Debian Codegraph @

src/qd8-f32-qc8w-igemm/gen/qd8-f32-qc8w-igemm-4x8-minmax-neon-mlal-lane.c

raw
1 // Auto-generated file. Do not edit!
2 // Template: src/qs8-igemm/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
11 #include <assert.h>
12
13 #include <arm_neon.h>
14
15 #include "xnnpack/common.h"
16 #include "xnnpack/igemm.h"
17
18
19 void xnn_qd8_f32_qc8w_igemm_minmax_ukernel_4x8__neon_mlal_lane(
20 size_t mr,
21 size_t nc,
22 size_t kc,
23 size_t ks,
24 const int8_t** restrict a,
25 const void* restrict w,
26 float* restrict c,
27 size_t cm_stride,
28 size_t cn_stride,
29 size_t a_offset,
30 const int8_t* zero,
31 const int8_t* zero_data,
32 const union xnn_f32_minmax_params params[restrict XNN_MIN_ELEMENTS(1)],
33 const struct xnn_qd8_quantization_params quantization_params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
34 {
35 assert(mr != 0);
36 assert(mr <= 4);
37 assert(nc != 0);
38 assert(kc != 0);
39 assert(ks != 0);
40 assert(ks % (4 * sizeof(void*)) == 0);
41 assert(a_offset % sizeof(int8_t) == 0);
42 assert(a != NULL);
43 assert(w != NULL);
44 assert(c != NULL);
45
46 float* c0 = c;
47 float* c1 = (float*) ((uintptr_t) c0 + cm_stride);
48 if XNN_UNPREDICTABLE(mr < 2) {
49 c1 = c0;
50 }
51 float* c2 = (float*) ((uintptr_t) c1 + cm_stride);
52 if XNN_UNPREDICTABLE(mr <= 2) {
53 c2 = c1;
54 }
55 float* c3 = (float*) ((uintptr_t) c2 + cm_stride);
56 if XNN_UNPREDICTABLE(mr != 4) {
57 c3 = c2;
58 }
59
60 do {
61 int32x4_t vksum0123 = vld1q_s32(w); w = (const int32_t*) w + 4;
62 int32x4_t vksum4567 = vld1q_s32(w); w = (const int32_t*) w + 4;
63 const int32x4_t vinput_zero_point = vld1q_dup_s32(&quantization_params->zero_point);
64 int32x4_t vacc0x0123 = vmulq_s32(vksum0123, vinput_zero_point);
65 int32x4_t vacc0x4567 = vmulq_s32(vksum4567, vinput_zero_point);
66 int32x4_t vacc1x0123 = vacc0x0123;
67 int32x4_t vacc1x4567 = vacc0x4567;
68 int32x4_t vacc2x0123 = vacc0x0123;
69 int32x4_t vacc2x4567 = vacc0x4567;
70 int32x4_t vacc3x0123 = vacc0x0123;
71 int32x4_t vacc3x4567 = vacc0x4567;
72
73 size_t p = ks;
74 do {
75 const int8_t* restrict a0 = a[0];
76 if XNN_UNPREDICTABLE(a0 != zero) {
77 a0 = (const int8_t*) ((uintptr_t) a0 + a_offset);
78 } else {
79 a0 = zero_data;
80 }
81 const int8_t* restrict a1 = a[1];
82 if XNN_UNPREDICTABLE(a1 != zero) {
83 a1 = (const int8_t*) ((uintptr_t) a1 + a_offset);
84 } else {
85 a1 = zero_data;
86 }
87 const int8_t* restrict a2 = a[2];
88 if XNN_UNPREDICTABLE(a2 != zero) {
89 a2 = (const int8_t*) ((uintptr_t) a2 + a_offset);
90 } else {
91 a2 = zero_data;
92 }
93 const int8_t* restrict a3 = a[3];
94 if XNN_UNPREDICTABLE(a3 != zero) {
95 a3 = (const int8_t*) ((uintptr_t) a3 + a_offset);
96 } else {
97 a3 = zero_data;
98 }
99 a += 4;
100
101 size_t k = kc;
102 while (k >= 8 * sizeof(int8_t)) {
103 const int8x8_t va0 = vld1_s8(a0); a0 += 8;
