xref: /aosp_15_r20/external/XNNPACK/src/f32-prelu/wasm.c.in (revision 4bdc94577ba0e567308109d787f7fec7b531ce36) 
1// Copyright 2020 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$assert CHANNEL_TILE > 0
7$assert ROW_TILE >= 1
8$ABC = "0123456789ABCDEFGHIJKLMNOPQRSTUVWXYZ"
9#include <assert.h>
10
11#include <xnnpack/math.h>
12#include <xnnpack/prelu.h>
13
14
15void xnn_f32_prelu_ukernel__wasm_${ROW_TILE}x${CHANNEL_TILE}(
16    size_t rows,
17    size_t channels,
18    const float*restrict input,
19    size_t input_stride,
20    const float*restrict weights,
21    float*restrict output,
22    size_t output_stride)
23{
24  assert(rows != 0);
25  assert(channels != 0);
26  assert(channels % sizeof(float) == 0);
27
28  const float* i0 = input;
29  float* o0 = output;
30  $for M in range(1, ROW_TILE):
31    const float* i${M} = (const float*) ((uintptr_t) i${M-1} + input_stride);
32    float* o${M} = (float*) ((uintptr_t) o${M-1} + output_stride);
33
34  const size_t input_increment = input_stride * ${ROW_TILE} - channels;
35  const size_t output_increment = output_stride * ${ROW_TILE} - channels;
36
37  const float vzero = 0.0f;
38  do {
39    $for M in range(1, ROW_TILE):
40      $if M % 2 == 0:
41        if XNN_UNPREDICTABLE(rows <= ${M}) {
42          i${M} = i${M-1};
43          o${M} = o${M-1};
44        }
45      $else:
46        if XNN_UNPREDICTABLE(rows < ${M+1}) {
47          i${M} = i${M-1};
48          o${M} = o${M-1};
49        }
50
51    const float* w = weights;
52    size_t c = channels;
53    $if CHANNEL_TILE > 1:
54      for (; c >= ${CHANNEL_TILE} * sizeof(float); c -= ${CHANNEL_TILE} * sizeof(float)) {
55        $for C in range(CHANNEL_TILE):
56          const float vw${ABC[C]} = w[${C}];
57
58        $for M in range(ROW_TILE):
59          $for C in range(CHANNEL_TILE):
60            float vi${M}x${ABC[C]} = i${M}[${C}];
61          i${M} += ${CHANNEL_TILE};
62
63        $for M in range(ROW_TILE):
64          $for C in range(CHANNEL_TILE):
65            float vacc${M}x${ABC[C]} = __builtin_wasm_max_f32(vi${M}x${ABC[C]}, vzero);
66            vi${M}x${ABC[C]} = __builtin_wasm_min_f32(vi${M}x${ABC[C]}, vzero);
67
68        $for M in range(ROW_TILE):
69          $for C in range(CHANNEL_TILE):
70            vacc${M}x${ABC[C]} += vi${M}x${ABC[C]} * vw${ABC[C]};
71
72        $for M in range(ROW_TILE):
73          $for C in range(CHANNEL_TILE):
74            o${M}[${C}] = vacc${M}x${ABC[C]};
75          o${M} += ${CHANNEL_TILE};
76
77        w += ${CHANNEL_TILE};
78      }
79      for (; c != 0; c -= sizeof(float)) {
80        const float vw = *w++;
81
82        $for M in range(ROW_TILE):
83          float vi${M} = *i${M}++;
84
85        $for M in range(ROW_TILE):
86          float vacc${M} = __builtin_wasm_max_f32(vi${M}, vzero);
87          vi${M} = __builtin_wasm_min_f32(vi${M}, vzero);
88
89        $for M in range(ROW_TILE):
90          vacc${M} += vi${M} * vw;
91
92        $for M in range(ROW_TILE):
93          *o${M}++ = vacc${M};
94      }
95    $else:
96      do {
97        const float vw = *w++;
98
99        $for M in range(ROW_TILE):
100          float vi${M} = *i${M}++;
101
102        $for M in range(ROW_TILE):
103          float vacc${M} = __builtin_wasm_max_f32(vi${M}, vzero);
104          vi${M} = __builtin_wasm_min_f32(vi${M}, vzero);
105
106        $for M in range(ROW_TILE):
107          vacc${M} += vi${M} * vw;
108
109        $for M in range(ROW_TILE):
110          *o${M}++ = vacc${M};
111
112        c -= sizeof(float);
113      } while (c != 0);
114    $for M in range(ROW_TILE):
115      i${M} = (const float*) ((uintptr_t) i${M} + input_increment);
116      o${M} = (float*) ((uintptr_t) o${M} + output_increment);
117    rows = doz(rows, ${ROW_TILE});
118  } while (rows != 0);
119}
120