Debian Codegraph @

src/qd8-f16-qb4w-gemm/gen/qd8-f16-qb4w-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_qb4w_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_qb4w_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 size_t bl = params->scalar.blocksize;
42 assert(bl <= kc);
43 assert(bl != 0);
44 assert(bl % 32 == 0);
45 const int8_t* a0 = a;
46 uint16_t* c0 = (uint16_t*) c;
47 const int8_t* a1 = (const int8_t*) ((uintptr_t) a0 + a_stride);
48 uint16_t* c1 = (uint16_t*) ((uintptr_t) c0 + cm_stride);
49 if XNN_UNPREDICTABLE(mr < 2) {
50 a1 = a0;
51 c1 = c0;
52 }
53 const int8_t* a2 = (const int8_t*) ((uintptr_t) a1 + a_stride);
54 uint16_t* c2 = (uint16_t*) ((uintptr_t) c1 + cm_stride);
55 if XNN_UNPREDICTABLE(mr <= 2) {
56 a2 = a1;
57 c2 = c1;
58 }
59 const int8_t* a3 = (const int8_t*) ((uintptr_t) a2 + a_stride);
60 uint16_t* c3 = (uint16_t*) ((uintptr_t) c2 + cm_stride);
61 if XNN_UNPREDICTABLE(mr < 4) {
62 a3 = a2;
63 c3 = c2;
64 }
65 const int8_t* a4 = (const int8_t*) ((uintptr_t) a3 + a_stride);
66 uint16_t* c4 = (uint16_t*) ((uintptr_t) c3 + cm_stride);
67 if XNN_UNPREDICTABLE(mr <= 4) {
68 a4 = a3;
69 c4 = c3;
70 }
71 const int8x16_t vmask = vmovq_n_s8(INT8_C(0xF0));
72
73 // Loop over groups of 8 columns.
74 do {
75 // Initialize accumulators with scaled vksum. 8 scaled vksum values are loaded from the
76 // weight matrix, at the start of the group of 8 columns.
77 const float32x4_t vinput_zero_point01 = vcvtq_f32_s32(vld1q_s32(&quantization_params[0].zero_point));
78 const float32x4_t vksum0123 = vld1q_f32(w); w = (const float*) w + 4;
79 float32x4_t vout0x0123 = vmulq_lane_f32(vksum0123, vget_low_f32(vinput_zero_point01), 0);
80 float32x4_t vout1x0123 = vmulq_lane_f32(vksum0123, vget_high_f32(vinput_zero_point01), 0);
81 const float32x4_t vksum4567 = vld1q_f32(w); w = (const float*) w + 4;
82 float32x4_t vout0x4567 = vmulq_lane_f32(vksum4567, vget_low_f32(vinput_zero_point01), 0);
83 float32x4_t vout1x4567 = vmulq_lane_f32(vksum4567, vget_high_f32(vinput_zero_point01), 0);
84 const float32x4_t vinput_zero_point23 = vcvtq_f32_s32(vld1q_s32(&quantization_params[2].zero_point));
85 float32x4_t vout2x0123 = vmulq_lane_f32(vksum0123, vget_low_f32(vinput_zero_point23), 0);
86 float32x4_t vout3x0123 = vmulq_lane_f32(vksum0123, vget_high_f32(vinput_zero_point23), 0);
87 float32x4_t vout2x4567 = vmulq_lane_f32(vksum4567, vget_low_f32(vinput_zero_point23), 0);
88 float32x4_t vout3x4567 = vmulq_lane_f32(vksum4567, vget_high_f32(vinput_zero_point23), 0);
89 const float32x4_t vinput_zero_point45 = vcvtq_f32_s32(vld1q_dup_s32(&quantization_params[4].zero_point));
90 float32x4_t vout4x0123 = vmulq_f32(vksum0123, vinput_zero_point45);
91 float32x4_t vout4x4567 = vmulq_f32(vksum4567, vinput_zero_point45);
92
93
94 for (size_t kb=0; kb < kc; kb += bl) {
95 int32x4_t vacc01x01 = vdupq_n_s32(0);
96 int32x4_t vacc01x23 = vdupq_n_s32(0);
97 int32x4_t vacc01x45 = vdupq_n_s32(0);
98 int32x4_t vacc01x67 = vdupq_n_s32(0);
99 int32x4_t vacc23x01 = vdupq_n_s32(0);
100 int32x4_t vacc23x23 = vdupq_n_s32(0);
101 int32x4_t vacc23x45 = vdupq_n_s32(0);
102 int32x4_t vacc23x67 = vdupq_n_s32(0);
103 int32x4_t vacc45x01 = vdupq_n_s32(0);
104 int32x4_t vacc45x23 = vdupq_n_s32(0);
105 int32x4_t vacc45x45 = vdupq_n_s32(0);
106 int32x4_t vacc45x67 = vdupq_n_s32(0);
107
108 size_t k = bl;
109 // 2x partial unrolled loop to load 8 bytes at a time.
