Debian Codegraph @

src/qd8-f32-qc8w-gemm/gen/qd8-f32-qc8w-gemm-3x8-minmax-neon-mlal-lane.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
17
18 void xnn_qd8_f32_qc8w_gemm_minmax_ukernel_3x8__neon_mlal_lane(
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 float* restrict c,
26 size_t cm_stride,
27 size_t cn_stride,
28 const union xnn_f32_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 <= 3);
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 const int8_t* a0 = a;
41 float* c0 = c;
42 const int8_t* a1 = (const int8_t*) ((uintptr_t) a0 + a_stride);
43 float* c1 = (float*) ((uintptr_t) c0 + cm_stride);
44 if XNN_UNPREDICTABLE(mr < 2) {
45 a1 = a0;
46 c1 = c0;
47 }
48 const int8_t* a2 = (const int8_t*) ((uintptr_t) a1 + a_stride);
49 float* c2 = (float*) ((uintptr_t) c1 + cm_stride);
50 if XNN_UNPREDICTABLE(mr <= 2) {
51 a2 = a1;
52 c2 = c1;
53 }
54
55 do {
56 int32x4_t vksum0123 = vld1q_s32(w); w = (const int32_t*) w + 4;
57 int32x4_t vksum4567 = vld1q_s32(w); w = (const int32_t*) w + 4;
58 const int32x4_t vzp01 = vld1q_s32(&quantization_params[0].zero_point);
59 int32x4_t vacc0x0123 = vmulq_lane_s32(vksum0123, vget_low_s32(vzp01), 0);
60 int32x4_t vacc1x0123 = vmulq_lane_s32(vksum0123, vget_high_s32(vzp01), 0);
61 int32x4_t vacc0x4567 = vmulq_lane_s32(vksum4567, vget_low_s32(vzp01), 0);
62 int32x4_t vacc1x4567 = vmulq_lane_s32(vksum4567, vget_high_s32(vzp01), 0);
63 const int32x4_t vzp2 = vld1q_dup_s32(&quantization_params[2].zero_point);
64 int32x4_t vacc2x0123 = vmulq_s32(vksum0123, vzp2);
65 int32x4_t vacc2x4567 = vmulq_s32(vksum4567, vzp2);
66
67 size_t k = kc;
68 while (k >= 8 * sizeof(int8_t)) {
69 const int8x8_t va0 = vld1_s8(a0); a0 += 8;
70 const int16x8_t vxa0 = vmovl_s8(va0);
71 const int8x8_t va1 = vld1_s8(a1); a1 += 8;
72 const int16x8_t vxa1 = vmovl_s8(va1);
73 const int8x8_t va2 = vld1_s8(a2); a2 += 8;
74 const int16x8_t vxa2 = vmovl_s8(va2);
75
76 const int8x8_t vb01234567c0 = vld1_s8(w); w = (const int8_t*) w + 8;
77 const int8x8_t vb01234567c1 = vld1_s8(w); w = (const int8_t*) w + 8;
78 const int8x8_t vb01234567c2 = vld1_s8(w); w = (const int8_t*) w + 8;
79 const int8x8_t vb01234567c3 = vld1_s8(w); w = (const int8_t*) w + 8;
80 const int8x8_t vb01234567c4 = vld1_s8(w); w = (const int8_t*) w + 8;
81 const int8x8_t vb01234567c5 = vld1_s8(w); w = (const int8_t*) w + 8;
82 const int8x8_t vb01234567c6 = vld1_s8(w); w = (const int8_t*) w + 8;
83 const int8x8_t vb01234567c7 = vld1_s8(w); w = (const int8_t*) w + 8;
84
85 const int16x8_t vxb01234567c0 = vmovl_s8(vb01234567c0);
86 const int16x8_t vxb01234567c1 = vmovl_s8(vb01234567c1);
87 const int16x8_t vxb01234567c2 = vmovl_s8(vb01234567c2);
88 const int16x8_t vxb01234567c3 = vmovl_s8(vb01234567c3);
89 const int16x8_t vxb01234567c4 = vmovl_s8(vb01234567c4);
90 const int16x8_t vxb01234567c5 = vmovl_s8(vb01234567c5);
