Debian Codegraph @

src/qd8-f32-qb4w-gemm/gen/qd8-f32-qb4w-gemm-1x16-minmax-neon-mlal-lane-prfm.c

raw
1 // Auto-generated file. Do not edit!
2 // Template: src/qs8-gemm/neon-mlal-lane.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/common.h"
15 #include "xnnpack/gemm.h"
16 #include "xnnpack/math.h"
17 #include "xnnpack/prefetch.h"
18
19
20 void xnn_qd8_f32_qb4w_gemm_minmax_ukernel_1x16__neon_mlal_lane_prfm(
21 size_t mr,
22 size_t nc,
23 size_t kc,
24 const int8_t* restrict a,
25 size_t a_stride,
26 const void* restrict w,
27 float* restrict c,
28 size_t cm_stride,
29 size_t cn_stride,
30 const struct xnn_f32_qb4w_minmax_params params[restrict XNN_MIN_ELEMENTS(1)],
31 const struct xnn_qd8_quantization_params quantization_params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
32 {
33 assert(mr != 0);
34 assert(mr <= 1);
35 assert(nc != 0);
36 assert(kc != 0);
37 assert(kc % sizeof(int8_t) == 0);
38 assert(a != NULL);
39 assert(w != NULL);
40 assert(c != NULL);
41
42 size_t bl = params->scalar.blocksize;
43 assert(bl <= round_up_po2(kc, 2));
44 assert(bl != 0);
45 assert(bl % 32 == 0);
46 const int8_t* a0 = a;
47 float* c0 = c;
48
49 const int8x8_t vmask = vmov_n_s8(INT8_C(0xF0));
50 kc = round_up_po2(kc, 2);
51 do {
52 float32x4_t vksum0123 = vld1q_f32(w); w = (const float*) w + 4;
53 float32x4_t vksum4567 = vld1q_f32(w); w = (const float*) w + 4;
54 float32x4_t vksum89AB = vld1q_f32(w); w = (const float*) w + 4;
55 float32x4_t vksumCDEF = vld1q_f32(w); w = (const float*) w + 4;
56 const float32x4_t vzp0 = vcvtq_f32_s32(vld1q_dup_s32(&quantization_params[0].zero_point));
57 float32x4_t vout0x0123 = vmulq_f32(vksum0123, vzp0);
58 float32x4_t vout0x4567 = vmulq_f32(vksum4567, vzp0);
59 float32x4_t vout0x89AB = vmulq_f32(vksum89AB, vzp0);
60 float32x4_t vout0xCDEF = vmulq_f32(vksumCDEF, vzp0);
61
62 for (size_t kb=0; kb < kc; kb += bl) {
63 int32x4_t vacc0x0123 = vdupq_n_s32(0);
64 int32x4_t vacc0x4567 = vdupq_n_s32(0);
65 int32x4_t vacc0x89AB = vdupq_n_s32(0);
66 int32x4_t vacc0xCDEF = vdupq_n_s32(0);
67
68 size_t k = bl;
69 while (k >= 8 * sizeof(int8_t)) {
70 const int8x8_t va0 = vld1_s8(a0); a0 += 8;
71 const int16x8_t vxa0 = vmovl_s8(va0);
72
73 const int8x8_t vb01234567c01 = vld1_s8(w); w = (const uint8_t*) w + 8;
74 const int8x8_t vb89ABCDEFc01 = vld1_s8(w); w = (const uint8_t*) w + 8;
75 const int8x8_t vb01234567c23 = vld1_s8(w); w = (const uint8_t*) w + 8;
76 const int8x8_t vb89ABCDEFc23 = vld1_s8(w); w = (const uint8_t*) w + 8;
77 const int8x8_t vb01234567c45 = vld1_s8(w); w = (const uint8_t*) w + 8;
78 const int8x8_t vb89ABCDEFc45 = vld1_s8(w); w = (const uint8_t*) w + 8;
79 const int8x8_t vb01234567c67 = vld1_s8(w); w = (const uint8_t*) w + 8;
