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_3x32c4__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_3x32c4__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 <= 3);
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
55 const uint8x8_t va_zero_point = vld1_dup_u8(¶ms->rndnu_neon.kernel_zero_point[0]);
56
57 // Loop over groups of 32 columns.
58 do {
59 // Initialize accumulators with bias. 32 bias values are loaded from the
60 // weight matrix, at the start of the group of 32 columns.
61 uint32x4_t vpacc0x0123 = vld1q_u32(w); w = (const void*) ((const uint32_t*) w + 4);
62 uint32x4_t vpacc0x4567 = vld1q_u32(w); w = (const void*) ((const uint32_t*) w + 4);
63 uint32x4_t vpacc0x89AB = vld1q_u32(w); w = (const void*) ((const uint32_t*) w + 4);
64 uint32x4_t vpacc0xCDEF = vld1q_u32(w); w = (const void*) ((const uint32_t*) w + 4);
65 uint32x4_t vpacc0xGHIJ = vld1q_u32(w); w = (const void*) ((const uint32_t*) w + 4);
66 uint32x4_t vpacc0xKLMN = vld1q_u32(w); w = (const void*) ((const uint32_t*) w + 4);
67 uint32x4_t vpacc0xOPQR = vld1q_u32(w); w = (const void*) ((const uint32_t*) w + 4);
68 uint32x4_t vpacc0xSTUV = vld1q_u32(w); w = (const void*) ((const uint32_t*) w + 4);
69 uint32x4_t vpacc1x0123 = vpacc0x0123;
70 uint32x4_t vpacc1x4567 = vpacc0x4567;
71 uint32x4_t vpacc1x89AB = vpacc0x89AB;
72 uint32x4_t vpacc1xCDEF = vpacc0xCDEF;
73 uint32x4_t vpacc1xGHIJ = vpacc0xGHIJ;
74 uint32x4_t vpacc1xKLMN = vpacc0xKLMN;
75 uint32x4_t vpacc1xOPQR = vpacc0xOPQR;
76 uint32x4_t vpacc1xSTUV = vpacc0xSTUV;
77 uint32x4_t vpacc2x0123 = vpacc0x0123;
78 uint32x4_t vpacc2x4567 = vpacc0x4567;
79 uint32x4_t vpacc2x89AB = vpacc0x89AB;
80 uint32x4_t vpacc2xCDEF = vpacc0xCDEF;
81 uint32x4_t vpacc2xGHIJ = vpacc0xGHIJ;
82 uint32x4_t vpacc2xKLMN = vpacc0xKLMN;
83 uint32x4_t vpacc2xOPQR = vpacc0xOPQR;
84 uint32x4_t vpacc2xSTUV = vpacc0xSTUV;
85 uint32x2_t vnacc0 = vmov_n_u32(0);
86 uint32x2_t vnacc1 = vmov_n_u32(0);
87 uint32x2_t vnacc2 = vmov_n_u32(0);
88
89 // Inner accumulation loop along the 32 columns.
90 size_t k = kc;
91 // 2x partial unrolled loop to load 8 bytes at a time.
92 while (k >= 8 * sizeof(uint8_t)) {
93 // Load a 3x8 block of activations.
94 const uint8x8_t va0x01234567 = vld1_u8(a0); a0 += 8;
95 const uint8x8_t va1x01234567 = vld1_u8(a1); a1 += 8;
96 const uint8x8_t va2x01234567 = vld1_u8(a2); a2 += 8;
97
98 // Load a 8x32 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 vb0123xGHIJ = vld1q_u8(w); w = (const void*) ((const uint8_t*) w + 16);
104 const uint8x16_t vb0123xKLMN = vld1q_u8(w); w = (const void*) ((const uint8_t*) w + 16);
105 const uint8x16_t vb0123xOPQR = vld1q_u8(w); w = (const void*) ((const uint8_t*) w + 16);
106 const uint8x16_t vb0123xSTUV = vld1q_u8(w); w = (const void*) ((const uint8_t*) w + 16);
107 const uint8x16_t vb4567x0123 = vld1q_u8(w); w = (const void*) ((const uint8_t*) w + 16);
108 const uint8x16_t vb4567x4567 = vld1q_u8(w); w = (const void*) ((const uint8_t*) w + 16);
109 const uint8x16_t vb4567x89AB = vld1q_u8(w); w = (const void*) ((const uint8_t*) w + 16);
110 const uint8x16_t vb4567xCDEF = vld1q_u8(w); w = (const void*) ((const uint8_t*) w + 16);
111 const uint8x16_t vb4567xGHIJ = vld1q_u8(w); w = (const void*) ((const uint8_t*) w + 16);
112 const uint8x16_t vb4567xKLMN = vld1q_u8(w); w = (const void*) ((const uint8_t*) w + 16);
113 const uint8x16_t vb4567xOPQR = vld1q_u8(w); w = (const void*) ((const uint8_t*) w + 16);
114 const uint8x16_t vb4567xSTUV = vld1q_u8(w); w = (const void*) ((const uint8_t*) w + 16);
115
116 // Multiply-accumulate: 3x8 * 8x32 --> 3x32.
