1 // Auto-generated file. Do not edit!
2 // Template: src/qs8-gemm/c8-neon-mull.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
17
xnn_qs8_gemm_minmax_rndnu_ukernel_2x16c8__neon_mull(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_qs8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS (1)])18 void xnn_qs8_gemm_minmax_rndnu_ukernel_2x16c8__neon_mull(
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_qs8_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, 8 * 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 vacc0x0 = vld1q_lane_s32(w, vmovq_n_s32(0), 0); w = (const void*) ((uintptr_t) w + sizeof(int32_t));
51 int32x4_t vacc0x1 = vld1q_lane_s32(w, vmovq_n_s32(0), 0); w = (const void*) ((uintptr_t) w + sizeof(int32_t));
52 int32x4_t vacc0x2 = vld1q_lane_s32(w, vmovq_n_s32(0), 0); w = (const void*) ((uintptr_t) w + sizeof(int32_t));
53 int32x4_t vacc0x3 = vld1q_lane_s32(w, vmovq_n_s32(0), 0); w = (const void*) ((uintptr_t) w + sizeof(int32_t));
54 int32x4_t vacc0x4 = vld1q_lane_s32(w, vmovq_n_s32(0), 0); w = (const void*) ((uintptr_t) w + sizeof(int32_t));
55 int32x4_t vacc0x5 = vld1q_lane_s32(w, vmovq_n_s32(0), 0); w = (const void*) ((uintptr_t) w + sizeof(int32_t));
56 int32x4_t vacc0x6 = vld1q_lane_s32(w, vmovq_n_s32(0), 0); w = (const void*) ((uintptr_t) w + sizeof(int32_t));
57 int32x4_t vacc0x7 = vld1q_lane_s32(w, vmovq_n_s32(0), 0); w = (const void*) ((uintptr_t) w + sizeof(int32_t));
58 int32x4_t vacc0x8 = vld1q_lane_s32(w, vmovq_n_s32(0), 0); w = (const void*) ((uintptr_t) w + sizeof(int32_t));
59 int32x4_t vacc0x9 = vld1q_lane_s32(w, vmovq_n_s32(0), 0); w = (const void*) ((uintptr_t) w + sizeof(int32_t));
60 int32x4_t vacc0x10 = vld1q_lane_s32(w, vmovq_n_s32(0), 0); w = (const void*) ((uintptr_t) w + sizeof(int32_t));
61 int32x4_t vacc0x11 = vld1q_lane_s32(w, vmovq_n_s32(0), 0); w = (const void*) ((uintptr_t) w + sizeof(int32_t));
62 int32x4_t vacc0x12 = vld1q_lane_s32(w, vmovq_n_s32(0), 0); w = (const void*) ((uintptr_t) w + sizeof(int32_t));
63 int32x4_t vacc0x13 = vld1q_lane_s32(w, vmovq_n_s32(0), 0); w = (const void*) ((uintptr_t) w + sizeof(int32_t));
64 int32x4_t vacc0x14 = vld1q_lane_s32(w, vmovq_n_s32(0), 0); w = (const void*) ((uintptr_t) w + sizeof(int32_t));
65 int32x4_t vacc0x15 = vld1q_lane_s32(w, vmovq_n_s32(0), 0); w = (const void*) ((uintptr_t) w + sizeof(int32_t));
66 int32x4_t vacc1x0 = vacc0x0;
67 int32x4_t vacc1x1 = vacc0x1;
68 int32x4_t vacc1x2 = vacc0x2;
69 int32x4_t vacc1x3 = vacc0x3;
70 int32x4_t vacc1x4 = vacc0x4;
71 int32x4_t vacc1x5 = vacc0x5;
72 int32x4_t vacc1x6 = vacc0x6;
73 int32x4_t vacc1x7 = vacc0x7;
74 int32x4_t vacc1x8 = vacc0x8;
75 int32x4_t vacc1x9 = vacc0x9;
76 int32x4_t vacc1x10 = vacc0x10;
77 int32x4_t vacc1x11 = vacc0x11;
78 int32x4_t vacc1x12 = vacc0x12;
79 int32x4_t vacc1x13 = vacc0x13;
80 int32x4_t vacc1x14 = vacc0x14;
81 int32x4_t vacc1x15 = vacc0x15;
82
83 size_t k = kc;
84
85 // Handle 8 bytes at a time using MUL.