104 const int16x8_t vxa0 = vmovl_s8(va0);
105 const int8x8_t va1 = vld1_s8(a1); a1 += 8;
106 const int16x8_t vxa1 = vmovl_s8(va1);
107 const int8x8_t va2 = vld1_s8(a2); a2 += 8;
108 const int16x8_t vxa2 = vmovl_s8(va2);
109 const int8x8_t va3 = vld1_s8(a3); a3 += 8;
110 const int16x8_t vxa3 = vmovl_s8(va3);
111
112 const int8x8_t vb01234567c0 = vld1_s8(w); w = (const int8_t*) w + 8;
113 const int16x8_t vxb01234567c0 = vmovl_s8(vb01234567c0);
114
115 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c0), vget_low_s16(vxa0), 0);
116 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c0), vget_low_s16(vxa0), 0);
117 vacc1x0123 = vmlal_lane_s16(vacc1x0123, vget_low_s16(vxb01234567c0), vget_low_s16(vxa1), 0);
118 vacc1x4567 = vmlal_lane_s16(vacc1x4567, vget_high_s16(vxb01234567c0), vget_low_s16(vxa1), 0);
119 vacc2x0123 = vmlal_lane_s16(vacc2x0123, vget_low_s16(vxb01234567c0), vget_low_s16(vxa2), 0);
120 vacc2x4567 = vmlal_lane_s16(vacc2x4567, vget_high_s16(vxb01234567c0), vget_low_s16(vxa2), 0);
121 vacc3x0123 = vmlal_lane_s16(vacc3x0123, vget_low_s16(vxb01234567c0), vget_low_s16(vxa3), 0);
122 vacc3x4567 = vmlal_lane_s16(vacc3x4567, vget_high_s16(vxb01234567c0), vget_low_s16(vxa3), 0);
123 const int8x8_t vb01234567c1 = vld1_s8(w); w = (const int8_t*) w + 8;
124 const int16x8_t vxb01234567c1 = vmovl_s8(vb01234567c1);
125
126 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c1), vget_low_s16(vxa0), 1);
127 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c1), vget_low_s16(vxa0), 1);
128 vacc1x0123 = vmlal_lane_s16(vacc1x0123, vget_low_s16(vxb01234567c1), vget_low_s16(vxa1), 1);
129 vacc1x4567 = vmlal_lane_s16(vacc1x4567, vget_high_s16(vxb01234567c1), vget_low_s16(vxa1), 1);
130 vacc2x0123 = vmlal_lane_s16(vacc2x0123, vget_low_s16(vxb01234567c1), vget_low_s16(vxa2), 1);
131 vacc2x4567 = vmlal_lane_s16(vacc2x4567, vget_high_s16(vxb01234567c1), vget_low_s16(vxa2), 1);
132 vacc3x0123 = vmlal_lane_s16(vacc3x0123, vget_low_s16(vxb01234567c1), vget_low_s16(vxa3), 1);
133 vacc3x4567 = vmlal_lane_s16(vacc3x4567, vget_high_s16(vxb01234567c1), vget_low_s16(vxa3), 1);
134 const int8x8_t vb01234567c2 = vld1_s8(w); w = (const int8_t*) w + 8;
135 const int16x8_t vxb01234567c2 = vmovl_s8(vb01234567c2);
136
137 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c2), vget_low_s16(vxa0), 2);
138 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c2), vget_low_s16(vxa0), 2);
139 vacc1x0123 = vmlal_lane_s16(vacc1x0123, vget_low_s16(vxb01234567c2), vget_low_s16(vxa1), 2);
140 vacc1x4567 = vmlal_lane_s16(vacc1x4567, vget_high_s16(vxb01234567c2), vget_low_s16(vxa1), 2);
141 vacc2x0123 = vmlal_lane_s16(vacc2x0123, vget_low_s16(vxb01234567c2), vget_low_s16(vxa2), 2);
142 vacc2x4567 = vmlal_lane_s16(vacc2x4567, vget_high_s16(vxb01234567c2), vget_low_s16(vxa2), 2);
143 vacc3x0123 = vmlal_lane_s16(vacc3x0123, vget_low_s16(vxb01234567c2), vget_low_s16(vxa3), 2);
