Debian Codegraph @

src/qd8-f32-qc8w-igemm/gen/qd8-f32-qc8w-igemm-4x8-minmax-neon-mlal-lane-prfm.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 #include "xnnpack/prefetch.h"
18
19
20 void xnn_qd8_f32_qc8w_igemm_minmax_ukernel_4x8__neon_mlal_lane_prfm(
21 size_t mr,
22 size_t nc,
23 size_t kc,
24 size_t ks,
25 const int8_t** restrict a,
26 const void* restrict w,
27 float* restrict c,
28 size_t cm_stride,
29 size_t cn_stride,
30 size_t a_offset,
31 const int8_t* zero,
32 const int8_t* zero_data,
33 const union xnn_f32_minmax_params params[restrict XNN_MIN_ELEMENTS(1)],
34 const struct xnn_qd8_quantization_params quantization_params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
35 {
36 assert(mr != 0);
37 assert(mr <= 4);
38 assert(nc != 0);
39 assert(kc != 0);
40 assert(ks != 0);
41 assert(ks % (4 * sizeof(void*)) == 0);
42 assert(a_offset % sizeof(int8_t) == 0);
43 assert(a != NULL);
44 assert(w != NULL);
45 assert(c != NULL);
46
47 float* c0 = c;
48 float* c1 = (float*) ((uintptr_t) c0 + cm_stride);
49 if XNN_UNPREDICTABLE(mr < 2) {
50 c1 = c0;
51 }
52 float* c2 = (float*) ((uintptr_t) c1 + cm_stride);
53 if XNN_UNPREDICTABLE(mr <= 2) {
54 c2 = c1;
55 }
56 float* c3 = (float*) ((uintptr_t) c2 + cm_stride);
57 if XNN_UNPREDICTABLE(mr != 4) {
58 c3 = c2;
59 }
60
61 do {
62 int32x4_t vksum0123 = vld1q_s32(w); w = (const int32_t*) w + 4;
63 int32x4_t vksum4567 = vld1q_s32(w); w = (const int32_t*) w + 4;
64 const int32x4_t vinput_zero_point = vld1q_dup_s32(&quantization_params->zero_point);
65 int32x4_t vacc0x0123 = vmulq_s32(vksum0123, vinput_zero_point);
66 int32x4_t vacc0x4567 = vmulq_s32(vksum4567, vinput_zero_point);
67 int32x4_t vacc1x0123 = vacc0x0123;
68 int32x4_t vacc1x4567 = vacc0x4567;
69 int32x4_t vacc2x0123 = vacc0x0123;
70 int32x4_t vacc2x4567 = vacc0x4567;
71 int32x4_t vacc3x0123 = vacc0x0123;
72 int32x4_t vacc3x4567 = vacc0x4567;
73
74 size_t p = ks;
75 do {
76 const int8_t* restrict a0 = a[0];
77 if XNN_UNPREDICTABLE(a0 != zero) {
78 a0 = (const int8_t*) ((uintptr_t) a0 + a_offset);
79 } else {
80 a0 = zero_data;
81 }
82 const int8_t* restrict a1 = a[1];
83 if XNN_UNPREDICTABLE(a1 != zero) {
84 a1 = (const int8_t*) ((uintptr_t) a1 + a_offset);
85 } else {
86 a1 = zero_data;
87 }
88 const int8_t* restrict a2 = a[2];
89 if XNN_UNPREDICTABLE(a2 != zero) {
90 a2 = (const int8_t*) ((uintptr_t) a2 + a_offset);
91 } else {
92 a2 = zero_data;
93 }
94 const int8_t* restrict a3 = a[3];
95 if XNN_UNPREDICTABLE(a3 != zero) {
96 a3 = (const int8_t*) ((uintptr_t) a3 + a_offset);
97 } else {
98 a3 = zero_data;
99 }
100 a += 4;
101
102 size_t k = kc;
103 while (k >= 8 * sizeof(int8_t)) {
104 const int8x8_t va0 = vld1_s8(a0); a0 += 8;
105 const int16x8_t vxa0 = vmovl_s8(va0);
106 const int8x8_t va1 = vld1_s8(a1); a1 += 8;
107 const int16x8_t vxa1 = vmovl_s8(va1);
