1 // Auto-generated file. Do not edit!
2 // Template: src/qs8-igemm/c4-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
17
xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c4__neon_mlal_dup(size_t mr,size_t nc,size_t kc,size_t ks,const int8_t ** restrict a,const void * restrict w,int8_t * restrict c,size_t cm_stride,size_t cn_stride,size_t a_offset,const int8_t * zero,const union xnn_qs8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS (1)])18 void xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c4__neon_mlal_dup(
19 size_t mr,
20 size_t nc,
21 size_t kc,
22 size_t ks,
23 const int8_t** restrict a,
24 const void* restrict w,
25 int8_t* restrict c,
26 size_t cm_stride,
27 size_t cn_stride,
28 size_t a_offset,
29 const int8_t* zero,
30 const union xnn_qs8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
31 {
32 assert(mr != 0);
33 assert(mr <= 4);
34 assert(nc != 0);
35 assert(kc != 0);
36 assert(ks != 0);
37 assert(ks % (4 * sizeof(void*)) == 0);
38 assert(a_offset % sizeof(int8_t) == 0);
39 assert(a != NULL);
40 assert(w != NULL);
41 assert(c != NULL);
42
43 kc = round_up_po2(kc, 4 * sizeof(int8_t));
44 int8_t* c0 = c;
45 int8_t* c1 = (int8_t*) ((uintptr_t) c0 + cm_stride);
46 if XNN_UNPREDICTABLE(mr < 2) {
47 c1 = c0;
48 }
49 int8_t* c2 = (int8_t*) ((uintptr_t) c1 + cm_stride);
50 if XNN_UNPREDICTABLE(mr <= 2) {
51 c2 = c1;
52 }
53 int8_t* c3 = (int8_t*) ((uintptr_t) c2 + cm_stride);
54 if XNN_UNPREDICTABLE(mr != 4) {
55 c3 = c2;
56 }
57
58 do {
59 int32x4_t vacc0x01 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
60 int32x4_t vacc0x23 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
61 int32x4_t vacc0x45 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
62 int32x4_t vacc0x67 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
63 int32x4_t vacc1x01 = vacc0x01;
64 int32x4_t vacc1x23 = vacc0x23;
65 int32x4_t vacc1x45 = vacc0x45;
66 int32x4_t vacc1x67 = vacc0x67;
67 int32x4_t vacc2x01 = vacc0x01;
68 int32x4_t vacc2x23 = vacc0x23;
69 int32x4_t vacc2x45 = vacc0x45;
70 int32x4_t vacc2x67 = vacc0x67;
71 int32x4_t vacc3x01 = vacc0x01;
72 int32x4_t vacc3x23 = vacc0x23;
73 int32x4_t vacc3x45 = vacc0x45;
74 int32x4_t vacc3x67 = vacc0x67;
75
76 size_t p = ks;
77 do {
78 const int8_t* restrict a0 = a[0];
79 if XNN_UNPREDICTABLE(a0 != zero) {
80 a0 = (const int8_t*) ((uintptr_t) a0 + a_offset);
81 }
82 const int8_t* restrict a1 = a[1];
83 if XNN_UNPREDICTABLE(a1 != zero) {
84 a1 = (const int8_t*) ((uintptr_t) a1 + a_offset);
85 }
86 const int8_t* restrict a2 = a[2];
87 if XNN_UNPREDICTABLE(a2 != zero) {
88 a2 = (const int8_t*) ((uintptr_t) a2 + a_offset);
89 }
90 const int8_t* restrict a3 = a[3];
91 if XNN_UNPREDICTABLE(a3 != zero) {
92 a3 = (const int8_t*) ((uintptr_t) a3 + a_offset);
93 }
94 a += 4;
95
96 size_t k = kc;
97
98 while (k >= 16 * sizeof(int8_t)) {
99 const int8x8_t va0x0 = vld1_s8(a0); a0 += 8;
100 const int8x8_t va0x1 = vld1_s8(a0); a0 += 8;
101 const int8x8_t va1x0 = vld1_s8(a1); a1 += 8;
102 const int8x8_t va1x1 = vld1_s8(a1); a1 += 8;
103 const int8x8_t va2x0 = vld1_s8(a2); a2 += 8;
104 const int8x8_t va2x1 = vld1_s8(a2); a2 += 8;
105 const int8x8_t va3x0 = vld1_s8(a3); a3 += 8;
106 const int8x8_t va3x1 = vld1_s8(a3); a3 += 8;
107
108 const int8x8_t vb01c0x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
109 const int8x8_t vb23c0x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
110 const int8x8_t vb45c0x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
