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_gemminc_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 float * restrict acc,const union xnn_f32_minmax_params params[restrict XNN_MIN_ELEMENTS (1)])17 void xnn_f32_gemminc_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 float*restrict acc,
28     const union xnn_f32_minmax_params params[restrict XNN_MIN_ELEMENTS(1)])
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       const __m128 va0 = _mm_loadu_ps(a0);
75       a0 += 4;
76       const __m128 va1 = _mm_loadu_ps(a1);
77       a1 += 4;
78       const __m128 va2 = _mm_loadu_ps(a2);
79       a2 += 4;
80       const __m128 va3 = _mm_loadu_ps(a3);
81       a3 += 4;
82 
83 
84       const __m128 va0c0000 = _mm_shuffle_ps(va0, va0, _MM_SHUFFLE(0, 0, 0, 0));
85       const __m128 va1c0000 = _mm_shuffle_ps(va1, va1, _MM_SHUFFLE(0, 0, 0, 0));
86       const __m128 va2c0000 = _mm_shuffle_ps(va2, va2, _MM_SHUFFLE(0, 0, 0, 0));
87       const __m128 va3c0000 = _mm_shuffle_ps(va3, va3, _MM_SHUFFLE(0, 0, 0, 0));
88 
89       const __m128 vb0123c0 = _mm_load_ps(w + 0);
90       const __m128 vb4567c0 = _mm_load_ps(w + 4);
91 
92       vacc0x0123 = _mm_add_ps(vacc0x0123, _mm_mul_ps(va0c0000, vb0123c0));
93       vacc1x0123 = _mm_add_ps(vacc1x0123, _mm_mul_ps(va1c0000, vb0123c0));
94       vacc2x0123 = _mm_add_ps(vacc2x0123, _mm_mul_ps(va2c0000, vb0123c0));
95       vacc3x0123 = _mm_add_ps(vacc3x0123, _mm_mul_ps(va3c0000, vb0123c0));
96       vacc0x4567 = _mm_add_ps(vacc0x4567, _mm_mul_ps(va0c0000, vb4567c0));
97       vacc1x4567 = _mm_add_ps(vacc1x4567, _mm_mul_ps(va1c0000, vb4567c0));
98       vacc2x4567 = _mm_add_ps(vacc2x4567, _mm_mul_ps(va2c0000, vb4567c0));
99       vacc3x4567 = _mm_add_ps(vacc3x4567, _mm_mul_ps(va3c0000, vb4567c0));
100 
101       const __m128 va0c1111 = _mm_shuffle_ps(va0, va0, _MM_SHUFFLE(1, 1, 1, 1));
102       const __m128 va1c1111 = _mm_shuffle_ps(va1, va1, _MM_SHUFFLE(1, 1, 1, 1));
103       const __m128 va2c1111 = _mm_shuffle_ps(va2, va2, _MM_SHUFFLE(1, 1, 1, 1));
104       const __m128 va3c1111 = _mm_shuffle_ps(va3, va3, _MM_SHUFFLE(1, 1, 1, 1));
105 
106       const __m128 vb0123c1 = _mm_load_ps(w + 8);
107       const __m128 vb4567c1 = _mm_load_ps(w + 12);
108 
109       vacc0x0123 = _mm_add_ps(vacc0x0123, _mm_mul_ps(va0c1111, vb0123c1));
110       vacc1x0123 = _mm_add_ps(vacc1x0123, _mm_mul_ps(va1c1111, vb0123c1));
111       vacc2x0123 = _mm_add_ps(vacc2x0123, _mm_mul_ps(va2c1111, vb0123c1));
112       vacc3x0123 = _mm_add_ps(vacc3x0123, _mm_mul_ps(va3c1111, vb0123c1));
113       vacc0x4567 = _mm_add_ps(vacc0x4567, _mm_mul_ps(va0c1111, vb4567c1));
114       vacc1x4567 = _mm_add_ps(vacc1x4567, _mm_mul_ps(va1c1111, vb4567c1));
115       vacc2x4567 = _mm_add_ps(vacc2x4567, _mm_mul_ps(va2c1111, vb4567c1));
116       vacc3x4567 = _mm_add_ps(vacc3x4567, _mm_mul_ps(va3c1111, vb4567c1));
117 
118       const __m128 va0c2222 = _mm_shuffle_ps(va0, va0, _MM_SHUFFLE(2, 2, 2, 2));
119       const __m128 va1c2222 = _mm_shuffle_ps(va1, va1, _MM_SHUFFLE(2, 2, 2, 2));
120       const __m128 va2c2222 = _mm_shuffle_ps(va2, va2, _MM_SHUFFLE(2, 2, 2, 2));
