Gitiles
Code Review Sign In
gerrit-public.fairphone.software / platform / external / XNNPACK / 27bf92c9919f6dc6b5d0dece78d958563e2d79b7 / . / src / qs8-igemm / gen / 2x8c2s4-minmax-fp32-neon-mlal.c
blob: 4c647cc03124957c5daed0301a43744ed8a9455c [file] [log] [blame]
Frank Barchard494cd2b2021-10-28 17:36:37 -0700[diff] [blame]1// Auto-generated file. Do not edit!
Frank Barcharde22685a2021-11-12 11:36:58 -0800[diff] [blame]2// Template: src/qs8-igemm/c2-neon-mull-shuffle.c.in
Frank Barchard494cd2b2021-10-28 17:36:37 -0700[diff] [blame]3// Generator: tools/xngen
4//
5// Copyright 2021 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
Frank Barcharde22685a2021-11-12 11:36:58 -0800[diff] [blame]18void xnn_qs8_igemm_minmax_fp32_ukernel_2x8c2s4__neon_mlal(
Frank Barchard494cd2b2021-10-28 17:36:37 -0700[diff] [blame]19 size_t mr,
20 size_t nc,
21 size_t kc,
22 size_t ks,
23 const int8_t** restrict a,
24 const void* restrict w,
25 int8_t* restrict c,
26 size_t cm_stride,
27 size_t cn_stride,
28 size_t a_offset,
29 const int8_t* zero,
30 const union xnn_qs8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_DISABLE_TSAN XNN_DISABLE_MSAN
31{
32 assert(mr != 0);
33 assert(mr <= 2);
34 assert(nc != 0);
35 assert(kc != 0);
36 assert(ks != 0);
37 assert(ks % (2 * sizeof(void*)) == 0);
38 assert(a_offset % sizeof(int8_t) == 0);
39 assert(a != NULL);
40 assert(w != NULL);
41 assert(c != NULL);
42
43 int8_t* c0 = c;
44 int8_t* c1 = (int8_t*) ((uintptr_t) c0 + cm_stride);
45 if XNN_UNPREDICTABLE(mr != 2) {
46 c1 = c0;
47 }
48
49 do {
50 int32x4_t vacc0x0123 = vld1q_s32(w); w = (const void*) ((uintptr_t) w + 4 * sizeof(int32_t));
51 int32x4_t vacc0x4567 = vld1q_s32(w); w = (const void*) ((uintptr_t) w + 4 * sizeof(int32_t));
52 int32x4_t vacc1x0123 = vacc0x0123;
53 int32x4_t vacc1x4567 = vacc0x4567;
54
55 size_t p = ks;
56 do {
57 const int8_t* restrict a0 = a[0];
58 if XNN_UNPREDICTABLE(a0 != zero) {
59 a0 = (const int8_t*) ((uintptr_t) a0 + a_offset);
60 }
61 const int8_t* restrict a1 = a[1];
62 if XNN_UNPREDICTABLE(a1 != zero) {
63 a1 = (const int8_t*) ((uintptr_t) a1 + a_offset);
64 }
65 a += 2;
66
67 size_t k = kc;
68
69 while (k >= 16 * sizeof(int8_t)) {
70 int8x8_t va0x0 = vld1_s8(a0); a0 += 8;
71 int8x8_t va0x1 = vld1_s8(a0); a0 += 8;
72 int8x8_t va1x0 = vld1_s8(a1); a1 += 8;
73 int8x8_t va1x1 = vld1_s8(a1); a1 += 8;
74
75 const int8x8_t vb0123c0x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
76 const int8x8_t vb4567c0x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
77 const int8x8_t vb0123c1x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
78 const int8x8_t vb4567c1x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
79 const int8x8_t vb0123c2x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
80 const int8x8_t vb4567c2x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
81 const int8x8_t vb0123c3x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
82 const int8x8_t vb4567c3x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
83
84 int16x8_t vprod0x0123c0 = vmull_s8(vb0123c0x0, va0x0);
85 int16x8_t vprod1x0123c0 = vmull_s8(vb0123c0x0, va1x0);
86 const int8x8_t vb0123c0x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
87 vprod0x0123c0 = vmlal_s8(vprod0x0123c0, vb0123c0x1, va0x1);
88 vprod1x0123c0 = vmlal_s8(vprod1x0123c0, vb0123c0x1, va1x1);
89 vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c0);
90 vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c0);
91 int16x8_t vprod0x4567c0 = vmull_s8(vb4567c0x0, va0x0);
92 int16x8_t vprod1x4567c0 = vmull_s8(vb4567c0x0, va1x0);
93 const int8x8_t vb4567c0x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
94 vprod0x4567c0 = vmlal_s8(vprod0x4567c0, vb4567c0x1, va0x1);
95 vprod1x4567c0 = vmlal_s8(vprod1x4567c0, vb4567c0x1, va1x1);
96 vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c0);
