• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
1 // Auto-generated file. Do not edit!
2 //   Template: src/bf16-gemm/c8-neon.c.in
3 //   Generator: tools/xngen
4 //
5 // Copyright 2022 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 
11 #include <assert.h>
12 
13 #include <arm_neon.h>
14 
15 #include <xnnpack/gemm.h>
16 
17 
xnn_bf16_gemm_minmax_ukernel_5x4c8__neonfma_shland(size_t mr,size_t nc,size_t kc,const void * restrict a,size_t a_stride,const void * restrict w_ptr,void * restrict c,size_t cm_stride,size_t cn_stride,const union xnn_bf16_minmax_params params[restrict XNN_MIN_ELEMENTS (1)])18 void xnn_bf16_gemm_minmax_ukernel_5x4c8__neonfma_shland(
19     size_t mr,
20     size_t nc,
21     size_t kc,
22     const void* restrict a,
23     size_t a_stride,
24     const void* restrict w_ptr,
25     void* restrict c,
26     size_t cm_stride,
27     size_t cn_stride,
28     const union xnn_bf16_minmax_params params[restrict XNN_MIN_ELEMENTS(1)])
29 {
30   assert(mr != 0);
31   assert(mr <= 5);
32   assert(nc != 0);
33   assert(kc != 0);
34   assert(kc % sizeof(uint16_t) == 0);
35   assert(a != NULL);
36   assert(w_ptr != NULL);
37   assert(c != NULL);
38 
39   const uint16_t* a0 = (const uint16_t*) a;
40   uint16_t* c0 = (uint16_t*) c;
41   const uint16_t* a1 = (const uint16_t*) ((uintptr_t) a0 + a_stride);
42   uint16_t* c1 = (uint16_t*) ((uintptr_t) c0 + cm_stride);
43   if XNN_UNPREDICTABLE(mr < 2) {
44     a1 = a0;
45     c1 = c0;
46   }
47   const uint16_t* a2 = (const uint16_t*) ((uintptr_t) a1 + a_stride);
48   uint16_t* c2 = (uint16_t*) ((uintptr_t) c1 + cm_stride);
49   if XNN_UNPREDICTABLE(mr <= 2) {
50     a2 = a1;
51     c2 = c1;
52   }
53   const uint16_t* a3 = (const uint16_t*) ((uintptr_t) a2 + a_stride);
54   uint16_t* c3 = (uint16_t*) ((uintptr_t) c2 + cm_stride);
55   if XNN_UNPREDICTABLE(mr < 4) {
56     a3 = a2;
57     c3 = c2;
58   }
59   const uint16_t* a4 = (const uint16_t*) ((uintptr_t) a3 + a_stride);
60   uint16_t* c4 = (uint16_t*) ((uintptr_t) c3 + cm_stride);
61   if XNN_UNPREDICTABLE(mr <= 4) {
62     a4 = a3;
63     c4 = c3;
64   }
65 
66   const uint16_t* w = (const uint16_t*) w_ptr;
67   const uint16x8_t vmask = vreinterpretq_u16_u32(vmovq_n_u32(UINT32_C(0xFFFF0000)));
68   do {
69     float32x4_t vacc0x0 = vreinterpretq_f32_u32(vshll_n_u16(vld1_lane_u16(w, vdup_n_u16(0), 0), 16)); w += 1;
70     float32x4_t vacc0x1 = vreinterpretq_f32_u32(vshll_n_u16(vld1_lane_u16(w, vdup_n_u16(0), 0), 16)); w += 1;
71     float32x4_t vacc0x2 = vreinterpretq_f32_u32(vshll_n_u16(vld1_lane_u16(w, vdup_n_u16(0), 0), 16)); w += 1;
72     float32x4_t vacc0x3 = vreinterpretq_f32_u32(vshll_n_u16(vld1_lane_u16(w, vdup_n_u16(0), 0), 16)); w += 1;
73     float32x4_t vacc1x0 = vacc0x0;
74     float32x4_t vacc1x1 = vacc0x1;
75     float32x4_t vacc1x2 = vacc0x2;
76     float32x4_t vacc1x3 = vacc0x3;
