Debian Codegraph @

src/qs8-qc8w-igemm/gen/qs8-qc8w-igemm-6x16c8-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_6x16c8__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 <= 6);
35 assert(nc != 0);
36 assert(kc != 0);
37 assert(ks != 0);
38 assert(ks % (6 * 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 int8_t* c3 = (int8_t*) ((uintptr_t) c2 + cm_stride);
55 if XNN_UNPREDICTABLE(mr < 4) {
56 c3 = c2;
57 }
58 int8_t* c4 = (int8_t*) ((uintptr_t) c3 + cm_stride);
59 if XNN_UNPREDICTABLE(mr <= 4) {
60 c4 = c3;
61 }
62 int8_t* c5 = (int8_t*) ((uintptr_t) c4 + cm_stride);
63 if XNN_UNPREDICTABLE(mr != 6) {
64 c5 = c4;
65 }
66
67 do {
68 const uint64x2x2_t vbias01x0123 = vld2q_dup_u64(w); w = (const int32_t*) w + 4;
69 const uint64x2x2_t vbias01x4567 = vld2q_dup_u64(w); w = (const int32_t*) w + 4;
70 const uint64x2x2_t vbias01x89AB = vld2q_dup_u64(w); w = (const int32_t*) w + 4;
71 const uint64x2x2_t vbias01xCDEF = vld2q_dup_u64(w); w = (const int32_t*) w + 4;
72 int32x4_t vacc01x01 = vreinterpretq_s32_u64(vbias01x0123.val[0]);
73 int32x4_t vacc01x23 = vreinterpretq_s32_u64(vbias01x0123.val[1]);
74 int32x4_t vacc01x45 = vreinterpretq_s32_u64(vbias01x4567.val[0]);
75 int32x4_t vacc01x67 = vreinterpretq_s32_u64(vbias01x4567.val[1]);
76 int32x4_t vacc01x89 = vreinterpretq_s32_u64(vbias01x89AB.val[0]);
77 int32x4_t vacc01xAB = vreinterpretq_s32_u64(vbias01x89AB.val[1]);
78 int32x4_t vacc01xCD = vreinterpretq_s32_u64(vbias01xCDEF.val[0]);
79 int32x4_t vacc01xEF = vreinterpretq_s32_u64(vbias01xCDEF.val[1]);
80 int32x4_t vacc23x01 = vreinterpretq_s32_u64(vbias01x0123.val[0]);
81 int32x4_t vacc23x23 = vreinterpretq_s32_u64(vbias01x0123.val[1]);
82 int32x4_t vacc23x45 = vreinterpretq_s32_u64(vbias01x4567.val[0]);
83 int32x4_t vacc23x67 = vreinterpretq_s32_u64(vbias01x4567.val[1]);
84 int32x4_t vacc23x89 = vreinterpretq_s32_u64(vbias01x89AB.val[0]);
85 int32x4_t vacc23xAB = vreinterpretq_s32_u64(vbias01x89AB.val[1]);
86 int32x4_t vacc23xCD = vreinterpretq_s32_u64(vbias01xCDEF.val[0]);
87 int32x4_t vacc23xEF = vreinterpretq_s32_u64(vbias01xCDEF.val[1]);
88 int32x4_t vacc45x01 = vreinterpretq_s32_u64(vbias01x0123.val[0]);
89 int32x4_t vacc45x23 = vreinterpretq_s32_u64(vbias01x0123.val[1]);
90 int32x4_t vacc45x45 = vreinterpretq_s32_u64(vbias01x4567.val[0]);
91 int32x4_t vacc45x67 = vreinterpretq_s32_u64(vbias01x4567.val[1]);
92 int32x4_t vacc45x89 = vreinterpretq_s32_u64(vbias01x89AB.val[0]);
93 int32x4_t vacc45xAB = vreinterpretq_s32_u64(vbias01x89AB.val[1]);
94 int32x4_t vacc45xCD = vreinterpretq_s32_u64(vbias01xCDEF.val[0]);
95 int32x4_t vacc45xEF = vreinterpretq_s32_u64(vbias01xCDEF.val[1]);
96
97 size_t p = ks;
98 do {
99 const int8_t* restrict a0 = a[0];
100 if XNN_UNPREDICTABLE(a0 != zero) {
101 a0 = (const int8_t*) ((uintptr_t) a0 + a_offset);
102 }
103 const int8_t* restrict a1 = a[1];
104 if XNN_UNPREDICTABLE(a1 != zero) {
105 a1 = (const int8_t*) ((uintptr_t) a1 + a_offset);
106 }
107 const int8_t* restrict a2 = a[2];
108 if XNN_UNPREDICTABLE(a2 != zero) {
109 a2 = (const int8_t*) ((uintptr_t) a2 + a_offset);
110 }
111 const int8_t* restrict a3 = a[3];
112 if XNN_UNPREDICTABLE(a3 != zero) {
113 a3 = (const int8_t*) ((uintptr_t) a3 + a_offset);
114 }
115 const int8_t* restrict a4 = a[4];
116 if XNN_UNPREDICTABLE(a4 != zero) {
117 a4 = (const int8_t*) ((uintptr_t) a4 + a_offset);
118 }
119 const int8_t* restrict a5 = a[5];
120 if XNN_UNPREDICTABLE(a5 != zero) {
121 a5 = (const int8_t*) ((uintptr_t) a5 + a_offset);
122 }
123 a += 6;
124
125 uint64x2x2_t va01x0123456789ABCDEF;
126 va01x0123456789ABCDEF.val[0] = vdupq_n_u64(0);
127 va01x0123456789ABCDEF.val[1] = vdupq_n_u64(0);
128 uint64x2x2_t va23x0123456789ABCDEF;
129 va23x0123456789ABCDEF.val[0] = vdupq_n_u64(0);
130 va23x0123456789ABCDEF.val[1] = vdupq_n_u64(0);
131 uint64x2x2_t va45x0123456789ABCDEF;
132 va45x0123456789ABCDEF.val[0] = vdupq_n_u64(0);
133 va45x0123456789ABCDEF.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 6x16 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 va45x0123456789ABCDEF = vld2q_lane_u64((const void*) a4, va45x0123456789ABCDEF, 0); a4 += 16;
143 va01x0123456789ABCDEF = vld2q_lane_u64((const void*) a1, va01x0123456789ABCDEF, 1); a1 += 16;
144 va23x0123456789ABCDEF = vld2q_lane_u64((const void*) a3, va23x0123456789ABCDEF, 1); a3 += 16;
145 va45x0123456789ABCDEF = vld2q_lane_u64((const void*) a5, va45x0123456789ABCDEF, 1); a5 += 16;
146
147 // Load a 16x16 block of weights.
