File size: 12,336 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
264
265
266
267
268
// Auto-generated file. Do not edit!
//   Template: src/f16-raddstoreexpminusmax/neonfp16arith-rr2-p2.c.in
//   Generator: tools/xngen
//
// Copyright 2022 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 <arm_neon.h>

#include <xnnpack/common.h>
#include <xnnpack/raddstoreexpminusmax.h>


void xnn_f16_raddstoreexpminusmax_ukernel__neonfp16arith_rr2_p2_x80_acc2(
    size_t batch,
    const void* input,
    const void* max,
    void* output,
    void* sum,
    const union xnn_f16_expminus_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
{
  assert(batch != 0);
  assert(batch % sizeof(uint16_t) == 0);
  assert(input != NULL);
  assert(max != NULL);
  assert(output != NULL);
  assert(sum != NULL);

  const float16x8_t vi_max = vreinterpretq_f16_u16(vld1q_dup_u16(max));
  const float16x8_t vlog2e = vreinterpretq_f16_u16(vld1q_dup_u16(&params->fp16arith_rr2_p2.log2e));
  const float16x8_t vmagic_bias = vreinterpretq_f16_u16(vld1q_dup_u16(&params->fp16arith_rr2_p2.magic_bias));
  const float16x8_t vminus_ln2_hi = vreinterpretq_f16_u16(vld1q_dup_u16(&params->fp16arith_rr2_p2.minus_ln2_hi));
  const float16x8_t vminus_ln2_lo = vreinterpretq_f16_u16(vld1q_dup_u16(&params->fp16arith_rr2_p2.minus_ln2_lo));
  const float16x8_t vc2 = vreinterpretq_f16_u16(vld1q_dup_u16(&params->fp16arith_rr2_p2.c2));
  const float16x8_t vc1 = vreinterpretq_f16_u16(vld1q_dup_u16(&params->fp16arith_rr2_p2.c1));
  const float16x8_t vdenorm_cutoff = vreinterpretq_f16_u16(vld1q_dup_u16(&params->fp16arith_rr2_p2.denorm_cutoff));

  const uint16_t* i = (const uint16_t*) input;
  uint16_t* o = (uint16_t*) output;
  float16x8_t vacc0 = vreinterpretq_f16_u16(vmovq_n_u16(0));
  float16x8_t vacc1 = vreinterpretq_f16_u16(vmovq_n_u16(0));
  for (; batch >= 80 * sizeof(uint16_t); batch -= 80 * sizeof(uint16_t)) {
    const float16x8_t vi0 = vreinterpretq_f16_u16(vld1q_u16(i)); i += 8;
    const float16x8_t vi1 = vreinterpretq_f16_u16(vld1q_u16(i)); i += 8;
    const float16x8_t vi2 = vreinterpretq_f16_u16(vld1q_u16(i)); i += 8;
    const float16x8_t vi3 = vreinterpretq_f16_u16(vld1q_u16(i)); i += 8;
    const float16x8_t vi4 = vreinterpretq_f16_u16(vld1q_u16(i)); i += 8;
    const float16x8_t vi5 = vreinterpretq_f16_u16(vld1q_u16(i)); i += 8;
    const float16x8_t vi6 = vreinterpretq_f16_u16(vld1q_u16(i)); i += 8;
    const float16x8_t vi7 = vreinterpretq_f16_u16(vld1q_u16(i)); i += 8;
    const float16x8_t vi8 = vreinterpretq_f16_u16(vld1q_u16(i)); i += 8;
    const float16x8_t vi9 = vreinterpretq_f16_u16(vld1q_u16(i)); i += 8;

