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_4x4c8__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_4x4c8__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 <= 4);
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
60 const uint16_t* w = (const uint16_t*) w_ptr;
61 const uint16x8_t vmask = vreinterpretq_u16_u32(vmovq_n_u32(UINT32_C(0xFFFF0000)));
62 do {
63 float32x4_t vacc0x0 = vreinterpretq_f32_u32(vshll_n_u16(vld1_lane_u16(w, vdup_n_u16(0), 0), 16)); w += 1;
64 float32x4_t vacc0x1 = vreinterpretq_f32_u32(vshll_n_u16(vld1_lane_u16(w, vdup_n_u16(0), 0), 16)); w += 1;
65 float32x4_t vacc0x2 = vreinterpretq_f32_u32(vshll_n_u16(vld1_lane_u16(w, vdup_n_u16(0), 0), 16)); w += 1;
66 float32x4_t vacc0x3 = vreinterpretq_f32_u32(vshll_n_u16(vld1_lane_u16(w, vdup_n_u16(0), 0), 16)); w += 1;
67 float32x4_t vacc1x0 = vacc0x0;
68 float32x4_t vacc1x1 = vacc0x1;
69 float32x4_t vacc1x2 = vacc0x2;
70 float32x4_t vacc1x3 = vacc0x3;
71 float32x4_t vacc2x0 = vacc0x0;
72 float32x4_t vacc2x1 = vacc0x1;
73 float32x4_t vacc2x2 = vacc0x2;
74 float32x4_t vacc2x3 = vacc0x3;
75 float32x4_t vacc3x0 = vacc0x0;
76 float32x4_t vacc3x1 = vacc0x1;
77 float32x4_t vacc3x2 = vacc0x2;
78 float32x4_t vacc3x3 = vacc0x3;
79
80 size_t k = kc;
81 for (; k >= 8 * sizeof(uint16_t); k -= 8 * sizeof(uint16_t)) {
82 const uint16x8_t va0 = vld1q_u16(a0); a0 += 8;
83 const uint16x8_t va1 = vld1q_u16(a1); a1 += 8;
84 const uint16x8_t va2 = vld1q_u16(a2); a2 += 8;
85 const uint16x8_t va3 = vld1q_u16(a3); a3 += 8;
86
87 const uint16x8_t vb0 = vld1q_u16(w); w += 8;
88 const uint16x8_t vb1 = vld1q_u16(w); w += 8;
89 const uint16x8_t vb2 = vld1q_u16(w); w += 8;
90 const uint16x8_t vb3 = vld1q_u16(w); w += 8;
91
92 const float32x4_t va0e = vreinterpretq_f32_u32(vshlq_n_u32(vreinterpretq_u32_u16(va0), 16));
93 const float32x4_t va1e = vreinterpretq_f32_u32(vshlq_n_u32(vreinterpretq_u32_u16(va1), 16));
94 const float32x4_t va2e = vreinterpretq_f32_u32(vshlq_n_u32(vreinterpretq_u32_u16(va2), 16));
95 const float32x4_t va3e = vreinterpretq_f32_u32(vshlq_n_u32(vreinterpretq_u32_u16(va3), 16));
96
97 const float32x4_t vb0e = vreinterpretq_f32_u32(vshlq_n_u32(vreinterpretq_u32_u16(vb0), 16));
98 const float32x4_t vb1e = vreinterpretq_f32_u32(vshlq_n_u32(vreinterpretq_u32_u16(vb1), 16));
99 const float32x4_t vb2e = vreinterpretq_f32_u32(vshlq_n_u32(vreinterpretq_u32_u16(vb2), 16));
100 const float32x4_t vb3e = vreinterpretq_f32_u32(vshlq_n_u32(vreinterpretq_u32_u16(vb3), 16));
101
102 vacc0x0 = vfmaq_f32(vacc0x0, va0e, vb0e);
103 vacc1x0 = vfmaq_f32(vacc1x0, va1e, vb0e);
104 vacc2x0 = vfmaq_f32(vacc2x0, va2e, vb0e);
105 vacc3x0 = vfmaq_f32(vacc3x0, va3e, vb0e);
106 vacc0x1 = vfmaq_f32(vacc0x1, va0e, vb1e);
