Debian Codegraph @

src/qd8-f16-qc8w-gemm/gen/qd8-f16-qc8w-gemm-8x8c8-minmax-neoni8mm.c

raw
1 // Auto-generated file. Do not edit!
2 // Template: src/qs8-gemm/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/gemm.h"
15 #include "xnnpack/math.h"
16
17
18 void xnn_qd8_f16_qc8w_gemm_minmax_ukernel_8x8c8__neoni8mm(
19 size_t mr,
20 size_t nc,
21 size_t kc,
22 const int8_t* restrict a,
23 size_t a_stride,
24 const void* restrict w,
25 xnn_float16* restrict c,
26 size_t cm_stride,
27 size_t cn_stride,
28 const union xnn_f16_minmax_params params[restrict XNN_MIN_ELEMENTS(1)],
29 const struct xnn_qd8_quantization_params quantization_params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
30 {
31 assert(mr != 0);
32 assert(mr <= 8);
33 assert(nc != 0);
34 assert(kc != 0);
35 assert(kc % sizeof(int8_t) == 0);
36 assert(a != NULL);
37 assert(w != NULL);
38 assert(c != NULL);
39
40 kc = round_up_po2(kc, 8 * sizeof(int8_t));
41 const int8_t* a0 = a;
42 uint16_t* c0 = (uint16_t*) c;
43 const int8_t* a1 = (const int8_t*) ((uintptr_t) a0 + a_stride);
44 uint16_t* c1 = (uint16_t*) ((uintptr_t) c0 + cm_stride);
45 if XNN_UNPREDICTABLE(mr < 2) {
46 a1 = a0;
47 c1 = c0;
48 }
49 const int8_t* a2 = (const int8_t*) ((uintptr_t) a1 + a_stride);
50 uint16_t* c2 = (uint16_t*) ((uintptr_t) c1 + cm_stride);
51 if XNN_UNPREDICTABLE(mr <= 2) {
52 a2 = a1;
53 c2 = c1;
54 }
55 const int8_t* a3 = (const int8_t*) ((uintptr_t) a2 + a_stride);
56 uint16_t* c3 = (uint16_t*) ((uintptr_t) c2 + cm_stride);
57 if XNN_UNPREDICTABLE(mr < 4) {
58 a3 = a2;
59 c3 = c2;
60 }
61 const int8_t* a4 = (const int8_t*) ((uintptr_t) a3 + a_stride);
62 uint16_t* c4 = (uint16_t*) ((uintptr_t) c3 + cm_stride);
63 if XNN_UNPREDICTABLE(mr <= 4) {
64 a4 = a3;
65 c4 = c3;
66 }
67 const int8_t* a5 = (const int8_t*) ((uintptr_t) a4 + a_stride);
68 uint16_t* c5 = (uint16_t*) ((uintptr_t) c4 + cm_stride);
69 if XNN_UNPREDICTABLE(mr < 6) {
70 a5 = a4;
71 c5 = c4;
72 }
73 const int8_t* a6 = (const int8_t*) ((uintptr_t) a5 + a_stride);
74 uint16_t* c6 = (uint16_t*) ((uintptr_t) c5 + cm_stride);
75 if XNN_UNPREDICTABLE(mr <= 6) {
76 a6 = a5;
77 c6 = c5;
78 }
79 const int8_t* a7 = (const int8_t*) ((uintptr_t) a6 + a_stride);
80 uint16_t* c7 = (uint16_t*) ((uintptr_t) c6 + cm_stride);
81 if XNN_UNPREDICTABLE(mr != 8) {
82 a7 = a6;
83 c7 = c6;
84 }
85
86 // Loop over groups of 8 columns.
87 do {
88 // Initialize accumulators with bias. 8 bias values are loaded from the
89 // weight matrix, at the start of the group of 8 columns.
