Debian Codegraph @

src/qd8-f16-qc4w-gemm/gen/qd8-f16-qc4w-gemm-1x32c8-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_qc4w_gemm_minmax_ukernel_1x32c8__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 struct xnn_f16_qc4w_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 <= 1);
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 int8x16_t vmask = vmovq_n_s8(INT8_C(0xF0));
44
45 // Loop over groups of 32 columns.
46 do {
47 // Initialize accumulators with bias. 32 bias values are loaded from the
48 // weight matrix, at the start of the group of 32 columns.
49 const int32x4_t vinput_zero_point01 = vld1q_dup_s32(&quantization_params[0].zero_point);
50 const int32x4_t vksum0123 = vld1q_s32(w); w = (const int32_t*) w + 4;
51 const int32x4_t vksumzp0x0123 = vmulq_s32(vksum0123, vinput_zero_point01);
52 const int32x4_t vksum4567 = vld1q_s32(w); w = (const int32_t*) w + 4;
53 const int32x4_t vksumzp0x4567 = vmulq_s32(vksum4567, vinput_zero_point01);
54 const int32x4_t vksum89AB = vld1q_s32(w); w = (const int32_t*) w + 4;
55 const int32x4_t vksumzp0x89AB = vmulq_s32(vksum89AB, vinput_zero_point01);
56 const int32x4_t vksumCDEF = vld1q_s32(w); w = (const int32_t*) w + 4;
57 const int32x4_t vksumzp0xCDEF = vmulq_s32(vksumCDEF, vinput_zero_point01);
58 const int32x4_t vksumGHIJ = vld1q_s32(w); w = (const int32_t*) w + 4;
59 const int32x4_t vksumzp0xGHIJ = vmulq_s32(vksumGHIJ, vinput_zero_point01);
60 const int32x4_t vksumKLMN = vld1q_s32(w); w = (const int32_t*) w + 4;
61 const int32x4_t vksumzp0xKLMN = vmulq_s32(vksumKLMN, vinput_zero_point01);
62 const int32x4_t vksumOPQR = vld1q_s32(w); w = (const int32_t*) w + 4;
63 const int32x4_t vksumzp0xOPQR = vmulq_s32(vksumOPQR, vinput_zero_point01);
64 const int32x4_t vksumSTUV = vld1q_s32(w); w = (const int32_t*) w + 4;
65 const int32x4_t vksumzp0xSTUV = vmulq_s32(vksumSTUV, vinput_zero_point01);
66
67 int32x4_t vacc01x01 = vreinterpretq_s32_u64(vtrn1q_u64(vreinterpretq_u64_s32(vksumzp0x0123), vreinterpretq_u64_s32(vksumzp0x0123)));
68 int32x4_t vacc01x23 = vreinterpretq_s32_u64(vtrn2q_u64(vreinterpretq_u64_s32(vksumzp0x0123), vreinterpretq_u64_s32(vksumzp0x0123)));
69 int32x4_t vacc01x45 = vreinterpretq_s32_u64(vtrn1q_u64(vreinterpretq_u64_s32(vksumzp0x4567), vreinterpretq_u64_s32(vksumzp0x4567)));
70 int32x4_t vacc01x67 = vreinterpretq_s32_u64(vtrn2q_u64(vreinterpretq_u64_s32(vksumzp0x4567), vreinterpretq_u64_s32(vksumzp0x4567)));
71 int32x4_t vacc01x89 = vreinterpretq_s32_u64(vtrn1q_u64(vreinterpretq_u64_s32(vksumzp0x89AB), vreinterpretq_u64_s32(vksumzp0x89AB)));
72 int32x4_t vacc01xAB = vreinterpretq_s32_u64(vtrn2q_u64(vreinterpretq_u64_s32(vksumzp0x89AB), vreinterpretq_u64_s32(vksumzp0x89AB)));
73 int32x4_t vacc01xCD = vreinterpretq_s32_u64(vtrn1q_u64(vreinterpretq_u64_s32(vksumzp0xCDEF), vreinterpretq_u64_s32(vksumzp0xCDEF)));
