xref: /aosp_15_r20/external/XNNPACK/src/f32-prelu/neon.c.in (revision 4bdc94577ba0e567308109d787f7fec7b531ce36) 
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_OOB_READS
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
37  const size_t input_increment = input_stride * ${ROW_TILE} - channels;
38  const size_t output_increment = output_stride * ${ROW_TILE} - channels;
39
40  do {
41    $for M in range(1, ROW_TILE):
42      $if M % 2 == 0:
43        if XNN_UNPREDICTABLE(rows <= ${M}) {
44          i${M} = i${M-1};
45          o${M} = o${M-1};
46        }
47      $else:
48        if XNN_UNPREDICTABLE(rows < ${M+1}) {
49          i${M} = i${M-1};
50          o${M} = o${M-1};
51        }
52
53    const float* w = weights;
54    size_t c = channels;
55    for (; c >= ${CHANNEL_TILE} * sizeof(float); c -= ${CHANNEL_TILE} * sizeof(float)) {
56      $for C in range(0, CHANNEL_TILE, 4):
57        const float32x4_t vw${ABC[C:C+4]} = vld1q_f32(w); w += 4;
58
59      $for M in range(ROW_TILE):
60        $for C in range(0, CHANNEL_TILE, 4):
61          const float32x4_t vi${M}x${ABC[C:C+4]} = vld1q_f32(i${M}); i${M} += 4;
62
63      $for M in range(ROW_TILE):
64        $for C in range(0, CHANNEL_TILE, 4):
65          float32x4_t vacc${M}x${ABC[C:C+4]} = vmulq_f32(vi${M}x${ABC[C:C+4]}, vw${ABC[C:C+4]});
66          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));
67
68      $for M in range(ROW_TILE):
69        $for C in range(0, CHANNEL_TILE, 4):
70          vacc${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]});
71
72      $for M in range(ROW_TILE):
73        $for C in range(0, CHANNEL_TILE, 4):
74          vst1q_f32(o${M}, vacc${M}x${ABC[C:C+4]}); o${M} += 4;
75    }
76    $if CHANNEL_TILE != 4:
77      for (; c >= 4 * sizeof(float); c -= 4 * sizeof(float)) {
78        const float32x4_t vw0123 = vld1q_f32(w); w += 4;
79
80        $for M in range(ROW_TILE):
81          const float32x4_t vi${M}x0123 = vld1q_f32(i${M});
82          i${M} += 4;
83
84        $for M in range(ROW_TILE):
85          float32x4_t vacc${M}x0123 = vmulq_f32(vi${M}x0123, vw0123);
86          const uint32x4_t vm${M}x0123 = vcltq_s32(vreinterpretq_s32_f32(vi${M}x0123), vmovq_n_s32(0));
87
88        $for M in range(ROW_TILE):
89          vacc${M}x0123 = vbslq_f32(vm${M}x0123, vacc${M}x0123, vi${M}x0123);
90
91        $for M in range(ROW_TILE):
92          vst1q_f32(o${M}, vacc${M}x0123); o${M} += 4;
93      }
94    if XNN_UNLIKELY(c != 0) {
95      const float32x4_t vw0123 = vld1q_f32(w); w += 4;
96
97      $for M in range(ROW_TILE):
98        const float32x4_t vi${M}x0123 = vld1q_f32(i${M});
99        i${M} = (const float*) ((uintptr_t) i${M} + c);
100
101      $for M in range(ROW_TILE):
102        float32x4_t vacc${M}x0123 = vmulq_f32(vi${M}x0123, vw0123);
103        const uint32x4_t vm${M}x0123 = vcltq_s32(vreinterpretq_s32_f32(vi${M}x0123), vmovq_n_s32(0));
104
105      $for M in range(ROW_TILE):
106        vacc${M}x0123 = vbslq_f32(vm${M}x0123, vacc${M}x0123, vi${M}x0123);
107
108      $for M in range(ROW_TILE):
109        float32x2_t vacc${M}x01 = vget_low_f32(vacc${M}x0123);
110      if (c & (2 * sizeof(float))) {
111        $for M in range(ROW_TILE):
112          vst1_f32(o${M}, vacc${M}x01); o${M} += 2;
113
114        $for M in range(ROW_TILE):
115          vacc${M}x01 = vget_high_f32(vacc${M}x0123);
116      }
117      if (c & (1 * sizeof(float))) {
118        $for M in range(ROW_TILE):
119          vst1_lane_f32(o${M}, vacc${M}x01, 0); o${M} += 1;
120      }
121    }
122    $for M in range(ROW_TILE):
123      i${M} = (const float*) ((uintptr_t) i${M} + input_increment);
124      o${M} = (float*) ((uintptr_t) o${M} + output_increment);
125    rows = doz(rows, ${ROW_TILE});
126  } while (rows != 0);
127}
128