144 vacc3x4567 = vmlal_lane_s16(vacc3x4567, vget_high_s16(vxb01234567c2), vget_low_s16(vxa3), 2);
145 const int8x8_t vb01234567c3 = vld1_s8(w); w = (const int8_t*) w + 8;
146 const int16x8_t vxb01234567c3 = vmovl_s8(vb01234567c3);
147
148 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c3), vget_low_s16(vxa0), 3);
149 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c3), vget_low_s16(vxa0), 3);
150 vacc1x0123 = vmlal_lane_s16(vacc1x0123, vget_low_s16(vxb01234567c3), vget_low_s16(vxa1), 3);
151 vacc1x4567 = vmlal_lane_s16(vacc1x4567, vget_high_s16(vxb01234567c3), vget_low_s16(vxa1), 3);
152 vacc2x0123 = vmlal_lane_s16(vacc2x0123, vget_low_s16(vxb01234567c3), vget_low_s16(vxa2), 3);
153 vacc2x4567 = vmlal_lane_s16(vacc2x4567, vget_high_s16(vxb01234567c3), vget_low_s16(vxa2), 3);
154 vacc3x0123 = vmlal_lane_s16(vacc3x0123, vget_low_s16(vxb01234567c3), vget_low_s16(vxa3), 3);
155 vacc3x4567 = vmlal_lane_s16(vacc3x4567, vget_high_s16(vxb01234567c3), vget_low_s16(vxa3), 3);
156
157
158 const int8x8_t vb01234567c4 = vld1_s8(w); w = (const int8_t*) w + 8;
159 const int16x8_t vxb01234567c4 = vmovl_s8(vb01234567c4);
160
161 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c4), vget_high_s16(vxa0), 0);
162 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c4), vget_high_s16(vxa0), 0);
163 vacc1x0123 = vmlal_lane_s16(vacc1x0123, vget_low_s16(vxb01234567c4), vget_high_s16(vxa1), 0);
164 vacc1x4567 = vmlal_lane_s16(vacc1x4567, vget_high_s16(vxb01234567c4), vget_high_s16(vxa1), 0);
165 vacc2x0123 = vmlal_lane_s16(vacc2x0123, vget_low_s16(vxb01234567c4), vget_high_s16(vxa2), 0);
166 vacc2x4567 = vmlal_lane_s16(vacc2x4567, vget_high_s16(vxb01234567c4), vget_high_s16(vxa2), 0);
167 vacc3x0123 = vmlal_lane_s16(vacc3x0123, vget_low_s16(vxb01234567c4), vget_high_s16(vxa3), 0);
168 vacc3x4567 = vmlal_lane_s16(vacc3x4567, vget_high_s16(vxb01234567c4), vget_high_s16(vxa3), 0);
169 const int8x8_t vb01234567c5 = vld1_s8(w); w = (const int8_t*) w + 8;
170 const int16x8_t vxb01234567c5 = vmovl_s8(vb01234567c5);
171
172 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c5), vget_high_s16(vxa0), 1);
173 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c5), vget_high_s16(vxa0), 1);
174 vacc1x0123 = vmlal_lane_s16(vacc1x0123, vget_low_s16(vxb01234567c5), vget_high_s16(vxa1), 1);
175 vacc1x4567 = vmlal_lane_s16(vacc1x4567, vget_high_s16(vxb01234567c5), vget_high_s16(vxa1), 1);
176 vacc2x0123 = vmlal_lane_s16(vacc2x0123, vget_low_s16(vxb01234567c5), vget_high_s16(vxa2), 1);
177 vacc2x4567 = vmlal_lane_s16(vacc2x4567, vget_high_s16(vxb01234567c5), vget_high_s16(vxa2), 1);
178 vacc3x0123 = vmlal_lane_s16(vacc3x0123, vget_low_s16(vxb01234567c5), vget_high_s16(vxa3), 1);
179 vacc3x4567 = vmlal_lane_s16(vacc3x4567, vget_high_s16(vxb01234567c5), vget_high_s16(vxa3), 1);