91 const int16x8_t vxb01234567c6 = vmovl_s8(vb01234567c6);
92 const int16x8_t vxb01234567c7 = vmovl_s8(vb01234567c7);
93
94
95 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c0), vget_low_s16(vxa0), 0);
96 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c0), vget_low_s16(vxa0), 0);
97 vacc1x0123 = vmlal_lane_s16(vacc1x0123, vget_low_s16(vxb01234567c0), vget_low_s16(vxa1), 0);
98 vacc1x4567 = vmlal_lane_s16(vacc1x4567, vget_high_s16(vxb01234567c0), vget_low_s16(vxa1), 0);
99 vacc2x0123 = vmlal_lane_s16(vacc2x0123, vget_low_s16(vxb01234567c0), vget_low_s16(vxa2), 0);
100 vacc2x4567 = vmlal_lane_s16(vacc2x4567, vget_high_s16(vxb01234567c0), vget_low_s16(vxa2), 0);
101 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c1), vget_low_s16(vxa0), 1);
102 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c1), vget_low_s16(vxa0), 1);
103 vacc1x0123 = vmlal_lane_s16(vacc1x0123, vget_low_s16(vxb01234567c1), vget_low_s16(vxa1), 1);
104 vacc1x4567 = vmlal_lane_s16(vacc1x4567, vget_high_s16(vxb01234567c1), vget_low_s16(vxa1), 1);
105 vacc2x0123 = vmlal_lane_s16(vacc2x0123, vget_low_s16(vxb01234567c1), vget_low_s16(vxa2), 1);
106 vacc2x4567 = vmlal_lane_s16(vacc2x4567, vget_high_s16(vxb01234567c1), vget_low_s16(vxa2), 1);
107 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c2), vget_low_s16(vxa0), 2);
108 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c2), vget_low_s16(vxa0), 2);
109 vacc1x0123 = vmlal_lane_s16(vacc1x0123, vget_low_s16(vxb01234567c2), vget_low_s16(vxa1), 2);
110 vacc1x4567 = vmlal_lane_s16(vacc1x4567, vget_high_s16(vxb01234567c2), vget_low_s16(vxa1), 2);
111 vacc2x0123 = vmlal_lane_s16(vacc2x0123, vget_low_s16(vxb01234567c2), vget_low_s16(vxa2), 2);
112 vacc2x4567 = vmlal_lane_s16(vacc2x4567, vget_high_s16(vxb01234567c2), vget_low_s16(vxa2), 2);
113 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c3), vget_low_s16(vxa0), 3);
114 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c3), vget_low_s16(vxa0), 3);
115 vacc1x0123 = vmlal_lane_s16(vacc1x0123, vget_low_s16(vxb01234567c3), vget_low_s16(vxa1), 3);
116 vacc1x4567 = vmlal_lane_s16(vacc1x4567, vget_high_s16(vxb01234567c3), vget_low_s16(vxa1), 3);
117 vacc2x0123 = vmlal_lane_s16(vacc2x0123, vget_low_s16(vxb01234567c3), vget_low_s16(vxa2), 3);
118 vacc2x4567 = vmlal_lane_s16(vacc2x4567, vget_high_s16(vxb01234567c3), vget_low_s16(vxa2), 3);
119
120 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c4), vget_high_s16(vxa0), 0);
121 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c4), vget_high_s16(vxa0), 0);
122 vacc1x0123 = vmlal_lane_s16(vacc1x0123, vget_low_s16(vxb01234567c4), vget_high_s16(vxa1), 0);
123 vacc1x4567 = vmlal_lane_s16(vacc1x4567, vget_high_s16(vxb01234567c4), vget_high_s16(vxa1), 0);
124 vacc2x0123 = vmlal_lane_s16(vacc2x0123, vget_low_s16(vxb01234567c4), vget_high_s16(vxa2), 0);
125 vacc2x4567 = vmlal_lane_s16(vacc2x4567, vget_high_s16(vxb01234567c4), vget_high_s16(vxa2), 0);