90 const int32x4_t vinput_zero_point01 = vld1q_s32(&quantization_params[0].zero_point);
91 const int32x4_t vksum0123 = vld1q_s32(w); w = (const int32_t*) w + 4;
92 const int32x4_t vksumzp0x0123 = vmulq_lane_s32(vksum0123, vget_low_s32(vinput_zero_point01), 0);
93 const int32x4_t vksumzp1x0123 = vmulq_lane_s32(vksum0123, vget_high_s32(vinput_zero_point01), 0);
94 const int32x4_t vksum4567 = vld1q_s32(w); w = (const int32_t*) w + 4;
95 const int32x4_t vksumzp0x4567 = vmulq_lane_s32(vksum4567, vget_low_s32(vinput_zero_point01), 0);
96 const int32x4_t vksumzp1x4567 = vmulq_lane_s32(vksum4567, vget_high_s32(vinput_zero_point01), 0);
97 const int32x4_t vinput_zero_point23 = vld1q_s32(&quantization_params[2].zero_point);
98 const int32x4_t vksumzp2x0123 = vmulq_lane_s32(vksum0123, vget_low_s32(vinput_zero_point23), 0);
99 const int32x4_t vksumzp3x0123 = vmulq_lane_s32(vksum0123, vget_high_s32(vinput_zero_point23), 0);
100 const int32x4_t vksumzp2x4567 = vmulq_lane_s32(vksum4567, vget_low_s32(vinput_zero_point23), 0);
101 const int32x4_t vksumzp3x4567 = vmulq_lane_s32(vksum4567, vget_high_s32(vinput_zero_point23), 0);
102 const int32x4_t vinput_zero_point45 = vld1q_s32(&quantization_params[4].zero_point);
103 const int32x4_t vksumzp4x0123 = vmulq_lane_s32(vksum0123, vget_low_s32(vinput_zero_point45), 0);
104 const int32x4_t vksumzp5x0123 = vmulq_lane_s32(vksum0123, vget_high_s32(vinput_zero_point45), 0);
105 const int32x4_t vksumzp4x4567 = vmulq_lane_s32(vksum4567, vget_low_s32(vinput_zero_point45), 0);
106 const int32x4_t vksumzp5x4567 = vmulq_lane_s32(vksum4567, vget_high_s32(vinput_zero_point45), 0);
107 const int32x4_t vinput_zero_point67 = vld1q_s32(&quantization_params[6].zero_point);
108 const int32x4_t vksumzp6x0123 = vmulq_lane_s32(vksum0123, vget_low_s32(vinput_zero_point67), 0);
109 const int32x4_t vksumzp7x0123 = vmulq_lane_s32(vksum0123, vget_high_s32(vinput_zero_point67), 0);
110 const int32x4_t vksumzp6x4567 = vmulq_lane_s32(vksum4567, vget_low_s32(vinput_zero_point67), 0);
111 const int32x4_t vksumzp7x4567 = vmulq_lane_s32(vksum4567, vget_high_s32(vinput_zero_point67), 0);
112
113 int32x4_t vacc01x01 = vreinterpretq_s32_u64(vtrn1q_u64(vreinterpretq_u64_s32(vksumzp0x0123), vreinterpretq_u64_s32(vksumzp1x0123)));
114 int32x4_t vacc01x23 = vreinterpretq_s32_u64(vtrn2q_u64(vreinterpretq_u64_s32(vksumzp0x0123), vreinterpretq_u64_s32(vksumzp1x0123)));
115 int32x4_t vacc01x45 = vreinterpretq_s32_u64(vtrn1q_u64(vreinterpretq_u64_s32(vksumzp0x4567), vreinterpretq_u64_s32(vksumzp1x4567)));
116 int32x4_t vacc01x67 = vreinterpretq_s32_u64(vtrn2q_u64(vreinterpretq_u64_s32(vksumzp0x4567), vreinterpretq_u64_s32(vksumzp1x4567)));
117 int32x4_t vacc23x01 = vreinterpretq_s32_u64(vtrn1q_u64(vreinterpretq_u64_s32(vksumzp2x0123), vreinterpretq_u64_s32(vksumzp3x0123)));
118 int32x4_t vacc23x23 = vreinterpretq_s32_u64(vtrn2q_u64(vreinterpretq_u64_s32(vksumzp2x0123), vreinterpretq_u64_s32(vksumzp3x0123)));
119 int32x4_t vacc23x45 = vreinterpretq_s32_u64(vtrn1q_u64(vreinterpretq_u64_s32(vksumzp2x4567), vreinterpretq_u64_s32(vksumzp3x4567)));
120 int32x4_t vacc23x67 = vreinterpretq_s32_u64(vtrn2q_u64(vreinterpretq_u64_s32(vksumzp2x4567), vreinterpretq_u64_s32(vksumzp3x4567)));
121 int32x4_t vacc45x01 = vreinterpretq_s32_u64(vtrn1q_u64(vreinterpretq_u64_s32(vksumzp4x0123), vreinterpretq_u64_s32(vksumzp5x0123)));
122 int32x4_t vacc45x23 = vreinterpretq_s32_u64(vtrn2q_u64(vreinterpretq_u64_s32(vksumzp4x0123), vreinterpretq_u64_s32(vksumzp5x0123)));
123 int32x4_t vacc45x45 = vreinterpretq_s32_u64(vtrn1q_u64(vreinterpretq_u64_s32(vksumzp4x4567), vreinterpretq_u64_s32(vksumzp5x4567)));
124 int32x4_t vacc45x67 = vreinterpretq_s32_u64(vtrn2q_u64(vreinterpretq_u64_s32(vksumzp4x4567), vreinterpretq_u64_s32(vksumzp5x4567)));
125 int32x4_t vacc67x01 = vreinterpretq_s32_u64(vtrn1q_u64(vreinterpretq_u64_s32(vksumzp6x0123), vreinterpretq_u64_s32(vksumzp7x0123)));
126 int32x4_t vacc67x23 = vreinterpretq_s32_u64(vtrn2q_u64(vreinterpretq_u64_s32(vksumzp6x0123), vreinterpretq_u64_s32(vksumzp7x0123)));
127 int32x4_t vacc67x45 = vreinterpretq_s32_u64(vtrn1q_u64(vreinterpretq_u64_s32(vksumzp6x4567), vreinterpretq_u64_s32(vksumzp7x4567)));
128 int32x4_t vacc67x67 = vreinterpretq_s32_u64(vtrn2q_u64(vreinterpretq_u64_s32(vksumzp6x4567), vreinterpretq_u64_s32(vksumzp7x4567)));
129
130 // Inner accumulation loop along the 8 columns.
131 size_t k = kc;
132 // 2x partial unrolled loop to load 8 bytes at a time.
