File size: 13,182 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
269
270
271
272
273
274
275
276
277
278
279
280
281
// Auto-generated file. Do not edit!
//   Template: src/f16-raddstoreexpminusmax/avx2-rr1-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 <immintrin.h>

#include <xnnpack/intrinsics-polyfill.h>
#include <xnnpack/raddstoreexpminusmax.h>


void xnn_f16_raddstoreexpminusmax_ukernel__avx2_rr1_p2_x96_acc3(
    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 __m256 vi_max = _mm256_cvtph_ps(_mm_set1_epi16((short) *((const uint16_t*) max)));
  const __m256 vlog2e = _mm256_load_ps(params->avx2_rr1_p2.log2e);
  const __m256 vmagic_bias = _mm256_load_ps(params->avx2_rr1_p2.magic_bias);
  const __m256 vminus_ln2 = _mm256_load_ps(params->avx2_rr1_p2.minus_ln2);
  const __m256 vc2 = _mm256_load_ps(params->avx2_rr1_p2.c2);
  const __m256 vc1 = _mm256_load_ps(params->avx2_rr1_p2.c1);
  const __m256 vdenorm_cutoff = _mm256_load_ps(params->avx2_rr1_p2.denorm_cutoff);

  const uint16_t* i = (const uint16_t*) input;
  uint16_t* o = (uint16_t*) output;
  __m256 vacc0 = _mm256_setzero_ps();
  __m256 vacc1 = _mm256_setzero_ps();
  __m256 vacc2 = _mm256_setzero_ps();
  for (; batch >= 96 * sizeof(uint16_t); batch -= 96 * sizeof(uint16_t)) {
    const __m256 vi0 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i));
    const __m256 vi1 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) (i + 8)));
    const __m256 vi2 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) (i + 16)));
    const __m256 vi3 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) (i + 24)));
    const __m256 vi4 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) (i + 32)));
    const __m256 vi5 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) (i + 40)));
    const __m256 vi6 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) (i + 48)));
    const __m256 vi7 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) (i + 56)));
    const __m256 vi8 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) (i + 64)));
    const __m256 vi9 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) (i + 72)));
    const __m256 viA = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) (i + 80)));
    const __m256 viB = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) (i + 88)));
    i += 96;

    const __m256 vx0 = _mm256_sub_ps(vi0, vi_max);
    const __m256 vx1 = _mm256_sub_ps(vi1, vi_max);
    const __m256 vx2 = _mm256_sub_ps(vi2, vi_max);
    const __m256 vx3 = _mm256_sub_ps(vi3, vi_max);
    const __m256 vx4 = _mm256_sub_ps(vi4, vi_max);
    const __m256 vx5 = _mm256_sub_ps(vi5, vi_max);
    const __m256 vx6 = _mm256_sub_ps(vi6, vi_max);
    const __m256 vx7 = _mm256_sub_ps(vi7, vi_max);
    const __m256 vx8 = _mm256_sub_ps(vi8, vi_max);
    const __m256 vx9 = _mm256_sub_ps(vi9, vi_max);
    const __m256 vxA = _mm256_sub_ps(viA, vi_max);
    const __m256 vxB = _mm256_sub_ps(viB, vi_max);

    __m256 vn0 = _mm256_fmadd_ps(vx0, vlog2e, vmagic_bias);
    __m256 vn1 = _mm256_fmadd_ps(vx1, vlog2e, vmagic_bias);
    __m256 vn2 = _mm256_fmadd_ps(vx2, vlog2e, vmagic_bias);
    __m256 vn3 = _mm256_fmadd_ps(vx3, vlog2e, vmagic_bias);
    __m256 vn4 = _mm256_fmadd_ps(vx4, vlog2e, vmagic_bias);
    __m256 vn5 = _mm256_fmadd_ps(vx5, vlog2e, vmagic_bias);
    __m256 vn6 = _mm256_fmadd_ps(vx6, vlog2e, vmagic_bias);
    __m256 vn7 = _mm256_fmadd_ps(vx7, vlog2e, vmagic_bias);
    __m256 vn8 = _mm256_fmadd_ps(vx8, vlog2e, vmagic_bias);
    __m256 vn9 = _mm256_fmadd_ps(vx9, vlog2e, vmagic_bias);
    __m256 vnA = _mm256_fmadd_ps(vxA, vlog2e, vmagic_bias);
    __m256 vnB = _mm256_fmadd_ps(vxB, vlog2e, vmagic_bias);

