xref: /aosp_15_r20/external/XNNPACK/src/bf16-gemm/gen/4x8c2-minmax-neonbf16-bfdot-lane-ld128.c (revision 4bdc94577ba0e567308109d787f7fec7b531ce36)
1 // Auto-generated file. Do not edit!
2 //   Template: src/bf16-gemm/c2-neonbf16-bfdot-lane-ld128.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_4x8c2__neonbf16_bfdot_lane_ld128(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_4x8c2__neonbf16_bfdot_lane_ld128(
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 <= 4);
32   assert(nc != 0);
33   assert(kc != 0);
34   assert(kc % sizeof(bfloat16_t) == 0);
35   assert(a != NULL);
36   assert(w_ptr != NULL);
37   assert(c != NULL);
38 
39   const bfloat16_t* a0 = (const bfloat16_t*) a;
40   bfloat16_t* c0 = (bfloat16_t*) c;
41   const bfloat16_t* a1 = (const bfloat16_t*) ((uintptr_t) a0 + a_stride);
42   bfloat16_t* c1 = (bfloat16_t*) ((uintptr_t) c0 + cm_stride);
43   if XNN_UNPREDICTABLE(mr < 2) {
44     a1 = a0;
45     c1 = c0;
46   }
47   const bfloat16_t* a2 = (const bfloat16_t*) ((uintptr_t) a1 + a_stride);
48   bfloat16_t* c2 = (bfloat16_t*) ((uintptr_t) c1 + cm_stride);
49   if XNN_UNPREDICTABLE(mr <= 2) {
50     a2 = a1;
51     c2 = c1;
52   }
53   const bfloat16_t* a3 = (const bfloat16_t*) ((uintptr_t) a2 + a_stride);
54   bfloat16_t* c3 = (bfloat16_t*) ((uintptr_t) c2 + cm_stride);
55   if XNN_UNPREDICTABLE(mr != 4) {
56     a3 = a2;
57     c3 = c2;
58   }
59 
60   const bfloat16_t* w = (const bfloat16_t*) w_ptr;
61   do {
62     float32x4_t vacc0x0123 = vcvt_f32_bf16(vld1_bf16(w)); w += 4;
63     float32x4_t vacc0x4567 = vcvt_f32_bf16(vld1_bf16(w)); w += 4;
64     float32x4_t vacc1x0123 = vacc0x0123;
65     float32x4_t vacc1x4567 = vacc0x4567;
66     float32x4_t vacc2x0123 = vacc0x0123;
67     float32x4_t vacc2x4567 = vacc0x4567;
68     float32x4_t vacc3x0123 = vacc0x0123;
69     float32x4_t vacc3x4567 = vacc0x4567;
70 
71     size_t k = kc;
72     for (; k >= 8 * sizeof(bfloat16_t); k -= 8 * sizeof(bfloat16_t)) {
73       const bfloat16x8_t va0 = vld1q_bf16(a0); a0 += 8;
74       const bfloat16x8_t va1 = vld1q_bf16(a1); a1 += 8;
75       const bfloat16x8_t va2 = vld1q_bf16(a2); a2 += 8;
76       const bfloat16x8_t va3 = vld1q_bf16(a3); a3 += 8;
77 
78       const bfloat16x8_t vb0123c01 = vld1q_bf16(w); w += 8;
79       const bfloat16x8_t vb4567c01 = vld1q_bf16(w); w += 8;
80 
81       vacc0x0123 = vbfdotq_laneq_f32(vacc0x0123, vb0123c01, va0, 0);
82       vacc1x0123 = vbfdotq_laneq_f32(vacc1x0123, vb0123c01, va1, 0);
83       vacc2x0123 = vbfdotq_laneq_f32(vacc2x0123, vb0123c01, va2, 0);
84       vacc3x0123 = vbfdotq_laneq_f32(vacc3x0123, vb0123c01, va3, 0);
