1 // Auto-generated file. Do not edit!
2 // Template: src/qu8-igemm/c4-neondot.c.in
3 // Generator: tools/xngen
4 //
5 // Copyright 2020 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/igemm.h>
15 #include <xnnpack/math.h>
16
17
xnn_qu8_igemm_minmax_rndnu_ukernel_5x16c4__neondot(size_t mr,size_t nc,size_t kc,size_t ks,const uint8_t ** restrict a,const void * restrict w,uint8_t * restrict c,size_t cm_stride,size_t cn_stride,size_t a_offset,const uint8_t * zero,const union xnn_qu8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS (1)])18 void xnn_qu8_igemm_minmax_rndnu_ukernel_5x16c4__neondot(
19 size_t mr,
20 size_t nc,
21 size_t kc,
22 size_t ks,
23 const uint8_t** restrict a,
24 const void* restrict w,
25 uint8_t* restrict c,
26 size_t cm_stride,
27 size_t cn_stride,
28 size_t a_offset,
29 const uint8_t* zero,
30 const union xnn_qu8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
31 {
32 assert(mr != 0);
33 assert(mr <= 5);
34 assert(nc != 0);
35 assert(kc != 0);
36 assert(ks != 0);
37 assert(ks % (5 * sizeof(void*)) == 0);
38 assert(a_offset % sizeof(uint8_t) == 0);
39 assert(a != NULL);
40 assert(w != NULL);
41 assert(c != NULL);
42
43 kc = round_up_po2(kc, 4 * sizeof(uint8_t));
44 uint8_t* c0 = c;
45 uint8_t* c1 = (uint8_t*) ((uintptr_t) c0 + cm_stride);
46 if XNN_UNPREDICTABLE(mr < 2) {
47 c1 = c0;
48 }
49 uint8_t* c2 = (uint8_t*) ((uintptr_t) c1 + cm_stride);
50 if XNN_UNPREDICTABLE(mr <= 2) {
51 c2 = c1;
52 }
53 uint8_t* c3 = (uint8_t*) ((uintptr_t) c2 + cm_stride);
54 if XNN_UNPREDICTABLE(mr < 4) {
55 c3 = c2;
56 }
57 uint8_t* c4 = (uint8_t*) ((uintptr_t) c3 + cm_stride);
58 if XNN_UNPREDICTABLE(mr <= 4) {
59 c4 = c3;
60 }
61
62 const uint8x8_t va_zero_point = vld1_dup_u8(¶ms->rndnu_neon.kernel_zero_point[0]);
63
64 do {
65 // Initialize accumulators with bias. 16 bias values are loaded from the
66 // weight matrix, at the start of the group of 16 columns.
67 uint32x4_t vpacc0x0123 = vld1q_u32(w); w = (const void*) ((const uint32_t*) w + 4);
68 uint32x4_t vpacc0x4567 = vld1q_u32(w); w = (const void*) ((const uint32_t*) w + 4);
69 uint32x4_t vpacc0x89AB = vld1q_u32(w); w = (const void*) ((const uint32_t*) w + 4);
70 uint32x4_t vpacc0xCDEF = vld1q_u32(w); w = (const void*) ((const uint32_t*) w + 4);
71 uint32x4_t vpacc1x0123 = vpacc0x0123;
72 uint32x4_t vpacc1x4567 = vpacc0x4567;
73 uint32x4_t vpacc1x89AB = vpacc0x89AB;
74 uint32x4_t vpacc1xCDEF = vpacc0xCDEF;
75 uint32x4_t vpacc2x0123 = vpacc0x0123;
76 uint32x4_t vpacc2x4567 = vpacc0x4567;
77 uint32x4_t vpacc2x89AB = vpacc0x89AB;
78 uint32x4_t vpacc2xCDEF = vpacc0xCDEF;
79 uint32x4_t vpacc3x0123 = vpacc0x0123;
80 uint32x4_t vpacc3x4567 = vpacc0x4567;
81 uint32x4_t vpacc3x89AB = vpacc0x89AB;
82 uint32x4_t vpacc3xCDEF = vpacc0xCDEF;
83 uint32x4_t vpacc4x0123 = vpacc0x0123;
84 uint32x4_t vpacc4x4567 = vpacc0x4567;
85 uint32x4_t vpacc4x89AB = vpacc0x89AB;
86 uint32x4_t vpacc4xCDEF = vpacc0xCDEF;
87 uint32x2_t vnacc0 = vmov_n_u32(0);
88 uint32x2_t vnacc1 = vmov_n_u32(0);
89 uint32x2_t vnacc2 = vmov_n_u32(0);
90 uint32x2_t vnacc3 = vmov_n_u32(0);
91 uint32x2_t vnacc4 = vmov_n_u32(0);
92
93 size_t p = ks;
94 do {
95 const uint8_t* restrict a0 = a[0];
96 if XNN_UNPREDICTABLE(a0 != zero) {
97 a0 = (const uint8_t*) ((uintptr_t) a0 + a_offset);
98 }
99 const uint8_t* restrict a1 = a[1];
100 if XNN_UNPREDICTABLE(a1 != zero) {
101 a1 = (const uint8_t*) ((uintptr_t) a1 + a_offset);
102 }
103 const uint8_t* restrict a2 = a[2];
104 if XNN_UNPREDICTABLE(a2 != zero) {
105 a2 = (const uint8_t*) ((uintptr_t) a2 + a_offset);
106 }
107 const uint8_t* restrict a3 = a[3];
108 if XNN_UNPREDICTABLE(a3 != zero) {
109 a3 = (const uint8_t*) ((uintptr_t) a3 + a_offset);
110 }
111 const uint8_t* restrict a4 = a[4];
112 if XNN_UNPREDICTABLE(a4 != zero) {
113 a4 = (const uint8_t*) ((uintptr_t) a4 + a_offset);
114 }
115 a += 5;
116
117 // Inner accumulation loop along the 16 columns.
