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_1x16c8__neon_mlal(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_1x16c8__neon_mlal(
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 <= 1);
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
43 do {
44 int32x4_t vacc0x0 = vld1q_lane_s32(w, vmovq_n_s32(0), 0); w = (const void*) ((uintptr_t) w + sizeof(int32_t));
45 int32x4_t vacc0x1 = vld1q_lane_s32(w, vmovq_n_s32(0), 0); w = (const void*) ((uintptr_t) w + sizeof(int32_t));
46 int32x4_t vacc0x2 = vld1q_lane_s32(w, vmovq_n_s32(0), 0); w = (const void*) ((uintptr_t) w + sizeof(int32_t));
47 int32x4_t vacc0x3 = vld1q_lane_s32(w, vmovq_n_s32(0), 0); w = (const void*) ((uintptr_t) w + sizeof(int32_t));
48 int32x4_t vacc0x4 = vld1q_lane_s32(w, vmovq_n_s32(0), 0); w = (const void*) ((uintptr_t) w + sizeof(int32_t));
49 int32x4_t vacc0x5 = vld1q_lane_s32(w, vmovq_n_s32(0), 0); w = (const void*) ((uintptr_t) w + sizeof(int32_t));
50 int32x4_t vacc0x6 = vld1q_lane_s32(w, vmovq_n_s32(0), 0); w = (const void*) ((uintptr_t) w + sizeof(int32_t));
51 int32x4_t vacc0x7 = vld1q_lane_s32(w, vmovq_n_s32(0), 0); w = (const void*) ((uintptr_t) w + sizeof(int32_t));
52 int32x4_t vacc0x8 = vld1q_lane_s32(w, vmovq_n_s32(0), 0); w = (const void*) ((uintptr_t) w + sizeof(int32_t));
53 int32x4_t vacc0x9 = vld1q_lane_s32(w, vmovq_n_s32(0), 0); w = (const void*) ((uintptr_t) w + sizeof(int32_t));
54 int32x4_t vacc0x10 = vld1q_lane_s32(w, vmovq_n_s32(0), 0); w = (const void*) ((uintptr_t) w + sizeof(int32_t));
55 int32x4_t vacc0x11 = vld1q_lane_s32(w, vmovq_n_s32(0), 0); w = (const void*) ((uintptr_t) w + sizeof(int32_t));
56 int32x4_t vacc0x12 = vld1q_lane_s32(w, vmovq_n_s32(0), 0); w = (const void*) ((uintptr_t) w + sizeof(int32_t));
57 int32x4_t vacc0x13 = vld1q_lane_s32(w, vmovq_n_s32(0), 0); w = (const void*) ((uintptr_t) w + sizeof(int32_t));
58 int32x4_t vacc0x14 = vld1q_lane_s32(w, vmovq_n_s32(0), 0); w = (const void*) ((uintptr_t) w + sizeof(int32_t));
59 int32x4_t vacc0x15 = vld1q_lane_s32(w, vmovq_n_s32(0), 0); w = (const void*) ((uintptr_t) w + sizeof(int32_t));
60
61 size_t k = kc;
62 // 2x partial unrolled loop to load 16 bytes at a time using MLA.
63 while (k >= 16 * sizeof(int8_t)) {
64 const int8x8_t va0x0 = vld1_s8(a0); a0 += 8;
65 const int8x8_t va0x1 = vld1_s8(a0); a0 += 8;
66
67 const int8x8_t vb0x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof( int8_t));
68 const int8x8_t vb1x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof( int8_t));
69 const int8x8_t vb2x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof( int8_t));
70 const int8x8_t vb3x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof( int8_t));
71 const int8x8_t vb4x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof( int8_t));
72 const int8x8_t vb5x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof( int8_t));
73 const int8x8_t vb6x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof( int8_t));
74 const int8x8_t vb7x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof( int8_t));
75 const int8x8_t vb8x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof( int8_t));
76 const int8x8_t vb9x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof( int8_t));
77 const int8x8_t vb10x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof( int8_t));
78 const int8x8_t vb11x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof( int8_t));
79 const int8x8_t vb12x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof( int8_t));
80 const int8x8_t vb13x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof( int8_t));
81 const int8x8_t vb14x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof( int8_t));
82 const int8x8_t vb15x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof( int8_t));
83
84 const int8x8_t vb0x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof( int8_t));
