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_qs8_gemm_minmax_rndnu_ukernel_4x8c2__neon_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_qs8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS (1)])17 void xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c2__neon_mlal_ld2r(
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_qs8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
28 {
29 assert(mr != 0);
30 assert(mr <= 4);
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 const int8_t* a2 = (const int8_t*) ((uintptr_t) a1 + a_stride);
48 int8_t* c2 = (int8_t*) ((uintptr_t) c1 + cm_stride);
49 if XNN_UNPREDICTABLE(mr <= 2) {
50 a2 = a1;
51 c2 = c1;
52 }
53 const int8_t* a3 = (const int8_t*) ((uintptr_t) a2 + a_stride);
54 int8_t* c3 = (int8_t*) ((uintptr_t) c2 + cm_stride);
55 if XNN_UNPREDICTABLE(mr != 4) {
56 a3 = a2;
57 c3 = c2;
58 }
59
60 do {
61 int32x4_t vacc0x0123 = vld1q_s32(w); w = (const void*) ((uintptr_t) w + 4 * sizeof(int32_t));
62 int32x4_t vacc0x4567 = vld1q_s32(w); w = (const void*) ((uintptr_t) w + 4 * sizeof(int32_t));
63 int32x4_t vacc1x0123 = vacc0x0123;
64 int32x4_t vacc1x4567 = vacc0x4567;
65 int32x4_t vacc2x0123 = vacc0x0123;
66 int32x4_t vacc2x4567 = vacc0x4567;
67 int32x4_t vacc3x0123 = vacc0x0123;
68 int32x4_t vacc3x4567 = vacc0x4567;
69
70 size_t k = kc;
71
72 while (k >= 16 * sizeof(int8_t)) {
73 const int16x4x2_t va00x0 = vld2_dup_s16((const void*)a0);
74 const int16x4x2_t va01x0 = vld2_dup_s16((const void*)(a0 + 4)); a0 += 8;
75 const int16x4x2_t va00x1 = vld2_dup_s16((const void*)a0);
76 const int16x4x2_t va01x1 = vld2_dup_s16((const void*)(a0 + 4)); a0 += 8;
77 const int16x4x2_t va10x0 = vld2_dup_s16((const void*)a1);
78 const int16x4x2_t va11x0 = vld2_dup_s16((const void*)(a1 + 4)); a1 += 8;
79 const int16x4x2_t va10x1 = vld2_dup_s16((const void*)a1);
80 const int16x4x2_t va11x1 = vld2_dup_s16((const void*)(a1 + 4)); a1 += 8;
81 const int16x4x2_t va20x0 = vld2_dup_s16((const void*)a2);
82 const int16x4x2_t va21x0 = vld2_dup_s16((const void*)(a2 + 4)); a2 += 8;
83 const int16x4x2_t va20x1 = vld2_dup_s16((const void*)a2);
84 const int16x4x2_t va21x1 = vld2_dup_s16((const void*)(a2 + 4)); a2 += 8;
85 const int16x4x2_t va30x0 = vld2_dup_s16((const void*)a3);
86 const int16x4x2_t va31x0 = vld2_dup_s16((const void*)(a3 + 4)); a3 += 8;
87 const int16x4x2_t va30x1 = vld2_dup_s16((const void*)a3);
88 const int16x4x2_t va31x1 = vld2_dup_s16((const void*)(a3 + 4)); a3 += 8;
89 const int8x8_t vb0123c0x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
90 const int8x8_t vb4567c0x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
91 const int8x8_t vb0123c1x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
92 const int8x8_t vb4567c1x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
93 const int8x8_t vb0123c2x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
94 const int8x8_t vb4567c2x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
95 const int8x8_t vb0123c3x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
96 const int8x8_t vb4567c3x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
97
98 const int8x8_t va0c0x0 = vreinterpret_s8_s16(va00x0.val[0]);
99 const int8x8_t va0c0x1 = vreinterpret_s8_s16(va00x1.val[0]);
100 const int8x8_t va1c0x0 = vreinterpret_s8_s16(va10x0.val[0]);
101 const int8x8_t va1c0x1 = vreinterpret_s8_s16(va10x1.val[0]);
102 const int8x8_t va2c0x0 = vreinterpret_s8_s16(va20x0.val[0]);
103 const int8x8_t va2c0x1 = vreinterpret_s8_s16(va20x1.val[0]);
104 const int8x8_t va3c0x0 = vreinterpret_s8_s16(va30x0.val[0]);
105 const int8x8_t va3c0x1 = vreinterpret_s8_s16(va30x1.val[0]);
106
