Debian Codegraph @

src/qs8-qc8w-igemm/gen/qs8-qc8w-igemm-3x16c8-minmax-fp32-neoni8mm.c

raw
1 // Auto-generated file. Do not edit!
2 // Template: src/qs8-igemm/c8-neoni8mm.c.in
3 // Generator: tools/xngen
4 //
5 // Copyright 2023 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/igemm.h"
15 #include "xnnpack/intrinsics-polyfill.h"
16 #include "xnnpack/math.h"
17
18
19 void xnn_qs8_qc8w_igemm_minmax_fp32_ukernel_3x16c8__neoni8mm(
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 int8_t* restrict c,
27 size_t cm_stride,
28 size_t cn_stride,
29 size_t a_offset,
30 const int8_t* zero,
31 const union xnn_qs8_qc8w_conv_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
32 {
33 assert(mr != 0);
34 assert(mr <= 3);
35 assert(nc != 0);
36 assert(kc != 0);
37 assert(ks != 0);
38 assert(ks % (3 * sizeof(void*)) == 0);
39 assert(a_offset % sizeof(int8_t) == 0);
40 assert(a != NULL);
41 assert(w != NULL);
42 assert(c != NULL);
43
44 kc = round_up_po2(kc, 8 * sizeof(int8_t));
45 int8_t* c0 = c;
46 int8_t* c1 = (int8_t*) ((uintptr_t) c0 + cm_stride);
47 if XNN_UNPREDICTABLE(mr < 2) {
48 c1 = c0;
49 }
50 int8_t* c2 = (int8_t*) ((uintptr_t) c1 + cm_stride);
51 if XNN_UNPREDICTABLE(mr <= 2) {
52 c2 = c1;
53 }
54
55 do {
56 const uint64x2x2_t vbias01x0123 = vld2q_dup_u64(w); w = (const int32_t*) w + 4;
57 const uint64x2x2_t vbias01x4567 = vld2q_dup_u64(w); w = (const int32_t*) w + 4;
58 const uint64x2x2_t vbias01x89AB = vld2q_dup_u64(w); w = (const int32_t*) w + 4;
59 const uint64x2x2_t vbias01xCDEF = vld2q_dup_u64(w); w = (const int32_t*) w + 4;
60 int32x4_t vacc01x01 = vreinterpretq_s32_u64(vbias01x0123.val[0]);
61 int32x4_t vacc01x23 = vreinterpretq_s32_u64(vbias01x0123.val[1]);
62 int32x4_t vacc01x45 = vreinterpretq_s32_u64(vbias01x4567.val[0]);
63 int32x4_t vacc01x67 = vreinterpretq_s32_u64(vbias01x4567.val[1]);
64 int32x4_t vacc01x89 = vreinterpretq_s32_u64(vbias01x89AB.val[0]);
65 int32x4_t vacc01xAB = vreinterpretq_s32_u64(vbias01x89AB.val[1]);
66 int32x4_t vacc01xCD = vreinterpretq_s32_u64(vbias01xCDEF.val[0]);
67 int32x4_t vacc01xEF = vreinterpretq_s32_u64(vbias01xCDEF.val[1]);
68 int32x4_t vacc23x01 = vreinterpretq_s32_u64(vbias01x0123.val[0]);
69 int32x4_t vacc23x23 = vreinterpretq_s32_u64(vbias01x0123.val[1]);
70 int32x4_t vacc23x45 = vreinterpretq_s32_u64(vbias01x4567.val[0]);
71 int32x4_t vacc23x67 = vreinterpretq_s32_u64(vbias01x4567.val[1]);
72 int32x4_t vacc23x89 = vreinterpretq_s32_u64(vbias01x89AB.val[0]);
73 int32x4_t vacc23xAB = vreinterpretq_s32_u64(vbias01x89AB.val[1]);
74 int32x4_t vacc23xCD = vreinterpretq_s32_u64(vbias01xCDEF.val[0]);
75 int32x4_t vacc23xEF = vreinterpretq_s32_u64(vbias01xCDEF.val[1]);
76
77 size_t p = ks;
78 do {
79 const int8_t* restrict a0 = a[0];
80 if XNN_UNPREDICTABLE(a0 != zero) {
81 a0 = (const int8_t*) ((uintptr_t) a0 + a_offset);
82 }
83 const int8_t* restrict a1 = a[1];
84 if XNN_UNPREDICTABLE(a1 != zero) {
85 a1 = (const int8_t*) ((uintptr_t) a1 + a_offset);
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 }
91 a += 3;
92
93 uint64x2x2_t va01x0123456789ABCDEF;
94 va01x0123456789ABCDEF.val[0] = vdupq_n_u64(0);
95 va01x0123456789ABCDEF.val[1] = vdupq_n_u64(0);
96 uint64x2x2_t va23x0123456789ABCDEF;
97 va23x0123456789ABCDEF.val[0] = vdupq_n_u64(0);
98 va23x0123456789ABCDEF.val[1] = vdupq_n_u64(0);
99
100 // Inner accumulation loop along the 16 columns.