126 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c5), vget_high_s16(vxa0), 1);
127 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c5), vget_high_s16(vxa0), 1);
128 vacc1x0123 = vmlal_lane_s16(vacc1x0123, vget_low_s16(vxb01234567c5), vget_high_s16(vxa1), 1);
129 vacc1x4567 = vmlal_lane_s16(vacc1x4567, vget_high_s16(vxb01234567c5), vget_high_s16(vxa1), 1);
130 vacc2x0123 = vmlal_lane_s16(vacc2x0123, vget_low_s16(vxb01234567c5), vget_high_s16(vxa2), 1);
131 vacc2x4567 = vmlal_lane_s16(vacc2x4567, vget_high_s16(vxb01234567c5), vget_high_s16(vxa2), 1);
132 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c6), vget_high_s16(vxa0), 2);
133 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c6), vget_high_s16(vxa0), 2);
134 vacc1x0123 = vmlal_lane_s16(vacc1x0123, vget_low_s16(vxb01234567c6), vget_high_s16(vxa1), 2);
135 vacc1x4567 = vmlal_lane_s16(vacc1x4567, vget_high_s16(vxb01234567c6), vget_high_s16(vxa1), 2);
136 vacc2x0123 = vmlal_lane_s16(vacc2x0123, vget_low_s16(vxb01234567c6), vget_high_s16(vxa2), 2);
137 vacc2x4567 = vmlal_lane_s16(vacc2x4567, vget_high_s16(vxb01234567c6), vget_high_s16(vxa2), 2);
138 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c7), vget_high_s16(vxa0), 3);
139 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c7), vget_high_s16(vxa0), 3);
140 vacc1x0123 = vmlal_lane_s16(vacc1x0123, vget_low_s16(vxb01234567c7), vget_high_s16(vxa1), 3);
141 vacc1x4567 = vmlal_lane_s16(vacc1x4567, vget_high_s16(vxb01234567c7), vget_high_s16(vxa1), 3);
142 vacc2x0123 = vmlal_lane_s16(vacc2x0123, vget_low_s16(vxb01234567c7), vget_high_s16(vxa2), 3);
143 vacc2x4567 = vmlal_lane_s16(vacc2x4567, vget_high_s16(vxb01234567c7), vget_high_s16(vxa2), 3);
144
145 k -= 8 * sizeof(int8_t);
146 }
147 if XNN_UNLIKELY(k != 0) {
148 const int8x8_t va0 = vld1_s8(a0); a0 = (const int8_t*) ((uintptr_t) a0 + k);
149 const int16x8_t vxa0 = vmovl_s8(va0);
150 const int8x8_t va1 = vld1_s8(a1); a1 = (const int8_t*) ((uintptr_t) a1 + k);
151 const int16x8_t vxa1 = vmovl_s8(va1);
152 const int8x8_t va2 = vld1_s8(a2); a2 = (const int8_t*) ((uintptr_t) a2 + k);
153 const int16x8_t vxa2 = vmovl_s8(va2);
154
155 const int8x8_t vb01234567c0 = vld1_s8(w); w = (const int8_t*) w + 8;
156
157 const int16x8_t vxb01234567c0 = vmovl_s8(vb01234567c0);
158
159 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c0), vget_low_s16(vxa0), 0);
160 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c0), vget_low_s16(vxa0), 0);
161 vacc1x0123 = vmlal_lane_s16(vacc1x0123, vget_low_s16(vxb01234567c0), vget_low_s16(vxa1), 0);
162 vacc1x4567 = vmlal_lane_s16(vacc1x4567, vget_high_s16(vxb01234567c0), vget_low_s16(vxa1), 0);
163 vacc2x0123 = vmlal_lane_s16(vacc2x0123, vget_low_s16(vxb01234567c0), vget_low_s16(vxa2), 0);
164 vacc2x4567 = vmlal_lane_s16(vacc2x4567, vget_high_s16(vxb01234567c0), vget_low_s16(vxa2), 0);
