Debian Codegraph @

src/qd8-f16-qb4w-gemm/gen/qd8-f16-qb4w-gemm-7x16c8-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_qb4w_gemm_minmax_ukernel_7x16c8__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_qb4w_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 size_t bl = params->scalar.blocksize;
42 assert(bl <= kc);
43 assert(bl != 0);
44 assert(bl % 32 == 0);
45 const int8_t* a0 = a;
46 uint16_t* c0 = (uint16_t*) c;
47 const int8_t* a1 = (const int8_t*) ((uintptr_t) a0 + a_stride);
48 uint16_t* c1 = (uint16_t*) ((uintptr_t) c0 + cm_stride);
49 if XNN_UNPREDICTABLE(mr < 2) {
50 a1 = a0;
51 c1 = c0;
52 }
53 const int8_t* a2 = (const int8_t*) ((uintptr_t) a1 + a_stride);
54 uint16_t* c2 = (uint16_t*) ((uintptr_t) c1 + cm_stride);
55 if XNN_UNPREDICTABLE(mr <= 2) {
56 a2 = a1;
57 c2 = c1;
58 }
59 const int8_t* a3 = (const int8_t*) ((uintptr_t) a2 + a_stride);
60 uint16_t* c3 = (uint16_t*) ((uintptr_t) c2 + cm_stride);
61 if XNN_UNPREDICTABLE(mr < 4) {
62 a3 = a2;
63 c3 = c2;
64 }
65 const int8_t* a4 = (const int8_t*) ((uintptr_t) a3 + a_stride);
66 uint16_t* c4 = (uint16_t*) ((uintptr_t) c3 + cm_stride);
67 if XNN_UNPREDICTABLE(mr <= 4) {
68 a4 = a3;
69 c4 = c3;
70 }
71 const int8_t* a5 = (const int8_t*) ((uintptr_t) a4 + a_stride);
72 uint16_t* c5 = (uint16_t*) ((uintptr_t) c4 + cm_stride);
73 if XNN_UNPREDICTABLE(mr < 6) {
74 a5 = a4;
75 c5 = c4;
76 }
77 const int8_t* a6 = (const int8_t*) ((uintptr_t) a5 + a_stride);
78 uint16_t* c6 = (uint16_t*) ((uintptr_t) c5 + cm_stride);
79 if XNN_UNPREDICTABLE(mr <= 6) {
80 a6 = a5;
81 c6 = c5;
82 }
83 const int8x16_t vmask = vmovq_n_s8(INT8_C(0xF0));
84
85 // Loop over groups of 16 columns.
86 do {
87 // Initialize accumulators with scaled vksum. 16 scaled vksum values are loaded from the
88 // weight matrix, at the start of the group of 16 columns.
89 const float32x4_t vinput_zero_point01 = vcvtq_f32_s32(vld1q_s32(&quantization_params[0].zero_point));
90 const float32x4_t vksum0123 = vld1q_f32(w); w = (const float*) w + 4;
91 float32x4_t vout0x0123 = vmulq_lane_f32(vksum0123, vget_low_f32(vinput_zero_point01), 0);
92 float32x4_t vout1x0123 = vmulq_lane_f32(vksum0123, vget_high_f32(vinput_zero_point01), 0);
93 const float32x4_t vksum4567 = vld1q_f32(w); w = (const float*) w + 4;
94 float32x4_t vout0x4567 = vmulq_lane_f32(vksum4567, vget_low_f32(vinput_zero_point01), 0);
95 float32x4_t vout1x4567 = vmulq_lane_f32(vksum4567, vget_high_f32(vinput_zero_point01), 0);
96 const float32x4_t vksum89AB = vld1q_f32(w); w = (const float*) w + 4;
97 float32x4_t vout0x89AB = vmulq_lane_f32(vksum89AB, vget_low_f32(vinput_zero_point01), 0);
98 float32x4_t vout1x89AB = vmulq_lane_f32(vksum89AB, vget_high_f32(vinput_zero_point01), 0);
99 const float32x4_t vksumCDEF = vld1q_f32(w); w = (const float*) w + 4;
100 float32x4_t vout0xCDEF = vmulq_lane_f32(vksumCDEF, vget_low_f32(vinput_zero_point01), 0);
101 float32x4_t vout1xCDEF = vmulq_lane_f32(vksumCDEF, vget_high_f32(vinput_zero_point01), 0);
102 const float32x4_t vinput_zero_point23 = vcvtq_f32_s32(vld1q_s32(&quantization_params[2].zero_point));
103 float32x4_t vout2x0123 = vmulq_lane_f32(vksum0123, vget_low_f32(vinput_zero_point23), 0);
104 float32x4_t vout3x0123 = vmulq_lane_f32(vksum0123, vget_high_f32(vinput_zero_point23), 0);
105 float32x4_t vout2x4567 = vmulq_lane_f32(vksum4567, vget_low_f32(vinput_zero_point23), 0);
106 float32x4_t vout3x4567 = vmulq_lane_f32(vksum4567, vget_high_f32(vinput_zero_point23), 0);
107 float32x4_t vout2x89AB = vmulq_lane_f32(vksum89AB, vget_low_f32(vinput_zero_point23), 0);
108 float32x4_t vout3x89AB = vmulq_lane_f32(vksum89AB, vget_high_f32(vinput_zero_point23), 0);