80 const int8x8_t vb89ABCDEFc67 = vld1_s8(w); w = (const uint8_t*) w + 8;
81 const int8x8_t vb01234567c0 = vshl_n_s8(vb01234567c01, 4);
82 const int8x8_t vb01234567c1 = vand_s8(vb01234567c01, vmask);
83 const int8x8_t vb89ABCDEFc0 = vshl_n_s8(vb89ABCDEFc01, 4);
84 const int8x8_t vb89ABCDEFc1 = vand_s8(vb89ABCDEFc01, vmask);
85 const int8x8_t vb01234567c2 = vshl_n_s8(vb01234567c23, 4);
86 const int8x8_t vb01234567c3 = vand_s8(vb01234567c23, vmask);
87 const int8x8_t vb89ABCDEFc2 = vshl_n_s8(vb89ABCDEFc23, 4);
88 const int8x8_t vb89ABCDEFc3 = vand_s8(vb89ABCDEFc23, vmask);
89 const int8x8_t vb01234567c4 = vshl_n_s8(vb01234567c45, 4);
90 const int8x8_t vb01234567c5 = vand_s8(vb01234567c45, vmask);
91 const int8x8_t vb89ABCDEFc4 = vshl_n_s8(vb89ABCDEFc45, 4);
92 const int8x8_t vb89ABCDEFc5 = vand_s8(vb89ABCDEFc45, vmask);
93 const int8x8_t vb01234567c6 = vshl_n_s8(vb01234567c67, 4);
94 const int8x8_t vb01234567c7 = vand_s8(vb01234567c67, vmask);
95 const int8x8_t vb89ABCDEFc6 = vshl_n_s8(vb89ABCDEFc67, 4);
96 const int8x8_t vb89ABCDEFc7 = vand_s8(vb89ABCDEFc67, vmask);
97
98 const int16x8_t vxb01234567c0 = vmovl_s8(vb01234567c0);
99 const int16x8_t vxb89ABCDEFc0 = vmovl_s8(vb89ABCDEFc0);
100 const int16x8_t vxb01234567c1 = vmovl_s8(vb01234567c1);
101 const int16x8_t vxb89ABCDEFc1 = vmovl_s8(vb89ABCDEFc1);
102 const int16x8_t vxb01234567c2 = vmovl_s8(vb01234567c2);
103 const int16x8_t vxb89ABCDEFc2 = vmovl_s8(vb89ABCDEFc2);
104 const int16x8_t vxb01234567c3 = vmovl_s8(vb01234567c3);
105 const int16x8_t vxb89ABCDEFc3 = vmovl_s8(vb89ABCDEFc3);
106 const int16x8_t vxb01234567c4 = vmovl_s8(vb01234567c4);
107 const int16x8_t vxb89ABCDEFc4 = vmovl_s8(vb89ABCDEFc4);
108 const int16x8_t vxb01234567c5 = vmovl_s8(vb01234567c5);
109 const int16x8_t vxb89ABCDEFc5 = vmovl_s8(vb89ABCDEFc5);
110 const int16x8_t vxb01234567c6 = vmovl_s8(vb01234567c6);
111 const int16x8_t vxb89ABCDEFc6 = vmovl_s8(vb89ABCDEFc6);
112 const int16x8_t vxb01234567c7 = vmovl_s8(vb01234567c7);
113 const int16x8_t vxb89ABCDEFc7 = vmovl_s8(vb89ABCDEFc7);
114
115 xnn_prefetch_to_l1((const int8_t*) w + 448);
116 xnn_prefetch_to_l1((const int8_t*) w + 512);
117
118 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c0), vget_low_s16(vxa0), 0);
119 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c0), vget_low_s16(vxa0), 0);
120 vacc0x89AB = vmlal_lane_s16(vacc0x89AB, vget_low_s16(vxb89ABCDEFc0), vget_low_s16(vxa0), 0);
121 vacc0xCDEF = vmlal_lane_s16(vacc0xCDEF, vget_high_s16(vxb89ABCDEFc0), vget_low_s16(vxa0), 0);
122 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c1), vget_low_s16(vxa0), 1);
123 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c1), vget_low_s16(vxa0), 1);
124 vacc0x89AB = vmlal_lane_s16(vacc0x89AB, vget_low_s16(vxb89ABCDEFc1), vget_low_s16(vxa0), 1);