121       const __m128 va3c2222 = _mm_shuffle_ps(va3, va3, _MM_SHUFFLE(2, 2, 2, 2));
122 
123       const __m128 vb0123c2 = _mm_load_ps(w + 16);
124       const __m128 vb4567c2 = _mm_load_ps(w + 20);
125 
126       vacc0x0123 = _mm_add_ps(vacc0x0123, _mm_mul_ps(va0c2222, vb0123c2));
127       vacc1x0123 = _mm_add_ps(vacc1x0123, _mm_mul_ps(va1c2222, vb0123c2));
128       vacc2x0123 = _mm_add_ps(vacc2x0123, _mm_mul_ps(va2c2222, vb0123c2));
129       vacc3x0123 = _mm_add_ps(vacc3x0123, _mm_mul_ps(va3c2222, vb0123c2));
130       vacc0x4567 = _mm_add_ps(vacc0x4567, _mm_mul_ps(va0c2222, vb4567c2));
131       vacc1x4567 = _mm_add_ps(vacc1x4567, _mm_mul_ps(va1c2222, vb4567c2));
132       vacc2x4567 = _mm_add_ps(vacc2x4567, _mm_mul_ps(va2c2222, vb4567c2));
133       vacc3x4567 = _mm_add_ps(vacc3x4567, _mm_mul_ps(va3c2222, vb4567c2));
134 
135       const __m128 va0c3333 = _mm_shuffle_ps(va0, va0, _MM_SHUFFLE(3, 3, 3, 3));
136       const __m128 va1c3333 = _mm_shuffle_ps(va1, va1, _MM_SHUFFLE(3, 3, 3, 3));
137       const __m128 va2c3333 = _mm_shuffle_ps(va2, va2, _MM_SHUFFLE(3, 3, 3, 3));
138       const __m128 va3c3333 = _mm_shuffle_ps(va3, va3, _MM_SHUFFLE(3, 3, 3, 3));
139 
140       const __m128 vb0123c3 = _mm_load_ps(w + 24);
141       const __m128 vb4567c3 = _mm_load_ps(w + 28);
142 
143       vacc0x0123 = _mm_add_ps(vacc0x0123, _mm_mul_ps(va0c3333, vb0123c3));
144       vacc1x0123 = _mm_add_ps(vacc1x0123, _mm_mul_ps(va1c3333, vb0123c3));
145       vacc2x0123 = _mm_add_ps(vacc2x0123, _mm_mul_ps(va2c3333, vb0123c3));
146       vacc3x0123 = _mm_add_ps(vacc3x0123, _mm_mul_ps(va3c3333, vb0123c3));
147       vacc0x4567 = _mm_add_ps(vacc0x4567, _mm_mul_ps(va0c3333, vb4567c3));
148       vacc1x4567 = _mm_add_ps(vacc1x4567, _mm_mul_ps(va1c3333, vb4567c3));
149       vacc2x4567 = _mm_add_ps(vacc2x4567, _mm_mul_ps(va2c3333, vb4567c3));
150       vacc3x4567 = _mm_add_ps(vacc3x4567, _mm_mul_ps(va3c3333, vb4567c3));
151 
152       w += 32;
153       k -= 4 * sizeof(float);
154     }
155     if XNN_UNLIKELY(k != 0) {
156       do {
157         const __m128 va0 = _mm_load1_ps(a0);
158         a0 += 1;
159         const __m128 va1 = _mm_load1_ps(a1);
160         a1 += 1;
161         const __m128 va2 = _mm_load1_ps(a2);
162         a2 += 1;
163         const __m128 va3 = _mm_load1_ps(a3);
164         a3 += 1;
165 
166         const __m128 vb0123 = _mm_load_ps(w);
167         const __m128 vb4567 = _mm_load_ps(w + 4);
168         w += 8;
169 
170         vacc0x0123 = _mm_add_ps(vacc0x0123, _mm_mul_ps(va0, vb0123));
171         vacc1x0123 = _mm_add_ps(vacc1x0123, _mm_mul_ps(va1, vb0123));
172         vacc2x0123 = _mm_add_ps(vacc2x0123, _mm_mul_ps(va2, vb0123));
173         vacc3x0123 = _mm_add_ps(vacc3x0123, _mm_mul_ps(va3, vb0123));
174         vacc0x4567 = _mm_add_ps(vacc0x4567, _mm_mul_ps(va0, vb4567));
175         vacc1x4567 = _mm_add_ps(vacc1x4567, _mm_mul_ps(va1, vb4567));
176         vacc2x4567 = _mm_add_ps(vacc2x4567, _mm_mul_ps(va2, vb4567));
177         vacc3x4567 = _mm_add_ps(vacc3x4567, _mm_mul_ps(va3, vb4567));
178 
179         k -= sizeof(float);
180       } while (k != 0);
