1 // Auto-generated file. Do not edit!
2 // Template: src/qs8-gemm/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_gemm_minmax_rndnu_ukernel_4x8c4__neon_mlal_ld1r(size_t mr,size_t nc,size_t kc,const int8_t * restrict a,size_t a_stride,const void * restrict w,int8_t * restrict c,size_t cm_stride,size_t cn_stride,const union xnn_qs8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS (1)])18 void xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c4__neon_mlal_ld1r(
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 <= 4);
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, 4 * 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 const int8_t* a2 = (const int8_t*) ((uintptr_t) a1 + a_stride);
49 int8_t* c2 = (int8_t*) ((uintptr_t) c1 + cm_stride);
50 if XNN_UNPREDICTABLE(mr <= 2) {
51 a2 = a1;
52 c2 = c1;
53 }
54 const int8_t* a3 = (const int8_t*) ((uintptr_t) a2 + a_stride);
55 int8_t* c3 = (int8_t*) ((uintptr_t) c2 + cm_stride);
56 if XNN_UNPREDICTABLE(mr != 4) {
57 a3 = a2;
58 c3 = c2;
59 }
60
61 do {
62 int32x4_t vacc0x01 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
63 int32x4_t vacc0x23 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
64 int32x4_t vacc0x45 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
65 int32x4_t vacc0x67 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
66 int32x4_t vacc1x01 = vacc0x01;
67 int32x4_t vacc1x23 = vacc0x23;
68 int32x4_t vacc1x45 = vacc0x45;
69 int32x4_t vacc1x67 = vacc0x67;
70 int32x4_t vacc2x01 = vacc0x01;
71 int32x4_t vacc2x23 = vacc0x23;
72 int32x4_t vacc2x45 = vacc0x45;
73 int32x4_t vacc2x67 = vacc0x67;
74 int32x4_t vacc3x01 = vacc0x01;
75 int32x4_t vacc3x23 = vacc0x23;
76 int32x4_t vacc3x45 = vacc0x45;
77 int32x4_t vacc3x67 = vacc0x67;
78
79 size_t k = kc;
80
81 while (k >= 16 * sizeof(int8_t)) {
82 const int32x2_t va00x0 = vld1_dup_s32((const void*)a0);
83 const int32x2_t va01x0 = vld1_dup_s32((const void*)(a0 + 4)); a0 += 8;
84 const int32x2_t va00x1 = vld1_dup_s32((const void*)a0);
85 const int32x2_t va01x1 = vld1_dup_s32((const void*)(a0 + 4)); a0 += 8;
86 const int32x2_t va10x0 = vld1_dup_s32((const void*)a1);
87 const int32x2_t va11x0 = vld1_dup_s32((const void*)(a1 + 4)); a1 += 8;
88 const int32x2_t va10x1 = vld1_dup_s32((const void*)a1);
89 const int32x2_t va11x1 = vld1_dup_s32((const void*)(a1 + 4)); a1 += 8;
90 const int32x2_t va20x0 = vld1_dup_s32((const void*)a2);
91 const int32x2_t va21x0 = vld1_dup_s32((const void*)(a2 + 4)); a2 += 8;
92 const int32x2_t va20x1 = vld1_dup_s32((const void*)a2);
93 const int32x2_t va21x1 = vld1_dup_s32((const void*)(a2 + 4)); a2 += 8;
94 const int32x2_t va30x0 = vld1_dup_s32((const void*)a3);
95 const int32x2_t va31x0 = vld1_dup_s32((const void*)(a3 + 4)); a3 += 8;
96 const int32x2_t va30x1 = vld1_dup_s32((const void*)a3);
97 const int32x2_t va31x1 = vld1_dup_s32((const void*)(a3 + 4)); a3 += 8;
98
99 const int8x8_t vb01c0x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
100 const int8x8_t vb23c0x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
101 const int8x8_t vb45c0x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
102 const int8x8_t vb67c0x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
103 const int8x8_t vb01c1x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
104 const int8x8_t vb23c1x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
105 const int8x8_t vb45c1x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
