Debian Codegraph @

src/qd8-f16-qc4w-gemm/gen/qd8-f16-qc4w-gemm-5x16c8-minmax-neoni8mm.c

raw
1 // Auto-generated file. Do not edit!
2 // Template: src/qs8-gemm/c8-neoni8mm.c.in
3 // Generator: tools/xngen
4 //
5 // Copyright 2023 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
18 void xnn_qd8_f16_qc4w_gemm_minmax_ukernel_5x16c8__neoni8mm(
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 xnn_float16* restrict c,
26 size_t cm_stride,
27 size_t cn_stride,
28 const struct xnn_f16_qc4w_minmax_params params[restrict XNN_MIN_ELEMENTS(1)],
29 const struct xnn_qd8_quantization_params quantization_params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
30 {
31 assert(mr != 0);
32 assert(mr <= 5);
33 assert(nc != 0);
34 assert(kc != 0);
35 assert(kc % sizeof(int8_t) == 0);
36 assert(a != NULL);
37 assert(w != NULL);
38 assert(c != NULL);
39
40 kc = round_up_po2(kc, 8 * sizeof(int8_t));
41 const int8_t* a0 = a;
42 uint16_t* c0 = (uint16_t*) c;
43 const int8_t* a1 = (const int8_t*) ((uintptr_t) a0 + a_stride);
44 uint16_t* c1 = (uint16_t*) ((uintptr_t) c0 + cm_stride);
45 if XNN_UNPREDICTABLE(mr < 2) {
46 a1 = a0;
47 c1 = c0;
48 }
49 const int8_t* a2 = (const int8_t*) ((uintptr_t) a1 + a_stride);
50 uint16_t* c2 = (uint16_t*) ((uintptr_t) c1 + cm_stride);
51 if XNN_UNPREDICTABLE(mr <= 2) {
52 a2 = a1;
53 c2 = c1;
54 }
55 const int8_t* a3 = (const int8_t*) ((uintptr_t) a2 + a_stride);
56 uint16_t* c3 = (uint16_t*) ((uintptr_t) c2 + cm_stride);
57 if XNN_UNPREDICTABLE(mr < 4) {
58 a3 = a2;
59 c3 = c2;
60 }
61 const int8_t* a4 = (const int8_t*) ((uintptr_t) a3 + a_stride);
62 uint16_t* c4 = (uint16_t*) ((uintptr_t) c3 + cm_stride);
63 if XNN_UNPREDICTABLE(mr <= 4) {
64 a4 = a3;
65 c4 = c3;
66 }
67 const int8x16_t vmask = vmovq_n_s8(INT8_C(0xF0));
68
69 // Loop over groups of 16 columns.
70 do {
71 // Initialize accumulators with bias. 16 bias values are loaded from the
72 // weight matrix, at the start of the group of 16 columns.
73 const int32x4_t vinput_zero_point01 = vld1q_s32(&quantization_params[0].zero_point);
74 const int32x4_t vksum0123 = vld1q_s32(w); w = (const int32_t*) w + 4;
75 const int32x4_t vksumzp0x0123 = vmulq_lane_s32(vksum0123, vget_low_s32(vinput_zero_point01), 0);
76 const int32x4_t vksumzp1x0123 = vmulq_lane_s32(vksum0123, vget_high_s32(vinput_zero_point01), 0);
77 const int32x4_t vksum4567 = vld1q_s32(w); w = (const int32_t*) w + 4;
78 const int32x4_t vksumzp0x4567 = vmulq_lane_s32(vksum4567, vget_low_s32(vinput_zero_point01), 0);
79 const int32x4_t vksumzp1x4567 = vmulq_lane_s32(vksum4567, vget_high_s32(vinput_zero_point01), 0);
80 const int32x4_t vksum89AB = vld1q_s32(w); w = (const int32_t*) w + 4;
81 const int32x4_t vksumzp0x89AB = vmulq_lane_s32(vksum89AB, vget_low_s32(vinput_zero_point01), 0);
82 const int32x4_t vksumzp1x89AB = vmulq_lane_s32(vksum89AB, vget_high_s32(vinput_zero_point01), 0);
83 const int32x4_t vksumCDEF = vld1q_s32(w); w = (const int32_t*) w + 4;
84 const int32x4_t vksumzp0xCDEF = vmulq_lane_s32(vksumCDEF, vget_low_s32(vinput_zero_point01), 0);
85 const int32x4_t vksumzp1xCDEF = vmulq_lane_s32(vksumCDEF, vget_high_s32(vinput_zero_point01), 0);
86 const int32x4_t vinput_zero_point23 = vld1q_s32(&quantization_params[2].zero_point);
87 const int32x4_t vksumzp2x0123 = vmulq_lane_s32(vksum0123, vget_low_s32(vinput_zero_point23), 0);
88 const int32x4_t vksumzp3x0123 = vmulq_lane_s32(vksum0123, vget_high_s32(vinput_zero_point23), 0);
89 const int32x4_t vksumzp2x4567 = vmulq_lane_s32(vksum4567, vget_low_s32(vinput_zero_point23), 0);
90 const int32x4_t vksumzp3x4567 = vmulq_lane_s32(vksum4567, vget_high_s32(vinput_zero_point23), 0);
91 const int32x4_t vksumzp2x89AB = vmulq_lane_s32(vksum89AB, vget_low_s32(vinput_zero_point23), 0);
92 const int32x4_t vksumzp3x89AB = vmulq_lane_s32(vksum89AB, vget_high_s32(vinput_zero_point23), 0);
93 const int32x4_t vksumzp2xCDEF = vmulq_lane_s32(vksumCDEF, vget_low_s32(vinput_zero_point23), 0);
94 const int32x4_t vksumzp3xCDEF = vmulq_lane_s32(vksumCDEF, vget_high_s32(vinput_zero_point23), 0);
95 const int32x4_t vinput_zero_point45 = vld1q_dup_s32(&quantization_params[4].zero_point);
96 const int32x4_t vksumzp4x0123 = vmulq_s32(vksum0123, vinput_zero_point45);