74 int32x4_t vacc01xEF = vreinterpretq_s32_u64(vtrn2q_u64(vreinterpretq_u64_s32(vksumzp0xCDEF), vreinterpretq_u64_s32(vksumzp0xCDEF)));
75 int32x4_t vacc01xGH = vreinterpretq_s32_u64(vtrn1q_u64(vreinterpretq_u64_s32(vksumzp0xGHIJ), vreinterpretq_u64_s32(vksumzp0xGHIJ)));
76 int32x4_t vacc01xIJ = vreinterpretq_s32_u64(vtrn2q_u64(vreinterpretq_u64_s32(vksumzp0xGHIJ), vreinterpretq_u64_s32(vksumzp0xGHIJ)));
77 int32x4_t vacc01xKL = vreinterpretq_s32_u64(vtrn1q_u64(vreinterpretq_u64_s32(vksumzp0xKLMN), vreinterpretq_u64_s32(vksumzp0xKLMN)));
78 int32x4_t vacc01xMN = vreinterpretq_s32_u64(vtrn2q_u64(vreinterpretq_u64_s32(vksumzp0xKLMN), vreinterpretq_u64_s32(vksumzp0xKLMN)));
79 int32x4_t vacc01xOP = vreinterpretq_s32_u64(vtrn1q_u64(vreinterpretq_u64_s32(vksumzp0xOPQR), vreinterpretq_u64_s32(vksumzp0xOPQR)));
80 int32x4_t vacc01xQR = vreinterpretq_s32_u64(vtrn2q_u64(vreinterpretq_u64_s32(vksumzp0xOPQR), vreinterpretq_u64_s32(vksumzp0xOPQR)));
81 int32x4_t vacc01xST = vreinterpretq_s32_u64(vtrn1q_u64(vreinterpretq_u64_s32(vksumzp0xSTUV), vreinterpretq_u64_s32(vksumzp0xSTUV)));
82 int32x4_t vacc01xUV = vreinterpretq_s32_u64(vtrn2q_u64(vreinterpretq_u64_s32(vksumzp0xSTUV), vreinterpretq_u64_s32(vksumzp0xSTUV)));
83
84 // Inner accumulation loop along the 32 columns.
85 size_t k = kc;
86 // 2x partial unrolled loop to load 8 bytes at a time.
87
88 uint64x2x2_t va01x0123456789ABCDEF;
89 va01x0123456789ABCDEF.val[0] = vdupq_n_u64(0);
90 va01x0123456789ABCDEF.val[1] = vdupq_n_u64(0);
91
92 while (k >= 16 * sizeof(int8_t)) {
93 // Load a 1x16 block of activations.
94 va01x0123456789ABCDEF = vld2q_lane_u64((const void*) a0, va01x0123456789ABCDEF, 0); a0 += 16;
95
96 const int8x16_t va01x01234567 = vreinterpretq_s8_u64(va01x0123456789ABCDEF.val[0]);
97 const int8x16_t va01x89ABCDEF = vreinterpretq_s8_u64(va01x0123456789ABCDEF.val[1]);
98
99 // Load a 16x32 block of weights.
100 const int8x16_t vb01x0123456789ABCDEF = vld1q_s8(w); w = (const int8_t*) w + 16;
101 const int8x16_t vb23x0123456789ABCDEF = vld1q_s8(w); w = (const int8_t*) w + 16;
102 const int8x16_t vb45x0123456789ABCDEF = vld1q_s8(w); w = (const int8_t*) w + 16;
103 const int8x16_t vb67x0123456789ABCDEF = vld1q_s8(w); w = (const int8_t*) w + 16;
104 const int8x16_t vb89x0123456789ABCDEF = vld1q_s8(w); w = (const int8_t*) w + 16;
105 const int8x16_t vbABx0123456789ABCDEF = vld1q_s8(w); w = (const int8_t*) w + 16;
106 const int8x16_t vbCDx0123456789ABCDEF = vld1q_s8(w); w = (const int8_t*) w + 16;
107 const int8x16_t vbEFx0123456789ABCDEF = vld1q_s8(w); w = (const int8_t*) w + 16;
108 const int8x16_t vbGHx0123456789ABCDEF = vld1q_s8(w); w = (const int8_t*) w + 16;
109 const int8x16_t vbIJx0123456789ABCDEF = vld1q_s8(w); w = (const int8_t*) w + 16;
110 const int8x16_t vbKLx0123456789ABCDEF = vld1q_s8(w); w = (const int8_t*) w + 16;
