1 // Auto-generated file. Do not edit!
2 // Template: src/qs8-gemm/neon-mlal-lane.c.in
3 // Generator: tools/xngen
4 //
5 // Copyright 2020 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/common.h>
15 #include <xnnpack/gemm.h>
16
17
xnn_qc8_gemm_minmax_fp32_ukernel_6x16__neon_mlal_lane(size_t mr,size_t nc,size_t kc,const int8_t * restrict a,size_t a_stride,const void * restrict w,int8_t * restrict c,size_t cm_stride,size_t cn_stride,const union xnn_qc8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS (1)])18 void xnn_qc8_gemm_minmax_fp32_ukernel_6x16__neon_mlal_lane(
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 int8_t* restrict c,
26 size_t cm_stride,
27 size_t cn_stride,
28 const union xnn_qc8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
29 {
30 assert(mr != 0);
31 assert(mr <= 6);
32 assert(nc != 0);
33 assert(kc != 0);
34 assert(kc % sizeof(int8_t) == 0);
35 assert(a != NULL);
36 assert(w != NULL);
37 assert(c != NULL);
38
39 const int8_t* a0 = a;
40 int8_t* c0 = c;
41 const int8_t* a1 = (const int8_t*) ((uintptr_t) a0 + a_stride);
42 int8_t* c1 = (int8_t*) ((uintptr_t) c0 + cm_stride);
43 if XNN_UNPREDICTABLE(mr < 2) {
44 a1 = a0;
45 c1 = c0;
46 }
47 const int8_t* a2 = (const int8_t*) ((uintptr_t) a1 + a_stride);
48 int8_t* c2 = (int8_t*) ((uintptr_t) c1 + cm_stride);
49 if XNN_UNPREDICTABLE(mr <= 2) {
50 a2 = a1;
51 c2 = c1;
52 }
53 const int8_t* a3 = (const int8_t*) ((uintptr_t) a2 + a_stride);
54 int8_t* c3 = (int8_t*) ((uintptr_t) c2 + cm_stride);
55 if XNN_UNPREDICTABLE(mr < 4) {
56 a3 = a2;
57 c3 = c2;
58 }
59 const int8_t* a4 = (const int8_t*) ((uintptr_t) a3 + a_stride);
60 int8_t* c4 = (int8_t*) ((uintptr_t) c3 + cm_stride);
61 if XNN_UNPREDICTABLE(mr <= 4) {
62 a4 = a3;
63 c4 = c3;
64 }
65 const int8_t* a5 = (const int8_t*) ((uintptr_t) a4 + a_stride);
66 int8_t* c5 = (int8_t*) ((uintptr_t) c4 + cm_stride);
67 if XNN_UNPREDICTABLE(mr != 6) {
68 a5 = a4;
69 c5 = c4;
70 }
71
72 do {
73 int32x4_t vacc0x0123 = vld1q_s32(w); w = (const void*) ((const int32_t*) w + 4);
74 int32x4_t vacc0x4567 = vld1q_s32(w); w = (const void*) ((const int32_t*) w + 4);
75 int32x4_t vacc0x89AB = vld1q_s32(w); w = (const void*) ((const int32_t*) w + 4);
76 int32x4_t vacc0xCDEF = vld1q_s32(w); w = (const void*) ((const int32_t*) w + 4);
77 int32x4_t vacc1x0123 = vacc0x0123;
78 int32x4_t vacc1x4567 = vacc0x4567;
79 int32x4_t vacc1x89AB = vacc0x89AB;
80 int32x4_t vacc1xCDEF = vacc0xCDEF;
81 int32x4_t vacc2x0123 = vacc0x0123;
82 int32x4_t vacc2x4567 = vacc0x4567;
83 int32x4_t vacc2x89AB = vacc0x89AB;
84 int32x4_t vacc2xCDEF = vacc0xCDEF;
85 int32x4_t vacc3x0123 = vacc0x0123;
86 int32x4_t vacc3x4567 = vacc0x4567;
87 int32x4_t vacc3x89AB = vacc0x89AB;
88 int32x4_t vacc3xCDEF = vacc0xCDEF;
89 int32x4_t vacc4x0123 = vacc0x0123;
90 int32x4_t vacc4x4567 = vacc0x4567;
91 int32x4_t vacc4x89AB = vacc0x89AB;
92 int32x4_t vacc4xCDEF = vacc0xCDEF;
93 int32x4_t vacc5x0123 = vacc0x0123;
94 int32x4_t vacc5x4567 = vacc0x4567;
95 int32x4_t vacc5x89AB = vacc0x89AB;
96 int32x4_t vacc5xCDEF = vacc0xCDEF;
97
98 size_t k = kc;
99 while (k >= 8 * sizeof(int8_t)) {
100 const int8x8_t va0 = vld1_s8(a0); a0 += 8;
101 const int16x8_t vxa0 = vmovl_s8(va0);
102 const int8x8_t va1 = vld1_s8(a1); a1 += 8;
103 const int16x8_t vxa1 = vmovl_s8(va1);
104 const int8x8_t va2 = vld1_s8(a2); a2 += 8;
105 const int16x8_t vxa2 = vmovl_s8(va2);
106 const int8x8_t va3 = vld1_s8(a3); a3 += 8;
107 const int16x8_t vxa3 = vmovl_s8(va3);
108 const int8x8_t va4 = vld1_s8(a4); a4 += 8;
109 const int16x8_t vxa4 = vmovl_s8(va4);
110 const int8x8_t va5 = vld1_s8(a5); a5 += 8;
111 const int16x8_t vxa5 = vmovl_s8(va5);
112
113 const int8x8_t vb01234567c0 = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
114 const int16x8_t vxb01234567c0 = vmovl_s8(vb01234567c0);
115
116 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c0), vget_low_s16(vxa0), 0);
117 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c0), vget_low_s16(vxa0), 0);
118 vacc1x0123 = vmlal_lane_s16(vacc1x0123, vget_low_s16(vxb01234567c0), vget_low_s16(vxa1), 0);
119 vacc1x4567 = vmlal_lane_s16(vacc1x4567, vget_high_s16(vxb01234567c0), vget_low_s16(vxa1), 0);
120 vacc2x0123 = vmlal_lane_s16(vacc2x0123, vget_low_s16(vxb01234567c0), vget_low_s16(vxa2), 0);
121 vacc2x4567 = vmlal_lane_s16(vacc2x4567, vget_high_s16(vxb01234567c0), vget_low_s16(vxa2), 0);
122 vacc3x0123 = vmlal_lane_s16(vacc3x0123, vget_low_s16(vxb01234567c0), vget_low_s16(vxa3), 0);
123 vacc3x4567 = vmlal_lane_s16(vacc3x4567, vget_high_s16(vxb01234567c0), vget_low_s16(vxa3), 0);
124 vacc4x0123 = vmlal_lane_s16(vacc4x0123, vget_low_s16(vxb01234567c0), vget_low_s16(vxa4), 0);
125 vacc4x4567 = vmlal_lane_s16(vacc4x4567, vget_high_s16(vxb01234567c0), vget_low_s16(vxa4), 0);
126 vacc5x0123 = vmlal_lane_s16(vacc5x0123, vget_low_s16(vxb01234567c0), vget_low_s16(vxa5), 0);
127 vacc5x4567 = vmlal_lane_s16(vacc5x4567, vget_high_s16(vxb01234567c0), vget_low_s16(vxa5), 0);
128 const int8x8_t vb89ABCDEFc0 = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
129 const int16x8_t vxb89ABCDEFc0 = vmovl_s8(vb89ABCDEFc0);
130
131 vacc0x89AB = vmlal_lane_s16(vacc0x89AB, vget_low_s16(vxb89ABCDEFc0), vget_low_s16(vxa0), 0);
132 vacc0xCDEF = vmlal_lane_s16(vacc0xCDEF, vget_high_s16(vxb89ABCDEFc0), vget_low_s16(vxa0), 0);
133 vacc1x89AB = vmlal_lane_s16(vacc1x89AB, vget_low_s16(vxb89ABCDEFc0), vget_low_s16(vxa1), 0);
134 vacc1xCDEF = vmlal_lane_s16(vacc1xCDEF, vget_high_s16(vxb89ABCDEFc0), vget_low_s16(vxa1), 0);
135 vacc2x89AB = vmlal_lane_s16(vacc2x89AB, vget_low_s16(vxb89ABCDEFc0), vget_low_s16(vxa2), 0);
136 vacc2xCDEF = vmlal_lane_s16(vacc2xCDEF, vget_high_s16(vxb89ABCDEFc0), vget_low_s16(vxa2), 0);
137 vacc3x89AB = vmlal_lane_s16(vacc3x89AB, vget_low_s16(vxb89ABCDEFc0), vget_low_s16(vxa3), 0);
138 vacc3xCDEF = vmlal_lane_s16(vacc3xCDEF, vget_high_s16(vxb89ABCDEFc0), vget_low_s16(vxa3), 0);
139 vacc4x89AB = vmlal_lane_s16(vacc4x89AB, vget_low_s16(vxb89ABCDEFc0), vget_low_s16(vxa4), 0);
140 vacc4xCDEF = vmlal_lane_s16(vacc4xCDEF, vget_high_s16(vxb89ABCDEFc0), vget_low_s16(vxa4), 0);
141 vacc5x89AB = vmlal_lane_s16(vacc5x89AB, vget_low_s16(vxb89ABCDEFc0), vget_low_s16(vxa5), 0);
142 vacc5xCDEF = vmlal_lane_s16(vacc5xCDEF, vget_high_s16(vxb89ABCDEFc0), vget_low_s16(vxa5), 0);
143 const int8x8_t vb01234567c1 = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
144 const int16x8_t vxb01234567c1 = vmovl_s8(vb01234567c1);
145
146 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c1), vget_low_s16(vxa0), 1);
147 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c1), vget_low_s16(vxa0), 1);
148 vacc1x0123 = vmlal_lane_s16(vacc1x0123, vget_low_s16(vxb01234567c1), vget_low_s16(vxa1), 1);
149 vacc1x4567 = vmlal_lane_s16(vacc1x4567, vget_high_s16(vxb01234567c1), vget_low_s16(vxa1), 1);
150 vacc2x0123 = vmlal_lane_s16(vacc2x0123, vget_low_s16(vxb01234567c1), vget_low_s16(vxa2), 1);
151 vacc2x4567 = vmlal_lane_s16(vacc2x4567, vget_high_s16(vxb01234567c1), vget_low_s16(vxa2), 1);
152 vacc3x0123 = vmlal_lane_s16(vacc3x0123, vget_low_s16(vxb01234567c1), vget_low_s16(vxa3), 1);
153 vacc3x4567 = vmlal_lane_s16(vacc3x4567, vget_high_s16(vxb01234567c1), vget_low_s16(vxa3), 1);
154 vacc4x0123 = vmlal_lane_s16(vacc4x0123, vget_low_s16(vxb01234567c1), vget_low_s16(vxa4), 1);
155 vacc4x4567 = vmlal_lane_s16(vacc4x4567, vget_high_s16(vxb01234567c1), vget_low_s16(vxa4), 1);
156 vacc5x0123 = vmlal_lane_s16(vacc5x0123, vget_low_s16(vxb01234567c1), vget_low_s16(vxa5), 1);