106 const int8x8_t vb67c1x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
107
108 const int8x8_t va0c0x0 = vreinterpret_s8_s32(va00x0);
109 const int8x8_t va0c0x1 = vreinterpret_s8_s32(va00x1);
110 const int8x8_t va1c0x0 = vreinterpret_s8_s32(va10x0);
111 const int8x8_t va1c0x1 = vreinterpret_s8_s32(va10x1);
112 const int8x8_t va2c0x0 = vreinterpret_s8_s32(va20x0);
113 const int8x8_t va2c0x1 = vreinterpret_s8_s32(va20x1);
114 const int8x8_t va3c0x0 = vreinterpret_s8_s32(va30x0);
115 const int8x8_t va3c0x1 = vreinterpret_s8_s32(va30x1);
116
117 int16x8_t vprod0x01c0 = vmull_s8(vb01c0x0, va0c0x0);
118 int16x8_t vprod1x01c0 = vmull_s8(vb01c0x0, va1c0x0);
119 int16x8_t vprod2x01c0 = vmull_s8(vb01c0x0, va2c0x0);
120 int16x8_t vprod3x01c0 = vmull_s8(vb01c0x0, va3c0x0);
121 const int8x8_t vb01c0x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
122 vprod0x01c0 = vmlal_s8(vprod0x01c0, vb01c0x1, va0c0x1);
123 vprod1x01c0 = vmlal_s8(vprod1x01c0, vb01c0x1, va1c0x1);
124 vprod2x01c0 = vmlal_s8(vprod2x01c0, vb01c0x1, va2c0x1);
125 vprod3x01c0 = vmlal_s8(vprod3x01c0, vb01c0x1, va3c0x1);
126 vacc0x01 = vpadalq_s16(vacc0x01, vprod0x01c0);
127 vacc1x01 = vpadalq_s16(vacc1x01, vprod1x01c0);
128 vacc2x01 = vpadalq_s16(vacc2x01, vprod2x01c0);
129 vacc3x01 = vpadalq_s16(vacc3x01, vprod3x01c0);
130 int16x8_t vprod0x23c0 = vmull_s8(vb23c0x0, va0c0x0);
131 int16x8_t vprod1x23c0 = vmull_s8(vb23c0x0, va1c0x0);
132 int16x8_t vprod2x23c0 = vmull_s8(vb23c0x0, va2c0x0);
133 int16x8_t vprod3x23c0 = vmull_s8(vb23c0x0, va3c0x0);
134 const int8x8_t vb23c0x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
135 vprod0x23c0 = vmlal_s8(vprod0x23c0, vb23c0x1, va0c0x1);
136 vprod1x23c0 = vmlal_s8(vprod1x23c0, vb23c0x1, va1c0x1);
137 vprod2x23c0 = vmlal_s8(vprod2x23c0, vb23c0x1, va2c0x1);
138 vprod3x23c0 = vmlal_s8(vprod3x23c0, vb23c0x1, va3c0x1);
139 vacc0x23 = vpadalq_s16(vacc0x23, vprod0x23c0);
140 vacc1x23 = vpadalq_s16(vacc1x23, vprod1x23c0);
141 vacc2x23 = vpadalq_s16(vacc2x23, vprod2x23c0);
142 vacc3x23 = vpadalq_s16(vacc3x23, vprod3x23c0);
143 int16x8_t vprod0x45c0 = vmull_s8(vb45c0x0, va0c0x0);
144 int16x8_t vprod1x45c0 = vmull_s8(vb45c0x0, va1c0x0);
145 int16x8_t vprod2x45c0 = vmull_s8(vb45c0x0, va2c0x0);
146 int16x8_t vprod3x45c0 = vmull_s8(vb45c0x0, va3c0x0);
147 const int8x8_t vb45c0x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
148 vprod0x45c0 = vmlal_s8(vprod0x45c0, vb45c0x1, va0c0x1);
149 vprod1x45c0 = vmlal_s8(vprod1x45c0, vb45c0x1, va1c0x1);
150 vprod2x45c0 = vmlal_s8(vprod2x45c0, vb45c0x1, va2c0x1);
151 vprod3x45c0 = vmlal_s8(vprod3x45c0, vb45c0x1, va3c0x1);
152 vacc0x45 = vpadalq_s16(vacc0x45, vprod0x45c0);
153 vacc1x45 = vpadalq_s16(vacc1x45, vprod1x45c0);
154 vacc2x45 = vpadalq_s16(vacc2x45, vprod2x45c0);
155 vacc3x45 = vpadalq_s16(vacc3x45, vprod3x45c0);
156 int16x8_t vprod0x67c0 = vmull_s8(vb67c0x0, va0c0x0);
157 int16x8_t vprod1x67c0 = vmull_s8(vb67c0x0, va1c0x0);
158 int16x8_t vprod2x67c0 = vmull_s8(vb67c0x0, va2c0x0);
159 int16x8_t vprod3x67c0 = vmull_s8(vb67c0x0, va3c0x0);
160 const int8x8_t vb67c0x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
161 vprod0x67c0 = vmlal_s8(vprod0x67c0, vb67c0x1, va0c0x1);
162 vprod1x67c0 = vmlal_s8(vprod1x67c0, vb67c0x1, va1c0x1);
163 vprod2x67c0 = vmlal_s8(vprod2x67c0, vb67c0x1, va2c0x1);
164 vprod3x67c0 = vmlal_s8(vprod3x67c0, vb67c0x1, va3c0x1);