    const __m256 vs0 = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(vn0), 23));
    const __m256 vs1 = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(vn1), 23));
    const __m256 vs2 = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(vn2), 23));
    const __m256 vs3 = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(vn3), 23));
    const __m256 vs4 = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(vn4), 23));
    const __m256 vs5 = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(vn5), 23));
    const __m256 vs6 = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(vn6), 23));
    const __m256 vs7 = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(vn7), 23));
    const __m256 vs8 = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(vn8), 23));
    const __m256 vs9 = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(vn9), 23));
    const __m256 vsA = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(vnA), 23));
    const __m256 vsB = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(vnB), 23));

    vn0 = _mm256_sub_ps(vn0, vmagic_bias);
    vn1 = _mm256_sub_ps(vn1, vmagic_bias);
    vn2 = _mm256_sub_ps(vn2, vmagic_bias);
    vn3 = _mm256_sub_ps(vn3, vmagic_bias);
    vn4 = _mm256_sub_ps(vn4, vmagic_bias);
    vn5 = _mm256_sub_ps(vn5, vmagic_bias);
    vn6 = _mm256_sub_ps(vn6, vmagic_bias);
    vn7 = _mm256_sub_ps(vn7, vmagic_bias);
    vn8 = _mm256_sub_ps(vn8, vmagic_bias);
    vn9 = _mm256_sub_ps(vn9, vmagic_bias);
    vnA = _mm256_sub_ps(vnA, vmagic_bias);
    vnB = _mm256_sub_ps(vnB, vmagic_bias);

    __m256 vt0 = _mm256_fmadd_ps(vn0, vminus_ln2, vx0);
    __m256 vt1 = _mm256_fmadd_ps(vn1, vminus_ln2, vx1);
    __m256 vt2 = _mm256_fmadd_ps(vn2, vminus_ln2, vx2);
    __m256 vt3 = _mm256_fmadd_ps(vn3, vminus_ln2, vx3);
    __m256 vt4 = _mm256_fmadd_ps(vn4, vminus_ln2, vx4);
    __m256 vt5 = _mm256_fmadd_ps(vn5, vminus_ln2, vx5);
    __m256 vt6 = _mm256_fmadd_ps(vn6, vminus_ln2, vx6);
    __m256 vt7 = _mm256_fmadd_ps(vn7, vminus_ln2, vx7);
    __m256 vt8 = _mm256_fmadd_ps(vn8, vminus_ln2, vx8);
    __m256 vt9 = _mm256_fmadd_ps(vn9, vminus_ln2, vx9);
    __m256 vtA = _mm256_fmadd_ps(vnA, vminus_ln2, vxA);
    __m256 vtB = _mm256_fmadd_ps(vnB, vminus_ln2, vxB);

    const __m256 vp0 = _mm256_fmadd_ps(vc2, vt0, vc1);
    const __m256 vp1 = _mm256_fmadd_ps(vc2, vt1, vc1);
    const __m256 vp2 = _mm256_fmadd_ps(vc2, vt2, vc1);
    const __m256 vp3 = _mm256_fmadd_ps(vc2, vt3, vc1);
    const __m256 vp4 = _mm256_fmadd_ps(vc2, vt4, vc1);
    const __m256 vp5 = _mm256_fmadd_ps(vc2, vt5, vc1);
    const __m256 vp6 = _mm256_fmadd_ps(vc2, vt6, vc1);
    const __m256 vp7 = _mm256_fmadd_ps(vc2, vt7, vc1);
    const __m256 vp8 = _mm256_fmadd_ps(vc2, vt8, vc1);
    const __m256 vp9 = _mm256_fmadd_ps(vc2, vt9, vc1);
    const __m256 vpA = _mm256_fmadd_ps(vc2, vtA, vc1);
    const __m256 vpB = _mm256_fmadd_ps(vc2, vtB, vc1);

    vt0 = _mm256_mul_ps(vt0, vs0);
    vt1 = _mm256_mul_ps(vt1, vs1);
    vt2 = _mm256_mul_ps(vt2, vs2);
    vt3 = _mm256_mul_ps(vt3, vs3);
    vt4 = _mm256_mul_ps(vt4, vs4);
    vt5 = _mm256_mul_ps(vt5, vs5);
    vt6 = _mm256_mul_ps(vt6, vs6);
    vt7 = _mm256_mul_ps(vt7, vs7);
    vt8 = _mm256_mul_ps(vt8, vs8);
    vt9 = _mm256_mul_ps(vt9, vs9);
    vtA = _mm256_mul_ps(vtA, vsA);
    vtB = _mm256_mul_ps(vtB, vsB);

