Debian Codegraph @

src/qd8-f16-qc4w-gemm/gen/qd8-f16-qc4w-gemm-7x8c8-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_7x8c8__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 <= 7);
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 int8x16_t vmask = vmovq_n_s8(INT8_C(0xF0));
80
81 // Loop over groups of 8 columns.
82 do {
83 // Initialize accumulators with bias. 8 bias values are loaded from the
84 // weight matrix, at the start of the group of 8 columns.
85 const int32x4_t vinput_zero_point01 = vld1q_s32(&quantization_params[0].zero_point);
86 const int32x4_t vksum0123 = vld1q_s32(w); w = (const int32_t*) w + 4;
87 const int32x4_t vksumzp0x0123 = vmulq_lane_s32(vksum0123, vget_low_s32(vinput_zero_point01), 0);
88 const int32x4_t vksumzp1x0123 = vmulq_lane_s32(vksum0123, vget_high_s32(vinput_zero_point01), 0);
89 const int32x4_t vksum4567 = vld1q_s32(w); w = (const int32_t*) w + 4;
90 const int32x4_t vksumzp0x4567 = vmulq_lane_s32(vksum4567, vget_low_s32(vinput_zero_point01), 0);
91 const int32x4_t vksumzp1x4567 = vmulq_lane_s32(vksum4567, vget_high_s32(vinput_zero_point01), 0);
92 const int32x4_t vinput_zero_point23 = vld1q_s32(&quantization_params[2].zero_point);
93 const int32x4_t vksumzp2x0123 = vmulq_lane_s32(vksum0123, vget_low_s32(vinput_zero_point23), 0);
94 const int32x4_t vksumzp3x0123 = vmulq_lane_s32(vksum0123, vget_high_s32(vinput_zero_point23), 0);
95 const int32x4_t vksumzp2x4567 = vmulq_lane_s32(vksum4567, vget_low_s32(vinput_zero_point23), 0);
96 const int32x4_t vksumzp3x4567 = vmulq_lane_s32(vksum4567, vget_high_s32(vinput_zero_point23), 0);
97 const int32x4_t vinput_zero_point45 = vld1q_s32(&quantization_params[4].zero_point);
98 const int32x4_t vksumzp4x0123 = vmulq_lane_s32(vksum0123, vget_low_s32(vinput_zero_point45), 0);
99 const int32x4_t vksumzp5x0123 = vmulq_lane_s32(vksum0123, vget_high_s32(vinput_zero_point45), 0);
100 const int32x4_t vksumzp4x4567 = vmulq_lane_s32(vksum4567, vget_low_s32(vinput_zero_point45), 0);
101 const int32x4_t vksumzp5x4567 = vmulq_lane_s32(vksum4567, vget_high_s32(vinput_zero_point45), 0);
102 const int32x4_t vinput_zero_point67 = vld1q_dup_s32(&quantization_params[6].zero_point);
103 const int32x4_t vksumzp6x0123 = vmulq_s32(vksum0123, vinput_zero_point67);
104 const int32x4_t vksumzp6x4567 = vmulq_s32(vksum4567, vinput_zero_point67);
105
106 int32x4_t vacc01x01 = vreinterpretq_s32_u64(vtrn1q_u64(vreinterpretq_u64_s32(vksumzp0x0123), vreinterpretq_u64_s32(vksumzp1x0123)));
107 int32x4_t vacc01x23 = vreinterpretq_s32_u64(vtrn2q_u64(vreinterpretq_u64_s32(vksumzp0x0123), vreinterpretq_u64_s32(vksumzp1x0123)));
108 int32x4_t vacc01x45 = vreinterpretq_s32_u64(vtrn1q_u64(vreinterpretq_u64_s32(vksumzp0x4567), vreinterpretq_u64_s32(vksumzp1x4567)));
109 int32x4_t vacc01x67 = vreinterpretq_s32_u64(vtrn2q_u64(vreinterpretq_u64_s32(vksumzp0x4567), vreinterpretq_u64_s32(vksumzp1x4567)));
110 int32x4_t vacc23x01 = vreinterpretq_s32_u64(vtrn1q_u64(vreinterpretq_u64_s32(vksumzp2x0123), vreinterpretq_u64_s32(vksumzp3x0123)));
111 int32x4_t vacc23x23 = vreinterpretq_s32_u64(vtrn2q_u64(vreinterpretq_u64_s32(vksumzp2x0123), vreinterpretq_u64_s32(vksumzp3x0123)));
112 int32x4_t vacc23x45 = vreinterpretq_s32_u64(vtrn1q_u64(vreinterpretq_u64_s32(vksumzp2x4567), vreinterpretq_u64_s32(vksumzp3x4567)));
113 int32x4_t vacc23x67 = vreinterpretq_s32_u64(vtrn2q_u64(vreinterpretq_u64_s32(vksumzp2x4567), vreinterpretq_u64_s32(vksumzp3x4567)));
