xref: /aosp_15_r20/external/XNNPACK/src/f32-prelu/wasmsimd-bitselect.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$assert CHANNEL_TILE % 4 == 0
7$assert CHANNEL_TILE >= 4
8$assert ROW_TILE >= 1
9$ABC = "0123456789ABCDEFGHIJKLMNOPQRSTUVWXYZ"
10#include <assert.h>
11
12#include <wasm_simd128.h>
13
14#include <xnnpack/math.h>
15#include <xnnpack/prelu.h>
16
17
18void xnn_f32_prelu_ukernel__wasmsimd_bitselect_${ROW_TILE}x${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 -= ${CHANNEL_TILE} * sizeof(float)) {
56      const v128_t vw${ABC[0:4]} = wasm_v128_load(w);
57      $for C in range(4, CHANNEL_TILE, 4):
58        const v128_t vw${ABC[C:C+4]} = wasm_v128_load(w + ${C});
59      w += ${CHANNEL_TILE};
60
61      $for M in range(ROW_TILE):
62        const v128_t vi${M}x${ABC[0:4]} = wasm_v128_load(i${M});
63        $for C in range(4, CHANNEL_TILE, 4):
64          const v128_t vi${M}x${ABC[C:C+4]} = wasm_v128_load(i${M} + ${C});
65        i${M} += ${CHANNEL_TILE};
66
67      $for M in range(ROW_TILE):
68        $for C in range(0, CHANNEL_TILE, 4):
69          v128_t vacc${M}x${ABC[C:C+4]} = wasm_f32x4_mul(vi${M}x${ABC[C:C+4]}, vw${ABC[C:C+4]});
70          const v128_t vmask${M}x${ABC[C:C+4]} = wasm_i32x4_shr(vi${M}x${ABC[C:C+4]}, 31);
71
72      $for M in range(ROW_TILE):
73        $for C in range(0, CHANNEL_TILE, 4):
74          vacc${M}x${ABC[C:C+4]} = wasm_v128_bitselect(vacc${M}x${ABC[C:C+4]}, vi${M}x${ABC[C:C+4]}, vmask${M}x${ABC[C:C+4]});
75
76      $for M in range(ROW_TILE):
77        wasm_v128_store(o${M}, vacc${M}x${ABC[0:4]});
78        $for C in range(4, CHANNEL_TILE, 4):
79          wasm_v128_store(o${M} + ${C}, vacc${M}x${ABC[C:C+4]});
80        o${M} += ${CHANNEL_TILE};
81    }
82    $if CHANNEL_TILE > 4:
83      for (; c >= 4 * sizeof(float); c -= 4 * sizeof(float)) {
84        const v128_t vw0123 = wasm_v128_load(w);
85        w += 4;
86
87        $for M in range(ROW_TILE):
88          const v128_t vi${M}x0123 = wasm_v128_load(i${M});
89          i${M} += 4;
90
91        $for M in range(ROW_TILE):
92          v128_t vacc${M}x0123 = wasm_f32x4_mul(vi${M}x0123, vw0123);
93          const v128_t vmask${M}x0123 = wasm_i32x4_shr(vi${M}x0123, 31);
94
95        $for M in range(ROW_TILE):
96          vacc${M}x0123 = wasm_v128_bitselect(vacc${M}x0123, vi${M}x0123, vmask${M}x0123);
97
98        $for M in range(ROW_TILE):
99          wasm_v128_store(o${M}, vacc${M}x0123);
100          o${M} += 4;
101      }
102    if XNN_UNLIKELY(c != 0) {
103      const v128_t vw0123 = wasm_v128_load(w);
104      w = (const float*) ((uintptr_t) w + c);
105
106      $for M in range(ROW_TILE):
107        const v128_t vi${M}x0123 = wasm_v128_load(i${M});
108        i${M} = (const float*) ((uintptr_t) i${M} + c);
109
110      $for M in range(ROW_TILE):
111        v128_t vacc${M}x0123 = wasm_f32x4_mul(vi${M}x0123, vw0123);
112        const v128_t vmask${M}x0123 = wasm_i32x4_shr(vi${M}x0123, 31);
113
114      $for M in range(ROW_TILE):
115        vacc${M}x0123 = wasm_v128_bitselect(vacc${M}x0123, vi${M}x0123, vmask${M}x0123);
116
117      if (c & (2 * sizeof(float))) {
118        $for M in range(ROW_TILE):
119          *((double*) o${M}) = wasm_f64x2_extract_lane(vacc${M}x0123, 0);
120
121        $for M in range(ROW_TILE):
122          vacc${M}x0123 = wasm_v32x4_shuffle(vacc${M}x0123, vacc${M}x0123, 2, 3, 2, 3);
123
124        $for M in range(ROW_TILE):
125          o${M} += 2;
126      }
127      if (c & (1 * sizeof(float))) {
128        $for M in range(ROW_TILE):
129          *o${M} = wasm_f32x4_extract_lane(vacc${M}x0123, 0);
130
131        $for M in range(ROW_TILE):
132          o${M} += 1;
133      }
134    }
135    $for M in range(ROW_TILE):
136      i${M} = (const float*) ((uintptr_t) i${M} + input_increment);
137      o${M} = (float*) ((uintptr_t) o${M} + output_increment);
138    rows = doz(rows, ${ROW_TILE});
139  } while (rows != 0);
140}
141