107 int16x8_t vprod0x0123c0 = vmull_s8(vb0123c0x0, va0c0x0);
108 int16x8_t vprod1x0123c0 = vmull_s8(vb0123c0x0, va1c0x0);
109 int16x8_t vprod2x0123c0 = vmull_s8(vb0123c0x0, va2c0x0);
110 int16x8_t vprod3x0123c0 = vmull_s8(vb0123c0x0, va3c0x0);
111 const int8x8_t vb0123c0x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
112 vprod0x0123c0 = vmlal_s8(vprod0x0123c0, vb0123c0x1, va0c0x1);
113 vprod1x0123c0 = vmlal_s8(vprod1x0123c0, vb0123c0x1, va1c0x1);
114 vprod2x0123c0 = vmlal_s8(vprod2x0123c0, vb0123c0x1, va2c0x1);
115 vprod3x0123c0 = vmlal_s8(vprod3x0123c0, vb0123c0x1, va3c0x1);
116 vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c0);
117 vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c0);
118 vacc2x0123 = vpadalq_s16(vacc2x0123, vprod2x0123c0);
119 vacc3x0123 = vpadalq_s16(vacc3x0123, vprod3x0123c0);
120 int16x8_t vprod0x4567c0 = vmull_s8(vb4567c0x0, va0c0x0);
121 int16x8_t vprod1x4567c0 = vmull_s8(vb4567c0x0, va1c0x0);
122 int16x8_t vprod2x4567c0 = vmull_s8(vb4567c0x0, va2c0x0);
123 int16x8_t vprod3x4567c0 = vmull_s8(vb4567c0x0, va3c0x0);
124 const int8x8_t vb4567c0x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
125 vprod0x4567c0 = vmlal_s8(vprod0x4567c0, vb4567c0x1, va0c0x1);
126 vprod1x4567c0 = vmlal_s8(vprod1x4567c0, vb4567c0x1, va1c0x1);
127 vprod2x4567c0 = vmlal_s8(vprod2x4567c0, vb4567c0x1, va2c0x1);
128 vprod3x4567c0 = vmlal_s8(vprod3x4567c0, vb4567c0x1, va3c0x1);
129 vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c0);
130 vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c0);
131 vacc2x4567 = vpadalq_s16(vacc2x4567, vprod2x4567c0);
132 vacc3x4567 = vpadalq_s16(vacc3x4567, vprod3x4567c0);
133 const int8x8_t va0c1x0 = vreinterpret_s8_s16(va00x0.val[1]);
134 const int8x8_t va0c1x1 = vreinterpret_s8_s16(va00x1.val[1]);
135 const int8x8_t va1c1x0 = vreinterpret_s8_s16(va10x0.val[1]);
136 const int8x8_t va1c1x1 = vreinterpret_s8_s16(va10x1.val[1]);
137 const int8x8_t va2c1x0 = vreinterpret_s8_s16(va20x0.val[1]);
138 const int8x8_t va2c1x1 = vreinterpret_s8_s16(va20x1.val[1]);
139 const int8x8_t va3c1x0 = vreinterpret_s8_s16(va30x0.val[1]);
140 const int8x8_t va3c1x1 = vreinterpret_s8_s16(va30x1.val[1]);
141
142 int16x8_t vprod0x0123c1 = vmull_s8(vb0123c1x0, va0c1x0);
143 int16x8_t vprod1x0123c1 = vmull_s8(vb0123c1x0, va1c1x0);
144 int16x8_t vprod2x0123c1 = vmull_s8(vb0123c1x0, va2c1x0);
145 int16x8_t vprod3x0123c1 = vmull_s8(vb0123c1x0, va3c1x0);
146 const int8x8_t vb0123c1x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
147 vprod0x0123c1 = vmlal_s8(vprod0x0123c1, vb0123c1x1, va0c1x1);
148 vprod1x0123c1 = vmlal_s8(vprod1x0123c1, vb0123c1x1, va1c1x1);
149 vprod2x0123c1 = vmlal_s8(vprod2x0123c1, vb0123c1x1, va2c1x1);
150 vprod3x0123c1 = vmlal_s8(vprod3x0123c1, vb0123c1x1, va3c1x1);
151 vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c1);
152 vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c1);
153 vacc2x0123 = vpadalq_s16(vacc2x0123, vprod2x0123c1);
154 vacc3x0123 = vpadalq_s16(vacc3x0123, vprod3x0123c1);
155 int16x8_t vprod0x4567c1 = vmull_s8(vb4567c1x0, va0c1x0);
156 int16x8_t vprod1x4567c1 = vmull_s8(vb4567c1x0, va1c1x0);
157 int16x8_t vprod2x4567c1 = vmull_s8(vb4567c1x0, va2c1x0);
158 int16x8_t vprod3x4567c1 = vmull_s8(vb4567c1x0, va3c1x0);
159 const int8x8_t vb4567c1x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
160 vprod0x4567c1 = vmlal_s8(vprod0x4567c1, vb4567c1x1, va0c1x1);
161 vprod1x4567c1 = vmlal_s8(vprod1x4567c1, vb4567c1x1, va1c1x1);
162 vprod2x4567c1 = vmlal_s8(vprod2x4567c1, vb4567c1x1, va2c1x1);
163 vprod3x4567c1 = vmlal_s8(vprod3x4567c1, vb4567c1x1, va3c1x1);
164 vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c1);