86 while (k != 0) {
87 const int8x8_t va0 = vld1_s8(a0); a0 += 8;
88 const int8x8_t va1 = vld1_s8(a1); a1 += 8;
89
90 const int8x8_t vb0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
91 const int16x8_t vprod0x0 = vmull_s8(vb0, va0);
92 const int16x8_t vprod1x0 = vmull_s8(vb0, va1);
93 vacc0x0 = vpadalq_s16(vacc0x0, vprod0x0);
94 vacc1x0 = vpadalq_s16(vacc1x0, vprod1x0);
95 const int8x8_t vb1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
96 const int16x8_t vprod0x1 = vmull_s8(vb1, va0);
97 const int16x8_t vprod1x1 = vmull_s8(vb1, va1);
98 vacc0x1 = vpadalq_s16(vacc0x1, vprod0x1);
99 vacc1x1 = vpadalq_s16(vacc1x1, vprod1x1);
100 const int8x8_t vb2 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
101 const int16x8_t vprod0x2 = vmull_s8(vb2, va0);
102 const int16x8_t vprod1x2 = vmull_s8(vb2, va1);
103 vacc0x2 = vpadalq_s16(vacc0x2, vprod0x2);
104 vacc1x2 = vpadalq_s16(vacc1x2, vprod1x2);
105 const int8x8_t vb3 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
106 const int16x8_t vprod0x3 = vmull_s8(vb3, va0);
107 const int16x8_t vprod1x3 = vmull_s8(vb3, va1);
108 vacc0x3 = vpadalq_s16(vacc0x3, vprod0x3);
109 vacc1x3 = vpadalq_s16(vacc1x3, vprod1x3);
110 const int8x8_t vb4 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
111 const int16x8_t vprod0x4 = vmull_s8(vb4, va0);
112 const int16x8_t vprod1x4 = vmull_s8(vb4, va1);
113 vacc0x4 = vpadalq_s16(vacc0x4, vprod0x4);
114 vacc1x4 = vpadalq_s16(vacc1x4, vprod1x4);
115 const int8x8_t vb5 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
116 const int16x8_t vprod0x5 = vmull_s8(vb5, va0);
117 const int16x8_t vprod1x5 = vmull_s8(vb5, va1);
118 vacc0x5 = vpadalq_s16(vacc0x5, vprod0x5);
119 vacc1x5 = vpadalq_s16(vacc1x5, vprod1x5);
120 const int8x8_t vb6 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
121 const int16x8_t vprod0x6 = vmull_s8(vb6, va0);
122 const int16x8_t vprod1x6 = vmull_s8(vb6, va1);
123 vacc0x6 = vpadalq_s16(vacc0x6, vprod0x6);
124 vacc1x6 = vpadalq_s16(vacc1x6, vprod1x6);
125 const int8x8_t vb7 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
126 const int16x8_t vprod0x7 = vmull_s8(vb7, va0);
127 const int16x8_t vprod1x7 = vmull_s8(vb7, va1);
128 vacc0x7 = vpadalq_s16(vacc0x7, vprod0x7);
129 vacc1x7 = vpadalq_s16(vacc1x7, vprod1x7);
130 const int8x8_t vb8 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
131 const int16x8_t vprod0x8 = vmull_s8(vb8, va0);
132 const int16x8_t vprod1x8 = vmull_s8(vb8, va1);
133 vacc0x8 = vpadalq_s16(vacc0x8, vprod0x8);
134 vacc1x8 = vpadalq_s16(vacc1x8, vprod1x8);
135 const int8x8_t vb9 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
136 const int16x8_t vprod0x9 = vmull_s8(vb9, va0);