133
134 uint64x2x2_t va01x0123456789ABCDEF;
135 va01x0123456789ABCDEF.val[0] = vdupq_n_u64(0);
136 va01x0123456789ABCDEF.val[1] = vdupq_n_u64(0);
137 uint64x2x2_t va23x0123456789ABCDEF;
138 va23x0123456789ABCDEF.val[0] = vdupq_n_u64(0);
139 va23x0123456789ABCDEF.val[1] = vdupq_n_u64(0);
140 uint64x2x2_t va45x0123456789ABCDEF;
141 va45x0123456789ABCDEF.val[0] = vdupq_n_u64(0);
142 va45x0123456789ABCDEF.val[1] = vdupq_n_u64(0);
143 uint64x2x2_t va67x0123456789ABCDEF;
144 va67x0123456789ABCDEF.val[0] = vdupq_n_u64(0);
145 va67x0123456789ABCDEF.val[1] = vdupq_n_u64(0);
146
147 while (k >= 16 * sizeof(int8_t)) {
148 // Load a 8x16 block of activations.
149 va01x0123456789ABCDEF = vld2q_lane_u64((const void*) a0, va01x0123456789ABCDEF, 0); a0 += 16;
150 va23x0123456789ABCDEF = vld2q_lane_u64((const void*) a2, va23x0123456789ABCDEF, 0); a2 += 16;
151 va45x0123456789ABCDEF = vld2q_lane_u64((const void*) a4, va45x0123456789ABCDEF, 0); a4 += 16;
152 va67x0123456789ABCDEF = vld2q_lane_u64((const void*) a6, va67x0123456789ABCDEF, 0); a6 += 16;
153 va01x0123456789ABCDEF = vld2q_lane_u64((const void*) a1, va01x0123456789ABCDEF, 1); a1 += 16;
154 va23x0123456789ABCDEF = vld2q_lane_u64((const void*) a3, va23x0123456789ABCDEF, 1); a3 += 16;
155 va45x0123456789ABCDEF = vld2q_lane_u64((const void*) a5, va45x0123456789ABCDEF, 1); a5 += 16;
156 va67x0123456789ABCDEF = vld2q_lane_u64((const void*) a7, va67x0123456789ABCDEF, 1); a7 += 16;
157
158 const int8x16_t va01x01234567 = vreinterpretq_s8_u64(va01x0123456789ABCDEF.val[0]);
159 const int8x16_t va01x89ABCDEF = vreinterpretq_s8_u64(va01x0123456789ABCDEF.val[1]);
160 const int8x16_t va23x01234567 = vreinterpretq_s8_u64(va23x0123456789ABCDEF.val[0]);
161 const int8x16_t va23x89ABCDEF = vreinterpretq_s8_u64(va23x0123456789ABCDEF.val[1]);
162 const int8x16_t va45x01234567 = vreinterpretq_s8_u64(va45x0123456789ABCDEF.val[0]);
163 const int8x16_t va45x89ABCDEF = vreinterpretq_s8_u64(va45x0123456789ABCDEF.val[1]);
164 const int8x16_t va67x01234567 = vreinterpretq_s8_u64(va67x0123456789ABCDEF.val[0]);
165 const int8x16_t va67x89ABCDEF = vreinterpretq_s8_u64(va67x0123456789ABCDEF.val[1]);
166
167 // Load a 16x8 block of weights.
168 const int8x16_t vb01x01234567 = vld1q_s8(w); w = (const int8_t*) w + 16;
169 const int8x16_t vb23x01234567 = vld1q_s8(w); w = (const int8_t*) w + 16;
170 const int8x16_t vb45x01234567 = vld1q_s8(w); w = (const int8_t*) w + 16;
171 const int8x16_t vb67x01234567 = vld1q_s8(w); w = (const int8_t*) w + 16;
172 const int8x16_t vb01x89ABCDEF = vld1q_s8(w); w = (const int8_t*) w + 16;
173 const int8x16_t vb23x89ABCDEF = vld1q_s8(w); w = (const int8_t*) w + 16;
174 const int8x16_t vb45x89ABCDEF = vld1q_s8(w); w = (const int8_t*) w + 16;
175 const int8x16_t vb67x89ABCDEF = vld1q_s8(w); w = (const int8_t*) w + 16;
176
177 // Multiply-accumulate: 8x8 * 8x8 --> 8x8.
