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$assert BATCH_TILE % 16 == 0
7$assert BATCH_TILE >= 16
8$assert RR_STEPS in [1, 2]
9$assert DIV_ALGO in ["div", "nr1fma", "nr1fma1adj"]
10$ABC = "0123456789ABCDEFGHIJKLMNOPQRSTUVWXYZ"
11$SIMD_TILE = BATCH_TILE // 16
12#include <assert.h>
13
14#include <immintrin.h>
15
16#include <xnnpack/common.h>
17#include <xnnpack/intrinsics-polyfill.h>
18#include <xnnpack/vunary.h>
19
20
21void xnn_f32_sigmoid_ukernel__avx512f_rr${RR_STEPS}_lut32_p2_perm2_scalef_${DIV_ALGO}_x${BATCH_TILE}(
22    size_t n,
23    const float* x,
24    float* y,
25    const void* params)
26{
27  assert(n % sizeof(float) == 0);
28
29  const __m512i vsign_mask = _mm512_set1_epi32(0x80000000);
30  const __m512 vmagic_bias = _mm512_set1_ps(0x1.800000p18f);
31  const __m512 vlog2e = _mm512_set1_ps(0x1.715476p0f);
32  const __m512 vtable_hi = _mm512_set_ps(
33    0x1.F50766p+0f, 0x1.EA4AFAp+0f, 0x1.DFC974p+0f, 0x1.D5818Ep+0f,
34    0x1.CB720Ep+0f, 0x1.C199BEp+0f, 0x1.B7F770p+0f, 0x1.AE89FAp+0f,
35    0x1.A5503Cp+0f, 0x1.9C4918p+0f, 0x1.93737Cp+0f, 0x1.8ACE54p+0f,
36    0x1.82589Ap+0f, 0x1.7A1148p+0f, 0x1.71F75Ep+0f, 0x1.6A09E6p+0f);
37  const __m512 vtable_lo = _mm512_set_ps(
38    0x1.6247ECp+0f, 0x1.5AB07Ep+0f, 0x1.5342B6p+0f, 0x1.4BFDAEp+0f,
39    0x1.44E086p+0f, 0x1.3DEA64p+0f, 0x1.371A74p+0f, 0x1.306FE0p+0f,
40    0x1.29E9E0p+0f, 0x1.2387A6p+0f, 0x1.1D4874p+0f, 0x1.172B84p+0f,
41    0x1.11301Ep+0f, 0x1.0B5586p+0f, 0x1.059B0Ep+0f, 0x1.000000p+0f);
42  $if RR_STEPS == 1:
43    const __m512 vminus_ln2 = _mm512_set1_ps(-0x1.62E43p-1f);
44  $else:
45    const __m512 vminus_ln2_hi = _mm512_set1_ps(-0x1.62E43p-1f);
46    const __m512 vminus_ln2_lo = _mm512_set1_ps(0x1.05C61p-29f);
47  const __m512 vc2 = _mm512_set1_ps(0x1.000000p-1f);
48  const __m512 vc1 = _mm512_set1_ps(0x1.0000F6p-0f);
49  const __m512 vone = _mm512_set1_ps(1.0f);
50
51  $if BATCH_TILE > 16:
52    for (; n >= ${BATCH_TILE} * sizeof(float); n -= ${BATCH_TILE} * sizeof(float)) {
53      const __m512 vx${ABC[0]} = _mm512_loadu_ps(x);
54      $for N in range(1, SIMD_TILE):
55        const __m512 vx${ABC[N]} = _mm512_loadu_ps(x + ${N * 16});
56      x += ${BATCH_TILE};
57
58      $for N in range(SIMD_TILE):
59        const __m512 vz${ABC[N]} = _mm512_castsi512_ps(_mm512_or_epi32(_mm512_castps_si512(vx${ABC[N]}), vsign_mask));
60
61      $for N in range(SIMD_TILE):
62        __m512 vn${ABC[N]} = _mm512_fmadd_ps(vz${ABC[N]}, vlog2e, vmagic_bias);
63
64      $for N in range(SIMD_TILE):
65        const __m512 vl${ABC[N]} = _mm512_permutex2var_ps(vtable_lo, _mm512_castps_si512(vn${ABC[N]}), vtable_hi);
66
67      $for N in range(SIMD_TILE):