165 vacc0x67 = vpadalq_s16(vacc0x67, vprod0x67c0);
166 vacc1x67 = vpadalq_s16(vacc1x67, vprod1x67c0);
167 vacc2x67 = vpadalq_s16(vacc2x67, vprod2x67c0);
168 vacc3x67 = vpadalq_s16(vacc3x67, vprod3x67c0);
169 const int8x8_t va0c1x0 = vreinterpret_s8_s32(va01x0);
170 const int8x8_t va0c1x1 = vreinterpret_s8_s32(va01x1);
171 const int8x8_t va1c1x0 = vreinterpret_s8_s32(va11x0);
172 const int8x8_t va1c1x1 = vreinterpret_s8_s32(va11x1);
173 const int8x8_t va2c1x0 = vreinterpret_s8_s32(va21x0);
174 const int8x8_t va2c1x1 = vreinterpret_s8_s32(va21x1);
175 const int8x8_t va3c1x0 = vreinterpret_s8_s32(va31x0);
176 const int8x8_t va3c1x1 = vreinterpret_s8_s32(va31x1);
177
178 int16x8_t vprod0x01c1 = vmull_s8(vb01c1x0, va0c1x0);
179 int16x8_t vprod1x01c1 = vmull_s8(vb01c1x0, va1c1x0);
180 int16x8_t vprod2x01c1 = vmull_s8(vb01c1x0, va2c1x0);
181 int16x8_t vprod3x01c1 = vmull_s8(vb01c1x0, va3c1x0);
182 const int8x8_t vb01c1x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
183 vprod0x01c1 = vmlal_s8(vprod0x01c1, vb01c1x1, va0c1x1);
184 vprod1x01c1 = vmlal_s8(vprod1x01c1, vb01c1x1, va1c1x1);
185 vprod2x01c1 = vmlal_s8(vprod2x01c1, vb01c1x1, va2c1x1);
186 vprod3x01c1 = vmlal_s8(vprod3x01c1, vb01c1x1, va3c1x1);
187 vacc0x01 = vpadalq_s16(vacc0x01, vprod0x01c1);
188 vacc1x01 = vpadalq_s16(vacc1x01, vprod1x01c1);
189 vacc2x01 = vpadalq_s16(vacc2x01, vprod2x01c1);
190 vacc3x01 = vpadalq_s16(vacc3x01, vprod3x01c1);
191 int16x8_t vprod0x23c1 = vmull_s8(vb23c1x0, va0c1x0);
192 int16x8_t vprod1x23c1 = vmull_s8(vb23c1x0, va1c1x0);
193 int16x8_t vprod2x23c1 = vmull_s8(vb23c1x0, va2c1x0);
194 int16x8_t vprod3x23c1 = vmull_s8(vb23c1x0, va3c1x0);
195 const int8x8_t vb23c1x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
196 vprod0x23c1 = vmlal_s8(vprod0x23c1, vb23c1x1, va0c1x1);
197 vprod1x23c1 = vmlal_s8(vprod1x23c1, vb23c1x1, va1c1x1);
198 vprod2x23c1 = vmlal_s8(vprod2x23c1, vb23c1x1, va2c1x1);
199 vprod3x23c1 = vmlal_s8(vprod3x23c1, vb23c1x1, va3c1x1);
200 vacc0x23 = vpadalq_s16(vacc0x23, vprod0x23c1);
201 vacc1x23 = vpadalq_s16(vacc1x23, vprod1x23c1);
202 vacc2x23 = vpadalq_s16(vacc2x23, vprod2x23c1);
203 vacc3x23 = vpadalq_s16(vacc3x23, vprod3x23c1);
204 int16x8_t vprod0x45c1 = vmull_s8(vb45c1x0, va0c1x0);
205 int16x8_t vprod1x45c1 = vmull_s8(vb45c1x0, va1c1x0);
206 int16x8_t vprod2x45c1 = vmull_s8(vb45c1x0, va2c1x0);
207 int16x8_t vprod3x45c1 = vmull_s8(vb45c1x0, va3c1x0);
208 const int8x8_t vb45c1x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
209 vprod0x45c1 = vmlal_s8(vprod0x45c1, vb45c1x1, va0c1x1);
210 vprod1x45c1 = vmlal_s8(vprod1x45c1, vb45c1x1, va1c1x1);
211 vprod2x45c1 = vmlal_s8(vprod2x45c1, vb45c1x1, va2c1x1);
212 vprod3x45c1 = vmlal_s8(vprod3x45c1, vb45c1x1, va3c1x1);
213 vacc0x45 = vpadalq_s16(vacc0x45, vprod0x45c1);
214 vacc1x45 = vpadalq_s16(vacc1x45, vprod1x45c1);
215 vacc2x45 = vpadalq_s16(vacc2x45, vprod2x45c1);
216 vacc3x45 = vpadalq_s16(vacc3x45, vprod3x45c1);
217 int16x8_t vprod0x67c1 = vmull_s8(vb67c1x0, va0c1x0);
218 int16x8_t vprod1x67c1 = vmull_s8(vb67c1x0, va1c1x0);
219 int16x8_t vprod2x67c1 = vmull_s8(vb67c1x0, va2c1x0);
220 int16x8_t vprod3x67c1 = vmull_s8(vb67c1x0, va3c1x0);
221 const int8x8_t vb67c1x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
222 vprod0x67c1 = vmlal_s8(vprod0x67c1, vb67c1x1, va0c1x1);
223 vprod1x67c1 = vmlal_s8(vprod1x67c1, vb67c1x1, va1c1x1);
224 vprod2x67c1 = vmlal_s8(vprod2x67c1, vb67c1x1, va2c1x1);
