• 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$assert CHANNEL_TILE % 4 == 0
7$assert CHANNEL_TILE >= 4
8$assert ROW_TILE >= 1
9$ABC = "0123456789ABCDEFGHIJKLMNOPQRSTUVWXYZ"
10#include <assert.h>
11
12#include <arm_neon.h>
13
14#include <xnnpack/math.h>
15#include <xnnpack/prelu.h>
16
17
18void xnn_f32_prelu_ukernel__neon_${ROW_TILE}x${CHANNEL_TILE}(
19    size_t rows,
20    size_t channels,
21    const float*restrict input,
22    size_t input_stride,
23    const float*restrict weights,
24    float*restrict output,
25    size_t output_stride) XNN_DISABLE_TSAN
26{
27  assert(rows != 0);
28  assert(channels != 0);
29  assert(channels % sizeof(float) == 0);
30
31  const float* i0 = input;
32  float* o0 = output;
33  $for M in range(1, ROW_TILE):
34    const float* i${M} = (const float*) ((uintptr_t) i${M-1} + input_stride);
35    float* o${M} = (float*) ((uintptr_t) o${M-1} + output_stride);
36    $if M % 2 == 0:
37      if XNN_UNPREDICTABLE(rows <= ${M}) {
38        i${M} = i${M-1};
39        o${M} = o${M-1};
40      }
41    $else:
42      if XNN_UNPREDICTABLE(rows < ${M+1}) {
43        i${M} = i${M-1};
44        o${M} = o${M-1};
45      }
46
47  const size_t input_increment = input_stride * ${ROW_TILE} - channels;
48  const size_t output_increment = output_stride * ${ROW_TILE} - channels;
49
50  do {
51    const float* w = weights;
52    size_t c = channels;
53    for (; c >= ${CHANNEL_TILE} * sizeof(float); c -= ${CHANNEL_TILE} * sizeof(float)) {
54      $for C in range(0, CHANNEL_TILE, 4):
55        const float32x4_t vw${ABC[C:C+4]} = vld1q_f32(w); w += 4;
56
57      $for M in range(ROW_TILE):
58        $for C in range(0, CHANNEL_TILE, 4):
59          const float32x4_t vi${M}x${ABC[C:C+4]} = vld1q_f32(i${M}); i${M} += 4;
60
61      $for M in range(ROW_TILE):
62        $for C in range(0, CHANNEL_TILE, 4):
63          float32x4_t vacc${M}x${ABC[C:C+4]} = vmulq_f32(vi${M}x${ABC[C:C+4]}, vw${ABC[C:C+4]});
64          const uint32x4_t vm${M}x${ABC[C:C+4]} = vcltq_s32(vreinterpretq_s32_f32(vi${M}x${ABC[C:C+4]}), vmovq_n_s32(0));
65
66      $for M in range(ROW_TILE):
67        $for C in range(0, CHANNEL_TILE, 4):
68          vacc${M}x${ABC[C:C+4]} = vbslq_f32(vm${M}x${ABC[C:C+4]}, vacc${M}x${ABC[C:C+4]}, vi${M}x${ABC[C:C+4]});
69
70      $for M in range(ROW_TILE):
71        $for C in range(0, CHANNEL_TILE, 4):
72          vst1q_f32(o${M}, vacc${M}x${ABC[C:C+4]}); o${M} += 4;
73    }
74    $if CHANNEL_TILE != 4:
75      for (; c >= 4 * sizeof(float); c -= 4 * sizeof(float)) {
76        const float32x4_t vw0123 = vld1q_f32(w); w += 4;
77
78        $for M in range(ROW_TILE):
79          const float32x4_t vi${M}x0123 = vld1q_f32(i${M});
80          i${M} += 4;
81
82        $for M in range(ROW_TILE):
83          float32x4_t vacc${M}x0123 = vmulq_f32(vi${M}x0123, vw0123);
84          const uint32x4_t vm${M}x0123 = vcltq_s32(vreinterpretq_s32_f32(vi${M}x0123), vmovq_n_s32(0));
85
86        $for M in range(ROW_TILE):
87          vacc${M}x0123 = vbslq_f32(vm${M}x0123, vacc${M}x0123, vi${M}x0123);
88
89        $for M in range(ROW_TILE):
90          vst1q_f32(o${M}, vacc${M}x0123); o${M} += 4;
91      }
92    if XNN_UNLIKELY(c != 0) {
93      const float32x4_t vw0123 = vld1q_f32(w); w += 4;
94
95      $for M in range(ROW_TILE):
96        const float32x4_t vi${M}x0123 = vld1q_f32(i${M});
97        i${M} = (const float*) ((uintptr_t) i${M} + c);
98
99      $for M in range(ROW_TILE):
100        float32x4_t vacc${M}x0123 = vmulq_f32(vi${M}x0123, vw0123);
101        const uint32x4_t vm${M}x0123 = vcltq_s32(vreinterpretq_s32_f32(vi${M}x0123), vmovq_n_s32(0));
102
103      $for M in range(ROW_TILE):
104        vacc${M}x0123 = vbslq_f32(vm${M}x0123, vacc${M}x0123, vi${M}x0123);
105
106      $for M in range(ROW_TILE):
107        float32x2_t vacc${M}x01 = vget_low_f32(vacc${M}x0123);
108      if (c & (2 * sizeof(float))) {
109        $for M in range(ROW_TILE):
110          vst1_f32(o${M}, vacc${M}x01); o${M} += 2;
111
112        $for M in range(ROW_TILE):
113          vacc${M}x01 = vget_high_f32(vacc${M}x0123);
114      }
115      if (c & (1 * sizeof(float))) {
116        $for M in range(ROW_TILE):
117          vst1_lane_f32(o${M}, vacc${M}x01, 0); o${M} += 1;
118      }
119    }
120    $for M in range(ROW_TILE):
121      i${M} = (const float*) ((uintptr_t) i${M} + input_increment);
122      o${M} = (float*) ((uintptr_t) o${M} + output_increment);
123      $if M % 2 == 1:
124        if XNN_UNPREDICTABLE(rows < ${ROW_TILE+M+1}) {
125          i${M} = i${M-1};
126          o${M} = o${M-1};
127        }
128      $elif M != 0:
129        if XNN_UNPREDICTABLE(rows <= ${ROW_TILE+M}) {
130          i${M} = i${M-1};
131          o${M} = o${M-1};
132        }
133    rows = doz(rows, ${ROW_TILE});
134  } while (rows != 0);
135}
136