110
111 uint64x2x2_t va01x0123456789ABCDEF;
112 va01x0123456789ABCDEF.val[0] = vdupq_n_u64(0);
113 va01x0123456789ABCDEF.val[1] = vdupq_n_u64(0);
114 uint64x2x2_t va23x0123456789ABCDEF;
115 va23x0123456789ABCDEF.val[0] = vdupq_n_u64(0);
116 va23x0123456789ABCDEF.val[1] = vdupq_n_u64(0);
117 uint64x2x2_t va45x0123456789ABCDEF;
118 va45x0123456789ABCDEF.val[0] = vdupq_n_u64(0);
119 va45x0123456789ABCDEF.val[1] = vdupq_n_u64(0);
120
121 while (k >= 16 * sizeof(int8_t)) {
122 // Load a 5x16 block of activations.
123 va01x0123456789ABCDEF = vld2q_lane_u64((const void*) a0, va01x0123456789ABCDEF, 0); a0 += 16;
124 va23x0123456789ABCDEF = vld2q_lane_u64((const void*) a2, va23x0123456789ABCDEF, 0); a2 += 16;
125 va45x0123456789ABCDEF = vld2q_lane_u64((const void*) a4, va45x0123456789ABCDEF, 0); a4 += 16;
126 va01x0123456789ABCDEF = vld2q_lane_u64((const void*) a1, va01x0123456789ABCDEF, 1); a1 += 16;
127 va23x0123456789ABCDEF = vld2q_lane_u64((const void*) a3, va23x0123456789ABCDEF, 1); a3 += 16;
128
129 const int8x16_t va01x01234567 = vreinterpretq_s8_u64(va01x0123456789ABCDEF.val[0]);
130 const int8x16_t va01x89ABCDEF = vreinterpretq_s8_u64(va01x0123456789ABCDEF.val[1]);
131 const int8x16_t va23x01234567 = vreinterpretq_s8_u64(va23x0123456789ABCDEF.val[0]);
132 const int8x16_t va23x89ABCDEF = vreinterpretq_s8_u64(va23x0123456789ABCDEF.val[1]);
133 const int8x16_t va45x01234567 = vreinterpretq_s8_u64(va45x0123456789ABCDEF.val[0]);
134 const int8x16_t va45x89ABCDEF = vreinterpretq_s8_u64(va45x0123456789ABCDEF.val[1]);
135
136 // Load a 16x8 block of weights.
137 const int8x16_t vb01x0123456789ABCDEF = vld1q_s8(w); w = (const int8_t*) w + 16;
138 const int8x16_t vb23x0123456789ABCDEF = vld1q_s8(w); w = (const int8_t*) w + 16;
139 const int8x16_t vb45x0123456789ABCDEF = vld1q_s8(w); w = (const int8_t*) w + 16;
140 const int8x16_t vb67x0123456789ABCDEF = vld1q_s8(w); w = (const int8_t*) w + 16;
141 const int8x16_t vb01x01234567 = vshlq_n_s8(vb01x0123456789ABCDEF, 4);
142 const int8x16_t vb23x01234567 = vshlq_n_s8(vb23x0123456789ABCDEF, 4);
143 const int8x16_t vb45x01234567 = vshlq_n_s8(vb45x0123456789ABCDEF, 4);
144 const int8x16_t vb67x01234567 = vshlq_n_s8(vb67x0123456789ABCDEF, 4);
145 const int8x16_t vb01x89ABCDEF = vandq_s8(vb01x0123456789ABCDEF, vmask);
146 const int8x16_t vb23x89ABCDEF = vandq_s8(vb23x0123456789ABCDEF, vmask);
147 const int8x16_t vb45x89ABCDEF = vandq_s8(vb45x0123456789ABCDEF, vmask);
148 const int8x16_t vb67x89ABCDEF = vandq_s8(vb67x0123456789ABCDEF, vmask);
149
150 // Multiply-accumulate: 5x8 * 8x8 --> 5x8.
