• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
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$assert BATCH_TILE % 8 == 0
7$assert BATCH_TILE >= 8
8$SIMD_TILE = BATCH_TILE // 8
9#include <assert.h>
10
11#include <immintrin.h>
12
13#include <xnnpack/common.h>
14#include <xnnpack/vunary.h>
15
16
17extern XNN_INTERNAL const int xnn_table_exp2minus_k_over_16[16];
18
19static const int32_t mask_table[14] = {-1, -1, -1, -1, -1, -1, -1, 0, 0, 0, 0, 0, 0, 0};
20
21void xnn_f32_velu_ukernel__avx2_rr1_lut16_p3_gather_x${BATCH_TILE}(
22    size_t n,
23    const float* x,
24    float* y,
25    const union xnn_f32_elu_params params[restrict XNN_MIN_ELEMENTS(1)])
26{
27  assert(n % sizeof(float) == 0);
28
29  const __m256 vprescale = _mm256_broadcast_ps((const __m128*) params->sse.prescale);
30  const __m256 valpha = _mm256_broadcast_ps((const __m128*) params->sse.alpha);
31  const __m256 vbeta = _mm256_broadcast_ps((const __m128*) params->sse.beta);
32
33  const __m256 vsat_cutoff = _mm256_set1_ps(-0x1.154246p+4f);
34  const __m256 vmagic_bias = _mm256_set1_ps(0x1.800000p19f);
35  const __m256 vlog2e = _mm256_set1_ps(0x1.715476p+0f);
36  const __m256i vindex_mask = _mm256_set1_epi32(0xF);
37  const __m256 vminus_ln2 = _mm256_set1_ps(-0x1.62E43p-1f);
38  const __m256 vc3 = _mm256_set1_ps(0x1.55561Cp-3f);
39  const __m256 vc2 = _mm256_set1_ps(0x1.0001ECp-1f);
40
41  $if BATCH_TILE > 8:
42    for (; n >= ${BATCH_TILE} * sizeof(float); n -= ${BATCH_TILE} * sizeof(float)) {
43      __m256 vx0 = _mm256_loadu_ps(x);
44      $for N in range(1, SIMD_TILE):
45        __m256 vx${N} = _mm256_loadu_ps(x + ${N * 8});
46      x += ${BATCH_TILE};
47
48      $for N in range(SIMD_TILE):
49        const __m256 vz${N} = _mm256_max_ps(vsat_cutoff, _mm256_mul_ps(vx${N}, vprescale));
50
51      $for N in range(SIMD_TILE):
52        __m256 vn${N} = _mm256_fmadd_ps(vz${N}, vlog2e, vmagic_bias);
53
54      $for N in range(SIMD_TILE):
55        const __m256i vidx${N} = _mm256_and_si256(_mm256_castps_si256(vn${N}), vindex_mask);
56        const __m256i vl${N} = _mm256_i32gather_epi32(xnn_table_exp2minus_k_over_16, vidx${N}, sizeof(float));
57
58      $for N in range(SIMD_TILE):
59        const __m256i ven${N} = _mm256_slli_epi32(_mm256_castps_si256(vn${N}), 19);
60        vn${N} = _mm256_sub_ps(vn${N}, vmagic_bias);
61
62      $for N in range(SIMD_TILE):
63        __m256 vs${N} = _mm256_castsi256_ps(_mm256_add_epi32(vl${N}, ven${N}));
64        __m256 vt${N} = _mm256_fmadd_ps(vn${N}, vminus_ln2, vz${N});
65
66      $for N in range(SIMD_TILE):
67        __m256 vp${N} = _mm256_fmadd_ps(vc3, vt${N}, vc2);
68
69      $for N in range(SIMD_TILE):
70        vp${N} = _mm256_mul_ps(vp${N}, vt${N});
71        vt${N} = _mm256_mul_ps(vt${N}, vs${N});
72
73      $for N in range(SIMD_TILE):
74        vs${N} = _mm256_fmsub_ps(vs${N}, valpha, valpha);
75        vp${N} = _mm256_fmadd_ps(vp${N}, vt${N}, vt${N});
76
77      $for N in range(SIMD_TILE):
78        const __m256 ve${N} = _mm256_fmadd_ps(vp${N}, valpha, vs${N});
79        vx${N} = _mm256_mul_ps(vx${N}, vbeta);
80
81      $for N in range(SIMD_TILE):