109 float32x4_t vout2xCDEF = vmulq_lane_f32(vksumCDEF, vget_low_f32(vinput_zero_point23), 0);
110 float32x4_t vout3xCDEF = vmulq_lane_f32(vksumCDEF, vget_high_f32(vinput_zero_point23), 0);
111 const float32x4_t vinput_zero_point45 = vcvtq_f32_s32(vld1q_s32(&quantization_params[4].zero_point));
112 float32x4_t vout4x0123 = vmulq_lane_f32(vksum0123, vget_low_f32(vinput_zero_point45), 0);
113 float32x4_t vout5x0123 = vmulq_lane_f32(vksum0123, vget_high_f32(vinput_zero_point45), 0);
114 float32x4_t vout4x4567 = vmulq_lane_f32(vksum4567, vget_low_f32(vinput_zero_point45), 0);
115 float32x4_t vout5x4567 = vmulq_lane_f32(vksum4567, vget_high_f32(vinput_zero_point45), 0);
116 float32x4_t vout4x89AB = vmulq_lane_f32(vksum89AB, vget_low_f32(vinput_zero_point45), 0);
117 float32x4_t vout5x89AB = vmulq_lane_f32(vksum89AB, vget_high_f32(vinput_zero_point45), 0);
118 float32x4_t vout4xCDEF = vmulq_lane_f32(vksumCDEF, vget_low_f32(vinput_zero_point45), 0);
119 float32x4_t vout5xCDEF = vmulq_lane_f32(vksumCDEF, vget_high_f32(vinput_zero_point45), 0);
120 const float32x4_t vinput_zero_point67 = vcvtq_f32_s32(vld1q_dup_s32(&quantization_params[6].zero_point));
121 float32x4_t vout6x0123 = vmulq_f32(vksum0123, vinput_zero_point67);
122 float32x4_t vout6x4567 = vmulq_f32(vksum4567, vinput_zero_point67);
123 float32x4_t vout6x89AB = vmulq_f32(vksum89AB, vinput_zero_point67);
124 float32x4_t vout6xCDEF = vmulq_f32(vksumCDEF, vinput_zero_point67);
125
126
127 for (size_t kb=0; kb < kc; kb += bl) {
128 int32x4_t vacc01x01 = vdupq_n_s32(0);
129 int32x4_t vacc01x23 = vdupq_n_s32(0);
130 int32x4_t vacc01x45 = vdupq_n_s32(0);
131 int32x4_t vacc01x67 = vdupq_n_s32(0);
132 int32x4_t vacc01x89 = vdupq_n_s32(0);
133 int32x4_t vacc01xAB = vdupq_n_s32(0);
134 int32x4_t vacc01xCD = vdupq_n_s32(0);
135 int32x4_t vacc01xEF = vdupq_n_s32(0);
136 int32x4_t vacc23x01 = vdupq_n_s32(0);
137 int32x4_t vacc23x23 = vdupq_n_s32(0);
138 int32x4_t vacc23x45 = vdupq_n_s32(0);
139 int32x4_t vacc23x67 = vdupq_n_s32(0);
140 int32x4_t vacc23x89 = vdupq_n_s32(0);
141 int32x4_t vacc23xAB = vdupq_n_s32(0);
142 int32x4_t vacc23xCD = vdupq_n_s32(0);
143 int32x4_t vacc23xEF = vdupq_n_s32(0);
144 int32x4_t vacc45x01 = vdupq_n_s32(0);
145 int32x4_t vacc45x23 = vdupq_n_s32(0);
146 int32x4_t vacc45x45 = vdupq_n_s32(0);
147 int32x4_t vacc45x67 = vdupq_n_s32(0);
148 int32x4_t vacc45x89 = vdupq_n_s32(0);
149 int32x4_t vacc45xAB = vdupq_n_s32(0);
150 int32x4_t vacc45xCD = vdupq_n_s32(0);
151 int32x4_t vacc45xEF = vdupq_n_s32(0);
152 int32x4_t vacc67x01 = vdupq_n_s32(0);
153 int32x4_t vacc67x23 = vdupq_n_s32(0);
154 int32x4_t vacc67x45 = vdupq_n_s32(0);
155 int32x4_t vacc67x67 = vdupq_n_s32(0);
156 int32x4_t vacc67x89 = vdupq_n_s32(0);
157 int32x4_t vacc67xAB = vdupq_n_s32(0);
158 int32x4_t vacc67xCD = vdupq_n_s32(0);
159 int32x4_t vacc67xEF = vdupq_n_s32(0);
160
161 size_t k = bl;
162 // 2x partial unrolled loop to load 8 bytes at a time.
163
164 uint64x2x2_t va01x0123456789ABCDEF;
165 va01x0123456789ABCDEF.val[0] = vdupq_n_u64(0);
166 va01x0123456789ABCDEF.val[1] = vdupq_n_u64(0);
167 uint64x2x2_t va23x0123456789ABCDEF;
168 va23x0123456789ABCDEF.val[0] = vdupq_n_u64(0);
169 va23x0123456789ABCDEF.val[1] = vdupq_n_u64(0);
170 uint64x2x2_t va45x0123456789ABCDEF;
171 va45x0123456789ABCDEF.val[0] = vdupq_n_u64(0);
172 va45x0123456789ABCDEF.val[1] = vdupq_n_u64(0);
173 uint64x2x2_t va67x0123456789ABCDEF;
174 va67x0123456789ABCDEF.val[0] = vdupq_n_u64(0);
175 va67x0123456789ABCDEF.val[1] = vdupq_n_u64(0);
176
177 while (k >= 16 * sizeof(int8_t)) {
178 // Load a 7x16 block of activations.
