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 #include <assert.h>
7 #include <stddef.h>
8
9 #include <immintrin.h>
10
11 #include <xnnpack/common.h>
12 #include <xnnpack/math-stubs.h>
13
14
15 // Table of exp2(k / 16) values decremented (as integer) by (k << 19), k = 0..15
16 extern XNN_INTERNAL const float xnn_table_exp2minus_k_over_16[16];
17
xnn_math_f32_expm1minus__avx_rr2_lut16_p3(size_t n,const float * input,float * output)18 void xnn_math_f32_expm1minus__avx_rr2_lut16_p3(
19 size_t n,
20 const float* input,
21 float* output)
22 {
23 assert(n % (8 * sizeof(float)) == 0);
24
25 // The largest x for which expm1f(x) is saturated at -1.0f.
26 const __m256 vsat_cutoff = _mm256_set1_ps(-0x1.154246p+4f);
27 // Large number such that ulp(magic bias) == exp2(-4)
28 const __m256 vmagic_bias = _mm256_set1_ps(0x1.800000p19f);
29 const __m256 vlog2e = _mm256_set1_ps(0x1.715476p+0f);
30 // Mask for the lowest 4 bits
31 const __m256 vindex_mask = _mm256_castsi256_ps(_mm256_set1_epi32(0xF));
32 // Last 9 bits are zeroes
33 const __m256 vminus_ln2_hi = _mm256_set1_ps(-0x1.62E400p-1f);
34 const __m256 vminus_ln2_lo = _mm256_set1_ps(-0x1.7F7D1Cp-20f);
35 // Coefficient of polynomial approximation
36 // exp(t) - 1 ~ t * (1 + t * (c2 + t * c3))
37 // on [-log(2)/32, log(2)/32]
38 const __m256 vc3 = _mm256_set1_ps(0x1.55561Cp-3f);
39 const __m256 vc2 = _mm256_set1_ps(0x1.0001ECp-1f);
40 const __m256 vone = _mm256_set1_ps(1.0f);
41
42 for (; n != 0; n -= 8 * sizeof(float)) {
43 __m256 vx = _mm256_loadu_ps(input);
44
45 // The function saturates at -1 for large negative inputs: expm1f(x) == -1.0f for x <= sat_cutoff ~= -17.328680.
46 // To guarantee this behaviour, we clip input at sat_cutoff, and leverage the fact that for our implementation
47 // expm1f(sat_cutoff) == -1.0f. The order of operands in the [V]MAXPS instruction matters: it ensures that NaN
48 // inputs are passed unchanged.
49 vx = _mm256_max_ps(vsat_cutoff, vx);
50
51 // Compute reduced argument n := round(x / log(2), 4).
52 // We do it by adding a large number (magic bias), which cause rounding of the result to 4 fractional bits, then
53 // subtracing the large number back. The trick with adding large number is valid only within certain bounds
54 // (|x / log(2)| <= 2**18, i.e. |x| <= 0x1.62E43p+17 = 181704.375), but that is acceptable, because inputs x are
55 // restricted to [-17.328680, 0].
56 // Note that addition-subtraction of the large number doesn't cause overflow for inputs in this range.
57 __m256 vn = _mm256_add_ps(_mm256_mul_ps(vx, vlog2e), vmagic_bias);
58
59 // Create a floating-point number s (scale) such that s := 2**n for valid inputs, i.e. -17.328680 <= x <= 0.0. As n
60 // has 4 fractional bits, we split s == 2**n = 2**int(n) * 2**frac(n). We create s in two steps:
61 // 1. Fetch 2**frac(n) from the table using the 4 low bits of n, as integer. Note that the fetched values are in
62 // the [1.0, 2.0) range, i.e. their floating-point exponent is 0.
63 // 2. Adjust fecthed value by addition of int(n) to its floating-point exponent. The result is always a normalized
64 // number, because for -17.328680 <= x <= 0.0 we have -25 <= int(n) <= 0, and thus the adjusted exponent is not
65 // lower than -25.
66 //
67 // Shift bits 4:12 into 23:31 (position of floating-point exponent).
68 const __m128i ven_lo = _mm_slli_epi32(_mm_castps_si128(_mm256_castps256_ps128(vn)), 19);
69 const __m128i ven_hi = _mm_slli_epi32(_mm_castps_si128(_mm256_extractf128_ps(vn, 1)), 19);
70
71 // Use bits 0:4 bits of n, as integer, as an index for table lookup of l := 2**frac(n).