225 vprod3x67c1 = vmlal_s8(vprod3x67c1, vb67c1x1, va3c1x1);
226 vacc0x67 = vpadalq_s16(vacc0x67, vprod0x67c1);
227 vacc1x67 = vpadalq_s16(vacc1x67, vprod1x67c1);
228 vacc2x67 = vpadalq_s16(vacc2x67, vprod2x67c1);
229 vacc3x67 = vpadalq_s16(vacc3x67, vprod3x67c1);
230
231 k -= 16 * sizeof(int8_t);
232 }
233
234 if (k >= 8 * sizeof(int8_t)) {
235 const int32x2_t va00 = vld1_dup_s32((const void*)a0);
236 const int32x2_t va01 = vld1_dup_s32((const void*)(a0 + 4)); a0 += 8;
237 const int32x2_t va10 = vld1_dup_s32((const void*)a1);
238 const int32x2_t va11 = vld1_dup_s32((const void*)(a1 + 4)); a1 += 8;
239 const int32x2_t va20 = vld1_dup_s32((const void*)a2);
240 const int32x2_t va21 = vld1_dup_s32((const void*)(a2 + 4)); a2 += 8;
241 const int32x2_t va30 = vld1_dup_s32((const void*)a3);
242 const int32x2_t va31 = vld1_dup_s32((const void*)(a3 + 4)); a3 += 8;
243
244 const int8x8_t vb01c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
245 const int8x8_t vb23c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
246 const int8x8_t vb45c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
247 const int8x8_t vb67c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
248 const int8x8_t vb01c1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
249 const int8x8_t vb23c1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
250 const int8x8_t vb45c1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
251 const int8x8_t vb67c1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
252
253 const int8x8_t va0c0 = vreinterpret_s8_s32(va00);
254 const int8x8_t va1c0 = vreinterpret_s8_s32(va10);
255 const int8x8_t va2c0 = vreinterpret_s8_s32(va20);
256 const int8x8_t va3c0 = vreinterpret_s8_s32(va30);
257
258 const int16x8_t vprod0x01c0 = vmull_s8(vb01c0, va0c0);
259 const int16x8_t vprod1x01c0 = vmull_s8(vb01c0, va1c0);
260 const int16x8_t vprod2x01c0 = vmull_s8(vb01c0, va2c0);
261 const int16x8_t vprod3x01c0 = vmull_s8(vb01c0, va3c0);
262 vacc0x01 = vpadalq_s16(vacc0x01, vprod0x01c0);
263 vacc1x01 = vpadalq_s16(vacc1x01, vprod1x01c0);
264 vacc2x01 = vpadalq_s16(vacc2x01, vprod2x01c0);
265 vacc3x01 = vpadalq_s16(vacc3x01, vprod3x01c0);
266 const int16x8_t vprod0x23c0 = vmull_s8(vb23c0, va0c0);
267 const int16x8_t vprod1x23c0 = vmull_s8(vb23c0, va1c0);
268 const int16x8_t vprod2x23c0 = vmull_s8(vb23c0, va2c0);
269 const int16x8_t vprod3x23c0 = vmull_s8(vb23c0, va3c0);
270 vacc0x23 = vpadalq_s16(vacc0x23, vprod0x23c0);
271 vacc1x23 = vpadalq_s16(vacc1x23, vprod1x23c0);
272 vacc2x23 = vpadalq_s16(vacc2x23, vprod2x23c0);
273 vacc3x23 = vpadalq_s16(vacc3x23, vprod3x23c0);
274 const int16x8_t vprod0x45c0 = vmull_s8(vb45c0, va0c0);
275 const int16x8_t vprod1x45c0 = vmull_s8(vb45c0, va1c0);
276 const int16x8_t vprod2x45c0 = vmull_s8(vb45c0, va2c0);
277 const int16x8_t vprod3x45c0 = vmull_s8(vb45c0, va3c0);
278 vacc0x45 = vpadalq_s16(vacc0x45, vprod0x45c0);
279 vacc1x45 = vpadalq_s16(vacc1x45, vprod1x45c0);
280 vacc2x45 = vpadalq_s16(vacc2x45, vprod2x45c0);
281 vacc3x45 = vpadalq_s16(vacc3x45, vprod3x45c0);
282 const int16x8_t vprod0x67c0 = vmull_s8(vb67c0, va0c0);
283 const int16x8_t vprod1x67c0 = vmull_s8(vb67c0, va1c0);
284 const int16x8_t vprod2x67c0 = vmull_s8(vb67c0, va2c0);
285 const int16x8_t vprod3x67c0 = vmull_s8(vb67c0, va3c0);
286 vacc0x67 = vpadalq_s16(vacc0x67, vprod0x67c0);
287 vacc1x67 = vpadalq_s16(vacc1x67, vprod1x67c0);
288 vacc2x67 = vpadalq_s16(vacc2x67, vprod2x67c0);