118 size_t k = kc;
119 // 2x partial unrolled loop to load 8 bytes at a time.
120 while (k >= 8 * sizeof(uint8_t)) {
121 // Load a 5x8 block of activations.
122 const uint8x8_t va0x01234567 = vld1_u8(a0); a0 += 8;
123 const uint8x8_t va1x01234567 = vld1_u8(a1); a1 += 8;
124 const uint8x8_t va2x01234567 = vld1_u8(a2); a2 += 8;
125 const uint8x8_t va3x01234567 = vld1_u8(a3); a3 += 8;
126 const uint8x8_t va4x01234567 = vld1_u8(a4); a4 += 8;
127
128 // Load a 8x16 block of weights.
129 const uint8x16_t vb0123x0123 = vld1q_u8(w); w = (const void*) ((const uint8_t*) w + 16);
130 const uint8x16_t vb0123x4567 = vld1q_u8(w); w = (const void*) ((const uint8_t*) w + 16);
131 const uint8x16_t vb0123x89AB = vld1q_u8(w); w = (const void*) ((const uint8_t*) w + 16);
132 const uint8x16_t vb0123xCDEF = vld1q_u8(w); w = (const void*) ((const uint8_t*) w + 16);
133 const uint8x16_t vb4567x0123 = vld1q_u8(w); w = (const void*) ((const uint8_t*) w + 16);
134 const uint8x16_t vb4567x4567 = vld1q_u8(w); w = (const void*) ((const uint8_t*) w + 16);
135 const uint8x16_t vb4567x89AB = vld1q_u8(w); w = (const void*) ((const uint8_t*) w + 16);
136 const uint8x16_t vb4567xCDEF = vld1q_u8(w); w = (const void*) ((const uint8_t*) w + 16);
137
138 // Multiply-accumulate: 5x8 * 8x16 --> 5x16.
139 vnacc0 = vdot_u32(vnacc0, va_zero_point, va0x01234567);
140 vpacc0x0123 = vdotq_lane_u32(vpacc0x0123, vb0123x0123, va0x01234567, 0);
141 vpacc0x4567 = vdotq_lane_u32(vpacc0x4567, vb0123x4567, va0x01234567, 0);
142 vpacc0x89AB = vdotq_lane_u32(vpacc0x89AB, vb0123x89AB, va0x01234567, 0);
143 vpacc0xCDEF = vdotq_lane_u32(vpacc0xCDEF, vb0123xCDEF, va0x01234567, 0);
144 vpacc0x0123 = vdotq_lane_u32(vpacc0x0123, vb4567x0123, va0x01234567, 1);
145 vpacc0x4567 = vdotq_lane_u32(vpacc0x4567, vb4567x4567, va0x01234567, 1);
146 vpacc0x89AB = vdotq_lane_u32(vpacc0x89AB, vb4567x89AB, va0x01234567, 1);
147 vpacc0xCDEF = vdotq_lane_u32(vpacc0xCDEF, vb4567xCDEF, va0x01234567, 1);
148 vnacc1 = vdot_u32(vnacc1, va_zero_point, va1x01234567);
149 vpacc1x0123 = vdotq_lane_u32(vpacc1x0123, vb0123x0123, va1x01234567, 0);
150 vpacc1x4567 = vdotq_lane_u32(vpacc1x4567, vb0123x4567, va1x01234567, 0);
151 vpacc1x89AB = vdotq_lane_u32(vpacc1x89AB, vb0123x89AB, va1x01234567, 0);
152 vpacc1xCDEF = vdotq_lane_u32(vpacc1xCDEF, vb0123xCDEF, va1x01234567, 0);
153 vpacc1x0123 = vdotq_lane_u32(vpacc1x0123, vb4567x0123, va1x01234567, 1);
154 vpacc1x4567 = vdotq_lane_u32(vpacc1x4567, vb4567x4567, va1x01234567, 1);
155 vpacc1x89AB = vdotq_lane_u32(vpacc1x89AB, vb4567x89AB, va1x01234567, 1);
156 vpacc1xCDEF = vdotq_lane_u32(vpacc1xCDEF, vb4567xCDEF, va1x01234567, 1);
157 vnacc2 = vdot_u32(vnacc2, va_zero_point, va2x01234567);
