xref: /aosp_15_r20/external/XNNPACK/src/f32-gemm/gen-inc/3x8inc-minmax-sse-dup.c (revision 4bdc94577ba0e567308109d787f7fec7b531ce36)
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_3x8__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_3x8__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 <= 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       const __m128 va0 = _mm_loadu_ps(a0);
67       a0 += 4;
68       const __m128 va1 = _mm_loadu_ps(a1);
69       a1 += 4;
70       const __m128 va2 = _mm_loadu_ps(a2);
71       a2 += 4;
72 
73 
74       const __m128 va0c0000 = _mm_shuffle_ps(va0, va0, _MM_SHUFFLE(0, 0, 0, 0));
75       const __m128 va1c0000 = _mm_shuffle_ps(va1, va1, _MM_SHUFFLE(0, 0, 0, 0));
76       const __m128 va2c0000 = _mm_shuffle_ps(va2, va2, _MM_SHUFFLE(0, 0, 0, 0));
77 
78       const __m128 vb0123c0 = _mm_load_ps(w + 0);
79       const __m128 vb4567c0 = _mm_load_ps(w + 4);
80 
81       vacc0x0123 = _mm_add_ps(vacc0x0123, _mm_mul_ps(va0c0000, vb0123c0));
82       vacc1x0123 = _mm_add_ps(vacc1x0123, _mm_mul_ps(va1c0000, vb0123c0));
83       vacc2x0123 = _mm_add_ps(vacc2x0123, _mm_mul_ps(va2c0000, vb0123c0));
84       vacc0x4567 = _mm_add_ps(vacc0x4567, _mm_mul_ps(va0c0000, vb4567c0));
85       vacc1x4567 = _mm_add_ps(vacc1x4567, _mm_mul_ps(va1c0000, vb4567c0));
86       vacc2x4567 = _mm_add_ps(vacc2x4567, _mm_mul_ps(va2c0000, vb4567c0));
87 
88       const __m128 va0c1111 = _mm_shuffle_ps(va0, va0, _MM_SHUFFLE(1, 1, 1, 1));
89       const __m128 va1c1111 = _mm_shuffle_ps(va1, va1, _MM_SHUFFLE(1, 1, 1, 1));
90       const __m128 va2c1111 = _mm_shuffle_ps(va2, va2, _MM_SHUFFLE(1, 1, 1, 1));
91 
92       const __m128 vb0123c1 = _mm_load_ps(w + 8);
93       const __m128 vb4567c1 = _mm_load_ps(w + 12);
94 
95       vacc0x0123 = _mm_add_ps(vacc0x0123, _mm_mul_ps(va0c1111, vb0123c1));
96       vacc1x0123 = _mm_add_ps(vacc1x0123, _mm_mul_ps(va1c1111, vb0123c1));
97       vacc2x0123 = _mm_add_ps(vacc2x0123, _mm_mul_ps(va2c1111, vb0123c1));
98       vacc0x4567 = _mm_add_ps(vacc0x4567, _mm_mul_ps(va0c1111, vb4567c1));
99       vacc1x4567 = _mm_add_ps(vacc1x4567, _mm_mul_ps(va1c1111, vb4567c1));
100       vacc2x4567 = _mm_add_ps(vacc2x4567, _mm_mul_ps(va2c1111, vb4567c1));
101 
102       const __m128 va0c2222 = _mm_shuffle_ps(va0, va0, _MM_SHUFFLE(2, 2, 2, 2));
103       const __m128 va1c2222 = _mm_shuffle_ps(va1, va1, _MM_SHUFFLE(2, 2, 2, 2));
104       const __m128 va2c2222 = _mm_shuffle_ps(va2, va2, _MM_SHUFFLE(2, 2, 2, 2));
105 
106       const __m128 vb0123c2 = _mm_load_ps(w + 16);
107       const __m128 vb4567c2 = _mm_load_ps(w + 20);
108 
109       vacc0x0123 = _mm_add_ps(vacc0x0123, _mm_mul_ps(va0c2222, vb0123c2));
110       vacc1x0123 = _mm_add_ps(vacc1x0123, _mm_mul_ps(va1c2222, vb0123c2));
111       vacc2x0123 = _mm_add_ps(vacc2x0123, _mm_mul_ps(va2c2222, vb0123c2));
112       vacc0x4567 = _mm_add_ps(vacc0x4567, _mm_mul_ps(va0c2222, vb4567c2));
113       vacc1x4567 = _mm_add_ps(vacc1x4567, _mm_mul_ps(va1c2222, vb4567c2));
114       vacc2x4567 = _mm_add_ps(vacc2x4567, _mm_mul_ps(va2c2222, vb4567c2));
115 
116       const __m128 va0c3333 = _mm_shuffle_ps(va0, va0, _MM_SHUFFLE(3, 3, 3, 3));
117       const __m128 va1c3333 = _mm_shuffle_ps(va1, va1, _MM_SHUFFLE(3, 3, 3, 3));
118       const __m128 va2c3333 = _mm_shuffle_ps(va2, va2, _MM_SHUFFLE(3, 3, 3, 3));
119 
120       const __m128 vb0123c3 = _mm_load_ps(w + 24);
121       const __m128 vb4567c3 = _mm_load_ps(w + 28);
122 
123       vacc0x0123 = _mm_add_ps(vacc0x0123, _mm_mul_ps(va0c3333, vb0123c3));
124       vacc1x0123 = _mm_add_ps(vacc1x0123, _mm_mul_ps(va1c3333, vb0123c3));
125       vacc2x0123 = _mm_add_ps(vacc2x0123, _mm_mul_ps(va2c3333, vb0123c3));
