1*4bdc9457SAndroid Build Coastguard Worker // Auto-generated file. Do not edit!
2*4bdc9457SAndroid Build Coastguard Worker // Template: src/f32-velu/sse-rr2-lut16-p3.c.in
3*4bdc9457SAndroid Build Coastguard Worker // Generator: tools/xngen
4*4bdc9457SAndroid Build Coastguard Worker //
5*4bdc9457SAndroid Build Coastguard Worker // Copyright 2020 Google LLC
6*4bdc9457SAndroid Build Coastguard Worker //
7*4bdc9457SAndroid Build Coastguard Worker // This source code is licensed under the BSD-style license found in the
8*4bdc9457SAndroid Build Coastguard Worker // LICENSE file in the root directory of this source tree.
9*4bdc9457SAndroid Build Coastguard Worker
10*4bdc9457SAndroid Build Coastguard Worker #include <assert.h>
11*4bdc9457SAndroid Build Coastguard Worker
12*4bdc9457SAndroid Build Coastguard Worker #include <emmintrin.h>
13*4bdc9457SAndroid Build Coastguard Worker
14*4bdc9457SAndroid Build Coastguard Worker #include <xnnpack/vunary.h>
15*4bdc9457SAndroid Build Coastguard Worker #include <xnnpack/common.h>
16*4bdc9457SAndroid Build Coastguard Worker
17*4bdc9457SAndroid Build Coastguard Worker
18*4bdc9457SAndroid Build Coastguard Worker extern XNN_INTERNAL const float xnn_table_exp2minus_k_over_16[16];
19*4bdc9457SAndroid Build Coastguard Worker
xnn_f32_velu_ukernel__sse2_rr2_lut16_p3_x16(size_t n,const float * x,float * y,const union xnn_f32_elu_params params[restrict XNN_MIN_ELEMENTS (1)])20*4bdc9457SAndroid Build Coastguard Worker void xnn_f32_velu_ukernel__sse2_rr2_lut16_p3_x16(
21*4bdc9457SAndroid Build Coastguard Worker size_t n,
22*4bdc9457SAndroid Build Coastguard Worker const float* x,
23*4bdc9457SAndroid Build Coastguard Worker float* y,
24*4bdc9457SAndroid Build Coastguard Worker const union xnn_f32_elu_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
25*4bdc9457SAndroid Build Coastguard Worker {
26*4bdc9457SAndroid Build Coastguard Worker assert(n != 0);
27*4bdc9457SAndroid Build Coastguard Worker assert(n % sizeof(float) == 0);
28*4bdc9457SAndroid Build Coastguard Worker assert(x != NULL);
29*4bdc9457SAndroid Build Coastguard Worker assert(y != NULL);
30*4bdc9457SAndroid Build Coastguard Worker
31*4bdc9457SAndroid Build Coastguard Worker const __m128 vprescale = _mm_load_ps(params->sse2_rr2_lut16_p3.prescale);
32*4bdc9457SAndroid Build Coastguard Worker const __m128 valpha = _mm_load_ps(params->sse2_rr2_lut16_p3.alpha);
33*4bdc9457SAndroid Build Coastguard Worker const __m128 vbeta = _mm_load_ps(params->sse2_rr2_lut16_p3.beta);
34*4bdc9457SAndroid Build Coastguard Worker const __m128 vsat_cutoff = _mm_load_ps(params->sse2_rr2_lut16_p3.sat_cutoff);
35*4bdc9457SAndroid Build Coastguard Worker const __m128 vmagic_bias = _mm_load_ps(params->sse2_rr2_lut16_p3.magic_bias);
36*4bdc9457SAndroid Build Coastguard Worker const __m128 vlog2e = _mm_load_ps(params->sse2_rr2_lut16_p3.log2e);
37*4bdc9457SAndroid Build Coastguard Worker const __m128i vindex_mask = _mm_load_si128((const __m128i*) params->sse2_rr2_lut16_p3.index_mask);
38*4bdc9457SAndroid Build Coastguard Worker const __m128 vminus_ln2_hi = _mm_load_ps(params->sse2_rr2_lut16_p3.minus_ln2_hi);
39*4bdc9457SAndroid Build Coastguard Worker const __m128 vminus_ln2_lo = _mm_load_ps(params->sse2_rr2_lut16_p3.minus_ln2_lo);
40*4bdc9457SAndroid Build Coastguard Worker const __m128 vc3 = _mm_load_ps(params->sse2_rr2_lut16_p3.c3);
41*4bdc9457SAndroid Build Coastguard Worker const __m128 vc2 = _mm_load_ps(params->sse2_rr2_lut16_p3.c2);
42*4bdc9457SAndroid Build Coastguard Worker const __m128 vone = _mm_load_ps(params->sse2_rr2_lut16_p3.one);
43*4bdc9457SAndroid Build Coastguard Worker
44*4bdc9457SAndroid Build Coastguard Worker for (; n >= 16 * sizeof(float); n -= 16 * sizeof(float)) {
45*4bdc9457SAndroid Build Coastguard Worker __m128 vx0123 = _mm_loadu_ps(x);
46*4bdc9457SAndroid Build Coastguard Worker __m128 vx4567 = _mm_loadu_ps(x + 4);
47*4bdc9457SAndroid Build Coastguard Worker __m128 vx89AB = _mm_loadu_ps(x + 8);