165
166 if (k >= 2 * sizeof(int8_t)) {
167 const int8x8_t vb01234567c1 = vld1_s8(w); w = (const int8_t*) w + 8;
168 const int16x8_t vxb01234567c1 = vmovl_s8(vb01234567c1);
169
170 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c1), vget_low_s16(vxa0), 1);
171 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c1), vget_low_s16(vxa0), 1);
172 vacc1x0123 = vmlal_lane_s16(vacc1x0123, vget_low_s16(vxb01234567c1), vget_low_s16(vxa1), 1);
173 vacc1x4567 = vmlal_lane_s16(vacc1x4567, vget_high_s16(vxb01234567c1), vget_low_s16(vxa1), 1);
174 vacc2x0123 = vmlal_lane_s16(vacc2x0123, vget_low_s16(vxb01234567c1), vget_low_s16(vxa2), 1);
175 vacc2x4567 = vmlal_lane_s16(vacc2x4567, vget_high_s16(vxb01234567c1), vget_low_s16(vxa2), 1);
176
177 if (k > 2 * sizeof(int8_t)) {
178 const int8x8_t vb01234567c2 = vld1_s8(w); w = (const int8_t*) w + 8;
179 const int16x8_t vxb01234567c2 = vmovl_s8(vb01234567c2);
180
181 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c2), vget_low_s16(vxa0), 2);
182 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c2), vget_low_s16(vxa0), 2);
183 vacc1x0123 = vmlal_lane_s16(vacc1x0123, vget_low_s16(vxb01234567c2), vget_low_s16(vxa1), 2);
184 vacc1x4567 = vmlal_lane_s16(vacc1x4567, vget_high_s16(vxb01234567c2), vget_low_s16(vxa1), 2);
185 vacc2x0123 = vmlal_lane_s16(vacc2x0123, vget_low_s16(vxb01234567c2), vget_low_s16(vxa2), 2);
186 vacc2x4567 = vmlal_lane_s16(vacc2x4567, vget_high_s16(vxb01234567c2), vget_low_s16(vxa2), 2);
187
188 if (k >= 4 * sizeof(int8_t)) {
189 const int8x8_t vb01234567c3 = vld1_s8(w); w = (const int8_t*) w + 8;
190 const int16x8_t vxb01234567c3 = vmovl_s8(vb01234567c3);
191
192 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c3), vget_low_s16(vxa0), 3);
193 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c3), vget_low_s16(vxa0), 3);
194 vacc1x0123 = vmlal_lane_s16(vacc1x0123, vget_low_s16(vxb01234567c3), vget_low_s16(vxa1), 3);
195 vacc1x4567 = vmlal_lane_s16(vacc1x4567, vget_high_s16(vxb01234567c3), vget_low_s16(vxa1), 3);
196 vacc2x0123 = vmlal_lane_s16(vacc2x0123, vget_low_s16(vxb01234567c3), vget_low_s16(vxa2), 3);
197 vacc2x4567 = vmlal_lane_s16(vacc2x4567, vget_high_s16(vxb01234567c3), vget_low_s16(vxa2), 3);
198
199 if (k > 4 * sizeof(int8_t)) {
200 const int8x8_t vb01234567c4 = vld1_s8(w); w = (const int8_t*) w + 8;
201 const int16x8_t vxb01234567c4 = vmovl_s8(vb01234567c4);
202
203 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c4), vget_high_s16(vxa0), 0);
204 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c4), vget_high_s16(vxa0), 0);
205 vacc1x0123 = vmlal_lane_s16(vacc1x0123, vget_low_s16(vxb01234567c4), vget_high_s16(vxa1), 0);
206 vacc1x4567 = vmlal_lane_s16(vacc1x4567, vget_high_s16(vxb01234567c4), vget_high_s16(vxa1), 0);
207 vacc2x0123 = vmlal_lane_s16(vacc2x0123, vget_low_s16(vxb01234567c4), vget_high_s16(vxa2), 0);