108 const int8x8_t va2 = vld1_s8(a2); a2 += 8;
109 const int16x8_t vxa2 = vmovl_s8(va2);
110 const int8x8_t va3 = vld1_s8(a3); a3 += 8;
111 const int16x8_t vxa3 = vmovl_s8(va3);
112
113 const int8x8_t vb01234567c0 = vld1_s8(w); w = (const int8_t*) w + 8;
114 const int16x8_t vxb01234567c0 = vmovl_s8(vb01234567c0);
115
116 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c0), vget_low_s16(vxa0), 0);
117 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c0), vget_low_s16(vxa0), 0);
118 vacc1x0123 = vmlal_lane_s16(vacc1x0123, vget_low_s16(vxb01234567c0), vget_low_s16(vxa1), 0);
119 vacc1x4567 = vmlal_lane_s16(vacc1x4567, vget_high_s16(vxb01234567c0), vget_low_s16(vxa1), 0);
120 vacc2x0123 = vmlal_lane_s16(vacc2x0123, vget_low_s16(vxb01234567c0), vget_low_s16(vxa2), 0);
121 vacc2x4567 = vmlal_lane_s16(vacc2x4567, vget_high_s16(vxb01234567c0), vget_low_s16(vxa2), 0);
122 vacc3x0123 = vmlal_lane_s16(vacc3x0123, vget_low_s16(vxb01234567c0), vget_low_s16(vxa3), 0);
123 vacc3x4567 = vmlal_lane_s16(vacc3x4567, vget_high_s16(vxb01234567c0), vget_low_s16(vxa3), 0);
124 const int8x8_t vb01234567c1 = vld1_s8(w); w = (const int8_t*) w + 8;
125 const int16x8_t vxb01234567c1 = vmovl_s8(vb01234567c1);
126
127 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c1), vget_low_s16(vxa0), 1);
128 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c1), vget_low_s16(vxa0), 1);
129 vacc1x0123 = vmlal_lane_s16(vacc1x0123, vget_low_s16(vxb01234567c1), vget_low_s16(vxa1), 1);
130 vacc1x4567 = vmlal_lane_s16(vacc1x4567, vget_high_s16(vxb01234567c1), vget_low_s16(vxa1), 1);
131 vacc2x0123 = vmlal_lane_s16(vacc2x0123, vget_low_s16(vxb01234567c1), vget_low_s16(vxa2), 1);
132 vacc2x4567 = vmlal_lane_s16(vacc2x4567, vget_high_s16(vxb01234567c1), vget_low_s16(vxa2), 1);
133 vacc3x0123 = vmlal_lane_s16(vacc3x0123, vget_low_s16(vxb01234567c1), vget_low_s16(vxa3), 1);
134 vacc3x4567 = vmlal_lane_s16(vacc3x4567, vget_high_s16(vxb01234567c1), vget_low_s16(vxa3), 1);
135 const int8x8_t vb01234567c2 = vld1_s8(w); w = (const int8_t*) w + 8;
136 const int16x8_t vxb01234567c2 = vmovl_s8(vb01234567c2);
137
138 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c2), vget_low_s16(vxa0), 2);
139 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c2), vget_low_s16(vxa0), 2);
140 vacc1x0123 = vmlal_lane_s16(vacc1x0123, vget_low_s16(vxb01234567c2), vget_low_s16(vxa1), 2);
141 vacc1x4567 = vmlal_lane_s16(vacc1x4567, vget_high_s16(vxb01234567c2), vget_low_s16(vxa1), 2);
142 vacc2x0123 = vmlal_lane_s16(vacc2x0123, vget_low_s16(vxb01234567c2), vget_low_s16(vxa2), 2);
143 vacc2x4567 = vmlal_lane_s16(vacc2x4567, vget_high_s16(vxb01234567c2), vget_low_s16(vxa2), 2);
144 vacc3x0123 = vmlal_lane_s16(vacc3x0123, vget_low_s16(vxb01234567c2), vget_low_s16(vxa3), 2);
145 vacc3x4567 = vmlal_lane_s16(vacc3x4567, vget_high_s16(vxb01234567c2), vget_low_s16(vxa3), 2);