126       vacc0x4567 = _mm_add_ps(vacc0x4567, _mm_mul_ps(va0c3333, vb4567c3));
127       vacc1x4567 = _mm_add_ps(vacc1x4567, _mm_mul_ps(va1c3333, vb4567c3));
128       vacc2x4567 = _mm_add_ps(vacc2x4567, _mm_mul_ps(va2c3333, vb4567c3));
129 
130       w += 32;
131       k -= 4 * sizeof(float);
132     }
133     if XNN_UNLIKELY(k != 0) {
134       do {
135         const __m128 va0 = _mm_load1_ps(a0);
136         a0 += 1;
137         const __m128 va1 = _mm_load1_ps(a1);
138         a1 += 1;
139         const __m128 va2 = _mm_load1_ps(a2);
140         a2 += 1;
141 
142         const __m128 vb0123 = _mm_load_ps(w);
143         const __m128 vb4567 = _mm_load_ps(w + 4);
144         w += 8;
145 
146         vacc0x0123 = _mm_add_ps(vacc0x0123, _mm_mul_ps(va0, vb0123));
147         vacc1x0123 = _mm_add_ps(vacc1x0123, _mm_mul_ps(va1, vb0123));
148         vacc2x0123 = _mm_add_ps(vacc2x0123, _mm_mul_ps(va2, vb0123));
149         vacc0x4567 = _mm_add_ps(vacc0x4567, _mm_mul_ps(va0, vb4567));
150         vacc1x4567 = _mm_add_ps(vacc1x4567, _mm_mul_ps(va1, vb4567));
151         vacc2x4567 = _mm_add_ps(vacc2x4567, _mm_mul_ps(va2, vb4567));
152 
153         k -= sizeof(float);
154       } while (k != 0);
155     }
156 
157     const __m128 vmax = _mm_load_ps(params->sse.max);
158     vacc0x0123 = _mm_min_ps(vacc0x0123, vmax);
159     vacc1x0123 = _mm_min_ps(vacc1x0123, vmax);
160     vacc2x0123 = _mm_min_ps(vacc2x0123, vmax);
161     vacc0x4567 = _mm_min_ps(vacc0x4567, vmax);
162     vacc1x4567 = _mm_min_ps(vacc1x4567, vmax);
163     vacc2x4567 = _mm_min_ps(vacc2x4567, vmax);
164 
165     const __m128 vmin = _mm_load_ps(params->sse.min);
166     vacc0x0123 = _mm_max_ps(vacc0x0123, vmin);
167     vacc1x0123 = _mm_max_ps(vacc1x0123, vmin);
168     vacc2x0123 = _mm_max_ps(vacc2x0123, vmin);
169     vacc0x4567 = _mm_max_ps(vacc0x4567, vmin);
170     vacc1x4567 = _mm_max_ps(vacc1x4567, vmin);
171     vacc2x4567 = _mm_max_ps(vacc2x4567, vmin);
172 
173     if XNN_LIKELY(nc >= 8) {
174       _mm_storeu_ps(c2, vacc2x0123);
175       _mm_storeu_ps(c2 + 4, vacc2x4567);
176       c2 = (float*) ((uintptr_t) c2 + cn_stride);
177       _mm_storeu_ps(c1, vacc1x0123);
178       _mm_storeu_ps(c1 + 4, vacc1x4567);
179       c1 = (float*) ((uintptr_t) c1 + cn_stride);
180       _mm_storeu_ps(c0, vacc0x0123);
181       _mm_storeu_ps(c0 + 4, vacc0x4567);
182       c0 = (float*) ((uintptr_t) c0 + cn_stride);
183 
184       a2 = (const float*) ((uintptr_t) a2 - kc);
185       a1 = (const float*) ((uintptr_t) a1 - kc);
186       a0 = (const float*) ((uintptr_t) a0 - kc);
187 
188       nc -= 8;
189     } else {
190       if (nc & 4) {
191         _mm_storeu_ps(c2, vacc2x0123);
192         _mm_storeu_ps(c1, vacc1x0123);
193         _mm_storeu_ps(c0, vacc0x0123);
194 
195         vacc2x0123 = vacc2x4567;
196         vacc1x0123 = vacc1x4567;
197         vacc0x0123 = vacc0x4567;
198 
199         c2 += 4;
200         c1 += 4;
201         c0 += 4;
202       }
203       if (nc & 2) {
204         _mm_storel_pi((__m64*) c2, vacc2x0123);
205         _mm_storel_pi((__m64*) c1, vacc1x0123);
206         _mm_storel_pi((__m64*) c0, vacc0x0123);
207 
208         vacc2x0123 = _mm_movehl_ps(vacc2x0123, vacc2x0123);
209         vacc1x0123 = _mm_movehl_ps(vacc1x0123, vacc1x0123);
210         vacc0x0123 = _mm_movehl_ps(vacc0x0123, vacc0x0123);
211 
212         c2 += 2;
213         c1 += 2;
214         c0 += 2;
215       }
216       if (nc & 1) {
217         _mm_store_ss(c2, vacc2x0123);
218         _mm_store_ss(c1, vacc1x0123);
219         _mm_store_ss(c0, vacc0x0123);
220       }
221 
222       nc = 0;
223     }
224   } while (nc != 0);
225 }
226