Debian Codegraph @

src/qd8-f16-qb4w-gemm/gen/qd8-f16-qb4w-gemm-4x16c4-minmax-neondotfp16arith.c

raw
1 // Auto-generated file. Do not edit!
2 // Template: src/qs8-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
18 void xnn_qd8_f16_qb4w_gemm_minmax_ukernel_4x16c4__neondotfp16arith(
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 <= 4);
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, 4 * 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
62 size_t bl = params->scalar.blocksize;
63 assert(bl <= kc);
64 assert(bl != 0);
65 assert(bl % 32 == 0);
66 const int8x16_t vmask = vmovq_n_s8(INT8_C(0xF0));
67 // Loop over groups of 16 columns.
68 do {
69 // Initialize accumulators with bias. 16 bias values are loaded from the
70 // weight matrix, at the start of the group of 16 columns.
71 const float32x4_t vinput_zero_point01 = vcvtq_f32_s32(vld1q_s32(&quantization_params[0].zero_point));
72 const float32x4_t vksum0123 = vld1q_f32(w); w = (const float*) w + 4;
73 float32x4_t vout0x0123 = vmulq_lane_f32(vksum0123, vget_low_f32(vinput_zero_point01), 0);
74 float32x4_t vout1x0123 = vmulq_lane_f32(vksum0123, vget_high_f32(vinput_zero_point01), 0);
75 const float32x4_t vksum4567 = vld1q_f32(w); w = (const float*) w + 4;
76 float32x4_t vout0x4567 = vmulq_lane_f32(vksum4567, vget_low_f32(vinput_zero_point01), 0);
77 float32x4_t vout1x4567 = vmulq_lane_f32(vksum4567, vget_high_f32(vinput_zero_point01), 0);
78 const float32x4_t vksum89AB = vld1q_f32(w); w = (const float*) w + 4;
79 float32x4_t vout0x89AB = vmulq_lane_f32(vksum89AB, vget_low_f32(vinput_zero_point01), 0);
80 float32x4_t vout1x89AB = vmulq_lane_f32(vksum89AB, vget_high_f32(vinput_zero_point01), 0);
81 const float32x4_t vksumCDEF = vld1q_f32(w); w = (const float*) w + 4;
82 float32x4_t vout0xCDEF = vmulq_lane_f32(vksumCDEF, vget_low_f32(vinput_zero_point01), 0);
83 float32x4_t vout1xCDEF = vmulq_lane_f32(vksumCDEF, 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 float32x4_t vout2x89AB = vmulq_lane_f32(vksum89AB, vget_low_f32(vinput_zero_point23), 0);
90 float32x4_t vout3x89AB = vmulq_lane_f32(vksum89AB, vget_high_f32(vinput_zero_point23), 0);
91 float32x4_t vout2xCDEF = vmulq_lane_f32(vksumCDEF, vget_low_f32(vinput_zero_point23), 0);
92 float32x4_t vout3xCDEF = vmulq_lane_f32(vksumCDEF, vget_high_f32(vinput_zero_point23), 0);
93
94 for (size_t kb=0; kb < kc; kb += bl) {
95 int32x4_t vacc0x0123 = vdupq_n_s32(0);
96 int32x4_t vacc1x0123 = vdupq_n_s32(0);
97 int32x4_t vacc0x4567 = vdupq_n_s32(0);
98 int32x4_t vacc1x4567 = vdupq_n_s32(0);
99 int32x4_t vacc0x89AB = vdupq_n_s32(0);
100 int32x4_t vacc1x89AB = vdupq_n_s32(0);
101 int32x4_t vacc0xCDEF = vdupq_n_s32(0);
102 int32x4_t vacc1xCDEF = vdupq_n_s32(0);
103 int32x4_t vacc2x0123 = vdupq_n_s32(0);
104 int32x4_t vacc3x0123 = vdupq_n_s32(0);
105 int32x4_t vacc2x4567 = vdupq_n_s32(0);
106 int32x4_t vacc3x4567 = vdupq_n_s32(0);
107 int32x4_t vacc2x89AB = vdupq_n_s32(0);
108 int32x4_t vacc3x89AB = vdupq_n_s32(0);
109 int32x4_t vacc2xCDEF = vdupq_n_s32(0);
110 int32x4_t vacc3xCDEF = vdupq_n_s32(0);
111 // Inner accumulation loop along the 16 columns.