107 vacc1x1 = vfmaq_f32(vacc1x1, va1e, vb1e);
108 vacc2x1 = vfmaq_f32(vacc2x1, va2e, vb1e);
109 vacc3x1 = vfmaq_f32(vacc3x1, va3e, vb1e);
110 vacc0x2 = vfmaq_f32(vacc0x2, va0e, vb2e);
111 vacc1x2 = vfmaq_f32(vacc1x2, va1e, vb2e);
112 vacc2x2 = vfmaq_f32(vacc2x2, va2e, vb2e);
113 vacc3x2 = vfmaq_f32(vacc3x2, va3e, vb2e);
114 vacc0x3 = vfmaq_f32(vacc0x3, va0e, vb3e);
115 vacc1x3 = vfmaq_f32(vacc1x3, va1e, vb3e);
116 vacc2x3 = vfmaq_f32(vacc2x3, va2e, vb3e);
117 vacc3x3 = vfmaq_f32(vacc3x3, va3e, vb3e);
118
119 const float32x4_t va0o = vreinterpretq_f32_u16(vandq_u16(va0, vmask));
120 const float32x4_t va1o = vreinterpretq_f32_u16(vandq_u16(va1, vmask));
121 const float32x4_t va2o = vreinterpretq_f32_u16(vandq_u16(va2, vmask));
122 const float32x4_t va3o = vreinterpretq_f32_u16(vandq_u16(va3, vmask));
123
124 const float32x4_t vb0o = vreinterpretq_f32_u16(vandq_u16(vb0, vmask));
125 const float32x4_t vb1o = vreinterpretq_f32_u16(vandq_u16(vb1, vmask));
126 const float32x4_t vb2o = vreinterpretq_f32_u16(vandq_u16(vb2, vmask));
127 const float32x4_t vb3o = vreinterpretq_f32_u16(vandq_u16(vb3, vmask));
128
129 vacc0x0 = vfmaq_f32(vacc0x0, va0o, vb0o);
130 vacc1x0 = vfmaq_f32(vacc1x0, va1o, vb0o);
131 vacc2x0 = vfmaq_f32(vacc2x0, va2o, vb0o);
132 vacc3x0 = vfmaq_f32(vacc3x0, va3o, vb0o);
133 vacc0x1 = vfmaq_f32(vacc0x1, va0o, vb1o);
134 vacc1x1 = vfmaq_f32(vacc1x1, va1o, vb1o);
135 vacc2x1 = vfmaq_f32(vacc2x1, va2o, vb1o);
136 vacc3x1 = vfmaq_f32(vacc3x1, va3o, vb1o);
137 vacc0x2 = vfmaq_f32(vacc0x2, va0o, vb2o);
138 vacc1x2 = vfmaq_f32(vacc1x2, va1o, vb2o);
139 vacc2x2 = vfmaq_f32(vacc2x2, va2o, vb2o);
140 vacc3x2 = vfmaq_f32(vacc3x2, va3o, vb2o);
141 vacc0x3 = vfmaq_f32(vacc0x3, va0o, vb3o);
142 vacc1x3 = vfmaq_f32(vacc1x3, va1o, vb3o);
143 vacc2x3 = vfmaq_f32(vacc2x3, va2o, vb3o);
144 vacc3x3 = vfmaq_f32(vacc3x3, va3o, vb3o);
145 }
146 if XNN_UNLIKELY(k != 0) {
147 const uint16x8_t va0 = vld1q_u16(a0); a0 = (const uint16_t*) ((uintptr_t) a0 + k);
148 const uint16x8_t va1 = vld1q_u16(a1); a1 = (const uint16_t*) ((uintptr_t) a1 + k);
149 const uint16x8_t va2 = vld1q_u16(a2); a2 = (const uint16_t*) ((uintptr_t) a2 + k);
150 const uint16x8_t va3 = vld1q_u16(a3); a3 = (const uint16_t*) ((uintptr_t) a3 + k);
151
152 const uint16x8_t vb0 = vld1q_u16(w); w += 8;
153 const uint16x8_t vb1 = vld1q_u16(w); w += 8;
154 const uint16x8_t vb2 = vld1q_u16(w); w += 8;
155 const uint16x8_t vb3 = vld1q_u16(w); w += 8;
156
157 const uint16x8_t vm0 = vceqq_u16(vb0, vmovq_n_u16(0));
158 const uint16x8_t vm1 = vceqq_u16(vb1, vmovq_n_u16(0));
159 const uint16x8_t vm2 = vceqq_u16(vb2, vmovq_n_u16(0));
160 const uint16x8_t vm3 = vceqq_u16(vb3, vmovq_n_u16(0));
