Debian Codegraph @

src/qd8-f16-qc4w-gemm/gen/qd8-f16-qc4w-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_qc4w_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 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 <= 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 const int8x16_t vmask = vmovq_n_s8(INT8_C(0xF0));
86
87 // Loop over groups of 8 columns.
88 do {
89 // Initialize accumulators with bias. 8 bias values are loaded from the
90 // weight matrix, at the start of the group of 8 columns.
91 const int32x4_t vinput_zero_point01 = vld1q_s32(&quantization_params[0].zero_point);
92 const int32x4_t vksum0123 = vld1q_s32(w); w = (const int32_t*) w + 4;
93 const int32x4_t vksumzp0x0123 = vmulq_lane_s32(vksum0123, vget_low_s32(vinput_zero_point01), 0);
94 const int32x4_t vksumzp1x0123 = vmulq_lane_s32(vksum0123, vget_high_s32(vinput_zero_point01), 0);
95 const int32x4_t vksum4567 = vld1q_s32(w); w = (const int32_t*) w + 4;
96 const int32x4_t vksumzp0x4567 = vmulq_lane_s32(vksum4567, vget_low_s32(vinput_zero_point01), 0);
97 const int32x4_t vksumzp1x4567 = vmulq_lane_s32(vksum4567, vget_high_s32(vinput_zero_point01), 0);
98 const int32x4_t vinput_zero_point23 = vld1q_s32(&quantization_params[2].zero_point);
99 const int32x4_t vksumzp2x0123 = vmulq_lane_s32(vksum0123, vget_low_s32(vinput_zero_point23), 0);
100 const int32x4_t vksumzp3x0123 = vmulq_lane_s32(vksum0123, vget_high_s32(vinput_zero_point23), 0);
101 const int32x4_t vksumzp2x4567 = vmulq_lane_s32(vksum4567, vget_low_s32(vinput_zero_point23), 0);
102 const int32x4_t vksumzp3x4567 = vmulq_lane_s32(vksum4567, vget_high_s32(vinput_zero_point23), 0);
103 const int32x4_t vinput_zero_point45 = vld1q_s32(&quantization_params[4].zero_point);
104 const int32x4_t vksumzp4x0123 = vmulq_lane_s32(vksum0123, vget_low_s32(vinput_zero_point45), 0);
105 const int32x4_t vksumzp5x0123 = vmulq_lane_s32(vksum0123, vget_high_s32(vinput_zero_point45), 0);
106 const int32x4_t vksumzp4x4567 = vmulq_lane_s32(vksum4567, vget_low_s32(vinput_zero_point45), 0);
107 const int32x4_t vksumzp5x4567 = vmulq_lane_s32(vksum4567, vget_high_s32(vinput_zero_point45), 0);
108 const int32x4_t vinput_zero_point67 = vld1q_s32(&quantization_params[6].zero_point);
109 const int32x4_t vksumzp6x0123 = vmulq_lane_s32(vksum0123, vget_low_s32(vinput_zero_point67), 0);
110 const int32x4_t vksumzp7x0123 = vmulq_lane_s32(vksum0123, vget_high_s32(vinput_zero_point67), 0);
111 const int32x4_t vksumzp6x4567 = vmulq_lane_s32(vksum4567, vget_low_s32(vinput_zero_point67), 0);
112 const int32x4_t vksumzp7x4567 = vmulq_lane_s32(vksum4567, vget_high_s32(vinput_zero_point67), 0);
113
114 int32x4_t vacc01x01 = vreinterpretq_s32_u64(vtrn1q_u64(vreinterpretq_u64_s32(vksumzp0x0123), vreinterpretq_u64_s32(vksumzp1x0123)));
115 int32x4_t vacc01x23 = vreinterpretq_s32_u64(vtrn2q_u64(vreinterpretq_u64_s32(vksumzp0x0123), vreinterpretq_u64_s32(vksumzp1x0123)));
116 int32x4_t vacc01x45 = vreinterpretq_s32_u64(vtrn1q_u64(vreinterpretq_u64_s32(vksumzp0x4567), vreinterpretq_u64_s32(vksumzp1x4567)));
117 int32x4_t vacc01x67 = vreinterpretq_s32_u64(vtrn2q_u64(vreinterpretq_u64_s32(vksumzp0x4567), vreinterpretq_u64_s32(vksumzp1x4567)));
118 int32x4_t vacc23x01 = vreinterpretq_s32_u64(vtrn1q_u64(vreinterpretq_u64_s32(vksumzp2x0123), vreinterpretq_u64_s32(vksumzp3x0123)));
119 int32x4_t vacc23x23 = vreinterpretq_s32_u64(vtrn2q_u64(vreinterpretq_u64_s32(vksumzp2x0123), vreinterpretq_u64_s32(vksumzp3x0123)));
120 int32x4_t vacc23x45 = vreinterpretq_s32_u64(vtrn1q_u64(vreinterpretq_u64_s32(vksumzp2x4567), vreinterpretq_u64_s32(vksumzp3x4567)));