114 int32x4_t vacc45x01 = vreinterpretq_s32_u64(vtrn1q_u64(vreinterpretq_u64_s32(vksumzp4x0123), vreinterpretq_u64_s32(vksumzp5x0123)));
115 int32x4_t vacc45x23 = vreinterpretq_s32_u64(vtrn2q_u64(vreinterpretq_u64_s32(vksumzp4x0123), vreinterpretq_u64_s32(vksumzp5x0123)));
116 int32x4_t vacc45x45 = vreinterpretq_s32_u64(vtrn1q_u64(vreinterpretq_u64_s32(vksumzp4x4567), vreinterpretq_u64_s32(vksumzp5x4567)));
117 int32x4_t vacc45x67 = vreinterpretq_s32_u64(vtrn2q_u64(vreinterpretq_u64_s32(vksumzp4x4567), vreinterpretq_u64_s32(vksumzp5x4567)));
118 int32x4_t vacc67x01 = vreinterpretq_s32_u64(vtrn1q_u64(vreinterpretq_u64_s32(vksumzp6x0123), vreinterpretq_u64_s32(vksumzp6x0123)));
119 int32x4_t vacc67x23 = vreinterpretq_s32_u64(vtrn2q_u64(vreinterpretq_u64_s32(vksumzp6x0123), vreinterpretq_u64_s32(vksumzp6x0123)));
120 int32x4_t vacc67x45 = vreinterpretq_s32_u64(vtrn1q_u64(vreinterpretq_u64_s32(vksumzp6x4567), vreinterpretq_u64_s32(vksumzp6x4567)));
121 int32x4_t vacc67x67 = vreinterpretq_s32_u64(vtrn2q_u64(vreinterpretq_u64_s32(vksumzp6x4567), vreinterpretq_u64_s32(vksumzp6x4567)));
122
123 // Inner accumulation loop along the 8 columns.
124 size_t k = kc;
125 // 2x partial unrolled loop to load 8 bytes at a time.
126
127 uint64x2x2_t va01x0123456789ABCDEF;
128 va01x0123456789ABCDEF.val[0] = vdupq_n_u64(0);
129 va01x0123456789ABCDEF.val[1] = vdupq_n_u64(0);
130 uint64x2x2_t va23x0123456789ABCDEF;
131 va23x0123456789ABCDEF.val[0] = vdupq_n_u64(0);
132 va23x0123456789ABCDEF.val[1] = vdupq_n_u64(0);
133 uint64x2x2_t va45x0123456789ABCDEF;
134 va45x0123456789ABCDEF.val[0] = vdupq_n_u64(0);
135 va45x0123456789ABCDEF.val[1] = vdupq_n_u64(0);
136 uint64x2x2_t va67x0123456789ABCDEF;
137 va67x0123456789ABCDEF.val[0] = vdupq_n_u64(0);
138 va67x0123456789ABCDEF.val[1] = vdupq_n_u64(0);
139
140 while (k >= 16 * sizeof(int8_t)) {
141 // Load a 7x16 block of activations.
142 va01x0123456789ABCDEF = vld2q_lane_u64((const void*) a0, va01x0123456789ABCDEF, 0); a0 += 16;
143 va23x0123456789ABCDEF = vld2q_lane_u64((const void*) a2, va23x0123456789ABCDEF, 0); a2 += 16;
144 va45x0123456789ABCDEF = vld2q_lane_u64((const void*) a4, va45x0123456789ABCDEF, 0); a4 += 16;
145 va67x0123456789ABCDEF = vld2q_lane_u64((const void*) a6, va67x0123456789ABCDEF, 0); a6 += 16;
146 va01x0123456789ABCDEF = vld2q_lane_u64((const void*) a1, va01x0123456789ABCDEF, 1); a1 += 16;
147 va23x0123456789ABCDEF = vld2q_lane_u64((const void*) a3, va23x0123456789ABCDEF, 1); a3 += 16;
148 va45x0123456789ABCDEF = vld2q_lane_u64((const void*) a5, va45x0123456789ABCDEF, 1); a5 += 16;
149
150 const int8x16_t va01x01234567 = vreinterpretq_s8_u64(va01x0123456789ABCDEF.val[0]);
151 const int8x16_t va01x89ABCDEF = vreinterpretq_s8_u64(va01x0123456789ABCDEF.val[1]);
152 const int8x16_t va23x01234567 = vreinterpretq_s8_u64(va23x0123456789ABCDEF.val[0]);
153 const int8x16_t va23x89ABCDEF = vreinterpretq_s8_u64(va23x0123456789ABCDEF.val[1]);
154 const int8x16_t va45x01234567 = vreinterpretq_s8_u64(va45x0123456789ABCDEF.val[0]);
155 const int8x16_t va45x89ABCDEF = vreinterpretq_s8_u64(va45x0123456789ABCDEF.val[1]);
156 const int8x16_t va67x01234567 = vreinterpretq_s8_u64(va67x0123456789ABCDEF.val[0]);
157 const int8x16_t va67x89ABCDEF = vreinterpretq_s8_u64(va67x0123456789ABCDEF.val[1]);
158
159 // Load a 16x8 block of weights.