101 size_t k = kc;
102 // 2x partial unrolled loop to load 8 bytes at a time.
103 while (k >= 16 * sizeof(int8_t)) {
104 // Load a 3x16 block of activations.
105 va01x0123456789ABCDEF = vld2q_lane_u64((const void*) a0, va01x0123456789ABCDEF, 0); a0 += 16;
106 va23x0123456789ABCDEF = vld2q_lane_u64((const void*) a2, va23x0123456789ABCDEF, 0); a2 += 16;
107 va01x0123456789ABCDEF = vld2q_lane_u64((const void*) a1, va01x0123456789ABCDEF, 1); a1 += 16;
108
109 // Load a 16x16 block of weights.
110 const int8x16_t vb01x01234567 = vld1q_s8(w); w = (const int8_t*) w + 16;
111 const int8x16_t vb23x01234567 = vld1q_s8(w); w = (const int8_t*) w + 16;
112 const int8x16_t vb45x01234567 = vld1q_s8(w); w = (const int8_t*) w + 16;
113 const int8x16_t vb67x01234567 = vld1q_s8(w); w = (const int8_t*) w + 16;
114 const int8x16_t vb89x01234567 = vld1q_s8(w); w = (const int8_t*) w + 16;
115 const int8x16_t vbABx01234567 = vld1q_s8(w); w = (const int8_t*) w + 16;
116 const int8x16_t vbCDx01234567 = vld1q_s8(w); w = (const int8_t*) w + 16;
117 const int8x16_t vbEFx01234567 = vld1q_s8(w); w = (const int8_t*) w + 16;
118 const int8x16_t vb01x89ABCDEF = vld1q_s8(w); w = (const int8_t*) w + 16;
119 const int8x16_t vb23x89ABCDEF = vld1q_s8(w); w = (const int8_t*) w + 16;
120 const int8x16_t vb45x89ABCDEF = vld1q_s8(w); w = (const int8_t*) w + 16;
121 const int8x16_t vb67x89ABCDEF = vld1q_s8(w); w = (const int8_t*) w + 16;
122 const int8x16_t vb89x89ABCDEF = vld1q_s8(w); w = (const int8_t*) w + 16;
123 const int8x16_t vbABx89ABCDEF = vld1q_s8(w); w = (const int8_t*) w + 16;
124 const int8x16_t vbCDx89ABCDEF = vld1q_s8(w); w = (const int8_t*) w + 16;
125 const int8x16_t vbEFx89ABCDEF = vld1q_s8(w); w = (const int8_t*) w + 16;
126
127 // Multiply-accumulate: 3x16 * 16x16 --> 3x16.
