• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
1// Copyright 2019 Google LLC
2//
3// This source code is licensed under the BSD-style license found in the
4// LICENSE file in the root directory of this source tree.
5
6#include <assert.h>
7
8#include <xnnpack/math.h>
9#include <xnnpack/ppmm.h>
10
11
12void xnn_f32_ppmm_minmax_ukernel_${MR}x${NR}__scalar(
13  size_t mr,
14  size_t nc,
15  size_t kc,
16  const float*restrict a,
17  const float*restrict w,
18  float*restrict c,
19  size_t cm_stride,
20  size_t cn_stride,
21  const union xnn_f32_minmax_params params[restrict XNN_MIN_ELEMENTS(1)])
22{
23  assert(mr != 0);
24  assert(mr <= ${MR});
25  assert(nc != 0);
26  assert(kc != 0);
27  assert(kc % sizeof(float) == 0);
28
29  float* c0 = c;
30  $for M in range(1, MR):
31    float* c${M} = (float*) ((uintptr_t) c${M-1} + cm_stride);
32    $if M % 2 == 0:
33      if XNN_UNPREDICTABLE(mr <= ${M}) {
34        c${M} = c${M-1};
35      }
36    $elif M + 1 == MR:
37      if XNN_UNPREDICTABLE(mr != ${M+1}) {
38        c${M} = c${M-1};
39      }
40    $else:
41      if XNN_UNPREDICTABLE(mr < ${M+1}) {
42        c${M} = c${M-1};
43      }
44
45  do {
46    $for N in range(NR):
47      float vacc0x${N} = w[${N}];
48    $for M in range(1, MR):
49      $for N in range(NR):
50        float vacc${M}x${N} = vacc0x${N};
51    w += ${NR};
52
53    size_t k = kc;
54    do {
55      $for M in range(MR):
56        const float va${M} = a[${M}];
57      a += ${MR};
58
59      $for N in range(NR):
60        const float vb${N} = w[${N}];
61      w += ${NR};
62
63      $for N in range(NR):
64        $for M in range(MR):
65          vacc${M}x${N} += va${M} * vb${N};
66
67      k -= sizeof(float);
68    } while (k != 0);
69
70    const float vmax = params->scalar.max;
71    $for N in range(NR):
72      $for M in range(MR):
73        vacc${M}x${N} = math_min_f32(vacc${M}x${N}, vmax);
74
75    const float vmin = params->scalar.min;
76    $for N in range(NR):
77      $for M in range(MR):
78        vacc${M}x${N} = math_max_f32(vacc${M}x${N}, vmin);
79
80    if XNN_LIKELY(nc >= ${NR}) {
81      $for M in reversed(range(MR)):
82        $for N in range(NR):
83          c${M}[${N}] = vacc${M}x${N};
84
85      a = (const float*) ((uintptr_t) a - kc * ${MR});
86
87      $for M in reversed(range(MR)):
88        c${M} = (float*) ((uintptr_t) c${M} + cn_stride);
89
90      nc -= ${NR};
91    } else {
92      $for LOG2N in reversed(range(NR.bit_length())):
93        $if NR != 1 << LOG2N:
94          if (nc & ${1 << LOG2N}) {
95            $if LOG2N != 0:
96              $for M in reversed(range(MR)):
97                $for N in range(1 << LOG2N):
98                  c${M}[${N}] = vacc${M}x${N};
99
100              $for M in reversed(range(MR)):
101                $for N in range(1 << (LOG2N - 1)):
102                  vacc${M}x${N} = vacc${M}x${N + (1 << LOG2N)};
103
104              $for M in reversed(range(MR)):
105                c${M} += ${1 << LOG2N};
106            $else:
107              $for M in reversed(range(MR)):
108                *c${M} = vacc${M}x0;
109          }
110
111      nc = 0;
112    }
113  } while (nc != 0);
114}
115