Debian Codegraph @

src/qd8-f32-qc8w-igemm/gen/qd8-f32-qc8w-igemm-3x16c8-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_f32_qc8w_igemm_minmax_ukernel_3x16c8__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 float* 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_f32_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 <= 3);
36 assert(nc != 0);
37 assert(kc != 0);
38 assert(ks != 0);
39 assert(ks % (3 * 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 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
56 do {
57 const int32x4_t vinput_zero_point = vld1q_dup_s32(&quantization_params->zero_point);
58 const int32x4_t vksum0123 = vld1q_s32(w); w = (const int32_t*) w + 4;
59 const int32x4_t vksumzp0x0123 = vmulq_s32(vksum0123, vinput_zero_point);
60 const int32x4_t vksum4567 = vld1q_s32(w); w = (const int32_t*) w + 4;
61 const int32x4_t vksumzp0x4567 = vmulq_s32(vksum4567, vinput_zero_point);
62 const int32x4_t vksum89AB = vld1q_s32(w); w = (const int32_t*) w + 4;
63 const int32x4_t vksumzp0x89AB = vmulq_s32(vksum89AB, vinput_zero_point);
64 const int32x4_t vksumCDEF = vld1q_s32(w); w = (const int32_t*) w + 4;
65 const int32x4_t vksumzp0xCDEF = vmulq_s32(vksumCDEF, vinput_zero_point);
66 const int32x4_t vksumzp1x0123 = vksumzp0x0123;
67 const int32x4_t vksumzp1x4567 = vksumzp0x4567;
68 const int32x4_t vksumzp1x89AB = vksumzp0x89AB;
69 const int32x4_t vksumzp1xCDEF = vksumzp0xCDEF;
70 const int32x4_t vksumzp2x0123 = vksumzp0x0123;
71 const int32x4_t vksumzp2x4567 = vksumzp0x4567;
72 const int32x4_t vksumzp2x89AB = vksumzp0x89AB;
73 const int32x4_t vksumzp2xCDEF = vksumzp0xCDEF;
74
75 int32x4_t vacc01x01 = vreinterpretq_s32_u64(vtrn1q_u64(vreinterpretq_u64_s32(vksumzp0x0123), vreinterpretq_u64_s32(vksumzp1x0123)));
76 int32x4_t vacc01x23 = vreinterpretq_s32_u64(vtrn2q_u64(vreinterpretq_u64_s32(vksumzp0x0123), vreinterpretq_u64_s32(vksumzp1x0123)));
77 int32x4_t vacc01x45 = vreinterpretq_s32_u64(vtrn1q_u64(vreinterpretq_u64_s32(vksumzp0x4567), vreinterpretq_u64_s32(vksumzp1x4567)));
78 int32x4_t vacc01x67 = vreinterpretq_s32_u64(vtrn2q_u64(vreinterpretq_u64_s32(vksumzp0x4567), vreinterpretq_u64_s32(vksumzp1x4567)));
79 int32x4_t vacc01x89 = vreinterpretq_s32_u64(vtrn1q_u64(vreinterpretq_u64_s32(vksumzp0x89AB), vreinterpretq_u64_s32(vksumzp1x89AB)));
80 int32x4_t vacc01xAB = vreinterpretq_s32_u64(vtrn2q_u64(vreinterpretq_u64_s32(vksumzp0x89AB), vreinterpretq_u64_s32(vksumzp1x89AB)));
81 int32x4_t vacc01xCD = vreinterpretq_s32_u64(vtrn1q_u64(vreinterpretq_u64_s32(vksumzp0xCDEF), vreinterpretq_u64_s32(vksumzp1xCDEF)));