165 vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c1);
166 vacc2x4567 = vpadalq_s16(vacc2x4567, vprod2x4567c1);
167 vacc3x4567 = vpadalq_s16(vacc3x4567, vprod3x4567c1);
168 const int8x8_t va0c2x0 = vreinterpret_s8_s16(va01x0.val[0]);
169 const int8x8_t va0c2x1 = vreinterpret_s8_s16(va01x1.val[0]);
170 const int8x8_t va1c2x0 = vreinterpret_s8_s16(va11x0.val[0]);
171 const int8x8_t va1c2x1 = vreinterpret_s8_s16(va11x1.val[0]);
172 const int8x8_t va2c2x0 = vreinterpret_s8_s16(va21x0.val[0]);
173 const int8x8_t va2c2x1 = vreinterpret_s8_s16(va21x1.val[0]);
174 const int8x8_t va3c2x0 = vreinterpret_s8_s16(va31x0.val[0]);
175 const int8x8_t va3c2x1 = vreinterpret_s8_s16(va31x1.val[0]);
176
177 int16x8_t vprod0x0123c2 = vmull_s8(vb0123c2x0, va0c2x0);
178 int16x8_t vprod1x0123c2 = vmull_s8(vb0123c2x0, va1c2x0);
179 int16x8_t vprod2x0123c2 = vmull_s8(vb0123c2x0, va2c2x0);
180 int16x8_t vprod3x0123c2 = vmull_s8(vb0123c2x0, va3c2x0);
181 const int8x8_t vb0123c2x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
182 vprod0x0123c2 = vmlal_s8(vprod0x0123c2, vb0123c2x1, va0c2x1);
183 vprod1x0123c2 = vmlal_s8(vprod1x0123c2, vb0123c2x1, va1c2x1);
184 vprod2x0123c2 = vmlal_s8(vprod2x0123c2, vb0123c2x1, va2c2x1);
185 vprod3x0123c2 = vmlal_s8(vprod3x0123c2, vb0123c2x1, va3c2x1);
186 vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c2);
187 vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c2);
188 vacc2x0123 = vpadalq_s16(vacc2x0123, vprod2x0123c2);
189 vacc3x0123 = vpadalq_s16(vacc3x0123, vprod3x0123c2);
190 int16x8_t vprod0x4567c2 = vmull_s8(vb4567c2x0, va0c2x0);
191 int16x8_t vprod1x4567c2 = vmull_s8(vb4567c2x0, va1c2x0);
192 int16x8_t vprod2x4567c2 = vmull_s8(vb4567c2x0, va2c2x0);
193 int16x8_t vprod3x4567c2 = vmull_s8(vb4567c2x0, va3c2x0);
194 const int8x8_t vb4567c2x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
195 vprod0x4567c2 = vmlal_s8(vprod0x4567c2, vb4567c2x1, va0c2x1);
196 vprod1x4567c2 = vmlal_s8(vprod1x4567c2, vb4567c2x1, va1c2x1);
197 vprod2x4567c2 = vmlal_s8(vprod2x4567c2, vb4567c2x1, va2c2x1);
198 vprod3x4567c2 = vmlal_s8(vprod3x4567c2, vb4567c2x1, va3c2x1);
199 vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c2);
200 vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c2);
201 vacc2x4567 = vpadalq_s16(vacc2x4567, vprod2x4567c2);
202 vacc3x4567 = vpadalq_s16(vacc3x4567, vprod3x4567c2);
203 const int8x8_t va0c3x0 = vreinterpret_s8_s16(va01x0.val[1]);
204 const int8x8_t va0c3x1 = vreinterpret_s8_s16(va01x1.val[1]);
205 const int8x8_t va1c3x0 = vreinterpret_s8_s16(va11x0.val[1]);
206 const int8x8_t va1c3x1 = vreinterpret_s8_s16(va11x1.val[1]);
207 const int8x8_t va2c3x0 = vreinterpret_s8_s16(va21x0.val[1]);
208 const int8x8_t va2c3x1 = vreinterpret_s8_s16(va21x1.val[1]);
209 const int8x8_t va3c3x0 = vreinterpret_s8_s16(va31x0.val[1]);
210 const int8x8_t va3c3x1 = vreinterpret_s8_s16(va31x1.val[1]);
211
212 int16x8_t vprod0x0123c3 = vmull_s8(vb0123c3x0, va0c3x0);
213 int16x8_t vprod1x0123c3 = vmull_s8(vb0123c3x0, va1c3x0);
214 int16x8_t vprod2x0123c3 = vmull_s8(vb0123c3x0, va2c3x0);
215 int16x8_t vprod3x0123c3 = vmull_s8(vb0123c3x0, va3c3x0);
216 const int8x8_t vb0123c3x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
217 vprod0x0123c3 = vmlal_s8(vprod0x0123c3, vb0123c3x1, va0c3x1);
218 vprod1x0123c3 = vmlal_s8(vprod1x0123c3, vb0123c3x1, va1c3x1);
219 vprod2x0123c3 = vmlal_s8(vprod2x0123c3, vb0123c3x1, va2c3x1);
220 vprod3x0123c3 = vmlal_s8(vprod3x0123c3, vb0123c3x1, va3c3x1);
221 vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c3);
222 vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c3);
223 vacc2x0123 = vpadalq_s16(vacc2x0123, vprod2x0123c3);
