1 // Auto-generated file. Do not edit!
2 // Template: src/qs8-gemm/c2-neon-mull-dup.c.in
3 // Generator: tools/xngen
4 //
5 // Copyright 2021 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/gemm.h>
15 #include <xnnpack/intrinsics-polyfill.h>
16 #include <xnnpack/math.h>
17
xnn_qc8_gemm_minmax_fp32_ukernel_2x8c2__neonv8_mlal_ld2r(size_t mr,size_t nc,size_t kc,const int8_t * restrict a,size_t a_stride,const void * restrict w,int8_t * restrict c,size_t cm_stride,size_t cn_stride,const union xnn_qc8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS (1)])18 void xnn_qc8_gemm_minmax_fp32_ukernel_2x8c2__neonv8_mlal_ld2r(
19 size_t mr,
20 size_t nc,
21 size_t kc,
22 const int8_t* restrict a,
23 size_t a_stride,
24 const void* restrict w,
25 int8_t* restrict c,
26 size_t cm_stride,
27 size_t cn_stride,
28 const union xnn_qc8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
29 {
30 assert(mr != 0);
31 assert(mr <= 2);
32 assert(nc != 0);
33 assert(kc != 0);
34 assert(kc % sizeof(int8_t) == 0);
35 assert(a != NULL);
36 assert(w != NULL);
37 assert(c != NULL);
38
39 kc = round_up_po2(kc, 2 * sizeof(int8_t));
40 const int8_t* a0 = a;
41 int8_t* c0 = c;
42 const int8_t* a1 = (const int8_t*) ((uintptr_t) a0 + a_stride);
43 int8_t* c1 = (int8_t*) ((uintptr_t) c0 + cm_stride);
44 if XNN_UNPREDICTABLE(mr != 2) {
45 a1 = a0;
46 c1 = c0;
47 }
48
49 do {
50 int32x4_t vacc0x0123 = vld1q_s32(w); w = (const void*) ((uintptr_t) w + 4 * sizeof(int32_t));
51 int32x4_t vacc0x4567 = vld1q_s32(w); w = (const void*) ((uintptr_t) w + 4 * sizeof(int32_t));
52 int32x4_t vacc1x0123 = vacc0x0123;
53 int32x4_t vacc1x4567 = vacc0x4567;
54
55 size_t k = kc;
56
57 while (k >= 16 * sizeof(int8_t)) {
58 const int16x4x2_t va00x0 = vld2_dup_s16((const void*)a0);
59 const int16x4x2_t va01x0 = vld2_dup_s16((const void*)(a0 + 4)); a0 += 8;
60 const int16x4x2_t va00x1 = vld2_dup_s16((const void*)a0);
61 const int16x4x2_t va01x1 = vld2_dup_s16((const void*)(a0 + 4)); a0 += 8;
62 const int16x4x2_t va10x0 = vld2_dup_s16((const void*)a1);
63 const int16x4x2_t va11x0 = vld2_dup_s16((const void*)(a1 + 4)); a1 += 8;
64 const int16x4x2_t va10x1 = vld2_dup_s16((const void*)a1);
65 const int16x4x2_t va11x1 = vld2_dup_s16((const void*)(a1 + 4)); a1 += 8;
66 const int8x8_t vb0123c0x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
67 const int8x8_t vb4567c0x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
68 const int8x8_t vb0123c1x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
69 const int8x8_t vb4567c1x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
70 const int8x8_t vb0123c2x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
71 const int8x8_t vb4567c2x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
72 const int8x8_t vb0123c3x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
73 const int8x8_t vb4567c3x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
74
75 const int8x8_t va0c0x0 = vreinterpret_s8_s16(va00x0.val[0]);
76 const int8x8_t va0c0x1 = vreinterpret_s8_s16(va00x1.val[0]);
77 const int8x8_t va1c0x0 = vreinterpret_s8_s16(va10x0.val[0]);
78 const int8x8_t va1c0x1 = vreinterpret_s8_s16(va10x1.val[0]);
79
80 int16x8_t vprod0x0123c0 = vmull_s8(vb0123c0x0, va0c0x0);
81 int16x8_t vprod1x0123c0 = vmull_s8(vb0123c0x0, va1c0x0);
82 const int8x8_t vb0123c0x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
83 vprod0x0123c0 = vmlal_s8(vprod0x0123c0, vb0123c0x1, va0c0x1);
84 vprod1x0123c0 = vmlal_s8(vprod1x0123c0, vb0123c0x1, va1c0x1);
