xref: /aosp_15_r20/external/XNNPACK/src/f32-raddextexp/avx2-p5.c.in (revision 4bdc94577ba0e567308109d787f7fec7b531ce36)
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$assert ELEMENTS_TILE % 8 == 0
7$assert ELEMENTS_TILE >= 8
8$SIMD_TILE = ELEMENTS_TILE // 8
9$ABC = "0123456789ABCDEFGHIJKLMNOPQRSTUVWXYZ"
10#include <assert.h>
11#include <math.h>
12
13#include <immintrin.h>
14
15#include <xnnpack/raddextexp.h>
16
17
18static const int32_t mask_table[14] = {-1, -1, -1, -1, -1, -1, -1, 0, 0, 0, 0, 0, 0, 0};
19
20void xnn_f32_raddextexp_ukernel__avx2_p5_x${ELEMENTS_TILE}${"" if ACCUMULATORS == 1 else "_acc%d" % ACCUMULATORS}(
21    size_t elements,
22    const float* x,
23    float* sum)
24{
25  assert(elements % sizeof(float) == 0);
26
27  const __m256 vlog2e = _mm256_set1_ps(0x1.715476p+0f);
28  const __m256 vminus_ln2_hi = _mm256_set1_ps(-0x1.62E43p-1f);
29  const __m256 vminus_ln2_lo = _mm256_set1_ps(0x1.05C61p-29f);
30
31  // The smallest elements such that 2**elements is considered non-negligible.
32  // For smaller elements, 2**elements is replaced with zero.
33  const __m256 vmin_exponent = _mm256_set1_ps(-127.0f);
34  const __m256 vmagic_bias = _mm256_set1_ps(0x1.8000FEp23f);
35  const __m256 vminus_inf = _mm256_set1_ps(-INFINITY);
36
37  const __m256 vc0 = _mm256_set1_ps(1.0f);
38  const __m256 vc1 = _mm256_set1_ps(0x1.FFFFF6p-1f);
39  const __m256 vc2 = _mm256_set1_ps(0x1.FFFDC6p-2f);
40  const __m256 vc3 = _mm256_set1_ps(0x1.555A80p-3f);
41  const __m256 vc4 = _mm256_set1_ps(0x1.573A1Ap-5f);
42  const __m256 vc5 = _mm256_set1_ps(0x1.0F9F9Cp-7f);
43
44  $for K in range(ACCUMULATORS):
45    __m256 vaccv${K} = _mm256_setzero_ps();
46  $for K in range(ACCUMULATORS):
47    __m256 vacce${K} = vminus_inf;
48  for (; elements >= ${ELEMENTS_TILE} * sizeof(float); elements -= ${ELEMENTS_TILE} * sizeof(float)) {
49    // Load ${ELEMENTS_TILE} (${SIMD_TILE}x8) inputs at a time.
50    const __m256 vx0 = _mm256_loadu_ps(x);
51    $for N in range(1, SIMD_TILE):
52      const __m256 vx${N} = _mm256_loadu_ps(x + ${N * 8});
53    x += ${ELEMENTS_TILE};
54
55    // Compute reduced argument elements := round(x / log(2)).
56    $for N in range(SIMD_TILE):
57      const __m256 vn${N} = _mm256_round_ps(_mm256_mul_ps(vx${N}, vlog2e), _MM_FROUND_TO_NEAREST_INT | _MM_FROUND_NO_EXC);
58
59    // Compute reduced argument t := x - elements * log(2).
60    // Use Cody-Waite range reduction method (note two constants to represent log(2)) to improve accuracy.
61    $for N in range(SIMD_TILE):
62      __m256 vt${N} = _mm256_fmadd_ps(vn${N}, vminus_ln2_hi, vx${N});
63
64    $for N in range(SIMD_TILE):
65      vt${N} = _mm256_fmadd_ps(vn${N}, vminus_ln2_lo, vt${N});
66
67    // Compute degree-5 polynomial approximation for exp(t) on [-log(2)/2, log(2)/2].
