Debian Codegraph @

src/qd8-f16-qc8w-igemm/gen/qd8-f16-qc8w-igemm-4x16c8-minmax-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/math.h"
16
17
18 void xnn_qd8_f16_qc8w_igemm_minmax_ukernel_4x16c8__neoni8mm(
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 xnn_float16* restrict c,
26 size_t cm_stride,
27 size_t cn_stride,
28 size_t a_offset,
29 const int8_t* zero,
30 const int8_t* zero_data,
31 const union xnn_f16_minmax_params params[restrict XNN_MIN_ELEMENTS(1)],
32 const struct xnn_qd8_quantization_params quantization_params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
33 {
34 assert(mr != 0);
35 assert(mr <= 4);
36 assert(nc != 0);
37 assert(kc != 0);
38 assert(ks != 0);
39 assert(ks % (4 * sizeof(void*)) == 0);
40 assert(a_offset % sizeof(int8_t) == 0);
41 assert(a != NULL);
42 assert(w != NULL);
43 assert(c != NULL);
44
45 kc = round_up_po2(kc, 8 * sizeof(int8_t));
46 uint16_t* c0 = (uint16_t*) c;
47 uint16_t* c1 = (uint16_t*) ((uintptr_t) c0 + cm_stride);
48 if XNN_UNPREDICTABLE(mr < 2) {
49 c1 = c0;
50 }
51 uint16_t* c2 = (uint16_t*) ((uintptr_t) c1 + cm_stride);
52 if XNN_UNPREDICTABLE(mr <= 2) {
53 c2 = c1;
54 }
55 uint16_t* c3 = (uint16_t*) ((uintptr_t) c2 + cm_stride);
56 if XNN_UNPREDICTABLE(mr != 4) {
57 c3 = c2;
58 }
59
60 do {
61 const int32x4_t vinput_zero_point = vld1q_dup_s32(&quantization_params->zero_point);
62 const int32x4_t vksum0123 = vld1q_s32(w); w = (const int32_t*) w + 4;
63 const int32x4_t vksumzp0x0123 = vmulq_s32(vksum0123, vinput_zero_point);
64 const int32x4_t vksum4567 = vld1q_s32(w); w = (const int32_t*) w + 4;
65 const int32x4_t vksumzp0x4567 = vmulq_s32(vksum4567, vinput_zero_point);
66 const int32x4_t vksum89AB = vld1q_s32(w); w = (const int32_t*) w + 4;
67 const int32x4_t vksumzp0x89AB = vmulq_s32(vksum89AB, vinput_zero_point);
68 const int32x4_t vksumCDEF = vld1q_s32(w); w = (const int32_t*) w + 4;
69 const int32x4_t vksumzp0xCDEF = vmulq_s32(vksumCDEF, vinput_zero_point);
70 const int32x4_t vksumzp1x0123 = vksumzp0x0123;
71 const int32x4_t vksumzp1x4567 = vksumzp0x4567;
72 const int32x4_t vksumzp1x89AB = vksumzp0x89AB;
73 const int32x4_t vksumzp1xCDEF = vksumzp0xCDEF;
74 const int32x4_t vksumzp2x0123 = vksumzp0x0123;
75 const int32x4_t vksumzp2x4567 = vksumzp0x4567;
76 const int32x4_t vksumzp2x89AB = vksumzp0x89AB;
77 const int32x4_t vksumzp2xCDEF = vksumzp0xCDEF;
78 const int32x4_t vksumzp3x0123 = vksumzp0x0123;
79 const int32x4_t vksumzp3x4567 = vksumzp0x4567;
80 const int32x4_t vksumzp3x89AB = vksumzp0x89AB;
81 const int32x4_t vksumzp3xCDEF = vksumzp0xCDEF;
82
83 int32x4_t vacc01x01 = vreinterpretq_s32_u64(vtrn1q_u64(vreinterpretq_u64_s32(vksumzp0x0123), vreinterpretq_u64_s32(vksumzp1x0123)));
84 int32x4_t vacc01x23 = vreinterpretq_s32_u64(vtrn2q_u64(vreinterpretq_u64_s32(vksumzp0x0123), vreinterpretq_u64_s32(vksumzp1x0123)));
85 int32x4_t vacc01x45 = vreinterpretq_s32_u64(vtrn1q_u64(vreinterpretq_u64_s32(vksumzp0x4567), vreinterpretq_u64_s32(vksumzp1x4567)));