289 vacc3x67 = vpadalq_s16(vacc3x67, vprod3x67c0);
290 const int8x8_t va0c1 = vreinterpret_s8_s32(va01);
291 const int8x8_t va1c1 = vreinterpret_s8_s32(va11);
292 const int8x8_t va2c1 = vreinterpret_s8_s32(va21);
293 const int8x8_t va3c1 = vreinterpret_s8_s32(va31);
294
295 const int16x8_t vprod0x01c1 = vmull_s8(vb01c1, va0c1);
296 const int16x8_t vprod1x01c1 = vmull_s8(vb01c1, va1c1);
297 const int16x8_t vprod2x01c1 = vmull_s8(vb01c1, va2c1);
298 const int16x8_t vprod3x01c1 = vmull_s8(vb01c1, va3c1);
299 vacc0x01 = vpadalq_s16(vacc0x01, vprod0x01c1);
300 vacc1x01 = vpadalq_s16(vacc1x01, vprod1x01c1);
301 vacc2x01 = vpadalq_s16(vacc2x01, vprod2x01c1);
302 vacc3x01 = vpadalq_s16(vacc3x01, vprod3x01c1);
303 const int16x8_t vprod0x23c1 = vmull_s8(vb23c1, va0c1);
304 const int16x8_t vprod1x23c1 = vmull_s8(vb23c1, va1c1);
305 const int16x8_t vprod2x23c1 = vmull_s8(vb23c1, va2c1);
306 const int16x8_t vprod3x23c1 = vmull_s8(vb23c1, va3c1);
307 vacc0x23 = vpadalq_s16(vacc0x23, vprod0x23c1);
308 vacc1x23 = vpadalq_s16(vacc1x23, vprod1x23c1);
309 vacc2x23 = vpadalq_s16(vacc2x23, vprod2x23c1);
310 vacc3x23 = vpadalq_s16(vacc3x23, vprod3x23c1);
311 const int16x8_t vprod0x45c1 = vmull_s8(vb45c1, va0c1);
312 const int16x8_t vprod1x45c1 = vmull_s8(vb45c1, va1c1);
313 const int16x8_t vprod2x45c1 = vmull_s8(vb45c1, va2c1);
314 const int16x8_t vprod3x45c1 = vmull_s8(vb45c1, va3c1);
315 vacc0x45 = vpadalq_s16(vacc0x45, vprod0x45c1);
316 vacc1x45 = vpadalq_s16(vacc1x45, vprod1x45c1);
317 vacc2x45 = vpadalq_s16(vacc2x45, vprod2x45c1);
318 vacc3x45 = vpadalq_s16(vacc3x45, vprod3x45c1);
319 const int16x8_t vprod0x67c1 = vmull_s8(vb67c1, va0c1);
320 const int16x8_t vprod1x67c1 = vmull_s8(vb67c1, va1c1);
321 const int16x8_t vprod2x67c1 = vmull_s8(vb67c1, va2c1);
322 const int16x8_t vprod3x67c1 = vmull_s8(vb67c1, va3c1);
323 vacc0x67 = vpadalq_s16(vacc0x67, vprod0x67c1);
324 vacc1x67 = vpadalq_s16(vacc1x67, vprod1x67c1);
325 vacc2x67 = vpadalq_s16(vacc2x67, vprod2x67c1);
326 vacc3x67 = vpadalq_s16(vacc3x67, vprod3x67c1);
327
328 k -= 8 * sizeof(int8_t);
329 }
330
331 if XNN_UNLIKELY(k != 0) {
332 const int8x8_t va0 = vld1_s8(a0); a0 = (const int8_t*) ((uintptr_t) a0 + k);
333 const int8x8_t va1 = vld1_s8(a1); a1 = (const int8_t*) ((uintptr_t) a1 + k);
334 const int8x8_t va2 = vld1_s8(a2); a2 = (const int8_t*) ((uintptr_t) a2 + k);
335 const int8x8_t va3 = vld1_s8(a3); a3 = (const int8_t*) ((uintptr_t) a3 + k);
336
337 const int8x8_t vb01c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
338 const int8x8_t vb23c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
339 const int8x8_t vb45c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
340 const int8x8_t vb67c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
341
342 const int8x8_t va0c0 = vreinterpret_s8_s32(vdup_lane_s32(vreinterpret_s32_s8(va0), 0));
343 const int16x8_t vprod0x01c0 = vmull_s8(vb01c0, va0c0);
344 vacc0x01 = vpadalq_s16(vacc0x01, vprod0x01c0);
345 const int16x8_t vprod0x23c0 = vmull_s8(vb23c0, va0c0);
346 vacc0x23 = vpadalq_s16(vacc0x23, vprod0x23c0);
347 const int16x8_t vprod0x45c0 = vmull_s8(vb45c0, va0c0);
348 vacc0x45 = vpadalq_s16(vacc0x45, vprod0x45c0);
349 const int16x8_t vprod0x67c0 = vmull_s8(vb67c0, va0c0);
350 vacc0x67 = vpadalq_s16(vacc0x67, vprod0x67c0);
351 const int8x8_t va1c0 = vreinterpret_s8_s32(vdup_lane_s32(vreinterpret_s32_s8(va1), 0));