117 vnacc0 = vdot_u32(vnacc0, va_zero_point, va0x01234567);
118 vpacc0x0123 = vdotq_lane_u32(vpacc0x0123, vb0123x0123, va0x01234567, 0);
119 vpacc0x4567 = vdotq_lane_u32(vpacc0x4567, vb0123x4567, va0x01234567, 0);
120 vpacc0x89AB = vdotq_lane_u32(vpacc0x89AB, vb0123x89AB, va0x01234567, 0);
121 vpacc0xCDEF = vdotq_lane_u32(vpacc0xCDEF, vb0123xCDEF, va0x01234567, 0);
122 vpacc0xGHIJ = vdotq_lane_u32(vpacc0xGHIJ, vb0123xGHIJ, va0x01234567, 0);
123 vpacc0xKLMN = vdotq_lane_u32(vpacc0xKLMN, vb0123xKLMN, va0x01234567, 0);
124 vpacc0xOPQR = vdotq_lane_u32(vpacc0xOPQR, vb0123xOPQR, va0x01234567, 0);
125 vpacc0xSTUV = vdotq_lane_u32(vpacc0xSTUV, vb0123xSTUV, va0x01234567, 0);
126 vpacc0x0123 = vdotq_lane_u32(vpacc0x0123, vb4567x0123, va0x01234567, 1);
127 vpacc0x4567 = vdotq_lane_u32(vpacc0x4567, vb4567x4567, va0x01234567, 1);
128 vpacc0x89AB = vdotq_lane_u32(vpacc0x89AB, vb4567x89AB, va0x01234567, 1);
129 vpacc0xCDEF = vdotq_lane_u32(vpacc0xCDEF, vb4567xCDEF, va0x01234567, 1);
130 vpacc0xGHIJ = vdotq_lane_u32(vpacc0xGHIJ, vb4567xGHIJ, va0x01234567, 1);
131 vpacc0xKLMN = vdotq_lane_u32(vpacc0xKLMN, vb4567xKLMN, va0x01234567, 1);
132 vpacc0xOPQR = vdotq_lane_u32(vpacc0xOPQR, vb4567xOPQR, va0x01234567, 1);
133 vpacc0xSTUV = vdotq_lane_u32(vpacc0xSTUV, vb4567xSTUV, va0x01234567, 1);
134 vnacc1 = vdot_u32(vnacc1, va_zero_point, va1x01234567);
135 vpacc1x0123 = vdotq_lane_u32(vpacc1x0123, vb0123x0123, va1x01234567, 0);
136 vpacc1x4567 = vdotq_lane_u32(vpacc1x4567, vb0123x4567, va1x01234567, 0);
137 vpacc1x89AB = vdotq_lane_u32(vpacc1x89AB, vb0123x89AB, va1x01234567, 0);
138 vpacc1xCDEF = vdotq_lane_u32(vpacc1xCDEF, vb0123xCDEF, va1x01234567, 0);
139 vpacc1xGHIJ = vdotq_lane_u32(vpacc1xGHIJ, vb0123xGHIJ, va1x01234567, 0);
140 vpacc1xKLMN = vdotq_lane_u32(vpacc1xKLMN, vb0123xKLMN, va1x01234567, 0);
141 vpacc1xOPQR = vdotq_lane_u32(vpacc1xOPQR, vb0123xOPQR, va1x01234567, 0);
142 vpacc1xSTUV = vdotq_lane_u32(vpacc1xSTUV, vb0123xSTUV, va1x01234567, 0);
143 vpacc1x0123 = vdotq_lane_u32(vpacc1x0123, vb4567x0123, va1x01234567, 1);
144 vpacc1x4567 = vdotq_lane_u32(vpacc1x4567, vb4567x4567, va1x01234567, 1);
145 vpacc1x89AB = vdotq_lane_u32(vpacc1x89AB, vb4567x89AB, va1x01234567, 1);
146 vpacc1xCDEF = vdotq_lane_u32(vpacc1xCDEF, vb4567xCDEF, va1x01234567, 1);
147 vpacc1xGHIJ = vdotq_lane_u32(vpacc1xGHIJ, vb4567xGHIJ, va1x01234567, 1);
148 vpacc1xKLMN = vdotq_lane_u32(vpacc1xKLMN, vb4567xKLMN, va1x01234567, 1);
149 vpacc1xOPQR = vdotq_lane_u32(vpacc1xOPQR, vb4567xOPQR, va1x01234567, 1);
150 vpacc1xSTUV = vdotq_lane_u32(vpacc1xSTUV, vb4567xSTUV, va1x01234567, 1);
