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$assert CHANNEL_TILE % 8 == 0
7$assert CHANNEL_TILE >= 8
8$assert ROW_TILE >= 1
9$ABC = "0123456789ABCDEFGHIJKLMNOPQRSTUVWXYZ"
10#include <assert.h>
11
12#include <immintrin.h>
13
14#include <xnnpack/math.h>
15#include <xnnpack/prelu.h>
16
17
18static const int32_t mask_table[14] = {-1, -1, -1, -1, -1, -1, -1, 0, 0, 0, 0, 0, 0, 0};
19
20void xnn_f32_prelu_ukernel__avx_${ROW_TILE}x${CHANNEL_TILE}(
21    size_t rows,
22    size_t channels,
23    const float*restrict input,
24    size_t input_stride,
25    const float*restrict weights,
26    float*restrict output,
27    size_t output_stride)
28{
29  assert(rows != 0);
30  assert(channels != 0);
31  assert(channels % sizeof(float) == 0);
32
33  const float* i0 = input;
34  float* o0 = output;
35  $for M in range(1, ROW_TILE):
36    const float* i${M} = (const float*) ((uintptr_t) i${M-1} + input_stride);
37    float* o${M} = (float*) ((uintptr_t) o${M-1} + output_stride);
38    $if M % 2 == 0:
39      if XNN_UNPREDICTABLE(rows <= ${M}) {
40        i${M} = i${M-1};
41        o${M} = o${M-1};
42      }
43    $else:
44      if XNN_UNPREDICTABLE(rows < ${M+1}) {
45        i${M} = i${M-1};
46        o${M} = o${M-1};
47      }
48
49  const size_t input_increment = input_stride * ${ROW_TILE} - channels;
50  const size_t output_increment = output_stride * ${ROW_TILE} - channels;
51
52  do {
53    const float* w = weights;
54    size_t c = channels;
55    for (; c >= ${CHANNEL_TILE} * sizeof(float); c -= ${CHANNEL_TILE} * sizeof(float)) {
56      const __m256 vw${ABC[0:8]} = _mm256_load_ps(w);
57      $for C in range(8, CHANNEL_TILE, 8):
58        const __m256 vw${ABC[C:C+8]} = _mm256_load_ps(w + ${C});
59      w += ${CHANNEL_TILE};
60
61      $for M in range(ROW_TILE):
62        const __m256 vi${M}x${ABC[0:8]} = _mm256_loadu_ps(i${M});
63        $for C in range(8, CHANNEL_TILE, 8):
64          const __m256 vi${M}x${ABC[C:C+8]} = _mm256_loadu_ps(i${M} + ${C});
65        i${M} += ${CHANNEL_TILE};
66
67      $for M in range(ROW_TILE):
68        $for C in range(0, CHANNEL_TILE, 8):
69          const __m256 vprod${M}x${ABC[C:C+8]} = _mm256_mul_ps(vi${M}x${ABC[C:C+8]}, vw${ABC[C:C+8]});
70
71      $for M in range(ROW_TILE):
72        $for C in range(0, CHANNEL_TILE, 8):
73          const __m256 vacc${M}x${ABC[C:C+8]} = _mm256_blendv_ps(vi${M}x${ABC[C:C+8]}, vprod${M}x${ABC[C:C+8]}, vi${M}x${ABC[C:C+8]});
74
75      $for M in range(ROW_TILE):
76        _mm256_storeu_ps(o${M}, vacc${M}x${ABC[0:8]});
77        $for C in range(8, CHANNEL_TILE, 8):
78          _mm256_storeu_ps(o${M} + ${C}, vacc${M}x${ABC[C:C+8]});
79        o${M} += ${CHANNEL_TILE};
80    }
81    $if CHANNEL_TILE > 8:
82      for (; c >= 8 * sizeof(float); c -= 8 * sizeof(float)) {
83        const __m256 vw = _mm256_load_ps(w);
84        w += 8;
85
86        $for M in range(ROW_TILE):
87          const __m256 vi${M} = _mm256_loadu_ps(i${M});
88          i${M} += 8;
89
90        $for M in range(ROW_TILE):
91          const __m256 vprod${M} = _mm256_mul_ps(vi${M}, vw);
92
93        $for M in range(ROW_TILE):
94          const __m256 vacc${M} = _mm256_blendv_ps(vi${M}, vprod${M}, vi${M});
95
96        $for M in range(ROW_TILE):
97          _mm256_storeu_ps(o${M}, vacc${M});
98          o${M} += 8;
99      }
100    if XNN_UNLIKELY(c != 0) {
101      assert(c >= 1 * sizeof(float));
102      assert(c <= 7 * sizeof(float));
103      __m256i vmask = _mm256_loadu_si256((const __m256i*) ((uintptr_t) &mask_table[7] - c));
104
105      const __m256 vw = _mm256_maskload_ps(w, vmask);
106
107      $for M in range(ROW_TILE):
108        const __m256 vi${M} = _mm256_maskload_ps(i${M}, vmask);
109        i${M} = (const float*) ((uintptr_t) i${M} + c);
110
111      $for M in range(ROW_TILE):
112        const __m256 vprod${M} = _mm256_mul_ps(vi${M}, vw);
113
114      $for M in range(ROW_TILE):
115        __m256 vacc${M} = _mm256_blendv_ps(vi${M}, vprod${M}, vi${M});
116
117      // _mm256_maskstore_ps(o${M}, vmask, vacc${M}) could be used here, but triggers msan failures (probably an msan bug).
118      $for M in range(ROW_TILE):
119        __m128 vacc${M}_lo = _mm256_castps256_ps128(vacc${M});
120      if (c & (4 * sizeof(float))) {
121        $for M in range(ROW_TILE):
122          _mm_storeu_ps(o${M}, vacc${M}_lo);
123
124        $for M in range(ROW_TILE):
125          vacc${M}_lo = _mm256_extractf128_ps(vacc${M}, 1);
126
127        $for M in range(ROW_TILE):
128          o${M} += 4;
129      }
130      if (c & (2 * sizeof(float))) {
131        $for M in range(ROW_TILE):
132          _mm_storel_pi((__m64*) o${M}, vacc${M}_lo);
133
134        $for M in range(ROW_TILE):
135          vacc${M}_lo = _mm_movehl_ps(vacc${M}_lo, vacc${M}_lo);
136
137        $for M in range(ROW_TILE):
138          o${M} += 2;
139      }
140      if (c & (1 * sizeof(float))) {
141        $for M in range(ROW_TILE):
142          _mm_store_ss(o${M}, vacc${M}_lo);
143
144        $for M in range(ROW_TILE):
145          o${M} += 1;
146      }
147    }
148    $for M in range(ROW_TILE):
149      i${M} = (const float*) ((uintptr_t) i${M} + input_increment);
150      o${M} = (float*) ((uintptr_t) o${M} + output_increment);
151      $if M % 2 == 1:
152        if XNN_UNPREDICTABLE(rows < ${ROW_TILE + M+1}) {
153          i${M} = i${M-1};
154          o${M} = o${M-1};
155        }
156      $elif M != 0:
157        if XNN_UNPREDICTABLE(rows <= ${ROW_TILE + M}) {
158          i${M} = i${M-1};
159          o${M} = o${M-1};
160        }
161    rows = doz(rows, ${ROW_TILE});
162  } while (rows != 0);
163}
164