Debian Codegraph @

src/qd8-f16-qb4w-gemm/gen/qd8-f16-qb4w-gemm-6x8c4-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_6x8c4__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 <= 6);
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 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 int8_t* a5 = (const int8_t*) ((uintptr_t) a4 + a_stride);
68 uint16_t* c5 = (uint16_t*) ((uintptr_t) c4 + cm_stride);
69 if XNN_UNPREDICTABLE(mr != 6) {
70 a5 = a4;
71 c5 = c4;
72 }
73
74 size_t bl = params->scalar.blocksize;
75 assert(bl <= kc);
76 assert(bl != 0);
77 assert(bl % 32 == 0);
78 const int8x16_t vmask = vmovq_n_s8(INT8_C(0xF0));
79 // Loop over groups of 8 columns.
80 do {
81 // Initialize accumulators with bias. 8 bias values are loaded from the
82 // weight matrix, at the start of the group of 8 columns.
83 const float32x4_t vinput_zero_point01 = vcvtq_f32_s32(vld1q_s32(&quantization_params[0].zero_point));
84 const float32x4_t vksum0123 = vld1q_f32(w); w = (const float*) w + 4;
85 float32x4_t vout0x0123 = vmulq_lane_f32(vksum0123, vget_low_f32(vinput_zero_point01), 0);
86 float32x4_t vout1x0123 = vmulq_lane_f32(vksum0123, vget_high_f32(vinput_zero_point01), 0);
87 const float32x4_t vksum4567 = vld1q_f32(w); w = (const float*) w + 4;
88 float32x4_t vout0x4567 = vmulq_lane_f32(vksum4567, vget_low_f32(vinput_zero_point01), 0);
89 float32x4_t vout1x4567 = vmulq_lane_f32(vksum4567, vget_high_f32(vinput_zero_point01), 0);
90 const float32x4_t vinput_zero_point23 = vcvtq_f32_s32(vld1q_s32(&quantization_params[2].zero_point));
91 float32x4_t vout2x0123 = vmulq_lane_f32(vksum0123, vget_low_f32(vinput_zero_point23), 0);
92 float32x4_t vout3x0123 = vmulq_lane_f32(vksum0123, vget_high_f32(vinput_zero_point23), 0);
93 float32x4_t vout2x4567 = vmulq_lane_f32(vksum4567, vget_low_f32(vinput_zero_point23), 0);
94 float32x4_t vout3x4567 = vmulq_lane_f32(vksum4567, vget_high_f32(vinput_zero_point23), 0);
95 const float32x4_t vinput_zero_point45 = vcvtq_f32_s32(vld1q_s32(&quantization_params[4].zero_point));
96 float32x4_t vout4x0123 = vmulq_lane_f32(vksum0123, vget_low_f32(vinput_zero_point45), 0);
97 float32x4_t vout5x0123 = vmulq_lane_f32(vksum0123, vget_high_f32(vinput_zero_point45), 0);
98 float32x4_t vout4x4567 = vmulq_lane_f32(vksum4567, vget_low_f32(vinput_zero_point45), 0);
99 float32x4_t vout5x4567 = vmulq_lane_f32(vksum4567, vget_high_f32(vinput_zero_point45), 0);
100
101 for (size_t kb=0; kb < kc; kb += bl) {
102 int32x4_t vacc0x0123 = vdupq_n_s32(0);
103 int32x4_t vacc1x0123 = vdupq_n_s32(0);
104 int32x4_t vacc0x4567 = vdupq_n_s32(0);
105 int32x4_t vacc1x4567 = vdupq_n_s32(0);
106 int32x4_t vacc2x0123 = vdupq_n_s32(0);
107 int32x4_t vacc3x0123 = vdupq_n_s32(0);
108 int32x4_t vacc2x4567 = vdupq_n_s32(0);
109 int32x4_t vacc3x4567 = vdupq_n_s32(0);
110 int32x4_t vacc4x0123 = vdupq_n_s32(0);
111 int32x4_t vacc5x0123 = vdupq_n_s32(0);
112 int32x4_t vacc4x4567 = vdupq_n_s32(0);
113 int32x4_t vacc5x4567 = vdupq_n_s32(0);
114 // Inner accumulation loop along the 8 columns.