148 const int8x16_t vb01x01234567 = vld1q_s8(w); w = (const int8_t*) w + 16;
149 const int8x16_t vb23x01234567 = vld1q_s8(w); w = (const int8_t*) w + 16;
150 const int8x16_t vb45x01234567 = vld1q_s8(w); w = (const int8_t*) w + 16;
151 const int8x16_t vb67x01234567 = vld1q_s8(w); w = (const int8_t*) w + 16;
152 const int8x16_t vb89x01234567 = vld1q_s8(w); w = (const int8_t*) w + 16;
153 const int8x16_t vbABx01234567 = vld1q_s8(w); w = (const int8_t*) w + 16;
154 const int8x16_t vbCDx01234567 = vld1q_s8(w); w = (const int8_t*) w + 16;
155 const int8x16_t vbEFx01234567 = vld1q_s8(w); w = (const int8_t*) w + 16;
156 const int8x16_t vb01x89ABCDEF = vld1q_s8(w); w = (const int8_t*) w + 16;
157 const int8x16_t vb23x89ABCDEF = vld1q_s8(w); w = (const int8_t*) w + 16;
158 const int8x16_t vb45x89ABCDEF = vld1q_s8(w); w = (const int8_t*) w + 16;
159 const int8x16_t vb67x89ABCDEF = vld1q_s8(w); w = (const int8_t*) w + 16;
160 const int8x16_t vb89x89ABCDEF = vld1q_s8(w); w = (const int8_t*) w + 16;
161 const int8x16_t vbABx89ABCDEF = vld1q_s8(w); w = (const int8_t*) w + 16;
162 const int8x16_t vbCDx89ABCDEF = vld1q_s8(w); w = (const int8_t*) w + 16;
163 const int8x16_t vbEFx89ABCDEF = vld1q_s8(w); w = (const int8_t*) w + 16;
164
165 // Multiply-accumulate: 6x16 * 16x16 --> 6x16.
166 vacc01x01 = vmmlaq_s32(vacc01x01, vreinterpretq_s8_u64(va01x0123456789ABCDEF.val[0]), vb01x01234567);
167 vacc01x23 = vmmlaq_s32(vacc01x23, vreinterpretq_s8_u64(va01x0123456789ABCDEF.val[0]), vb23x01234567);
168 vacc01x45 = vmmlaq_s32(vacc01x45, vreinterpretq_s8_u64(va01x0123456789ABCDEF.val[0]), vb45x01234567);
169 vacc01x67 = vmmlaq_s32(vacc01x67, vreinterpretq_s8_u64(va01x0123456789ABCDEF.val[0]), vb67x01234567);
170 vacc01x89 = vmmlaq_s32(vacc01x89, vreinterpretq_s8_u64(va01x0123456789ABCDEF.val[0]), vb89x01234567);
171 vacc01xAB = vmmlaq_s32(vacc01xAB, vreinterpretq_s8_u64(va01x0123456789ABCDEF.val[0]), vbABx01234567);
172 vacc01xCD = vmmlaq_s32(vacc01xCD, vreinterpretq_s8_u64(va01x0123456789ABCDEF.val[0]), vbCDx01234567);
173 vacc01xEF = vmmlaq_s32(vacc01xEF, vreinterpretq_s8_u64(va01x0123456789ABCDEF.val[0]), vbEFx01234567);
174 vacc23x01 = vmmlaq_s32(vacc23x01, vreinterpretq_s8_u64(va23x0123456789ABCDEF.val[0]), vb01x01234567);
175 vacc23x23 = vmmlaq_s32(vacc23x23, vreinterpretq_s8_u64(va23x0123456789ABCDEF.val[0]), vb23x01234567);
176 vacc23x45 = vmmlaq_s32(vacc23x45, vreinterpretq_s8_u64(va23x0123456789ABCDEF.val[0]), vb45x01234567);
177 vacc23x67 = vmmlaq_s32(vacc23x67, vreinterpretq_s8_u64(va23x0123456789ABCDEF.val[0]), vb67x01234567);
178 vacc23x89 = vmmlaq_s32(vacc23x89, vreinterpretq_s8_u64(va23x0123456789ABCDEF.val[0]), vb89x01234567);
179 vacc23xAB = vmmlaq_s32(vacc23xAB, vreinterpretq_s8_u64(va23x0123456789ABCDEF.val[0]), vbABx01234567);
180 vacc23xCD = vmmlaq_s32(vacc23xCD, vreinterpretq_s8_u64(va23x0123456789ABCDEF.val[0]), vbCDx01234567);
181 vacc23xEF = vmmlaq_s32(vacc23xEF, vreinterpretq_s8_u64(va23x0123456789ABCDEF.val[0]), vbEFx01234567);
182 vacc45x01 = vmmlaq_s32(vacc45x01, vreinterpretq_s8_u64(va45x0123456789ABCDEF.val[0]), vb01x01234567);
183 vacc45x23 = vmmlaq_s32(vacc45x23, vreinterpretq_s8_u64(va45x0123456789ABCDEF.val[0]), vb23x01234567);
184 vacc45x45 = vmmlaq_s32(vacc45x45, vreinterpretq_s8_u64(va45x0123456789ABCDEF.val[0]), vb45x01234567);
