1 // Auto-generated file. Do not edit!
2 // Template: src/f32-gemm/sse-dup.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_4x8__sse_dup(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_4x8__sse_dup(
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)])
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 const __m128 va0 = _mm_loadu_ps(a0);
73 a0 += 4;
74 const __m128 va1 = _mm_loadu_ps(a1);
75 a1 += 4;
76 const __m128 va2 = _mm_loadu_ps(a2);
77 a2 += 4;
78 const __m128 va3 = _mm_loadu_ps(a3);
79 a3 += 4;
80
81
82 const __m128 va0c0000 = _mm_shuffle_ps(va0, va0, _MM_SHUFFLE(0, 0, 0, 0));
83 const __m128 va1c0000 = _mm_shuffle_ps(va1, va1, _MM_SHUFFLE(0, 0, 0, 0));
84 const __m128 va2c0000 = _mm_shuffle_ps(va2, va2, _MM_SHUFFLE(0, 0, 0, 0));
85 const __m128 va3c0000 = _mm_shuffle_ps(va3, va3, _MM_SHUFFLE(0, 0, 0, 0));
86
87 const __m128 vb0123c0 = _mm_load_ps(w + 0);
88 const __m128 vb4567c0 = _mm_load_ps(w + 4);
89
90 vacc0x0123 = _mm_add_ps(vacc0x0123, _mm_mul_ps(va0c0000, vb0123c0));
91 vacc1x0123 = _mm_add_ps(vacc1x0123, _mm_mul_ps(va1c0000, vb0123c0));
92 vacc2x0123 = _mm_add_ps(vacc2x0123, _mm_mul_ps(va2c0000, vb0123c0));
93 vacc3x0123 = _mm_add_ps(vacc3x0123, _mm_mul_ps(va3c0000, vb0123c0));
94 vacc0x4567 = _mm_add_ps(vacc0x4567, _mm_mul_ps(va0c0000, vb4567c0));
95 vacc1x4567 = _mm_add_ps(vacc1x4567, _mm_mul_ps(va1c0000, vb4567c0));
96 vacc2x4567 = _mm_add_ps(vacc2x4567, _mm_mul_ps(va2c0000, vb4567c0));
97 vacc3x4567 = _mm_add_ps(vacc3x4567, _mm_mul_ps(va3c0000, vb4567c0));
98
99 const __m128 va0c1111 = _mm_shuffle_ps(va0, va0, _MM_SHUFFLE(1, 1, 1, 1));
100 const __m128 va1c1111 = _mm_shuffle_ps(va1, va1, _MM_SHUFFLE(1, 1, 1, 1));
101 const __m128 va2c1111 = _mm_shuffle_ps(va2, va2, _MM_SHUFFLE(1, 1, 1, 1));
102 const __m128 va3c1111 = _mm_shuffle_ps(va3, va3, _MM_SHUFFLE(1, 1, 1, 1));
103
104 const __m128 vb0123c1 = _mm_load_ps(w + 8);
105 const __m128 vb4567c1 = _mm_load_ps(w + 12);
106
107 vacc0x0123 = _mm_add_ps(vacc0x0123, _mm_mul_ps(va0c1111, vb0123c1));
108 vacc1x0123 = _mm_add_ps(vacc1x0123, _mm_mul_ps(va1c1111, vb0123c1));
109 vacc2x0123 = _mm_add_ps(vacc2x0123, _mm_mul_ps(va2c1111, vb0123c1));
110 vacc3x0123 = _mm_add_ps(vacc3x0123, _mm_mul_ps(va3c1111, vb0123c1));
111 vacc0x4567 = _mm_add_ps(vacc0x4567, _mm_mul_ps(va0c1111, vb4567c1));
112 vacc1x4567 = _mm_add_ps(vacc1x4567, _mm_mul_ps(va1c1111, vb4567c1));
113 vacc2x4567 = _mm_add_ps(vacc2x4567, _mm_mul_ps(va2c1111, vb4567c1));
114 vacc3x4567 = _mm_add_ps(vacc3x4567, _mm_mul_ps(va3c1111, vb4567c1));
115
116 const __m128 va0c2222 = _mm_shuffle_ps(va0, va0, _MM_SHUFFLE(2, 2, 2, 2));
117 const __m128 va1c2222 = _mm_shuffle_ps(va1, va1, _MM_SHUFFLE(2, 2, 2, 2));
118 const __m128 va2c2222 = _mm_shuffle_ps(va2, va2, _MM_SHUFFLE(2, 2, 2, 2));