111 const int8x16_t vbMNx0123456789ABCDEF = vld1q_s8(w); w = (const int8_t*) w + 16;
112 const int8x16_t vbOPx0123456789ABCDEF = vld1q_s8(w); w = (const int8_t*) w + 16;
113 const int8x16_t vbQRx0123456789ABCDEF = vld1q_s8(w); w = (const int8_t*) w + 16;
114 const int8x16_t vbSTx0123456789ABCDEF = vld1q_s8(w); w = (const int8_t*) w + 16;
115 const int8x16_t vbUVx0123456789ABCDEF = vld1q_s8(w); w = (const int8_t*) w + 16;
116 const int8x16_t vb01x01234567 = vshlq_n_s8(vb01x0123456789ABCDEF, 4);
117 const int8x16_t vb23x01234567 = vshlq_n_s8(vb23x0123456789ABCDEF, 4);
118 const int8x16_t vb45x01234567 = vshlq_n_s8(vb45x0123456789ABCDEF, 4);
119 const int8x16_t vb67x01234567 = vshlq_n_s8(vb67x0123456789ABCDEF, 4);
120 const int8x16_t vb89x01234567 = vshlq_n_s8(vb89x0123456789ABCDEF, 4);
121 const int8x16_t vbABx01234567 = vshlq_n_s8(vbABx0123456789ABCDEF, 4);
122 const int8x16_t vbCDx01234567 = vshlq_n_s8(vbCDx0123456789ABCDEF, 4);
123 const int8x16_t vbEFx01234567 = vshlq_n_s8(vbEFx0123456789ABCDEF, 4);
124 const int8x16_t vbGHx01234567 = vshlq_n_s8(vbGHx0123456789ABCDEF, 4);
125 const int8x16_t vbIJx01234567 = vshlq_n_s8(vbIJx0123456789ABCDEF, 4);
126 const int8x16_t vbKLx01234567 = vshlq_n_s8(vbKLx0123456789ABCDEF, 4);
127 const int8x16_t vbMNx01234567 = vshlq_n_s8(vbMNx0123456789ABCDEF, 4);
128 const int8x16_t vbOPx01234567 = vshlq_n_s8(vbOPx0123456789ABCDEF, 4);
129 const int8x16_t vbQRx01234567 = vshlq_n_s8(vbQRx0123456789ABCDEF, 4);
130 const int8x16_t vbSTx01234567 = vshlq_n_s8(vbSTx0123456789ABCDEF, 4);
131 const int8x16_t vbUVx01234567 = vshlq_n_s8(vbUVx0123456789ABCDEF, 4);
132 const int8x16_t vb01x89ABCDEF = vandq_s8(vb01x0123456789ABCDEF, vmask);
133 const int8x16_t vb23x89ABCDEF = vandq_s8(vb23x0123456789ABCDEF, vmask);
134 const int8x16_t vb45x89ABCDEF = vandq_s8(vb45x0123456789ABCDEF, vmask);
135 const int8x16_t vb67x89ABCDEF = vandq_s8(vb67x0123456789ABCDEF, vmask);
136 const int8x16_t vb89x89ABCDEF = vandq_s8(vb89x0123456789ABCDEF, vmask);
137 const int8x16_t vbABx89ABCDEF = vandq_s8(vbABx0123456789ABCDEF, vmask);
138 const int8x16_t vbCDx89ABCDEF = vandq_s8(vbCDx0123456789ABCDEF, vmask);
139 const int8x16_t vbEFx89ABCDEF = vandq_s8(vbEFx0123456789ABCDEF, vmask);
140 const int8x16_t vbGHx89ABCDEF = vandq_s8(vbGHx0123456789ABCDEF, vmask);
141 const int8x16_t vbIJx89ABCDEF = vandq_s8(vbIJx0123456789ABCDEF, vmask);
142 const int8x16_t vbKLx89ABCDEF = vandq_s8(vbKLx0123456789ABCDEF, vmask);
143 const int8x16_t vbMNx89ABCDEF = vandq_s8(vbMNx0123456789ABCDEF, vmask);
144 const int8x16_t vbOPx89ABCDEF = vandq_s8(vbOPx0123456789ABCDEF, vmask);
145 const int8x16_t vbQRx89ABCDEF = vandq_s8(vbQRx0123456789ABCDEF, vmask);
146 const int8x16_t vbSTx89ABCDEF = vandq_s8(vbSTx0123456789ABCDEF, vmask);
147 const int8x16_t vbUVx89ABCDEF = vandq_s8(vbUVx0123456789ABCDEF, vmask);
148
149 // Multiply-accumulate: 1x8 * 8x32 --> 1x32.