112 size_t k = bl;
113 // 2x partial unrolled loop to load 8 bytes at a time.
114 while (k >= 8 * sizeof(int8_t)) {
115 // Load a 4x8 block of activations.
116 const int8x8_t va0x01234567 = vld1_s8(a0); a0 += 8;
117 const int8x8_t va1x01234567 = vld1_s8(a1); a1 += 8;
118 const int8x8_t va2x01234567 = vld1_s8(a2); a2 += 8;
119 const int8x8_t va3x01234567 = vld1_s8(a3); a3 += 8;
120
121 // Load a 8x16 block of weights.
122 const int8x16_t vb01234567x0123 = vld1q_s8(w); w = (const int8_t*) w + 16;
123 const int8x16_t vb01234567x4567 = vld1q_s8(w); w = (const int8_t*) w + 16;
124 const int8x16_t vb01234567x89AB = vld1q_s8(w); w = (const int8_t*) w + 16;
125 const int8x16_t vb01234567xCDEF = vld1q_s8(w); w = (const int8_t*) w + 16;
126 const int8x16_t vb0123x0123 = vshlq_n_s8(vb01234567x0123, 4);
127 const int8x16_t vb0123x4567 = vshlq_n_s8(vb01234567x4567, 4);
128 const int8x16_t vb0123x89AB = vshlq_n_s8(vb01234567x89AB, 4);
129 const int8x16_t vb0123xCDEF = vshlq_n_s8(vb01234567xCDEF, 4);
130 const int8x16_t vb4567x0123 = vandq_s8(vb01234567x0123, vmask);
131 const int8x16_t vb4567x4567 = vandq_s8(vb01234567x4567, vmask);
132 const int8x16_t vb4567x89AB = vandq_s8(vb01234567x89AB, vmask);
133 const int8x16_t vb4567xCDEF = vandq_s8(vb01234567xCDEF, vmask);
134
135 // Multiply-accumulate: 4x8 * 8x16 --> 4x16.
136 vacc0x0123 = vdotq_lane_s32(vacc0x0123, vb0123x0123, va0x01234567, 0);
137 vacc0x4567 = vdotq_lane_s32(vacc0x4567, vb0123x4567, va0x01234567, 0);
138 vacc0x89AB = vdotq_lane_s32(vacc0x89AB, vb0123x89AB, va0x01234567, 0);
139 vacc0xCDEF = vdotq_lane_s32(vacc0xCDEF, vb0123xCDEF, va0x01234567, 0);
140 vacc1x0123 = vdotq_lane_s32(vacc1x0123, vb0123x0123, va1x01234567, 0);
141 vacc1x4567 = vdotq_lane_s32(vacc1x4567, vb0123x4567, va1x01234567, 0);
142 vacc1x89AB = vdotq_lane_s32(vacc1x89AB, vb0123x89AB, va1x01234567, 0);
143 vacc1xCDEF = vdotq_lane_s32(vacc1xCDEF, vb0123xCDEF, va1x01234567, 0);
144 vacc2x0123 = vdotq_lane_s32(vacc2x0123, vb0123x0123, va2x01234567, 0);
145 vacc2x4567 = vdotq_lane_s32(vacc2x4567, vb0123x4567, va2x01234567, 0);
146 vacc2x89AB = vdotq_lane_s32(vacc2x89AB, vb0123x89AB, va2x01234567, 0);
147 vacc2xCDEF = vdotq_lane_s32(vacc2xCDEF, vb0123xCDEF, va2x01234567, 0);