137 const int16x8_t vprod1x9 = vmull_s8(vb9, va1);
138 vacc0x9 = vpadalq_s16(vacc0x9, vprod0x9);
139 vacc1x9 = vpadalq_s16(vacc1x9, vprod1x9);
140 const int8x8_t vb10 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
141 const int16x8_t vprod0x10 = vmull_s8(vb10, va0);
142 const int16x8_t vprod1x10 = vmull_s8(vb10, va1);
143 vacc0x10 = vpadalq_s16(vacc0x10, vprod0x10);
144 vacc1x10 = vpadalq_s16(vacc1x10, vprod1x10);
145 const int8x8_t vb11 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
146 const int16x8_t vprod0x11 = vmull_s8(vb11, va0);
147 const int16x8_t vprod1x11 = vmull_s8(vb11, va1);
148 vacc0x11 = vpadalq_s16(vacc0x11, vprod0x11);
149 vacc1x11 = vpadalq_s16(vacc1x11, vprod1x11);
150 const int8x8_t vb12 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
151 const int16x8_t vprod0x12 = vmull_s8(vb12, va0);
152 const int16x8_t vprod1x12 = vmull_s8(vb12, va1);
153 vacc0x12 = vpadalq_s16(vacc0x12, vprod0x12);
154 vacc1x12 = vpadalq_s16(vacc1x12, vprod1x12);
155 const int8x8_t vb13 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
156 const int16x8_t vprod0x13 = vmull_s8(vb13, va0);
157 const int16x8_t vprod1x13 = vmull_s8(vb13, va1);
158 vacc0x13 = vpadalq_s16(vacc0x13, vprod0x13);
159 vacc1x13 = vpadalq_s16(vacc1x13, vprod1x13);
160 const int8x8_t vb14 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
161 const int16x8_t vprod0x14 = vmull_s8(vb14, va0);
162 const int16x8_t vprod1x14 = vmull_s8(vb14, va1);
163 vacc0x14 = vpadalq_s16(vacc0x14, vprod0x14);
164 vacc1x14 = vpadalq_s16(vacc1x14, vprod1x14);
165 const int8x8_t vb15 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
166 const int16x8_t vprod0x15 = vmull_s8(vb15, va0);
167 const int16x8_t vprod1x15 = vmull_s8(vb15, va1);
168 vacc0x15 = vpadalq_s16(vacc0x15, vprod0x15);
169 vacc1x15 = vpadalq_s16(vacc1x15, vprod1x15);
170
171 k -= 8 * sizeof(int8_t);
172 }
173
174 #if XNN_ARCH_ARM64
175 const int32x4_t vsum0x01 = vpaddq_s32(vacc0x0, vacc0x1);
176 const int32x4_t vsum0x23 = vpaddq_s32(vacc0x2, vacc0x3);
177 const int32x4_t vsum0x45 = vpaddq_s32(vacc0x4, vacc0x5);
178 const int32x4_t vsum0x67 = vpaddq_s32(vacc0x6, vacc0x7);
179 const int32x4_t vsum0x89 = vpaddq_s32(vacc0x8, vacc0x9);
180 const int32x4_t vsum0xAB = vpaddq_s32(vacc0x10, vacc0x11);
181 const int32x4_t vsum0xCD = vpaddq_s32(vacc0x12, vacc0x13);
182 const int32x4_t vsum0xEF = vpaddq_s32(vacc0x14, vacc0x15);
183 const int32x4_t vsum1x01 = vpaddq_s32(vacc1x0, vacc1x1);
184 const int32x4_t vsum1x23 = vpaddq_s32(vacc1x2, vacc1x3);
185 const int32x4_t vsum1x45 = vpaddq_s32(vacc1x4, vacc1x5);
186 const int32x4_t vsum1x67 = vpaddq_s32(vacc1x6, vacc1x7);