85 vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c0);
86 vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c0);
87 int16x8_t vprod0x4567c0 = vmull_s8(vb4567c0x0, va0c0x0);
88 int16x8_t vprod1x4567c0 = vmull_s8(vb4567c0x0, va1c0x0);
89 const int8x8_t vb4567c0x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
90 vprod0x4567c0 = vmlal_s8(vprod0x4567c0, vb4567c0x1, va0c0x1);
91 vprod1x4567c0 = vmlal_s8(vprod1x4567c0, vb4567c0x1, va1c0x1);
92 vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c0);
93 vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c0);
94 const int8x8_t va0c1x0 = vreinterpret_s8_s16(va00x0.val[1]);
95 const int8x8_t va0c1x1 = vreinterpret_s8_s16(va00x1.val[1]);
96 const int8x8_t va1c1x0 = vreinterpret_s8_s16(va10x0.val[1]);
97 const int8x8_t va1c1x1 = vreinterpret_s8_s16(va10x1.val[1]);
98
99 int16x8_t vprod0x0123c1 = vmull_s8(vb0123c1x0, va0c1x0);
100 int16x8_t vprod1x0123c1 = vmull_s8(vb0123c1x0, va1c1x0);
101 const int8x8_t vb0123c1x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
102 vprod0x0123c1 = vmlal_s8(vprod0x0123c1, vb0123c1x1, va0c1x1);
103 vprod1x0123c1 = vmlal_s8(vprod1x0123c1, vb0123c1x1, va1c1x1);
104 vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c1);
105 vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c1);
106 int16x8_t vprod0x4567c1 = vmull_s8(vb4567c1x0, va0c1x0);
107 int16x8_t vprod1x4567c1 = vmull_s8(vb4567c1x0, va1c1x0);
108 const int8x8_t vb4567c1x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
109 vprod0x4567c1 = vmlal_s8(vprod0x4567c1, vb4567c1x1, va0c1x1);
110 vprod1x4567c1 = vmlal_s8(vprod1x4567c1, vb4567c1x1, va1c1x1);
111 vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c1);
112 vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c1);
113 const int8x8_t va0c2x0 = vreinterpret_s8_s16(va01x0.val[0]);
114 const int8x8_t va0c2x1 = vreinterpret_s8_s16(va01x1.val[0]);
115 const int8x8_t va1c2x0 = vreinterpret_s8_s16(va11x0.val[0]);
116 const int8x8_t va1c2x1 = vreinterpret_s8_s16(va11x1.val[0]);
117
118 int16x8_t vprod0x0123c2 = vmull_s8(vb0123c2x0, va0c2x0);
119 int16x8_t vprod1x0123c2 = vmull_s8(vb0123c2x0, va1c2x0);
120 const int8x8_t vb0123c2x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
121 vprod0x0123c2 = vmlal_s8(vprod0x0123c2, vb0123c2x1, va0c2x1);
122 vprod1x0123c2 = vmlal_s8(vprod1x0123c2, vb0123c2x1, va1c2x1);
123 vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c2);
124 vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c2);
125 int16x8_t vprod0x4567c2 = vmull_s8(vb4567c2x0, va0c2x0);
126 int16x8_t vprod1x4567c2 = vmull_s8(vb4567c2x0, va1c2x0);
127 const int8x8_t vb4567c2x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
128 vprod0x4567c2 = vmlal_s8(vprod0x4567c2, vb4567c2x1, va0c2x1);
129 vprod1x4567c2 = vmlal_s8(vprod1x4567c2, vb4567c2x1, va1c2x1);
130 vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c2);
131 vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c2);
132 const int8x8_t va0c3x0 = vreinterpret_s8_s16(va01x0.val[1]);
133 const int8x8_t va0c3x1 = vreinterpret_s8_s16(va01x1.val[1]);
134 const int8x8_t va1c3x0 = vreinterpret_s8_s16(va11x0.val[1]);
135 const int8x8_t va1c3x1 = vreinterpret_s8_s16(va11x1.val[1]);
136
137 int16x8_t vprod0x0123c3 = vmull_s8(vb0123c3x0, va0c3x0);
138 int16x8_t vprod1x0123c3 = vmull_s8(vb0123c3x0, va1c3x0);
139 const int8x8_t vb0123c3x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
140 vprod0x0123c3 = vmlal_s8(vprod0x0123c3, vb0123c3x1, va0c3x1);
