Debian Codegraph @

src/qd8-f16-qc4w-gemm/gen/qd8-f16-qc4w-gemm-2x32c8-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_2x32c8__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 <= 2);
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 int8x16_t vmask = vmovq_n_s8(INT8_C(0xF0));
50
51 // Loop over groups of 32 columns.
52 do {
53 // Initialize accumulators with bias. 32 bias values are loaded from the
54 // weight matrix, at the start of the group of 32 columns.
55 const int32x4_t vinput_zero_point01 = vld1q_s32(&quantization_params[0].zero_point);
56 const int32x4_t vksum0123 = vld1q_s32(w); w = (const int32_t*) w + 4;
57 const int32x4_t vksumzp0x0123 = vmulq_lane_s32(vksum0123, vget_low_s32(vinput_zero_point01), 0);
58 const int32x4_t vksumzp1x0123 = vmulq_lane_s32(vksum0123, vget_high_s32(vinput_zero_point01), 0);
59 const int32x4_t vksum4567 = vld1q_s32(w); w = (const int32_t*) w + 4;
60 const int32x4_t vksumzp0x4567 = vmulq_lane_s32(vksum4567, vget_low_s32(vinput_zero_point01), 0);
61 const int32x4_t vksumzp1x4567 = vmulq_lane_s32(vksum4567, vget_high_s32(vinput_zero_point01), 0);
62 const int32x4_t vksum89AB = vld1q_s32(w); w = (const int32_t*) w + 4;
63 const int32x4_t vksumzp0x89AB = vmulq_lane_s32(vksum89AB, vget_low_s32(vinput_zero_point01), 0);
64 const int32x4_t vksumzp1x89AB = vmulq_lane_s32(vksum89AB, vget_high_s32(vinput_zero_point01), 0);
65 const int32x4_t vksumCDEF = vld1q_s32(w); w = (const int32_t*) w + 4;
66 const int32x4_t vksumzp0xCDEF = vmulq_lane_s32(vksumCDEF, vget_low_s32(vinput_zero_point01), 0);
67 const int32x4_t vksumzp1xCDEF = vmulq_lane_s32(vksumCDEF, vget_high_s32(vinput_zero_point01), 0);
68 const int32x4_t vksumGHIJ = vld1q_s32(w); w = (const int32_t*) w + 4;
69 const int32x4_t vksumzp0xGHIJ = vmulq_lane_s32(vksumGHIJ, vget_low_s32(vinput_zero_point01), 0);
70 const int32x4_t vksumzp1xGHIJ = vmulq_lane_s32(vksumGHIJ, vget_high_s32(vinput_zero_point01), 0);
71 const int32x4_t vksumKLMN = vld1q_s32(w); w = (const int32_t*) w + 4;
72 const int32x4_t vksumzp0xKLMN = vmulq_lane_s32(vksumKLMN, vget_low_s32(vinput_zero_point01), 0);
73 const int32x4_t vksumzp1xKLMN = vmulq_lane_s32(vksumKLMN, vget_high_s32(vinput_zero_point01), 0);
74 const int32x4_t vksumOPQR = vld1q_s32(w); w = (const int32_t*) w + 4;
75 const int32x4_t vksumzp0xOPQR = vmulq_lane_s32(vksumOPQR, vget_low_s32(vinput_zero_point01), 0);
76 const int32x4_t vksumzp1xOPQR = vmulq_lane_s32(vksumOPQR, vget_high_s32(vinput_zero_point01), 0);
77 const int32x4_t vksumSTUV = vld1q_s32(w); w = (const int32_t*) w + 4;
78 const int32x4_t vksumzp0xSTUV = vmulq_lane_s32(vksumSTUV, vget_low_s32(vinput_zero_point01), 0);
79 const int32x4_t vksumzp1xSTUV = vmulq_lane_s32(vksumSTUV, vget_high_s32(vinput_zero_point01), 0);
80
81 int32x4_t vacc01x01 = vreinterpretq_s32_u64(vtrn1q_u64(vreinterpretq_u64_s32(vksumzp0x0123), vreinterpretq_u64_s32(vksumzp1x0123)));
82 int32x4_t vacc01x23 = vreinterpretq_s32_u64(vtrn2q_u64(vreinterpretq_u64_s32(vksumzp0x0123), vreinterpretq_u64_s32(vksumzp1x0123)));
83 int32x4_t vacc01x45 = vreinterpretq_s32_u64(vtrn1q_u64(vreinterpretq_u64_s32(vksumzp0x4567), vreinterpretq_u64_s32(vksumzp1x4567)));
84 int32x4_t vacc01x67 = vreinterpretq_s32_u64(vtrn2q_u64(vreinterpretq_u64_s32(vksumzp0x4567), vreinterpretq_u64_s32(vksumzp1x4567)));
85 int32x4_t vacc01x89 = vreinterpretq_s32_u64(vtrn1q_u64(vreinterpretq_u64_s32(vksumzp0x89AB), vreinterpretq_u64_s32(vksumzp1x89AB)));
86 int32x4_t vacc01xAB = vreinterpretq_s32_u64(vtrn2q_u64(vreinterpretq_u64_s32(vksumzp0x89AB), vreinterpretq_u64_s32(vksumzp1x89AB)));
87 int32x4_t vacc01xCD = vreinterpretq_s32_u64(vtrn1q_u64(vreinterpretq_u64_s32(vksumzp0xCDEF), vreinterpretq_u64_s32(vksumzp1xCDEF)));