97 const int32x4_t vksumzp4x4567 = vmulq_s32(vksum4567, vinput_zero_point45);
98 const int32x4_t vksumzp4x89AB = vmulq_s32(vksum89AB, vinput_zero_point45);
99 const int32x4_t vksumzp4xCDEF = vmulq_s32(vksumCDEF, vinput_zero_point45);
100
101 int32x4_t vacc01x01 = vreinterpretq_s32_u64(vtrn1q_u64(vreinterpretq_u64_s32(vksumzp0x0123), vreinterpretq_u64_s32(vksumzp1x0123)));
102 int32x4_t vacc01x23 = vreinterpretq_s32_u64(vtrn2q_u64(vreinterpretq_u64_s32(vksumzp0x0123), vreinterpretq_u64_s32(vksumzp1x0123)));
103 int32x4_t vacc01x45 = vreinterpretq_s32_u64(vtrn1q_u64(vreinterpretq_u64_s32(vksumzp0x4567), vreinterpretq_u64_s32(vksumzp1x4567)));
104 int32x4_t vacc01x67 = vreinterpretq_s32_u64(vtrn2q_u64(vreinterpretq_u64_s32(vksumzp0x4567), vreinterpretq_u64_s32(vksumzp1x4567)));
105 int32x4_t vacc01x89 = vreinterpretq_s32_u64(vtrn1q_u64(vreinterpretq_u64_s32(vksumzp0x89AB), vreinterpretq_u64_s32(vksumzp1x89AB)));
106 int32x4_t vacc01xAB = vreinterpretq_s32_u64(vtrn2q_u64(vreinterpretq_u64_s32(vksumzp0x89AB), vreinterpretq_u64_s32(vksumzp1x89AB)));
107 int32x4_t vacc01xCD = vreinterpretq_s32_u64(vtrn1q_u64(vreinterpretq_u64_s32(vksumzp0xCDEF), vreinterpretq_u64_s32(vksumzp1xCDEF)));
108 int32x4_t vacc01xEF = vreinterpretq_s32_u64(vtrn2q_u64(vreinterpretq_u64_s32(vksumzp0xCDEF), vreinterpretq_u64_s32(vksumzp1xCDEF)));
109 int32x4_t vacc23x01 = vreinterpretq_s32_u64(vtrn1q_u64(vreinterpretq_u64_s32(vksumzp2x0123), vreinterpretq_u64_s32(vksumzp3x0123)));
110 int32x4_t vacc23x23 = vreinterpretq_s32_u64(vtrn2q_u64(vreinterpretq_u64_s32(vksumzp2x0123), vreinterpretq_u64_s32(vksumzp3x0123)));
111 int32x4_t vacc23x45 = vreinterpretq_s32_u64(vtrn1q_u64(vreinterpretq_u64_s32(vksumzp2x4567), vreinterpretq_u64_s32(vksumzp3x4567)));
112 int32x4_t vacc23x67 = vreinterpretq_s32_u64(vtrn2q_u64(vreinterpretq_u64_s32(vksumzp2x4567), vreinterpretq_u64_s32(vksumzp3x4567)));
113 int32x4_t vacc23x89 = vreinterpretq_s32_u64(vtrn1q_u64(vreinterpretq_u64_s32(vksumzp2x89AB), vreinterpretq_u64_s32(vksumzp3x89AB)));
114 int32x4_t vacc23xAB = vreinterpretq_s32_u64(vtrn2q_u64(vreinterpretq_u64_s32(vksumzp2x89AB), vreinterpretq_u64_s32(vksumzp3x89AB)));
115 int32x4_t vacc23xCD = vreinterpretq_s32_u64(vtrn1q_u64(vreinterpretq_u64_s32(vksumzp2xCDEF), vreinterpretq_u64_s32(vksumzp3xCDEF)));
116 int32x4_t vacc23xEF = vreinterpretq_s32_u64(vtrn2q_u64(vreinterpretq_u64_s32(vksumzp2xCDEF), vreinterpretq_u64_s32(vksumzp3xCDEF)));
117 int32x4_t vacc45x01 = vreinterpretq_s32_u64(vtrn1q_u64(vreinterpretq_u64_s32(vksumzp4x0123), vreinterpretq_u64_s32(vksumzp4x0123)));
118 int32x4_t vacc45x23 = vreinterpretq_s32_u64(vtrn2q_u64(vreinterpretq_u64_s32(vksumzp4x0123), vreinterpretq_u64_s32(vksumzp4x0123)));
119 int32x4_t vacc45x45 = vreinterpretq_s32_u64(vtrn1q_u64(vreinterpretq_u64_s32(vksumzp4x4567), vreinterpretq_u64_s32(vksumzp4x4567)));
120 int32x4_t vacc45x67 = vreinterpretq_s32_u64(vtrn2q_u64(vreinterpretq_u64_s32(vksumzp4x4567), vreinterpretq_u64_s32(vksumzp4x4567)));
121 int32x4_t vacc45x89 = vreinterpretq_s32_u64(vtrn1q_u64(vreinterpretq_u64_s32(vksumzp4x89AB), vreinterpretq_u64_s32(vksumzp4x89AB)));
122 int32x4_t vacc45xAB = vreinterpretq_s32_u64(vtrn2q_u64(vreinterpretq_u64_s32(vksumzp4x89AB), vreinterpretq_u64_s32(vksumzp4x89AB)));
123 int32x4_t vacc45xCD = vreinterpretq_s32_u64(vtrn1q_u64(vreinterpretq_u64_s32(vksumzp4xCDEF), vreinterpretq_u64_s32(vksumzp4xCDEF)));
124 int32x4_t vacc45xEF = vreinterpretq_s32_u64(vtrn2q_u64(vreinterpretq_u64_s32(vksumzp4xCDEF), vreinterpretq_u64_s32(vksumzp4xCDEF)));
125
126 // Inner accumulation loop along the 16 columns.
127 size_t k = kc;
128 // 2x partial unrolled loop to load 8 bytes at a time.
129
130 uint64x2x2_t va01x0123456789ABCDEF;
131 va01x0123456789ABCDEF.val[0] = vdupq_n_u64(0);
132 va01x0123456789ABCDEF.val[1] = vdupq_n_u64(0);
133 uint64x2x2_t va23x0123456789ABCDEF;
134 va23x0123456789ABCDEF.val[0] = vdupq_n_u64(0);
135 va23x0123456789ABCDEF.val[1] = vdupq_n_u64(0);
136 uint64x2x2_t va45x0123456789ABCDEF;
137 va45x0123456789ABCDEF.val[0] = vdupq_n_u64(0);
138 va45x0123456789ABCDEF.val[1] = vdupq_n_u64(0);
139
140 while (k >= 16 * sizeof(int8_t)) {
141 // Load a 5x16 block of activations.