160 const int8x16_t vb01x0123456789ABCDEF = vld1q_s8(w); w = (const int8_t*) w + 16;
161 const int8x16_t vb23x0123456789ABCDEF = vld1q_s8(w); w = (const int8_t*) w + 16;
162 const int8x16_t vb45x0123456789ABCDEF = vld1q_s8(w); w = (const int8_t*) w + 16;
163 const int8x16_t vb67x0123456789ABCDEF = vld1q_s8(w); w = (const int8_t*) w + 16;
164 const int8x16_t vb01x01234567 = vshlq_n_s8(vb01x0123456789ABCDEF, 4);
165 const int8x16_t vb23x01234567 = vshlq_n_s8(vb23x0123456789ABCDEF, 4);
166 const int8x16_t vb45x01234567 = vshlq_n_s8(vb45x0123456789ABCDEF, 4);
167 const int8x16_t vb67x01234567 = vshlq_n_s8(vb67x0123456789ABCDEF, 4);
168 const int8x16_t vb01x89ABCDEF = vandq_s8(vb01x0123456789ABCDEF, vmask);
169 const int8x16_t vb23x89ABCDEF = vandq_s8(vb23x0123456789ABCDEF, vmask);
170 const int8x16_t vb45x89ABCDEF = vandq_s8(vb45x0123456789ABCDEF, vmask);
171 const int8x16_t vb67x89ABCDEF = vandq_s8(vb67x0123456789ABCDEF, vmask);
172
173 // Multiply-accumulate: 7x8 * 8x8 --> 7x8.
174 vacc01x01 = vmmlaq_s32(vacc01x01, va01x01234567, vb01x01234567);
175 vacc01x23 = vmmlaq_s32(vacc01x23, va01x01234567, vb23x01234567);
176 vacc01x45 = vmmlaq_s32(vacc01x45, va01x01234567, vb45x01234567);
177 vacc01x67 = vmmlaq_s32(vacc01x67, va01x01234567, vb67x01234567);
178 vacc23x01 = vmmlaq_s32(vacc23x01, va23x01234567, vb01x01234567);
179 vacc23x23 = vmmlaq_s32(vacc23x23, va23x01234567, vb23x01234567);
180 vacc23x45 = vmmlaq_s32(vacc23x45, va23x01234567, vb45x01234567);
181 vacc23x67 = vmmlaq_s32(vacc23x67, va23x01234567, vb67x01234567);
182 vacc45x01 = vmmlaq_s32(vacc45x01, va45x01234567, vb01x01234567);
183 vacc45x23 = vmmlaq_s32(vacc45x23, va45x01234567, vb23x01234567);
184 vacc45x45 = vmmlaq_s32(vacc45x45, va45x01234567, vb45x01234567);
185 vacc45x67 = vmmlaq_s32(vacc45x67, va45x01234567, vb67x01234567);
186 vacc67x01 = vmmlaq_s32(vacc67x01, va67x01234567, vb01x01234567);
187 vacc67x23 = vmmlaq_s32(vacc67x23, va67x01234567, vb23x01234567);
188 vacc67x45 = vmmlaq_s32(vacc67x45, va67x01234567, vb45x01234567);
189 vacc67x67 = vmmlaq_s32(vacc67x67, va67x01234567, vb67x01234567);
190 vacc01x01 = vmmlaq_s32(vacc01x01, va01x89ABCDEF, vb01x89ABCDEF);
191 vacc01x23 = vmmlaq_s32(vacc01x23, va01x89ABCDEF, vb23x89ABCDEF);
192 vacc01x45 = vmmlaq_s32(vacc01x45, va01x89ABCDEF, vb45x89ABCDEF);
193 vacc01x67 = vmmlaq_s32(vacc01x67, va01x89ABCDEF, vb67x89ABCDEF);
194 vacc23x01 = vmmlaq_s32(vacc23x01, va23x89ABCDEF, vb01x89ABCDEF);
195 vacc23x23 = vmmlaq_s32(vacc23x23, va23x89ABCDEF, vb23x89ABCDEF);
196 vacc23x45 = vmmlaq_s32(vacc23x45, va23x89ABCDEF, vb45x89ABCDEF);
197 vacc23x67 = vmmlaq_s32(vacc23x67, va23x89ABCDEF, vb67x89ABCDEF);
198 vacc45x01 = vmmlaq_s32(vacc45x01, va45x89ABCDEF, vb01x89ABCDEF);
199 vacc45x23 = vmmlaq_s32(vacc45x23, va45x89ABCDEF, vb23x89ABCDEF);
200 vacc45x45 = vmmlaq_s32(vacc45x45, va45x89ABCDEF, vb45x89ABCDEF);
201 vacc45x67 = vmmlaq_s32(vacc45x67, va45x89ABCDEF, vb67x89ABCDEF);
202 vacc67x01 = vmmlaq_s32(vacc67x01, va67x89ABCDEF, vb01x89ABCDEF);
203 vacc67x23 = vmmlaq_s32(vacc67x23, va67x89ABCDEF, vb23x89ABCDEF);
204 vacc67x45 = vmmlaq_s32(vacc67x45, va67x89ABCDEF, vb45x89ABCDEF);
205 vacc67x67 = vmmlaq_s32(vacc67x67, va67x89ABCDEF, vb67x89ABCDEF);
206
207 k -= 16 * sizeof(int8_t);
208 }
209 // Handle up to 8 final positions of `k`
210 if XNN_UNLIKELY(k != 0) {
211 // Load a 7x8 block of activations.