158 vpacc2x0123 = vdotq_lane_u32(vpacc2x0123, vb0123x0123, va2x01234567, 0);
159 vpacc2x4567 = vdotq_lane_u32(vpacc2x4567, vb0123x4567, va2x01234567, 0);
160 vpacc2x89AB = vdotq_lane_u32(vpacc2x89AB, vb0123x89AB, va2x01234567, 0);
161 vpacc2xCDEF = vdotq_lane_u32(vpacc2xCDEF, vb0123xCDEF, va2x01234567, 0);
162 vpacc2x0123 = vdotq_lane_u32(vpacc2x0123, vb4567x0123, va2x01234567, 1);
163 vpacc2x4567 = vdotq_lane_u32(vpacc2x4567, vb4567x4567, va2x01234567, 1);
164 vpacc2x89AB = vdotq_lane_u32(vpacc2x89AB, vb4567x89AB, va2x01234567, 1);
165 vpacc2xCDEF = vdotq_lane_u32(vpacc2xCDEF, vb4567xCDEF, va2x01234567, 1);
166 vnacc3 = vdot_u32(vnacc3, va_zero_point, va3x01234567);
167 vpacc3x0123 = vdotq_lane_u32(vpacc3x0123, vb0123x0123, va3x01234567, 0);
168 vpacc3x4567 = vdotq_lane_u32(vpacc3x4567, vb0123x4567, va3x01234567, 0);
169 vpacc3x89AB = vdotq_lane_u32(vpacc3x89AB, vb0123x89AB, va3x01234567, 0);
170 vpacc3xCDEF = vdotq_lane_u32(vpacc3xCDEF, vb0123xCDEF, va3x01234567, 0);
171 vpacc3x0123 = vdotq_lane_u32(vpacc3x0123, vb4567x0123, va3x01234567, 1);
172 vpacc3x4567 = vdotq_lane_u32(vpacc3x4567, vb4567x4567, va3x01234567, 1);
173 vpacc3x89AB = vdotq_lane_u32(vpacc3x89AB, vb4567x89AB, va3x01234567, 1);
174 vpacc3xCDEF = vdotq_lane_u32(vpacc3xCDEF, vb4567xCDEF, va3x01234567, 1);
175 vnacc4 = vdot_u32(vnacc4, va_zero_point, va4x01234567);
176 vpacc4x0123 = vdotq_lane_u32(vpacc4x0123, vb0123x0123, va4x01234567, 0);
177 vpacc4x4567 = vdotq_lane_u32(vpacc4x4567, vb0123x4567, va4x01234567, 0);
178 vpacc4x89AB = vdotq_lane_u32(vpacc4x89AB, vb0123x89AB, va4x01234567, 0);
179 vpacc4xCDEF = vdotq_lane_u32(vpacc4xCDEF, vb0123xCDEF, va4x01234567, 0);
180 vpacc4x0123 = vdotq_lane_u32(vpacc4x0123, vb4567x0123, va4x01234567, 1);
181 vpacc4x4567 = vdotq_lane_u32(vpacc4x4567, vb4567x4567, va4x01234567, 1);
182 vpacc4x89AB = vdotq_lane_u32(vpacc4x89AB, vb4567x89AB, va4x01234567, 1);
183 vpacc4xCDEF = vdotq_lane_u32(vpacc4xCDEF, vb4567xCDEF, va4x01234567, 1);
184
185 k -= 8 * sizeof(uint8_t);
186 }
187 // Handle up to 4 final positions of `k`
188 if XNN_UNLIKELY(k != 0) {
189 // Load a 5x4 block of activations.
190 const uint8x8_t va0x01234567 = vreinterpret_u8_u32(vld1_lane_u32((const void*) a0, vmov_n_u32(0), 0)); a0 += 4;
191 const uint8x8_t va1x01234567 = vreinterpret_u8_u32(vld1_lane_u32((const void*) a1, vmov_n_u32(0), 0)); a1 += 4;
192 const uint8x8_t va2x01234567 = vreinterpret_u8_u32(vld1_lane_u32((const void*) a2, vmov_n_u32(0), 0)); a2 += 4;
193 const uint8x8_t va3x01234567 = vreinterpret_u8_u32(vld1_lane_u32((const void*) a3, vmov_n_u32(0), 0)); a3 += 4;
194 const uint8x8_t va4x01234567 = vreinterpret_u8_u32(vld1_lane_u32((const void*) a4, vmov_n_u32(0), 0)); a4 += 4;
195
196 // Load a 4x16 block of weights.