82 int32x4_t vacc01xEF = vreinterpretq_s32_u64(vtrn2q_u64(vreinterpretq_u64_s32(vksumzp0xCDEF), vreinterpretq_u64_s32(vksumzp1xCDEF)));
83 int32x4_t vacc23x01 = vreinterpretq_s32_u64(vtrn1q_u64(vreinterpretq_u64_s32(vksumzp2x0123), vreinterpretq_u64_s32(vksumzp2x0123)));
84 int32x4_t vacc23x23 = vreinterpretq_s32_u64(vtrn2q_u64(vreinterpretq_u64_s32(vksumzp2x0123), vreinterpretq_u64_s32(vksumzp2x0123)));
85 int32x4_t vacc23x45 = vreinterpretq_s32_u64(vtrn1q_u64(vreinterpretq_u64_s32(vksumzp2x4567), vreinterpretq_u64_s32(vksumzp2x4567)));
86 int32x4_t vacc23x67 = vreinterpretq_s32_u64(vtrn2q_u64(vreinterpretq_u64_s32(vksumzp2x4567), vreinterpretq_u64_s32(vksumzp2x4567)));
87 int32x4_t vacc23x89 = vreinterpretq_s32_u64(vtrn1q_u64(vreinterpretq_u64_s32(vksumzp2x89AB), vreinterpretq_u64_s32(vksumzp2x89AB)));
88 int32x4_t vacc23xAB = vreinterpretq_s32_u64(vtrn2q_u64(vreinterpretq_u64_s32(vksumzp2x89AB), vreinterpretq_u64_s32(vksumzp2x89AB)));
89 int32x4_t vacc23xCD = vreinterpretq_s32_u64(vtrn1q_u64(vreinterpretq_u64_s32(vksumzp2xCDEF), vreinterpretq_u64_s32(vksumzp2xCDEF)));
90 int32x4_t vacc23xEF = vreinterpretq_s32_u64(vtrn2q_u64(vreinterpretq_u64_s32(vksumzp2xCDEF), vreinterpretq_u64_s32(vksumzp2xCDEF)));
91
92 size_t p = ks;
93 do {
94 const int8_t* restrict a0 = a[0];
95 if XNN_UNPREDICTABLE(a0 != zero) {
96 a0 = (const int8_t*) ((uintptr_t) a0 + a_offset);
97 } else {
98 a0 = zero_data;
99 }
100 const int8_t* restrict a1 = a[1];
101 if XNN_UNPREDICTABLE(a1 != zero) {
102 a1 = (const int8_t*) ((uintptr_t) a1 + a_offset);
103 } else {
104 a1 = zero_data;
105 }
106 const int8_t* restrict a2 = a[2];
107 if XNN_UNPREDICTABLE(a2 != zero) {
108 a2 = (const int8_t*) ((uintptr_t) a2 + a_offset);
109 } else {
110 a2 = zero_data;
111 }
112 a += 3;
113
114 uint64x2x2_t va01x0123456789ABCDEF;
115 va01x0123456789ABCDEF.val[0] = vdupq_n_u64(0);
116 va01x0123456789ABCDEF.val[1] = vdupq_n_u64(0);
117 uint64x2x2_t va23x0123456789ABCDEF;
118 va23x0123456789ABCDEF.val[0] = vdupq_n_u64(0);
119 va23x0123456789ABCDEF.val[1] = vdupq_n_u64(0);
120
121 // Inner accumulation loop along the 16 columns.
122 size_t k = kc;
123 // 2x partial unrolled loop to load 8 bytes at a time.
124 while (k >= 16 * sizeof(int8_t)) {
125 // Load a 3x16 block of activations.
126 va01x0123456789ABCDEF = vld2q_lane_u64((const void*) a0, va01x0123456789ABCDEF, 0); a0 += 16;
127 va23x0123456789ABCDEF = vld2q_lane_u64((const void*) a2, va23x0123456789ABCDEF, 0); a2 += 16;
128 va01x0123456789ABCDEF = vld2q_lane_u64((const void*) a1, va01x0123456789ABCDEF, 1); a1 += 16;
129
130 // Load a 16x16 block of weights.