86 int32x4_t vacc01x67 = vreinterpretq_s32_u64(vtrn2q_u64(vreinterpretq_u64_s32(vksumzp0x4567), vreinterpretq_u64_s32(vksumzp1x4567)));
87 int32x4_t vacc01x89 = vreinterpretq_s32_u64(vtrn1q_u64(vreinterpretq_u64_s32(vksumzp0x89AB), vreinterpretq_u64_s32(vksumzp1x89AB)));
88 int32x4_t vacc01xAB = vreinterpretq_s32_u64(vtrn2q_u64(vreinterpretq_u64_s32(vksumzp0x89AB), vreinterpretq_u64_s32(vksumzp1x89AB)));
89 int32x4_t vacc01xCD = vreinterpretq_s32_u64(vtrn1q_u64(vreinterpretq_u64_s32(vksumzp0xCDEF), vreinterpretq_u64_s32(vksumzp1xCDEF)));
90 int32x4_t vacc01xEF = vreinterpretq_s32_u64(vtrn2q_u64(vreinterpretq_u64_s32(vksumzp0xCDEF), vreinterpretq_u64_s32(vksumzp1xCDEF)));
91 int32x4_t vacc23x01 = vreinterpretq_s32_u64(vtrn1q_u64(vreinterpretq_u64_s32(vksumzp2x0123), vreinterpretq_u64_s32(vksumzp3x0123)));
92 int32x4_t vacc23x23 = vreinterpretq_s32_u64(vtrn2q_u64(vreinterpretq_u64_s32(vksumzp2x0123), vreinterpretq_u64_s32(vksumzp3x0123)));
93 int32x4_t vacc23x45 = vreinterpretq_s32_u64(vtrn1q_u64(vreinterpretq_u64_s32(vksumzp2x4567), vreinterpretq_u64_s32(vksumzp3x4567)));
94 int32x4_t vacc23x67 = vreinterpretq_s32_u64(vtrn2q_u64(vreinterpretq_u64_s32(vksumzp2x4567), vreinterpretq_u64_s32(vksumzp3x4567)));
95 int32x4_t vacc23x89 = vreinterpretq_s32_u64(vtrn1q_u64(vreinterpretq_u64_s32(vksumzp2x89AB), vreinterpretq_u64_s32(vksumzp3x89AB)));
96 int32x4_t vacc23xAB = vreinterpretq_s32_u64(vtrn2q_u64(vreinterpretq_u64_s32(vksumzp2x89AB), vreinterpretq_u64_s32(vksumzp3x89AB)));
97 int32x4_t vacc23xCD = vreinterpretq_s32_u64(vtrn1q_u64(vreinterpretq_u64_s32(vksumzp2xCDEF), vreinterpretq_u64_s32(vksumzp3xCDEF)));
98 int32x4_t vacc23xEF = vreinterpretq_s32_u64(vtrn2q_u64(vreinterpretq_u64_s32(vksumzp2xCDEF), vreinterpretq_u64_s32(vksumzp3xCDEF)));
99
100 size_t p = ks;
101 do {
102 const int8_t* restrict a0 = a[0];
103 if XNN_UNPREDICTABLE(a0 != zero) {
104 a0 = (const int8_t*) ((uintptr_t) a0 + a_offset);
105 } else {
106 a0 = zero_data;
107 }
108 const int8_t* restrict a1 = a[1];
109 if XNN_UNPREDICTABLE(a1 != zero) {
110 a1 = (const int8_t*) ((uintptr_t) a1 + a_offset);
111 } else {
112 a1 = zero_data;
113 }
114 const int8_t* restrict a2 = a[2];
115 if XNN_UNPREDICTABLE(a2 != zero) {
116 a2 = (const int8_t*) ((uintptr_t) a2 + a_offset);
117 } else {
118 a2 = zero_data;
119 }
120 const int8_t* restrict a3 = a[3];
121 if XNN_UNPREDICTABLE(a3 != zero) {
122 a3 = (const int8_t*) ((uintptr_t) a3 + a_offset);
123 } else {
124 a3 = zero_data;
125 }
126 a += 4;
127
128 uint64x2x2_t va01x0123456789ABCDEF;
129 va01x0123456789ABCDEF.val[0] = vdupq_n_u64(0);
130 va01x0123456789ABCDEF.val[1] = vdupq_n_u64(0);
131 uint64x2x2_t va23x0123456789ABCDEF;
132 va23x0123456789ABCDEF.val[0] = vdupq_n_u64(0);
133 va23x0123456789ABCDEF.val[1] = vdupq_n_u64(0);
134
135 // Inner accumulation loop along the 16 columns.
136 size_t k = kc;
137 // 2x partial unrolled loop to load 8 bytes at a time.
138 while (k >= 16 * sizeof(int8_t)) {
139 // Load a 4x16 block of activations.