212 uint64x2_t va01x01234567 = vld1q_dup_u64((const void*) a0); a0 += 8;
213 uint64x2_t va23x01234567 = vld1q_dup_u64((const void*) a2); a2 += 8;
214 uint64x2_t va45x01234567 = vld1q_dup_u64((const void*) a4); a4 += 8;
215 uint64x2_t va67x01234567 = vld1q_dup_u64((const void*) a6); a6 += 8;
216 va01x01234567 = vld1q_lane_u64((const void*) a1, va01x01234567, 1); a1 += 8;
217 va23x01234567 = vld1q_lane_u64((const void*) a3, va23x01234567, 1); a3 += 8;
218 va45x01234567 = vld1q_lane_u64((const void*) a5, va45x01234567, 1); a5 += 8;
219
220 // Load a 16x8 block of weights.
221 const int8x16_t vb01x0123456789ABCDEF = vld1q_s8(w); w = (const int8_t*) w + 16;
222 const int8x16_t vb23x0123456789ABCDEF = vld1q_s8(w); w = (const int8_t*) w + 16;
223 const int8x16_t vb45x0123456789ABCDEF = vld1q_s8(w); w = (const int8_t*) w + 16;
224 const int8x16_t vb67x0123456789ABCDEF = vld1q_s8(w); w = (const int8_t*) w + 16;
225 const int8x16_t vb01x01234567 = vshlq_n_s8(vb01x0123456789ABCDEF, 4);
226 const int8x16_t vb23x01234567 = vshlq_n_s8(vb23x0123456789ABCDEF, 4);
227 const int8x16_t vb45x01234567 = vshlq_n_s8(vb45x0123456789ABCDEF, 4);
228 const int8x16_t vb67x01234567 = vshlq_n_s8(vb67x0123456789ABCDEF, 4);
229
230 // Multiply-accumulate: 7x4 * 4x8 --> 7x8.
231 vacc01x01 = vmmlaq_s32(vacc01x01, vreinterpretq_s8_u64(va01x01234567), vb01x01234567);
232 vacc01x23 = vmmlaq_s32(vacc01x23, vreinterpretq_s8_u64(va01x01234567), vb23x01234567);
233 vacc01x45 = vmmlaq_s32(vacc01x45, vreinterpretq_s8_u64(va01x01234567), vb45x01234567);
234 vacc01x67 = vmmlaq_s32(vacc01x67, vreinterpretq_s8_u64(va01x01234567), vb67x01234567);
235 vacc23x01 = vmmlaq_s32(vacc23x01, vreinterpretq_s8_u64(va23x01234567), vb01x01234567);
236 vacc23x23 = vmmlaq_s32(vacc23x23, vreinterpretq_s8_u64(va23x01234567), vb23x01234567);
237 vacc23x45 = vmmlaq_s32(vacc23x45, vreinterpretq_s8_u64(va23x01234567), vb45x01234567);
238 vacc23x67 = vmmlaq_s32(vacc23x67, vreinterpretq_s8_u64(va23x01234567), vb67x01234567);
239 vacc45x01 = vmmlaq_s32(vacc45x01, vreinterpretq_s8_u64(va45x01234567), vb01x01234567);
240 vacc45x23 = vmmlaq_s32(vacc45x23, vreinterpretq_s8_u64(va45x01234567), vb23x01234567);
241 vacc45x45 = vmmlaq_s32(vacc45x45, vreinterpretq_s8_u64(va45x01234567), vb45x01234567);
242 vacc45x67 = vmmlaq_s32(vacc45x67, vreinterpretq_s8_u64(va45x01234567), vb67x01234567);
243 vacc67x01 = vmmlaq_s32(vacc67x01, vreinterpretq_s8_u64(va67x01234567), vb01x01234567);
244 vacc67x23 = vmmlaq_s32(vacc67x23, vreinterpretq_s8_u64(va67x01234567), vb23x01234567);
245 vacc67x45 = vmmlaq_s32(vacc67x45, vreinterpretq_s8_u64(va67x01234567), vb45x01234567);
246 vacc67x67 = vmmlaq_s32(vacc67x67, vreinterpretq_s8_u64(va67x01234567), vb67x01234567);
247 }
248
249 int32x4_t vacc0x0123 = vreinterpretq_s32_u64(vtrn1q_u64(vreinterpretq_u64_s32(vacc01x01), vreinterpretq_u64_s32(vacc01x23)));