128 vacc01x01 = vmmlaq_s32(vacc01x01, vreinterpretq_s8_u64(va01x0123456789ABCDEF.val[0]), vb01x01234567);
129 vacc01x23 = vmmlaq_s32(vacc01x23, vreinterpretq_s8_u64(va01x0123456789ABCDEF.val[0]), vb23x01234567);
130 vacc01x45 = vmmlaq_s32(vacc01x45, vreinterpretq_s8_u64(va01x0123456789ABCDEF.val[0]), vb45x01234567);
131 vacc01x67 = vmmlaq_s32(vacc01x67, vreinterpretq_s8_u64(va01x0123456789ABCDEF.val[0]), vb67x01234567);
132 vacc01x89 = vmmlaq_s32(vacc01x89, vreinterpretq_s8_u64(va01x0123456789ABCDEF.val[0]), vb89x01234567);
133 vacc01xAB = vmmlaq_s32(vacc01xAB, vreinterpretq_s8_u64(va01x0123456789ABCDEF.val[0]), vbABx01234567);
134 vacc01xCD = vmmlaq_s32(vacc01xCD, vreinterpretq_s8_u64(va01x0123456789ABCDEF.val[0]), vbCDx01234567);
135 vacc01xEF = vmmlaq_s32(vacc01xEF, vreinterpretq_s8_u64(va01x0123456789ABCDEF.val[0]), vbEFx01234567);
136 vacc23x01 = vmmlaq_s32(vacc23x01, vreinterpretq_s8_u64(va23x0123456789ABCDEF.val[0]), vb01x01234567);
137 vacc23x23 = vmmlaq_s32(vacc23x23, vreinterpretq_s8_u64(va23x0123456789ABCDEF.val[0]), vb23x01234567);
138 vacc23x45 = vmmlaq_s32(vacc23x45, vreinterpretq_s8_u64(va23x0123456789ABCDEF.val[0]), vb45x01234567);
139 vacc23x67 = vmmlaq_s32(vacc23x67, vreinterpretq_s8_u64(va23x0123456789ABCDEF.val[0]), vb67x01234567);
140 vacc23x89 = vmmlaq_s32(vacc23x89, vreinterpretq_s8_u64(va23x0123456789ABCDEF.val[0]), vb89x01234567);
141 vacc23xAB = vmmlaq_s32(vacc23xAB, vreinterpretq_s8_u64(va23x0123456789ABCDEF.val[0]), vbABx01234567);
142 vacc23xCD = vmmlaq_s32(vacc23xCD, vreinterpretq_s8_u64(va23x0123456789ABCDEF.val[0]), vbCDx01234567);
143 vacc23xEF = vmmlaq_s32(vacc23xEF, vreinterpretq_s8_u64(va23x0123456789ABCDEF.val[0]), vbEFx01234567);
144 vacc01x01 = vmmlaq_s32(vacc01x01, vreinterpretq_s8_u64(va01x0123456789ABCDEF.val[1]), vb01x89ABCDEF);
145 vacc01x23 = vmmlaq_s32(vacc01x23, vreinterpretq_s8_u64(va01x0123456789ABCDEF.val[1]), vb23x89ABCDEF);
146 vacc01x45 = vmmlaq_s32(vacc01x45, vreinterpretq_s8_u64(va01x0123456789ABCDEF.val[1]), vb45x89ABCDEF);
147 vacc01x67 = vmmlaq_s32(vacc01x67, vreinterpretq_s8_u64(va01x0123456789ABCDEF.val[1]), vb67x89ABCDEF);
148 vacc01x89 = vmmlaq_s32(vacc01x89, vreinterpretq_s8_u64(va01x0123456789ABCDEF.val[1]), vb89x89ABCDEF);
149 vacc01xAB = vmmlaq_s32(vacc01xAB, vreinterpretq_s8_u64(va01x0123456789ABCDEF.val[1]), vbABx89ABCDEF);
150 vacc01xCD = vmmlaq_s32(vacc01xCD, vreinterpretq_s8_u64(va01x0123456789ABCDEF.val[1]), vbCDx89ABCDEF);
151 vacc01xEF = vmmlaq_s32(vacc01xEF, vreinterpretq_s8_u64(va01x0123456789ABCDEF.val[1]), vbEFx89ABCDEF);
152 vacc23x01 = vmmlaq_s32(vacc23x01, vreinterpretq_s8_u64(va23x0123456789ABCDEF.val[1]), vb01x89ABCDEF);
153 vacc23x23 = vmmlaq_s32(vacc23x23, vreinterpretq_s8_u64(va23x0123456789ABCDEF.val[1]), vb23x89ABCDEF);
154 vacc23x45 = vmmlaq_s32(vacc23x45, vreinterpretq_s8_u64(va23x0123456789ABCDEF.val[1]), vb45x89ABCDEF);
155 vacc23x67 = vmmlaq_s32(vacc23x67, vreinterpretq_s8_u64(va23x0123456789ABCDEF.val[1]), vb67x89ABCDEF);
156 vacc23x89 = vmmlaq_s32(vacc23x89, vreinterpretq_s8_u64(va23x0123456789ABCDEF.val[1]), vb89x89ABCDEF);
157 vacc23xAB = vmmlaq_s32(vacc23xAB, vreinterpretq_s8_u64(va23x0123456789ABCDEF.val[1]), vbABx89ABCDEF);
158 vacc23xCD = vmmlaq_s32(vacc23xCD, vreinterpretq_s8_u64(va23x0123456789ABCDEF.val[1]), vbCDx89ABCDEF);
159 vacc23xEF = vmmlaq_s32(vacc23xEF, vreinterpretq_s8_u64(va23x0123456789ABCDEF.val[1]), vbEFx89ABCDEF);
160
161 k -= 16 * sizeof(int8_t);
162 }
163 // Handle up to 8 final positions of `k`
164 if XNN_UNLIKELY(k != 0) {
165 // Load a 3x8 block of activations.