151 vnacc2 = vdot_u32(vnacc2, va_zero_point, va2x01234567);
152 vpacc2x0123 = vdotq_lane_u32(vpacc2x0123, vb0123x0123, va2x01234567, 0);
153 vpacc2x4567 = vdotq_lane_u32(vpacc2x4567, vb0123x4567, va2x01234567, 0);
154 vpacc2x89AB = vdotq_lane_u32(vpacc2x89AB, vb0123x89AB, va2x01234567, 0);
155 vpacc2xCDEF = vdotq_lane_u32(vpacc2xCDEF, vb0123xCDEF, va2x01234567, 0);
156 vpacc2xGHIJ = vdotq_lane_u32(vpacc2xGHIJ, vb0123xGHIJ, va2x01234567, 0);
157 vpacc2xKLMN = vdotq_lane_u32(vpacc2xKLMN, vb0123xKLMN, va2x01234567, 0);
158 vpacc2xOPQR = vdotq_lane_u32(vpacc2xOPQR, vb0123xOPQR, va2x01234567, 0);
159 vpacc2xSTUV = vdotq_lane_u32(vpacc2xSTUV, vb0123xSTUV, va2x01234567, 0);
160 vpacc2x0123 = vdotq_lane_u32(vpacc2x0123, vb4567x0123, va2x01234567, 1);
161 vpacc2x4567 = vdotq_lane_u32(vpacc2x4567, vb4567x4567, va2x01234567, 1);
162 vpacc2x89AB = vdotq_lane_u32(vpacc2x89AB, vb4567x89AB, va2x01234567, 1);
163 vpacc2xCDEF = vdotq_lane_u32(vpacc2xCDEF, vb4567xCDEF, va2x01234567, 1);
164 vpacc2xGHIJ = vdotq_lane_u32(vpacc2xGHIJ, vb4567xGHIJ, va2x01234567, 1);
165 vpacc2xKLMN = vdotq_lane_u32(vpacc2xKLMN, vb4567xKLMN, va2x01234567, 1);
166 vpacc2xOPQR = vdotq_lane_u32(vpacc2xOPQR, vb4567xOPQR, va2x01234567, 1);
167 vpacc2xSTUV = vdotq_lane_u32(vpacc2xSTUV, vb4567xSTUV, va2x01234567, 1);
168
169 k -= 8 * sizeof(uint8_t);
170 }
171 // Handle up to 4 final positions of `k`
172 if XNN_UNLIKELY(k != 0) {
173 // Load a 3x4 block of activations.
174 const uint8x8_t va0x01234567 = vreinterpret_u8_u32(vld1_lane_u32((const void*) a0, vmov_n_u32(0), 0)); a0 += 4;
175 const uint8x8_t va1x01234567 = vreinterpret_u8_u32(vld1_lane_u32((const void*) a1, vmov_n_u32(0), 0)); a1 += 4;
176 const uint8x8_t va2x01234567 = vreinterpret_u8_u32(vld1_lane_u32((const void*) a2, vmov_n_u32(0), 0)); a2 += 4;
177
178 // Load a 4x32 block of weights.
179 const uint8x16_t vb0123x0123 = vld1q_u8(w); w = (const void*) ((const uint8_t*) w + 16);
180 const uint8x16_t vb0123x4567 = vld1q_u8(w); w = (const void*) ((const uint8_t*) w + 16);
181 const uint8x16_t vb0123x89AB = vld1q_u8(w); w = (const void*) ((const uint8_t*) w + 16);
182 const uint8x16_t vb0123xCDEF = vld1q_u8(w); w = (const void*) ((const uint8_t*) w + 16);
183 const uint8x16_t vb0123xGHIJ = vld1q_u8(w); w = (const void*) ((const uint8_t*) w + 16);
184 const uint8x16_t vb0123xKLMN = vld1q_u8(w); w = (const void*) ((const uint8_t*) w + 16);
185 const uint8x16_t vb0123xOPQR = vld1q_u8(w); w = (const void*) ((const uint8_t*) w + 16);
186 const uint8x16_t vb0123xSTUV = vld1q_u8(w); w = (const void*) ((const uint8_t*) w + 16);
187
188 // Multiply-accumulate: 3x4 * 4x32 --> 3x32.