250 int32x4_t vacc1x0123 = vreinterpretq_s32_u64(vtrn2q_u64(vreinterpretq_u64_s32(vacc01x01), vreinterpretq_u64_s32(vacc01x23)));
251 int32x4_t vacc0x4567 = vreinterpretq_s32_u64(vtrn1q_u64(vreinterpretq_u64_s32(vacc01x45), vreinterpretq_u64_s32(vacc01x67)));
252 int32x4_t vacc1x4567 = vreinterpretq_s32_u64(vtrn2q_u64(vreinterpretq_u64_s32(vacc01x45), vreinterpretq_u64_s32(vacc01x67)));
253 int32x4_t vacc2x0123 = vreinterpretq_s32_u64(vtrn1q_u64(vreinterpretq_u64_s32(vacc23x01), vreinterpretq_u64_s32(vacc23x23)));
254 int32x4_t vacc3x0123 = vreinterpretq_s32_u64(vtrn2q_u64(vreinterpretq_u64_s32(vacc23x01), vreinterpretq_u64_s32(vacc23x23)));
255 int32x4_t vacc2x4567 = vreinterpretq_s32_u64(vtrn1q_u64(vreinterpretq_u64_s32(vacc23x45), vreinterpretq_u64_s32(vacc23x67)));
256 int32x4_t vacc3x4567 = vreinterpretq_s32_u64(vtrn2q_u64(vreinterpretq_u64_s32(vacc23x45), vreinterpretq_u64_s32(vacc23x67)));
257 int32x4_t vacc4x0123 = vreinterpretq_s32_u64(vtrn1q_u64(vreinterpretq_u64_s32(vacc45x01), vreinterpretq_u64_s32(vacc45x23)));
258 int32x4_t vacc5x0123 = vreinterpretq_s32_u64(vtrn2q_u64(vreinterpretq_u64_s32(vacc45x01), vreinterpretq_u64_s32(vacc45x23)));
259 int32x4_t vacc4x4567 = vreinterpretq_s32_u64(vtrn1q_u64(vreinterpretq_u64_s32(vacc45x45), vreinterpretq_u64_s32(vacc45x67)));
260 int32x4_t vacc5x4567 = vreinterpretq_s32_u64(vtrn2q_u64(vreinterpretq_u64_s32(vacc45x45), vreinterpretq_u64_s32(vacc45x67)));
261 int32x4_t vacc6x0123 = vreinterpretq_s32_u64(vtrn1q_u64(vreinterpretq_u64_s32(vacc67x01), vreinterpretq_u64_s32(vacc67x23)));
262 int32x4_t vacc6x4567 = vreinterpretq_s32_u64(vtrn1q_u64(vreinterpretq_u64_s32(vacc67x45), vreinterpretq_u64_s32(vacc67x67)));
263 float32x4_t vout0x0123 = vcvtq_n_f32_s32(vacc0x0123, 4);
264 float32x4_t vout0x4567 = vcvtq_n_f32_s32(vacc0x4567, 4);
265 float32x4_t vout1x0123 = vcvtq_n_f32_s32(vacc1x0123, 4);
266 float32x4_t vout1x4567 = vcvtq_n_f32_s32(vacc1x4567, 4);
267 float32x4_t vout2x0123 = vcvtq_n_f32_s32(vacc2x0123, 4);
268 float32x4_t vout2x4567 = vcvtq_n_f32_s32(vacc2x4567, 4);
269 float32x4_t vout3x0123 = vcvtq_n_f32_s32(vacc3x0123, 4);
270 float32x4_t vout3x4567 = vcvtq_n_f32_s32(vacc3x4567, 4);
271 float32x4_t vout4x0123 = vcvtq_n_f32_s32(vacc4x0123, 4);
272 float32x4_t vout4x4567 = vcvtq_n_f32_s32(vacc4x4567, 4);
273 float32x4_t vout5x0123 = vcvtq_n_f32_s32(vacc5x0123, 4);
274 float32x4_t vout5x4567 = vcvtq_n_f32_s32(vacc5x4567, 4);
275 float32x4_t vout6x0123 = vcvtq_n_f32_s32(vacc6x0123, 4);
276 float32x4_t vout6x4567 = vcvtq_n_f32_s32(vacc6x4567, 4);
277
278 const float32x4_t vinput_scale01 = vreinterpretq_f32_s32(vld1q_s32(&quantization_params[0].zero_point));
279 vout0x0123 = vmulq_lane_f32(vout0x0123, vget_low_f32(vinput_scale01), 1);
280 vout1x0123 = vmulq_lane_f32(vout1x0123, vget_high_f32(vinput_scale01), 1);
281 vout0x4567 = vmulq_lane_f32(vout0x4567, vget_low_f32(vinput_scale01), 1);
