File size: 5,791 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
// 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.

$assert NR % 8 == 0
$ABC = "0123456789ABCDEFGHIJKLMNOPQRSTUVWXYZ"
$assert ACCTYPE in ["F16", "F32"]
$ACC_SUFFIX = "_f32acc" if ACCTYPE == "F32" else ""
#include <assert.h>

#include <immintrin.h>

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


void xnn_f16${ACC_SUFFIX}_gemm_minmax_ukernel_${MR}x${NR}__avx2_broadcast(
    size_t mr,
    size_t nc,
    size_t kc,
    const void* restrict a,
    size_t a_stride,
    const void* restrict w,
    void* restrict c,
    size_t cm_stride,
    size_t cn_stride,
    const union xnn_f16_minmax_params params[restrict XNN_MIN_ELEMENTS(1)])
{
  assert(mr != 0);
  assert(mr <= ${MR});
  assert(nc != 0);
  assert(kc != 0);
  assert(kc % sizeof(uint16_t) == 0);
  assert(a != NULL);
  assert(w != NULL);
  assert(c != NULL);

  const uint16_t* a0 = a;
  uint16_t* c0 = c;
  $for M in range(1, MR):
    const uint16_t* a${M} = (const uint16_t*) ((uintptr_t) a${M-1} + a_stride);
    uint16_t* c${M} = (uint16_t*) ((uintptr_t) c${M-1} + cm_stride);
    $if M % 2 == 0:
      if XNN_UNPREDICTABLE(mr <= ${M}) {
        a${M} = a${M-1};
        c${M} = c${M-1};
      }
    $elif M + 1 == MR:
      if XNN_UNPREDICTABLE(mr != ${M+1}) {
        a${M} = a${M-1};
        c${M} = c${M-1};
      }
    $else:
      if XNN_UNPREDICTABLE(mr < ${M+1}) {
        a${M} = a${M-1};
        c${M} = c${M-1};
      }

  do {
    __m256 vacc0x${ABC[0:8]} = _mm256_cvtph_ps(_mm_load_si128((const __m128i*) w));
    $for N in range(8, NR, 8):
      __m256 vacc0x${ABC[N:N+8]} = _mm256_cvtph_ps(_mm_load_si128((const __m128i*) ((const uint16_t*) w + ${N})));
    $for M in range(1, MR):
      $for N in range(0, NR, 8):
        __m256 vacc${M}x${ABC[N:N+8]} = vacc0x${ABC[N:N+8]};
    w = (const uint16_t*) w + ${NR};

    size_t k = kc;
    do {
      $for M in range(MR):
        const __m256 va${M} = _mm256_cvtph_ps(_mm_set1_epi16((short) *a${M}));
        a${M} += 1;

      const __m256 vb${ABC[0:8]} = _mm256_cvtph_ps(_mm_load_si128((const __m128i*) w));
      $for N in range(8, NR, 8):
        const __m256 vb${ABC[N:N+8]} = _mm256_cvtph_ps(_mm_load_si128((const __m128i*) ((const uint16_t*) w + ${N})));
      w = (const uint16_t*) w + ${NR};

      $for N in range(0, NR, 8):
        $for M in range(MR):
          $if ACCTYPE == "F32":
            vacc${M}x${ABC[N:N+8]} = _mm256_fmadd_ps(va${M}, vb${ABC[N:N+8]}, vacc${M}x${ABC[N:N+8]});
          $else:
            vacc${M}x${ABC[N:N+8]} = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_fmadd_ps(va${M}, vb${ABC[N:N+8]}, vacc${M}x${ABC[N:N+8]}), _MM_FROUND_TO_NEAREST_INT));

      k -= sizeof(uint16_t);
    } while (k != 0);

    const __m256 vmin = _mm256_load_ps(params->avx.min);
    $for N in range(0, NR, 8):
      $for M in range(MR):
        vacc${M}x${ABC[N:N+8]} = _mm256_max_ps(vacc${M}x${ABC[N:N+8]}, vmin);

    const __m256 vmax = _mm256_load_ps(params->avx.max);
    $for N in range(0, NR, 8):
      $for M in range(MR):
        vacc${M}x${ABC[N:N+8]} = _mm256_min_ps(vacc${M}x${ABC[N:N+8]}, vmax);

    if XNN_LIKELY(nc >= ${NR}) {
      $for M in range(MR):
        _mm_storeu_si128((__m128i*) c${M}, _mm256_cvtps_ph(vacc${M}x${ABC[0:8]}, _MM_FROUND_TO_NEAREST_INT));
        $for N in range(8, NR, 8):
          _mm_storeu_si128((__m128i*) (c${M} + ${N}), _mm256_cvtps_ph(vacc${M}x${ABC[N:N+8]}, _MM_FROUND_TO_NEAREST_INT));
        c${M} = (uint16_t*) ((uintptr_t) c${M} + cn_stride);

      $for M in range(MR):
        a${M} = (const uint16_t*) ((uintptr_t) a${M} - kc);

      nc -= ${NR};
    } else {
      $for LOG2N in reversed(range(NR.bit_length())):
        $if LOG2N == 3:
          $for M in range(MR):
            __m128i vh${M}x${ABC[0:8]} = _mm256_cvtps_ph(vacc${M}x${ABC[0:8]}, _MM_FROUND_TO_NEAREST_INT);
        $if NR != 1 << LOG2N:
          if (nc & ${1 << LOG2N}) {
            $if LOG2N >= 4:
              $for M in range(MR):
                _mm_storeu_si128((__m128i*) c${M}, _mm256_cvtps_ph(vacc${M}x${ABC[0:8]}, _MM_FROUND_TO_NEAREST_INT));
                $for N in range(8, 1 << LOG2N, 8):
                  _mm_storeu_si128((__m128i*) (c${M} + ${N}), _mm256_cvtps_ph(vacc${M}x${ABC[N:N+8]}, _MM_FROUND_TO_NEAREST_INT));

              $for M in range(MR):
                $for N in range(0, NR - (1 << LOG2N), 8):
                  vacc${M}x${ABC[N:N+8]} = vacc${M}x${ABC[N + (1 << LOG2N):N + (1 << LOG2N)+8]};

              $for M in range(MR):
                c${M} += ${1 << LOG2N};
            $elif LOG2N == 3:
              $for M in range(MR):
                _mm_storeu_si128((__m128i*) c${M}, vh${M}x${ABC[0:8]});

              $for M in range(MR):
                vh${M}x${ABC[0:8]} = _mm256_cvtps_ph(vacc${M}x${ABC[8:16]}, _MM_FROUND_TO_NEAREST_INT);

              $for M in range(MR):
                c${M} += ${1 << LOG2N};
            $elif LOG2N == 2:
              $for M in range(MR):
                _mm_storel_epi64((__m128i*) c${M}, vh${M}x${ABC[0:8]});

              $for M in range(MR):
                vh${M}x${ABC[0:8]} = _mm_unpackhi_epi64(vh${M}x${ABC[0:8]}, vh${M}x${ABC[0:8]});

              $for M in range(MR):
                c${M} += 4;
            $elif LOG2N == 1:
              $for M in range(MR):
                _mm_storeu_si32(c${M}, vh${M}x${ABC[0:8]});

              $for M in range(MR):
                vh${M}x${ABC[0:8]} = _mm_srli_epi64(vh${M}x${ABC[0:8]}, 32);

              $for M in range(MR):
                c${M} += 2;
            $elif LOG2N == 0:
              $for M in range(MR):
                *c${M} = (uint16_t) _mm_extract_epi16(vh${M}x${ABC[0:8]}, 0);
          }

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