File size: 9,942 Bytes
8b7c501
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
// Auto-generated file. Do not edit!
//   Template: src/f32-gemm/sse-dup.c.in
//   Generator: tools/xngen
//
// Copyright 2019 Google LLC
//
// This source code is licensed under the BSD-style license found in the
// LICENSE file in the root directory of this source tree.

#include <assert.h>

#include <immintrin.h>

#include <xnnpack/gemm.h>


void 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)])
{
  assert(mr != 0);
  assert(mr <= 3);
  assert(nc != 0);
  assert(kc != 0);
  assert(kc % sizeof(float) == 0);
  assert(a != NULL);
  assert(w != NULL);
  assert(c != NULL);
  assert(acc != NULL);

  const float* a0 = a;
  float* c0 = c;
  const float* a1 = (const float*) ((uintptr_t) a0 + a_stride);
  float* c1 = (float*) ((uintptr_t) c0 + cm_stride);
  if XNN_UNPREDICTABLE(mr < 2) {
    a1 = a0;
    c1 = c0;
  }
  const float* a2 = (const float*) ((uintptr_t) a1 + a_stride);
  float* c2 = (float*) ((uintptr_t) c1 + cm_stride);
  if XNN_UNPREDICTABLE(mr <= 2) {
    a2 = a1;
    c2 = c1;
  }

  do {
    __m128 vacc0x0123 = _mm_load_ps(acc + 0);
    __m128 vacc0x4567 = _mm_load_ps(acc + 4);
    __m128 vacc1x0123 = _mm_load_ps(acc + 8);
    __m128 vacc1x4567 = _mm_load_ps(acc + 12);
    __m128 vacc2x0123 = _mm_load_ps(acc + 16);
    __m128 vacc2x4567 = _mm_load_ps(acc + 20);
    acc += 24;

    size_t k = kc;
    for (; k >= 4 * sizeof(float); k -= 4 * sizeof(float)) {
      const __m128 va0 = _mm_loadu_ps(a0);
      a0 += 4;
      const __m128 va1 = _mm_loadu_ps(a1);
      a1 += 4;
      const __m128 va2 = _mm_loadu_ps(a2);
      a2 += 4;


      const __m128 va0c0000 = _mm_shuffle_ps(va0, va0, _MM_SHUFFLE(0, 0, 0, 0));
      const __m128 va1c0000 = _mm_shuffle_ps(va1, va1, _MM_SHUFFLE(0, 0, 0, 0));
      const __m128 va2c0000 = _mm_shuffle_ps(va2, va2, _MM_SHUFFLE(0, 0, 0, 0));

      const __m128 vb0123c0 = _mm_load_ps(w + 0);
      const __m128 vb4567c0 = _mm_load_ps(w + 4);

      vacc0x0123 = _mm_add_ps(vacc0x0123, _mm_mul_ps(va0c0000, vb0123c0));
      vacc1x0123 = _mm_add_ps(vacc1x0123, _mm_mul_ps(va1c0000, vb0123c0));
      vacc2x0123 = _mm_add_ps(vacc2x0123, _mm_mul_ps(va2c0000, vb0123c0));
      vacc0x4567 = _mm_add_ps(vacc0x4567, _mm_mul_ps(va0c0000, vb4567c0));
      vacc1x4567 = _mm_add_ps(vacc1x4567, _mm_mul_ps(va1c0000, vb4567c0));
      vacc2x4567 = _mm_add_ps(vacc2x4567, _mm_mul_ps(va2c0000, vb4567c0));

      const __m128 va0c1111 = _mm_shuffle_ps(va0, va0, _MM_SHUFFLE(1, 1, 1, 1));
      const __m128 va1c1111 = _mm_shuffle_ps(va1, va1, _MM_SHUFFLE(1, 1, 1, 1));
      const __m128 va2c1111 = _mm_shuffle_ps(va2, va2, _MM_SHUFFLE(1, 1, 1, 1));

      const __m128 vb0123c1 = _mm_load_ps(w + 8);
      const __m128 vb4567c1 = _mm_load_ps(w + 12);

      vacc0x0123 = _mm_add_ps(vacc0x0123, _mm_mul_ps(va0c1111, vb0123c1));
      vacc1x0123 = _mm_add_ps(vacc1x0123, _mm_mul_ps(va1c1111, vb0123c1));
      vacc2x0123 = _mm_add_ps(vacc2x0123, _mm_mul_ps(va2c1111, vb0123c1));
      vacc0x4567 = _mm_add_ps(vacc0x4567, _mm_mul_ps(va0c1111, vb4567c1));
      vacc1x4567 = _mm_add_ps(vacc1x4567, _mm_mul_ps(va1c1111, vb4567c1));
      vacc2x4567 = _mm_add_ps(vacc2x4567, _mm_mul_ps(va2c1111, vb4567c1));

      const __m128 va0c2222 = _mm_shuffle_ps(va0, va0, _MM_SHUFFLE(2, 2, 2, 2));
      const __m128 va1c2222 = _mm_shuffle_ps(va1, va1, _MM_SHUFFLE(2, 2, 2, 2));
      const __m128 va2c2222 = _mm_shuffle_ps(va2, va2, _MM_SHUFFLE(2, 2, 2, 2));