187 const int32x4_t vsum1x89 = vpaddq_s32(vacc1x8, vacc1x9);
188 const int32x4_t vsum1xAB = vpaddq_s32(vacc1x10, vacc1x11);
189 const int32x4_t vsum1xCD = vpaddq_s32(vacc1x12, vacc1x13);
190 const int32x4_t vsum1xEF = vpaddq_s32(vacc1x14, vacc1x15);
191
192 int32x4_t vacc0x0123 = vpaddq_s32(vsum0x01, vsum0x23);
193 int32x4_t vacc0x4567 = vpaddq_s32(vsum0x45, vsum0x67);
194 int32x4_t vacc0x89AB = vpaddq_s32(vsum0x89, vsum0xAB);
195 int32x4_t vacc0xCDEF = vpaddq_s32(vsum0xCD, vsum0xEF);
196 int32x4_t vacc1x0123 = vpaddq_s32(vsum1x01, vsum1x23);
197 int32x4_t vacc1x4567 = vpaddq_s32(vsum1x45, vsum1x67);
198 int32x4_t vacc1x89AB = vpaddq_s32(vsum1x89, vsum1xAB);
199 int32x4_t vacc1xCDEF = vpaddq_s32(vsum1xCD, vsum1xEF);
200 #else
201 const int32x2_t vpsum0x0 = vadd_s32(vget_low_s32(vacc0x0), vget_high_s32(vacc0x0));
202 const int32x2_t vpsum0x1 = vadd_s32(vget_low_s32(vacc0x1), vget_high_s32(vacc0x1));
203 const int32x2_t vpsum0x2 = vadd_s32(vget_low_s32(vacc0x2), vget_high_s32(vacc0x2));
204 const int32x2_t vpsum0x3 = vadd_s32(vget_low_s32(vacc0x3), vget_high_s32(vacc0x3));
205 const int32x2_t vsum0x01 = vpadd_s32(vpsum0x0, vpsum0x1);
206 const int32x2_t vsum0x23 = vpadd_s32(vpsum0x2, vpsum0x3);
207 int32x4_t vacc0x0123 = vcombine_s32(vsum0x01, vsum0x23 );
208 const int32x2_t vpsum0x4 = vadd_s32(vget_low_s32(vacc0x4), vget_high_s32(vacc0x4));
209 const int32x2_t vpsum0x5 = vadd_s32(vget_low_s32(vacc0x5), vget_high_s32(vacc0x5));
210 const int32x2_t vpsum0x6 = vadd_s32(vget_low_s32(vacc0x6), vget_high_s32(vacc0x6));
211 const int32x2_t vpsum0x7 = vadd_s32(vget_low_s32(vacc0x7), vget_high_s32(vacc0x7));
212 const int32x2_t vsum0x45 = vpadd_s32(vpsum0x4, vpsum0x5);
213 const int32x2_t vsum0x67 = vpadd_s32(vpsum0x6, vpsum0x7);
214 int32x4_t vacc0x4567 = vcombine_s32(vsum0x45, vsum0x67 );
215 const int32x2_t vpsum0x8 = vadd_s32(vget_low_s32(vacc0x8), vget_high_s32(vacc0x8));
216 const int32x2_t vpsum0x9 = vadd_s32(vget_low_s32(vacc0x9), vget_high_s32(vacc0x9));
217 const int32x2_t vpsum0xA = vadd_s32(vget_low_s32(vacc0x10), vget_high_s32(vacc0x10));
218 const int32x2_t vpsum0xB = vadd_s32(vget_low_s32(vacc0x11), vget_high_s32(vacc0x11));
219 const int32x2_t vsum0x89 = vpadd_s32(vpsum0x8, vpsum0x9);
220 const int32x2_t vsum0xAB = vpadd_s32(vpsum0xA, vpsum0xB);
221 int32x4_t vacc0x89AB = vcombine_s32(vsum0x89, vsum0xAB );
222 const int32x2_t vpsum0xC = vadd_s32(vget_low_s32(vacc0x12), vget_high_s32(vacc0x12));
223 const int32x2_t vpsum0xD = vadd_s32(vget_low_s32(vacc0x13), vget_high_s32(vacc0x13));
224 const int32x2_t vpsum0xE = vadd_s32(vget_low_s32(vacc0x14), vget_high_s32(vacc0x14));
225 const int32x2_t vpsum0xF = vadd_s32(vget_low_s32(vacc0x15), vget_high_s32(vacc0x15));