180 const int8x8_t vb01234567c6 = vld1_s8(w); w = (const int8_t*) w + 8;
181 const int16x8_t vxb01234567c6 = vmovl_s8(vb01234567c6);
182
183 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c6), vget_high_s16(vxa0), 2);
184 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c6), vget_high_s16(vxa0), 2);
185 vacc1x0123 = vmlal_lane_s16(vacc1x0123, vget_low_s16(vxb01234567c6), vget_high_s16(vxa1), 2);
186 vacc1x4567 = vmlal_lane_s16(vacc1x4567, vget_high_s16(vxb01234567c6), vget_high_s16(vxa1), 2);
187 vacc2x0123 = vmlal_lane_s16(vacc2x0123, vget_low_s16(vxb01234567c6), vget_high_s16(vxa2), 2);
188 vacc2x4567 = vmlal_lane_s16(vacc2x4567, vget_high_s16(vxb01234567c6), vget_high_s16(vxa2), 2);
189 vacc3x0123 = vmlal_lane_s16(vacc3x0123, vget_low_s16(vxb01234567c6), vget_high_s16(vxa3), 2);
190 vacc3x4567 = vmlal_lane_s16(vacc3x4567, vget_high_s16(vxb01234567c6), vget_high_s16(vxa3), 2);
191 const int8x8_t vb01234567c7 = vld1_s8(w); w = (const int8_t*) w + 8;
192 const int16x8_t vxb01234567c7 = vmovl_s8(vb01234567c7);
193
194 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c7), vget_high_s16(vxa0), 3);
195 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c7), vget_high_s16(vxa0), 3);
196 vacc1x0123 = vmlal_lane_s16(vacc1x0123, vget_low_s16(vxb01234567c7), vget_high_s16(vxa1), 3);
197 vacc1x4567 = vmlal_lane_s16(vacc1x4567, vget_high_s16(vxb01234567c7), vget_high_s16(vxa1), 3);
198 vacc2x0123 = vmlal_lane_s16(vacc2x0123, vget_low_s16(vxb01234567c7), vget_high_s16(vxa2), 3);
199 vacc2x4567 = vmlal_lane_s16(vacc2x4567, vget_high_s16(vxb01234567c7), vget_high_s16(vxa2), 3);
200 vacc3x0123 = vmlal_lane_s16(vacc3x0123, vget_low_s16(vxb01234567c7), vget_high_s16(vxa3), 3);
201 vacc3x4567 = vmlal_lane_s16(vacc3x4567, vget_high_s16(vxb01234567c7), vget_high_s16(vxa3), 3);
202
203 k -= 8 * sizeof(int8_t);
204 }
205 if XNN_UNLIKELY(k != 0) {
206 const int8x8_t va0 = vld1_s8(a0); a0 = (const int8_t*) ((uintptr_t) a0 + k);
207 const int16x8_t vxa0 = vmovl_s8(va0);
208 const int8x8_t va1 = vld1_s8(a1); a1 = (const int8_t*) ((uintptr_t) a1 + k);
209 const int16x8_t vxa1 = vmovl_s8(va1);
210 const int8x8_t va2 = vld1_s8(a2); a2 = (const int8_t*) ((uintptr_t) a2 + k);
211 const int16x8_t vxa2 = vmovl_s8(va2);
212 const int8x8_t va3 = vld1_s8(a3); a3 = (const int8_t*) ((uintptr_t) a3 + k);
213 const int16x8_t vxa3 = vmovl_s8(va3);
214
215 const int8x8_t vb01234567c0 = vld1_s8(w); w = (const int8_t*) w + 8;
216 const int16x8_t vxb01234567c0 = vmovl_s8(vb01234567c0);
217
218 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c0), vget_low_s16(vxa0), 0);
219 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c0), vget_low_s16(vxa0), 0);
220 vacc1x0123 = vmlal_lane_s16(vacc1x0123, vget_low_s16(vxb01234567c0), vget_low_s16(vxa1), 0);
221 vacc1x4567 = vmlal_lane_s16(vacc1x4567, vget_high_s16(vxb01234567c0), vget_low_s16(vxa1), 0);