88 int32x4_t vacc01xEF = vreinterpretq_s32_u64(vtrn2q_u64(vreinterpretq_u64_s32(vksumzp0xCDEF), vreinterpretq_u64_s32(vksumzp1xCDEF)));
89 int32x4_t vacc01xGH = vreinterpretq_s32_u64(vtrn1q_u64(vreinterpretq_u64_s32(vksumzp0xGHIJ), vreinterpretq_u64_s32(vksumzp1xGHIJ)));
90 int32x4_t vacc01xIJ = vreinterpretq_s32_u64(vtrn2q_u64(vreinterpretq_u64_s32(vksumzp0xGHIJ), vreinterpretq_u64_s32(vksumzp1xGHIJ)));
91 int32x4_t vacc01xKL = vreinterpretq_s32_u64(vtrn1q_u64(vreinterpretq_u64_s32(vksumzp0xKLMN), vreinterpretq_u64_s32(vksumzp1xKLMN)));
92 int32x4_t vacc01xMN = vreinterpretq_s32_u64(vtrn2q_u64(vreinterpretq_u64_s32(vksumzp0xKLMN), vreinterpretq_u64_s32(vksumzp1xKLMN)));
93 int32x4_t vacc01xOP = vreinterpretq_s32_u64(vtrn1q_u64(vreinterpretq_u64_s32(vksumzp0xOPQR), vreinterpretq_u64_s32(vksumzp1xOPQR)));
94 int32x4_t vacc01xQR = vreinterpretq_s32_u64(vtrn2q_u64(vreinterpretq_u64_s32(vksumzp0xOPQR), vreinterpretq_u64_s32(vksumzp1xOPQR)));
95 int32x4_t vacc01xST = vreinterpretq_s32_u64(vtrn1q_u64(vreinterpretq_u64_s32(vksumzp0xSTUV), vreinterpretq_u64_s32(vksumzp1xSTUV)));
96 int32x4_t vacc01xUV = vreinterpretq_s32_u64(vtrn2q_u64(vreinterpretq_u64_s32(vksumzp0xSTUV), vreinterpretq_u64_s32(vksumzp1xSTUV)));
97
98 // Inner accumulation loop along the 32 columns.
99 size_t k = kc;
100 // 2x partial unrolled loop to load 8 bytes at a time.
101
102 uint64x2x2_t va01x0123456789ABCDEF;
103 va01x0123456789ABCDEF.val[0] = vdupq_n_u64(0);
104 va01x0123456789ABCDEF.val[1] = vdupq_n_u64(0);
105
106 while (k >= 16 * sizeof(int8_t)) {
107 // Load a 2x16 block of activations.
108 va01x0123456789ABCDEF = vld2q_lane_u64((const void*) a0, va01x0123456789ABCDEF, 0); a0 += 16;
109 va01x0123456789ABCDEF = vld2q_lane_u64((const void*) a1, va01x0123456789ABCDEF, 1); a1 += 16;
110
111 const int8x16_t va01x01234567 = vreinterpretq_s8_u64(va01x0123456789ABCDEF.val[0]);
112 const int8x16_t va01x89ABCDEF = vreinterpretq_s8_u64(va01x0123456789ABCDEF.val[1]);
113
114 // Load a 16x32 block of weights.
115 const int8x16_t vb01x0123456789ABCDEF = vld1q_s8(w); w = (const int8_t*) w + 16;
116 const int8x16_t vb23x0123456789ABCDEF = vld1q_s8(w); w = (const int8_t*) w + 16;
117 const int8x16_t vb45x0123456789ABCDEF = vld1q_s8(w); w = (const int8_t*) w + 16;
118 const int8x16_t vb67x0123456789ABCDEF = vld1q_s8(w); w = (const int8_t*) w + 16;
119 const int8x16_t vb89x0123456789ABCDEF = vld1q_s8(w); w = (const int8_t*) w + 16;
120 const int8x16_t vbABx0123456789ABCDEF = vld1q_s8(w); w = (const int8_t*) w + 16;
121 const int8x16_t vbCDx0123456789ABCDEF = vld1q_s8(w); w = (const int8_t*) w + 16;
122 const int8x16_t vbEFx0123456789ABCDEF = vld1q_s8(w); w = (const int8_t*) w + 16;
123 const int8x16_t vbGHx0123456789ABCDEF = vld1q_s8(w); w = (const int8_t*) w + 16;
124 const int8x16_t vbIJx0123456789ABCDEF = vld1q_s8(w); w = (const int8_t*) w + 16;
125 const int8x16_t vbKLx0123456789ABCDEF = vld1q_s8(w); w = (const int8_t*) w + 16;
126 const int8x16_t vbMNx0123456789ABCDEF = vld1q_s8(w); w = (const int8_t*) w + 16;
127 const int8x16_t vbOPx0123456789ABCDEF = vld1q_s8(w); w = (const int8_t*) w + 16;
128 const int8x16_t vbQRx0123456789ABCDEF = vld1q_s8(w); w = (const int8_t*) w + 16;
129 const int8x16_t vbSTx0123456789ABCDEF = vld1q_s8(w); w = (const int8_t*) w + 16;
130 const int8x16_t vbUVx0123456789ABCDEF = vld1q_s8(w); w = (const int8_t*) w + 16;
131 const int8x16_t vb01x01234567 = vshlq_n_s8(vb01x0123456789ABCDEF, 4);
132 const int8x16_t vb23x01234567 = vshlq_n_s8(vb23x0123456789ABCDEF, 4);
133 const int8x16_t vb45x01234567 = vshlq_n_s8(vb45x0123456789ABCDEF, 4);
134 const int8x16_t vb67x01234567 = vshlq_n_s8(vb67x0123456789ABCDEF, 4);