85       vacc0x4567 = vbfdotq_laneq_f32(vacc0x4567, vb4567c01, va0, 0);
86       vacc1x4567 = vbfdotq_laneq_f32(vacc1x4567, vb4567c01, va1, 0);
87       vacc2x4567 = vbfdotq_laneq_f32(vacc2x4567, vb4567c01, va2, 0);
88       vacc3x4567 = vbfdotq_laneq_f32(vacc3x4567, vb4567c01, va3, 0);
89       const bfloat16x8_t vb0123c23 = vld1q_bf16(w); w += 8;
90       const bfloat16x8_t vb4567c23 = vld1q_bf16(w); w += 8;
91 
92       vacc0x0123 = vbfdotq_laneq_f32(vacc0x0123, vb0123c23, va0, 1);
93       vacc1x0123 = vbfdotq_laneq_f32(vacc1x0123, vb0123c23, va1, 1);
94       vacc2x0123 = vbfdotq_laneq_f32(vacc2x0123, vb0123c23, va2, 1);
95       vacc3x0123 = vbfdotq_laneq_f32(vacc3x0123, vb0123c23, va3, 1);
96       vacc0x4567 = vbfdotq_laneq_f32(vacc0x4567, vb4567c23, va0, 1);
97       vacc1x4567 = vbfdotq_laneq_f32(vacc1x4567, vb4567c23, va1, 1);
98       vacc2x4567 = vbfdotq_laneq_f32(vacc2x4567, vb4567c23, va2, 1);
99       vacc3x4567 = vbfdotq_laneq_f32(vacc3x4567, vb4567c23, va3, 1);
100       const bfloat16x8_t vb0123c45 = vld1q_bf16(w); w += 8;
101       const bfloat16x8_t vb4567c45 = vld1q_bf16(w); w += 8;
102 
103       vacc0x0123 = vbfdotq_laneq_f32(vacc0x0123, vb0123c45, va0, 2);
104       vacc1x0123 = vbfdotq_laneq_f32(vacc1x0123, vb0123c45, va1, 2);
105       vacc2x0123 = vbfdotq_laneq_f32(vacc2x0123, vb0123c45, va2, 2);
106       vacc3x0123 = vbfdotq_laneq_f32(vacc3x0123, vb0123c45, va3, 2);
107       vacc0x4567 = vbfdotq_laneq_f32(vacc0x4567, vb4567c45, va0, 2);
108       vacc1x4567 = vbfdotq_laneq_f32(vacc1x4567, vb4567c45, va1, 2);
109       vacc2x4567 = vbfdotq_laneq_f32(vacc2x4567, vb4567c45, va2, 2);
110       vacc3x4567 = vbfdotq_laneq_f32(vacc3x4567, vb4567c45, va3, 2);
111       const bfloat16x8_t vb0123c67 = vld1q_bf16(w); w += 8;
112       const bfloat16x8_t vb4567c67 = vld1q_bf16(w); w += 8;
113 
114       vacc0x0123 = vbfdotq_laneq_f32(vacc0x0123, vb0123c67, va0, 3);
115       vacc1x0123 = vbfdotq_laneq_f32(vacc1x0123, vb0123c67, va1, 3);
116       vacc2x0123 = vbfdotq_laneq_f32(vacc2x0123, vb0123c67, va2, 3);
117       vacc3x0123 = vbfdotq_laneq_f32(vacc3x0123, vb0123c67, va3, 3);
118       vacc0x4567 = vbfdotq_laneq_f32(vacc0x4567, vb4567c67, va0, 3);
119       vacc1x4567 = vbfdotq_laneq_f32(vacc1x4567, vb4567c67, va1, 3);
120       vacc2x4567 = vbfdotq_laneq_f32(vacc2x4567, vb4567c67, va2, 3);
121       vacc3x4567 = vbfdotq_laneq_f32(vacc3x4567, vb4567c67, va3, 3);
122     }
123     if XNN_UNLIKELY(k != 0) {
124       const bfloat16x8_t va0 = vld1q_bf16(a0); a0 = (const bfloat16_t*) ((uintptr_t) a0 + k);
125       const bfloat16x8_t va1 = vld1q_bf16(a1); a1 = (const bfloat16_t*) ((uintptr_t) a1 + k);