222 vacc2x0123 = vmlal_lane_s16(vacc2x0123, vget_low_s16(vxb01234567c0), vget_low_s16(vxa2), 0);
223 vacc2x4567 = vmlal_lane_s16(vacc2x4567, vget_high_s16(vxb01234567c0), vget_low_s16(vxa2), 0);
224 vacc3x0123 = vmlal_lane_s16(vacc3x0123, vget_low_s16(vxb01234567c0), vget_low_s16(vxa3), 0);
225 vacc3x4567 = vmlal_lane_s16(vacc3x4567, vget_high_s16(vxb01234567c0), vget_low_s16(vxa3), 0);
226
227 if (k >= 2 * sizeof(int8_t)) {
228 const int8x8_t vb01234567c1 = vld1_s8(w); w = (const int8_t*) w + 8;
229 const int16x8_t vxb01234567c1 = vmovl_s8(vb01234567c1);
230
231 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c1), vget_low_s16(vxa0), 1);
232 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c1), vget_low_s16(vxa0), 1);
233 vacc1x0123 = vmlal_lane_s16(vacc1x0123, vget_low_s16(vxb01234567c1), vget_low_s16(vxa1), 1);
234 vacc1x4567 = vmlal_lane_s16(vacc1x4567, vget_high_s16(vxb01234567c1), vget_low_s16(vxa1), 1);
235 vacc2x0123 = vmlal_lane_s16(vacc2x0123, vget_low_s16(vxb01234567c1), vget_low_s16(vxa2), 1);
236 vacc2x4567 = vmlal_lane_s16(vacc2x4567, vget_high_s16(vxb01234567c1), vget_low_s16(vxa2), 1);
237 vacc3x0123 = vmlal_lane_s16(vacc3x0123, vget_low_s16(vxb01234567c1), vget_low_s16(vxa3), 1);
238 vacc3x4567 = vmlal_lane_s16(vacc3x4567, vget_high_s16(vxb01234567c1), vget_low_s16(vxa3), 1);
239
240 if (k > 2 * sizeof(int8_t)) {
241 const int8x8_t vb01234567c2 = vld1_s8(w); w = (const int8_t*) w + 8;
242 const int16x8_t vxb01234567c2 = vmovl_s8(vb01234567c2);
243
244 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c2), vget_low_s16(vxa0), 2);
245 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c2), vget_low_s16(vxa0), 2);
246 vacc1x0123 = vmlal_lane_s16(vacc1x0123, vget_low_s16(vxb01234567c2), vget_low_s16(vxa1), 2);
247 vacc1x4567 = vmlal_lane_s16(vacc1x4567, vget_high_s16(vxb01234567c2), vget_low_s16(vxa1), 2);
248 vacc2x0123 = vmlal_lane_s16(vacc2x0123, vget_low_s16(vxb01234567c2), vget_low_s16(vxa2), 2);
249 vacc2x4567 = vmlal_lane_s16(vacc2x4567, vget_high_s16(vxb01234567c2), vget_low_s16(vxa2), 2);
250 vacc3x0123 = vmlal_lane_s16(vacc3x0123, vget_low_s16(vxb01234567c2), vget_low_s16(vxa3), 2);
251 vacc3x4567 = vmlal_lane_s16(vacc3x4567, vget_high_s16(vxb01234567c2), vget_low_s16(vxa3), 2);
252
253 if (k >= 4 * sizeof(int8_t)) {
254 const int8x8_t vb01234567c3 = vld1_s8(w); w = (const int8_t*) w + 8;
255 const int16x8_t vxb01234567c3 = vmovl_s8(vb01234567c3);
256
257 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c3), vget_low_s16(vxa0), 3);
258 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c3), vget_low_s16(vxa0), 3);
259 vacc1x0123 = vmlal_lane_s16(vacc1x0123, vget_low_s16(vxb01234567c3), vget_low_s16(vxa1), 3);
260 vacc1x4567 = vmlal_lane_s16(vacc1x4567, vget_high_s16(vxb01234567c3), vget_low_s16(vxa1), 3);