185 vacc45x67 = vmmlaq_s32(vacc45x67, vreinterpretq_s8_u64(va45x0123456789ABCDEF.val[0]), vb67x01234567);
186 vacc45x89 = vmmlaq_s32(vacc45x89, vreinterpretq_s8_u64(va45x0123456789ABCDEF.val[0]), vb89x01234567);
187 vacc45xAB = vmmlaq_s32(vacc45xAB, vreinterpretq_s8_u64(va45x0123456789ABCDEF.val[0]), vbABx01234567);
188 vacc45xCD = vmmlaq_s32(vacc45xCD, vreinterpretq_s8_u64(va45x0123456789ABCDEF.val[0]), vbCDx01234567);
189 vacc45xEF = vmmlaq_s32(vacc45xEF, vreinterpretq_s8_u64(va45x0123456789ABCDEF.val[0]), vbEFx01234567);
190 vacc01x01 = vmmlaq_s32(vacc01x01, vreinterpretq_s8_u64(va01x0123456789ABCDEF.val[1]), vb01x89ABCDEF);
191 vacc01x23 = vmmlaq_s32(vacc01x23, vreinterpretq_s8_u64(va01x0123456789ABCDEF.val[1]), vb23x89ABCDEF);
192 vacc01x45 = vmmlaq_s32(vacc01x45, vreinterpretq_s8_u64(va01x0123456789ABCDEF.val[1]), vb45x89ABCDEF);
193 vacc01x67 = vmmlaq_s32(vacc01x67, vreinterpretq_s8_u64(va01x0123456789ABCDEF.val[1]), vb67x89ABCDEF);
194 vacc01x89 = vmmlaq_s32(vacc01x89, vreinterpretq_s8_u64(va01x0123456789ABCDEF.val[1]), vb89x89ABCDEF);
195 vacc01xAB = vmmlaq_s32(vacc01xAB, vreinterpretq_s8_u64(va01x0123456789ABCDEF.val[1]), vbABx89ABCDEF);
196 vacc01xCD = vmmlaq_s32(vacc01xCD, vreinterpretq_s8_u64(va01x0123456789ABCDEF.val[1]), vbCDx89ABCDEF);
197 vacc01xEF = vmmlaq_s32(vacc01xEF, vreinterpretq_s8_u64(va01x0123456789ABCDEF.val[1]), vbEFx89ABCDEF);
198 vacc23x01 = vmmlaq_s32(vacc23x01, vreinterpretq_s8_u64(va23x0123456789ABCDEF.val[1]), vb01x89ABCDEF);
199 vacc23x23 = vmmlaq_s32(vacc23x23, vreinterpretq_s8_u64(va23x0123456789ABCDEF.val[1]), vb23x89ABCDEF);
200 vacc23x45 = vmmlaq_s32(vacc23x45, vreinterpretq_s8_u64(va23x0123456789ABCDEF.val[1]), vb45x89ABCDEF);
201 vacc23x67 = vmmlaq_s32(vacc23x67, vreinterpretq_s8_u64(va23x0123456789ABCDEF.val[1]), vb67x89ABCDEF);
202 vacc23x89 = vmmlaq_s32(vacc23x89, vreinterpretq_s8_u64(va23x0123456789ABCDEF.val[1]), vb89x89ABCDEF);
203 vacc23xAB = vmmlaq_s32(vacc23xAB, vreinterpretq_s8_u64(va23x0123456789ABCDEF.val[1]), vbABx89ABCDEF);
204 vacc23xCD = vmmlaq_s32(vacc23xCD, vreinterpretq_s8_u64(va23x0123456789ABCDEF.val[1]), vbCDx89ABCDEF);
205 vacc23xEF = vmmlaq_s32(vacc23xEF, vreinterpretq_s8_u64(va23x0123456789ABCDEF.val[1]), vbEFx89ABCDEF);
206 vacc45x01 = vmmlaq_s32(vacc45x01, vreinterpretq_s8_u64(va45x0123456789ABCDEF.val[1]), vb01x89ABCDEF);
207 vacc45x23 = vmmlaq_s32(vacc45x23, vreinterpretq_s8_u64(va45x0123456789ABCDEF.val[1]), vb23x89ABCDEF);
208 vacc45x45 = vmmlaq_s32(vacc45x45, vreinterpretq_s8_u64(va45x0123456789ABCDEF.val[1]), vb45x89ABCDEF);
209 vacc45x67 = vmmlaq_s32(vacc45x67, vreinterpretq_s8_u64(va45x0123456789ABCDEF.val[1]), vb67x89ABCDEF);
210 vacc45x89 = vmmlaq_s32(vacc45x89, vreinterpretq_s8_u64(va45x0123456789ABCDEF.val[1]), vb89x89ABCDEF);
211 vacc45xAB = vmmlaq_s32(vacc45xAB, vreinterpretq_s8_u64(va45x0123456789ABCDEF.val[1]), vbABx89ABCDEF);
212 vacc45xCD = vmmlaq_s32(vacc45xCD, vreinterpretq_s8_u64(va45x0123456789ABCDEF.val[1]), vbCDx89ABCDEF);
213 vacc45xEF = vmmlaq_s32(vacc45xEF, vreinterpretq_s8_u64(va45x0123456789ABCDEF.val[1]), vbEFx89ABCDEF);
214
215 k -= 16 * sizeof(int8_t);
216 }
217 // Handle up to 8 final positions of `k`
218 if XNN_UNLIKELY(k != 0) {
219 // Load a 6x8 block of activations.