150 vacc01x01 = vmmlaq_s32(vacc01x01, va01x01234567, vb01x01234567);
151 vacc01x23 = vmmlaq_s32(vacc01x23, va01x01234567, vb23x01234567);
152 vacc01x45 = vmmlaq_s32(vacc01x45, va01x01234567, vb45x01234567);
153 vacc01x67 = vmmlaq_s32(vacc01x67, va01x01234567, vb67x01234567);
154 vacc01x89 = vmmlaq_s32(vacc01x89, va01x01234567, vb89x01234567);
155 vacc01xAB = vmmlaq_s32(vacc01xAB, va01x01234567, vbABx01234567);
156 vacc01xCD = vmmlaq_s32(vacc01xCD, va01x01234567, vbCDx01234567);
157 vacc01xEF = vmmlaq_s32(vacc01xEF, va01x01234567, vbEFx01234567);
158 vacc01xGH = vmmlaq_s32(vacc01xGH, va01x01234567, vbGHx01234567);
159 vacc01xIJ = vmmlaq_s32(vacc01xIJ, va01x01234567, vbIJx01234567);
160 vacc01xKL = vmmlaq_s32(vacc01xKL, va01x01234567, vbKLx01234567);
161 vacc01xMN = vmmlaq_s32(vacc01xMN, va01x01234567, vbMNx01234567);
162 vacc01xOP = vmmlaq_s32(vacc01xOP, va01x01234567, vbOPx01234567);
163 vacc01xQR = vmmlaq_s32(vacc01xQR, va01x01234567, vbQRx01234567);
164 vacc01xST = vmmlaq_s32(vacc01xST, va01x01234567, vbSTx01234567);
165 vacc01xUV = vmmlaq_s32(vacc01xUV, va01x01234567, vbUVx01234567);
166 vacc01x01 = vmmlaq_s32(vacc01x01, va01x89ABCDEF, vb01x89ABCDEF);
167 vacc01x23 = vmmlaq_s32(vacc01x23, va01x89ABCDEF, vb23x89ABCDEF);
168 vacc01x45 = vmmlaq_s32(vacc01x45, va01x89ABCDEF, vb45x89ABCDEF);
169 vacc01x67 = vmmlaq_s32(vacc01x67, va01x89ABCDEF, vb67x89ABCDEF);
170 vacc01x89 = vmmlaq_s32(vacc01x89, va01x89ABCDEF, vb89x89ABCDEF);
171 vacc01xAB = vmmlaq_s32(vacc01xAB, va01x89ABCDEF, vbABx89ABCDEF);
172 vacc01xCD = vmmlaq_s32(vacc01xCD, va01x89ABCDEF, vbCDx89ABCDEF);
173 vacc01xEF = vmmlaq_s32(vacc01xEF, va01x89ABCDEF, vbEFx89ABCDEF);
174 vacc01xGH = vmmlaq_s32(vacc01xGH, va01x89ABCDEF, vbGHx89ABCDEF);
175 vacc01xIJ = vmmlaq_s32(vacc01xIJ, va01x89ABCDEF, vbIJx89ABCDEF);
176 vacc01xKL = vmmlaq_s32(vacc01xKL, va01x89ABCDEF, vbKLx89ABCDEF);
177 vacc01xMN = vmmlaq_s32(vacc01xMN, va01x89ABCDEF, vbMNx89ABCDEF);
178 vacc01xOP = vmmlaq_s32(vacc01xOP, va01x89ABCDEF, vbOPx89ABCDEF);
179 vacc01xQR = vmmlaq_s32(vacc01xQR, va01x89ABCDEF, vbQRx89ABCDEF);
180 vacc01xST = vmmlaq_s32(vacc01xST, va01x89ABCDEF, vbSTx89ABCDEF);
181 vacc01xUV = vmmlaq_s32(vacc01xUV, va01x89ABCDEF, vbUVx89ABCDEF);
182
183 k -= 16 * sizeof(int8_t);
184 }
185 // Handle up to 8 final positions of `k`
186 if XNN_UNLIKELY(k != 0) {
187 // Load a 1x8 block of activations.