140 va01x0123456789ABCDEF = vld2q_lane_u64((const void*) a0, va01x0123456789ABCDEF, 0); a0 += 16;
141 va23x0123456789ABCDEF = vld2q_lane_u64((const void*) a2, va23x0123456789ABCDEF, 0); a2 += 16;
142 va01x0123456789ABCDEF = vld2q_lane_u64((const void*) a1, va01x0123456789ABCDEF, 1); a1 += 16;
143 va23x0123456789ABCDEF = vld2q_lane_u64((const void*) a3, va23x0123456789ABCDEF, 1); a3 += 16;
144
145 // Load a 16x16 block of weights.
146 const int8x16_t vb01x01234567 = vld1q_s8(w); w = (const int8_t*) w + 16;
147 const int8x16_t vb23x01234567 = vld1q_s8(w); w = (const int8_t*) w + 16;
148 const int8x16_t vb45x01234567 = vld1q_s8(w); w = (const int8_t*) w + 16;
149 const int8x16_t vb67x01234567 = vld1q_s8(w); w = (const int8_t*) w + 16;
150 const int8x16_t vb89x01234567 = vld1q_s8(w); w = (const int8_t*) w + 16;
151 const int8x16_t vbABx01234567 = vld1q_s8(w); w = (const int8_t*) w + 16;
152 const int8x16_t vbCDx01234567 = vld1q_s8(w); w = (const int8_t*) w + 16;
153 const int8x16_t vbEFx01234567 = vld1q_s8(w); w = (const int8_t*) w + 16;
154 const int8x16_t vb01x89ABCDEF = vld1q_s8(w); w = (const int8_t*) w + 16;
155 const int8x16_t vb23x89ABCDEF = vld1q_s8(w); w = (const int8_t*) w + 16;
156 const int8x16_t vb45x89ABCDEF = vld1q_s8(w); w = (const int8_t*) w + 16;
157 const int8x16_t vb67x89ABCDEF = vld1q_s8(w); w = (const int8_t*) w + 16;
158 const int8x16_t vb89x89ABCDEF = vld1q_s8(w); w = (const int8_t*) w + 16;
159 const int8x16_t vbABx89ABCDEF = vld1q_s8(w); w = (const int8_t*) w + 16;
160 const int8x16_t vbCDx89ABCDEF = vld1q_s8(w); w = (const int8_t*) w + 16;
161 const int8x16_t vbEFx89ABCDEF = vld1q_s8(w); w = (const int8_t*) w + 16;
162
163 // Multiply-accumulate: 4x16 * 16x16 --> 4x16.
164 vacc01x01 = vmmlaq_s32(vacc01x01, vreinterpretq_s8_u64(va01x0123456789ABCDEF.val[0]), vb01x01234567);
165 vacc01x23 = vmmlaq_s32(vacc01x23, vreinterpretq_s8_u64(va01x0123456789ABCDEF.val[0]), vb23x01234567);
166 vacc01x45 = vmmlaq_s32(vacc01x45, vreinterpretq_s8_u64(va01x0123456789ABCDEF.val[0]), vb45x01234567);
167 vacc01x67 = vmmlaq_s32(vacc01x67, vreinterpretq_s8_u64(va01x0123456789ABCDEF.val[0]), vb67x01234567);
168 vacc01x89 = vmmlaq_s32(vacc01x89, vreinterpretq_s8_u64(va01x0123456789ABCDEF.val[0]), vb89x01234567);
169 vacc01xAB = vmmlaq_s32(vacc01xAB, vreinterpretq_s8_u64(va01x0123456789ABCDEF.val[0]), vbABx01234567);
170 vacc01xCD = vmmlaq_s32(vacc01xCD, vreinterpretq_s8_u64(va01x0123456789ABCDEF.val[0]), vbCDx01234567);
171 vacc01xEF = vmmlaq_s32(vacc01xEF, vreinterpretq_s8_u64(va01x0123456789ABCDEF.val[0]), vbEFx01234567);
172 vacc23x01 = vmmlaq_s32(vacc23x01, vreinterpretq_s8_u64(va23x0123456789ABCDEF.val[0]), vb01x01234567);
173 vacc23x23 = vmmlaq_s32(vacc23x23, vreinterpretq_s8_u64(va23x0123456789ABCDEF.val[0]), vb23x01234567);
174 vacc23x45 = vmmlaq_s32(vacc23x45, vreinterpretq_s8_u64(va23x0123456789ABCDEF.val[0]), vb45x01234567);
175 vacc23x67 = vmmlaq_s32(vacc23x67, vreinterpretq_s8_u64(va23x0123456789ABCDEF.val[0]), vb67x01234567);
176 vacc23x89 = vmmlaq_s32(vacc23x89, vreinterpretq_s8_u64(va23x0123456789ABCDEF.val[0]), vb89x01234567);