142 va01x0123456789ABCDEF = vld2q_lane_u64((const void*) a0, va01x0123456789ABCDEF, 0); a0 += 16;
143 va23x0123456789ABCDEF = vld2q_lane_u64((const void*) a2, va23x0123456789ABCDEF, 0); a2 += 16;
144 va45x0123456789ABCDEF = vld2q_lane_u64((const void*) a4, va45x0123456789ABCDEF, 0); a4 += 16;
145 va01x0123456789ABCDEF = vld2q_lane_u64((const void*) a1, va01x0123456789ABCDEF, 1); a1 += 16;
146 va23x0123456789ABCDEF = vld2q_lane_u64((const void*) a3, va23x0123456789ABCDEF, 1); a3 += 16;
147
148 const int8x16_t va01x01234567 = vreinterpretq_s8_u64(va01x0123456789ABCDEF.val[0]);
149 const int8x16_t va01x89ABCDEF = vreinterpretq_s8_u64(va01x0123456789ABCDEF.val[1]);
150 const int8x16_t va23x01234567 = vreinterpretq_s8_u64(va23x0123456789ABCDEF.val[0]);
151 const int8x16_t va23x89ABCDEF = vreinterpretq_s8_u64(va23x0123456789ABCDEF.val[1]);
152 const int8x16_t va45x01234567 = vreinterpretq_s8_u64(va45x0123456789ABCDEF.val[0]);
153 const int8x16_t va45x89ABCDEF = vreinterpretq_s8_u64(va45x0123456789ABCDEF.val[1]);
154
155 // Load a 16x16 block of weights.
156 const int8x16_t vb01x0123456789ABCDEF = vld1q_s8(w); w = (const int8_t*) w + 16;
157 const int8x16_t vb23x0123456789ABCDEF = vld1q_s8(w); w = (const int8_t*) w + 16;
158 const int8x16_t vb45x0123456789ABCDEF = vld1q_s8(w); w = (const int8_t*) w + 16;
159 const int8x16_t vb67x0123456789ABCDEF = vld1q_s8(w); w = (const int8_t*) w + 16;
160 const int8x16_t vb89x0123456789ABCDEF = vld1q_s8(w); w = (const int8_t*) w + 16;
161 const int8x16_t vbABx0123456789ABCDEF = vld1q_s8(w); w = (const int8_t*) w + 16;
162 const int8x16_t vbCDx0123456789ABCDEF = vld1q_s8(w); w = (const int8_t*) w + 16;
163 const int8x16_t vbEFx0123456789ABCDEF = vld1q_s8(w); w = (const int8_t*) w + 16;
164 const int8x16_t vb01x01234567 = vshlq_n_s8(vb01x0123456789ABCDEF, 4);
165 const int8x16_t vb23x01234567 = vshlq_n_s8(vb23x0123456789ABCDEF, 4);
166 const int8x16_t vb45x01234567 = vshlq_n_s8(vb45x0123456789ABCDEF, 4);
167 const int8x16_t vb67x01234567 = vshlq_n_s8(vb67x0123456789ABCDEF, 4);
168 const int8x16_t vb89x01234567 = vshlq_n_s8(vb89x0123456789ABCDEF, 4);
169 const int8x16_t vbABx01234567 = vshlq_n_s8(vbABx0123456789ABCDEF, 4);
170 const int8x16_t vbCDx01234567 = vshlq_n_s8(vbCDx0123456789ABCDEF, 4);
171 const int8x16_t vbEFx01234567 = vshlq_n_s8(vbEFx0123456789ABCDEF, 4);
172 const int8x16_t vb01x89ABCDEF = vandq_s8(vb01x0123456789ABCDEF, vmask);
173 const int8x16_t vb23x89ABCDEF = vandq_s8(vb23x0123456789ABCDEF, vmask);
174 const int8x16_t vb45x89ABCDEF = vandq_s8(vb45x0123456789ABCDEF, vmask);
175 const int8x16_t vb67x89ABCDEF = vandq_s8(vb67x0123456789ABCDEF, vmask);
176 const int8x16_t vb89x89ABCDEF = vandq_s8(vb89x0123456789ABCDEF, vmask);
177 const int8x16_t vbABx89ABCDEF = vandq_s8(vbABx0123456789ABCDEF, vmask);
178 const int8x16_t vbCDx89ABCDEF = vandq_s8(vbCDx0123456789ABCDEF, vmask);
179 const int8x16_t vbEFx89ABCDEF = vandq_s8(vbEFx0123456789ABCDEF, vmask);
180
181 // Multiply-accumulate: 5x8 * 8x16 --> 5x16.
182 vacc01x01 = vmmlaq_s32(vacc01x01, va01x01234567, vb01x01234567);
183 vacc01x23 = vmmlaq_s32(vacc01x23, va01x01234567, vb23x01234567);
184 vacc01x45 = vmmlaq_s32(vacc01x45, va01x01234567, vb45x01234567);
185 vacc01x67 = vmmlaq_s32(vacc01x67, va01x01234567, vb67x01234567);
186 vacc01x89 = vmmlaq_s32(vacc01x89, va01x01234567, vb89x01234567);
187 vacc01xAB = vmmlaq_s32(vacc01xAB, va01x01234567, vbABx01234567);
188 vacc01xCD = vmmlaq_s32(vacc01xCD, va01x01234567, vbCDx01234567);
189 vacc01xEF = vmmlaq_s32(vacc01xEF, va01x01234567, vbEFx01234567);
190 vacc23x01 = vmmlaq_s32(vacc23x01, va23x01234567, vb01x01234567);
191 vacc23x23 = vmmlaq_s32(vacc23x23, va23x01234567, vb23x01234567);
192 vacc23x45 = vmmlaq_s32(vacc23x45, va23x01234567, vb45x01234567);
193 vacc23x67 = vmmlaq_s32(vacc23x67, va23x01234567, vb67x01234567);
194 vacc23x89 = vmmlaq_s32(vacc23x89, va23x01234567, vb89x01234567);
195 vacc23xAB = vmmlaq_s32(vacc23xAB, va23x01234567, vbABx01234567);
196 vacc23xCD = vmmlaq_s32(vacc23xCD, va23x01234567, vbCDx01234567);
197 vacc23xEF = vmmlaq_s32(vacc23xEF, va23x01234567, vbEFx01234567);