97 vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c0);
98 va0x0 = vext_s8(va0x0, va0x0, 2);
99 va0x1 = vext_s8(va0x1, va0x1, 2);
100 va1x0 = vext_s8(va1x0, va1x0, 2);
101 va1x1 = vext_s8(va1x1, va1x1, 2);
102 int16x8_t vprod0x0123c1 = vmull_s8(vb0123c1x0, va0x0);
103 int16x8_t vprod1x0123c1 = vmull_s8(vb0123c1x0, va1x0);
104 const int8x8_t vb0123c1x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
105 vprod0x0123c1 = vmlal_s8(vprod0x0123c1, vb0123c1x1, va0x1);
106 vprod1x0123c1 = vmlal_s8(vprod1x0123c1, vb0123c1x1, va1x1);
107 vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c1);
108 vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c1);
109 int16x8_t vprod0x4567c1 = vmull_s8(vb4567c1x0, va0x0);
110 int16x8_t vprod1x4567c1 = vmull_s8(vb4567c1x0, va1x0);
111 const int8x8_t vb4567c1x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
112 vprod0x4567c1 = vmlal_s8(vprod0x4567c1, vb4567c1x1, va0x1);
113 vprod1x4567c1 = vmlal_s8(vprod1x4567c1, vb4567c1x1, va1x1);
114 vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c1);
115 vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c1);
116 va0x0 = vext_s8(va0x0, va0x0, 2);
117 va0x1 = vext_s8(va0x1, va0x1, 2);
118 va1x0 = vext_s8(va1x0, va1x0, 2);
119 va1x1 = vext_s8(va1x1, va1x1, 2);
120 int16x8_t vprod0x0123c2 = vmull_s8(vb0123c2x0, va0x0);
121 int16x8_t vprod1x0123c2 = vmull_s8(vb0123c2x0, va1x0);
122 const int8x8_t vb0123c2x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
123 vprod0x0123c2 = vmlal_s8(vprod0x0123c2, vb0123c2x1, va0x1);
124 vprod1x0123c2 = vmlal_s8(vprod1x0123c2, vb0123c2x1, va1x1);
125 vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c2);
126 vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c2);
127 int16x8_t vprod0x4567c2 = vmull_s8(vb4567c2x0, va0x0);
128 int16x8_t vprod1x4567c2 = vmull_s8(vb4567c2x0, va1x0);
129 const int8x8_t vb4567c2x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
130 vprod0x4567c2 = vmlal_s8(vprod0x4567c2, vb4567c2x1, va0x1);
131 vprod1x4567c2 = vmlal_s8(vprod1x4567c2, vb4567c2x1, va1x1);
132 vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c2);
133 vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c2);
134 va0x0 = vext_s8(va0x0, va0x0, 2);
135 va0x1 = vext_s8(va0x1, va0x1, 2);
136 va1x0 = vext_s8(va1x0, va1x0, 2);
137 va1x1 = vext_s8(va1x1, va1x1, 2);
138 int16x8_t vprod0x0123c3 = vmull_s8(vb0123c3x0, va0x0);
139 int16x8_t vprod1x0123c3 = vmull_s8(vb0123c3x0, va1x0);
140 const int8x8_t vb0123c3x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
141 vprod0x0123c3 = vmlal_s8(vprod0x0123c3, vb0123c3x1, va0x1);
142 vprod1x0123c3 = vmlal_s8(vprod1x0123c3, vb0123c3x1, va1x1);
143 vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c3);
144 vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c3);
145 int16x8_t vprod0x4567c3 = vmull_s8(vb4567c3x0, va0x0);
146 int16x8_t vprod1x4567c3 = vmull_s8(vb4567c3x0, va1x0);
147 const int8x8_t vb4567c3x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
148 vprod0x4567c3 = vmlal_s8(vprod0x4567c3, vb4567c3x1, va0x1);
149 vprod1x4567c3 = vmlal_s8(vprod1x4567c3, vb4567c3x1, va1x1);
150 vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c3);
151 vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c3);
152
153 k -= 16 * sizeof(int8_t);
154 }
155
156 if (k >= 8 * sizeof(int8_t)) {
157 int8x8_t va0x0 = vld1_s8(a0); a0 += 8;
158 int8x8_t va1x0 = vld1_s8(a1); a1 += 8;
159
160 const int8x8_t vb0123c0x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
161 const int8x8_t vb4567c0x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
162 const int8x8_t vb0123c1x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
163 const int8x8_t vb4567c1x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
164 const int8x8_t vb0123c2x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
165 const int8x8_t vb4567c2x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
