1 // Auto-generated file. Do not edit!
2 // Template: src/qu8-gemm/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/gemm.h>
15 #include <xnnpack/math.h>
16
17
xnn_qu8_gemm_minmax_rndnu_ukernel_4x16c4__neondot(size_t mr,size_t nc,size_t kc,const uint8_t * restrict a,size_t a_stride,const void * restrict w,uint8_t * restrict c,size_t cm_stride,size_t cn_stride,const union xnn_qu8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS (1)])18 void xnn_qu8_gemm_minmax_rndnu_ukernel_4x16c4__neondot(
19 size_t mr,
20 size_t nc,
21 size_t kc,
22 const uint8_t* restrict a,
23 size_t a_stride,
24 const void* restrict w,
25 uint8_t* restrict c,
26 size_t cm_stride,
27 size_t cn_stride,
28 const union xnn_qu8_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(uint8_t) == 0);
35 assert(a != NULL);
36 assert(w != NULL);
37 assert(c != NULL);
38
39 kc = round_up_po2(kc, 4 * sizeof(uint8_t));
40 const uint8_t* a0 = a;
41 uint8_t* c0 = c;
42 const uint8_t* a1 = (const uint8_t*) ((uintptr_t) a0 + a_stride);
43 uint8_t* c1 = (uint8_t*) ((uintptr_t) c0 + cm_stride);
44 if XNN_UNPREDICTABLE(mr < 2) {
45 a1 = a0;
46 c1 = c0;
47 }
48 const uint8_t* a2 = (const uint8_t*) ((uintptr_t) a1 + a_stride);
49 uint8_t* c2 = (uint8_t*) ((uintptr_t) c1 + cm_stride);
50 if XNN_UNPREDICTABLE(mr <= 2) {
51 a2 = a1;
52 c2 = c1;
53 }
54 const uint8_t* a3 = (const uint8_t*) ((uintptr_t) a2 + a_stride);
55 uint8_t* c3 = (uint8_t*) ((uintptr_t) c2 + cm_stride);
56 if XNN_UNPREDICTABLE(mr != 4) {
57 a3 = a2;
58 c3 = c2;
59 }
60
61 const uint8x8_t va_zero_point = vld1_dup_u8(¶ms->rndnu_neon.kernel_zero_point[0]);
62
63 // Loop over groups of 16 columns.
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 uint32x2_t vnacc0 = vmov_n_u32(0);
84 uint32x2_t vnacc1 = vmov_n_u32(0);
85 uint32x2_t vnacc2 = vmov_n_u32(0);
86 uint32x2_t vnacc3 = vmov_n_u32(0);
87
88 // Inner accumulation loop along the 16 columns.
89 size_t k = kc;
90 // 2x partial unrolled loop to load 8 bytes at a time.
91 while (k >= 8 * sizeof(uint8_t)) {
92 // Load a 4x8 block of activations.
93 const uint8x8_t va0x01234567 = vld1_u8(a0); a0 += 8;
94 const uint8x8_t va1x01234567 = vld1_u8(a1); a1 += 8;
95 const uint8x8_t va2x01234567 = vld1_u8(a2); a2 += 8;
96 const uint8x8_t va3x01234567 = vld1_u8(a3); a3 += 8;
97
98 // Load a 8x16 block of weights.
99 const uint8x16_t vb0123x0123 = vld1q_u8(w); w = (const void*) ((const uint8_t*) w + 16);
100 const uint8x16_t vb0123x4567 = vld1q_u8(w); w = (const void*) ((const uint8_t*) w + 16);
101 const uint8x16_t vb0123x89AB = vld1q_u8(w); w = (const void*) ((const uint8_t*) w + 16);
102 const uint8x16_t vb0123xCDEF = vld1q_u8(w); w = (const void*) ((const uint8_t*) w + 16);
103 const uint8x16_t vb4567x0123 = vld1q_u8(w); w = (const void*) ((const uint8_t*) w + 16);
104 const uint8x16_t vb4567x4567 = vld1q_u8(w); w = (const void*) ((const uint8_t*) w + 16);
105 const uint8x16_t vb4567x89AB = vld1q_u8(w); w = (const void*) ((const uint8_t*) w + 16);
106 const uint8x16_t vb4567xCDEF = vld1q_u8(w); w = (const void*) ((const uint8_t*) w + 16);
107
108 // Multiply-accumulate: 4x8 * 8x16 --> 4x16.