198 vacc45x01 = vmmlaq_s32(vacc45x01, va45x01234567, vb01x01234567);
199 vacc45x23 = vmmlaq_s32(vacc45x23, va45x01234567, vb23x01234567);
200 vacc45x45 = vmmlaq_s32(vacc45x45, va45x01234567, vb45x01234567);
201 vacc45x67 = vmmlaq_s32(vacc45x67, va45x01234567, vb67x01234567);
202 vacc45x89 = vmmlaq_s32(vacc45x89, va45x01234567, vb89x01234567);
203 vacc45xAB = vmmlaq_s32(vacc45xAB, va45x01234567, vbABx01234567);
204 vacc45xCD = vmmlaq_s32(vacc45xCD, va45x01234567, vbCDx01234567);
205 vacc45xEF = vmmlaq_s32(vacc45xEF, va45x01234567, vbEFx01234567);
206 vacc01x01 = vmmlaq_s32(vacc01x01, va01x89ABCDEF, vb01x89ABCDEF);
207 vacc01x23 = vmmlaq_s32(vacc01x23, va01x89ABCDEF, vb23x89ABCDEF);
208 vacc01x45 = vmmlaq_s32(vacc01x45, va01x89ABCDEF, vb45x89ABCDEF);
209 vacc01x67 = vmmlaq_s32(vacc01x67, va01x89ABCDEF, vb67x89ABCDEF);
210 vacc01x89 = vmmlaq_s32(vacc01x89, va01x89ABCDEF, vb89x89ABCDEF);
211 vacc01xAB = vmmlaq_s32(vacc01xAB, va01x89ABCDEF, vbABx89ABCDEF);
212 vacc01xCD = vmmlaq_s32(vacc01xCD, va01x89ABCDEF, vbCDx89ABCDEF);
213 vacc01xEF = vmmlaq_s32(vacc01xEF, va01x89ABCDEF, vbEFx89ABCDEF);
214 vacc23x01 = vmmlaq_s32(vacc23x01, va23x89ABCDEF, vb01x89ABCDEF);
215 vacc23x23 = vmmlaq_s32(vacc23x23, va23x89ABCDEF, vb23x89ABCDEF);
216 vacc23x45 = vmmlaq_s32(vacc23x45, va23x89ABCDEF, vb45x89ABCDEF);
217 vacc23x67 = vmmlaq_s32(vacc23x67, va23x89ABCDEF, vb67x89ABCDEF);
218 vacc23x89 = vmmlaq_s32(vacc23x89, va23x89ABCDEF, vb89x89ABCDEF);
219 vacc23xAB = vmmlaq_s32(vacc23xAB, va23x89ABCDEF, vbABx89ABCDEF);
220 vacc23xCD = vmmlaq_s32(vacc23xCD, va23x89ABCDEF, vbCDx89ABCDEF);
221 vacc23xEF = vmmlaq_s32(vacc23xEF, va23x89ABCDEF, vbEFx89ABCDEF);
222 vacc45x01 = vmmlaq_s32(vacc45x01, va45x89ABCDEF, vb01x89ABCDEF);
223 vacc45x23 = vmmlaq_s32(vacc45x23, va45x89ABCDEF, vb23x89ABCDEF);
224 vacc45x45 = vmmlaq_s32(vacc45x45, va45x89ABCDEF, vb45x89ABCDEF);
225 vacc45x67 = vmmlaq_s32(vacc45x67, va45x89ABCDEF, vb67x89ABCDEF);
226 vacc45x89 = vmmlaq_s32(vacc45x89, va45x89ABCDEF, vb89x89ABCDEF);
227 vacc45xAB = vmmlaq_s32(vacc45xAB, va45x89ABCDEF, vbABx89ABCDEF);
228 vacc45xCD = vmmlaq_s32(vacc45xCD, va45x89ABCDEF, vbCDx89ABCDEF);
229 vacc45xEF = vmmlaq_s32(vacc45xEF, va45x89ABCDEF, vbEFx89ABCDEF);
230
231 k -= 16 * sizeof(int8_t);
232 }
233 // Handle up to 8 final positions of `k`
234 if XNN_UNLIKELY(k != 0) {
235 // Load a 5x8 block of activations.
236 uint64x2_t va01x01234567 = vld1q_dup_u64((const void*) a0); a0 += 8;
237 uint64x2_t va23x01234567 = vld1q_dup_u64((const void*) a2); a2 += 8;
238 uint64x2_t va45x01234567 = vld1q_dup_u64((const void*) a4); a4 += 8;
239 va01x01234567 = vld1q_lane_u64((const void*) a1, va01x01234567, 1); a1 += 8;
240 va23x01234567 = vld1q_lane_u64((const void*) a3, va23x01234567, 1); a3 += 8;
241
242 // Load a 16x16 block of weights.
243 const int8x16_t vb01x0123456789ABCDEF = vld1q_s8(w); w = (const int8_t*) w + 16;
244 const int8x16_t vb23x0123456789ABCDEF = vld1q_s8(w); w = (const int8_t*) w + 16;
245 const int8x16_t vb45x0123456789ABCDEF = vld1q_s8(w); w = (const int8_t*) w + 16;
246 const int8x16_t vb67x0123456789ABCDEF = vld1q_s8(w); w = (const int8_t*) w + 16;
247 const int8x16_t vb89x0123456789ABCDEF = vld1q_s8(w); w = (const int8_t*) w + 16;
248 const int8x16_t vbABx0123456789ABCDEF = vld1q_s8(w); w = (const int8_t*) w + 16;
249 const int8x16_t vbCDx0123456789ABCDEF = vld1q_s8(w); w = (const int8_t*) w + 16;
250 const int8x16_t vbEFx0123456789ABCDEF = vld1q_s8(w); w = (const int8_t*) w + 16;
251 const int8x16_t vb01x01234567 = vshlq_n_s8(vb01x0123456789ABCDEF, 4);
252 const int8x16_t vb23x01234567 = vshlq_n_s8(vb23x0123456789ABCDEF, 4);
253 const int8x16_t vb45x01234567 = vshlq_n_s8(vb45x0123456789ABCDEF, 4);
254 const int8x16_t vb67x01234567 = vshlq_n_s8(vb67x0123456789ABCDEF, 4);
255 const int8x16_t vb89x01234567 = vshlq_n_s8(vb89x0123456789ABCDEF, 4);
256 const int8x16_t vbABx01234567 = vshlq_n_s8(vbABx0123456789ABCDEF, 4);
257 const int8x16_t vbCDx01234567 = vshlq_n_s8(vbCDx0123456789ABCDEF, 4);
258 const int8x16_t vbEFx01234567 = vshlq_n_s8(vbEFx0123456789ABCDEF, 4);
259
260 // Multiply-accumulate: 5x4 * 4x16 --> 5x16.