119 const __m128 va3c2222 = _mm_shuffle_ps(va3, va3, _MM_SHUFFLE(2, 2, 2, 2));
120
121 const __m128 vb0123c2 = _mm_load_ps(w + 16);
122 const __m128 vb4567c2 = _mm_load_ps(w + 20);
123
124 vacc0x0123 = _mm_add_ps(vacc0x0123, _mm_mul_ps(va0c2222, vb0123c2));
125 vacc1x0123 = _mm_add_ps(vacc1x0123, _mm_mul_ps(va1c2222, vb0123c2));
126 vacc2x0123 = _mm_add_ps(vacc2x0123, _mm_mul_ps(va2c2222, vb0123c2));
127 vacc3x0123 = _mm_add_ps(vacc3x0123, _mm_mul_ps(va3c2222, vb0123c2));
128 vacc0x4567 = _mm_add_ps(vacc0x4567, _mm_mul_ps(va0c2222, vb4567c2));
129 vacc1x4567 = _mm_add_ps(vacc1x4567, _mm_mul_ps(va1c2222, vb4567c2));
130 vacc2x4567 = _mm_add_ps(vacc2x4567, _mm_mul_ps(va2c2222, vb4567c2));
131 vacc3x4567 = _mm_add_ps(vacc3x4567, _mm_mul_ps(va3c2222, vb4567c2));
132
133 const __m128 va0c3333 = _mm_shuffle_ps(va0, va0, _MM_SHUFFLE(3, 3, 3, 3));
134 const __m128 va1c3333 = _mm_shuffle_ps(va1, va1, _MM_SHUFFLE(3, 3, 3, 3));
135 const __m128 va2c3333 = _mm_shuffle_ps(va2, va2, _MM_SHUFFLE(3, 3, 3, 3));
136 const __m128 va3c3333 = _mm_shuffle_ps(va3, va3, _MM_SHUFFLE(3, 3, 3, 3));
137
138 const __m128 vb0123c3 = _mm_load_ps(w + 24);
139 const __m128 vb4567c3 = _mm_load_ps(w + 28);
140
141 vacc0x0123 = _mm_add_ps(vacc0x0123, _mm_mul_ps(va0c3333, vb0123c3));
142 vacc1x0123 = _mm_add_ps(vacc1x0123, _mm_mul_ps(va1c3333, vb0123c3));
143 vacc2x0123 = _mm_add_ps(vacc2x0123, _mm_mul_ps(va2c3333, vb0123c3));
144 vacc3x0123 = _mm_add_ps(vacc3x0123, _mm_mul_ps(va3c3333, vb0123c3));
145 vacc0x4567 = _mm_add_ps(vacc0x4567, _mm_mul_ps(va0c3333, vb4567c3));
146 vacc1x4567 = _mm_add_ps(vacc1x4567, _mm_mul_ps(va1c3333, vb4567c3));
147 vacc2x4567 = _mm_add_ps(vacc2x4567, _mm_mul_ps(va2c3333, vb4567c3));
148 vacc3x4567 = _mm_add_ps(vacc3x4567, _mm_mul_ps(va3c3333, vb4567c3));
149
150 w += 32;
151 k -= 4 * sizeof(float);
152 }
153 if XNN_UNLIKELY(k != 0) {
154 do {
155 const __m128 va0 = _mm_load1_ps(a0);
156 a0 += 1;
157 const __m128 va1 = _mm_load1_ps(a1);
158 a1 += 1;
159 const __m128 va2 = _mm_load1_ps(a2);
160 a2 += 1;
161 const __m128 va3 = _mm_load1_ps(a3);
162 a3 += 1;
163
164 const __m128 vb0123 = _mm_load_ps(w);
165 const __m128 vb4567 = _mm_load_ps(w + 4);
166 w += 8;
167
168 vacc0x0123 = _mm_add_ps(vacc0x0123, _mm_mul_ps(va0, vb0123));
169 vacc1x0123 = _mm_add_ps(vacc1x0123, _mm_mul_ps(va1, vb0123));
170 vacc2x0123 = _mm_add_ps(vacc2x0123, _mm_mul_ps(va2, vb0123));
171 vacc3x0123 = _mm_add_ps(vacc3x0123, _mm_mul_ps(va3, vb0123));
172 vacc0x4567 = _mm_add_ps(vacc0x4567, _mm_mul_ps(va0, vb4567));
173 vacc1x4567 = _mm_add_ps(vacc1x4567, _mm_mul_ps(va1, vb4567));
174 vacc2x4567 = _mm_add_ps(vacc2x4567, _mm_mul_ps(va2, vb4567));
175 vacc3x4567 = _mm_add_ps(vacc3x4567, _mm_mul_ps(va3, vb4567));