109 vnacc0 = vdot_u32(vnacc0, va_zero_point, va0x01234567);
110 vpacc0x0123 = vdotq_lane_u32(vpacc0x0123, vb0123x0123, va0x01234567, 0);
111 vpacc0x4567 = vdotq_lane_u32(vpacc0x4567, vb0123x4567, va0x01234567, 0);
112 vpacc0x89AB = vdotq_lane_u32(vpacc0x89AB, vb0123x89AB, va0x01234567, 0);
113 vpacc0xCDEF = vdotq_lane_u32(vpacc0xCDEF, vb0123xCDEF, va0x01234567, 0);
114 vpacc0x0123 = vdotq_lane_u32(vpacc0x0123, vb4567x0123, va0x01234567, 1);
115 vpacc0x4567 = vdotq_lane_u32(vpacc0x4567, vb4567x4567, va0x01234567, 1);
116 vpacc0x89AB = vdotq_lane_u32(vpacc0x89AB, vb4567x89AB, va0x01234567, 1);
117 vpacc0xCDEF = vdotq_lane_u32(vpacc0xCDEF, vb4567xCDEF, va0x01234567, 1);
118 vnacc1 = vdot_u32(vnacc1, va_zero_point, va1x01234567);
119 vpacc1x0123 = vdotq_lane_u32(vpacc1x0123, vb0123x0123, va1x01234567, 0);
120 vpacc1x4567 = vdotq_lane_u32(vpacc1x4567, vb0123x4567, va1x01234567, 0);
121 vpacc1x89AB = vdotq_lane_u32(vpacc1x89AB, vb0123x89AB, va1x01234567, 0);
122 vpacc1xCDEF = vdotq_lane_u32(vpacc1xCDEF, vb0123xCDEF, va1x01234567, 0);
123 vpacc1x0123 = vdotq_lane_u32(vpacc1x0123, vb4567x0123, va1x01234567, 1);
124 vpacc1x4567 = vdotq_lane_u32(vpacc1x4567, vb4567x4567, va1x01234567, 1);
125 vpacc1x89AB = vdotq_lane_u32(vpacc1x89AB, vb4567x89AB, va1x01234567, 1);
126 vpacc1xCDEF = vdotq_lane_u32(vpacc1xCDEF, vb4567xCDEF, va1x01234567, 1);
127 vnacc2 = vdot_u32(vnacc2, va_zero_point, va2x01234567);
128 vpacc2x0123 = vdotq_lane_u32(vpacc2x0123, vb0123x0123, va2x01234567, 0);
129 vpacc2x4567 = vdotq_lane_u32(vpacc2x4567, vb0123x4567, va2x01234567, 0);
130 vpacc2x89AB = vdotq_lane_u32(vpacc2x89AB, vb0123x89AB, va2x01234567, 0);
131 vpacc2xCDEF = vdotq_lane_u32(vpacc2xCDEF, vb0123xCDEF, va2x01234567, 0);
132 vpacc2x0123 = vdotq_lane_u32(vpacc2x0123, vb4567x0123, va2x01234567, 1);
133 vpacc2x4567 = vdotq_lane_u32(vpacc2x4567, vb4567x4567, va2x01234567, 1);
134 vpacc2x89AB = vdotq_lane_u32(vpacc2x89AB, vb4567x89AB, va2x01234567, 1);
135 vpacc2xCDEF = vdotq_lane_u32(vpacc2xCDEF, vb4567xCDEF, va2x01234567, 1);
136 vnacc3 = vdot_u32(vnacc3, va_zero_point, va3x01234567);
137 vpacc3x0123 = vdotq_lane_u32(vpacc3x0123, vb0123x0123, va3x01234567, 0);
138 vpacc3x4567 = vdotq_lane_u32(vpacc3x4567, vb0123x4567, va3x01234567, 0);
139 vpacc3x89AB = vdotq_lane_u32(vpacc3x89AB, vb0123x89AB, va3x01234567, 0);
140 vpacc3xCDEF = vdotq_lane_u32(vpacc3xCDEF, vb0123xCDEF, va3x01234567, 0);
141 vpacc3x0123 = vdotq_lane_u32(vpacc3x0123, vb4567x0123, va3x01234567, 1);
142 vpacc3x4567 = vdotq_lane_u32(vpacc3x4567, vb4567x4567, va3x01234567, 1);
143 vpacc3x89AB = vdotq_lane_u32(vpacc3x89AB, vb4567x89AB, va3x01234567, 1);
144 vpacc3xCDEF = vdotq_lane_u32(vpacc3xCDEF, vb4567xCDEF, va3x01234567, 1);
145
146 k -= 8 * sizeof(uint8_t);
147 }
148 // Handle up to 4 final positions of `k`
149 if XNN_UNLIKELY(k != 0) {
150 // Load a 4x4 block of activations.
