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_ukernel_4x8s4__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_output_params params[restrict static1])17 void xnn_f32_gemminc_ukernel_4x8s4__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_output_params params[restrict static 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 __m128 va0 = _mm_loadu_ps(a0);
75 a0 += 4;
76 __m128 va1 = _mm_loadu_ps(a1);
77 a1 += 4;
78 __m128 va2 = _mm_loadu_ps(a2);
79 a2 += 4;
80 __m128 va3 = _mm_loadu_ps(a3);
81 a3 += 4;
82
83
84 const __m128 vb0123c0 = _mm_load_ps(w + 0);
85 const __m128 vb4567c0 = _mm_load_ps(w + 4);
86
87 vacc0x0123 = _mm_add_ps(vacc0x0123, _mm_mul_ps(va0, vb0123c0));
88 vacc1x0123 = _mm_add_ps(vacc1x0123, _mm_mul_ps(va1, vb0123c0));
89 vacc2x0123 = _mm_add_ps(vacc2x0123, _mm_mul_ps(va2, vb0123c0));
90 vacc3x0123 = _mm_add_ps(vacc3x0123, _mm_mul_ps(va3, vb0123c0));
91 vacc0x4567 = _mm_add_ps(vacc0x4567, _mm_mul_ps(va0, vb4567c0));
92 vacc1x4567 = _mm_add_ps(vacc1x4567, _mm_mul_ps(va1, vb4567c0));
93 vacc2x4567 = _mm_add_ps(vacc2x4567, _mm_mul_ps(va2, vb4567c0));
94 vacc3x4567 = _mm_add_ps(vacc3x4567, _mm_mul_ps(va3, vb4567c0));
95
96 va0 = _mm_shuffle_ps(va0, va0, _MM_SHUFFLE(0, 3, 2, 1));
97 va1 = _mm_shuffle_ps(va1, va1, _MM_SHUFFLE(0, 3, 2, 1));
98 va2 = _mm_shuffle_ps(va2, va2, _MM_SHUFFLE(0, 3, 2, 1));
99 va3 = _mm_shuffle_ps(va3, va3, _MM_SHUFFLE(0, 3, 2, 1));
100
101 const __m128 vb0123c1 = _mm_load_ps(w + 8);
102 const __m128 vb4567c1 = _mm_load_ps(w + 12);
103
104 vacc0x0123 = _mm_add_ps(vacc0x0123, _mm_mul_ps(va0, vb0123c1));
105 vacc1x0123 = _mm_add_ps(vacc1x0123, _mm_mul_ps(va1, vb0123c1));
106 vacc2x0123 = _mm_add_ps(vacc2x0123, _mm_mul_ps(va2, vb0123c1));
107 vacc3x0123 = _mm_add_ps(vacc3x0123, _mm_mul_ps(va3, vb0123c1));
108 vacc0x4567 = _mm_add_ps(vacc0x4567, _mm_mul_ps(va0, vb4567c1));
109 vacc1x4567 = _mm_add_ps(vacc1x4567, _mm_mul_ps(va1, vb4567c1));
110 vacc2x4567 = _mm_add_ps(vacc2x4567, _mm_mul_ps(va2, vb4567c1));
111 vacc3x4567 = _mm_add_ps(vacc3x4567, _mm_mul_ps(va3, vb4567c1));
112
113 va0 = _mm_shuffle_ps(va0, va0, _MM_SHUFFLE(0, 3, 2, 1));
114 va1 = _mm_shuffle_ps(va1, va1, _MM_SHUFFLE(0, 3, 2, 1));
115 va2 = _mm_shuffle_ps(va2, va2, _MM_SHUFFLE(0, 3, 2, 1));
116 va3 = _mm_shuffle_ps(va3, va3, _MM_SHUFFLE(0, 3, 2, 1));
117
118 const __m128 vb0123c2 = _mm_load_ps(w + 16);
119 const __m128 vb4567c2 = _mm_load_ps(w + 20);
120
121 vacc0x0123 = _mm_add_ps(vacc0x0123, _mm_mul_ps(va0, vb0123c2));
