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/math.h>
16
xnn_qc8_gemm_minmax_fp32_ukernel_2x8c2__neon_mlal_ld1r(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)])17 void xnn_qc8_gemm_minmax_fp32_ukernel_2x8c2__neon_mlal_ld1r(
18 size_t mr,
19 size_t nc,
20 size_t kc,
21 const int8_t* restrict a,
22 size_t a_stride,
23 const void* restrict w,
24 int8_t* restrict c,
25 size_t cm_stride,
26 size_t cn_stride,
27 const union xnn_qc8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
28 {
29 assert(mr != 0);
30 assert(mr <= 2);
31 assert(nc != 0);
32 assert(kc != 0);
33 assert(kc % sizeof(int8_t) == 0);
34 assert(a != NULL);
35 assert(w != NULL);
36 assert(c != NULL);
37
38 kc = round_up_po2(kc, 2 * sizeof(int8_t));
39 const int8_t* a0 = a;
40 int8_t* c0 = c;
41 const int8_t* a1 = (const int8_t*) ((uintptr_t) a0 + a_stride);
42 int8_t* c1 = (int8_t*) ((uintptr_t) c0 + cm_stride);
43 if XNN_UNPREDICTABLE(mr != 2) {
44 a1 = a0;
45 c1 = c0;
46 }
47
48 do {
49 int32x4_t vacc0x0123 = vld1q_s32(w); w = (const void*) ((uintptr_t) w + 4 * sizeof(int32_t));
50 int32x4_t vacc0x4567 = vld1q_s32(w); w = (const void*) ((uintptr_t) w + 4 * sizeof(int32_t));
51 int32x4_t vacc1x0123 = vacc0x0123;
52 int32x4_t vacc1x4567 = vacc0x4567;
53
54 size_t k = kc;
55
56 while (k >= 16 * sizeof(int8_t)) {
57 const int16x4_t va00x0 = vld1_dup_s16((const void*)a0);
58 const int16x4_t va01x0 = vld1_dup_s16((const void*)(a0 + 2));
59 const int16x4_t va02x0 = vld1_dup_s16((const void*)(a0 + 4));
60 const int16x4_t va03x0 = vld1_dup_s16((const void*)(a0 + 6)); a0 += 8;
61 const int16x4_t va00x1 = vld1_dup_s16((const void*)a0);
62 const int16x4_t va01x1 = vld1_dup_s16((const void*)(a0 + 2));
63 const int16x4_t va02x1 = vld1_dup_s16((const void*)(a0 + 4));
64 const int16x4_t va03x1 = vld1_dup_s16((const void*)(a0 + 6)); a0 += 8;
65 const int16x4_t va10x0 = vld1_dup_s16((const void*)a1);
66 const int16x4_t va11x0 = vld1_dup_s16((const void*)(a1 + 2));
67 const int16x4_t va12x0 = vld1_dup_s16((const void*)(a1 + 4));
68 const int16x4_t va13x0 = vld1_dup_s16((const void*)(a1 + 6)); a1 += 8;
69 const int16x4_t va10x1 = vld1_dup_s16((const void*)a1);
70 const int16x4_t va11x1 = vld1_dup_s16((const void*)(a1 + 2));
71 const int16x4_t va12x1 = vld1_dup_s16((const void*)(a1 + 4));
72 const int16x4_t va13x1 = vld1_dup_s16((const void*)(a1 + 6)); a1 += 8;
73 const int8x8_t vb0123c0x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
74 const int8x8_t vb4567c0x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
75 const int8x8_t vb0123c1x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
76 const int8x8_t vb4567c1x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
77 const int8x8_t vb0123c2x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
78 const int8x8_t vb4567c2x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
79 const int8x8_t vb0123c3x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
80 const int8x8_t vb4567c3x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
81
82 const int8x8_t va0c0x0 = vreinterpret_s8_s16(va00x0);
83 const int8x8_t va0c0x1 = vreinterpret_s8_s16(va00x1);
84 const int8x8_t va1c0x0 = vreinterpret_s8_s16(va10x0);
85 const int8x8_t va1c0x1 = vreinterpret_s8_s16(va10x1);
86
87 int16x8_t vprod0x0123c0 = vmull_s8(vb0123c0x0, va0c0x0);
88 int16x8_t vprod1x0123c0 = vmull_s8(vb0123c0x0, va1c0x0);
89 const int8x8_t vb0123c0x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
90 vprod0x0123c0 = vmlal_s8(vprod0x0123c0, vb0123c0x1, va0c0x1);
91 vprod1x0123c0 = vmlal_s8(vprod1x0123c0, vb0123c0x1, va1c0x1);