157 vacc5x4567 = vmlal_lane_s16(vacc5x4567, vget_high_s16(vxb01234567c1), vget_low_s16(vxa5), 1);
158 const int8x8_t vb89ABCDEFc1 = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
159 const int16x8_t vxb89ABCDEFc1 = vmovl_s8(vb89ABCDEFc1);
160
161 vacc0x89AB = vmlal_lane_s16(vacc0x89AB, vget_low_s16(vxb89ABCDEFc1), vget_low_s16(vxa0), 1);
162 vacc0xCDEF = vmlal_lane_s16(vacc0xCDEF, vget_high_s16(vxb89ABCDEFc1), vget_low_s16(vxa0), 1);
163 vacc1x89AB = vmlal_lane_s16(vacc1x89AB, vget_low_s16(vxb89ABCDEFc1), vget_low_s16(vxa1), 1);
164 vacc1xCDEF = vmlal_lane_s16(vacc1xCDEF, vget_high_s16(vxb89ABCDEFc1), vget_low_s16(vxa1), 1);
165 vacc2x89AB = vmlal_lane_s16(vacc2x89AB, vget_low_s16(vxb89ABCDEFc1), vget_low_s16(vxa2), 1);
166 vacc2xCDEF = vmlal_lane_s16(vacc2xCDEF, vget_high_s16(vxb89ABCDEFc1), vget_low_s16(vxa2), 1);
167 vacc3x89AB = vmlal_lane_s16(vacc3x89AB, vget_low_s16(vxb89ABCDEFc1), vget_low_s16(vxa3), 1);
168 vacc3xCDEF = vmlal_lane_s16(vacc3xCDEF, vget_high_s16(vxb89ABCDEFc1), vget_low_s16(vxa3), 1);
169 vacc4x89AB = vmlal_lane_s16(vacc4x89AB, vget_low_s16(vxb89ABCDEFc1), vget_low_s16(vxa4), 1);
170 vacc4xCDEF = vmlal_lane_s16(vacc4xCDEF, vget_high_s16(vxb89ABCDEFc1), vget_low_s16(vxa4), 1);
171 vacc5x89AB = vmlal_lane_s16(vacc5x89AB, vget_low_s16(vxb89ABCDEFc1), vget_low_s16(vxa5), 1);
172 vacc5xCDEF = vmlal_lane_s16(vacc5xCDEF, vget_high_s16(vxb89ABCDEFc1), vget_low_s16(vxa5), 1);
173 const int8x8_t vb01234567c2 = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
174 const int16x8_t vxb01234567c2 = vmovl_s8(vb01234567c2);
175
176 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c2), vget_low_s16(vxa0), 2);
177 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c2), vget_low_s16(vxa0), 2);
178 vacc1x0123 = vmlal_lane_s16(vacc1x0123, vget_low_s16(vxb01234567c2), vget_low_s16(vxa1), 2);
179 vacc1x4567 = vmlal_lane_s16(vacc1x4567, vget_high_s16(vxb01234567c2), vget_low_s16(vxa1), 2);
180 vacc2x0123 = vmlal_lane_s16(vacc2x0123, vget_low_s16(vxb01234567c2), vget_low_s16(vxa2), 2);
181 vacc2x4567 = vmlal_lane_s16(vacc2x4567, vget_high_s16(vxb01234567c2), vget_low_s16(vxa2), 2);
182 vacc3x0123 = vmlal_lane_s16(vacc3x0123, vget_low_s16(vxb01234567c2), vget_low_s16(vxa3), 2);
183 vacc3x4567 = vmlal_lane_s16(vacc3x4567, vget_high_s16(vxb01234567c2), vget_low_s16(vxa3), 2);
184 vacc4x0123 = vmlal_lane_s16(vacc4x0123, vget_low_s16(vxb01234567c2), vget_low_s16(vxa4), 2);
185 vacc4x4567 = vmlal_lane_s16(vacc4x4567, vget_high_s16(vxb01234567c2), vget_low_s16(vxa4), 2);
186 vacc5x0123 = vmlal_lane_s16(vacc5x0123, vget_low_s16(vxb01234567c2), vget_low_s16(vxa5), 2);
187 vacc5x4567 = vmlal_lane_s16(vacc5x4567, vget_high_s16(vxb01234567c2), vget_low_s16(vxa5), 2);
188 const int8x8_t vb89ABCDEFc2 = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
189 const int16x8_t vxb89ABCDEFc2 = vmovl_s8(vb89ABCDEFc2);
190
191 vacc0x89AB = vmlal_lane_s16(vacc0x89AB, vget_low_s16(vxb89ABCDEFc2), vget_low_s16(vxa0), 2);
192 vacc0xCDEF = vmlal_lane_s16(vacc0xCDEF, vget_high_s16(vxb89ABCDEFc2), vget_low_s16(vxa0), 2);
193 vacc1x89AB = vmlal_lane_s16(vacc1x89AB, vget_low_s16(vxb89ABCDEFc2), vget_low_s16(vxa1), 2);
194 vacc1xCDEF = vmlal_lane_s16(vacc1xCDEF, vget_high_s16(vxb89ABCDEFc2), vget_low_s16(vxa1), 2);
195 vacc2x89AB = vmlal_lane_s16(vacc2x89AB, vget_low_s16(vxb89ABCDEFc2), vget_low_s16(vxa2), 2);
196 vacc2xCDEF = vmlal_lane_s16(vacc2xCDEF, vget_high_s16(vxb89ABCDEFc2), vget_low_s16(vxa2), 2);
197 vacc3x89AB = vmlal_lane_s16(vacc3x89AB, vget_low_s16(vxb89ABCDEFc2), vget_low_s16(vxa3), 2);
198 vacc3xCDEF = vmlal_lane_s16(vacc3xCDEF, vget_high_s16(vxb89ABCDEFc2), vget_low_s16(vxa3), 2);
199 vacc4x89AB = vmlal_lane_s16(vacc4x89AB, vget_low_s16(vxb89ABCDEFc2), vget_low_s16(vxa4), 2);
200 vacc4xCDEF = vmlal_lane_s16(vacc4xCDEF, vget_high_s16(vxb89ABCDEFc2), vget_low_s16(vxa4), 2);
201 vacc5x89AB = vmlal_lane_s16(vacc5x89AB, vget_low_s16(vxb89ABCDEFc2), vget_low_s16(vxa5), 2);
202 vacc5xCDEF = vmlal_lane_s16(vacc5xCDEF, vget_high_s16(vxb89ABCDEFc2), vget_low_s16(vxa5), 2);
203 const int8x8_t vb01234567c3 = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
204 const int16x8_t vxb01234567c3 = vmovl_s8(vb01234567c3);
205
206 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c3), vget_low_s16(vxa0), 3);
207 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c3), vget_low_s16(vxa0), 3);
208 vacc1x0123 = vmlal_lane_s16(vacc1x0123, vget_low_s16(vxb01234567c3), vget_low_s16(vxa1), 3);
209 vacc1x4567 = vmlal_lane_s16(vacc1x4567, vget_high_s16(vxb01234567c3), vget_low_s16(vxa1), 3);
210 vacc2x0123 = vmlal_lane_s16(vacc2x0123, vget_low_s16(vxb01234567c3), vget_low_s16(vxa2), 3);
211 vacc2x4567 = vmlal_lane_s16(vacc2x4567, vget_high_s16(vxb01234567c3), vget_low_s16(vxa2), 3);
212 vacc3x0123 = vmlal_lane_s16(vacc3x0123, vget_low_s16(vxb01234567c3), vget_low_s16(vxa3), 3);
213 vacc3x4567 = vmlal_lane_s16(vacc3x4567, vget_high_s16(vxb01234567c3), vget_low_s16(vxa3), 3);
214 vacc4x0123 = vmlal_lane_s16(vacc4x0123, vget_low_s16(vxb01234567c3), vget_low_s16(vxa4), 3);
215 vacc4x4567 = vmlal_lane_s16(vacc4x4567, vget_high_s16(vxb01234567c3), vget_low_s16(vxa4), 3);
216 vacc5x0123 = vmlal_lane_s16(vacc5x0123, vget_low_s16(vxb01234567c3), vget_low_s16(vxa5), 3);
217 vacc5x4567 = vmlal_lane_s16(vacc5x4567, vget_high_s16(vxb01234567c3), vget_low_s16(vxa5), 3);
218 const int8x8_t vb89ABCDEFc3 = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
219 const int16x8_t vxb89ABCDEFc3 = vmovl_s8(vb89ABCDEFc3);
220
221 vacc0x89AB = vmlal_lane_s16(vacc0x89AB, vget_low_s16(vxb89ABCDEFc3), vget_low_s16(vxa0), 3);
222 vacc0xCDEF = vmlal_lane_s16(vacc0xCDEF, vget_high_s16(vxb89ABCDEFc3), vget_low_s16(vxa0), 3);
223 vacc1x89AB = vmlal_lane_s16(vacc1x89AB, vget_low_s16(vxb89ABCDEFc3), vget_low_s16(vxa1), 3);
224 vacc1xCDEF = vmlal_lane_s16(vacc1xCDEF, vget_high_s16(vxb89ABCDEFc3), vget_low_s16(vxa1), 3);
225 vacc2x89AB = vmlal_lane_s16(vacc2x89AB, vget_low_s16(vxb89ABCDEFc3), vget_low_s16(vxa2), 3);
226 vacc2xCDEF = vmlal_lane_s16(vacc2xCDEF, vget_high_s16(vxb89ABCDEFc3), vget_low_s16(vxa2), 3);
227 vacc3x89AB = vmlal_lane_s16(vacc3x89AB, vget_low_s16(vxb89ABCDEFc3), vget_low_s16(vxa3), 3);
228 vacc3xCDEF = vmlal_lane_s16(vacc3xCDEF, vget_high_s16(vxb89ABCDEFc3), vget_low_s16(vxa3), 3);
229 vacc4x89AB = vmlal_lane_s16(vacc4x89AB, vget_low_s16(vxb89ABCDEFc3), vget_low_s16(vxa4), 3);
230 vacc4xCDEF = vmlal_lane_s16(vacc4xCDEF, vget_high_s16(vxb89ABCDEFc3), vget_low_s16(vxa4), 3);
231 vacc5x89AB = vmlal_lane_s16(vacc5x89AB, vget_low_s16(vxb89ABCDEFc3), vget_low_s16(vxa5), 3);
232 vacc5xCDEF = vmlal_lane_s16(vacc5xCDEF, vget_high_s16(vxb89ABCDEFc3), vget_low_s16(vxa5), 3);
233
234
235 const int8x8_t vb01234567c4 = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
236 const int16x8_t vxb01234567c4 = vmovl_s8(vb01234567c4);
237
238 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c4), vget_high_s16(vxa0), 0);
239 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c4), vget_high_s16(vxa0), 0);
240 vacc1x0123 = vmlal_lane_s16(vacc1x0123, vget_low_s16(vxb01234567c4), vget_high_s16(vxa1), 0);
241 vacc1x4567 = vmlal_lane_s16(vacc1x4567, vget_high_s16(vxb01234567c4), vget_high_s16(vxa1), 0);
242 vacc2x0123 = vmlal_lane_s16(vacc2x0123, vget_low_s16(vxb01234567c4), vget_high_s16(vxa2), 0);
243 vacc2x4567 = vmlal_lane_s16(vacc2x4567, vget_high_s16(vxb01234567c4), vget_high_s16(vxa2), 0);
244 vacc3x0123 = vmlal_lane_s16(vacc3x0123, vget_low_s16(vxb01234567c4), vget_high_s16(vxa3), 0);
245 vacc3x4567 = vmlal_lane_s16(vacc3x4567, vget_high_s16(vxb01234567c4), vget_high_s16(vxa3), 0);