224 vacc3x0123 = vpadalq_s16(vacc3x0123, vprod3x0123c3);
225 int16x8_t vprod0x4567c3 = vmull_s8(vb4567c3x0, va0c3x0);
226 int16x8_t vprod1x4567c3 = vmull_s8(vb4567c3x0, va1c3x0);
227 int16x8_t vprod2x4567c3 = vmull_s8(vb4567c3x0, va2c3x0);
228 int16x8_t vprod3x4567c3 = vmull_s8(vb4567c3x0, va3c3x0);
229 const int8x8_t vb4567c3x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
230 vprod0x4567c3 = vmlal_s8(vprod0x4567c3, vb4567c3x1, va0c3x1);
231 vprod1x4567c3 = vmlal_s8(vprod1x4567c3, vb4567c3x1, va1c3x1);
232 vprod2x4567c3 = vmlal_s8(vprod2x4567c3, vb4567c3x1, va2c3x1);
233 vprod3x4567c3 = vmlal_s8(vprod3x4567c3, vb4567c3x1, va3c3x1);
234 vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c3);
235 vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c3);
236 vacc2x4567 = vpadalq_s16(vacc2x4567, vprod2x4567c3);
237 vacc3x4567 = vpadalq_s16(vacc3x4567, vprod3x4567c3);
238
239 k -= 16 * sizeof(int8_t);
240 }
241
242 if (k >= 8 * sizeof(int8_t)) {
243 const int16x4x2_t va00 = vld2_dup_s16((const void*)a0);
244 const int16x4x2_t va01 = vld2_dup_s16((const void*)(a0 + 4)); a0 += 8;
245 const int16x4x2_t va10 = vld2_dup_s16((const void*)a1);
246 const int16x4x2_t va11 = vld2_dup_s16((const void*)(a1 + 4)); a1 += 8;
247 const int16x4x2_t va20 = vld2_dup_s16((const void*)a2);
248 const int16x4x2_t va21 = vld2_dup_s16((const void*)(a2 + 4)); a2 += 8;
249 const int16x4x2_t va30 = vld2_dup_s16((const void*)a3);
250 const int16x4x2_t va31 = vld2_dup_s16((const void*)(a3 + 4)); a3 += 8;
251
252 const int8x8_t vb0123c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
253 const int8x8_t vb4567c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
254 const int8x8_t vb0123c1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
255 const int8x8_t vb4567c1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
256 const int8x8_t vb0123c2 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
257 const int8x8_t vb4567c2 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
258 const int8x8_t vb0123c3 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
259 const int8x8_t vb4567c3 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
260
261 const int8x8_t va0c0 = vreinterpret_s8_s16(va00.val[0]);
262 const int8x8_t va1c0 = vreinterpret_s8_s16(va10.val[0]);
263 const int8x8_t va2c0 = vreinterpret_s8_s16(va20.val[0]);
264 const int8x8_t va3c0 = vreinterpret_s8_s16(va30.val[0]);
265
266 const int16x8_t vprod0x0123c0 = vmull_s8(vb0123c0, va0c0);
267 const int16x8_t vprod1x0123c0 = vmull_s8(vb0123c0, va1c0);
268 const int16x8_t vprod2x0123c0 = vmull_s8(vb0123c0, va2c0);
269 const int16x8_t vprod3x0123c0 = vmull_s8(vb0123c0, va3c0);
270 vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c0);
271 vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c0);
272 vacc2x0123 = vpadalq_s16(vacc2x0123, vprod2x0123c0);
273 vacc3x0123 = vpadalq_s16(vacc3x0123, vprod3x0123c0);
274 const int16x8_t vprod0x4567c0 = vmull_s8(vb4567c0, va0c0);
275 const int16x8_t vprod1x4567c0 = vmull_s8(vb4567c0, va1c0);
276 const int16x8_t vprod2x4567c0 = vmull_s8(vb4567c0, va2c0);
277 const int16x8_t vprod3x4567c0 = vmull_s8(vb4567c0, va3c0);
278 vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c0);
279 vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c0);
280 vacc2x4567 = vpadalq_s16(vacc2x4567, vprod2x4567c0);
281 vacc3x4567 = vpadalq_s16(vacc3x4567, vprod3x4567c0);
282 const int8x8_t va0c1 = vreinterpret_s8_s16(va00.val[1]);
283 const int8x8_t va1c1 = vreinterpret_s8_s16(va10.val[1]);
284 const int8x8_t va2c1 = vreinterpret_s8_s16(va20.val[1]);
285 const int8x8_t va3c1 = vreinterpret_s8_s16(va30.val[1]);
286
287 const int16x8_t vprod0x0123c1 = vmull_s8(vb0123c1, va0c1);