197 const uint8x16_t vb0123x0123 = vld1q_u8(w); w = (const void*) ((const uint8_t*) w + 16);
198 const uint8x16_t vb0123x4567 = vld1q_u8(w); w = (const void*) ((const uint8_t*) w + 16);
199 const uint8x16_t vb0123x89AB = vld1q_u8(w); w = (const void*) ((const uint8_t*) w + 16);
200 const uint8x16_t vb0123xCDEF = vld1q_u8(w); w = (const void*) ((const uint8_t*) w + 16);
201
202 // Multiply-accumulate: 5x4 * 4x16 --> 5x16.
203 vnacc0 = vdot_u32(vnacc0, va_zero_point, va0x01234567);
204 vpacc0x0123 = vdotq_lane_u32(vpacc0x0123, vb0123x0123, va0x01234567, 0);
205 vpacc0x4567 = vdotq_lane_u32(vpacc0x4567, vb0123x4567, va0x01234567, 0);
206 vpacc0x89AB = vdotq_lane_u32(vpacc0x89AB, vb0123x89AB, va0x01234567, 0);
207 vpacc0xCDEF = vdotq_lane_u32(vpacc0xCDEF, vb0123xCDEF, va0x01234567, 0);
208 vnacc1 = vdot_u32(vnacc1, va_zero_point, va1x01234567);
209 vpacc1x0123 = vdotq_lane_u32(vpacc1x0123, vb0123x0123, va1x01234567, 0);
210 vpacc1x4567 = vdotq_lane_u32(vpacc1x4567, vb0123x4567, va1x01234567, 0);
211 vpacc1x89AB = vdotq_lane_u32(vpacc1x89AB, vb0123x89AB, va1x01234567, 0);
212 vpacc1xCDEF = vdotq_lane_u32(vpacc1xCDEF, vb0123xCDEF, va1x01234567, 0);
213 vnacc2 = vdot_u32(vnacc2, va_zero_point, va2x01234567);
214 vpacc2x0123 = vdotq_lane_u32(vpacc2x0123, vb0123x0123, va2x01234567, 0);
215 vpacc2x4567 = vdotq_lane_u32(vpacc2x4567, vb0123x4567, va2x01234567, 0);
216 vpacc2x89AB = vdotq_lane_u32(vpacc2x89AB, vb0123x89AB, va2x01234567, 0);
217 vpacc2xCDEF = vdotq_lane_u32(vpacc2xCDEF, vb0123xCDEF, va2x01234567, 0);
218 vnacc3 = vdot_u32(vnacc3, va_zero_point, va3x01234567);
219 vpacc3x0123 = vdotq_lane_u32(vpacc3x0123, vb0123x0123, va3x01234567, 0);
220 vpacc3x4567 = vdotq_lane_u32(vpacc3x4567, vb0123x4567, va3x01234567, 0);
221 vpacc3x89AB = vdotq_lane_u32(vpacc3x89AB, vb0123x89AB, va3x01234567, 0);
222 vpacc3xCDEF = vdotq_lane_u32(vpacc3xCDEF, vb0123xCDEF, va3x01234567, 0);
223 vnacc4 = vdot_u32(vnacc4, va_zero_point, va4x01234567);
224 vpacc4x0123 = vdotq_lane_u32(vpacc4x0123, vb0123x0123, va4x01234567, 0);
225 vpacc4x4567 = vdotq_lane_u32(vpacc4x4567, vb0123x4567, va4x01234567, 0);
226 vpacc4x89AB = vdotq_lane_u32(vpacc4x89AB, vb0123x89AB, va4x01234567, 0);
227 vpacc4xCDEF = vdotq_lane_u32(vpacc4xCDEF, vb0123xCDEF, va4x01234567, 0);
228 }
229 p -= 5 * sizeof(void*);
230 } while (p != 0);
231
232 // Subtract zero point from accumulators.