121 int32x4_t vacc23x67 = vreinterpretq_s32_u64(vtrn2q_u64(vreinterpretq_u64_s32(vksumzp2x4567), vreinterpretq_u64_s32(vksumzp3x4567)));
122 int32x4_t vacc45x01 = vreinterpretq_s32_u64(vtrn1q_u64(vreinterpretq_u64_s32(vksumzp4x0123), vreinterpretq_u64_s32(vksumzp5x0123)));
123 int32x4_t vacc45x23 = vreinterpretq_s32_u64(vtrn2q_u64(vreinterpretq_u64_s32(vksumzp4x0123), vreinterpretq_u64_s32(vksumzp5x0123)));
124 int32x4_t vacc45x45 = vreinterpretq_s32_u64(vtrn1q_u64(vreinterpretq_u64_s32(vksumzp4x4567), vreinterpretq_u64_s32(vksumzp5x4567)));
125 int32x4_t vacc45x67 = vreinterpretq_s32_u64(vtrn2q_u64(vreinterpretq_u64_s32(vksumzp4x4567), vreinterpretq_u64_s32(vksumzp5x4567)));
126 int32x4_t vacc67x01 = vreinterpretq_s32_u64(vtrn1q_u64(vreinterpretq_u64_s32(vksumzp6x0123), vreinterpretq_u64_s32(vksumzp7x0123)));
127 int32x4_t vacc67x23 = vreinterpretq_s32_u64(vtrn2q_u64(vreinterpretq_u64_s32(vksumzp6x0123), vreinterpretq_u64_s32(vksumzp7x0123)));
128 int32x4_t vacc67x45 = vreinterpretq_s32_u64(vtrn1q_u64(vreinterpretq_u64_s32(vksumzp6x4567), vreinterpretq_u64_s32(vksumzp7x4567)));
129 int32x4_t vacc67x67 = vreinterpretq_s32_u64(vtrn2q_u64(vreinterpretq_u64_s32(vksumzp6x4567), vreinterpretq_u64_s32(vksumzp7x4567)));
130
131 // Inner accumulation loop along the 8 columns.
132 size_t k = kc;
133 // 2x partial unrolled loop to load 8 bytes at a time.
134
135 uint64x2x2_t va01x0123456789ABCDEF;
136 va01x0123456789ABCDEF.val[0] = vdupq_n_u64(0);
137 va01x0123456789ABCDEF.val[1] = vdupq_n_u64(0);
138 uint64x2x2_t va23x0123456789ABCDEF;
139 va23x0123456789ABCDEF.val[0] = vdupq_n_u64(0);
140 va23x0123456789ABCDEF.val[1] = vdupq_n_u64(0);
141 uint64x2x2_t va45x0123456789ABCDEF;
142 va45x0123456789ABCDEF.val[0] = vdupq_n_u64(0);
143 va45x0123456789ABCDEF.val[1] = vdupq_n_u64(0);
144 uint64x2x2_t va67x0123456789ABCDEF;
145 va67x0123456789ABCDEF.val[0] = vdupq_n_u64(0);
146 va67x0123456789ABCDEF.val[1] = vdupq_n_u64(0);
147
148 while (k >= 16 * sizeof(int8_t)) {
149 // Load a 8x16 block of activations.
150 va01x0123456789ABCDEF = vld2q_lane_u64((const void*) a0, va01x0123456789ABCDEF, 0); a0 += 16;
151 va23x0123456789ABCDEF = vld2q_lane_u64((const void*) a2, va23x0123456789ABCDEF, 0); a2 += 16;
152 va45x0123456789ABCDEF = vld2q_lane_u64((const void*) a4, va45x0123456789ABCDEF, 0); a4 += 16;
153 va67x0123456789ABCDEF = vld2q_lane_u64((const void*) a6, va67x0123456789ABCDEF, 0); a6 += 16;
154 va01x0123456789ABCDEF = vld2q_lane_u64((const void*) a1, va01x0123456789ABCDEF, 1); a1 += 16;
155 va23x0123456789ABCDEF = vld2q_lane_u64((const void*) a3, va23x0123456789ABCDEF, 1); a3 += 16;
156 va45x0123456789ABCDEF = vld2q_lane_u64((const void*) a5, va45x0123456789ABCDEF, 1); a5 += 16;
157 va67x0123456789ABCDEF = vld2q_lane_u64((const void*) a7, va67x0123456789ABCDEF, 1); a7 += 16;
158
159 const int8x16_t va01x01234567 = vreinterpretq_s8_u64(va01x0123456789ABCDEF.val[0]);
160 const int8x16_t va01x89ABCDEF = vreinterpretq_s8_u64(va01x0123456789ABCDEF.val[1]);
161 const int8x16_t va23x01234567 = vreinterpretq_s8_u64(va23x0123456789ABCDEF.val[0]);
162 const int8x16_t va23x89ABCDEF = vreinterpretq_s8_u64(va23x0123456789ABCDEF.val[1]);
163 const int8x16_t va45x01234567 = vreinterpretq_s8_u64(va45x0123456789ABCDEF.val[0]);
164 const int8x16_t va45x89ABCDEF = vreinterpretq_s8_u64(va45x0123456789ABCDEF.val[1]);
165 const int8x16_t va67x01234567 = vreinterpretq_s8_u64(va67x0123456789ABCDEF.val[0]);
166 const int8x16_t va67x89ABCDEF = vreinterpretq_s8_u64(va67x0123456789ABCDEF.val[1]);
167
168 // Load a 16x8 block of weights.