261 vacc01x01 = vmmlaq_s32(vacc01x01, vreinterpretq_s8_u64(va01x01234567), vb01x01234567);
262 vacc01x23 = vmmlaq_s32(vacc01x23, vreinterpretq_s8_u64(va01x01234567), vb23x01234567);
263 vacc01x45 = vmmlaq_s32(vacc01x45, vreinterpretq_s8_u64(va01x01234567), vb45x01234567);
264 vacc01x67 = vmmlaq_s32(vacc01x67, vreinterpretq_s8_u64(va01x01234567), vb67x01234567);
265 vacc01x89 = vmmlaq_s32(vacc01x89, vreinterpretq_s8_u64(va01x01234567), vb89x01234567);
266 vacc01xAB = vmmlaq_s32(vacc01xAB, vreinterpretq_s8_u64(va01x01234567), vbABx01234567);
267 vacc01xCD = vmmlaq_s32(vacc01xCD, vreinterpretq_s8_u64(va01x01234567), vbCDx01234567);
268 vacc01xEF = vmmlaq_s32(vacc01xEF, vreinterpretq_s8_u64(va01x01234567), vbEFx01234567);
269 vacc23x01 = vmmlaq_s32(vacc23x01, vreinterpretq_s8_u64(va23x01234567), vb01x01234567);
270 vacc23x23 = vmmlaq_s32(vacc23x23, vreinterpretq_s8_u64(va23x01234567), vb23x01234567);
271 vacc23x45 = vmmlaq_s32(vacc23x45, vreinterpretq_s8_u64(va23x01234567), vb45x01234567);
272 vacc23x67 = vmmlaq_s32(vacc23x67, vreinterpretq_s8_u64(va23x01234567), vb67x01234567);
273 vacc23x89 = vmmlaq_s32(vacc23x89, vreinterpretq_s8_u64(va23x01234567), vb89x01234567);
274 vacc23xAB = vmmlaq_s32(vacc23xAB, vreinterpretq_s8_u64(va23x01234567), vbABx01234567);
275 vacc23xCD = vmmlaq_s32(vacc23xCD, vreinterpretq_s8_u64(va23x01234567), vbCDx01234567);
276 vacc23xEF = vmmlaq_s32(vacc23xEF, vreinterpretq_s8_u64(va23x01234567), vbEFx01234567);
277 vacc45x01 = vmmlaq_s32(vacc45x01, vreinterpretq_s8_u64(va45x01234567), vb01x01234567);
278 vacc45x23 = vmmlaq_s32(vacc45x23, vreinterpretq_s8_u64(va45x01234567), vb23x01234567);
279 vacc45x45 = vmmlaq_s32(vacc45x45, vreinterpretq_s8_u64(va45x01234567), vb45x01234567);
280 vacc45x67 = vmmlaq_s32(vacc45x67, vreinterpretq_s8_u64(va45x01234567), vb67x01234567);
281 vacc45x89 = vmmlaq_s32(vacc45x89, vreinterpretq_s8_u64(va45x01234567), vb89x01234567);
282 vacc45xAB = vmmlaq_s32(vacc45xAB, vreinterpretq_s8_u64(va45x01234567), vbABx01234567);
283 vacc45xCD = vmmlaq_s32(vacc45xCD, vreinterpretq_s8_u64(va45x01234567), vbCDx01234567);
284 vacc45xEF = vmmlaq_s32(vacc45xEF, vreinterpretq_s8_u64(va45x01234567), vbEFx01234567);
285 }
286
287 int32x4_t vacc0x0123 = vreinterpretq_s32_u64(vtrn1q_u64(vreinterpretq_u64_s32(vacc01x01), vreinterpretq_u64_s32(vacc01x23)));
288 int32x4_t vacc1x0123 = vreinterpretq_s32_u64(vtrn2q_u64(vreinterpretq_u64_s32(vacc01x01), vreinterpretq_u64_s32(vacc01x23)));
289 int32x4_t vacc0x4567 = vreinterpretq_s32_u64(vtrn1q_u64(vreinterpretq_u64_s32(vacc01x45), vreinterpretq_u64_s32(vacc01x67)));
290 int32x4_t vacc1x4567 = vreinterpretq_s32_u64(vtrn2q_u64(vreinterpretq_u64_s32(vacc01x45), vreinterpretq_u64_s32(vacc01x67)));
291 int32x4_t vacc0x89AB = vreinterpretq_s32_u64(vtrn1q_u64(vreinterpretq_u64_s32(vacc01x89), vreinterpretq_u64_s32(vacc01xAB)));
292 int32x4_t vacc1x89AB = vreinterpretq_s32_u64(vtrn2q_u64(vreinterpretq_u64_s32(vacc01x89), vreinterpretq_u64_s32(vacc01xAB)));
293 int32x4_t vacc0xCDEF = vreinterpretq_s32_u64(vtrn1q_u64(vreinterpretq_u64_s32(vacc01xCD), vreinterpretq_u64_s32(vacc01xEF)));
294 int32x4_t vacc1xCDEF = vreinterpretq_s32_u64(vtrn2q_u64(vreinterpretq_u64_s32(vacc01xCD), vreinterpretq_u64_s32(vacc01xEF)));
295 int32x4_t vacc2x0123 = vreinterpretq_s32_u64(vtrn1q_u64(vreinterpretq_u64_s32(vacc23x01), vreinterpretq_u64_s32(vacc23x23)));
296 int32x4_t vacc3x0123 = vreinterpretq_s32_u64(vtrn2q_u64(vreinterpretq_u64_s32(vacc23x01), vreinterpretq_u64_s32(vacc23x23)));
297 int32x4_t vacc2x4567 = vreinterpretq_s32_u64(vtrn1q_u64(vreinterpretq_u64_s32(vacc23x45), vreinterpretq_u64_s32(vacc23x67)));
298 int32x4_t vacc3x4567 = vreinterpretq_s32_u64(vtrn2q_u64(vreinterpretq_u64_s32(vacc23x45), vreinterpretq_u64_s32(vacc23x67)));
299 int32x4_t vacc2x89AB = vreinterpretq_s32_u64(vtrn1q_u64(vreinterpretq_u64_s32(vacc23x89), vreinterpretq_u64_s32(vacc23xAB)));
300 int32x4_t vacc3x89AB = vreinterpretq_s32_u64(vtrn2q_u64(vreinterpretq_u64_s32(vacc23x89), vreinterpretq_u64_s32(vacc23xAB)));
301 int32x4_t vacc2xCDEF = vreinterpretq_s32_u64(vtrn1q_u64(vreinterpretq_u64_s32(vacc23xCD), vreinterpretq_u64_s32(vacc23xEF)));