148 vacc3x0123 = vdotq_lane_s32(vacc3x0123, vb0123x0123, va3x01234567, 0);
149 vacc3x4567 = vdotq_lane_s32(vacc3x4567, vb0123x4567, va3x01234567, 0);
150 vacc3x89AB = vdotq_lane_s32(vacc3x89AB, vb0123x89AB, va3x01234567, 0);
151 vacc3xCDEF = vdotq_lane_s32(vacc3xCDEF, vb0123xCDEF, va3x01234567, 0);
152 vacc0x0123 = vdotq_lane_s32(vacc0x0123, vb4567x0123, va0x01234567, 1);
153 vacc0x4567 = vdotq_lane_s32(vacc0x4567, vb4567x4567, va0x01234567, 1);
154 vacc0x89AB = vdotq_lane_s32(vacc0x89AB, vb4567x89AB, va0x01234567, 1);
155 vacc0xCDEF = vdotq_lane_s32(vacc0xCDEF, vb4567xCDEF, va0x01234567, 1);
156 vacc1x0123 = vdotq_lane_s32(vacc1x0123, vb4567x0123, va1x01234567, 1);
157 vacc1x4567 = vdotq_lane_s32(vacc1x4567, vb4567x4567, va1x01234567, 1);
158 vacc1x89AB = vdotq_lane_s32(vacc1x89AB, vb4567x89AB, va1x01234567, 1);
159 vacc1xCDEF = vdotq_lane_s32(vacc1xCDEF, vb4567xCDEF, va1x01234567, 1);
160 vacc2x0123 = vdotq_lane_s32(vacc2x0123, vb4567x0123, va2x01234567, 1);
161 vacc2x4567 = vdotq_lane_s32(vacc2x4567, vb4567x4567, va2x01234567, 1);
162 vacc2x89AB = vdotq_lane_s32(vacc2x89AB, vb4567x89AB, va2x01234567, 1);
163 vacc2xCDEF = vdotq_lane_s32(vacc2xCDEF, vb4567xCDEF, va2x01234567, 1);
164 vacc3x0123 = vdotq_lane_s32(vacc3x0123, vb4567x0123, va3x01234567, 1);
165 vacc3x4567 = vdotq_lane_s32(vacc3x4567, vb4567x4567, va3x01234567, 1);
166 vacc3x89AB = vdotq_lane_s32(vacc3x89AB, vb4567x89AB, va3x01234567, 1);
167 vacc3xCDEF = vdotq_lane_s32(vacc3xCDEF, vb4567xCDEF, va3x01234567, 1);
168
169 k -= 8 * sizeof(int8_t);
170 }
171 // Handle up to 4 final positions of `k`
172 if XNN_UNLIKELY(k != 0) {
173 // Load a 4x4 block of activations.
174 const int8x8_t va0x01234567 = vld1_s8(a0); a0 += 4;
175 const int8x8_t va1x01234567 = vld1_s8(a1); a1 += 4;
176 const int8x8_t va2x01234567 = vld1_s8(a2); a2 += 4;
177 const int8x8_t va3x01234567 = vld1_s8(a3); a3 += 4;
178
179 // Load a 4x16 block of weights.
180 const int8x16_t vb0123x0123 = vld1q_s8(w); w = (const int8_t*) w + 16;
181 const int8x16_t vb0123x4567 = vld1q_s8(w); w = (const int8_t*) w + 16;
182 const int8x16_t vb0123x89AB = vld1q_s8(w); w = (const int8_t*) w + 16;
183 const int8x16_t vb0123xCDEF = vld1q_s8(w); w = (const int8_t*) w + 16;
184
185 // Multiply-accumulate: 4x4 * 4x16 --> 4x16.