169 const int8x16_t vb01x0123456789ABCDEF = vld1q_s8(w); w = (const int8_t*) w + 16;
170 const int8x16_t vb23x0123456789ABCDEF = vld1q_s8(w); w = (const int8_t*) w + 16;
171 const int8x16_t vb45x0123456789ABCDEF = vld1q_s8(w); w = (const int8_t*) w + 16;
172 const int8x16_t vb67x0123456789ABCDEF = vld1q_s8(w); w = (const int8_t*) w + 16;
173 const int8x16_t vb01x01234567 = vshlq_n_s8(vb01x0123456789ABCDEF, 4);
174 const int8x16_t vb23x01234567 = vshlq_n_s8(vb23x0123456789ABCDEF, 4);
175 const int8x16_t vb45x01234567 = vshlq_n_s8(vb45x0123456789ABCDEF, 4);
176 const int8x16_t vb67x01234567 = vshlq_n_s8(vb67x0123456789ABCDEF, 4);
177 const int8x16_t vb01x89ABCDEF = vandq_s8(vb01x0123456789ABCDEF, vmask);
178 const int8x16_t vb23x89ABCDEF = vandq_s8(vb23x0123456789ABCDEF, vmask);
179 const int8x16_t vb45x89ABCDEF = vandq_s8(vb45x0123456789ABCDEF, vmask);
180 const int8x16_t vb67x89ABCDEF = vandq_s8(vb67x0123456789ABCDEF, vmask);
181
182 // Multiply-accumulate: 8x8 * 8x8 --> 8x8.
183 vacc01x01 = vmmlaq_s32(vacc01x01, va01x01234567, vb01x01234567);
184 vacc01x23 = vmmlaq_s32(vacc01x23, va01x01234567, vb23x01234567);
185 vacc01x45 = vmmlaq_s32(vacc01x45, va01x01234567, vb45x01234567);
186 vacc01x67 = vmmlaq_s32(vacc01x67, va01x01234567, vb67x01234567);
187 vacc23x01 = vmmlaq_s32(vacc23x01, va23x01234567, vb01x01234567);
188 vacc23x23 = vmmlaq_s32(vacc23x23, va23x01234567, vb23x01234567);
189 vacc23x45 = vmmlaq_s32(vacc23x45, va23x01234567, vb45x01234567);
190 vacc23x67 = vmmlaq_s32(vacc23x67, va23x01234567, vb67x01234567);
191 vacc45x01 = vmmlaq_s32(vacc45x01, va45x01234567, vb01x01234567);
192 vacc45x23 = vmmlaq_s32(vacc45x23, va45x01234567, vb23x01234567);
193 vacc45x45 = vmmlaq_s32(vacc45x45, va45x01234567, vb45x01234567);
194 vacc45x67 = vmmlaq_s32(vacc45x67, va45x01234567, vb67x01234567);
195 vacc67x01 = vmmlaq_s32(vacc67x01, va67x01234567, vb01x01234567);
196 vacc67x23 = vmmlaq_s32(vacc67x23, va67x01234567, vb23x01234567);
197 vacc67x45 = vmmlaq_s32(vacc67x45, va67x01234567, vb45x01234567);
198 vacc67x67 = vmmlaq_s32(vacc67x67, va67x01234567, vb67x01234567);
199 vacc01x01 = vmmlaq_s32(vacc01x01, va01x89ABCDEF, vb01x89ABCDEF);
200 vacc01x23 = vmmlaq_s32(vacc01x23, va01x89ABCDEF, vb23x89ABCDEF);
201 vacc01x45 = vmmlaq_s32(vacc01x45, va01x89ABCDEF, vb45x89ABCDEF);
202 vacc01x67 = vmmlaq_s32(vacc01x67, va01x89ABCDEF, vb67x89ABCDEF);