151 const uint8x8_t va0x01234567 = vreinterpret_u8_u32(vld1_lane_u32((const void*) a0, vmov_n_u32(0), 0)); a0 += 4;
152 const uint8x8_t va1x01234567 = vreinterpret_u8_u32(vld1_lane_u32((const void*) a1, vmov_n_u32(0), 0)); a1 += 4;
153 const uint8x8_t va2x01234567 = vreinterpret_u8_u32(vld1_lane_u32((const void*) a2, vmov_n_u32(0), 0)); a2 += 4;
154 const uint8x8_t va3x01234567 = vreinterpret_u8_u32(vld1_lane_u32((const void*) a3, vmov_n_u32(0), 0)); a3 += 4;
155
156 // Load a 4x16 block of weights.
157 const uint8x16_t vb0123x0123 = vld1q_u8(w); w = (const void*) ((const uint8_t*) w + 16);
158 const uint8x16_t vb0123x4567 = vld1q_u8(w); w = (const void*) ((const uint8_t*) w + 16);
159 const uint8x16_t vb0123x89AB = vld1q_u8(w); w = (const void*) ((const uint8_t*) w + 16);
160 const uint8x16_t vb0123xCDEF = vld1q_u8(w); w = (const void*) ((const uint8_t*) w + 16);
161
162 // Multiply-accumulate: 4x4 * 4x16 --> 4x16.
163 vnacc0 = vdot_u32(vnacc0, va_zero_point, va0x01234567);
164 vpacc0x0123 = vdotq_lane_u32(vpacc0x0123, vb0123x0123, va0x01234567, 0);
165 vpacc0x4567 = vdotq_lane_u32(vpacc0x4567, vb0123x4567, va0x01234567, 0);
166 vpacc0x89AB = vdotq_lane_u32(vpacc0x89AB, vb0123x89AB, va0x01234567, 0);
167 vpacc0xCDEF = vdotq_lane_u32(vpacc0xCDEF, vb0123xCDEF, va0x01234567, 0);
168 vnacc1 = vdot_u32(vnacc1, va_zero_point, va1x01234567);
169 vpacc1x0123 = vdotq_lane_u32(vpacc1x0123, vb0123x0123, va1x01234567, 0);
170 vpacc1x4567 = vdotq_lane_u32(vpacc1x4567, vb0123x4567, va1x01234567, 0);
171 vpacc1x89AB = vdotq_lane_u32(vpacc1x89AB, vb0123x89AB, va1x01234567, 0);
172 vpacc1xCDEF = vdotq_lane_u32(vpacc1xCDEF, vb0123xCDEF, va1x01234567, 0);
173 vnacc2 = vdot_u32(vnacc2, va_zero_point, va2x01234567);
174 vpacc2x0123 = vdotq_lane_u32(vpacc2x0123, vb0123x0123, va2x01234567, 0);
175 vpacc2x4567 = vdotq_lane_u32(vpacc2x4567, vb0123x4567, va2x01234567, 0);
176 vpacc2x89AB = vdotq_lane_u32(vpacc2x89AB, vb0123x89AB, va2x01234567, 0);
177 vpacc2xCDEF = vdotq_lane_u32(vpacc2xCDEF, vb0123xCDEF, va2x01234567, 0);
178 vnacc3 = vdot_u32(vnacc3, va_zero_point, va3x01234567);
179 vpacc3x0123 = vdotq_lane_u32(vpacc3x0123, vb0123x0123, va3x01234567, 0);
180 vpacc3x4567 = vdotq_lane_u32(vpacc3x4567, vb0123x4567, va3x01234567, 0);
181 vpacc3x89AB = vdotq_lane_u32(vpacc3x89AB, vb0123x89AB, va3x01234567, 0);
182 vpacc3xCDEF = vdotq_lane_u32(vpacc3xCDEF, vb0123xCDEF, va3x01234567, 0);
183 }
184
185 // Subtract zero point from accumulators.