68    $for N in range(SIMD_TILE):
69      __m256 vp${N} = _mm256_fmadd_ps(vc5, vt${N}, vc4);
70
71    $for N in range(SIMD_TILE):
72      vp${N} = _mm256_fmadd_ps(vp${N}, vt${N}, vc3);
73
74    $for N in range(SIMD_TILE):
75      vp${N} = _mm256_fmadd_ps(vp${N}, vt${N}, vc2);
76
77    $for N in range(SIMD_TILE):
78      vp${N} = _mm256_fmadd_ps(vp${N}, vt${N}, vc1);
79
80    $for N in range(SIMD_TILE):
81      vp${N} = _mm256_fmadd_ps(vp${N}, vt${N}, vc0);
82
83    // Accumulate "extended" floating-point numbers in ("mantissa", "exponent") representation where
84    //  - vnX is "exponent"
85    //  - vpX is "mantissa"
86    //
87    // exp2(ae) * av + exp2(be) * bv =
88    //   = exp2(max(ae, be)) * exp2(ae - max(ae, be)) * av + exp2(max(ae, be)) * exp2(be - max(ae, be)) * bv
89    //   = exp2(max_e) * (exp2(ae - max_e) * av + exp2(be - max_e) * bv)
90    //   = exp2(max_e) * (exp2(delta_ae) * av + exp2(delta_be) * bv)
91    //
92    // For computational efficiency we may add several "extended" floating-point numbers at a time.
93    $for N in range(SIMD_TILE):
94      $if N < ACCUMULATORS:
95        __m256 vmax_e${N} = _mm256_max_ps(vacce${N}, vn${N});
96      $else:
97        vmax_e${N % ACCUMULATORS} = _mm256_max_ps(vmax_e${N % ACCUMULATORS}, vn${N});
98
99    // For computational efficiency, replace exp2(delta_e) with 0.0f when delta_e <= -127.0.
100    // This replacement is done in two steps:
101    // 1. Clamp minimum delta_e at -127.0.
102    // 2. Map delta_e to scale factor 0.0 when delta_e == -127.0
103    $for K in range(ACCUMULATORS):
104      const __m256 vdelta_acce${K} = _mm256_max_ps(_mm256_sub_ps(vacce${K}, vmax_e${K}), vmin_exponent);
105    $for N in range(SIMD_TILE):
106      const __m256 vdelta_e${N} = _mm256_max_ps(_mm256_sub_ps(vn${N}, vmax_e${N % ACCUMULATORS}), vmin_exponent);
107
108    // Convert delta-exponents into scale factors:
109    // - s = exp2(delta_e) when delta_e > -127.0
110    // - s = 0.0 when delta_e <= -127.0
111    //
112    // Note: delta-exponents can not exceed 0.0, thus scale factors can not exceed 1.0.
113    $for K in range(ACCUMULATORS):
114      const __m256 vaccs${K} = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(_mm256_add_ps(vdelta_acce${K}, vmagic_bias)), 23));
115    $for N in range(SIMD_TILE):
116      const __m256 vs${N} = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(_mm256_add_ps(vdelta_e${N}, vmagic_bias)), 23));
117
118    // Update accumulated "mantissa" and "exponent" values
119    $for K in range(ACCUMULATORS):
120      vaccv${K} = _mm256_mul_ps(vaccv${K}, vaccs${K});
121    $for N in range(SIMD_TILE):
122      vaccv${N % ACCUMULATORS} = _mm256_fmadd_ps(vp${N}, vs${N}, vaccv${N % ACCUMULATORS});
123
124    $for K in range(ACCUMULATORS):
125      vacce${K} = vmax_e${K};
126  }
127
128  // Reduce partial sums of "extended" floating-point numbers into a single "extended" SIMD vector of sums.