77     float32x4_t vacc2x0 = vacc0x0;
78     float32x4_t vacc2x1 = vacc0x1;
79     float32x4_t vacc2x2 = vacc0x2;
80     float32x4_t vacc2x3 = vacc0x3;
81     float32x4_t vacc3x0 = vacc0x0;
82     float32x4_t vacc3x1 = vacc0x1;
83     float32x4_t vacc3x2 = vacc0x2;
84     float32x4_t vacc3x3 = vacc0x3;
85     float32x4_t vacc4x0 = vacc0x0;
86     float32x4_t vacc4x1 = vacc0x1;
87     float32x4_t vacc4x2 = vacc0x2;
88     float32x4_t vacc4x3 = vacc0x3;
89 
90     size_t k = kc;
91     for (; k >= 8 * sizeof(uint16_t); k -= 8 * sizeof(uint16_t)) {
92       const uint16x8_t va0 = vld1q_u16(a0); a0 += 8;
93       const uint16x8_t va1 = vld1q_u16(a1); a1 += 8;
94       const uint16x8_t va2 = vld1q_u16(a2); a2 += 8;
95       const uint16x8_t va3 = vld1q_u16(a3); a3 += 8;
96       const uint16x8_t va4 = vld1q_u16(a4); a4 += 8;
97 
98       const uint16x8_t vb0 = vld1q_u16(w); w += 8;
99       const uint16x8_t vb1 = vld1q_u16(w); w += 8;
100       const uint16x8_t vb2 = vld1q_u16(w); w += 8;
101       const uint16x8_t vb3 = vld1q_u16(w); w += 8;
102 
103       const float32x4_t va0e = vreinterpretq_f32_u32(vshlq_n_u32(vreinterpretq_u32_u16(va0), 16));
104       const float32x4_t va1e = vreinterpretq_f32_u32(vshlq_n_u32(vreinterpretq_u32_u16(va1), 16));
105       const float32x4_t va2e = vreinterpretq_f32_u32(vshlq_n_u32(vreinterpretq_u32_u16(va2), 16));
106       const float32x4_t va3e = vreinterpretq_f32_u32(vshlq_n_u32(vreinterpretq_u32_u16(va3), 16));
107       const float32x4_t va4e = vreinterpretq_f32_u32(vshlq_n_u32(vreinterpretq_u32_u16(va4), 16));
108 
109       const float32x4_t vb0e = vreinterpretq_f32_u32(vshlq_n_u32(vreinterpretq_u32_u16(vb0), 16));
110       const float32x4_t vb1e = vreinterpretq_f32_u32(vshlq_n_u32(vreinterpretq_u32_u16(vb1), 16));
111       const float32x4_t vb2e = vreinterpretq_f32_u32(vshlq_n_u32(vreinterpretq_u32_u16(vb2), 16));
112       const float32x4_t vb3e = vreinterpretq_f32_u32(vshlq_n_u32(vreinterpretq_u32_u16(vb3), 16));
113 
114       vacc0x0 = vfmaq_f32(vacc0x0, va0e, vb0e);
115       vacc1x0 = vfmaq_f32(vacc1x0, va1e, vb0e);
116       vacc2x0 = vfmaq_f32(vacc2x0, va2e, vb0e);
117       vacc3x0 = vfmaq_f32(vacc3x0, va3e, vb0e);
118       vacc4x0 = vfmaq_f32(vacc4x0, va4e, vb0e);
119       vacc0x1 = vfmaq_f32(vacc0x1, va0e, vb1e);
120       vacc1x1 = vfmaq_f32(vacc1x1, va1e, vb1e);
121       vacc2x1 = vfmaq_f32(vacc2x1, va2e, vb1e);
122       vacc3x1 = vfmaq_f32(vacc3x1, va3e, vb1e);
123       vacc4x1 = vfmaq_f32(vacc4x1, va4e, vb1e);
124       vacc0x2 = vfmaq_f32(vacc0x2, va0e, vb2e);
125       vacc1x2 = vfmaq_f32(vacc1x2, va1e, vb2e);
126       vacc2x2 = vfmaq_f32(vacc2x2, va2e, vb2e);
127       vacc3x2 = vfmaq_f32(vacc3x2, va3e, vb2e);
128       vacc4x2 = vfmaq_f32(vacc4x2, va4e, vb2e);
129       vacc0x3 = vfmaq_f32(vacc0x3, va0e, vb3e);
130       vacc1x3 = vfmaq_f32(vacc1x3, va1e, vb3e);