125 vacc0xCDEF = vmlal_lane_s16(vacc0xCDEF, vget_high_s16(vxb89ABCDEFc1), vget_low_s16(vxa0), 1);
126 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c2), vget_low_s16(vxa0), 2);
127 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c2), vget_low_s16(vxa0), 2);
128 vacc0x89AB = vmlal_lane_s16(vacc0x89AB, vget_low_s16(vxb89ABCDEFc2), vget_low_s16(vxa0), 2);
129 vacc0xCDEF = vmlal_lane_s16(vacc0xCDEF, vget_high_s16(vxb89ABCDEFc2), vget_low_s16(vxa0), 2);
130 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c3), vget_low_s16(vxa0), 3);
131 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c3), vget_low_s16(vxa0), 3);
132 vacc0x89AB = vmlal_lane_s16(vacc0x89AB, vget_low_s16(vxb89ABCDEFc3), vget_low_s16(vxa0), 3);
133 vacc0xCDEF = vmlal_lane_s16(vacc0xCDEF, vget_high_s16(vxb89ABCDEFc3), vget_low_s16(vxa0), 3);
134
135 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c4), vget_high_s16(vxa0), 0);
136 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c4), vget_high_s16(vxa0), 0);
137 vacc0x89AB = vmlal_lane_s16(vacc0x89AB, vget_low_s16(vxb89ABCDEFc4), vget_high_s16(vxa0), 0);
138 vacc0xCDEF = vmlal_lane_s16(vacc0xCDEF, vget_high_s16(vxb89ABCDEFc4), vget_high_s16(vxa0), 0);
139 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c5), vget_high_s16(vxa0), 1);
140 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c5), vget_high_s16(vxa0), 1);
141 vacc0x89AB = vmlal_lane_s16(vacc0x89AB, vget_low_s16(vxb89ABCDEFc5), vget_high_s16(vxa0), 1);
142 vacc0xCDEF = vmlal_lane_s16(vacc0xCDEF, vget_high_s16(vxb89ABCDEFc5), vget_high_s16(vxa0), 1);
143 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c6), vget_high_s16(vxa0), 2);
144 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c6), vget_high_s16(vxa0), 2);
145 vacc0x89AB = vmlal_lane_s16(vacc0x89AB, vget_low_s16(vxb89ABCDEFc6), vget_high_s16(vxa0), 2);
146 vacc0xCDEF = vmlal_lane_s16(vacc0xCDEF, vget_high_s16(vxb89ABCDEFc6), vget_high_s16(vxa0), 2);
147 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c7), vget_high_s16(vxa0), 3);
148 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c7), vget_high_s16(vxa0), 3);
149 vacc0x89AB = vmlal_lane_s16(vacc0x89AB, vget_low_s16(vxb89ABCDEFc7), vget_high_s16(vxa0), 3);
150 vacc0xCDEF = vmlal_lane_s16(vacc0xCDEF, vget_high_s16(vxb89ABCDEFc7), vget_high_s16(vxa0), 3);
151
152 k -= 8 * sizeof(int8_t);
153 }
154 if XNN_UNLIKELY(k != 0) {
155 const int8x8_t va0 = vld1_s8(a0); a0 = (const int8_t*) ((uintptr_t) a0 + k);
156 const int16x8_t vxa0 = vmovl_s8(va0);
157
158 const int8x8_t vb01234567c01 = vld1_s8(w); w = (const int8_t*) w + 8;
159 const int8x8_t vb89ABCDEFc01 = vld1_s8(w); w = (const int8_t*) w + 8;