226 const int32x2_t vsum0xCD = vpadd_s32(vpsum0xC, vpsum0xD);
227 const int32x2_t vsum0xEF = vpadd_s32(vpsum0xE, vpsum0xF);
228 int32x4_t vacc0xCDEF = vcombine_s32(vsum0xCD, vsum0xEF );
229 const int32x2_t vpsum1x0 = vadd_s32(vget_low_s32(vacc1x0), vget_high_s32(vacc1x0));
230 const int32x2_t vpsum1x1 = vadd_s32(vget_low_s32(vacc1x1), vget_high_s32(vacc1x1));
231 const int32x2_t vpsum1x2 = vadd_s32(vget_low_s32(vacc1x2), vget_high_s32(vacc1x2));
232 const int32x2_t vpsum1x3 = vadd_s32(vget_low_s32(vacc1x3), vget_high_s32(vacc1x3));
233 const int32x2_t vsum1x01 = vpadd_s32(vpsum1x0, vpsum1x1);
234 const int32x2_t vsum1x23 = vpadd_s32(vpsum1x2, vpsum1x3);
235 int32x4_t vacc1x0123 = vcombine_s32(vsum1x01, vsum1x23 );
236 const int32x2_t vpsum1x4 = vadd_s32(vget_low_s32(vacc1x4), vget_high_s32(vacc1x4));
237 const int32x2_t vpsum1x5 = vadd_s32(vget_low_s32(vacc1x5), vget_high_s32(vacc1x5));
238 const int32x2_t vpsum1x6 = vadd_s32(vget_low_s32(vacc1x6), vget_high_s32(vacc1x6));
239 const int32x2_t vpsum1x7 = vadd_s32(vget_low_s32(vacc1x7), vget_high_s32(vacc1x7));
240 const int32x2_t vsum1x45 = vpadd_s32(vpsum1x4, vpsum1x5);
241 const int32x2_t vsum1x67 = vpadd_s32(vpsum1x6, vpsum1x7);
242 int32x4_t vacc1x4567 = vcombine_s32(vsum1x45, vsum1x67 );
243 const int32x2_t vpsum1x8 = vadd_s32(vget_low_s32(vacc1x8), vget_high_s32(vacc1x8));
244 const int32x2_t vpsum1x9 = vadd_s32(vget_low_s32(vacc1x9), vget_high_s32(vacc1x9));
245 const int32x2_t vpsum1xA = vadd_s32(vget_low_s32(vacc1x10), vget_high_s32(vacc1x10));
246 const int32x2_t vpsum1xB = vadd_s32(vget_low_s32(vacc1x11), vget_high_s32(vacc1x11));
247 const int32x2_t vsum1x89 = vpadd_s32(vpsum1x8, vpsum1x9);
248 const int32x2_t vsum1xAB = vpadd_s32(vpsum1xA, vpsum1xB);
249 int32x4_t vacc1x89AB = vcombine_s32(vsum1x89, vsum1xAB );
250 const int32x2_t vpsum1xC = vadd_s32(vget_low_s32(vacc1x12), vget_high_s32(vacc1x12));
251 const int32x2_t vpsum1xD = vadd_s32(vget_low_s32(vacc1x13), vget_high_s32(vacc1x13));
252 const int32x2_t vpsum1xE = vadd_s32(vget_low_s32(vacc1x14), vget_high_s32(vacc1x14));
253 const int32x2_t vpsum1xF = vadd_s32(vget_low_s32(vacc1x15), vget_high_s32(vacc1x15));
254 const int32x2_t vsum1xCD = vpadd_s32(vpsum1xC, vpsum1xD);
255 const int32x2_t vsum1xEF = vpadd_s32(vpsum1xE, vpsum1xF);
256 int32x4_t vacc1xCDEF = vcombine_s32(vsum1xCD, vsum1xEF );
257 #endif
258
259 const int32x4_t vright_pre_shift = vld1q_dup_s32(¶ms->rndnu_neon.right_pre_shift);
260 const int32x4_t vmultiplier = vld1q_dup_s32(¶ms->rndnu_neon.multiplier);
261 const int32x4_t vright_post_shift = vld1q_dup_s32(¶ms->rndnu_neon.right_post_shift);
262