186 vnacc0 = vpadd_u32(vnacc0, vnacc0);
187 const uint32x4_t vnacc0x0123 = vcombine_u32(vnacc0, vnacc0);
188 int32x4_t vacc0x0123 = vreinterpretq_s32_u32(vsubq_u32(vpacc0x0123, vnacc0x0123));
189 int32x4_t vacc0x4567 = vreinterpretq_s32_u32(vsubq_u32(vpacc0x4567, vnacc0x0123));
190 int32x4_t vacc0x89AB = vreinterpretq_s32_u32(vsubq_u32(vpacc0x89AB, vnacc0x0123));
191 int32x4_t vacc0xCDEF = vreinterpretq_s32_u32(vsubq_u32(vpacc0xCDEF, vnacc0x0123));
192 vnacc1 = vpadd_u32(vnacc1, vnacc1);
193 const uint32x4_t vnacc1x0123 = vcombine_u32(vnacc1, vnacc1);
194 int32x4_t vacc1x0123 = vreinterpretq_s32_u32(vsubq_u32(vpacc1x0123, vnacc1x0123));
195 int32x4_t vacc1x4567 = vreinterpretq_s32_u32(vsubq_u32(vpacc1x4567, vnacc1x0123));
196 int32x4_t vacc1x89AB = vreinterpretq_s32_u32(vsubq_u32(vpacc1x89AB, vnacc1x0123));
197 int32x4_t vacc1xCDEF = vreinterpretq_s32_u32(vsubq_u32(vpacc1xCDEF, vnacc1x0123));
198 vnacc2 = vpadd_u32(vnacc2, vnacc2);
199 const uint32x4_t vnacc2x0123 = vcombine_u32(vnacc2, vnacc2);
200 int32x4_t vacc2x0123 = vreinterpretq_s32_u32(vsubq_u32(vpacc2x0123, vnacc2x0123));
201 int32x4_t vacc2x4567 = vreinterpretq_s32_u32(vsubq_u32(vpacc2x4567, vnacc2x0123));
202 int32x4_t vacc2x89AB = vreinterpretq_s32_u32(vsubq_u32(vpacc2x89AB, vnacc2x0123));
203 int32x4_t vacc2xCDEF = vreinterpretq_s32_u32(vsubq_u32(vpacc2xCDEF, vnacc2x0123));
204 vnacc3 = vpadd_u32(vnacc3, vnacc3);
205 const uint32x4_t vnacc3x0123 = vcombine_u32(vnacc3, vnacc3);
206 int32x4_t vacc3x0123 = vreinterpretq_s32_u32(vsubq_u32(vpacc3x0123, vnacc3x0123));
207 int32x4_t vacc3x4567 = vreinterpretq_s32_u32(vsubq_u32(vpacc3x4567, vnacc3x0123));
208 int32x4_t vacc3x89AB = vreinterpretq_s32_u32(vsubq_u32(vpacc3x89AB, vnacc3x0123));
209 int32x4_t vacc3xCDEF = vreinterpretq_s32_u32(vsubq_u32(vpacc3xCDEF, vnacc3x0123));
210
211 const int32x4_t vright_pre_shift = vld1q_dup_s32(¶ms->rndnu_neon.right_pre_shift);
212 const int32x4_t vmultiplier = vld1q_dup_s32(¶ms->rndnu_neon.multiplier);
213 const int32x4_t vright_post_shift = vld1q_dup_s32(¶ms->rndnu_neon.right_post_shift);
214
215 vacc0x0123 = vshlq_s32(vacc0x0123, vright_pre_shift);
216 vacc0x4567 = vshlq_s32(vacc0x4567, vright_pre_shift);
217 vacc0x89AB = vshlq_s32(vacc0x89AB, vright_pre_shift);
218 vacc0xCDEF = vshlq_s32(vacc0xCDEF, vright_pre_shift);
219 vacc1x0123 = vshlq_s32(vacc1x0123, vright_pre_shift);
220 vacc1x4567 = vshlq_s32(vacc1x4567, vright_pre_shift);
221 vacc1x89AB = vshlq_s32(vacc1x89AB, vright_pre_shift);
222 vacc1xCDEF = vshlq_s32(vacc1xCDEF, vright_pre_shift);
223 vacc2x0123 = vshlq_s32(vacc2x0123, vright_pre_shift);
224 vacc2x4567 = vshlq_s32(vacc2x4567, vright_pre_shift);
225 vacc2x89AB = vshlq_s32(vacc2x89AB, vright_pre_shift);
226 vacc2xCDEF = vshlq_s32(vacc2xCDEF, vright_pre_shift);
227 vacc3x0123 = vshlq_s32(vacc3x0123, vright_pre_shift);
228 vacc3x4567 = vshlq_s32(vacc3x4567, vright_pre_shift);
229 vacc3x89AB = vshlq_s32(vacc3x89AB, vright_pre_shift);