179 va01x0123456789ABCDEF = vld2q_lane_u64((const void*) a0, va01x0123456789ABCDEF, 0); a0 += 16;
180 va23x0123456789ABCDEF = vld2q_lane_u64((const void*) a2, va23x0123456789ABCDEF, 0); a2 += 16;
181 va45x0123456789ABCDEF = vld2q_lane_u64((const void*) a4, va45x0123456789ABCDEF, 0); a4 += 16;
182 va67x0123456789ABCDEF = vld2q_lane_u64((const void*) a6, va67x0123456789ABCDEF, 0); a6 += 16;
183 va01x0123456789ABCDEF = vld2q_lane_u64((const void*) a1, va01x0123456789ABCDEF, 1); a1 += 16;
184 va23x0123456789ABCDEF = vld2q_lane_u64((const void*) a3, va23x0123456789ABCDEF, 1); a3 += 16;
185 va45x0123456789ABCDEF = vld2q_lane_u64((const void*) a5, va45x0123456789ABCDEF, 1); a5 += 16;
186
187 const int8x16_t va01x01234567 = vreinterpretq_s8_u64(va01x0123456789ABCDEF.val[0]);
188 const int8x16_t va01x89ABCDEF = vreinterpretq_s8_u64(va01x0123456789ABCDEF.val[1]);
189 const int8x16_t va23x01234567 = vreinterpretq_s8_u64(va23x0123456789ABCDEF.val[0]);
190 const int8x16_t va23x89ABCDEF = vreinterpretq_s8_u64(va23x0123456789ABCDEF.val[1]);
191 const int8x16_t va45x01234567 = vreinterpretq_s8_u64(va45x0123456789ABCDEF.val[0]);
192 const int8x16_t va45x89ABCDEF = vreinterpretq_s8_u64(va45x0123456789ABCDEF.val[1]);
193 const int8x16_t va67x01234567 = vreinterpretq_s8_u64(va67x0123456789ABCDEF.val[0]);
194 const int8x16_t va67x89ABCDEF = vreinterpretq_s8_u64(va67x0123456789ABCDEF.val[1]);
195
196 // Load a 16x16 block of weights.
197 const int8x16_t vb01x0123456789ABCDEF = vld1q_s8(w); w = (const int8_t*) w + 16;
198 const int8x16_t vb23x0123456789ABCDEF = vld1q_s8(w); w = (const int8_t*) w + 16;
199 const int8x16_t vb45x0123456789ABCDEF = vld1q_s8(w); w = (const int8_t*) w + 16;
200 const int8x16_t vb67x0123456789ABCDEF = vld1q_s8(w); w = (const int8_t*) w + 16;
201 const int8x16_t vb89x0123456789ABCDEF = vld1q_s8(w); w = (const int8_t*) w + 16;
202 const int8x16_t vbABx0123456789ABCDEF = vld1q_s8(w); w = (const int8_t*) w + 16;
203 const int8x16_t vbCDx0123456789ABCDEF = vld1q_s8(w); w = (const int8_t*) w + 16;
204 const int8x16_t vbEFx0123456789ABCDEF = vld1q_s8(w); w = (const int8_t*) w + 16;
205 const int8x16_t vb01x01234567 = vshlq_n_s8(vb01x0123456789ABCDEF, 4);
206 const int8x16_t vb23x01234567 = vshlq_n_s8(vb23x0123456789ABCDEF, 4);
207 const int8x16_t vb45x01234567 = vshlq_n_s8(vb45x0123456789ABCDEF, 4);
208 const int8x16_t vb67x01234567 = vshlq_n_s8(vb67x0123456789ABCDEF, 4);
209 const int8x16_t vb89x01234567 = vshlq_n_s8(vb89x0123456789ABCDEF, 4);
210 const int8x16_t vbABx01234567 = vshlq_n_s8(vbABx0123456789ABCDEF, 4);
211 const int8x16_t vbCDx01234567 = vshlq_n_s8(vbCDx0123456789ABCDEF, 4);
212 const int8x16_t vbEFx01234567 = vshlq_n_s8(vbEFx0123456789ABCDEF, 4);
213 const int8x16_t vb01x89ABCDEF = vandq_s8(vb01x0123456789ABCDEF, vmask);
214 const int8x16_t vb23x89ABCDEF = vandq_s8(vb23x0123456789ABCDEF, vmask);
215 const int8x16_t vb45x89ABCDEF = vandq_s8(vb45x0123456789ABCDEF, vmask);
216 const int8x16_t vb67x89ABCDEF = vandq_s8(vb67x0123456789ABCDEF, vmask);
217 const int8x16_t vb89x89ABCDEF = vandq_s8(vb89x0123456789ABCDEF, vmask);
218 const int8x16_t vbABx89ABCDEF = vandq_s8(vbABx0123456789ABCDEF, vmask);
219 const int8x16_t vbCDx89ABCDEF = vandq_s8(vbCDx0123456789ABCDEF, vmask);
220 const int8x16_t vbEFx89ABCDEF = vandq_s8(vbEFx0123456789ABCDEF, vmask);
221
222 // Multiply-accumulate: 7x8 * 8x16 --> 7x16.