203 vacc23x01 = vmmlaq_s32(vacc23x01, va23x89ABCDEF, vb01x89ABCDEF);
204 vacc23x23 = vmmlaq_s32(vacc23x23, va23x89ABCDEF, vb23x89ABCDEF);
205 vacc23x45 = vmmlaq_s32(vacc23x45, va23x89ABCDEF, vb45x89ABCDEF);
206 vacc23x67 = vmmlaq_s32(vacc23x67, va23x89ABCDEF, vb67x89ABCDEF);
207 vacc45x01 = vmmlaq_s32(vacc45x01, va45x89ABCDEF, vb01x89ABCDEF);
208 vacc45x23 = vmmlaq_s32(vacc45x23, va45x89ABCDEF, vb23x89ABCDEF);
209 vacc45x45 = vmmlaq_s32(vacc45x45, va45x89ABCDEF, vb45x89ABCDEF);
210 vacc45x67 = vmmlaq_s32(vacc45x67, va45x89ABCDEF, vb67x89ABCDEF);
211 vacc67x01 = vmmlaq_s32(vacc67x01, va67x89ABCDEF, vb01x89ABCDEF);
212 vacc67x23 = vmmlaq_s32(vacc67x23, va67x89ABCDEF, vb23x89ABCDEF);
213 vacc67x45 = vmmlaq_s32(vacc67x45, va67x89ABCDEF, vb45x89ABCDEF);
214 vacc67x67 = vmmlaq_s32(vacc67x67, va67x89ABCDEF, vb67x89ABCDEF);
215
216 k -= 16 * sizeof(int8_t);
217 }
218 // Handle up to 8 final positions of `k`
219 if XNN_UNLIKELY(k != 0) {
220 // Load a 8x8 block of activations.
221 uint64x2_t va01x01234567 = vld1q_dup_u64((const void*) a0); a0 += 8;
222 uint64x2_t va23x01234567 = vld1q_dup_u64((const void*) a2); a2 += 8;
223 uint64x2_t va45x01234567 = vld1q_dup_u64((const void*) a4); a4 += 8;
224 uint64x2_t va67x01234567 = vld1q_dup_u64((const void*) a6); a6 += 8;
225 va01x01234567 = vld1q_lane_u64((const void*) a1, va01x01234567, 1); a1 += 8;
226 va23x01234567 = vld1q_lane_u64((const void*) a3, va23x01234567, 1); a3 += 8;
227 va45x01234567 = vld1q_lane_u64((const void*) a5, va45x01234567, 1); a5 += 8;
228 va67x01234567 = vld1q_lane_u64((const void*) a7, va67x01234567, 1); a7 += 8;
229
230 // Load a 16x8 block of weights.
231 const int8x16_t vb01x0123456789ABCDEF = vld1q_s8(w); w = (const int8_t*) w + 16;
232 const int8x16_t vb23x0123456789ABCDEF = vld1q_s8(w); w = (const int8_t*) w + 16;
233 const int8x16_t vb45x0123456789ABCDEF = vld1q_s8(w); w = (const int8_t*) w + 16;
234 const int8x16_t vb67x0123456789ABCDEF = vld1q_s8(w); w = (const int8_t*) w + 16;
235 const int8x16_t vb01x01234567 = vshlq_n_s8(vb01x0123456789ABCDEF, 4);
236 const int8x16_t vb23x01234567 = vshlq_n_s8(vb23x0123456789ABCDEF, 4);
237 const int8x16_t vb45x01234567 = vshlq_n_s8(vb45x0123456789ABCDEF, 4);
238 const int8x16_t vb67x01234567 = vshlq_n_s8(vb67x0123456789ABCDEF, 4);
239
240 // Multiply-accumulate: 8x4 * 4x8 --> 8x8.