220 uint64x2_t va01x01234567 = vld1q_dup_u64((const void*) a0); a0 += 8;
221 uint64x2_t va23x01234567 = vld1q_dup_u64((const void*) a2); a2 += 8;
222 uint64x2_t va45x01234567 = vld1q_dup_u64((const void*) a4); a4 += 8;
223 va01x01234567 = vld1q_lane_u64((const void*) a1, va01x01234567, 1); a1 += 8;
224 va23x01234567 = vld1q_lane_u64((const void*) a3, va23x01234567, 1); a3 += 8;
225 va45x01234567 = vld1q_lane_u64((const void*) a5, va45x01234567, 1); a5 += 8;
226
227 // Load a 16x16 block of weights.
228 const int8x16_t vb01x01234567 = vld1q_s8(w); w = (const int8_t*) w + 16;
229 const int8x16_t vb23x01234567 = vld1q_s8(w); w = (const int8_t*) w + 16;
230 const int8x16_t vb45x01234567 = vld1q_s8(w); w = (const int8_t*) w + 16;
231 const int8x16_t vb67x01234567 = vld1q_s8(w); w = (const int8_t*) w + 16;
232 const int8x16_t vb89x01234567 = vld1q_s8(w); w = (const int8_t*) w + 16;
233 const int8x16_t vbABx01234567 = vld1q_s8(w); w = (const int8_t*) w + 16;
234 const int8x16_t vbCDx01234567 = vld1q_s8(w); w = (const int8_t*) w + 16;
235 const int8x16_t vbEFx01234567 = vld1q_s8(w); w = (const int8_t*) w + 16;
236
237 // Multiply-accumulate: 6x4 * 4x16 --> 6x16.
238 vacc01x01 = vmmlaq_s32(vacc01x01, vreinterpretq_s8_u64(va01x01234567), vb01x01234567);
239 vacc01x23 = vmmlaq_s32(vacc01x23, vreinterpretq_s8_u64(va01x01234567), vb23x01234567);
240 vacc01x45 = vmmlaq_s32(vacc01x45, vreinterpretq_s8_u64(va01x01234567), vb45x01234567);
241 vacc01x67 = vmmlaq_s32(vacc01x67, vreinterpretq_s8_u64(va01x01234567), vb67x01234567);
242 vacc01x89 = vmmlaq_s32(vacc01x89, vreinterpretq_s8_u64(va01x01234567), vb89x01234567);
243 vacc01xAB = vmmlaq_s32(vacc01xAB, vreinterpretq_s8_u64(va01x01234567), vbABx01234567);
244 vacc01xCD = vmmlaq_s32(vacc01xCD, vreinterpretq_s8_u64(va01x01234567), vbCDx01234567);
245 vacc01xEF = vmmlaq_s32(vacc01xEF, vreinterpretq_s8_u64(va01x01234567), vbEFx01234567);
246 vacc23x01 = vmmlaq_s32(vacc23x01, vreinterpretq_s8_u64(va23x01234567), vb01x01234567);
247 vacc23x23 = vmmlaq_s32(vacc23x23, vreinterpretq_s8_u64(va23x01234567), vb23x01234567);
248 vacc23x45 = vmmlaq_s32(vacc23x45, vreinterpretq_s8_u64(va23x01234567), vb45x01234567);
249 vacc23x67 = vmmlaq_s32(vacc23x67, vreinterpretq_s8_u64(va23x01234567), vb67x01234567);
250 vacc23x89 = vmmlaq_s32(vacc23x89, vreinterpretq_s8_u64(va23x01234567), vb89x01234567);
251 vacc23xAB = vmmlaq_s32(vacc23xAB, vreinterpretq_s8_u64(va23x01234567), vbABx01234567);
252 vacc23xCD = vmmlaq_s32(vacc23xCD, vreinterpretq_s8_u64(va23x01234567), vbCDx01234567);
253 vacc23xEF = vmmlaq_s32(vacc23xEF, vreinterpretq_s8_u64(va23x01234567), vbEFx01234567);
254 vacc45x01 = vmmlaq_s32(vacc45x01, vreinterpretq_s8_u64(va45x01234567), vb01x01234567);
255 vacc45x23 = vmmlaq_s32(vacc45x23, vreinterpretq_s8_u64(va45x01234567), vb23x01234567);
256 vacc45x45 = vmmlaq_s32(vacc45x45, vreinterpretq_s8_u64(va45x01234567), vb45x01234567);
257 vacc45x67 = vmmlaq_s32(vacc45x67, vreinterpretq_s8_u64(va45x01234567), vb67x01234567);
258 vacc45x89 = vmmlaq_s32(vacc45x89, vreinterpretq_s8_u64(va45x01234567), vb89x01234567);
259 vacc45xAB = vmmlaq_s32(vacc45xAB, vreinterpretq_s8_u64(va45x01234567), vbABx01234567);
260 vacc45xCD = vmmlaq_s32(vacc45xCD, vreinterpretq_s8_u64(va45x01234567), vbCDx01234567);
261 vacc45xEF = vmmlaq_s32(vacc45xEF, vreinterpretq_s8_u64(va45x01234567), vbEFx01234567);
262 }
263
264 p -= 6 * sizeof(void*);
265 } while (p != 0);
266
267 int32x4_t vacc0x0123 = vcombine_s32(vget_low_s32(vacc01x01), vget_low_s32(vacc01x23));
268 int32x4_t vacc1x0123 = vcombine_s32(vget_high_s32(vacc01x01), vget_high_s32(vacc01x23));
269 int32x4_t vacc0x4567 = vcombine_s32(vget_low_s32(vacc01x45), vget_low_s32(vacc01x67));
270 int32x4_t vacc1x4567 = vcombine_s32(vget_high_s32(vacc01x45), vget_high_s32(vacc01x67));
271 int32x4_t vacc0x89AB = vcombine_s32(vget_low_s32(vacc01x89), vget_low_s32(vacc01xAB));
272 int32x4_t vacc1x89AB = vcombine_s32(vget_high_s32(vacc01x89), vget_high_s32(vacc01xAB));