246 vacc4x0123 = vmlal_lane_s16(vacc4x0123, vget_low_s16(vxb01234567c4), vget_high_s16(vxa4), 0);
247 vacc4x4567 = vmlal_lane_s16(vacc4x4567, vget_high_s16(vxb01234567c4), vget_high_s16(vxa4), 0);
248 vacc5x0123 = vmlal_lane_s16(vacc5x0123, vget_low_s16(vxb01234567c4), vget_high_s16(vxa5), 0);
249 vacc5x4567 = vmlal_lane_s16(vacc5x4567, vget_high_s16(vxb01234567c4), vget_high_s16(vxa5), 0);
250 const int8x8_t vb89ABCDEFc4 = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
251 const int16x8_t vxb89ABCDEFc4 = vmovl_s8(vb89ABCDEFc4);
252
253 vacc0x89AB = vmlal_lane_s16(vacc0x89AB, vget_low_s16(vxb89ABCDEFc4), vget_high_s16(vxa0), 0);
254 vacc0xCDEF = vmlal_lane_s16(vacc0xCDEF, vget_high_s16(vxb89ABCDEFc4), vget_high_s16(vxa0), 0);
255 vacc1x89AB = vmlal_lane_s16(vacc1x89AB, vget_low_s16(vxb89ABCDEFc4), vget_high_s16(vxa1), 0);
256 vacc1xCDEF = vmlal_lane_s16(vacc1xCDEF, vget_high_s16(vxb89ABCDEFc4), vget_high_s16(vxa1), 0);
257 vacc2x89AB = vmlal_lane_s16(vacc2x89AB, vget_low_s16(vxb89ABCDEFc4), vget_high_s16(vxa2), 0);
258 vacc2xCDEF = vmlal_lane_s16(vacc2xCDEF, vget_high_s16(vxb89ABCDEFc4), vget_high_s16(vxa2), 0);
259 vacc3x89AB = vmlal_lane_s16(vacc3x89AB, vget_low_s16(vxb89ABCDEFc4), vget_high_s16(vxa3), 0);
260 vacc3xCDEF = vmlal_lane_s16(vacc3xCDEF, vget_high_s16(vxb89ABCDEFc4), vget_high_s16(vxa3), 0);
261 vacc4x89AB = vmlal_lane_s16(vacc4x89AB, vget_low_s16(vxb89ABCDEFc4), vget_high_s16(vxa4), 0);
262 vacc4xCDEF = vmlal_lane_s16(vacc4xCDEF, vget_high_s16(vxb89ABCDEFc4), vget_high_s16(vxa4), 0);
263 vacc5x89AB = vmlal_lane_s16(vacc5x89AB, vget_low_s16(vxb89ABCDEFc4), vget_high_s16(vxa5), 0);
264 vacc5xCDEF = vmlal_lane_s16(vacc5xCDEF, vget_high_s16(vxb89ABCDEFc4), vget_high_s16(vxa5), 0);
265 const int8x8_t vb01234567c5 = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
266 const int16x8_t vxb01234567c5 = vmovl_s8(vb01234567c5);
267
268 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c5), vget_high_s16(vxa0), 1);
269 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c5), vget_high_s16(vxa0), 1);
270 vacc1x0123 = vmlal_lane_s16(vacc1x0123, vget_low_s16(vxb01234567c5), vget_high_s16(vxa1), 1);
271 vacc1x4567 = vmlal_lane_s16(vacc1x4567, vget_high_s16(vxb01234567c5), vget_high_s16(vxa1), 1);
272 vacc2x0123 = vmlal_lane_s16(vacc2x0123, vget_low_s16(vxb01234567c5), vget_high_s16(vxa2), 1);
273 vacc2x4567 = vmlal_lane_s16(vacc2x4567, vget_high_s16(vxb01234567c5), vget_high_s16(vxa2), 1);
274 vacc3x0123 = vmlal_lane_s16(vacc3x0123, vget_low_s16(vxb01234567c5), vget_high_s16(vxa3), 1);
275 vacc3x4567 = vmlal_lane_s16(vacc3x4567, vget_high_s16(vxb01234567c5), vget_high_s16(vxa3), 1);
276 vacc4x0123 = vmlal_lane_s16(vacc4x0123, vget_low_s16(vxb01234567c5), vget_high_s16(vxa4), 1);
277 vacc4x4567 = vmlal_lane_s16(vacc4x4567, vget_high_s16(vxb01234567c5), vget_high_s16(vxa4), 1);
278 vacc5x0123 = vmlal_lane_s16(vacc5x0123, vget_low_s16(vxb01234567c5), vget_high_s16(vxa5), 1);
279 vacc5x4567 = vmlal_lane_s16(vacc5x4567, vget_high_s16(vxb01234567c5), vget_high_s16(vxa5), 1);
280 const int8x8_t vb89ABCDEFc5 = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
281 const int16x8_t vxb89ABCDEFc5 = vmovl_s8(vb89ABCDEFc5);
282
283 vacc0x89AB = vmlal_lane_s16(vacc0x89AB, vget_low_s16(vxb89ABCDEFc5), vget_high_s16(vxa0), 1);
284 vacc0xCDEF = vmlal_lane_s16(vacc0xCDEF, vget_high_s16(vxb89ABCDEFc5), vget_high_s16(vxa0), 1);
285 vacc1x89AB = vmlal_lane_s16(vacc1x89AB, vget_low_s16(vxb89ABCDEFc5), vget_high_s16(vxa1), 1);
286 vacc1xCDEF = vmlal_lane_s16(vacc1xCDEF, vget_high_s16(vxb89ABCDEFc5), vget_high_s16(vxa1), 1);
287 vacc2x89AB = vmlal_lane_s16(vacc2x89AB, vget_low_s16(vxb89ABCDEFc5), vget_high_s16(vxa2), 1);
288 vacc2xCDEF = vmlal_lane_s16(vacc2xCDEF, vget_high_s16(vxb89ABCDEFc5), vget_high_s16(vxa2), 1);
289 vacc3x89AB = vmlal_lane_s16(vacc3x89AB, vget_low_s16(vxb89ABCDEFc5), vget_high_s16(vxa3), 1);
290 vacc3xCDEF = vmlal_lane_s16(vacc3xCDEF, vget_high_s16(vxb89ABCDEFc5), vget_high_s16(vxa3), 1);
291 vacc4x89AB = vmlal_lane_s16(vacc4x89AB, vget_low_s16(vxb89ABCDEFc5), vget_high_s16(vxa4), 1);
292 vacc4xCDEF = vmlal_lane_s16(vacc4xCDEF, vget_high_s16(vxb89ABCDEFc5), vget_high_s16(vxa4), 1);
293 vacc5x89AB = vmlal_lane_s16(vacc5x89AB, vget_low_s16(vxb89ABCDEFc5), vget_high_s16(vxa5), 1);
294 vacc5xCDEF = vmlal_lane_s16(vacc5xCDEF, vget_high_s16(vxb89ABCDEFc5), vget_high_s16(vxa5), 1);
295 const int8x8_t vb01234567c6 = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
296 const int16x8_t vxb01234567c6 = vmovl_s8(vb01234567c6);
297
298 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c6), vget_high_s16(vxa0), 2);
299 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c6), vget_high_s16(vxa0), 2);
300 vacc1x0123 = vmlal_lane_s16(vacc1x0123, vget_low_s16(vxb01234567c6), vget_high_s16(vxa1), 2);
301 vacc1x4567 = vmlal_lane_s16(vacc1x4567, vget_high_s16(vxb01234567c6), vget_high_s16(vxa1), 2);
302 vacc2x0123 = vmlal_lane_s16(vacc2x0123, vget_low_s16(vxb01234567c6), vget_high_s16(vxa2), 2);
303 vacc2x4567 = vmlal_lane_s16(vacc2x4567, vget_high_s16(vxb01234567c6), vget_high_s16(vxa2), 2);
304 vacc3x0123 = vmlal_lane_s16(vacc3x0123, vget_low_s16(vxb01234567c6), vget_high_s16(vxa3), 2);
305 vacc3x4567 = vmlal_lane_s16(vacc3x4567, vget_high_s16(vxb01234567c6), vget_high_s16(vxa3), 2);
306 vacc4x0123 = vmlal_lane_s16(vacc4x0123, vget_low_s16(vxb01234567c6), vget_high_s16(vxa4), 2);
307 vacc4x4567 = vmlal_lane_s16(vacc4x4567, vget_high_s16(vxb01234567c6), vget_high_s16(vxa4), 2);
308 vacc5x0123 = vmlal_lane_s16(vacc5x0123, vget_low_s16(vxb01234567c6), vget_high_s16(vxa5), 2);
309 vacc5x4567 = vmlal_lane_s16(vacc5x4567, vget_high_s16(vxb01234567c6), vget_high_s16(vxa5), 2);
310 const int8x8_t vb89ABCDEFc6 = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
311 const int16x8_t vxb89ABCDEFc6 = vmovl_s8(vb89ABCDEFc6);
312
313 vacc0x89AB = vmlal_lane_s16(vacc0x89AB, vget_low_s16(vxb89ABCDEFc6), vget_high_s16(vxa0), 2);
314 vacc0xCDEF = vmlal_lane_s16(vacc0xCDEF, vget_high_s16(vxb89ABCDEFc6), vget_high_s16(vxa0), 2);
315 vacc1x89AB = vmlal_lane_s16(vacc1x89AB, vget_low_s16(vxb89ABCDEFc6), vget_high_s16(vxa1), 2);
316 vacc1xCDEF = vmlal_lane_s16(vacc1xCDEF, vget_high_s16(vxb89ABCDEFc6), vget_high_s16(vxa1), 2);
317 vacc2x89AB = vmlal_lane_s16(vacc2x89AB, vget_low_s16(vxb89ABCDEFc6), vget_high_s16(vxa2), 2);
318 vacc2xCDEF = vmlal_lane_s16(vacc2xCDEF, vget_high_s16(vxb89ABCDEFc6), vget_high_s16(vxa2), 2);
319 vacc3x89AB = vmlal_lane_s16(vacc3x89AB, vget_low_s16(vxb89ABCDEFc6), vget_high_s16(vxa3), 2);
320 vacc3xCDEF = vmlal_lane_s16(vacc3xCDEF, vget_high_s16(vxb89ABCDEFc6), vget_high_s16(vxa3), 2);
321 vacc4x89AB = vmlal_lane_s16(vacc4x89AB, vget_low_s16(vxb89ABCDEFc6), vget_high_s16(vxa4), 2);
322 vacc4xCDEF = vmlal_lane_s16(vacc4xCDEF, vget_high_s16(vxb89ABCDEFc6), vget_high_s16(vxa4), 2);
323 vacc5x89AB = vmlal_lane_s16(vacc5x89AB, vget_low_s16(vxb89ABCDEFc6), vget_high_s16(vxa5), 2);
324 vacc5xCDEF = vmlal_lane_s16(vacc5xCDEF, vget_high_s16(vxb89ABCDEFc6), vget_high_s16(vxa5), 2);
325 const int8x8_t vb01234567c7 = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
326 const int16x8_t vxb01234567c7 = vmovl_s8(vb01234567c7);
327
328 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c7), vget_high_s16(vxa0), 3);
329 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c7), vget_high_s16(vxa0), 3);
330 vacc1x0123 = vmlal_lane_s16(vacc1x0123, vget_low_s16(vxb01234567c7), vget_high_s16(vxa1), 3);