115 size_t k = bl;
116 // 2x partial unrolled loop to load 8 bytes at a time.
117 while (k >= 8 * sizeof(int8_t)) {
118 // Load a 6x8 block of activations.
119 const int8x8_t va0x01234567 = vld1_s8(a0); a0 += 8;
120 const int8x8_t va1x01234567 = vld1_s8(a1); a1 += 8;
121 const int8x8_t va2x01234567 = vld1_s8(a2); a2 += 8;
122 const int8x8_t va3x01234567 = vld1_s8(a3); a3 += 8;
123 const int8x8_t va4x01234567 = vld1_s8(a4); a4 += 8;
124 const int8x8_t va5x01234567 = vld1_s8(a5); a5 += 8;
125
126 // Load a 8x8 block of weights.
127 const int8x16_t vb01234567x0123 = vld1q_s8(w); w = (const int8_t*) w + 16;
128 const int8x16_t vb01234567x4567 = vld1q_s8(w); w = (const int8_t*) w + 16;
129 const int8x16_t vb0123x0123 = vshlq_n_s8(vb01234567x0123, 4);
130 const int8x16_t vb0123x4567 = vshlq_n_s8(vb01234567x4567, 4);
131 const int8x16_t vb4567x0123 = vandq_s8(vb01234567x0123, vmask);
132 const int8x16_t vb4567x4567 = vandq_s8(vb01234567x4567, vmask);
133
134 // Multiply-accumulate: 6x8 * 8x8 --> 6x8.
135 vacc0x0123 = vdotq_lane_s32(vacc0x0123, vb0123x0123, va0x01234567, 0);
136 vacc0x4567 = vdotq_lane_s32(vacc0x4567, vb0123x4567, va0x01234567, 0);
137 vacc1x0123 = vdotq_lane_s32(vacc1x0123, vb0123x0123, va1x01234567, 0);
138 vacc1x4567 = vdotq_lane_s32(vacc1x4567, vb0123x4567, va1x01234567, 0);
139 vacc2x0123 = vdotq_lane_s32(vacc2x0123, vb0123x0123, va2x01234567, 0);
140 vacc2x4567 = vdotq_lane_s32(vacc2x4567, vb0123x4567, va2x01234567, 0);
141 vacc3x0123 = vdotq_lane_s32(vacc3x0123, vb0123x0123, va3x01234567, 0);
142 vacc3x4567 = vdotq_lane_s32(vacc3x4567, vb0123x4567, va3x01234567, 0);
143 vacc4x0123 = vdotq_lane_s32(vacc4x0123, vb0123x0123, va4x01234567, 0);
144 vacc4x4567 = vdotq_lane_s32(vacc4x4567, vb0123x4567, va4x01234567, 0);
145 vacc5x0123 = vdotq_lane_s32(vacc5x0123, vb0123x0123, va5x01234567, 0);
146 vacc5x4567 = vdotq_lane_s32(vacc5x4567, vb0123x4567, va5x01234567, 0);
147 vacc0x0123 = vdotq_lane_s32(vacc0x0123, vb4567x0123, va0x01234567, 1);
148 vacc0x4567 = vdotq_lane_s32(vacc0x4567, vb4567x4567, va0x01234567, 1);
149 vacc1x0123 = vdotq_lane_s32(vacc1x0123, vb4567x0123, va1x01234567, 1);
150 vacc1x4567 = vdotq_lane_s32(vacc1x4567, vb4567x4567, va1x01234567, 1);
151 vacc2x0123 = vdotq_lane_s32(vacc2x0123, vb4567x0123, va2x01234567, 1);
152 vacc2x4567 = vdotq_lane_s32(vacc2x4567, vb4567x4567, va2x01234567, 1);
153 vacc3x0123 = vdotq_lane_s32(vacc3x0123, vb4567x0123, va3x01234567, 1);
154 vacc3x4567 = vdotq_lane_s32(vacc3x4567, vb4567x4567, va3x01234567, 1);
155 vacc4x0123 = vdotq_lane_s32(vacc4x0123, vb4567x0123, va4x01234567, 1);
156 vacc4x4567 = vdotq_lane_s32(vacc4x4567, vb4567x4567, va4x01234567, 1);
157 vacc5x0123 = vdotq_lane_s32(vacc5x0123, vb4567x0123, va5x01234567, 1);
158 vacc5x4567 = vdotq_lane_s32(vacc5x4567, vb4567x4567, va5x01234567, 1);
159
160 k -= 8 * sizeof(int8_t);
161 }
162 // Handle up to 4 final positions of `k`
163 if XNN_UNLIKELY(k != 0) {
164 // Load a 6x4 block of activations.