178 vacc01x01 = vmmlaq_s32(vacc01x01, va01x01234567, vb01x01234567);
179 vacc01x23 = vmmlaq_s32(vacc01x23, va01x01234567, vb23x01234567);
180 vacc01x45 = vmmlaq_s32(vacc01x45, va01x01234567, vb45x01234567);
181 vacc01x67 = vmmlaq_s32(vacc01x67, va01x01234567, vb67x01234567);
182 vacc23x01 = vmmlaq_s32(vacc23x01, va23x01234567, vb01x01234567);
183 vacc23x23 = vmmlaq_s32(vacc23x23, va23x01234567, vb23x01234567);
184 vacc23x45 = vmmlaq_s32(vacc23x45, va23x01234567, vb45x01234567);
185 vacc23x67 = vmmlaq_s32(vacc23x67, va23x01234567, vb67x01234567);
186 vacc45x01 = vmmlaq_s32(vacc45x01, va45x01234567, vb01x01234567);
187 vacc45x23 = vmmlaq_s32(vacc45x23, va45x01234567, vb23x01234567);
188 vacc45x45 = vmmlaq_s32(vacc45x45, va45x01234567, vb45x01234567);
189 vacc45x67 = vmmlaq_s32(vacc45x67, va45x01234567, vb67x01234567);
190 vacc67x01 = vmmlaq_s32(vacc67x01, va67x01234567, vb01x01234567);
191 vacc67x23 = vmmlaq_s32(vacc67x23, va67x01234567, vb23x01234567);
192 vacc67x45 = vmmlaq_s32(vacc67x45, va67x01234567, vb45x01234567);
193 vacc67x67 = vmmlaq_s32(vacc67x67, va67x01234567, vb67x01234567);
194 vacc01x01 = vmmlaq_s32(vacc01x01, va01x89ABCDEF, vb01x89ABCDEF);
195 vacc01x23 = vmmlaq_s32(vacc01x23, va01x89ABCDEF, vb23x89ABCDEF);
196 vacc01x45 = vmmlaq_s32(vacc01x45, va01x89ABCDEF, vb45x89ABCDEF);
197 vacc01x67 = vmmlaq_s32(vacc01x67, va01x89ABCDEF, vb67x89ABCDEF);
198 vacc23x01 = vmmlaq_s32(vacc23x01, va23x89ABCDEF, vb01x89ABCDEF);
199 vacc23x23 = vmmlaq_s32(vacc23x23, va23x89ABCDEF, vb23x89ABCDEF);
200 vacc23x45 = vmmlaq_s32(vacc23x45, va23x89ABCDEF, vb45x89ABCDEF);
201 vacc23x67 = vmmlaq_s32(vacc23x67, va23x89ABCDEF, vb67x89ABCDEF);
202 vacc45x01 = vmmlaq_s32(vacc45x01, va45x89ABCDEF, vb01x89ABCDEF);
203 vacc45x23 = vmmlaq_s32(vacc45x23, va45x89ABCDEF, vb23x89ABCDEF);
204 vacc45x45 = vmmlaq_s32(vacc45x45, va45x89ABCDEF, vb45x89ABCDEF);
205 vacc45x67 = vmmlaq_s32(vacc45x67, va45x89ABCDEF, vb67x89ABCDEF);
206 vacc67x01 = vmmlaq_s32(vacc67x01, va67x89ABCDEF, vb01x89ABCDEF);
207 vacc67x23 = vmmlaq_s32(vacc67x23, va67x89ABCDEF, vb23x89ABCDEF);
208 vacc67x45 = vmmlaq_s32(vacc67x45, va67x89ABCDEF, vb45x89ABCDEF);
209 vacc67x67 = vmmlaq_s32(vacc67x67, va67x89ABCDEF, vb67x89ABCDEF);
210
211 k -= 16 * sizeof(int8_t);
212 }
213 // Handle up to 8 final positions of `k`
214 if XNN_UNLIKELY(k != 0) {
215 // Load a 8x8 block of activations.
216 uint64x2_t va01x01234567 = vld1q_dup_u64((const void*) a0); a0 += 8;
217 uint64x2_t va23x01234567 = vld1q_dup_u64((const void*) a2); a2 += 8;
218 uint64x2_t va45x01234567 = vld1q_dup_u64((const void*) a4); a4 += 8;
219 uint64x2_t va67x01234567 = vld1q_dup_u64((const void*) a6); a6 += 8;
220 va01x01234567 = vld1q_lane_u64((const void*) a1, va01x01234567, 1); a1 += 8;
221 va23x01234567 = vld1q_lane_u64((const void*) a3, va23x01234567, 1); a3 += 8;
222 va45x01234567 = vld1q_lane_u64((const void*) a5, va45x01234567, 1); a5 += 8;
223 va67x01234567 = vld1q_lane_u64((const void*) a7, va67x01234567, 1); a7 += 8;
224
225 // Load a 16x8 block of weights.
226 const int8x16_t vb01x01234567 = vld1q_s8(w); w = (const int8_t*) w + 16;
227 const int8x16_t vb23x01234567 = vld1q_s8(w); w = (const int8_t*) w + 16;
228 const int8x16_t vb45x01234567 = vld1q_s8(w); w = (const int8_t*) w + 16;
229 const int8x16_t vb67x01234567 = vld1q_s8(w); w = (const int8_t*) w + 16;
230
231 // Multiply-accumulate: 8x4 * 4x8 --> 8x8.