331 vacc1x4567 = vmlal_lane_s16(vacc1x4567, vget_high_s16(vxb01234567c7), vget_high_s16(vxa1), 3);
332 vacc2x0123 = vmlal_lane_s16(vacc2x0123, vget_low_s16(vxb01234567c7), vget_high_s16(vxa2), 3);
333 vacc2x4567 = vmlal_lane_s16(vacc2x4567, vget_high_s16(vxb01234567c7), vget_high_s16(vxa2), 3);
334 vacc3x0123 = vmlal_lane_s16(vacc3x0123, vget_low_s16(vxb01234567c7), vget_high_s16(vxa3), 3);
335 vacc3x4567 = vmlal_lane_s16(vacc3x4567, vget_high_s16(vxb01234567c7), vget_high_s16(vxa3), 3);
336 vacc4x0123 = vmlal_lane_s16(vacc4x0123, vget_low_s16(vxb01234567c7), vget_high_s16(vxa4), 3);
337 vacc4x4567 = vmlal_lane_s16(vacc4x4567, vget_high_s16(vxb01234567c7), vget_high_s16(vxa4), 3);
338 vacc5x0123 = vmlal_lane_s16(vacc5x0123, vget_low_s16(vxb01234567c7), vget_high_s16(vxa5), 3);
339 vacc5x4567 = vmlal_lane_s16(vacc5x4567, vget_high_s16(vxb01234567c7), vget_high_s16(vxa5), 3);
340 const int8x8_t vb89ABCDEFc7 = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
341 const int16x8_t vxb89ABCDEFc7 = vmovl_s8(vb89ABCDEFc7);
342
343 vacc0x89AB = vmlal_lane_s16(vacc0x89AB, vget_low_s16(vxb89ABCDEFc7), vget_high_s16(vxa0), 3);
344 vacc0xCDEF = vmlal_lane_s16(vacc0xCDEF, vget_high_s16(vxb89ABCDEFc7), vget_high_s16(vxa0), 3);
345 vacc1x89AB = vmlal_lane_s16(vacc1x89AB, vget_low_s16(vxb89ABCDEFc7), vget_high_s16(vxa1), 3);
346 vacc1xCDEF = vmlal_lane_s16(vacc1xCDEF, vget_high_s16(vxb89ABCDEFc7), vget_high_s16(vxa1), 3);
347 vacc2x89AB = vmlal_lane_s16(vacc2x89AB, vget_low_s16(vxb89ABCDEFc7), vget_high_s16(vxa2), 3);
348 vacc2xCDEF = vmlal_lane_s16(vacc2xCDEF, vget_high_s16(vxb89ABCDEFc7), vget_high_s16(vxa2), 3);
349 vacc3x89AB = vmlal_lane_s16(vacc3x89AB, vget_low_s16(vxb89ABCDEFc7), vget_high_s16(vxa3), 3);
350 vacc3xCDEF = vmlal_lane_s16(vacc3xCDEF, vget_high_s16(vxb89ABCDEFc7), vget_high_s16(vxa3), 3);
351 vacc4x89AB = vmlal_lane_s16(vacc4x89AB, vget_low_s16(vxb89ABCDEFc7), vget_high_s16(vxa4), 3);
352 vacc4xCDEF = vmlal_lane_s16(vacc4xCDEF, vget_high_s16(vxb89ABCDEFc7), vget_high_s16(vxa4), 3);
353 vacc5x89AB = vmlal_lane_s16(vacc5x89AB, vget_low_s16(vxb89ABCDEFc7), vget_high_s16(vxa5), 3);
354 vacc5xCDEF = vmlal_lane_s16(vacc5xCDEF, vget_high_s16(vxb89ABCDEFc7), vget_high_s16(vxa5), 3);
355
356 k -= 8 * sizeof(int8_t);
357 }
358 if XNN_UNLIKELY(k != 0) {
359 const int8x8_t va0 = vld1_s8(a0); a0 = (const int8_t*) ((uintptr_t) a0 + k);
360 const int16x8_t vxa0 = vmovl_s8(va0);
361 const int8x8_t va1 = vld1_s8(a1); a1 = (const int8_t*) ((uintptr_t) a1 + k);
362 const int16x8_t vxa1 = vmovl_s8(va1);
363 const int8x8_t va2 = vld1_s8(a2); a2 = (const int8_t*) ((uintptr_t) a2 + k);
364 const int16x8_t vxa2 = vmovl_s8(va2);
365 const int8x8_t va3 = vld1_s8(a3); a3 = (const int8_t*) ((uintptr_t) a3 + k);
366 const int16x8_t vxa3 = vmovl_s8(va3);
367 const int8x8_t va4 = vld1_s8(a4); a4 = (const int8_t*) ((uintptr_t) a4 + k);
368 const int16x8_t vxa4 = vmovl_s8(va4);
369 const int8x8_t va5 = vld1_s8(a5); a5 = (const int8_t*) ((uintptr_t) a5 + k);
370 const int16x8_t vxa5 = vmovl_s8(va5);
371
372 const int8x8_t vb01234567c0 = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
373 const int16x8_t vxb01234567c0 = vmovl_s8(vb01234567c0);
374 const int8x8_t vb89ABCDEFc0 = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
375 const int16x8_t vxb89ABCDEFc0 = vmovl_s8(vb89ABCDEFc0);
376
377 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c0), vget_low_s16(vxa0), 0);
378 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c0), vget_low_s16(vxa0), 0);
379 vacc0x89AB = vmlal_lane_s16(vacc0x89AB, vget_low_s16(vxb89ABCDEFc0), vget_low_s16(vxa0), 0);
380 vacc0xCDEF = vmlal_lane_s16(vacc0xCDEF, vget_high_s16(vxb89ABCDEFc0), vget_low_s16(vxa0), 0);
381 vacc1x0123 = vmlal_lane_s16(vacc1x0123, vget_low_s16(vxb01234567c0), vget_low_s16(vxa1), 0);
382 vacc1x4567 = vmlal_lane_s16(vacc1x4567, vget_high_s16(vxb01234567c0), vget_low_s16(vxa1), 0);
383 vacc1x89AB = vmlal_lane_s16(vacc1x89AB, vget_low_s16(vxb89ABCDEFc0), vget_low_s16(vxa1), 0);
384 vacc1xCDEF = vmlal_lane_s16(vacc1xCDEF, vget_high_s16(vxb89ABCDEFc0), vget_low_s16(vxa1), 0);
385 vacc2x0123 = vmlal_lane_s16(vacc2x0123, vget_low_s16(vxb01234567c0), vget_low_s16(vxa2), 0);
386 vacc2x4567 = vmlal_lane_s16(vacc2x4567, vget_high_s16(vxb01234567c0), vget_low_s16(vxa2), 0);
387 vacc2x89AB = vmlal_lane_s16(vacc2x89AB, vget_low_s16(vxb89ABCDEFc0), vget_low_s16(vxa2), 0);
388 vacc2xCDEF = vmlal_lane_s16(vacc2xCDEF, vget_high_s16(vxb89ABCDEFc0), vget_low_s16(vxa2), 0);
389 vacc3x0123 = vmlal_lane_s16(vacc3x0123, vget_low_s16(vxb01234567c0), vget_low_s16(vxa3), 0);
390 vacc3x4567 = vmlal_lane_s16(vacc3x4567, vget_high_s16(vxb01234567c0), vget_low_s16(vxa3), 0);
391 vacc3x89AB = vmlal_lane_s16(vacc3x89AB, vget_low_s16(vxb89ABCDEFc0), vget_low_s16(vxa3), 0);
392 vacc3xCDEF = vmlal_lane_s16(vacc3xCDEF, vget_high_s16(vxb89ABCDEFc0), vget_low_s16(vxa3), 0);
393 vacc4x0123 = vmlal_lane_s16(vacc4x0123, vget_low_s16(vxb01234567c0), vget_low_s16(vxa4), 0);
394 vacc4x4567 = vmlal_lane_s16(vacc4x4567, vget_high_s16(vxb01234567c0), vget_low_s16(vxa4), 0);
395 vacc4x89AB = vmlal_lane_s16(vacc4x89AB, vget_low_s16(vxb89ABCDEFc0), vget_low_s16(vxa4), 0);
396 vacc4xCDEF = vmlal_lane_s16(vacc4xCDEF, vget_high_s16(vxb89ABCDEFc0), vget_low_s16(vxa4), 0);
397 vacc5x0123 = vmlal_lane_s16(vacc5x0123, vget_low_s16(vxb01234567c0), vget_low_s16(vxa5), 0);
398 vacc5x4567 = vmlal_lane_s16(vacc5x4567, vget_high_s16(vxb01234567c0), vget_low_s16(vxa5), 0);
399 vacc5x89AB = vmlal_lane_s16(vacc5x89AB, vget_low_s16(vxb89ABCDEFc0), vget_low_s16(vxa5), 0);
400 vacc5xCDEF = vmlal_lane_s16(vacc5xCDEF, vget_high_s16(vxb89ABCDEFc0), vget_low_s16(vxa5), 0);
401
402 if (k >= 2 * sizeof(int8_t)) {
403 const int8x8_t vb01234567c1 = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
404 const int16x8_t vxb01234567c1 = vmovl_s8(vb01234567c1);
405 const int8x8_t vb89ABCDEFc1 = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
406 const int16x8_t vxb89ABCDEFc1 = vmovl_s8(vb89ABCDEFc1);
407
408 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c1), vget_low_s16(vxa0), 1);
409 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c1), vget_low_s16(vxa0), 1);
410 vacc0x89AB = vmlal_lane_s16(vacc0x89AB, vget_low_s16(vxb89ABCDEFc1), vget_low_s16(vxa0), 1);
411 vacc0xCDEF = vmlal_lane_s16(vacc0xCDEF, vget_high_s16(vxb89ABCDEFc1), vget_low_s16(vxa0), 1);
412 vacc1x0123 = vmlal_lane_s16(vacc1x0123, vget_low_s16(vxb01234567c1), vget_low_s16(vxa1), 1);
413 vacc1x4567 = vmlal_lane_s16(vacc1x4567, vget_high_s16(vxb01234567c1), vget_low_s16(vxa1), 1);
414 vacc1x89AB = vmlal_lane_s16(vacc1x89AB, vget_low_s16(vxb89ABCDEFc1), vget_low_s16(vxa1), 1);
415 vacc1xCDEF = vmlal_lane_s16(vacc1xCDEF, vget_high_s16(vxb89ABCDEFc1), vget_low_s16(vxa1), 1);
416 vacc2x0123 = vmlal_lane_s16(vacc2x0123, vget_low_s16(vxb01234567c1), vget_low_s16(vxa2), 1);
417 vacc2x4567 = vmlal_lane_s16(vacc2x4567, vget_high_s16(vxb01234567c1), vget_low_s16(vxa2), 1);
418 vacc2x89AB = vmlal_lane_s16(vacc2x89AB, vget_low_s16(vxb89ABCDEFc1), vget_low_s16(vxa2), 1);
419 vacc2xCDEF = vmlal_lane_s16(vacc2xCDEF, vget_high_s16(vxb89ABCDEFc1), vget_low_s16(vxa2), 1);
420 vacc3x0123 = vmlal_lane_s16(vacc3x0123, vget_low_s16(vxb01234567c1), vget_low_s16(vxa3), 1);
421 vacc3x4567 = vmlal_lane_s16(vacc3x4567, vget_high_s16(vxb01234567c1), vget_low_s16(vxa3), 1);
422 vacc3x89AB = vmlal_lane_s16(vacc3x89AB, vget_low_s16(vxb89ABCDEFc1), vget_low_s16(vxa3), 1);
423 vacc3xCDEF = vmlal_lane_s16(vacc3xCDEF, vget_high_s16(vxb89ABCDEFc1), vget_low_s16(vxa3), 1);