223 vacc01x01 = vmmlaq_s32(vacc01x01, va01x01234567, vb01x01234567);
224 vacc01x23 = vmmlaq_s32(vacc01x23, va01x01234567, vb23x01234567);
225 vacc01x45 = vmmlaq_s32(vacc01x45, va01x01234567, vb45x01234567);
226 vacc01x67 = vmmlaq_s32(vacc01x67, va01x01234567, vb67x01234567);
227 vacc01x89 = vmmlaq_s32(vacc01x89, va01x01234567, vb89x01234567);
228 vacc01xAB = vmmlaq_s32(vacc01xAB, va01x01234567, vbABx01234567);
229 vacc01xCD = vmmlaq_s32(vacc01xCD, va01x01234567, vbCDx01234567);
230 vacc01xEF = vmmlaq_s32(vacc01xEF, va01x01234567, vbEFx01234567);
231 vacc23x01 = vmmlaq_s32(vacc23x01, va23x01234567, vb01x01234567);
232 vacc23x23 = vmmlaq_s32(vacc23x23, va23x01234567, vb23x01234567);
233 vacc23x45 = vmmlaq_s32(vacc23x45, va23x01234567, vb45x01234567);
234 vacc23x67 = vmmlaq_s32(vacc23x67, va23x01234567, vb67x01234567);
235 vacc23x89 = vmmlaq_s32(vacc23x89, va23x01234567, vb89x01234567);
236 vacc23xAB = vmmlaq_s32(vacc23xAB, va23x01234567, vbABx01234567);
237 vacc23xCD = vmmlaq_s32(vacc23xCD, va23x01234567, vbCDx01234567);
238 vacc23xEF = vmmlaq_s32(vacc23xEF, va23x01234567, vbEFx01234567);
239 vacc45x01 = vmmlaq_s32(vacc45x01, va45x01234567, vb01x01234567);
240 vacc45x23 = vmmlaq_s32(vacc45x23, va45x01234567, vb23x01234567);
241 vacc45x45 = vmmlaq_s32(vacc45x45, va45x01234567, vb45x01234567);
242 vacc45x67 = vmmlaq_s32(vacc45x67, va45x01234567, vb67x01234567);
243 vacc45x89 = vmmlaq_s32(vacc45x89, va45x01234567, vb89x01234567);
244 vacc45xAB = vmmlaq_s32(vacc45xAB, va45x01234567, vbABx01234567);
245 vacc45xCD = vmmlaq_s32(vacc45xCD, va45x01234567, vbCDx01234567);
246 vacc45xEF = vmmlaq_s32(vacc45xEF, va45x01234567, vbEFx01234567);
247 vacc67x01 = vmmlaq_s32(vacc67x01, va67x01234567, vb01x01234567);
248 vacc67x23 = vmmlaq_s32(vacc67x23, va67x01234567, vb23x01234567);
249 vacc67x45 = vmmlaq_s32(vacc67x45, va67x01234567, vb45x01234567);
250 vacc67x67 = vmmlaq_s32(vacc67x67, va67x01234567, vb67x01234567);
251 vacc67x89 = vmmlaq_s32(vacc67x89, va67x01234567, vb89x01234567);
252 vacc67xAB = vmmlaq_s32(vacc67xAB, va67x01234567, vbABx01234567);
253 vacc67xCD = vmmlaq_s32(vacc67xCD, va67x01234567, vbCDx01234567);
254 vacc67xEF = vmmlaq_s32(vacc67xEF, va67x01234567, vbEFx01234567);
255 vacc01x01 = vmmlaq_s32(vacc01x01, va01x89ABCDEF, vb01x89ABCDEF);
256 vacc01x23 = vmmlaq_s32(vacc01x23, va01x89ABCDEF, vb23x89ABCDEF);
257 vacc01x45 = vmmlaq_s32(vacc01x45, va01x89ABCDEF, vb45x89ABCDEF);
258 vacc01x67 = vmmlaq_s32(vacc01x67, va01x89ABCDEF, vb67x89ABCDEF);
259 vacc01x89 = vmmlaq_s32(vacc01x89, va01x89ABCDEF, vb89x89ABCDEF);
260 vacc01xAB = vmmlaq_s32(vacc01xAB, va01x89ABCDEF, vbABx89ABCDEF);
261 vacc01xCD = vmmlaq_s32(vacc01xCD, va01x89ABCDEF, vbCDx89ABCDEF);
262 vacc01xEF = vmmlaq_s32(vacc01xEF, va01x89ABCDEF, vbEFx89ABCDEF);
263 vacc23x01 = vmmlaq_s32(vacc23x01, va23x89ABCDEF, vb01x89ABCDEF);
264 vacc23x23 = vmmlaq_s32(vacc23x23, va23x89ABCDEF, vb23x89ABCDEF);
265 vacc23x45 = vmmlaq_s32(vacc23x45, va23x89ABCDEF, vb45x89ABCDEF);
266 vacc23x67 = vmmlaq_s32(vacc23x67, va23x89ABCDEF, vb67x89ABCDEF);
267 vacc23x89 = vmmlaq_s32(vacc23x89, va23x89ABCDEF, vb89x89ABCDEF);
268 vacc23xAB = vmmlaq_s32(vacc23xAB, va23x89ABCDEF, vbABx89ABCDEF);
269 vacc23xCD = vmmlaq_s32(vacc23xCD, va23x89ABCDEF, vbCDx89ABCDEF);