122 vacc1x0123 = _mm_add_ps(vacc1x0123, _mm_mul_ps(va1, vb0123c2));
123 vacc2x0123 = _mm_add_ps(vacc2x0123, _mm_mul_ps(va2, vb0123c2));
124 vacc3x0123 = _mm_add_ps(vacc3x0123, _mm_mul_ps(va3, vb0123c2));
125 vacc0x4567 = _mm_add_ps(vacc0x4567, _mm_mul_ps(va0, vb4567c2));
126 vacc1x4567 = _mm_add_ps(vacc1x4567, _mm_mul_ps(va1, vb4567c2));
127 vacc2x4567 = _mm_add_ps(vacc2x4567, _mm_mul_ps(va2, vb4567c2));
128 vacc3x4567 = _mm_add_ps(vacc3x4567, _mm_mul_ps(va3, vb4567c2));
129
130 va0 = _mm_shuffle_ps(va0, va0, _MM_SHUFFLE(0, 3, 2, 1));
131 va1 = _mm_shuffle_ps(va1, va1, _MM_SHUFFLE(0, 3, 2, 1));
132 va2 = _mm_shuffle_ps(va2, va2, _MM_SHUFFLE(0, 3, 2, 1));
133 va3 = _mm_shuffle_ps(va3, va3, _MM_SHUFFLE(0, 3, 2, 1));
134
135 const __m128 vb0123c3 = _mm_load_ps(w + 24);
136 const __m128 vb4567c3 = _mm_load_ps(w + 28);
137
138 vacc0x0123 = _mm_add_ps(vacc0x0123, _mm_mul_ps(va0, vb0123c3));
139 vacc1x0123 = _mm_add_ps(vacc1x0123, _mm_mul_ps(va1, vb0123c3));
140 vacc2x0123 = _mm_add_ps(vacc2x0123, _mm_mul_ps(va2, vb0123c3));
141 vacc3x0123 = _mm_add_ps(vacc3x0123, _mm_mul_ps(va3, vb0123c3));
142 vacc0x4567 = _mm_add_ps(vacc0x4567, _mm_mul_ps(va0, vb4567c3));
143 vacc1x4567 = _mm_add_ps(vacc1x4567, _mm_mul_ps(va1, vb4567c3));
144 vacc2x4567 = _mm_add_ps(vacc2x4567, _mm_mul_ps(va2, vb4567c3));
145 vacc3x4567 = _mm_add_ps(vacc3x4567, _mm_mul_ps(va3, vb4567c3));
146
147
148 w += 32;
149 k -= 4 * sizeof(float);
150 }
151 if XNN_UNLIKELY(k != 0) {
152 do {
153 const __m128 va0 = _mm_load1_ps(a0);
154 a0 += 1;
155 const __m128 va1 = _mm_load1_ps(a1);
156 a1 += 1;
157 const __m128 va2 = _mm_load1_ps(a2);
158 a2 += 1;
159 const __m128 va3 = _mm_load1_ps(a3);
160 a3 += 1;
161
162 const __m128 vb0123 = _mm_load_ps(w);
163 const __m128 vb4567 = _mm_load_ps(w + 4);
164 w += 8;
165
166 vacc0x0123 = _mm_add_ps(vacc0x0123, _mm_mul_ps(va0, vb0123));
167 vacc1x0123 = _mm_add_ps(vacc1x0123, _mm_mul_ps(va1, vb0123));
168 vacc2x0123 = _mm_add_ps(vacc2x0123, _mm_mul_ps(va2, vb0123));
169 vacc3x0123 = _mm_add_ps(vacc3x0123, _mm_mul_ps(va3, vb0123));
170 vacc0x4567 = _mm_add_ps(vacc0x4567, _mm_mul_ps(va0, vb4567));
171 vacc1x4567 = _mm_add_ps(vacc1x4567, _mm_mul_ps(va1, vb4567));
172 vacc2x4567 = _mm_add_ps(vacc2x4567, _mm_mul_ps(va2, vb4567));
173 vacc3x4567 = _mm_add_ps(vacc3x4567, _mm_mul_ps(va3, vb4567));
174
175 k -= sizeof(float);
176 } while (k != 0);
177 }
178
179 const __m128 vmax = _mm_load_ps(params->sse.max);
180 vacc0x0123 = _mm_min_ps(vacc0x0123, vmax);
181 vacc1x0123 = _mm_min_ps(vacc1x0123, vmax);
