Debian Codegraph @

src/qd8-f16-qc4w-gemm/gen/qd8-f16-qc4w-gemm-5x8c8-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_5x8c8__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 8 columns.
70 do {
71 // Initialize accumulators with bias. 8 bias values are loaded from the
72 // weight matrix, at the start of the group of 8 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 vinput_zero_point23 = vld1q_s32(&quantization_params[2].zero_point);
81 const int32x4_t vksumzp2x0123 = vmulq_lane_s32(vksum0123, vget_low_s32(vinput_zero_point23), 0);
82 const int32x4_t vksumzp3x0123 = vmulq_lane_s32(vksum0123, vget_high_s32(vinput_zero_point23), 0);
83 const int32x4_t vksumzp2x4567 = vmulq_lane_s32(vksum4567, vget_low_s32(vinput_zero_point23), 0);
84 const int32x4_t vksumzp3x4567 = vmulq_lane_s32(vksum4567, vget_high_s32(vinput_zero_point23), 0);
85 const int32x4_t vinput_zero_point45 = vld1q_dup_s32(&quantization_params[4].zero_point);
86 const int32x4_t vksumzp4x0123 = vmulq_s32(vksum0123, vinput_zero_point45);
87 const int32x4_t vksumzp4x4567 = vmulq_s32(vksum4567, vinput_zero_point45);
88
89 int32x4_t vacc01x01 = vreinterpretq_s32_u64(vtrn1q_u64(vreinterpretq_u64_s32(vksumzp0x0123), vreinterpretq_u64_s32(vksumzp1x0123)));
90 int32x4_t vacc01x23 = vreinterpretq_s32_u64(vtrn2q_u64(vreinterpretq_u64_s32(vksumzp0x0123), vreinterpretq_u64_s32(vksumzp1x0123)));
91 int32x4_t vacc01x45 = vreinterpretq_s32_u64(vtrn1q_u64(vreinterpretq_u64_s32(vksumzp0x4567), vreinterpretq_u64_s32(vksumzp1x4567)));
92 int32x4_t vacc01x67 = vreinterpretq_s32_u64(vtrn2q_u64(vreinterpretq_u64_s32(vksumzp0x4567), vreinterpretq_u64_s32(vksumzp1x4567)));
93 int32x4_t vacc23x01 = vreinterpretq_s32_u64(vtrn1q_u64(vreinterpretq_u64_s32(vksumzp2x0123), vreinterpretq_u64_s32(vksumzp3x0123)));
94 int32x4_t vacc23x23 = vreinterpretq_s32_u64(vtrn2q_u64(vreinterpretq_u64_s32(vksumzp2x0123), vreinterpretq_u64_s32(vksumzp3x0123)));
95 int32x4_t vacc23x45 = vreinterpretq_s32_u64(vtrn1q_u64(vreinterpretq_u64_s32(vksumzp2x4567), vreinterpretq_u64_s32(vksumzp3x4567)));
96 int32x4_t vacc23x67 = vreinterpretq_s32_u64(vtrn2q_u64(vreinterpretq_u64_s32(vksumzp2x4567), vreinterpretq_u64_s32(vksumzp3x4567)));
97 int32x4_t vacc45x01 = vreinterpretq_s32_u64(vtrn1q_u64(vreinterpretq_u64_s32(vksumzp4x0123), vreinterpretq_u64_s32(vksumzp4x0123)));
98 int32x4_t vacc45x23 = vreinterpretq_s32_u64(vtrn2q_u64(vreinterpretq_u64_s32(vksumzp4x0123), vreinterpretq_u64_s32(vksumzp4x0123)));
99 int32x4_t vacc45x45 = vreinterpretq_s32_u64(vtrn1q_u64(vreinterpretq_u64_s32(vksumzp4x4567), vreinterpretq_u64_s32(vksumzp4x4567)));
100 int32x4_t vacc45x67 = vreinterpretq_s32_u64(vtrn2q_u64(vreinterpretq_u64_s32(vksumzp4x4567), vreinterpretq_u64_s32(vksumzp4x4567)));
101
102 // Inner accumulation loop along the 8 columns.
