• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
1 // Auto-generated file. Do not edit!
2 //   Template: src/f32-raddstoreexpminusmax/neon-lut64-p2.c.in
3 //   Generator: tools/xngen
4 //
5 // Copyright 2020 Google LLC
6 //
7 // This source code is licensed under the BSD-style license found in the
8 // LICENSE file in the root directory of this source tree.
9 
10 #include <assert.h>
11 
12 #include <arm_neon.h>
13 
14 #include <xnnpack/common.h>
15 #include <xnnpack/raddstoreexpminusmax.h>
16 
17 
18 extern XNN_INTERNAL const float xnn_table_exp2_k_over_64[64];
19 
xnn_f32_raddstoreexpminusmax_ukernel__neonfma_rr1_lut64_p2_x8_acc2(size_t elements,const float * input,const float * max,float * output,float * sum,const union xnn_f32_expminus_params params[restrict XNN_MIN_ELEMENTS (1)])20 void xnn_f32_raddstoreexpminusmax_ukernel__neonfma_rr1_lut64_p2_x8_acc2(
21     size_t elements,
22     const float* input,
23     const float* max,
24     float* output,
25     float* sum,
26     const union xnn_f32_expminus_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
27 {
28   assert(elements % sizeof(float) == 0);
29 
30   const float32x4_t vi_max = vld1q_dup_f32(max);
31   const float32x4_t vlog2e = vld1q_dup_f32(&params->neonfma_rr1_lut64_p2.log2e);
32   const float32x4_t vmagic_bias = vld1q_dup_f32(&params->neonfma_rr1_lut64_p2.magic_bias);
33   const int32x4_t vindex_mask = vmovq_n_s32(INT32_C(0x3F));
34   const float32x4_t vminus_ln2 = vld1q_dup_f32(&params->neonfma_rr1_lut64_p2.minus_ln2);
35   const float32x4_t vc2 = vld1q_dup_f32(&params->neonfma_rr1_lut64_p2.c2);
36   const float32x4_t vdenorm_cutoff = vld1q_dup_f32(&params->neonfma_rr1_lut64_p2.denorm_cutoff);
37 
38   float32x4_t vacc0 = vmovq_n_f32(0.0f);
39   float32x4_t vacc1 = vmovq_n_f32(0.0f);
40   for (; elements >= 8 * sizeof(float); elements -= 8 * sizeof(float)) {
41     const float32x4_t vi0123 = vld1q_f32(input); input += 4;
42     const float32x4_t vi4567 = vld1q_f32(input); input += 4;
43 
44     const float32x4_t vx0123 = vsubq_f32(vi0123, vi_max);
45     const float32x4_t vx4567 = vsubq_f32(vi4567, vi_max);
46 
47     float32x4_t vn0123 = vfmaq_f32(vmagic_bias, vx0123, vlog2e);
48     float32x4_t vn4567 = vfmaq_f32(vmagic_bias, vx4567, vlog2e);
49 
50     const int32x4_t ve0123 = vshlq_n_s32(vbicq_s32(vreinterpretq_s32_f32(vn0123), vmovq_n_s32(INT32_C(0x3F))), 17);
51     const int32x4_t ve4567 = vshlq_n_s32(vbicq_s32(vreinterpretq_s32_f32(vn4567), vmovq_n_s32(INT32_C(0x3F))), 17);
52 
53     const uint64x2_t vidx0123 = vreinterpretq_u64_s32(vandq_s32(vreinterpretq_s32_f32(vn0123), vindex_mask));
54     const uint64_t vidx01 = vgetq_lane_u64(vidx0123, 0);
55     const uint64_t vidx23 = vgetq_lane_u64(vidx0123, 1);
56     const uint64x2_t vidx4567 = vreinterpretq_u64_s32(vandq_s32(vreinterpretq_s32_f32(vn4567), vindex_mask));
57     const uint64_t vidx45 = vgetq_lane_u64(vidx4567, 0);
58     const uint64_t vidx67 = vgetq_lane_u64(vidx4567, 1);
59 
60     float32x2_t vl01 = vld1_dup_f32(&xnn_table_exp2_k_over_64[(uint32_t) vidx01]);