232 vacc01x01 = vmmlaq_s32(vacc01x01, vreinterpretq_s8_u64(va01x01234567), vb01x01234567);
233 vacc01x23 = vmmlaq_s32(vacc01x23, vreinterpretq_s8_u64(va01x01234567), vb23x01234567);
234 vacc01x45 = vmmlaq_s32(vacc01x45, vreinterpretq_s8_u64(va01x01234567), vb45x01234567);
235 vacc01x67 = vmmlaq_s32(vacc01x67, vreinterpretq_s8_u64(va01x01234567), vb67x01234567);
236 vacc23x01 = vmmlaq_s32(vacc23x01, vreinterpretq_s8_u64(va23x01234567), vb01x01234567);
237 vacc23x23 = vmmlaq_s32(vacc23x23, vreinterpretq_s8_u64(va23x01234567), vb23x01234567);
238 vacc23x45 = vmmlaq_s32(vacc23x45, vreinterpretq_s8_u64(va23x01234567), vb45x01234567);
239 vacc23x67 = vmmlaq_s32(vacc23x67, vreinterpretq_s8_u64(va23x01234567), vb67x01234567);
240 vacc45x01 = vmmlaq_s32(vacc45x01, vreinterpretq_s8_u64(va45x01234567), vb01x01234567);
241 vacc45x23 = vmmlaq_s32(vacc45x23, vreinterpretq_s8_u64(va45x01234567), vb23x01234567);
242 vacc45x45 = vmmlaq_s32(vacc45x45, vreinterpretq_s8_u64(va45x01234567), vb45x01234567);
243 vacc45x67 = vmmlaq_s32(vacc45x67, vreinterpretq_s8_u64(va45x01234567), vb67x01234567);
244 vacc67x01 = vmmlaq_s32(vacc67x01, vreinterpretq_s8_u64(va67x01234567), vb01x01234567);
245 vacc67x23 = vmmlaq_s32(vacc67x23, vreinterpretq_s8_u64(va67x01234567), vb23x01234567);
246 vacc67x45 = vmmlaq_s32(vacc67x45, vreinterpretq_s8_u64(va67x01234567), vb45x01234567);
247 vacc67x67 = vmmlaq_s32(vacc67x67, vreinterpretq_s8_u64(va67x01234567), vb67x01234567);
248 }
249
250 int32x4_t vacc0x0123 = vreinterpretq_s32_u64(vtrn1q_u64(vreinterpretq_u64_s32(vacc01x01), vreinterpretq_u64_s32(vacc01x23)));
251 int32x4_t vacc1x0123 = vreinterpretq_s32_u64(vtrn2q_u64(vreinterpretq_u64_s32(vacc01x01), vreinterpretq_u64_s32(vacc01x23)));
252 int32x4_t vacc0x4567 = vreinterpretq_s32_u64(vtrn1q_u64(vreinterpretq_u64_s32(vacc01x45), vreinterpretq_u64_s32(vacc01x67)));
253 int32x4_t vacc1x4567 = vreinterpretq_s32_u64(vtrn2q_u64(vreinterpretq_u64_s32(vacc01x45), vreinterpretq_u64_s32(vacc01x67)));
254 int32x4_t vacc2x0123 = vreinterpretq_s32_u64(vtrn1q_u64(vreinterpretq_u64_s32(vacc23x01), vreinterpretq_u64_s32(vacc23x23)));
255 int32x4_t vacc3x0123 = vreinterpretq_s32_u64(vtrn2q_u64(vreinterpretq_u64_s32(vacc23x01), vreinterpretq_u64_s32(vacc23x23)));
256 int32x4_t vacc2x4567 = vreinterpretq_s32_u64(vtrn1q_u64(vreinterpretq_u64_s32(vacc23x45), vreinterpretq_u64_s32(vacc23x67)));
257 int32x4_t vacc3x4567 = vreinterpretq_s32_u64(vtrn2q_u64(vreinterpretq_u64_s32(vacc23x45), vreinterpretq_u64_s32(vacc23x67)));
258 int32x4_t vacc4x0123 = vreinterpretq_s32_u64(vtrn1q_u64(vreinterpretq_u64_s32(vacc45x01), vreinterpretq_u64_s32(vacc45x23)));
259 int32x4_t vacc5x0123 = vreinterpretq_s32_u64(vtrn2q_u64(vreinterpretq_u64_s32(vacc45x01), vreinterpretq_u64_s32(vacc45x23)));
260 int32x4_t vacc4x4567 = vreinterpretq_s32_u64(vtrn1q_u64(vreinterpretq_u64_s32(vacc45x45), vreinterpretq_u64_s32(vacc45x67)));
261 int32x4_t vacc5x4567 = vreinterpretq_s32_u64(vtrn2q_u64(vreinterpretq_u64_s32(vacc45x45), vreinterpretq_u64_s32(vacc45x67)));
262 int32x4_t vacc6x0123 = vreinterpretq_s32_u64(vtrn1q_u64(vreinterpretq_u64_s32(vacc67x01), vreinterpretq_u64_s32(vacc67x23)));
263 int32x4_t vacc7x0123 = vreinterpretq_s32_u64(vtrn2q_u64(vreinterpretq_u64_s32(vacc67x01), vreinterpretq_u64_s32(vacc67x23)));
264 int32x4_t vacc6x4567 = vreinterpretq_s32_u64(vtrn1q_u64(vreinterpretq_u64_s32(vacc67x45), vreinterpretq_u64_s32(vacc67x67)));
265 int32x4_t vacc7x4567 = vreinterpretq_s32_u64(vtrn2q_u64(vreinterpretq_u64_s32(vacc67x45), vreinterpretq_u64_s32(vacc67x67)));
266 float32x4_t vout0x0123 = vcvtq_f32_s32(vacc0x0123);
267 float32x4_t vout0x4567 = vcvtq_f32_s32(vacc0x4567);