    __m256 vf0 = _mm256_fmadd_ps(vt0, vp0, vs0);
    __m256 vf1 = _mm256_fmadd_ps(vt1, vp1, vs1);
    __m256 vf2 = _mm256_fmadd_ps(vt2, vp2, vs2);
    __m256 vf3 = _mm256_fmadd_ps(vt3, vp3, vs3);
    __m256 vf4 = _mm256_fmadd_ps(vt4, vp4, vs4);
    __m256 vf5 = _mm256_fmadd_ps(vt5, vp5, vs5);
    __m256 vf6 = _mm256_fmadd_ps(vt6, vp6, vs6);
    __m256 vf7 = _mm256_fmadd_ps(vt7, vp7, vs7);
    __m256 vf8 = _mm256_fmadd_ps(vt8, vp8, vs8);
    __m256 vf9 = _mm256_fmadd_ps(vt9, vp9, vs9);
    __m256 vfA = _mm256_fmadd_ps(vtA, vpA, vsA);
    __m256 vfB = _mm256_fmadd_ps(vtB, vpB, vsB);

    vf0 = _mm256_andnot_ps(_mm256_cmp_ps(vx0, vdenorm_cutoff, _CMP_LT_OS), vf0);
    vf1 = _mm256_andnot_ps(_mm256_cmp_ps(vx1, vdenorm_cutoff, _CMP_LT_OS), vf1);
    vf2 = _mm256_andnot_ps(_mm256_cmp_ps(vx2, vdenorm_cutoff, _CMP_LT_OS), vf2);
    vf3 = _mm256_andnot_ps(_mm256_cmp_ps(vx3, vdenorm_cutoff, _CMP_LT_OS), vf3);
    vf4 = _mm256_andnot_ps(_mm256_cmp_ps(vx4, vdenorm_cutoff, _CMP_LT_OS), vf4);
    vf5 = _mm256_andnot_ps(_mm256_cmp_ps(vx5, vdenorm_cutoff, _CMP_LT_OS), vf5);
    vf6 = _mm256_andnot_ps(_mm256_cmp_ps(vx6, vdenorm_cutoff, _CMP_LT_OS), vf6);
    vf7 = _mm256_andnot_ps(_mm256_cmp_ps(vx7, vdenorm_cutoff, _CMP_LT_OS), vf7);
    vf8 = _mm256_andnot_ps(_mm256_cmp_ps(vx8, vdenorm_cutoff, _CMP_LT_OS), vf8);
    vf9 = _mm256_andnot_ps(_mm256_cmp_ps(vx9, vdenorm_cutoff, _CMP_LT_OS), vf9);
    vfA = _mm256_andnot_ps(_mm256_cmp_ps(vxA, vdenorm_cutoff, _CMP_LT_OS), vfA);
    vfB = _mm256_andnot_ps(_mm256_cmp_ps(vxB, vdenorm_cutoff, _CMP_LT_OS), vfB);

    _mm_storeu_si128((__m128i*) o, _mm256_cvtps_ph(vf0, _MM_FROUND_TO_NEAREST_INT));
    _mm_storeu_si128((__m128i*) (o + 8), _mm256_cvtps_ph(vf1, _MM_FROUND_TO_NEAREST_INT));
    _mm_storeu_si128((__m128i*) (o + 16), _mm256_cvtps_ph(vf2, _MM_FROUND_TO_NEAREST_INT));
    _mm_storeu_si128((__m128i*) (o + 24), _mm256_cvtps_ph(vf3, _MM_FROUND_TO_NEAREST_INT));
    _mm_storeu_si128((__m128i*) (o + 32), _mm256_cvtps_ph(vf4, _MM_FROUND_TO_NEAREST_INT));
    _mm_storeu_si128((__m128i*) (o + 40), _mm256_cvtps_ph(vf5, _MM_FROUND_TO_NEAREST_INT));
    _mm_storeu_si128((__m128i*) (o + 48), _mm256_cvtps_ph(vf6, _MM_FROUND_TO_NEAREST_INT));
    _mm_storeu_si128((__m128i*) (o + 56), _mm256_cvtps_ph(vf7, _MM_FROUND_TO_NEAREST_INT));
    _mm_storeu_si128((__m128i*) (o + 64), _mm256_cvtps_ph(vf8, _MM_FROUND_TO_NEAREST_INT));
    _mm_storeu_si128((__m128i*) (o + 72), _mm256_cvtps_ph(vf9, _MM_FROUND_TO_NEAREST_INT));
    _mm_storeu_si128((__m128i*) (o + 80), _mm256_cvtps_ph(vfA, _MM_FROUND_TO_NEAREST_INT));
    _mm_storeu_si128((__m128i*) (o + 88), _mm256_cvtps_ph(vfB, _MM_FROUND_TO_NEAREST_INT));
    o += 96;