166 uint64x2_t va01x01234567 = vld1q_dup_u64((const void*) a0); a0 += 8;
167 uint64x2_t va23x01234567 = vld1q_dup_u64((const void*) a2); a2 += 8;
168 va01x01234567 = vld1q_lane_u64((const void*) a1, va01x01234567, 1); a1 += 8;
169
170 // Load a 16x16 block of weights.
171 const int8x16_t vb01x01234567 = vld1q_s8(w); w = (const int8_t*) w + 16;
172 const int8x16_t vb23x01234567 = vld1q_s8(w); w = (const int8_t*) w + 16;
173 const int8x16_t vb45x01234567 = vld1q_s8(w); w = (const int8_t*) w + 16;
174 const int8x16_t vb67x01234567 = vld1q_s8(w); w = (const int8_t*) w + 16;
175 const int8x16_t vb89x01234567 = vld1q_s8(w); w = (const int8_t*) w + 16;
176 const int8x16_t vbABx01234567 = vld1q_s8(w); w = (const int8_t*) w + 16;
177 const int8x16_t vbCDx01234567 = vld1q_s8(w); w = (const int8_t*) w + 16;
178 const int8x16_t vbEFx01234567 = vld1q_s8(w); w = (const int8_t*) w + 16;
179
180 // Multiply-accumulate: 3x4 * 4x16 --> 3x16.
181 vacc01x01 = vmmlaq_s32(vacc01x01, vreinterpretq_s8_u64(va01x01234567), vb01x01234567);
182 vacc01x23 = vmmlaq_s32(vacc01x23, vreinterpretq_s8_u64(va01x01234567), vb23x01234567);
183 vacc01x45 = vmmlaq_s32(vacc01x45, vreinterpretq_s8_u64(va01x01234567), vb45x01234567);
184 vacc01x67 = vmmlaq_s32(vacc01x67, vreinterpretq_s8_u64(va01x01234567), vb67x01234567);
185 vacc01x89 = vmmlaq_s32(vacc01x89, vreinterpretq_s8_u64(va01x01234567), vb89x01234567);
186 vacc01xAB = vmmlaq_s32(vacc01xAB, vreinterpretq_s8_u64(va01x01234567), vbABx01234567);
187 vacc01xCD = vmmlaq_s32(vacc01xCD, vreinterpretq_s8_u64(va01x01234567), vbCDx01234567);
188 vacc01xEF = vmmlaq_s32(vacc01xEF, vreinterpretq_s8_u64(va01x01234567), vbEFx01234567);
189 vacc23x01 = vmmlaq_s32(vacc23x01, vreinterpretq_s8_u64(va23x01234567), vb01x01234567);
190 vacc23x23 = vmmlaq_s32(vacc23x23, vreinterpretq_s8_u64(va23x01234567), vb23x01234567);
191 vacc23x45 = vmmlaq_s32(vacc23x45, vreinterpretq_s8_u64(va23x01234567), vb45x01234567);
192 vacc23x67 = vmmlaq_s32(vacc23x67, vreinterpretq_s8_u64(va23x01234567), vb67x01234567);
193 vacc23x89 = vmmlaq_s32(vacc23x89, vreinterpretq_s8_u64(va23x01234567), vb89x01234567);
194 vacc23xAB = vmmlaq_s32(vacc23xAB, vreinterpretq_s8_u64(va23x01234567), vbABx01234567);
195 vacc23xCD = vmmlaq_s32(vacc23xCD, vreinterpretq_s8_u64(va23x01234567), vbCDx01234567);
196 vacc23xEF = vmmlaq_s32(vacc23xEF, vreinterpretq_s8_u64(va23x01234567), vbEFx01234567);
197 }
198
199 p -= 3 * sizeof(void*);
200 } while (p != 0);
201
202 int32x4_t vacc0x0123 = vcombine_s32(vget_low_s32(vacc01x01), vget_low_s32(vacc01x23));
203 int32x4_t vacc1x0123 = vcombine_s32(vget_high_s32(vacc01x01), vget_high_s32(vacc01x23));
204 int32x4_t vacc0x4567 = vcombine_s32(vget_low_s32(vacc01x45), vget_low_s32(vacc01x67));