48*4bdc9457SAndroid Build Coastguard Worker __m128 vxCDEF = _mm_loadu_ps(x + 12);
49*4bdc9457SAndroid Build Coastguard Worker x += 16;
50*4bdc9457SAndroid Build Coastguard Worker
51*4bdc9457SAndroid Build Coastguard Worker const __m128 vz0123 = _mm_max_ps(vsat_cutoff, _mm_mul_ps(vx0123, vprescale));
52*4bdc9457SAndroid Build Coastguard Worker const __m128 vz4567 = _mm_max_ps(vsat_cutoff, _mm_mul_ps(vx4567, vprescale));
53*4bdc9457SAndroid Build Coastguard Worker const __m128 vz89AB = _mm_max_ps(vsat_cutoff, _mm_mul_ps(vx89AB, vprescale));
54*4bdc9457SAndroid Build Coastguard Worker const __m128 vzCDEF = _mm_max_ps(vsat_cutoff, _mm_mul_ps(vxCDEF, vprescale));
55*4bdc9457SAndroid Build Coastguard Worker
56*4bdc9457SAndroid Build Coastguard Worker __m128 vn0123 = _mm_add_ps(_mm_mul_ps(vz0123, vlog2e), vmagic_bias);
57*4bdc9457SAndroid Build Coastguard Worker __m128 vn4567 = _mm_add_ps(_mm_mul_ps(vz4567, vlog2e), vmagic_bias);
58*4bdc9457SAndroid Build Coastguard Worker __m128 vn89AB = _mm_add_ps(_mm_mul_ps(vz89AB, vlog2e), vmagic_bias);
59*4bdc9457SAndroid Build Coastguard Worker __m128 vnCDEF = _mm_add_ps(_mm_mul_ps(vzCDEF, vlog2e), vmagic_bias);
60*4bdc9457SAndroid Build Coastguard Worker
61*4bdc9457SAndroid Build Coastguard Worker const __m128i vidx0123 = _mm_slli_epi32(_mm_and_si128(_mm_castps_si128(vn0123), vindex_mask), 2);
62*4bdc9457SAndroid Build Coastguard Worker const __m128i ven0123 = _mm_slli_epi32(_mm_castps_si128(vn0123), 19);
63*4bdc9457SAndroid Build Coastguard Worker const __m128i vidx4567 = _mm_slli_epi32(_mm_and_si128(_mm_castps_si128(vn4567), vindex_mask), 2);
64*4bdc9457SAndroid Build Coastguard Worker const __m128i ven4567 = _mm_slli_epi32(_mm_castps_si128(vn4567), 19);
65*4bdc9457SAndroid Build Coastguard Worker const __m128i vidx89AB = _mm_slli_epi32(_mm_and_si128(_mm_castps_si128(vn89AB), vindex_mask), 2);
66*4bdc9457SAndroid Build Coastguard Worker const __m128i ven89AB = _mm_slli_epi32(_mm_castps_si128(vn89AB), 19);
67*4bdc9457SAndroid Build Coastguard Worker const __m128i vidxCDEF = _mm_slli_epi32(_mm_and_si128(_mm_castps_si128(vnCDEF), vindex_mask), 2);
68*4bdc9457SAndroid Build Coastguard Worker const __m128i venCDEF = _mm_slli_epi32(_mm_castps_si128(vnCDEF), 19);
69*4bdc9457SAndroid Build Coastguard Worker
70*4bdc9457SAndroid Build Coastguard Worker #if XNN_ARCH_X86_64
71*4bdc9457SAndroid Build Coastguard Worker const uint64_t vidx01 = (uint64_t) _mm_cvtsi128_si64(vidx0123);
72*4bdc9457SAndroid Build Coastguard Worker const uint64_t vidx23 = (uint64_t) _mm_cvtsi128_si64(_mm_unpackhi_epi64(vidx0123, vidx0123));
73*4bdc9457SAndroid Build Coastguard Worker const __m128i vl0 = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) vidx01)));
74*4bdc9457SAndroid Build Coastguard Worker const __m128i vl2 = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) vidx23)));
75*4bdc9457SAndroid Build Coastguard Worker const __m128i vl1 = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) (vidx01 >> 32))));
76*4bdc9457SAndroid Build Coastguard Worker const __m128i vl01 = _mm_unpacklo_epi32(vl0, vl1);
77*4bdc9457SAndroid Build Coastguard Worker const __m128i vl3 = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) (vidx23 >> 32))));
78*4bdc9457SAndroid Build Coastguard Worker const __m128i vl23 = _mm_unpacklo_epi32(vl2, vl3);
79*4bdc9457SAndroid Build Coastguard Worker const __m128i vl0123 = _mm_unpacklo_epi64(vl01, vl23);
80*4bdc9457SAndroid Build Coastguard Worker const uint64_t vidx45 = (uint64_t) _mm_cvtsi128_si64(vidx4567);
81*4bdc9457SAndroid Build Coastguard Worker const uint64_t vidx67 = (uint64_t) _mm_cvtsi128_si64(_mm_unpackhi_epi64(vidx4567, vidx4567));
82*4bdc9457SAndroid Build Coastguard Worker const __m128i vl4 = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) vidx45)));
83*4bdc9457SAndroid Build Coastguard Worker const __m128i vl6 = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) vidx67)));