103 size_t k = kc;
104 // 2x partial unrolled loop to load 8 bytes at a time.
105
106 uint64x2x2_t va01x0123456789ABCDEF;
107 va01x0123456789ABCDEF.val[0] = vdupq_n_u64(0);
108 va01x0123456789ABCDEF.val[1] = vdupq_n_u64(0);
109 uint64x2x2_t va23x0123456789ABCDEF;
110 va23x0123456789ABCDEF.val[0] = vdupq_n_u64(0);
111 va23x0123456789ABCDEF.val[1] = vdupq_n_u64(0);
112 uint64x2x2_t va45x0123456789ABCDEF;
113 va45x0123456789ABCDEF.val[0] = vdupq_n_u64(0);
114 va45x0123456789ABCDEF.val[1] = vdupq_n_u64(0);
115
116 while (k >= 16 * sizeof(int8_t)) {
117 // Load a 5x16 block of activations.
118 va01x0123456789ABCDEF = vld2q_lane_u64((const void*) a0, va01x0123456789ABCDEF, 0); a0 += 16;
119 va23x0123456789ABCDEF = vld2q_lane_u64((const void*) a2, va23x0123456789ABCDEF, 0); a2 += 16;
120 va45x0123456789ABCDEF = vld2q_lane_u64((const void*) a4, va45x0123456789ABCDEF, 0); a4 += 16;
121 va01x0123456789ABCDEF = vld2q_lane_u64((const void*) a1, va01x0123456789ABCDEF, 1); a1 += 16;
122 va23x0123456789ABCDEF = vld2q_lane_u64((const void*) a3, va23x0123456789ABCDEF, 1); a3 += 16;
123
124 const int8x16_t va01x01234567 = vreinterpretq_s8_u64(va01x0123456789ABCDEF.val[0]);
125 const int8x16_t va01x89ABCDEF = vreinterpretq_s8_u64(va01x0123456789ABCDEF.val[1]);
126 const int8x16_t va23x01234567 = vreinterpretq_s8_u64(va23x0123456789ABCDEF.val[0]);
127 const int8x16_t va23x89ABCDEF = vreinterpretq_s8_u64(va23x0123456789ABCDEF.val[1]);
128 const int8x16_t va45x01234567 = vreinterpretq_s8_u64(va45x0123456789ABCDEF.val[0]);
129 const int8x16_t va45x89ABCDEF = vreinterpretq_s8_u64(va45x0123456789ABCDEF.val[1]);
130
131 // Load a 16x8 block of weights.
132 const int8x16_t vb01x0123456789ABCDEF = vld1q_s8(w); w = (const int8_t*) w + 16;
133 const int8x16_t vb23x0123456789ABCDEF = vld1q_s8(w); w = (const int8_t*) w + 16;
134 const int8x16_t vb45x0123456789ABCDEF = vld1q_s8(w); w = (const int8_t*) w + 16;
135 const int8x16_t vb67x0123456789ABCDEF = vld1q_s8(w); w = (const int8_t*) w + 16;
136 const int8x16_t vb01x01234567 = vshlq_n_s8(vb01x0123456789ABCDEF, 4);
137 const int8x16_t vb23x01234567 = vshlq_n_s8(vb23x0123456789ABCDEF, 4);
138 const int8x16_t vb45x01234567 = vshlq_n_s8(vb45x0123456789ABCDEF, 4);
139 const int8x16_t vb67x01234567 = vshlq_n_s8(vb67x0123456789ABCDEF, 4);
140 const int8x16_t vb01x89ABCDEF = vandq_s8(vb01x0123456789ABCDEF, vmask);
141 const int8x16_t vb23x89ABCDEF = vandq_s8(vb23x0123456789ABCDEF, vmask);
142 const int8x16_t vb45x89ABCDEF = vandq_s8(vb45x0123456789ABCDEF, vmask);
143 const int8x16_t vb67x89ABCDEF = vandq_s8(vb67x0123456789ABCDEF, vmask);
144
145 // Multiply-accumulate: 5x8 * 8x8 --> 5x8.