      const __m128 vb0123c2 = _mm_load_ps(w + 16);
      const __m128 vb4567c2 = _mm_load_ps(w + 20);

      vacc0x0123 = _mm_add_ps(vacc0x0123, _mm_mul_ps(va0c2222, vb0123c2));
      vacc1x0123 = _mm_add_ps(vacc1x0123, _mm_mul_ps(va1c2222, vb0123c2));
      vacc2x0123 = _mm_add_ps(vacc2x0123, _mm_mul_ps(va2c2222, vb0123c2));
      vacc0x4567 = _mm_add_ps(vacc0x4567, _mm_mul_ps(va0c2222, vb4567c2));
      vacc1x4567 = _mm_add_ps(vacc1x4567, _mm_mul_ps(va1c2222, vb4567c2));
      vacc2x4567 = _mm_add_ps(vacc2x4567, _mm_mul_ps(va2c2222, vb4567c2));

      const __m128 va0c3333 = _mm_shuffle_ps(va0, va0, _MM_SHUFFLE(3, 3, 3, 3));
      const __m128 va1c3333 = _mm_shuffle_ps(va1, va1, _MM_SHUFFLE(3, 3, 3, 3));
      const __m128 va2c3333 = _mm_shuffle_ps(va2, va2, _MM_SHUFFLE(3, 3, 3, 3));

      const __m128 vb0123c3 = _mm_load_ps(w + 24);
      const __m128 vb4567c3 = _mm_load_ps(w + 28);

      vacc0x0123 = _mm_add_ps(vacc0x0123, _mm_mul_ps(va0c3333, vb0123c3));
      vacc1x0123 = _mm_add_ps(vacc1x0123, _mm_mul_ps(va1c3333, vb0123c3));
      vacc2x0123 = _mm_add_ps(vacc2x0123, _mm_mul_ps(va2c3333, vb0123c3));
      vacc0x4567 = _mm_add_ps(vacc0x4567, _mm_mul_ps(va0c3333, vb4567c3));
      vacc1x4567 = _mm_add_ps(vacc1x4567, _mm_mul_ps(va1c3333, vb4567c3));
      vacc2x4567 = _mm_add_ps(vacc2x4567, _mm_mul_ps(va2c3333, vb4567c3));


      w += 32;
    }
    if XNN_UNLIKELY(k >= 2 * sizeof(float)) {
      const __m128 va0 = _mm_loadl_pi(_mm_undefined_ps(), (const __m64*) a0);
      a0 += 2;
      const __m128 va1 = _mm_loadl_pi(_mm_undefined_ps(), (const __m64*) a1);
      a1 += 2;
      const __m128 va2 = _mm_loadl_pi(_mm_undefined_ps(), (const __m64*) a2);
      a2 += 2;

      const __m128 va0c0000 = _mm_shuffle_ps(va0, va0, _MM_SHUFFLE(0, 0, 0, 0));
      const __m128 va1c0000 = _mm_shuffle_ps(va1, va1, _MM_SHUFFLE(0, 0, 0, 0));
      const __m128 va2c0000 = _mm_shuffle_ps(va2, va2, _MM_SHUFFLE(0, 0, 0, 0));

      const __m128 vb0123c0 = _mm_load_ps(w + 0);
      const __m128 vb4567c0 = _mm_load_ps(w + 4);

      vacc0x0123 = _mm_add_ps(vacc0x0123, _mm_mul_ps(va0c0000, vb0123c0));
      vacc1x0123 = _mm_add_ps(vacc1x0123, _mm_mul_ps(va1c0000, vb0123c0));
      vacc2x0123 = _mm_add_ps(vacc2x0123, _mm_mul_ps(va2c0000, vb0123c0));
      vacc0x4567 = _mm_add_ps(vacc0x4567, _mm_mul_ps(va0c0000, vb4567c0));
      vacc1x4567 = _mm_add_ps(vacc1x4567, _mm_mul_ps(va1c0000, vb4567c0));
      vacc2x4567 = _mm_add_ps(vacc2x4567, _mm_mul_ps(va2c0000, vb4567c0));

      const __m128 va0c1111 = _mm_shuffle_ps(va0, va0, _MM_SHUFFLE(1, 1, 1, 1));
      const __m128 va1c1111 = _mm_shuffle_ps(va1, va1, _MM_SHUFFLE(1, 1, 1, 1));
      const __m128 va2c1111 = _mm_shuffle_ps(va2, va2, _MM_SHUFFLE(1, 1, 1, 1));

      const __m128 vb0123c1 = _mm_load_ps(w + 8);
      const __m128 vb4567c1 = _mm_load_ps(w + 12);

