xref: /aosp_15_r20/external/XNNPACK/src/f32-prelu/wasmsimd-minmax.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_minmax_${ROW_TILE}x${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  const v128_t vzero = wasm_i32x4_const_splat(0);
41  do {
42    $for M in range(1, ROW_TILE):
43      $if M % 2 == 0:
44        if XNN_UNPREDICTABLE(rows <= ${M}) {
45          i${M} = i${M-1};
46          o${M} = o${M-1};
47        }
48      $else:
49        if XNN_UNPREDICTABLE(rows < ${M+1}) {
50          i${M} = i${M-1};
51          o${M} = o${M-1};
52        }
53
54    const float* w = weights;
55    size_t c = channels;
56    for (; c >= ${CHANNEL_TILE} * sizeof(float); c -= ${CHANNEL_TILE} * sizeof(float)) {
57      const v128_t vw${ABC[0:4]} = wasm_v128_load(w);
58      $for C in range(4, CHANNEL_TILE, 4):
59        const v128_t vw${ABC[C:C+4]} = wasm_v128_load(w + ${C});
60      w += ${CHANNEL_TILE};
61
62      $for M in range(ROW_TILE):
63        v128_t vi${M}x${ABC[0:4]} = wasm_v128_load(i${M});
64        $for C in range(4, CHANNEL_TILE, 4):
65          v128_t vi${M}x${ABC[C:C+4]} = wasm_v128_load(i${M} + ${C});
66        i${M} += ${CHANNEL_TILE};
67
68      $for M in range(ROW_TILE):
69        $for C in range(0, CHANNEL_TILE, 4):
70          v128_t vacc${M}x${ABC[C:C+4]} = wasm_i32x4_max(vi${M}x${ABC[C:C+4]}, vzero);
71          vi${M}x${ABC[C:C+4]} = wasm_i32x4_min(vi${M}x${ABC[C:C+4]}, vzero);
72
73      $for M in range(ROW_TILE):
74        $for C in range(0, CHANNEL_TILE, 4):
75          vacc${M}x${ABC[C:C+4]} = wasm_f32x4_add(vacc${M}x${ABC[C:C+4]}, wasm_f32x4_mul(vi${M}x${ABC[C:C+4]}, vw${ABC[C:C+4]}));
76
77      $for M in range(ROW_TILE):
78        wasm_v128_store(o${M}, vacc${M}x${ABC[0:4]});
79        $for C in range(4, CHANNEL_TILE, 4):
80          wasm_v128_store(o${M} + ${C}, vacc${M}x${ABC[C:C+4]});
81        o${M} += ${CHANNEL_TILE};
82    }
83    $if CHANNEL_TILE > 4:
84      for (; c >= 4 * sizeof(float); c -= 4 * sizeof(float)) {
85        const v128_t vw0123 = wasm_v128_load(w);
86        w += 4;
87
88        $for M in range(ROW_TILE):
89          v128_t vi${M}x0123 = wasm_v128_load(i${M});
90          i${M} += 4;
91
92        $for M in range(ROW_TILE):
93          v128_t vacc${M}x0123 = wasm_i32x4_max(vi${M}x0123, vzero);
94          vi${M}x0123 = wasm_i32x4_min(vi${M}x0123, vzero);
95
96        $for M in range(ROW_TILE):
97          vacc${M}x0123 = wasm_f32x4_add(vacc${M}x0123, wasm_f32x4_mul(vi${M}x0123, vw0123));
98
99        $for M in range(ROW_TILE):
100          wasm_v128_store(o${M}, vacc${M}x0123);
101          o${M} += 4;
102      }
103    if XNN_UNLIKELY(c != 0) {
104      const v128_t vw0123 = wasm_v128_load(w);
105      w = (const float*) ((uintptr_t) w + c);
106
107      $for M in range(ROW_TILE):
108        v128_t vi${M}x0123 = wasm_v128_load(i${M});
109        i${M} = (const float*) ((uintptr_t) i${M} + c);
110
111      $for M in range(ROW_TILE):
112        v128_t vacc${M}x0123 = wasm_i32x4_max(vi${M}x0123, vzero);
113        vi${M}x0123 = wasm_i32x4_min(vi${M}x0123, vzero);
114
115      $for M in range(ROW_TILE):
116        vacc${M}x0123 = wasm_f32x4_add(vacc${M}x0123, wasm_f32x4_mul(vi${M}x0123, vw0123));
117
118      if (c & (2 * sizeof(float))) {
119        $for M in range(ROW_TILE):
120          *((double*) o${M}) = wasm_f64x2_extract_lane(vacc${M}x0123, 0);
121
122        $for M in range(ROW_TILE):
123          vacc${M}x0123 = wasm_v32x4_shuffle(vacc${M}x0123, vacc${M}x0123, 2, 3, 2, 3);
124
125        $for M in range(ROW_TILE):
126          o${M} += 2;
127      }
128      if (c & (1 * sizeof(float))) {
129        $for M in range(ROW_TILE):
130          *o${M} = wasm_f32x4_extract_lane(vacc${M}x0123, 0);
131
132        $for M in range(ROW_TILE):
133          o${M} += 1;
134      }
135    }
136    $for M in range(ROW_TILE):
137      i${M} = (const float*) ((uintptr_t) i${M} + input_increment);
138      o${M} = (float*) ((uintptr_t) o${M} + output_increment);
139    rows = doz(rows, ${ROW_TILE});
140  } while (rows != 0);
141}
142