166 const int8x8_t vb0123c3x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
167 const int8x8_t vb4567c3x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
168
169 int16x8_t vprod0x0123c0 = vmull_s8(vb0123c0x0, va0x0);
170 int16x8_t vprod1x0123c0 = vmull_s8(vb0123c0x0, va1x0);
171 vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c0);
172 vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c0);
173 int16x8_t vprod0x4567c0 = vmull_s8(vb4567c0x0, va0x0);
174 int16x8_t vprod1x4567c0 = vmull_s8(vb4567c0x0, va1x0);
175 vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c0);
176 vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c0);
177 va0x0 = vext_s8(va0x0, va0x0, 2);
178 va1x0 = vext_s8(va1x0, va1x0, 2);
179 int16x8_t vprod0x0123c1 = vmull_s8(vb0123c1x0, va0x0);
180 int16x8_t vprod1x0123c1 = vmull_s8(vb0123c1x0, va1x0);
181 vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c1);
182 vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c1);
183 int16x8_t vprod0x4567c1 = vmull_s8(vb4567c1x0, va0x0);
184 int16x8_t vprod1x4567c1 = vmull_s8(vb4567c1x0, va1x0);
185 vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c1);
186 vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c1);
187 va0x0 = vext_s8(va0x0, va0x0, 2);
188 va1x0 = vext_s8(va1x0, va1x0, 2);
189 int16x8_t vprod0x0123c2 = vmull_s8(vb0123c2x0, va0x0);
190 int16x8_t vprod1x0123c2 = vmull_s8(vb0123c2x0, va1x0);
191 vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c2);
192 vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c2);
193 int16x8_t vprod0x4567c2 = vmull_s8(vb4567c2x0, va0x0);
194 int16x8_t vprod1x4567c2 = vmull_s8(vb4567c2x0, va1x0);
195 vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c2);
196 vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c2);
197 va0x0 = vext_s8(va0x0, va0x0, 2);
198 va1x0 = vext_s8(va1x0, va1x0, 2);
199 int16x8_t vprod0x0123c3 = vmull_s8(vb0123c3x0, va0x0);
200 int16x8_t vprod1x0123c3 = vmull_s8(vb0123c3x0, va1x0);
201 vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c3);
202 vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c3);
203 int16x8_t vprod0x4567c3 = vmull_s8(vb4567c3x0, va0x0);
204 int16x8_t vprod1x4567c3 = vmull_s8(vb4567c3x0, va1x0);
205 vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c3);
206 vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c3);
207
208 k -= 8 * sizeof(int8_t);
209 }
210
211 if XNN_UNLIKELY(k != 0) {
212 const int8x8_t va0 = vld1_s8(a0); a0 = (const int8_t*) ((uintptr_t) a0 + k);
213 const int8x8_t va1 = vld1_s8(a1); a1 = (const int8_t*) ((uintptr_t) a1 + k);
214
215 const int8x8_t vb0123c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
216 const int8x8_t vb4567c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
217
218 const int16x8_t vprod0x0123c0 = vmull_s8(vb0123c0, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0), 0)));
219 vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c0);
220 const int16x8_t vprod0x4567c0 = vmull_s8(vb4567c0, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0), 0)));
221 vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c0);
222 const int16x8_t vprod1x0123c0 = vmull_s8(vb0123c0, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va1), 0)));
223 vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c0);
224 const int16x8_t vprod1x4567c0 = vmull_s8(vb4567c0, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va1), 0)));
225 vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c0);
226
227 if (k > 2 * sizeof(int8_t)) {
228 const int8x8_t vb0123c1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
229 const int8x8_t vb4567c1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
230
231 const int16x8_t vprod0x0123c1 = vmull_s8(vb0123c1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0), 1)));
232 vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c1);