146 const int8x8_t vb01234567c3 = vld1_s8(w); w = (const int8_t*) w + 8;
147 const int16x8_t vxb01234567c3 = vmovl_s8(vb01234567c3);
148
149 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c3), vget_low_s16(vxa0), 3);
150 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c3), vget_low_s16(vxa0), 3);
151 vacc1x0123 = vmlal_lane_s16(vacc1x0123, vget_low_s16(vxb01234567c3), vget_low_s16(vxa1), 3);
152 vacc1x4567 = vmlal_lane_s16(vacc1x4567, vget_high_s16(vxb01234567c3), vget_low_s16(vxa1), 3);
153 vacc2x0123 = vmlal_lane_s16(vacc2x0123, vget_low_s16(vxb01234567c3), vget_low_s16(vxa2), 3);
154 vacc2x4567 = vmlal_lane_s16(vacc2x4567, vget_high_s16(vxb01234567c3), vget_low_s16(vxa2), 3);
155 vacc3x0123 = vmlal_lane_s16(vacc3x0123, vget_low_s16(vxb01234567c3), vget_low_s16(vxa3), 3);
156 vacc3x4567 = vmlal_lane_s16(vacc3x4567, vget_high_s16(vxb01234567c3), vget_low_s16(vxa3), 3);
157
158 xnn_prefetch_to_l1((const int8_t*) w + 448);
159
160 const int8x8_t vb01234567c4 = vld1_s8(w); w = (const int8_t*) w + 8;
161 const int16x8_t vxb01234567c4 = vmovl_s8(vb01234567c4);
162
163 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c4), vget_high_s16(vxa0), 0);
164 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c4), vget_high_s16(vxa0), 0);
165 vacc1x0123 = vmlal_lane_s16(vacc1x0123, vget_low_s16(vxb01234567c4), vget_high_s16(vxa1), 0);
166 vacc1x4567 = vmlal_lane_s16(vacc1x4567, vget_high_s16(vxb01234567c4), vget_high_s16(vxa1), 0);
167 vacc2x0123 = vmlal_lane_s16(vacc2x0123, vget_low_s16(vxb01234567c4), vget_high_s16(vxa2), 0);
168 vacc2x4567 = vmlal_lane_s16(vacc2x4567, vget_high_s16(vxb01234567c4), vget_high_s16(vxa2), 0);
169 vacc3x0123 = vmlal_lane_s16(vacc3x0123, vget_low_s16(vxb01234567c4), vget_high_s16(vxa3), 0);
170 vacc3x4567 = vmlal_lane_s16(vacc3x4567, vget_high_s16(vxb01234567c4), vget_high_s16(vxa3), 0);
171 const int8x8_t vb01234567c5 = vld1_s8(w); w = (const int8_t*) w + 8;
172 const int16x8_t vxb01234567c5 = vmovl_s8(vb01234567c5);
173
174 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c5), vget_high_s16(vxa0), 1);
175 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c5), vget_high_s16(vxa0), 1);
176 vacc1x0123 = vmlal_lane_s16(vacc1x0123, vget_low_s16(vxb01234567c5), vget_high_s16(vxa1), 1);
177 vacc1x4567 = vmlal_lane_s16(vacc1x4567, vget_high_s16(vxb01234567c5), vget_high_s16(vxa1), 1);
178 vacc2x0123 = vmlal_lane_s16(vacc2x0123, vget_low_s16(vxb01234567c5), vget_high_s16(vxa2), 1);
179 vacc2x4567 = vmlal_lane_s16(vacc2x4567, vget_high_s16(vxb01234567c5), vget_high_s16(vxa2), 1);
180 vacc3x0123 = vmlal_lane_s16(vacc3x0123, vget_low_s16(vxb01234567c5), vget_high_s16(vxa3), 1);
181 vacc3x4567 = vmlal_lane_s16(vacc3x4567, vget_high_s16(vxb01234567c5), vget_high_s16(vxa3), 1);
182 const int8x8_t vb01234567c6 = vld1_s8(w); w = (const int8_t*) w + 8;