129  $if ACCUMULATORS > 1:
130    $for A in range(0, ACCUMULATORS, 2):
131      $if A + 1 < ACCUMULATORS:
132        const __m256 vmax_acce${ABC[A:A+2]} = _mm256_max_ps(vacce${A}, vacce${A+1});
133      $else:
134        const __m256 vmax_acce${ABC[A]} = vacce${A};
135    $ACC_SLICE = 2
136    $while ACC_SLICE < ACCUMULATORS:
137      $for A in range(0, ACCUMULATORS, ACC_SLICE * 2):
138        $if A + ACC_SLICE < ACCUMULATORS:
139          const __m256 vmax_acce${ABC[A:min(A+ACC_SLICE*2, ACCUMULATORS)]} = _mm256_max_ps(vmax_acce${ABC[A:A+ACC_SLICE]}, vmax_acce${ABC[A+ACC_SLICE:min(ACCUMULATORS,A+ACC_SLICE*2)]});
140      $ACC_SLICE *= 2
141
142    $for K in range(ACCUMULATORS):
143      const __m256 vdelta_acce${K} = _mm256_max_ps(_mm256_sub_ps(vacce${K}, vmax_acce${ABC[0:ACCUMULATORS]}), vmin_exponent);
144
145    $for K in range(ACCUMULATORS):
146      const __m256 vaccs${K} = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(_mm256_add_ps(vdelta_acce${K}, vmagic_bias)), 23));
147
148    __m256 vaccv = _mm256_mul_ps(vaccv0, vaccs0);
149    $for K in range(1, ACCUMULATORS):
150      vaccv = _mm256_fmadd_ps(vaccv${K}, vaccs${K}, vaccv);
151    __m256 vacce = vmax_acce${ABC[0:ACCUMULATORS]};
152  $else:
153    __m256 vaccv = vaccv0;
154    __m256 vacce = vacce0;
155
156  for (; elements >= 8 * sizeof(float); elements -= 8 * sizeof(float)) {
157    // Load 8 inputs at a time.
158    const __m256 vx = _mm256_loadu_ps(x);
159    x += 8;
160
161    // Compute reduced argument elements := round(x / log(2)).
162    const __m256 vn = _mm256_round_ps(_mm256_mul_ps(vx, vlog2e), _MM_FROUND_TO_NEAREST_INT | _MM_FROUND_NO_EXC);
163
164    // Compute reduced argument t := x - elements * log(2).
165    // Use Cody-Waite range reduction method (note two constants to represent log(2)) to improve accuracy.
166    __m256 vt = _mm256_fmadd_ps(vn, vminus_ln2_hi, vx);
167    vt = _mm256_fmadd_ps(vn, vminus_ln2_lo, vt);
168
169    // Compute degree-5 polynomial approximation for exp(t) on [-log(2)/2, log(2)/2].
170    __m256 vp = _mm256_fmadd_ps(vc5, vt, vc4);
171    vp = _mm256_fmadd_ps(vp, vt, vc3);
172    vp = _mm256_fmadd_ps(vp, vt, vc2);
173    vp = _mm256_fmadd_ps(vp, vt, vc1);
174    vp = _mm256_fmadd_ps(vp, vt, vc0);
175
176    // Accumulate "extended" floating-point numbers in ("mantissa", "exponent") representation.
177    const __m256 vmax_e = _mm256_max_ps(vacce, vn);
178
179    // For computational efficiency, clamp minimum exp2(delta_e) at -127.0. It will be mapped to 0.0 scale factor later.
180    const __m256 vdelta_acce = _mm256_max_ps(_mm256_sub_ps(vacce, vmax_e), vmin_exponent);
181    const __m256 vdelta_e = _mm256_max_ps(_mm256_sub_ps(vn, vmax_e), vmin_exponent);
182
183    // Convert exponents into scale factors.
184    const __m256 vaccs = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(_mm256_add_ps(vdelta_acce, vmagic_bias)), 23));
185    const __m256 vs = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(_mm256_add_ps(vdelta_e, vmagic_bias)), 23));
186
187    // Update accumulated "mantissa" and "exponent" values.
188    vaccv = _mm256_mul_ps(vaccv, vaccs);
189    vaccv = _mm256_fmadd_ps(vp, vs, vaccv);
190
191    vacce = vmax_e;
192  }
193  if XNN_UNLIKELY(elements != 0) {
194    assert(elements >= 1 * sizeof(float));
195    assert(elements <= 7 * sizeof(float));
196    const __m256i vmask = _mm256_loadu_si256((const __m256i*) ((uintptr_t) &mask_table[7] - elements));
197
198    // Load up to 7 inputs at a time.