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 vacc23x01 = vmmlaq_s32(vacc23x01, vreinterpretq_s8_u64(va23x01234567), vb01x01234567);
246 vacc23x23 = vmmlaq_s32(vacc23x23, vreinterpretq_s8_u64(va23x01234567), vb23x01234567);
247 vacc23x45 = vmmlaq_s32(vacc23x45, vreinterpretq_s8_u64(va23x01234567), vb45x01234567);
248 vacc23x67 = vmmlaq_s32(vacc23x67, vreinterpretq_s8_u64(va23x01234567), vb67x01234567);
249 vacc45x01 = vmmlaq_s32(vacc45x01, vreinterpretq_s8_u64(va45x01234567), vb01x01234567);
250 vacc45x23 = vmmlaq_s32(vacc45x23, vreinterpretq_s8_u64(va45x01234567), vb23x01234567);
251 vacc45x45 = vmmlaq_s32(vacc45x45, vreinterpretq_s8_u64(va45x01234567), vb45x01234567);
252 vacc45x67 = vmmlaq_s32(vacc45x67, vreinterpretq_s8_u64(va45x01234567), vb67x01234567);
253 vacc67x01 = vmmlaq_s32(vacc67x01, vreinterpretq_s8_u64(va67x01234567), vb01x01234567);
254 vacc67x23 = vmmlaq_s32(vacc67x23, vreinterpretq_s8_u64(va67x01234567), vb23x01234567);
255 vacc67x45 = vmmlaq_s32(vacc67x45, vreinterpretq_s8_u64(va67x01234567), vb45x01234567);
256 vacc67x67 = vmmlaq_s32(vacc67x67, vreinterpretq_s8_u64(va67x01234567), vb67x01234567);
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 vacc2x0123 = vreinterpretq_s32_u64(vtrn1q_u64(vreinterpretq_u64_s32(vacc23x01), vreinterpretq_u64_s32(vacc23x23)));
264 int32x4_t vacc3x0123 = vreinterpretq_s32_u64(vtrn2q_u64(vreinterpretq_u64_s32(vacc23x01), vreinterpretq_u64_s32(vacc23x23)));
265 int32x4_t vacc2x4567 = vreinterpretq_s32_u64(vtrn1q_u64(vreinterpretq_u64_s32(vacc23x45), vreinterpretq_u64_s32(vacc23x67)));
266 int32x4_t vacc3x4567 = vreinterpretq_s32_u64(vtrn2q_u64(vreinterpretq_u64_s32(vacc23x45), vreinterpretq_u64_s32(vacc23x67)));
267 int32x4_t vacc4x0123 = vreinterpretq_s32_u64(vtrn1q_u64(vreinterpretq_u64_s32(vacc45x01), vreinterpretq_u64_s32(vacc45x23)));
268 int32x4_t vacc5x0123 = vreinterpretq_s32_u64(vtrn2q_u64(vreinterpretq_u64_s32(vacc45x01), vreinterpretq_u64_s32(vacc45x23)));
269 int32x4_t vacc4x4567 = vreinterpretq_s32_u64(vtrn1q_u64(vreinterpretq_u64_s32(vacc45x45), vreinterpretq_u64_s32(vacc45x67)));
270 int32x4_t vacc5x4567 = vreinterpretq_s32_u64(vtrn2q_u64(vreinterpretq_u64_s32(vacc45x45), vreinterpretq_u64_s32(vacc45x67)));
271 int32x4_t vacc6x0123 = vreinterpretq_s32_u64(vtrn1q_u64(vreinterpretq_u64_s32(vacc67x01), vreinterpretq_u64_s32(vacc67x23)));
272 int32x4_t vacc7x0123 = vreinterpretq_s32_u64(vtrn2q_u64(vreinterpretq_u64_s32(vacc67x01), vreinterpretq_u64_s32(vacc67x23)));
273 int32x4_t vacc6x4567 = vreinterpretq_s32_u64(vtrn1q_u64(vreinterpretq_u64_s32(vacc67x45), vreinterpretq_u64_s32(vacc67x67)));
274 int32x4_t vacc7x4567 = vreinterpretq_s32_u64(vtrn2q_u64(vreinterpretq_u64_s32(vacc67x45), vreinterpretq_u64_s32(vacc67x67)));
275 float32x4_t vout0x0123 = vcvtq_n_f32_s32(vacc0x0123, 4);
276 float32x4_t vout0x4567 = vcvtq_n_f32_s32(vacc0x4567, 4);
277 float32x4_t vout1x0123 = vcvtq_n_f32_s32(vacc1x0123, 4);
278 float32x4_t vout1x4567 = vcvtq_n_f32_s32(vacc1x4567, 4);
279 float32x4_t vout2x0123 = vcvtq_n_f32_s32(vacc2x0123, 4);
280 float32x4_t vout2x4567 = vcvtq_n_f32_s32(vacc2x4567, 4);
281 float32x4_t vout3x0123 = vcvtq_n_f32_s32(vacc3x0123, 4);
282 float32x4_t vout3x4567 = vcvtq_n_f32_s32(vacc3x4567, 4);
283 float32x4_t vout4x0123 = vcvtq_n_f32_s32(vacc4x0123, 4);
284 float32x4_t vout4x4567 = vcvtq_n_f32_s32(vacc4x4567, 4);