131 const int8x16_t vb01x01234567 = vld1q_s8(w); w = (const int8_t*) w + 16;
132 const int8x16_t vb23x01234567 = vld1q_s8(w); w = (const int8_t*) w + 16;
133 const int8x16_t vb45x01234567 = vld1q_s8(w); w = (const int8_t*) w + 16;
134 const int8x16_t vb67x01234567 = vld1q_s8(w); w = (const int8_t*) w + 16;
135 const int8x16_t vb89x01234567 = vld1q_s8(w); w = (const int8_t*) w + 16;
136 const int8x16_t vbABx01234567 = vld1q_s8(w); w = (const int8_t*) w + 16;
137 const int8x16_t vbCDx01234567 = vld1q_s8(w); w = (const int8_t*) w + 16;
138 const int8x16_t vbEFx01234567 = vld1q_s8(w); w = (const int8_t*) w + 16;
139 const int8x16_t vb01x89ABCDEF = vld1q_s8(w); w = (const int8_t*) w + 16;
140 const int8x16_t vb23x89ABCDEF = vld1q_s8(w); w = (const int8_t*) w + 16;
141 const int8x16_t vb45x89ABCDEF = vld1q_s8(w); w = (const int8_t*) w + 16;
142 const int8x16_t vb67x89ABCDEF = vld1q_s8(w); w = (const int8_t*) w + 16;
143 const int8x16_t vb89x89ABCDEF = vld1q_s8(w); w = (const int8_t*) w + 16;
144 const int8x16_t vbABx89ABCDEF = vld1q_s8(w); w = (const int8_t*) w + 16;
145 const int8x16_t vbCDx89ABCDEF = vld1q_s8(w); w = (const int8_t*) w + 16;
146 const int8x16_t vbEFx89ABCDEF = vld1q_s8(w); w = (const int8_t*) w + 16;
147
148 // Multiply-accumulate: 3x16 * 16x16 --> 3x16.
149 vacc01x01 = vmmlaq_s32(vacc01x01, vreinterpretq_s8_u64(va01x0123456789ABCDEF.val[0]), vb01x01234567);
150 vacc01x23 = vmmlaq_s32(vacc01x23, vreinterpretq_s8_u64(va01x0123456789ABCDEF.val[0]), vb23x01234567);
151 vacc01x45 = vmmlaq_s32(vacc01x45, vreinterpretq_s8_u64(va01x0123456789ABCDEF.val[0]), vb45x01234567);
152 vacc01x67 = vmmlaq_s32(vacc01x67, vreinterpretq_s8_u64(va01x0123456789ABCDEF.val[0]), vb67x01234567);
153 vacc01x89 = vmmlaq_s32(vacc01x89, vreinterpretq_s8_u64(va01x0123456789ABCDEF.val[0]), vb89x01234567);
154 vacc01xAB = vmmlaq_s32(vacc01xAB, vreinterpretq_s8_u64(va01x0123456789ABCDEF.val[0]), vbABx01234567);
155 vacc01xCD = vmmlaq_s32(vacc01xCD, vreinterpretq_s8_u64(va01x0123456789ABCDEF.val[0]), vbCDx01234567);
156 vacc01xEF = vmmlaq_s32(vacc01xEF, vreinterpretq_s8_u64(va01x0123456789ABCDEF.val[0]), vbEFx01234567);
157 vacc23x01 = vmmlaq_s32(vacc23x01, vreinterpretq_s8_u64(va23x0123456789ABCDEF.val[0]), vb01x01234567);
158 vacc23x23 = vmmlaq_s32(vacc23x23, vreinterpretq_s8_u64(va23x0123456789ABCDEF.val[0]), vb23x01234567);
159 vacc23x45 = vmmlaq_s32(vacc23x45, vreinterpretq_s8_u64(va23x0123456789ABCDEF.val[0]), vb45x01234567);