    vacc0 = _mm256_add_ps(vacc0, vf0);
    vacc1 = _mm256_add_ps(vacc1, vf1);
    vacc2 = _mm256_add_ps(vacc2, vf2);
    vacc0 = _mm256_add_ps(vacc0, vf3);
    vacc1 = _mm256_add_ps(vacc1, vf4);
    vacc2 = _mm256_add_ps(vacc2, vf5);
    vacc0 = _mm256_add_ps(vacc0, vf6);
    vacc1 = _mm256_add_ps(vacc1, vf7);
    vacc2 = _mm256_add_ps(vacc2, vf8);
    vacc0 = _mm256_add_ps(vacc0, vf9);
    vacc1 = _mm256_add_ps(vacc1, vfA);
    vacc2 = _mm256_add_ps(vacc2, vfB);
  }
  vacc0 = _mm256_add_ps(vacc0, vacc1);
  vacc0 = _mm256_add_ps(vacc0, vacc2);

  __m256 vacc = vacc0;
  for (; batch >= 8 * sizeof(uint16_t); batch -= 8 * sizeof(uint16_t)) {
    const __m256 vi = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i));
    i += 8;

    const __m256 vx = _mm256_sub_ps(vi, vi_max);

    __m256 vn = _mm256_fmadd_ps(vx, vlog2e, vmagic_bias);

    const __m256 vs = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(vn), 23));

    vn = _mm256_sub_ps(vn, vmagic_bias);

    __m256 vt = _mm256_fmadd_ps(vn, vminus_ln2, vx);

    const __m256 vp = _mm256_fmadd_ps(vc2, vt, vc1);
    vt = _mm256_mul_ps(vt, vs);
    __m256 vf = _mm256_fmadd_ps(vt, vp, vs);
    vf = _mm256_andnot_ps(_mm256_cmp_ps(vx, vdenorm_cutoff, _CMP_LT_OS), vf);

    _mm_storeu_si128((__m128i*) o, _mm256_cvtps_ph(vf, _MM_FROUND_TO_NEAREST_INT));
    o += 8;

    vacc = _mm256_add_ps(vacc, vf);
  }
  __m128 vacc_lo = _mm_add_ps(_mm256_castps256_ps128(vacc), _mm256_extractf128_ps(vacc, 1));
  if (batch != 0) {
    assert(batch >= 1 * sizeof(uint16_t));
    assert(batch <= 7 * sizeof(uint16_t));

    const __m256 vi = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i));

    const __m256 vx = _mm256_sub_ps(vi, vi_max);

    __m256 vn = _mm256_fmadd_ps(vx, vlog2e, vmagic_bias);

    const __m256 vs = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(vn), 23));

    vn = _mm256_sub_ps(vn, vmagic_bias);

    __m256 vt = _mm256_fmadd_ps(vn, vminus_ln2, vx);

    const __m256 vp = _mm256_fmadd_ps(vc2, vt, vc1);
    vt = _mm256_mul_ps(vt, vs);
    __m256 vf = _mm256_fmadd_ps(vt, vp, vs);
    vf = _mm256_andnot_ps(_mm256_cmp_ps(vx, vdenorm_cutoff, _CMP_LT_OS), vf);

    __m128i vh = _mm256_cvtps_ph(vf, _MM_FROUND_TO_NEAREST_INT);
    __m128 vf_lo = _mm256_castps256_ps128(vf);
    if (batch & (4 * sizeof(uint16_t))) {
      _mm_storel_epi64((__m128i*) o, vh);
      vh = _mm_unpackhi_epi64(vh, vh);
      vacc_lo = _mm_add_ps(vacc_lo, vf_lo);
      vf_lo = _mm256_extractf128_ps(vf, 1);
      o += 4;
    }
    if (batch & (2 * sizeof(uint16_t))) {
      _mm_storeu_si32(o, vh);
      vh = _mm_srli_epi64(vh, 32);
      vacc_lo = _mm_blend_ps(_mm_add_ps(vacc_lo, vf_lo), vacc_lo, 0xC);
      vf_lo = _mm_movehl_ps(vf_lo, vf_lo);
      o += 2;
    }
    if (batch & (1 * sizeof(uint16_t))) {
      *o = (uint16_t) _mm_extract_epi16(vh, 0);
      vacc_lo = _mm_add_ss(vacc_lo, vf_lo);
    }
  }
  vacc_lo = _mm_add_ps(vacc_lo, _mm_movehl_ps(vacc_lo, vacc_lo));
  vacc_lo = _mm_add_ss(vacc_lo, _mm_movehdup_ps(vacc_lo));
  *((uint16_t*) sum) = (uint16_t) _mm_extract_epi16(_mm_cvtps_ph(vacc_lo, _MM_FROUND_TO_NEAREST_INT), 0);
  _mm256_zeroupper();
}