111 const int8x8_t vb67c0x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
112 const int8x8_t vb01c1x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
113 const int8x8_t vb23c1x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
114 const int8x8_t vb45c1x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
115 const int8x8_t vb67c1x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
116
117 const int8x8_t va0c0x0 = vreinterpret_s8_s32(vdup_lane_s32(vreinterpret_s32_s8(va0x0), 0));
118 const int8x8_t va0c0x1 = vreinterpret_s8_s32(vdup_lane_s32(vreinterpret_s32_s8(va0x1), 0));
119 const int8x8_t va1c0x0 = vreinterpret_s8_s32(vdup_lane_s32(vreinterpret_s32_s8(va1x0), 0));
120 const int8x8_t va1c0x1 = vreinterpret_s8_s32(vdup_lane_s32(vreinterpret_s32_s8(va1x1), 0));
121 const int8x8_t va2c0x0 = vreinterpret_s8_s32(vdup_lane_s32(vreinterpret_s32_s8(va2x0), 0));
122 const int8x8_t va2c0x1 = vreinterpret_s8_s32(vdup_lane_s32(vreinterpret_s32_s8(va2x1), 0));
123 const int8x8_t va3c0x0 = vreinterpret_s8_s32(vdup_lane_s32(vreinterpret_s32_s8(va3x0), 0));
124 const int8x8_t va3c0x1 = vreinterpret_s8_s32(vdup_lane_s32(vreinterpret_s32_s8(va3x1), 0));
125
126 int16x8_t vprod0x01c0 = vmull_s8(vb01c0x0, va0c0x0);
127 int16x8_t vprod1x01c0 = vmull_s8(vb01c0x0, va1c0x0);
128 int16x8_t vprod2x01c0 = vmull_s8(vb01c0x0, va2c0x0);
129 int16x8_t vprod3x01c0 = vmull_s8(vb01c0x0, va3c0x0);
130 const int8x8_t vb01c0x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
131 vprod0x01c0 = vmlal_s8(vprod0x01c0, vb01c0x1, va0c0x1);
132 vprod1x01c0 = vmlal_s8(vprod1x01c0, vb01c0x1, va1c0x1);
133 vprod2x01c0 = vmlal_s8(vprod2x01c0, vb01c0x1, va2c0x1);
134 vprod3x01c0 = vmlal_s8(vprod3x01c0, vb01c0x1, va3c0x1);
135 vacc0x01 = vpadalq_s16(vacc0x01, vprod0x01c0);
136 vacc1x01 = vpadalq_s16(vacc1x01, vprod1x01c0);
137 vacc2x01 = vpadalq_s16(vacc2x01, vprod2x01c0);
138 vacc3x01 = vpadalq_s16(vacc3x01, vprod3x01c0);
139 int16x8_t vprod0x23c0 = vmull_s8(vb23c0x0, va0c0x0);
140 int16x8_t vprod1x23c0 = vmull_s8(vb23c0x0, va1c0x0);
141 int16x8_t vprod2x23c0 = vmull_s8(vb23c0x0, va2c0x0);
142 int16x8_t vprod3x23c0 = vmull_s8(vb23c0x0, va3c0x0);
143 const int8x8_t vb23c0x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
144 vprod0x23c0 = vmlal_s8(vprod0x23c0, vb23c0x1, va0c0x1);
145 vprod1x23c0 = vmlal_s8(vprod1x23c0, vb23c0x1, va1c0x1);
146 vprod2x23c0 = vmlal_s8(vprod2x23c0, vb23c0x1, va2c0x1);
147 vprod3x23c0 = vmlal_s8(vprod3x23c0, vb23c0x1, va3c0x1);
148 vacc0x23 = vpadalq_s16(vacc0x23, vprod0x23c0);
149 vacc1x23 = vpadalq_s16(vacc1x23, vprod1x23c0);
150 vacc2x23 = vpadalq_s16(vacc2x23, vprod2x23c0);
151 vacc3x23 = vpadalq_s16(vacc3x23, vprod3x23c0);
152 int16x8_t vprod0x45c0 = vmull_s8(vb45c0x0, va0c0x0);
153 int16x8_t vprod1x45c0 = vmull_s8(vb45c0x0, va1c0x0);
154 int16x8_t vprod2x45c0 = vmull_s8(vb45c0x0, va2c0x0);
155 int16x8_t vprod3x45c0 = vmull_s8(vb45c0x0, va3c0x0);
156 const int8x8_t vb45c0x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
157 vprod0x45c0 = vmlal_s8(vprod0x45c0, vb45c0x1, va0c0x1);
158 vprod1x45c0 = vmlal_s8(vprod1x45c0, vb45c0x1, va1c0x1);
159 vprod2x45c0 = vmlal_s8(vprod2x45c0, vb45c0x1, va2c0x1);
160 vprod3x45c0 = vmlal_s8(vprod3x45c0, vb45c0x1, va3c0x1);
161 vacc0x45 = vpadalq_s16(vacc0x45, vprod0x45c0);
162 vacc1x45 = vpadalq_s16(vacc1x45, vprod1x45c0);
163 vacc2x45 = vpadalq_s16(vacc2x45, vprod2x45c0);
164 vacc3x45 = vpadalq_s16(vacc3x45, vprod3x45c0);
165 int16x8_t vprod0x67c0 = vmull_s8(vb67c0x0, va0c0x0);
166 int16x8_t vprod1x67c0 = vmull_s8(vb67c0x0, va1c0x0);