141 vprod1x0123c3 = vmlal_s8(vprod1x0123c3, vb0123c3x1, va1c3x1);
142 vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c3);
143 vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c3);
144 int16x8_t vprod0x4567c3 = vmull_s8(vb4567c3x0, va0c3x0);
145 int16x8_t vprod1x4567c3 = vmull_s8(vb4567c3x0, va1c3x0);
146 const int8x8_t vb4567c3x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
147 vprod0x4567c3 = vmlal_s8(vprod0x4567c3, vb4567c3x1, va0c3x1);
148 vprod1x4567c3 = vmlal_s8(vprod1x4567c3, vb4567c3x1, va1c3x1);
149 vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c3);
150 vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c3);
151
152 k -= 16 * sizeof(int8_t);
153 }
154
155 if (k >= 8 * sizeof(int8_t)) {
156 const int16x4x2_t va00 = vld2_dup_s16((const void*)a0);
157 const int16x4x2_t va01 = vld2_dup_s16((const void*)(a0 + 4)); a0 += 8;
158 const int16x4x2_t va10 = vld2_dup_s16((const void*)a1);
159 const int16x4x2_t va11 = vld2_dup_s16((const void*)(a1 + 4)); a1 += 8;
160
161 const int8x8_t vb0123c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
162 const int8x8_t vb4567c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
163 const int8x8_t vb0123c1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
164 const int8x8_t vb4567c1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
165 const int8x8_t vb0123c2 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
166 const int8x8_t vb4567c2 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
167 const int8x8_t vb0123c3 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
168 const int8x8_t vb4567c3 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
169
170 const int8x8_t va0c0 = vreinterpret_s8_s16(va00.val[0]);
171 const int8x8_t va1c0 = vreinterpret_s8_s16(va10.val[0]);
172
173 const int16x8_t vprod0x0123c0 = vmull_s8(vb0123c0, va0c0);
174 const int16x8_t vprod1x0123c0 = vmull_s8(vb0123c0, va1c0);
175 vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c0);
176 vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c0);
177 const int16x8_t vprod0x4567c0 = vmull_s8(vb4567c0, va0c0);
178 const int16x8_t vprod1x4567c0 = vmull_s8(vb4567c0, va1c0);
179 vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c0);
180 vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c0);
181 const int8x8_t va0c1 = vreinterpret_s8_s16(va00.val[1]);
182 const int8x8_t va1c1 = vreinterpret_s8_s16(va10.val[1]);
183
184 const int16x8_t vprod0x0123c1 = vmull_s8(vb0123c1, va0c1);
185 const int16x8_t vprod1x0123c1 = vmull_s8(vb0123c1, va1c1);
186 vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c1);
187 vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c1);
188 const int16x8_t vprod0x4567c1 = vmull_s8(vb4567c1, va0c1);
189 const int16x8_t vprod1x4567c1 = vmull_s8(vb4567c1, va1c1);
190 vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c1);
191 vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c1);
192 const int8x8_t va0c2 = vreinterpret_s8_s16(va01.val[0]);
193 const int8x8_t va1c2 = vreinterpret_s8_s16(va11.val[0]);
194
195 const int16x8_t vprod0x0123c2 = vmull_s8(vb0123c2, va0c2);
196 const int16x8_t vprod1x0123c2 = vmull_s8(vb0123c2, va1c2);
197 vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c2);
198 vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c2);
199 const int16x8_t vprod0x4567c2 = vmull_s8(vb4567c2, va0c2);
200 const int16x8_t vprod1x4567c2 = vmull_s8(vb4567c2, va1c2);
201 vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c2);