92 vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c0);
93 vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c0);
94 int16x8_t vprod0x4567c0 = vmull_s8(vb4567c0x0, va0c0x0);
95 int16x8_t vprod1x4567c0 = vmull_s8(vb4567c0x0, va1c0x0);
96 const int8x8_t vb4567c0x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
97 vprod0x4567c0 = vmlal_s8(vprod0x4567c0, vb4567c0x1, va0c0x1);
98 vprod1x4567c0 = vmlal_s8(vprod1x4567c0, vb4567c0x1, va1c0x1);
99 vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c0);
100 vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c0);
101 const int8x8_t va0c1x0 = vreinterpret_s8_s16(va01x0);
102 const int8x8_t va0c1x1 = vreinterpret_s8_s16(va01x1);
103 const int8x8_t va1c1x0 = vreinterpret_s8_s16(va11x0);
104 const int8x8_t va1c1x1 = vreinterpret_s8_s16(va11x1);
105
106 int16x8_t vprod0x0123c1 = vmull_s8(vb0123c1x0, va0c1x0);
107 int16x8_t vprod1x0123c1 = vmull_s8(vb0123c1x0, va1c1x0);
108 const int8x8_t vb0123c1x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
109 vprod0x0123c1 = vmlal_s8(vprod0x0123c1, vb0123c1x1, va0c1x1);
110 vprod1x0123c1 = vmlal_s8(vprod1x0123c1, vb0123c1x1, va1c1x1);
111 vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c1);
112 vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c1);
113 int16x8_t vprod0x4567c1 = vmull_s8(vb4567c1x0, va0c1x0);
114 int16x8_t vprod1x4567c1 = vmull_s8(vb4567c1x0, va1c1x0);
115 const int8x8_t vb4567c1x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
116 vprod0x4567c1 = vmlal_s8(vprod0x4567c1, vb4567c1x1, va0c1x1);
117 vprod1x4567c1 = vmlal_s8(vprod1x4567c1, vb4567c1x1, va1c1x1);
118 vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c1);
119 vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c1);
120 const int8x8_t va0c2x0 = vreinterpret_s8_s16(va02x0);
121 const int8x8_t va0c2x1 = vreinterpret_s8_s16(va02x1);
122 const int8x8_t va1c2x0 = vreinterpret_s8_s16(va12x0);
123 const int8x8_t va1c2x1 = vreinterpret_s8_s16(va12x1);
124
125 int16x8_t vprod0x0123c2 = vmull_s8(vb0123c2x0, va0c2x0);
126 int16x8_t vprod1x0123c2 = vmull_s8(vb0123c2x0, va1c2x0);
127 const int8x8_t vb0123c2x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
128 vprod0x0123c2 = vmlal_s8(vprod0x0123c2, vb0123c2x1, va0c2x1);
129 vprod1x0123c2 = vmlal_s8(vprod1x0123c2, vb0123c2x1, va1c2x1);
130 vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c2);
131 vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c2);
132 int16x8_t vprod0x4567c2 = vmull_s8(vb4567c2x0, va0c2x0);
133 int16x8_t vprod1x4567c2 = vmull_s8(vb4567c2x0, va1c2x0);
134 const int8x8_t vb4567c2x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
135 vprod0x4567c2 = vmlal_s8(vprod0x4567c2, vb4567c2x1, va0c2x1);
136 vprod1x4567c2 = vmlal_s8(vprod1x4567c2, vb4567c2x1, va1c2x1);
137 vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c2);
138 vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c2);
139 const int8x8_t va0c3x0 = vreinterpret_s8_s16(va03x0);
140 const int8x8_t va0c3x1 = vreinterpret_s8_s16(va03x1);
141 const int8x8_t va1c3x0 = vreinterpret_s8_s16(va13x0);
142 const int8x8_t va1c3x1 = vreinterpret_s8_s16(va13x1);
143
144 int16x8_t vprod0x0123c3 = vmull_s8(vb0123c3x0, va0c3x0);
145 int16x8_t vprod1x0123c3 = vmull_s8(vb0123c3x0, va1c3x0);
146 const int8x8_t vb0123c3x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
147 vprod0x0123c3 = vmlal_s8(vprod0x0123c3, vb0123c3x1, va0c3x1);
148 vprod1x0123c3 = vmlal_s8(vprod1x0123c3, vb0123c3x1, va1c3x1);
149 vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c3);
150 vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c3);