167 int16x8_t vprod2x67c0 = vmull_s8(vb67c0x0, va2c0x0);
168 int16x8_t vprod3x67c0 = vmull_s8(vb67c0x0, va3c0x0);
169 const int8x8_t vb67c0x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
170 vprod0x67c0 = vmlal_s8(vprod0x67c0, vb67c0x1, va0c0x1);
171 vprod1x67c0 = vmlal_s8(vprod1x67c0, vb67c0x1, va1c0x1);
172 vprod2x67c0 = vmlal_s8(vprod2x67c0, vb67c0x1, va2c0x1);
173 vprod3x67c0 = vmlal_s8(vprod3x67c0, vb67c0x1, va3c0x1);
174 vacc0x67 = vpadalq_s16(vacc0x67, vprod0x67c0);
175 vacc1x67 = vpadalq_s16(vacc1x67, vprod1x67c0);
176 vacc2x67 = vpadalq_s16(vacc2x67, vprod2x67c0);
177 vacc3x67 = vpadalq_s16(vacc3x67, vprod3x67c0);
178 const int8x8_t va0c1x0 = vreinterpret_s8_s32(vdup_lane_s32(vreinterpret_s32_s8(va0x0), 1));
179 const int8x8_t va0c1x1 = vreinterpret_s8_s32(vdup_lane_s32(vreinterpret_s32_s8(va0x1), 1));
180 const int8x8_t va1c1x0 = vreinterpret_s8_s32(vdup_lane_s32(vreinterpret_s32_s8(va1x0), 1));
181 const int8x8_t va1c1x1 = vreinterpret_s8_s32(vdup_lane_s32(vreinterpret_s32_s8(va1x1), 1));
182 const int8x8_t va2c1x0 = vreinterpret_s8_s32(vdup_lane_s32(vreinterpret_s32_s8(va2x0), 1));
183 const int8x8_t va2c1x1 = vreinterpret_s8_s32(vdup_lane_s32(vreinterpret_s32_s8(va2x1), 1));
184 const int8x8_t va3c1x0 = vreinterpret_s8_s32(vdup_lane_s32(vreinterpret_s32_s8(va3x0), 1));
185 const int8x8_t va3c1x1 = vreinterpret_s8_s32(vdup_lane_s32(vreinterpret_s32_s8(va3x1), 1));
186
187 int16x8_t vprod0x01c1 = vmull_s8(vb01c1x0, va0c1x0);
188 int16x8_t vprod1x01c1 = vmull_s8(vb01c1x0, va1c1x0);
189 int16x8_t vprod2x01c1 = vmull_s8(vb01c1x0, va2c1x0);
190 int16x8_t vprod3x01c1 = vmull_s8(vb01c1x0, va3c1x0);
191 const int8x8_t vb01c1x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
192 vprod0x01c1 = vmlal_s8(vprod0x01c1, vb01c1x1, va0c1x1);
193 vprod1x01c1 = vmlal_s8(vprod1x01c1, vb01c1x1, va1c1x1);
194 vprod2x01c1 = vmlal_s8(vprod2x01c1, vb01c1x1, va2c1x1);
195 vprod3x01c1 = vmlal_s8(vprod3x01c1, vb01c1x1, va3c1x1);
196 vacc0x01 = vpadalq_s16(vacc0x01, vprod0x01c1);
197 vacc1x01 = vpadalq_s16(vacc1x01, vprod1x01c1);
198 vacc2x01 = vpadalq_s16(vacc2x01, vprod2x01c1);
199 vacc3x01 = vpadalq_s16(vacc3x01, vprod3x01c1);
200 int16x8_t vprod0x23c1 = vmull_s8(vb23c1x0, va0c1x0);
201 int16x8_t vprod1x23c1 = vmull_s8(vb23c1x0, va1c1x0);
202 int16x8_t vprod2x23c1 = vmull_s8(vb23c1x0, va2c1x0);
203 int16x8_t vprod3x23c1 = vmull_s8(vb23c1x0, va3c1x0);
204 const int8x8_t vb23c1x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
205 vprod0x23c1 = vmlal_s8(vprod0x23c1, vb23c1x1, va0c1x1);
206 vprod1x23c1 = vmlal_s8(vprod1x23c1, vb23c1x1, va1c1x1);
207 vprod2x23c1 = vmlal_s8(vprod2x23c1, vb23c1x1, va2c1x1);
208 vprod3x23c1 = vmlal_s8(vprod3x23c1, vb23c1x1, va3c1x1);
209 vacc0x23 = vpadalq_s16(vacc0x23, vprod0x23c1);
210 vacc1x23 = vpadalq_s16(vacc1x23, vprod1x23c1);
211 vacc2x23 = vpadalq_s16(vacc2x23, vprod2x23c1);
212 vacc3x23 = vpadalq_s16(vacc3x23, vprod3x23c1);
213 int16x8_t vprod0x45c1 = vmull_s8(vb45c1x0, va0c1x0);
214 int16x8_t vprod1x45c1 = vmull_s8(vb45c1x0, va1c1x0);
215 int16x8_t vprod2x45c1 = vmull_s8(vb45c1x0, va2c1x0);
216 int16x8_t vprod3x45c1 = vmull_s8(vb45c1x0, va3c1x0);
217 const int8x8_t vb45c1x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
218 vprod0x45c1 = vmlal_s8(vprod0x45c1, vb45c1x1, va0c1x1);
219 vprod1x45c1 = vmlal_s8(vprod1x45c1, vb45c1x1, va1c1x1);
220 vprod2x45c1 = vmlal_s8(vprod2x45c1, vb45c1x1, va2c1x1);
221 vprod3x45c1 = vmlal_s8(vprod3x45c1, vb45c1x1, va3c1x1);