151 vacc01x01 = vmmlaq_s32(vacc01x01, va01x01234567, vb01x01234567);
152 vacc01x23 = vmmlaq_s32(vacc01x23, va01x01234567, vb23x01234567);
153 vacc01x45 = vmmlaq_s32(vacc01x45, va01x01234567, vb45x01234567);
154 vacc01x67 = vmmlaq_s32(vacc01x67, va01x01234567, vb67x01234567);
155 vacc23x01 = vmmlaq_s32(vacc23x01, va23x01234567, vb01x01234567);
156 vacc23x23 = vmmlaq_s32(vacc23x23, va23x01234567, vb23x01234567);
157 vacc23x45 = vmmlaq_s32(vacc23x45, va23x01234567, vb45x01234567);
158 vacc23x67 = vmmlaq_s32(vacc23x67, va23x01234567, vb67x01234567);
159 vacc45x01 = vmmlaq_s32(vacc45x01, va45x01234567, vb01x01234567);
160 vacc45x23 = vmmlaq_s32(vacc45x23, va45x01234567, vb23x01234567);
161 vacc45x45 = vmmlaq_s32(vacc45x45, va45x01234567, vb45x01234567);
162 vacc45x67 = vmmlaq_s32(vacc45x67, va45x01234567, vb67x01234567);
163 vacc01x01 = vmmlaq_s32(vacc01x01, va01x89ABCDEF, vb01x89ABCDEF);
164 vacc01x23 = vmmlaq_s32(vacc01x23, va01x89ABCDEF, vb23x89ABCDEF);
165 vacc01x45 = vmmlaq_s32(vacc01x45, va01x89ABCDEF, vb45x89ABCDEF);
166 vacc01x67 = vmmlaq_s32(vacc01x67, va01x89ABCDEF, vb67x89ABCDEF);
167 vacc23x01 = vmmlaq_s32(vacc23x01, va23x89ABCDEF, vb01x89ABCDEF);
168 vacc23x23 = vmmlaq_s32(vacc23x23, va23x89ABCDEF, vb23x89ABCDEF);
169 vacc23x45 = vmmlaq_s32(vacc23x45, va23x89ABCDEF, vb45x89ABCDEF);
170 vacc23x67 = vmmlaq_s32(vacc23x67, va23x89ABCDEF, vb67x89ABCDEF);
171 vacc45x01 = vmmlaq_s32(vacc45x01, va45x89ABCDEF, vb01x89ABCDEF);
172 vacc45x23 = vmmlaq_s32(vacc45x23, va45x89ABCDEF, vb23x89ABCDEF);
173 vacc45x45 = vmmlaq_s32(vacc45x45, va45x89ABCDEF, vb45x89ABCDEF);
174 vacc45x67 = vmmlaq_s32(vacc45x67, va45x89ABCDEF, vb67x89ABCDEF);
175
176 k -= 16 * sizeof(int8_t);
177 }
178 // Handle up to 8 final positions of `k`
179 if XNN_UNLIKELY(k != 0) {
180 // Load a 5x8 block of activations.
181 uint64x2_t va01x01234567 = vld1q_dup_u64((const void*) a0); a0 += 8;
182 uint64x2_t va23x01234567 = vld1q_dup_u64((const void*) a2); a2 += 8;
183 uint64x2_t va45x01234567 = vld1q_dup_u64((const void*) a4); a4 += 8;
184 va01x01234567 = vld1q_lane_u64((const void*) a1, va01x01234567, 1); a1 += 8;
185 va23x01234567 = vld1q_lane_u64((const void*) a3, va23x01234567, 1); a3 += 8;
186
187 // Load a 16x8 block of weights.