165 const int8x8_t va0x01234567 = vld1_s8(a0); a0 += 4;
166 const int8x8_t va1x01234567 = vld1_s8(a1); a1 += 4;
167 const int8x8_t va2x01234567 = vld1_s8(a2); a2 += 4;
168 const int8x8_t va3x01234567 = vld1_s8(a3); a3 += 4;
169 const int8x8_t va4x01234567 = vld1_s8(a4); a4 += 4;
170 const int8x8_t va5x01234567 = vld1_s8(a5); a5 += 4;
171
172 // Load a 4x8 block of weights.
173 const int8x16_t vb0123x0123 = vld1q_s8(w); w = (const int8_t*) w + 16;
174 const int8x16_t vb0123x4567 = vld1q_s8(w); w = (const int8_t*) w + 16;
175
176 // Multiply-accumulate: 6x4 * 4x8 --> 6x8.
177 vacc0x0123 = vdotq_lane_s32(vacc0x0123, vb0123x0123, va0x01234567, 0);
178 vacc0x4567 = vdotq_lane_s32(vacc0x4567, vb0123x4567, va0x01234567, 0);
179 vacc1x0123 = vdotq_lane_s32(vacc1x0123, vb0123x0123, va1x01234567, 0);
180 vacc1x4567 = vdotq_lane_s32(vacc1x4567, vb0123x4567, va1x01234567, 0);
181 vacc2x0123 = vdotq_lane_s32(vacc2x0123, vb0123x0123, va2x01234567, 0);
182 vacc2x4567 = vdotq_lane_s32(vacc2x4567, vb0123x4567, va2x01234567, 0);
183 vacc3x0123 = vdotq_lane_s32(vacc3x0123, vb0123x0123, va3x01234567, 0);
184 vacc3x4567 = vdotq_lane_s32(vacc3x4567, vb0123x4567, va3x01234567, 0);
185 vacc4x0123 = vdotq_lane_s32(vacc4x0123, vb0123x0123, va4x01234567, 0);
186 vacc4x4567 = vdotq_lane_s32(vacc4x4567, vb0123x4567, va4x01234567, 0);
187 vacc5x0123 = vdotq_lane_s32(vacc5x0123, vb0123x0123, va5x01234567, 0);
188 vacc5x4567 = vdotq_lane_s32(vacc5x4567, vb0123x4567, va5x01234567, 0);
189 }
190 const float32x4_t vfilter_output_scale0123 = vreinterpretq_f32_u32(vshll_n_u16(vld1_u16(w), 16)); w = (const uint16_t*) w + 4;
191 const float32x4_t vfilter_output_scale4567 = vreinterpretq_f32_u32(vshll_n_u16(vld1_u16(w), 16)); w = (const uint16_t*) w + 4;
192
193 float32x4_t vf0x0123 = vcvtq_f32_s32(vacc0x0123);
194 float32x4_t vf1x0123 = vcvtq_f32_s32(vacc1x0123);
195 vout0x0123 = vfmaq_f32(vout0x0123, vf0x0123, vfilter_output_scale0123);
196 vout1x0123 = vfmaq_f32(vout1x0123, vf1x0123, vfilter_output_scale0123);
197 float32x4_t vf0x4567 = vcvtq_f32_s32(vacc0x4567);
198 float32x4_t vf1x4567 = vcvtq_f32_s32(vacc1x4567);
199 vout0x4567 = vfmaq_f32(vout0x4567, vf0x4567, vfilter_output_scale4567);
200 vout1x4567 = vfmaq_f32(vout1x4567, vf1x4567, vfilter_output_scale4567);
201 float32x4_t vf2x0123 = vcvtq_f32_s32(vacc2x0123);
202 float32x4_t vf3x0123 = vcvtq_f32_s32(vacc3x0123);
203 vout2x0123 = vfmaq_f32(vout2x0123, vf2x0123, vfilter_output_scale0123);
204 vout3x0123 = vfmaq_f32(vout3x0123, vf3x0123, vfilter_output_scale0123);
205 float32x4_t vf2x4567 = vcvtq_f32_s32(vacc2x4567);