126       const bfloat16x8_t va2 = vld1q_bf16(a2); a2 = (const bfloat16_t*) ((uintptr_t) a2 + k);
127       const bfloat16x8_t va3 = vld1q_bf16(a3); a3 = (const bfloat16_t*) ((uintptr_t) a3 + k);
128 
129       const bfloat16x8_t vb0123c01 = vld1q_bf16(w); w += 8;
130       const bfloat16x8_t vb4567c01 = vld1q_bf16(w); w += 8;
131 
132       const uint32x4_t va0c01 = vdupq_lane_u32(vreinterpret_u32_bf16(vget_low_bf16(va0)), 0);
133       const uint32x4_t va1c01 = vdupq_lane_u32(vreinterpret_u32_bf16(vget_low_bf16(va1)), 0);
134       const uint32x4_t va2c01 = vdupq_lane_u32(vreinterpret_u32_bf16(vget_low_bf16(va2)), 0);
135       const uint32x4_t va3c01 = vdupq_lane_u32(vreinterpret_u32_bf16(vget_low_bf16(va3)), 0);
136 
137       const uint32x4_t vm0123c01 = vreinterpretq_u32_u16(vceqq_u16(vreinterpretq_u16_bf16(vb0123c01), vmovq_n_u16(0)));
138       const uint32x4_t vm4567c01 = vreinterpretq_u32_u16(vceqq_u16(vreinterpretq_u16_bf16(vb4567c01), vmovq_n_u16(0)));
139 
140       const uint32x4_t va0x0123c01 = vbicq_u32(va0c01, vm0123c01);
141       vacc0x0123 = vbfdotq_f32(vacc0x0123, vb0123c01, vreinterpretq_bf16_u32(va0x0123c01));
142       const uint32x4_t va1x0123c01 = vbicq_u32(va1c01, vm0123c01);
143       vacc1x0123 = vbfdotq_f32(vacc1x0123, vb0123c01, vreinterpretq_bf16_u32(va1x0123c01));
144       const uint32x4_t va2x0123c01 = vbicq_u32(va2c01, vm0123c01);
145       vacc2x0123 = vbfdotq_f32(vacc2x0123, vb0123c01, vreinterpretq_bf16_u32(va2x0123c01));
146       const uint32x4_t va3x0123c01 = vbicq_u32(va3c01, vm0123c01);
147       vacc3x0123 = vbfdotq_f32(vacc3x0123, vb0123c01, vreinterpretq_bf16_u32(va3x0123c01));
148       const uint32x4_t va0x4567c01 = vbicq_u32(va0c01, vm4567c01);
149       vacc0x4567 = vbfdotq_f32(vacc0x4567, vb4567c01, vreinterpretq_bf16_u32(va0x4567c01));
150       const uint32x4_t va1x4567c01 = vbicq_u32(va1c01, vm4567c01);
151       vacc1x4567 = vbfdotq_f32(vacc1x4567, vb4567c01, vreinterpretq_bf16_u32(va1x4567c01));
152       const uint32x4_t va2x4567c01 = vbicq_u32(va2c01, vm4567c01);
153       vacc2x4567 = vbfdotq_f32(vacc2x4567, vb4567c01, vreinterpretq_bf16_u32(va2x4567c01));
154       const uint32x4_t va3x4567c01 = vbicq_u32(va3c01, vm4567c01);
155       vacc3x4567 = vbfdotq_f32(vacc3x4567, vb4567c01, vreinterpretq_bf16_u32(va3x4567c01));
156 
157       if (k > 2 * sizeof(bfloat16_t)) {
158         const bfloat16x8_t vb0123c23 = vld1q_bf16(w); w += 8;
159         const bfloat16x8_t vb4567c23 = vld1q_bf16(w); w += 8;
160 
161         const uint32x4_t va0c23 = vdupq_lane_u32(vreinterpret_u32_bf16(vget_low_bf16(va0)), 1);