288 const int16x8_t vprod1x0123c1 = vmull_s8(vb0123c1, va1c1);
289 const int16x8_t vprod2x0123c1 = vmull_s8(vb0123c1, va2c1);
290 const int16x8_t vprod3x0123c1 = vmull_s8(vb0123c1, va3c1);
291 vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c1);
292 vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c1);
293 vacc2x0123 = vpadalq_s16(vacc2x0123, vprod2x0123c1);
294 vacc3x0123 = vpadalq_s16(vacc3x0123, vprod3x0123c1);
295 const int16x8_t vprod0x4567c1 = vmull_s8(vb4567c1, va0c1);
296 const int16x8_t vprod1x4567c1 = vmull_s8(vb4567c1, va1c1);
297 const int16x8_t vprod2x4567c1 = vmull_s8(vb4567c1, va2c1);
298 const int16x8_t vprod3x4567c1 = vmull_s8(vb4567c1, va3c1);
299 vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c1);
300 vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c1);
301 vacc2x4567 = vpadalq_s16(vacc2x4567, vprod2x4567c1);
302 vacc3x4567 = vpadalq_s16(vacc3x4567, vprod3x4567c1);
303 const int8x8_t va0c2 = vreinterpret_s8_s16(va01.val[0]);
304 const int8x8_t va1c2 = vreinterpret_s8_s16(va11.val[0]);
305 const int8x8_t va2c2 = vreinterpret_s8_s16(va21.val[0]);
306 const int8x8_t va3c2 = vreinterpret_s8_s16(va31.val[0]);
307
308 const int16x8_t vprod0x0123c2 = vmull_s8(vb0123c2, va0c2);
309 const int16x8_t vprod1x0123c2 = vmull_s8(vb0123c2, va1c2);
310 const int16x8_t vprod2x0123c2 = vmull_s8(vb0123c2, va2c2);
311 const int16x8_t vprod3x0123c2 = vmull_s8(vb0123c2, va3c2);
312 vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c2);
313 vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c2);
314 vacc2x0123 = vpadalq_s16(vacc2x0123, vprod2x0123c2);
315 vacc3x0123 = vpadalq_s16(vacc3x0123, vprod3x0123c2);
316 const int16x8_t vprod0x4567c2 = vmull_s8(vb4567c2, va0c2);
317 const int16x8_t vprod1x4567c2 = vmull_s8(vb4567c2, va1c2);
318 const int16x8_t vprod2x4567c2 = vmull_s8(vb4567c2, va2c2);
319 const int16x8_t vprod3x4567c2 = vmull_s8(vb4567c2, va3c2);
320 vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c2);
321 vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c2);
322 vacc2x4567 = vpadalq_s16(vacc2x4567, vprod2x4567c2);
323 vacc3x4567 = vpadalq_s16(vacc3x4567, vprod3x4567c2);
324 const int8x8_t va0c3 = vreinterpret_s8_s16(va01.val[1]);
325 const int8x8_t va1c3 = vreinterpret_s8_s16(va11.val[1]);
326 const int8x8_t va2c3 = vreinterpret_s8_s16(va21.val[1]);
327 const int8x8_t va3c3 = vreinterpret_s8_s16(va31.val[1]);
328
329 const int16x8_t vprod0x0123c3 = vmull_s8(vb0123c3, va0c3);
330 const int16x8_t vprod1x0123c3 = vmull_s8(vb0123c3, va1c3);
331 const int16x8_t vprod2x0123c3 = vmull_s8(vb0123c3, va2c3);
332 const int16x8_t vprod3x0123c3 = vmull_s8(vb0123c3, va3c3);
333 vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c3);
334 vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c3);
335 vacc2x0123 = vpadalq_s16(vacc2x0123, vprod2x0123c3);
336 vacc3x0123 = vpadalq_s16(vacc3x0123, vprod3x0123c3);
337 const int16x8_t vprod0x4567c3 = vmull_s8(vb4567c3, va0c3);
338 const int16x8_t vprod1x4567c3 = vmull_s8(vb4567c3, va1c3);
339 const int16x8_t vprod2x4567c3 = vmull_s8(vb4567c3, va2c3);
340 const int16x8_t vprod3x4567c3 = vmull_s8(vb4567c3, va3c3);
341 vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c3);
342 vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c3);
343 vacc2x4567 = vpadalq_s16(vacc2x4567, vprod2x4567c3);
344 vacc3x4567 = vpadalq_s16(vacc3x4567, vprod3x4567c3);
345
346 k -= 8 * sizeof(int8_t);
347 }
348
349 if XNN_UNLIKELY(k != 0) {
350 const int8x8_t va0 = vld1_s8(a0); a0 = (const int8_t*) ((uintptr_t) a0 + k);
351 const int8x8_t va1 = vld1_s8(a1); a1 = (const int8_t*) ((uintptr_t) a1 + k);