135 const int8x16_t vb89x01234567 = vshlq_n_s8(vb89x0123456789ABCDEF, 4);
136 const int8x16_t vbABx01234567 = vshlq_n_s8(vbABx0123456789ABCDEF, 4);
137 const int8x16_t vbCDx01234567 = vshlq_n_s8(vbCDx0123456789ABCDEF, 4);
138 const int8x16_t vbEFx01234567 = vshlq_n_s8(vbEFx0123456789ABCDEF, 4);
139 const int8x16_t vbGHx01234567 = vshlq_n_s8(vbGHx0123456789ABCDEF, 4);
140 const int8x16_t vbIJx01234567 = vshlq_n_s8(vbIJx0123456789ABCDEF, 4);
141 const int8x16_t vbKLx01234567 = vshlq_n_s8(vbKLx0123456789ABCDEF, 4);
142 const int8x16_t vbMNx01234567 = vshlq_n_s8(vbMNx0123456789ABCDEF, 4);
143 const int8x16_t vbOPx01234567 = vshlq_n_s8(vbOPx0123456789ABCDEF, 4);
144 const int8x16_t vbQRx01234567 = vshlq_n_s8(vbQRx0123456789ABCDEF, 4);
145 const int8x16_t vbSTx01234567 = vshlq_n_s8(vbSTx0123456789ABCDEF, 4);
146 const int8x16_t vbUVx01234567 = vshlq_n_s8(vbUVx0123456789ABCDEF, 4);
147 const int8x16_t vb01x89ABCDEF = vandq_s8(vb01x0123456789ABCDEF, vmask);
148 const int8x16_t vb23x89ABCDEF = vandq_s8(vb23x0123456789ABCDEF, vmask);
149 const int8x16_t vb45x89ABCDEF = vandq_s8(vb45x0123456789ABCDEF, vmask);
150 const int8x16_t vb67x89ABCDEF = vandq_s8(vb67x0123456789ABCDEF, vmask);
151 const int8x16_t vb89x89ABCDEF = vandq_s8(vb89x0123456789ABCDEF, vmask);
152 const int8x16_t vbABx89ABCDEF = vandq_s8(vbABx0123456789ABCDEF, vmask);
153 const int8x16_t vbCDx89ABCDEF = vandq_s8(vbCDx0123456789ABCDEF, vmask);
154 const int8x16_t vbEFx89ABCDEF = vandq_s8(vbEFx0123456789ABCDEF, vmask);
155 const int8x16_t vbGHx89ABCDEF = vandq_s8(vbGHx0123456789ABCDEF, vmask);
156 const int8x16_t vbIJx89ABCDEF = vandq_s8(vbIJx0123456789ABCDEF, vmask);
157 const int8x16_t vbKLx89ABCDEF = vandq_s8(vbKLx0123456789ABCDEF, vmask);
158 const int8x16_t vbMNx89ABCDEF = vandq_s8(vbMNx0123456789ABCDEF, vmask);
159 const int8x16_t vbOPx89ABCDEF = vandq_s8(vbOPx0123456789ABCDEF, vmask);
160 const int8x16_t vbQRx89ABCDEF = vandq_s8(vbQRx0123456789ABCDEF, vmask);
161 const int8x16_t vbSTx89ABCDEF = vandq_s8(vbSTx0123456789ABCDEF, vmask);
162 const int8x16_t vbUVx89ABCDEF = vandq_s8(vbUVx0123456789ABCDEF, vmask);
163
164 // Multiply-accumulate: 2x8 * 8x32 --> 2x32.
165 vacc01x01 = vmmlaq_s32(vacc01x01, va01x01234567, vb01x01234567);
166 vacc01x23 = vmmlaq_s32(vacc01x23, va01x01234567, vb23x01234567);
167 vacc01x45 = vmmlaq_s32(vacc01x45, va01x01234567, vb45x01234567);
168 vacc01x67 = vmmlaq_s32(vacc01x67, va01x01234567, vb67x01234567);
169 vacc01x89 = vmmlaq_s32(vacc01x89, va01x01234567, vb89x01234567);
170 vacc01xAB = vmmlaq_s32(vacc01xAB, va01x01234567, vbABx01234567);
171 vacc01xCD = vmmlaq_s32(vacc01xCD, va01x01234567, vbCDx01234567);
172 vacc01xEF = vmmlaq_s32(vacc01xEF, va01x01234567, vbEFx01234567);
173 vacc01xGH = vmmlaq_s32(vacc01xGH, va01x01234567, vbGHx01234567);
174 vacc01xIJ = vmmlaq_s32(vacc01xIJ, va01x01234567, vbIJx01234567);
175 vacc01xKL = vmmlaq_s32(vacc01xKL, va01x01234567, vbKLx01234567);
176 vacc01xMN = vmmlaq_s32(vacc01xMN, va01x01234567, vbMNx01234567);
177 vacc01xOP = vmmlaq_s32(vacc01xOP, va01x01234567, vbOPx01234567);
178 vacc01xQR = vmmlaq_s32(vacc01xQR, va01x01234567, vbQRx01234567);
179 vacc01xST = vmmlaq_s32(vacc01xST, va01x01234567, vbSTx01234567);
180 vacc01xUV = vmmlaq_s32(vacc01xUV, va01x01234567, vbUVx01234567);
181 vacc01x01 = vmmlaq_s32(vacc01x01, va01x89ABCDEF, vb01x89ABCDEF);
182 vacc01x23 = vmmlaq_s32(vacc01x23, va01x89ABCDEF, vb23x89ABCDEF);
183 vacc01x45 = vmmlaq_s32(vacc01x45, va01x89ABCDEF, vb45x89ABCDEF);
184 vacc01x67 = vmmlaq_s32(vacc01x67, va01x89ABCDEF, vb67x89ABCDEF);
185 vacc01x89 = vmmlaq_s32(vacc01x89, va01x89ABCDEF, vb89x89ABCDEF);
186 vacc01xAB = vmmlaq_s32(vacc01xAB, va01x89ABCDEF, vbABx89ABCDEF);