263 vacc0x0123 = vqshlq_s32(vacc0x0123, vright_pre_shift);
264 vacc0x4567 = vqshlq_s32(vacc0x4567, vright_pre_shift);
265 vacc0x89AB = vqshlq_s32(vacc0x89AB, vright_pre_shift);
266 vacc0xCDEF = vqshlq_s32(vacc0xCDEF, vright_pre_shift);
267 vacc1x0123 = vqshlq_s32(vacc1x0123, vright_pre_shift);
268 vacc1x4567 = vqshlq_s32(vacc1x4567, vright_pre_shift);
269 vacc1x89AB = vqshlq_s32(vacc1x89AB, vright_pre_shift);
270 vacc1xCDEF = vqshlq_s32(vacc1xCDEF, vright_pre_shift);
271
272 vacc0x0123 = vqdmulhq_s32(vacc0x0123, vmultiplier);
273 vacc0x4567 = vqdmulhq_s32(vacc0x4567, vmultiplier);
274 vacc0x89AB = vqdmulhq_s32(vacc0x89AB, vmultiplier);
275 vacc0xCDEF = vqdmulhq_s32(vacc0xCDEF, vmultiplier);
276 vacc1x0123 = vqdmulhq_s32(vacc1x0123, vmultiplier);
277 vacc1x4567 = vqdmulhq_s32(vacc1x4567, vmultiplier);
278 vacc1x89AB = vqdmulhq_s32(vacc1x89AB, vmultiplier);
279 vacc1xCDEF = vqdmulhq_s32(vacc1xCDEF, vmultiplier);
280
281 vacc0x0123 = vrshlq_s32(vacc0x0123, vright_post_shift);
282 vacc0x4567 = vrshlq_s32(vacc0x4567, vright_post_shift);
283 vacc0x89AB = vrshlq_s32(vacc0x89AB, vright_post_shift);
284 vacc0xCDEF = vrshlq_s32(vacc0xCDEF, vright_post_shift);
285 vacc1x0123 = vrshlq_s32(vacc1x0123, vright_post_shift);
286 vacc1x4567 = vrshlq_s32(vacc1x4567, vright_post_shift);
287 vacc1x89AB = vrshlq_s32(vacc1x89AB, vright_post_shift);
288 vacc1xCDEF = vrshlq_s32(vacc1xCDEF, vright_post_shift);
289
290 const int16x8_t voutput_zero_point = vld1q_dup_s16(¶ms->rndnu_neon.output_zero_point);
291 #if XNN_ARCH_ARM64
292 int16x8_t vacc0x01234567 = vqmovn_high_s32(vqmovn_s32(vacc0x0123), vacc0x4567);
293 int16x8_t vacc0x89ABCDEF = vqmovn_high_s32(vqmovn_s32(vacc0x89AB), vacc0xCDEF);
294 int16x8_t vacc1x01234567 = vqmovn_high_s32(vqmovn_s32(vacc1x0123), vacc1x4567);
295 int16x8_t vacc1x89ABCDEF = vqmovn_high_s32(vqmovn_s32(vacc1x89AB), vacc1xCDEF);
296
297 vacc0x01234567 = vqaddq_s16(vacc0x01234567, voutput_zero_point);
298 vacc0x89ABCDEF = vqaddq_s16(vacc0x89ABCDEF, voutput_zero_point);
299 vacc1x01234567 = vqaddq_s16(vacc1x01234567, voutput_zero_point);
300 vacc1x89ABCDEF = vqaddq_s16(vacc1x89ABCDEF, voutput_zero_point);
301
302 int8x16_t vout0x0123456789ABCDEF = vqmovn_high_s16(vqmovn_s16(vacc0x01234567), vacc0x89ABCDEF);
303 int8x16_t vout1x0123456789ABCDEF = vqmovn_high_s16(vqmovn_s16(vacc1x01234567), vacc1x89ABCDEF);
304 #else
305 int16x8_t vacc0x01234567 = vcombine_s16(vqmovn_s32(vacc0x0123), vqmovn_s32(vacc0x4567));
306 int16x8_t vacc0x89ABCDEF = vcombine_s16(vqmovn_s32(vacc0x89AB), vqmovn_s32(vacc0xCDEF));
307 int16x8_t vacc1x01234567 = vcombine_s16(vqmovn_s32(vacc1x0123), vqmovn_s32(vacc1x4567));