189 vnacc0 = vdot_u32(vnacc0, va_zero_point, va0x01234567);
190 vpacc0x0123 = vdotq_lane_u32(vpacc0x0123, vb0123x0123, va0x01234567, 0);
191 vpacc0x4567 = vdotq_lane_u32(vpacc0x4567, vb0123x4567, va0x01234567, 0);
192 vpacc0x89AB = vdotq_lane_u32(vpacc0x89AB, vb0123x89AB, va0x01234567, 0);
193 vpacc0xCDEF = vdotq_lane_u32(vpacc0xCDEF, vb0123xCDEF, va0x01234567, 0);
194 vpacc0xGHIJ = vdotq_lane_u32(vpacc0xGHIJ, vb0123xGHIJ, va0x01234567, 0);
195 vpacc0xKLMN = vdotq_lane_u32(vpacc0xKLMN, vb0123xKLMN, va0x01234567, 0);
196 vpacc0xOPQR = vdotq_lane_u32(vpacc0xOPQR, vb0123xOPQR, va0x01234567, 0);
197 vpacc0xSTUV = vdotq_lane_u32(vpacc0xSTUV, vb0123xSTUV, va0x01234567, 0);
198 vnacc1 = vdot_u32(vnacc1, va_zero_point, va1x01234567);
199 vpacc1x0123 = vdotq_lane_u32(vpacc1x0123, vb0123x0123, va1x01234567, 0);
200 vpacc1x4567 = vdotq_lane_u32(vpacc1x4567, vb0123x4567, va1x01234567, 0);
201 vpacc1x89AB = vdotq_lane_u32(vpacc1x89AB, vb0123x89AB, va1x01234567, 0);
202 vpacc1xCDEF = vdotq_lane_u32(vpacc1xCDEF, vb0123xCDEF, va1x01234567, 0);
203 vpacc1xGHIJ = vdotq_lane_u32(vpacc1xGHIJ, vb0123xGHIJ, va1x01234567, 0);
204 vpacc1xKLMN = vdotq_lane_u32(vpacc1xKLMN, vb0123xKLMN, va1x01234567, 0);
205 vpacc1xOPQR = vdotq_lane_u32(vpacc1xOPQR, vb0123xOPQR, va1x01234567, 0);
206 vpacc1xSTUV = vdotq_lane_u32(vpacc1xSTUV, vb0123xSTUV, va1x01234567, 0);
207 vnacc2 = vdot_u32(vnacc2, va_zero_point, va2x01234567);
208 vpacc2x0123 = vdotq_lane_u32(vpacc2x0123, vb0123x0123, va2x01234567, 0);
209 vpacc2x4567 = vdotq_lane_u32(vpacc2x4567, vb0123x4567, va2x01234567, 0);
210 vpacc2x89AB = vdotq_lane_u32(vpacc2x89AB, vb0123x89AB, va2x01234567, 0);
211 vpacc2xCDEF = vdotq_lane_u32(vpacc2xCDEF, vb0123xCDEF, va2x01234567, 0);
212 vpacc2xGHIJ = vdotq_lane_u32(vpacc2xGHIJ, vb0123xGHIJ, va2x01234567, 0);
213 vpacc2xKLMN = vdotq_lane_u32(vpacc2xKLMN, vb0123xKLMN, va2x01234567, 0);
214 vpacc2xOPQR = vdotq_lane_u32(vpacc2xOPQR, vb0123xOPQR, va2x01234567, 0);
215 vpacc2xSTUV = vdotq_lane_u32(vpacc2xSTUV, vb0123xSTUV, va2x01234567, 0);
216 }
217
218 // Subtract zero point from accumulators.
219 vnacc0 = vpadd_u32(vnacc0, vnacc0);
220 const uint32x4_t vnacc0x0123 = vcombine_u32(vnacc0, vnacc0);
221 int32x4_t vacc0x0123 = vreinterpretq_s32_u32(vsubq_u32(vpacc0x0123, vnacc0x0123));
222 int32x4_t vacc0x4567 = vreinterpretq_s32_u32(vsubq_u32(vpacc0x4567, vnacc0x0123));
223 int32x4_t vacc0x89AB = vreinterpretq_s32_u32(vsubq_u32(vpacc0x89AB, vnacc0x0123));
224 int32x4_t vacc0xCDEF = vreinterpretq_s32_u32(vsubq_u32(vpacc0xCDEF, vnacc0x0123));
225 int32x4_t vacc0xGHIJ = vreinterpretq_s32_u32(vsubq_u32(vpacc0xGHIJ, vnacc0x0123));
226 int32x4_t vacc0xKLMN = vreinterpretq_s32_u32(vsubq_u32(vpacc0xKLMN, vnacc0x0123));
227 int32x4_t vacc0xOPQR = vreinterpretq_s32_u32(vsubq_u32(vpacc0xOPQR, vnacc0x0123));
228 int32x4_t vacc0xSTUV = vreinterpretq_s32_u32(vsubq_u32(vpacc0xSTUV, vnacc0x0123));
229 vnacc1 = vpadd_u32(vnacc1, vnacc1);
230 const uint32x4_t vnacc1x0123 = vcombine_u32(vnacc1, vnacc1);
231 int32x4_t vacc1x0123 = vreinterpretq_s32_u32(vsubq_u32(vpacc1x0123, vnacc1x0123));
232 int32x4_t vacc1x4567 = vreinterpretq_s32_u32(vsubq_u32(vpacc1x4567, vnacc1x0123));
233 int32x4_t vacc1x89AB = vreinterpretq_s32_u32(vsubq_u32(vpacc1x89AB, vnacc1x0123));
234 int32x4_t vacc1xCDEF = vreinterpretq_s32_u32(vsubq_u32(vpacc1xCDEF, vnacc1x0123));
235 int32x4_t vacc1xGHIJ = vreinterpretq_s32_u32(vsubq_u32(vpacc1xGHIJ, vnacc1x0123));