352 const int16x8_t vprod1x01c0 = vmull_s8(vb01c0, va1c0);
353 vacc1x01 = vpadalq_s16(vacc1x01, vprod1x01c0);
354 const int16x8_t vprod1x23c0 = vmull_s8(vb23c0, va1c0);
355 vacc1x23 = vpadalq_s16(vacc1x23, vprod1x23c0);
356 const int16x8_t vprod1x45c0 = vmull_s8(vb45c0, va1c0);
357 vacc1x45 = vpadalq_s16(vacc1x45, vprod1x45c0);
358 const int16x8_t vprod1x67c0 = vmull_s8(vb67c0, va1c0);
359 vacc1x67 = vpadalq_s16(vacc1x67, vprod1x67c0);
360 const int8x8_t va2c0 = vreinterpret_s8_s32(vdup_lane_s32(vreinterpret_s32_s8(va2), 0));
361 const int16x8_t vprod2x01c0 = vmull_s8(vb01c0, va2c0);
362 vacc2x01 = vpadalq_s16(vacc2x01, vprod2x01c0);
363 const int16x8_t vprod2x23c0 = vmull_s8(vb23c0, va2c0);
364 vacc2x23 = vpadalq_s16(vacc2x23, vprod2x23c0);
365 const int16x8_t vprod2x45c0 = vmull_s8(vb45c0, va2c0);
366 vacc2x45 = vpadalq_s16(vacc2x45, vprod2x45c0);
367 const int16x8_t vprod2x67c0 = vmull_s8(vb67c0, va2c0);
368 vacc2x67 = vpadalq_s16(vacc2x67, vprod2x67c0);
369 const int8x8_t va3c0 = vreinterpret_s8_s32(vdup_lane_s32(vreinterpret_s32_s8(va3), 0));
370 const int16x8_t vprod3x01c0 = vmull_s8(vb01c0, va3c0);
371 vacc3x01 = vpadalq_s16(vacc3x01, vprod3x01c0);
372 const int16x8_t vprod3x23c0 = vmull_s8(vb23c0, va3c0);
373 vacc3x23 = vpadalq_s16(vacc3x23, vprod3x23c0);
374 const int16x8_t vprod3x45c0 = vmull_s8(vb45c0, va3c0);
375 vacc3x45 = vpadalq_s16(vacc3x45, vprod3x45c0);
376 const int16x8_t vprod3x67c0 = vmull_s8(vb67c0, va3c0);
377 vacc3x67 = vpadalq_s16(vacc3x67, vprod3x67c0);
378 }
379
380 #if XNN_ARCH_ARM64
381 int32x4_t vacc0x0123 = vpaddq_s32(vacc0x01, vacc0x23);
382 int32x4_t vacc0x4567 = vpaddq_s32(vacc0x45, vacc0x67);
383 int32x4_t vacc1x0123 = vpaddq_s32(vacc1x01, vacc1x23);
384 int32x4_t vacc1x4567 = vpaddq_s32(vacc1x45, vacc1x67);
385 int32x4_t vacc2x0123 = vpaddq_s32(vacc2x01, vacc2x23);
386 int32x4_t vacc2x4567 = vpaddq_s32(vacc2x45, vacc2x67);
387 int32x4_t vacc3x0123 = vpaddq_s32(vacc3x01, vacc3x23);
388 int32x4_t vacc3x4567 = vpaddq_s32(vacc3x45, vacc3x67);
389 #else
390 const int32x2_t vsum0x01 = vpadd_s32(vget_low_s32(vacc0x01), vget_high_s32(vacc0x01));
391 const int32x2_t vsum0x23 = vpadd_s32(vget_low_s32(vacc0x23), vget_high_s32(vacc0x23));
392 int32x4_t vacc0x0123 = vcombine_s32(vsum0x01, vsum0x23);
393 const int32x2_t vsum0x45 = vpadd_s32(vget_low_s32(vacc0x45), vget_high_s32(vacc0x45));
394 const int32x2_t vsum0x67 = vpadd_s32(vget_low_s32(vacc0x67), vget_high_s32(vacc0x67));
395 int32x4_t vacc0x4567 = vcombine_s32(vsum0x45, vsum0x67);
396 const int32x2_t vsum1x01 = vpadd_s32(vget_low_s32(vacc1x01), vget_high_s32(vacc1x01));
397 const int32x2_t vsum1x23 = vpadd_s32(vget_low_s32(vacc1x23), vget_high_s32(vacc1x23));
398 int32x4_t vacc1x0123 = vcombine_s32(vsum1x01, vsum1x23);
399 const int32x2_t vsum1x45 = vpadd_s32(vget_low_s32(vacc1x45), vget_high_s32(vacc1x45));
400 const int32x2_t vsum1x67 = vpadd_s32(vget_low_s32(vacc1x67), vget_high_s32(vacc1x67));
401 int32x4_t vacc1x4567 = vcombine_s32(vsum1x45, vsum1x67);
402 const int32x2_t vsum2x01 = vpadd_s32(vget_low_s32(vacc2x01), vget_high_s32(vacc2x01));
403 const int32x2_t vsum2x23 = vpadd_s32(vget_low_s32(vacc2x23), vget_high_s32(vacc2x23));
404 int32x4_t vacc2x0123 = vcombine_s32(vsum2x01, vsum2x23);
405 const int32x2_t vsum2x45 = vpadd_s32(vget_low_s32(vacc2x45), vget_high_s32(vacc2x45));
406 const int32x2_t vsum2x67 = vpadd_s32(vget_low_s32(vacc2x67), vget_high_s32(vacc2x67));