273 int32x4_t vacc0xCDEF = vcombine_s32(vget_low_s32(vacc01xCD), vget_low_s32(vacc01xEF));
274 int32x4_t vacc1xCDEF = vcombine_s32(vget_high_s32(vacc01xCD), vget_high_s32(vacc01xEF));
275 int32x4_t vacc2x0123 = vcombine_s32(vget_low_s32(vacc23x01), vget_low_s32(vacc23x23));
276 int32x4_t vacc3x0123 = vcombine_s32(vget_high_s32(vacc23x01), vget_high_s32(vacc23x23));
277 int32x4_t vacc2x4567 = vcombine_s32(vget_low_s32(vacc23x45), vget_low_s32(vacc23x67));
278 int32x4_t vacc3x4567 = vcombine_s32(vget_high_s32(vacc23x45), vget_high_s32(vacc23x67));
279 int32x4_t vacc2x89AB = vcombine_s32(vget_low_s32(vacc23x89), vget_low_s32(vacc23xAB));
280 int32x4_t vacc3x89AB = vcombine_s32(vget_high_s32(vacc23x89), vget_high_s32(vacc23xAB));
281 int32x4_t vacc2xCDEF = vcombine_s32(vget_low_s32(vacc23xCD), vget_low_s32(vacc23xEF));
282 int32x4_t vacc3xCDEF = vcombine_s32(vget_high_s32(vacc23xCD), vget_high_s32(vacc23xEF));
283 int32x4_t vacc4x0123 = vcombine_s32(vget_low_s32(vacc45x01), vget_low_s32(vacc45x23));
284 int32x4_t vacc5x0123 = vcombine_s32(vget_high_s32(vacc45x01), vget_high_s32(vacc45x23));
285 int32x4_t vacc4x4567 = vcombine_s32(vget_low_s32(vacc45x45), vget_low_s32(vacc45x67));
286 int32x4_t vacc5x4567 = vcombine_s32(vget_high_s32(vacc45x45), vget_high_s32(vacc45x67));
287 int32x4_t vacc4x89AB = vcombine_s32(vget_low_s32(vacc45x89), vget_low_s32(vacc45xAB));
288 int32x4_t vacc5x89AB = vcombine_s32(vget_high_s32(vacc45x89), vget_high_s32(vacc45xAB));
289 int32x4_t vacc4xCDEF = vcombine_s32(vget_low_s32(vacc45xCD), vget_low_s32(vacc45xEF));
290 int32x4_t vacc5xCDEF = vcombine_s32(vget_high_s32(vacc45xCD), vget_high_s32(vacc45xEF));
291
292 float32x4_t vfpacc0x0123 = vcvtq_f32_s32(vacc0x0123);
293 float32x4_t vfpacc0x4567 = vcvtq_f32_s32(vacc0x4567);
294 float32x4_t vfpacc0x89AB = vcvtq_f32_s32(vacc0x89AB);
295 float32x4_t vfpacc0xCDEF = vcvtq_f32_s32(vacc0xCDEF);
296 float32x4_t vfpacc1x0123 = vcvtq_f32_s32(vacc1x0123);
297 float32x4_t vfpacc1x4567 = vcvtq_f32_s32(vacc1x4567);
298 float32x4_t vfpacc1x89AB = vcvtq_f32_s32(vacc1x89AB);
299 float32x4_t vfpacc1xCDEF = vcvtq_f32_s32(vacc1xCDEF);
300 float32x4_t vfpacc2x0123 = vcvtq_f32_s32(vacc2x0123);
301 float32x4_t vfpacc2x4567 = vcvtq_f32_s32(vacc2x4567);
302 float32x4_t vfpacc2x89AB = vcvtq_f32_s32(vacc2x89AB);
303 float32x4_t vfpacc2xCDEF = vcvtq_f32_s32(vacc2xCDEF);
304 float32x4_t vfpacc3x0123 = vcvtq_f32_s32(vacc3x0123);
305 float32x4_t vfpacc3x4567 = vcvtq_f32_s32(vacc3x4567);
306 float32x4_t vfpacc3x89AB = vcvtq_f32_s32(vacc3x89AB);
307 float32x4_t vfpacc3xCDEF = vcvtq_f32_s32(vacc3xCDEF);
308 float32x4_t vfpacc4x0123 = vcvtq_f32_s32(vacc4x0123);
309 float32x4_t vfpacc4x4567 = vcvtq_f32_s32(vacc4x4567);
310 float32x4_t vfpacc4x89AB = vcvtq_f32_s32(vacc4x89AB);
311 float32x4_t vfpacc4xCDEF = vcvtq_f32_s32(vacc4xCDEF);
312 float32x4_t vfpacc5x0123 = vcvtq_f32_s32(vacc5x0123);
313 float32x4_t vfpacc5x4567 = vcvtq_f32_s32(vacc5x4567);
314 float32x4_t vfpacc5x89AB = vcvtq_f32_s32(vacc5x89AB);
315 float32x4_t vfpacc5xCDEF = vcvtq_f32_s32(vacc5xCDEF);
316
317 const float32x4_t vscale0123 = vld1q_f32((const float*) w); w = (const float*) w + 4;
318 vfpacc0x0123 = vmulq_f32(vfpacc0x0123, vscale0123);
319 vfpacc1x0123 = vmulq_f32(vfpacc1x0123, vscale0123);
320 vfpacc2x0123 = vmulq_f32(vfpacc2x0123, vscale0123);
321 vfpacc3x0123 = vmulq_f32(vfpacc3x0123, vscale0123);
322 vfpacc4x0123 = vmulq_f32(vfpacc4x0123, vscale0123);
323 vfpacc5x0123 = vmulq_f32(vfpacc5x0123, vscale0123);
324 const float32x4_t vscale4567 = vld1q_f32((const float*) w); w = (const float*) w + 4;
325 vfpacc0x4567 = vmulq_f32(vfpacc0x4567, vscale4567);
326 vfpacc1x4567 = vmulq_f32(vfpacc1x4567, vscale4567);
327 vfpacc2x4567 = vmulq_f32(vfpacc2x4567, vscale4567);
328 vfpacc3x4567 = vmulq_f32(vfpacc3x4567, vscale4567);
329 vfpacc4x4567 = vmulq_f32(vfpacc4x4567, vscale4567);
330 vfpacc5x4567 = vmulq_f32(vfpacc5x4567, vscale4567);