270 vacc23xEF = vmmlaq_s32(vacc23xEF, va23x89ABCDEF, vbEFx89ABCDEF);
271 vacc45x01 = vmmlaq_s32(vacc45x01, va45x89ABCDEF, vb01x89ABCDEF);
272 vacc45x23 = vmmlaq_s32(vacc45x23, va45x89ABCDEF, vb23x89ABCDEF);
273 vacc45x45 = vmmlaq_s32(vacc45x45, va45x89ABCDEF, vb45x89ABCDEF);
274 vacc45x67 = vmmlaq_s32(vacc45x67, va45x89ABCDEF, vb67x89ABCDEF);
275 vacc45x89 = vmmlaq_s32(vacc45x89, va45x89ABCDEF, vb89x89ABCDEF);
276 vacc45xAB = vmmlaq_s32(vacc45xAB, va45x89ABCDEF, vbABx89ABCDEF);
277 vacc45xCD = vmmlaq_s32(vacc45xCD, va45x89ABCDEF, vbCDx89ABCDEF);
278 vacc45xEF = vmmlaq_s32(vacc45xEF, va45x89ABCDEF, vbEFx89ABCDEF);
279 vacc67x01 = vmmlaq_s32(vacc67x01, va67x89ABCDEF, vb01x89ABCDEF);
280 vacc67x23 = vmmlaq_s32(vacc67x23, va67x89ABCDEF, vb23x89ABCDEF);
281 vacc67x45 = vmmlaq_s32(vacc67x45, va67x89ABCDEF, vb45x89ABCDEF);
282 vacc67x67 = vmmlaq_s32(vacc67x67, va67x89ABCDEF, vb67x89ABCDEF);
283 vacc67x89 = vmmlaq_s32(vacc67x89, va67x89ABCDEF, vb89x89ABCDEF);
284 vacc67xAB = vmmlaq_s32(vacc67xAB, va67x89ABCDEF, vbABx89ABCDEF);
285 vacc67xCD = vmmlaq_s32(vacc67xCD, va67x89ABCDEF, vbCDx89ABCDEF);
286 vacc67xEF = vmmlaq_s32(vacc67xEF, va67x89ABCDEF, vbEFx89ABCDEF);
287
288 k -= 16 * sizeof(int8_t);
289 }
290 // Handle up to 8 final positions of `k`
291 if XNN_UNLIKELY(k != 0) {
292 // Load a 7x8 block of activations.
293 uint64x2_t va01x01234567 = vld1q_dup_u64((const void*) a0); a0 += 8;
294 uint64x2_t va23x01234567 = vld1q_dup_u64((const void*) a2); a2 += 8;
295 uint64x2_t va45x01234567 = vld1q_dup_u64((const void*) a4); a4 += 8;
296 uint64x2_t va67x01234567 = vld1q_dup_u64((const void*) a6); a6 += 8;
297 va01x01234567 = vld1q_lane_u64((const void*) a1, va01x01234567, 1); a1 += 8;
298 va23x01234567 = vld1q_lane_u64((const void*) a3, va23x01234567, 1); a3 += 8;
299 va45x01234567 = vld1q_lane_u64((const void*) a5, va45x01234567, 1); a5 += 8;
300
301 // Load a 16x16 block of weights.
302 const int8x16_t vb01x0123456789ABCDEF = vld1q_s8(w); w = (const int8_t*) w + 16;
303 const int8x16_t vb23x0123456789ABCDEF = vld1q_s8(w); w = (const int8_t*) w + 16;
304 const int8x16_t vb45x0123456789ABCDEF = vld1q_s8(w); w = (const int8_t*) w + 16;
305 const int8x16_t vb67x0123456789ABCDEF = vld1q_s8(w); w = (const int8_t*) w + 16;
306 const int8x16_t vb89x0123456789ABCDEF = vld1q_s8(w); w = (const int8_t*) w + 16;
307 const int8x16_t vbABx0123456789ABCDEF = vld1q_s8(w); w = (const int8_t*) w + 16;
308 const int8x16_t vbCDx0123456789ABCDEF = vld1q_s8(w); w = (const int8_t*) w + 16;
309 const int8x16_t vbEFx0123456789ABCDEF = vld1q_s8(w); w = (const int8_t*) w + 16;
310 const int8x16_t vb01x01234567 = vshlq_n_s8(vb01x0123456789ABCDEF, 4);
311 const int8x16_t vb23x01234567 = vshlq_n_s8(vb23x0123456789ABCDEF, 4);
312 const int8x16_t vb45x01234567 = vshlq_n_s8(vb45x0123456789ABCDEF, 4);
313 const int8x16_t vb67x01234567 = vshlq_n_s8(vb67x0123456789ABCDEF, 4);
314 const int8x16_t vb89x01234567 = vshlq_n_s8(vb89x0123456789ABCDEF, 4);
315 const int8x16_t vbABx01234567 = vshlq_n_s8(vbABx0123456789ABCDEF, 4);
316 const int8x16_t vbCDx01234567 = vshlq_n_s8(vbCDx0123456789ABCDEF, 4);
317 const int8x16_t vbEFx01234567 = vshlq_n_s8(vbEFx0123456789ABCDEF, 4);
318
319 // Multiply-accumulate: 7x4 * 4x16 --> 7x16.