236 int32x4_t vacc1xKLMN = vreinterpretq_s32_u32(vsubq_u32(vpacc1xKLMN, vnacc1x0123));
237 int32x4_t vacc1xOPQR = vreinterpretq_s32_u32(vsubq_u32(vpacc1xOPQR, vnacc1x0123));
238 int32x4_t vacc1xSTUV = vreinterpretq_s32_u32(vsubq_u32(vpacc1xSTUV, vnacc1x0123));
239 vnacc2 = vpadd_u32(vnacc2, vnacc2);
240 const uint32x4_t vnacc2x0123 = vcombine_u32(vnacc2, vnacc2);
241 int32x4_t vacc2x0123 = vreinterpretq_s32_u32(vsubq_u32(vpacc2x0123, vnacc2x0123));
242 int32x4_t vacc2x4567 = vreinterpretq_s32_u32(vsubq_u32(vpacc2x4567, vnacc2x0123));
243 int32x4_t vacc2x89AB = vreinterpretq_s32_u32(vsubq_u32(vpacc2x89AB, vnacc2x0123));
244 int32x4_t vacc2xCDEF = vreinterpretq_s32_u32(vsubq_u32(vpacc2xCDEF, vnacc2x0123));
245 int32x4_t vacc2xGHIJ = vreinterpretq_s32_u32(vsubq_u32(vpacc2xGHIJ, vnacc2x0123));
246 int32x4_t vacc2xKLMN = vreinterpretq_s32_u32(vsubq_u32(vpacc2xKLMN, vnacc2x0123));
247 int32x4_t vacc2xOPQR = vreinterpretq_s32_u32(vsubq_u32(vpacc2xOPQR, vnacc2x0123));
248 int32x4_t vacc2xSTUV = vreinterpretq_s32_u32(vsubq_u32(vpacc2xSTUV, vnacc2x0123));
249
250 const int32x4_t vright_pre_shift = vld1q_dup_s32(¶ms->rndnu_neon.right_pre_shift);
251 const int32x4_t vmultiplier = vld1q_dup_s32(¶ms->rndnu_neon.multiplier);
252 const int32x4_t vright_post_shift = vld1q_dup_s32(¶ms->rndnu_neon.right_post_shift);
253
254 vacc0x0123 = vshlq_s32(vacc0x0123, vright_pre_shift);
255 vacc0x4567 = vshlq_s32(vacc0x4567, vright_pre_shift);
256 vacc0x89AB = vshlq_s32(vacc0x89AB, vright_pre_shift);
257 vacc0xCDEF = vshlq_s32(vacc0xCDEF, vright_pre_shift);
258 vacc0xGHIJ = vshlq_s32(vacc0xGHIJ, vright_pre_shift);
259 vacc0xKLMN = vshlq_s32(vacc0xKLMN, vright_pre_shift);
260 vacc0xOPQR = vshlq_s32(vacc0xOPQR, vright_pre_shift);
261 vacc0xSTUV = vshlq_s32(vacc0xSTUV, vright_pre_shift);
262 vacc1x0123 = vshlq_s32(vacc1x0123, vright_pre_shift);
263 vacc1x4567 = vshlq_s32(vacc1x4567, vright_pre_shift);
264 vacc1x89AB = vshlq_s32(vacc1x89AB, vright_pre_shift);
265 vacc1xCDEF = vshlq_s32(vacc1xCDEF, vright_pre_shift);
266 vacc1xGHIJ = vshlq_s32(vacc1xGHIJ, vright_pre_shift);
267 vacc1xKLMN = vshlq_s32(vacc1xKLMN, vright_pre_shift);
268 vacc1xOPQR = vshlq_s32(vacc1xOPQR, vright_pre_shift);
269 vacc1xSTUV = vshlq_s32(vacc1xSTUV, vright_pre_shift);
270 vacc2x0123 = vshlq_s32(vacc2x0123, vright_pre_shift);
271 vacc2x4567 = vshlq_s32(vacc2x4567, vright_pre_shift);
272 vacc2x89AB = vshlq_s32(vacc2x89AB, vright_pre_shift);
273 vacc2xCDEF = vshlq_s32(vacc2xCDEF, vright_pre_shift);
274 vacc2xGHIJ = vshlq_s32(vacc2xGHIJ, vright_pre_shift);
275 vacc2xKLMN = vshlq_s32(vacc2xKLMN, vright_pre_shift);
276 vacc2xOPQR = vshlq_s32(vacc2xOPQR, vright_pre_shift);
277 vacc2xSTUV = vshlq_s32(vacc2xSTUV, vright_pre_shift);
278
279 vacc0x0123 = vqdmulhq_s32(vacc0x0123, vmultiplier);
280 vacc0x4567 = vqdmulhq_s32(vacc0x4567, vmultiplier);
281 vacc0x89AB = vqdmulhq_s32(vacc0x89AB, vmultiplier);
282 vacc0xCDEF = vqdmulhq_s32(vacc0xCDEF, vmultiplier);
283 vacc0xGHIJ = vqdmulhq_s32(vacc0xGHIJ, vmultiplier);
284 vacc0xKLMN = vqdmulhq_s32(vacc0xKLMN, vmultiplier);
285 vacc0xOPQR = vqdmulhq_s32(vacc0xOPQR, vmultiplier);
286 vacc0xSTUV = vqdmulhq_s32(vacc0xSTUV, vmultiplier);
287 vacc1x0123 = vqdmulhq_s32(vacc1x0123, vmultiplier);
288 vacc1x4567 = vqdmulhq_s32(vacc1x4567, vmultiplier);
289 vacc1x89AB = vqdmulhq_s32(vacc1x89AB, vmultiplier);
290 vacc1xCDEF = vqdmulhq_s32(vacc1xCDEF, vmultiplier);
291 vacc1xGHIJ = vqdmulhq_s32(vacc1xGHIJ, vmultiplier);