151 int16x8_t vprod0x4567c3 = vmull_s8(vb4567c3x0, va0c3x0);
152 int16x8_t vprod1x4567c3 = vmull_s8(vb4567c3x0, va1c3x0);
153 const int8x8_t vb4567c3x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
154 vprod0x4567c3 = vmlal_s8(vprod0x4567c3, vb4567c3x1, va0c3x1);
155 vprod1x4567c3 = vmlal_s8(vprod1x4567c3, vb4567c3x1, va1c3x1);
156 vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c3);
157 vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c3);
158
159 k -= 16 * sizeof(int8_t);
160 }
161
162 if (k >= 8 * sizeof(int8_t)) {
163 const int16x4_t va00 = vld1_dup_s16((const void*)a0);
164 const int16x4_t va01 = vld1_dup_s16((const void*)(a0 + 2));
165 const int16x4_t va02 = vld1_dup_s16((const void*)(a0 + 4));
166 const int16x4_t va03 = vld1_dup_s16((const void*)(a0 + 6)); a0 += 8;
167 const int16x4_t va10 = vld1_dup_s16((const void*)a1);
168 const int16x4_t va11 = vld1_dup_s16((const void*)(a1 + 2));
169 const int16x4_t va12 = vld1_dup_s16((const void*)(a1 + 4));
170 const int16x4_t va13 = vld1_dup_s16((const void*)(a1 + 6)); a1 += 8;
171
172 const int8x8_t vb0123c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
173 const int8x8_t vb4567c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
174 const int8x8_t vb0123c1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
175 const int8x8_t vb4567c1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
176 const int8x8_t vb0123c2 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
177 const int8x8_t vb4567c2 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
178 const int8x8_t vb0123c3 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
179 const int8x8_t vb4567c3 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
180
181 const int8x8_t va0c0 = vreinterpret_s8_s16(va00);
182 const int8x8_t va1c0 = vreinterpret_s8_s16(va10);
183
184 const int16x8_t vprod0x0123c0 = vmull_s8(vb0123c0, va0c0);
185 const int16x8_t vprod1x0123c0 = vmull_s8(vb0123c0, va1c0);
186 vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c0);
187 vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c0);
188 const int16x8_t vprod0x4567c0 = vmull_s8(vb4567c0, va0c0);
189 const int16x8_t vprod1x4567c0 = vmull_s8(vb4567c0, va1c0);
190 vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c0);
191 vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c0);
192 const int8x8_t va0c1 = vreinterpret_s8_s16(va01);
193 const int8x8_t va1c1 = vreinterpret_s8_s16(va11);
194
195 const int16x8_t vprod0x0123c1 = vmull_s8(vb0123c1, va0c1);
196 const int16x8_t vprod1x0123c1 = vmull_s8(vb0123c1, va1c1);
197 vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c1);
198 vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c1);
199 const int16x8_t vprod0x4567c1 = vmull_s8(vb4567c1, va0c1);
200 const int16x8_t vprod1x4567c1 = vmull_s8(vb4567c1, va1c1);
201 vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c1);
202 vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c1);
203 const int8x8_t va0c2 = vreinterpret_s8_s16(va02);
204 const int8x8_t va1c2 = vreinterpret_s8_s16(va12);
205
206 const int16x8_t vprod0x0123c2 = vmull_s8(vb0123c2, va0c2);
207 const int16x8_t vprod1x0123c2 = vmull_s8(vb0123c2, va1c2);
208 vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c2);
209 vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c2);
210 const int16x8_t vprod0x4567c2 = vmull_s8(vb4567c2, va0c2);
211 const int16x8_t vprod1x4567c2 = vmull_s8(vb4567c2, va1c2);
212 vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c2);
213 vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c2);