186 vacc0x0123 = vdotq_lane_s32(vacc0x0123, vb0123x0123, va0x01234567, 0);
187 vacc0x4567 = vdotq_lane_s32(vacc0x4567, vb0123x4567, va0x01234567, 0);
188 vacc0x89AB = vdotq_lane_s32(vacc0x89AB, vb0123x89AB, va0x01234567, 0);
189 vacc0xCDEF = vdotq_lane_s32(vacc0xCDEF, vb0123xCDEF, va0x01234567, 0);
190 vacc1x0123 = vdotq_lane_s32(vacc1x0123, vb0123x0123, va1x01234567, 0);
191 vacc1x4567 = vdotq_lane_s32(vacc1x4567, vb0123x4567, va1x01234567, 0);
192 vacc1x89AB = vdotq_lane_s32(vacc1x89AB, vb0123x89AB, va1x01234567, 0);
193 vacc1xCDEF = vdotq_lane_s32(vacc1xCDEF, vb0123xCDEF, va1x01234567, 0);
194 vacc2x0123 = vdotq_lane_s32(vacc2x0123, vb0123x0123, va2x01234567, 0);
195 vacc2x4567 = vdotq_lane_s32(vacc2x4567, vb0123x4567, va2x01234567, 0);
196 vacc2x89AB = vdotq_lane_s32(vacc2x89AB, vb0123x89AB, va2x01234567, 0);
197 vacc2xCDEF = vdotq_lane_s32(vacc2xCDEF, vb0123xCDEF, va2x01234567, 0);
198 vacc3x0123 = vdotq_lane_s32(vacc3x0123, vb0123x0123, va3x01234567, 0);
199 vacc3x4567 = vdotq_lane_s32(vacc3x4567, vb0123x4567, va3x01234567, 0);
200 vacc3x89AB = vdotq_lane_s32(vacc3x89AB, vb0123x89AB, va3x01234567, 0);
201 vacc3xCDEF = vdotq_lane_s32(vacc3xCDEF, vb0123xCDEF, va3x01234567, 0);
202 }
203 const float32x4_t vfilter_output_scale0123 = vreinterpretq_f32_u32(vshll_n_u16(vld1_u16(w), 16)); w = (const uint16_t*) w + 4;
204 const float32x4_t vfilter_output_scale4567 = vreinterpretq_f32_u32(vshll_n_u16(vld1_u16(w), 16)); w = (const uint16_t*) w + 4;
205 const float32x4_t vfilter_output_scale89AB = vreinterpretq_f32_u32(vshll_n_u16(vld1_u16(w), 16)); w = (const uint16_t*) w + 4;
206 const float32x4_t vfilter_output_scaleCDEF = vreinterpretq_f32_u32(vshll_n_u16(vld1_u16(w), 16)); w = (const uint16_t*) w + 4;
207
208 float32x4_t vf0x0123 = vcvtq_f32_s32(vacc0x0123);
209 float32x4_t vf1x0123 = vcvtq_f32_s32(vacc1x0123);
210 vout0x0123 = vfmaq_f32(vout0x0123, vf0x0123, vfilter_output_scale0123);
211 vout1x0123 = vfmaq_f32(vout1x0123, vf1x0123, vfilter_output_scale0123);
212 float32x4_t vf0x4567 = vcvtq_f32_s32(vacc0x4567);
213 float32x4_t vf1x4567 = vcvtq_f32_s32(vacc1x4567);
214 vout0x4567 = vfmaq_f32(vout0x4567, vf0x4567, vfilter_output_scale4567);
215 vout1x4567 = vfmaq_f32(vout1x4567, vf1x4567, vfilter_output_scale4567);
216 float32x4_t vf0x89AB = vcvtq_f32_s32(vacc0x89AB);
217 float32x4_t vf1x89AB = vcvtq_f32_s32(vacc1x89AB);
218 vout0x89AB = vfmaq_f32(vout0x89AB, vf0x89AB, vfilter_output_scale89AB);
219 vout1x89AB = vfmaq_f32(vout1x89AB, vf1x89AB, vfilter_output_scale89AB);
220 float32x4_t vf0xCDEF = vcvtq_f32_s32(vacc0xCDEF);
221 float32x4_t vf1xCDEF = vcvtq_f32_s32(vacc1xCDEF);
222 vout0xCDEF = vfmaq_f32(vout0xCDEF, vf0xCDEF, vfilter_output_scaleCDEF);
223 vout1xCDEF = vfmaq_f32(vout1xCDEF, vf1xCDEF, vfilter_output_scaleCDEF);