162         const uint32x4_t va1c23 = vdupq_lane_u32(vreinterpret_u32_bf16(vget_low_bf16(va1)), 1);
163         const uint32x4_t va2c23 = vdupq_lane_u32(vreinterpret_u32_bf16(vget_low_bf16(va2)), 1);
164         const uint32x4_t va3c23 = vdupq_lane_u32(vreinterpret_u32_bf16(vget_low_bf16(va3)), 1);
165 
166         const uint32x4_t vm0123c23 = vreinterpretq_u32_u16(vceqq_u16(vreinterpretq_u16_bf16(vb0123c23), vmovq_n_u16(0)));
167         const uint32x4_t vm4567c23 = vreinterpretq_u32_u16(vceqq_u16(vreinterpretq_u16_bf16(vb4567c23), vmovq_n_u16(0)));
168 
169         const uint32x4_t va0x0123c23 = vbicq_u32(va0c23, vm0123c23);
170         vacc0x0123 = vbfdotq_f32(vacc0x0123, vb0123c23, vreinterpretq_bf16_u32(va0x0123c23));
171         const uint32x4_t va1x0123c23 = vbicq_u32(va1c23, vm0123c23);
172         vacc1x0123 = vbfdotq_f32(vacc1x0123, vb0123c23, vreinterpretq_bf16_u32(va1x0123c23));
173         const uint32x4_t va2x0123c23 = vbicq_u32(va2c23, vm0123c23);
174         vacc2x0123 = vbfdotq_f32(vacc2x0123, vb0123c23, vreinterpretq_bf16_u32(va2x0123c23));
175         const uint32x4_t va3x0123c23 = vbicq_u32(va3c23, vm0123c23);
176         vacc3x0123 = vbfdotq_f32(vacc3x0123, vb0123c23, vreinterpretq_bf16_u32(va3x0123c23));
177         const uint32x4_t va0x4567c23 = vbicq_u32(va0c23, vm4567c23);
178         vacc0x4567 = vbfdotq_f32(vacc0x4567, vb4567c23, vreinterpretq_bf16_u32(va0x4567c23));
179         const uint32x4_t va1x4567c23 = vbicq_u32(va1c23, vm4567c23);
180         vacc1x4567 = vbfdotq_f32(vacc1x4567, vb4567c23, vreinterpretq_bf16_u32(va1x4567c23));
181         const uint32x4_t va2x4567c23 = vbicq_u32(va2c23, vm4567c23);
182         vacc2x4567 = vbfdotq_f32(vacc2x4567, vb4567c23, vreinterpretq_bf16_u32(va2x4567c23));
183         const uint32x4_t va3x4567c23 = vbicq_u32(va3c23, vm4567c23);
184         vacc3x4567 = vbfdotq_f32(vacc3x4567, vb4567c23, vreinterpretq_bf16_u32(va3x4567c23));
185 
186         if (k > 4 * sizeof(bfloat16_t)) {
187           const bfloat16x8_t vb0123c45 = vld1q_bf16(w); w += 8;
188           const bfloat16x8_t vb4567c45 = vld1q_bf16(w); w += 8;
189 
190           const uint32x4_t va0c45 = vdupq_lane_u32(vreinterpret_u32_bf16(vget_high_bf16(va0)), 0);
191           const uint32x4_t va1c45 = vdupq_lane_u32(vreinterpret_u32_bf16(vget_high_bf16(va1)), 0);
192           const uint32x4_t va2c45 = vdupq_lane_u32(vreinterpret_u32_bf16(vget_high_bf16(va2)), 0);
193           const uint32x4_t va3c45 = vdupq_lane_u32(vreinterpret_u32_bf16(vget_high_bf16(va3)), 0);
194 