187 vacc01xCD = vmmlaq_s32(vacc01xCD, va01x89ABCDEF, vbCDx89ABCDEF);
188 vacc01xEF = vmmlaq_s32(vacc01xEF, va01x89ABCDEF, vbEFx89ABCDEF);
189 vacc01xGH = vmmlaq_s32(vacc01xGH, va01x89ABCDEF, vbGHx89ABCDEF);
190 vacc01xIJ = vmmlaq_s32(vacc01xIJ, va01x89ABCDEF, vbIJx89ABCDEF);
191 vacc01xKL = vmmlaq_s32(vacc01xKL, va01x89ABCDEF, vbKLx89ABCDEF);
192 vacc01xMN = vmmlaq_s32(vacc01xMN, va01x89ABCDEF, vbMNx89ABCDEF);
193 vacc01xOP = vmmlaq_s32(vacc01xOP, va01x89ABCDEF, vbOPx89ABCDEF);
194 vacc01xQR = vmmlaq_s32(vacc01xQR, va01x89ABCDEF, vbQRx89ABCDEF);
195 vacc01xST = vmmlaq_s32(vacc01xST, va01x89ABCDEF, vbSTx89ABCDEF);
196 vacc01xUV = vmmlaq_s32(vacc01xUV, va01x89ABCDEF, vbUVx89ABCDEF);
197
198 k -= 16 * sizeof(int8_t);
199 }
200 // Handle up to 8 final positions of `k`
201 if XNN_UNLIKELY(k != 0) {
202 // Load a 2x8 block of activations.
203 uint64x2_t va01x01234567 = vld1q_dup_u64((const void*) a0); a0 += 8;
204 va01x01234567 = vld1q_lane_u64((const void*) a1, va01x01234567, 1); a1 += 8;
205
206 // Load a 16x32 block of weights.
207 const int8x16_t vb01x0123456789ABCDEF = vld1q_s8(w); w = (const int8_t*) w + 16;
208 const int8x16_t vb23x0123456789ABCDEF = vld1q_s8(w); w = (const int8_t*) w + 16;
209 const int8x16_t vb45x0123456789ABCDEF = vld1q_s8(w); w = (const int8_t*) w + 16;
210 const int8x16_t vb67x0123456789ABCDEF = vld1q_s8(w); w = (const int8_t*) w + 16;
211 const int8x16_t vb89x0123456789ABCDEF = vld1q_s8(w); w = (const int8_t*) w + 16;
212 const int8x16_t vbABx0123456789ABCDEF = vld1q_s8(w); w = (const int8_t*) w + 16;
213 const int8x16_t vbCDx0123456789ABCDEF = vld1q_s8(w); w = (const int8_t*) w + 16;
214 const int8x16_t vbEFx0123456789ABCDEF = vld1q_s8(w); w = (const int8_t*) w + 16;
215 const int8x16_t vbGHx0123456789ABCDEF = vld1q_s8(w); w = (const int8_t*) w + 16;
216 const int8x16_t vbIJx0123456789ABCDEF = vld1q_s8(w); w = (const int8_t*) w + 16;
217 const int8x16_t vbKLx0123456789ABCDEF = vld1q_s8(w); w = (const int8_t*) w + 16;
218 const int8x16_t vbMNx0123456789ABCDEF = vld1q_s8(w); w = (const int8_t*) w + 16;
219 const int8x16_t vbOPx0123456789ABCDEF = vld1q_s8(w); w = (const int8_t*) w + 16;
220 const int8x16_t vbQRx0123456789ABCDEF = vld1q_s8(w); w = (const int8_t*) w + 16;
221 const int8x16_t vbSTx0123456789ABCDEF = vld1q_s8(w); w = (const int8_t*) w + 16;
222 const int8x16_t vbUVx0123456789ABCDEF = vld1q_s8(w); w = (const int8_t*) w + 16;
223 const int8x16_t vb01x01234567 = vshlq_n_s8(vb01x0123456789ABCDEF, 4);
224 const int8x16_t vb23x01234567 = vshlq_n_s8(vb23x0123456789ABCDEF, 4);
225 const int8x16_t vb45x01234567 = vshlq_n_s8(vb45x0123456789ABCDEF, 4);
226 const int8x16_t vb67x01234567 = vshlq_n_s8(vb67x0123456789ABCDEF, 4);
227 const int8x16_t vb89x01234567 = vshlq_n_s8(vb89x0123456789ABCDEF, 4);
228 const int8x16_t vbABx01234567 = vshlq_n_s8(vbABx0123456789ABCDEF, 4);
229 const int8x16_t vbCDx01234567 = vshlq_n_s8(vbCDx0123456789ABCDEF, 4);
230 const int8x16_t vbEFx01234567 = vshlq_n_s8(vbEFx0123456789ABCDEF, 4);
231 const int8x16_t vbGHx01234567 = vshlq_n_s8(vbGHx0123456789ABCDEF, 4);
232 const int8x16_t vbIJx01234567 = vshlq_n_s8(vbIJx0123456789ABCDEF, 4);
233 const int8x16_t vbKLx01234567 = vshlq_n_s8(vbKLx0123456789ABCDEF, 4);
234 const int8x16_t vbMNx01234567 = vshlq_n_s8(vbMNx0123456789ABCDEF, 4);
235 const int8x16_t vbOPx01234567 = vshlq_n_s8(vbOPx0123456789ABCDEF, 4);
236 const int8x16_t vbQRx01234567 = vshlq_n_s8(vbQRx0123456789ABCDEF, 4);
237 const int8x16_t vbSTx01234567 = vshlq_n_s8(vbSTx0123456789ABCDEF, 4);
238 const int8x16_t vbUVx01234567 = vshlq_n_s8(vbUVx0123456789ABCDEF, 4);
239
240 // Multiply-accumulate: 2x4 * 4x32 --> 2x32.