131       vacc2x3 = vfmaq_f32(vacc2x3, va2e, vb3e);
132       vacc3x3 = vfmaq_f32(vacc3x3, va3e, vb3e);
133       vacc4x3 = vfmaq_f32(vacc4x3, va4e, vb3e);
134 
135       const float32x4_t va0o = vreinterpretq_f32_u16(vandq_u16(va0, vmask));
136       const float32x4_t va1o = vreinterpretq_f32_u16(vandq_u16(va1, vmask));
137       const float32x4_t va2o = vreinterpretq_f32_u16(vandq_u16(va2, vmask));
138       const float32x4_t va3o = vreinterpretq_f32_u16(vandq_u16(va3, vmask));
139       const float32x4_t va4o = vreinterpretq_f32_u16(vandq_u16(va4, vmask));
140 
141       const float32x4_t vb0o = vreinterpretq_f32_u16(vandq_u16(vb0, vmask));
142       const float32x4_t vb1o = vreinterpretq_f32_u16(vandq_u16(vb1, vmask));
143       const float32x4_t vb2o = vreinterpretq_f32_u16(vandq_u16(vb2, vmask));
144       const float32x4_t vb3o = vreinterpretq_f32_u16(vandq_u16(vb3, vmask));
145 
146       vacc0x0 = vfmaq_f32(vacc0x0, va0o, vb0o);
147       vacc1x0 = vfmaq_f32(vacc1x0, va1o, vb0o);
148       vacc2x0 = vfmaq_f32(vacc2x0, va2o, vb0o);
149       vacc3x0 = vfmaq_f32(vacc3x0, va3o, vb0o);
150       vacc4x0 = vfmaq_f32(vacc4x0, va4o, vb0o);
151       vacc0x1 = vfmaq_f32(vacc0x1, va0o, vb1o);
152       vacc1x1 = vfmaq_f32(vacc1x1, va1o, vb1o);
153       vacc2x1 = vfmaq_f32(vacc2x1, va2o, vb1o);
154       vacc3x1 = vfmaq_f32(vacc3x1, va3o, vb1o);
155       vacc4x1 = vfmaq_f32(vacc4x1, va4o, vb1o);
156       vacc0x2 = vfmaq_f32(vacc0x2, va0o, vb2o);
157       vacc1x2 = vfmaq_f32(vacc1x2, va1o, vb2o);
158       vacc2x2 = vfmaq_f32(vacc2x2, va2o, vb2o);
159       vacc3x2 = vfmaq_f32(vacc3x2, va3o, vb2o);
160       vacc4x2 = vfmaq_f32(vacc4x2, va4o, vb2o);
161       vacc0x3 = vfmaq_f32(vacc0x3, va0o, vb3o);
162       vacc1x3 = vfmaq_f32(vacc1x3, va1o, vb3o);
163       vacc2x3 = vfmaq_f32(vacc2x3, va2o, vb3o);
164       vacc3x3 = vfmaq_f32(vacc3x3, va3o, vb3o);
165       vacc4x3 = vfmaq_f32(vacc4x3, va4o, vb3o);
166     }
167     if XNN_UNLIKELY(k != 0) {
168       const uint16x8_t va0 = vld1q_u16(a0); a0 = (const uint16_t*) ((uintptr_t) a0 + k);
169       const uint16x8_t va1 = vld1q_u16(a1); a1 = (const uint16_t*) ((uintptr_t) a1 + k);
170       const uint16x8_t va2 = vld1q_u16(a2); a2 = (const uint16_t*) ((uintptr_t) a2 + k);
171       const uint16x8_t va3 = vld1q_u16(a3); a3 = (const uint16_t*) ((uintptr_t) a3 + k);
172       const uint16x8_t va4 = vld1q_u16(a4); a4 = (const uint16_t*) ((uintptr_t) a4 + k);
173 
174       const uint16x8_t vb0 = vld1q_u16(w); w += 8;
175       const uint16x8_t vb1 = vld1q_u16(w); w += 8;
176       const uint16x8_t vb2 = vld1q_u16(w); w += 8;
177       const uint16x8_t vb3 = vld1q_u16(w); w += 8;
178 
179       const uint16x8_t vm0 = vceqq_u16(vb0, vmovq_n_u16(0));
180       const uint16x8_t vm1 = vceqq_u16(vb1, vmovq_n_u16(0));
181       const uint16x8_t vm2 = vceqq_u16(vb2, vmovq_n_u16(0));