224 float32x4_t vf2x0123 = vcvtq_f32_s32(vacc2x0123);
225 float32x4_t vf3x0123 = vcvtq_f32_s32(vacc3x0123);
226 vout2x0123 = vfmaq_f32(vout2x0123, vf2x0123, vfilter_output_scale0123);
227 vout3x0123 = vfmaq_f32(vout3x0123, vf3x0123, vfilter_output_scale0123);
228 float32x4_t vf2x4567 = vcvtq_f32_s32(vacc2x4567);
229 float32x4_t vf3x4567 = vcvtq_f32_s32(vacc3x4567);
230 vout2x4567 = vfmaq_f32(vout2x4567, vf2x4567, vfilter_output_scale4567);
231 vout3x4567 = vfmaq_f32(vout3x4567, vf3x4567, vfilter_output_scale4567);
232 float32x4_t vf2x89AB = vcvtq_f32_s32(vacc2x89AB);
233 float32x4_t vf3x89AB = vcvtq_f32_s32(vacc3x89AB);
234 vout2x89AB = vfmaq_f32(vout2x89AB, vf2x89AB, vfilter_output_scale89AB);
235 vout3x89AB = vfmaq_f32(vout3x89AB, vf3x89AB, vfilter_output_scale89AB);
236 float32x4_t vf2xCDEF = vcvtq_f32_s32(vacc2xCDEF);
237 float32x4_t vf3xCDEF = vcvtq_f32_s32(vacc3xCDEF);
238 vout2xCDEF = vfmaq_f32(vout2xCDEF, vf2xCDEF, vfilter_output_scaleCDEF);
239 vout3xCDEF = vfmaq_f32(vout3xCDEF, vf3xCDEF, vfilter_output_scaleCDEF);
240 }
241
242 const float32x4_t vinput_scale01 = vreinterpretq_f32_s32(vld1q_s32(&quantization_params[0].zero_point));
243 vout0x0123 = vmulq_lane_f32(vout0x0123, vget_low_f32(vinput_scale01), 1);
244 vout1x0123 = vmulq_lane_f32(vout1x0123, vget_high_f32(vinput_scale01), 1);
245 vout0x4567 = vmulq_lane_f32(vout0x4567, vget_low_f32(vinput_scale01), 1);
246 vout1x4567 = vmulq_lane_f32(vout1x4567, vget_high_f32(vinput_scale01), 1);
247 vout0x89AB = vmulq_lane_f32(vout0x89AB, vget_low_f32(vinput_scale01), 1);
248 vout1x89AB = vmulq_lane_f32(vout1x89AB, vget_high_f32(vinput_scale01), 1);
249 vout0xCDEF = vmulq_lane_f32(vout0xCDEF, vget_low_f32(vinput_scale01), 1);
250 vout1xCDEF = vmulq_lane_f32(vout1xCDEF, 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 vout2x89AB = vmulq_lane_f32(vout2x89AB, vget_low_f32(vinput_scale23), 1);
257 vout3x89AB = vmulq_lane_f32(vout3x89AB, vget_high_f32(vinput_scale23), 1);
258 vout2xCDEF = vmulq_lane_f32(vout2xCDEF, vget_low_f32(vinput_scale23), 1);
259 vout3xCDEF = vmulq_lane_f32(vout3xCDEF, vget_high_f32(vinput_scale23), 1);
260
261
262 const float32x4_t vbias0123 = vld1q_f32(w); w = (const float*) w + 4;
263 vout0x0123 = vaddq_f32(vbias0123, vout0x0123);
264 vout1x0123 = vaddq_f32(vbias0123, vout1x0123);
265 vout2x0123 = vaddq_f32(vbias0123, vout2x0123);
266 vout3x0123 = vaddq_f32(vbias0123, vout3x0123);
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 const float32x4_t vbias89AB = vld1q_f32(w); w = (const float*) w + 4;
273 vout0x89AB = vaddq_f32(vbias89AB, vout0x89AB);
274 vout1x89AB = vaddq_f32(vbias89AB, vout1x89AB);
275 vout2x89AB = vaddq_f32(vbias89AB, vout2x89AB);
276 vout3x89AB = vaddq_f32(vbias89AB, vout3x89AB);
277 const float32x4_t vbiasCDEF = vld1q_f32(w); w = (const float*) w + 4;