230 vacc3xCDEF = vshlq_s32(vacc3xCDEF, vright_pre_shift);
231
232 vacc0x0123 = vqdmulhq_s32(vacc0x0123, vmultiplier);
233 vacc0x4567 = vqdmulhq_s32(vacc0x4567, vmultiplier);
234 vacc0x89AB = vqdmulhq_s32(vacc0x89AB, vmultiplier);
235 vacc0xCDEF = vqdmulhq_s32(vacc0xCDEF, vmultiplier);
236 vacc1x0123 = vqdmulhq_s32(vacc1x0123, vmultiplier);
237 vacc1x4567 = vqdmulhq_s32(vacc1x4567, vmultiplier);
238 vacc1x89AB = vqdmulhq_s32(vacc1x89AB, vmultiplier);
239 vacc1xCDEF = vqdmulhq_s32(vacc1xCDEF, vmultiplier);
240 vacc2x0123 = vqdmulhq_s32(vacc2x0123, vmultiplier);
241 vacc2x4567 = vqdmulhq_s32(vacc2x4567, vmultiplier);
242 vacc2x89AB = vqdmulhq_s32(vacc2x89AB, vmultiplier);
243 vacc2xCDEF = vqdmulhq_s32(vacc2xCDEF, vmultiplier);
244 vacc3x0123 = vqdmulhq_s32(vacc3x0123, vmultiplier);
245 vacc3x4567 = vqdmulhq_s32(vacc3x4567, vmultiplier);
246 vacc3x89AB = vqdmulhq_s32(vacc3x89AB, vmultiplier);
247 vacc3xCDEF = vqdmulhq_s32(vacc3xCDEF, vmultiplier);
248
249 vacc0x0123 = vrshlq_s32(vacc0x0123, vright_post_shift);
250 vacc0x4567 = vrshlq_s32(vacc0x4567, vright_post_shift);
251 vacc0x89AB = vrshlq_s32(vacc0x89AB, vright_post_shift);
252 vacc0xCDEF = vrshlq_s32(vacc0xCDEF, vright_post_shift);
253 vacc1x0123 = vrshlq_s32(vacc1x0123, vright_post_shift);
254 vacc1x4567 = vrshlq_s32(vacc1x4567, vright_post_shift);
255 vacc1x89AB = vrshlq_s32(vacc1x89AB, vright_post_shift);
256 vacc1xCDEF = vrshlq_s32(vacc1xCDEF, vright_post_shift);
257 vacc2x0123 = vrshlq_s32(vacc2x0123, vright_post_shift);
258 vacc2x4567 = vrshlq_s32(vacc2x4567, vright_post_shift);
259 vacc2x89AB = vrshlq_s32(vacc2x89AB, vright_post_shift);
260 vacc2xCDEF = vrshlq_s32(vacc2xCDEF, vright_post_shift);
261 vacc3x0123 = vrshlq_s32(vacc3x0123, vright_post_shift);
262 vacc3x4567 = vrshlq_s32(vacc3x4567, vright_post_shift);
263 vacc3x89AB = vrshlq_s32(vacc3x89AB, vright_post_shift);
264 vacc3xCDEF = vrshlq_s32(vacc3xCDEF, vright_post_shift);
265
266 const int16x8_t voutput_zero_point = vld1q_dup_s16(¶ms->rndnu_neon.output_zero_point);
267 #if XNN_ARCH_ARM64
268 const int16x8_t vacc0x01234567 = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc0x0123), vacc0x4567), voutput_zero_point);
269 const int16x8_t vacc0x89ABCDEF = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc0x89AB), vacc0xCDEF), voutput_zero_point);
270 const int16x8_t vacc1x01234567 = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc1x0123), vacc1x4567), voutput_zero_point);
271 const int16x8_t vacc1x89ABCDEF = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc1x89AB), vacc1xCDEF), voutput_zero_point);
272 const int16x8_t vacc2x01234567 = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc2x0123), vacc2x4567), voutput_zero_point);
273 const int16x8_t vacc2x89ABCDEF = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc2x89AB), vacc2xCDEF), voutput_zero_point);
274 const int16x8_t vacc3x01234567 = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc3x0123), vacc3x4567), voutput_zero_point);