233 const int16x8_t vprod0x4567c1 = vmull_s8(vb4567c1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0), 1)));
234 vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c1);
235 const int16x8_t vprod1x0123c1 = vmull_s8(vb0123c1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va1), 1)));
236 vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c1);
237 const int16x8_t vprod1x4567c1 = vmull_s8(vb4567c1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va1), 1)));
238 vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c1);
239
240 if (k > 4 * sizeof(int8_t)) {
241 const int8x8_t vb0123c2 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
242 const int8x8_t vb4567c2 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
243
244 const int16x8_t vprod0x0123c2 = vmull_s8(vb0123c2, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0), 2)));
245 vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c2);
246 const int16x8_t vprod0x4567c2 = vmull_s8(vb4567c2, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0), 2)));
247 vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c2);
248 const int16x8_t vprod1x0123c2 = vmull_s8(vb0123c2, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va1), 2)));
249 vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c2);
250 const int16x8_t vprod1x4567c2 = vmull_s8(vb4567c2, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va1), 2)));
251 vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c2);
252
253 if (k > 6 * sizeof(int8_t)) {
254 const int8x8_t vb0123c2 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
255 const int8x8_t vb4567c2 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
256
257 const int16x8_t vprod0x0123c2 = vmull_s8(vb0123c2, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0), 3)));
258 vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c2);
259 const int16x8_t vprod0x4567c2 = vmull_s8(vb4567c2, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0), 3)));
260 vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c2);
261 const int16x8_t vprod1x0123c2 = vmull_s8(vb0123c2, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va1), 3)));
262 vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c2);
263 const int16x8_t vprod1x4567c2 = vmull_s8(vb4567c2, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va1), 3)));
264 vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c2);
265 }
266 }
267 }
268 }
269 p -= 2 * sizeof(void*);
270 } while (p != 0);
271
Frank Barchardc7a032d2021-11-10 12:37:49 -0800[diff] [blame]272 float32x4_t vfpacc0x0123 = vcvtq_f32_s32(vacc0x0123);
273 float32x4_t vfpacc0x4567 = vcvtq_f32_s32(vacc0x4567);
274 float32x4_t vfpacc1x0123 = vcvtq_f32_s32(vacc1x0123);
275 float32x4_t vfpacc1x4567 = vcvtq_f32_s32(vacc1x4567);
Frank Barchard494cd2b2021-10-28 17:36:37 -0700[diff] [blame]276
Frank Barchardc7a032d2021-11-10 12:37:49 -0800[diff] [blame]277 const float32x4_t vscale = vld1q_dup_f32(&params->fp32_neon.scale);
278 vfpacc0x0123 = vmulq_f32(vfpacc0x0123, vscale);
279 vfpacc0x4567 = vmulq_f32(vfpacc0x4567, vscale);
280 vfpacc1x0123 = vmulq_f32(vfpacc1x0123, vscale);
281 vfpacc1x4567 = vmulq_f32(vfpacc1x4567, vscale);
Frank Barchard494cd2b2021-10-28 17:36:37 -0700[diff] [blame]282
Frank Barchardc7a032d2021-11-10 12:37:49 -0800[diff] [blame]283 const float32x4_t voutput_min_less_zero_point = vld1q_dup_f32(&params->fp32_neon.output_min_less_zero_point);
284 vfpacc0x0123 = vmaxq_f32(vfpacc0x0123, voutput_min_less_zero_point);
285 vfpacc0x4567 = vmaxq_f32(vfpacc0x4567, voutput_min_less_zero_point);
286 vfpacc1x0123 = vmaxq_f32(vfpacc1x0123, voutput_min_less_zero_point);
287 vfpacc1x4567 = vmaxq_f32(vfpacc1x4567, voutput_min_less_zero_point);
Frank Barchard494cd2b2021-10-28 17:36:37 -0700[diff] [blame]288