407 int32x4_t vacc2x4567 = vcombine_s32(vsum2x45, vsum2x67);
408 const int32x2_t vsum3x01 = vpadd_s32(vget_low_s32(vacc3x01), vget_high_s32(vacc3x01));
409 const int32x2_t vsum3x23 = vpadd_s32(vget_low_s32(vacc3x23), vget_high_s32(vacc3x23));
410 int32x4_t vacc3x0123 = vcombine_s32(vsum3x01, vsum3x23);
411 const int32x2_t vsum3x45 = vpadd_s32(vget_low_s32(vacc3x45), vget_high_s32(vacc3x45));
412 const int32x2_t vsum3x67 = vpadd_s32(vget_low_s32(vacc3x67), vget_high_s32(vacc3x67));
413 int32x4_t vacc3x4567 = vcombine_s32(vsum3x45, vsum3x67);
414 #endif
415
416 const int32x4_t vright_pre_shift = vld1q_dup_s32(¶ms->rndnu_neon.right_pre_shift);
417 const int32x4_t vmultiplier = vld1q_dup_s32(¶ms->rndnu_neon.multiplier);
418 const int32x4_t vright_post_shift = vld1q_dup_s32(¶ms->rndnu_neon.right_post_shift);
419
420 vacc0x0123 = vqshlq_s32(vacc0x0123, vright_pre_shift);
421 vacc0x4567 = vqshlq_s32(vacc0x4567, vright_pre_shift);
422 vacc1x0123 = vqshlq_s32(vacc1x0123, vright_pre_shift);
423 vacc1x4567 = vqshlq_s32(vacc1x4567, vright_pre_shift);
424 vacc2x0123 = vqshlq_s32(vacc2x0123, vright_pre_shift);
425 vacc2x4567 = vqshlq_s32(vacc2x4567, vright_pre_shift);
426 vacc3x0123 = vqshlq_s32(vacc3x0123, vright_pre_shift);
427 vacc3x4567 = vqshlq_s32(vacc3x4567, vright_pre_shift);
428
429 vacc0x0123 = vqdmulhq_s32(vacc0x0123, vmultiplier);
430 vacc0x4567 = vqdmulhq_s32(vacc0x4567, vmultiplier);
431 vacc1x0123 = vqdmulhq_s32(vacc1x0123, vmultiplier);
432 vacc1x4567 = vqdmulhq_s32(vacc1x4567, vmultiplier);
433 vacc2x0123 = vqdmulhq_s32(vacc2x0123, vmultiplier);
434 vacc2x4567 = vqdmulhq_s32(vacc2x4567, vmultiplier);
435 vacc3x0123 = vqdmulhq_s32(vacc3x0123, vmultiplier);
436 vacc3x4567 = vqdmulhq_s32(vacc3x4567, vmultiplier);
437
438 vacc0x0123 = vrshlq_s32(vacc0x0123, vright_post_shift);
439 vacc0x4567 = vrshlq_s32(vacc0x4567, vright_post_shift);
440 vacc1x0123 = vrshlq_s32(vacc1x0123, vright_post_shift);
441 vacc1x4567 = vrshlq_s32(vacc1x4567, vright_post_shift);
442 vacc2x0123 = vrshlq_s32(vacc2x0123, vright_post_shift);
443 vacc2x4567 = vrshlq_s32(vacc2x4567, vright_post_shift);
444 vacc3x0123 = vrshlq_s32(vacc3x0123, vright_post_shift);
445 vacc3x4567 = vrshlq_s32(vacc3x4567, vright_post_shift);
446
447 const int16x8_t voutput_zero_point = vld1q_dup_s16(¶ms->rndnu_neon.output_zero_point);
448 #if XNN_ARCH_ARM64
449 int16x8_t vacc0x01234567 = vqmovn_high_s32(vqmovn_s32(vacc0x0123), vacc0x4567);
450 int16x8_t vacc1x01234567 = vqmovn_high_s32(vqmovn_s32(vacc1x0123), vacc1x4567);
451 int16x8_t vacc2x01234567 = vqmovn_high_s32(vqmovn_s32(vacc2x0123), vacc2x4567);
452 int16x8_t vacc3x01234567 = vqmovn_high_s32(vqmovn_s32(vacc3x0123), vacc3x4567);
453
454 vacc0x01234567 = vqaddq_s16(vacc0x01234567, voutput_zero_point);
455 vacc1x01234567 = vqaddq_s16(vacc1x01234567, voutput_zero_point);
456 vacc2x01234567 = vqaddq_s16(vacc2x01234567, voutput_zero_point);
457 vacc3x01234567 = vqaddq_s16(vacc3x01234567, voutput_zero_point);
458
459 int8x16_t vout0x01234567_1x01234567 = vqmovn_high_s16(vqmovn_s16(vacc0x01234567), vacc1x01234567);
460 int8x16_t vout2x01234567_3x01234567 = vqmovn_high_s16(vqmovn_s16(vacc2x01234567), vacc3x01234567);
461 #else
462 int16x8_t vacc0x01234567 = vcombine_s16(vqmovn_s32(vacc0x0123), vqmovn_s32(vacc0x4567));
463 int16x8_t vacc1x01234567 = vcombine_s16(vqmovn_s32(vacc1x0123), vqmovn_s32(vacc1x4567));