302 int32x4_t vacc3xCDEF = vreinterpretq_s32_u64(vtrn2q_u64(vreinterpretq_u64_s32(vacc23xCD), vreinterpretq_u64_s32(vacc23xEF)));
303 int32x4_t vacc4x0123 = vreinterpretq_s32_u64(vtrn1q_u64(vreinterpretq_u64_s32(vacc45x01), vreinterpretq_u64_s32(vacc45x23)));
304 int32x4_t vacc4x4567 = vreinterpretq_s32_u64(vtrn1q_u64(vreinterpretq_u64_s32(vacc45x45), vreinterpretq_u64_s32(vacc45x67)));
305 int32x4_t vacc4x89AB = vreinterpretq_s32_u64(vtrn1q_u64(vreinterpretq_u64_s32(vacc45x89), vreinterpretq_u64_s32(vacc45xAB)));
306 int32x4_t vacc4xCDEF = vreinterpretq_s32_u64(vtrn1q_u64(vreinterpretq_u64_s32(vacc45xCD), vreinterpretq_u64_s32(vacc45xEF)));
307 float32x4_t vout0x0123 = vcvtq_n_f32_s32(vacc0x0123, 4);
308 float32x4_t vout0x4567 = vcvtq_n_f32_s32(vacc0x4567, 4);
309 float32x4_t vout0x89AB = vcvtq_n_f32_s32(vacc0x89AB, 4);
310 float32x4_t vout0xCDEF = vcvtq_n_f32_s32(vacc0xCDEF, 4);
311 float32x4_t vout1x0123 = vcvtq_n_f32_s32(vacc1x0123, 4);
312 float32x4_t vout1x4567 = vcvtq_n_f32_s32(vacc1x4567, 4);
313 float32x4_t vout1x89AB = vcvtq_n_f32_s32(vacc1x89AB, 4);
314 float32x4_t vout1xCDEF = vcvtq_n_f32_s32(vacc1xCDEF, 4);
315 float32x4_t vout2x0123 = vcvtq_n_f32_s32(vacc2x0123, 4);
316 float32x4_t vout2x4567 = vcvtq_n_f32_s32(vacc2x4567, 4);
317 float32x4_t vout2x89AB = vcvtq_n_f32_s32(vacc2x89AB, 4);
318 float32x4_t vout2xCDEF = vcvtq_n_f32_s32(vacc2xCDEF, 4);
319 float32x4_t vout3x0123 = vcvtq_n_f32_s32(vacc3x0123, 4);
320 float32x4_t vout3x4567 = vcvtq_n_f32_s32(vacc3x4567, 4);
321 float32x4_t vout3x89AB = vcvtq_n_f32_s32(vacc3x89AB, 4);
322 float32x4_t vout3xCDEF = vcvtq_n_f32_s32(vacc3xCDEF, 4);
323 float32x4_t vout4x0123 = vcvtq_n_f32_s32(vacc4x0123, 4);
324 float32x4_t vout4x4567 = vcvtq_n_f32_s32(vacc4x4567, 4);
325 float32x4_t vout4x89AB = vcvtq_n_f32_s32(vacc4x89AB, 4);
326 float32x4_t vout4xCDEF = vcvtq_n_f32_s32(vacc4xCDEF, 4);
327
328 const float32x4_t vinput_scale01 = vreinterpretq_f32_s32(vld1q_s32(&quantization_params[0].zero_point));
329 vout0x0123 = vmulq_lane_f32(vout0x0123, vget_low_f32(vinput_scale01), 1);
330 vout1x0123 = vmulq_lane_f32(vout1x0123, vget_high_f32(vinput_scale01), 1);
331 vout0x4567 = vmulq_lane_f32(vout0x4567, vget_low_f32(vinput_scale01), 1);
332 vout1x4567 = vmulq_lane_f32(vout1x4567, vget_high_f32(vinput_scale01), 1);
333 vout0x89AB = vmulq_lane_f32(vout0x89AB, vget_low_f32(vinput_scale01), 1);
334 vout1x89AB = vmulq_lane_f32(vout1x89AB, vget_high_f32(vinput_scale01), 1);
335 vout0xCDEF = vmulq_lane_f32(vout0xCDEF, vget_low_f32(vinput_scale01), 1);
336 vout1xCDEF = vmulq_lane_f32(vout1xCDEF, vget_high_f32(vinput_scale01), 1);
337 const float32x4_t vinput_scale23 = vreinterpretq_f32_s32(vld1q_s32(&quantization_params[2].zero_point));
338 vout2x0123 = vmulq_lane_f32(vout2x0123, vget_low_f32(vinput_scale23), 1);
339 vout3x0123 = vmulq_lane_f32(vout3x0123, vget_high_f32(vinput_scale23), 1);
340 vout2x4567 = vmulq_lane_f32(vout2x4567, vget_low_f32(vinput_scale23), 1);
341 vout3x4567 = vmulq_lane_f32(vout3x4567, vget_high_f32(vinput_scale23), 1);
342 vout2x89AB = vmulq_lane_f32(vout2x89AB, vget_low_f32(vinput_scale23), 1);
343 vout3x89AB = vmulq_lane_f32(vout3x89AB, vget_high_f32(vinput_scale23), 1);
344 vout2xCDEF = vmulq_lane_f32(vout2xCDEF, vget_low_f32(vinput_scale23), 1);
345 vout3xCDEF = vmulq_lane_f32(vout3xCDEF, vget_high_f32(vinput_scale23), 1);
346 const float32x4_t vinput_scale4 = vld1q_dup_f32(&quantization_params[4].inv_scale);
347 vout4x0123 = vmulq_f32(vout4x0123, vinput_scale4);
348 vout4x4567 = vmulq_f32(vout4x4567, vinput_scale4);
349 vout4x89AB = vmulq_f32(vout4x89AB, vinput_scale4);
350 vout4xCDEF = vmulq_f32(vout4xCDEF, vinput_scale4);
351
352 const float32x4_t vfilter_output_scale0123 = vld1q_f32(w); w = (const float*) w + 4;
353 const float32x4_t vfilter_output_scale4567 = vld1q_f32(w); w = (const float*) w + 4;
354 const float32x4_t vfilter_output_scale89AB = vld1q_f32(w); w = (const float*) w + 4;
355 const float32x4_t vfilter_output_scaleCDEF = vld1q_f32(w); w = (const float*) w + 4;
356
357 const float32x4_t vbias0123 = vld1q_f32(w); w = (const float*) w + 4;
358 vout0x0123 = vfmaq_f32(vbias0123, vout0x0123, vfilter_output_scale0123);