188 const int8x16_t vb01x0123456789ABCDEF = vld1q_s8(w); w = (const int8_t*) w + 16;
189 const int8x16_t vb23x0123456789ABCDEF = vld1q_s8(w); w = (const int8_t*) w + 16;
190 const int8x16_t vb45x0123456789ABCDEF = vld1q_s8(w); w = (const int8_t*) w + 16;
191 const int8x16_t vb67x0123456789ABCDEF = vld1q_s8(w); w = (const int8_t*) w + 16;
192 const int8x16_t vb01x01234567 = vshlq_n_s8(vb01x0123456789ABCDEF, 4);
193 const int8x16_t vb23x01234567 = vshlq_n_s8(vb23x0123456789ABCDEF, 4);
194 const int8x16_t vb45x01234567 = vshlq_n_s8(vb45x0123456789ABCDEF, 4);
195 const int8x16_t vb67x01234567 = vshlq_n_s8(vb67x0123456789ABCDEF, 4);
196
197 // Multiply-accumulate: 5x4 * 4x8 --> 5x8.
198 vacc01x01 = vmmlaq_s32(vacc01x01, vreinterpretq_s8_u64(va01x01234567), vb01x01234567);
199 vacc01x23 = vmmlaq_s32(vacc01x23, vreinterpretq_s8_u64(va01x01234567), vb23x01234567);
200 vacc01x45 = vmmlaq_s32(vacc01x45, vreinterpretq_s8_u64(va01x01234567), vb45x01234567);
201 vacc01x67 = vmmlaq_s32(vacc01x67, vreinterpretq_s8_u64(va01x01234567), vb67x01234567);
202 vacc23x01 = vmmlaq_s32(vacc23x01, vreinterpretq_s8_u64(va23x01234567), vb01x01234567);
203 vacc23x23 = vmmlaq_s32(vacc23x23, vreinterpretq_s8_u64(va23x01234567), vb23x01234567);
204 vacc23x45 = vmmlaq_s32(vacc23x45, vreinterpretq_s8_u64(va23x01234567), vb45x01234567);
205 vacc23x67 = vmmlaq_s32(vacc23x67, vreinterpretq_s8_u64(va23x01234567), vb67x01234567);
206 vacc45x01 = vmmlaq_s32(vacc45x01, vreinterpretq_s8_u64(va45x01234567), vb01x01234567);
207 vacc45x23 = vmmlaq_s32(vacc45x23, vreinterpretq_s8_u64(va45x01234567), vb23x01234567);
208 vacc45x45 = vmmlaq_s32(vacc45x45, vreinterpretq_s8_u64(va45x01234567), vb45x01234567);
209 vacc45x67 = vmmlaq_s32(vacc45x67, vreinterpretq_s8_u64(va45x01234567), vb67x01234567);
210 }
211
212 int32x4_t vacc0x0123 = vreinterpretq_s32_u64(vtrn1q_u64(vreinterpretq_u64_s32(vacc01x01), vreinterpretq_u64_s32(vacc01x23)));
213 int32x4_t vacc1x0123 = vreinterpretq_s32_u64(vtrn2q_u64(vreinterpretq_u64_s32(vacc01x01), vreinterpretq_u64_s32(vacc01x23)));
214 int32x4_t vacc0x4567 = vreinterpretq_s32_u64(vtrn1q_u64(vreinterpretq_u64_s32(vacc01x45), vreinterpretq_u64_s32(vacc01x67)));
215 int32x4_t vacc1x4567 = vreinterpretq_s32_u64(vtrn2q_u64(vreinterpretq_u64_s32(vacc01x45), vreinterpretq_u64_s32(vacc01x67)));
216 int32x4_t vacc2x0123 = vreinterpretq_s32_u64(vtrn1q_u64(vreinterpretq_u64_s32(vacc23x01), vreinterpretq_u64_s32(vacc23x23)));
217 int32x4_t vacc3x0123 = vreinterpretq_s32_u64(vtrn2q_u64(vreinterpretq_u64_s32(vacc23x01), vreinterpretq_u64_s32(vacc23x23)));
218 int32x4_t vacc2x4567 = vreinterpretq_s32_u64(vtrn1q_u64(vreinterpretq_u64_s32(vacc23x45), vreinterpretq_u64_s32(vacc23x67)));