183 const int16x8_t vxb01234567c6 = vmovl_s8(vb01234567c6);
184
185 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c6), vget_high_s16(vxa0), 2);
186 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c6), vget_high_s16(vxa0), 2);
187 vacc1x0123 = vmlal_lane_s16(vacc1x0123, vget_low_s16(vxb01234567c6), vget_high_s16(vxa1), 2);
188 vacc1x4567 = vmlal_lane_s16(vacc1x4567, vget_high_s16(vxb01234567c6), vget_high_s16(vxa1), 2);
189 vacc2x0123 = vmlal_lane_s16(vacc2x0123, vget_low_s16(vxb01234567c6), vget_high_s16(vxa2), 2);
190 vacc2x4567 = vmlal_lane_s16(vacc2x4567, vget_high_s16(vxb01234567c6), vget_high_s16(vxa2), 2);
191 vacc3x0123 = vmlal_lane_s16(vacc3x0123, vget_low_s16(vxb01234567c6), vget_high_s16(vxa3), 2);
192 vacc3x4567 = vmlal_lane_s16(vacc3x4567, vget_high_s16(vxb01234567c6), vget_high_s16(vxa3), 2);
193 const int8x8_t vb01234567c7 = vld1_s8(w); w = (const int8_t*) w + 8;
194 const int16x8_t vxb01234567c7 = vmovl_s8(vb01234567c7);
195
196 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c7), vget_high_s16(vxa0), 3);
197 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c7), vget_high_s16(vxa0), 3);
198 vacc1x0123 = vmlal_lane_s16(vacc1x0123, vget_low_s16(vxb01234567c7), vget_high_s16(vxa1), 3);
199 vacc1x4567 = vmlal_lane_s16(vacc1x4567, vget_high_s16(vxb01234567c7), vget_high_s16(vxa1), 3);
200 vacc2x0123 = vmlal_lane_s16(vacc2x0123, vget_low_s16(vxb01234567c7), vget_high_s16(vxa2), 3);
201 vacc2x4567 = vmlal_lane_s16(vacc2x4567, vget_high_s16(vxb01234567c7), vget_high_s16(vxa2), 3);
202 vacc3x0123 = vmlal_lane_s16(vacc3x0123, vget_low_s16(vxb01234567c7), vget_high_s16(vxa3), 3);
203 vacc3x4567 = vmlal_lane_s16(vacc3x4567, vget_high_s16(vxb01234567c7), vget_high_s16(vxa3), 3);
204
205 k -= 8 * sizeof(int8_t);
206 }
207 if XNN_UNLIKELY(k != 0) {
208 const int8x8_t va0 = vld1_s8(a0); a0 = (const int8_t*) ((uintptr_t) a0 + k);
209 const int16x8_t vxa0 = vmovl_s8(va0);
210 const int8x8_t va1 = vld1_s8(a1); a1 = (const int8_t*) ((uintptr_t) a1 + k);
211 const int16x8_t vxa1 = vmovl_s8(va1);
212 const int8x8_t va2 = vld1_s8(a2); a2 = (const int8_t*) ((uintptr_t) a2 + k);
213 const int16x8_t vxa2 = vmovl_s8(va2);
214 const int8x8_t va3 = vld1_s8(a3); a3 = (const int8_t*) ((uintptr_t) a3 + k);
215 const int16x8_t vxa3 = vmovl_s8(va3);
216
217 const int8x8_t vb01234567c0 = vld1_s8(w); w = (const int8_t*) w + 8;
218 const int16x8_t vxb01234567c0 = vmovl_s8(vb01234567c0);
219
220 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c0), vget_low_s16(vxa0), 0);
221 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c0), vget_low_s16(vxa0), 0);
222 vacc1x0123 = vmlal_lane_s16(vacc1x0123, vget_low_s16(vxb01234567c0), vget_low_s16(vxa1), 0);
223 vacc1x4567 = vmlal_lane_s16(vacc1x4567, vget_high_s16(vxb01234567c0), vget_low_s16(vxa1), 0);