199    const __m256 vx = _mm256_maskload_ps(x, vmask);
200
201    // Compute reduced argument elements := round(x / log(2)).
202    __m256 vn = _mm256_round_ps(_mm256_mul_ps(vx, vlog2e), _MM_FROUND_TO_NEAREST_INT | _MM_FROUND_NO_EXC);
203
204    // Compute reduced argument t := x - elements * log(2).
205    // Use Cody-Waite range reduction method (note two constants to represent log(2)) to improve accuracy.
206    __m256 vt = _mm256_fmadd_ps(vn, vminus_ln2_hi, vx);
207    vt = _mm256_fmadd_ps(vn, vminus_ln2_lo, vt);
208
209    // Correct reduced argument elements for masked out elements.
210    vn = _mm256_blendv_ps(vacce, vn, _mm256_castsi256_ps(vmask));
211
212    // Compute degree-5 polynomial approximation for exp(t) on [-log(2)/2, log(2)/2].
213    __m256 vp = _mm256_fmadd_ps(vc5, vt, vc4);
214    vp = _mm256_fmadd_ps(vp, vt, vc3);
215    vp = _mm256_fmadd_ps(vp, vt, vc2);
216    vp = _mm256_fmadd_ps(vp, vt, vc1);
217    vp = _mm256_fmadd_ps(vp, vt, vc0);
218    vp = _mm256_and_ps(vp, _mm256_castsi256_ps(vmask));
219
220    // Accumulate "extended" floating-point numbers in ("mantissa", "exponent") representation.
221    const __m256 vmax_e = _mm256_max_ps(vacce, vn);
222
223    // For computational efficiency, clamp minimum exp2(delta_e) at -127.0. It will be mapped to 0.0 scale factor later.
224    const __m256 vdelta_e = _mm256_max_ps(_mm256_sub_ps(vn, vmax_e), vmin_exponent);
225    const __m256 vdelta_acce = _mm256_max_ps(_mm256_sub_ps(vacce, vmax_e), vmin_exponent);
226
227    // Convert exponents into scale factors.
228    const __m256 vs = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(_mm256_add_ps(vdelta_e, vmagic_bias)), 23));
229    const __m256 vaccs = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(_mm256_add_ps(vdelta_acce, vmagic_bias)), 23));
230
231    // Update accumulated "mantissa" and "exponent" values.
232    vaccv = _mm256_mul_ps(vaccv, vaccs);
233    vaccv = _mm256_fmadd_ps(vp, vs, vaccv);
234
235    vacce = vmax_e;
236  }
237
238  // Reduce partial sums of "extended" floating-point numbers into a single "extended" floating-point sum.
239  __m256 vmax_acce = _mm256_max_ps(vacce, _mm256_permute2f128_ps(vacce, vacce, 1));
240  vmax_acce = _mm256_max_ps(vmax_acce, _mm256_shuffle_ps(vmax_acce, vmax_acce, _MM_SHUFFLE(1, 0, 3, 2)));
241  vmax_acce = _mm256_max_ps(vmax_acce, _mm256_shuffle_ps(vmax_acce, vmax_acce, _MM_SHUFFLE(2, 3, 0, 1)));
242  const __m256 vdelta_acce = _mm256_max_ps(_mm256_sub_ps(vacce, vmax_acce), vmin_exponent);
243  const __m256 vaccs = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(_mm256_add_ps(vdelta_acce, vmagic_bias)), 23));
244
245  vaccv = _mm256_mul_ps(vaccv, vaccs);
246  __m128 vaccv_sum = _mm_add_ps(_mm256_castps256_ps128(vaccv), _mm256_extractf128_ps(vaccv, 1));
247  vaccv_sum = _mm_add_ps(vaccv_sum, _mm_movehl_ps(vaccv_sum, vaccv_sum));
248  vaccv_sum = _mm_add_ss(vaccv_sum, _mm_movehdup_ps(vaccv_sum));
249
250  _mm_store_ss(&sum[0], vaccv_sum);
251  _mm_store_ss(&sum[1], _mm256_castps256_ps128(vmax_acce));
252
253  _mm256_zeroupper();
254}
255