      vacc0x0123 = _mm_add_ps(vacc0x0123, _mm_mul_ps(va0c1111, vb0123c1));
      vacc1x0123 = _mm_add_ps(vacc1x0123, _mm_mul_ps(va1c1111, vb0123c1));
      vacc2x0123 = _mm_add_ps(vacc2x0123, _mm_mul_ps(va2c1111, vb0123c1));
      vacc0x4567 = _mm_add_ps(vacc0x4567, _mm_mul_ps(va0c1111, vb4567c1));
      vacc1x4567 = _mm_add_ps(vacc1x4567, _mm_mul_ps(va1c1111, vb4567c1));
      vacc2x4567 = _mm_add_ps(vacc2x4567, _mm_mul_ps(va2c1111, vb4567c1));

      w += 16;
      k -= 2 * sizeof(float);
    }
    if XNN_UNLIKELY(k != 0) {
      const __m128 va0 = _mm_load1_ps(a0);
      a0 += 1;
      const __m128 va1 = _mm_load1_ps(a1);
      a1 += 1;
      const __m128 va2 = _mm_load1_ps(a2);
      a2 += 1;

      const __m128 vb0123 = _mm_load_ps(w);
      const __m128 vb4567 = _mm_load_ps(w + 4);
      w += 8;

      vacc0x0123 = _mm_add_ps(vacc0x0123, _mm_mul_ps(va0, vb0123));
      vacc1x0123 = _mm_add_ps(vacc1x0123, _mm_mul_ps(va1, vb0123));
      vacc2x0123 = _mm_add_ps(vacc2x0123, _mm_mul_ps(va2, vb0123));
      vacc0x4567 = _mm_add_ps(vacc0x4567, _mm_mul_ps(va0, vb4567));
      vacc1x4567 = _mm_add_ps(vacc1x4567, _mm_mul_ps(va1, vb4567));
      vacc2x4567 = _mm_add_ps(vacc2x4567, _mm_mul_ps(va2, vb4567));

      k -= sizeof(float);
    }

    const __m128 vmax = _mm_load_ps(params->sse.max);
    vacc0x0123 = _mm_min_ps(vacc0x0123, vmax);
    vacc1x0123 = _mm_min_ps(vacc1x0123, vmax);
    vacc2x0123 = _mm_min_ps(vacc2x0123, vmax);
    vacc0x4567 = _mm_min_ps(vacc0x4567, vmax);
    vacc1x4567 = _mm_min_ps(vacc1x4567, vmax);
    vacc2x4567 = _mm_min_ps(vacc2x4567, vmax);

    const __m128 vmin = _mm_load_ps(params->sse.min);
    vacc0x0123 = _mm_max_ps(vacc0x0123, vmin);
    vacc1x0123 = _mm_max_ps(vacc1x0123, vmin);
    vacc2x0123 = _mm_max_ps(vacc2x0123, vmin);
    vacc0x4567 = _mm_max_ps(vacc0x4567, vmin);
    vacc1x4567 = _mm_max_ps(vacc1x4567, vmin);
    vacc2x4567 = _mm_max_ps(vacc2x4567, vmin);

    if XNN_LIKELY(nc >= 8) {
      _mm_storeu_ps(c2, vacc2x0123);
      _mm_storeu_ps(c2 + 4, vacc2x4567);
      c2 = (float*) ((uintptr_t) c2 + cn_stride);
      _mm_storeu_ps(c1, vacc1x0123);
      _mm_storeu_ps(c1 + 4, vacc1x4567);
      c1 = (float*) ((uintptr_t) c1 + cn_stride);
      _mm_storeu_ps(c0, vacc0x0123);
      _mm_storeu_ps(c0 + 4, vacc0x4567);
      c0 = (float*) ((uintptr_t) c0 + cn_stride);

      a2 = (const float*) ((uintptr_t) a2 - kc);
      a1 = (const float*) ((uintptr_t) a1 - kc);
      a0 = (const float*) ((uintptr_t) a0 - kc);

      nc -= 8;
    } else {
      if (nc & 4) {
        _mm_storeu_ps(c2, vacc2x0123);
        _mm_storeu_ps(c1, vacc1x0123);
        _mm_storeu_ps(c0, vacc0x0123);

        vacc2x0123 = vacc2x4567;
        vacc1x0123 = vacc1x4567;
        vacc0x0123 = vacc0x4567;

        c2 += 4;
        c1 += 4;
        c0 += 4;
      }
      if (nc & 2) {
        _mm_storel_pi((__m64*) c2, vacc2x0123);
        _mm_storel_pi((__m64*) c1, vacc1x0123);
        _mm_storel_pi((__m64*) c0, vacc0x0123);

        vacc2x0123 = _mm_movehl_ps(vacc2x0123, vacc2x0123);
        vacc1x0123 = _mm_movehl_ps(vacc1x0123, vacc1x0123);
        vacc0x0123 = _mm_movehl_ps(vacc0x0123, vacc0x0123);

        c2 += 2;
        c1 += 2;
        c0 += 2;
      }
      if (nc & 1) {
        _mm_store_ss(c2, vacc2x0123);
        _mm_store_ss(c1, vacc1x0123);
        _mm_store_ss(c0, vacc0x0123);
      }

      nc = 0;
    }
  } while (nc != 0);
}