424 vacc4x0123 = vmlal_lane_s16(vacc4x0123, vget_low_s16(vxb01234567c1), vget_low_s16(vxa4), 1);
425 vacc4x4567 = vmlal_lane_s16(vacc4x4567, vget_high_s16(vxb01234567c1), vget_low_s16(vxa4), 1);
426 vacc4x89AB = vmlal_lane_s16(vacc4x89AB, vget_low_s16(vxb89ABCDEFc1), vget_low_s16(vxa4), 1);
427 vacc4xCDEF = vmlal_lane_s16(vacc4xCDEF, vget_high_s16(vxb89ABCDEFc1), vget_low_s16(vxa4), 1);
428 vacc5x0123 = vmlal_lane_s16(vacc5x0123, vget_low_s16(vxb01234567c1), vget_low_s16(vxa5), 1);
429 vacc5x4567 = vmlal_lane_s16(vacc5x4567, vget_high_s16(vxb01234567c1), vget_low_s16(vxa5), 1);
430 vacc5x89AB = vmlal_lane_s16(vacc5x89AB, vget_low_s16(vxb89ABCDEFc1), vget_low_s16(vxa5), 1);
431 vacc5xCDEF = vmlal_lane_s16(vacc5xCDEF, vget_high_s16(vxb89ABCDEFc1), vget_low_s16(vxa5), 1);
432
433 if (k > 2 * sizeof(int8_t)) {
434 const int8x8_t vb01234567c2 = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
435 const int16x8_t vxb01234567c2 = vmovl_s8(vb01234567c2);
436 const int8x8_t vb89ABCDEFc2 = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
437 const int16x8_t vxb89ABCDEFc2 = vmovl_s8(vb89ABCDEFc2);
438
439 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c2), vget_low_s16(vxa0), 2);
440 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c2), vget_low_s16(vxa0), 2);
441 vacc0x89AB = vmlal_lane_s16(vacc0x89AB, vget_low_s16(vxb89ABCDEFc2), vget_low_s16(vxa0), 2);
442 vacc0xCDEF = vmlal_lane_s16(vacc0xCDEF, vget_high_s16(vxb89ABCDEFc2), vget_low_s16(vxa0), 2);
443 vacc1x0123 = vmlal_lane_s16(vacc1x0123, vget_low_s16(vxb01234567c2), vget_low_s16(vxa1), 2);
444 vacc1x4567 = vmlal_lane_s16(vacc1x4567, vget_high_s16(vxb01234567c2), vget_low_s16(vxa1), 2);
445 vacc1x89AB = vmlal_lane_s16(vacc1x89AB, vget_low_s16(vxb89ABCDEFc2), vget_low_s16(vxa1), 2);
446 vacc1xCDEF = vmlal_lane_s16(vacc1xCDEF, vget_high_s16(vxb89ABCDEFc2), vget_low_s16(vxa1), 2);
447 vacc2x0123 = vmlal_lane_s16(vacc2x0123, vget_low_s16(vxb01234567c2), vget_low_s16(vxa2), 2);
448 vacc2x4567 = vmlal_lane_s16(vacc2x4567, vget_high_s16(vxb01234567c2), vget_low_s16(vxa2), 2);
449 vacc2x89AB = vmlal_lane_s16(vacc2x89AB, vget_low_s16(vxb89ABCDEFc2), vget_low_s16(vxa2), 2);
450 vacc2xCDEF = vmlal_lane_s16(vacc2xCDEF, vget_high_s16(vxb89ABCDEFc2), vget_low_s16(vxa2), 2);
451 vacc3x0123 = vmlal_lane_s16(vacc3x0123, vget_low_s16(vxb01234567c2), vget_low_s16(vxa3), 2);
452 vacc3x4567 = vmlal_lane_s16(vacc3x4567, vget_high_s16(vxb01234567c2), vget_low_s16(vxa3), 2);
453 vacc3x89AB = vmlal_lane_s16(vacc3x89AB, vget_low_s16(vxb89ABCDEFc2), vget_low_s16(vxa3), 2);
454 vacc3xCDEF = vmlal_lane_s16(vacc3xCDEF, vget_high_s16(vxb89ABCDEFc2), vget_low_s16(vxa3), 2);
455 vacc4x0123 = vmlal_lane_s16(vacc4x0123, vget_low_s16(vxb01234567c2), vget_low_s16(vxa4), 2);
456 vacc4x4567 = vmlal_lane_s16(vacc4x4567, vget_high_s16(vxb01234567c2), vget_low_s16(vxa4), 2);
457 vacc4x89AB = vmlal_lane_s16(vacc4x89AB, vget_low_s16(vxb89ABCDEFc2), vget_low_s16(vxa4), 2);
458 vacc4xCDEF = vmlal_lane_s16(vacc4xCDEF, vget_high_s16(vxb89ABCDEFc2), vget_low_s16(vxa4), 2);
459 vacc5x0123 = vmlal_lane_s16(vacc5x0123, vget_low_s16(vxb01234567c2), vget_low_s16(vxa5), 2);
460 vacc5x4567 = vmlal_lane_s16(vacc5x4567, vget_high_s16(vxb01234567c2), vget_low_s16(vxa5), 2);
461 vacc5x89AB = vmlal_lane_s16(vacc5x89AB, vget_low_s16(vxb89ABCDEFc2), vget_low_s16(vxa5), 2);
462 vacc5xCDEF = vmlal_lane_s16(vacc5xCDEF, vget_high_s16(vxb89ABCDEFc2), vget_low_s16(vxa5), 2);
463
464 if (k >= 4 * sizeof(int8_t)) {
465 const int8x8_t vb01234567c3 = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
466 const int16x8_t vxb01234567c3 = vmovl_s8(vb01234567c3);
467 const int8x8_t vb89ABCDEFc3 = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
468 const int16x8_t vxb89ABCDEFc3 = vmovl_s8(vb89ABCDEFc3);
469
470 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c3), vget_low_s16(vxa0), 3);
471 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c3), vget_low_s16(vxa0), 3);
472 vacc0x89AB = vmlal_lane_s16(vacc0x89AB, vget_low_s16(vxb89ABCDEFc3), vget_low_s16(vxa0), 3);
473 vacc0xCDEF = vmlal_lane_s16(vacc0xCDEF, vget_high_s16(vxb89ABCDEFc3), vget_low_s16(vxa0), 3);
474 vacc1x0123 = vmlal_lane_s16(vacc1x0123, vget_low_s16(vxb01234567c3), vget_low_s16(vxa1), 3);
475 vacc1x4567 = vmlal_lane_s16(vacc1x4567, vget_high_s16(vxb01234567c3), vget_low_s16(vxa1), 3);
476 vacc1x89AB = vmlal_lane_s16(vacc1x89AB, vget_low_s16(vxb89ABCDEFc3), vget_low_s16(vxa1), 3);
477 vacc1xCDEF = vmlal_lane_s16(vacc1xCDEF, vget_high_s16(vxb89ABCDEFc3), vget_low_s16(vxa1), 3);
478 vacc2x0123 = vmlal_lane_s16(vacc2x0123, vget_low_s16(vxb01234567c3), vget_low_s16(vxa2), 3);
479 vacc2x4567 = vmlal_lane_s16(vacc2x4567, vget_high_s16(vxb01234567c3), vget_low_s16(vxa2), 3);
480 vacc2x89AB = vmlal_lane_s16(vacc2x89AB, vget_low_s16(vxb89ABCDEFc3), vget_low_s16(vxa2), 3);
481 vacc2xCDEF = vmlal_lane_s16(vacc2xCDEF, vget_high_s16(vxb89ABCDEFc3), vget_low_s16(vxa2), 3);
482 vacc3x0123 = vmlal_lane_s16(vacc3x0123, vget_low_s16(vxb01234567c3), vget_low_s16(vxa3), 3);
483 vacc3x4567 = vmlal_lane_s16(vacc3x4567, vget_high_s16(vxb01234567c3), vget_low_s16(vxa3), 3);
484 vacc3x89AB = vmlal_lane_s16(vacc3x89AB, vget_low_s16(vxb89ABCDEFc3), vget_low_s16(vxa3), 3);
485 vacc3xCDEF = vmlal_lane_s16(vacc3xCDEF, vget_high_s16(vxb89ABCDEFc3), vget_low_s16(vxa3), 3);
486 vacc4x0123 = vmlal_lane_s16(vacc4x0123, vget_low_s16(vxb01234567c3), vget_low_s16(vxa4), 3);
487 vacc4x4567 = vmlal_lane_s16(vacc4x4567, vget_high_s16(vxb01234567c3), vget_low_s16(vxa4), 3);
488 vacc4x89AB = vmlal_lane_s16(vacc4x89AB, vget_low_s16(vxb89ABCDEFc3), vget_low_s16(vxa4), 3);
489 vacc4xCDEF = vmlal_lane_s16(vacc4xCDEF, vget_high_s16(vxb89ABCDEFc3), vget_low_s16(vxa4), 3);
490 vacc5x0123 = vmlal_lane_s16(vacc5x0123, vget_low_s16(vxb01234567c3), vget_low_s16(vxa5), 3);
491 vacc5x4567 = vmlal_lane_s16(vacc5x4567, vget_high_s16(vxb01234567c3), vget_low_s16(vxa5), 3);
492 vacc5x89AB = vmlal_lane_s16(vacc5x89AB, vget_low_s16(vxb89ABCDEFc3), vget_low_s16(vxa5), 3);
493 vacc5xCDEF = vmlal_lane_s16(vacc5xCDEF, vget_high_s16(vxb89ABCDEFc3), vget_low_s16(vxa5), 3);
494
495 if (k > 4 * sizeof(int8_t)) {
496 const int8x8_t vb01234567c4 = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
497 const int16x8_t vxb01234567c4 = vmovl_s8(vb01234567c4);
498 const int8x8_t vb89ABCDEFc4 = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
499 const int16x8_t vxb89ABCDEFc4 = vmovl_s8(vb89ABCDEFc4);
500
501 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c4), vget_high_s16(vxa0), 0);
502 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c4), vget_high_s16(vxa0), 0);
503 vacc0x89AB = vmlal_lane_s16(vacc0x89AB, vget_low_s16(vxb89ABCDEFc4), vget_high_s16(vxa0), 0);
504 vacc0xCDEF = vmlal_lane_s16(vacc0xCDEF, vget_high_s16(vxb89ABCDEFc4), vget_high_s16(vxa0), 0);
505 vacc1x0123 = vmlal_lane_s16(vacc1x0123, vget_low_s16(vxb01234567c4), vget_high_s16(vxa1), 0);
506 vacc1x4567 = vmlal_lane_s16(vacc1x4567, vget_high_s16(vxb01234567c4), vget_high_s16(vxa1), 0);
507 vacc1x89AB = vmlal_lane_s16(vacc1x89AB, vget_low_s16(vxb89ABCDEFc4), vget_high_s16(vxa1), 0);
508 vacc1xCDEF = vmlal_lane_s16(vacc1xCDEF, vget_high_s16(vxb89ABCDEFc4), vget_high_s16(vxa1), 0);
509 vacc2x0123 = vmlal_lane_s16(vacc2x0123, vget_low_s16(vxb01234567c4), vget_high_s16(vxa2), 0);
510 vacc2x4567 = vmlal_lane_s16(vacc2x4567, vget_high_s16(vxb01234567c4), vget_high_s16(vxa2), 0);
511 vacc2x89AB = vmlal_lane_s16(vacc2x89AB, vget_low_s16(vxb89ABCDEFc4), vget_high_s16(vxa2), 0);