182       const uint16x8_t vm3 = vceqq_u16(vb3, vmovq_n_u16(0));
183 
184       const float32x4_t vb0e = vreinterpretq_f32_u32(vshlq_n_u32(vreinterpretq_u32_u16(vb0), 16));
185       const float32x4_t vb1e = vreinterpretq_f32_u32(vshlq_n_u32(vreinterpretq_u32_u16(vb1), 16));
186       const float32x4_t vb2e = vreinterpretq_f32_u32(vshlq_n_u32(vreinterpretq_u32_u16(vb2), 16));
187       const float32x4_t vb3e = vreinterpretq_f32_u32(vshlq_n_u32(vreinterpretq_u32_u16(vb3), 16));
188 
189       const uint16x8_t va0x0 = vbicq_u16(va0, vm0);
190       const uint16x8_t va1x0 = vbicq_u16(va1, vm0);
191       const uint16x8_t va2x0 = vbicq_u16(va2, vm0);
192       const uint16x8_t va3x0 = vbicq_u16(va3, vm0);
193       const uint16x8_t va4x0 = vbicq_u16(va4, vm0);
194       const uint16x8_t va0x1 = vbicq_u16(va0, vm1);
195       const uint16x8_t va1x1 = vbicq_u16(va1, vm1);
196       const uint16x8_t va2x1 = vbicq_u16(va2, vm1);
197       const uint16x8_t va3x1 = vbicq_u16(va3, vm1);
198       const uint16x8_t va4x1 = vbicq_u16(va4, vm1);
199       const uint16x8_t va0x2 = vbicq_u16(va0, vm2);
200       const uint16x8_t va1x2 = vbicq_u16(va1, vm2);
201       const uint16x8_t va2x2 = vbicq_u16(va2, vm2);
202       const uint16x8_t va3x2 = vbicq_u16(va3, vm2);
203       const uint16x8_t va4x2 = vbicq_u16(va4, vm2);
204       const uint16x8_t va0x3 = vbicq_u16(va0, vm3);
205       const uint16x8_t va1x3 = vbicq_u16(va1, vm3);
206       const uint16x8_t va2x3 = vbicq_u16(va2, vm3);
207       const uint16x8_t va3x3 = vbicq_u16(va3, vm3);
208       const uint16x8_t va4x3 = vbicq_u16(va4, vm3);
209 
210       const float32x4_t va0x0e = vreinterpretq_f32_u32(vshlq_n_u32(vreinterpretq_u32_u16(va0x0), 16));
211       const float32x4_t va1x0e = vreinterpretq_f32_u32(vshlq_n_u32(vreinterpretq_u32_u16(va1x0), 16));
212       const float32x4_t va2x0e = vreinterpretq_f32_u32(vshlq_n_u32(vreinterpretq_u32_u16(va2x0), 16));
213       const float32x4_t va3x0e = vreinterpretq_f32_u32(vshlq_n_u32(vreinterpretq_u32_u16(va3x0), 16));
214       const float32x4_t va4x0e = vreinterpretq_f32_u32(vshlq_n_u32(vreinterpretq_u32_u16(va4x0), 16));
215       const float32x4_t va0x1e = vreinterpretq_f32_u32(vshlq_n_u32(vreinterpretq_u32_u16(va0x1), 16));
216       const float32x4_t va1x1e = vreinterpretq_f32_u32(vshlq_n_u32(vreinterpretq_u32_u16(va1x1), 16));
217       const float32x4_t va2x1e = vreinterpretq_f32_u32(vshlq_n_u32(vreinterpretq_u32_u16(va2x1), 16));
218       const float32x4_t va3x1e = vreinterpretq_f32_u32(vshlq_n_u32(vreinterpretq_u32_u16(va3x1), 16));
219       const float32x4_t va4x1e = vreinterpretq_f32_u32(vshlq_n_u32(vreinterpretq_u32_u16(va4x1), 16));
220       const float32x4_t va0x2e = vreinterpretq_f32_u32(vshlq_n_u32(vreinterpretq_u32_u16(va0x2), 16));
221       const float32x4_t va1x2e = vreinterpretq_f32_u32(vshlq_n_u32(vreinterpretq_u32_u16(va1x2), 16));