160 vacc23x67 = vmmlaq_s32(vacc23x67, vreinterpretq_s8_u64(va23x0123456789ABCDEF.val[0]), vb67x01234567);
161 vacc23x89 = vmmlaq_s32(vacc23x89, vreinterpretq_s8_u64(va23x0123456789ABCDEF.val[0]), vb89x01234567);
162 vacc23xAB = vmmlaq_s32(vacc23xAB, vreinterpretq_s8_u64(va23x0123456789ABCDEF.val[0]), vbABx01234567);
163 vacc23xCD = vmmlaq_s32(vacc23xCD, vreinterpretq_s8_u64(va23x0123456789ABCDEF.val[0]), vbCDx01234567);
164 vacc23xEF = vmmlaq_s32(vacc23xEF, vreinterpretq_s8_u64(va23x0123456789ABCDEF.val[0]), vbEFx01234567);
165 vacc01x01 = vmmlaq_s32(vacc01x01, vreinterpretq_s8_u64(va01x0123456789ABCDEF.val[1]), vb01x89ABCDEF);
166 vacc01x23 = vmmlaq_s32(vacc01x23, vreinterpretq_s8_u64(va01x0123456789ABCDEF.val[1]), vb23x89ABCDEF);
167 vacc01x45 = vmmlaq_s32(vacc01x45, vreinterpretq_s8_u64(va01x0123456789ABCDEF.val[1]), vb45x89ABCDEF);
168 vacc01x67 = vmmlaq_s32(vacc01x67, vreinterpretq_s8_u64(va01x0123456789ABCDEF.val[1]), vb67x89ABCDEF);
169 vacc01x89 = vmmlaq_s32(vacc01x89, vreinterpretq_s8_u64(va01x0123456789ABCDEF.val[1]), vb89x89ABCDEF);
170 vacc01xAB = vmmlaq_s32(vacc01xAB, vreinterpretq_s8_u64(va01x0123456789ABCDEF.val[1]), vbABx89ABCDEF);
171 vacc01xCD = vmmlaq_s32(vacc01xCD, vreinterpretq_s8_u64(va01x0123456789ABCDEF.val[1]), vbCDx89ABCDEF);
172 vacc01xEF = vmmlaq_s32(vacc01xEF, vreinterpretq_s8_u64(va01x0123456789ABCDEF.val[1]), vbEFx89ABCDEF);
173 vacc23x01 = vmmlaq_s32(vacc23x01, vreinterpretq_s8_u64(va23x0123456789ABCDEF.val[1]), vb01x89ABCDEF);
174 vacc23x23 = vmmlaq_s32(vacc23x23, vreinterpretq_s8_u64(va23x0123456789ABCDEF.val[1]), vb23x89ABCDEF);
175 vacc23x45 = vmmlaq_s32(vacc23x45, vreinterpretq_s8_u64(va23x0123456789ABCDEF.val[1]), vb45x89ABCDEF);
176 vacc23x67 = vmmlaq_s32(vacc23x67, vreinterpretq_s8_u64(va23x0123456789ABCDEF.val[1]), vb67x89ABCDEF);
177 vacc23x89 = vmmlaq_s32(vacc23x89, vreinterpretq_s8_u64(va23x0123456789ABCDEF.val[1]), vb89x89ABCDEF);
178 vacc23xAB = vmmlaq_s32(vacc23xAB, vreinterpretq_s8_u64(va23x0123456789ABCDEF.val[1]), vbABx89ABCDEF);
179 vacc23xCD = vmmlaq_s32(vacc23xCD, vreinterpretq_s8_u64(va23x0123456789ABCDEF.val[1]), vbCDx89ABCDEF);
180 vacc23xEF = vmmlaq_s32(vacc23xEF, vreinterpretq_s8_u64(va23x0123456789ABCDEF.val[1]), vbEFx89ABCDEF);
181
182 k -= 16 * sizeof(int8_t);
183 }
184 // Handle up to 8 final positions of `k`
185 if XNN_UNLIKELY(k != 0) {
186 // Load a 3x8 block of activations.