146 vacc01x01 = vmmlaq_s32(vacc01x01, va01x01234567, vb01x01234567);
147 vacc01x23 = vmmlaq_s32(vacc01x23, va01x01234567, vb23x01234567);
148 vacc01x45 = vmmlaq_s32(vacc01x45, va01x01234567, vb45x01234567);
149 vacc01x67 = vmmlaq_s32(vacc01x67, va01x01234567, vb67x01234567);
150 vacc23x01 = vmmlaq_s32(vacc23x01, va23x01234567, vb01x01234567);
151 vacc23x23 = vmmlaq_s32(vacc23x23, va23x01234567, vb23x01234567);
152 vacc23x45 = vmmlaq_s32(vacc23x45, va23x01234567, vb45x01234567);
153 vacc23x67 = vmmlaq_s32(vacc23x67, va23x01234567, vb67x01234567);
154 vacc45x01 = vmmlaq_s32(vacc45x01, va45x01234567, vb01x01234567);
155 vacc45x23 = vmmlaq_s32(vacc45x23, va45x01234567, vb23x01234567);
156 vacc45x45 = vmmlaq_s32(vacc45x45, va45x01234567, vb45x01234567);
157 vacc45x67 = vmmlaq_s32(vacc45x67, va45x01234567, vb67x01234567);
158 vacc01x01 = vmmlaq_s32(vacc01x01, va01x89ABCDEF, vb01x89ABCDEF);
159 vacc01x23 = vmmlaq_s32(vacc01x23, va01x89ABCDEF, vb23x89ABCDEF);
160 vacc01x45 = vmmlaq_s32(vacc01x45, va01x89ABCDEF, vb45x89ABCDEF);
161 vacc01x67 = vmmlaq_s32(vacc01x67, va01x89ABCDEF, vb67x89ABCDEF);
162 vacc23x01 = vmmlaq_s32(vacc23x01, va23x89ABCDEF, vb01x89ABCDEF);
163 vacc23x23 = vmmlaq_s32(vacc23x23, va23x89ABCDEF, vb23x89ABCDEF);
164 vacc23x45 = vmmlaq_s32(vacc23x45, va23x89ABCDEF, vb45x89ABCDEF);
165 vacc23x67 = vmmlaq_s32(vacc23x67, va23x89ABCDEF, vb67x89ABCDEF);
166 vacc45x01 = vmmlaq_s32(vacc45x01, va45x89ABCDEF, vb01x89ABCDEF);
167 vacc45x23 = vmmlaq_s32(vacc45x23, va45x89ABCDEF, vb23x89ABCDEF);
168 vacc45x45 = vmmlaq_s32(vacc45x45, va45x89ABCDEF, vb45x89ABCDEF);
169 vacc45x67 = vmmlaq_s32(vacc45x67, va45x89ABCDEF, vb67x89ABCDEF);
170
171 k -= 16 * sizeof(int8_t);
172 }
173 // Handle up to 8 final positions of `k`
174 if XNN_UNLIKELY(k != 0) {
175 // Load a 5x8 block of activations.
176 uint64x2_t va01x01234567 = vld1q_dup_u64((const void*) a0); a0 += 8;
177 uint64x2_t va23x01234567 = vld1q_dup_u64((const void*) a2); a2 += 8;
178 uint64x2_t va45x01234567 = vld1q_dup_u64((const void*) a4); a4 += 8;
179 va01x01234567 = vld1q_lane_u64((const void*) a1, va01x01234567, 1); a1 += 8;
180 va23x01234567 = vld1q_lane_u64((const void*) a3, va23x01234567, 1); a3 += 8;
181
182 // Load a 16x8 block of weights.