275 const int16x8_t vacc3x89ABCDEF = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc3x89AB), vacc3xCDEF), voutput_zero_point);
276
277 uint8x16_t vout0x0123456789ABCDEF = vqmovun_high_s16(vqmovun_s16(vacc0x01234567), vacc0x89ABCDEF);
278 uint8x16_t vout1x0123456789ABCDEF = vqmovun_high_s16(vqmovun_s16(vacc1x01234567), vacc1x89ABCDEF);
279 uint8x16_t vout2x0123456789ABCDEF = vqmovun_high_s16(vqmovun_s16(vacc2x01234567), vacc2x89ABCDEF);
280 uint8x16_t vout3x0123456789ABCDEF = vqmovun_high_s16(vqmovun_s16(vacc3x01234567), vacc3x89ABCDEF);
281 #else
282 const int16x8_t vacc0x01234567 = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc0x0123), vqmovn_s32(vacc0x4567)), voutput_zero_point);
283 const int16x8_t vacc0x89ABCDEF = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc0x89AB), vqmovn_s32(vacc0xCDEF)), voutput_zero_point);
284 const int16x8_t vacc1x01234567 = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc1x0123), vqmovn_s32(vacc1x4567)), voutput_zero_point);
285 const int16x8_t vacc1x89ABCDEF = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc1x89AB), vqmovn_s32(vacc1xCDEF)), voutput_zero_point);
286 const int16x8_t vacc2x01234567 = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc2x0123), vqmovn_s32(vacc2x4567)), voutput_zero_point);
287 const int16x8_t vacc2x89ABCDEF = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc2x89AB), vqmovn_s32(vacc2xCDEF)), voutput_zero_point);
288 const int16x8_t vacc3x01234567 = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc3x0123), vqmovn_s32(vacc3x4567)), voutput_zero_point);
289 const int16x8_t vacc3x89ABCDEF = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc3x89AB), vqmovn_s32(vacc3xCDEF)), voutput_zero_point);
290
291 uint8x16_t vout0x0123456789ABCDEF = vcombine_u8(vqmovun_s16(vacc0x01234567), vqmovun_s16(vacc0x89ABCDEF));
292 uint8x16_t vout1x0123456789ABCDEF = vcombine_u8(vqmovun_s16(vacc1x01234567), vqmovun_s16(vacc1x89ABCDEF));
293 uint8x16_t vout2x0123456789ABCDEF = vcombine_u8(vqmovun_s16(vacc2x01234567), vqmovun_s16(vacc2x89ABCDEF));
294 uint8x16_t vout3x0123456789ABCDEF = vcombine_u8(vqmovun_s16(vacc3x01234567), vqmovun_s16(vacc3x89ABCDEF));
295 #endif
296 const uint8x16_t voutput_min = vld1q_dup_u8(¶ms->rndnu_neon.output_min);
297 const uint8x16_t voutput_max = vld1q_dup_u8(¶ms->rndnu_neon.output_max);
298
299 vout0x0123456789ABCDEF = vmaxq_u8(vout0x0123456789ABCDEF, voutput_min);
300 vout1x0123456789ABCDEF = vmaxq_u8(vout1x0123456789ABCDEF, voutput_min);
301 vout2x0123456789ABCDEF = vmaxq_u8(vout2x0123456789ABCDEF, voutput_min);
302 vout3x0123456789ABCDEF = vmaxq_u8(vout3x0123456789ABCDEF, voutput_min);
303
304 vout0x0123456789ABCDEF = vminq_u8(vout0x0123456789ABCDEF, voutput_max);
305 vout1x0123456789ABCDEF = vminq_u8(vout1x0123456789ABCDEF, voutput_max);
306 vout2x0123456789ABCDEF = vminq_u8(vout2x0123456789ABCDEF, voutput_max);
307 vout3x0123456789ABCDEF = vminq_u8(vout3x0123456789ABCDEF, voutput_max);