195           const uint32x4_t vm0123c45 = vreinterpretq_u32_u16(vceqq_u16(vreinterpretq_u16_bf16(vb0123c45), vmovq_n_u16(0)));
196           const uint32x4_t vm4567c45 = vreinterpretq_u32_u16(vceqq_u16(vreinterpretq_u16_bf16(vb4567c45), vmovq_n_u16(0)));
197 
198           const uint32x4_t va0x0123c45 = vbicq_u32(va0c45, vm0123c45);
199           vacc0x0123 = vbfdotq_f32(vacc0x0123, vb0123c45, vreinterpretq_bf16_u32(va0x0123c45));
200           const uint32x4_t va1x0123c45 = vbicq_u32(va1c45, vm0123c45);
201           vacc1x0123 = vbfdotq_f32(vacc1x0123, vb0123c45, vreinterpretq_bf16_u32(va1x0123c45));
202           const uint32x4_t va2x0123c45 = vbicq_u32(va2c45, vm0123c45);
203           vacc2x0123 = vbfdotq_f32(vacc2x0123, vb0123c45, vreinterpretq_bf16_u32(va2x0123c45));
204           const uint32x4_t va3x0123c45 = vbicq_u32(va3c45, vm0123c45);
205           vacc3x0123 = vbfdotq_f32(vacc3x0123, vb0123c45, vreinterpretq_bf16_u32(va3x0123c45));
206           const uint32x4_t va0x4567c45 = vbicq_u32(va0c45, vm4567c45);
207           vacc0x4567 = vbfdotq_f32(vacc0x4567, vb4567c45, vreinterpretq_bf16_u32(va0x4567c45));
208           const uint32x4_t va1x4567c45 = vbicq_u32(va1c45, vm4567c45);
209           vacc1x4567 = vbfdotq_f32(vacc1x4567, vb4567c45, vreinterpretq_bf16_u32(va1x4567c45));
210           const uint32x4_t va2x4567c45 = vbicq_u32(va2c45, vm4567c45);
211           vacc2x4567 = vbfdotq_f32(vacc2x4567, vb4567c45, vreinterpretq_bf16_u32(va2x4567c45));
212           const uint32x4_t va3x4567c45 = vbicq_u32(va3c45, vm4567c45);
213           vacc3x4567 = vbfdotq_f32(vacc3x4567, vb4567c45, vreinterpretq_bf16_u32(va3x4567c45));
214 
215           if (k > 6 * sizeof(bfloat16_t)) {
216             const bfloat16x8_t vb0123c67 = vld1q_bf16(w); w += 8;
217             const bfloat16x8_t vb4567c67 = vld1q_bf16(w); w += 8;
218 
219             const uint32x4_t va0c67 = vdupq_lane_u32(vreinterpret_u32_bf16(vget_high_bf16(va0)), 1);
220             const uint32x4_t va1c67 = vdupq_lane_u32(vreinterpret_u32_bf16(vget_high_bf16(va1)), 1);
221             const uint32x4_t va2c67 = vdupq_lane_u32(vreinterpret_u32_bf16(vget_high_bf16(va2)), 1);
222             const uint32x4_t va3c67 = vdupq_lane_u32(vreinterpret_u32_bf16(vget_high_bf16(va3)), 1);
223 
224             const uint32x4_t vm0123c67 = vreinterpretq_u32_u16(vceqq_u16(vreinterpretq_u16_bf16(vb0123c67), vmovq_n_u16(0)));
225             const uint32x4_t vm4567c67 = vreinterpretq_u32_u16(vceqq_u16(vreinterpretq_u16_bf16(vb4567c67), vmovq_n_u16(0)));
226 
227             const uint32x4_t va0x0123c67 = vbicq_u32(va0c67, vm0123c67);