331 const float32x4_t vscale89AB = vld1q_f32((const float*) w); w = (const float*) w + 4;
332 vfpacc0x89AB = vmulq_f32(vfpacc0x89AB, vscale89AB);
333 vfpacc1x89AB = vmulq_f32(vfpacc1x89AB, vscale89AB);
334 vfpacc2x89AB = vmulq_f32(vfpacc2x89AB, vscale89AB);
335 vfpacc3x89AB = vmulq_f32(vfpacc3x89AB, vscale89AB);
336 vfpacc4x89AB = vmulq_f32(vfpacc4x89AB, vscale89AB);
337 vfpacc5x89AB = vmulq_f32(vfpacc5x89AB, vscale89AB);
338 const float32x4_t vscaleCDEF = vld1q_f32((const float*) w); w = (const float*) w + 4;
339 vfpacc0xCDEF = vmulq_f32(vfpacc0xCDEF, vscaleCDEF);
340 vfpacc1xCDEF = vmulq_f32(vfpacc1xCDEF, vscaleCDEF);
341 vfpacc2xCDEF = vmulq_f32(vfpacc2xCDEF, vscaleCDEF);
342 vfpacc3xCDEF = vmulq_f32(vfpacc3xCDEF, vscaleCDEF);
343 vfpacc4xCDEF = vmulq_f32(vfpacc4xCDEF, vscaleCDEF);
344 vfpacc5xCDEF = vmulq_f32(vfpacc5xCDEF, vscaleCDEF);
345
346 vacc0x0123 = vcvtnq_s32_f32(vfpacc0x0123);
347 vacc0x4567 = vcvtnq_s32_f32(vfpacc0x4567);
348 vacc0x89AB = vcvtnq_s32_f32(vfpacc0x89AB);
349 vacc0xCDEF = vcvtnq_s32_f32(vfpacc0xCDEF);
350 vacc1x0123 = vcvtnq_s32_f32(vfpacc1x0123);
351 vacc1x4567 = vcvtnq_s32_f32(vfpacc1x4567);
352 vacc1x89AB = vcvtnq_s32_f32(vfpacc1x89AB);
353 vacc1xCDEF = vcvtnq_s32_f32(vfpacc1xCDEF);
354 vacc2x0123 = vcvtnq_s32_f32(vfpacc2x0123);
355 vacc2x4567 = vcvtnq_s32_f32(vfpacc2x4567);
356 vacc2x89AB = vcvtnq_s32_f32(vfpacc2x89AB);
357 vacc2xCDEF = vcvtnq_s32_f32(vfpacc2xCDEF);
358 vacc3x0123 = vcvtnq_s32_f32(vfpacc3x0123);
359 vacc3x4567 = vcvtnq_s32_f32(vfpacc3x4567);
360 vacc3x89AB = vcvtnq_s32_f32(vfpacc3x89AB);
361 vacc3xCDEF = vcvtnq_s32_f32(vfpacc3xCDEF);
362 vacc4x0123 = vcvtnq_s32_f32(vfpacc4x0123);
363 vacc4x4567 = vcvtnq_s32_f32(vfpacc4x4567);
364 vacc4x89AB = vcvtnq_s32_f32(vfpacc4x89AB);
365 vacc4xCDEF = vcvtnq_s32_f32(vfpacc4xCDEF);
366 vacc5x0123 = vcvtnq_s32_f32(vfpacc5x0123);
367 vacc5x4567 = vcvtnq_s32_f32(vfpacc5x4567);
368 vacc5x89AB = vcvtnq_s32_f32(vfpacc5x89AB);
369 vacc5xCDEF = vcvtnq_s32_f32(vfpacc5xCDEF);
370
371 const int16x8_t voutput_zero_point = vld1q_dup_s16(&params->fp32_neonv8.output_zero_point);
372 #if XNN_ARCH_ARM64
373 const int16x8_t vacc0x01234567 = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc0x0123), vacc0x4567), voutput_zero_point);
374 const int16x8_t vacc0x89ABCDEF = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc0x89AB), vacc0xCDEF), voutput_zero_point);
375 const int16x8_t vacc1x01234567 = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc1x0123), vacc1x4567), voutput_zero_point);
376 const int16x8_t vacc1x89ABCDEF = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc1x89AB), vacc1xCDEF), voutput_zero_point);
377 const int16x8_t vacc2x01234567 = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc2x0123), vacc2x4567), voutput_zero_point);
378 const int16x8_t vacc2x89ABCDEF = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc2x89AB), vacc2xCDEF), voutput_zero_point);
379 const int16x8_t vacc3x01234567 = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc3x0123), vacc3x4567), voutput_zero_point);
380 const int16x8_t vacc3x89ABCDEF = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc3x89AB), vacc3xCDEF), voutput_zero_point);
381 const int16x8_t vacc4x01234567 = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc4x0123), vacc4x4567), voutput_zero_point);
382 const int16x8_t vacc4x89ABCDEF = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc4x89AB), vacc4xCDEF), voutput_zero_point);
383 const int16x8_t vacc5x01234567 = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc5x0123), vacc5x4567), voutput_zero_point);
384 const int16x8_t vacc5x89ABCDEF = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc5x89AB), vacc5xCDEF), voutput_zero_point);
385
386 int8x16_t vout0x0123456789ABCDEF = vqmovn_high_s16(vqmovn_s16(vacc0x01234567), vacc0x89ABCDEF);
387 int8x16_t vout1x0123456789ABCDEF = vqmovn_high_s16(vqmovn_s16(vacc1x01234567), vacc1x89ABCDEF);