61     float32x2_t vl23 = vld1_dup_f32(&xnn_table_exp2_k_over_64[(uint32_t) vidx23]);
62     float32x2_t vl45 = vld1_dup_f32(&xnn_table_exp2_k_over_64[(uint32_t) vidx45]);
63     float32x2_t vl67 = vld1_dup_f32(&xnn_table_exp2_k_over_64[(uint32_t) vidx67]);
64 
65     vl01 = vld1_lane_f32(&xnn_table_exp2_k_over_64[(uint32_t) (vidx01 >> 32)], vl01, 1);
66     vl23 = vld1_lane_f32(&xnn_table_exp2_k_over_64[(uint32_t) (vidx23 >> 32)], vl23, 1);
67     const float32x4_t vl0123 = vcombine_f32(vl01, vl23);
68     vl45 = vld1_lane_f32(&xnn_table_exp2_k_over_64[(uint32_t) (vidx45 >> 32)], vl45, 1);
69     vl67 = vld1_lane_f32(&xnn_table_exp2_k_over_64[(uint32_t) (vidx67 >> 32)], vl67, 1);
70     const float32x4_t vl4567 = vcombine_f32(vl45, vl67);
71 
72     const float32x4_t vs0123 = vreinterpretq_f32_s32(vaddq_s32(vreinterpretq_s32_f32(vl0123), ve0123));
73     const float32x4_t vs4567 = vreinterpretq_f32_s32(vaddq_s32(vreinterpretq_s32_f32(vl4567), ve4567));
74 
75     vn0123 = vsubq_f32(vn0123, vmagic_bias);
76     vn4567 = vsubq_f32(vn4567, vmagic_bias);
77 
78     float32x4_t vt0123 = vfmaq_f32(vx0123, vn0123, vminus_ln2);
79     float32x4_t vt4567 = vfmaq_f32(vx4567, vn4567, vminus_ln2);
80 
81     float32x4_t vp0123 = vmulq_f32(vt0123, vc2);
82     float32x4_t vp4567 = vmulq_f32(vt4567, vc2);
83 
84     vp0123 = vfmaq_f32(vt0123, vt0123, vp0123);
85     vp4567 = vfmaq_f32(vt4567, vt4567, vp4567);
86 
87     float32x4_t vf0123 = vfmaq_f32(vs0123, vs0123, vp0123);
88     float32x4_t vf4567 = vfmaq_f32(vs4567, vs4567, vp4567);
89 
90     vf0123 = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(vf0123), vcltq_f32(vx0123, vdenorm_cutoff)));
91     vf4567 = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(vf4567), vcltq_f32(vx4567, vdenorm_cutoff)));
92 
93     vst1q_f32(output, vf0123); output += 4;
94     vst1q_f32(output, vf4567); output += 4;
95 
96     vacc0 = vaddq_f32(vacc0, vf0123);
97     vacc0 = vaddq_f32(vacc0, vf4567);
98   }
99   vacc0 = vaddq_f32(vacc0, vacc1);
100 
101   float32x4_t vacc = vacc0;
102   for (; elements >= 4 * sizeof(float); elements -= 4 * sizeof(float)) {
103     const float32x4_t vi = vld1q_f32(input); input += 4;
104 
105     const float32x4_t vx = vsubq_f32(vi, vi_max);
106 
107     float32x4_t vn = vfmaq_f32(vmagic_bias, vx, vlog2e);
108 
109     const int32x4_t ve = vshlq_n_s32(vbicq_s32(vreinterpretq_s32_f32(vn), vmovq_n_s32(INT32_C(0x3F))), 17);
110 
111     const uint64x2_t vidx = vreinterpretq_u64_s32(vandq_s32(vreinterpretq_s32_f32(vn), vindex_mask));
112     const uint64_t vidx_lo = vgetq_lane_u64(vidx, 0);
113     const uint64_t vidx_hi = vgetq_lane_u64(vidx, 1);
114     float32x2_t vl_lo = vld1_dup_f32(&xnn_table_exp2_k_over_64[(uint32_t) vidx_lo]);
115     float32x2_t vl_hi = vld1_dup_f32(&xnn_table_exp2_k_over_64[(uint32_t) vidx_hi]);
116     vl_lo = vld1_lane_f32(&xnn_table_exp2_k_over_64[(uint32_t) (vidx_lo >> 32)], vl_lo, 1);
117     vl_hi = vld1_lane_f32(&xnn_table_exp2_k_over_64[(uint32_t) (vidx_hi >> 32)], vl_hi, 1);
118     const float32x4_t vl = vcombine_f32(vl_lo, vl_hi);