224 vacc2x0123 = vmlal_lane_s16(vacc2x0123, vget_low_s16(vxb01234567c0), vget_low_s16(vxa2), 0);
225 vacc2x4567 = vmlal_lane_s16(vacc2x4567, vget_high_s16(vxb01234567c0), vget_low_s16(vxa2), 0);
226 vacc3x0123 = vmlal_lane_s16(vacc3x0123, vget_low_s16(vxb01234567c0), vget_low_s16(vxa3), 0);
227 vacc3x4567 = vmlal_lane_s16(vacc3x4567, vget_high_s16(vxb01234567c0), vget_low_s16(vxa3), 0);
228
229 if (k >= 2 * sizeof(int8_t)) {
230 const int8x8_t vb01234567c1 = vld1_s8(w); w = (const int8_t*) w + 8;
231 const int16x8_t vxb01234567c1 = vmovl_s8(vb01234567c1);
232
233 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c1), vget_low_s16(vxa0), 1);
234 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c1), vget_low_s16(vxa0), 1);
235 vacc1x0123 = vmlal_lane_s16(vacc1x0123, vget_low_s16(vxb01234567c1), vget_low_s16(vxa1), 1);
236 vacc1x4567 = vmlal_lane_s16(vacc1x4567, vget_high_s16(vxb01234567c1), vget_low_s16(vxa1), 1);
237 vacc2x0123 = vmlal_lane_s16(vacc2x0123, vget_low_s16(vxb01234567c1), vget_low_s16(vxa2), 1);
238 vacc2x4567 = vmlal_lane_s16(vacc2x4567, vget_high_s16(vxb01234567c1), vget_low_s16(vxa2), 1);
239 vacc3x0123 = vmlal_lane_s16(vacc3x0123, vget_low_s16(vxb01234567c1), vget_low_s16(vxa3), 1);
240 vacc3x4567 = vmlal_lane_s16(vacc3x4567, vget_high_s16(vxb01234567c1), vget_low_s16(vxa3), 1);
241
242 if (k > 2 * sizeof(int8_t)) {
243 const int8x8_t vb01234567c2 = vld1_s8(w); w = (const int8_t*) w + 8;
244 const int16x8_t vxb01234567c2 = vmovl_s8(vb01234567c2);
245
246 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c2), vget_low_s16(vxa0), 2);
247 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c2), vget_low_s16(vxa0), 2);
248 vacc1x0123 = vmlal_lane_s16(vacc1x0123, vget_low_s16(vxb01234567c2), vget_low_s16(vxa1), 2);
249 vacc1x4567 = vmlal_lane_s16(vacc1x4567, vget_high_s16(vxb01234567c2), vget_low_s16(vxa1), 2);
250 vacc2x0123 = vmlal_lane_s16(vacc2x0123, vget_low_s16(vxb01234567c2), vget_low_s16(vxa2), 2);
251 vacc2x4567 = vmlal_lane_s16(vacc2x4567, vget_high_s16(vxb01234567c2), vget_low_s16(vxa2), 2);
252 vacc3x0123 = vmlal_lane_s16(vacc3x0123, vget_low_s16(vxb01234567c2), vget_low_s16(vxa3), 2);
253 vacc3x4567 = vmlal_lane_s16(vacc3x4567, vget_high_s16(vxb01234567c2), vget_low_s16(vxa3), 2);
254
255 if (k >= 4 * sizeof(int8_t)) {
256 const int8x8_t vb01234567c3 = vld1_s8(w); w = (const int8_t*) w + 8;
257 const int16x8_t vxb01234567c3 = vmovl_s8(vb01234567c3);
258
259 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c3), vget_low_s16(vxa0), 3);
260 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c3), vget_low_s16(vxa0), 3);
261 vacc1x0123 = vmlal_lane_s16(vacc1x0123, vget_low_s16(vxb01234567c3), vget_low_s16(vxa1), 3);
262 vacc1x4567 = vmlal_lane_s16(vacc1x4567, vget_high_s16(vxb01234567c3), vget_low_s16(vxa1), 3);
263 vacc2x0123 = vmlal_lane_s16(vacc2x0123, vget_low_s16(vxb01234567c3), vget_low_s16(vxa2), 3);