    const float16x8_t vx0 = vsubq_f16(vi0, vi_max);
    const float16x8_t vx1 = vsubq_f16(vi1, vi_max);
    const float16x8_t vx2 = vsubq_f16(vi2, vi_max);
    const float16x8_t vx3 = vsubq_f16(vi3, vi_max);
    const float16x8_t vx4 = vsubq_f16(vi4, vi_max);
    const float16x8_t vx5 = vsubq_f16(vi5, vi_max);
    const float16x8_t vx6 = vsubq_f16(vi6, vi_max);
    const float16x8_t vx7 = vsubq_f16(vi7, vi_max);
    const float16x8_t vx8 = vsubq_f16(vi8, vi_max);
    const float16x8_t vx9 = vsubq_f16(vi9, vi_max);

    float16x8_t vn0 = vfmaq_f16(vmagic_bias, vx0, vlog2e);
    float16x8_t vn1 = vfmaq_f16(vmagic_bias, vx1, vlog2e);
    float16x8_t vn2 = vfmaq_f16(vmagic_bias, vx2, vlog2e);
    float16x8_t vn3 = vfmaq_f16(vmagic_bias, vx3, vlog2e);
    float16x8_t vn4 = vfmaq_f16(vmagic_bias, vx4, vlog2e);
    float16x8_t vn5 = vfmaq_f16(vmagic_bias, vx5, vlog2e);
    float16x8_t vn6 = vfmaq_f16(vmagic_bias, vx6, vlog2e);
    float16x8_t vn7 = vfmaq_f16(vmagic_bias, vx7, vlog2e);
    float16x8_t vn8 = vfmaq_f16(vmagic_bias, vx8, vlog2e);
    float16x8_t vn9 = vfmaq_f16(vmagic_bias, vx9, vlog2e);

    const float16x8_t vs0 = vreinterpretq_f16_s16(vshlq_n_s16(vreinterpretq_s16_f16(vn0), 10));
    const float16x8_t vs1 = vreinterpretq_f16_s16(vshlq_n_s16(vreinterpretq_s16_f16(vn1), 10));
    const float16x8_t vs2 = vreinterpretq_f16_s16(vshlq_n_s16(vreinterpretq_s16_f16(vn2), 10));
    const float16x8_t vs3 = vreinterpretq_f16_s16(vshlq_n_s16(vreinterpretq_s16_f16(vn3), 10));
    const float16x8_t vs4 = vreinterpretq_f16_s16(vshlq_n_s16(vreinterpretq_s16_f16(vn4), 10));
    const float16x8_t vs5 = vreinterpretq_f16_s16(vshlq_n_s16(vreinterpretq_s16_f16(vn5), 10));
    const float16x8_t vs6 = vreinterpretq_f16_s16(vshlq_n_s16(vreinterpretq_s16_f16(vn6), 10));
    const float16x8_t vs7 = vreinterpretq_f16_s16(vshlq_n_s16(vreinterpretq_s16_f16(vn7), 10));
    const float16x8_t vs8 = vreinterpretq_f16_s16(vshlq_n_s16(vreinterpretq_s16_f16(vn8), 10));
    const float16x8_t vs9 = vreinterpretq_f16_s16(vshlq_n_s16(vreinterpretq_s16_f16(vn9), 10));

    vn0 = vsubq_f16(vn0, vmagic_bias);
    vn1 = vsubq_f16(vn1, vmagic_bias);
    vn2 = vsubq_f16(vn2, vmagic_bias);
    vn3 = vsubq_f16(vn3, vmagic_bias);
    vn4 = vsubq_f16(vn4, vmagic_bias);
    vn5 = vsubq_f16(vn5, vmagic_bias);
    vn6 = vsubq_f16(vn6, vmagic_bias);
    vn7 = vsubq_f16(vn7, vmagic_bias);
    vn8 = vsubq_f16(vn8, vmagic_bias);
    vn9 = vsubq_f16(vn9, vmagic_bias);