222       const float32x4_t va2x2e = vreinterpretq_f32_u32(vshlq_n_u32(vreinterpretq_u32_u16(va2x2), 16));
223       const float32x4_t va3x2e = vreinterpretq_f32_u32(vshlq_n_u32(vreinterpretq_u32_u16(va3x2), 16));
224       const float32x4_t va4x2e = vreinterpretq_f32_u32(vshlq_n_u32(vreinterpretq_u32_u16(va4x2), 16));
225       const float32x4_t va0x3e = vreinterpretq_f32_u32(vshlq_n_u32(vreinterpretq_u32_u16(va0x3), 16));
226       const float32x4_t va1x3e = vreinterpretq_f32_u32(vshlq_n_u32(vreinterpretq_u32_u16(va1x3), 16));
227       const float32x4_t va2x3e = vreinterpretq_f32_u32(vshlq_n_u32(vreinterpretq_u32_u16(va2x3), 16));
228       const float32x4_t va3x3e = vreinterpretq_f32_u32(vshlq_n_u32(vreinterpretq_u32_u16(va3x3), 16));
229       const float32x4_t va4x3e = vreinterpretq_f32_u32(vshlq_n_u32(vreinterpretq_u32_u16(va4x3), 16));
230 
231       vacc0x0 = vfmaq_f32(vacc0x0, va0x0e, vb0e);
232       vacc1x0 = vfmaq_f32(vacc1x0, va1x0e, vb0e);
233       vacc2x0 = vfmaq_f32(vacc2x0, va2x0e, vb0e);
234       vacc3x0 = vfmaq_f32(vacc3x0, va3x0e, vb0e);
235       vacc4x0 = vfmaq_f32(vacc4x0, va4x0e, vb0e);
236       vacc0x1 = vfmaq_f32(vacc0x1, va0x1e, vb1e);
237       vacc1x1 = vfmaq_f32(vacc1x1, va1x1e, vb1e);
238       vacc2x1 = vfmaq_f32(vacc2x1, va2x1e, vb1e);
239       vacc3x1 = vfmaq_f32(vacc3x1, va3x1e, vb1e);
240       vacc4x1 = vfmaq_f32(vacc4x1, va4x1e, vb1e);
241       vacc0x2 = vfmaq_f32(vacc0x2, va0x2e, vb2e);
242       vacc1x2 = vfmaq_f32(vacc1x2, va1x2e, vb2e);
243       vacc2x2 = vfmaq_f32(vacc2x2, va2x2e, vb2e);
244       vacc3x2 = vfmaq_f32(vacc3x2, va3x2e, vb2e);
245       vacc4x2 = vfmaq_f32(vacc4x2, va4x2e, vb2e);
246       vacc0x3 = vfmaq_f32(vacc0x3, va0x3e, vb3e);
247       vacc1x3 = vfmaq_f32(vacc1x3, va1x3e, vb3e);
248       vacc2x3 = vfmaq_f32(vacc2x3, va2x3e, vb3e);
249       vacc3x3 = vfmaq_f32(vacc3x3, va3x3e, vb3e);
250       vacc4x3 = vfmaq_f32(vacc4x3, va4x3e, vb3e);
251 
252       const float32x4_t vb0o = vreinterpretq_f32_u16(vandq_u16(vb0, vmask));
253       const float32x4_t vb1o = vreinterpretq_f32_u16(vandq_u16(vb1, vmask));
254       const float32x4_t vb2o = vreinterpretq_f32_u16(vandq_u16(vb2, vmask));
255       const float32x4_t vb3o = vreinterpretq_f32_u16(vandq_u16(vb3, vmask));
256 
257       const float32x4_t va0x0o = vreinterpretq_f32_u16(vandq_u16(va0x0, vmask));
258       const float32x4_t va1x0o = vreinterpretq_f32_u16(vandq_u16(va1x0, vmask));
259       const float32x4_t va2x0o = vreinterpretq_f32_u16(vandq_u16(va2x0, vmask));
260       const float32x4_t va3x0o = vreinterpretq_f32_u16(vandq_u16(va3x0, vmask));
261       const float32x4_t va4x0o = vreinterpretq_f32_u16(vandq_u16(va4x0, vmask));
262       const float32x4_t va0x1o = vreinterpretq_f32_u16(vandq_u16(va0x1, vmask));
263       const float32x4_t va1x1o = vreinterpretq_f32_u16(vandq_u16(va1x1, vmask));