268 float32x4_t vout1x0123 = vcvtq_f32_s32(vacc1x0123);
269 float32x4_t vout1x4567 = vcvtq_f32_s32(vacc1x4567);
270 float32x4_t vout2x0123 = vcvtq_f32_s32(vacc2x0123);
271 float32x4_t vout2x4567 = vcvtq_f32_s32(vacc2x4567);
272 float32x4_t vout3x0123 = vcvtq_f32_s32(vacc3x0123);
273 float32x4_t vout3x4567 = vcvtq_f32_s32(vacc3x4567);
274 float32x4_t vout4x0123 = vcvtq_f32_s32(vacc4x0123);
275 float32x4_t vout4x4567 = vcvtq_f32_s32(vacc4x4567);
276 float32x4_t vout5x0123 = vcvtq_f32_s32(vacc5x0123);
277 float32x4_t vout5x4567 = vcvtq_f32_s32(vacc5x4567);
278 float32x4_t vout6x0123 = vcvtq_f32_s32(vacc6x0123);
279 float32x4_t vout6x4567 = vcvtq_f32_s32(vacc6x4567);
280 float32x4_t vout7x0123 = vcvtq_f32_s32(vacc7x0123);
281 float32x4_t vout7x4567 = vcvtq_f32_s32(vacc7x4567);
282
283 const float32x4_t vinput_scale01 = vreinterpretq_f32_s32(vld1q_s32(&quantization_params[0].zero_point));
284 vout0x0123 = vmulq_lane_f32(vout0x0123, vget_low_f32(vinput_scale01), 1);
285 vout1x0123 = vmulq_lane_f32(vout1x0123, vget_high_f32(vinput_scale01), 1);
286 vout0x4567 = vmulq_lane_f32(vout0x4567, vget_low_f32(vinput_scale01), 1);
287 vout1x4567 = vmulq_lane_f32(vout1x4567, vget_high_f32(vinput_scale01), 1);
288 const float32x4_t vinput_scale23 = vreinterpretq_f32_s32(vld1q_s32(&quantization_params[2].zero_point));
289 vout2x0123 = vmulq_lane_f32(vout2x0123, vget_low_f32(vinput_scale23), 1);
290 vout3x0123 = vmulq_lane_f32(vout3x0123, vget_high_f32(vinput_scale23), 1);
291 vout2x4567 = vmulq_lane_f32(vout2x4567, vget_low_f32(vinput_scale23), 1);
292 vout3x4567 = vmulq_lane_f32(vout3x4567, vget_high_f32(vinput_scale23), 1);
293 const float32x4_t vinput_scale45 = vreinterpretq_f32_s32(vld1q_s32(&quantization_params[4].zero_point));
294 vout4x0123 = vmulq_lane_f32(vout4x0123, vget_low_f32(vinput_scale45), 1);
295 vout5x0123 = vmulq_lane_f32(vout5x0123, vget_high_f32(vinput_scale45), 1);
296 vout4x4567 = vmulq_lane_f32(vout4x4567, vget_low_f32(vinput_scale45), 1);
297 vout5x4567 = vmulq_lane_f32(vout5x4567, vget_high_f32(vinput_scale45), 1);
298 const float32x4_t vinput_scale67 = vreinterpretq_f32_s32(vld1q_s32(&quantization_params[6].zero_point));
299 vout6x0123 = vmulq_lane_f32(vout6x0123, vget_low_f32(vinput_scale67), 1);
300 vout7x0123 = vmulq_lane_f32(vout7x0123, vget_high_f32(vinput_scale67), 1);
301 vout6x4567 = vmulq_lane_f32(vout6x4567, vget_low_f32(vinput_scale67), 1);
302 vout7x4567 = vmulq_lane_f32(vout7x4567, vget_high_f32(vinput_scale67), 1);
303
304 const float32x4_t vfilter_output_scale0123 = vld1q_f32(w); w = (const float*) w + 4;
305 const float32x4_t vfilter_output_scale4567 = vld1q_f32(w); w = (const float*) w + 4;
306
307 const float32x4_t vbias0123 = vld1q_f32(w); w = (const float*) w + 4;
308 vout0x0123 = vfmaq_f32(vbias0123, vout0x0123, vfilter_output_scale0123);
309 vout1x0123 = vfmaq_f32(vbias0123, vout1x0123, vfilter_output_scale0123);
310 vout2x0123 = vfmaq_f32(vbias0123, vout2x0123, vfilter_output_scale0123);
311 vout3x0123 = vfmaq_f32(vbias0123, vout3x0123, vfilter_output_scale0123);
312 vout4x0123 = vfmaq_f32(vbias0123, vout4x0123, vfilter_output_scale0123);
313 vout5x0123 = vfmaq_f32(vbias0123, vout5x0123, vfilter_output_scale0123);
314 vout6x0123 = vfmaq_f32(vbias0123, vout6x0123, vfilter_output_scale0123);
315 vout7x0123 = vfmaq_f32(vbias0123, vout7x0123, vfilter_output_scale0123);
316 const float32x4_t vbias4567 = vld1q_f32(w); w = (const float*) w + 4;
317 vout0x4567 = vfmaq_f32(vbias4567, vout0x4567, vfilter_output_scale4567);
318 vout1x4567 = vfmaq_f32(vbias4567, vout1x4567, vfilter_output_scale4567);