119     const float32x4_t vs = vreinterpretq_f32_s32(vaddq_s32(vreinterpretq_s32_f32(vl), ve));
120 
121     vn = vsubq_f32(vn, vmagic_bias);
122 
123     float32x4_t vt = vfmaq_f32(vx, vn, vminus_ln2);
124 
125     float32x4_t vp = vmulq_f32(vt, vc2);
126     vp = vfmaq_f32(vt, vt, vp);
127 
128     float32x4_t vf = vfmaq_f32(vs, vs, vp);
129 
130     vf = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(vf), vcltq_f32(vx, vdenorm_cutoff)));
131 
132     vst1q_f32(output, vf); output += 4;
133 
134     vacc = vaddq_f32(vacc, vf);
135   }
136 #if XNN_ARCH_ARM64
137   float vacc_lo = vaddvq_f32(vacc);
138 #else
139   float32x2_t vacc_lo = vadd_f32(vget_high_f32(vacc), vget_low_f32(vacc));
140 #endif
141   if (elements != 0) {
142     assert(elements >= 1 * sizeof(float));
143     assert(elements <= 3 * sizeof(float));
144     const float32x4_t vi = vld1q_f32(input); input += 4;
145 
146     const float32x4_t vx = vsubq_f32(vi, vi_max);
147 
148     float32x4_t vn = vfmaq_f32(vmagic_bias, vx, vlog2e);
149 
150     const int32x4_t ve = vshlq_n_s32(vbicq_s32(vreinterpretq_s32_f32(vn), vmovq_n_s32(INT32_C(0x3F))), 17);
151 
152     const uint64x2_t vidx = vreinterpretq_u64_s32(vandq_s32(vreinterpretq_s32_f32(vn), vindex_mask));
153     const uint64_t vidx_lo = vgetq_lane_u64(vidx, 0);
154     const uint64_t vidx_hi = vgetq_lane_u64(vidx, 1);
155     float32x2_t vl_lo = vld1_dup_f32(&xnn_table_exp2_k_over_64[(uint32_t) vidx_lo]);
156     float32x2_t vl_hi = vld1_dup_f32(&xnn_table_exp2_k_over_64[(uint32_t) vidx_hi]);
157     vl_lo = vld1_lane_f32(&xnn_table_exp2_k_over_64[(uint32_t) (vidx_lo >> 32)], vl_lo, 1);
158     vl_hi = vld1_lane_f32(&xnn_table_exp2_k_over_64[(uint32_t) (vidx_hi >> 32)], vl_hi, 1);
159     const float32x4_t vl = vcombine_f32(vl_lo, vl_hi);
160     const float32x4_t vs = vreinterpretq_f32_s32(vaddq_s32(vreinterpretq_s32_f32(vl), ve));
161 
162     vn = vsubq_f32(vn, vmagic_bias);
163 
164     float32x4_t vt = vfmaq_f32(vx, vn, vminus_ln2);
165 
166     float32x4_t vp = vmulq_f32(vt, vc2);
167     vp = vfmaq_f32(vt, vt, vp);
168 
169     float32x4_t vf = vfmaq_f32(vs, vs, vp);
170 
171     vf = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(vf), vcltq_f32(vx, vdenorm_cutoff)));
172 
173     float32x2_t vf_lo = vget_low_f32(vf);
174     if (elements & (2 * sizeof(float))) {
175       vst1_f32(output, vf_lo); output += 2;
176 
177       #if XNN_ARCH_ARM64
178         vacc_lo += vaddv_f32(vf_lo);
179       #else
180         vacc_lo = vadd_f32(vacc_lo, vf_lo);
181       #endif
182 
183       vf_lo = vget_high_f32(vf);
184     }
185     if (elements & (1 * sizeof(float))) {
186       vst1_lane_f32(output, vf_lo, 0);
187 
188       #if XNN_ARCH_ARM64
189         vacc_lo += vget_lane_f32(vf_lo, 0);
190       #else
191         vacc_lo = vadd_f32(vacc_lo, vreinterpret_f32_u64(vshl_n_u64(vreinterpret_u64_f32(vf_lo), 32)));
192       #endif
193     }
194   }
195 #if XNN_ARCH_ARM64
196   *sum = vacc_lo;
197 #else
198   vst1_lane_f32(sum, vpadd_f32(vacc_lo, vacc_lo), 0);
199 #endif
200 }
201