228             vacc0x0123 = vbfdotq_f32(vacc0x0123, vb0123c67, vreinterpretq_bf16_u32(va0x0123c67));
229             const uint32x4_t va1x0123c67 = vbicq_u32(va1c67, vm0123c67);
230             vacc1x0123 = vbfdotq_f32(vacc1x0123, vb0123c67, vreinterpretq_bf16_u32(va1x0123c67));
231             const uint32x4_t va2x0123c67 = vbicq_u32(va2c67, vm0123c67);
232             vacc2x0123 = vbfdotq_f32(vacc2x0123, vb0123c67, vreinterpretq_bf16_u32(va2x0123c67));
233             const uint32x4_t va3x0123c67 = vbicq_u32(va3c67, vm0123c67);
234             vacc3x0123 = vbfdotq_f32(vacc3x0123, vb0123c67, vreinterpretq_bf16_u32(va3x0123c67));
235             const uint32x4_t va0x4567c67 = vbicq_u32(va0c67, vm4567c67);
236             vacc0x4567 = vbfdotq_f32(vacc0x4567, vb4567c67, vreinterpretq_bf16_u32(va0x4567c67));
237             const uint32x4_t va1x4567c67 = vbicq_u32(va1c67, vm4567c67);
238             vacc1x4567 = vbfdotq_f32(vacc1x4567, vb4567c67, vreinterpretq_bf16_u32(va1x4567c67));
239             const uint32x4_t va2x4567c67 = vbicq_u32(va2c67, vm4567c67);
240             vacc2x4567 = vbfdotq_f32(vacc2x4567, vb4567c67, vreinterpretq_bf16_u32(va2x4567c67));
241             const uint32x4_t va3x4567c67 = vbicq_u32(va3c67, vm4567c67);
242             vacc3x4567 = vbfdotq_f32(vacc3x4567, vb4567c67, vreinterpretq_bf16_u32(va3x4567c67));
243           }
244         }
245       }
246     }
247 
248     const float32x4_t vmax = vld1q_dup_f32(&params->scalar.max);
249     vacc0x0123 = vminq_f32(vacc0x0123, vmax);
250     vacc1x0123 = vminq_f32(vacc1x0123, vmax);
251     vacc2x0123 = vminq_f32(vacc2x0123, vmax);
252     vacc3x0123 = vminq_f32(vacc3x0123, vmax);
253     vacc0x4567 = vminq_f32(vacc0x4567, vmax);
254     vacc1x4567 = vminq_f32(vacc1x4567, vmax);
255     vacc2x4567 = vminq_f32(vacc2x4567, vmax);
256     vacc3x4567 = vminq_f32(vacc3x4567, vmax);
257 
258     const float32x4_t vmin = vld1q_dup_f32(&params->scalar.min);
259     vacc0x0123 = vmaxq_f32(vacc0x0123, vmin);
260     vacc1x0123 = vmaxq_f32(vacc1x0123, vmin);
261     vacc2x0123 = vmaxq_f32(vacc2x0123, vmin);
262     vacc3x0123 = vmaxq_f32(vacc3x0123, vmin);
263     vacc0x4567 = vmaxq_f32(vacc0x4567, vmin);
264     vacc1x4567 = vmaxq_f32(vacc1x4567, vmin);
265     vacc2x4567 = vmaxq_f32(vacc2x4567, vmin);
266     vacc3x4567 = vmaxq_f32(vacc3x4567, vmin);
267 
268     bfloat16x4_t vout0x0123 = vcvt_bf16_f32(vacc0x0123);
269     bfloat16x4_t vout1x0123 = vcvt_bf16_f32(vacc1x0123);
270     bfloat16x4_t vout2x0123 = vcvt_bf16_f32(vacc2x0123);
271     bfloat16x4_t vout3x0123 = vcvt_bf16_f32(vacc3x0123);
272     bfloat16x4_t vout0x4567 = vcvt_bf16_f32(vacc0x4567);