182 vacc2x0123 = _mm_min_ps(vacc2x0123, vmax);
183 vacc3x0123 = _mm_min_ps(vacc3x0123, vmax);
184 vacc0x4567 = _mm_min_ps(vacc0x4567, vmax);
185 vacc1x4567 = _mm_min_ps(vacc1x4567, vmax);
186 vacc2x4567 = _mm_min_ps(vacc2x4567, vmax);
187 vacc3x4567 = _mm_min_ps(vacc3x4567, vmax);
188
189 const __m128 vmin = _mm_load_ps(params->sse.min);
190 vacc0x0123 = _mm_max_ps(vacc0x0123, vmin);
191 vacc1x0123 = _mm_max_ps(vacc1x0123, vmin);
192 vacc2x0123 = _mm_max_ps(vacc2x0123, vmin);
193 vacc3x0123 = _mm_max_ps(vacc3x0123, vmin);
194 vacc0x4567 = _mm_max_ps(vacc0x4567, vmin);
195 vacc1x4567 = _mm_max_ps(vacc1x4567, vmin);
196 vacc2x4567 = _mm_max_ps(vacc2x4567, vmin);
197 vacc3x4567 = _mm_max_ps(vacc3x4567, vmin);
198
199 if XNN_LIKELY(nc >= 8) {
200 _mm_storeu_ps(c3, vacc3x0123);
201 _mm_storeu_ps(c3 + 4, vacc3x4567);
202 c3 = (float*) ((uintptr_t) c3 + cn_stride);
203 _mm_storeu_ps(c2, vacc2x0123);
204 _mm_storeu_ps(c2 + 4, vacc2x4567);
205 c2 = (float*) ((uintptr_t) c2 + cn_stride);
206 _mm_storeu_ps(c1, vacc1x0123);
207 _mm_storeu_ps(c1 + 4, vacc1x4567);
208 c1 = (float*) ((uintptr_t) c1 + cn_stride);
209 _mm_storeu_ps(c0, vacc0x0123);
210 _mm_storeu_ps(c0 + 4, vacc0x4567);
211 c0 = (float*) ((uintptr_t) c0 + cn_stride);
212
213 a3 = (const float*) ((uintptr_t) a3 - kc);
214 a2 = (const float*) ((uintptr_t) a2 - kc);
215 a1 = (const float*) ((uintptr_t) a1 - kc);
216 a0 = (const float*) ((uintptr_t) a0 - kc);
217
218 nc -= 8;
219 } else {
220 if (nc & 4) {
221 _mm_storeu_ps(c3, vacc3x0123);
222 _mm_storeu_ps(c2, vacc2x0123);
223 _mm_storeu_ps(c1, vacc1x0123);
224 _mm_storeu_ps(c0, vacc0x0123);
225
226 vacc3x0123 = vacc3x4567;
227 vacc2x0123 = vacc2x4567;
228 vacc1x0123 = vacc1x4567;
229 vacc0x0123 = vacc0x4567;
230
231 c3 += 4;
232 c2 += 4;
233 c1 += 4;
234 c0 += 4;
235 }
236 if (nc & 2) {
237 _mm_storel_pi((__m64*) c3, vacc3x0123);
238 _mm_storel_pi((__m64*) c2, vacc2x0123);
239 _mm_storel_pi((__m64*) c1, vacc1x0123);
240 _mm_storel_pi((__m64*) c0, vacc0x0123);
241
242 vacc3x0123 = _mm_movehl_ps(vacc3x0123, vacc3x0123);
243 vacc2x0123 = _mm_movehl_ps(vacc2x0123, vacc2x0123);
244 vacc1x0123 = _mm_movehl_ps(vacc1x0123, vacc1x0123);
245 vacc0x0123 = _mm_movehl_ps(vacc0x0123, vacc0x0123);
246
247 c3 += 2;
248 c2 += 2;
249 c1 += 2;
250 c0 += 2;
251 }
252 if (nc & 1) {
253 _mm_store_ss(c3, vacc3x0123);
254 _mm_store_ss(c2, vacc2x0123);
255 _mm_store_ss(c1, vacc1x0123);
256 _mm_store_ss(c0, vacc0x0123);
257 }
258
259 nc = 0;
260 }
261 } while (nc != 0);
262 }
263