308
309 if (nc >= 16) {
310 vst1q_u8(c0 + 0, vout0x0123456789ABCDEF);
311 vst1q_u8(c1 + 0, vout1x0123456789ABCDEF);
312 vst1q_u8(c2 + 0, vout2x0123456789ABCDEF);
313 vst1q_u8(c3 + 0, vout3x0123456789ABCDEF);
314
315 c0 = (uint8_t*) ((uintptr_t) c0 + cn_stride);
316 c1 = (uint8_t*) ((uintptr_t) c1 + cn_stride);
317 c2 = (uint8_t*) ((uintptr_t) c2 + cn_stride);
318 c3 = (uint8_t*) ((uintptr_t) c3 + cn_stride);
319
320 a0 = (const uint8_t*) ((uintptr_t) a0 - kc);
321 a1 = (const uint8_t*) ((uintptr_t) a1 - kc);
322 a2 = (const uint8_t*) ((uintptr_t) a2 - kc);
323 a3 = (const uint8_t*) ((uintptr_t) a3 - kc);
324
325 nc -= 16;
326 } else {
327 uint8x16_t vout0x01234567_1x01234567 = vcombine_u8(vget_low_u8(vout0x0123456789ABCDEF), vget_low_u8(vout1x0123456789ABCDEF));
328 uint8x16_t vout2x01234567_3x01234567 = vcombine_u8(vget_low_u8(vout2x0123456789ABCDEF), vget_low_u8(vout3x0123456789ABCDEF));
329 if (nc & 8) {
330 vst1_u8(c0, vget_low_u8(vout0x01234567_1x01234567)); c0 += 8;
331 vst1_u8(c1, vget_high_u8(vout0x01234567_1x01234567)); c1 += 8;
332 vst1_u8(c2, vget_low_u8(vout2x01234567_3x01234567)); c2 += 8;
333 vst1_u8(c3, vget_high_u8(vout2x01234567_3x01234567)); c3 += 8;
334 vout0x01234567_1x01234567 = vcombine_u8(vget_high_u8(vout0x0123456789ABCDEF), vget_high_u8(vout1x0123456789ABCDEF));
335 vout2x01234567_3x01234567 = vcombine_u8(vget_high_u8(vout2x0123456789ABCDEF), vget_high_u8(vout3x0123456789ABCDEF));
336 }
337 if (nc & 4) {
338 vst1q_lane_u32((void*) c0, vreinterpretq_u32_u8(vout0x01234567_1x01234567), 0); c0 += 4;
339 vst1q_lane_u32((void*) c1, vreinterpretq_u32_u8(vout0x01234567_1x01234567), 2); c1 += 4;
340 vst1q_lane_u32((void*) c2, vreinterpretq_u32_u8(vout2x01234567_3x01234567), 0); c2 += 4;
341 vst1q_lane_u32((void*) c3, vreinterpretq_u32_u8(vout2x01234567_3x01234567), 2); c3 += 4;
342 vout0x01234567_1x01234567 = vextq_u8(vout0x01234567_1x01234567, vout0x01234567_1x01234567, 4);
343 vout2x01234567_3x01234567 = vextq_u8(vout2x01234567_3x01234567, vout2x01234567_3x01234567, 4);
344 }
345 if (nc & 2) {
346 vst1q_lane_u16((void*) c0, vreinterpretq_u16_u8(vout0x01234567_1x01234567), 0); c0 += 2;
347 vst1q_lane_u16((void*) c1, vreinterpretq_u16_u8(vout0x01234567_1x01234567), 4); c1 += 2;
348 vst1q_lane_u16((void*) c2, vreinterpretq_u16_u8(vout2x01234567_3x01234567), 0); c2 += 2;
349 vst1q_lane_u16((void*) c3, vreinterpretq_u16_u8(vout2x01234567_3x01234567), 4); c3 += 2;
350 vout0x01234567_1x01234567 = vextq_u8(vout0x01234567_1x01234567, vout0x01234567_1x01234567, 2);
351 vout2x01234567_3x01234567 = vextq_u8(vout2x01234567_3x01234567, vout2x01234567_3x01234567, 2);
352 }
353 if (nc & 1) {
354 vst1q_lane_u8(c0, vout0x01234567_1x01234567, 0);
355 vst1q_lane_u8(c1, vout0x01234567_1x01234567, 8);
356 vst1q_lane_u8(c2, vout2x01234567_3x01234567, 0);
357 vst1q_lane_u8(c3, vout2x01234567_3x01234567, 8);
358 }
359
360 nc = 0;
361 }
362 } while (nc != 0);
363 }
364