222 vacc0x45 = vpadalq_s16(vacc0x45, vprod0x45c1);
223 vacc1x45 = vpadalq_s16(vacc1x45, vprod1x45c1);
224 vacc2x45 = vpadalq_s16(vacc2x45, vprod2x45c1);
225 vacc3x45 = vpadalq_s16(vacc3x45, vprod3x45c1);
226 int16x8_t vprod0x67c1 = vmull_s8(vb67c1x0, va0c1x0);
227 int16x8_t vprod1x67c1 = vmull_s8(vb67c1x0, va1c1x0);
228 int16x8_t vprod2x67c1 = vmull_s8(vb67c1x0, va2c1x0);
229 int16x8_t vprod3x67c1 = vmull_s8(vb67c1x0, va3c1x0);
230 const int8x8_t vb67c1x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
231 vprod0x67c1 = vmlal_s8(vprod0x67c1, vb67c1x1, va0c1x1);
232 vprod1x67c1 = vmlal_s8(vprod1x67c1, vb67c1x1, va1c1x1);
233 vprod2x67c1 = vmlal_s8(vprod2x67c1, vb67c1x1, va2c1x1);
234 vprod3x67c1 = vmlal_s8(vprod3x67c1, vb67c1x1, va3c1x1);
235 vacc0x67 = vpadalq_s16(vacc0x67, vprod0x67c1);
236 vacc1x67 = vpadalq_s16(vacc1x67, vprod1x67c1);
237 vacc2x67 = vpadalq_s16(vacc2x67, vprod2x67c1);
238 vacc3x67 = vpadalq_s16(vacc3x67, vprod3x67c1);
239
240 k -= 16 * sizeof(int8_t);
241 }
242
243 if (k >= 8 * sizeof(int8_t)) {
244 const int8x8_t va0 = vld1_s8(a0); a0 += 8;
245 const int8x8_t va1 = vld1_s8(a1); a1 += 8;
246 const int8x8_t va2 = vld1_s8(a2); a2 += 8;
247 const int8x8_t va3 = vld1_s8(a3); a3 += 8;
248
249 const int8x8_t vb01c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
250 const int8x8_t vb23c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
251 const int8x8_t vb45c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
252 const int8x8_t vb67c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
253 const int8x8_t vb01c1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
254 const int8x8_t vb23c1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
255 const int8x8_t vb45c1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
256 const int8x8_t vb67c1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
257
258 const int8x8_t va0c0 = vreinterpret_s8_s32(vdup_lane_s32(vreinterpret_s32_s8(va0), 0));
259 const int8x8_t va1c0 = vreinterpret_s8_s32(vdup_lane_s32(vreinterpret_s32_s8(va1), 0));
260 const int8x8_t va2c0 = vreinterpret_s8_s32(vdup_lane_s32(vreinterpret_s32_s8(va2), 0));
261 const int8x8_t va3c0 = vreinterpret_s8_s32(vdup_lane_s32(vreinterpret_s32_s8(va3), 0));
262
263 const int16x8_t vprod0x01c0 = vmull_s8(vb01c0, va0c0);
264 const int16x8_t vprod1x01c0 = vmull_s8(vb01c0, va1c0);
265 const int16x8_t vprod2x01c0 = vmull_s8(vb01c0, va2c0);
266 const int16x8_t vprod3x01c0 = vmull_s8(vb01c0, va3c0);
267 vacc0x01 = vpadalq_s16(vacc0x01, vprod0x01c0);
268 vacc1x01 = vpadalq_s16(vacc1x01, vprod1x01c0);
269 vacc2x01 = vpadalq_s16(vacc2x01, vprod2x01c0);
270 vacc3x01 = vpadalq_s16(vacc3x01, vprod3x01c0);
271 const int16x8_t vprod0x23c0 = vmull_s8(vb23c0, va0c0);
272 const int16x8_t vprod1x23c0 = vmull_s8(vb23c0, va1c0);
273 const int16x8_t vprod2x23c0 = vmull_s8(vb23c0, va2c0);
274 const int16x8_t vprod3x23c0 = vmull_s8(vb23c0, va3c0);
275 vacc0x23 = vpadalq_s16(vacc0x23, vprod0x23c0);
276 vacc1x23 = vpadalq_s16(vacc1x23, vprod1x23c0);
277 vacc2x23 = vpadalq_s16(vacc2x23, vprod2x23c0);
278 vacc3x23 = vpadalq_s16(vacc3x23, vprod3x23c0);
279 const int16x8_t vprod0x45c0 = vmull_s8(vb45c0, va0c0);
280 const int16x8_t vprod1x45c0 = vmull_s8(vb45c0, va1c0);
281 const int16x8_t vprod2x45c0 = vmull_s8(vb45c0, va2c0);
282 const int16x8_t vprod3x45c0 = vmull_s8(vb45c0, va3c0);
283 vacc0x45 = vpadalq_s16(vacc0x45, vprod0x45c0);
284 vacc1x45 = vpadalq_s16(vacc1x45, vprod1x45c0);
285 vacc2x45 = vpadalq_s16(vacc2x45, vprod2x45c0);
