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)])
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 do {
132 const __m128 va0 = _mm_load1_ps(a0);
133 a0 += 1;
134 const __m128 va1 = _mm_load1_ps(a1);
135 a1 += 1;
136 const __m128 va2 = _mm_load1_ps(a2);
137 a2 += 1;
138
139 const __m128 vb0123 = _mm_load_ps(w);
140 const __m128 vb4567 = _mm_load_ps(w + 4);
141 w += 8;
142
143 vacc0x0123 = _mm_add_ps(vacc0x0123, _mm_mul_ps(va0, vb0123));
144 vacc1x0123 = _mm_add_ps(vacc1x0123, _mm_mul_ps(va1, vb0123));
145 vacc2x0123 = _mm_add_ps(vacc2x0123, _mm_mul_ps(va2, vb0123));
146 vacc0x4567 = _mm_add_ps(vacc0x4567, _mm_mul_ps(va0, vb4567));
147 vacc1x4567 = _mm_add_ps(vacc1x4567, _mm_mul_ps(va1, vb4567));
148 vacc2x4567 = _mm_add_ps(vacc2x4567, _mm_mul_ps(va2, vb4567));
149
150 k -= sizeof(float);
151 } while (k != 0);
152 }
153
154 const __m128 vmax = _mm_load_ps(params->sse.max);
155 vacc0x0123 = _mm_min_ps(vacc0x0123, vmax);
156 vacc1x0123 = _mm_min_ps(vacc1x0123, vmax);
157 vacc2x0123 = _mm_min_ps(vacc2x0123, vmax);
158 vacc0x4567 = _mm_min_ps(vacc0x4567, vmax);
159 vacc1x4567 = _mm_min_ps(vacc1x4567, vmax);
160 vacc2x4567 = _mm_min_ps(vacc2x4567, vmax);
161
162 const __m128 vmin = _mm_load_ps(params->sse.min);
163 vacc0x0123 = _mm_max_ps(vacc0x0123, vmin);
164 vacc1x0123 = _mm_max_ps(vacc1x0123, vmin);
165 vacc2x0123 = _mm_max_ps(vacc2x0123, vmin);
166 vacc0x4567 = _mm_max_ps(vacc0x4567, vmin);
167 vacc1x4567 = _mm_max_ps(vacc1x4567, vmin);
168 vacc2x4567 = _mm_max_ps(vacc2x4567, vmin);
169
170 if XNN_LIKELY(nc >= 8) {
171 _mm_storeu_ps(c2, vacc2x0123);
172 _mm_storeu_ps(c2 + 4, vacc2x4567);
173 c2 = (float*) ((uintptr_t) c2 + cn_stride);
174 _mm_storeu_ps(c1, vacc1x0123);
175 _mm_storeu_ps(c1 + 4, vacc1x4567);
176 c1 = (float*) ((uintptr_t) c1 + cn_stride);
177 _mm_storeu_ps(c0, vacc0x0123);
178 _mm_storeu_ps(c0 + 4, vacc0x4567);
179 c0 = (float*) ((uintptr_t) c0 + cn_stride);
180
181 a2 = (const float*) ((uintptr_t) a2 - kc);
182 a1 = (const float*) ((uintptr_t) a1 - kc);
183 a0 = (const float*) ((uintptr_t) a0 - kc);
184
185 nc -= 8;
186 } else {
187 if (nc & 4) {
188 _mm_storeu_ps(c2, vacc2x0123);
189 _mm_storeu_ps(c1, vacc1x0123);
190 _mm_storeu_ps(c0, vacc0x0123);
191
192 vacc2x0123 = vacc2x4567;
193 vacc1x0123 = vacc1x4567;
194 vacc0x0123 = vacc0x4567;
195
196 c2 += 4;
197 c1 += 4;
198 c0 += 4;
199 }
200 if (nc & 2) {
201 _mm_storel_pi((__m64*) c2, vacc2x0123);
202 _mm_storel_pi((__m64*) c1, vacc1x0123);
203 _mm_storel_pi((__m64*) c0, vacc0x0123);
204
205 vacc2x0123 = _mm_movehl_ps(vacc2x0123, vacc2x0123);
206 vacc1x0123 = _mm_movehl_ps(vacc1x0123, vacc1x0123);
207 vacc0x0123 = _mm_movehl_ps(vacc0x0123, vacc0x0123);
208
209 c2 += 2;
210 c1 += 2;
211 c0 += 2;
212 }
213 if (nc & 1) {
214 _mm_store_ss(c2, vacc2x0123);
215 _mm_store_ss(c1, vacc1x0123);
216 _mm_store_ss(c0, vacc0x0123);
217 }
218
219 nc = 0;
220 }
221 } while (nc != 0);
222 }
223