388 int8x16_t vout2x0123456789ABCDEF = vqmovn_high_s16(vqmovn_s16(vacc2x01234567), vacc2x89ABCDEF);
389 int8x16_t vout3x0123456789ABCDEF = vqmovn_high_s16(vqmovn_s16(vacc3x01234567), vacc3x89ABCDEF);
390 int8x16_t vout4x0123456789ABCDEF = vqmovn_high_s16(vqmovn_s16(vacc4x01234567), vacc4x89ABCDEF);
391 int8x16_t vout5x0123456789ABCDEF = vqmovn_high_s16(vqmovn_s16(vacc5x01234567), vacc5x89ABCDEF);
392 #else
393 const int16x8_t vacc0x01234567 = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc0x0123), vqmovn_s32(vacc0x4567)), voutput_zero_point);
394 const int16x8_t vacc0x89ABCDEF = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc0x89AB), vqmovn_s32(vacc0xCDEF)), voutput_zero_point);
395 const int16x8_t vacc1x01234567 = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc1x0123), vqmovn_s32(vacc1x4567)), voutput_zero_point);
396 const int16x8_t vacc1x89ABCDEF = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc1x89AB), vqmovn_s32(vacc1xCDEF)), voutput_zero_point);
397 const int16x8_t vacc2x01234567 = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc2x0123), vqmovn_s32(vacc2x4567)), voutput_zero_point);
398 const int16x8_t vacc2x89ABCDEF = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc2x89AB), vqmovn_s32(vacc2xCDEF)), voutput_zero_point);
399 const int16x8_t vacc3x01234567 = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc3x0123), vqmovn_s32(vacc3x4567)), voutput_zero_point);
400 const int16x8_t vacc3x89ABCDEF = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc3x89AB), vqmovn_s32(vacc3xCDEF)), voutput_zero_point);
401 const int16x8_t vacc4x01234567 = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc4x0123), vqmovn_s32(vacc4x4567)), voutput_zero_point);
402 const int16x8_t vacc4x89ABCDEF = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc4x89AB), vqmovn_s32(vacc4xCDEF)), voutput_zero_point);
403 const int16x8_t vacc5x01234567 = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc5x0123), vqmovn_s32(vacc5x4567)), voutput_zero_point);
404 const int16x8_t vacc5x89ABCDEF = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc5x89AB), vqmovn_s32(vacc5xCDEF)), voutput_zero_point);
405
406 int8x16_t vout0x0123456789ABCDEF = vcombine_s8(vqmovn_s16(vacc0x01234567), vqmovn_s16(vacc0x89ABCDEF));
407 int8x16_t vout1x0123456789ABCDEF = vcombine_s8(vqmovn_s16(vacc1x01234567), vqmovn_s16(vacc1x89ABCDEF));
408 int8x16_t vout2x0123456789ABCDEF = vcombine_s8(vqmovn_s16(vacc2x01234567), vqmovn_s16(vacc2x89ABCDEF));
409 int8x16_t vout3x0123456789ABCDEF = vcombine_s8(vqmovn_s16(vacc3x01234567), vqmovn_s16(vacc3x89ABCDEF));
410 int8x16_t vout4x0123456789ABCDEF = vcombine_s8(vqmovn_s16(vacc4x01234567), vqmovn_s16(vacc4x89ABCDEF));
411 int8x16_t vout5x0123456789ABCDEF = vcombine_s8(vqmovn_s16(vacc5x01234567), vqmovn_s16(vacc5x89ABCDEF));
412 #endif
413 const int8x16_t voutput_min = vld1q_dup_s8(&params->fp32_neonv8.output_min);
414 const int8x16_t voutput_max = vld1q_dup_s8(&params->fp32_neonv8.output_max);
415
416 vout5x0123456789ABCDEF = vmaxq_s8(vout5x0123456789ABCDEF, voutput_min);
417 vout4x0123456789ABCDEF = vmaxq_s8(vout4x0123456789ABCDEF, voutput_min);
418 vout3x0123456789ABCDEF = vmaxq_s8(vout3x0123456789ABCDEF, voutput_min);
419 vout2x0123456789ABCDEF = vmaxq_s8(vout2x0123456789ABCDEF, voutput_min);
420 vout1x0123456789ABCDEF = vmaxq_s8(vout1x0123456789ABCDEF, voutput_min);
421 vout0x0123456789ABCDEF = vmaxq_s8(vout0x0123456789ABCDEF, voutput_min);
422
423 vout5x0123456789ABCDEF = vminq_s8(vout5x0123456789ABCDEF, voutput_max);
424 vout4x0123456789ABCDEF = vminq_s8(vout4x0123456789ABCDEF, voutput_max);
425 vout3x0123456789ABCDEF = vminq_s8(vout3x0123456789ABCDEF, voutput_max);
426 vout2x0123456789ABCDEF = vminq_s8(vout2x0123456789ABCDEF, voutput_max);
427 vout1x0123456789ABCDEF = vminq_s8(vout1x0123456789ABCDEF, voutput_max);
428 vout0x0123456789ABCDEF = vminq_s8(vout0x0123456789ABCDEF, voutput_max);
429
430 if (nc >= 16) {
431 vst1q_s8(c5, vout5x0123456789ABCDEF);