177 vacc23xAB = vmmlaq_s32(vacc23xAB, vreinterpretq_s8_u64(va23x0123456789ABCDEF.val[0]), vbABx01234567);
178 vacc23xCD = vmmlaq_s32(vacc23xCD, vreinterpretq_s8_u64(va23x0123456789ABCDEF.val[0]), vbCDx01234567);
179 vacc23xEF = vmmlaq_s32(vacc23xEF, vreinterpretq_s8_u64(va23x0123456789ABCDEF.val[0]), vbEFx01234567);
180 vacc01x01 = vmmlaq_s32(vacc01x01, vreinterpretq_s8_u64(va01x0123456789ABCDEF.val[1]), vb01x89ABCDEF);
181 vacc01x23 = vmmlaq_s32(vacc01x23, vreinterpretq_s8_u64(va01x0123456789ABCDEF.val[1]), vb23x89ABCDEF);
182 vacc01x45 = vmmlaq_s32(vacc01x45, vreinterpretq_s8_u64(va01x0123456789ABCDEF.val[1]), vb45x89ABCDEF);
183 vacc01x67 = vmmlaq_s32(vacc01x67, vreinterpretq_s8_u64(va01x0123456789ABCDEF.val[1]), vb67x89ABCDEF);
184 vacc01x89 = vmmlaq_s32(vacc01x89, vreinterpretq_s8_u64(va01x0123456789ABCDEF.val[1]), vb89x89ABCDEF);
185 vacc01xAB = vmmlaq_s32(vacc01xAB, vreinterpretq_s8_u64(va01x0123456789ABCDEF.val[1]), vbABx89ABCDEF);
186 vacc01xCD = vmmlaq_s32(vacc01xCD, vreinterpretq_s8_u64(va01x0123456789ABCDEF.val[1]), vbCDx89ABCDEF);
187 vacc01xEF = vmmlaq_s32(vacc01xEF, vreinterpretq_s8_u64(va01x0123456789ABCDEF.val[1]), vbEFx89ABCDEF);
188 vacc23x01 = vmmlaq_s32(vacc23x01, vreinterpretq_s8_u64(va23x0123456789ABCDEF.val[1]), vb01x89ABCDEF);
189 vacc23x23 = vmmlaq_s32(vacc23x23, vreinterpretq_s8_u64(va23x0123456789ABCDEF.val[1]), vb23x89ABCDEF);
190 vacc23x45 = vmmlaq_s32(vacc23x45, vreinterpretq_s8_u64(va23x0123456789ABCDEF.val[1]), vb45x89ABCDEF);
191 vacc23x67 = vmmlaq_s32(vacc23x67, vreinterpretq_s8_u64(va23x0123456789ABCDEF.val[1]), vb67x89ABCDEF);
192 vacc23x89 = vmmlaq_s32(vacc23x89, vreinterpretq_s8_u64(va23x0123456789ABCDEF.val[1]), vb89x89ABCDEF);
193 vacc23xAB = vmmlaq_s32(vacc23xAB, vreinterpretq_s8_u64(va23x0123456789ABCDEF.val[1]), vbABx89ABCDEF);
194 vacc23xCD = vmmlaq_s32(vacc23xCD, vreinterpretq_s8_u64(va23x0123456789ABCDEF.val[1]), vbCDx89ABCDEF);
195 vacc23xEF = vmmlaq_s32(vacc23xEF, vreinterpretq_s8_u64(va23x0123456789ABCDEF.val[1]), vbEFx89ABCDEF);
196
197 k -= 16 * sizeof(int8_t);
198 }
199 // Handle up to 8 final positions of `k`
200 if XNN_UNLIKELY(k != 0) {
201 // Load a 4x8 block of activations.
202 uint64x2_t va01x01234567 = vld1q_dup_u64((const void*) a0); a0 += 8;
203 uint64x2_t va23x01234567 = vld1q_dup_u64((const void*) a2); a2 += 8;
204 va01x01234567 = vld1q_lane_u64((const void*) a1, va01x01234567, 1); a1 += 8;
205 va23x01234567 = vld1q_lane_u64((const void*) a3, va23x01234567, 1); a3 += 8;
206
207 // Load a 16x16 block of weights.
208 const int8x16_t vb01x01234567 = vld1q_s8(w); w = (const int8_t*) w + 16;
209 const int8x16_t vb23x01234567 = vld1q_s8(w); w = (const int8_t*) w + 16;
210 const int8x16_t vb45x01234567 = vld1q_s8(w); w = (const int8_t*) w + 16;
211 const int8x16_t vb67x01234567 = vld1q_s8(w); w = (const int8_t*) w + 16;
212 const int8x16_t vb89x01234567 = vld1q_s8(w); w = (const int8_t*) w + 16;
213 const int8x16_t vbABx01234567 = vld1q_s8(w); w = (const int8_t*) w + 16;
214 const int8x16_t vbCDx01234567 = vld1q_s8(w); w = (const int8_t*) w + 16;
215 const int8x16_t vbEFx01234567 = vld1q_s8(w); w = (const int8_t*) w + 16;
216
217 // Multiply-accumulate: 4x4 * 4x16 --> 4x16.