188 uint64x2_t va01x01234567 = vld1q_dup_u64((const void*) a0); a0 += 8;
189
190 // Load a 16x32 block of weights.
191 const int8x16_t vb01x0123456789ABCDEF = vld1q_s8(w); w = (const int8_t*) w + 16;
192 const int8x16_t vb23x0123456789ABCDEF = vld1q_s8(w); w = (const int8_t*) w + 16;
193 const int8x16_t vb45x0123456789ABCDEF = vld1q_s8(w); w = (const int8_t*) w + 16;
194 const int8x16_t vb67x0123456789ABCDEF = vld1q_s8(w); w = (const int8_t*) w + 16;
195 const int8x16_t vb89x0123456789ABCDEF = vld1q_s8(w); w = (const int8_t*) w + 16;
196 const int8x16_t vbABx0123456789ABCDEF = vld1q_s8(w); w = (const int8_t*) w + 16;
197 const int8x16_t vbCDx0123456789ABCDEF = vld1q_s8(w); w = (const int8_t*) w + 16;
198 const int8x16_t vbEFx0123456789ABCDEF = vld1q_s8(w); w = (const int8_t*) w + 16;
199 const int8x16_t vbGHx0123456789ABCDEF = vld1q_s8(w); w = (const int8_t*) w + 16;
200 const int8x16_t vbIJx0123456789ABCDEF = vld1q_s8(w); w = (const int8_t*) w + 16;
201 const int8x16_t vbKLx0123456789ABCDEF = vld1q_s8(w); w = (const int8_t*) w + 16;
202 const int8x16_t vbMNx0123456789ABCDEF = vld1q_s8(w); w = (const int8_t*) w + 16;
203 const int8x16_t vbOPx0123456789ABCDEF = vld1q_s8(w); w = (const int8_t*) w + 16;
204 const int8x16_t vbQRx0123456789ABCDEF = vld1q_s8(w); w = (const int8_t*) w + 16;
205 const int8x16_t vbSTx0123456789ABCDEF = vld1q_s8(w); w = (const int8_t*) w + 16;
206 const int8x16_t vbUVx0123456789ABCDEF = vld1q_s8(w); w = (const int8_t*) w + 16;
207 const int8x16_t vb01x01234567 = vshlq_n_s8(vb01x0123456789ABCDEF, 4);
208 const int8x16_t vb23x01234567 = vshlq_n_s8(vb23x0123456789ABCDEF, 4);
209 const int8x16_t vb45x01234567 = vshlq_n_s8(vb45x0123456789ABCDEF, 4);
210 const int8x16_t vb67x01234567 = vshlq_n_s8(vb67x0123456789ABCDEF, 4);
211 const int8x16_t vb89x01234567 = vshlq_n_s8(vb89x0123456789ABCDEF, 4);
212 const int8x16_t vbABx01234567 = vshlq_n_s8(vbABx0123456789ABCDEF, 4);
213 const int8x16_t vbCDx01234567 = vshlq_n_s8(vbCDx0123456789ABCDEF, 4);
214 const int8x16_t vbEFx01234567 = vshlq_n_s8(vbEFx0123456789ABCDEF, 4);
215 const int8x16_t vbGHx01234567 = vshlq_n_s8(vbGHx0123456789ABCDEF, 4);
216 const int8x16_t vbIJx01234567 = vshlq_n_s8(vbIJx0123456789ABCDEF, 4);
217 const int8x16_t vbKLx01234567 = vshlq_n_s8(vbKLx0123456789ABCDEF, 4);
218 const int8x16_t vbMNx01234567 = vshlq_n_s8(vbMNx0123456789ABCDEF, 4);
219 const int8x16_t vbOPx01234567 = vshlq_n_s8(vbOPx0123456789ABCDEF, 4);
220 const int8x16_t vbQRx01234567 = vshlq_n_s8(vbQRx0123456789ABCDEF, 4);
221 const int8x16_t vbSTx01234567 = vshlq_n_s8(vbSTx0123456789ABCDEF, 4);
222 const int8x16_t vbUVx01234567 = vshlq_n_s8(vbUVx0123456789ABCDEF, 4);
223
224 // Multiply-accumulate: 1x4 * 4x32 --> 1x32.