183 const int8x16_t vb01x0123456789ABCDEF = vld1q_s8(w); w = (const int8_t*) w + 16;
184 const int8x16_t vb23x0123456789ABCDEF = vld1q_s8(w); w = (const int8_t*) w + 16;
185 const int8x16_t vb45x0123456789ABCDEF = vld1q_s8(w); w = (const int8_t*) w + 16;
186 const int8x16_t vb67x0123456789ABCDEF = vld1q_s8(w); w = (const int8_t*) w + 16;
187 const int8x16_t vb01x01234567 = vshlq_n_s8(vb01x0123456789ABCDEF, 4);
188 const int8x16_t vb23x01234567 = vshlq_n_s8(vb23x0123456789ABCDEF, 4);
189 const int8x16_t vb45x01234567 = vshlq_n_s8(vb45x0123456789ABCDEF, 4);
190 const int8x16_t vb67x01234567 = vshlq_n_s8(vb67x0123456789ABCDEF, 4);
191
192 // Multiply-accumulate: 5x4 * 4x8 --> 5x8.
193 vacc01x01 = vmmlaq_s32(vacc01x01, vreinterpretq_s8_u64(va01x01234567), vb01x01234567);
194 vacc01x23 = vmmlaq_s32(vacc01x23, vreinterpretq_s8_u64(va01x01234567), vb23x01234567);
195 vacc01x45 = vmmlaq_s32(vacc01x45, vreinterpretq_s8_u64(va01x01234567), vb45x01234567);
196 vacc01x67 = vmmlaq_s32(vacc01x67, vreinterpretq_s8_u64(va01x01234567), vb67x01234567);
197 vacc23x01 = vmmlaq_s32(vacc23x01, vreinterpretq_s8_u64(va23x01234567), vb01x01234567);
198 vacc23x23 = vmmlaq_s32(vacc23x23, vreinterpretq_s8_u64(va23x01234567), vb23x01234567);
199 vacc23x45 = vmmlaq_s32(vacc23x45, vreinterpretq_s8_u64(va23x01234567), vb45x01234567);
200 vacc23x67 = vmmlaq_s32(vacc23x67, vreinterpretq_s8_u64(va23x01234567), vb67x01234567);
201 vacc45x01 = vmmlaq_s32(vacc45x01, vreinterpretq_s8_u64(va45x01234567), vb01x01234567);
202 vacc45x23 = vmmlaq_s32(vacc45x23, vreinterpretq_s8_u64(va45x01234567), vb23x01234567);
203 vacc45x45 = vmmlaq_s32(vacc45x45, vreinterpretq_s8_u64(va45x01234567), vb45x01234567);
204 vacc45x67 = vmmlaq_s32(vacc45x67, vreinterpretq_s8_u64(va45x01234567), vb67x01234567);
205 }
206
207 int32x4_t vacc0x0123 = vreinterpretq_s32_u64(vtrn1q_u64(vreinterpretq_u64_s32(vacc01x01), vreinterpretq_u64_s32(vacc01x23)));
208 int32x4_t vacc1x0123 = vreinterpretq_s32_u64(vtrn2q_u64(vreinterpretq_u64_s32(vacc01x01), vreinterpretq_u64_s32(vacc01x23)));
209 int32x4_t vacc0x4567 = vreinterpretq_s32_u64(vtrn1q_u64(vreinterpretq_u64_s32(vacc01x45), vreinterpretq_u64_s32(vacc01x67)));
210 int32x4_t vacc1x4567 = vreinterpretq_s32_u64(vtrn2q_u64(vreinterpretq_u64_s32(vacc01x45), vreinterpretq_u64_s32(vacc01x67)));
211 int32x4_t vacc2x0123 = vreinterpretq_s32_u64(vtrn1q_u64(vreinterpretq_u64_s32(vacc23x01), vreinterpretq_u64_s32(vacc23x23)));
212 int32x4_t vacc3x0123 = vreinterpretq_s32_u64(vtrn2q_u64(vreinterpretq_u64_s32(vacc23x01), vreinterpretq_u64_s32(vacc23x23)));