85 int16x8_t vprod0x0 = vmull_s8(vb0x0, va0x0);
86 vprod0x0 = vmlal_s8(vprod0x0, vb0x1, va0x1);
87 vacc0x0 = vpadalq_s16(vacc0x0, vprod0x0);
88 const int8x8_t vb1x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof( int8_t));
89 int16x8_t vprod0x1 = vmull_s8(vb1x0, va0x0);
90 vprod0x1 = vmlal_s8(vprod0x1, vb1x1, va0x1);
91 vacc0x1 = vpadalq_s16(vacc0x1, vprod0x1);
92 const int8x8_t vb2x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof( int8_t));
93 int16x8_t vprod0x2 = vmull_s8(vb2x0, va0x0);
94 vprod0x2 = vmlal_s8(vprod0x2, vb2x1, va0x1);
95 vacc0x2 = vpadalq_s16(vacc0x2, vprod0x2);
96 const int8x8_t vb3x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof( int8_t));
97 int16x8_t vprod0x3 = vmull_s8(vb3x0, va0x0);
98 vprod0x3 = vmlal_s8(vprod0x3, vb3x1, va0x1);
99 vacc0x3 = vpadalq_s16(vacc0x3, vprod0x3);
100 const int8x8_t vb4x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof( int8_t));
101 int16x8_t vprod0x4 = vmull_s8(vb4x0, va0x0);
102 vprod0x4 = vmlal_s8(vprod0x4, vb4x1, va0x1);
103 vacc0x4 = vpadalq_s16(vacc0x4, vprod0x4);
104 const int8x8_t vb5x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof( int8_t));
105 int16x8_t vprod0x5 = vmull_s8(vb5x0, va0x0);
106 vprod0x5 = vmlal_s8(vprod0x5, vb5x1, va0x1);
107 vacc0x5 = vpadalq_s16(vacc0x5, vprod0x5);
108 const int8x8_t vb6x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof( int8_t));
109 int16x8_t vprod0x6 = vmull_s8(vb6x0, va0x0);
110 vprod0x6 = vmlal_s8(vprod0x6, vb6x1, va0x1);
111 vacc0x6 = vpadalq_s16(vacc0x6, vprod0x6);
112 const int8x8_t vb7x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof( int8_t));
113 int16x8_t vprod0x7 = vmull_s8(vb7x0, va0x0);
114 vprod0x7 = vmlal_s8(vprod0x7, vb7x1, va0x1);
115 vacc0x7 = vpadalq_s16(vacc0x7, vprod0x7);
116 const int8x8_t vb8x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof( int8_t));
117 int16x8_t vprod0x8 = vmull_s8(vb8x0, va0x0);
118 vprod0x8 = vmlal_s8(vprod0x8, vb8x1, va0x1);
119 vacc0x8 = vpadalq_s16(vacc0x8, vprod0x8);
120 const int8x8_t vb9x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof( int8_t));
121 int16x8_t vprod0x9 = vmull_s8(vb9x0, va0x0);
122 vprod0x9 = vmlal_s8(vprod0x9, vb9x1, va0x1);
123 vacc0x9 = vpadalq_s16(vacc0x9, vprod0x9);
124 const int8x8_t vb10x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof( int8_t));
125 int16x8_t vprod0x10 = vmull_s8(vb10x0, va0x0);
126 vprod0x10 = vmlal_s8(vprod0x10, vb10x1, va0x1);
127 vacc0x10 = vpadalq_s16(vacc0x10, vprod0x10);
128 const int8x8_t vb11x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof( int8_t));
129 int16x8_t vprod0x11 = vmull_s8(vb11x0, va0x0);
130 vprod0x11 = vmlal_s8(vprod0x11, vb11x1, va0x1);
131 vacc0x11 = vpadalq_s16(vacc0x11, vprod0x11);
132 const int8x8_t vb12x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof( int8_t));
133 int16x8_t vprod0x12 = vmull_s8(vb12x0, va0x0);
134 vprod0x12 = vmlal_s8(vprod0x12, vb12x1, va0x1);
135 vacc0x12 = vpadalq_s16(vacc0x12, vprod0x12);
136 const int8x8_t vb13x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof( int8_t));
137 int16x8_t vprod0x13 = vmull_s8(vb13x0, va0x0);
138 vprod0x13 = vmlal_s8(vprod0x13, vb13x1, va0x1);
139 vacc0x13 = vpadalq_s16(vacc0x13, vprod0x13);
140 const int8x8_t vb14x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof( int8_t));
141 int16x8_t vprod0x14 = vmull_s8(vb14x0, va0x0);
142 vprod0x14 = vmlal_s8(vprod0x14, vb14x1, va0x1);
143 vacc0x14 = vpadalq_s16(vacc0x14, vprod0x14);
144 const int8x8_t vb15x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof( int8_t));
145 int16x8_t vprod0x15 = vmull_s8(vb15x0, va0x0);
146 vprod0x15 = vmlal_s8(vprod0x15, vb15x1, va0x1);
147 vacc0x15 = vpadalq_s16(vacc0x15, vprod0x15);
148
149 k -= 16 * sizeof(int8_t);
150 }
151
152 // Handle 8 bytes at a time using MUL.