359 vout1x0123 = vfmaq_f32(vbias0123, vout1x0123, vfilter_output_scale0123);
360 vout2x0123 = vfmaq_f32(vbias0123, vout2x0123, vfilter_output_scale0123);
361 vout3x0123 = vfmaq_f32(vbias0123, vout3x0123, vfilter_output_scale0123);
362 vout4x0123 = vfmaq_f32(vbias0123, vout4x0123, vfilter_output_scale0123);
363 const float32x4_t vbias4567 = vld1q_f32(w); w = (const float*) w + 4;
364 vout0x4567 = vfmaq_f32(vbias4567, vout0x4567, vfilter_output_scale4567);
365 vout1x4567 = vfmaq_f32(vbias4567, vout1x4567, vfilter_output_scale4567);
366 vout2x4567 = vfmaq_f32(vbias4567, vout2x4567, vfilter_output_scale4567);
367 vout3x4567 = vfmaq_f32(vbias4567, vout3x4567, vfilter_output_scale4567);
368 vout4x4567 = vfmaq_f32(vbias4567, vout4x4567, vfilter_output_scale4567);
369 const float32x4_t vbias89AB = vld1q_f32(w); w = (const float*) w + 4;
370 vout0x89AB = vfmaq_f32(vbias89AB, vout0x89AB, vfilter_output_scale89AB);
371 vout1x89AB = vfmaq_f32(vbias89AB, vout1x89AB, vfilter_output_scale89AB);
372 vout2x89AB = vfmaq_f32(vbias89AB, vout2x89AB, vfilter_output_scale89AB);
373 vout3x89AB = vfmaq_f32(vbias89AB, vout3x89AB, vfilter_output_scale89AB);
374 vout4x89AB = vfmaq_f32(vbias89AB, vout4x89AB, vfilter_output_scale89AB);
375 const float32x4_t vbiasCDEF = vld1q_f32(w); w = (const float*) w + 4;
376 vout0xCDEF = vfmaq_f32(vbiasCDEF, vout0xCDEF, vfilter_output_scaleCDEF);
377 vout1xCDEF = vfmaq_f32(vbiasCDEF, vout1xCDEF, vfilter_output_scaleCDEF);
378 vout2xCDEF = vfmaq_f32(vbiasCDEF, vout2xCDEF, vfilter_output_scaleCDEF);
379 vout3xCDEF = vfmaq_f32(vbiasCDEF, vout3xCDEF, vfilter_output_scaleCDEF);
380 vout4xCDEF = vfmaq_f32(vbiasCDEF, vout4xCDEF, vfilter_output_scaleCDEF);
381
382 float16x8_t vfp16out0x01234567 = vcombine_f16(vcvt_f16_f32(vout0x0123), vcvt_f16_f32(vout0x4567));
383 float16x8_t vfp16out0x89ABCDEF = vcombine_f16(vcvt_f16_f32(vout0x89AB), vcvt_f16_f32(vout0xCDEF));
384 float16x8_t vfp16out1x01234567 = vcombine_f16(vcvt_f16_f32(vout1x0123), vcvt_f16_f32(vout1x4567));
385 float16x8_t vfp16out1x89ABCDEF = vcombine_f16(vcvt_f16_f32(vout1x89AB), vcvt_f16_f32(vout1xCDEF));
386 float16x8_t vfp16out2x01234567 = vcombine_f16(vcvt_f16_f32(vout2x0123), vcvt_f16_f32(vout2x4567));
387 float16x8_t vfp16out2x89ABCDEF = vcombine_f16(vcvt_f16_f32(vout2x89AB), vcvt_f16_f32(vout2xCDEF));
388 float16x8_t vfp16out3x01234567 = vcombine_f16(vcvt_f16_f32(vout3x0123), vcvt_f16_f32(vout3x4567));
389 float16x8_t vfp16out3x89ABCDEF = vcombine_f16(vcvt_f16_f32(vout3x89AB), vcvt_f16_f32(vout3xCDEF));
390 float16x8_t vfp16out4x01234567 = vcombine_f16(vcvt_f16_f32(vout4x0123), vcvt_f16_f32(vout4x4567));
391 float16x8_t vfp16out4x89ABCDEF = vcombine_f16(vcvt_f16_f32(vout4x89AB), vcvt_f16_f32(vout4xCDEF));
392
393 const float16x8_t voutput_min = vreinterpretq_f16_u16(vld1q_dup_u16((const uint16_t*) &params->scalar.min));
394 vfp16out0x01234567 = vmaxq_f16(vfp16out0x01234567, voutput_min);
395 vfp16out0x89ABCDEF = vmaxq_f16(vfp16out0x89ABCDEF, voutput_min);
396 vfp16out1x01234567 = vmaxq_f16(vfp16out1x01234567, voutput_min);
397 vfp16out1x89ABCDEF = vmaxq_f16(vfp16out1x89ABCDEF, voutput_min);
398 vfp16out2x01234567 = vmaxq_f16(vfp16out2x01234567, voutput_min);
399 vfp16out2x89ABCDEF = vmaxq_f16(vfp16out2x89ABCDEF, voutput_min);
400 vfp16out3x01234567 = vmaxq_f16(vfp16out3x01234567, voutput_min);
401 vfp16out3x89ABCDEF = vmaxq_f16(vfp16out3x89ABCDEF, voutput_min);
402 vfp16out4x01234567 = vmaxq_f16(vfp16out4x01234567, voutput_min);
403 vfp16out4x89ABCDEF = vmaxq_f16(vfp16out4x89ABCDEF, voutput_min);
404 const float16x8_t voutput_max = vreinterpretq_f16_u16(vld1q_dup_u16((const uint16_t*) &params->scalar.max));
405 vfp16out0x01234567 = vminq_f16(vfp16out0x01234567, voutput_max);
406 vfp16out0x89ABCDEF = vminq_f16(vfp16out0x89ABCDEF, voutput_max);
407 vfp16out1x01234567 = vminq_f16(vfp16out1x01234567, voutput_max);
408 vfp16out1x89ABCDEF = vminq_f16(vfp16out1x89ABCDEF, voutput_max);
409 vfp16out2x01234567 = vminq_f16(vfp16out2x01234567, voutput_max);
410 vfp16out2x89ABCDEF = vminq_f16(vfp16out2x89ABCDEF, voutput_max);
411 vfp16out3x01234567 = vminq_f16(vfp16out3x01234567, voutput_max);