261 vacc2x0123 = vmlal_lane_s16(vacc2x0123, vget_low_s16(vxb01234567c3), vget_low_s16(vxa2), 3);
262 vacc2x4567 = vmlal_lane_s16(vacc2x4567, vget_high_s16(vxb01234567c3), vget_low_s16(vxa2), 3);
263 vacc3x0123 = vmlal_lane_s16(vacc3x0123, vget_low_s16(vxb01234567c3), vget_low_s16(vxa3), 3);
264 vacc3x4567 = vmlal_lane_s16(vacc3x4567, vget_high_s16(vxb01234567c3), vget_low_s16(vxa3), 3);
265
266 if (k > 4 * sizeof(int8_t)) {
267 const int8x8_t vb01234567c4 = vld1_s8(w); w = (const int8_t*) w + 8;
268 const int16x8_t vxb01234567c4 = vmovl_s8(vb01234567c4);
269
270 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c4), vget_high_s16(vxa0), 0);
271 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c4), vget_high_s16(vxa0), 0);
272 vacc1x0123 = vmlal_lane_s16(vacc1x0123, vget_low_s16(vxb01234567c4), vget_high_s16(vxa1), 0);
273 vacc1x4567 = vmlal_lane_s16(vacc1x4567, vget_high_s16(vxb01234567c4), vget_high_s16(vxa1), 0);
274 vacc2x0123 = vmlal_lane_s16(vacc2x0123, vget_low_s16(vxb01234567c4), vget_high_s16(vxa2), 0);
275 vacc2x4567 = vmlal_lane_s16(vacc2x4567, vget_high_s16(vxb01234567c4), vget_high_s16(vxa2), 0);
276 vacc3x0123 = vmlal_lane_s16(vacc3x0123, vget_low_s16(vxb01234567c4), vget_high_s16(vxa3), 0);
277 vacc3x4567 = vmlal_lane_s16(vacc3x4567, vget_high_s16(vxb01234567c4), vget_high_s16(vxa3), 0);
278
279 if (k >= 6 * sizeof(int8_t)) {
280 const int8x8_t vb01234567c5 = vld1_s8(w); w = (const int8_t*) w + 8;
281 const int16x8_t vxb01234567c5 = vmovl_s8(vb01234567c5);
282
283 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c5), vget_high_s16(vxa0), 1);
284 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c5), vget_high_s16(vxa0), 1);
285 vacc1x0123 = vmlal_lane_s16(vacc1x0123, vget_low_s16(vxb01234567c5), vget_high_s16(vxa1), 1);
286 vacc1x4567 = vmlal_lane_s16(vacc1x4567, vget_high_s16(vxb01234567c5), vget_high_s16(vxa1), 1);
287 vacc2x0123 = vmlal_lane_s16(vacc2x0123, vget_low_s16(vxb01234567c5), vget_high_s16(vxa2), 1);
288 vacc2x4567 = vmlal_lane_s16(vacc2x4567, vget_high_s16(vxb01234567c5), vget_high_s16(vxa2), 1);
289 vacc3x0123 = vmlal_lane_s16(vacc3x0123, vget_low_s16(vxb01234567c5), vget_high_s16(vxa3), 1);
290 vacc3x4567 = vmlal_lane_s16(vacc3x4567, vget_high_s16(vxb01234567c5), vget_high_s16(vxa3), 1);
291
292 if (k > 6 * sizeof(int8_t)) {
293 const int8x8_t vb01234567c6 = vld1_s8(w); w = (const int8_t*) w + 8;
294 const int16x8_t vxb01234567c6 = vmovl_s8(vb01234567c6);
295
296 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c6), vget_high_s16(vxa0), 2);
297 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c6), vget_high_s16(vxa0), 2);
298 vacc1x0123 = vmlal_lane_s16(vacc1x0123, vget_low_s16(vxb01234567c6), vget_high_s16(vxa1), 2);
299 vacc1x4567 = vmlal_lane_s16(vacc1x4567, vget_high_s16(vxb01234567c6), vget_high_s16(vxa1), 2);
300 vacc2x0123 = vmlal_lane_s16(vacc2x0123, vget_low_s16(vxb01234567c6), vget_high_s16(vxa2), 2);