264 vacc2x4567 = vmlal_lane_s16(vacc2x4567, vget_high_s16(vxb01234567c3), vget_low_s16(vxa2), 3);
265 vacc3x0123 = vmlal_lane_s16(vacc3x0123, vget_low_s16(vxb01234567c3), vget_low_s16(vxa3), 3);
266 vacc3x4567 = vmlal_lane_s16(vacc3x4567, vget_high_s16(vxb01234567c3), vget_low_s16(vxa3), 3);
267
268 if (k > 4 * sizeof(int8_t)) {
269 const int8x8_t vb01234567c4 = vld1_s8(w); w = (const int8_t*) w + 8;
270 const int16x8_t vxb01234567c4 = vmovl_s8(vb01234567c4);
271
272 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c4), vget_high_s16(vxa0), 0);
273 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c4), vget_high_s16(vxa0), 0);
274 vacc1x0123 = vmlal_lane_s16(vacc1x0123, vget_low_s16(vxb01234567c4), vget_high_s16(vxa1), 0);
275 vacc1x4567 = vmlal_lane_s16(vacc1x4567, vget_high_s16(vxb01234567c4), vget_high_s16(vxa1), 0);
276 vacc2x0123 = vmlal_lane_s16(vacc2x0123, vget_low_s16(vxb01234567c4), vget_high_s16(vxa2), 0);
277 vacc2x4567 = vmlal_lane_s16(vacc2x4567, vget_high_s16(vxb01234567c4), vget_high_s16(vxa2), 0);
278 vacc3x0123 = vmlal_lane_s16(vacc3x0123, vget_low_s16(vxb01234567c4), vget_high_s16(vxa3), 0);
279 vacc3x4567 = vmlal_lane_s16(vacc3x4567, vget_high_s16(vxb01234567c4), vget_high_s16(vxa3), 0);
280
281 if (k >= 6 * sizeof(int8_t)) {
282 const int8x8_t vb01234567c5 = vld1_s8(w); w = (const int8_t*) w + 8;
283 const int16x8_t vxb01234567c5 = vmovl_s8(vb01234567c5);
284
285 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c5), vget_high_s16(vxa0), 1);
286 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c5), vget_high_s16(vxa0), 1);
287 vacc1x0123 = vmlal_lane_s16(vacc1x0123, vget_low_s16(vxb01234567c5), vget_high_s16(vxa1), 1);
288 vacc1x4567 = vmlal_lane_s16(vacc1x4567, vget_high_s16(vxb01234567c5), vget_high_s16(vxa1), 1);
289 vacc2x0123 = vmlal_lane_s16(vacc2x0123, vget_low_s16(vxb01234567c5), vget_high_s16(vxa2), 1);
290 vacc2x4567 = vmlal_lane_s16(vacc2x4567, vget_high_s16(vxb01234567c5), vget_high_s16(vxa2), 1);
291 vacc3x0123 = vmlal_lane_s16(vacc3x0123, vget_low_s16(vxb01234567c5), vget_high_s16(vxa3), 1);
292 vacc3x4567 = vmlal_lane_s16(vacc3x4567, vget_high_s16(vxb01234567c5), vget_high_s16(vxa3), 1);
293
294 if (k > 6 * sizeof(int8_t)) {
295 const int8x8_t vb01234567c6 = vld1_s8(w); w = (const int8_t*) w + 8;
296 const int16x8_t vxb01234567c6 = vmovl_s8(vb01234567c6);
297
298 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c6), vget_high_s16(vxa0), 2);
299 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c6), vget_high_s16(vxa0), 2);
300 vacc1x0123 = vmlal_lane_s16(vacc1x0123, vget_low_s16(vxb01234567c6), vget_high_s16(vxa1), 2);
301 vacc1x4567 = vmlal_lane_s16(vacc1x4567, vget_high_s16(vxb01234567c6), vget_high_s16(vxa1), 2);
302 vacc2x0123 = vmlal_lane_s16(vacc2x0123, vget_low_s16(vxb01234567c6), vget_high_s16(vxa2), 2);