264       const float32x4_t va2x1o = vreinterpretq_f32_u16(vandq_u16(va2x1, vmask));
265       const float32x4_t va3x1o = vreinterpretq_f32_u16(vandq_u16(va3x1, vmask));
266       const float32x4_t va4x1o = vreinterpretq_f32_u16(vandq_u16(va4x1, vmask));
267       const float32x4_t va0x2o = vreinterpretq_f32_u16(vandq_u16(va0x2, vmask));
268       const float32x4_t va1x2o = vreinterpretq_f32_u16(vandq_u16(va1x2, vmask));
269       const float32x4_t va2x2o = vreinterpretq_f32_u16(vandq_u16(va2x2, vmask));
270       const float32x4_t va3x2o = vreinterpretq_f32_u16(vandq_u16(va3x2, vmask));
271       const float32x4_t va4x2o = vreinterpretq_f32_u16(vandq_u16(va4x2, vmask));
272       const float32x4_t va0x3o = vreinterpretq_f32_u16(vandq_u16(va0x3, vmask));
273       const float32x4_t va1x3o = vreinterpretq_f32_u16(vandq_u16(va1x3, vmask));
274       const float32x4_t va2x3o = vreinterpretq_f32_u16(vandq_u16(va2x3, vmask));
275       const float32x4_t va3x3o = vreinterpretq_f32_u16(vandq_u16(va3x3, vmask));
276       const float32x4_t va4x3o = vreinterpretq_f32_u16(vandq_u16(va4x3, vmask));
277 
278       vacc0x0 = vfmaq_f32(vacc0x0, va0x0o, vb0o);
279       vacc1x0 = vfmaq_f32(vacc1x0, va1x0o, vb0o);
280       vacc2x0 = vfmaq_f32(vacc2x0, va2x0o, vb0o);
281       vacc3x0 = vfmaq_f32(vacc3x0, va3x0o, vb0o);
282       vacc4x0 = vfmaq_f32(vacc4x0, va4x0o, vb0o);
283       vacc0x1 = vfmaq_f32(vacc0x1, va0x1o, vb1o);
284       vacc1x1 = vfmaq_f32(vacc1x1, va1x1o, vb1o);
285       vacc2x1 = vfmaq_f32(vacc2x1, va2x1o, vb1o);
286       vacc3x1 = vfmaq_f32(vacc3x1, va3x1o, vb1o);
287       vacc4x1 = vfmaq_f32(vacc4x1, va4x1o, vb1o);
288       vacc0x2 = vfmaq_f32(vacc0x2, va0x2o, vb2o);
289       vacc1x2 = vfmaq_f32(vacc1x2, va1x2o, vb2o);
290       vacc2x2 = vfmaq_f32(vacc2x2, va2x2o, vb2o);
291       vacc3x2 = vfmaq_f32(vacc3x2, va3x2o, vb2o);
292       vacc4x2 = vfmaq_f32(vacc4x2, va4x2o, vb2o);
293       vacc0x3 = vfmaq_f32(vacc0x3, va0x3o, vb3o);
294       vacc1x3 = vfmaq_f32(vacc1x3, va1x3o, vb3o);
295       vacc2x3 = vfmaq_f32(vacc2x3, va2x3o, vb3o);
296       vacc3x3 = vfmaq_f32(vacc3x3, va3x3o, vb3o);
297       vacc4x3 = vfmaq_f32(vacc4x3, va4x3o, vb3o);
298     }
299 
300 #if XNN_ARCH_ARM64
301     const float32x4_t vacc0x01 = vpaddq_f32(vacc0x0, vacc0x1);
302     const float32x4_t vacc1x01 = vpaddq_f32(vacc1x0, vacc1x1);
303     const float32x4_t vacc2x01 = vpaddq_f32(vacc2x0, vacc2x1);
304     const float32x4_t vacc3x01 = vpaddq_f32(vacc3x0, vacc3x1);
305     const float32x4_t vacc4x01 = vpaddq_f32(vacc4x0, vacc4x1);
306     const float32x4_t vacc0x23 = vpaddq_f32(vacc0x2, vacc0x3);
307     const float32x4_t vacc1x23 = vpaddq_f32(vacc1x2, vacc1x3);
308     const float32x4_t vacc2x23 = vpaddq_f32(vacc2x2, vacc2x3);
309     const float32x4_t vacc3x23 = vpaddq_f32(vacc3x2, vacc3x3);
310     const float32x4_t vacc4x23 = vpaddq_f32(vacc4x2, vacc4x3);