286 vacc3x45 = vpadalq_s16(vacc3x45, vprod3x45c0);
287 const int16x8_t vprod0x67c0 = vmull_s8(vb67c0, va0c0);
288 const int16x8_t vprod1x67c0 = vmull_s8(vb67c0, va1c0);
289 const int16x8_t vprod2x67c0 = vmull_s8(vb67c0, va2c0);
290 const int16x8_t vprod3x67c0 = vmull_s8(vb67c0, va3c0);
291 vacc0x67 = vpadalq_s16(vacc0x67, vprod0x67c0);
292 vacc1x67 = vpadalq_s16(vacc1x67, vprod1x67c0);
293 vacc2x67 = vpadalq_s16(vacc2x67, vprod2x67c0);
294 vacc3x67 = vpadalq_s16(vacc3x67, vprod3x67c0);
295 const int8x8_t va0c1 = vreinterpret_s8_s32(vdup_lane_s32(vreinterpret_s32_s8(va0), 1));
296 const int8x8_t va1c1 = vreinterpret_s8_s32(vdup_lane_s32(vreinterpret_s32_s8(va1), 1));
297 const int8x8_t va2c1 = vreinterpret_s8_s32(vdup_lane_s32(vreinterpret_s32_s8(va2), 1));
298 const int8x8_t va3c1 = vreinterpret_s8_s32(vdup_lane_s32(vreinterpret_s32_s8(va3), 1));
299
300 const int16x8_t vprod0x01c1 = vmull_s8(vb01c1, va0c1);
301 const int16x8_t vprod1x01c1 = vmull_s8(vb01c1, va1c1);
302 const int16x8_t vprod2x01c1 = vmull_s8(vb01c1, va2c1);
303 const int16x8_t vprod3x01c1 = vmull_s8(vb01c1, va3c1);
304 vacc0x01 = vpadalq_s16(vacc0x01, vprod0x01c1);
305 vacc1x01 = vpadalq_s16(vacc1x01, vprod1x01c1);
306 vacc2x01 = vpadalq_s16(vacc2x01, vprod2x01c1);
307 vacc3x01 = vpadalq_s16(vacc3x01, vprod3x01c1);
308 const int16x8_t vprod0x23c1 = vmull_s8(vb23c1, va0c1);
309 const int16x8_t vprod1x23c1 = vmull_s8(vb23c1, va1c1);
310 const int16x8_t vprod2x23c1 = vmull_s8(vb23c1, va2c1);
311 const int16x8_t vprod3x23c1 = vmull_s8(vb23c1, va3c1);
312 vacc0x23 = vpadalq_s16(vacc0x23, vprod0x23c1);
313 vacc1x23 = vpadalq_s16(vacc1x23, vprod1x23c1);
314 vacc2x23 = vpadalq_s16(vacc2x23, vprod2x23c1);
315 vacc3x23 = vpadalq_s16(vacc3x23, vprod3x23c1);
316 const int16x8_t vprod0x45c1 = vmull_s8(vb45c1, va0c1);
317 const int16x8_t vprod1x45c1 = vmull_s8(vb45c1, va1c1);
318 const int16x8_t vprod2x45c1 = vmull_s8(vb45c1, va2c1);
319 const int16x8_t vprod3x45c1 = vmull_s8(vb45c1, va3c1);
320 vacc0x45 = vpadalq_s16(vacc0x45, vprod0x45c1);
321 vacc1x45 = vpadalq_s16(vacc1x45, vprod1x45c1);
322 vacc2x45 = vpadalq_s16(vacc2x45, vprod2x45c1);
323 vacc3x45 = vpadalq_s16(vacc3x45, vprod3x45c1);
324 const int16x8_t vprod0x67c1 = vmull_s8(vb67c1, va0c1);
325 const int16x8_t vprod1x67c1 = vmull_s8(vb67c1, va1c1);
326 const int16x8_t vprod2x67c1 = vmull_s8(vb67c1, va2c1);
327 const int16x8_t vprod3x67c1 = vmull_s8(vb67c1, va3c1);
328 vacc0x67 = vpadalq_s16(vacc0x67, vprod0x67c1);
329 vacc1x67 = vpadalq_s16(vacc1x67, vprod1x67c1);
330 vacc2x67 = vpadalq_s16(vacc2x67, vprod2x67c1);
331 vacc3x67 = vpadalq_s16(vacc3x67, vprod3x67c1);
332
333 k -= 8 * sizeof(int8_t);
334 }
335
336 if XNN_UNLIKELY(k != 0) {
337 const int8x8_t va0 = vld1_s8(a0); a0 = (const int8_t*) ((uintptr_t) a0 + k);
338 const int8x8_t va1 = vld1_s8(a1); a1 = (const int8_t*) ((uintptr_t) a1 + k);
339 const int8x8_t va2 = vld1_s8(a2); a2 = (const int8_t*) ((uintptr_t) a2 + k);
340 const int8x8_t va3 = vld1_s8(a3); a3 = (const int8_t*) ((uintptr_t) a3 + k);
341
342 const int8x8_t vb01c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
343 const int8x8_t vb23c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
344 const int8x8_t vb45c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
345 const int8x8_t vb67c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
346
347 const int8x8_t va0c0 = vreinterpret_s8_s32(vdup_lane_s32(vreinterpret_s32_s8(va0), 0));
348 const int16x8_t vprod0x01c0 = vmull_s8(vb01c0, va0c0);
349 vacc0x01 = vpadalq_s16(vacc0x01, vprod0x01c0);