303 vacc2x4567 = vmlal_lane_s16(vacc2x4567, vget_high_s16(vxb01234567c6), vget_high_s16(vxa2), 2);
304 vacc3x0123 = vmlal_lane_s16(vacc3x0123, vget_low_s16(vxb01234567c6), vget_high_s16(vxa3), 2);
305 vacc3x4567 = vmlal_lane_s16(vacc3x4567, vget_high_s16(vxb01234567c6), vget_high_s16(vxa3), 2);
306 }
307 }
308 }
309 }
310 }
311 }
312 }
313 p -= 4 * sizeof(void*);
314 } while (p != 0);
315
316 float32x4_t vout0x0123 = vcvtq_f32_s32(vacc0x0123);
317 float32x4_t vout0x4567 = vcvtq_f32_s32(vacc0x4567);
318 float32x4_t vout1x0123 = vcvtq_f32_s32(vacc1x0123);
319 float32x4_t vout1x4567 = vcvtq_f32_s32(vacc1x4567);
320 float32x4_t vout2x0123 = vcvtq_f32_s32(vacc2x0123);
321 float32x4_t vout2x4567 = vcvtq_f32_s32(vacc2x4567);
322 float32x4_t vout3x0123 = vcvtq_f32_s32(vacc3x0123);
323 float32x4_t vout3x4567 = vcvtq_f32_s32(vacc3x4567);
324
325 const float32x4_t vinput_scale = vld1q_dup_f32(&quantization_params->inv_scale);
326 vout0x0123 = vmulq_f32(vout0x0123, vinput_scale);
327 vout0x4567 = vmulq_f32(vout0x4567, vinput_scale);
328 vout1x0123 = vmulq_f32(vout1x0123, vinput_scale);
329 vout1x4567 = vmulq_f32(vout1x4567, vinput_scale);
330 vout2x0123 = vmulq_f32(vout2x0123, vinput_scale);
331 vout2x4567 = vmulq_f32(vout2x4567, vinput_scale);
332 vout3x0123 = vmulq_f32(vout3x0123, vinput_scale);
333 vout3x4567 = vmulq_f32(vout3x4567, vinput_scale);
334
335 const float32x4_t vfilter_output_scale0123 = vld1q_f32(w); w = (const float*) w + 4;
336 const float32x4_t vfilter_output_scale4567 = vld1q_f32(w); w = (const float*) w + 4;
337
338 const float32x4_t vbias0123 = vld1q_f32(w); w = (const float*) w + 4;
339 #if XNN_ARCH_ARM64
340 vout0x0123 = vfmaq_f32(vbias0123, vout0x0123, vfilter_output_scale0123);
341 vout1x0123 = vfmaq_f32(vbias0123, vout1x0123, vfilter_output_scale0123);
342 vout2x0123 = vfmaq_f32(vbias0123, vout2x0123, vfilter_output_scale0123);
343 vout3x0123 = vfmaq_f32(vbias0123, vout3x0123, vfilter_output_scale0123);
344 #else
345 vout0x0123 = vmlaq_f32(vbias0123, vout0x0123, vfilter_output_scale0123);
346 vout1x0123 = vmlaq_f32(vbias0123, vout1x0123, vfilter_output_scale0123);
347 vout2x0123 = vmlaq_f32(vbias0123, vout2x0123, vfilter_output_scale0123);
348 vout3x0123 = vmlaq_f32(vbias0123, vout3x0123, vfilter_output_scale0123);
349 #endif
350 const float32x4_t vbias4567 = vld1q_f32(w); w = (const float*) w + 4;
351 #if XNN_ARCH_ARM64
352 vout0x4567 = vfmaq_f32(vbias4567, vout0x4567, vfilter_output_scale4567);
353 vout1x4567 = vfmaq_f32(vbias4567, vout1x4567, vfilter_output_scale4567);
354 vout2x4567 = vfmaq_f32(vbias4567, vout2x4567, vfilter_output_scale4567);
355 vout3x4567 = vfmaq_f32(vbias4567, vout3x4567, vfilter_output_scale4567);
356 #else
357 vout0x4567 = vmlaq_f32(vbias4567, vout0x4567, vfilter_output_scale4567);
358 vout1x4567 = vmlaq_f32(vbias4567, vout1x4567, vfilter_output_scale4567);