282 vout1x4567 = vmulq_lane_f32(vout1x4567, vget_high_f32(vinput_scale01), 1);
283 const float32x4_t vinput_scale23 = vreinterpretq_f32_s32(vld1q_s32(&quantization_params[2].zero_point));
284 vout2x0123 = vmulq_lane_f32(vout2x0123, vget_low_f32(vinput_scale23), 1);
285 vout3x0123 = vmulq_lane_f32(vout3x0123, vget_high_f32(vinput_scale23), 1);
286 vout2x4567 = vmulq_lane_f32(vout2x4567, vget_low_f32(vinput_scale23), 1);
287 vout3x4567 = vmulq_lane_f32(vout3x4567, vget_high_f32(vinput_scale23), 1);
288 const float32x4_t vinput_scale45 = vreinterpretq_f32_s32(vld1q_s32(&quantization_params[4].zero_point));
289 vout4x0123 = vmulq_lane_f32(vout4x0123, vget_low_f32(vinput_scale45), 1);
290 vout5x0123 = vmulq_lane_f32(vout5x0123, vget_high_f32(vinput_scale45), 1);
291 vout4x4567 = vmulq_lane_f32(vout4x4567, vget_low_f32(vinput_scale45), 1);
292 vout5x4567 = vmulq_lane_f32(vout5x4567, vget_high_f32(vinput_scale45), 1);
293 const float32x4_t vinput_scale6 = vld1q_dup_f32(&quantization_params[6].inv_scale);
294 vout6x0123 = vmulq_f32(vout6x0123, vinput_scale6);
295 vout6x4567 = vmulq_f32(vout6x4567, vinput_scale6);
296
297 const float32x4_t vfilter_output_scale0123 = vld1q_f32(w); w = (const float*) w + 4;
298 const float32x4_t vfilter_output_scale4567 = vld1q_f32(w); w = (const float*) w + 4;
299
300 const float32x4_t vbias0123 = vld1q_f32(w); w = (const float*) w + 4;
301 vout0x0123 = vfmaq_f32(vbias0123, vout0x0123, vfilter_output_scale0123);
302 vout1x0123 = vfmaq_f32(vbias0123, vout1x0123, vfilter_output_scale0123);
303 vout2x0123 = vfmaq_f32(vbias0123, vout2x0123, vfilter_output_scale0123);
304 vout3x0123 = vfmaq_f32(vbias0123, vout3x0123, vfilter_output_scale0123);
305 vout4x0123 = vfmaq_f32(vbias0123, vout4x0123, vfilter_output_scale0123);
306 vout5x0123 = vfmaq_f32(vbias0123, vout5x0123, vfilter_output_scale0123);
307 vout6x0123 = vfmaq_f32(vbias0123, vout6x0123, vfilter_output_scale0123);
308 const float32x4_t vbias4567 = vld1q_f32(w); w = (const float*) w + 4;
309 vout0x4567 = vfmaq_f32(vbias4567, vout0x4567, vfilter_output_scale4567);
310 vout1x4567 = vfmaq_f32(vbias4567, vout1x4567, vfilter_output_scale4567);
311 vout2x4567 = vfmaq_f32(vbias4567, vout2x4567, vfilter_output_scale4567);
312 vout3x4567 = vfmaq_f32(vbias4567, vout3x4567, vfilter_output_scale4567);
313 vout4x4567 = vfmaq_f32(vbias4567, vout4x4567, vfilter_output_scale4567);
314 vout5x4567 = vfmaq_f32(vbias4567, vout5x4567, vfilter_output_scale4567);
315 vout6x4567 = vfmaq_f32(vbias4567, vout6x4567, vfilter_output_scale4567);
316
317 float16x8_t vfp16out0x01234567 = vcombine_f16(vcvt_f16_f32(vout0x0123), vcvt_f16_f32(vout0x4567));
318 float16x8_t vfp16out1x01234567 = vcombine_f16(vcvt_f16_f32(vout1x0123), vcvt_f16_f32(vout1x4567));
319 float16x8_t vfp16out2x01234567 = vcombine_f16(vcvt_f16_f32(vout2x0123), vcvt_f16_f32(vout2x4567));