308 int16x8_t vacc1x89ABCDEF = vcombine_s16(vqmovn_s32(vacc1x89AB), vqmovn_s32(vacc1xCDEF));
309
310 vacc0x01234567 = vqaddq_s16(vacc0x01234567, voutput_zero_point);
311 vacc0x89ABCDEF = vqaddq_s16(vacc0x89ABCDEF, voutput_zero_point);
312 vacc1x01234567 = vqaddq_s16(vacc1x01234567, voutput_zero_point);
313 vacc1x89ABCDEF = vqaddq_s16(vacc1x89ABCDEF, voutput_zero_point);
314
315 int8x16_t vout0x0123456789ABCDEF = vcombine_s8(vqmovn_s16(vacc0x01234567), vqmovn_s16(vacc0x89ABCDEF));
316 int8x16_t vout1x0123456789ABCDEF = vcombine_s8(vqmovn_s16(vacc1x01234567), vqmovn_s16(vacc1x89ABCDEF));
317 #endif
318
319 const int8x16_t voutput_min = vld1q_dup_s8(¶ms->rndnu_neon.output_min);
320 vout0x0123456789ABCDEF = vmaxq_s8(vout0x0123456789ABCDEF, voutput_min);
321 vout1x0123456789ABCDEF = vmaxq_s8(vout1x0123456789ABCDEF, voutput_min);
322
323 const int8x16_t voutput_max = vld1q_dup_s8(¶ms->rndnu_neon.output_max);
324 vout0x0123456789ABCDEF = vminq_s8(vout0x0123456789ABCDEF, voutput_max);
325 vout1x0123456789ABCDEF = vminq_s8(vout1x0123456789ABCDEF, voutput_max);
326
327 if (nc >= 16) {
328 vst1q_s8(c0 + 0, vout0x0123456789ABCDEF);
329 vst1q_s8(c1 + 0, vout1x0123456789ABCDEF);
330
331 c0 = (int8_t*) ((uintptr_t) c0 + cn_stride);
332 c1 = (int8_t*) ((uintptr_t) c1 + cn_stride);
333
334 a0 = (const int8_t*) ((uintptr_t) a0 - kc);
335 a1 = (const int8_t*) ((uintptr_t) a1 - kc);
336
337 nc -= 16;
338 } else {
339 // Final case where not all of the 16 columns fit in the destination.
340 int8x16_t vout0x01234567_1x01234567 = vcombine_s8(vget_low_s8(vout0x0123456789ABCDEF), vget_low_s8(vout1x0123456789ABCDEF));
341 if (nc & 8) {
342 vst1_s8(c0, vget_low_s8(vout0x01234567_1x01234567)); c0 += 8;
343 vst1_s8(c1, vget_high_s8(vout0x01234567_1x01234567)); c1 += 8;
344 vout0x01234567_1x01234567 = vcombine_s8(vget_high_s8(vout0x0123456789ABCDEF), vget_high_s8(vout1x0123456789ABCDEF));
345 }
346 if (nc & 4) {
347 vst1q_lane_u32((void*) c0, vreinterpretq_u32_s8(vout0x01234567_1x01234567), 0); c0 += 4;
348 vst1q_lane_u32((void*) c1, vreinterpretq_u32_s8(vout0x01234567_1x01234567), 2); c1 += 4;
349 vout0x01234567_1x01234567 = vextq_s8(vout0x01234567_1x01234567, vout0x01234567_1x01234567, 4);
350 }
351 if (nc & 2) {
352 vst1q_lane_u16((void*) c0, vreinterpretq_u16_s8(vout0x01234567_1x01234567), 0); c0 += 2;
353 vst1q_lane_u16((void*) c1, vreinterpretq_u16_s8(vout0x01234567_1x01234567), 4); c1 += 2;
354 vout0x01234567_1x01234567 = vextq_s8(vout0x01234567_1x01234567, vout0x01234567_1x01234567, 2);
355 }
356 if (nc & 1) {
357 vst1q_lane_s8(c0, vout0x01234567_1x01234567, 0);
358 vst1q_lane_s8(c1, vout0x01234567_1x01234567, 8);
359 }
360
361 nc = 0;
362 }
363 } while (nc != 0);
364 }
365