84*4bdc9457SAndroid Build Coastguard Worker const __m128i vl5 = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) (vidx45 >> 32))));
85*4bdc9457SAndroid Build Coastguard Worker const __m128i vl45 = _mm_unpacklo_epi32(vl4, vl5);
86*4bdc9457SAndroid Build Coastguard Worker const __m128i vl7 = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) (vidx67 >> 32))));
87*4bdc9457SAndroid Build Coastguard Worker const __m128i vl67 = _mm_unpacklo_epi32(vl6, vl7);
88*4bdc9457SAndroid Build Coastguard Worker const __m128i vl4567 = _mm_unpacklo_epi64(vl45, vl67);
89*4bdc9457SAndroid Build Coastguard Worker const uint64_t vidx89 = (uint64_t) _mm_cvtsi128_si64(vidx89AB);
90*4bdc9457SAndroid Build Coastguard Worker const uint64_t vidxAB = (uint64_t) _mm_cvtsi128_si64(_mm_unpackhi_epi64(vidx89AB, vidx89AB));
91*4bdc9457SAndroid Build Coastguard Worker const __m128i vl8 = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) vidx89)));
92*4bdc9457SAndroid Build Coastguard Worker const __m128i vlA = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) vidxAB)));
93*4bdc9457SAndroid Build Coastguard Worker const __m128i vl9 = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) (vidx89 >> 32))));
94*4bdc9457SAndroid Build Coastguard Worker const __m128i vl89 = _mm_unpacklo_epi32(vl8, vl9);
95*4bdc9457SAndroid Build Coastguard Worker const __m128i vlB = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) (vidxAB >> 32))));
96*4bdc9457SAndroid Build Coastguard Worker const __m128i vlAB = _mm_unpacklo_epi32(vlA, vlB);
97*4bdc9457SAndroid Build Coastguard Worker const __m128i vl89AB = _mm_unpacklo_epi64(vl89, vlAB);
98*4bdc9457SAndroid Build Coastguard Worker const uint64_t vidxCD = (uint64_t) _mm_cvtsi128_si64(vidxCDEF);
99*4bdc9457SAndroid Build Coastguard Worker const uint64_t vidxEF = (uint64_t) _mm_cvtsi128_si64(_mm_unpackhi_epi64(vidxCDEF, vidxCDEF));
100*4bdc9457SAndroid Build Coastguard Worker const __m128i vlC = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) vidxCD)));
101*4bdc9457SAndroid Build Coastguard Worker const __m128i vlE = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) vidxEF)));
102*4bdc9457SAndroid Build Coastguard Worker const __m128i vlD = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) (vidxCD >> 32))));
103*4bdc9457SAndroid Build Coastguard Worker const __m128i vlCD = _mm_unpacklo_epi32(vlC, vlD);
104*4bdc9457SAndroid Build Coastguard Worker const __m128i vlF = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) (vidxEF >> 32))));
105*4bdc9457SAndroid Build Coastguard Worker const __m128i vlEF = _mm_unpacklo_epi32(vlE, vlF);
106*4bdc9457SAndroid Build Coastguard Worker const __m128i vlCDEF = _mm_unpacklo_epi64(vlCD, vlEF);
107*4bdc9457SAndroid Build Coastguard Worker #else // !XNN_ARCH_X86_64
108*4bdc9457SAndroid Build Coastguard Worker const uint32_t vidx0 = (uint32_t) _mm_cvtsi128_si32(vidx0123);
109*4bdc9457SAndroid Build Coastguard Worker const uint32_t vidx1 = (uint32_t) _mm_extract_epi16(vidx0123, 2);
110*4bdc9457SAndroid Build Coastguard Worker const uint32_t vidx2 = (uint32_t) _mm_extract_epi16(vidx0123, 4);
111*4bdc9457SAndroid Build Coastguard Worker const uint32_t vidx3 = (uint32_t) _mm_extract_epi16(vidx0123, 6);
112*4bdc9457SAndroid Build Coastguard Worker const __m128i vl0 = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + vidx0)));
113*4bdc9457SAndroid Build Coastguard Worker const __m128i vl2 = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + vidx2)));
114*4bdc9457SAndroid Build Coastguard Worker const __m128i vl1 = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + vidx1)));
115*4bdc9457SAndroid Build Coastguard Worker const __m128i vl01 = _mm_unpacklo_epi32(vl0, vl1);
116*4bdc9457SAndroid Build Coastguard Worker const __m128i vl3 = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + vidx3)));
117*4bdc9457SAndroid Build Coastguard Worker const __m128i vl23 = _mm_unpacklo_epi32(vl2, vl3);