218 vacc01x01 = vmmlaq_s32(vacc01x01, vreinterpretq_s8_u64(va01x01234567), vb01x01234567);
219 vacc01x23 = vmmlaq_s32(vacc01x23, vreinterpretq_s8_u64(va01x01234567), vb23x01234567);
220 vacc01x45 = vmmlaq_s32(vacc01x45, vreinterpretq_s8_u64(va01x01234567), vb45x01234567);
221 vacc01x67 = vmmlaq_s32(vacc01x67, vreinterpretq_s8_u64(va01x01234567), vb67x01234567);
222 vacc01x89 = vmmlaq_s32(vacc01x89, vreinterpretq_s8_u64(va01x01234567), vb89x01234567);
223 vacc01xAB = vmmlaq_s32(vacc01xAB, vreinterpretq_s8_u64(va01x01234567), vbABx01234567);
224 vacc01xCD = vmmlaq_s32(vacc01xCD, vreinterpretq_s8_u64(va01x01234567), vbCDx01234567);
225 vacc01xEF = vmmlaq_s32(vacc01xEF, vreinterpretq_s8_u64(va01x01234567), vbEFx01234567);
226 vacc23x01 = vmmlaq_s32(vacc23x01, vreinterpretq_s8_u64(va23x01234567), vb01x01234567);
227 vacc23x23 = vmmlaq_s32(vacc23x23, vreinterpretq_s8_u64(va23x01234567), vb23x01234567);
228 vacc23x45 = vmmlaq_s32(vacc23x45, vreinterpretq_s8_u64(va23x01234567), vb45x01234567);
229 vacc23x67 = vmmlaq_s32(vacc23x67, vreinterpretq_s8_u64(va23x01234567), vb67x01234567);
230 vacc23x89 = vmmlaq_s32(vacc23x89, vreinterpretq_s8_u64(va23x01234567), vb89x01234567);
231 vacc23xAB = vmmlaq_s32(vacc23xAB, vreinterpretq_s8_u64(va23x01234567), vbABx01234567);
232 vacc23xCD = vmmlaq_s32(vacc23xCD, vreinterpretq_s8_u64(va23x01234567), vbCDx01234567);
233 vacc23xEF = vmmlaq_s32(vacc23xEF, vreinterpretq_s8_u64(va23x01234567), vbEFx01234567);
234 }
235
236 p -= 4 * sizeof(void*);
237 } while (p != 0);
238
239 int32x4_t vacc0x0123 = vcombine_s32(vget_low_s32(vacc01x01), vget_low_s32(vacc01x23));
240 int32x4_t vacc1x0123 = vcombine_s32(vget_high_s32(vacc01x01), vget_high_s32(vacc01x23));
241 int32x4_t vacc0x4567 = vcombine_s32(vget_low_s32(vacc01x45), vget_low_s32(vacc01x67));
242 int32x4_t vacc1x4567 = vcombine_s32(vget_high_s32(vacc01x45), vget_high_s32(vacc01x67));
243 int32x4_t vacc0x89AB = vcombine_s32(vget_low_s32(vacc01x89), vget_low_s32(vacc01xAB));
244 int32x4_t vacc1x89AB = vcombine_s32(vget_high_s32(vacc01x89), vget_high_s32(vacc01xAB));
245 int32x4_t vacc0xCDEF = vcombine_s32(vget_low_s32(vacc01xCD), vget_low_s32(vacc01xEF));
246 int32x4_t vacc1xCDEF = vcombine_s32(vget_high_s32(vacc01xCD), vget_high_s32(vacc01xEF));
247 int32x4_t vacc2x0123 = vcombine_s32(vget_low_s32(vacc23x01), vget_low_s32(vacc23x23));
248 int32x4_t vacc3x0123 = vcombine_s32(vget_high_s32(vacc23x01), vget_high_s32(vacc23x23));
249 int32x4_t vacc2x4567 = vcombine_s32(vget_low_s32(vacc23x45), vget_low_s32(vacc23x67));
250 int32x4_t vacc3x4567 = vcombine_s32(vget_high_s32(vacc23x45), vget_high_s32(vacc23x67));
251 int32x4_t vacc2x89AB = vcombine_s32(vget_low_s32(vacc23x89), vget_low_s32(vacc23xAB));
252 int32x4_t vacc3x89AB = vcombine_s32(vget_high_s32(vacc23x89), vget_high_s32(vacc23xAB));
253 int32x4_t vacc2xCDEF = vcombine_s32(vget_low_s32(vacc23xCD), vget_low_s32(vacc23xEF));
254 int32x4_t vacc3xCDEF = vcombine_s32(vget_high_s32(vacc23xCD), vget_high_s32(vacc23xEF));