352 const int8x8_t va2 = vld1_s8(a2); a2 = (const int8_t*) ((uintptr_t) a2 + k);
353 const int8x8_t va3 = vld1_s8(a3); a3 = (const int8_t*) ((uintptr_t) a3 + k);
354
355 const int8x8_t vb0123c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
356 const int8x8_t vb4567c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
357
358 const int8x8_t va0c0 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0), 0));
359 const int16x8_t vprod0x0123c0 = vmull_s8(vb0123c0, va0c0);
360 vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c0);
361 const int16x8_t vprod0x4567c0 = vmull_s8(vb4567c0, va0c0);
362 vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c0);
363 const int8x8_t va1c0 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va1), 0));
364 const int16x8_t vprod1x0123c0 = vmull_s8(vb0123c0, va1c0);
365 vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c0);
366 const int16x8_t vprod1x4567c0 = vmull_s8(vb4567c0, va1c0);
367 vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c0);
368 const int8x8_t va2c0 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va2), 0));
369 const int16x8_t vprod2x0123c0 = vmull_s8(vb0123c0, va2c0);
370 vacc2x0123 = vpadalq_s16(vacc2x0123, vprod2x0123c0);
371 const int16x8_t vprod2x4567c0 = vmull_s8(vb4567c0, va2c0);
372 vacc2x4567 = vpadalq_s16(vacc2x4567, vprod2x4567c0);
373 const int8x8_t va3c0 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va3), 0));
374 const int16x8_t vprod3x0123c0 = vmull_s8(vb0123c0, va3c0);
375 vacc3x0123 = vpadalq_s16(vacc3x0123, vprod3x0123c0);
376 const int16x8_t vprod3x4567c0 = vmull_s8(vb4567c0, va3c0);
377 vacc3x4567 = vpadalq_s16(vacc3x4567, vprod3x4567c0);
378
379 if (k > 2 * sizeof(int8_t)) {
380 const int8x8_t vb0123c1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
381 const int8x8_t vb4567c1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
382
383 const int8x8_t va0c1 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0), 1));
384 const int16x8_t vprod0x0123c1 = vmull_s8(vb0123c1, va0c1);
385 vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c1);
386 const int16x8_t vprod0x4567c1 = vmull_s8(vb4567c1, va0c1);
387 vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c1);
388 const int8x8_t va1c1 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va1), 1));
389 const int16x8_t vprod1x0123c1 = vmull_s8(vb0123c1, va1c1);
390 vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c1);
391 const int16x8_t vprod1x4567c1 = vmull_s8(vb4567c1, va1c1);
392 vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c1);
393 const int8x8_t va2c1 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va2), 1));
394 const int16x8_t vprod2x0123c1 = vmull_s8(vb0123c1, va2c1);
395 vacc2x0123 = vpadalq_s16(vacc2x0123, vprod2x0123c1);
396 const int16x8_t vprod2x4567c1 = vmull_s8(vb4567c1, va2c1);
397 vacc2x4567 = vpadalq_s16(vacc2x4567, vprod2x4567c1);
398 const int8x8_t va3c1 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va3), 1));
399 const int16x8_t vprod3x0123c1 = vmull_s8(vb0123c1, va3c1);
400 vacc3x0123 = vpadalq_s16(vacc3x0123, vprod3x0123c1);
401 const int16x8_t vprod3x4567c1 = vmull_s8(vb4567c1, va3c1);
402 vacc3x4567 = vpadalq_s16(vacc3x4567, vprod3x4567c1);
403
404 if (k > 4 * sizeof(int8_t)) {
405 const int8x8_t vb0123c2 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
406 const int8x8_t vb4567c2 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
407
408 const int8x8_t va0c2 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0), 2));
409 const int16x8_t vprod0x0123c2 = vmull_s8(vb0123c2, va0c2);
410 vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c2);
411 const int16x8_t vprod0x4567c2 = vmull_s8(vb4567c2, va0c2);