464 int16x8_t vacc2x01234567 = vcombine_s16(vqmovn_s32(vacc2x0123), vqmovn_s32(vacc2x4567));
465 int16x8_t vacc3x01234567 = vcombine_s16(vqmovn_s32(vacc3x0123), vqmovn_s32(vacc3x4567));
466
467 vacc0x01234567 = vqaddq_s16(vacc0x01234567, voutput_zero_point);
468 vacc1x01234567 = vqaddq_s16(vacc1x01234567, voutput_zero_point);
469 vacc2x01234567 = vqaddq_s16(vacc2x01234567, voutput_zero_point);
470 vacc3x01234567 = vqaddq_s16(vacc3x01234567, voutput_zero_point);
471
472 int8x16_t vout0x01234567_1x01234567 = vcombine_s8(vqmovn_s16(vacc0x01234567), vqmovn_s16(vacc1x01234567));
473 int8x16_t vout2x01234567_3x01234567 = vcombine_s8(vqmovn_s16(vacc2x01234567), vqmovn_s16(vacc3x01234567));
474 #endif
475
476 const int8x16_t voutput_min = vld1q_dup_s8(¶ms->rndnu_neon.output_min);
477 vout0x01234567_1x01234567 = vmaxq_s8(vout0x01234567_1x01234567, voutput_min);
478 vout2x01234567_3x01234567 = vmaxq_s8(vout2x01234567_3x01234567, voutput_min);
479
480 const int8x16_t voutput_max = vld1q_dup_s8(¶ms->rndnu_neon.output_max);
481 vout0x01234567_1x01234567 = vminq_s8(vout0x01234567_1x01234567, voutput_max);
482 vout2x01234567_3x01234567 = vminq_s8(vout2x01234567_3x01234567, voutput_max);
483
484 if (nc >= 8) {
485 vst1_s8(c0 + 0, vget_low_s8(vout0x01234567_1x01234567));
486 vst1_s8(c1 + 0, vget_high_s8(vout0x01234567_1x01234567));
487 vst1_s8(c2 + 0, vget_low_s8(vout2x01234567_3x01234567));
488 vst1_s8(c3 + 0, vget_high_s8(vout2x01234567_3x01234567));
489
490 c0 = (int8_t*) ((uintptr_t) c0 + cn_stride);
491 c1 = (int8_t*) ((uintptr_t) c1 + cn_stride);
492 c2 = (int8_t*) ((uintptr_t) c2 + cn_stride);
493 c3 = (int8_t*) ((uintptr_t) c3 + cn_stride);
494
495 a0 = (const int8_t*) ((uintptr_t) a0 - kc);
496 a1 = (const int8_t*) ((uintptr_t) a1 - kc);
497 a2 = (const int8_t*) ((uintptr_t) a2 - kc);
498 a3 = (const int8_t*) ((uintptr_t) a3 - kc);
499
500 nc -= 8;
501 } else {
502 // Final case where not all of the 8 columns fit in the destination.
503 if (nc & 4) {
504 vst1q_lane_u32((void*) c0, vreinterpretq_u32_s8(vout0x01234567_1x01234567), 0); c0 += 4;
505 vst1q_lane_u32((void*) c1, vreinterpretq_u32_s8(vout0x01234567_1x01234567), 2); c1 += 4;
506 vst1q_lane_u32((void*) c2, vreinterpretq_u32_s8(vout2x01234567_3x01234567), 0); c2 += 4;
507 vst1q_lane_u32((void*) c3, vreinterpretq_u32_s8(vout2x01234567_3x01234567), 2); c3 += 4;
508 vout0x01234567_1x01234567 = vextq_s8(vout0x01234567_1x01234567, vout0x01234567_1x01234567, 4);
509 vout2x01234567_3x01234567 = vextq_s8(vout2x01234567_3x01234567, vout2x01234567_3x01234567, 4);
510 }
511 if (nc & 2) {
512 vst1q_lane_u16((void*) c0, vreinterpretq_u16_s8(vout0x01234567_1x01234567), 0); c0 += 2;
513 vst1q_lane_u16((void*) c1, vreinterpretq_u16_s8(vout0x01234567_1x01234567), 4); c1 += 2;
514 vst1q_lane_u16((void*) c2, vreinterpretq_u16_s8(vout2x01234567_3x01234567), 0); c2 += 2;
515 vst1q_lane_u16((void*) c3, vreinterpretq_u16_s8(vout2x01234567_3x01234567), 4); c3 += 2;
516 vout0x01234567_1x01234567 = vextq_s8(vout0x01234567_1x01234567, vout0x01234567_1x01234567, 2);
517 vout2x01234567_3x01234567 = vextq_s8(vout2x01234567_3x01234567, vout2x01234567_3x01234567, 2);
518 }
519 if (nc & 1) {
520 vst1q_lane_s8(c0, vout0x01234567_1x01234567, 0);
521 vst1q_lane_s8(c1, vout0x01234567_1x01234567, 8);
522 vst1q_lane_s8(c2, vout2x01234567_3x01234567, 0);
523 vst1q_lane_s8(c3, vout2x01234567_3x01234567, 8);
524 }
525
526 nc = 0;
527 }
528 } while (nc != 0);
529 }
530