311 
312     float32x4_t vacc0x0123 = vpaddq_f32(vacc0x01, vacc0x23);
313     float32x4_t vacc1x0123 = vpaddq_f32(vacc1x01, vacc1x23);
314     float32x4_t vacc2x0123 = vpaddq_f32(vacc2x01, vacc2x23);
315     float32x4_t vacc3x0123 = vpaddq_f32(vacc3x01, vacc3x23);
316     float32x4_t vacc4x0123 = vpaddq_f32(vacc4x01, vacc4x23);
317 #else
318     const float32x2_t vsum0x0 = vadd_f32(vget_low_f32(vacc0x0), vget_high_f32(vacc0x0));
319     const float32x2_t vsum1x0 = vadd_f32(vget_low_f32(vacc1x0), vget_high_f32(vacc1x0));
320     const float32x2_t vsum2x0 = vadd_f32(vget_low_f32(vacc2x0), vget_high_f32(vacc2x0));
321     const float32x2_t vsum3x0 = vadd_f32(vget_low_f32(vacc3x0), vget_high_f32(vacc3x0));
322     const float32x2_t vsum4x0 = vadd_f32(vget_low_f32(vacc4x0), vget_high_f32(vacc4x0));
323     const float32x2_t vsum0x1 = vadd_f32(vget_low_f32(vacc0x1), vget_high_f32(vacc0x1));
324     const float32x2_t vsum1x1 = vadd_f32(vget_low_f32(vacc1x1), vget_high_f32(vacc1x1));
325     const float32x2_t vsum2x1 = vadd_f32(vget_low_f32(vacc2x1), vget_high_f32(vacc2x1));
326     const float32x2_t vsum3x1 = vadd_f32(vget_low_f32(vacc3x1), vget_high_f32(vacc3x1));
327     const float32x2_t vsum4x1 = vadd_f32(vget_low_f32(vacc4x1), vget_high_f32(vacc4x1));
328     const float32x2_t vsum0x2 = vadd_f32(vget_low_f32(vacc0x2), vget_high_f32(vacc0x2));
329     const float32x2_t vsum1x2 = vadd_f32(vget_low_f32(vacc1x2), vget_high_f32(vacc1x2));
330     const float32x2_t vsum2x2 = vadd_f32(vget_low_f32(vacc2x2), vget_high_f32(vacc2x2));
331     const float32x2_t vsum3x2 = vadd_f32(vget_low_f32(vacc3x2), vget_high_f32(vacc3x2));
332     const float32x2_t vsum4x2 = vadd_f32(vget_low_f32(vacc4x2), vget_high_f32(vacc4x2));
333     const float32x2_t vsum0x3 = vadd_f32(vget_low_f32(vacc0x3), vget_high_f32(vacc0x3));
334     const float32x2_t vsum1x3 = vadd_f32(vget_low_f32(vacc1x3), vget_high_f32(vacc1x3));
335     const float32x2_t vsum2x3 = vadd_f32(vget_low_f32(vacc2x3), vget_high_f32(vacc2x3));
336     const float32x2_t vsum3x3 = vadd_f32(vget_low_f32(vacc3x3), vget_high_f32(vacc3x3));
337     const float32x2_t vsum4x3 = vadd_f32(vget_low_f32(vacc4x3), vget_high_f32(vacc4x3));
338 
339     float32x4_t vacc0x0123 = vcombine_f32(vpadd_f32(vsum0x0, vsum0x1), vpadd_f32(vsum0x2, vsum0x3));
340     float32x4_t vacc1x0123 = vcombine_f32(vpadd_f32(vsum1x0, vsum1x1), vpadd_f32(vsum1x2, vsum1x3));
341     float32x4_t vacc2x0123 = vcombine_f32(vpadd_f32(vsum2x0, vsum2x1), vpadd_f32(vsum2x2, vsum2x3));
342     float32x4_t vacc3x0123 = vcombine_f32(vpadd_f32(vsum3x0, vsum3x1), vpadd_f32(vsum3x2, vsum3x3));
343     float32x4_t vacc4x0123 = vcombine_f32(vpadd_f32(vsum4x0, vsum4x1), vpadd_f32(vsum4x2, vsum4x3));
344 #endif
345 
346     const float32x4_t vmax = vld1q_dup_f32(&params->scalar.max);
347     vacc0x0123 = vminq_f32(vacc0x0123, vmax);