205 int32x4_t vacc1x4567 = vcombine_s32(vget_high_s32(vacc01x45), vget_high_s32(vacc01x67));
206 int32x4_t vacc0x89AB = vcombine_s32(vget_low_s32(vacc01x89), vget_low_s32(vacc01xAB));
207 int32x4_t vacc1x89AB = vcombine_s32(vget_high_s32(vacc01x89), vget_high_s32(vacc01xAB));
208 int32x4_t vacc0xCDEF = vcombine_s32(vget_low_s32(vacc01xCD), vget_low_s32(vacc01xEF));
209 int32x4_t vacc1xCDEF = vcombine_s32(vget_high_s32(vacc01xCD), vget_high_s32(vacc01xEF));
210 int32x4_t vacc2x0123 = vcombine_s32(vget_low_s32(vacc23x01), vget_low_s32(vacc23x23));
211 int32x4_t vacc2x4567 = vcombine_s32(vget_low_s32(vacc23x45), vget_low_s32(vacc23x67));
212 int32x4_t vacc2x89AB = vcombine_s32(vget_low_s32(vacc23x89), vget_low_s32(vacc23xAB));
213 int32x4_t vacc2xCDEF = vcombine_s32(vget_low_s32(vacc23xCD), vget_low_s32(vacc23xEF));
214
215 float32x4_t vfpacc0x0123 = vcvtq_f32_s32(vacc0x0123);
216 float32x4_t vfpacc0x4567 = vcvtq_f32_s32(vacc0x4567);
217 float32x4_t vfpacc0x89AB = vcvtq_f32_s32(vacc0x89AB);
218 float32x4_t vfpacc0xCDEF = vcvtq_f32_s32(vacc0xCDEF);
219 float32x4_t vfpacc1x0123 = vcvtq_f32_s32(vacc1x0123);
220 float32x4_t vfpacc1x4567 = vcvtq_f32_s32(vacc1x4567);
221 float32x4_t vfpacc1x89AB = vcvtq_f32_s32(vacc1x89AB);
222 float32x4_t vfpacc1xCDEF = vcvtq_f32_s32(vacc1xCDEF);
223 float32x4_t vfpacc2x0123 = vcvtq_f32_s32(vacc2x0123);
224 float32x4_t vfpacc2x4567 = vcvtq_f32_s32(vacc2x4567);
225 float32x4_t vfpacc2x89AB = vcvtq_f32_s32(vacc2x89AB);
226 float32x4_t vfpacc2xCDEF = vcvtq_f32_s32(vacc2xCDEF);
227
228 const float32x4_t vscale0123 = vld1q_f32((const float*) w); w = (const float*) w + 4;
229 vfpacc0x0123 = vmulq_f32(vfpacc0x0123, vscale0123);
230 vfpacc1x0123 = vmulq_f32(vfpacc1x0123, vscale0123);
231 vfpacc2x0123 = vmulq_f32(vfpacc2x0123, vscale0123);
232 const float32x4_t vscale4567 = vld1q_f32((const float*) w); w = (const float*) w + 4;
233 vfpacc0x4567 = vmulq_f32(vfpacc0x4567, vscale4567);
234 vfpacc1x4567 = vmulq_f32(vfpacc1x4567, vscale4567);
235 vfpacc2x4567 = vmulq_f32(vfpacc2x4567, vscale4567);
236 const float32x4_t vscale89AB = vld1q_f32((const float*) w); w = (const float*) w + 4;
237 vfpacc0x89AB = vmulq_f32(vfpacc0x89AB, vscale89AB);
238 vfpacc1x89AB = vmulq_f32(vfpacc1x89AB, vscale89AB);
239 vfpacc2x89AB = vmulq_f32(vfpacc2x89AB, vscale89AB);
240 const float32x4_t vscaleCDEF = vld1q_f32((const float*) w); w = (const float*) w + 4;
241 vfpacc0xCDEF = vmulq_f32(vfpacc0xCDEF, vscaleCDEF);
242 vfpacc1xCDEF = vmulq_f32(vfpacc1xCDEF, vscaleCDEF);
243 vfpacc2xCDEF = vmulq_f32(vfpacc2xCDEF, vscaleCDEF);
244
245 vacc0x0123 = vcvtnq_s32_f32(vfpacc0x0123);
246 vacc0x4567 = vcvtnq_s32_f32(vfpacc0x4567);
247 vacc0x89AB = vcvtnq_s32_f32(vfpacc0x89AB);
248 vacc0xCDEF = vcvtnq_s32_f32(vfpacc0xCDEF);
249 vacc1x0123 = vcvtnq_s32_f32(vfpacc1x0123);
