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$assert BATCH_TILE % 4 == 0
7$assert BATCH_TILE >= 4
8$ABC = "0123456789ABCDEFGHIJKLMNOPQRSTUVWXYZ"
9$assert OP in ["ADD", "DIV", "MAX", "MIN", "MUL", "SUB", "SQRDIFF"]
10$assert ACTIVATION in ["LINEAR", "MINMAX"]
11#include <assert.h>
12
13#include <xmmintrin.h>
14
15#include <xnnpack/common.h>
16#include <xnnpack/intrinsics-polyfill.h>
17#include <xnnpack/vbinary.h>
18
19
20$_MM_OP_PS = {
21$  "ADD": lambda x, y: "_mm_add_ps(%s, %s)" % (x, y),
22$  "DIV": lambda x, y: "_mm_div_ps(%s, %s)" % (x, y),
23$  "MAX": lambda x, y: "_mm_max_ps(%s, %s)" % (x, y),
24$  "MIN": lambda x, y: "_mm_min_ps(%s, %s)" % (x, y),
25$  "MUL": lambda x, y: "_mm_mul_ps(%s, %s)" % (x, y),
26$  "SUB": lambda x, y: "_mm_sub_ps(%s, %s)" % (x, y),
27$  "SQRDIFF": lambda x, y: "_mm_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()}${SUFFIX}_ukernel__sse_x${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)]) XNN_DISABLE_TSAN
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 __m128 vy_min = _mm_load_ps(params->sse.min);
46    const __m128 vy_max = _mm_load_ps(params->sse.max);
47
48  for (; n >= ${BATCH_TILE} * sizeof(float); n -= ${BATCH_TILE} * sizeof(float)) {
49    const __m128 va${ABC[0:4]} = _mm_loadu_ps(a);
50    $for N in range(4, BATCH_TILE, 4):
51      const __m128 va${ABC[N:N+4]} = _mm_loadu_ps(a + ${N});
52    a += ${BATCH_TILE};
53
54    const __m128 vb${ABC[0:4]} = _mm_loadu_ps(b);
55    $for N in range(4, BATCH_TILE, 4):
56      const __m128 vb${ABC[N:N+4]} = _mm_loadu_ps(b + ${N});
57    b += ${BATCH_TILE};
58
59    $for N in range(0, BATCH_TILE, 4):
60      __m128 vy${ABC[N:N+4]} = ${_MM_OP_PS("va" + ABC[N:N+4], "vb" + ABC[N:N+4])};
61
62    $if OP == "SQRDIFF":
63      $for N in range(0, BATCH_TILE, 4):
64        vy${ABC[N:N+4]} = _mm_mul_ps(vy${ABC[N:N+4]}, vy${ABC[N:N+4]});
65
66    $if ACTIVATION == "MINMAX":
67      $for N in range(0, BATCH_TILE, 4):
68        vy${ABC[N:N+4]} = _mm_max_ps(vy${ABC[N:N+4]}, vy_min);
69
70      $for N in range(0, BATCH_TILE, 4):
71        vy${ABC[N:N+4]} = _mm_min_ps(vy${ABC[N:N+4]}, vy_max);
72
73    _mm_storeu_ps(y, vy${ABC[0:4]});
74    $for N in range(4, BATCH_TILE, 4):
75      _mm_storeu_ps(y + ${N}, vy${ABC[N:N+4]});
76    y += ${BATCH_TILE};
77  }
78  $if BATCH_TILE > 4:
79    for (; n >= 4 * sizeof(float); n -= 4 * sizeof(float)) {
80      const __m128 va0123 = _mm_loadu_ps(a);
81      a += 4;
82
83      const __m128 vb0123 = _mm_loadu_ps(b);
84      b += 4;
85
86      __m128 vy0123 = ${_MM_OP_PS("va0123", "vb0123")};
87      $if OP == "SQRDIFF":
88        vy0123 = _mm_mul_ps(vy0123, vy0123);
89      $if ACTIVATION == "MINMAX":
90        vy0123 = _mm_max_ps(vy0123, vy_min);
91        vy0123 = _mm_min_ps(vy0123, vy_max);
92      _mm_storeu_ps(y, vy0123);
93      y += 4;
94    }
95  if XNN_UNLIKELY(n != 0) {
96    const __m128 va0123 = _mm_loadu_ps(a);
97    const __m128 vb0123 = _mm_loadu_ps(b);
98
99    __m128 vy0123 = ${_MM_OP_PS("va0123", "vb0123")};
100    $if OP == "SQRDIFF":
101      vy0123 = _mm_mul_ps(vy0123, vy0123);
102    $if ACTIVATION == "MINMAX":
103      vy0123 = _mm_max_ps(vy0123, vy_min);
104      vy0123 = _mm_min_ps(vy0123, vy_max);
105    if (n & (2 * sizeof(float))) {
106      _mm_storel_pi((__m64*) y, vy0123);
107      vy0123 = _mm_movehl_ps(vy0123, vy0123);
108      y += 2;
109    }
110    if (n & (1 * sizeof(float))) {
111      _mm_store_ss(y, vy0123);
112    }
113  }
114}
115