153 if (k != 0) {
154 const int8x8_t va0 = vld1_s8(a0); a0 += 8;
155
156 const int8x8_t vb0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
157 const int16x8_t vprod0x0 = vmull_s8(vb0, va0);
158 vacc0x0 = vpadalq_s16(vacc0x0, vprod0x0);
159 const int8x8_t vb1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
160 const int16x8_t vprod0x1 = vmull_s8(vb1, va0);
161 vacc0x1 = vpadalq_s16(vacc0x1, vprod0x1);
162 const int8x8_t vb2 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
163 const int16x8_t vprod0x2 = vmull_s8(vb2, va0);
164 vacc0x2 = vpadalq_s16(vacc0x2, vprod0x2);
165 const int8x8_t vb3 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
166 const int16x8_t vprod0x3 = vmull_s8(vb3, va0);
167 vacc0x3 = vpadalq_s16(vacc0x3, vprod0x3);
168 const int8x8_t vb4 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
169 const int16x8_t vprod0x4 = vmull_s8(vb4, va0);
170 vacc0x4 = vpadalq_s16(vacc0x4, vprod0x4);
171 const int8x8_t vb5 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
172 const int16x8_t vprod0x5 = vmull_s8(vb5, va0);
173 vacc0x5 = vpadalq_s16(vacc0x5, vprod0x5);
174 const int8x8_t vb6 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
175 const int16x8_t vprod0x6 = vmull_s8(vb6, va0);
176 vacc0x6 = vpadalq_s16(vacc0x6, vprod0x6);
177 const int8x8_t vb7 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
178 const int16x8_t vprod0x7 = vmull_s8(vb7, va0);
179 vacc0x7 = vpadalq_s16(vacc0x7, vprod0x7);
180 const int8x8_t vb8 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
181 const int16x8_t vprod0x8 = vmull_s8(vb8, va0);
182 vacc0x8 = vpadalq_s16(vacc0x8, vprod0x8);
183 const int8x8_t vb9 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
184 const int16x8_t vprod0x9 = vmull_s8(vb9, va0);
185 vacc0x9 = vpadalq_s16(vacc0x9, vprod0x9);
186 const int8x8_t vb10 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
187 const int16x8_t vprod0x10 = vmull_s8(vb10, va0);
188 vacc0x10 = vpadalq_s16(vacc0x10, vprod0x10);
189 const int8x8_t vb11 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
190 const int16x8_t vprod0x11 = vmull_s8(vb11, va0);
191 vacc0x11 = vpadalq_s16(vacc0x11, vprod0x11);
192 const int8x8_t vb12 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
193 const int16x8_t vprod0x12 = vmull_s8(vb12, va0);
194 vacc0x12 = vpadalq_s16(vacc0x12, vprod0x12);
195 const int8x8_t vb13 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
196 const int16x8_t vprod0x13 = vmull_s8(vb13, va0);
197 vacc0x13 = vpadalq_s16(vacc0x13, vprod0x13);
198 const int8x8_t vb14 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
199 const int16x8_t vprod0x14 = vmull_s8(vb14, va0);
200 vacc0x14 = vpadalq_s16(vacc0x14, vprod0x14);
201 const int8x8_t vb15 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
202 const int16x8_t vprod0x15 = vmull_s8(vb15, va0);
203 vacc0x15 = vpadalq_s16(vacc0x15, vprod0x15);
204
205 k -= 8 * sizeof(int8_t);
206 }
207
208 #if XNN_ARCH_ARM64
209 const int32x4_t vsum0x01 = vpaddq_s32(vacc0x0, vacc0x1);