214 const int8x8_t va0c3 = vreinterpret_s8_s16(va03);
215 const int8x8_t va1c3 = vreinterpret_s8_s16(va13);
216
217 const int16x8_t vprod0x0123c3 = vmull_s8(vb0123c3, va0c3);
218 const int16x8_t vprod1x0123c3 = vmull_s8(vb0123c3, va1c3);
219 vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c3);
220 vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c3);
221 const int16x8_t vprod0x4567c3 = vmull_s8(vb4567c3, va0c3);
222 const int16x8_t vprod1x4567c3 = vmull_s8(vb4567c3, va1c3);
223 vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c3);
224 vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c3);
225
226 k -= 8 * sizeof(int8_t);
227 }
228
229 if XNN_UNLIKELY(k != 0) {
230 const int8x8_t va0 = vld1_s8(a0); a0 = (const int8_t*) ((uintptr_t) a0 + k);
231 const int8x8_t va1 = vld1_s8(a1); a1 = (const int8_t*) ((uintptr_t) a1 + k);
232
233 const int8x8_t vb0123c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
234 const int8x8_t vb4567c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
235
236 const int8x8_t va0c0 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0), 0));
237 const int16x8_t vprod0x0123c0 = vmull_s8(vb0123c0, va0c0);
238 vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c0);
239 const int16x8_t vprod0x4567c0 = vmull_s8(vb4567c0, va0c0);
240 vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c0);
241 const int8x8_t va1c0 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va1), 0));
242 const int16x8_t vprod1x0123c0 = vmull_s8(vb0123c0, va1c0);
243 vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c0);
244 const int16x8_t vprod1x4567c0 = vmull_s8(vb4567c0, va1c0);
245 vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c0);
246
247 if (k > 2 * sizeof(int8_t)) {
248 const int8x8_t vb0123c1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
249 const int8x8_t vb4567c1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
250
251 const int8x8_t va0c1 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0), 1));
252 const int16x8_t vprod0x0123c1 = vmull_s8(vb0123c1, va0c1);
253 vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c1);
254 const int16x8_t vprod0x4567c1 = vmull_s8(vb4567c1, va0c1);
255 vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c1);
256 const int8x8_t va1c1 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va1), 1));
257 const int16x8_t vprod1x0123c1 = vmull_s8(vb0123c1, va1c1);
258 vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c1);
259 const int16x8_t vprod1x4567c1 = vmull_s8(vb4567c1, va1c1);
260 vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c1);
261
262 if (k > 4 * sizeof(int8_t)) {
263 const int8x8_t vb0123c2 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
264 const int8x8_t vb4567c2 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
265
266 const int8x8_t va0c2 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0), 2));
267 const int16x8_t vprod0x0123c2 = vmull_s8(vb0123c2, va0c2);
268 vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c2);
269 const int16x8_t vprod0x4567c2 = vmull_s8(vb4567c2, va0c2);
270 vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c2);
271 const int8x8_t va1c2 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va1), 2));
272 const int16x8_t vprod1x0123c2 = vmull_s8(vb0123c2, va1c2);
273 vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c2);
274 const int16x8_t vprod1x4567c2 = vmull_s8(vb4567c2, va1c2);
275 vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c2);
276 }
277 }
278 }
279
280 float32x4_t vfpacc0x0123 = vcvtq_f32_s32(vacc0x0123);
281 float32x4_t vfpacc0x4567 = vcvtq_f32_s32(vacc0x4567);