118*4bdc9457SAndroid Build Coastguard Worker const __m128i vl0123 = _mm_unpacklo_epi64(vl01, vl23);
119*4bdc9457SAndroid Build Coastguard Worker const uint32_t vidx4 = (uint32_t) _mm_cvtsi128_si32(vidx4567);
120*4bdc9457SAndroid Build Coastguard Worker const uint32_t vidx5 = (uint32_t) _mm_extract_epi16(vidx4567, 2);
121*4bdc9457SAndroid Build Coastguard Worker const uint32_t vidx6 = (uint32_t) _mm_extract_epi16(vidx4567, 4);
122*4bdc9457SAndroid Build Coastguard Worker const uint32_t vidx7 = (uint32_t) _mm_extract_epi16(vidx4567, 6);
123*4bdc9457SAndroid Build Coastguard Worker const __m128i vl4 = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + vidx4)));
124*4bdc9457SAndroid Build Coastguard Worker const __m128i vl6 = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + vidx6)));
125*4bdc9457SAndroid Build Coastguard Worker const __m128i vl5 = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + vidx5)));
126*4bdc9457SAndroid Build Coastguard Worker const __m128i vl45 = _mm_unpacklo_epi32(vl4, vl5);
127*4bdc9457SAndroid Build Coastguard Worker const __m128i vl7 = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + vidx7)));
128*4bdc9457SAndroid Build Coastguard Worker const __m128i vl67 = _mm_unpacklo_epi32(vl6, vl7);
129*4bdc9457SAndroid Build Coastguard Worker const __m128i vl4567 = _mm_unpacklo_epi64(vl45, vl67);
130*4bdc9457SAndroid Build Coastguard Worker const uint32_t vidx8 = (uint32_t) _mm_cvtsi128_si32(vidx89AB);
131*4bdc9457SAndroid Build Coastguard Worker const uint32_t vidx9 = (uint32_t) _mm_extract_epi16(vidx89AB, 2);
132*4bdc9457SAndroid Build Coastguard Worker const uint32_t vidxA = (uint32_t) _mm_extract_epi16(vidx89AB, 4);
133*4bdc9457SAndroid Build Coastguard Worker const uint32_t vidxB = (uint32_t) _mm_extract_epi16(vidx89AB, 6);
134*4bdc9457SAndroid Build Coastguard Worker const __m128i vl8 = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + vidx8)));
135*4bdc9457SAndroid Build Coastguard Worker const __m128i vlA = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + vidxA)));
136*4bdc9457SAndroid Build Coastguard Worker const __m128i vl9 = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + vidx9)));
137*4bdc9457SAndroid Build Coastguard Worker const __m128i vl89 = _mm_unpacklo_epi32(vl8, vl9);
138*4bdc9457SAndroid Build Coastguard Worker const __m128i vlB = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + vidxB)));
139*4bdc9457SAndroid Build Coastguard Worker const __m128i vlAB = _mm_unpacklo_epi32(vlA, vlB);
140*4bdc9457SAndroid Build Coastguard Worker const __m128i vl89AB = _mm_unpacklo_epi64(vl89, vlAB);
141*4bdc9457SAndroid Build Coastguard Worker const uint32_t vidxC = (uint32_t) _mm_cvtsi128_si32(vidxCDEF);
142*4bdc9457SAndroid Build Coastguard Worker const uint32_t vidxD = (uint32_t) _mm_extract_epi16(vidxCDEF, 2);
143*4bdc9457SAndroid Build Coastguard Worker const uint32_t vidxE = (uint32_t) _mm_extract_epi16(vidxCDEF, 4);
144*4bdc9457SAndroid Build Coastguard Worker const uint32_t vidxF = (uint32_t) _mm_extract_epi16(vidxCDEF, 6);
145*4bdc9457SAndroid Build Coastguard Worker const __m128i vlC = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + vidxC)));
146*4bdc9457SAndroid Build Coastguard Worker const __m128i vlE = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + vidxE)));
147*4bdc9457SAndroid Build Coastguard Worker const __m128i vlD = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + vidxD)));
148*4bdc9457SAndroid Build Coastguard Worker const __m128i vlCD = _mm_unpacklo_epi32(vlC, vlD);
149*4bdc9457SAndroid Build Coastguard Worker const __m128i vlF = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + vidxF)));
150*4bdc9457SAndroid Build Coastguard Worker const __m128i vlEF = _mm_unpacklo_epi32(vlE, vlF);
151*4bdc9457SAndroid Build Coastguard Worker const __m128i vlCDEF = _mm_unpacklo_epi64(vlCD, vlEF);
152*4bdc9457SAndroid Build Coastguard Worker #endif // XNN_ARCH_X86_64