301 vacc2x4567 = vmlal_lane_s16(vacc2x4567, vget_high_s16(vxb01234567c6), vget_high_s16(vxa2), 2);
302 vacc3x0123 = vmlal_lane_s16(vacc3x0123, vget_low_s16(vxb01234567c6), vget_high_s16(vxa3), 2);
303 vacc3x4567 = vmlal_lane_s16(vacc3x4567, vget_high_s16(vxb01234567c6), vget_high_s16(vxa3), 2);
304 }
305 }
306 }
307 }
308 }
309 }
310 }
311 p -= 4 * sizeof(void*);
312 } while (p != 0);
313
314 float32x4_t vout0x0123 = vcvtq_f32_s32(vacc0x0123);
315 float32x4_t vout0x4567 = vcvtq_f32_s32(vacc0x4567);
316 float32x4_t vout1x0123 = vcvtq_f32_s32(vacc1x0123);
317 float32x4_t vout1x4567 = vcvtq_f32_s32(vacc1x4567);
318 float32x4_t vout2x0123 = vcvtq_f32_s32(vacc2x0123);
319 float32x4_t vout2x4567 = vcvtq_f32_s32(vacc2x4567);
320 float32x4_t vout3x0123 = vcvtq_f32_s32(vacc3x0123);
321 float32x4_t vout3x4567 = vcvtq_f32_s32(vacc3x4567);
322
323 const float32x4_t vinput_scale = vld1q_dup_f32(&quantization_params->inv_scale);
324 vout0x0123 = vmulq_f32(vout0x0123, vinput_scale);
325 vout0x4567 = vmulq_f32(vout0x4567, vinput_scale);
326 vout1x0123 = vmulq_f32(vout1x0123, vinput_scale);
327 vout1x4567 = vmulq_f32(vout1x4567, vinput_scale);
328 vout2x0123 = vmulq_f32(vout2x0123, vinput_scale);
329 vout2x4567 = vmulq_f32(vout2x4567, vinput_scale);
330 vout3x0123 = vmulq_f32(vout3x0123, vinput_scale);
331 vout3x4567 = vmulq_f32(vout3x4567, vinput_scale);
332
333 const float32x4_t vfilter_output_scale0123 = vld1q_f32(w); w = (const float*) w + 4;
334 const float32x4_t vfilter_output_scale4567 = vld1q_f32(w); w = (const float*) w + 4;
335
336 const float32x4_t vbias0123 = vld1q_f32(w); w = (const float*) w + 4;
337 #if XNN_ARCH_ARM64
338 vout0x0123 = vfmaq_f32(vbias0123, vout0x0123, vfilter_output_scale0123);
339 vout1x0123 = vfmaq_f32(vbias0123, vout1x0123, vfilter_output_scale0123);
340 vout2x0123 = vfmaq_f32(vbias0123, vout2x0123, vfilter_output_scale0123);
341 vout3x0123 = vfmaq_f32(vbias0123, vout3x0123, vfilter_output_scale0123);
342 #else
343 vout0x0123 = vmlaq_f32(vbias0123, vout0x0123, vfilter_output_scale0123);
344 vout1x0123 = vmlaq_f32(vbias0123, vout1x0123, vfilter_output_scale0123);
345 vout2x0123 = vmlaq_f32(vbias0123, vout2x0123, vfilter_output_scale0123);
346 vout3x0123 = vmlaq_f32(vbias0123, vout3x0123, vfilter_output_scale0123);
347 #endif
348 const float32x4_t vbias4567 = vld1q_f32(w); w = (const float*) w + 4;
349 #if XNN_ARCH_ARM64
350 vout0x4567 = vfmaq_f32(vbias4567, vout0x4567, vfilter_output_scale4567);
351 vout1x4567 = vfmaq_f32(vbias4567, vout1x4567, vfilter_output_scale4567);
352 vout2x4567 = vfmaq_f32(vbias4567, vout2x4567, vfilter_output_scale4567);
353 vout3x4567 = vfmaq_f32(vbias4567, vout3x4567, vfilter_output_scale4567);
354 #else
355 vout0x4567 = vmlaq_f32(vbias4567, vout0x4567, vfilter_output_scale4567);
356 vout1x4567 = vmlaq_f32(vbias4567, vout1x4567, vfilter_output_scale4567);