202 vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c2);
203 const int8x8_t va0c3 = vreinterpret_s8_s16(va01.val[1]);
204 const int8x8_t va1c3 = vreinterpret_s8_s16(va11.val[1]);
205
206 const int16x8_t vprod0x0123c3 = vmull_s8(vb0123c3, va0c3);
207 const int16x8_t vprod1x0123c3 = vmull_s8(vb0123c3, va1c3);
208 vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c3);
209 vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c3);
210 const int16x8_t vprod0x4567c3 = vmull_s8(vb4567c3, va0c3);
211 const int16x8_t vprod1x4567c3 = vmull_s8(vb4567c3, va1c3);
212 vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c3);
213 vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c3);
214
215 k -= 8 * sizeof(int8_t);
216 }
217
218 if XNN_UNLIKELY(k != 0) {
219 const int8x8_t va0 = vld1_s8(a0); a0 = (const int8_t*) ((uintptr_t) a0 + k);
220 const int8x8_t va1 = vld1_s8(a1); a1 = (const int8_t*) ((uintptr_t) a1 + k);
221
222 const int8x8_t vb0123c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
223 const int8x8_t vb4567c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
224
225 const int8x8_t va0c0 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0), 0));
226 const int16x8_t vprod0x0123c0 = vmull_s8(vb0123c0, va0c0);
227 vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c0);
228 const int16x8_t vprod0x4567c0 = vmull_s8(vb4567c0, va0c0);
229 vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c0);
230 const int8x8_t va1c0 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va1), 0));
231 const int16x8_t vprod1x0123c0 = vmull_s8(vb0123c0, va1c0);
232 vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c0);
233 const int16x8_t vprod1x4567c0 = vmull_s8(vb4567c0, va1c0);
234 vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c0);
235
236 if (k > 2 * sizeof(int8_t)) {
237 const int8x8_t vb0123c1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
238 const int8x8_t vb4567c1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
239
240 const int8x8_t va0c1 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0), 1));
241 const int16x8_t vprod0x0123c1 = vmull_s8(vb0123c1, va0c1);
242 vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c1);
243 const int16x8_t vprod0x4567c1 = vmull_s8(vb4567c1, va0c1);
244 vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c1);
245 const int8x8_t va1c1 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va1), 1));
246 const int16x8_t vprod1x0123c1 = vmull_s8(vb0123c1, va1c1);
247 vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c1);
248 const int16x8_t vprod1x4567c1 = vmull_s8(vb4567c1, va1c1);
249 vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c1);
250
251 if (k > 4 * sizeof(int8_t)) {
252 const int8x8_t vb0123c2 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
253 const int8x8_t vb4567c2 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
254
255 const int8x8_t va0c2 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0), 2));
256 const int16x8_t vprod0x0123c2 = vmull_s8(vb0123c2, va0c2);
257 vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c2);
258 const int16x8_t vprod0x4567c2 = vmull_s8(vb4567c2, va0c2);
259 vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c2);
260 const int8x8_t va1c2 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va1), 2));
261 const int16x8_t vprod1x0123c2 = vmull_s8(vb0123c2, va1c2);
262 vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c2);
263 const int16x8_t vprod1x4567c2 = vmull_s8(vb4567c2, va1c2);
264 vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c2);
