File size: 5,657 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
// Auto-generated file. Do not edit!
//   Template: src/f32-gemm/avx-shuffle4.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_gemm_minmax_ukernel_1x16s4__fma3_broadcast(
    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 union xnn_f32_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
{
  assert(mr != 0);
  assert(mr <= 1);
  assert(nc != 0);
  assert(kc != 0);
  assert(kc % sizeof(float) == 0);
  assert(a != NULL);
  assert(w != NULL);
  assert(c != NULL);

  const float* a0 = a;
  float* c0 = c;

  do {
    __m256 vacc0x01234567 = _mm256_load_ps(w + 0);
    __m256 vacc0x89ABCDEF = _mm256_load_ps(w + 8);
    w += 16;

    size_t k = kc;
    while (k >= 4 * sizeof(float)) {
      __m256 va0 = _mm256_broadcast_ps((const __m128*) a0);
      a0 += 4;

      const __m256 vb01234567c0 = _mm256_load_ps(w + 0);
      const __m256 vb89ABCDEFc0 = _mm256_load_ps(w + 8);

      vacc0x01234567 = _mm256_fmadd_ps(va0, vb01234567c0, vacc0x01234567);
      vacc0x89ABCDEF = _mm256_fmadd_ps(va0, vb89ABCDEFc0, vacc0x89ABCDEF);

      va0 = _mm256_permute_ps(va0, _MM_SHUFFLE(0, 3, 2, 1));
      const __m256 vb01234567c1 = _mm256_load_ps(w + 16);
      const __m256 vb89ABCDEFc1 = _mm256_load_ps(w + 24);

      vacc0x01234567 = _mm256_fmadd_ps(va0, vb01234567c1, vacc0x01234567);
      vacc0x89ABCDEF = _mm256_fmadd_ps(va0, vb89ABCDEFc1, vacc0x89ABCDEF);

      va0 = _mm256_permute_ps(va0, _MM_SHUFFLE(0, 3, 2, 1));
      const __m256 vb01234567c2 = _mm256_load_ps(w + 32);
      const __m256 vb89ABCDEFc2 = _mm256_load_ps(w + 40);

      vacc0x01234567 = _mm256_fmadd_ps(va0, vb01234567c2, vacc0x01234567);
      vacc0x89ABCDEF = _mm256_fmadd_ps(va0, vb89ABCDEFc2, vacc0x89ABCDEF);

      va0 = _mm256_permute_ps(va0, _MM_SHUFFLE(0, 3, 2, 1));
      const __m256 vb01234567c3 = _mm256_load_ps(w + 48);
      const __m256 vb89ABCDEFc3 = _mm256_load_ps(w + 56);

      vacc0x01234567 = _mm256_fmadd_ps(va0, vb01234567c3, vacc0x01234567);
      vacc0x89ABCDEF = _mm256_fmadd_ps(va0, vb89ABCDEFc3, vacc0x89ABCDEF);


      w += 64;
      k -= 4 * sizeof(float);
    }
    if XNN_UNLIKELY(k != 0) {
      __m256 va0 = _mm256_broadcast_ps((const __m128*) a0);
      a0 = (const float*) ((uintptr_t) a0 + k);

      const __m256 vzero = _mm256_setzero_ps();
      const __m256 vb01234567c0 = _mm256_load_ps(w + 0);
      const __m256 vb89ABCDEFc0 = _mm256_load_ps(w + 8);

      vacc0x01234567 = _mm256_fmadd_ps(_mm256_and_ps(va0, _mm256_cmp_ps(vb01234567c0, vzero, _CMP_NEQ_OQ)), vb01234567c0, vacc0x01234567);
      vacc0x89ABCDEF = _mm256_fmadd_ps(_mm256_and_ps(va0, _mm256_cmp_ps(vb89ABCDEFc0, vzero, _CMP_NEQ_OQ)), vb89ABCDEFc0, vacc0x89ABCDEF);

      va0 = _mm256_permute_ps(va0, _MM_SHUFFLE(0, 3, 2, 1));
      const __m256 vb01234567c1 = _mm256_load_ps(w + 16);
      const __m256 vb89ABCDEFc1 = _mm256_load_ps(w + 24);

      vacc0x01234567 = _mm256_fmadd_ps(_mm256_and_ps(va0, _mm256_cmp_ps(vb01234567c1, vzero, _CMP_NEQ_OQ)), vb01234567c1, vacc0x01234567);
      vacc0x89ABCDEF = _mm256_fmadd_ps(_mm256_and_ps(va0, _mm256_cmp_ps(vb89ABCDEFc1, vzero, _CMP_NEQ_OQ)), vb89ABCDEFc1, vacc0x89ABCDEF);

      va0 = _mm256_permute_ps(va0, _MM_SHUFFLE(0, 3, 2, 1));
      const __m256 vb01234567c2 = _mm256_load_ps(w + 32);
      const __m256 vb89ABCDEFc2 = _mm256_load_ps(w + 40);

      vacc0x01234567 = _mm256_fmadd_ps(_mm256_and_ps(va0, _mm256_cmp_ps(vb01234567c2, vzero, _CMP_NEQ_OQ)), vb01234567c2, vacc0x01234567);
      vacc0x89ABCDEF = _mm256_fmadd_ps(_mm256_and_ps(va0, _mm256_cmp_ps(vb89ABCDEFc2, vzero, _CMP_NEQ_OQ)), vb89ABCDEFc2, vacc0x89ABCDEF);

      va0 = _mm256_permute_ps(va0, _MM_SHUFFLE(0, 3, 2, 1));
      const __m256 vb01234567c3 = _mm256_load_ps(w + 48);
      const __m256 vb89ABCDEFc3 = _mm256_load_ps(w + 56);

      vacc0x01234567 = _mm256_fmadd_ps(_mm256_and_ps(va0, _mm256_cmp_ps(vb01234567c3, vzero, _CMP_NEQ_OQ)), vb01234567c3, vacc0x01234567);
      vacc0x89ABCDEF = _mm256_fmadd_ps(_mm256_and_ps(va0, _mm256_cmp_ps(vb89ABCDEFc3, vzero, _CMP_NEQ_OQ)), vb89ABCDEFc3, vacc0x89ABCDEF);


      w += 64;
    }

    const __m256 vmin = _mm256_load_ps(params->avx.min);
    vacc0x01234567 = _mm256_max_ps(vmin, vacc0x01234567);
    vacc0x89ABCDEF = _mm256_max_ps(vmin, vacc0x89ABCDEF);

    const __m256 vmax = _mm256_load_ps(params->avx.max);
    vacc0x01234567 = _mm256_min_ps(vmax, vacc0x01234567);
    vacc0x89ABCDEF = _mm256_min_ps(vmax, vacc0x89ABCDEF);

    if XNN_LIKELY(nc >= 16) {
      _mm256_storeu_ps(c0, vacc0x01234567);
      _mm256_storeu_ps(c0 + 8, vacc0x89ABCDEF);
      c0 = (float*) ((uintptr_t) c0 + cn_stride);

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

      nc -= 16;
    } else {
      if (nc & 8) {
        _mm256_storeu_ps(c0, vacc0x01234567);

        vacc0x01234567 = vacc0x89ABCDEF;

        c0 += 8;
      }
      __m128 vacc0x0123 = _mm256_castps256_ps128(vacc0x01234567);
      if (nc & 4) {
        _mm_storeu_ps(c0, vacc0x0123);

        vacc0x0123 = _mm256_extractf128_ps(vacc0x01234567, 1);

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

        vacc0x0123 = _mm_movehl_ps(vacc0x0123, vacc0x0123);

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

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