250 vacc1x4567 = vcvtnq_s32_f32(vfpacc1x4567);
251 vacc1x89AB = vcvtnq_s32_f32(vfpacc1x89AB);
252 vacc1xCDEF = vcvtnq_s32_f32(vfpacc1xCDEF);
253 vacc2x0123 = vcvtnq_s32_f32(vfpacc2x0123);
254 vacc2x4567 = vcvtnq_s32_f32(vfpacc2x4567);
255 vacc2x89AB = vcvtnq_s32_f32(vfpacc2x89AB);
256 vacc2xCDEF = vcvtnq_s32_f32(vfpacc2xCDEF);
257
258 const int16x8_t voutput_zero_point = vld1q_dup_s16(&params->fp32_neonv8.output_zero_point);
259 #if XNN_ARCH_ARM64
260 const int16x8_t vacc0x01234567 = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc0x0123), vacc0x4567), voutput_zero_point);
261 const int16x8_t vacc0x89ABCDEF = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc0x89AB), vacc0xCDEF), voutput_zero_point);
262 const int16x8_t vacc1x01234567 = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc1x0123), vacc1x4567), voutput_zero_point);
263 const int16x8_t vacc1x89ABCDEF = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc1x89AB), vacc1xCDEF), voutput_zero_point);
264 const int16x8_t vacc2x01234567 = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc2x0123), vacc2x4567), voutput_zero_point);
265 const int16x8_t vacc2x89ABCDEF = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc2x89AB), vacc2xCDEF), voutput_zero_point);
266
267 int8x16_t vout0x0123456789ABCDEF = vqmovn_high_s16(vqmovn_s16(vacc0x01234567), vacc0x89ABCDEF);
268 int8x16_t vout1x0123456789ABCDEF = vqmovn_high_s16(vqmovn_s16(vacc1x01234567), vacc1x89ABCDEF);
269 int8x16_t vout2x0123456789ABCDEF = vqmovn_high_s16(vqmovn_s16(vacc2x01234567), vacc2x89ABCDEF);
270 #else
271 const int16x8_t vacc0x01234567 = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc0x0123), vqmovn_s32(vacc0x4567)), voutput_zero_point);
272 const int16x8_t vacc0x89ABCDEF = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc0x89AB), vqmovn_s32(vacc0xCDEF)), voutput_zero_point);
273 const int16x8_t vacc1x01234567 = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc1x0123), vqmovn_s32(vacc1x4567)), voutput_zero_point);
274 const int16x8_t vacc1x89ABCDEF = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc1x89AB), vqmovn_s32(vacc1xCDEF)), voutput_zero_point);
275 const int16x8_t vacc2x01234567 = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc2x0123), vqmovn_s32(vacc2x4567)), voutput_zero_point);
276 const int16x8_t vacc2x89ABCDEF = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc2x89AB), vqmovn_s32(vacc2xCDEF)), voutput_zero_point);
277
278 int8x16_t vout0x0123456789ABCDEF = vcombine_s8(vqmovn_s16(vacc0x01234567), vqmovn_s16(vacc0x89ABCDEF));
279 int8x16_t vout1x0123456789ABCDEF = vcombine_s8(vqmovn_s16(vacc1x01234567), vqmovn_s16(vacc1x89ABCDEF));
280 int8x16_t vout2x0123456789ABCDEF = vcombine_s8(vqmovn_s16(vacc2x01234567), vqmovn_s16(vacc2x89ABCDEF));