255
256 const float32x4_t vinput_scale = vld1q_dup_f32(&quantization_params->inv_scale);
257 float32x4_t vout0x0123 = vcvtq_f32_s32(vacc0x0123);
258 float32x4_t vout0x4567 = vcvtq_f32_s32(vacc0x4567);
259 float32x4_t vout0x89AB = vcvtq_f32_s32(vacc0x89AB);
260 float32x4_t vout0xCDEF = vcvtq_f32_s32(vacc0xCDEF);
261 float32x4_t vout1x0123 = vcvtq_f32_s32(vacc1x0123);
262 float32x4_t vout1x4567 = vcvtq_f32_s32(vacc1x4567);
263 float32x4_t vout1x89AB = vcvtq_f32_s32(vacc1x89AB);
264 float32x4_t vout1xCDEF = vcvtq_f32_s32(vacc1xCDEF);
265 float32x4_t vout2x0123 = vcvtq_f32_s32(vacc2x0123);
266 float32x4_t vout2x4567 = vcvtq_f32_s32(vacc2x4567);
267 float32x4_t vout2x89AB = vcvtq_f32_s32(vacc2x89AB);
268 float32x4_t vout2xCDEF = vcvtq_f32_s32(vacc2xCDEF);
269 float32x4_t vout3x0123 = vcvtq_f32_s32(vacc3x0123);
270 float32x4_t vout3x4567 = vcvtq_f32_s32(vacc3x4567);
271 float32x4_t vout3x89AB = vcvtq_f32_s32(vacc3x89AB);
272 float32x4_t vout3xCDEF = vcvtq_f32_s32(vacc3xCDEF);
273 vout0x0123 = vmulq_f32(vout0x0123, vinput_scale);
274 vout1x0123 = vmulq_f32(vout1x0123, vinput_scale);
275 vout0x4567 = vmulq_f32(vout0x4567, vinput_scale);
276 vout1x4567 = vmulq_f32(vout1x4567, vinput_scale);
277 vout0x89AB = vmulq_f32(vout0x89AB, vinput_scale);
278 vout1x89AB = vmulq_f32(vout1x89AB, vinput_scale);
279 vout0xCDEF = vmulq_f32(vout0xCDEF, vinput_scale);
280 vout1xCDEF = vmulq_f32(vout1xCDEF, vinput_scale);
281 vout2x0123 = vmulq_f32(vout2x0123, vinput_scale);
282 vout3x0123 = vmulq_f32(vout3x0123, vinput_scale);
283 vout2x4567 = vmulq_f32(vout2x4567, vinput_scale);
284 vout3x4567 = vmulq_f32(vout3x4567, vinput_scale);
285 vout2x89AB = vmulq_f32(vout2x89AB, vinput_scale);
286 vout3x89AB = vmulq_f32(vout3x89AB, vinput_scale);
287 vout2xCDEF = vmulq_f32(vout2xCDEF, vinput_scale);
288 vout3xCDEF = vmulq_f32(vout3xCDEF, vinput_scale);
289
290 const float32x4_t vfilter_output_scale0123 = vld1q_f32(w); w = (const float*) w + 4;
291 const float32x4_t vfilter_output_scale4567 = vld1q_f32(w); w = (const float*) w + 4;
292 const float32x4_t vfilter_output_scale89AB = vld1q_f32(w); w = (const float*) w + 4;
293 const float32x4_t vfilter_output_scaleCDEF = vld1q_f32(w); w = (const float*) w + 4;
294
295 const float32x4_t vbias0123 = vld1q_f32(w); w = (const float*) w + 4;
296 vout0x0123 = vfmaq_f32(vbias0123, vout0x0123, vfilter_output_scale0123);
297 vout1x0123 = vfmaq_f32(vbias0123, vout1x0123, vfilter_output_scale0123);
298 vout2x0123 = vfmaq_f32(vbias0123, vout2x0123, vfilter_output_scale0123);
299 vout3x0123 = vfmaq_f32(vbias0123, vout3x0123, vfilter_output_scale0123);
300 const float32x4_t vbias4567 = vld1q_f32(w); w = (const float*) w + 4;
301 vout0x4567 = vfmaq_f32(vbias4567, vout0x4567, vfilter_output_scale4567);
302 vout1x4567 = vfmaq_f32(vbias4567, vout1x4567, vfilter_output_scale4567);
303 vout2x4567 = vfmaq_f32(vbias4567, vout2x4567, vfilter_output_scale4567);
304 vout3x4567 = vfmaq_f32(vbias4567, vout3x4567, vfilter_output_scale4567);