187 uint64x2_t va01x01234567 = vld1q_dup_u64((const void*) a0); a0 += 8;
188 uint64x2_t va23x01234567 = vld1q_dup_u64((const void*) a2); a2 += 8;
189 va01x01234567 = vld1q_lane_u64((const void*) a1, va01x01234567, 1); a1 += 8;
190
191 // Load a 16x16 block of weights.
192 const int8x16_t vb01x01234567 = vld1q_s8(w); w = (const int8_t*) w + 16;
193 const int8x16_t vb23x01234567 = vld1q_s8(w); w = (const int8_t*) w + 16;
194 const int8x16_t vb45x01234567 = vld1q_s8(w); w = (const int8_t*) w + 16;
195 const int8x16_t vb67x01234567 = vld1q_s8(w); w = (const int8_t*) w + 16;
196 const int8x16_t vb89x01234567 = vld1q_s8(w); w = (const int8_t*) w + 16;
197 const int8x16_t vbABx01234567 = vld1q_s8(w); w = (const int8_t*) w + 16;
198 const int8x16_t vbCDx01234567 = vld1q_s8(w); w = (const int8_t*) w + 16;
199 const int8x16_t vbEFx01234567 = vld1q_s8(w); w = (const int8_t*) w + 16;
200
201 // Multiply-accumulate: 3x4 * 4x16 --> 3x16.
202 vacc01x01 = vmmlaq_s32(vacc01x01, vreinterpretq_s8_u64(va01x01234567), vb01x01234567);
203 vacc01x23 = vmmlaq_s32(vacc01x23, vreinterpretq_s8_u64(va01x01234567), vb23x01234567);
204 vacc01x45 = vmmlaq_s32(vacc01x45, vreinterpretq_s8_u64(va01x01234567), vb45x01234567);
205 vacc01x67 = vmmlaq_s32(vacc01x67, vreinterpretq_s8_u64(va01x01234567), vb67x01234567);
206 vacc01x89 = vmmlaq_s32(vacc01x89, vreinterpretq_s8_u64(va01x01234567), vb89x01234567);
207 vacc01xAB = vmmlaq_s32(vacc01xAB, vreinterpretq_s8_u64(va01x01234567), vbABx01234567);
208 vacc01xCD = vmmlaq_s32(vacc01xCD, vreinterpretq_s8_u64(va01x01234567), vbCDx01234567);
209 vacc01xEF = vmmlaq_s32(vacc01xEF, vreinterpretq_s8_u64(va01x01234567), vbEFx01234567);
210 vacc23x01 = vmmlaq_s32(vacc23x01, vreinterpretq_s8_u64(va23x01234567), vb01x01234567);
211 vacc23x23 = vmmlaq_s32(vacc23x23, vreinterpretq_s8_u64(va23x01234567), vb23x01234567);
212 vacc23x45 = vmmlaq_s32(vacc23x45, vreinterpretq_s8_u64(va23x01234567), vb45x01234567);
213 vacc23x67 = vmmlaq_s32(vacc23x67, vreinterpretq_s8_u64(va23x01234567), vb67x01234567);
214 vacc23x89 = vmmlaq_s32(vacc23x89, vreinterpretq_s8_u64(va23x01234567), vb89x01234567);
215 vacc23xAB = vmmlaq_s32(vacc23xAB, vreinterpretq_s8_u64(va23x01234567), vbABx01234567);
216 vacc23xCD = vmmlaq_s32(vacc23xCD, vreinterpretq_s8_u64(va23x01234567), vbCDx01234567);
217 vacc23xEF = vmmlaq_s32(vacc23xEF, vreinterpretq_s8_u64(va23x01234567), vbEFx01234567);
218 }
219
220 p -= 3 * sizeof(void*);
221 } while (p != 0);
222