320 vacc01x01 = vmmlaq_s32(vacc01x01, vreinterpretq_s8_u64(va01x01234567), vb01x01234567);
321 vacc01x23 = vmmlaq_s32(vacc01x23, vreinterpretq_s8_u64(va01x01234567), vb23x01234567);
322 vacc01x45 = vmmlaq_s32(vacc01x45, vreinterpretq_s8_u64(va01x01234567), vb45x01234567);
323 vacc01x67 = vmmlaq_s32(vacc01x67, vreinterpretq_s8_u64(va01x01234567), vb67x01234567);
324 vacc01x89 = vmmlaq_s32(vacc01x89, vreinterpretq_s8_u64(va01x01234567), vb89x01234567);
325 vacc01xAB = vmmlaq_s32(vacc01xAB, vreinterpretq_s8_u64(va01x01234567), vbABx01234567);
326 vacc01xCD = vmmlaq_s32(vacc01xCD, vreinterpretq_s8_u64(va01x01234567), vbCDx01234567);
327 vacc01xEF = vmmlaq_s32(vacc01xEF, vreinterpretq_s8_u64(va01x01234567), vbEFx01234567);
328 vacc23x01 = vmmlaq_s32(vacc23x01, vreinterpretq_s8_u64(va23x01234567), vb01x01234567);
329 vacc23x23 = vmmlaq_s32(vacc23x23, vreinterpretq_s8_u64(va23x01234567), vb23x01234567);
330 vacc23x45 = vmmlaq_s32(vacc23x45, vreinterpretq_s8_u64(va23x01234567), vb45x01234567);
331 vacc23x67 = vmmlaq_s32(vacc23x67, vreinterpretq_s8_u64(va23x01234567), vb67x01234567);
332 vacc23x89 = vmmlaq_s32(vacc23x89, vreinterpretq_s8_u64(va23x01234567), vb89x01234567);
333 vacc23xAB = vmmlaq_s32(vacc23xAB, vreinterpretq_s8_u64(va23x01234567), vbABx01234567);
334 vacc23xCD = vmmlaq_s32(vacc23xCD, vreinterpretq_s8_u64(va23x01234567), vbCDx01234567);
335 vacc23xEF = vmmlaq_s32(vacc23xEF, vreinterpretq_s8_u64(va23x01234567), vbEFx01234567);
336 vacc45x01 = vmmlaq_s32(vacc45x01, vreinterpretq_s8_u64(va45x01234567), vb01x01234567);
337 vacc45x23 = vmmlaq_s32(vacc45x23, vreinterpretq_s8_u64(va45x01234567), vb23x01234567);
338 vacc45x45 = vmmlaq_s32(vacc45x45, vreinterpretq_s8_u64(va45x01234567), vb45x01234567);
339 vacc45x67 = vmmlaq_s32(vacc45x67, vreinterpretq_s8_u64(va45x01234567), vb67x01234567);
340 vacc45x89 = vmmlaq_s32(vacc45x89, vreinterpretq_s8_u64(va45x01234567), vb89x01234567);
341 vacc45xAB = vmmlaq_s32(vacc45xAB, vreinterpretq_s8_u64(va45x01234567), vbABx01234567);
342 vacc45xCD = vmmlaq_s32(vacc45xCD, vreinterpretq_s8_u64(va45x01234567), vbCDx01234567);
343 vacc45xEF = vmmlaq_s32(vacc45xEF, vreinterpretq_s8_u64(va45x01234567), vbEFx01234567);
344 vacc67x01 = vmmlaq_s32(vacc67x01, vreinterpretq_s8_u64(va67x01234567), vb01x01234567);
345 vacc67x23 = vmmlaq_s32(vacc67x23, vreinterpretq_s8_u64(va67x01234567), vb23x01234567);
346 vacc67x45 = vmmlaq_s32(vacc67x45, vreinterpretq_s8_u64(va67x01234567), vb45x01234567);
347 vacc67x67 = vmmlaq_s32(vacc67x67, vreinterpretq_s8_u64(va67x01234567), vb67x01234567);
348 vacc67x89 = vmmlaq_s32(vacc67x89, vreinterpretq_s8_u64(va67x01234567), vb89x01234567);
349 vacc67xAB = vmmlaq_s32(vacc67xAB, vreinterpretq_s8_u64(va67x01234567), vbABx01234567);
350 vacc67xCD = vmmlaq_s32(vacc67xCD, vreinterpretq_s8_u64(va67x01234567), vbCDx01234567);
351 vacc67xEF = vmmlaq_s32(vacc67xEF, vreinterpretq_s8_u64(va67x01234567), vbEFx01234567);
352 }
353
354 int32x4_t vacc0x0123 = vreinterpretq_s32_u64(vtrn1q_u64(vreinterpretq_u64_s32(vacc01x01), vreinterpretq_u64_s32(vacc01x23)));
355 int32x4_t vacc1x0123 = vreinterpretq_s32_u64(vtrn2q_u64(vreinterpretq_u64_s32(vacc01x01), vreinterpretq_u64_s32(vacc01x23)));
356 int32x4_t vacc0x4567 = vreinterpretq_s32_u64(vtrn1q_u64(vreinterpretq_u64_s32(vacc01x45), vreinterpretq_u64_s32(vacc01x67)));
357 int32x4_t vacc1x4567 = vreinterpretq_s32_u64(vtrn2q_u64(vreinterpretq_u64_s32(vacc01x45), vreinterpretq_u64_s32(vacc01x67)));