305 const float32x4_t vbias89AB = vld1q_f32(w); w = (const float*) w + 4;
306 vout0x89AB = vfmaq_f32(vbias89AB, vout0x89AB, vfilter_output_scale89AB);
307 vout1x89AB = vfmaq_f32(vbias89AB, vout1x89AB, vfilter_output_scale89AB);
308 vout2x89AB = vfmaq_f32(vbias89AB, vout2x89AB, vfilter_output_scale89AB);
309 vout3x89AB = vfmaq_f32(vbias89AB, vout3x89AB, vfilter_output_scale89AB);
310 const float32x4_t vbiasCDEF = vld1q_f32(w); w = (const float*) w + 4;
311 vout0xCDEF = vfmaq_f32(vbiasCDEF, vout0xCDEF, vfilter_output_scaleCDEF);
312 vout1xCDEF = vfmaq_f32(vbiasCDEF, vout1xCDEF, vfilter_output_scaleCDEF);
313 vout2xCDEF = vfmaq_f32(vbiasCDEF, vout2xCDEF, vfilter_output_scaleCDEF);
314 vout3xCDEF = vfmaq_f32(vbiasCDEF, vout3xCDEF, vfilter_output_scaleCDEF);
315
316 float16x8_t vfp16out3x01234567 = vcombine_f16(vcvt_f16_f32(vout3x0123), vcvt_f16_f32(vout3x4567));
317 float16x8_t vfp16out3x89ABCDEF = vcombine_f16(vcvt_f16_f32(vout3x89AB), vcvt_f16_f32(vout3xCDEF));
318 float16x8_t vfp16out2x01234567 = vcombine_f16(vcvt_f16_f32(vout2x0123), vcvt_f16_f32(vout2x4567));
319 float16x8_t vfp16out2x89ABCDEF = vcombine_f16(vcvt_f16_f32(vout2x89AB), vcvt_f16_f32(vout2xCDEF));
320 float16x8_t vfp16out1x01234567 = vcombine_f16(vcvt_f16_f32(vout1x0123), vcvt_f16_f32(vout1x4567));
321 float16x8_t vfp16out1x89ABCDEF = vcombine_f16(vcvt_f16_f32(vout1x89AB), vcvt_f16_f32(vout1xCDEF));
322 float16x8_t vfp16out0x01234567 = vcombine_f16(vcvt_f16_f32(vout0x0123), vcvt_f16_f32(vout0x4567));
323 float16x8_t vfp16out0x89ABCDEF = vcombine_f16(vcvt_f16_f32(vout0x89AB), vcvt_f16_f32(vout0xCDEF));
324 const float16x8_t voutput_min = vreinterpretq_f16_u16(vld1q_dup_u16((const uint16_t*) &params->scalar.min));
325 vfp16out3x01234567 = vmaxq_f16(vfp16out3x01234567, voutput_min);
326 vfp16out3x89ABCDEF = vmaxq_f16(vfp16out3x89ABCDEF, voutput_min);
327 vfp16out2x01234567 = vmaxq_f16(vfp16out2x01234567, voutput_min);
328 vfp16out2x89ABCDEF = vmaxq_f16(vfp16out2x89ABCDEF, voutput_min);
329 vfp16out1x01234567 = vmaxq_f16(vfp16out1x01234567, voutput_min);
330 vfp16out1x89ABCDEF = vmaxq_f16(vfp16out1x89ABCDEF, voutput_min);
331 vfp16out0x01234567 = vmaxq_f16(vfp16out0x01234567, voutput_min);
332 vfp16out0x89ABCDEF = vmaxq_f16(vfp16out0x89ABCDEF, voutput_min);
333 const float16x8_t voutput_max = vreinterpretq_f16_u16(vld1q_dup_u16((const uint16_t*) &params->scalar.max));
334 vfp16out3x01234567 = vminq_f16(vfp16out3x01234567, voutput_max);
335 vfp16out3x89ABCDEF = vminq_f16(vfp16out3x89ABCDEF, voutput_max);
336 vfp16out2x01234567 = vminq_f16(vfp16out2x01234567, voutput_max);
337 vfp16out2x89ABCDEF = vminq_f16(vfp16out2x89ABCDEF, voutput_max);
338 vfp16out1x01234567 = vminq_f16(vfp16out1x01234567, voutput_max);
339 vfp16out1x89ABCDEF = vminq_f16(vfp16out1x89ABCDEF, voutput_max);
340 vfp16out0x01234567 = vminq_f16(vfp16out0x01234567, voutput_max);
341 vfp16out0x89ABCDEF = vminq_f16(vfp16out0x89ABCDEF, voutput_max);
342
343 if XNN_LIKELY(nc >= 16) {