72 const __m256 vidx = _mm256_and_ps(vn, vindex_mask);
73 const __m128i vidx_lo = _mm_slli_epi32(_mm_castps_si128(_mm256_castps256_ps128(vidx)), 2);
74 const __m128i vidx_hi = _mm_slli_epi32(_mm_castps_si128(_mm256_extractf128_ps(vidx, 1)), 2);
75 #if XNN_ARCH_X86_64
76 const uint64_t vidx_ll = (uint64_t) _mm_cvtsi128_si64(vidx_lo);
77 const uint64_t vidx_lh = (uint64_t) _mm_extract_epi64(vidx_lo, 1);
78 const uint64_t vidx_hl = (uint64_t) _mm_cvtsi128_si64(vidx_hi);
79 const uint64_t vidx_hh = (uint64_t) _mm_extract_epi64(vidx_hi, 1);
80 __m128i vl_ll = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) vidx_ll)));
81 __m128i vl_lh = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) vidx_lh)));
82 __m128i vl_hl = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) vidx_hl)));
83 __m128i vl_hh = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) vidx_hh)));
84 vl_ll = _mm_insert_epi32(vl_ll, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) (vidx_ll >> 32))), 1);
85 vl_lh = _mm_insert_epi32(vl_lh, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) (vidx_lh >> 32))), 1);
86 vl_hl = _mm_insert_epi32(vl_hl, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) (vidx_hl >> 32))), 1);
87 vl_hh = _mm_insert_epi32(vl_hh, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) (vidx_hh >> 32))), 1);
88 #else
89 __m128i vl_ll = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) _mm_cvtsi128_si32(vidx_lo))));
90 __m128i vl_lh = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) _mm_extract_epi32(vidx_lo, 2))));
91 __m128i vl_hl = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) _mm_cvtsi128_si32(vidx_hi))));
92 __m128i vl_hh = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) _mm_extract_epi32(vidx_hi, 2))));
93 vl_ll = _mm_insert_epi32(vl_ll, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) _mm_extract_epi32(vidx_lo, 1))), 1);
94 vl_lh = _mm_insert_epi32(vl_lh, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) _mm_extract_epi32(vidx_lo, 3))), 1);
95 vl_hl = _mm_insert_epi32(vl_hl, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) _mm_extract_epi32(vidx_hi, 1))), 1);
96 vl_hh = _mm_insert_epi32(vl_hh, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) _mm_extract_epi32(vidx_hi, 3))), 1);
97 #endif
98 const __m128i vl_lo = _mm_unpacklo_epi64(vl_ll, vl_lh);
99 const __m128i vl_hi = _mm_unpacklo_epi64(vl_hl, vl_hh);
100 // Adjust exponent of the value l fetched from the table to get the final s value.
101 const __m128 vs_lo = _mm_castsi128_ps(_mm_add_epi32(vl_lo, ven_lo));
102 const __m128 vs_hi = _mm_castsi128_ps(_mm_add_epi32(vl_hi, ven_hi));
103 const __m256 vs = _mm256_insertf128_ps(_mm256_castps128_ps256(vs_lo), vs_hi, 1);
104
105 // Subtract the large number back to get final n := round(x / log(2), 4).
106 vn = _mm256_sub_ps(vn, vmagic_bias);
107
108 // Compute reduced argument t := x - n * log(2).
109 // Use Cody-Waite range reduction method (note two constants to represent log(2)) to improve accuracy.
110 __m256 vt = _mm256_add_ps(_mm256_mul_ps(vn, vminus_ln2_hi), vx);
111 vt = _mm256_add_ps(_mm256_mul_ps(vn, vminus_ln2_lo), vt);
112
113 // Compute degree-3 polynomial approximation for exp(t) - 1 on [-log(2)/32, log(2)/32].
114 // P(t) = t * (1 + t * (c2 + t * c3)) = t + t * (t * (c2 + t * c3)) = t + t * p
115 __m256 vp = _mm256_add_ps(_mm256_mul_ps(vc3, vt), vc2);
116 vp = _mm256_mul_ps(vp, vt);
117
118 // Reconstruct the exp(x) - 1 value:
119 // exp(x) - 1 = s * (1 + t * (1 + t * (c2 + t * c3))) - 1
120 // = (s - 1) + s * (t + t * p)
121 // = ((t * s) + (t * s) * p) + (s - 1)
122 vt = _mm256_mul_ps(vt, vs);
123 const __m256 vsm1 = _mm256_sub_ps(vs, vone);
124 vp = _mm256_add_ps(_mm256_mul_ps(vp, vt), vt);
125 const __m256 vf = _mm256_add_ps(vp, vsm1);
126
127 _mm256_storeu_ps(output, vf);
128
129 input += 8;
130 output += 8;
131 }
132 }
133