265 }
266 }
267 }
268
269 float32x4_t vfpacc0x0123 = vcvtq_f32_s32(vacc0x0123);
270 float32x4_t vfpacc0x4567 = vcvtq_f32_s32(vacc0x4567);
271 float32x4_t vfpacc1x0123 = vcvtq_f32_s32(vacc1x0123);
272 float32x4_t vfpacc1x4567 = vcvtq_f32_s32(vacc1x4567);
273
274 const float32x4_t vscale0123 = vld1q_f32((const float*) w); w = (const void*) ((const float*) w + 4);
275 vfpacc0x0123 = vmulq_f32(vfpacc0x0123, vscale0123);
276 vfpacc1x0123 = vmulq_f32(vfpacc1x0123, vscale0123);
277 const float32x4_t vscale4567 = vld1q_f32((const float*) w); w = (const void*) ((const float*) w + 4);
278 vfpacc0x4567 = vmulq_f32(vfpacc0x4567, vscale4567);
279 vfpacc1x4567 = vmulq_f32(vfpacc1x4567, vscale4567);
280
281 vacc0x0123 = vcvtnq_s32_f32(vfpacc0x0123);
282 vacc0x4567 = vcvtnq_s32_f32(vfpacc0x4567);
283 vacc1x0123 = vcvtnq_s32_f32(vfpacc1x0123);
284 vacc1x4567 = vcvtnq_s32_f32(vfpacc1x4567);
285
286 const int16x8_t voutput_zero_point = vld1q_dup_s16(¶ms->fp32_neonv8.output_zero_point);
287 #if XNN_ARCH_ARM64
288 int16x8_t vacc0x01234567 = vqmovn_high_s32(vqmovn_s32(vacc0x0123), vacc0x4567);
289 int16x8_t vacc1x01234567 = vqmovn_high_s32(vqmovn_s32(vacc1x0123), vacc1x4567);
290
291 vacc0x01234567 = vqaddq_s16(vacc0x01234567, voutput_zero_point);
292 vacc1x01234567 = vqaddq_s16(vacc1x01234567, voutput_zero_point);
293
294 int8x16_t vout0x01234567_1x01234567 = vqmovn_high_s16(vqmovn_s16(vacc0x01234567), vacc1x01234567);
295 #else
296 int16x8_t vacc0x01234567 = vcombine_s16(vqmovn_s32(vacc0x0123), vqmovn_s32(vacc0x4567));
297 int16x8_t vacc1x01234567 = vcombine_s16(vqmovn_s32(vacc1x0123), vqmovn_s32(vacc1x4567));
298
299 vacc0x01234567 = vqaddq_s16(vacc0x01234567, voutput_zero_point);
300 vacc1x01234567 = vqaddq_s16(vacc1x01234567, voutput_zero_point);
301
302 int8x16_t vout0x01234567_1x01234567 = vcombine_s8(vqmovn_s16(vacc0x01234567), vqmovn_s16(vacc1x01234567));
303 #endif
304
305 const int8x16_t voutput_min = vld1q_dup_s8(¶ms->fp32_neonv8.output_min);
306 vout0x01234567_1x01234567 = vmaxq_s8(vout0x01234567_1x01234567, voutput_min);
307
308 const int8x16_t voutput_max = vld1q_dup_s8(¶ms->fp32_neonv8.output_max);
309 vout0x01234567_1x01234567 = vminq_s8(vout0x01234567_1x01234567, voutput_max);
310
311 if (nc >= 8) {
312 vst1_s8(c0 + 0, vget_low_s8(vout0x01234567_1x01234567));
313 vst1_s8(c1 + 0, vget_high_s8(vout0x01234567_1x01234567));
314
315 c0 = (int8_t*) ((uintptr_t) c0 + cn_stride);
316 c1 = (int8_t*) ((uintptr_t) c1 + cn_stride);
317
318 a0 = (const int8_t*) ((uintptr_t) a0 - kc);
319 a1 = (const int8_t*) ((uintptr_t) a1 - kc);
320
321 nc -= 8;
322 } else {
323 // Final case where not all of the 8 columns fit in the destination.
324 if (nc & 4) {
325 vst1q_lane_u32((void*) c0, vreinterpretq_u32_s8(vout0x01234567_1x01234567), 0); c0 += 4;
326 vst1q_lane_u32((void*) c1, vreinterpretq_u32_s8(vout0x01234567_1x01234567), 2); c1 += 4;
327 vout0x01234567_1x01234567 = vextq_s8(vout0x01234567_1x01234567, vout0x01234567_1x01234567, 4);
328 }
329 if (nc & 2) {
330 vst1q_lane_u16((void*) c0, vreinterpretq_u16_s8(vout0x01234567_1x01234567), 0); c0 += 2;
331 vst1q_lane_u16((void*) c1, vreinterpretq_u16_s8(vout0x01234567_1x01234567), 4); c1 += 2;
332 vout0x01234567_1x01234567 = vextq_s8(vout0x01234567_1x01234567, vout0x01234567_1x01234567, 2);
333 }
334 if (nc & 1) {
335 vst1q_lane_s8(c0, vout0x01234567_1x01234567, 0);
336 vst1q_lane_s8(c1, vout0x01234567_1x01234567, 8);
337 }
338
339 nc = 0;
340 }
341 } while (nc != 0);
342 }
343