• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
1// Copyright 2019 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$assert BATCH_TILE % 16 == 0
7$assert BATCH_TILE >= 16
8$ABC = "0123456789ABCDEFGHIJKLMNOPQRSTUVWXYZ"
9$assert OP in ["ADD", "DIV", "MAX", "MIN", "MUL", "SUB", "SQRDIFF"]
10$assert ACTIVATION in ["LINEAR", "MINMAX"]
11#include <assert.h>
12
13#include <immintrin.h>
14
15#include <xnnpack/common.h>
16#include <xnnpack/intrinsics-polyfill.h>
17#include <xnnpack/vbinary.h>
18
19
20$_MM512_OP_PS = {
21$  "ADD": lambda x, y: "_mm512_add_ps(%s, %s)" % (x, y),
22$  "DIV": lambda x, y: "_mm512_div_ps(%s, %s)" % (x, y),
23$  "MAX": lambda x, y: "_mm512_max_ps(%s, %s)" % (x, y),
24$  "MIN": lambda x, y: "_mm512_min_ps(%s, %s)" % (x, y),
25$  "MUL": lambda x, y: "_mm512_mul_ps(%s, %s)" % (x, y),
26$  "SUB": lambda x, y: "_mm512_sub_ps(%s, %s)" % (x, y),
27$  "SQRDIFF": lambda x, y: "_mm512_sub_ps(%s, %s)" % (x, y),
28$}[OP]
29$SUFFIX = {"LINEAR": "", "MINMAX": "_minmax"}[ACTIVATION]
30$PARAMS = {"LINEAR": "xnn_f32_default_params", "MINMAX": "xnn_f32_minmax_params"}[ACTIVATION]
31void xnn_f32_v${OP.lower()}${SUFFIX}_ukernel__avx512f_x${BATCH_TILE}(
32    size_t n,
33    const float* a,
34    const float* b,
35    float* y,
36    const union ${PARAMS} params[restrict XNN_MIN_ELEMENTS(1)])
37{
38  assert(n != 0);
39  assert(n % sizeof(float) == 0);
40  assert(a != NULL);
41  assert(b != NULL);
42  assert(y != NULL);
43
44  $if ACTIVATION == "MINMAX":
45    const __m512 vy_min = _mm512_broadcast_f32x4(_mm_load_ps(params->sse.min));
46    const __m512 vy_max = _mm512_broadcast_f32x4(_mm_load_ps(params->sse.max));
47
48  for (; n >= ${BATCH_TILE} * sizeof(float); n -= ${BATCH_TILE} * sizeof(float)) {
49    const __m512 va${ABC[0:16]} = _mm512_loadu_ps(a);
50    $for N in range(16, BATCH_TILE, 16):
51      const __m512 va${ABC[N:N+16]} = _mm512_loadu_ps(a + ${N});
52    a += ${BATCH_TILE};
53
54    const __m512 vb${ABC[0:16]} = _mm512_loadu_ps(b);
55    $for N in range(16, BATCH_TILE, 16):
56      const __m512 vb${ABC[N:N+16]} = _mm512_loadu_ps(b + ${N});
57    b += ${BATCH_TILE};
58
59    $for N in range(0, BATCH_TILE, 16):
60      __m512 vy${ABC[N:N+16]} = ${_MM512_OP_PS("va" + ABC[N:N+16], "vb" + ABC[N:N+16])};
61
62    $if OP == "SQRDIFF":
63      $for N in range(0, BATCH_TILE, 16):
64        vy${ABC[N:N+16]} = _mm512_mul_ps(vy${ABC[N:N+16]}, vy${ABC[N:N+16]});
65
66    $if ACTIVATION == "MINMAX":
67      $for N in range(0, BATCH_TILE, 16):
68        vy${ABC[N:N+16]} = _mm512_max_ps(vy${ABC[N:N+16]}, vy_min);
69
70      $for N in range(0, BATCH_TILE, 16):
71        vy${ABC[N:N+16]} = _mm512_min_ps(vy${ABC[N:N+16]}, vy_max);
72
73    _mm512_storeu_ps(y, vy${ABC[0:16]});
74    $for N in range(16, BATCH_TILE, 16):
75      _mm512_storeu_ps(y + ${N}, vy${ABC[N:N+16]});
76    y += ${BATCH_TILE};
77  }
78  $if BATCH_TILE > 16:
79    for (; n >= 16 * sizeof(float); n -= 16 * sizeof(float)) {
80      const __m512 va = _mm512_loadu_ps(a);
81      a += 16;
82
83      const __m512 vb = _mm512_loadu_ps(b);
84      b += 16;
85
86      __m512 vy = ${_MM512_OP_PS("va", "vb")};
87      $if OP == "SQRDIFF":
88        vy = _mm512_mul_ps(vy, vy);
89      $if ACTIVATION == "MINMAX":
90        vy = _mm512_max_ps(vy, vy_min);
91        vy = _mm512_min_ps(vy, vy_max);
92      _mm512_storeu_ps(y, vy);
93      y += 16;
94    }
95  if XNN_UNLIKELY(n != 0) {
96    assert(n >= 1 * sizeof(float));
97    assert(n <= 15 * sizeof(float));
98    // Prepare mask for valid 32-bit elements (depends on n).
99    n >>= 2 /* log2(sizeof(float)) */;
100    const __mmask16 vmask = _cvtu32_mask16((uint16_t) ((uint32_t) (UINT32_C(1) << n) - UINT32_C(1)));
101
102    const __m512 va = _mm512_maskz_loadu_ps(vmask, a);
103    const __m512 vb = _mm512_maskz_loadu_ps(vmask, b);
104
105    __m512 vy = ${_MM512_OP_PS("va", "vb")};
106    $if OP == "SQRDIFF":
107      vy = _mm512_mul_ps(vy, vy);
108    $if ACTIVATION == "MINMAX":
109      vy = _mm512_max_ps(vy, vy_min);
110      vy = _mm512_min_ps(vy, vy_max);
111    _mm512_mask_storeu_ps(y, vmask, vy);
112  }
113}
114