223 int32x4_t vacc0x0123 = vcombine_s32(vget_low_s32(vacc01x01), vget_low_s32(vacc01x23));
224 int32x4_t vacc1x0123 = vcombine_s32(vget_high_s32(vacc01x01), vget_high_s32(vacc01x23));
225 int32x4_t vacc0x4567 = vcombine_s32(vget_low_s32(vacc01x45), vget_low_s32(vacc01x67));
226 int32x4_t vacc1x4567 = vcombine_s32(vget_high_s32(vacc01x45), vget_high_s32(vacc01x67));
227 int32x4_t vacc0x89AB = vcombine_s32(vget_low_s32(vacc01x89), vget_low_s32(vacc01xAB));
228 int32x4_t vacc1x89AB = vcombine_s32(vget_high_s32(vacc01x89), vget_high_s32(vacc01xAB));
229 int32x4_t vacc0xCDEF = vcombine_s32(vget_low_s32(vacc01xCD), vget_low_s32(vacc01xEF));
230 int32x4_t vacc1xCDEF = vcombine_s32(vget_high_s32(vacc01xCD), vget_high_s32(vacc01xEF));
231 int32x4_t vacc2x0123 = vcombine_s32(vget_low_s32(vacc23x01), vget_low_s32(vacc23x23));
232 int32x4_t vacc2x4567 = vcombine_s32(vget_low_s32(vacc23x45), vget_low_s32(vacc23x67));
233 int32x4_t vacc2x89AB = vcombine_s32(vget_low_s32(vacc23x89), vget_low_s32(vacc23xAB));
234 int32x4_t vacc2xCDEF = vcombine_s32(vget_low_s32(vacc23xCD), vget_low_s32(vacc23xEF));
235
236 const float32x4_t vinput_scale = vld1q_dup_f32(&quantization_params->inv_scale);
237 float32x4_t vout0x0123 = vcvtq_f32_s32(vacc0x0123);
238 float32x4_t vout0x4567 = vcvtq_f32_s32(vacc0x4567);
239 float32x4_t vout0x89AB = vcvtq_f32_s32(vacc0x89AB);
240 float32x4_t vout0xCDEF = vcvtq_f32_s32(vacc0xCDEF);
241 float32x4_t vout1x0123 = vcvtq_f32_s32(vacc1x0123);
242 float32x4_t vout1x4567 = vcvtq_f32_s32(vacc1x4567);
243 float32x4_t vout1x89AB = vcvtq_f32_s32(vacc1x89AB);
244 float32x4_t vout1xCDEF = vcvtq_f32_s32(vacc1xCDEF);
245 float32x4_t vout2x0123 = vcvtq_f32_s32(vacc2x0123);
246 float32x4_t vout2x4567 = vcvtq_f32_s32(vacc2x4567);
247 float32x4_t vout2x89AB = vcvtq_f32_s32(vacc2x89AB);
248 float32x4_t vout2xCDEF = vcvtq_f32_s32(vacc2xCDEF);
249 vout0x0123 = vmulq_f32(vout0x0123, vinput_scale);
250 vout1x0123 = vmulq_f32(vout1x0123, vinput_scale);
251 vout0x4567 = vmulq_f32(vout0x4567, vinput_scale);
252 vout1x4567 = vmulq_f32(vout1x4567, vinput_scale);
253 vout0x89AB = vmulq_f32(vout0x89AB, vinput_scale);
254 vout1x89AB = vmulq_f32(vout1x89AB, vinput_scale);
255 vout0xCDEF = vmulq_f32(vout0xCDEF, vinput_scale);
256 vout1xCDEF = vmulq_f32(vout1xCDEF, vinput_scale);
257 vout2x0123 = vmulq_f32(vout2x0123, vinput_scale);
258 vout2x4567 = vmulq_f32(vout2x4567, vinput_scale);
259 vout2x89AB = vmulq_f32(vout2x89AB, vinput_scale);
260 vout2xCDEF = vmulq_f32(vout2xCDEF, vinput_scale);