412 vfp16out3x89ABCDEF = vminq_f16(vfp16out3x89ABCDEF, voutput_max);
413 vfp16out4x01234567 = vminq_f16(vfp16out4x01234567, voutput_max);
414 vfp16out4x89ABCDEF = vminq_f16(vfp16out4x89ABCDEF, voutput_max);
415 if XNN_LIKELY(nc >= 16) {
416 vst1q_u16(c0, vreinterpretq_u16_f16(vfp16out0x01234567));
417 vst1q_u16(c0 + 8, vreinterpretq_u16_f16(vfp16out0x89ABCDEF));
418 vst1q_u16(c1, vreinterpretq_u16_f16(vfp16out1x01234567));
419 vst1q_u16(c1 + 8, vreinterpretq_u16_f16(vfp16out1x89ABCDEF));
420 vst1q_u16(c2, vreinterpretq_u16_f16(vfp16out2x01234567));
421 vst1q_u16(c2 + 8, vreinterpretq_u16_f16(vfp16out2x89ABCDEF));
422 vst1q_u16(c3, vreinterpretq_u16_f16(vfp16out3x01234567));
423 vst1q_u16(c3 + 8, vreinterpretq_u16_f16(vfp16out3x89ABCDEF));
424 vst1q_u16(c4, vreinterpretq_u16_f16(vfp16out4x01234567));
425 vst1q_u16(c4 + 8, vreinterpretq_u16_f16(vfp16out4x89ABCDEF));
426
427 a0 = (const int8_t*) ((uintptr_t) a0 - kc);
428 a1 = (const int8_t*) ((uintptr_t) a1 - kc);
429 a2 = (const int8_t*) ((uintptr_t) a2 - kc);
430 a3 = (const int8_t*) ((uintptr_t) a3 - kc);
431 a4 = (const int8_t*) ((uintptr_t) a4 - kc);
432
433 c0 = (uint16_t*) ((uintptr_t) c0 + cn_stride);
434 c1 = (uint16_t*) ((uintptr_t) c1 + cn_stride);
435 c2 = (uint16_t*) ((uintptr_t) c2 + cn_stride);
436 c3 = (uint16_t*) ((uintptr_t) c3 + cn_stride);
437 c4 = (uint16_t*) ((uintptr_t) c4 + cn_stride);
438
439 nc -= 16;
440 } else {
441 if (nc & 8) {
442 vst1q_u16(c0, vreinterpretq_u16_f16(vfp16out0x01234567)); c0 += 8;
443 vfp16out0x01234567 = vfp16out0x89ABCDEF;
444 vst1q_u16(c1, vreinterpretq_u16_f16(vfp16out1x01234567)); c1 += 8;
445 vfp16out1x01234567 = vfp16out1x89ABCDEF;
446 vst1q_u16(c2, vreinterpretq_u16_f16(vfp16out2x01234567)); c2 += 8;
447 vfp16out2x01234567 = vfp16out2x89ABCDEF;
448 vst1q_u16(c3, vreinterpretq_u16_f16(vfp16out3x01234567)); c3 += 8;
449 vfp16out3x01234567 = vfp16out3x89ABCDEF;
450 vst1q_u16(c4, vreinterpretq_u16_f16(vfp16out4x01234567)); c4 += 8;
451 vfp16out4x01234567 = vfp16out4x89ABCDEF;
452 }
453 float16x4_t vfp16out0x0123 = vget_low_f16(vfp16out0x01234567);
454 float16x4_t vfp16out1x0123 = vget_low_f16(vfp16out1x01234567);
455 float16x4_t vfp16out2x0123 = vget_low_f16(vfp16out2x01234567);
456 float16x4_t vfp16out3x0123 = vget_low_f16(vfp16out3x01234567);
457 float16x4_t vfp16out4x0123 = vget_low_f16(vfp16out4x01234567);
458 if (nc & 4) {
459 vst1_u16(c0, vreinterpret_u16_f16(vfp16out0x0123)); c0 += 4;
460 vst1_u16(c1, vreinterpret_u16_f16(vfp16out1x0123)); c1 += 4;
461 vst1_u16(c2, vreinterpret_u16_f16(vfp16out2x0123)); c2 += 4;
462 vst1_u16(c3, vreinterpret_u16_f16(vfp16out3x0123)); c3 += 4;
463 vst1_u16(c4, vreinterpret_u16_f16(vfp16out4x0123)); c4 += 4;
464 vfp16out0x0123 = vget_high_f16(vfp16out0x01234567);
465 vfp16out1x0123 = vget_high_f16(vfp16out1x01234567);
466 vfp16out2x0123 = vget_high_f16(vfp16out2x01234567);
467 vfp16out3x0123 = vget_high_f16(vfp16out3x01234567);
468 vfp16out4x0123 = vget_high_f16(vfp16out4x01234567);
469 }
470 if (nc & 2) {
471 vst1_lane_u32((void*) c0, vreinterpret_u32_f16(vfp16out0x0123), 0); c0 += 2;
472 vst1_lane_u32((void*) c1, vreinterpret_u32_f16(vfp16out1x0123), 0); c1 += 2;
473 vst1_lane_u32((void*) c2, vreinterpret_u32_f16(vfp16out2x0123), 0); c2 += 2;
474 vst1_lane_u32((void*) c3, vreinterpret_u32_f16(vfp16out3x0123), 0); c3 += 2;
475 vst1_lane_u32((void*) c4, vreinterpret_u32_f16(vfp16out4x0123), 0); c4 += 2;
476 vfp16out0x0123 = vext_f16(vfp16out0x0123, vfp16out0x0123, 2);
477 vfp16out1x0123 = vext_f16(vfp16out1x0123, vfp16out1x0123, 2);
478 vfp16out2x0123 = vext_f16(vfp16out2x0123, vfp16out2x0123, 2);
479 vfp16out3x0123 = vext_f16(vfp16out3x0123, vfp16out3x0123, 2);
480 vfp16out4x0123 = vext_f16(vfp16out4x0123, vfp16out4x0123, 2);
481 }
482 if (nc & 1) {
483 vst1_lane_u16(c0, vreinterpret_u16_f16(vfp16out0x0123), 0);
484 vst1_lane_u16(c1, vreinterpret_u16_f16(vfp16out1x0123), 0);
485 vst1_lane_u16(c2, vreinterpret_u16_f16(vfp16out2x0123), 0);
486 vst1_lane_u16(c3, vreinterpret_u16_f16(vfp16out3x0123), 0);
487 vst1_lane_u16(c4, vreinterpret_u16_f16(vfp16out4x0123), 0);
488 }
489 nc = 0;
490 }
491 } while (nc != 0);
492 }