285 float32x4_t vout5x0123 = vcvtq_n_f32_s32(vacc5x0123, 4);
286 float32x4_t vout5x4567 = vcvtq_n_f32_s32(vacc5x4567, 4);
287 float32x4_t vout6x0123 = vcvtq_n_f32_s32(vacc6x0123, 4);
288 float32x4_t vout6x4567 = vcvtq_n_f32_s32(vacc6x4567, 4);
289 float32x4_t vout7x0123 = vcvtq_n_f32_s32(vacc7x0123, 4);
290 float32x4_t vout7x4567 = vcvtq_n_f32_s32(vacc7x4567, 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 const float32x4_t vinput_scale23 = vreinterpretq_f32_s32(vld1q_s32(&quantization_params[2].zero_point));
298 vout2x0123 = vmulq_lane_f32(vout2x0123, vget_low_f32(vinput_scale23), 1);
299 vout3x0123 = vmulq_lane_f32(vout3x0123, vget_high_f32(vinput_scale23), 1);
300 vout2x4567 = vmulq_lane_f32(vout2x4567, vget_low_f32(vinput_scale23), 1);
301 vout3x4567 = vmulq_lane_f32(vout3x4567, vget_high_f32(vinput_scale23), 1);
302 const float32x4_t vinput_scale45 = vreinterpretq_f32_s32(vld1q_s32(&quantization_params[4].zero_point));
303 vout4x0123 = vmulq_lane_f32(vout4x0123, vget_low_f32(vinput_scale45), 1);
304 vout5x0123 = vmulq_lane_f32(vout5x0123, vget_high_f32(vinput_scale45), 1);
305 vout4x4567 = vmulq_lane_f32(vout4x4567, vget_low_f32(vinput_scale45), 1);
306 vout5x4567 = vmulq_lane_f32(vout5x4567, vget_high_f32(vinput_scale45), 1);
307 const float32x4_t vinput_scale67 = vreinterpretq_f32_s32(vld1q_s32(&quantization_params[6].zero_point));
308 vout6x0123 = vmulq_lane_f32(vout6x0123, vget_low_f32(vinput_scale67), 1);
309 vout7x0123 = vmulq_lane_f32(vout7x0123, vget_high_f32(vinput_scale67), 1);
310 vout6x4567 = vmulq_lane_f32(vout6x4567, vget_low_f32(vinput_scale67), 1);
311 vout7x4567 = vmulq_lane_f32(vout7x4567, vget_high_f32(vinput_scale67), 1);
312
313 const float32x4_t vfilter_output_scale0123 = vld1q_f32(w); w = (const float*) w + 4;
314 const float32x4_t vfilter_output_scale4567 = vld1q_f32(w); w = (const float*) w + 4;
315
316 const float32x4_t vbias0123 = vld1q_f32(w); w = (const float*) w + 4;
317 vout0x0123 = vfmaq_f32(vbias0123, vout0x0123, vfilter_output_scale0123);
318 vout1x0123 = vfmaq_f32(vbias0123, vout1x0123, vfilter_output_scale0123);
319 vout2x0123 = vfmaq_f32(vbias0123, vout2x0123, vfilter_output_scale0123);
320 vout3x0123 = vfmaq_f32(vbias0123, vout3x0123, vfilter_output_scale0123);
321 vout4x0123 = vfmaq_f32(vbias0123, vout4x0123, vfilter_output_scale0123);
322 vout5x0123 = vfmaq_f32(vbias0123, vout5x0123, vfilter_output_scale0123);
323 vout6x0123 = vfmaq_f32(vbias0123, vout6x0123, vfilter_output_scale0123);
324 vout7x0123 = vfmaq_f32(vbias0123, vout7x0123, vfilter_output_scale0123);
325 const float32x4_t vbias4567 = vld1q_f32(w); w = (const float*) w + 4;
326 vout0x4567 = vfmaq_f32(vbias4567, vout0x4567, vfilter_output_scale4567);
327 vout1x4567 = vfmaq_f32(vbias4567, vout1x4567, vfilter_output_scale4567);
328 vout2x4567 = vfmaq_f32(vbias4567, vout2x4567, vfilter_output_scale4567);
329 vout3x4567 = vfmaq_f32(vbias4567, vout3x4567, vfilter_output_scale4567);
330 vout4x4567 = vfmaq_f32(vbias4567, vout4x4567, vfilter_output_scale4567);
331 vout5x4567 = vfmaq_f32(vbias4567, vout5x4567, vfilter_output_scale4567);
332 vout6x4567 = vfmaq_f32(vbias4567, vout6x4567, vfilter_output_scale4567);
333 vout7x4567 = vfmaq_f32(vbias4567, vout7x4567, vfilter_output_scale4567);
334
335 float16x8_t vfp16out0x01234567 = vcombine_f16(vcvt_f16_f32(vout0x0123), vcvt_f16_f32(vout0x4567));