319 vout2x4567 = vfmaq_f32(vbias4567, vout2x4567, vfilter_output_scale4567);
320 vout3x4567 = vfmaq_f32(vbias4567, vout3x4567, vfilter_output_scale4567);
321 vout4x4567 = vfmaq_f32(vbias4567, vout4x4567, vfilter_output_scale4567);
322 vout5x4567 = vfmaq_f32(vbias4567, vout5x4567, vfilter_output_scale4567);
323 vout6x4567 = vfmaq_f32(vbias4567, vout6x4567, vfilter_output_scale4567);
324 vout7x4567 = vfmaq_f32(vbias4567, vout7x4567, vfilter_output_scale4567);
325
326 float16x8_t vfp16out0x01234567 = vcombine_f16(vcvt_f16_f32(vout0x0123), vcvt_f16_f32(vout0x4567));
327 float16x8_t vfp16out1x01234567 = vcombine_f16(vcvt_f16_f32(vout1x0123), vcvt_f16_f32(vout1x4567));
328 float16x8_t vfp16out2x01234567 = vcombine_f16(vcvt_f16_f32(vout2x0123), vcvt_f16_f32(vout2x4567));
329 float16x8_t vfp16out3x01234567 = vcombine_f16(vcvt_f16_f32(vout3x0123), vcvt_f16_f32(vout3x4567));
330 float16x8_t vfp16out4x01234567 = vcombine_f16(vcvt_f16_f32(vout4x0123), vcvt_f16_f32(vout4x4567));
331 float16x8_t vfp16out5x01234567 = vcombine_f16(vcvt_f16_f32(vout5x0123), vcvt_f16_f32(vout5x4567));
332 float16x8_t vfp16out6x01234567 = vcombine_f16(vcvt_f16_f32(vout6x0123), vcvt_f16_f32(vout6x4567));
333 float16x8_t vfp16out7x01234567 = vcombine_f16(vcvt_f16_f32(vout7x0123), vcvt_f16_f32(vout7x4567));
334
335 const float16x8_t voutput_min = vreinterpretq_f16_u16(vld1q_dup_u16((const uint16_t*) &params->scalar.min));
336 vfp16out0x01234567 = vmaxq_f16(vfp16out0x01234567, voutput_min);
337 vfp16out1x01234567 = vmaxq_f16(vfp16out1x01234567, voutput_min);
338 vfp16out2x01234567 = vmaxq_f16(vfp16out2x01234567, voutput_min);
339 vfp16out3x01234567 = vmaxq_f16(vfp16out3x01234567, voutput_min);
340 vfp16out4x01234567 = vmaxq_f16(vfp16out4x01234567, voutput_min);
341 vfp16out5x01234567 = vmaxq_f16(vfp16out5x01234567, voutput_min);
342 vfp16out6x01234567 = vmaxq_f16(vfp16out6x01234567, voutput_min);
343 vfp16out7x01234567 = vmaxq_f16(vfp16out7x01234567, voutput_min);
344 const float16x8_t voutput_max = vreinterpretq_f16_u16(vld1q_dup_u16((const uint16_t*) &params->scalar.max));
345 vfp16out0x01234567 = vminq_f16(vfp16out0x01234567, voutput_max);
346 vfp16out1x01234567 = vminq_f16(vfp16out1x01234567, voutput_max);
347 vfp16out2x01234567 = vminq_f16(vfp16out2x01234567, voutput_max);
348 vfp16out3x01234567 = vminq_f16(vfp16out3x01234567, voutput_max);
349 vfp16out4x01234567 = vminq_f16(vfp16out4x01234567, voutput_max);
350 vfp16out5x01234567 = vminq_f16(vfp16out5x01234567, voutput_max);
351 vfp16out6x01234567 = vminq_f16(vfp16out6x01234567, voutput_max);
352 vfp16out7x01234567 = vminq_f16(vfp16out7x01234567, voutput_max);
353 if XNN_LIKELY(nc >= 8) {
354 vst1q_u16(c0, vreinterpretq_u16_f16(vfp16out0x01234567));
355 vst1q_u16(c1, vreinterpretq_u16_f16(vfp16out1x01234567));
356 vst1q_u16(c2, vreinterpretq_u16_f16(vfp16out2x01234567));
357 vst1q_u16(c3, vreinterpretq_u16_f16(vfp16out3x01234567));
358 vst1q_u16(c4, vreinterpretq_u16_f16(vfp16out4x01234567));
359 vst1q_u16(c5, vreinterpretq_u16_f16(vfp16out5x01234567));
360 vst1q_u16(c6, vreinterpretq_u16_f16(vfp16out6x01234567));
361 vst1q_u16(c7, vreinterpretq_u16_f16(vfp16out7x01234567));
362
363 a0 = (const int8_t*) ((uintptr_t) a0 - kc);
364 a1 = (const int8_t*) ((uintptr_t) a1 - kc);
365 a2 = (const int8_t*) ((uintptr_t) a2 - kc);
366 a3 = (const int8_t*) ((uintptr_t) a3 - kc);
367 a4 = (const int8_t*) ((uintptr_t) a4 - kc);
368 a5 = (const int8_t*) ((uintptr_t) a5 - kc);
369 a6 = (const int8_t*) ((uintptr_t) a6 - kc);
370 a7 = (const int8_t*) ((uintptr_t) a7 - kc);
371
372 c0 = (uint16_t*) ((uintptr_t) c0 + cn_stride);
373 c1 = (uint16_t*) ((uintptr_t) c1 + cn_stride);