412 vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c2);
413 const int8x8_t va1c2 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va1), 2));
414 const int16x8_t vprod1x0123c2 = vmull_s8(vb0123c2, va1c2);
415 vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c2);
416 const int16x8_t vprod1x4567c2 = vmull_s8(vb4567c2, va1c2);
417 vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c2);
418 const int8x8_t va2c2 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va2), 2));
419 const int16x8_t vprod2x0123c2 = vmull_s8(vb0123c2, va2c2);
420 vacc2x0123 = vpadalq_s16(vacc2x0123, vprod2x0123c2);
421 const int16x8_t vprod2x4567c2 = vmull_s8(vb4567c2, va2c2);
422 vacc2x4567 = vpadalq_s16(vacc2x4567, vprod2x4567c2);
423 const int8x8_t va3c2 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va3), 2));
424 const int16x8_t vprod3x0123c2 = vmull_s8(vb0123c2, va3c2);
425 vacc3x0123 = vpadalq_s16(vacc3x0123, vprod3x0123c2);
426 const int16x8_t vprod3x4567c2 = vmull_s8(vb4567c2, va3c2);
427 vacc3x4567 = vpadalq_s16(vacc3x4567, vprod3x4567c2);
428 }
429 }
430 }
431
432 const int32x4_t vright_pre_shift = vld1q_dup_s32(¶ms->rndnu_neon.right_pre_shift);
433 const int32x4_t vmultiplier = vld1q_dup_s32(¶ms->rndnu_neon.multiplier);
434 const int32x4_t vright_post_shift = vld1q_dup_s32(¶ms->rndnu_neon.right_post_shift);
435
436 vacc0x0123 = vqshlq_s32(vacc0x0123, vright_pre_shift);
437 vacc0x4567 = vqshlq_s32(vacc0x4567, vright_pre_shift);
438 vacc1x0123 = vqshlq_s32(vacc1x0123, vright_pre_shift);
439 vacc1x4567 = vqshlq_s32(vacc1x4567, vright_pre_shift);
440 vacc2x0123 = vqshlq_s32(vacc2x0123, vright_pre_shift);
441 vacc2x4567 = vqshlq_s32(vacc2x4567, vright_pre_shift);
442 vacc3x0123 = vqshlq_s32(vacc3x0123, vright_pre_shift);
443 vacc3x4567 = vqshlq_s32(vacc3x4567, vright_pre_shift);
444
445 vacc0x0123 = vqdmulhq_s32(vacc0x0123, vmultiplier);
446 vacc0x4567 = vqdmulhq_s32(vacc0x4567, vmultiplier);
447 vacc1x0123 = vqdmulhq_s32(vacc1x0123, vmultiplier);
448 vacc1x4567 = vqdmulhq_s32(vacc1x4567, vmultiplier);
449 vacc2x0123 = vqdmulhq_s32(vacc2x0123, vmultiplier);
450 vacc2x4567 = vqdmulhq_s32(vacc2x4567, vmultiplier);
451 vacc3x0123 = vqdmulhq_s32(vacc3x0123, vmultiplier);
452 vacc3x4567 = vqdmulhq_s32(vacc3x4567, vmultiplier);
453
454 vacc0x0123 = vrshlq_s32(vacc0x0123, vright_post_shift);
455 vacc0x4567 = vrshlq_s32(vacc0x4567, vright_post_shift);
456 vacc1x0123 = vrshlq_s32(vacc1x0123, vright_post_shift);
457 vacc1x4567 = vrshlq_s32(vacc1x4567, vright_post_shift);
458 vacc2x0123 = vrshlq_s32(vacc2x0123, vright_post_shift);
459 vacc2x4567 = vrshlq_s32(vacc2x4567, vright_post_shift);
460 vacc3x0123 = vrshlq_s32(vacc3x0123, vright_post_shift);
461 vacc3x4567 = vrshlq_s32(vacc3x4567, vright_post_shift);
462
463 const int16x8_t voutput_zero_point = vld1q_dup_s16(¶ms->rndnu_neon.output_zero_point);
464 #if XNN_ARCH_ARM64
465 int16x8_t vacc0x01234567 = vqmovn_high_s32(vqmovn_s32(vacc0x0123), vacc0x4567);
466 int16x8_t vacc1x01234567 = vqmovn_high_s32(vqmovn_s32(vacc1x0123), vacc1x4567);
467 int16x8_t vacc2x01234567 = vqmovn_high_s32(vqmovn_s32(vacc2x0123), vacc2x4567);
468 int16x8_t vacc3x01234567 = vqmovn_high_s32(vqmovn_s32(vacc3x0123), vacc3x4567);
469
470 vacc0x01234567 = vqaddq_s16(vacc0x01234567, voutput_zero_point);
471 vacc1x01234567 = vqaddq_s16(vacc1x01234567, voutput_zero_point);
472 vacc2x01234567 = vqaddq_s16(vacc2x01234567, voutput_zero_point);
473 vacc3x01234567 = vqaddq_s16(vacc3x01234567, voutput_zero_point);
474
475 int8x16_t vout0x01234567_1x01234567 = vqmovn_high_s16(vqmovn_s16(vacc0x01234567), vacc1x01234567);
476 int8x16_t vout2x01234567_3x01234567 = vqmovn_high_s16(vqmovn_s16(vacc2x01234567), vacc3x01234567);