225 vacc01x01 = vmmlaq_s32(vacc01x01, vreinterpretq_s8_u64(va01x01234567), vb01x01234567);
226 vacc01x23 = vmmlaq_s32(vacc01x23, vreinterpretq_s8_u64(va01x01234567), vb23x01234567);
227 vacc01x45 = vmmlaq_s32(vacc01x45, vreinterpretq_s8_u64(va01x01234567), vb45x01234567);
228 vacc01x67 = vmmlaq_s32(vacc01x67, vreinterpretq_s8_u64(va01x01234567), vb67x01234567);
229 vacc01x89 = vmmlaq_s32(vacc01x89, vreinterpretq_s8_u64(va01x01234567), vb89x01234567);
230 vacc01xAB = vmmlaq_s32(vacc01xAB, vreinterpretq_s8_u64(va01x01234567), vbABx01234567);
231 vacc01xCD = vmmlaq_s32(vacc01xCD, vreinterpretq_s8_u64(va01x01234567), vbCDx01234567);
232 vacc01xEF = vmmlaq_s32(vacc01xEF, vreinterpretq_s8_u64(va01x01234567), vbEFx01234567);
233 vacc01xGH = vmmlaq_s32(vacc01xGH, vreinterpretq_s8_u64(va01x01234567), vbGHx01234567);
234 vacc01xIJ = vmmlaq_s32(vacc01xIJ, vreinterpretq_s8_u64(va01x01234567), vbIJx01234567);
235 vacc01xKL = vmmlaq_s32(vacc01xKL, vreinterpretq_s8_u64(va01x01234567), vbKLx01234567);
236 vacc01xMN = vmmlaq_s32(vacc01xMN, vreinterpretq_s8_u64(va01x01234567), vbMNx01234567);
237 vacc01xOP = vmmlaq_s32(vacc01xOP, vreinterpretq_s8_u64(va01x01234567), vbOPx01234567);
238 vacc01xQR = vmmlaq_s32(vacc01xQR, vreinterpretq_s8_u64(va01x01234567), vbQRx01234567);
239 vacc01xST = vmmlaq_s32(vacc01xST, vreinterpretq_s8_u64(va01x01234567), vbSTx01234567);
240 vacc01xUV = vmmlaq_s32(vacc01xUV, vreinterpretq_s8_u64(va01x01234567), vbUVx01234567);
241 }
242
243 int32x4_t vacc0x0123 = vreinterpretq_s32_u64(vtrn1q_u64(vreinterpretq_u64_s32(vacc01x01), vreinterpretq_u64_s32(vacc01x23)));
244 int32x4_t vacc0x4567 = vreinterpretq_s32_u64(vtrn1q_u64(vreinterpretq_u64_s32(vacc01x45), vreinterpretq_u64_s32(vacc01x67)));
245 int32x4_t vacc0x89AB = vreinterpretq_s32_u64(vtrn1q_u64(vreinterpretq_u64_s32(vacc01x89), vreinterpretq_u64_s32(vacc01xAB)));
246 int32x4_t vacc0xCDEF = vreinterpretq_s32_u64(vtrn1q_u64(vreinterpretq_u64_s32(vacc01xCD), vreinterpretq_u64_s32(vacc01xEF)));
247 int32x4_t vacc0xGHIJ = vreinterpretq_s32_u64(vtrn1q_u64(vreinterpretq_u64_s32(vacc01xGH), vreinterpretq_u64_s32(vacc01xIJ)));
248 int32x4_t vacc0xKLMN = vreinterpretq_s32_u64(vtrn1q_u64(vreinterpretq_u64_s32(vacc01xKL), vreinterpretq_u64_s32(vacc01xMN)));
249 int32x4_t vacc0xOPQR = vreinterpretq_s32_u64(vtrn1q_u64(vreinterpretq_u64_s32(vacc01xOP), vreinterpretq_u64_s32(vacc01xQR)));
250 int32x4_t vacc0xSTUV = vreinterpretq_s32_u64(vtrn1q_u64(vreinterpretq_u64_s32(vacc01xST), vreinterpretq_u64_s32(vacc01xUV)));
251 float32x4_t vout0x0123 = vcvtq_n_f32_s32(vacc0x0123, 4);
252 float32x4_t vout0x4567 = vcvtq_n_f32_s32(vacc0x4567, 4);
253 float32x4_t vout0x89AB = vcvtq_n_f32_s32(vacc0x89AB, 4);