233 vnacc0 = vpadd_u32(vnacc0, vnacc0);
234 const uint32x4_t vnacc0x0123 = vcombine_u32(vnacc0, vnacc0);
235 int32x4_t vacc0x0123 = vreinterpretq_s32_u32(vsubq_u32(vpacc0x0123, vnacc0x0123));
236 int32x4_t vacc0x4567 = vreinterpretq_s32_u32(vsubq_u32(vpacc0x4567, vnacc0x0123));
237 int32x4_t vacc0x89AB = vreinterpretq_s32_u32(vsubq_u32(vpacc0x89AB, vnacc0x0123));
238 int32x4_t vacc0xCDEF = vreinterpretq_s32_u32(vsubq_u32(vpacc0xCDEF, vnacc0x0123));
239 vnacc1 = vpadd_u32(vnacc1, vnacc1);
240 const uint32x4_t vnacc1x0123 = vcombine_u32(vnacc1, vnacc1);
241 int32x4_t vacc1x0123 = vreinterpretq_s32_u32(vsubq_u32(vpacc1x0123, vnacc1x0123));
242 int32x4_t vacc1x4567 = vreinterpretq_s32_u32(vsubq_u32(vpacc1x4567, vnacc1x0123));
243 int32x4_t vacc1x89AB = vreinterpretq_s32_u32(vsubq_u32(vpacc1x89AB, vnacc1x0123));
244 int32x4_t vacc1xCDEF = vreinterpretq_s32_u32(vsubq_u32(vpacc1xCDEF, vnacc1x0123));
245 vnacc2 = vpadd_u32(vnacc2, vnacc2);
246 const uint32x4_t vnacc2x0123 = vcombine_u32(vnacc2, vnacc2);
247 int32x4_t vacc2x0123 = vreinterpretq_s32_u32(vsubq_u32(vpacc2x0123, vnacc2x0123));
248 int32x4_t vacc2x4567 = vreinterpretq_s32_u32(vsubq_u32(vpacc2x4567, vnacc2x0123));
249 int32x4_t vacc2x89AB = vreinterpretq_s32_u32(vsubq_u32(vpacc2x89AB, vnacc2x0123));
250 int32x4_t vacc2xCDEF = vreinterpretq_s32_u32(vsubq_u32(vpacc2xCDEF, vnacc2x0123));
251 vnacc3 = vpadd_u32(vnacc3, vnacc3);
252 const uint32x4_t vnacc3x0123 = vcombine_u32(vnacc3, vnacc3);
253 int32x4_t vacc3x0123 = vreinterpretq_s32_u32(vsubq_u32(vpacc3x0123, vnacc3x0123));
254 int32x4_t vacc3x4567 = vreinterpretq_s32_u32(vsubq_u32(vpacc3x4567, vnacc3x0123));
255 int32x4_t vacc3x89AB = vreinterpretq_s32_u32(vsubq_u32(vpacc3x89AB, vnacc3x0123));
256 int32x4_t vacc3xCDEF = vreinterpretq_s32_u32(vsubq_u32(vpacc3xCDEF, vnacc3x0123));
257 vnacc4 = vpadd_u32(vnacc4, vnacc4);
258 const uint32x4_t vnacc4x0123 = vcombine_u32(vnacc4, vnacc4);
259 int32x4_t vacc4x0123 = vreinterpretq_s32_u32(vsubq_u32(vpacc4x0123, vnacc4x0123));
260 int32x4_t vacc4x4567 = vreinterpretq_s32_u32(vsubq_u32(vpacc4x4567, vnacc4x0123));
261 int32x4_t vacc4x89AB = vreinterpretq_s32_u32(vsubq_u32(vpacc4x89AB, vnacc4x0123));
262 int32x4_t vacc4xCDEF = vreinterpretq_s32_u32(vsubq_u32(vpacc4xCDEF, vnacc4x0123));
263
264 const int32x4_t vright_pre_shift = vld1q_dup_s32(¶ms->rndnu_neon.right_pre_shift);
265 const int32x4_t vmultiplier = vld1q_dup_s32(¶ms->rndnu_neon.multiplier);
266 const int32x4_t vright_post_shift = vld1q_dup_s32(¶ms->rndnu_neon.right_post_shift);
267
268 vacc0x0123 = vshlq_s32(vacc0x0123, vright_pre_shift);
269 vacc0x4567 = vshlq_s32(vacc0x4567, vright_pre_shift);
270 vacc0x89AB = vshlq_s32(vacc0x89AB, vright_pre_shift);
271 vacc0xCDEF = vshlq_s32(vacc0xCDEF, vright_pre_shift);
272 vacc1x0123 = vshlq_s32(vacc1x0123, vright_pre_shift);
273 vacc1x4567 = vshlq_s32(vacc1x4567, vright_pre_shift);
274 vacc1x89AB = vshlq_s32(vacc1x89AB, vright_pre_shift);
275 vacc1xCDEF = vshlq_s32(vacc1xCDEF, vright_pre_shift);
276 vacc2x0123 = vshlq_s32(vacc2x0123, vright_pre_shift);
277 vacc2x4567 = vshlq_s32(vacc2x4567, vright_pre_shift);
278 vacc2x89AB = vshlq_s32(vacc2x89AB, vright_pre_shift);
279 vacc2xCDEF = vshlq_s32(vacc2xCDEF, vright_pre_shift);
280 vacc3x0123 = vshlq_s32(vacc3x0123, vright_pre_shift);
281 vacc3x4567 = vshlq_s32(vacc3x4567, vright_pre_shift);
282 vacc3x89AB = vshlq_s32(vacc3x89AB, vright_pre_shift);
283 vacc3xCDEF = vshlq_s32(vacc3xCDEF, vright_pre_shift);
284 vacc4x0123 = vshlq_s32(vacc4x0123, vright_pre_shift);
285 vacc4x4567 = vshlq_s32(vacc4x4567, vright_pre_shift);