160 const int8x8_t vb01234567c0 = vshl_n_s8(vb01234567c01, 4);
161 const int8x8_t vb01234567c1 = vand_s8(vb01234567c01, vmask);
162 const int8x8_t vb89ABCDEFc0 = vshl_n_s8(vb89ABCDEFc01, 4);
163 const int8x8_t vb89ABCDEFc1 = vand_s8(vb89ABCDEFc01, vmask);
164
165 const int16x8_t vxb01234567c0 = vmovl_s8(vb01234567c0);
166 const int16x8_t vxb89ABCDEFc0 = vmovl_s8(vb89ABCDEFc0);
167
168 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c0), vget_low_s16(vxa0), 0);
169 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c0), vget_low_s16(vxa0), 0);
170 vacc0x89AB = vmlal_lane_s16(vacc0x89AB, vget_low_s16(vxb89ABCDEFc0), vget_low_s16(vxa0), 0);
171 vacc0xCDEF = vmlal_lane_s16(vacc0xCDEF, vget_high_s16(vxb89ABCDEFc0), vget_low_s16(vxa0), 0);
172
173 if (k >= 2 * sizeof(int8_t)) {
174 const int16x8_t vxb01234567c1 = vmovl_s8(vb01234567c1);
175 const int16x8_t vxb89ABCDEFc1 = vmovl_s8(vb89ABCDEFc1);
176
177 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c1), vget_low_s16(vxa0), 1);
178 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c1), vget_low_s16(vxa0), 1);
179 vacc0x89AB = vmlal_lane_s16(vacc0x89AB, vget_low_s16(vxb89ABCDEFc1), vget_low_s16(vxa0), 1);
180 vacc0xCDEF = vmlal_lane_s16(vacc0xCDEF, vget_high_s16(vxb89ABCDEFc1), vget_low_s16(vxa0), 1);
181
182 if (k > 2 * sizeof(int8_t)) {
183 const int8x8_t vb01234567c23 = vld1_s8(w); w = (const int8_t*) w + 8;
184 const int8x8_t vb89ABCDEFc23 = vld1_s8(w); w = (const int8_t*) w + 8;
185 const int8x8_t vb01234567c2 = vshl_n_s8(vb01234567c23, 4);
186 const int8x8_t vb01234567c3 = vand_s8(vb01234567c23, vmask);
187 const int8x8_t vb89ABCDEFc2 = vshl_n_s8(vb89ABCDEFc23, 4);
188 const int8x8_t vb89ABCDEFc3 = vand_s8(vb89ABCDEFc23, vmask);
189 const int16x8_t vxb01234567c2 = vmovl_s8(vb01234567c2);
190 const int16x8_t vxb89ABCDEFc2 = vmovl_s8(vb89ABCDEFc2);
191
192 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c2), vget_low_s16(vxa0), 2);
193 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c2), vget_low_s16(vxa0), 2);
194 vacc0x89AB = vmlal_lane_s16(vacc0x89AB, vget_low_s16(vxb89ABCDEFc2), vget_low_s16(vxa0), 2);
195 vacc0xCDEF = vmlal_lane_s16(vacc0xCDEF, vget_high_s16(vxb89ABCDEFc2), vget_low_s16(vxa0), 2);
196
197 if (k >= 4 * sizeof(int8_t)) {
198 const int16x8_t vxb01234567c3 = vmovl_s8(vb01234567c3);
199 const int16x8_t vxb89ABCDEFc3 = vmovl_s8(vb89ABCDEFc3);
200
201 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c3), vget_low_s16(vxa0), 3);
202 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c3), vget_low_s16(vxa0), 3);
203 vacc0x89AB = vmlal_lane_s16(vacc0x89AB, vget_low_s16(vxb89ABCDEFc3), vget_low_s16(vxa0), 3);
204 vacc0xCDEF = vmlal_lane_s16(vacc0xCDEF, vget_high_s16(vxb89ABCDEFc3), vget_low_s16(vxa0), 3);
205