254 float32x4_t vout0xCDEF = vcvtq_n_f32_s32(vacc0xCDEF, 4);
255 float32x4_t vout0xGHIJ = vcvtq_n_f32_s32(vacc0xGHIJ, 4);
256 float32x4_t vout0xKLMN = vcvtq_n_f32_s32(vacc0xKLMN, 4);
257 float32x4_t vout0xOPQR = vcvtq_n_f32_s32(vacc0xOPQR, 4);
258 float32x4_t vout0xSTUV = vcvtq_n_f32_s32(vacc0xSTUV, 4);
259
260 const float32x4_t vinput_scale0 = vld1q_dup_f32(&quantization_params[0].inv_scale);
261 vout0x0123 = vmulq_f32(vout0x0123, vinput_scale0);
262 vout0x4567 = vmulq_f32(vout0x4567, vinput_scale0);
263 vout0x89AB = vmulq_f32(vout0x89AB, vinput_scale0);
264 vout0xCDEF = vmulq_f32(vout0xCDEF, vinput_scale0);
265 vout0xGHIJ = vmulq_f32(vout0xGHIJ, vinput_scale0);
266 vout0xKLMN = vmulq_f32(vout0xKLMN, vinput_scale0);
267 vout0xOPQR = vmulq_f32(vout0xOPQR, vinput_scale0);
268 vout0xSTUV = vmulq_f32(vout0xSTUV, vinput_scale0);
269
270 const float32x4_t vfilter_output_scale0123 = vld1q_f32(w); w = (const float*) w + 4;
271 const float32x4_t vfilter_output_scale4567 = vld1q_f32(w); w = (const float*) w + 4;
272 const float32x4_t vfilter_output_scale89AB = vld1q_f32(w); w = (const float*) w + 4;
273 const float32x4_t vfilter_output_scaleCDEF = vld1q_f32(w); w = (const float*) w + 4;
274 const float32x4_t vfilter_output_scaleGHIJ = vld1q_f32(w); w = (const float*) w + 4;
275 const float32x4_t vfilter_output_scaleKLMN = vld1q_f32(w); w = (const float*) w + 4;
276 const float32x4_t vfilter_output_scaleOPQR = vld1q_f32(w); w = (const float*) w + 4;
277 const float32x4_t vfilter_output_scaleSTUV = vld1q_f32(w); w = (const float*) w + 4;
278
279 const float32x4_t vbias0123 = vld1q_f32(w); w = (const float*) w + 4;
280 vout0x0123 = vfmaq_f32(vbias0123, vout0x0123, vfilter_output_scale0123);
281 const float32x4_t vbias4567 = vld1q_f32(w); w = (const float*) w + 4;
282 vout0x4567 = vfmaq_f32(vbias4567, vout0x4567, vfilter_output_scale4567);
283 const float32x4_t vbias89AB = vld1q_f32(w); w = (const float*) w + 4;
284 vout0x89AB = vfmaq_f32(vbias89AB, vout0x89AB, vfilter_output_scale89AB);
285 const float32x4_t vbiasCDEF = vld1q_f32(w); w = (const float*) w + 4;
286 vout0xCDEF = vfmaq_f32(vbiasCDEF, vout0xCDEF, vfilter_output_scaleCDEF);
287 const float32x4_t vbiasGHIJ = vld1q_f32(w); w = (const float*) w + 4;
288 vout0xGHIJ = vfmaq_f32(vbiasGHIJ, vout0xGHIJ, vfilter_output_scaleGHIJ);
289 const float32x4_t vbiasKLMN = vld1q_f32(w); w = (const float*) w + 4;
290 vout0xKLMN = vfmaq_f32(vbiasKLMN, vout0xKLMN, vfilter_output_scaleKLMN);
291 const float32x4_t vbiasOPQR = vld1q_f32(w); w = (const float*) w + 4;
292 vout0xOPQR = vfmaq_f32(vbiasOPQR, vout0xOPQR, vfilter_output_scaleOPQR);