153*4bdc9457SAndroid Build Coastguard Worker
154*4bdc9457SAndroid Build Coastguard Worker vn0123 = _mm_sub_ps(vn0123, vmagic_bias);
155*4bdc9457SAndroid Build Coastguard Worker __m128 vs0123 = _mm_castsi128_ps(_mm_add_epi32(vl0123, ven0123));
156*4bdc9457SAndroid Build Coastguard Worker vn4567 = _mm_sub_ps(vn4567, vmagic_bias);
157*4bdc9457SAndroid Build Coastguard Worker __m128 vs4567 = _mm_castsi128_ps(_mm_add_epi32(vl4567, ven4567));
158*4bdc9457SAndroid Build Coastguard Worker vn89AB = _mm_sub_ps(vn89AB, vmagic_bias);
159*4bdc9457SAndroid Build Coastguard Worker __m128 vs89AB = _mm_castsi128_ps(_mm_add_epi32(vl89AB, ven89AB));
160*4bdc9457SAndroid Build Coastguard Worker vnCDEF = _mm_sub_ps(vnCDEF, vmagic_bias);
161*4bdc9457SAndroid Build Coastguard Worker __m128 vsCDEF = _mm_castsi128_ps(_mm_add_epi32(vlCDEF, venCDEF));
162*4bdc9457SAndroid Build Coastguard Worker
163*4bdc9457SAndroid Build Coastguard Worker __m128 vt0123 = _mm_add_ps(_mm_mul_ps(vn0123, vminus_ln2_hi), vz0123);
164*4bdc9457SAndroid Build Coastguard Worker __m128 vt4567 = _mm_add_ps(_mm_mul_ps(vn4567, vminus_ln2_hi), vz4567);
165*4bdc9457SAndroid Build Coastguard Worker __m128 vt89AB = _mm_add_ps(_mm_mul_ps(vn89AB, vminus_ln2_hi), vz89AB);
166*4bdc9457SAndroid Build Coastguard Worker __m128 vtCDEF = _mm_add_ps(_mm_mul_ps(vnCDEF, vminus_ln2_hi), vzCDEF);
167*4bdc9457SAndroid Build Coastguard Worker
168*4bdc9457SAndroid Build Coastguard Worker vt0123 = _mm_add_ps(_mm_mul_ps(vn0123, vminus_ln2_lo), vt0123);
169*4bdc9457SAndroid Build Coastguard Worker vt4567 = _mm_add_ps(_mm_mul_ps(vn4567, vminus_ln2_lo), vt4567);
170*4bdc9457SAndroid Build Coastguard Worker vt89AB = _mm_add_ps(_mm_mul_ps(vn89AB, vminus_ln2_lo), vt89AB);
171*4bdc9457SAndroid Build Coastguard Worker vtCDEF = _mm_add_ps(_mm_mul_ps(vnCDEF, vminus_ln2_lo), vtCDEF);
172*4bdc9457SAndroid Build Coastguard Worker
173*4bdc9457SAndroid Build Coastguard Worker __m128 vp0123 = _mm_add_ps(_mm_mul_ps(vc3, vt0123), vc2);
174*4bdc9457SAndroid Build Coastguard Worker __m128 vp4567 = _mm_add_ps(_mm_mul_ps(vc3, vt4567), vc2);
175*4bdc9457SAndroid Build Coastguard Worker __m128 vp89AB = _mm_add_ps(_mm_mul_ps(vc3, vt89AB), vc2);
176*4bdc9457SAndroid Build Coastguard Worker __m128 vpCDEF = _mm_add_ps(_mm_mul_ps(vc3, vtCDEF), vc2);
177*4bdc9457SAndroid Build Coastguard Worker
178*4bdc9457SAndroid Build Coastguard Worker vp0123 = _mm_mul_ps(vp0123, vt0123);
179*4bdc9457SAndroid Build Coastguard Worker vp4567 = _mm_mul_ps(vp4567, vt4567);
180*4bdc9457SAndroid Build Coastguard Worker vp89AB = _mm_mul_ps(vp89AB, vt89AB);
181*4bdc9457SAndroid Build Coastguard Worker vpCDEF = _mm_mul_ps(vpCDEF, vtCDEF);
182*4bdc9457SAndroid Build Coastguard Worker
183*4bdc9457SAndroid Build Coastguard Worker vt0123 = _mm_mul_ps(vt0123, vs0123);
184*4bdc9457SAndroid Build Coastguard Worker vs0123 = _mm_sub_ps(vs0123, vone);
185*4bdc9457SAndroid Build Coastguard Worker vt4567 = _mm_mul_ps(vt4567, vs4567);
186*4bdc9457SAndroid Build Coastguard Worker vs4567 = _mm_sub_ps(vs4567, vone);
187*4bdc9457SAndroid Build Coastguard Worker vt89AB = _mm_mul_ps(vt89AB, vs89AB);
188*4bdc9457SAndroid Build Coastguard Worker vs89AB = _mm_sub_ps(vs89AB, vone);
189*4bdc9457SAndroid Build Coastguard Worker vtCDEF = _mm_mul_ps(vtCDEF, vsCDEF);
190*4bdc9457SAndroid Build Coastguard Worker vsCDEF = _mm_sub_ps(vsCDEF, vone);
191*4bdc9457SAndroid Build Coastguard Worker
192*4bdc9457SAndroid Build Coastguard Worker vp0123 = _mm_add_ps(_mm_mul_ps(vp0123, vt0123), vt0123);
193*4bdc9457SAndroid Build Coastguard Worker vp4567 = _mm_add_ps(_mm_mul_ps(vp4567, vt4567), vt4567);
194*4bdc9457SAndroid Build Coastguard Worker vp89AB = _mm_add_ps(_mm_mul_ps(vp89AB, vt89AB), vt89AB);
195*4bdc9457SAndroid Build Coastguard Worker vpCDEF = _mm_add_ps(_mm_mul_ps(vpCDEF, vtCDEF), vtCDEF);
196*4bdc9457SAndroid Build Coastguard Worker
197*4bdc9457SAndroid Build Coastguard Worker const __m128 ve0123 = _mm_mul_ps(_mm_add_ps(vp0123, vs0123), valpha);