512 vacc2xCDEF = vmlal_lane_s16(vacc2xCDEF, vget_high_s16(vxb89ABCDEFc4), vget_high_s16(vxa2), 0);
513 vacc3x0123 = vmlal_lane_s16(vacc3x0123, vget_low_s16(vxb01234567c4), vget_high_s16(vxa3), 0);
514 vacc3x4567 = vmlal_lane_s16(vacc3x4567, vget_high_s16(vxb01234567c4), vget_high_s16(vxa3), 0);
515 vacc3x89AB = vmlal_lane_s16(vacc3x89AB, vget_low_s16(vxb89ABCDEFc4), vget_high_s16(vxa3), 0);
516 vacc3xCDEF = vmlal_lane_s16(vacc3xCDEF, vget_high_s16(vxb89ABCDEFc4), vget_high_s16(vxa3), 0);
517 vacc4x0123 = vmlal_lane_s16(vacc4x0123, vget_low_s16(vxb01234567c4), vget_high_s16(vxa4), 0);
518 vacc4x4567 = vmlal_lane_s16(vacc4x4567, vget_high_s16(vxb01234567c4), vget_high_s16(vxa4), 0);
519 vacc4x89AB = vmlal_lane_s16(vacc4x89AB, vget_low_s16(vxb89ABCDEFc4), vget_high_s16(vxa4), 0);
520 vacc4xCDEF = vmlal_lane_s16(vacc4xCDEF, vget_high_s16(vxb89ABCDEFc4), vget_high_s16(vxa4), 0);
521 vacc5x0123 = vmlal_lane_s16(vacc5x0123, vget_low_s16(vxb01234567c4), vget_high_s16(vxa5), 0);
522 vacc5x4567 = vmlal_lane_s16(vacc5x4567, vget_high_s16(vxb01234567c4), vget_high_s16(vxa5), 0);
523 vacc5x89AB = vmlal_lane_s16(vacc5x89AB, vget_low_s16(vxb89ABCDEFc4), vget_high_s16(vxa5), 0);
524 vacc5xCDEF = vmlal_lane_s16(vacc5xCDEF, vget_high_s16(vxb89ABCDEFc4), vget_high_s16(vxa5), 0);
525
526 if (k >= 6 * sizeof(int8_t)) {
527 const int8x8_t vb01234567c5 = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
528 const int16x8_t vxb01234567c5 = vmovl_s8(vb01234567c5);
529 const int8x8_t vb89ABCDEFc5 = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
530 const int16x8_t vxb89ABCDEFc5 = vmovl_s8(vb89ABCDEFc5);
531
532 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c5), vget_high_s16(vxa0), 1);
533 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c5), vget_high_s16(vxa0), 1);
534 vacc0x89AB = vmlal_lane_s16(vacc0x89AB, vget_low_s16(vxb89ABCDEFc5), vget_high_s16(vxa0), 1);
535 vacc0xCDEF = vmlal_lane_s16(vacc0xCDEF, vget_high_s16(vxb89ABCDEFc5), vget_high_s16(vxa0), 1);
536 vacc1x0123 = vmlal_lane_s16(vacc1x0123, vget_low_s16(vxb01234567c5), vget_high_s16(vxa1), 1);
537 vacc1x4567 = vmlal_lane_s16(vacc1x4567, vget_high_s16(vxb01234567c5), vget_high_s16(vxa1), 1);
538 vacc1x89AB = vmlal_lane_s16(vacc1x89AB, vget_low_s16(vxb89ABCDEFc5), vget_high_s16(vxa1), 1);
539 vacc1xCDEF = vmlal_lane_s16(vacc1xCDEF, vget_high_s16(vxb89ABCDEFc5), vget_high_s16(vxa1), 1);
540 vacc2x0123 = vmlal_lane_s16(vacc2x0123, vget_low_s16(vxb01234567c5), vget_high_s16(vxa2), 1);
541 vacc2x4567 = vmlal_lane_s16(vacc2x4567, vget_high_s16(vxb01234567c5), vget_high_s16(vxa2), 1);
542 vacc2x89AB = vmlal_lane_s16(vacc2x89AB, vget_low_s16(vxb89ABCDEFc5), vget_high_s16(vxa2), 1);
543 vacc2xCDEF = vmlal_lane_s16(vacc2xCDEF, vget_high_s16(vxb89ABCDEFc5), vget_high_s16(vxa2), 1);
544 vacc3x0123 = vmlal_lane_s16(vacc3x0123, vget_low_s16(vxb01234567c5), vget_high_s16(vxa3), 1);
545 vacc3x4567 = vmlal_lane_s16(vacc3x4567, vget_high_s16(vxb01234567c5), vget_high_s16(vxa3), 1);
546 vacc3x89AB = vmlal_lane_s16(vacc3x89AB, vget_low_s16(vxb89ABCDEFc5), vget_high_s16(vxa3), 1);
547 vacc3xCDEF = vmlal_lane_s16(vacc3xCDEF, vget_high_s16(vxb89ABCDEFc5), vget_high_s16(vxa3), 1);
548 vacc4x0123 = vmlal_lane_s16(vacc4x0123, vget_low_s16(vxb01234567c5), vget_high_s16(vxa4), 1);
549 vacc4x4567 = vmlal_lane_s16(vacc4x4567, vget_high_s16(vxb01234567c5), vget_high_s16(vxa4), 1);
550 vacc4x89AB = vmlal_lane_s16(vacc4x89AB, vget_low_s16(vxb89ABCDEFc5), vget_high_s16(vxa4), 1);
551 vacc4xCDEF = vmlal_lane_s16(vacc4xCDEF, vget_high_s16(vxb89ABCDEFc5), vget_high_s16(vxa4), 1);
552 vacc5x0123 = vmlal_lane_s16(vacc5x0123, vget_low_s16(vxb01234567c5), vget_high_s16(vxa5), 1);
553 vacc5x4567 = vmlal_lane_s16(vacc5x4567, vget_high_s16(vxb01234567c5), vget_high_s16(vxa5), 1);
554 vacc5x89AB = vmlal_lane_s16(vacc5x89AB, vget_low_s16(vxb89ABCDEFc5), vget_high_s16(vxa5), 1);
555 vacc5xCDEF = vmlal_lane_s16(vacc5xCDEF, vget_high_s16(vxb89ABCDEFc5), vget_high_s16(vxa5), 1);
556
557 if (k > 6 * sizeof(int8_t)) {
558 const int8x8_t vb01234567c6 = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
559 const int16x8_t vxb01234567c6 = vmovl_s8(vb01234567c6);
560 const int8x8_t vb89ABCDEFc6 = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
561 const int16x8_t vxb89ABCDEFc6 = vmovl_s8(vb89ABCDEFc6);
562
563 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c6), vget_high_s16(vxa0), 2);
564 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c6), vget_high_s16(vxa0), 2);
565 vacc0x89AB = vmlal_lane_s16(vacc0x89AB, vget_low_s16(vxb89ABCDEFc6), vget_high_s16(vxa0), 2);
566 vacc0xCDEF = vmlal_lane_s16(vacc0xCDEF, vget_high_s16(vxb89ABCDEFc6), vget_high_s16(vxa0), 2);
567 vacc1x0123 = vmlal_lane_s16(vacc1x0123, vget_low_s16(vxb01234567c6), vget_high_s16(vxa1), 2);
568 vacc1x4567 = vmlal_lane_s16(vacc1x4567, vget_high_s16(vxb01234567c6), vget_high_s16(vxa1), 2);
569 vacc1x89AB = vmlal_lane_s16(vacc1x89AB, vget_low_s16(vxb89ABCDEFc6), vget_high_s16(vxa1), 2);
570 vacc1xCDEF = vmlal_lane_s16(vacc1xCDEF, vget_high_s16(vxb89ABCDEFc6), vget_high_s16(vxa1), 2);
571 vacc2x0123 = vmlal_lane_s16(vacc2x0123, vget_low_s16(vxb01234567c6), vget_high_s16(vxa2), 2);
572 vacc2x4567 = vmlal_lane_s16(vacc2x4567, vget_high_s16(vxb01234567c6), vget_high_s16(vxa2), 2);
573 vacc2x89AB = vmlal_lane_s16(vacc2x89AB, vget_low_s16(vxb89ABCDEFc6), vget_high_s16(vxa2), 2);
574 vacc2xCDEF = vmlal_lane_s16(vacc2xCDEF, vget_high_s16(vxb89ABCDEFc6), vget_high_s16(vxa2), 2);
575 vacc3x0123 = vmlal_lane_s16(vacc3x0123, vget_low_s16(vxb01234567c6), vget_high_s16(vxa3), 2);
576 vacc3x4567 = vmlal_lane_s16(vacc3x4567, vget_high_s16(vxb01234567c6), vget_high_s16(vxa3), 2);
577 vacc3x89AB = vmlal_lane_s16(vacc3x89AB, vget_low_s16(vxb89ABCDEFc6), vget_high_s16(vxa3), 2);
578 vacc3xCDEF = vmlal_lane_s16(vacc3xCDEF, vget_high_s16(vxb89ABCDEFc6), vget_high_s16(vxa3), 2);
579 vacc4x0123 = vmlal_lane_s16(vacc4x0123, vget_low_s16(vxb01234567c6), vget_high_s16(vxa4), 2);
580 vacc4x4567 = vmlal_lane_s16(vacc4x4567, vget_high_s16(vxb01234567c6), vget_high_s16(vxa4), 2);
581 vacc4x89AB = vmlal_lane_s16(vacc4x89AB, vget_low_s16(vxb89ABCDEFc6), vget_high_s16(vxa4), 2);
582 vacc4xCDEF = vmlal_lane_s16(vacc4xCDEF, vget_high_s16(vxb89ABCDEFc6), vget_high_s16(vxa4), 2);
583 vacc5x0123 = vmlal_lane_s16(vacc5x0123, vget_low_s16(vxb01234567c6), vget_high_s16(vxa5), 2);
584 vacc5x4567 = vmlal_lane_s16(vacc5x4567, vget_high_s16(vxb01234567c6), vget_high_s16(vxa5), 2);
585 vacc5x89AB = vmlal_lane_s16(vacc5x89AB, vget_low_s16(vxb89ABCDEFc6), vget_high_s16(vxa5), 2);
586 vacc5xCDEF = vmlal_lane_s16(vacc5xCDEF, vget_high_s16(vxb89ABCDEFc6), vget_high_s16(vxa5), 2);
587 }
588 }
589 }
590 }
591 }
592 }
593 }
594
595 float32x4_t vfpacc0x0123 = vcvtq_f32_s32(vacc0x0123);
596 float32x4_t vfpacc0x4567 = vcvtq_f32_s32(vacc0x4567);
597 float32x4_t vfpacc0x89AB = vcvtq_f32_s32(vacc0x89AB);
598 float32x4_t vfpacc0xCDEF = vcvtq_f32_s32(vacc0xCDEF);
599 float32x4_t vfpacc1x0123 = vcvtq_f32_s32(vacc1x0123);
600 float32x4_t vfpacc1x4567 = vcvtq_f32_s32(vacc1x4567);
601 float32x4_t vfpacc1x89AB = vcvtq_f32_s32(vacc1x89AB);
602 float32x4_t vfpacc1xCDEF = vcvtq_f32_s32(vacc1xCDEF);
603 float32x4_t vfpacc2x0123 = vcvtq_f32_s32(vacc2x0123);
604 float32x4_t vfpacc2x4567 = vcvtq_f32_s32(vacc2x4567);
605 float32x4_t vfpacc2x89AB = vcvtq_f32_s32(vacc2x89AB);
606 float32x4_t vfpacc2xCDEF = vcvtq_f32_s32(vacc2xCDEF);
607 float32x4_t vfpacc3x0123 = vcvtq_f32_s32(vacc3x0123);
608 float32x4_t vfpacc3x4567 = vcvtq_f32_s32(vacc3x4567);
609 float32x4_t vfpacc3x89AB = vcvtq_f32_s32(vacc3x89AB);