161
162 const float32x4_t vb0e = vreinterpretq_f32_u32(vshlq_n_u32(vreinterpretq_u32_u16(vb0), 16));
163 const float32x4_t vb1e = vreinterpretq_f32_u32(vshlq_n_u32(vreinterpretq_u32_u16(vb1), 16));
164 const float32x4_t vb2e = vreinterpretq_f32_u32(vshlq_n_u32(vreinterpretq_u32_u16(vb2), 16));
165 const float32x4_t vb3e = vreinterpretq_f32_u32(vshlq_n_u32(vreinterpretq_u32_u16(vb3), 16));
166
167 const uint16x8_t va0x0 = vbicq_u16(va0, vm0);
168 const uint16x8_t va1x0 = vbicq_u16(va1, vm0);
169 const uint16x8_t va2x0 = vbicq_u16(va2, vm0);
170 const uint16x8_t va3x0 = vbicq_u16(va3, vm0);
171 const uint16x8_t va0x1 = vbicq_u16(va0, vm1);
172 const uint16x8_t va1x1 = vbicq_u16(va1, vm1);
173 const uint16x8_t va2x1 = vbicq_u16(va2, vm1);
174 const uint16x8_t va3x1 = vbicq_u16(va3, vm1);
175 const uint16x8_t va0x2 = vbicq_u16(va0, vm2);
176 const uint16x8_t va1x2 = vbicq_u16(va1, vm2);
177 const uint16x8_t va2x2 = vbicq_u16(va2, vm2);
178 const uint16x8_t va3x2 = vbicq_u16(va3, vm2);
179 const uint16x8_t va0x3 = vbicq_u16(va0, vm3);
180 const uint16x8_t va1x3 = vbicq_u16(va1, vm3);
181 const uint16x8_t va2x3 = vbicq_u16(va2, vm3);
182 const uint16x8_t va3x3 = vbicq_u16(va3, vm3);
183
184 const float32x4_t va0x0e = vreinterpretq_f32_u32(vshlq_n_u32(vreinterpretq_u32_u16(va0x0), 16));
185 const float32x4_t va1x0e = vreinterpretq_f32_u32(vshlq_n_u32(vreinterpretq_u32_u16(va1x0), 16));
186 const float32x4_t va2x0e = vreinterpretq_f32_u32(vshlq_n_u32(vreinterpretq_u32_u16(va2x0), 16));
187 const float32x4_t va3x0e = vreinterpretq_f32_u32(vshlq_n_u32(vreinterpretq_u32_u16(va3x0), 16));
188 const float32x4_t va0x1e = vreinterpretq_f32_u32(vshlq_n_u32(vreinterpretq_u32_u16(va0x1), 16));
189 const float32x4_t va1x1e = vreinterpretq_f32_u32(vshlq_n_u32(vreinterpretq_u32_u16(va1x1), 16));
190 const float32x4_t va2x1e = vreinterpretq_f32_u32(vshlq_n_u32(vreinterpretq_u32_u16(va2x1), 16));
191 const float32x4_t va3x1e = vreinterpretq_f32_u32(vshlq_n_u32(vreinterpretq_u32_u16(va3x1), 16));
192 const float32x4_t va0x2e = vreinterpretq_f32_u32(vshlq_n_u32(vreinterpretq_u32_u16(va0x2), 16));
193 const float32x4_t va1x2e = vreinterpretq_f32_u32(vshlq_n_u32(vreinterpretq_u32_u16(va1x2), 16));
194 const float32x4_t va2x2e = vreinterpretq_f32_u32(vshlq_n_u32(vreinterpretq_u32_u16(va2x2), 16));
195 const float32x4_t va3x2e = vreinterpretq_f32_u32(vshlq_n_u32(vreinterpretq_u32_u16(va3x2), 16));
196 const float32x4_t va0x3e = vreinterpretq_f32_u32(vshlq_n_u32(vreinterpretq_u32_u16(va0x3), 16));
197 const float32x4_t va1x3e = vreinterpretq_f32_u32(vshlq_n_u32(vreinterpretq_u32_u16(va1x3), 16));
198 const float32x4_t va2x3e = vreinterpretq_f32_u32(vshlq_n_u32(vreinterpretq_u32_u16(va2x3), 16));