    float16x8_t vt0 = vfmaq_f16(vx0, vn0, vminus_ln2_hi);
    float16x8_t vt1 = vfmaq_f16(vx1, vn1, vminus_ln2_hi);
    float16x8_t vt2 = vfmaq_f16(vx2, vn2, vminus_ln2_hi);
    float16x8_t vt3 = vfmaq_f16(vx3, vn3, vminus_ln2_hi);
    float16x8_t vt4 = vfmaq_f16(vx4, vn4, vminus_ln2_hi);
    float16x8_t vt5 = vfmaq_f16(vx5, vn5, vminus_ln2_hi);
    float16x8_t vt6 = vfmaq_f16(vx6, vn6, vminus_ln2_hi);
    float16x8_t vt7 = vfmaq_f16(vx7, vn7, vminus_ln2_hi);
    float16x8_t vt8 = vfmaq_f16(vx8, vn8, vminus_ln2_hi);
    float16x8_t vt9 = vfmaq_f16(vx9, vn9, vminus_ln2_hi);

    vt0 = vfmaq_f16(vt0, vn0, vminus_ln2_lo);
    vt1 = vfmaq_f16(vt1, vn1, vminus_ln2_lo);
    vt2 = vfmaq_f16(vt2, vn2, vminus_ln2_lo);
    vt3 = vfmaq_f16(vt3, vn3, vminus_ln2_lo);
    vt4 = vfmaq_f16(vt4, vn4, vminus_ln2_lo);
    vt5 = vfmaq_f16(vt5, vn5, vminus_ln2_lo);
    vt6 = vfmaq_f16(vt6, vn6, vminus_ln2_lo);
    vt7 = vfmaq_f16(vt7, vn7, vminus_ln2_lo);
    vt8 = vfmaq_f16(vt8, vn8, vminus_ln2_lo);
    vt9 = vfmaq_f16(vt9, vn9, vminus_ln2_lo);

    const float16x8_t vp0 = vfmaq_f16(vc1, vc2, vt0);
    const float16x8_t vp1 = vfmaq_f16(vc1, vc2, vt1);
    const float16x8_t vp2 = vfmaq_f16(vc1, vc2, vt2);
    const float16x8_t vp3 = vfmaq_f16(vc1, vc2, vt3);
    const float16x8_t vp4 = vfmaq_f16(vc1, vc2, vt4);
    const float16x8_t vp5 = vfmaq_f16(vc1, vc2, vt5);
    const float16x8_t vp6 = vfmaq_f16(vc1, vc2, vt6);
    const float16x8_t vp7 = vfmaq_f16(vc1, vc2, vt7);
    const float16x8_t vp8 = vfmaq_f16(vc1, vc2, vt8);
    const float16x8_t vp9 = vfmaq_f16(vc1, vc2, vt9);

    vt0 = vmulq_f16(vt0, vs0);
    vt1 = vmulq_f16(vt1, vs1);
    vt2 = vmulq_f16(vt2, vs2);
    vt3 = vmulq_f16(vt3, vs3);
    vt4 = vmulq_f16(vt4, vs4);
    vt5 = vmulq_f16(vt5, vs5);
    vt6 = vmulq_f16(vt6, vs6);
    vt7 = vmulq_f16(vt7, vs7);
    vt8 = vmulq_f16(vt8, vs8);
    vt9 = vmulq_f16(vt9, vs9);