241 vacc01x01 = vmmlaq_s32(vacc01x01, vreinterpretq_s8_u64(va01x01234567), vb01x01234567);
242 vacc01x23 = vmmlaq_s32(vacc01x23, vreinterpretq_s8_u64(va01x01234567), vb23x01234567);
243 vacc01x45 = vmmlaq_s32(vacc01x45, vreinterpretq_s8_u64(va01x01234567), vb45x01234567);
244 vacc01x67 = vmmlaq_s32(vacc01x67, vreinterpretq_s8_u64(va01x01234567), vb67x01234567);
245 vacc01x89 = vmmlaq_s32(vacc01x89, vreinterpretq_s8_u64(va01x01234567), vb89x01234567);
246 vacc01xAB = vmmlaq_s32(vacc01xAB, vreinterpretq_s8_u64(va01x01234567), vbABx01234567);
247 vacc01xCD = vmmlaq_s32(vacc01xCD, vreinterpretq_s8_u64(va01x01234567), vbCDx01234567);
248 vacc01xEF = vmmlaq_s32(vacc01xEF, vreinterpretq_s8_u64(va01x01234567), vbEFx01234567);
249 vacc01xGH = vmmlaq_s32(vacc01xGH, vreinterpretq_s8_u64(va01x01234567), vbGHx01234567);
250 vacc01xIJ = vmmlaq_s32(vacc01xIJ, vreinterpretq_s8_u64(va01x01234567), vbIJx01234567);
251 vacc01xKL = vmmlaq_s32(vacc01xKL, vreinterpretq_s8_u64(va01x01234567), vbKLx01234567);
252 vacc01xMN = vmmlaq_s32(vacc01xMN, vreinterpretq_s8_u64(va01x01234567), vbMNx01234567);
253 vacc01xOP = vmmlaq_s32(vacc01xOP, vreinterpretq_s8_u64(va01x01234567), vbOPx01234567);
254 vacc01xQR = vmmlaq_s32(vacc01xQR, vreinterpretq_s8_u64(va01x01234567), vbQRx01234567);
255 vacc01xST = vmmlaq_s32(vacc01xST, vreinterpretq_s8_u64(va01x01234567), vbSTx01234567);
256 vacc01xUV = vmmlaq_s32(vacc01xUV, vreinterpretq_s8_u64(va01x01234567), vbUVx01234567);
257 }
258
259 int32x4_t vacc0x0123 = vreinterpretq_s32_u64(vtrn1q_u64(vreinterpretq_u64_s32(vacc01x01), vreinterpretq_u64_s32(vacc01x23)));
260 int32x4_t vacc1x0123 = vreinterpretq_s32_u64(vtrn2q_u64(vreinterpretq_u64_s32(vacc01x01), vreinterpretq_u64_s32(vacc01x23)));
261 int32x4_t vacc0x4567 = vreinterpretq_s32_u64(vtrn1q_u64(vreinterpretq_u64_s32(vacc01x45), vreinterpretq_u64_s32(vacc01x67)));
262 int32x4_t vacc1x4567 = vreinterpretq_s32_u64(vtrn2q_u64(vreinterpretq_u64_s32(vacc01x45), vreinterpretq_u64_s32(vacc01x67)));
263 int32x4_t vacc0x89AB = vreinterpretq_s32_u64(vtrn1q_u64(vreinterpretq_u64_s32(vacc01x89), vreinterpretq_u64_s32(vacc01xAB)));
264 int32x4_t vacc1x89AB = vreinterpretq_s32_u64(vtrn2q_u64(vreinterpretq_u64_s32(vacc01x89), vreinterpretq_u64_s32(vacc01xAB)));
265 int32x4_t vacc0xCDEF = vreinterpretq_s32_u64(vtrn1q_u64(vreinterpretq_u64_s32(vacc01xCD), vreinterpretq_u64_s32(vacc01xEF)));
266 int32x4_t vacc1xCDEF = vreinterpretq_s32_u64(vtrn2q_u64(vreinterpretq_u64_s32(vacc01xCD), vreinterpretq_u64_s32(vacc01xEF)));
267 int32x4_t vacc0xGHIJ = vreinterpretq_s32_u64(vtrn1q_u64(vreinterpretq_u64_s32(vacc01xGH), vreinterpretq_u64_s32(vacc01xIJ)));
268 int32x4_t vacc1xGHIJ = vreinterpretq_s32_u64(vtrn2q_u64(vreinterpretq_u64_s32(vacc01xGH), vreinterpretq_u64_s32(vacc01xIJ)));
269 int32x4_t vacc0xKLMN = vreinterpretq_s32_u64(vtrn1q_u64(vreinterpretq_u64_s32(vacc01xKL), vreinterpretq_u64_s32(vacc01xMN)));
270 int32x4_t vacc1xKLMN = vreinterpretq_s32_u64(vtrn2q_u64(vreinterpretq_u64_s32(vacc01xKL), vreinterpretq_u64_s32(vacc01xMN)));
271 int32x4_t vacc0xOPQR = vreinterpretq_s32_u64(vtrn1q_u64(vreinterpretq_u64_s32(vacc01xOP), vreinterpretq_u64_s32(vacc01xQR)));
272 int32x4_t vacc1xOPQR = vreinterpretq_s32_u64(vtrn2q_u64(vreinterpretq_u64_s32(vacc01xOP), vreinterpretq_u64_s32(vacc01xQR)));