348     vacc1x0123 = vminq_f32(vacc1x0123, vmax);
349     vacc2x0123 = vminq_f32(vacc2x0123, vmax);
350     vacc3x0123 = vminq_f32(vacc3x0123, vmax);
351     vacc4x0123 = vminq_f32(vacc4x0123, vmax);
352 
353     const float32x4_t vmin = vld1q_dup_f32(&params->scalar.min);
354     vacc0x0123 = vmaxq_f32(vacc0x0123, vmin);
355     vacc1x0123 = vmaxq_f32(vacc1x0123, vmin);
356     vacc2x0123 = vmaxq_f32(vacc2x0123, vmin);
357     vacc3x0123 = vmaxq_f32(vacc3x0123, vmin);
358     vacc4x0123 = vmaxq_f32(vacc4x0123, vmin);
359 
360     uint16x4_t vout0x0123 = vshrn_n_u32(vreinterpretq_u32_f32(vacc0x0123), 16);
361     uint16x4_t vout1x0123 = vshrn_n_u32(vreinterpretq_u32_f32(vacc1x0123), 16);
362     uint16x4_t vout2x0123 = vshrn_n_u32(vreinterpretq_u32_f32(vacc2x0123), 16);
363     uint16x4_t vout3x0123 = vshrn_n_u32(vreinterpretq_u32_f32(vacc3x0123), 16);
364     uint16x4_t vout4x0123 = vshrn_n_u32(vreinterpretq_u32_f32(vacc4x0123), 16);
365 
366     if XNN_LIKELY(nc >= 4) {
367       vst1_u16(c0, vout0x0123);
368       c0 = (uint16_t*) ((uintptr_t) c0 + cn_stride);
369       vst1_u16(c1, vout1x0123);
370       c1 = (uint16_t*) ((uintptr_t) c1 + cn_stride);
371       vst1_u16(c2, vout2x0123);
372       c2 = (uint16_t*) ((uintptr_t) c2 + cn_stride);
373       vst1_u16(c3, vout3x0123);
374       c3 = (uint16_t*) ((uintptr_t) c3 + cn_stride);
375       vst1_u16(c4, vout4x0123);
376       c4 = (uint16_t*) ((uintptr_t) c4 + cn_stride);
377 
378       a0 = (const uint16_t*) ((uintptr_t) a0 - kc);
379       a1 = (const uint16_t*) ((uintptr_t) a1 - kc);
380       a2 = (const uint16_t*) ((uintptr_t) a2 - kc);
381       a3 = (const uint16_t*) ((uintptr_t) a3 - kc);
382       a4 = (const uint16_t*) ((uintptr_t) a4 - kc);
383 
384       nc -= 4;
385     } else {
386       if (nc & 2) {
387         vst1_lane_u32((void*) c0, vreinterpret_u32_u16(vout0x0123), 0); c0 += 2;
388         vst1_lane_u32((void*) c1, vreinterpret_u32_u16(vout1x0123), 0); c1 += 2;
389         vst1_lane_u32((void*) c2, vreinterpret_u32_u16(vout2x0123), 0); c2 += 2;
390         vst1_lane_u32((void*) c3, vreinterpret_u32_u16(vout3x0123), 0); c3 += 2;
391         vst1_lane_u32((void*) c4, vreinterpret_u32_u16(vout4x0123), 0); c4 += 2;
392 
393         vout0x0123 = vext_u16(vout0x0123, vout0x0123, 2);
394         vout1x0123 = vext_u16(vout1x0123, vout1x0123, 2);
395         vout2x0123 = vext_u16(vout2x0123, vout2x0123, 2);
396         vout3x0123 = vext_u16(vout3x0123, vout3x0123, 2);
397         vout4x0123 = vext_u16(vout4x0123, vout4x0123, 2);
398       }
399       if (nc & 1) {
400         vst1_lane_u16(c0, vout0x0123, 0);
401         vst1_lane_u16(c1, vout1x0123, 0);
402         vst1_lane_u16(c2, vout2x0123, 0);
403         vst1_lane_u16(c3, vout3x0123, 0);
404         vst1_lane_u16(c4, vout4x0123, 0);
405       }
406 
407       nc = 0;
408     }
409   } while (nc != 0);
410 }
411