198*4bdc9457SAndroid Build Coastguard Worker const __m128 ve4567 = _mm_mul_ps(_mm_add_ps(vp4567, vs4567), valpha);
199*4bdc9457SAndroid Build Coastguard Worker const __m128 ve89AB = _mm_mul_ps(_mm_add_ps(vp89AB, vs89AB), valpha);
200*4bdc9457SAndroid Build Coastguard Worker const __m128 veCDEF = _mm_mul_ps(_mm_add_ps(vpCDEF, vsCDEF), valpha);
201*4bdc9457SAndroid Build Coastguard Worker
202*4bdc9457SAndroid Build Coastguard Worker const __m128 vm0123 = _mm_castsi128_ps(_mm_cmpgt_epi32(_mm_setzero_si128(), _mm_castps_si128(vx0123)));
203*4bdc9457SAndroid Build Coastguard Worker vx0123 = _mm_mul_ps(vx0123, vbeta);
204*4bdc9457SAndroid Build Coastguard Worker const __m128 vm4567 = _mm_castsi128_ps(_mm_cmpgt_epi32(_mm_setzero_si128(), _mm_castps_si128(vx4567)));
205*4bdc9457SAndroid Build Coastguard Worker vx4567 = _mm_mul_ps(vx4567, vbeta);
206*4bdc9457SAndroid Build Coastguard Worker const __m128 vm89AB = _mm_castsi128_ps(_mm_cmpgt_epi32(_mm_setzero_si128(), _mm_castps_si128(vx89AB)));
207*4bdc9457SAndroid Build Coastguard Worker vx89AB = _mm_mul_ps(vx89AB, vbeta);
208*4bdc9457SAndroid Build Coastguard Worker const __m128 vmCDEF = _mm_castsi128_ps(_mm_cmpgt_epi32(_mm_setzero_si128(), _mm_castps_si128(vxCDEF)));
209*4bdc9457SAndroid Build Coastguard Worker vxCDEF = _mm_mul_ps(vxCDEF, vbeta);
210*4bdc9457SAndroid Build Coastguard Worker
211*4bdc9457SAndroid Build Coastguard Worker const __m128 vy0123 = _mm_or_ps(_mm_and_ps(ve0123, vm0123), _mm_andnot_ps(vm0123, vx0123));
212*4bdc9457SAndroid Build Coastguard Worker const __m128 vy4567 = _mm_or_ps(_mm_and_ps(ve4567, vm4567), _mm_andnot_ps(vm4567, vx4567));
213*4bdc9457SAndroid Build Coastguard Worker const __m128 vy89AB = _mm_or_ps(_mm_and_ps(ve89AB, vm89AB), _mm_andnot_ps(vm89AB, vx89AB));
214*4bdc9457SAndroid Build Coastguard Worker const __m128 vyCDEF = _mm_or_ps(_mm_and_ps(veCDEF, vmCDEF), _mm_andnot_ps(vmCDEF, vxCDEF));
215*4bdc9457SAndroid Build Coastguard Worker
216*4bdc9457SAndroid Build Coastguard Worker _mm_storeu_ps(y, vy0123);
217*4bdc9457SAndroid Build Coastguard Worker _mm_storeu_ps(y + 4, vy4567);
218*4bdc9457SAndroid Build Coastguard Worker _mm_storeu_ps(y + 8, vy89AB);
219*4bdc9457SAndroid Build Coastguard Worker _mm_storeu_ps(y + 12, vyCDEF);
220*4bdc9457SAndroid Build Coastguard Worker y += 16;
221*4bdc9457SAndroid Build Coastguard Worker }
222*4bdc9457SAndroid Build Coastguard Worker for (; n >= 4 * sizeof(float); n -= 4 * sizeof(float)) {
223*4bdc9457SAndroid Build Coastguard Worker __m128 vx = _mm_loadu_ps(x);
224*4bdc9457SAndroid Build Coastguard Worker x += 4;
225*4bdc9457SAndroid Build Coastguard Worker
226*4bdc9457SAndroid Build Coastguard Worker const __m128 vz = _mm_max_ps(vsat_cutoff, _mm_mul_ps(vx, vprescale));
227*4bdc9457SAndroid Build Coastguard Worker
228*4bdc9457SAndroid Build Coastguard Worker __m128 vn = _mm_add_ps(_mm_mul_ps(vz, vlog2e), vmagic_bias);
229*4bdc9457SAndroid Build Coastguard Worker
230*4bdc9457SAndroid Build Coastguard Worker const __m128i ven = _mm_slli_epi32(_mm_castps_si128(vn), 19);
231*4bdc9457SAndroid Build Coastguard Worker const __m128i vidx = _mm_slli_epi32(_mm_and_si128(_mm_castps_si128(vn), vindex_mask), 2);
232*4bdc9457SAndroid Build Coastguard Worker #if XNN_ARCH_X86_64
233*4bdc9457SAndroid Build Coastguard Worker const uint64_t vidx_lo = (uint64_t) _mm_cvtsi128_si64(vidx);
234*4bdc9457SAndroid Build Coastguard Worker const uint64_t vidx_hi = (uint64_t) _mm_cvtsi128_si64(_mm_unpackhi_epi64(vidx, vidx));
235*4bdc9457SAndroid Build Coastguard Worker const __m128i vl_ll = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) vidx_lo)));
236*4bdc9457SAndroid Build Coastguard Worker const __m128i vl_hl = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) vidx_hi)));
237*4bdc9457SAndroid Build Coastguard Worker const __m128i vl_lh = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) (vidx_lo >> 32))));