206 float32x4_t vf3x4567 = vcvtq_f32_s32(vacc3x4567);
207 vout2x4567 = vfmaq_f32(vout2x4567, vf2x4567, vfilter_output_scale4567);
208 vout3x4567 = vfmaq_f32(vout3x4567, vf3x4567, vfilter_output_scale4567);
209 float32x4_t vf4x0123 = vcvtq_f32_s32(vacc4x0123);
210 float32x4_t vf5x0123 = vcvtq_f32_s32(vacc5x0123);
211 vout4x0123 = vfmaq_f32(vout4x0123, vf4x0123, vfilter_output_scale0123);
212 vout5x0123 = vfmaq_f32(vout5x0123, vf5x0123, vfilter_output_scale0123);
213 float32x4_t vf4x4567 = vcvtq_f32_s32(vacc4x4567);
214 float32x4_t vf5x4567 = vcvtq_f32_s32(vacc5x4567);
215 vout4x4567 = vfmaq_f32(vout4x4567, vf4x4567, vfilter_output_scale4567);
216 vout5x4567 = vfmaq_f32(vout5x4567, vf5x4567, vfilter_output_scale4567);
217 }
218
219 const float32x4_t vinput_scale01 = vreinterpretq_f32_s32(vld1q_s32(&quantization_params[0].zero_point));
220 vout0x0123 = vmulq_lane_f32(vout0x0123, vget_low_f32(vinput_scale01), 1);
221 vout1x0123 = vmulq_lane_f32(vout1x0123, vget_high_f32(vinput_scale01), 1);
222 vout0x4567 = vmulq_lane_f32(vout0x4567, vget_low_f32(vinput_scale01), 1);
223 vout1x4567 = vmulq_lane_f32(vout1x4567, vget_high_f32(vinput_scale01), 1);
224 const float32x4_t vinput_scale23 = vreinterpretq_f32_s32(vld1q_s32(&quantization_params[2].zero_point));
225 vout2x0123 = vmulq_lane_f32(vout2x0123, vget_low_f32(vinput_scale23), 1);
226 vout3x0123 = vmulq_lane_f32(vout3x0123, vget_high_f32(vinput_scale23), 1);
227 vout2x4567 = vmulq_lane_f32(vout2x4567, vget_low_f32(vinput_scale23), 1);
228 vout3x4567 = vmulq_lane_f32(vout3x4567, vget_high_f32(vinput_scale23), 1);
229 const float32x4_t vinput_scale45 = vreinterpretq_f32_s32(vld1q_s32(&quantization_params[4].zero_point));
230 vout4x0123 = vmulq_lane_f32(vout4x0123, vget_low_f32(vinput_scale45), 1);
231 vout5x0123 = vmulq_lane_f32(vout5x0123, vget_high_f32(vinput_scale45), 1);
232 vout4x4567 = vmulq_lane_f32(vout4x4567, vget_low_f32(vinput_scale45), 1);
233 vout5x4567 = vmulq_lane_f32(vout5x4567, vget_high_f32(vinput_scale45), 1);
234
235
236 const float32x4_t vbias0123 = vld1q_f32(w); w = (const float*) w + 4;
237 vout0x0123 = vaddq_f32(vbias0123, vout0x0123);
238 vout1x0123 = vaddq_f32(vbias0123, vout1x0123);
239 vout2x0123 = vaddq_f32(vbias0123, vout2x0123);
240 vout3x0123 = vaddq_f32(vbias0123, vout3x0123);
241 vout4x0123 = vaddq_f32(vbias0123, vout4x0123);
242 vout5x0123 = vaddq_f32(vbias0123, vout5x0123);
243 const float32x4_t vbias4567 = vld1q_f32(w); w = (const float*) w + 4;
244 vout0x4567 = vaddq_f32(vbias4567, vout0x4567);
245 vout1x4567 = vaddq_f32(vbias4567, vout1x4567);