350 const int16x8_t vprod0x23c0 = vmull_s8(vb23c0, va0c0);
351 vacc0x23 = vpadalq_s16(vacc0x23, vprod0x23c0);
352 const int16x8_t vprod0x45c0 = vmull_s8(vb45c0, va0c0);
353 vacc0x45 = vpadalq_s16(vacc0x45, vprod0x45c0);
354 const int16x8_t vprod0x67c0 = vmull_s8(vb67c0, va0c0);
355 vacc0x67 = vpadalq_s16(vacc0x67, vprod0x67c0);
356 const int8x8_t va1c0 = vreinterpret_s8_s32(vdup_lane_s32(vreinterpret_s32_s8(va1), 0));
357 const int16x8_t vprod1x01c0 = vmull_s8(vb01c0, va1c0);
358 vacc1x01 = vpadalq_s16(vacc1x01, vprod1x01c0);
359 const int16x8_t vprod1x23c0 = vmull_s8(vb23c0, va1c0);
360 vacc1x23 = vpadalq_s16(vacc1x23, vprod1x23c0);
361 const int16x8_t vprod1x45c0 = vmull_s8(vb45c0, va1c0);
362 vacc1x45 = vpadalq_s16(vacc1x45, vprod1x45c0);
363 const int16x8_t vprod1x67c0 = vmull_s8(vb67c0, va1c0);
364 vacc1x67 = vpadalq_s16(vacc1x67, vprod1x67c0);
365 const int8x8_t va2c0 = vreinterpret_s8_s32(vdup_lane_s32(vreinterpret_s32_s8(va2), 0));
366 const int16x8_t vprod2x01c0 = vmull_s8(vb01c0, va2c0);
367 vacc2x01 = vpadalq_s16(vacc2x01, vprod2x01c0);
368 const int16x8_t vprod2x23c0 = vmull_s8(vb23c0, va2c0);
369 vacc2x23 = vpadalq_s16(vacc2x23, vprod2x23c0);
370 const int16x8_t vprod2x45c0 = vmull_s8(vb45c0, va2c0);
371 vacc2x45 = vpadalq_s16(vacc2x45, vprod2x45c0);
372 const int16x8_t vprod2x67c0 = vmull_s8(vb67c0, va2c0);
373 vacc2x67 = vpadalq_s16(vacc2x67, vprod2x67c0);
374 const int8x8_t va3c0 = vreinterpret_s8_s32(vdup_lane_s32(vreinterpret_s32_s8(va3), 0));
375 const int16x8_t vprod3x01c0 = vmull_s8(vb01c0, va3c0);
376 vacc3x01 = vpadalq_s16(vacc3x01, vprod3x01c0);
377 const int16x8_t vprod3x23c0 = vmull_s8(vb23c0, va3c0);
378 vacc3x23 = vpadalq_s16(vacc3x23, vprod3x23c0);
379 const int16x8_t vprod3x45c0 = vmull_s8(vb45c0, va3c0);
380 vacc3x45 = vpadalq_s16(vacc3x45, vprod3x45c0);
381 const int16x8_t vprod3x67c0 = vmull_s8(vb67c0, va3c0);
382 vacc3x67 = vpadalq_s16(vacc3x67, vprod3x67c0);
383 }
384 p -= 4 * sizeof(void*);
385 } while (p != 0);
386
387 #if XNN_ARCH_ARM64
388 int32x4_t vacc0x0123 = vpaddq_s32(vacc0x01, vacc0x23);
389 int32x4_t vacc0x4567 = vpaddq_s32(vacc0x45, vacc0x67);
390 int32x4_t vacc1x0123 = vpaddq_s32(vacc1x01, vacc1x23);
391 int32x4_t vacc1x4567 = vpaddq_s32(vacc1x45, vacc1x67);
392 int32x4_t vacc2x0123 = vpaddq_s32(vacc2x01, vacc2x23);
393 int32x4_t vacc2x4567 = vpaddq_s32(vacc2x45, vacc2x67);
394 int32x4_t vacc3x0123 = vpaddq_s32(vacc3x01, vacc3x23);
395 int32x4_t vacc3x4567 = vpaddq_s32(vacc3x45, vacc3x67);
396 #else
397 const int32x2_t vsum0x01 = vpadd_s32(vget_low_s32(vacc0x01), vget_high_s32(vacc0x01));
398 const int32x2_t vsum0x23 = vpadd_s32(vget_low_s32(vacc0x23), vget_high_s32(vacc0x23));
399 int32x4_t vacc0x0123 = vcombine_s32(vsum0x01, vsum0x23);
400 const int32x2_t vsum0x45 = vpadd_s32(vget_low_s32(vacc0x45), vget_high_s32(vacc0x45));
401 const int32x2_t vsum0x67 = vpadd_s32(vget_low_s32(vacc0x67), vget_high_s32(vacc0x67));
402 int32x4_t vacc0x4567 = vcombine_s32(vsum0x45, vsum0x67);
403 const int32x2_t vsum1x01 = vpadd_s32(vget_low_s32(vacc1x01), vget_high_s32(vacc1x01));
404 const int32x2_t vsum1x23 = vpadd_s32(vget_low_s32(vacc1x23), vget_high_s32(vacc1x23));
405 int32x4_t vacc1x0123 = vcombine_s32(vsum1x01, vsum1x23);
406 const int32x2_t vsum1x45 = vpadd_s32(vget_low_s32(vacc1x45), vget_high_s32(vacc1x45));
407 const int32x2_t vsum1x67 = vpadd_s32(vget_low_s32(vacc1x67), vget_high_s32(vacc1x67));
408 int32x4_t vacc1x4567 = vcombine_s32(vsum1x45, vsum1x67);