213 int32x4_t vacc2x4567 = vreinterpretq_s32_u64(vtrn1q_u64(vreinterpretq_u64_s32(vacc23x45), vreinterpretq_u64_s32(vacc23x67)));
214 int32x4_t vacc3x4567 = vreinterpretq_s32_u64(vtrn2q_u64(vreinterpretq_u64_s32(vacc23x45), vreinterpretq_u64_s32(vacc23x67)));
215 int32x4_t vacc4x0123 = vreinterpretq_s32_u64(vtrn1q_u64(vreinterpretq_u64_s32(vacc45x01), vreinterpretq_u64_s32(vacc45x23)));
216 int32x4_t vacc4x4567 = vreinterpretq_s32_u64(vtrn1q_u64(vreinterpretq_u64_s32(vacc45x45), vreinterpretq_u64_s32(vacc45x67)));
217 float32x4_t vout0x0123 = vcvtq_n_f32_s32(vacc0x0123, 4);
218 float32x4_t vout0x4567 = vcvtq_n_f32_s32(vacc0x4567, 4);
219 float32x4_t vout1x0123 = vcvtq_n_f32_s32(vacc1x0123, 4);
220 float32x4_t vout1x4567 = vcvtq_n_f32_s32(vacc1x4567, 4);
221 float32x4_t vout2x0123 = vcvtq_n_f32_s32(vacc2x0123, 4);
222 float32x4_t vout2x4567 = vcvtq_n_f32_s32(vacc2x4567, 4);
223 float32x4_t vout3x0123 = vcvtq_n_f32_s32(vacc3x0123, 4);
224 float32x4_t vout3x4567 = vcvtq_n_f32_s32(vacc3x4567, 4);
225 float32x4_t vout4x0123 = vcvtq_n_f32_s32(vacc4x0123, 4);
226 float32x4_t vout4x4567 = vcvtq_n_f32_s32(vacc4x4567, 4);
227
228 const float32x4_t vinput_scale01 = vreinterpretq_f32_s32(vld1q_s32(&quantization_params[0].zero_point));
229 vout0x0123 = vmulq_lane_f32(vout0x0123, vget_low_f32(vinput_scale01), 1);
230 vout1x0123 = vmulq_lane_f32(vout1x0123, vget_high_f32(vinput_scale01), 1);
231 vout0x4567 = vmulq_lane_f32(vout0x4567, vget_low_f32(vinput_scale01), 1);
232 vout1x4567 = vmulq_lane_f32(vout1x4567, vget_high_f32(vinput_scale01), 1);
233 const float32x4_t vinput_scale23 = vreinterpretq_f32_s32(vld1q_s32(&quantization_params[2].zero_point));
234 vout2x0123 = vmulq_lane_f32(vout2x0123, vget_low_f32(vinput_scale23), 1);
235 vout3x0123 = vmulq_lane_f32(vout3x0123, vget_high_f32(vinput_scale23), 1);
236 vout2x4567 = vmulq_lane_f32(vout2x4567, vget_low_f32(vinput_scale23), 1);
237 vout3x4567 = vmulq_lane_f32(vout3x4567, vget_high_f32(vinput_scale23), 1);
238 const float32x4_t vinput_scale4 = vld1q_dup_f32(&quantization_params[4].inv_scale);
239 vout4x0123 = vmulq_f32(vout4x0123, vinput_scale4);
240 vout4x4567 = vmulq_f32(vout4x4567, vinput_scale4);
241
242 const float32x4_t vfilter_output_scale0123 = vld1q_f32(w); w = (const float*) w + 4;
243 const float32x4_t vfilter_output_scale4567 = vld1q_f32(w); w = (const float*) w + 4;
244
245 const float32x4_t vbias0123 = vld1q_f32(w); w = (const float*) w + 4;
246 vout0x0123 = vfmaq_f32(vbias0123, vout0x0123, vfilter_output_scale0123);
247 vout1x0123 = vfmaq_f32(vbias0123, vout1x0123, vfilter_output_scale0123);