181     }
182 
183     const __m128 vmax = _mm_load_ps(params->sse.max);
184     vacc0x0123 = _mm_min_ps(vacc0x0123, vmax);
185     vacc1x0123 = _mm_min_ps(vacc1x0123, vmax);
186     vacc2x0123 = _mm_min_ps(vacc2x0123, vmax);
187     vacc3x0123 = _mm_min_ps(vacc3x0123, vmax);
188     vacc0x4567 = _mm_min_ps(vacc0x4567, vmax);
189     vacc1x4567 = _mm_min_ps(vacc1x4567, vmax);
190     vacc2x4567 = _mm_min_ps(vacc2x4567, vmax);
191     vacc3x4567 = _mm_min_ps(vacc3x4567, vmax);
192 
193     const __m128 vmin = _mm_load_ps(params->sse.min);
194     vacc0x0123 = _mm_max_ps(vacc0x0123, vmin);
195     vacc1x0123 = _mm_max_ps(vacc1x0123, vmin);
196     vacc2x0123 = _mm_max_ps(vacc2x0123, vmin);
197     vacc3x0123 = _mm_max_ps(vacc3x0123, vmin);
198     vacc0x4567 = _mm_max_ps(vacc0x4567, vmin);
199     vacc1x4567 = _mm_max_ps(vacc1x4567, vmin);
200     vacc2x4567 = _mm_max_ps(vacc2x4567, vmin);
201     vacc3x4567 = _mm_max_ps(vacc3x4567, vmin);
202 
203     if XNN_LIKELY(nc >= 8) {
204       _mm_storeu_ps(c3, vacc3x0123);
205       _mm_storeu_ps(c3 + 4, vacc3x4567);
206       c3 = (float*) ((uintptr_t) c3 + cn_stride);
207       _mm_storeu_ps(c2, vacc2x0123);
208       _mm_storeu_ps(c2 + 4, vacc2x4567);
209       c2 = (float*) ((uintptr_t) c2 + cn_stride);
210       _mm_storeu_ps(c1, vacc1x0123);
211       _mm_storeu_ps(c1 + 4, vacc1x4567);
212       c1 = (float*) ((uintptr_t) c1 + cn_stride);
213       _mm_storeu_ps(c0, vacc0x0123);
214       _mm_storeu_ps(c0 + 4, vacc0x4567);
215       c0 = (float*) ((uintptr_t) c0 + cn_stride);
216 
217       a3 = (const float*) ((uintptr_t) a3 - kc);
218       a2 = (const float*) ((uintptr_t) a2 - kc);
219       a1 = (const float*) ((uintptr_t) a1 - kc);
220       a0 = (const float*) ((uintptr_t) a0 - kc);
221 
222       nc -= 8;
223     } else {
224       if (nc & 4) {
225         _mm_storeu_ps(c3, vacc3x0123);
226         _mm_storeu_ps(c2, vacc2x0123);
227         _mm_storeu_ps(c1, vacc1x0123);
228         _mm_storeu_ps(c0, vacc0x0123);
229 
230         vacc3x0123 = vacc3x4567;
231         vacc2x0123 = vacc2x4567;
232         vacc1x0123 = vacc1x4567;
233         vacc0x0123 = vacc0x4567;
234 
235         c3 += 4;
236         c2 += 4;
237         c1 += 4;
238         c0 += 4;
239       }
240       if (nc & 2) {
241         _mm_storel_pi((__m64*) c3, vacc3x0123);
242         _mm_storel_pi((__m64*) c2, vacc2x0123);
243         _mm_storel_pi((__m64*) c1, vacc1x0123);
244         _mm_storel_pi((__m64*) c0, vacc0x0123);
245 
246         vacc3x0123 = _mm_movehl_ps(vacc3x0123, vacc3x0123);
247         vacc2x0123 = _mm_movehl_ps(vacc2x0123, vacc2x0123);
248         vacc1x0123 = _mm_movehl_ps(vacc1x0123, vacc1x0123);
249         vacc0x0123 = _mm_movehl_ps(vacc0x0123, vacc0x0123);
250 
251         c3 += 2;
252         c2 += 2;
253         c1 += 2;
254         c0 += 2;
255       }
256       if (nc & 1) {
257         _mm_store_ss(c3, vacc3x0123);
258         _mm_store_ss(c2, vacc2x0123);
259         _mm_store_ss(c1, vacc1x0123);
260         _mm_store_ss(c0, vacc0x0123);
261       }
262 
263       nc = 0;
264     }
265   } while (nc != 0);
266 }
267