282 float32x4_t vfpacc1x0123 = vcvtq_f32_s32(vacc1x0123);
283 float32x4_t vfpacc1x4567 = vcvtq_f32_s32(vacc1x4567);
284
285 const float32x4_t vscale0123 = vld1q_f32((const float*) w); w = (const void*) ((const float*) w + 4);
286 vfpacc0x0123 = vmulq_f32(vfpacc0x0123, vscale0123);
287 vfpacc1x0123 = vmulq_f32(vfpacc1x0123, vscale0123);
288 const float32x4_t vscale4567 = vld1q_f32((const float*) w); w = (const void*) ((const float*) w + 4);
289 vfpacc0x4567 = vmulq_f32(vfpacc0x4567, vscale4567);
290 vfpacc1x4567 = vmulq_f32(vfpacc1x4567, vscale4567);
291
292 const float32x4_t vmagic_bias = vld1q_dup_f32(¶ms->fp32_neon.magic_bias);
293 vacc0x0123 = vreinterpretq_s32_f32(vaddq_f32(vfpacc0x0123, vmagic_bias));
294 vacc0x4567 = vreinterpretq_s32_f32(vaddq_f32(vfpacc0x4567, vmagic_bias));
295 vacc1x0123 = vreinterpretq_s32_f32(vaddq_f32(vfpacc1x0123, vmagic_bias));
296 vacc1x4567 = vreinterpretq_s32_f32(vaddq_f32(vfpacc1x4567, vmagic_bias));
297
298 const int32x4_t vmagic_bias_less_output_zero_point = vld1q_dup_s32(¶ms->fp32_neon.magic_bias_less_output_zero_point);
299 vacc0x0123 = vqsubq_s32(vacc0x0123, vmagic_bias_less_output_zero_point);
300 vacc0x4567 = vqsubq_s32(vacc0x4567, vmagic_bias_less_output_zero_point);
301 vacc1x0123 = vqsubq_s32(vacc1x0123, vmagic_bias_less_output_zero_point);
302 vacc1x4567 = vqsubq_s32(vacc1x4567, vmagic_bias_less_output_zero_point);
303
304 #if XNN_ARCH_ARM64
305 int16x8_t vacc0x01234567 = vqmovn_high_s32(vqmovn_s32(vacc0x0123), vacc0x4567);
306 int16x8_t vacc1x01234567 = vqmovn_high_s32(vqmovn_s32(vacc1x0123), vacc1x4567);
307
308
309 int8x16_t vout0x01234567_1x01234567 = vqmovn_high_s16(vqmovn_s16(vacc0x01234567), vacc1x01234567);
310 #else
311 int16x8_t vacc0x01234567 = vcombine_s16(vqmovn_s32(vacc0x0123), vqmovn_s32(vacc0x4567));
312 int16x8_t vacc1x01234567 = vcombine_s16(vqmovn_s32(vacc1x0123), vqmovn_s32(vacc1x4567));
313
314
315 int8x16_t vout0x01234567_1x01234567 = vcombine_s8(vqmovn_s16(vacc0x01234567), vqmovn_s16(vacc1x01234567));
316 #endif
317
318 const int8x16_t voutput_min = vld1q_dup_s8(¶ms->fp32_neon.output_min);
319 vout0x01234567_1x01234567 = vmaxq_s8(vout0x01234567_1x01234567, voutput_min);
320
321 const int8x16_t voutput_max = vld1q_dup_s8(¶ms->fp32_neon.output_max);
322 vout0x01234567_1x01234567 = vminq_s8(vout0x01234567_1x01234567, voutput_max);
323
324 if (nc >= 8) {
325 vst1_s8(c0 + 0, vget_low_s8(vout0x01234567_1x01234567));
326 vst1_s8(c1 + 0, vget_high_s8(vout0x01234567_1x01234567));
327
328 c0 = (int8_t*) ((uintptr_t) c0 + cn_stride);
329 c1 = (int8_t*) ((uintptr_t) c1 + cn_stride);
330
331 a0 = (const int8_t*) ((uintptr_t) a0 - kc);
332 a1 = (const int8_t*) ((uintptr_t) a1 - kc);
333
334 nc -= 8;
335 } else {
336 // Final case where not all of the 8 columns fit in the destination.
337 if (nc & 4) {
338 vst1q_lane_u32((void*) c0, vreinterpretq_u32_s8(vout0x01234567_1x01234567), 0); c0 += 4;
339 vst1q_lane_u32((void*) c1, vreinterpretq_u32_s8(vout0x01234567_1x01234567), 2); c1 += 4;
340 vout0x01234567_1x01234567 = vextq_s8(vout0x01234567_1x01234567, vout0x01234567_1x01234567, 4);
341 }
342 if (nc & 2) {
343 vst1q_lane_u16((void*) c0, vreinterpretq_u16_s8(vout0x01234567_1x01234567), 0); c0 += 2;
344 vst1q_lane_u16((void*) c1, vreinterpretq_u16_s8(vout0x01234567_1x01234567), 4); c1 += 2;
345 vout0x01234567_1x01234567 = vextq_s8(vout0x01234567_1x01234567, vout0x01234567_1x01234567, 2);
346 }
347 if (nc & 1) {
348 vst1q_lane_s8(c0, vout0x01234567_1x01234567, 0);
349 vst1q_lane_s8(c1, vout0x01234567_1x01234567, 8);
350 }
351
352 nc = 0;
353 }
354 } while (nc != 0);
355 }
356