273 int32x4_t vacc0xSTUV = vreinterpretq_s32_u64(vtrn1q_u64(vreinterpretq_u64_s32(vacc01xST), vreinterpretq_u64_s32(vacc01xUV)));
274 int32x4_t vacc1xSTUV = vreinterpretq_s32_u64(vtrn2q_u64(vreinterpretq_u64_s32(vacc01xST), vreinterpretq_u64_s32(vacc01xUV)));
275 float32x4_t vout0x0123 = vcvtq_n_f32_s32(vacc0x0123, 4);
276 float32x4_t vout0x4567 = vcvtq_n_f32_s32(vacc0x4567, 4);
277 float32x4_t vout0x89AB = vcvtq_n_f32_s32(vacc0x89AB, 4);
278 float32x4_t vout0xCDEF = vcvtq_n_f32_s32(vacc0xCDEF, 4);
279 float32x4_t vout0xGHIJ = vcvtq_n_f32_s32(vacc0xGHIJ, 4);
280 float32x4_t vout0xKLMN = vcvtq_n_f32_s32(vacc0xKLMN, 4);
281 float32x4_t vout0xOPQR = vcvtq_n_f32_s32(vacc0xOPQR, 4);
282 float32x4_t vout0xSTUV = vcvtq_n_f32_s32(vacc0xSTUV, 4);
283 float32x4_t vout1x0123 = vcvtq_n_f32_s32(vacc1x0123, 4);
284 float32x4_t vout1x4567 = vcvtq_n_f32_s32(vacc1x4567, 4);
285 float32x4_t vout1x89AB = vcvtq_n_f32_s32(vacc1x89AB, 4);
286 float32x4_t vout1xCDEF = vcvtq_n_f32_s32(vacc1xCDEF, 4);
287 float32x4_t vout1xGHIJ = vcvtq_n_f32_s32(vacc1xGHIJ, 4);
288 float32x4_t vout1xKLMN = vcvtq_n_f32_s32(vacc1xKLMN, 4);
289 float32x4_t vout1xOPQR = vcvtq_n_f32_s32(vacc1xOPQR, 4);
290 float32x4_t vout1xSTUV = vcvtq_n_f32_s32(vacc1xSTUV, 4);
291
292 const float32x4_t vinput_scale01 = vreinterpretq_f32_s32(vld1q_s32(&quantization_params[0].zero_point));
293 vout0x0123 = vmulq_lane_f32(vout0x0123, vget_low_f32(vinput_scale01), 1);
294 vout1x0123 = vmulq_lane_f32(vout1x0123, vget_high_f32(vinput_scale01), 1);
295 vout0x4567 = vmulq_lane_f32(vout0x4567, vget_low_f32(vinput_scale01), 1);
296 vout1x4567 = vmulq_lane_f32(vout1x4567, vget_high_f32(vinput_scale01), 1);
297 vout0x89AB = vmulq_lane_f32(vout0x89AB, vget_low_f32(vinput_scale01), 1);
298 vout1x89AB = vmulq_lane_f32(vout1x89AB, vget_high_f32(vinput_scale01), 1);
299 vout0xCDEF = vmulq_lane_f32(vout0xCDEF, vget_low_f32(vinput_scale01), 1);
300 vout1xCDEF = vmulq_lane_f32(vout1xCDEF, vget_high_f32(vinput_scale01), 1);
301 vout0xGHIJ = vmulq_lane_f32(vout0xGHIJ, vget_low_f32(vinput_scale01), 1);
302 vout1xGHIJ = vmulq_lane_f32(vout1xGHIJ, vget_high_f32(vinput_scale01), 1);
303 vout0xKLMN = vmulq_lane_f32(vout0xKLMN, vget_low_f32(vinput_scale01), 1);
304 vout1xKLMN = vmulq_lane_f32(vout1xKLMN, vget_high_f32(vinput_scale01), 1);
305 vout0xOPQR = vmulq_lane_f32(vout0xOPQR, vget_low_f32(vinput_scale01), 1);
306 vout1xOPQR = vmulq_lane_f32(vout1xOPQR, vget_high_f32(vinput_scale01), 1);
307 vout0xSTUV = vmulq_lane_f32(vout0xSTUV, vget_low_f32(vinput_scale01), 1);
308 vout1xSTUV = vmulq_lane_f32(vout1xSTUV, vget_high_f32(vinput_scale01), 1);
309
310 const float32x4_t vfilter_output_scale0123 = vld1q_f32(w); w = (const float*) w + 4;
311 const float32x4_t vfilter_output_scale4567 = vld1q_f32(w); w = (const float*) w + 4;
312 const float32x4_t vfilter_output_scale89AB = vld1q_f32(w); w = (const float*) w + 4;
313 const float32x4_t vfilter_output_scaleCDEF = vld1q_f32(w); w = (const float*) w + 4;
314 const float32x4_t vfilter_output_scaleGHIJ = vld1q_f32(w); w = (const float*) w + 4;
315 const float32x4_t vfilter_output_scaleKLMN = vld1q_f32(w); w = (const float*) w + 4;
316 const float32x4_t vfilter_output_scaleOPQR = vld1q_f32(w); w = (const float*) w + 4;