278 vout0xCDEF = vaddq_f32(vbiasCDEF, vout0xCDEF);
279 vout1xCDEF = vaddq_f32(vbiasCDEF, vout1xCDEF);
280 vout2xCDEF = vaddq_f32(vbiasCDEF, vout2xCDEF);
281 vout3xCDEF = vaddq_f32(vbiasCDEF, vout3xCDEF);
282
283 float16x8_t vfp16out0x01234567 = vcombine_f16(vcvt_f16_f32(vout0x0123), vcvt_f16_f32(vout0x4567));
284 float16x8_t vfp16out0x89ABCDEF = vcombine_f16(vcvt_f16_f32(vout0x89AB), vcvt_f16_f32(vout0xCDEF));
285 float16x8_t vfp16out1x01234567 = vcombine_f16(vcvt_f16_f32(vout1x0123), vcvt_f16_f32(vout1x4567));
286 float16x8_t vfp16out1x89ABCDEF = vcombine_f16(vcvt_f16_f32(vout1x89AB), vcvt_f16_f32(vout1xCDEF));
287 float16x8_t vfp16out2x01234567 = vcombine_f16(vcvt_f16_f32(vout2x0123), vcvt_f16_f32(vout2x4567));
288 float16x8_t vfp16out2x89ABCDEF = vcombine_f16(vcvt_f16_f32(vout2x89AB), vcvt_f16_f32(vout2xCDEF));
289 float16x8_t vfp16out3x01234567 = vcombine_f16(vcvt_f16_f32(vout3x0123), vcvt_f16_f32(vout3x4567));
290 float16x8_t vfp16out3x89ABCDEF = vcombine_f16(vcvt_f16_f32(vout3x89AB), vcvt_f16_f32(vout3xCDEF));
291
292 const float16x8_t voutput_min = vreinterpretq_f16_u16(vld1q_dup_u16((const uint16_t*) &params->scalar.min));
293 vfp16out0x01234567 = vmaxq_f16(vfp16out0x01234567, voutput_min);
294 vfp16out0x89ABCDEF = vmaxq_f16(vfp16out0x89ABCDEF, voutput_min);
295 vfp16out1x01234567 = vmaxq_f16(vfp16out1x01234567, voutput_min);
296 vfp16out1x89ABCDEF = vmaxq_f16(vfp16out1x89ABCDEF, voutput_min);
297 vfp16out2x01234567 = vmaxq_f16(vfp16out2x01234567, voutput_min);
298 vfp16out2x89ABCDEF = vmaxq_f16(vfp16out2x89ABCDEF, voutput_min);
299 vfp16out3x01234567 = vmaxq_f16(vfp16out3x01234567, voutput_min);
300 vfp16out3x89ABCDEF = vmaxq_f16(vfp16out3x89ABCDEF, voutput_min);
301 const float16x8_t voutput_max = vreinterpretq_f16_u16(vld1q_dup_u16((const uint16_t*) &params->scalar.max));
302 vfp16out0x01234567 = vminq_f16(vfp16out0x01234567, voutput_max);
303 vfp16out0x89ABCDEF = vminq_f16(vfp16out0x89ABCDEF, voutput_max);
304 vfp16out1x01234567 = vminq_f16(vfp16out1x01234567, voutput_max);
305 vfp16out1x89ABCDEF = vminq_f16(vfp16out1x89ABCDEF, voutput_max);
306 vfp16out2x01234567 = vminq_f16(vfp16out2x01234567, voutput_max);
307 vfp16out2x89ABCDEF = vminq_f16(vfp16out2x89ABCDEF, voutput_max);
308 vfp16out3x01234567 = vminq_f16(vfp16out3x01234567, voutput_max);
309 vfp16out3x89ABCDEF = vminq_f16(vfp16out3x89ABCDEF, voutput_max);
310 if XNN_LIKELY(nc >= 16) {
311 vst1q_u16(c0, vreinterpretq_u16_f16(vfp16out0x01234567));
312 vst1q_u16(c0 + 8, vreinterpretq_u16_f16(vfp16out0x89ABCDEF));
313 vst1q_u16(c1, vreinterpretq_u16_f16(vfp16out1x01234567));
314 vst1q_u16(c1 + 8, vreinterpretq_u16_f16(vfp16out1x89ABCDEF));
315 vst1q_u16(c2, vreinterpretq_u16_f16(vfp16out2x01234567));
316 vst1q_u16(c2 + 8, vreinterpretq_u16_f16(vfp16out2x89ABCDEF));