358 int32x4_t vacc0x89AB = vreinterpretq_s32_u64(vtrn1q_u64(vreinterpretq_u64_s32(vacc01x89), vreinterpretq_u64_s32(vacc01xAB)));
359 int32x4_t vacc1x89AB = vreinterpretq_s32_u64(vtrn2q_u64(vreinterpretq_u64_s32(vacc01x89), vreinterpretq_u64_s32(vacc01xAB)));
360 int32x4_t vacc0xCDEF = vreinterpretq_s32_u64(vtrn1q_u64(vreinterpretq_u64_s32(vacc01xCD), vreinterpretq_u64_s32(vacc01xEF)));
361 int32x4_t vacc1xCDEF = vreinterpretq_s32_u64(vtrn2q_u64(vreinterpretq_u64_s32(vacc01xCD), vreinterpretq_u64_s32(vacc01xEF)));
362 int32x4_t vacc2x0123 = vreinterpretq_s32_u64(vtrn1q_u64(vreinterpretq_u64_s32(vacc23x01), vreinterpretq_u64_s32(vacc23x23)));
363 int32x4_t vacc3x0123 = vreinterpretq_s32_u64(vtrn2q_u64(vreinterpretq_u64_s32(vacc23x01), vreinterpretq_u64_s32(vacc23x23)));
364 int32x4_t vacc2x4567 = vreinterpretq_s32_u64(vtrn1q_u64(vreinterpretq_u64_s32(vacc23x45), vreinterpretq_u64_s32(vacc23x67)));
365 int32x4_t vacc3x4567 = vreinterpretq_s32_u64(vtrn2q_u64(vreinterpretq_u64_s32(vacc23x45), vreinterpretq_u64_s32(vacc23x67)));
366 int32x4_t vacc2x89AB = vreinterpretq_s32_u64(vtrn1q_u64(vreinterpretq_u64_s32(vacc23x89), vreinterpretq_u64_s32(vacc23xAB)));
367 int32x4_t vacc3x89AB = vreinterpretq_s32_u64(vtrn2q_u64(vreinterpretq_u64_s32(vacc23x89), vreinterpretq_u64_s32(vacc23xAB)));
368 int32x4_t vacc2xCDEF = vreinterpretq_s32_u64(vtrn1q_u64(vreinterpretq_u64_s32(vacc23xCD), vreinterpretq_u64_s32(vacc23xEF)));
369 int32x4_t vacc3xCDEF = vreinterpretq_s32_u64(vtrn2q_u64(vreinterpretq_u64_s32(vacc23xCD), vreinterpretq_u64_s32(vacc23xEF)));
370 int32x4_t vacc4x0123 = vreinterpretq_s32_u64(vtrn1q_u64(vreinterpretq_u64_s32(vacc45x01), vreinterpretq_u64_s32(vacc45x23)));
371 int32x4_t vacc5x0123 = vreinterpretq_s32_u64(vtrn2q_u64(vreinterpretq_u64_s32(vacc45x01), vreinterpretq_u64_s32(vacc45x23)));
372 int32x4_t vacc4x4567 = vreinterpretq_s32_u64(vtrn1q_u64(vreinterpretq_u64_s32(vacc45x45), vreinterpretq_u64_s32(vacc45x67)));
373 int32x4_t vacc5x4567 = vreinterpretq_s32_u64(vtrn2q_u64(vreinterpretq_u64_s32(vacc45x45), vreinterpretq_u64_s32(vacc45x67)));
374 int32x4_t vacc4x89AB = vreinterpretq_s32_u64(vtrn1q_u64(vreinterpretq_u64_s32(vacc45x89), vreinterpretq_u64_s32(vacc45xAB)));
375 int32x4_t vacc5x89AB = vreinterpretq_s32_u64(vtrn2q_u64(vreinterpretq_u64_s32(vacc45x89), vreinterpretq_u64_s32(vacc45xAB)));
376 int32x4_t vacc4xCDEF = vreinterpretq_s32_u64(vtrn1q_u64(vreinterpretq_u64_s32(vacc45xCD), vreinterpretq_u64_s32(vacc45xEF)));
377 int32x4_t vacc5xCDEF = vreinterpretq_s32_u64(vtrn2q_u64(vreinterpretq_u64_s32(vacc45xCD), vreinterpretq_u64_s32(vacc45xEF)));
378 int32x4_t vacc6x0123 = vreinterpretq_s32_u64(vtrn1q_u64(vreinterpretq_u64_s32(vacc67x01), vreinterpretq_u64_s32(vacc67x23)));
379 int32x4_t vacc6x4567 = vreinterpretq_s32_u64(vtrn1q_u64(vreinterpretq_u64_s32(vacc67x45), vreinterpretq_u64_s32(vacc67x67)));
380 int32x4_t vacc6x89AB = vreinterpretq_s32_u64(vtrn1q_u64(vreinterpretq_u64_s32(vacc67x89), vreinterpretq_u64_s32(vacc67xAB)));
381 int32x4_t vacc6xCDEF = vreinterpretq_s32_u64(vtrn1q_u64(vreinterpretq_u64_s32(vacc67xCD), vreinterpretq_u64_s32(vacc67xEF)));
382 const float32x4_t vfilter_output_scale0123 = vreinterpretq_f32_u32(vshll_n_u16(vld1_u16(w), 16)); w = (const uint16_t*) w + 4;
383 const float32x4_t vfilter_output_scale4567 = vreinterpretq_f32_u32(vshll_n_u16(vld1_u16(w), 16)); w = (const uint16_t*) w + 4;
384 const float32x4_t vfilter_output_scale89AB = vreinterpretq_f32_u32(vshll_n_u16(vld1_u16(w), 16)); w = (const uint16_t*) w + 4;