286 vacc4x89AB = vshlq_s32(vacc4x89AB, vright_pre_shift);
287 vacc4xCDEF = vshlq_s32(vacc4xCDEF, vright_pre_shift);
288
289 vacc0x0123 = vqdmulhq_s32(vacc0x0123, vmultiplier);
290 vacc0x4567 = vqdmulhq_s32(vacc0x4567, vmultiplier);
291 vacc0x89AB = vqdmulhq_s32(vacc0x89AB, vmultiplier);
292 vacc0xCDEF = vqdmulhq_s32(vacc0xCDEF, vmultiplier);
293 vacc1x0123 = vqdmulhq_s32(vacc1x0123, vmultiplier);
294 vacc1x4567 = vqdmulhq_s32(vacc1x4567, vmultiplier);
295 vacc1x89AB = vqdmulhq_s32(vacc1x89AB, vmultiplier);
296 vacc1xCDEF = vqdmulhq_s32(vacc1xCDEF, vmultiplier);
297 vacc2x0123 = vqdmulhq_s32(vacc2x0123, vmultiplier);
298 vacc2x4567 = vqdmulhq_s32(vacc2x4567, vmultiplier);
299 vacc2x89AB = vqdmulhq_s32(vacc2x89AB, vmultiplier);
300 vacc2xCDEF = vqdmulhq_s32(vacc2xCDEF, vmultiplier);
301 vacc3x0123 = vqdmulhq_s32(vacc3x0123, vmultiplier);
302 vacc3x4567 = vqdmulhq_s32(vacc3x4567, vmultiplier);
303 vacc3x89AB = vqdmulhq_s32(vacc3x89AB, vmultiplier);
304 vacc3xCDEF = vqdmulhq_s32(vacc3xCDEF, vmultiplier);
305 vacc4x0123 = vqdmulhq_s32(vacc4x0123, vmultiplier);
306 vacc4x4567 = vqdmulhq_s32(vacc4x4567, vmultiplier);
307 vacc4x89AB = vqdmulhq_s32(vacc4x89AB, vmultiplier);
308 vacc4xCDEF = vqdmulhq_s32(vacc4xCDEF, vmultiplier);
309
310 vacc0x0123 = vrshlq_s32(vacc0x0123, vright_post_shift);
311 vacc0x4567 = vrshlq_s32(vacc0x4567, vright_post_shift);
312 vacc0x89AB = vrshlq_s32(vacc0x89AB, vright_post_shift);
313 vacc0xCDEF = vrshlq_s32(vacc0xCDEF, vright_post_shift);
314 vacc1x0123 = vrshlq_s32(vacc1x0123, vright_post_shift);
315 vacc1x4567 = vrshlq_s32(vacc1x4567, vright_post_shift);
316 vacc1x89AB = vrshlq_s32(vacc1x89AB, vright_post_shift);
317 vacc1xCDEF = vrshlq_s32(vacc1xCDEF, vright_post_shift);
318 vacc2x0123 = vrshlq_s32(vacc2x0123, vright_post_shift);
319 vacc2x4567 = vrshlq_s32(vacc2x4567, vright_post_shift);
320 vacc2x89AB = vrshlq_s32(vacc2x89AB, vright_post_shift);
321 vacc2xCDEF = vrshlq_s32(vacc2xCDEF, vright_post_shift);
322 vacc3x0123 = vrshlq_s32(vacc3x0123, vright_post_shift);
323 vacc3x4567 = vrshlq_s32(vacc3x4567, vright_post_shift);
324 vacc3x89AB = vrshlq_s32(vacc3x89AB, vright_post_shift);
325 vacc3xCDEF = vrshlq_s32(vacc3xCDEF, vright_post_shift);
326 vacc4x0123 = vrshlq_s32(vacc4x0123, vright_post_shift);
327 vacc4x4567 = vrshlq_s32(vacc4x4567, vright_post_shift);
328 vacc4x89AB = vrshlq_s32(vacc4x89AB, vright_post_shift);
329 vacc4xCDEF = vrshlq_s32(vacc4xCDEF, vright_post_shift);
330
331 const int16x8_t voutput_zero_point = vld1q_dup_s16(¶ms->rndnu_neon.output_zero_point);
332 #if XNN_ARCH_ARM64
333 const int16x8_t vacc0x01234567 = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc0x0123), vacc0x4567), voutput_zero_point);
334 const int16x8_t vacc0x89ABCDEF = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc0x89AB), vacc0xCDEF), voutput_zero_point);
335 const int16x8_t vacc1x01234567 = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc1x0123), vacc1x4567), voutput_zero_point);
336 const int16x8_t vacc1x89ABCDEF = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc1x89AB), vacc1xCDEF), voutput_zero_point);
337 const int16x8_t vacc2x01234567 = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc2x0123), vacc2x4567), voutput_zero_point);
338 const int16x8_t vacc2x89ABCDEF = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc2x89AB), vacc2xCDEF), voutput_zero_point);
339 const int16x8_t vacc3x01234567 = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc3x0123), vacc3x4567), voutput_zero_point);
340 const int16x8_t vacc3x89ABCDEF = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc3x89AB), vacc3xCDEF), voutput_zero_point);