317 const float32x4_t vfilter_output_scaleSTUV = vld1q_f32(w); w = (const float*) w + 4;
318
319 const float32x4_t vbias0123 = vld1q_f32(w); w = (const float*) w + 4;
320 vout0x0123 = vfmaq_f32(vbias0123, vout0x0123, vfilter_output_scale0123);
321 vout1x0123 = vfmaq_f32(vbias0123, vout1x0123, vfilter_output_scale0123);
322 const float32x4_t vbias4567 = vld1q_f32(w); w = (const float*) w + 4;
323 vout0x4567 = vfmaq_f32(vbias4567, vout0x4567, vfilter_output_scale4567);
324 vout1x4567 = vfmaq_f32(vbias4567, vout1x4567, vfilter_output_scale4567);
325 const float32x4_t vbias89AB = vld1q_f32(w); w = (const float*) w + 4;
326 vout0x89AB = vfmaq_f32(vbias89AB, vout0x89AB, vfilter_output_scale89AB);
327 vout1x89AB = vfmaq_f32(vbias89AB, vout1x89AB, vfilter_output_scale89AB);
328 const float32x4_t vbiasCDEF = vld1q_f32(w); w = (const float*) w + 4;
329 vout0xCDEF = vfmaq_f32(vbiasCDEF, vout0xCDEF, vfilter_output_scaleCDEF);
330 vout1xCDEF = vfmaq_f32(vbiasCDEF, vout1xCDEF, vfilter_output_scaleCDEF);
331 const float32x4_t vbiasGHIJ = vld1q_f32(w); w = (const float*) w + 4;
332 vout0xGHIJ = vfmaq_f32(vbiasGHIJ, vout0xGHIJ, vfilter_output_scaleGHIJ);
333 vout1xGHIJ = vfmaq_f32(vbiasGHIJ, vout1xGHIJ, vfilter_output_scaleGHIJ);
334 const float32x4_t vbiasKLMN = vld1q_f32(w); w = (const float*) w + 4;
335 vout0xKLMN = vfmaq_f32(vbiasKLMN, vout0xKLMN, vfilter_output_scaleKLMN);
336 vout1xKLMN = vfmaq_f32(vbiasKLMN, vout1xKLMN, vfilter_output_scaleKLMN);
337 const float32x4_t vbiasOPQR = vld1q_f32(w); w = (const float*) w + 4;
338 vout0xOPQR = vfmaq_f32(vbiasOPQR, vout0xOPQR, vfilter_output_scaleOPQR);
339 vout1xOPQR = vfmaq_f32(vbiasOPQR, vout1xOPQR, vfilter_output_scaleOPQR);
340 const float32x4_t vbiasSTUV = vld1q_f32(w); w = (const float*) w + 4;
341 vout0xSTUV = vfmaq_f32(vbiasSTUV, vout0xSTUV, vfilter_output_scaleSTUV);
342 vout1xSTUV = vfmaq_f32(vbiasSTUV, vout1xSTUV, vfilter_output_scaleSTUV);
343
344 float16x8_t vfp16out0x01234567 = vcombine_f16(vcvt_f16_f32(vout0x0123), vcvt_f16_f32(vout0x4567));
345 float16x8_t vfp16out0x89ABCDEF = vcombine_f16(vcvt_f16_f32(vout0x89AB), vcvt_f16_f32(vout0xCDEF));
346 float16x8_t vfp16out0xGHIJKLMN = vcombine_f16(vcvt_f16_f32(vout0xGHIJ), vcvt_f16_f32(vout0xKLMN));
347 float16x8_t vfp16out0xOPQRSTUV = vcombine_f16(vcvt_f16_f32(vout0xOPQR), vcvt_f16_f32(vout0xSTUV));
348 float16x8_t vfp16out1x01234567 = vcombine_f16(vcvt_f16_f32(vout1x0123), vcvt_f16_f32(vout1x4567));
349 float16x8_t vfp16out1x89ABCDEF = vcombine_f16(vcvt_f16_f32(vout1x89AB), vcvt_f16_f32(vout1xCDEF));
350 float16x8_t vfp16out1xGHIJKLMN = vcombine_f16(vcvt_f16_f32(vout1xGHIJ), vcvt_f16_f32(vout1xKLMN));
351 float16x8_t vfp16out1xOPQRSTUV = vcombine_f16(vcvt_f16_f32(vout1xOPQR), vcvt_f16_f32(vout1xSTUV));
352
353 const float16x8_t voutput_min = vreinterpretq_f16_u16(vld1q_dup_u16((const uint16_t*) &params->scalar.min));
354 vfp16out0x01234567 = vmaxq_f16(vfp16out0x01234567, voutput_min);
355 vfp16out0x89ABCDEF = vmaxq_f16(vfp16out0x89ABCDEF, voutput_min);
356 vfp16out0xGHIJKLMN = vmaxq_f16(vfp16out0xGHIJKLMN, voutput_min);
357 vfp16out0xOPQRSTUV = vmaxq_f16(vfp16out0xOPQRSTUV, voutput_min);
358 vfp16out1x01234567 = vmaxq_f16(vfp16out1x01234567, voutput_min);
359 vfp16out1x89ABCDEF = vmaxq_f16(vfp16out1x89ABCDEF, voutput_min);