176
177 k -= sizeof(float);
178 } while (k != 0);
179 }
180
181 const __m128 vmax = _mm_load_ps(params->sse.max);
182 vacc0x0123 = _mm_min_ps(vacc0x0123, vmax);
183 vacc1x0123 = _mm_min_ps(vacc1x0123, vmax);
184 vacc2x0123 = _mm_min_ps(vacc2x0123, vmax);
185 vacc3x0123 = _mm_min_ps(vacc3x0123, vmax);
186 vacc0x4567 = _mm_min_ps(vacc0x4567, vmax);
187 vacc1x4567 = _mm_min_ps(vacc1x4567, vmax);
188 vacc2x4567 = _mm_min_ps(vacc2x4567, vmax);
189 vacc3x4567 = _mm_min_ps(vacc3x4567, vmax);
190
191 const __m128 vmin = _mm_load_ps(params->sse.min);
192 vacc0x0123 = _mm_max_ps(vacc0x0123, vmin);
193 vacc1x0123 = _mm_max_ps(vacc1x0123, vmin);
194 vacc2x0123 = _mm_max_ps(vacc2x0123, vmin);
195 vacc3x0123 = _mm_max_ps(vacc3x0123, vmin);
196 vacc0x4567 = _mm_max_ps(vacc0x4567, vmin);
197 vacc1x4567 = _mm_max_ps(vacc1x4567, vmin);
198 vacc2x4567 = _mm_max_ps(vacc2x4567, vmin);
199 vacc3x4567 = _mm_max_ps(vacc3x4567, vmin);
200
201 if XNN_LIKELY(nc >= 8) {
202 _mm_storeu_ps(c3, vacc3x0123);
203 _mm_storeu_ps(c3 + 4, vacc3x4567);
204 c3 = (float*) ((uintptr_t) c3 + cn_stride);
205 _mm_storeu_ps(c2, vacc2x0123);
206 _mm_storeu_ps(c2 + 4, vacc2x4567);
207 c2 = (float*) ((uintptr_t) c2 + cn_stride);
208 _mm_storeu_ps(c1, vacc1x0123);
209 _mm_storeu_ps(c1 + 4, vacc1x4567);
210 c1 = (float*) ((uintptr_t) c1 + cn_stride);
211 _mm_storeu_ps(c0, vacc0x0123);
212 _mm_storeu_ps(c0 + 4, vacc0x4567);
213 c0 = (float*) ((uintptr_t) c0 + cn_stride);
214
215 a3 = (const float*) ((uintptr_t) a3 - kc);
216 a2 = (const float*) ((uintptr_t) a2 - kc);
217 a1 = (const float*) ((uintptr_t) a1 - kc);
218 a0 = (const float*) ((uintptr_t) a0 - kc);
219
220 nc -= 8;
221 } else {
222 if (nc & 4) {
223 _mm_storeu_ps(c3, vacc3x0123);
224 _mm_storeu_ps(c2, vacc2x0123);
225 _mm_storeu_ps(c1, vacc1x0123);
226 _mm_storeu_ps(c0, vacc0x0123);
227
228 vacc3x0123 = vacc3x4567;
229 vacc2x0123 = vacc2x4567;
230 vacc1x0123 = vacc1x4567;
231 vacc0x0123 = vacc0x4567;
232
233 c3 += 4;
234 c2 += 4;
235 c1 += 4;
236 c0 += 4;
237 }
238 if (nc & 2) {
239 _mm_storel_pi((__m64*) c3, vacc3x0123);
240 _mm_storel_pi((__m64*) c2, vacc2x0123);
241 _mm_storel_pi((__m64*) c1, vacc1x0123);
242 _mm_storel_pi((__m64*) c0, vacc0x0123);
243
244 vacc3x0123 = _mm_movehl_ps(vacc3x0123, vacc3x0123);
245 vacc2x0123 = _mm_movehl_ps(vacc2x0123, vacc2x0123);
246 vacc1x0123 = _mm_movehl_ps(vacc1x0123, vacc1x0123);
247 vacc0x0123 = _mm_movehl_ps(vacc0x0123, vacc0x0123);
248
249 c3 += 2;
250 c2 += 2;
251 c1 += 2;
252 c0 += 2;
253 }
254 if (nc & 1) {
255 _mm_store_ss(c3, vacc3x0123);
256 _mm_store_ss(c2, vacc2x0123);
257 _mm_store_ss(c1, vacc1x0123);
258 _mm_store_ss(c0, vacc0x0123);
259 }
260
261 nc = 0;
262 }
263 } while (nc != 0);
264 }
265