206 if (k > 4 * sizeof(int8_t)) {
207 const int8x8_t vb01234567c45 = vld1_s8(w); w = (const int8_t*) w + 8;
208 const int8x8_t vb89ABCDEFc45 = vld1_s8(w); w = (const int8_t*) w + 8;
209 const int8x8_t vb01234567c4 = vshl_n_s8(vb01234567c45, 4);
210 const int8x8_t vb01234567c5 = vand_s8(vb01234567c45, vmask);
211 const int8x8_t vb89ABCDEFc4 = vshl_n_s8(vb89ABCDEFc45, 4);
212 const int8x8_t vb89ABCDEFc5 = vand_s8(vb89ABCDEFc45, vmask);
213 const int16x8_t vxb01234567c4 = vmovl_s8(vb01234567c4);
214 const int16x8_t vxb89ABCDEFc4 = vmovl_s8(vb89ABCDEFc4);
215
216 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c4), vget_high_s16(vxa0), 0);
217 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c4), vget_high_s16(vxa0), 0);
218 vacc0x89AB = vmlal_lane_s16(vacc0x89AB, vget_low_s16(vxb89ABCDEFc4), vget_high_s16(vxa0), 0);
219 vacc0xCDEF = vmlal_lane_s16(vacc0xCDEF, vget_high_s16(vxb89ABCDEFc4), vget_high_s16(vxa0), 0);
220
221 if (k >= 6 * sizeof(int8_t)) {
222 const int16x8_t vxb01234567c5 = vmovl_s8(vb01234567c5);
223 const int16x8_t vxb89ABCDEFc5 = vmovl_s8(vb89ABCDEFc5);
224
225 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c5), vget_high_s16(vxa0), 1);
226 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c5), vget_high_s16(vxa0), 1);
227 vacc0x89AB = vmlal_lane_s16(vacc0x89AB, vget_low_s16(vxb89ABCDEFc5), vget_high_s16(vxa0), 1);
228 vacc0xCDEF = vmlal_lane_s16(vacc0xCDEF, vget_high_s16(vxb89ABCDEFc5), vget_high_s16(vxa0), 1);
229
230 }
231 }
232 }
233 }
234 }
235 }
236
237 const float32x4_t vfilter_output_scale0123 = vreinterpretq_f32_u32(vshll_n_u16(vld1_u16(w), 16)); w = (const uint16_t*) w + 4;
238 const float32x4_t vfilter_output_scale4567 = vreinterpretq_f32_u32(vshll_n_u16(vld1_u16(w), 16)); w = (const uint16_t*) w + 4;
239 const float32x4_t vfilter_output_scale89AB = vreinterpretq_f32_u32(vshll_n_u16(vld1_u16(w), 16)); w = (const uint16_t*) w + 4;
240 const float32x4_t vfilter_output_scaleCDEF = vreinterpretq_f32_u32(vshll_n_u16(vld1_u16(w), 16)); w = (const uint16_t*) w + 4;
241
242 const float32x4_t vf0x0123 = vcvtq_f32_s32(vacc0x0123);
243 const float32x4_t vf0x4567 = vcvtq_f32_s32(vacc0x4567);
244 const float32x4_t vf0x89AB = vcvtq_f32_s32(vacc0x89AB);
245 const float32x4_t vf0xCDEF = vcvtq_f32_s32(vacc0xCDEF);
246
247 #if XNN_ARCH_ARM64
248 vout0x0123 = vfmaq_f32(vout0x0123, vf0x0123, vfilter_output_scale0123);
249 vout0x4567 = vfmaq_f32(vout0x4567, vf0x4567, vfilter_output_scale4567);
250 vout0x89AB = vfmaq_f32(vout0x89AB, vf0x89AB, vfilter_output_scale89AB);
251 vout0xCDEF = vfmaq_f32(vout0xCDEF, vf0xCDEF, vfilter_output_scaleCDEF);
252 #else
253 vout0x0123 = vmlaq_f32(vout0x0123, vf0x0123, vfilter_output_scale0123);