293 const float32x4_t vbiasSTUV = vld1q_f32(w); w = (const float*) w + 4;
294 vout0xSTUV = vfmaq_f32(vbiasSTUV, vout0xSTUV, vfilter_output_scaleSTUV);
295
296 float16x8_t vfp16out0x01234567 = vcombine_f16(vcvt_f16_f32(vout0x0123), vcvt_f16_f32(vout0x4567));
297 float16x8_t vfp16out0x89ABCDEF = vcombine_f16(vcvt_f16_f32(vout0x89AB), vcvt_f16_f32(vout0xCDEF));
298 float16x8_t vfp16out0xGHIJKLMN = vcombine_f16(vcvt_f16_f32(vout0xGHIJ), vcvt_f16_f32(vout0xKLMN));
299 float16x8_t vfp16out0xOPQRSTUV = vcombine_f16(vcvt_f16_f32(vout0xOPQR), vcvt_f16_f32(vout0xSTUV));
300
301 const float16x8_t voutput_min = vreinterpretq_f16_u16(vld1q_dup_u16((const uint16_t*) &params->scalar.min));
302 vfp16out0x01234567 = vmaxq_f16(vfp16out0x01234567, voutput_min);
303 vfp16out0x89ABCDEF = vmaxq_f16(vfp16out0x89ABCDEF, voutput_min);
304 vfp16out0xGHIJKLMN = vmaxq_f16(vfp16out0xGHIJKLMN, voutput_min);
305 vfp16out0xOPQRSTUV = vmaxq_f16(vfp16out0xOPQRSTUV, voutput_min);
306 const float16x8_t voutput_max = vreinterpretq_f16_u16(vld1q_dup_u16((const uint16_t*) &params->scalar.max));
307 vfp16out0x01234567 = vminq_f16(vfp16out0x01234567, voutput_max);
308 vfp16out0x89ABCDEF = vminq_f16(vfp16out0x89ABCDEF, voutput_max);
309 vfp16out0xGHIJKLMN = vminq_f16(vfp16out0xGHIJKLMN, voutput_max);
310 vfp16out0xOPQRSTUV = vminq_f16(vfp16out0xOPQRSTUV, voutput_max);
311 if XNN_LIKELY(nc >= 32) {
312 vst1q_u16(c0, vreinterpretq_u16_f16(vfp16out0x01234567));
313 vst1q_u16(c0 + 8, vreinterpretq_u16_f16(vfp16out0x89ABCDEF));
314 vst1q_u16(c0 + 16, vreinterpretq_u16_f16(vfp16out0xGHIJKLMN));
315 vst1q_u16(c0 + 24, vreinterpretq_u16_f16(vfp16out0xOPQRSTUV));
316
317 a0 = (const int8_t*) ((uintptr_t) a0 - kc);
318
319 c0 = (uint16_t*) ((uintptr_t) c0 + cn_stride);
320
321 nc -= 32;
322 } else {
323 if (nc & 16) {
324 vst1q_u16(c0, vreinterpretq_u16_f16(vfp16out0x01234567)); c0 += 8;
325 vfp16out0x01234567 = vfp16out0xGHIJKLMN;
326 vst1q_u16(c0, vreinterpretq_u16_f16(vfp16out0x89ABCDEF)); c0 += 8;
327 vfp16out0x89ABCDEF = vfp16out0xOPQRSTUV;
328 }
329 if (nc & 8) {
330 vst1q_u16(c0, vreinterpretq_u16_f16(vfp16out0x01234567)); c0 += 8;
331 vfp16out0x01234567 = vfp16out0x89ABCDEF;
332 }
333 float16x4_t vfp16out0x0123 = vget_low_f16(vfp16out0x01234567);
334 if (nc & 4) {
335 vst1_u16(c0, vreinterpret_u16_f16(vfp16out0x0123)); c0 += 4;
336 vfp16out0x0123 = vget_high_f16(vfp16out0x01234567);
337 }
338 if (nc & 2) {
339 vst1_lane_u32((void*) c0, vreinterpret_u32_f16(vfp16out0x0123), 0); c0 += 2;
340 vfp16out0x0123 = vext_f16(vfp16out0x0123, vfp16out0x0123, 2);
341 }
342 if (nc & 1) {
343 vst1_lane_u16(c0, vreinterpret_u16_f16(vfp16out0x0123), 0);
344 }
345 nc = 0;
346 }
347 } while (nc != 0);
348 }