360 vfp16out1xGHIJKLMN = vmaxq_f16(vfp16out1xGHIJKLMN, voutput_min);
361 vfp16out1xOPQRSTUV = vmaxq_f16(vfp16out1xOPQRSTUV, voutput_min);
362 const float16x8_t voutput_max = vreinterpretq_f16_u16(vld1q_dup_u16((const uint16_t*) &params->scalar.max));
363 vfp16out0x01234567 = vminq_f16(vfp16out0x01234567, voutput_max);
364 vfp16out0x89ABCDEF = vminq_f16(vfp16out0x89ABCDEF, voutput_max);
365 vfp16out0xGHIJKLMN = vminq_f16(vfp16out0xGHIJKLMN, voutput_max);
366 vfp16out0xOPQRSTUV = vminq_f16(vfp16out0xOPQRSTUV, voutput_max);
367 vfp16out1x01234567 = vminq_f16(vfp16out1x01234567, voutput_max);
368 vfp16out1x89ABCDEF = vminq_f16(vfp16out1x89ABCDEF, voutput_max);
369 vfp16out1xGHIJKLMN = vminq_f16(vfp16out1xGHIJKLMN, voutput_max);
370 vfp16out1xOPQRSTUV = vminq_f16(vfp16out1xOPQRSTUV, voutput_max);
371 if XNN_LIKELY(nc >= 32) {
372 vst1q_u16(c0, vreinterpretq_u16_f16(vfp16out0x01234567));
373 vst1q_u16(c0 + 8, vreinterpretq_u16_f16(vfp16out0x89ABCDEF));
374 vst1q_u16(c0 + 16, vreinterpretq_u16_f16(vfp16out0xGHIJKLMN));
375 vst1q_u16(c0 + 24, vreinterpretq_u16_f16(vfp16out0xOPQRSTUV));
376 vst1q_u16(c1, vreinterpretq_u16_f16(vfp16out1x01234567));
377 vst1q_u16(c1 + 8, vreinterpretq_u16_f16(vfp16out1x89ABCDEF));
378 vst1q_u16(c1 + 16, vreinterpretq_u16_f16(vfp16out1xGHIJKLMN));
379 vst1q_u16(c1 + 24, vreinterpretq_u16_f16(vfp16out1xOPQRSTUV));
380
381 a0 = (const int8_t*) ((uintptr_t) a0 - kc);
382 a1 = (const int8_t*) ((uintptr_t) a1 - kc);
383
384 c0 = (uint16_t*) ((uintptr_t) c0 + cn_stride);
385 c1 = (uint16_t*) ((uintptr_t) c1 + cn_stride);
386
387 nc -= 32;
388 } else {
389 if (nc & 16) {
390 vst1q_u16(c0, vreinterpretq_u16_f16(vfp16out0x01234567)); c0 += 8;
391 vfp16out0x01234567 = vfp16out0xGHIJKLMN;
392 vst1q_u16(c1, vreinterpretq_u16_f16(vfp16out1x01234567)); c1 += 8;
393 vfp16out1x01234567 = vfp16out1xGHIJKLMN;
394 vst1q_u16(c0, vreinterpretq_u16_f16(vfp16out0x89ABCDEF)); c0 += 8;
395 vfp16out0x89ABCDEF = vfp16out0xOPQRSTUV;
396 vst1q_u16(c1, vreinterpretq_u16_f16(vfp16out1x89ABCDEF)); c1 += 8;
397 vfp16out1x89ABCDEF = vfp16out1xOPQRSTUV;
398 }
399 if (nc & 8) {
400 vst1q_u16(c0, vreinterpretq_u16_f16(vfp16out0x01234567)); c0 += 8;
401 vfp16out0x01234567 = vfp16out0x89ABCDEF;
402 vst1q_u16(c1, vreinterpretq_u16_f16(vfp16out1x01234567)); c1 += 8;
403 vfp16out1x01234567 = vfp16out1x89ABCDEF;
404 }
405 float16x4_t vfp16out0x0123 = vget_low_f16(vfp16out0x01234567);
406 float16x4_t vfp16out1x0123 = vget_low_f16(vfp16out1x01234567);
407 if (nc & 4) {
408 vst1_u16(c0, vreinterpret_u16_f16(vfp16out0x0123)); c0 += 4;
409 vst1_u16(c1, vreinterpret_u16_f16(vfp16out1x0123)); c1 += 4;
410 vfp16out0x0123 = vget_high_f16(vfp16out0x01234567);
411 vfp16out1x0123 = vget_high_f16(vfp16out1x01234567);
412 }
413 if (nc & 2) {
414 vst1_lane_u32((void*) c0, vreinterpret_u32_f16(vfp16out0x0123), 0); c0 += 2;
415 vst1_lane_u32((void*) c1, vreinterpret_u32_f16(vfp16out1x0123), 0); c1 += 2;
416 vfp16out0x0123 = vext_f16(vfp16out0x0123, vfp16out0x0123, 2);
417 vfp16out1x0123 = vext_f16(vfp16out1x0123, vfp16out1x0123, 2);
418 }
419 if (nc & 1) {
420 vst1_lane_u16(c0, vreinterpret_u16_f16(vfp16out0x0123), 0);
421 vst1_lane_u16(c1, vreinterpret_u16_f16(vfp16out1x0123), 0);
422 }
423 nc = 0;
424 }
425 } while (nc != 0);
426 }