xref: /aosp_15_r20/external/XNNPACK/src/f32-prelu/avx512f.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$assert CHANNEL_TILE % 16 == 0
7$assert CHANNEL_TILE >= 16
8$assert ROW_TILE >= 1
9$ABC = "0123456789ABCDEFGHIJKLMNOPQRSTUVWXYZ"
10#include <assert.h>
11
12#include <immintrin.h>
13
14#include <xnnpack/intrinsics-polyfill.h>
15#include <xnnpack/math.h>
16#include <xnnpack/prelu.h>
17
18
19void xnn_f32_prelu_ukernel__avx512f_${ROW_TILE}x${CHANNEL_TILE}(
20    size_t rows,
21    size_t channels,
22    const float*restrict input,
23    size_t input_stride,
24    const float*restrict weights,
25    float*restrict output,
26    size_t output_stride)
27{
28  assert(rows != 0);
29  assert(channels != 0);
30  assert(channels % sizeof(float) == 0);
31
32  const float* i0 = input;
33  float* o0 = output;
34  $for M in range(1, ROW_TILE):
35    const float* i${M} = (const float*) ((uintptr_t) i${M-1} + input_stride);
36    float* o${M} = (float*) ((uintptr_t) o${M-1} + output_stride);
37
38  const size_t input_increment = input_stride * ${ROW_TILE} - channels;
39  const size_t output_increment = output_stride * ${ROW_TILE} - channels;
40
41  const __m512 vzero = _mm512_setzero_ps();
42  do {
43    $for M in range(1, ROW_TILE):
44      $if M % 2 == 0:
45        if XNN_UNPREDICTABLE(rows <= ${M}) {
46          i${M} = i${M-1};
47          o${M} = o${M-1};
48        }
49      $else:
50        if XNN_UNPREDICTABLE(rows < ${M+1}) {
51          i${M} = i${M-1};
52          o${M} = o${M-1};
53        }
54
55    const float* w = weights;
56    size_t c = channels;
57    for (; c >= ${CHANNEL_TILE} * sizeof(float); c -= ${CHANNEL_TILE} * sizeof(float)) {
58      const __m512 vw${ABC[0:16]} = _mm512_load_ps(w);
59      $for C in range(16, CHANNEL_TILE, 16):
60        const __m512 vw${ABC[C:C+16]} = _mm512_load_ps(w + ${C});
61      w += ${CHANNEL_TILE};
62
63      $for M in range(ROW_TILE):
64        const __m512 vi${M}x${ABC[0:16]} = _mm512_loadu_ps(i${M});
65        $for C in range(16, CHANNEL_TILE, 16):
66          const __m512 vi${M}x${ABC[C:C+16]} = _mm512_loadu_ps(i${M} + ${C});
67        i${M} += ${CHANNEL_TILE};
68
69      $for M in range(ROW_TILE):
70        $for C in range(0, CHANNEL_TILE, 16):
71          const __mmask16 vsign${M}x${ABC[C:C+16]} = _mm512_cmp_ps_mask(vi${M}x${ABC[C:C+16]}, vzero, _CMP_LT_OQ);
72          const __m512 vacc${M}x${ABC[C:C+16]} = _mm512_mask_mul_ps(vi${M}x${ABC[C:C+16]}, vsign${M}x${ABC[C:C+16]}, vi${M}x${ABC[C:C+16]}, vw${ABC[C:C+16]});
73
74      $for M in range(ROW_TILE):
75        _mm512_storeu_ps(o${M}, vacc${M}x${ABC[0:16]});
76        $for C in range(16, CHANNEL_TILE, 16):
77          _mm512_storeu_ps(o${M} + ${C}, vacc${M}x${ABC[C:C+16]});
78        o${M} += ${CHANNEL_TILE};
79    }
80    $if CHANNEL_TILE > 16:
81      for (; c >= 16 * sizeof(float); c -= 16 * sizeof(float)) {
82        const __m512 vw = _mm512_load_ps(w);
83        w += 16;
84
85        $for M in range(ROW_TILE):
86          const __m512 vi${M} = _mm512_loadu_ps(i${M});
87          i${M} += 16;
88
89        $for M in range(ROW_TILE):
90          const __mmask16 vsign${M} = _mm512_cmp_ps_mask(vi${M}, vzero, _CMP_LT_OQ);
91          const __m512 vacc${M} = _mm512_mask_mul_ps(vi${M}, vsign${M}, vi${M}, vw);
92
93        $for M in range(ROW_TILE):
94          _mm512_storeu_ps(o${M}, vacc${M});
95          o${M} += 16;
96      }
97    if XNN_UNLIKELY(c != 0) {
98      assert(c >= 1 * sizeof(float));
99      assert(c <= 15 * sizeof(float));
100      // Prepare mask for valid 32-bit elements (depends on c).
101      const __mmask16 vmask = _cvtu32_mask16((uint16_t) ((uint32_t) (UINT32_C(1) << (c >> 2 /* log2(sizeof(float))*/)) - UINT32_C(1)));
102
103      const __m512 vw = _mm512_maskz_loadu_ps(vmask, w);
104
105      $for M in range(ROW_TILE):
106        const __m512 vi${M} = _mm512_maskz_loadu_ps(vmask, i${M});
107        i${M} = (const float*) ((uintptr_t) i${M} + c);
108
109      $for M in range(ROW_TILE):
110        const __mmask16 vsign${M} = _mm512_cmp_ps_mask(vi${M}, vzero, _CMP_LT_OQ);
111        const __m512 vacc${M} = _mm512_mask_mul_ps(vi${M}, vsign${M}, vi${M}, vw);
112
113      $for M in range(ROW_TILE):
114        _mm512_mask_storeu_ps(o${M}, vmask, vacc${M});
115        o${M} = (float*) ((uintptr_t) o${M} + c);
116    }
117    $for M in range(ROW_TILE):
118      i${M} = (const float*) ((uintptr_t) i${M} + input_increment);
119      o${M} = (float*) ((uintptr_t) o${M} + output_increment);
120    rows = doz(rows, ${ROW_TILE});
121  } while (rows != 0);
122}
123