    float16x8_t vf0 = vfmaq_f16(vs0, vp0, vt0);
    const uint16x8_t vm0 = vcltq_f16(vx0, vdenorm_cutoff);
    float16x8_t vf1 = vfmaq_f16(vs1, vp1, vt1);
    const uint16x8_t vm1 = vcltq_f16(vx1, vdenorm_cutoff);
    float16x8_t vf2 = vfmaq_f16(vs2, vp2, vt2);
    const uint16x8_t vm2 = vcltq_f16(vx2, vdenorm_cutoff);
    float16x8_t vf3 = vfmaq_f16(vs3, vp3, vt3);
    const uint16x8_t vm3 = vcltq_f16(vx3, vdenorm_cutoff);
    float16x8_t vf4 = vfmaq_f16(vs4, vp4, vt4);
    const uint16x8_t vm4 = vcltq_f16(vx4, vdenorm_cutoff);
    float16x8_t vf5 = vfmaq_f16(vs5, vp5, vt5);
    const uint16x8_t vm5 = vcltq_f16(vx5, vdenorm_cutoff);
    float16x8_t vf6 = vfmaq_f16(vs6, vp6, vt6);
    const uint16x8_t vm6 = vcltq_f16(vx6, vdenorm_cutoff);
    float16x8_t vf7 = vfmaq_f16(vs7, vp7, vt7);
    const uint16x8_t vm7 = vcltq_f16(vx7, vdenorm_cutoff);
    float16x8_t vf8 = vfmaq_f16(vs8, vp8, vt8);
    const uint16x8_t vm8 = vcltq_f16(vx8, vdenorm_cutoff);
    float16x8_t vf9 = vfmaq_f16(vs9, vp9, vt9);
    const uint16x8_t vm9 = vcltq_f16(vx9, vdenorm_cutoff);

    vf0 = vreinterpretq_f16_u16(vbicq_u16(vreinterpretq_u16_f16(vf0), vm0));
    vf1 = vreinterpretq_f16_u16(vbicq_u16(vreinterpretq_u16_f16(vf1), vm1));
    vf2 = vreinterpretq_f16_u16(vbicq_u16(vreinterpretq_u16_f16(vf2), vm2));
    vf3 = vreinterpretq_f16_u16(vbicq_u16(vreinterpretq_u16_f16(vf3), vm3));
    vf4 = vreinterpretq_f16_u16(vbicq_u16(vreinterpretq_u16_f16(vf4), vm4));
    vf5 = vreinterpretq_f16_u16(vbicq_u16(vreinterpretq_u16_f16(vf5), vm5));
    vf6 = vreinterpretq_f16_u16(vbicq_u16(vreinterpretq_u16_f16(vf6), vm6));
    vf7 = vreinterpretq_f16_u16(vbicq_u16(vreinterpretq_u16_f16(vf7), vm7));
    vf8 = vreinterpretq_f16_u16(vbicq_u16(vreinterpretq_u16_f16(vf8), vm8));
    vf9 = vreinterpretq_f16_u16(vbicq_u16(vreinterpretq_u16_f16(vf9), vm9));

    vst1q_u16(o, vreinterpretq_u16_f16(vf0)); o += 8;
    vst1q_u16(o, vreinterpretq_u16_f16(vf1)); o += 8;
    vst1q_u16(o, vreinterpretq_u16_f16(vf2)); o += 8;
    vst1q_u16(o, vreinterpretq_u16_f16(vf3)); o += 8;
    vst1q_u16(o, vreinterpretq_u16_f16(vf4)); o += 8;
    vst1q_u16(o, vreinterpretq_u16_f16(vf5)); o += 8;
    vst1q_u16(o, vreinterpretq_u16_f16(vf6)); o += 8;
    vst1q_u16(o, vreinterpretq_u16_f16(vf7)); o += 8;
    vst1q_u16(o, vreinterpretq_u16_f16(vf8)); o += 8;
    vst1q_u16(o, vreinterpretq_u16_f16(vf9)); o += 8;

    vacc0 = vaddq_f16(vacc0, vf0);
    vacc1 = vaddq_f16(vacc1, vf1);
    vacc0 = vaddq_f16(vacc0, vf2);
    vacc1 = vaddq_f16(vacc1, vf3);
    vacc0 = vaddq_f16(vacc0, vf4);
    vacc1 = vaddq_f16(vacc1, vf5);
    vacc0 = vaddq_f16(vacc0, vf6);
    vacc1 = vaddq_f16(vacc1, vf7);
    vacc0 = vaddq_f16(vacc0, vf8);
    vacc1 = vaddq_f16(vacc1, vf9);
  }
  vacc0 = vaddq_f16(vacc0, vacc1);