238*4bdc9457SAndroid Build Coastguard Worker const __m128i vl_lo = _mm_unpacklo_epi32(vl_ll, vl_lh);
239*4bdc9457SAndroid Build Coastguard Worker const __m128i vl_hh = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) (vidx_hi >> 32))));
240*4bdc9457SAndroid Build Coastguard Worker const __m128i vl_hi = _mm_unpacklo_epi32(vl_hl, vl_hh);
241*4bdc9457SAndroid Build Coastguard Worker #else // !XNN_ARCH_X86_64
242*4bdc9457SAndroid Build Coastguard Worker const __m128i vl_ll = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) _mm_cvtsi128_si32(vidx))));
243*4bdc9457SAndroid Build Coastguard Worker const __m128i vl_hl = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) _mm_extract_epi16(vidx, 4))));
244*4bdc9457SAndroid Build Coastguard Worker const __m128i vl_lh = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) _mm_extract_epi16(vidx, 2))));
245*4bdc9457SAndroid Build Coastguard Worker const __m128i vl_lo = _mm_unpacklo_epi32(vl_ll, vl_lh);
246*4bdc9457SAndroid Build Coastguard Worker const __m128i vl_hh = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) _mm_extract_epi16(vidx, 6))));
247*4bdc9457SAndroid Build Coastguard Worker const __m128i vl_hi = _mm_unpacklo_epi32(vl_hl, vl_hh);
248*4bdc9457SAndroid Build Coastguard Worker #endif // XNN_ARCH_X86_64
249*4bdc9457SAndroid Build Coastguard Worker const __m128i vl = _mm_unpacklo_epi64(vl_lo, vl_hi);
250*4bdc9457SAndroid Build Coastguard Worker __m128 vs = _mm_castsi128_ps(_mm_add_epi32(vl, ven));
251*4bdc9457SAndroid Build Coastguard Worker vn = _mm_sub_ps(vn, vmagic_bias);
252*4bdc9457SAndroid Build Coastguard Worker
253*4bdc9457SAndroid Build Coastguard Worker __m128 vt = _mm_add_ps(_mm_mul_ps(vn, vminus_ln2_hi), vz);
254*4bdc9457SAndroid Build Coastguard Worker vt = _mm_add_ps(_mm_mul_ps(vn, vminus_ln2_lo), vt);
255*4bdc9457SAndroid Build Coastguard Worker
256*4bdc9457SAndroid Build Coastguard Worker __m128 vp = _mm_add_ps(_mm_mul_ps(vc3, vt), vc2);
257*4bdc9457SAndroid Build Coastguard Worker vp = _mm_mul_ps(vp, vt);
258*4bdc9457SAndroid Build Coastguard Worker
259*4bdc9457SAndroid Build Coastguard Worker vt = _mm_mul_ps(vt, vs);
260*4bdc9457SAndroid Build Coastguard Worker vs = _mm_sub_ps(vs, vone);
261*4bdc9457SAndroid Build Coastguard Worker vp = _mm_add_ps(_mm_mul_ps(vp, vt), vt);
262*4bdc9457SAndroid Build Coastguard Worker const __m128 ve = _mm_mul_ps(_mm_add_ps(vp, vs), valpha);
263*4bdc9457SAndroid Build Coastguard Worker
264*4bdc9457SAndroid Build Coastguard Worker const __m128 vm = _mm_castsi128_ps(_mm_cmpgt_epi32(_mm_setzero_si128(), _mm_castps_si128(vx)));
265*4bdc9457SAndroid Build Coastguard Worker vx = _mm_mul_ps(vx, vbeta);
266*4bdc9457SAndroid Build Coastguard Worker const __m128 vy = _mm_or_ps(_mm_and_ps(ve, vm), _mm_andnot_ps(vm, vx));
267*4bdc9457SAndroid Build Coastguard Worker
268*4bdc9457SAndroid Build Coastguard Worker _mm_storeu_ps(y, vy);
269*4bdc9457SAndroid Build Coastguard Worker y += 4;
270*4bdc9457SAndroid Build Coastguard Worker }
271*4bdc9457SAndroid Build Coastguard Worker if XNN_UNLIKELY(n != 0) {
272*4bdc9457SAndroid Build Coastguard Worker __m128 vx = _mm_loadu_ps(x);
273*4bdc9457SAndroid Build Coastguard Worker
274*4bdc9457SAndroid Build Coastguard Worker const __m128 vz = _mm_max_ps(vsat_cutoff, _mm_mul_ps(vx, vprescale));
275*4bdc9457SAndroid Build Coastguard Worker
276*4bdc9457SAndroid Build Coastguard Worker __m128 vn = _mm_add_ps(_mm_mul_ps(vz, vlog2e), vmagic_bias);
277*4bdc9457SAndroid Build Coastguard Worker
278*4bdc9457SAndroid Build Coastguard Worker const __m128i ven = _mm_slli_epi32(_mm_castps_si128(vn), 19);
279*4bdc9457SAndroid Build Coastguard Worker const __m128i vidx = _mm_slli_epi32(_mm_and_si128(_mm_castps_si128(vn), vindex_mask), 2);
280*4bdc9457SAndroid Build Coastguard Worker #if XNN_ARCH_X86_64
281*4bdc9457SAndroid Build Coastguard Worker const uint64_t vidx_lo = (uint64_t) _mm_cvtsi128_si64(vidx);