292 vacc1xKLMN = vqdmulhq_s32(vacc1xKLMN, vmultiplier);
293 vacc1xOPQR = vqdmulhq_s32(vacc1xOPQR, vmultiplier);
294 vacc1xSTUV = vqdmulhq_s32(vacc1xSTUV, vmultiplier);
295 vacc2x0123 = vqdmulhq_s32(vacc2x0123, vmultiplier);
296 vacc2x4567 = vqdmulhq_s32(vacc2x4567, vmultiplier);
297 vacc2x89AB = vqdmulhq_s32(vacc2x89AB, vmultiplier);
298 vacc2xCDEF = vqdmulhq_s32(vacc2xCDEF, vmultiplier);
299 vacc2xGHIJ = vqdmulhq_s32(vacc2xGHIJ, vmultiplier);
300 vacc2xKLMN = vqdmulhq_s32(vacc2xKLMN, vmultiplier);
301 vacc2xOPQR = vqdmulhq_s32(vacc2xOPQR, vmultiplier);
302 vacc2xSTUV = vqdmulhq_s32(vacc2xSTUV, vmultiplier);
303
304 vacc0x0123 = vrshlq_s32(vacc0x0123, vright_post_shift);
305 vacc0x4567 = vrshlq_s32(vacc0x4567, vright_post_shift);
306 vacc0x89AB = vrshlq_s32(vacc0x89AB, vright_post_shift);
307 vacc0xCDEF = vrshlq_s32(vacc0xCDEF, vright_post_shift);
308 vacc0xGHIJ = vrshlq_s32(vacc0xGHIJ, vright_post_shift);
309 vacc0xKLMN = vrshlq_s32(vacc0xKLMN, vright_post_shift);
310 vacc0xOPQR = vrshlq_s32(vacc0xOPQR, vright_post_shift);
311 vacc0xSTUV = vrshlq_s32(vacc0xSTUV, vright_post_shift);
312 vacc1x0123 = vrshlq_s32(vacc1x0123, vright_post_shift);
313 vacc1x4567 = vrshlq_s32(vacc1x4567, vright_post_shift);
314 vacc1x89AB = vrshlq_s32(vacc1x89AB, vright_post_shift);
315 vacc1xCDEF = vrshlq_s32(vacc1xCDEF, vright_post_shift);
316 vacc1xGHIJ = vrshlq_s32(vacc1xGHIJ, vright_post_shift);
317 vacc1xKLMN = vrshlq_s32(vacc1xKLMN, vright_post_shift);
318 vacc1xOPQR = vrshlq_s32(vacc1xOPQR, vright_post_shift);
319 vacc1xSTUV = vrshlq_s32(vacc1xSTUV, vright_post_shift);
320 vacc2x0123 = vrshlq_s32(vacc2x0123, vright_post_shift);
321 vacc2x4567 = vrshlq_s32(vacc2x4567, vright_post_shift);
322 vacc2x89AB = vrshlq_s32(vacc2x89AB, vright_post_shift);
323 vacc2xCDEF = vrshlq_s32(vacc2xCDEF, vright_post_shift);
324 vacc2xGHIJ = vrshlq_s32(vacc2xGHIJ, vright_post_shift);
325 vacc2xKLMN = vrshlq_s32(vacc2xKLMN, vright_post_shift);
326 vacc2xOPQR = vrshlq_s32(vacc2xOPQR, vright_post_shift);
327 vacc2xSTUV = vrshlq_s32(vacc2xSTUV, vright_post_shift);
328
329 const int16x8_t voutput_zero_point = vld1q_dup_s16(¶ms->rndnu_neon.output_zero_point);
330 #if XNN_ARCH_ARM64
331 const int16x8_t vacc0x01234567 = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc0x0123), vacc0x4567), voutput_zero_point);
332 const int16x8_t vacc0x89ABCDEF = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc0x89AB), vacc0xCDEF), voutput_zero_point);
333 const int16x8_t vacc0xGHIJKLMN = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc0xGHIJ), vacc0xKLMN), voutput_zero_point);
334 const int16x8_t vacc0xOPQRSTUV = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc0xOPQR), vacc0xSTUV), 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 vacc1xGHIJKLMN = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc1xGHIJ), vacc1xKLMN), voutput_zero_point);
338 const int16x8_t vacc1xOPQRSTUV = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc1xOPQR), vacc1xSTUV), voutput_zero_point);
339 const int16x8_t vacc2x01234567 = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc2x0123), vacc2x4567), voutput_zero_point);
340 const int16x8_t vacc2x89ABCDEF = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc2x89AB), vacc2xCDEF), voutput_zero_point);
341 const int16x8_t vacc2xGHIJKLMN = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc2xGHIJ), vacc2xKLMN), voutput_zero_point);
342 const int16x8_t vacc2xOPQRSTUV = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc2xOPQR), vacc2xSTUV), voutput_zero_point);