82        const __m256 vy${N} = _mm256_blendv_ps(vx${N}, ve${N}, vx${N});
83
84      _mm256_storeu_ps(y, vy0);
85      $for N in range(1, SIMD_TILE):
86        _mm256_storeu_ps(y + ${N * 8}, vy${N});
87      y += ${BATCH_TILE};
88    }
89  for (; n >= 8 * sizeof(float); n -= 8 * sizeof(float)) {
90    __m256 vx = _mm256_loadu_ps(x);
91    x += 8;
92
93    const __m256 vz = _mm256_max_ps(vsat_cutoff, _mm256_mul_ps(vx, vprescale));
94
95    __m256 vn = _mm256_fmadd_ps(vz, vlog2e, vmagic_bias);
96    const __m256i vidx = _mm256_and_si256(_mm256_castps_si256(vn), vindex_mask);
97    const __m256i vl = _mm256_i32gather_epi32(xnn_table_exp2minus_k_over_16, vidx, sizeof(float));
98
99    const __m256i ven = _mm256_slli_epi32(_mm256_castps_si256(vn), 19);
100    vn = _mm256_sub_ps(vn, vmagic_bias);
101
102    __m256 vs = _mm256_castsi256_ps(_mm256_add_epi32(vl, ven));
103    __m256 vt = _mm256_fmadd_ps(vn, vminus_ln2, vz);
104
105    __m256 vp = _mm256_fmadd_ps(vc3, vt, vc2);
106    vp = _mm256_mul_ps(vp, vt);
107
108    vt = _mm256_mul_ps(vt, vs);
109    vs = _mm256_fmsub_ps(vs, valpha, valpha);
110    vp = _mm256_fmadd_ps(vp, vt, vt);
111    const __m256 ve = _mm256_fmadd_ps(vp, valpha, vs);
112
113    vx = _mm256_mul_ps(vx, vbeta);
114    const __m256 vy = _mm256_blendv_ps(vx, ve, vx);
115
116    _mm256_storeu_ps(y, vy);
117    y += 8;
118  }
119  if XNN_UNLIKELY(n != 0) {
120    assert(n >= 1 * sizeof(float));
121    assert(n <= 7 * sizeof(float));
122    __m256i vmask = _mm256_loadu_si256((const __m256i*) ((uintptr_t) &mask_table[7] - n));
123
124    __m256 vx = _mm256_maskload_ps(x, vmask);
125
126    const __m256 vz = _mm256_max_ps(vsat_cutoff, _mm256_mul_ps(vx, vprescale));
127
128    __m256 vn = _mm256_fmadd_ps(vz, vlog2e, vmagic_bias);
129    const __m256i vidx = _mm256_and_si256(_mm256_castps_si256(vn), vindex_mask);
130    const __m256i vl = _mm256_i32gather_epi32(xnn_table_exp2minus_k_over_16, vidx, sizeof(float));
131
132    const __m256i ven = _mm256_slli_epi32(_mm256_castps_si256(vn), 19);
133    vn = _mm256_sub_ps(vn, vmagic_bias);
134
135    __m256 vs = _mm256_castsi256_ps(_mm256_add_epi32(vl, ven));
136    __m256 vt = _mm256_fmadd_ps(vn, vminus_ln2, vz);
137
138    __m256 vp = _mm256_fmadd_ps(vc3, vt, vc2);
139    vp = _mm256_mul_ps(vp, vt);
140
141    vt = _mm256_mul_ps(vt, vs);
142    vs = _mm256_fmsub_ps(vs, valpha, valpha);
143    vp = _mm256_fmadd_ps(vp, vt, vt);
144    const __m256 ve = _mm256_fmadd_ps(vp, valpha, vs);
145
146    vx = _mm256_mul_ps(vx, vbeta);
147    const __m256 vy = _mm256_blendv_ps(vx, ve, vx);
148
149    // _mm256_maskstore_ps(y, vmask, vf) could be used here, but triggers msan failures (probably an msan bug).
150    __m128 vy_lo = _mm256_castps256_ps128(vy);
151    if (n & (4 * sizeof(float))) {
152      _mm_storeu_ps(y, vy_lo);
153      vy_lo = _mm256_extractf128_ps(vy, 1);
154      y += 4;
155    }
156    if (n & (2 * sizeof(float))) {
157      _mm_storel_pi((__m64*) y, vy_lo);
158      vy_lo = _mm_movehl_ps(vy_lo, vy_lo);
159      y += 2;
160    }
161    if (n & (1 * sizeof(float))) {
162      _mm_store_ss(y, vy_lo);
163    }
164  }
165}
166