357 vout2x4567 = vmlaq_f32(vbias4567, vout2x4567, vfilter_output_scale4567);
358 vout3x4567 = vmlaq_f32(vbias4567, vout3x4567, vfilter_output_scale4567);
359 #endif
360
361 const float32x4_t voutput_min = vld1q_dup_f32(&params->scalar.min);
362 vout0x0123 = vmaxq_f32(vout0x0123, voutput_min);
363 vout0x4567 = vmaxq_f32(vout0x4567, voutput_min);
364 vout1x0123 = vmaxq_f32(vout1x0123, voutput_min);
365 vout1x4567 = vmaxq_f32(vout1x4567, voutput_min);
366 vout2x0123 = vmaxq_f32(vout2x0123, voutput_min);
367 vout2x4567 = vmaxq_f32(vout2x4567, voutput_min);
368 vout3x0123 = vmaxq_f32(vout3x0123, voutput_min);
369 vout3x4567 = vmaxq_f32(vout3x4567, voutput_min);
370
371 const float32x4_t voutput_max = vld1q_dup_f32(&params->scalar.max);
372 vout0x0123 = vminq_f32(vout0x0123, voutput_max);
373 vout0x4567 = vminq_f32(vout0x4567, voutput_max);
374 vout1x0123 = vminq_f32(vout1x0123, voutput_max);
375 vout1x4567 = vminq_f32(vout1x4567, voutput_max);
376 vout2x0123 = vminq_f32(vout2x0123, voutput_max);
377 vout2x4567 = vminq_f32(vout2x4567, voutput_max);
378 vout3x0123 = vminq_f32(vout3x0123, voutput_max);
379 vout3x4567 = vminq_f32(vout3x4567, voutput_max);
380
381 if XNN_LIKELY(nc >= 8) {
382 vst1q_f32(&c3[0], vout3x0123);
383 vst1q_f32(&c3[4], vout3x4567);
384 vst1q_f32(&c2[0], vout2x0123);
385 vst1q_f32(&c2[4], vout2x4567);
386 vst1q_f32(&c1[0], vout1x0123);
387 vst1q_f32(&c1[4], vout1x4567);
388 vst1q_f32(&c0[0], vout0x0123);
389 vst1q_f32(&c0[4], vout0x4567);
390
391 a = (const int8_t**restrict) ((uintptr_t) a - ks);
392
393 c0 = (float*) ((uintptr_t) c0 + cn_stride);
394 c1 = (float*) ((uintptr_t) c1 + cn_stride);
395 c2 = (float*) ((uintptr_t) c2 + cn_stride);
396 c3 = (float*) ((uintptr_t) c3 + cn_stride);
397
398 nc -= 8;
399 } else {
400 if (nc & 4) {
401 vst1q_f32(c3, vout3x0123); c3 += 4;
402 vout3x0123 = vout3x4567;
403 vst1q_f32(c2, vout2x0123); c2 += 4;
404 vout2x0123 = vout2x4567;
405 vst1q_f32(c1, vout1x0123); c1 += 4;
406 vout1x0123 = vout1x4567;
407 vst1q_f32(c0, vout0x0123); c0 += 4;
408 vout0x0123 = vout0x4567;
409 }
410 float32x2_t vout3x01 = vget_low_f32(vout3x0123);
411 float32x2_t vout2x01 = vget_low_f32(vout2x0123);
412 float32x2_t vout1x01 = vget_low_f32(vout1x0123);
413 float32x2_t vout0x01 = vget_low_f32(vout0x0123);
414 if (nc & 2) {
415 vst1_f32(c3, vout3x01); c3 += 2;
416 vst1_f32(c2, vout2x01); c2 += 2;
417 vst1_f32(c1, vout1x01); c1 += 2;
418 vst1_f32(c0, vout0x01); c0 += 2;
419 vout3x01 = vget_high_f32(vout3x0123);
420 vout2x01 = vget_high_f32(vout2x0123);
421 vout1x01 = vget_high_f32(vout1x0123);
422 vout0x01 = vget_high_f32(vout0x0123);
423 }
424 if (nc & 1) {
425 vst1_lane_f32(c3, vout3x01, 0);
426 vst1_lane_f32(c2, vout2x01, 0);
427 vst1_lane_f32(c1, vout1x01, 0);
428 vst1_lane_f32(c0, vout0x01, 0);
429 }
430 nc = 0;
431 }
432 } while (nc != 0);
433 }