610 float32x4_t vfpacc3xCDEF = vcvtq_f32_s32(vacc3xCDEF);
611 float32x4_t vfpacc4x0123 = vcvtq_f32_s32(vacc4x0123);
612 float32x4_t vfpacc4x4567 = vcvtq_f32_s32(vacc4x4567);
613 float32x4_t vfpacc4x89AB = vcvtq_f32_s32(vacc4x89AB);
614 float32x4_t vfpacc4xCDEF = vcvtq_f32_s32(vacc4xCDEF);
615 float32x4_t vfpacc5x0123 = vcvtq_f32_s32(vacc5x0123);
616 float32x4_t vfpacc5x4567 = vcvtq_f32_s32(vacc5x4567);
617 float32x4_t vfpacc5x89AB = vcvtq_f32_s32(vacc5x89AB);
618 float32x4_t vfpacc5xCDEF = vcvtq_f32_s32(vacc5xCDEF);
619
620 const float32x4_t vscale0123 = vld1q_f32((const float*) w); w = (const void*) ((const float*) w + 4);
621 vfpacc0x0123 = vmulq_f32(vfpacc0x0123, vscale0123);
622 vfpacc1x0123 = vmulq_f32(vfpacc1x0123, vscale0123);
623 vfpacc2x0123 = vmulq_f32(vfpacc2x0123, vscale0123);
624 vfpacc3x0123 = vmulq_f32(vfpacc3x0123, vscale0123);
625 vfpacc4x0123 = vmulq_f32(vfpacc4x0123, vscale0123);
626 vfpacc5x0123 = vmulq_f32(vfpacc5x0123, vscale0123);
627 const float32x4_t vscale4567 = vld1q_f32((const float*) w); w = (const void*) ((const float*) w + 4);
628 vfpacc0x4567 = vmulq_f32(vfpacc0x4567, vscale4567);
629 vfpacc1x4567 = vmulq_f32(vfpacc1x4567, vscale4567);
630 vfpacc2x4567 = vmulq_f32(vfpacc2x4567, vscale4567);
631 vfpacc3x4567 = vmulq_f32(vfpacc3x4567, vscale4567);
632 vfpacc4x4567 = vmulq_f32(vfpacc4x4567, vscale4567);
633 vfpacc5x4567 = vmulq_f32(vfpacc5x4567, vscale4567);
634 const float32x4_t vscale89AB = vld1q_f32((const float*) w); w = (const void*) ((const float*) w + 4);
635 vfpacc0x89AB = vmulq_f32(vfpacc0x89AB, vscale89AB);
636 vfpacc1x89AB = vmulq_f32(vfpacc1x89AB, vscale89AB);
637 vfpacc2x89AB = vmulq_f32(vfpacc2x89AB, vscale89AB);
638 vfpacc3x89AB = vmulq_f32(vfpacc3x89AB, vscale89AB);
639 vfpacc4x89AB = vmulq_f32(vfpacc4x89AB, vscale89AB);
640 vfpacc5x89AB = vmulq_f32(vfpacc5x89AB, vscale89AB);
641 const float32x4_t vscaleCDEF = vld1q_f32((const float*) w); w = (const void*) ((const float*) w + 4);
642 vfpacc0xCDEF = vmulq_f32(vfpacc0xCDEF, vscaleCDEF);
643 vfpacc1xCDEF = vmulq_f32(vfpacc1xCDEF, vscaleCDEF);
644 vfpacc2xCDEF = vmulq_f32(vfpacc2xCDEF, vscaleCDEF);
645 vfpacc3xCDEF = vmulq_f32(vfpacc3xCDEF, vscaleCDEF);
646 vfpacc4xCDEF = vmulq_f32(vfpacc4xCDEF, vscaleCDEF);
647 vfpacc5xCDEF = vmulq_f32(vfpacc5xCDEF, vscaleCDEF);
648
649 const float32x4_t vmagic_bias = vld1q_dup_f32(¶ms->fp32_neon.magic_bias);
650 vacc0x0123 = vreinterpretq_s32_f32(vaddq_f32(vfpacc0x0123, vmagic_bias));
651 vacc0x4567 = vreinterpretq_s32_f32(vaddq_f32(vfpacc0x4567, vmagic_bias));
652 vacc0x89AB = vreinterpretq_s32_f32(vaddq_f32(vfpacc0x89AB, vmagic_bias));
653 vacc0xCDEF = vreinterpretq_s32_f32(vaddq_f32(vfpacc0xCDEF, vmagic_bias));
654 vacc1x0123 = vreinterpretq_s32_f32(vaddq_f32(vfpacc1x0123, vmagic_bias));
655 vacc1x4567 = vreinterpretq_s32_f32(vaddq_f32(vfpacc1x4567, vmagic_bias));
656 vacc1x89AB = vreinterpretq_s32_f32(vaddq_f32(vfpacc1x89AB, vmagic_bias));
657 vacc1xCDEF = vreinterpretq_s32_f32(vaddq_f32(vfpacc1xCDEF, vmagic_bias));
658 vacc2x0123 = vreinterpretq_s32_f32(vaddq_f32(vfpacc2x0123, vmagic_bias));
659 vacc2x4567 = vreinterpretq_s32_f32(vaddq_f32(vfpacc2x4567, vmagic_bias));
660 vacc2x89AB = vreinterpretq_s32_f32(vaddq_f32(vfpacc2x89AB, vmagic_bias));
661 vacc2xCDEF = vreinterpretq_s32_f32(vaddq_f32(vfpacc2xCDEF, vmagic_bias));
662 vacc3x0123 = vreinterpretq_s32_f32(vaddq_f32(vfpacc3x0123, vmagic_bias));
663 vacc3x4567 = vreinterpretq_s32_f32(vaddq_f32(vfpacc3x4567, vmagic_bias));
664 vacc3x89AB = vreinterpretq_s32_f32(vaddq_f32(vfpacc3x89AB, vmagic_bias));
665 vacc3xCDEF = vreinterpretq_s32_f32(vaddq_f32(vfpacc3xCDEF, vmagic_bias));
666 vacc4x0123 = vreinterpretq_s32_f32(vaddq_f32(vfpacc4x0123, vmagic_bias));
667 vacc4x4567 = vreinterpretq_s32_f32(vaddq_f32(vfpacc4x4567, vmagic_bias));
668 vacc4x89AB = vreinterpretq_s32_f32(vaddq_f32(vfpacc4x89AB, vmagic_bias));
669 vacc4xCDEF = vreinterpretq_s32_f32(vaddq_f32(vfpacc4xCDEF, vmagic_bias));
670 vacc5x0123 = vreinterpretq_s32_f32(vaddq_f32(vfpacc5x0123, vmagic_bias));
671 vacc5x4567 = vreinterpretq_s32_f32(vaddq_f32(vfpacc5x4567, vmagic_bias));
672 vacc5x89AB = vreinterpretq_s32_f32(vaddq_f32(vfpacc5x89AB, vmagic_bias));
673 vacc5xCDEF = vreinterpretq_s32_f32(vaddq_f32(vfpacc5xCDEF, vmagic_bias));
674
675 const int32x4_t vmagic_bias_less_output_zero_point = vld1q_dup_s32(¶ms->fp32_neon.magic_bias_less_output_zero_point);
676 vacc0x0123 = vqsubq_s32(vacc0x0123, vmagic_bias_less_output_zero_point);
677 vacc0x4567 = vqsubq_s32(vacc0x4567, vmagic_bias_less_output_zero_point);
678 vacc0x89AB = vqsubq_s32(vacc0x89AB, vmagic_bias_less_output_zero_point);
679 vacc0xCDEF = vqsubq_s32(vacc0xCDEF, vmagic_bias_less_output_zero_point);
680 vacc1x0123 = vqsubq_s32(vacc1x0123, vmagic_bias_less_output_zero_point);
681 vacc1x4567 = vqsubq_s32(vacc1x4567, vmagic_bias_less_output_zero_point);
682 vacc1x89AB = vqsubq_s32(vacc1x89AB, vmagic_bias_less_output_zero_point);
683 vacc1xCDEF = vqsubq_s32(vacc1xCDEF, vmagic_bias_less_output_zero_point);
684 vacc2x0123 = vqsubq_s32(vacc2x0123, vmagic_bias_less_output_zero_point);
685 vacc2x4567 = vqsubq_s32(vacc2x4567, vmagic_bias_less_output_zero_point);
686 vacc2x89AB = vqsubq_s32(vacc2x89AB, vmagic_bias_less_output_zero_point);
687 vacc2xCDEF = vqsubq_s32(vacc2xCDEF, vmagic_bias_less_output_zero_point);
688 vacc3x0123 = vqsubq_s32(vacc3x0123, vmagic_bias_less_output_zero_point);
689 vacc3x4567 = vqsubq_s32(vacc3x4567, vmagic_bias_less_output_zero_point);
690 vacc3x89AB = vqsubq_s32(vacc3x89AB, vmagic_bias_less_output_zero_point);
691 vacc3xCDEF = vqsubq_s32(vacc3xCDEF, vmagic_bias_less_output_zero_point);
692 vacc4x0123 = vqsubq_s32(vacc4x0123, vmagic_bias_less_output_zero_point);
693 vacc4x4567 = vqsubq_s32(vacc4x4567, vmagic_bias_less_output_zero_point);
694 vacc4x89AB = vqsubq_s32(vacc4x89AB, vmagic_bias_less_output_zero_point);
695 vacc4xCDEF = vqsubq_s32(vacc4xCDEF, vmagic_bias_less_output_zero_point);
696 vacc5x0123 = vqsubq_s32(vacc5x0123, vmagic_bias_less_output_zero_point);
697 vacc5x4567 = vqsubq_s32(vacc5x4567, vmagic_bias_less_output_zero_point);
698 vacc5x89AB = vqsubq_s32(vacc5x89AB, vmagic_bias_less_output_zero_point);
699 vacc5xCDEF = vqsubq_s32(vacc5xCDEF, vmagic_bias_less_output_zero_point);
700
701 #if XNN_ARCH_ARM64
702 int16x8_t vacc0x01234567 = vqmovn_high_s32(vqmovn_s32(vacc0x0123), vacc0x4567);
703 int16x8_t vacc0x89ABCDEF = vqmovn_high_s32(vqmovn_s32(vacc0x89AB), vacc0xCDEF);
704 int16x8_t vacc1x01234567 = vqmovn_high_s32(vqmovn_s32(vacc1x0123), vacc1x4567);
705 int16x8_t vacc1x89ABCDEF = vqmovn_high_s32(vqmovn_s32(vacc1x89AB), vacc1xCDEF);
706 int16x8_t vacc2x01234567 = vqmovn_high_s32(vqmovn_s32(vacc2x0123), vacc2x4567);
707 int16x8_t vacc2x89ABCDEF = vqmovn_high_s32(vqmovn_s32(vacc2x89AB), vacc2xCDEF);
708 int16x8_t vacc3x01234567 = vqmovn_high_s32(vqmovn_s32(vacc3x0123), vacc3x4567);
709 int16x8_t vacc3x89ABCDEF = vqmovn_high_s32(vqmovn_s32(vacc3x89AB), vacc3xCDEF);
710 int16x8_t vacc4x01234567 = vqmovn_high_s32(vqmovn_s32(vacc4x0123), vacc4x4567);
711 int16x8_t vacc4x89ABCDEF = vqmovn_high_s32(vqmovn_s32(vacc4x89AB), vacc4xCDEF);
712 int16x8_t vacc5x01234567 = vqmovn_high_s32(vqmovn_s32(vacc5x0123), vacc5x4567);
713 int16x8_t vacc5x89ABCDEF = vqmovn_high_s32(vqmovn_s32(vacc5x89AB), vacc5xCDEF);
714
715
716 int8x16_t vout0x0123456789ABCDEF = vqmovn_high_s16(vqmovn_s16(vacc0x01234567), vacc0x89ABCDEF);
717 int8x16_t vout1x0123456789ABCDEF = vqmovn_high_s16(vqmovn_s16(vacc1x01234567), vacc1x89ABCDEF);