432 vst1q_s8(c4, vout4x0123456789ABCDEF);
433 vst1q_s8(c3, vout3x0123456789ABCDEF);
434 vst1q_s8(c2, vout2x0123456789ABCDEF);
435 vst1q_s8(c1, vout1x0123456789ABCDEF);
436 vst1q_s8(c0, vout0x0123456789ABCDEF);
437
438 c5 = (int8_t*) ((uintptr_t) c5 + cn_stride);
439 c4 = (int8_t*) ((uintptr_t) c4 + cn_stride);
440 c3 = (int8_t*) ((uintptr_t) c3 + cn_stride);
441 c2 = (int8_t*) ((uintptr_t) c2 + cn_stride);
442 c1 = (int8_t*) ((uintptr_t) c1 + cn_stride);
443 c0 = (int8_t*) ((uintptr_t) c0 + cn_stride);
444
445 a = (const int8_t**restrict) ((uintptr_t) a - ks);
446
447 nc -= 16;
448 } else {
449 int8x16_t vout4x01234567_5x01234567 = vcombine_s8(vget_low_s8(vout4x0123456789ABCDEF), vget_low_s8(vout5x0123456789ABCDEF));
450 int8x16_t vout2x01234567_3x01234567 = vcombine_s8(vget_low_s8(vout2x0123456789ABCDEF), vget_low_s8(vout3x0123456789ABCDEF));
451 int8x16_t vout0x01234567_1x01234567 = vcombine_s8(vget_low_s8(vout0x0123456789ABCDEF), vget_low_s8(vout1x0123456789ABCDEF));
452 if (nc & 8) {
453 vst1_s8(c5, vget_high_s8(vout4x01234567_5x01234567)); c5 += 8;
454 vst1_s8(c4, vget_low_s8(vout4x01234567_5x01234567)); c4 += 8;
455 vst1_s8(c3, vget_high_s8(vout2x01234567_3x01234567)); c3 += 8;
456 vst1_s8(c2, vget_low_s8(vout2x01234567_3x01234567)); c2 += 8;
457 vst1_s8(c1, vget_high_s8(vout0x01234567_1x01234567)); c1 += 8;
458 vst1_s8(c0, vget_low_s8(vout0x01234567_1x01234567)); c0 += 8;
459 vout4x01234567_5x01234567 = vcombine_s8(vget_high_s8(vout4x0123456789ABCDEF), vget_high_s8(vout5x0123456789ABCDEF));
460 vout2x01234567_3x01234567 = vcombine_s8(vget_high_s8(vout2x0123456789ABCDEF), vget_high_s8(vout3x0123456789ABCDEF));
461 vout0x01234567_1x01234567 = vcombine_s8(vget_high_s8(vout0x0123456789ABCDEF), vget_high_s8(vout1x0123456789ABCDEF));
462 }
463 if (nc & 4) {
464 vst1q_lane_u32((void*) c5, vreinterpretq_u32_s8(vout4x01234567_5x01234567), 2); c5 += 4;
465 vst1q_lane_u32((void*) c4, vreinterpretq_u32_s8(vout4x01234567_5x01234567), 0); c4 += 4;
466 vst1q_lane_u32((void*) c3, vreinterpretq_u32_s8(vout2x01234567_3x01234567), 2); c3 += 4;
467 vst1q_lane_u32((void*) c2, vreinterpretq_u32_s8(vout2x01234567_3x01234567), 0); c2 += 4;
468 vst1q_lane_u32((void*) c1, vreinterpretq_u32_s8(vout0x01234567_1x01234567), 2); c1 += 4;
469 vst1q_lane_u32((void*) c0, vreinterpretq_u32_s8(vout0x01234567_1x01234567), 0); c0 += 4;
470 vout4x01234567_5x01234567 = vextq_s8(vout4x01234567_5x01234567, vout4x01234567_5x01234567, 4);
471 vout2x01234567_3x01234567 = vextq_s8(vout2x01234567_3x01234567, vout2x01234567_3x01234567, 4);
472 vout0x01234567_1x01234567 = vextq_s8(vout0x01234567_1x01234567, vout0x01234567_1x01234567, 4);
473 }
474 if (nc & 2) {
475 vst1q_lane_u16((void*) c5, vreinterpretq_u16_s8(vout4x01234567_5x01234567), 4); c5 += 2;
476 vst1q_lane_u16((void*) c4, vreinterpretq_u16_s8(vout4x01234567_5x01234567), 0); c4 += 2;
477 vst1q_lane_u16((void*) c3, vreinterpretq_u16_s8(vout2x01234567_3x01234567), 4); c3 += 2;
478 vst1q_lane_u16((void*) c2, vreinterpretq_u16_s8(vout2x01234567_3x01234567), 0); c2 += 2;
479 vst1q_lane_u16((void*) c1, vreinterpretq_u16_s8(vout0x01234567_1x01234567), 4); c1 += 2;
480 vst1q_lane_u16((void*) c0, vreinterpretq_u16_s8(vout0x01234567_1x01234567), 0); c0 += 2;
481 vout4x01234567_5x01234567 = vextq_s8(vout4x01234567_5x01234567, vout4x01234567_5x01234567, 2);
482 vout2x01234567_3x01234567 = vextq_s8(vout2x01234567_3x01234567, vout2x01234567_3x01234567, 2);
483 vout0x01234567_1x01234567 = vextq_s8(vout0x01234567_1x01234567, vout0x01234567_1x01234567, 2);
484 }
485 if (nc & 1) {
486 vst1q_lane_s8(c5, vout4x01234567_5x01234567, 8);
487 vst1q_lane_s8(c4, vout4x01234567_5x01234567, 0);
488 vst1q_lane_s8(c3, vout2x01234567_3x01234567, 8);
489 vst1q_lane_s8(c2, vout2x01234567_3x01234567, 0);
490 vst1q_lane_s8(c1, vout0x01234567_1x01234567, 8);
491 vst1q_lane_s8(c0, vout0x01234567_1x01234567, 0);
492 }
493
494 nc = 0;
495 }
496 } while (nc != 0);
497 }