344 vst1q_u16(c3, vreinterpretq_u16_f16(vfp16out3x01234567));
345 vst1q_u16(c3 + 8, vreinterpretq_u16_f16(vfp16out3x89ABCDEF));
346 vst1q_u16(c2, vreinterpretq_u16_f16(vfp16out2x01234567));
347 vst1q_u16(c2 + 8, vreinterpretq_u16_f16(vfp16out2x89ABCDEF));
348 vst1q_u16(c1, vreinterpretq_u16_f16(vfp16out1x01234567));
349 vst1q_u16(c1 + 8, vreinterpretq_u16_f16(vfp16out1x89ABCDEF));
350 vst1q_u16(c0, vreinterpretq_u16_f16(vfp16out0x01234567));
351 vst1q_u16(c0 + 8, vreinterpretq_u16_f16(vfp16out0x89ABCDEF));
352
353 c0 = (uint16_t*) ((uintptr_t) c0 + cn_stride);
354 c1 = (uint16_t*) ((uintptr_t) c1 + cn_stride);
355 c2 = (uint16_t*) ((uintptr_t) c2 + cn_stride);
356 c3 = (uint16_t*) ((uintptr_t) c3 + cn_stride);
357
358 a = (const int8_t**restrict) ((uintptr_t) a - ks);
359
360 nc -= 16;
361 } else {
362 if (nc & 8) {
363 vst1q_u16(c3, vreinterpretq_u16_f16(vfp16out3x01234567)); c3 += 8;
364 vfp16out3x01234567 = vfp16out3x89ABCDEF;
365 vst1q_u16(c2, vreinterpretq_u16_f16(vfp16out2x01234567)); c2 += 8;
366 vfp16out2x01234567 = vfp16out2x89ABCDEF;
367 vst1q_u16(c1, vreinterpretq_u16_f16(vfp16out1x01234567)); c1 += 8;
368 vfp16out1x01234567 = vfp16out1x89ABCDEF;
369 vst1q_u16(c0, vreinterpretq_u16_f16(vfp16out0x01234567)); c0 += 8;
370 vfp16out0x01234567 = vfp16out0x89ABCDEF;
371 }
372 float16x4_t vfp16out3x0123 = vget_low_f16(vfp16out3x01234567);
373 float16x4_t vfp16out2x0123 = vget_low_f16(vfp16out2x01234567);
374 float16x4_t vfp16out1x0123 = vget_low_f16(vfp16out1x01234567);
375 float16x4_t vfp16out0x0123 = vget_low_f16(vfp16out0x01234567);
376 if (nc & 4) {
377 vst1_u16(c3, vreinterpret_u16_f16(vfp16out3x0123)); c3 += 4;
378 vst1_u16(c2, vreinterpret_u16_f16(vfp16out2x0123)); c2 += 4;
379 vst1_u16(c1, vreinterpret_u16_f16(vfp16out1x0123)); c1 += 4;
380 vst1_u16(c0, vreinterpret_u16_f16(vfp16out0x0123)); c0 += 4;
381 vfp16out3x0123 = vget_high_f16(vfp16out3x01234567);
382 vfp16out2x0123 = vget_high_f16(vfp16out2x01234567);
383 vfp16out1x0123 = vget_high_f16(vfp16out1x01234567);
384 vfp16out0x0123 = vget_high_f16(vfp16out0x01234567);
385 }
386 if (nc & 2) {
387 vst1_lane_u32((void*) c3, vreinterpret_u32_f16(vfp16out3x0123), 0); c3 += 2;
388 vst1_lane_u32((void*) c2, vreinterpret_u32_f16(vfp16out2x0123), 0); c2 += 2;
389 vst1_lane_u32((void*) c1, vreinterpret_u32_f16(vfp16out1x0123), 0); c1 += 2;
390 vst1_lane_u32((void*) c0, vreinterpret_u32_f16(vfp16out0x0123), 0); c0 += 2;
391 vfp16out3x0123 = vext_f16(vfp16out3x0123, vfp16out3x0123, 2);
392 vfp16out2x0123 = vext_f16(vfp16out2x0123, vfp16out2x0123, 2);
393 vfp16out1x0123 = vext_f16(vfp16out1x0123, vfp16out1x0123, 2);
394 vfp16out0x0123 = vext_f16(vfp16out0x0123, vfp16out0x0123, 2);
395 }
396 if (nc & 1) {
397 vst1_lane_u16(c3, vreinterpret_u16_f16(vfp16out3x0123), 0);
398 vst1_lane_u16(c2, vreinterpret_u16_f16(vfp16out2x0123), 0);
399 vst1_lane_u16(c1, vreinterpret_u16_f16(vfp16out1x0123), 0);
400 vst1_lane_u16(c0, vreinterpret_u16_f16(vfp16out0x0123), 0);
401 }
402 nc = 0;
403 }
404 } while (nc != 0);
405 }