281 #endif
282 const int8x16_t voutput_min = vld1q_dup_s8(&params->fp32_neonv8.output_min);
283 const int8x16_t voutput_max = vld1q_dup_s8(&params->fp32_neonv8.output_max);
284
285 vout2x0123456789ABCDEF = vmaxq_s8(vout2x0123456789ABCDEF, voutput_min);
286 vout1x0123456789ABCDEF = vmaxq_s8(vout1x0123456789ABCDEF, voutput_min);
287 vout0x0123456789ABCDEF = vmaxq_s8(vout0x0123456789ABCDEF, voutput_min);
288
289 vout2x0123456789ABCDEF = vminq_s8(vout2x0123456789ABCDEF, voutput_max);
290 vout1x0123456789ABCDEF = vminq_s8(vout1x0123456789ABCDEF, voutput_max);
291 vout0x0123456789ABCDEF = vminq_s8(vout0x0123456789ABCDEF, voutput_max);
292
293 if (nc >= 16) {
294 vst1q_s8(c2, vout2x0123456789ABCDEF);
295 vst1q_s8(c1, vout1x0123456789ABCDEF);
296 vst1q_s8(c0, vout0x0123456789ABCDEF);
297
298 c2 = (int8_t*) ((uintptr_t) c2 + cn_stride);
299 c1 = (int8_t*) ((uintptr_t) c1 + cn_stride);
300 c0 = (int8_t*) ((uintptr_t) c0 + cn_stride);
301
302 a = (const int8_t**restrict) ((uintptr_t) a - ks);
303
304 nc -= 16;
305 } else {
306 int8x8_t vout2x01234567 = vget_low_s8(vout2x0123456789ABCDEF);
307 int8x16_t vout0x01234567_1x01234567 = vcombine_s8(vget_low_s8(vout0x0123456789ABCDEF), vget_low_s8(vout1x0123456789ABCDEF));
308 if (nc & 8) {
309 vst1_s8(c2, vout2x01234567); c2 += 8;
310 vst1_s8(c1, vget_high_s8(vout0x01234567_1x01234567)); c1 += 8;
311 vst1_s8(c0, vget_low_s8(vout0x01234567_1x01234567)); c0 += 8;
312 vout2x01234567 = vget_high_s8(vout2x0123456789ABCDEF);
313 vout0x01234567_1x01234567 = vcombine_s8(vget_high_s8(vout0x0123456789ABCDEF), vget_high_s8(vout1x0123456789ABCDEF));
314 }
315 if (nc & 4) {
316 vst1_lane_u32((void*) c2, vreinterpret_u32_s8(vout2x01234567), 0); c2 += 4;
317 vst1q_lane_u32((void*) c1, vreinterpretq_u32_s8(vout0x01234567_1x01234567), 2); c1 += 4;
318 vst1q_lane_u32((void*) c0, vreinterpretq_u32_s8(vout0x01234567_1x01234567), 0); c0 += 4;
319 vout2x01234567 = vext_s8(vout2x01234567, vout2x01234567, 4);
320 vout0x01234567_1x01234567 = vextq_s8(vout0x01234567_1x01234567, vout0x01234567_1x01234567, 4);
321 }
322 if (nc & 2) {
323 vst1_lane_u16((void*) c2, vreinterpret_u16_s8(vout2x01234567), 0); c2 += 2;
324 vst1q_lane_u16((void*) c1, vreinterpretq_u16_s8(vout0x01234567_1x01234567), 4); c1 += 2;
325 vst1q_lane_u16((void*) c0, vreinterpretq_u16_s8(vout0x01234567_1x01234567), 0); c0 += 2;
326 vout2x01234567 = vext_s8(vout2x01234567, vout2x01234567, 2);
327 vout0x01234567_1x01234567 = vextq_s8(vout0x01234567_1x01234567, vout0x01234567_1x01234567, 2);
328 }
329 if (nc & 1) {
330 vst1_lane_s8(c2, vout2x01234567, 0);
331 vst1q_lane_s8(c1, vout0x01234567_1x01234567, 8);
332 vst1q_lane_s8(c0, vout0x01234567_1x01234567, 0);
333 }
334
335 nc = 0;
336 }
337 } while (nc != 0);
338 }