261
262 const float32x4_t vfilter_output_scale0123 = vld1q_f32(w); w = (const float*) w + 4;
263 const float32x4_t vfilter_output_scale4567 = vld1q_f32(w); w = (const float*) w + 4;
264 const float32x4_t vfilter_output_scale89AB = vld1q_f32(w); w = (const float*) w + 4;
265 const float32x4_t vfilter_output_scaleCDEF = vld1q_f32(w); w = (const float*) w + 4;
266
267 const float32x4_t vbias0123 = vld1q_f32(w); w = (const float*) w + 4;
268 vout0x0123 = vfmaq_f32(vbias0123, vout0x0123, vfilter_output_scale0123);
269 vout1x0123 = vfmaq_f32(vbias0123, vout1x0123, vfilter_output_scale0123);
270 vout2x0123 = vfmaq_f32(vbias0123, vout2x0123, vfilter_output_scale0123);
271 const float32x4_t vbias4567 = vld1q_f32(w); w = (const float*) w + 4;
272 vout0x4567 = vfmaq_f32(vbias4567, vout0x4567, vfilter_output_scale4567);
273 vout1x4567 = vfmaq_f32(vbias4567, vout1x4567, vfilter_output_scale4567);
274 vout2x4567 = vfmaq_f32(vbias4567, vout2x4567, vfilter_output_scale4567);
275 const float32x4_t vbias89AB = vld1q_f32(w); w = (const float*) w + 4;
276 vout0x89AB = vfmaq_f32(vbias89AB, vout0x89AB, vfilter_output_scale89AB);
277 vout1x89AB = vfmaq_f32(vbias89AB, vout1x89AB, vfilter_output_scale89AB);
278 vout2x89AB = vfmaq_f32(vbias89AB, vout2x89AB, vfilter_output_scale89AB);
279 const float32x4_t vbiasCDEF = vld1q_f32(w); w = (const float*) w + 4;
280 vout0xCDEF = vfmaq_f32(vbiasCDEF, vout0xCDEF, vfilter_output_scaleCDEF);
281 vout1xCDEF = vfmaq_f32(vbiasCDEF, vout1xCDEF, vfilter_output_scaleCDEF);
282 vout2xCDEF = vfmaq_f32(vbiasCDEF, vout2xCDEF, vfilter_output_scaleCDEF);
283
284 const float32x4_t voutput_min = vld1q_dup_f32(&params->scalar.min);
285 vout0x0123 = vmaxq_f32(vout0x0123, voutput_min);
286 vout0x4567 = vmaxq_f32(vout0x4567, voutput_min);
287 vout0x89AB = vmaxq_f32(vout0x89AB, voutput_min);
288 vout0xCDEF = vmaxq_f32(vout0xCDEF, voutput_min);
289 vout1x0123 = vmaxq_f32(vout1x0123, voutput_min);
290 vout1x4567 = vmaxq_f32(vout1x4567, voutput_min);
291 vout1x89AB = vmaxq_f32(vout1x89AB, voutput_min);
292 vout1xCDEF = vmaxq_f32(vout1xCDEF, voutput_min);
293 vout2x0123 = vmaxq_f32(vout2x0123, voutput_min);
294 vout2x4567 = vmaxq_f32(vout2x4567, voutput_min);
295 vout2x89AB = vmaxq_f32(vout2x89AB, voutput_min);
296 vout2xCDEF = vmaxq_f32(vout2xCDEF, voutput_min);
297
298 const float32x4_t voutput_max = vld1q_dup_f32(&params->scalar.max);
299 vout0x0123 = vminq_f32(vout0x0123, voutput_max);
300 vout0x4567 = vminq_f32(vout0x4567, voutput_max);