219 int32x4_t vacc3x4567 = vreinterpretq_s32_u64(vtrn2q_u64(vreinterpretq_u64_s32(vacc23x45), vreinterpretq_u64_s32(vacc23x67)));
220 int32x4_t vacc4x0123 = vreinterpretq_s32_u64(vtrn1q_u64(vreinterpretq_u64_s32(vacc45x01), vreinterpretq_u64_s32(vacc45x23)));
221 int32x4_t vacc4x4567 = vreinterpretq_s32_u64(vtrn1q_u64(vreinterpretq_u64_s32(vacc45x45), vreinterpretq_u64_s32(vacc45x67)));
222 const float32x4_t vfilter_output_scale0123 = vreinterpretq_f32_u32(vshll_n_u16(vld1_u16(w), 16)); w = (const uint16_t*) w + 4;
223 const float32x4_t vfilter_output_scale4567 = vreinterpretq_f32_u32(vshll_n_u16(vld1_u16(w), 16)); w = (const uint16_t*) w + 4;
224 float32x4_t vf0x0123 = vcvtq_f32_s32(vacc0x0123);
225 vout0x0123 = vfmaq_f32(vout0x0123, vf0x0123, vfilter_output_scale0123);
226 float32x4_t vf0x4567 = vcvtq_f32_s32(vacc0x4567);
227 vout0x4567 = vfmaq_f32(vout0x4567, vf0x4567, vfilter_output_scale4567);
228 float32x4_t vf1x0123 = vcvtq_f32_s32(vacc1x0123);
229 vout1x0123 = vfmaq_f32(vout1x0123, vf1x0123, vfilter_output_scale0123);
230 float32x4_t vf1x4567 = vcvtq_f32_s32(vacc1x4567);
231 vout1x4567 = vfmaq_f32(vout1x4567, vf1x4567, vfilter_output_scale4567);
232 float32x4_t vf2x0123 = vcvtq_f32_s32(vacc2x0123);
233 vout2x0123 = vfmaq_f32(vout2x0123, vf2x0123, vfilter_output_scale0123);
234 float32x4_t vf2x4567 = vcvtq_f32_s32(vacc2x4567);
235 vout2x4567 = vfmaq_f32(vout2x4567, vf2x4567, vfilter_output_scale4567);
236 float32x4_t vf3x0123 = vcvtq_f32_s32(vacc3x0123);
237 vout3x0123 = vfmaq_f32(vout3x0123, vf3x0123, vfilter_output_scale0123);
238 float32x4_t vf3x4567 = vcvtq_f32_s32(vacc3x4567);
239 vout3x4567 = vfmaq_f32(vout3x4567, vf3x4567, vfilter_output_scale4567);
240 float32x4_t vf4x0123 = vcvtq_f32_s32(vacc4x0123);
241 vout4x0123 = vfmaq_f32(vout4x0123, vf4x0123, vfilter_output_scale0123);
242 float32x4_t vf4x4567 = vcvtq_f32_s32(vacc4x4567);
243 vout4x4567 = vfmaq_f32(vout4x4567, vf4x4567, vfilter_output_scale4567);
244 }
245
246 const float32x4_t vinput_scale01 = vreinterpretq_f32_s32(vld1q_s32(&quantization_params[0].zero_point));
247 vout0x0123 = vmulq_lane_f32(vout0x0123, vget_low_f32(vinput_scale01), 1);
248 vout1x0123 = vmulq_lane_f32(vout1x0123, vget_high_f32(vinput_scale01), 1);
249 vout0x4567 = vmulq_lane_f32(vout0x4567, vget_low_f32(vinput_scale01), 1);
250 vout1x4567 = vmulq_lane_f32(vout1x4567, vget_high_f32(vinput_scale01), 1);
251 const float32x4_t vinput_scale23 = vreinterpretq_f32_s32(vld1q_s32(&quantization_params[2].zero_point));
252 vout2x0123 = vmulq_lane_f32(vout2x0123, vget_low_f32(vinput_scale23), 1);
253 vout3x0123 = vmulq_lane_f32(vout3x0123, vget_high_f32(vinput_scale23), 1);
254 vout2x4567 = vmulq_lane_f32(vout2x4567, vget_low_f32(vinput_scale23), 1);