210 const int32x4_t vsum0x23 = vpaddq_s32(vacc0x2, vacc0x3);
211 const int32x4_t vsum0x45 = vpaddq_s32(vacc0x4, vacc0x5);
212 const int32x4_t vsum0x67 = vpaddq_s32(vacc0x6, vacc0x7);
213 const int32x4_t vsum0x89 = vpaddq_s32(vacc0x8, vacc0x9);
214 const int32x4_t vsum0xAB = vpaddq_s32(vacc0x10, vacc0x11);
215 const int32x4_t vsum0xCD = vpaddq_s32(vacc0x12, vacc0x13);
216 const int32x4_t vsum0xEF = vpaddq_s32(vacc0x14, vacc0x15);
217
218 int32x4_t vacc0x0123 = vpaddq_s32(vsum0x01, vsum0x23);
219 int32x4_t vacc0x4567 = vpaddq_s32(vsum0x45, vsum0x67);
220 int32x4_t vacc0x89AB = vpaddq_s32(vsum0x89, vsum0xAB);
221 int32x4_t vacc0xCDEF = vpaddq_s32(vsum0xCD, vsum0xEF);
222 #else
223 const int32x2_t vpsum0x0 = vadd_s32(vget_low_s32(vacc0x0), vget_high_s32(vacc0x0));
224 const int32x2_t vpsum0x1 = vadd_s32(vget_low_s32(vacc0x1), vget_high_s32(vacc0x1));
225 const int32x2_t vpsum0x2 = vadd_s32(vget_low_s32(vacc0x2), vget_high_s32(vacc0x2));
226 const int32x2_t vpsum0x3 = vadd_s32(vget_low_s32(vacc0x3), vget_high_s32(vacc0x3));
227 const int32x2_t vsum0x01 = vpadd_s32(vpsum0x0, vpsum0x1);
228 const int32x2_t vsum0x23 = vpadd_s32(vpsum0x2, vpsum0x3);
229 int32x4_t vacc0x0123 = vcombine_s32(vsum0x01, vsum0x23 );
230 const int32x2_t vpsum0x4 = vadd_s32(vget_low_s32(vacc0x4), vget_high_s32(vacc0x4));
231 const int32x2_t vpsum0x5 = vadd_s32(vget_low_s32(vacc0x5), vget_high_s32(vacc0x5));
232 const int32x2_t vpsum0x6 = vadd_s32(vget_low_s32(vacc0x6), vget_high_s32(vacc0x6));
233 const int32x2_t vpsum0x7 = vadd_s32(vget_low_s32(vacc0x7), vget_high_s32(vacc0x7));
234 const int32x2_t vsum0x45 = vpadd_s32(vpsum0x4, vpsum0x5);
235 const int32x2_t vsum0x67 = vpadd_s32(vpsum0x6, vpsum0x7);
236 int32x4_t vacc0x4567 = vcombine_s32(vsum0x45, vsum0x67 );
237 const int32x2_t vpsum0x8 = vadd_s32(vget_low_s32(vacc0x8), vget_high_s32(vacc0x8));
238 const int32x2_t vpsum0x9 = vadd_s32(vget_low_s32(vacc0x9), vget_high_s32(vacc0x9));
239 const int32x2_t vpsum0xA = vadd_s32(vget_low_s32(vacc0x10), vget_high_s32(vacc0x10));
240 const int32x2_t vpsum0xB = vadd_s32(vget_low_s32(vacc0x11), vget_high_s32(vacc0x11));
241 const int32x2_t vsum0x89 = vpadd_s32(vpsum0x8, vpsum0x9);
242 const int32x2_t vsum0xAB = vpadd_s32(vpsum0xA, vpsum0xB);
243 int32x4_t vacc0x89AB = vcombine_s32(vsum0x89, vsum0xAB );
244 const int32x2_t vpsum0xC = vadd_s32(vget_low_s32(vacc0x12), vget_high_s32(vacc0x12));
245 const int32x2_t vpsum0xD = vadd_s32(vget_low_s32(vacc0x13), vget_high_s32(vacc0x13));
246 const int32x2_t vpsum0xE = vadd_s32(vget_low_s32(vacc0x14), vget_high_s32(vacc0x14));
247 const int32x2_t vpsum0xF = vadd_s32(vget_low_s32(vacc0x15), vget_high_s32(vacc0x15));
248 const int32x2_t vsum0xCD = vpadd_s32(vpsum0xC, vpsum0xD);
249 const int32x2_t vsum0xEF = vpadd_s32(vpsum0xE, vpsum0xF);
250 int32x4_t vacc0xCDEF = vcombine_s32(vsum0xCD, vsum0xEF );
251 #endif
252
253 const int32x4_t vright_pre_shift = vld1q_dup_s32(¶ms->rndnu_neon.right_pre_shift);
254 const int32x4_t vmultiplier = vld1q_dup_s32(¶ms->rndnu_neon.multiplier);