336 float16x8_t vfp16out1x01234567 = vcombine_f16(vcvt_f16_f32(vout1x0123), vcvt_f16_f32(vout1x4567));
337 float16x8_t vfp16out2x01234567 = vcombine_f16(vcvt_f16_f32(vout2x0123), vcvt_f16_f32(vout2x4567));
338 float16x8_t vfp16out3x01234567 = vcombine_f16(vcvt_f16_f32(vout3x0123), vcvt_f16_f32(vout3x4567));
339 float16x8_t vfp16out4x01234567 = vcombine_f16(vcvt_f16_f32(vout4x0123), vcvt_f16_f32(vout4x4567));
340 float16x8_t vfp16out5x01234567 = vcombine_f16(vcvt_f16_f32(vout5x0123), vcvt_f16_f32(vout5x4567));
341 float16x8_t vfp16out6x01234567 = vcombine_f16(vcvt_f16_f32(vout6x0123), vcvt_f16_f32(vout6x4567));
342 float16x8_t vfp16out7x01234567 = vcombine_f16(vcvt_f16_f32(vout7x0123), vcvt_f16_f32(vout7x4567));
343
344 const float16x8_t voutput_min = vreinterpretq_f16_u16(vld1q_dup_u16((const uint16_t*) &params->scalar.min));
345 vfp16out0x01234567 = vmaxq_f16(vfp16out0x01234567, voutput_min);
346 vfp16out1x01234567 = vmaxq_f16(vfp16out1x01234567, voutput_min);
347 vfp16out2x01234567 = vmaxq_f16(vfp16out2x01234567, voutput_min);
348 vfp16out3x01234567 = vmaxq_f16(vfp16out3x01234567, voutput_min);
349 vfp16out4x01234567 = vmaxq_f16(vfp16out4x01234567, voutput_min);
350 vfp16out5x01234567 = vmaxq_f16(vfp16out5x01234567, voutput_min);
351 vfp16out6x01234567 = vmaxq_f16(vfp16out6x01234567, voutput_min);
352 vfp16out7x01234567 = vmaxq_f16(vfp16out7x01234567, voutput_min);
353 const float16x8_t voutput_max = vreinterpretq_f16_u16(vld1q_dup_u16((const uint16_t*) &params->scalar.max));
354 vfp16out0x01234567 = vminq_f16(vfp16out0x01234567, voutput_max);
355 vfp16out1x01234567 = vminq_f16(vfp16out1x01234567, voutput_max);
356 vfp16out2x01234567 = vminq_f16(vfp16out2x01234567, voutput_max);
357 vfp16out3x01234567 = vminq_f16(vfp16out3x01234567, voutput_max);
358 vfp16out4x01234567 = vminq_f16(vfp16out4x01234567, voutput_max);
359 vfp16out5x01234567 = vminq_f16(vfp16out5x01234567, voutput_max);
360 vfp16out6x01234567 = vminq_f16(vfp16out6x01234567, voutput_max);
361 vfp16out7x01234567 = vminq_f16(vfp16out7x01234567, voutput_max);
362 if XNN_LIKELY(nc >= 8) {
363 vst1q_u16(c0, vreinterpretq_u16_f16(vfp16out0x01234567));
364 vst1q_u16(c1, vreinterpretq_u16_f16(vfp16out1x01234567));
365 vst1q_u16(c2, vreinterpretq_u16_f16(vfp16out2x01234567));
366 vst1q_u16(c3, vreinterpretq_u16_f16(vfp16out3x01234567));
367 vst1q_u16(c4, vreinterpretq_u16_f16(vfp16out4x01234567));
368 vst1q_u16(c5, vreinterpretq_u16_f16(vfp16out5x01234567));
369 vst1q_u16(c6, vreinterpretq_u16_f16(vfp16out6x01234567));
370 vst1q_u16(c7, vreinterpretq_u16_f16(vfp16out7x01234567));
371
372 a0 = (const int8_t*) ((uintptr_t) a0 - kc);
373 a1 = (const int8_t*) ((uintptr_t) a1 - kc);
374 a2 = (const int8_t*) ((uintptr_t) a2 - kc);
375 a3 = (const int8_t*) ((uintptr_t) a3 - kc);
376 a4 = (const int8_t*) ((uintptr_t) a4 - kc);
377 a5 = (const int8_t*) ((uintptr_t) a5 - kc);
378 a6 = (const int8_t*) ((uintptr_t) a6 - kc);
379 a7 = (const int8_t*) ((uintptr_t) a7 - kc);
380
381 c0 = (uint16_t*) ((uintptr_t) c0 + cn_stride);
382 c1 = (uint16_t*) ((uintptr_t) c1 + cn_stride);
383 c2 = (uint16_t*) ((uintptr_t) c2 + cn_stride);
384 c3 = (uint16_t*) ((uintptr_t) c3 + cn_stride);
385 c4 = (uint16_t*) ((uintptr_t) c4 + cn_stride);
386 c5 = (uint16_t*) ((uintptr_t) c5 + cn_stride);
387 c6 = (uint16_t*) ((uintptr_t) c6 + cn_stride);
388 c7 = (uint16_t*) ((uintptr_t) c7 + cn_stride);