255 vout3x4567 = vmulq_lane_f32(vout3x4567, vget_high_f32(vinput_scale23), 1);
256 const float32x4_t vinput_scale4 = vld1q_dup_f32(&quantization_params[4].inv_scale);
257 vout4x0123 = vmulq_f32(vout4x0123, vinput_scale4);
258 vout4x4567 = vmulq_f32(vout4x4567, vinput_scale4);
259
260
261 const float32x4_t vbias0123 = vld1q_f32(w); w = (const float*) w + 4;
262 vout0x0123 = vaddq_f32(vbias0123, vout0x0123);
263 vout1x0123 = vaddq_f32(vbias0123, vout1x0123);
264 vout2x0123 = vaddq_f32(vbias0123, vout2x0123);
265 vout3x0123 = vaddq_f32(vbias0123, vout3x0123);
266 vout4x0123 = vaddq_f32(vbias0123, vout4x0123);
267 const float32x4_t vbias4567 = vld1q_f32(w); w = (const float*) w + 4;
268 vout0x4567 = vaddq_f32(vbias4567, vout0x4567);
269 vout1x4567 = vaddq_f32(vbias4567, vout1x4567);
270 vout2x4567 = vaddq_f32(vbias4567, vout2x4567);
271 vout3x4567 = vaddq_f32(vbias4567, vout3x4567);
272 vout4x4567 = vaddq_f32(vbias4567, vout4x4567);
273
274 float16x8_t vfp16out0x01234567 = vcombine_f16(vcvt_f16_f32(vout0x0123), vcvt_f16_f32(vout0x4567));
275 float16x8_t vfp16out1x01234567 = vcombine_f16(vcvt_f16_f32(vout1x0123), vcvt_f16_f32(vout1x4567));
276 float16x8_t vfp16out2x01234567 = vcombine_f16(vcvt_f16_f32(vout2x0123), vcvt_f16_f32(vout2x4567));
277 float16x8_t vfp16out3x01234567 = vcombine_f16(vcvt_f16_f32(vout3x0123), vcvt_f16_f32(vout3x4567));
278 float16x8_t vfp16out4x01234567 = vcombine_f16(vcvt_f16_f32(vout4x0123), vcvt_f16_f32(vout4x4567));
279
280 const float16x8_t voutput_min = vreinterpretq_f16_u16(vld1q_dup_u16((const uint16_t*) &params->scalar.min));
281 vfp16out0x01234567 = vmaxq_f16(vfp16out0x01234567, voutput_min);
282 vfp16out1x01234567 = vmaxq_f16(vfp16out1x01234567, voutput_min);
283 vfp16out2x01234567 = vmaxq_f16(vfp16out2x01234567, voutput_min);
284 vfp16out3x01234567 = vmaxq_f16(vfp16out3x01234567, voutput_min);
285 vfp16out4x01234567 = vmaxq_f16(vfp16out4x01234567, voutput_min);
286 const float16x8_t voutput_max = vreinterpretq_f16_u16(vld1q_dup_u16((const uint16_t*) &params->scalar.max));
287 vfp16out0x01234567 = vminq_f16(vfp16out0x01234567, voutput_max);
288 vfp16out1x01234567 = vminq_f16(vfp16out1x01234567, voutput_max);
289 vfp16out2x01234567 = vminq_f16(vfp16out2x01234567, voutput_max);
290 vfp16out3x01234567 = vminq_f16(vfp16out3x01234567, voutput_max);
291 vfp16out4x01234567 = vminq_f16(vfp16out4x01234567, voutput_max);
292 if XNN_LIKELY(nc >= 8) {
293 vst1q_u16(c0, vreinterpretq_u16_f16(vfp16out0x01234567));
294 vst1q_u16(c1, vreinterpretq_u16_f16(vfp16out1x01234567));
295 vst1q_u16(c2, vreinterpretq_u16_f16(vfp16out2x01234567));
296 vst1q_u16(c3, vreinterpretq_u16_f16(vfp16out3x01234567));