248 vout2x0123 = vfmaq_f32(vbias0123, vout2x0123, vfilter_output_scale0123);
249 vout3x0123 = vfmaq_f32(vbias0123, vout3x0123, vfilter_output_scale0123);
250 vout4x0123 = vfmaq_f32(vbias0123, vout4x0123, vfilter_output_scale0123);
251 const float32x4_t vbias4567 = vld1q_f32(w); w = (const float*) w + 4;
252 vout0x4567 = vfmaq_f32(vbias4567, vout0x4567, vfilter_output_scale4567);
253 vout1x4567 = vfmaq_f32(vbias4567, vout1x4567, vfilter_output_scale4567);
254 vout2x4567 = vfmaq_f32(vbias4567, vout2x4567, vfilter_output_scale4567);
255 vout3x4567 = vfmaq_f32(vbias4567, vout3x4567, vfilter_output_scale4567);
256 vout4x4567 = vfmaq_f32(vbias4567, vout4x4567, vfilter_output_scale4567);
257
258 float16x8_t vfp16out0x01234567 = vcombine_f16(vcvt_f16_f32(vout0x0123), vcvt_f16_f32(vout0x4567));
259 float16x8_t vfp16out1x01234567 = vcombine_f16(vcvt_f16_f32(vout1x0123), vcvt_f16_f32(vout1x4567));
260 float16x8_t vfp16out2x01234567 = vcombine_f16(vcvt_f16_f32(vout2x0123), vcvt_f16_f32(vout2x4567));
261 float16x8_t vfp16out3x01234567 = vcombine_f16(vcvt_f16_f32(vout3x0123), vcvt_f16_f32(vout3x4567));
262 float16x8_t vfp16out4x01234567 = vcombine_f16(vcvt_f16_f32(vout4x0123), vcvt_f16_f32(vout4x4567));
263
264 const float16x8_t voutput_min = vreinterpretq_f16_u16(vld1q_dup_u16((const uint16_t*) &params->scalar.min));
265 vfp16out0x01234567 = vmaxq_f16(vfp16out0x01234567, voutput_min);
266 vfp16out1x01234567 = vmaxq_f16(vfp16out1x01234567, voutput_min);
267 vfp16out2x01234567 = vmaxq_f16(vfp16out2x01234567, voutput_min);
268 vfp16out3x01234567 = vmaxq_f16(vfp16out3x01234567, voutput_min);
269 vfp16out4x01234567 = vmaxq_f16(vfp16out4x01234567, voutput_min);
270 const float16x8_t voutput_max = vreinterpretq_f16_u16(vld1q_dup_u16((const uint16_t*) &params->scalar.max));
271 vfp16out0x01234567 = vminq_f16(vfp16out0x01234567, voutput_max);
272 vfp16out1x01234567 = vminq_f16(vfp16out1x01234567, voutput_max);
273 vfp16out2x01234567 = vminq_f16(vfp16out2x01234567, voutput_max);
274 vfp16out3x01234567 = vminq_f16(vfp16out3x01234567, voutput_max);
275 vfp16out4x01234567 = vminq_f16(vfp16out4x01234567, voutput_max);
276 if XNN_LIKELY(nc >= 8) {
277 vst1q_u16(c0, vreinterpretq_u16_f16(vfp16out0x01234567));
278 vst1q_u16(c1, vreinterpretq_u16_f16(vfp16out1x01234567));
279 vst1q_u16(c2, vreinterpretq_u16_f16(vfp16out2x01234567));
280 vst1q_u16(c3, vreinterpretq_u16_f16(vfp16out3x01234567));
281 vst1q_u16(c4, vreinterpretq_u16_f16(vfp16out4x01234567));
282
283 a0 = (const int8_t*) ((uintptr_t) a0 - kc);
284 a1 = (const int8_t*) ((uintptr_t) a1 - kc);