389
390 nc -= 8;
391 } else {
392 float16x4_t vfp16out0x0123 = vget_low_f16(vfp16out0x01234567);
393 float16x4_t vfp16out1x0123 = vget_low_f16(vfp16out1x01234567);
394 float16x4_t vfp16out2x0123 = vget_low_f16(vfp16out2x01234567);
395 float16x4_t vfp16out3x0123 = vget_low_f16(vfp16out3x01234567);
396 float16x4_t vfp16out4x0123 = vget_low_f16(vfp16out4x01234567);
397 float16x4_t vfp16out5x0123 = vget_low_f16(vfp16out5x01234567);
398 float16x4_t vfp16out6x0123 = vget_low_f16(vfp16out6x01234567);
399 float16x4_t vfp16out7x0123 = vget_low_f16(vfp16out7x01234567);
400 if (nc & 4) {
401 vst1_u16(c0, vreinterpret_u16_f16(vfp16out0x0123)); c0 += 4;
402 vst1_u16(c1, vreinterpret_u16_f16(vfp16out1x0123)); c1 += 4;
403 vst1_u16(c2, vreinterpret_u16_f16(vfp16out2x0123)); c2 += 4;
404 vst1_u16(c3, vreinterpret_u16_f16(vfp16out3x0123)); c3 += 4;
405 vst1_u16(c4, vreinterpret_u16_f16(vfp16out4x0123)); c4 += 4;
406 vst1_u16(c5, vreinterpret_u16_f16(vfp16out5x0123)); c5 += 4;
407 vst1_u16(c6, vreinterpret_u16_f16(vfp16out6x0123)); c6 += 4;
408 vst1_u16(c7, vreinterpret_u16_f16(vfp16out7x0123)); c7 += 4;
409 vfp16out0x0123 = vget_high_f16(vfp16out0x01234567);
410 vfp16out1x0123 = vget_high_f16(vfp16out1x01234567);
411 vfp16out2x0123 = vget_high_f16(vfp16out2x01234567);
412 vfp16out3x0123 = vget_high_f16(vfp16out3x01234567);
413 vfp16out4x0123 = vget_high_f16(vfp16out4x01234567);
414 vfp16out5x0123 = vget_high_f16(vfp16out5x01234567);
415 vfp16out6x0123 = vget_high_f16(vfp16out6x01234567);
416 vfp16out7x0123 = vget_high_f16(vfp16out7x01234567);
417 }
418 if (nc & 2) {
419 vst1_lane_u32((void*) c0, vreinterpret_u32_f16(vfp16out0x0123), 0); c0 += 2;
420 vst1_lane_u32((void*) c1, vreinterpret_u32_f16(vfp16out1x0123), 0); c1 += 2;
421 vst1_lane_u32((void*) c2, vreinterpret_u32_f16(vfp16out2x0123), 0); c2 += 2;
422 vst1_lane_u32((void*) c3, vreinterpret_u32_f16(vfp16out3x0123), 0); c3 += 2;
423 vst1_lane_u32((void*) c4, vreinterpret_u32_f16(vfp16out4x0123), 0); c4 += 2;
424 vst1_lane_u32((void*) c5, vreinterpret_u32_f16(vfp16out5x0123), 0); c5 += 2;
425 vst1_lane_u32((void*) c6, vreinterpret_u32_f16(vfp16out6x0123), 0); c6 += 2;
426 vst1_lane_u32((void*) c7, vreinterpret_u32_f16(vfp16out7x0123), 0); c7 += 2;
427 vfp16out0x0123 = vext_f16(vfp16out0x0123, vfp16out0x0123, 2);
428 vfp16out1x0123 = vext_f16(vfp16out1x0123, vfp16out1x0123, 2);
429 vfp16out2x0123 = vext_f16(vfp16out2x0123, vfp16out2x0123, 2);
430 vfp16out3x0123 = vext_f16(vfp16out3x0123, vfp16out3x0123, 2);
431 vfp16out4x0123 = vext_f16(vfp16out4x0123, vfp16out4x0123, 2);
432 vfp16out5x0123 = vext_f16(vfp16out5x0123, vfp16out5x0123, 2);
433 vfp16out6x0123 = vext_f16(vfp16out6x0123, vfp16out6x0123, 2);
434 vfp16out7x0123 = vext_f16(vfp16out7x0123, vfp16out7x0123, 2);
435 }
436 if (nc & 1) {
437 vst1_lane_u16(c0, vreinterpret_u16_f16(vfp16out0x0123), 0);
438 vst1_lane_u16(c1, vreinterpret_u16_f16(vfp16out1x0123), 0);
439 vst1_lane_u16(c2, vreinterpret_u16_f16(vfp16out2x0123), 0);
440 vst1_lane_u16(c3, vreinterpret_u16_f16(vfp16out3x0123), 0);
441 vst1_lane_u16(c4, vreinterpret_u16_f16(vfp16out4x0123), 0);
442 vst1_lane_u16(c5, vreinterpret_u16_f16(vfp16out5x0123), 0);
443 vst1_lane_u16(c6, vreinterpret_u16_f16(vfp16out6x0123), 0);
444 vst1_lane_u16(c7, vreinterpret_u16_f16(vfp16out7x0123), 0);
445 }
446 nc = 0;
447 }
448 } while (nc != 0);
449 }