320 float16x8_t vfp16out3x01234567 = vcombine_f16(vcvt_f16_f32(vout3x0123), vcvt_f16_f32(vout3x4567));
321 float16x8_t vfp16out4x01234567 = vcombine_f16(vcvt_f16_f32(vout4x0123), vcvt_f16_f32(vout4x4567));
322 float16x8_t vfp16out5x01234567 = vcombine_f16(vcvt_f16_f32(vout5x0123), vcvt_f16_f32(vout5x4567));
323 float16x8_t vfp16out6x01234567 = vcombine_f16(vcvt_f16_f32(vout6x0123), vcvt_f16_f32(vout6x4567));
324
325 const float16x8_t voutput_min = vreinterpretq_f16_u16(vld1q_dup_u16((const uint16_t*) &params->scalar.min));
326 vfp16out0x01234567 = vmaxq_f16(vfp16out0x01234567, voutput_min);
327 vfp16out1x01234567 = vmaxq_f16(vfp16out1x01234567, voutput_min);
328 vfp16out2x01234567 = vmaxq_f16(vfp16out2x01234567, voutput_min);
329 vfp16out3x01234567 = vmaxq_f16(vfp16out3x01234567, voutput_min);
330 vfp16out4x01234567 = vmaxq_f16(vfp16out4x01234567, voutput_min);
331 vfp16out5x01234567 = vmaxq_f16(vfp16out5x01234567, voutput_min);
332 vfp16out6x01234567 = vmaxq_f16(vfp16out6x01234567, voutput_min);
333 const float16x8_t voutput_max = vreinterpretq_f16_u16(vld1q_dup_u16((const uint16_t*) &params->scalar.max));
334 vfp16out0x01234567 = vminq_f16(vfp16out0x01234567, voutput_max);
335 vfp16out1x01234567 = vminq_f16(vfp16out1x01234567, voutput_max);
336 vfp16out2x01234567 = vminq_f16(vfp16out2x01234567, voutput_max);
337 vfp16out3x01234567 = vminq_f16(vfp16out3x01234567, voutput_max);
338 vfp16out4x01234567 = vminq_f16(vfp16out4x01234567, voutput_max);
339 vfp16out5x01234567 = vminq_f16(vfp16out5x01234567, voutput_max);
340 vfp16out6x01234567 = vminq_f16(vfp16out6x01234567, voutput_max);
341 if XNN_LIKELY(nc >= 8) {
342 vst1q_u16(c0, vreinterpretq_u16_f16(vfp16out0x01234567));
343 vst1q_u16(c1, vreinterpretq_u16_f16(vfp16out1x01234567));
344 vst1q_u16(c2, vreinterpretq_u16_f16(vfp16out2x01234567));
345 vst1q_u16(c3, vreinterpretq_u16_f16(vfp16out3x01234567));
346 vst1q_u16(c4, vreinterpretq_u16_f16(vfp16out4x01234567));
347 vst1q_u16(c5, vreinterpretq_u16_f16(vfp16out5x01234567));
348 vst1q_u16(c6, vreinterpretq_u16_f16(vfp16out6x01234567));
349
350 a0 = (const int8_t*) ((uintptr_t) a0 - kc);
351 a1 = (const int8_t*) ((uintptr_t) a1 - kc);
352 a2 = (const int8_t*) ((uintptr_t) a2 - kc);
353 a3 = (const int8_t*) ((uintptr_t) a3 - kc);
354 a4 = (const int8_t*) ((uintptr_t) a4 - kc);
355 a5 = (const int8_t*) ((uintptr_t) a5 - kc);
356 a6 = (const int8_t*) ((uintptr_t) a6 - kc);
357
358 c0 = (uint16_t*) ((uintptr_t) c0 + cn_stride);
359 c1 = (uint16_t*) ((uintptr_t) c1 + cn_stride);
360 c2 = (uint16_t*) ((uintptr_t) c2 + cn_stride);
361 c3 = (uint16_t*) ((uintptr_t) c3 + cn_stride);
362 c4 = (uint16_t*) ((uintptr_t) c4 + cn_stride);
363 c5 = (uint16_t*) ((uintptr_t) c5 + cn_stride);
364 c6 = (uint16_t*) ((uintptr_t) c6 + cn_stride);
365
366 nc -= 8;
367 } else {
368 float16x4_t vfp16out0x0123 = vget_low_f16(vfp16out0x01234567);