341 const int16x8_t vacc4x01234567 = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc4x0123), vacc4x4567), voutput_zero_point);
342 const int16x8_t vacc4x89ABCDEF = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc4x89AB), vacc4xCDEF), voutput_zero_point);
343
344 uint8x16_t vout0x0123456789ABCDEF = vqmovun_high_s16(vqmovun_s16(vacc0x01234567), vacc0x89ABCDEF);
345 uint8x16_t vout1x0123456789ABCDEF = vqmovun_high_s16(vqmovun_s16(vacc1x01234567), vacc1x89ABCDEF);
346 uint8x16_t vout2x0123456789ABCDEF = vqmovun_high_s16(vqmovun_s16(vacc2x01234567), vacc2x89ABCDEF);
347 uint8x16_t vout3x0123456789ABCDEF = vqmovun_high_s16(vqmovun_s16(vacc3x01234567), vacc3x89ABCDEF);
348 uint8x16_t vout4x0123456789ABCDEF = vqmovun_high_s16(vqmovun_s16(vacc4x01234567), vacc4x89ABCDEF);
349 #else
350 const int16x8_t vacc0x01234567 = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc0x0123), vqmovn_s32(vacc0x4567)), voutput_zero_point);
351 const int16x8_t vacc0x89ABCDEF = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc0x89AB), vqmovn_s32(vacc0xCDEF)), voutput_zero_point);
352 const int16x8_t vacc1x01234567 = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc1x0123), vqmovn_s32(vacc1x4567)), voutput_zero_point);
353 const int16x8_t vacc1x89ABCDEF = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc1x89AB), vqmovn_s32(vacc1xCDEF)), voutput_zero_point);
354 const int16x8_t vacc2x01234567 = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc2x0123), vqmovn_s32(vacc2x4567)), voutput_zero_point);
355 const int16x8_t vacc2x89ABCDEF = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc2x89AB), vqmovn_s32(vacc2xCDEF)), voutput_zero_point);
356 const int16x8_t vacc3x01234567 = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc3x0123), vqmovn_s32(vacc3x4567)), voutput_zero_point);
357 const int16x8_t vacc3x89ABCDEF = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc3x89AB), vqmovn_s32(vacc3xCDEF)), voutput_zero_point);
358 const int16x8_t vacc4x01234567 = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc4x0123), vqmovn_s32(vacc4x4567)), voutput_zero_point);
359 const int16x8_t vacc4x89ABCDEF = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc4x89AB), vqmovn_s32(vacc4xCDEF)), voutput_zero_point);
360
361 uint8x16_t vout0x0123456789ABCDEF = vcombine_u8(vqmovun_s16(vacc0x01234567), vqmovun_s16(vacc0x89ABCDEF));
362 uint8x16_t vout1x0123456789ABCDEF = vcombine_u8(vqmovun_s16(vacc1x01234567), vqmovun_s16(vacc1x89ABCDEF));
363 uint8x16_t vout2x0123456789ABCDEF = vcombine_u8(vqmovun_s16(vacc2x01234567), vqmovun_s16(vacc2x89ABCDEF));
364 uint8x16_t vout3x0123456789ABCDEF = vcombine_u8(vqmovun_s16(vacc3x01234567), vqmovun_s16(vacc3x89ABCDEF));
365 uint8x16_t vout4x0123456789ABCDEF = vcombine_u8(vqmovun_s16(vacc4x01234567), vqmovun_s16(vacc4x89ABCDEF));
366 #endif
367 const uint8x16_t voutput_min = vld1q_dup_u8(¶ms->rndnu_neon.output_min);
368 const uint8x16_t voutput_max = vld1q_dup_u8(¶ms->rndnu_neon.output_max);
369
370 vout0x0123456789ABCDEF = vmaxq_u8(vout0x0123456789ABCDEF, voutput_min);
371 vout1x0123456789ABCDEF = vmaxq_u8(vout1x0123456789ABCDEF, voutput_min);
372 vout2x0123456789ABCDEF = vmaxq_u8(vout2x0123456789ABCDEF, voutput_min);
373 vout3x0123456789ABCDEF = vmaxq_u8(vout3x0123456789ABCDEF, voutput_min);
374 vout4x0123456789ABCDEF = vmaxq_u8(vout4x0123456789ABCDEF, voutput_min);
375
376 vout0x0123456789ABCDEF = vminq_u8(vout0x0123456789ABCDEF, voutput_max);
377 vout1x0123456789ABCDEF = vminq_u8(vout1x0123456789ABCDEF, voutput_max);