343
344 uint8x16_t vout0x0123456789ABCDEF = vqmovun_high_s16(vqmovun_s16(vacc0x01234567), vacc0x89ABCDEF);
345 uint8x16_t vout0xGHIJKLMNOPQRSTUV = vqmovun_high_s16(vqmovun_s16(vacc0xGHIJKLMN), vacc0xOPQRSTUV);
346 uint8x16_t vout1x0123456789ABCDEF = vqmovun_high_s16(vqmovun_s16(vacc1x01234567), vacc1x89ABCDEF);
347 uint8x16_t vout1xGHIJKLMNOPQRSTUV = vqmovun_high_s16(vqmovun_s16(vacc1xGHIJKLMN), vacc1xOPQRSTUV);
348 uint8x16_t vout2x0123456789ABCDEF = vqmovun_high_s16(vqmovun_s16(vacc2x01234567), vacc2x89ABCDEF);
349 uint8x16_t vout2xGHIJKLMNOPQRSTUV = vqmovun_high_s16(vqmovun_s16(vacc2xGHIJKLMN), vacc2xOPQRSTUV);
350 #else
351 const int16x8_t vacc0x01234567 = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc0x0123), vqmovn_s32(vacc0x4567)), voutput_zero_point);
352 const int16x8_t vacc0x89ABCDEF = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc0x89AB), vqmovn_s32(vacc0xCDEF)), voutput_zero_point);
353 const int16x8_t vacc0xGHIJKLMN = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc0xGHIJ), vqmovn_s32(vacc0xKLMN)), voutput_zero_point);
354 const int16x8_t vacc0xOPQRSTUV = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc0xOPQR), vqmovn_s32(vacc0xSTUV)), voutput_zero_point);
355 const int16x8_t vacc1x01234567 = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc1x0123), vqmovn_s32(vacc1x4567)), voutput_zero_point);
356 const int16x8_t vacc1x89ABCDEF = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc1x89AB), vqmovn_s32(vacc1xCDEF)), voutput_zero_point);
357 const int16x8_t vacc1xGHIJKLMN = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc1xGHIJ), vqmovn_s32(vacc1xKLMN)), voutput_zero_point);
358 const int16x8_t vacc1xOPQRSTUV = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc1xOPQR), vqmovn_s32(vacc1xSTUV)), voutput_zero_point);
359 const int16x8_t vacc2x01234567 = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc2x0123), vqmovn_s32(vacc2x4567)), voutput_zero_point);
360 const int16x8_t vacc2x89ABCDEF = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc2x89AB), vqmovn_s32(vacc2xCDEF)), voutput_zero_point);
361 const int16x8_t vacc2xGHIJKLMN = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc2xGHIJ), vqmovn_s32(vacc2xKLMN)), voutput_zero_point);
362 const int16x8_t vacc2xOPQRSTUV = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc2xOPQR), vqmovn_s32(vacc2xSTUV)), voutput_zero_point);
363
364 uint8x16_t vout0x0123456789ABCDEF = vcombine_u8(vqmovun_s16(vacc0x01234567), vqmovun_s16(vacc0x89ABCDEF));
365 uint8x16_t vout0xGHIJKLMNOPQRSTUV = vcombine_u8(vqmovun_s16(vacc0xGHIJKLMN), vqmovun_s16(vacc0xOPQRSTUV));
366 uint8x16_t vout1x0123456789ABCDEF = vcombine_u8(vqmovun_s16(vacc1x01234567), vqmovun_s16(vacc1x89ABCDEF));
367 uint8x16_t vout1xGHIJKLMNOPQRSTUV = vcombine_u8(vqmovun_s16(vacc1xGHIJKLMN), vqmovun_s16(vacc1xOPQRSTUV));
368 uint8x16_t vout2x0123456789ABCDEF = vcombine_u8(vqmovun_s16(vacc2x01234567), vqmovun_s16(vacc2x89ABCDEF));
369 uint8x16_t vout2xGHIJKLMNOPQRSTUV = vcombine_u8(vqmovun_s16(vacc2xGHIJKLMN), vqmovun_s16(vacc2xOPQRSTUV));
370 #endif
371 const uint8x16_t voutput_min = vld1q_dup_u8(¶ms->rndnu_neon.output_min);
372 const uint8x16_t voutput_max = vld1q_dup_u8(¶ms->rndnu_neon.output_max);
373
374 vout0x0123456789ABCDEF = vmaxq_u8(vout0x0123456789ABCDEF, voutput_min);
375 vout0xGHIJKLMNOPQRSTUV = vmaxq_u8(vout0xGHIJKLMNOPQRSTUV, voutput_min);