273     bfloat16x4_t vout1x4567 = vcvt_bf16_f32(vacc1x4567);
274     bfloat16x4_t vout2x4567 = vcvt_bf16_f32(vacc2x4567);
275     bfloat16x4_t vout3x4567 = vcvt_bf16_f32(vacc3x4567);
276 
277     if XNN_LIKELY(nc >= 8) {
278       vst1_bf16(c0, vout0x0123);
279       vst1_bf16(c0 + 4, vout0x4567);
280       c0 = (bfloat16_t*) ((uintptr_t) c0 + cn_stride);
281       vst1_bf16(c1, vout1x0123);
282       vst1_bf16(c1 + 4, vout1x4567);
283       c1 = (bfloat16_t*) ((uintptr_t) c1 + cn_stride);
284       vst1_bf16(c2, vout2x0123);
285       vst1_bf16(c2 + 4, vout2x4567);
286       c2 = (bfloat16_t*) ((uintptr_t) c2 + cn_stride);
287       vst1_bf16(c3, vout3x0123);
288       vst1_bf16(c3 + 4, vout3x4567);
289       c3 = (bfloat16_t*) ((uintptr_t) c3 + cn_stride);
290 
291       a0 = (const bfloat16_t*) ((uintptr_t) a0 - kc);
292       a1 = (const bfloat16_t*) ((uintptr_t) a1 - kc);
293       a2 = (const bfloat16_t*) ((uintptr_t) a2 - kc);
294       a3 = (const bfloat16_t*) ((uintptr_t) a3 - kc);
295 
296       nc -= 8;
297     } else {
298       if (nc & 4) {
299         vst1_bf16(c0, vout0x0123); c0 += 4;
300         vst1_bf16(c1, vout1x0123); c1 += 4;
301         vst1_bf16(c2, vout2x0123); c2 += 4;
302         vst1_bf16(c3, vout3x0123); c3 += 4;
303 
304         vout0x0123 = vout0x4567;
305         vout1x0123 = vout1x4567;
306         vout2x0123 = vout2x4567;
307         vout3x0123 = vout3x4567;
308       }
309       if (nc & 2) {
310         vst1_lane_u32((void*) c0, vreinterpret_u32_bf16(vout0x0123), 0); c0 += 2;
311         vst1_lane_u32((void*) c1, vreinterpret_u32_bf16(vout1x0123), 0); c1 += 2;
312         vst1_lane_u32((void*) c2, vreinterpret_u32_bf16(vout2x0123), 0); c2 += 2;
313         vst1_lane_u32((void*) c3, vreinterpret_u32_bf16(vout3x0123), 0); c3 += 2;
314 
315         vout0x0123 = vreinterpret_bf16_u16(vext_u16(vreinterpret_u16_bf16(vout0x0123), vreinterpret_u16_bf16(vout0x0123), 2));
316         vout1x0123 = vreinterpret_bf16_u16(vext_u16(vreinterpret_u16_bf16(vout1x0123), vreinterpret_u16_bf16(vout1x0123), 2));
317         vout2x0123 = vreinterpret_bf16_u16(vext_u16(vreinterpret_u16_bf16(vout2x0123), vreinterpret_u16_bf16(vout2x0123), 2));
318         vout3x0123 = vreinterpret_bf16_u16(vext_u16(vreinterpret_u16_bf16(vout3x0123), vreinterpret_u16_bf16(vout3x0123), 2));
319       }
320       if (nc & 1) {
321         vst1_lane_bf16(c0, vout0x0123, 0);
322         vst1_lane_bf16(c1, vout1x0123, 0);
323         vst1_lane_bf16(c2, vout2x0123, 0);
324         vst1_lane_bf16(c3, vout3x0123, 0);
325       }
326 
327       nc = 0;
328     }
329   } while (nc != 0);
330 }
331