208 vacc2x4567 = vmlal_lane_s16(vacc2x4567, vget_high_s16(vxb01234567c4), vget_high_s16(vxa2), 0);
209
210 if (k >= 6 * sizeof(int8_t)) {
211 const int8x8_t vb01234567c5 = vld1_s8(w); w = (const int8_t*) w + 8;
212 const int16x8_t vxb01234567c5 = vmovl_s8(vb01234567c5);
213
214 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c5), vget_high_s16(vxa0), 1);
215 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c5), vget_high_s16(vxa0), 1);
216 vacc1x0123 = vmlal_lane_s16(vacc1x0123, vget_low_s16(vxb01234567c5), vget_high_s16(vxa1), 1);
217 vacc1x4567 = vmlal_lane_s16(vacc1x4567, vget_high_s16(vxb01234567c5), vget_high_s16(vxa1), 1);
218 vacc2x0123 = vmlal_lane_s16(vacc2x0123, vget_low_s16(vxb01234567c5), vget_high_s16(vxa2), 1);
219 vacc2x4567 = vmlal_lane_s16(vacc2x4567, vget_high_s16(vxb01234567c5), vget_high_s16(vxa2), 1);
220
221 if (k > 6 * sizeof(int8_t)) {
222 const int8x8_t vb01234567c6 = vld1_s8(w); w = (const int8_t*) w + 8;
223 const int16x8_t vxb01234567c6 = vmovl_s8(vb01234567c6);
224
225 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c6), vget_high_s16(vxa0), 2);
226 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c6), vget_high_s16(vxa0), 2);
227 vacc1x0123 = vmlal_lane_s16(vacc1x0123, vget_low_s16(vxb01234567c6), vget_high_s16(vxa1), 2);
228 vacc1x4567 = vmlal_lane_s16(vacc1x4567, vget_high_s16(vxb01234567c6), vget_high_s16(vxa1), 2);
229 vacc2x0123 = vmlal_lane_s16(vacc2x0123, vget_low_s16(vxb01234567c6), vget_high_s16(vxa2), 2);
230 vacc2x4567 = vmlal_lane_s16(vacc2x4567, vget_high_s16(vxb01234567c6), vget_high_s16(vxa2), 2);
231 }
232 }
233 }
234 }
235 }
236 }
237 }
238
239 float32x4_t vout0x0123 = vcvtq_f32_s32(vacc0x0123);
240 float32x4_t vout0x4567 = vcvtq_f32_s32(vacc0x4567);
241 float32x4_t vout1x0123 = vcvtq_f32_s32(vacc1x0123);
242 float32x4_t vout1x4567 = vcvtq_f32_s32(vacc1x4567);
243 float32x4_t vout2x0123 = vcvtq_f32_s32(vacc2x0123);
244 float32x4_t vout2x4567 = vcvtq_f32_s32(vacc2x4567);
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_scale2 = vld1q_dup_f32(&quantization_params[2].inv_scale);
252 vout2x0123 = vmulq_f32(vout2x0123, vinput_scale2);
253 vout2x4567 = vmulq_f32(vout2x4567, vinput_scale2);
254
255 const float32x4_t vfilter_output_scale0123 = vld1q_f32(w); w = (const float*) w + 4;
256 const float32x4_t vfilter_output_scale4567 = vld1q_f32(w); w = (const float*) w + 4;
257
258 const float32x4_t vbias0123 = vld1q_f32(w); w = (const float*) w + 4;
259 #if XNN_ARCH_ARM64
260 vout0x0123 = vfmaq_f32(vbias0123, vout0x0123, vfilter_output_scale0123);
261 vout1x0123 = vfmaq_f32(vbias0123, vout1x0123, vfilter_output_scale0123);
262 vout2x0123 = vfmaq_f32(vbias0123, vout2x0123, vfilter_output_scale0123);
263 #else
264 vout0x0123 = vmlaq_f32(vbias0123, vout0x0123, vfilter_output_scale0123);
265 vout1x0123 = vmlaq_f32(vbias0123, vout1x0123, vfilter_output_scale0123);