246 vout2x4567 = vaddq_f32(vbias4567, vout2x4567);
247 vout3x4567 = vaddq_f32(vbias4567, vout3x4567);
248 vout4x4567 = vaddq_f32(vbias4567, vout4x4567);
249 vout5x4567 = vaddq_f32(vbias4567, vout5x4567);
250
251 float16x8_t vfp16out0x01234567 = vcombine_f16(vcvt_f16_f32(vout0x0123), vcvt_f16_f32(vout0x4567));
252 float16x8_t vfp16out1x01234567 = vcombine_f16(vcvt_f16_f32(vout1x0123), vcvt_f16_f32(vout1x4567));
253 float16x8_t vfp16out2x01234567 = vcombine_f16(vcvt_f16_f32(vout2x0123), vcvt_f16_f32(vout2x4567));
254 float16x8_t vfp16out3x01234567 = vcombine_f16(vcvt_f16_f32(vout3x0123), vcvt_f16_f32(vout3x4567));
255 float16x8_t vfp16out4x01234567 = vcombine_f16(vcvt_f16_f32(vout4x0123), vcvt_f16_f32(vout4x4567));
256 float16x8_t vfp16out5x01234567 = vcombine_f16(vcvt_f16_f32(vout5x0123), vcvt_f16_f32(vout5x4567));
257
258 const float16x8_t voutput_min = vreinterpretq_f16_u16(vld1q_dup_u16((const uint16_t*) &params->scalar.min));
259 vfp16out0x01234567 = vmaxq_f16(vfp16out0x01234567, voutput_min);
260 vfp16out1x01234567 = vmaxq_f16(vfp16out1x01234567, voutput_min);
261 vfp16out2x01234567 = vmaxq_f16(vfp16out2x01234567, voutput_min);
262 vfp16out3x01234567 = vmaxq_f16(vfp16out3x01234567, voutput_min);
263 vfp16out4x01234567 = vmaxq_f16(vfp16out4x01234567, voutput_min);
264 vfp16out5x01234567 = vmaxq_f16(vfp16out5x01234567, voutput_min);
265 const float16x8_t voutput_max = vreinterpretq_f16_u16(vld1q_dup_u16((const uint16_t*) &params->scalar.max));
266 vfp16out0x01234567 = vminq_f16(vfp16out0x01234567, voutput_max);
267 vfp16out1x01234567 = vminq_f16(vfp16out1x01234567, voutput_max);
268 vfp16out2x01234567 = vminq_f16(vfp16out2x01234567, voutput_max);
269 vfp16out3x01234567 = vminq_f16(vfp16out3x01234567, voutput_max);
270 vfp16out4x01234567 = vminq_f16(vfp16out4x01234567, voutput_max);
271 vfp16out5x01234567 = vminq_f16(vfp16out5x01234567, voutput_max);
272 if XNN_LIKELY(nc >= 8) {
273 vst1q_u16(c0, vreinterpretq_u16_f16(vfp16out0x01234567));
274 vst1q_u16(c1, vreinterpretq_u16_f16(vfp16out1x01234567));
275 vst1q_u16(c2, vreinterpretq_u16_f16(vfp16out2x01234567));
276 vst1q_u16(c3, vreinterpretq_u16_f16(vfp16out3x01234567));
277 vst1q_u16(c4, vreinterpretq_u16_f16(vfp16out4x01234567));
278 vst1q_u16(c5, vreinterpretq_u16_f16(vfp16out5x01234567));
279
280 a0 = (const int8_t*) ((uintptr_t) a0 - kc);
281 a1 = (const int8_t*) ((uintptr_t) a1 - kc);
282 a2 = (const int8_t*) ((uintptr_t) a2 - kc);
283 a3 = (const int8_t*) ((uintptr_t) a3 - kc);
284 a4 = (const int8_t*) ((uintptr_t) a4 - kc);
285 a5 = (const int8_t*) ((uintptr_t) a5 - kc);