68        vn${ABC[N]} = _mm512_sub_ps(vn${ABC[N]}, vmagic_bias);
69
70      $if RR_STEPS == 1:
71        $for N in range(SIMD_TILE):
72          __m512 vt${ABC[N]} = _mm512_fmadd_ps(vn${ABC[N]}, vminus_ln2, vz${ABC[N]});
73      $else:
74        $for N in range(SIMD_TILE):
75          __m512 vt${ABC[N]} = _mm512_fmadd_ps(vn${ABC[N]}, vminus_ln2_hi, vz${ABC[N]});
76
77        $for N in range(SIMD_TILE):
78          vt${ABC[N]} = _mm512_fmadd_ps(vn${ABC[N]}, vminus_ln2_lo, vt${ABC[N]});
79
80      $for N in range(SIMD_TILE):
81        __m512 vp${ABC[N]} = _mm512_fmadd_ps(vt${ABC[N]}, vc2, vc1);
82
83      $for N in range(SIMD_TILE):
84        vt${ABC[N]} = _mm512_mul_ps(vt${ABC[N]}, vl${ABC[N]});
85
86      $for N in range(SIMD_TILE):
87        vp${ABC[N]} = _mm512_fmadd_ps(vt${ABC[N]}, vp${ABC[N]}, vl${ABC[N]});
88
89      $for N in range(SIMD_TILE):
90        const __m512 ve${ABC[N]} = _mm512_scalef_ps(vp${ABC[N]}, vn${ABC[N]});
91
92      $for N in range(SIMD_TILE):
93        const __m512 vd${ABC[N]} = _mm512_add_ps(ve${ABC[N]}, vone);
94
95      $if DIV_ALGO == "div":
96        $for N in range(SIMD_TILE):
97          __m512 vf${ABC[N]} = _mm512_div_ps(ve${ABC[N]}, vd${ABC[N]});
98      $else:
99        $for N in range(SIMD_TILE):
100          __m512 vr${ABC[N]} = _mm512_rcp14_ps(vd${ABC[N]});
101
102        $for N in range(SIMD_TILE):
103          vr${ABC[N]} = _mm512_fmadd_ps(_mm512_fnmadd_ps(vr${ABC[N]}, vd${ABC[N]}, vone), vr${ABC[N]}, vr${ABC[N]});
104
105        $for N in range(SIMD_TILE):
106          __m512 vf${ABC[N]} = _mm512_mul_ps(ve${ABC[N]}, vr${ABC[N]});
107
108        $if DIV_ALGO == "nr1fma1adj":
109          $for N in range(SIMD_TILE):
110            vf${ABC[N]} = _mm512_fmadd_ps(_mm512_fnmadd_ps(vf${ABC[N]}, vd${ABC[N]}, ve${ABC[N]}), vr${ABC[N]}, vf${ABC[N]});
111
112      $for N in range(SIMD_TILE):
113        vf${ABC[N]} = _mm512_mask_sub_ps(vf${ABC[N]}, _mm512_testn_epi32_mask(_mm512_castps_si512(vx${ABC[N]}), vsign_mask), vone, vf${ABC[N]});
114
115      _mm512_storeu_ps(y, vf${ABC[0]});
116      $for N in range(1, SIMD_TILE):
117        _mm512_storeu_ps(y + ${N * 16}, vf${ABC[N]});
118      y += ${BATCH_TILE};
119    }
120  for (; n >= 16 * sizeof(float); n -= 16 * sizeof(float)) {
121    const __m512 vx = _mm512_loadu_ps(x);
122    x += 16;
123
124    const __m512 vz = _mm512_castsi512_ps(_mm512_or_epi32(_mm512_castps_si512(vx), vsign_mask));
125
126    __m512 vn = _mm512_fmadd_ps(vz, vlog2e, vmagic_bias);
127    const __m512 vl = _mm512_permutex2var_ps(vtable_lo, _mm512_castps_si512(vn), vtable_hi);
128    vn = _mm512_sub_ps(vn, vmagic_bias);
129
130    $if RR_STEPS == 1:
131      __m512 vt = _mm512_fmadd_ps(vn, vminus_ln2, vz);
132    $else:
133      __m512 vt = _mm512_fmadd_ps(vn, vminus_ln2_hi, vz);
134      vt = _mm512_fmadd_ps(vn, vminus_ln2_lo, vt);
135
136    __m512 vp = _mm512_fmadd_ps(vt, vc2, vc1);
137    vt = _mm512_mul_ps(vt, vl);
138    vp = _mm512_fmadd_ps(vt, vp, vl);
139
140    const __m512 ve = _mm512_scalef_ps(vp, vn);
141    const __m512 vd = _mm512_add_ps(ve, vone);
142
143    $if DIV_ALGO == "div":
144      __m512 vf = _mm512_div_ps(ve, vd);
145    $else:
146      __m512 vr = _mm512_rcp14_ps(vd);
147      vr = _mm512_fmadd_ps(_mm512_fnmadd_ps(vr, vd, vone), vr, vr);
148
149      __m512 vf = _mm512_mul_ps(ve, vr);
150      $if DIV_ALGO == "nr1fma1adj":
151        vf = _mm512_fmadd_ps(_mm512_fnmadd_ps(vf, vd, ve), vr, vf);
152
153    vf = _mm512_mask_sub_ps(vf, _mm512_testn_epi32_mask(_mm512_castps_si512(vx), vsign_mask), vone, vf);
154
155    _mm512_storeu_ps(y, vf);
156    y += 16;
157  }
158  if XNN_UNLIKELY(n != 0) {
159    assert(n >= 1 * sizeof(float));
160    assert(n <= 15 * sizeof(float));
161
162    // Prepare mask for valid 32-bit elements (depends on n).
163    n >>= 2 /* log2(sizeof(float)) */;
164    const __mmask16 vmask = _cvtu32_mask16((uint16_t) ((uint32_t) (UINT32_C(1) << n) - UINT32_C(1)));
165
166    const __m512 vx = _mm512_maskz_loadu_ps(vmask, x);
167    const __m512 vz = _mm512_castsi512_ps(_mm512_or_epi32(_mm512_castps_si512(vx), vsign_mask));
168
169    __m512 vn = _mm512_fmadd_ps(vz, vlog2e, vmagic_bias);
170    const __m512 vl = _mm512_permutex2var_ps(vtable_lo, _mm512_castps_si512(vn), vtable_hi);
171    vn = _mm512_sub_ps(vn, vmagic_bias);
172
173    $if RR_STEPS == 1:
174      __m512 vt = _mm512_fmadd_ps(vn, vminus_ln2, vz);
175    $else:
176      __m512 vt = _mm512_fmadd_ps(vn, vminus_ln2_hi, vz);
177      vt = _mm512_fmadd_ps(vn, vminus_ln2_lo, vt);
178
179    __m512 vp = _mm512_fmadd_ps(vt, vc2, vc1);
180    vt = _mm512_mul_ps(vt, vl);
181    vp = _mm512_fmadd_ps(vt, vp, vl);
182
183    const __m512 ve = _mm512_scalef_ps(vp, vn);
184    const __m512 vd = _mm512_add_ps(ve, vone);
185
186    $if DIV_ALGO == "div":
187      __m512 vf = _mm512_div_ps(ve, vd);
188    $else:
189      __m512 vr = _mm512_rcp14_ps(vd);
190      vr = _mm512_fmadd_ps(_mm512_fnmadd_ps(vr, vd, vone), vr, vr);
191
192      __m512 vf = _mm512_mul_ps(ve, vr);
193      $if DIV_ALGO == "nr1fma1adj":
194        vf = _mm512_fmadd_ps(_mm512_fnmadd_ps(vf, vd, ve), vr, vf);
195
196    vf = _mm512_mask_sub_ps(vf, _mm512_testn_epi32_mask(_mm512_castps_si512(vx), vsign_mask), vone, vf);
197
198    _mm512_mask_storeu_ps(y, vmask, vf);
199  }
200}
201