266 vout2x0123 = vmlaq_f32(vbias0123, vout2x0123, vfilter_output_scale0123);
267 #endif
268 const float32x4_t vbias4567 = vld1q_f32(w); w = (const float*) w + 4;
269 #if XNN_ARCH_ARM64
270 vout0x4567 = vfmaq_f32(vbias4567, vout0x4567, vfilter_output_scale4567);
271 vout1x4567 = vfmaq_f32(vbias4567, vout1x4567, vfilter_output_scale4567);
272 vout2x4567 = vfmaq_f32(vbias4567, vout2x4567, vfilter_output_scale4567);
273 #else
274 vout0x4567 = vmlaq_f32(vbias4567, vout0x4567, vfilter_output_scale4567);
275 vout1x4567 = vmlaq_f32(vbias4567, vout1x4567, vfilter_output_scale4567);
276 vout2x4567 = vmlaq_f32(vbias4567, vout2x4567, vfilter_output_scale4567);
277 #endif
278
279 const float32x4_t voutput_min = vld1q_dup_f32(&params->scalar.min);
280 vout0x0123 = vmaxq_f32(vout0x0123, voutput_min);
281 vout0x4567 = vmaxq_f32(vout0x4567, voutput_min);
282 vout1x0123 = vmaxq_f32(vout1x0123, voutput_min);
283 vout1x4567 = vmaxq_f32(vout1x4567, voutput_min);
284 vout2x0123 = vmaxq_f32(vout2x0123, voutput_min);
285 vout2x4567 = vmaxq_f32(vout2x4567, voutput_min);
286
287 const float32x4_t voutput_max = vld1q_dup_f32(&params->scalar.max);
288 vout0x0123 = vminq_f32(vout0x0123, voutput_max);
289 vout0x4567 = vminq_f32(vout0x4567, voutput_max);
290 vout1x0123 = vminq_f32(vout1x0123, voutput_max);
291 vout1x4567 = vminq_f32(vout1x4567, voutput_max);
292 vout2x0123 = vminq_f32(vout2x0123, voutput_max);
293 vout2x4567 = vminq_f32(vout2x4567, voutput_max);
294
295 if XNN_LIKELY(nc >= 8) {
296 vst1q_f32(&c0[0], vout0x0123);
297 vst1q_f32(&c0[4], vout0x4567);
298 vst1q_f32(&c1[0], vout1x0123);
299 vst1q_f32(&c1[4], vout1x4567);
300 vst1q_f32(&c2[0], vout2x0123);
301 vst1q_f32(&c2[4], vout2x4567);
302
303 a0 = (const int8_t*) ((uintptr_t) a0 - kc);
304 a1 = (const int8_t*) ((uintptr_t) a1 - kc);
305 a2 = (const int8_t*) ((uintptr_t) a2 - kc);
306
307 c0 = (float*) ((uintptr_t) c0 + cn_stride);
308 c1 = (float*) ((uintptr_t) c1 + cn_stride);
309 c2 = (float*) ((uintptr_t) c2 + cn_stride);
310
311 nc -= 8;
312 } else {
313 if (nc & 4) {
314 vst1q_f32(c0, vout0x0123); c0 += 4;
315 vout0x0123 = vout0x4567;
316 vst1q_f32(c1, vout1x0123); c1 += 4;
317 vout1x0123 = vout1x4567;
318 vst1q_f32(c2, vout2x0123); c2 += 4;
319 vout2x0123 = vout2x4567;
320 }
321 float32x2_t vout0x01 = vget_low_f32(vout0x0123);
322 float32x2_t vout1x01 = vget_low_f32(vout1x0123);
323 float32x2_t vout2x01 = vget_low_f32(vout2x0123);
324 if (nc & 2) {
325 vst1_f32(c0, vout0x01); c0 += 2;
326 vst1_f32(c1, vout1x01); c1 += 2;
327 vst1_f32(c2, vout2x01); c2 += 2;
328 vout0x01 = vget_high_f32(vout0x0123);
329 vout1x01 = vget_high_f32(vout1x0123);
330 vout2x01 = vget_high_f32(vout2x0123);
331 }
332 if (nc & 1) {
333 vst1_lane_f32(c0, vout0x01, 0);
334 vst1_lane_f32(c1, vout1x01, 0);
335 vst1_lane_f32(c2, vout2x01, 0);
336 }
337 nc = 0;
338 }
339 } while (nc != 0);
340 }