199 const float32x4_t va3x3e = vreinterpretq_f32_u32(vshlq_n_u32(vreinterpretq_u32_u16(va3x3), 16));
200
201 vacc0x0 = vfmaq_f32(vacc0x0, va0x0e, vb0e);
202 vacc1x0 = vfmaq_f32(vacc1x0, va1x0e, vb0e);
203 vacc2x0 = vfmaq_f32(vacc2x0, va2x0e, vb0e);
204 vacc3x0 = vfmaq_f32(vacc3x0, va3x0e, vb0e);
205 vacc0x1 = vfmaq_f32(vacc0x1, va0x1e, vb1e);
206 vacc1x1 = vfmaq_f32(vacc1x1, va1x1e, vb1e);
207 vacc2x1 = vfmaq_f32(vacc2x1, va2x1e, vb1e);
208 vacc3x1 = vfmaq_f32(vacc3x1, va3x1e, vb1e);
209 vacc0x2 = vfmaq_f32(vacc0x2, va0x2e, vb2e);
210 vacc1x2 = vfmaq_f32(vacc1x2, va1x2e, vb2e);
211 vacc2x2 = vfmaq_f32(vacc2x2, va2x2e, vb2e);
212 vacc3x2 = vfmaq_f32(vacc3x2, va3x2e, vb2e);
213 vacc0x3 = vfmaq_f32(vacc0x3, va0x3e, vb3e);
214 vacc1x3 = vfmaq_f32(vacc1x3, va1x3e, vb3e);
215 vacc2x3 = vfmaq_f32(vacc2x3, va2x3e, vb3e);
216 vacc3x3 = vfmaq_f32(vacc3x3, va3x3e, vb3e);
217
218 const float32x4_t vb0o = vreinterpretq_f32_u16(vandq_u16(vb0, vmask));
219 const float32x4_t vb1o = vreinterpretq_f32_u16(vandq_u16(vb1, vmask));
220 const float32x4_t vb2o = vreinterpretq_f32_u16(vandq_u16(vb2, vmask));
221 const float32x4_t vb3o = vreinterpretq_f32_u16(vandq_u16(vb3, vmask));
222
223 const float32x4_t va0x0o = vreinterpretq_f32_u16(vandq_u16(va0x0, vmask));
224 const float32x4_t va1x0o = vreinterpretq_f32_u16(vandq_u16(va1x0, vmask));
225 const float32x4_t va2x0o = vreinterpretq_f32_u16(vandq_u16(va2x0, vmask));
226 const float32x4_t va3x0o = vreinterpretq_f32_u16(vandq_u16(va3x0, vmask));
227 const float32x4_t va0x1o = vreinterpretq_f32_u16(vandq_u16(va0x1, vmask));
228 const float32x4_t va1x1o = vreinterpretq_f32_u16(vandq_u16(va1x1, vmask));
229 const float32x4_t va2x1o = vreinterpretq_f32_u16(vandq_u16(va2x1, vmask));
230 const float32x4_t va3x1o = vreinterpretq_f32_u16(vandq_u16(va3x1, vmask));
231 const float32x4_t va0x2o = vreinterpretq_f32_u16(vandq_u16(va0x2, vmask));
232 const float32x4_t va1x2o = vreinterpretq_f32_u16(vandq_u16(va1x2, vmask));
233 const float32x4_t va2x2o = vreinterpretq_f32_u16(vandq_u16(va2x2, vmask));
234 const float32x4_t va3x2o = vreinterpretq_f32_u16(vandq_u16(va3x2, vmask));
235 const float32x4_t va0x3o = vreinterpretq_f32_u16(vandq_u16(va0x3, vmask));
236 const float32x4_t va1x3o = vreinterpretq_f32_u16(vandq_u16(va1x3, vmask));
237 const float32x4_t va2x3o = vreinterpretq_f32_u16(vandq_u16(va2x3, vmask));
238 const float32x4_t va3x3o = vreinterpretq_f32_u16(vandq_u16(va3x3, vmask));
239
240 vacc0x0 = vfmaq_f32(vacc0x0, va0x0o, vb0o);
241 vacc1x0 = vfmaq_f32(vacc1x0, va1x0o, vb0o);
242 vacc2x0 = vfmaq_f32(vacc2x0, va2x0o, vb0o);
243 vacc3x0 = vfmaq_f32(vacc3x0, va3x0o, vb0o);