286
287 c0 = (uint16_t*) ((uintptr_t) c0 + cn_stride);
288 c1 = (uint16_t*) ((uintptr_t) c1 + cn_stride);
289 c2 = (uint16_t*) ((uintptr_t) c2 + cn_stride);
290 c3 = (uint16_t*) ((uintptr_t) c3 + cn_stride);
291 c4 = (uint16_t*) ((uintptr_t) c4 + cn_stride);
292 c5 = (uint16_t*) ((uintptr_t) c5 + cn_stride);
293
294 nc -= 8;
295 } else {
296 float16x4_t vfp16out0x0123 = vget_low_f16(vfp16out0x01234567);
297 float16x4_t vfp16out1x0123 = vget_low_f16(vfp16out1x01234567);
298 float16x4_t vfp16out2x0123 = vget_low_f16(vfp16out2x01234567);
299 float16x4_t vfp16out3x0123 = vget_low_f16(vfp16out3x01234567);
300 float16x4_t vfp16out4x0123 = vget_low_f16(vfp16out4x01234567);
301 float16x4_t vfp16out5x0123 = vget_low_f16(vfp16out5x01234567);
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 vst1_u16(c5, vreinterpret_u16_f16(vfp16out5x0123)); c5 += 4;
309 vfp16out0x0123 = vget_high_f16(vfp16out0x01234567);
310 vfp16out1x0123 = vget_high_f16(vfp16out1x01234567);
311 vfp16out2x0123 = vget_high_f16(vfp16out2x01234567);
312 vfp16out3x0123 = vget_high_f16(vfp16out3x01234567);
313 vfp16out4x0123 = vget_high_f16(vfp16out4x01234567);
314 vfp16out5x0123 = vget_high_f16(vfp16out5x01234567);
315 }
316 if (nc & 2) {
317 vst1_lane_u32((void*) c0, vreinterpret_u32_f16(vfp16out0x0123), 0); c0 += 2;
318 vst1_lane_u32((void*) c1, vreinterpret_u32_f16(vfp16out1x0123), 0); c1 += 2;
319 vst1_lane_u32((void*) c2, vreinterpret_u32_f16(vfp16out2x0123), 0); c2 += 2;
320 vst1_lane_u32((void*) c3, vreinterpret_u32_f16(vfp16out3x0123), 0); c3 += 2;
321 vst1_lane_u32((void*) c4, vreinterpret_u32_f16(vfp16out4x0123), 0); c4 += 2;
322 vst1_lane_u32((void*) c5, vreinterpret_u32_f16(vfp16out5x0123), 0); c5 += 2;
323 vfp16out0x0123 = vext_f16(vfp16out0x0123, vfp16out0x0123, 2);
324 vfp16out1x0123 = vext_f16(vfp16out1x0123, vfp16out1x0123, 2);
325 vfp16out2x0123 = vext_f16(vfp16out2x0123, vfp16out2x0123, 2);
326 vfp16out3x0123 = vext_f16(vfp16out3x0123, vfp16out3x0123, 2);
327 vfp16out4x0123 = vext_f16(vfp16out4x0123, vfp16out4x0123, 2);
328 vfp16out5x0123 = vext_f16(vfp16out5x0123, vfp16out5x0123, 2);
329 }
330 if (nc & 1) {
331 vst1_lane_u16(c0, vreinterpret_u16_f16(vfp16out0x0123), 0);
332 vst1_lane_u16(c1, vreinterpret_u16_f16(vfp16out1x0123), 0);
333 vst1_lane_u16(c2, vreinterpret_u16_f16(vfp16out2x0123), 0);
334 vst1_lane_u16(c3, vreinterpret_u16_f16(vfp16out3x0123), 0);
335 vst1_lane_u16(c4, vreinterpret_u16_f16(vfp16out4x0123), 0);
336 vst1_lane_u16(c5, vreinterpret_u16_f16(vfp16out5x0123), 0);
337 }
338 nc = 0;
339 }
340 } while (nc != 0);
341 }