369 float16x4_t vfp16out1x0123 = vget_low_f16(vfp16out1x01234567);
370 float16x4_t vfp16out2x0123 = vget_low_f16(vfp16out2x01234567);
371 float16x4_t vfp16out3x0123 = vget_low_f16(vfp16out3x01234567);
372 float16x4_t vfp16out4x0123 = vget_low_f16(vfp16out4x01234567);
373 float16x4_t vfp16out5x0123 = vget_low_f16(vfp16out5x01234567);
374 float16x4_t vfp16out6x0123 = vget_low_f16(vfp16out6x01234567);
375 if (nc & 4) {
376 vst1_u16(c0, vreinterpret_u16_f16(vfp16out0x0123)); c0 += 4;
377 vst1_u16(c1, vreinterpret_u16_f16(vfp16out1x0123)); c1 += 4;
378 vst1_u16(c2, vreinterpret_u16_f16(vfp16out2x0123)); c2 += 4;
379 vst1_u16(c3, vreinterpret_u16_f16(vfp16out3x0123)); c3 += 4;
380 vst1_u16(c4, vreinterpret_u16_f16(vfp16out4x0123)); c4 += 4;
381 vst1_u16(c5, vreinterpret_u16_f16(vfp16out5x0123)); c5 += 4;
382 vst1_u16(c6, vreinterpret_u16_f16(vfp16out6x0123)); c6 += 4;
383 vfp16out0x0123 = vget_high_f16(vfp16out0x01234567);
384 vfp16out1x0123 = vget_high_f16(vfp16out1x01234567);
385 vfp16out2x0123 = vget_high_f16(vfp16out2x01234567);
386 vfp16out3x0123 = vget_high_f16(vfp16out3x01234567);
387 vfp16out4x0123 = vget_high_f16(vfp16out4x01234567);
388 vfp16out5x0123 = vget_high_f16(vfp16out5x01234567);
389 vfp16out6x0123 = vget_high_f16(vfp16out6x01234567);
390 }
391 if (nc & 2) {
392 vst1_lane_u32((void*) c0, vreinterpret_u32_f16(vfp16out0x0123), 0); c0 += 2;
393 vst1_lane_u32((void*) c1, vreinterpret_u32_f16(vfp16out1x0123), 0); c1 += 2;
394 vst1_lane_u32((void*) c2, vreinterpret_u32_f16(vfp16out2x0123), 0); c2 += 2;
395 vst1_lane_u32((void*) c3, vreinterpret_u32_f16(vfp16out3x0123), 0); c3 += 2;
396 vst1_lane_u32((void*) c4, vreinterpret_u32_f16(vfp16out4x0123), 0); c4 += 2;
397 vst1_lane_u32((void*) c5, vreinterpret_u32_f16(vfp16out5x0123), 0); c5 += 2;
398 vst1_lane_u32((void*) c6, vreinterpret_u32_f16(vfp16out6x0123), 0); c6 += 2;
399 vfp16out0x0123 = vext_f16(vfp16out0x0123, vfp16out0x0123, 2);
400 vfp16out1x0123 = vext_f16(vfp16out1x0123, vfp16out1x0123, 2);
401 vfp16out2x0123 = vext_f16(vfp16out2x0123, vfp16out2x0123, 2);
402 vfp16out3x0123 = vext_f16(vfp16out3x0123, vfp16out3x0123, 2);
403 vfp16out4x0123 = vext_f16(vfp16out4x0123, vfp16out4x0123, 2);
404 vfp16out5x0123 = vext_f16(vfp16out5x0123, vfp16out5x0123, 2);
405 vfp16out6x0123 = vext_f16(vfp16out6x0123, vfp16out6x0123, 2);
406 }
407 if (nc & 1) {
408 vst1_lane_u16(c0, vreinterpret_u16_f16(vfp16out0x0123), 0);
409 vst1_lane_u16(c1, vreinterpret_u16_f16(vfp16out1x0123), 0);
410 vst1_lane_u16(c2, vreinterpret_u16_f16(vfp16out2x0123), 0);
411 vst1_lane_u16(c3, vreinterpret_u16_f16(vfp16out3x0123), 0);
412 vst1_lane_u16(c4, vreinterpret_u16_f16(vfp16out4x0123), 0);
413 vst1_lane_u16(c5, vreinterpret_u16_f16(vfp16out5x0123), 0);
414 vst1_lane_u16(c6, vreinterpret_u16_f16(vfp16out6x0123), 0);
415 }
416 nc = 0;
417 }
418 } while (nc != 0);
419 }