244 vacc0x1 = vfmaq_f32(vacc0x1, va0x1o, vb1o);
245 vacc1x1 = vfmaq_f32(vacc1x1, va1x1o, vb1o);
246 vacc2x1 = vfmaq_f32(vacc2x1, va2x1o, vb1o);
247 vacc3x1 = vfmaq_f32(vacc3x1, va3x1o, vb1o);
248 vacc0x2 = vfmaq_f32(vacc0x2, va0x2o, vb2o);
249 vacc1x2 = vfmaq_f32(vacc1x2, va1x2o, vb2o);
250 vacc2x2 = vfmaq_f32(vacc2x2, va2x2o, vb2o);
251 vacc3x2 = vfmaq_f32(vacc3x2, va3x2o, vb2o);
252 vacc0x3 = vfmaq_f32(vacc0x3, va0x3o, vb3o);
253 vacc1x3 = vfmaq_f32(vacc1x3, va1x3o, vb3o);
254 vacc2x3 = vfmaq_f32(vacc2x3, va2x3o, vb3o);
255 vacc3x3 = vfmaq_f32(vacc3x3, va3x3o, vb3o);
256 }
257
258 #if XNN_ARCH_ARM64
259 const float32x4_t vacc0x01 = vpaddq_f32(vacc0x0, vacc0x1);
260 const float32x4_t vacc1x01 = vpaddq_f32(vacc1x0, vacc1x1);
261 const float32x4_t vacc2x01 = vpaddq_f32(vacc2x0, vacc2x1);
262 const float32x4_t vacc3x01 = vpaddq_f32(vacc3x0, vacc3x1);
263 const float32x4_t vacc0x23 = vpaddq_f32(vacc0x2, vacc0x3);
264 const float32x4_t vacc1x23 = vpaddq_f32(vacc1x2, vacc1x3);
265 const float32x4_t vacc2x23 = vpaddq_f32(vacc2x2, vacc2x3);
266 const float32x4_t vacc3x23 = vpaddq_f32(vacc3x2, vacc3x3);
267
268 float32x4_t vacc0x0123 = vpaddq_f32(vacc0x01, vacc0x23);
269 float32x4_t vacc1x0123 = vpaddq_f32(vacc1x01, vacc1x23);
270 float32x4_t vacc2x0123 = vpaddq_f32(vacc2x01, vacc2x23);
271 float32x4_t vacc3x0123 = vpaddq_f32(vacc3x01, vacc3x23);
272 #else
273 const float32x2_t vsum0x0 = vadd_f32(vget_low_f32(vacc0x0), vget_high_f32(vacc0x0));
274 const float32x2_t vsum1x0 = vadd_f32(vget_low_f32(vacc1x0), vget_high_f32(vacc1x0));
275 const float32x2_t vsum2x0 = vadd_f32(vget_low_f32(vacc2x0), vget_high_f32(vacc2x0));
276 const float32x2_t vsum3x0 = vadd_f32(vget_low_f32(vacc3x0), vget_high_f32(vacc3x0));
277 const float32x2_t vsum0x1 = vadd_f32(vget_low_f32(vacc0x1), vget_high_f32(vacc0x1));
278 const float32x2_t vsum1x1 = vadd_f32(vget_low_f32(vacc1x1), vget_high_f32(vacc1x1));
279 const float32x2_t vsum2x1 = vadd_f32(vget_low_f32(vacc2x1), vget_high_f32(vacc2x1));
280 const float32x2_t vsum3x1 = vadd_f32(vget_low_f32(vacc3x1), vget_high_f32(vacc3x1));
281 const float32x2_t vsum0x2 = vadd_f32(vget_low_f32(vacc0x2), vget_high_f32(vacc0x2));
282 const float32x2_t vsum1x2 = vadd_f32(vget_low_f32(vacc1x2), vget_high_f32(vacc1x2));
283 const float32x2_t vsum2x2 = vadd_f32(vget_low_f32(vacc2x2), vget_high_f32(vacc2x2));
284 const float32x2_t vsum3x2 = vadd_f32(vget_low_f32(vacc3x2), vget_high_f32(vacc3x2));
285 const float32x2_t vsum0x3 = vadd_f32(vget_low_f32(vacc0x3), vget_high_f32(vacc0x3));
286 const float32x2_t vsum1x3 = vadd_f32(vget_low_f32(vacc1x3), vget_high_f32(vacc1x3));
287 const float32x2_t vsum2x3 = vadd_f32(vget_low_f32(vacc2x3), vget_high_f32(vacc2x3));