  float16x8_t vacc = vacc0;
  for (; batch >= 8 * sizeof(uint16_t); batch -= 8 * sizeof(uint16_t)) {
    const float16x8_t vi = vreinterpretq_f16_u16(vld1q_u16(i)); i += 8;

    const float16x8_t vx = vsubq_f16(vi, vi_max);

    float16x8_t vn = vfmaq_f16(vmagic_bias, vx, vlog2e);
    const float16x8_t vs = vreinterpretq_f16_s16(vshlq_n_s16(vreinterpretq_s16_f16(vn), 10));
    vn = vsubq_f16(vn, vmagic_bias);

    float16x8_t vt = vfmaq_f16(vx, vn, vminus_ln2_hi);
    vt = vfmaq_f16(vt, vn, vminus_ln2_lo);

    const float16x8_t vp = vfmaq_f16(vc1, vc2, vt);
    vt = vmulq_f16(vt, vs);

    float16x8_t vf = vfmaq_f16(vs, vp, vt);
    const uint16x8_t vm = vcltq_f16(vx, vdenorm_cutoff);
    vf = vreinterpretq_f16_u16(vbicq_u16(vreinterpretq_u16_f16(vf), vm));

    vst1q_u16(o, vreinterpretq_u16_f16(vf)); o += 8;

    vacc = vaddq_f16(vacc, vf);
  }
  float16x4_t vacc_lo = vadd_f16(vget_low_f16(vacc), vget_high_f16(vacc));
  if (batch != 0) {
    assert(batch >= 1 * sizeof(uint16_t));
    assert(batch <= 7 * sizeof(uint16_t));
    const float16x8_t vi = vreinterpretq_f16_u16(vld1q_u16(i));

    const float16x8_t vx = vsubq_f16(vi, vi_max);

    float16x8_t vn = vfmaq_f16(vmagic_bias, vx, vlog2e);
    const float16x8_t vs = vreinterpretq_f16_s16(vshlq_n_s16(vreinterpretq_s16_f16(vn), 10));
    vn = vsubq_f16(vn, vmagic_bias);

    float16x8_t vt = vfmaq_f16(vx, vn, vminus_ln2_hi);
    vt = vfmaq_f16(vt, vn, vminus_ln2_lo);

    const float16x8_t vp = vfmaq_f16(vc1, vc2, vt);
    vt = vmulq_f16(vt, vs);

    float16x8_t vf = vfmaq_f16(vs, vp, vt);
    const uint16x8_t vm = vcltq_f16(vx, vdenorm_cutoff);
    vf = vreinterpretq_f16_u16(vbicq_u16(vreinterpretq_u16_f16(vf), vm));

    float16x4_t vf_lo = vget_low_f16(vf);
    if (batch & (4 * sizeof(uint16_t))) {
      vst1_u16(o, vreinterpret_u16_f16(vf_lo)); o += 4;
      vacc_lo = vadd_f16(vacc_lo, vf_lo);
      vf_lo = vget_high_f16(vf);
    }
    if (batch & (2 * sizeof(uint16_t))) {
      vst1_lane_u32((void*) o, vreinterpret_u32_f16(vf_lo), 0); o += 2;
      vacc_lo = vadd_f16(vacc_lo, vreinterpret_f16_u64(vshl_n_u64(vreinterpret_u64_f16(vf_lo), 32)));
      vf_lo = vext_f16(vf_lo, vf_lo, 2);
    }
    if (batch & (1 * sizeof(uint16_t))) {
      vst1_lane_u16(o, vreinterpret_u16_f16(vf_lo), 0);
      vacc_lo = vadd_f16(vacc_lo, vreinterpret_f16_u64(vshl_n_u64(vreinterpret_u64_f16(vf_lo), 48)));
    }
  }
  vacc_lo = vpadd_f16(vacc_lo, vacc_lo);
  vacc_lo = vpadd_f16(vacc_lo, vacc_lo);
  vst1_lane_u16(sum, vreinterpret_u16_f16(vacc_lo), 0);
}