254 vout0x4567 = vmlaq_f32(vout0x4567, vf0x4567, vfilter_output_scale4567);
255 vout0x89AB = vmlaq_f32(vout0x89AB, vf0x89AB, vfilter_output_scale89AB);
256 vout0xCDEF = vmlaq_f32(vout0xCDEF, vf0xCDEF, vfilter_output_scaleCDEF);
257 #endif
258 }
259 const float32x4_t vinput_scale0 = vld1q_dup_f32(&quantization_params[0].inv_scale);
260
261 const float32x4_t vbias0123 = vld1q_f32(w); w = (const float*) w + 4;
262
263 #if XNN_ARCH_ARM64
264 vout0x0123 = vfmaq_f32(vbias0123, vout0x0123, vinput_scale0);
265 #else
266 vout0x0123 = vmlaq_f32(vbias0123, vout0x0123, vinput_scale0);
267 #endif
268 const float32x4_t vbias4567 = vld1q_f32(w); w = (const float*) w + 4;
269
270 #if XNN_ARCH_ARM64
271 vout0x4567 = vfmaq_f32(vbias4567, vout0x4567, vinput_scale0);
272 #else
273 vout0x4567 = vmlaq_f32(vbias4567, vout0x4567, vinput_scale0);
274 #endif
275 const float32x4_t vbias89AB = vld1q_f32(w); w = (const float*) w + 4;
276
277 #if XNN_ARCH_ARM64
278 vout0x89AB = vfmaq_f32(vbias89AB, vout0x89AB, vinput_scale0);
279 #else
280 vout0x89AB = vmlaq_f32(vbias89AB, vout0x89AB, vinput_scale0);
281 #endif
282 const float32x4_t vbiasCDEF = vld1q_f32(w); w = (const float*) w + 4;
283
284 #if XNN_ARCH_ARM64
285 vout0xCDEF = vfmaq_f32(vbiasCDEF, vout0xCDEF, vinput_scale0);
286 #else
287 vout0xCDEF = vmlaq_f32(vbiasCDEF, vout0xCDEF, vinput_scale0);
288 #endif
289
290 const float32x4_t voutput_min = vld1q_dup_f32(&params->scalar.min);
291 vout0x0123 = vmaxq_f32(vout0x0123, voutput_min);
292 vout0x4567 = vmaxq_f32(vout0x4567, voutput_min);
293 vout0x89AB = vmaxq_f32(vout0x89AB, voutput_min);
294 vout0xCDEF = vmaxq_f32(vout0xCDEF, voutput_min);
295
296 const float32x4_t voutput_max = vld1q_dup_f32(&params->scalar.max);
297 vout0x0123 = vminq_f32(vout0x0123, voutput_max);
298 vout0x4567 = vminq_f32(vout0x4567, voutput_max);
299 vout0x89AB = vminq_f32(vout0x89AB, voutput_max);
300 vout0xCDEF = vminq_f32(vout0xCDEF, voutput_max);
301
302 if XNN_LIKELY(nc >= 16) {
303 vst1q_f32(&c0[0], vout0x0123);
304 vst1q_f32(&c0[4], vout0x4567);
305 vst1q_f32(&c0[8], vout0x89AB);
306 vst1q_f32(&c0[12], vout0xCDEF);
307
308 a0 = (const int8_t*) ((uintptr_t) a0 - kc);
309
310 c0 = (float*) ((uintptr_t) c0 + cn_stride);
311
312 nc -= 16;
313 } else {
314 if (nc & 8) {
315 vst1q_f32(c0, vout0x0123); c0 += 4;
316 vout0x0123 = vout0x89AB;
317 vst1q_f32(c0, vout0x4567); c0 += 4;
318 vout0x4567 = vout0xCDEF;
319 }
320 if (nc & 4) {
321 vst1q_f32(c0, vout0x0123); c0 += 4;
322 vout0x0123 = vout0x4567;
323 }
324 float32x2_t vout0x01 = vget_low_f32(vout0x0123);
325 if (nc & 2) {
326 vst1_f32(c0, vout0x01); c0 += 2;
327 vout0x01 = vget_high_f32(vout0x0123);
328 }
329 if (nc & 1) {
330 vst1_lane_f32(c0, vout0x01, 0);
331 }
332 nc = 0;
333 }
334 } while (nc != 0);
335 }