409 const int32x2_t vsum2x01 = vpadd_s32(vget_low_s32(vacc2x01), vget_high_s32(vacc2x01));
410 const int32x2_t vsum2x23 = vpadd_s32(vget_low_s32(vacc2x23), vget_high_s32(vacc2x23));
411 int32x4_t vacc2x0123 = vcombine_s32(vsum2x01, vsum2x23);
412 const int32x2_t vsum2x45 = vpadd_s32(vget_low_s32(vacc2x45), vget_high_s32(vacc2x45));
413 const int32x2_t vsum2x67 = vpadd_s32(vget_low_s32(vacc2x67), vget_high_s32(vacc2x67));
414 int32x4_t vacc2x4567 = vcombine_s32(vsum2x45, vsum2x67);
415 const int32x2_t vsum3x01 = vpadd_s32(vget_low_s32(vacc3x01), vget_high_s32(vacc3x01));
416 const int32x2_t vsum3x23 = vpadd_s32(vget_low_s32(vacc3x23), vget_high_s32(vacc3x23));
417 int32x4_t vacc3x0123 = vcombine_s32(vsum3x01, vsum3x23);
418 const int32x2_t vsum3x45 = vpadd_s32(vget_low_s32(vacc3x45), vget_high_s32(vacc3x45));
419 const int32x2_t vsum3x67 = vpadd_s32(vget_low_s32(vacc3x67), vget_high_s32(vacc3x67));
420 int32x4_t vacc3x4567 = vcombine_s32(vsum3x45, vsum3x67);
421 #endif
422
423 const int32x4_t vright_pre_shift = vld1q_dup_s32(¶ms->rndnu_neon.right_pre_shift);
424 const int32x4_t vmultiplier = vld1q_dup_s32(¶ms->rndnu_neon.multiplier);
425 const int32x4_t vright_post_shift = vld1q_dup_s32(¶ms->rndnu_neon.right_post_shift);
426
427 vacc0x0123 = vqshlq_s32(vacc0x0123, vright_pre_shift);
428 vacc0x4567 = vqshlq_s32(vacc0x4567, vright_pre_shift);
429 vacc1x0123 = vqshlq_s32(vacc1x0123, vright_pre_shift);
430 vacc1x4567 = vqshlq_s32(vacc1x4567, vright_pre_shift);
431 vacc2x0123 = vqshlq_s32(vacc2x0123, vright_pre_shift);
432 vacc2x4567 = vqshlq_s32(vacc2x4567, vright_pre_shift);
433 vacc3x0123 = vqshlq_s32(vacc3x0123, vright_pre_shift);
434 vacc3x4567 = vqshlq_s32(vacc3x4567, vright_pre_shift);
435
436 vacc0x0123 = vqdmulhq_s32(vacc0x0123, vmultiplier);
437 vacc0x4567 = vqdmulhq_s32(vacc0x4567, vmultiplier);
438 vacc1x0123 = vqdmulhq_s32(vacc1x0123, vmultiplier);
439 vacc1x4567 = vqdmulhq_s32(vacc1x4567, vmultiplier);
440 vacc2x0123 = vqdmulhq_s32(vacc2x0123, vmultiplier);
441 vacc2x4567 = vqdmulhq_s32(vacc2x4567, vmultiplier);
442 vacc3x0123 = vqdmulhq_s32(vacc3x0123, vmultiplier);
443 vacc3x4567 = vqdmulhq_s32(vacc3x4567, vmultiplier);
444
445 vacc0x0123 = vrshlq_s32(vacc0x0123, vright_post_shift);
446 vacc0x4567 = vrshlq_s32(vacc0x4567, vright_post_shift);
447 vacc1x0123 = vrshlq_s32(vacc1x0123, vright_post_shift);
448 vacc1x4567 = vrshlq_s32(vacc1x4567, vright_post_shift);
449 vacc2x0123 = vrshlq_s32(vacc2x0123, vright_post_shift);
450 vacc2x4567 = vrshlq_s32(vacc2x4567, vright_post_shift);
451 vacc3x0123 = vrshlq_s32(vacc3x0123, vright_post_shift);
452 vacc3x4567 = vrshlq_s32(vacc3x4567, vright_post_shift);
453
454 const int16x8_t voutput_zero_point = vld1q_dup_s16(¶ms->rndnu_neon.output_zero_point);
455 #if XNN_ARCH_ARM64
456 int16x8_t vacc0x01234567 = vqmovn_high_s32(vqmovn_s32(vacc0x0123), vacc0x4567);
457 int16x8_t vacc1x01234567 = vqmovn_high_s32(vqmovn_s32(vacc1x0123), vacc1x4567);
458 int16x8_t vacc2x01234567 = vqmovn_high_s32(vqmovn_s32(vacc2x0123), vacc2x4567);
459 int16x8_t vacc3x01234567 = vqmovn_high_s32(vqmovn_s32(vacc3x0123), vacc3x4567);
460
461 vacc0x01234567 = vqaddq_s16(vacc0x01234567, voutput_zero_point);
462 vacc1x01234567 = vqaddq_s16(vacc1x01234567, voutput_zero_point);
463 vacc2x01234567 = vqaddq_s16(vacc2x01234567, voutput_zero_point);
464 vacc3x01234567 = vqaddq_s16(vacc3x01234567, voutput_zero_point);
465
466 int8x16_t vout0x01234567_1x01234567 = vqmovn_high_s16(vqmovn_s16(vacc0x01234567), vacc1x01234567);