255 const int32x4_t vright_post_shift = vld1q_dup_s32(¶ms->rndnu_neon.right_post_shift);
256
257 vacc0x0123 = vqshlq_s32(vacc0x0123, vright_pre_shift);
258 vacc0x4567 = vqshlq_s32(vacc0x4567, vright_pre_shift);
259 vacc0x89AB = vqshlq_s32(vacc0x89AB, vright_pre_shift);
260 vacc0xCDEF = vqshlq_s32(vacc0xCDEF, vright_pre_shift);
261
262 vacc0x0123 = vqdmulhq_s32(vacc0x0123, vmultiplier);
263 vacc0x4567 = vqdmulhq_s32(vacc0x4567, vmultiplier);
264 vacc0x89AB = vqdmulhq_s32(vacc0x89AB, vmultiplier);
265 vacc0xCDEF = vqdmulhq_s32(vacc0xCDEF, vmultiplier);
266
267 vacc0x0123 = vrshlq_s32(vacc0x0123, vright_post_shift);
268 vacc0x4567 = vrshlq_s32(vacc0x4567, vright_post_shift);
269 vacc0x89AB = vrshlq_s32(vacc0x89AB, vright_post_shift);
270 vacc0xCDEF = vrshlq_s32(vacc0xCDEF, vright_post_shift);
271
272 const int16x8_t voutput_zero_point = vld1q_dup_s16(¶ms->rndnu_neon.output_zero_point);
273 #if XNN_ARCH_ARM64
274 int16x8_t vacc0x01234567 = vqmovn_high_s32(vqmovn_s32(vacc0x0123), vacc0x4567);
275 int16x8_t vacc0x89ABCDEF = vqmovn_high_s32(vqmovn_s32(vacc0x89AB), vacc0xCDEF);
276
277 vacc0x01234567 = vqaddq_s16(vacc0x01234567, voutput_zero_point);
278 vacc0x89ABCDEF = vqaddq_s16(vacc0x89ABCDEF, voutput_zero_point);
279
280 int8x16_t vout0x0123456789ABCDEF = vqmovn_high_s16(vqmovn_s16(vacc0x01234567), vacc0x89ABCDEF);
281 #else
282 int16x8_t vacc0x01234567 = vcombine_s16(vqmovn_s32(vacc0x0123), vqmovn_s32(vacc0x4567));
283 int16x8_t vacc0x89ABCDEF = vcombine_s16(vqmovn_s32(vacc0x89AB), vqmovn_s32(vacc0xCDEF));
284
285 vacc0x01234567 = vqaddq_s16(vacc0x01234567, voutput_zero_point);
286 vacc0x89ABCDEF = vqaddq_s16(vacc0x89ABCDEF, voutput_zero_point);
287
288 int8x16_t vout0x0123456789ABCDEF = vcombine_s8(vqmovn_s16(vacc0x01234567), vqmovn_s16(vacc0x89ABCDEF));
289 #endif
290
291 const int8x16_t voutput_min = vld1q_dup_s8(¶ms->rndnu_neon.output_min);
292 vout0x0123456789ABCDEF = vmaxq_s8(vout0x0123456789ABCDEF, voutput_min);
293
294 const int8x16_t voutput_max = vld1q_dup_s8(¶ms->rndnu_neon.output_max);
295 vout0x0123456789ABCDEF = vminq_s8(vout0x0123456789ABCDEF, voutput_max);
296
297 if (nc >= 16) {
298 vst1q_s8(c0 + 0, vout0x0123456789ABCDEF);
299
300 c0 = (int8_t*) ((uintptr_t) c0 + cn_stride);
301
302 a0 = (const int8_t*) ((uintptr_t) a0 - kc);
303
304 nc -= 16;
305 } else {
306 // Final case where not all of the 16 columns fit in the destination.
307 int8x8_t vout0x01234567 = vget_low_s8(vout0x0123456789ABCDEF);
308 if (nc & 8) {
309 vst1_s8(c0, vout0x01234567); c0 += 8;
310 vout0x01234567 = vget_high_s8(vout0x0123456789ABCDEF);
311 }
312 if (nc & 4) {
313 vst1_lane_u32((void*) c0, vreinterpret_u32_s8(vout0x01234567), 0); c0 += 4;
314 vout0x01234567 = vext_s8(vout0x01234567, vout0x01234567, 4);
315 }
316 if (nc & 2) {
317 vst1_lane_u16((void*) c0, vreinterpret_u16_s8(vout0x01234567), 0); c0 += 2;
318 vout0x01234567 = vext_s8(vout0x01234567, vout0x01234567, 2);
319 }
320 if (nc & 1) {
321 vst1_lane_s8(c0, vout0x01234567, 0);
322 }
323
324 nc = 0;
325 }
326 } while (nc != 0);
327 }
328