378 vout2x0123456789ABCDEF = vminq_u8(vout2x0123456789ABCDEF, voutput_max);
379 vout3x0123456789ABCDEF = vminq_u8(vout3x0123456789ABCDEF, voutput_max);
380 vout4x0123456789ABCDEF = vminq_u8(vout4x0123456789ABCDEF, voutput_max);
381
382 if (nc >= 16) {
383 vst1q_u8(c4 + 0, vout4x0123456789ABCDEF);
384 vst1q_u8(c3 + 0, vout3x0123456789ABCDEF);
385 vst1q_u8(c2 + 0, vout2x0123456789ABCDEF);
386 vst1q_u8(c1 + 0, vout1x0123456789ABCDEF);
387 vst1q_u8(c0 + 0, vout0x0123456789ABCDEF);
388
389 c4 = (uint8_t*) ((uintptr_t) c4 + cn_stride);
390 c3 = (uint8_t*) ((uintptr_t) c3 + cn_stride);
391 c2 = (uint8_t*) ((uintptr_t) c2 + cn_stride);
392 c1 = (uint8_t*) ((uintptr_t) c1 + cn_stride);
393 c0 = (uint8_t*) ((uintptr_t) c0 + cn_stride);
394
395 a = (const uint8_t**restrict) ((uintptr_t) a - ks);
396
397 nc -= 16;
398 } else {
399 uint8x8_t vout4x01234567 = vget_low_u8(vout4x0123456789ABCDEF);
400 uint8x16_t vout2x01234567_3x01234567 = vcombine_u8(vget_low_u8(vout2x0123456789ABCDEF), vget_low_u8(vout3x0123456789ABCDEF));
401 uint8x16_t vout0x01234567_1x01234567 = vcombine_u8(vget_low_u8(vout0x0123456789ABCDEF), vget_low_u8(vout1x0123456789ABCDEF));
402 if (nc & 8) {
403 vst1_u8(c4, vout4x01234567); c4 += 8; // This line
404 vst1_u8(c3, vget_high_u8(vout2x01234567_3x01234567)); c3 += 8;
405 vst1_u8(c2, vget_low_u8(vout2x01234567_3x01234567)); c2 += 8;
406 vst1_u8(c1, vget_high_u8(vout0x01234567_1x01234567)); c1 += 8;
407 vst1_u8(c0, vget_low_u8(vout0x01234567_1x01234567)); c0 += 8;
408 vout4x01234567 = vget_high_u8(vout4x0123456789ABCDEF);
409 vout2x01234567_3x01234567 = vcombine_u8(vget_high_u8(vout2x0123456789ABCDEF), vget_high_u8(vout3x0123456789ABCDEF));
410 vout0x01234567_1x01234567 = vcombine_u8(vget_high_u8(vout0x0123456789ABCDEF), vget_high_u8(vout1x0123456789ABCDEF));
411 }
412 if (nc & 4) {
413 vst1_lane_u32((void*) c4, vreinterpret_u32_u8(vout4x01234567), 0); c4 += 4;
414 vst1q_lane_u32((void*) c3, vreinterpretq_u32_u8(vout2x01234567_3x01234567), 2); c3 += 4;
415 vst1q_lane_u32((void*) c2, vreinterpretq_u32_u8(vout2x01234567_3x01234567), 0); c2 += 4;
416 vst1q_lane_u32((void*) c1, vreinterpretq_u32_u8(vout0x01234567_1x01234567), 2); c1 += 4;
417 vst1q_lane_u32((void*) c0, vreinterpretq_u32_u8(vout0x01234567_1x01234567), 0); c0 += 4;
418 vout4x01234567 = vext_u8(vout4x01234567, vout4x01234567, 4);
419 vout2x01234567_3x01234567 = vextq_u8(vout2x01234567_3x01234567, vout2x01234567_3x01234567, 4);
420 vout0x01234567_1x01234567 = vextq_u8(vout0x01234567_1x01234567, vout0x01234567_1x01234567, 4);
421 }
422 if (nc & 2) {
423 vst1_lane_u16((void*) c4, vreinterpret_u16_u8(vout4x01234567), 0); c4 += 2;
424 vst1q_lane_u16((void*) c3, vreinterpretq_u16_u8(vout2x01234567_3x01234567), 4); c3 += 2;
425 vst1q_lane_u16((void*) c2, vreinterpretq_u16_u8(vout2x01234567_3x01234567), 0); c2 += 2;
426 vst1q_lane_u16((void*) c1, vreinterpretq_u16_u8(vout0x01234567_1x01234567), 4); c1 += 2;
427 vst1q_lane_u16((void*) c0, vreinterpretq_u16_u8(vout0x01234567_1x01234567), 0); c0 += 2;
428 vout4x01234567 = vext_u8(vout4x01234567, vout4x01234567, 2);
429 vout2x01234567_3x01234567 = vextq_u8(vout2x01234567_3x01234567, vout2x01234567_3x01234567, 2);
430 vout0x01234567_1x01234567 = vextq_u8(vout0x01234567_1x01234567, vout0x01234567_1x01234567, 2);
431 }
432 if (nc & 1) {
433 vst1_lane_u8(c4, vout4x01234567, 0);
434 vst1q_lane_u8(c3, vout2x01234567_3x01234567, 8);
435 vst1q_lane_u8(c2, vout2x01234567_3x01234567, 0);
436 vst1q_lane_u8(c1, vout0x01234567_1x01234567, 8);
437 vst1q_lane_u8(c0, vout0x01234567_1x01234567, 0);
438 }
439
440 nc = 0;
441 }
442 } while (nc != 0);
443 }
444