718 int8x16_t vout2x0123456789ABCDEF = vqmovn_high_s16(vqmovn_s16(vacc2x01234567), vacc2x89ABCDEF);
719 int8x16_t vout3x0123456789ABCDEF = vqmovn_high_s16(vqmovn_s16(vacc3x01234567), vacc3x89ABCDEF);
720 int8x16_t vout4x0123456789ABCDEF = vqmovn_high_s16(vqmovn_s16(vacc4x01234567), vacc4x89ABCDEF);
721 int8x16_t vout5x0123456789ABCDEF = vqmovn_high_s16(vqmovn_s16(vacc5x01234567), vacc5x89ABCDEF);
722 #else
723 int16x8_t vacc0x01234567 = vcombine_s16(vqmovn_s32(vacc0x0123), vqmovn_s32(vacc0x4567));
724 int16x8_t vacc0x89ABCDEF = vcombine_s16(vqmovn_s32(vacc0x89AB), vqmovn_s32(vacc0xCDEF));
725 int16x8_t vacc1x01234567 = vcombine_s16(vqmovn_s32(vacc1x0123), vqmovn_s32(vacc1x4567));
726 int16x8_t vacc1x89ABCDEF = vcombine_s16(vqmovn_s32(vacc1x89AB), vqmovn_s32(vacc1xCDEF));
727 int16x8_t vacc2x01234567 = vcombine_s16(vqmovn_s32(vacc2x0123), vqmovn_s32(vacc2x4567));
728 int16x8_t vacc2x89ABCDEF = vcombine_s16(vqmovn_s32(vacc2x89AB), vqmovn_s32(vacc2xCDEF));
729 int16x8_t vacc3x01234567 = vcombine_s16(vqmovn_s32(vacc3x0123), vqmovn_s32(vacc3x4567));
730 int16x8_t vacc3x89ABCDEF = vcombine_s16(vqmovn_s32(vacc3x89AB), vqmovn_s32(vacc3xCDEF));
731 int16x8_t vacc4x01234567 = vcombine_s16(vqmovn_s32(vacc4x0123), vqmovn_s32(vacc4x4567));
732 int16x8_t vacc4x89ABCDEF = vcombine_s16(vqmovn_s32(vacc4x89AB), vqmovn_s32(vacc4xCDEF));
733 int16x8_t vacc5x01234567 = vcombine_s16(vqmovn_s32(vacc5x0123), vqmovn_s32(vacc5x4567));
734 int16x8_t vacc5x89ABCDEF = vcombine_s16(vqmovn_s32(vacc5x89AB), vqmovn_s32(vacc5xCDEF));
735
736
737 int8x16_t vout0x0123456789ABCDEF = vcombine_s8(vqmovn_s16(vacc0x01234567), vqmovn_s16(vacc0x89ABCDEF));
738 int8x16_t vout1x0123456789ABCDEF = vcombine_s8(vqmovn_s16(vacc1x01234567), vqmovn_s16(vacc1x89ABCDEF));
739 int8x16_t vout2x0123456789ABCDEF = vcombine_s8(vqmovn_s16(vacc2x01234567), vqmovn_s16(vacc2x89ABCDEF));
740 int8x16_t vout3x0123456789ABCDEF = vcombine_s8(vqmovn_s16(vacc3x01234567), vqmovn_s16(vacc3x89ABCDEF));
741 int8x16_t vout4x0123456789ABCDEF = vcombine_s8(vqmovn_s16(vacc4x01234567), vqmovn_s16(vacc4x89ABCDEF));
742 int8x16_t vout5x0123456789ABCDEF = vcombine_s8(vqmovn_s16(vacc5x01234567), vqmovn_s16(vacc5x89ABCDEF));
743 #endif
744
745 const int8x16_t voutput_min = vld1q_dup_s8(¶ms->fp32_neon.output_min);
746 vout0x0123456789ABCDEF = vmaxq_s8(vout0x0123456789ABCDEF, voutput_min);
747 vout1x0123456789ABCDEF = vmaxq_s8(vout1x0123456789ABCDEF, voutput_min);
748 vout2x0123456789ABCDEF = vmaxq_s8(vout2x0123456789ABCDEF, voutput_min);
749 vout3x0123456789ABCDEF = vmaxq_s8(vout3x0123456789ABCDEF, voutput_min);
750 vout4x0123456789ABCDEF = vmaxq_s8(vout4x0123456789ABCDEF, voutput_min);
751 vout5x0123456789ABCDEF = vmaxq_s8(vout5x0123456789ABCDEF, voutput_min);
752
753 const int8x16_t voutput_max = vld1q_dup_s8(¶ms->fp32_neon.output_max);
754 vout0x0123456789ABCDEF = vminq_s8(vout0x0123456789ABCDEF, voutput_max);
755 vout1x0123456789ABCDEF = vminq_s8(vout1x0123456789ABCDEF, voutput_max);
756 vout2x0123456789ABCDEF = vminq_s8(vout2x0123456789ABCDEF, voutput_max);
757 vout3x0123456789ABCDEF = vminq_s8(vout3x0123456789ABCDEF, voutput_max);
758 vout4x0123456789ABCDEF = vminq_s8(vout4x0123456789ABCDEF, voutput_max);
759 vout5x0123456789ABCDEF = vminq_s8(vout5x0123456789ABCDEF, voutput_max);
760
761 if (nc >= 16) {
762 vst1q_s8(c0 + 0, vout0x0123456789ABCDEF);
763 vst1q_s8(c1 + 0, vout1x0123456789ABCDEF);
764 vst1q_s8(c2 + 0, vout2x0123456789ABCDEF);
765 vst1q_s8(c3 + 0, vout3x0123456789ABCDEF);
766 vst1q_s8(c4 + 0, vout4x0123456789ABCDEF);
767 vst1q_s8(c5 + 0, vout5x0123456789ABCDEF);
768
769 c0 = (int8_t*) ((uintptr_t) c0 + cn_stride);
770 c1 = (int8_t*) ((uintptr_t) c1 + cn_stride);
771 c2 = (int8_t*) ((uintptr_t) c2 + cn_stride);
772 c3 = (int8_t*) ((uintptr_t) c3 + cn_stride);
773 c4 = (int8_t*) ((uintptr_t) c4 + cn_stride);
774 c5 = (int8_t*) ((uintptr_t) c5 + cn_stride);
775
776 a0 = (const int8_t*) ((uintptr_t) a0 - kc);
777 a1 = (const int8_t*) ((uintptr_t) a1 - kc);
778 a2 = (const int8_t*) ((uintptr_t) a2 - kc);
779 a3 = (const int8_t*) ((uintptr_t) a3 - kc);
780 a4 = (const int8_t*) ((uintptr_t) a4 - kc);
781 a5 = (const int8_t*) ((uintptr_t) a5 - kc);
782
783 nc -= 16;
784 } else {
785 int8x16_t vout0x01234567_1x01234567 = vcombine_s8(vget_low_s8(vout0x0123456789ABCDEF), vget_low_s8(vout1x0123456789ABCDEF));
786 int8x16_t vout2x01234567_3x01234567 = vcombine_s8(vget_low_s8(vout2x0123456789ABCDEF), vget_low_s8(vout3x0123456789ABCDEF));
787 int8x16_t vout4x01234567_5x01234567 = vcombine_s8(vget_low_s8(vout4x0123456789ABCDEF), vget_low_s8(vout5x0123456789ABCDEF));
788 if (nc & 8) {
789 vst1_s8(c0, vget_low_s8(vout0x01234567_1x01234567)); c0 += 8;
790 vst1_s8(c1, vget_high_s8(vout0x01234567_1x01234567)); c1 += 8;
791 vst1_s8(c2, vget_low_s8(vout2x01234567_3x01234567)); c2 += 8;
792 vst1_s8(c3, vget_high_s8(vout2x01234567_3x01234567)); c3 += 8;
793 vst1_s8(c4, vget_low_s8(vout4x01234567_5x01234567)); c4 += 8;
794 vst1_s8(c5, vget_high_s8(vout4x01234567_5x01234567)); c5 += 8;
795 vout0x01234567_1x01234567 = vcombine_s8(vget_high_s8(vout0x0123456789ABCDEF), vget_high_s8(vout1x0123456789ABCDEF));
796 vout2x01234567_3x01234567 = vcombine_s8(vget_high_s8(vout2x0123456789ABCDEF), vget_high_s8(vout3x0123456789ABCDEF));
797 vout4x01234567_5x01234567 = vcombine_s8(vget_high_s8(vout4x0123456789ABCDEF), vget_high_s8(vout5x0123456789ABCDEF));
798 }
799 if (nc & 4) {
800 vst1q_lane_u32((void*) c0, vreinterpretq_u32_s8(vout0x01234567_1x01234567), 0); c0 += 4;
801 vst1q_lane_u32((void*) c1, vreinterpretq_u32_s8(vout0x01234567_1x01234567), 2); c1 += 4;
802 vst1q_lane_u32((void*) c2, vreinterpretq_u32_s8(vout2x01234567_3x01234567), 0); c2 += 4;
803 vst1q_lane_u32((void*) c3, vreinterpretq_u32_s8(vout2x01234567_3x01234567), 2); c3 += 4;
804 vst1q_lane_u32((void*) c4, vreinterpretq_u32_s8(vout4x01234567_5x01234567), 0); c4 += 4;
805 vst1q_lane_u32((void*) c5, vreinterpretq_u32_s8(vout4x01234567_5x01234567), 2); c5 += 4;
806 vout0x01234567_1x01234567 = vextq_s8(vout0x01234567_1x01234567, vout0x01234567_1x01234567, 4);
807 vout2x01234567_3x01234567 = vextq_s8(vout2x01234567_3x01234567, vout2x01234567_3x01234567, 4);
808 vout4x01234567_5x01234567 = vextq_s8(vout4x01234567_5x01234567, vout4x01234567_5x01234567, 4);
809 }
810 if (nc & 2) {
811 vst1q_lane_u16((void*) c0, vreinterpretq_u16_s8(vout0x01234567_1x01234567), 0); c0 += 2;
812 vst1q_lane_u16((void*) c1, vreinterpretq_u16_s8(vout0x01234567_1x01234567), 4); c1 += 2;
813 vst1q_lane_u16((void*) c2, vreinterpretq_u16_s8(vout2x01234567_3x01234567), 0); c2 += 2;
814 vst1q_lane_u16((void*) c3, vreinterpretq_u16_s8(vout2x01234567_3x01234567), 4); c3 += 2;
815 vst1q_lane_u16((void*) c4, vreinterpretq_u16_s8(vout4x01234567_5x01234567), 0); c4 += 2;
816 vst1q_lane_u16((void*) c5, vreinterpretq_u16_s8(vout4x01234567_5x01234567), 4); c5 += 2;
817 vout0x01234567_1x01234567 = vextq_s8(vout0x01234567_1x01234567, vout0x01234567_1x01234567, 2);
818 vout2x01234567_3x01234567 = vextq_s8(vout2x01234567_3x01234567, vout2x01234567_3x01234567, 2);
819 vout4x01234567_5x01234567 = vextq_s8(vout4x01234567_5x01234567, vout4x01234567_5x01234567, 2);
820 }
821 if (nc & 1) {
822 vst1q_lane_s8(c0, vout0x01234567_1x01234567, 0);
823 vst1q_lane_s8(c1, vout0x01234567_1x01234567, 8);
824 vst1q_lane_s8(c2, vout2x01234567_3x01234567, 0);
825 vst1q_lane_s8(c3, vout2x01234567_3x01234567, 8);
826 vst1q_lane_s8(c4, vout4x01234567_5x01234567, 0);
827 vst1q_lane_s8(c5, vout4x01234567_5x01234567, 8);
828 }
829
830 nc = 0;
831 }
832 } while (nc != 0);
833 }
834