374 c2 = (uint16_t*) ((uintptr_t) c2 + cn_stride);
375 c3 = (uint16_t*) ((uintptr_t) c3 + cn_stride);
376 c4 = (uint16_t*) ((uintptr_t) c4 + cn_stride);
377 c5 = (uint16_t*) ((uintptr_t) c5 + cn_stride);
378 c6 = (uint16_t*) ((uintptr_t) c6 + cn_stride);
379 c7 = (uint16_t*) ((uintptr_t) c7 + cn_stride);
380
381 nc -= 8;
382 } else {
383 float16x4_t vfp16out0x0123 = vget_low_f16(vfp16out0x01234567);
384 float16x4_t vfp16out1x0123 = vget_low_f16(vfp16out1x01234567);
385 float16x4_t vfp16out2x0123 = vget_low_f16(vfp16out2x01234567);
386 float16x4_t vfp16out3x0123 = vget_low_f16(vfp16out3x01234567);
387 float16x4_t vfp16out4x0123 = vget_low_f16(vfp16out4x01234567);
388 float16x4_t vfp16out5x0123 = vget_low_f16(vfp16out5x01234567);
389 float16x4_t vfp16out6x0123 = vget_low_f16(vfp16out6x01234567);
390 float16x4_t vfp16out7x0123 = vget_low_f16(vfp16out7x01234567);
391 if (nc & 4) {
392 vst1_u16(c0, vreinterpret_u16_f16(vfp16out0x0123)); c0 += 4;
393 vst1_u16(c1, vreinterpret_u16_f16(vfp16out1x0123)); c1 += 4;
394 vst1_u16(c2, vreinterpret_u16_f16(vfp16out2x0123)); c2 += 4;
395 vst1_u16(c3, vreinterpret_u16_f16(vfp16out3x0123)); c3 += 4;
396 vst1_u16(c4, vreinterpret_u16_f16(vfp16out4x0123)); c4 += 4;
397 vst1_u16(c5, vreinterpret_u16_f16(vfp16out5x0123)); c5 += 4;
398 vst1_u16(c6, vreinterpret_u16_f16(vfp16out6x0123)); c6 += 4;
399 vst1_u16(c7, vreinterpret_u16_f16(vfp16out7x0123)); c7 += 4;
400 vfp16out0x0123 = vget_high_f16(vfp16out0x01234567);
401 vfp16out1x0123 = vget_high_f16(vfp16out1x01234567);
402 vfp16out2x0123 = vget_high_f16(vfp16out2x01234567);
403 vfp16out3x0123 = vget_high_f16(vfp16out3x01234567);
404 vfp16out4x0123 = vget_high_f16(vfp16out4x01234567);
405 vfp16out5x0123 = vget_high_f16(vfp16out5x01234567);
406 vfp16out6x0123 = vget_high_f16(vfp16out6x01234567);
407 vfp16out7x0123 = vget_high_f16(vfp16out7x01234567);
408 }
409 if (nc & 2) {
410 vst1_lane_u32((void*) c0, vreinterpret_u32_f16(vfp16out0x0123), 0); c0 += 2;
411 vst1_lane_u32((void*) c1, vreinterpret_u32_f16(vfp16out1x0123), 0); c1 += 2;
412 vst1_lane_u32((void*) c2, vreinterpret_u32_f16(vfp16out2x0123), 0); c2 += 2;
413 vst1_lane_u32((void*) c3, vreinterpret_u32_f16(vfp16out3x0123), 0); c3 += 2;
414 vst1_lane_u32((void*) c4, vreinterpret_u32_f16(vfp16out4x0123), 0); c4 += 2;
415 vst1_lane_u32((void*) c5, vreinterpret_u32_f16(vfp16out5x0123), 0); c5 += 2;
416 vst1_lane_u32((void*) c6, vreinterpret_u32_f16(vfp16out6x0123), 0); c6 += 2;
417 vst1_lane_u32((void*) c7, vreinterpret_u32_f16(vfp16out7x0123), 0); c7 += 2;
418 vfp16out0x0123 = vext_f16(vfp16out0x0123, vfp16out0x0123, 2);
419 vfp16out1x0123 = vext_f16(vfp16out1x0123, vfp16out1x0123, 2);
420 vfp16out2x0123 = vext_f16(vfp16out2x0123, vfp16out2x0123, 2);
421 vfp16out3x0123 = vext_f16(vfp16out3x0123, vfp16out3x0123, 2);
422 vfp16out4x0123 = vext_f16(vfp16out4x0123, vfp16out4x0123, 2);
423 vfp16out5x0123 = vext_f16(vfp16out5x0123, vfp16out5x0123, 2);
424 vfp16out6x0123 = vext_f16(vfp16out6x0123, vfp16out6x0123, 2);
425 vfp16out7x0123 = vext_f16(vfp16out7x0123, vfp16out7x0123, 2);
426 }
427 if (nc & 1) {
428 vst1_lane_u16(c0, vreinterpret_u16_f16(vfp16out0x0123), 0);
429 vst1_lane_u16(c1, vreinterpret_u16_f16(vfp16out1x0123), 0);
430 vst1_lane_u16(c2, vreinterpret_u16_f16(vfp16out2x0123), 0);
431 vst1_lane_u16(c3, vreinterpret_u16_f16(vfp16out3x0123), 0);
432 vst1_lane_u16(c4, vreinterpret_u16_f16(vfp16out4x0123), 0);
433 vst1_lane_u16(c5, vreinterpret_u16_f16(vfp16out5x0123), 0);
434 vst1_lane_u16(c6, vreinterpret_u16_f16(vfp16out6x0123), 0);
435 vst1_lane_u16(c7, vreinterpret_u16_f16(vfp16out7x0123), 0);
436 }
437 nc = 0;
438 }
439 } while (nc != 0);
440 }