288 const float32x2_t vsum3x3 = vadd_f32(vget_low_f32(vacc3x3), vget_high_f32(vacc3x3));
289
290 float32x4_t vacc0x0123 = vcombine_f32(vpadd_f32(vsum0x0, vsum0x1), vpadd_f32(vsum0x2, vsum0x3));
291 float32x4_t vacc1x0123 = vcombine_f32(vpadd_f32(vsum1x0, vsum1x1), vpadd_f32(vsum1x2, vsum1x3));
292 float32x4_t vacc2x0123 = vcombine_f32(vpadd_f32(vsum2x0, vsum2x1), vpadd_f32(vsum2x2, vsum2x3));
293 float32x4_t vacc3x0123 = vcombine_f32(vpadd_f32(vsum3x0, vsum3x1), vpadd_f32(vsum3x2, vsum3x3));
294 #endif
295
296 const float32x4_t vmax = vld1q_dup_f32(¶ms->scalar.max);
297 vacc0x0123 = vminq_f32(vacc0x0123, vmax);
298 vacc1x0123 = vminq_f32(vacc1x0123, vmax);
299 vacc2x0123 = vminq_f32(vacc2x0123, vmax);
300 vacc3x0123 = vminq_f32(vacc3x0123, vmax);
301
302 const float32x4_t vmin = vld1q_dup_f32(¶ms->scalar.min);
303 vacc0x0123 = vmaxq_f32(vacc0x0123, vmin);
304 vacc1x0123 = vmaxq_f32(vacc1x0123, vmin);
305 vacc2x0123 = vmaxq_f32(vacc2x0123, vmin);
306 vacc3x0123 = vmaxq_f32(vacc3x0123, vmin);
307
308 uint16x4_t vout0x0123 = vshrn_n_u32(vreinterpretq_u32_f32(vacc0x0123), 16);
309 uint16x4_t vout1x0123 = vshrn_n_u32(vreinterpretq_u32_f32(vacc1x0123), 16);
310 uint16x4_t vout2x0123 = vshrn_n_u32(vreinterpretq_u32_f32(vacc2x0123), 16);
311 uint16x4_t vout3x0123 = vshrn_n_u32(vreinterpretq_u32_f32(vacc3x0123), 16);
312
313 if XNN_LIKELY(nc >= 4) {
314 vst1_u16(c0, vout0x0123);
315 c0 = (uint16_t*) ((uintptr_t) c0 + cn_stride);
316 vst1_u16(c1, vout1x0123);
317 c1 = (uint16_t*) ((uintptr_t) c1 + cn_stride);
318 vst1_u16(c2, vout2x0123);
319 c2 = (uint16_t*) ((uintptr_t) c2 + cn_stride);
320 vst1_u16(c3, vout3x0123);
321 c3 = (uint16_t*) ((uintptr_t) c3 + cn_stride);
322
323 a0 = (const uint16_t*) ((uintptr_t) a0 - kc);
324 a1 = (const uint16_t*) ((uintptr_t) a1 - kc);
325 a2 = (const uint16_t*) ((uintptr_t) a2 - kc);
326 a3 = (const uint16_t*) ((uintptr_t) a3 - kc);
327
328 nc -= 4;
329 } else {
330 if (nc & 2) {
331 vst1_lane_u32((void*) c0, vreinterpret_u32_u16(vout0x0123), 0); c0 += 2;
332 vst1_lane_u32((void*) c1, vreinterpret_u32_u16(vout1x0123), 0); c1 += 2;
333 vst1_lane_u32((void*) c2, vreinterpret_u32_u16(vout2x0123), 0); c2 += 2;
334 vst1_lane_u32((void*) c3, vreinterpret_u32_u16(vout3x0123), 0); c3 += 2;
335
336 vout0x0123 = vext_u16(vout0x0123, vout0x0123, 2);
337 vout1x0123 = vext_u16(vout1x0123, vout1x0123, 2);
338 vout2x0123 = vext_u16(vout2x0123, vout2x0123, 2);
339 vout3x0123 = vext_u16(vout3x0123, vout3x0123, 2);
340 }
341 if (nc & 1) {
342 vst1_lane_u16(c0, vout0x0123, 0);
343 vst1_lane_u16(c1, vout1x0123, 0);
344 vst1_lane_u16(c2, vout2x0123, 0);
345 vst1_lane_u16(c3, vout3x0123, 0);
346 }
347
348 nc = 0;
349 }
350 } while (nc != 0);
351 }
352