359 vout2x4567 = vmlaq_f32(vbias4567, vout2x4567, vfilter_output_scale4567);
360 vout3x4567 = vmlaq_f32(vbias4567, vout3x4567, vfilter_output_scale4567);
361 #endif
362
363 const float32x4_t voutput_min = vld1q_dup_f32(&params->scalar.min);
364 vout0x0123 = vmaxq_f32(vout0x0123, voutput_min);
365 vout0x4567 = vmaxq_f32(vout0x4567, voutput_min);
366 vout1x0123 = vmaxq_f32(vout1x0123, voutput_min);
367 vout1x4567 = vmaxq_f32(vout1x4567, voutput_min);
368 vout2x0123 = vmaxq_f32(vout2x0123, voutput_min);
369 vout2x4567 = vmaxq_f32(vout2x4567, voutput_min);
370 vout3x0123 = vmaxq_f32(vout3x0123, voutput_min);
371 vout3x4567 = vmaxq_f32(vout3x4567, voutput_min);
372
373 const float32x4_t voutput_max = vld1q_dup_f32(&params->scalar.max);
374 vout0x0123 = vminq_f32(vout0x0123, voutput_max);
375 vout0x4567 = vminq_f32(vout0x4567, voutput_max);
376 vout1x0123 = vminq_f32(vout1x0123, voutput_max);
377 vout1x4567 = vminq_f32(vout1x4567, voutput_max);
378 vout2x0123 = vminq_f32(vout2x0123, voutput_max);
379 vout2x4567 = vminq_f32(vout2x4567, voutput_max);
380 vout3x0123 = vminq_f32(vout3x0123, voutput_max);
381 vout3x4567 = vminq_f32(vout3x4567, voutput_max);
382
383 if XNN_LIKELY(nc >= 8) {
384 vst1q_f32(&c3[0], vout3x0123);
385 vst1q_f32(&c3[4], vout3x4567);
386 vst1q_f32(&c2[0], vout2x0123);
387 vst1q_f32(&c2[4], vout2x4567);
388 vst1q_f32(&c1[0], vout1x0123);
389 vst1q_f32(&c1[4], vout1x4567);
390 vst1q_f32(&c0[0], vout0x0123);
391 vst1q_f32(&c0[4], vout0x4567);
392
393 a = (const int8_t**restrict) ((uintptr_t) a - ks);
394
395 c0 = (float*) ((uintptr_t) c0 + cn_stride);
396 c1 = (float*) ((uintptr_t) c1 + cn_stride);
397 c2 = (float*) ((uintptr_t) c2 + cn_stride);
398 c3 = (float*) ((uintptr_t) c3 + cn_stride);
399
400 nc -= 8;
401 } else {
402 if (nc & 4) {
403 vst1q_f32(c3, vout3x0123); c3 += 4;
404 vout3x0123 = vout3x4567;
405 vst1q_f32(c2, vout2x0123); c2 += 4;
406 vout2x0123 = vout2x4567;
407 vst1q_f32(c1, vout1x0123); c1 += 4;
408 vout1x0123 = vout1x4567;
409 vst1q_f32(c0, vout0x0123); c0 += 4;
410 vout0x0123 = vout0x4567;
411 }
412 float32x2_t vout3x01 = vget_low_f32(vout3x0123);
413 float32x2_t vout2x01 = vget_low_f32(vout2x0123);
414 float32x2_t vout1x01 = vget_low_f32(vout1x0123);
415 float32x2_t vout0x01 = vget_low_f32(vout0x0123);
416 if (nc & 2) {
417 vst1_f32(c3, vout3x01); c3 += 2;
418 vst1_f32(c2, vout2x01); c2 += 2;
419 vst1_f32(c1, vout1x01); c1 += 2;
420 vst1_f32(c0, vout0x01); c0 += 2;
421 vout3x01 = vget_high_f32(vout3x0123);
422 vout2x01 = vget_high_f32(vout2x0123);
423 vout1x01 = vget_high_f32(vout1x0123);
424 vout0x01 = vget_high_f32(vout0x0123);
425 }
426 if (nc & 1) {
427 vst1_lane_f32(c3, vout3x01, 0);
428 vst1_lane_f32(c2, vout2x01, 0);
429 vst1_lane_f32(c1, vout1x01, 0);
430 vst1_lane_f32(c0, vout0x01, 0);
431 }
432 nc = 0;
433 }
434 } while (nc != 0);
435 }