376 vout1x0123456789ABCDEF = vmaxq_u8(vout1x0123456789ABCDEF, voutput_min);
377 vout1xGHIJKLMNOPQRSTUV = vmaxq_u8(vout1xGHIJKLMNOPQRSTUV, voutput_min);
378 vout2x0123456789ABCDEF = vmaxq_u8(vout2x0123456789ABCDEF, voutput_min);
379 vout2xGHIJKLMNOPQRSTUV = vmaxq_u8(vout2xGHIJKLMNOPQRSTUV, voutput_min);
380
381 vout0x0123456789ABCDEF = vminq_u8(vout0x0123456789ABCDEF, voutput_max);
382 vout0xGHIJKLMNOPQRSTUV = vminq_u8(vout0xGHIJKLMNOPQRSTUV, voutput_max);
383 vout1x0123456789ABCDEF = vminq_u8(vout1x0123456789ABCDEF, voutput_max);
384 vout1xGHIJKLMNOPQRSTUV = vminq_u8(vout1xGHIJKLMNOPQRSTUV, voutput_max);
385 vout2x0123456789ABCDEF = vminq_u8(vout2x0123456789ABCDEF, voutput_max);
386 vout2xGHIJKLMNOPQRSTUV = vminq_u8(vout2xGHIJKLMNOPQRSTUV, voutput_max);
387
388 if (nc >= 32) {
389 vst1q_u8(c0 + 0, vout0x0123456789ABCDEF);
390 vst1q_u8(c0 + 16, vout0xGHIJKLMNOPQRSTUV);
391 vst1q_u8(c1 + 0, vout1x0123456789ABCDEF);
392 vst1q_u8(c1 + 16, vout1xGHIJKLMNOPQRSTUV);
393 vst1q_u8(c2 + 0, vout2x0123456789ABCDEF);
394 vst1q_u8(c2 + 16, vout2xGHIJKLMNOPQRSTUV);
395
396 c0 = (uint8_t*) ((uintptr_t) c0 + cn_stride);
397 c1 = (uint8_t*) ((uintptr_t) c1 + cn_stride);
398 c2 = (uint8_t*) ((uintptr_t) c2 + cn_stride);
399
400 a0 = (const uint8_t*) ((uintptr_t) a0 - kc);
401 a1 = (const uint8_t*) ((uintptr_t) a1 - kc);
402 a2 = (const uint8_t*) ((uintptr_t) a2 - kc);
403
404 nc -= 32;
405 } else {
406 if (nc & 16) {
407 vst1q_u8(c0, vout0x0123456789ABCDEF); c0 += 16;
408 vst1q_u8(c1, vout1x0123456789ABCDEF); c1 += 16;
409 vst1q_u8(c2, vout2x0123456789ABCDEF); c2 += 16;
410
411 vout0x0123456789ABCDEF = vout0xGHIJKLMNOPQRSTUV;
412 vout1x0123456789ABCDEF = vout1xGHIJKLMNOPQRSTUV;
413 vout2x0123456789ABCDEF = vout2xGHIJKLMNOPQRSTUV;
414 }
415 uint8x16_t vout0x01234567_1x01234567 = vcombine_u8(vget_low_u8(vout0x0123456789ABCDEF), vget_low_u8(vout1x0123456789ABCDEF));
416 uint8x8_t vout2x01234567 = vget_low_u8(vout2x0123456789ABCDEF);
417 if (nc & 8) {
418 vst1_u8(c0, vget_low_u8(vout0x01234567_1x01234567)); c0 += 8;
419 vst1_u8(c1, vget_high_u8(vout0x01234567_1x01234567)); c1 += 8;
420 vst1_u8(c2, vout2x01234567); c2 += 8;
421 vout0x01234567_1x01234567 = vcombine_u8(vget_high_u8(vout0x0123456789ABCDEF), vget_high_u8(vout1x0123456789ABCDEF));
422 vout2x01234567 = vget_high_u8(vout2x0123456789ABCDEF);
423 }
424 if (nc & 4) {
425 vst1q_lane_u32((void*) c0, vreinterpretq_u32_u8(vout0x01234567_1x01234567), 0); c0 += 4;
426 vst1q_lane_u32((void*) c1, vreinterpretq_u32_u8(vout0x01234567_1x01234567), 2); c1 += 4;
427 vst1_lane_u32((void*) c2, vreinterpret_u32_u8(vout2x01234567), 0); c2 += 4;
428 vout0x01234567_1x01234567 = vextq_u8(vout0x01234567_1x01234567, vout0x01234567_1x01234567, 4);
429 vout2x01234567 = vext_u8(vout2x01234567, vout2x01234567, 4);
430 }
431 if (nc & 2) {
432 vst1q_lane_u16((void*) c0, vreinterpretq_u16_u8(vout0x01234567_1x01234567), 0); c0 += 2;
433 vst1q_lane_u16((void*) c1, vreinterpretq_u16_u8(vout0x01234567_1x01234567), 4); c1 += 2;
434 vst1_lane_u16((void*) c2, vreinterpret_u16_u8(vout2x01234567), 0); c2 += 2;
435 vout0x01234567_1x01234567 = vextq_u8(vout0x01234567_1x01234567, vout0x01234567_1x01234567, 2);
436 vout2x01234567 = vext_u8(vout2x01234567, vout2x01234567, 2);
437 }
438 if (nc & 1) {
439 vst1q_lane_u8(c0, vout0x01234567_1x01234567, 0);
440 vst1q_lane_u8(c1, vout0x01234567_1x01234567, 8);
441 vst1_lane_u8(c2, vout2x01234567, 0);
442 }
443
444 nc = 0;
445 }
446 } while (nc != 0);
447 }
448