467 int8x16_t vout2x01234567_3x01234567 = vqmovn_high_s16(vqmovn_s16(vacc2x01234567), vacc3x01234567);
468 #else
469 int16x8_t vacc0x01234567 = vcombine_s16(vqmovn_s32(vacc0x0123), vqmovn_s32(vacc0x4567));
470 int16x8_t vacc1x01234567 = vcombine_s16(vqmovn_s32(vacc1x0123), vqmovn_s32(vacc1x4567));
471 int16x8_t vacc2x01234567 = vcombine_s16(vqmovn_s32(vacc2x0123), vqmovn_s32(vacc2x4567));
472 int16x8_t vacc3x01234567 = vcombine_s16(vqmovn_s32(vacc3x0123), vqmovn_s32(vacc3x4567));
473
474 vacc0x01234567 = vqaddq_s16(vacc0x01234567, voutput_zero_point);
475 vacc1x01234567 = vqaddq_s16(vacc1x01234567, voutput_zero_point);
476 vacc2x01234567 = vqaddq_s16(vacc2x01234567, voutput_zero_point);
477 vacc3x01234567 = vqaddq_s16(vacc3x01234567, voutput_zero_point);
478
479 int8x16_t vout0x01234567_1x01234567 = vcombine_s8(vqmovn_s16(vacc0x01234567), vqmovn_s16(vacc1x01234567));
480 int8x16_t vout2x01234567_3x01234567 = vcombine_s8(vqmovn_s16(vacc2x01234567), vqmovn_s16(vacc3x01234567));
481 #endif
482
483 const int8x16_t voutput_min = vld1q_dup_s8(¶ms->rndnu_neon.output_min);
484 vout0x01234567_1x01234567 = vmaxq_s8(vout0x01234567_1x01234567, voutput_min);
485 vout2x01234567_3x01234567 = vmaxq_s8(vout2x01234567_3x01234567, voutput_min);
486
487 const int8x16_t voutput_max = vld1q_dup_s8(¶ms->rndnu_neon.output_max);
488 vout0x01234567_1x01234567 = vminq_s8(vout0x01234567_1x01234567, voutput_max);
489 vout2x01234567_3x01234567 = vminq_s8(vout2x01234567_3x01234567, voutput_max);
490
491 if (nc >= 8) {
492 vst1_s8(c3 + 0, vget_high_s8(vout2x01234567_3x01234567));
493 vst1_s8(c2 + 0, vget_low_s8(vout2x01234567_3x01234567));
494 vst1_s8(c1 + 0, vget_high_s8(vout0x01234567_1x01234567));
495 vst1_s8(c0 + 0, vget_low_s8(vout0x01234567_1x01234567));
496
497 c3 = (int8_t*) ((uintptr_t) c3 + cn_stride);
498 c2 = (int8_t*) ((uintptr_t) c2 + cn_stride);
499 c1 = (int8_t*) ((uintptr_t) c1 + cn_stride);
500 c0 = (int8_t*) ((uintptr_t) c0 + cn_stride);
501
502 a = (const int8_t**restrict) ((uintptr_t) a - ks);
503
504 nc -= 8;
505 } else {
506 if (nc & 4) {
507 vst1q_lane_u32((void*) c3, vreinterpretq_u32_s8(vout2x01234567_3x01234567), 2); c3 += 4;
508 vst1q_lane_u32((void*) c2, vreinterpretq_u32_s8(vout2x01234567_3x01234567), 0); c2 += 4;
509 vst1q_lane_u32((void*) c1, vreinterpretq_u32_s8(vout0x01234567_1x01234567), 2); c1 += 4;
510 vst1q_lane_u32((void*) c0, vreinterpretq_u32_s8(vout0x01234567_1x01234567), 0); c0 += 4;
511 vout2x01234567_3x01234567 = vextq_s8(vout2x01234567_3x01234567, vout2x01234567_3x01234567, 4);
512 vout0x01234567_1x01234567 = vextq_s8(vout0x01234567_1x01234567, vout0x01234567_1x01234567, 4);
513 }
514 if (nc & 2) {
515 vst1q_lane_u16((void*) c3, vreinterpretq_u16_s8(vout2x01234567_3x01234567), 4); c3 += 2;
516 vst1q_lane_u16((void*) c2, vreinterpretq_u16_s8(vout2x01234567_3x01234567), 0); c2 += 2;
517 vst1q_lane_u16((void*) c1, vreinterpretq_u16_s8(vout0x01234567_1x01234567), 4); c1 += 2;
518 vst1q_lane_u16((void*) c0, vreinterpretq_u16_s8(vout0x01234567_1x01234567), 0); c0 += 2;
519 vout2x01234567_3x01234567 = vextq_s8(vout2x01234567_3x01234567, vout2x01234567_3x01234567, 2);
520 vout0x01234567_1x01234567 = vextq_s8(vout0x01234567_1x01234567, vout0x01234567_1x01234567, 2);
521 }
522 if (nc & 1) {
523 vst1q_lane_s8(c3, vout2x01234567_3x01234567, 8);
524 vst1q_lane_s8(c2, vout2x01234567_3x01234567, 0);
525 vst1q_lane_s8(c1, vout0x01234567_1x01234567, 8);
526 vst1q_lane_s8(c0, vout0x01234567_1x01234567, 0);
527 }
528
529 nc = 0;
530 }
531 } while (nc != 0);
532 }
533