282*4bdc9457SAndroid Build Coastguard Worker const uint64_t vidx_hi = (uint64_t) _mm_cvtsi128_si64(_mm_unpackhi_epi64(vidx, vidx));
283*4bdc9457SAndroid Build Coastguard Worker const __m128i vl_ll = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) vidx_lo)));
284*4bdc9457SAndroid Build Coastguard Worker const __m128i vl_hl = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) vidx_hi)));
285*4bdc9457SAndroid Build Coastguard Worker const __m128i vl_lh = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) (vidx_lo >> 32))));
286*4bdc9457SAndroid Build Coastguard Worker const __m128i vl_lo = _mm_unpacklo_epi32(vl_ll, vl_lh);
287*4bdc9457SAndroid Build Coastguard Worker const __m128i vl_hh = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) (vidx_hi >> 32))));
288*4bdc9457SAndroid Build Coastguard Worker const __m128i vl_hi = _mm_unpacklo_epi32(vl_hl, vl_hh);
289*4bdc9457SAndroid Build Coastguard Worker #else // !XNN_ARCH_X86_64
290*4bdc9457SAndroid Build Coastguard Worker const __m128i vl_ll = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) _mm_cvtsi128_si32(vidx))));
291*4bdc9457SAndroid Build Coastguard Worker const __m128i vl_hl = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) _mm_extract_epi16(vidx, 4))));
292*4bdc9457SAndroid Build Coastguard Worker const __m128i vl_lh = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) _mm_extract_epi16(vidx, 2))));
293*4bdc9457SAndroid Build Coastguard Worker const __m128i vl_lo = _mm_unpacklo_epi32(vl_ll, vl_lh);
294*4bdc9457SAndroid Build Coastguard Worker const __m128i vl_hh = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) _mm_extract_epi16(vidx, 6))));
295*4bdc9457SAndroid Build Coastguard Worker const __m128i vl_hi = _mm_unpacklo_epi32(vl_hl, vl_hh);
296*4bdc9457SAndroid Build Coastguard Worker #endif // XNN_ARCH_X86_64
297*4bdc9457SAndroid Build Coastguard Worker const __m128i vl = _mm_unpacklo_epi64(vl_lo, vl_hi);
298*4bdc9457SAndroid Build Coastguard Worker __m128 vs = _mm_castsi128_ps(_mm_add_epi32(vl, ven));
299*4bdc9457SAndroid Build Coastguard Worker vn = _mm_sub_ps(vn, vmagic_bias);
300*4bdc9457SAndroid Build Coastguard Worker
301*4bdc9457SAndroid Build Coastguard Worker __m128 vt = _mm_add_ps(_mm_mul_ps(vn, vminus_ln2_hi), vz);
302*4bdc9457SAndroid Build Coastguard Worker vt = _mm_add_ps(_mm_mul_ps(vn, vminus_ln2_lo), vt);
303*4bdc9457SAndroid Build Coastguard Worker
304*4bdc9457SAndroid Build Coastguard Worker __m128 vp = _mm_add_ps(_mm_mul_ps(vc3, vt), vc2);
305*4bdc9457SAndroid Build Coastguard Worker vp = _mm_mul_ps(vp, vt);
306*4bdc9457SAndroid Build Coastguard Worker
307*4bdc9457SAndroid Build Coastguard Worker vt = _mm_mul_ps(vt, vs);
308*4bdc9457SAndroid Build Coastguard Worker vs = _mm_sub_ps(vs, vone);
309*4bdc9457SAndroid Build Coastguard Worker vp = _mm_add_ps(_mm_mul_ps(vp, vt), vt);
310*4bdc9457SAndroid Build Coastguard Worker const __m128 ve = _mm_mul_ps(_mm_add_ps(vp, vs), valpha);
311*4bdc9457SAndroid Build Coastguard Worker
312*4bdc9457SAndroid Build Coastguard Worker const __m128 vm = _mm_castsi128_ps(_mm_cmpgt_epi32(_mm_setzero_si128(), _mm_castps_si128(vx)));
313*4bdc9457SAndroid Build Coastguard Worker vx = _mm_mul_ps(vx, vbeta);
314*4bdc9457SAndroid Build Coastguard Worker __m128 vy = _mm_or_ps(_mm_and_ps(ve, vm), _mm_andnot_ps(vm, vx));
315*4bdc9457SAndroid Build Coastguard Worker
316*4bdc9457SAndroid Build Coastguard Worker if (n & (2 * sizeof(float))) {
317*4bdc9457SAndroid Build Coastguard Worker _mm_storel_pi((__m64*) y, vy);
318*4bdc9457SAndroid Build Coastguard Worker vy = _mm_movehl_ps(vy, vy);
319*4bdc9457SAndroid Build Coastguard Worker y += 2;
320*4bdc9457SAndroid Build Coastguard Worker }
321*4bdc9457SAndroid Build Coastguard Worker if (n & (1 * sizeof(float))) {
322*4bdc9457SAndroid Build Coastguard Worker _mm_store_ss(y, vy);
323*4bdc9457SAndroid Build Coastguard Worker }
324*4bdc9457SAndroid Build Coastguard Worker }
325*4bdc9457SAndroid Build Coastguard Worker }
326