297 vst1q_u16(c4, vreinterpretq_u16_f16(vfp16out4x01234567));
298
299 a0 = (const int8_t*) ((uintptr_t) a0 - kc);
300 a1 = (const int8_t*) ((uintptr_t) a1 - kc);
301 a2 = (const int8_t*) ((uintptr_t) a2 - kc);
302 a3 = (const int8_t*) ((uintptr_t) a3 - kc);
303 a4 = (const int8_t*) ((uintptr_t) a4 - kc);
304
305 c0 = (uint16_t*) ((uintptr_t) c0 + cn_stride);
306 c1 = (uint16_t*) ((uintptr_t) c1 + cn_stride);
307 c2 = (uint16_t*) ((uintptr_t) c2 + cn_stride);
308 c3 = (uint16_t*) ((uintptr_t) c3 + cn_stride);
309 c4 = (uint16_t*) ((uintptr_t) c4 + cn_stride);
310
311 nc -= 8;
312 } else {
313 float16x4_t vfp16out0x0123 = vget_low_f16(vfp16out0x01234567);
314 float16x4_t vfp16out1x0123 = vget_low_f16(vfp16out1x01234567);
315 float16x4_t vfp16out2x0123 = vget_low_f16(vfp16out2x01234567);
316 float16x4_t vfp16out3x0123 = vget_low_f16(vfp16out3x01234567);
317 float16x4_t vfp16out4x0123 = vget_low_f16(vfp16out4x01234567);
318 if (nc & 4) {
319 vst1_u16(c0, vreinterpret_u16_f16(vfp16out0x0123)); c0 += 4;
320 vst1_u16(c1, vreinterpret_u16_f16(vfp16out1x0123)); c1 += 4;
321 vst1_u16(c2, vreinterpret_u16_f16(vfp16out2x0123)); c2 += 4;
322 vst1_u16(c3, vreinterpret_u16_f16(vfp16out3x0123)); c3 += 4;
323 vst1_u16(c4, vreinterpret_u16_f16(vfp16out4x0123)); c4 += 4;
324 vfp16out0x0123 = vget_high_f16(vfp16out0x01234567);
325 vfp16out1x0123 = vget_high_f16(vfp16out1x01234567);
326 vfp16out2x0123 = vget_high_f16(vfp16out2x01234567);
327 vfp16out3x0123 = vget_high_f16(vfp16out3x01234567);
328 vfp16out4x0123 = vget_high_f16(vfp16out4x01234567);
329 }
330 if (nc & 2) {
331 vst1_lane_u32((void*) c0, vreinterpret_u32_f16(vfp16out0x0123), 0); c0 += 2;
332 vst1_lane_u32((void*) c1, vreinterpret_u32_f16(vfp16out1x0123), 0); c1 += 2;
333 vst1_lane_u32((void*) c2, vreinterpret_u32_f16(vfp16out2x0123), 0); c2 += 2;
334 vst1_lane_u32((void*) c3, vreinterpret_u32_f16(vfp16out3x0123), 0); c3 += 2;
335 vst1_lane_u32((void*) c4, vreinterpret_u32_f16(vfp16out4x0123), 0); c4 += 2;
336 vfp16out0x0123 = vext_f16(vfp16out0x0123, vfp16out0x0123, 2);
337 vfp16out1x0123 = vext_f16(vfp16out1x0123, vfp16out1x0123, 2);
338 vfp16out2x0123 = vext_f16(vfp16out2x0123, vfp16out2x0123, 2);
339 vfp16out3x0123 = vext_f16(vfp16out3x0123, vfp16out3x0123, 2);
340 vfp16out4x0123 = vext_f16(vfp16out4x0123, vfp16out4x0123, 2);
341 }
342 if (nc & 1) {
343 vst1_lane_u16(c0, vreinterpret_u16_f16(vfp16out0x0123), 0);
344 vst1_lane_u16(c1, vreinterpret_u16_f16(vfp16out1x0123), 0);
345 vst1_lane_u16(c2, vreinterpret_u16_f16(vfp16out2x0123), 0);
346 vst1_lane_u16(c3, vreinterpret_u16_f16(vfp16out3x0123), 0);
347 vst1_lane_u16(c4, vreinterpret_u16_f16(vfp16out4x0123), 0);
348 }
349 nc = 0;
350 }
351 } while (nc != 0);
352 }