel_pi((__m64*) c0, vacc0x0123);
245
246 vacc2x0123 = _mm_movehl_ps(vacc2x0123, vacc2x0123);
247 vacc1x0123 = _mm_movehl_ps(vacc1x0123, vacc1x0123);
248 vacc0x0123 = _mm_movehl_ps(vacc0x0123, vacc0x0123);
249
250 c2 += 2;
251 c1 += 2;
252 c0 += 2;
253 }
254 if (nc & 1) {
255 _mm_store_ss(c2, vacc2x0123);
256 _mm_store_ss(c1, vacc1x0123);
257 _mm_store_ss(c0, vacc0x0123);
258 }
259
260 nc = 0;
261 }
262 } while (nc != 0);
263 }
264