317 vst1q_u16(c3, vreinterpretq_u16_f16(vfp16out3x01234567));
318 vst1q_u16(c3 + 8, vreinterpretq_u16_f16(vfp16out3x89ABCDEF));
319
320 a0 = (const int8_t*) ((uintptr_t) a0 - kc);
321 a1 = (const int8_t*) ((uintptr_t) a1 - kc);
322 a2 = (const int8_t*) ((uintptr_t) a2 - kc);
323 a3 = (const int8_t*) ((uintptr_t) a3 - kc);
324
325 c0 = (uint16_t*) ((uintptr_t) c0 + cn_stride);
326 c1 = (uint16_t*) ((uintptr_t) c1 + cn_stride);
327 c2 = (uint16_t*) ((uintptr_t) c2 + cn_stride);
328 c3 = (uint16_t*) ((uintptr_t) c3 + cn_stride);
329
330 nc -= 16;
331 } else {
332 if (nc & 8) {
333 vst1q_u16(c0, vreinterpretq_u16_f16(vfp16out0x01234567)); c0 += 8;
334 vfp16out0x01234567 = vfp16out0x89ABCDEF;
335 vst1q_u16(c1, vreinterpretq_u16_f16(vfp16out1x01234567)); c1 += 8;
336 vfp16out1x01234567 = vfp16out1x89ABCDEF;
337 vst1q_u16(c2, vreinterpretq_u16_f16(vfp16out2x01234567)); c2 += 8;
338 vfp16out2x01234567 = vfp16out2x89ABCDEF;
339 vst1q_u16(c3, vreinterpretq_u16_f16(vfp16out3x01234567)); c3 += 8;
340 vfp16out3x01234567 = vfp16out3x89ABCDEF;
341 }
342 float16x4_t vfp16out0x0123 = vget_low_f16(vfp16out0x01234567);
343 float16x4_t vfp16out1x0123 = vget_low_f16(vfp16out1x01234567);
344 float16x4_t vfp16out2x0123 = vget_low_f16(vfp16out2x01234567);
345 float16x4_t vfp16out3x0123 = vget_low_f16(vfp16out3x01234567);
346 if (nc & 4) {
347 vst1_u16(c0, vreinterpret_u16_f16(vfp16out0x0123)); c0 += 4;
348 vst1_u16(c1, vreinterpret_u16_f16(vfp16out1x0123)); c1 += 4;
349 vst1_u16(c2, vreinterpret_u16_f16(vfp16out2x0123)); c2 += 4;
350 vst1_u16(c3, vreinterpret_u16_f16(vfp16out3x0123)); c3 += 4;
351 vfp16out0x0123 = vget_high_f16(vfp16out0x01234567);
352 vfp16out1x0123 = vget_high_f16(vfp16out1x01234567);
353 vfp16out2x0123 = vget_high_f16(vfp16out2x01234567);
354 vfp16out3x0123 = vget_high_f16(vfp16out3x01234567);
355 }
356 if (nc & 2) {
357 vst1_lane_u32((void*) c0, vreinterpret_u32_f16(vfp16out0x0123), 0); c0 += 2;
358 vst1_lane_u32((void*) c1, vreinterpret_u32_f16(vfp16out1x0123), 0); c1 += 2;
359 vst1_lane_u32((void*) c2, vreinterpret_u32_f16(vfp16out2x0123), 0); c2 += 2;
360 vst1_lane_u32((void*) c3, vreinterpret_u32_f16(vfp16out3x0123), 0); c3 += 2;
361 vfp16out0x0123 = vext_f16(vfp16out0x0123, vfp16out0x0123, 2);
362 vfp16out1x0123 = vext_f16(vfp16out1x0123, vfp16out1x0123, 2);
363 vfp16out2x0123 = vext_f16(vfp16out2x0123, vfp16out2x0123, 2);
364 vfp16out3x0123 = vext_f16(vfp16out3x0123, vfp16out3x0123, 2);
365 }
366 if (nc & 1) {
367 vst1_lane_u16(c0, vreinterpret_u16_f16(vfp16out0x0123), 0);
368 vst1_lane_u16(c1, vreinterpret_u16_f16(vfp16out1x0123), 0);
369 vst1_lane_u16(c2, vreinterpret_u16_f16(vfp16out2x0123), 0);
370 vst1_lane_u16(c3, vreinterpret_u16_f16(vfp16out3x0123), 0);
371 }
372 nc = 0;
373 }
374 } while (nc != 0);
375 }