Frank Barchardc7a032d2021-11-10 12:37:49 -0800[diff] [blame]289 const float32x4_t voutput_max_less_zero_point = vld1q_dup_f32(&params->fp32_neon.output_max_less_zero_point);
290 vfpacc0x0123 = vminq_f32(vfpacc0x0123, voutput_max_less_zero_point);
291 vfpacc0x4567 = vminq_f32(vfpacc0x4567, voutput_max_less_zero_point);
292 vfpacc1x0123 = vminq_f32(vfpacc1x0123, voutput_max_less_zero_point);
293 vfpacc1x4567 = vminq_f32(vfpacc1x4567, voutput_max_less_zero_point);
Frank Barchard494cd2b2021-10-28 17:36:37 -0700[diff] [blame]294
Frank Barchardc7a032d2021-11-10 12:37:49 -0800[diff] [blame]295 const float32x4_t vmagic_bias = vld1q_dup_f32(&params->fp32_neon.magic_bias);
296 vacc0x0123 = vreinterpretq_s32_f32(vaddq_f32(vfpacc0x0123, vmagic_bias));
297 vacc0x4567 = vreinterpretq_s32_f32(vaddq_f32(vfpacc0x4567, vmagic_bias));
298 vacc1x0123 = vreinterpretq_s32_f32(vaddq_f32(vfpacc1x0123, vmagic_bias));
299 vacc1x4567 = vreinterpretq_s32_f32(vaddq_f32(vfpacc1x4567, vmagic_bias));
300
301 const int32x4_t vmagic_bias_less_zero_point = vld1q_dup_s32(&params->fp32_neon.magic_bias_less_zero_point);
302 vacc0x0123 = vsubq_s32(vacc0x0123, vmagic_bias_less_zero_point);
303 vacc0x4567 = vsubq_s32(vacc0x4567, vmagic_bias_less_zero_point);
304 vacc1x0123 = vsubq_s32(vacc1x0123, vmagic_bias_less_zero_point);
305 vacc1x4567 = vsubq_s32(vacc1x4567, vmagic_bias_less_zero_point);
306
Frank Barchard494cd2b2021-10-28 17:36:37 -0700[diff] [blame]307#if XNN_ARCH_ARM64
Frank Barchardc7a032d2021-11-10 12:37:49 -0800[diff] [blame]308 const int16x8_t vacc0x01234567 = vuzp1q_s16(vreinterpretq_s16_s32(vacc0x0123), vreinterpretq_s16_s32(vacc0x4567));
309 const int16x8_t vacc1x01234567 = vuzp1q_s16(vreinterpretq_s16_s32(vacc1x0123), vreinterpretq_s16_s32(vacc1x4567));
Frank Barchard494cd2b2021-10-28 17:36:37 -0700[diff] [blame]310
Frank Barchardc7a032d2021-11-10 12:37:49 -0800[diff] [blame]311 int8x16_t vout0x01234567_1x01234567 = vuzp1q_s8(vreinterpretq_s8_s16(vacc0x01234567), vreinterpretq_s8_s16(vacc1x01234567));
Frank Barchard494cd2b2021-10-28 17:36:37 -0700[diff] [blame]312#else
Frank Barchardc7a032d2021-11-10 12:37:49 -0800[diff] [blame]313 const int16x8_t vacc0x01234567 = vcombine_s16(vmovn_s32(vacc0x0123), vmovn_s32(vacc0x4567));
314 const int16x8_t vacc1x01234567 = vcombine_s16(vmovn_s32(vacc1x0123), vmovn_s32(vacc1x4567));
Frank Barchard494cd2b2021-10-28 17:36:37 -0700[diff] [blame]315
Frank Barchardc7a032d2021-11-10 12:37:49 -0800[diff] [blame]316 int8x16_t vout0x01234567_1x01234567 = vcombine_s8(vmovn_s16(vacc0x01234567), vmovn_s16(vacc1x01234567));
Frank Barchard494cd2b2021-10-28 17:36:37 -0700[diff] [blame]317#endif
Frank Barchard494cd2b2021-10-28 17:36:37 -0700[diff] [blame]318
319 if (nc >= 8) {
320 vst1_s8(c1 + 0, vget_high_s8(vout0x01234567_1x01234567));
321 vst1_s8(c0 + 0, vget_low_s8(vout0x01234567_1x01234567));
322
323 c1 = (int8_t*) ((uintptr_t) c1 + cn_stride);
324 c0 = (int8_t*) ((uintptr_t) c0 + cn_stride);
325
326 a = (const int8_t**restrict) ((uintptr_t) a - ks);
327
328 nc -= 8;
329 } else {
330 if (nc & 4) {
331 vst1q_lane_u32(__builtin_assume_aligned(c1, 1), vreinterpretq_u32_s8(vout0x01234567_1x01234567), 2); c1 += 4;
332 vst1q_lane_u32(__builtin_assume_aligned(c0, 1), vreinterpretq_u32_s8(vout0x01234567_1x01234567), 0); c0 += 4;
333 vout0x01234567_1x01234567 = vextq_s8(vout0x01234567_1x01234567, vout0x01234567_1x01234567, 4);
334 }
335 if (nc & 2) {
336 vst1q_lane_u16(__builtin_assume_aligned(c1, 1), vreinterpretq_u16_s8(vout0x01234567_1x01234567), 4); c1 += 2;
337 vst1q_lane_u16(__builtin_assume_aligned(c0, 1), vreinterpretq_u16_s8(vout0x01234567_1x01234567), 0); c0 += 2;
338 vout0x01234567_1x01234567 = vextq_s8(vout0x01234567_1x01234567, vout0x01234567_1x01234567, 2);
339 }
340 if (nc & 1) {
341 vst1q_lane_s8(c1, vout0x01234567_1x01234567, 8);
342 vst1q_lane_s8(c0, vout0x01234567_1x01234567, 0);
343 }
344
345 nc = 0;
346 }
347 } while (nc != 0);
348}