477 #else
478 int16x8_t vacc0x01234567 = vcombine_s16(vqmovn_s32(vacc0x0123), vqmovn_s32(vacc0x4567));
479 int16x8_t vacc1x01234567 = vcombine_s16(vqmovn_s32(vacc1x0123), vqmovn_s32(vacc1x4567));
480 int16x8_t vacc2x01234567 = vcombine_s16(vqmovn_s32(vacc2x0123), vqmovn_s32(vacc2x4567));
481 int16x8_t vacc3x01234567 = vcombine_s16(vqmovn_s32(vacc3x0123), vqmovn_s32(vacc3x4567));
482
483 vacc0x01234567 = vqaddq_s16(vacc0x01234567, voutput_zero_point);
484 vacc1x01234567 = vqaddq_s16(vacc1x01234567, voutput_zero_point);
485 vacc2x01234567 = vqaddq_s16(vacc2x01234567, voutput_zero_point);
486 vacc3x01234567 = vqaddq_s16(vacc3x01234567, voutput_zero_point);
487
488 int8x16_t vout0x01234567_1x01234567 = vcombine_s8(vqmovn_s16(vacc0x01234567), vqmovn_s16(vacc1x01234567));
489 int8x16_t vout2x01234567_3x01234567 = vcombine_s8(vqmovn_s16(vacc2x01234567), vqmovn_s16(vacc3x01234567));
490 #endif
491
492 const int8x16_t voutput_min = vld1q_dup_s8(¶ms->rndnu_neon.output_min);
493 vout0x01234567_1x01234567 = vmaxq_s8(vout0x01234567_1x01234567, voutput_min);
494 vout2x01234567_3x01234567 = vmaxq_s8(vout2x01234567_3x01234567, voutput_min);
495
496 const int8x16_t voutput_max = vld1q_dup_s8(¶ms->rndnu_neon.output_max);
497 vout0x01234567_1x01234567 = vminq_s8(vout0x01234567_1x01234567, voutput_max);
498 vout2x01234567_3x01234567 = vminq_s8(vout2x01234567_3x01234567, voutput_max);
499
500 if (nc >= 8) {
501 vst1_s8(c0 + 0, vget_low_s8(vout0x01234567_1x01234567));
502 vst1_s8(c1 + 0, vget_high_s8(vout0x01234567_1x01234567));
503 vst1_s8(c2 + 0, vget_low_s8(vout2x01234567_3x01234567));
504 vst1_s8(c3 + 0, vget_high_s8(vout2x01234567_3x01234567));
505
506 c0 = (int8_t*) ((uintptr_t) c0 + cn_stride);
507 c1 = (int8_t*) ((uintptr_t) c1 + cn_stride);
508 c2 = (int8_t*) ((uintptr_t) c2 + cn_stride);
509 c3 = (int8_t*) ((uintptr_t) c3 + cn_stride);
510
511 a0 = (const int8_t*) ((uintptr_t) a0 - kc);
512 a1 = (const int8_t*) ((uintptr_t) a1 - kc);
513 a2 = (const int8_t*) ((uintptr_t) a2 - kc);
514 a3 = (const int8_t*) ((uintptr_t) a3 - kc);
515
516 nc -= 8;
517 } else {
518 // Final case where not all of the 8 columns fit in the destination.
519 if (nc & 4) {
520 vst1q_lane_u32((void*) c0, vreinterpretq_u32_s8(vout0x01234567_1x01234567), 0); c0 += 4;
521 vst1q_lane_u32((void*) c1, vreinterpretq_u32_s8(vout0x01234567_1x01234567), 2); c1 += 4;
522 vst1q_lane_u32((void*) c2, vreinterpretq_u32_s8(vout2x01234567_3x01234567), 0); c2 += 4;
523 vst1q_lane_u32((void*) c3, vreinterpretq_u32_s8(vout2x01234567_3x01234567), 2); c3 += 4;
524 vout0x01234567_1x01234567 = vextq_s8(vout0x01234567_1x01234567, vout0x01234567_1x01234567, 4);
525 vout2x01234567_3x01234567 = vextq_s8(vout2x01234567_3x01234567, vout2x01234567_3x01234567, 4);
526 }
527 if (nc & 2) {
528 vst1q_lane_u16((void*) c0, vreinterpretq_u16_s8(vout0x01234567_1x01234567), 0); c0 += 2;
529 vst1q_lane_u16((void*) c1, vreinterpretq_u16_s8(vout0x01234567_1x01234567), 4); c1 += 2;
530 vst1q_lane_u16((void*) c2, vreinterpretq_u16_s8(vout2x01234567_3x01234567), 0); c2 += 2;
531 vst1q_lane_u16((void*) c3, vreinterpretq_u16_s8(vout2x01234567_3x01234567), 4); c3 += 2;
532 vout0x01234567_1x01234567 = vextq_s8(vout0x01234567_1x01234567, vout0x01234567_1x01234567, 2);
533 vout2x01234567_3x01234567 = vextq_s8(vout2x01234567_3x01234567, vout2x01234567_3x01234567, 2);
534 }
535 if (nc & 1) {
536 vst1q_lane_s8(c0, vout0x01234567_1x01234567, 0);
537 vst1q_lane_s8(c1, vout0x01234567_1x01234567, 8);
538 vst1q_lane_s8(c2, vout2x01234567_3x01234567, 0);
539 vst1q_lane_s8(c3, vout2x01234567_3x01234567, 8);
540 }
541
542 nc = 0;
543 }
544 } while (nc != 0);
545 }
546