301 vout0x89AB = vminq_f32(vout0x89AB, voutput_max);
302 vout0xCDEF = vminq_f32(vout0xCDEF, voutput_max);
303 vout1x0123 = vminq_f32(vout1x0123, voutput_max);
304 vout1x4567 = vminq_f32(vout1x4567, voutput_max);
305 vout1x89AB = vminq_f32(vout1x89AB, voutput_max);
306 vout1xCDEF = vminq_f32(vout1xCDEF, voutput_max);
307 vout2x0123 = vminq_f32(vout2x0123, voutput_max);
308 vout2x4567 = vminq_f32(vout2x4567, voutput_max);
309 vout2x89AB = vminq_f32(vout2x89AB, voutput_max);
310 vout2xCDEF = vminq_f32(vout2xCDEF, voutput_max);
311
312 if XNN_LIKELY(nc >= 16) {
313 vst1q_f32(c2, vout2x0123);
314 vst1q_f32(c2 + 4, vout2x4567);
315 vst1q_f32(c2 + 8, vout2x89AB);
316 vst1q_f32(c2 + 12, vout2xCDEF);
317 vst1q_f32(c1, vout1x0123);
318 vst1q_f32(c1 + 4, vout1x4567);
319 vst1q_f32(c1 + 8, vout1x89AB);
320 vst1q_f32(c1 + 12, vout1xCDEF);
321 vst1q_f32(c0, vout0x0123);
322 vst1q_f32(c0 + 4, vout0x4567);
323 vst1q_f32(c0 + 8, vout0x89AB);
324 vst1q_f32(c0 + 12, vout0xCDEF);
325
326 c2 = (float*) ((uintptr_t) c2 + cn_stride);
327 c1 = (float*) ((uintptr_t) c1 + cn_stride);
328 c0 = (float*) ((uintptr_t) c0 + cn_stride);
329
330 a = (const int8_t**restrict) ((uintptr_t) a - ks);
331
332 nc -= 16;
333 } else {
334 if (nc & 8) {
335 vst1q_f32(c2, vout2x0123); c2 += 4;
336 vout2x0123 = vout2x89AB;
337 vst1q_f32(c1, vout1x0123); c1 += 4;
338 vout1x0123 = vout1x89AB;
339 vst1q_f32(c0, vout0x0123); c0 += 4;
340 vout0x0123 = vout0x89AB;
341 vst1q_f32(c2, vout2x4567); c2 += 4;
342 vout2x4567 = vout2xCDEF;
343 vst1q_f32(c1, vout1x4567); c1 += 4;
344 vout1x4567 = vout1xCDEF;
345 vst1q_f32(c0, vout0x4567); c0 += 4;
346 vout0x4567 = vout0xCDEF;
347 }
348 if (nc & 4) {
349 vst1q_f32(c2, vout2x0123); c2 += 4;
350 vout2x0123 = vout2x4567;
351 vst1q_f32(c1, vout1x0123); c1 += 4;
352 vout1x0123 = vout1x4567;
353 vst1q_f32(c0, vout0x0123); c0 += 4;
354 vout0x0123 = vout0x4567;
355 }
356 float32x2_t vout2x01 = vget_low_f32(vout2x0123);
357 float32x2_t vout1x01 = vget_low_f32(vout1x0123);
358 float32x2_t vout0x01 = vget_low_f32(vout0x0123);
359 if (nc & 2) {
360 vst1_f32(c2, vout2x01); c2 += 2;
361 vst1_f32(c1, vout1x01); c1 += 2;
362 vst1_f32(c0, vout0x01); c0 += 2;
363 vout2x01 = vget_high_f32(vout2x0123);
364 vout1x01 = vget_high_f32(vout1x0123);
365 vout0x01 = vget_high_f32(vout0x0123);
366 }
367 if (nc & 1) {
368 vst1_lane_f32(c2, vout2x01, 0);
369 vst1_lane_f32(c1, vout1x01, 0);
370 vst1_lane_f32(c0, vout0x01, 0);
371 }
372 nc = 0;
373 }
374 } while (nc != 0);
375 }