285 a2 = (const int8_t*) ((uintptr_t) a2 - kc);
286 a3 = (const int8_t*) ((uintptr_t) a3 - kc);
287 a4 = (const int8_t*) ((uintptr_t) a4 - kc);
288
289 c0 = (uint16_t*) ((uintptr_t) c0 + cn_stride);
290 c1 = (uint16_t*) ((uintptr_t) c1 + cn_stride);
291 c2 = (uint16_t*) ((uintptr_t) c2 + cn_stride);
292 c3 = (uint16_t*) ((uintptr_t) c3 + cn_stride);
293 c4 = (uint16_t*) ((uintptr_t) c4 + cn_stride);
294
295 nc -= 8;
296 } else {
297 float16x4_t vfp16out0x0123 = vget_low_f16(vfp16out0x01234567);
298 float16x4_t vfp16out1x0123 = vget_low_f16(vfp16out1x01234567);
299 float16x4_t vfp16out2x0123 = vget_low_f16(vfp16out2x01234567);
300 float16x4_t vfp16out3x0123 = vget_low_f16(vfp16out3x01234567);
301 float16x4_t vfp16out4x0123 = vget_low_f16(vfp16out4x01234567);
302 if (nc & 4) {
303 vst1_u16(c0, vreinterpret_u16_f16(vfp16out0x0123)); c0 += 4;
304 vst1_u16(c1, vreinterpret_u16_f16(vfp16out1x0123)); c1 += 4;
305 vst1_u16(c2, vreinterpret_u16_f16(vfp16out2x0123)); c2 += 4;
306 vst1_u16(c3, vreinterpret_u16_f16(vfp16out3x0123)); c3 += 4;
307 vst1_u16(c4, vreinterpret_u16_f16(vfp16out4x0123)); c4 += 4;
308 vfp16out0x0123 = vget_high_f16(vfp16out0x01234567);
309 vfp16out1x0123 = vget_high_f16(vfp16out1x01234567);
310 vfp16out2x0123 = vget_high_f16(vfp16out2x01234567);
311 vfp16out3x0123 = vget_high_f16(vfp16out3x01234567);
312 vfp16out4x0123 = vget_high_f16(vfp16out4x01234567);
313 }
314 if (nc & 2) {
315 vst1_lane_u32((void*) c0, vreinterpret_u32_f16(vfp16out0x0123), 0); c0 += 2;
316 vst1_lane_u32((void*) c1, vreinterpret_u32_f16(vfp16out1x0123), 0); c1 += 2;
317 vst1_lane_u32((void*) c2, vreinterpret_u32_f16(vfp16out2x0123), 0); c2 += 2;
318 vst1_lane_u32((void*) c3, vreinterpret_u32_f16(vfp16out3x0123), 0); c3 += 2;
319 vst1_lane_u32((void*) c4, vreinterpret_u32_f16(vfp16out4x0123), 0); c4 += 2;
320 vfp16out0x0123 = vext_f16(vfp16out0x0123, vfp16out0x0123, 2);
321 vfp16out1x0123 = vext_f16(vfp16out1x0123, vfp16out1x0123, 2);
322 vfp16out2x0123 = vext_f16(vfp16out2x0123, vfp16out2x0123, 2);
323 vfp16out3x0123 = vext_f16(vfp16out3x0123, vfp16out3x0123, 2);
324 vfp16out4x0123 = vext_f16(vfp16out4x0123, vfp16out4x0123, 2);
325 }
326 if (nc & 1) {
327 vst1_lane_u16(c0, vreinterpret_u16_f16(vfp16out0x0123), 0);
328 vst1_lane_u16(c1, vreinterpret_u16_f16(vfp16out1x0123), 0);
329 vst1_lane_u16(c2, vreinterpret_u16_f16(vfp16out2x0123), 0);
330 vst1_lane_u16(c3, vreinterpret_u16_f16(vfp16out3x0123), 0);
331 vst1_lane_u16(c4, vreinterpret_u16_f16(vfp16out4x0123), 0);
332 }
333 nc = 0;
334 }
335 } while (nc != 0);
336 }