385 const float32x4_t vfilter_output_scaleCDEF = vreinterpretq_f32_u32(vshll_n_u16(vld1_u16(w), 16)); w = (const uint16_t*) w + 4;
386 float32x4_t vf0x0123 = vcvtq_f32_s32(vacc0x0123);
387 vout0x0123 = vfmaq_f32(vout0x0123, vf0x0123, vfilter_output_scale0123);
388 float32x4_t vf0x4567 = vcvtq_f32_s32(vacc0x4567);
389 vout0x4567 = vfmaq_f32(vout0x4567, vf0x4567, vfilter_output_scale4567);
390 float32x4_t vf0x89AB = vcvtq_f32_s32(vacc0x89AB);
391 vout0x89AB = vfmaq_f32(vout0x89AB, vf0x89AB, vfilter_output_scale89AB);
392 float32x4_t vf0xCDEF = vcvtq_f32_s32(vacc0xCDEF);
393 vout0xCDEF = vfmaq_f32(vout0xCDEF, vf0xCDEF, vfilter_output_scaleCDEF);
394 float32x4_t vf1x0123 = vcvtq_f32_s32(vacc1x0123);
395 vout1x0123 = vfmaq_f32(vout1x0123, vf1x0123, vfilter_output_scale0123);
396 float32x4_t vf1x4567 = vcvtq_f32_s32(vacc1x4567);
397 vout1x4567 = vfmaq_f32(vout1x4567, vf1x4567, vfilter_output_scale4567);
398 float32x4_t vf1x89AB = vcvtq_f32_s32(vacc1x89AB);
399 vout1x89AB = vfmaq_f32(vout1x89AB, vf1x89AB, vfilter_output_scale89AB);
400 float32x4_t vf1xCDEF = vcvtq_f32_s32(vacc1xCDEF);
401 vout1xCDEF = vfmaq_f32(vout1xCDEF, vf1xCDEF, vfilter_output_scaleCDEF);
402 float32x4_t vf2x0123 = vcvtq_f32_s32(vacc2x0123);
403 vout2x0123 = vfmaq_f32(vout2x0123, vf2x0123, vfilter_output_scale0123);
404 float32x4_t vf2x4567 = vcvtq_f32_s32(vacc2x4567);
405 vout2x4567 = vfmaq_f32(vout2x4567, vf2x4567, vfilter_output_scale4567);
406 float32x4_t vf2x89AB = vcvtq_f32_s32(vacc2x89AB);
407 vout2x89AB = vfmaq_f32(vout2x89AB, vf2x89AB, vfilter_output_scale89AB);
408 float32x4_t vf2xCDEF = vcvtq_f32_s32(vacc2xCDEF);
409 vout2xCDEF = vfmaq_f32(vout2xCDEF, vf2xCDEF, vfilter_output_scaleCDEF);
410 float32x4_t vf3x0123 = vcvtq_f32_s32(vacc3x0123);
411 vout3x0123 = vfmaq_f32(vout3x0123, vf3x0123, vfilter_output_scale0123);
412 float32x4_t vf3x4567 = vcvtq_f32_s32(vacc3x4567);
413 vout3x4567 = vfmaq_f32(vout3x4567, vf3x4567, vfilter_output_scale4567);
414 float32x4_t vf3x89AB = vcvtq_f32_s32(vacc3x89AB);
415 vout3x89AB = vfmaq_f32(vout3x89AB, vf3x89AB, vfilter_output_scale89AB);
416 float32x4_t vf3xCDEF = vcvtq_f32_s32(vacc3xCDEF);
417 vout3xCDEF = vfmaq_f32(vout3xCDEF, vf3xCDEF, vfilter_output_scaleCDEF);
418 float32x4_t vf4x0123 = vcvtq_f32_s32(vacc4x0123);
419 vout4x0123 = vfmaq_f32(vout4x0123, vf4x0123, vfilter_output_scale0123);
420 float32x4_t vf4x4567 = vcvtq_f32_s32(vacc4x4567);
421 vout4x4567 = vfmaq_f32(vout4x4567, vf4x4567, vfilter_output_scale4567);
422 float32x4_t vf4x89AB = vcvtq_f32_s32(vacc4x89AB);
423 vout4x89AB = vfmaq_f32(vout4x89AB, vf4x89AB, vfilter_output_scale89AB);
424 float32x4_t vf4xCDEF = vcvtq_f32_s32(vacc4xCDEF);
425 vout4xCDEF = vfmaq_f32(vout4xCDEF, vf4xCDEF, vfilter_output_scaleCDEF);
426 float32x4_t vf5x0123 = vcvtq_f32_s32(vacc5x0123);
427 vout5x0123 = vfmaq_f32(vout5x0123, vf5x0123, vfilter_output_scale0123);
428 float32x4_t vf5x4567 = vcvtq_f32_s32(vacc5x4567);
429 vout5x4567 = vfmaq_f32(vout5x4567, vf5x4567, vfilter_output_scale4567);
430 float32x4_t vf5x89AB = vcvtq_f32_s32(vacc5x89AB);
431 vout5x89AB = vfmaq_f32(vout5x89AB, vf5x89AB, vfilter_output_scale89AB);
432 float32x4_t vf5xCDEF = vcvtq_f32_s32(vacc5xCDEF);
433 vout5xCDEF = vfmaq_f32(vout5xCDEF, vf5xCDEF, vfilter_output_scaleCDEF);
434 float32x4_t vf6x0123 = vcvtq_f32_s32(vacc6x0123);
435 vout6x0123 = vfmaq_f32(vout6x0123, vf6x0123, vfilter_output_scale0123);
436 float32x4_t vf6x4567 = vcvtq_f32_s32(vacc6x4567);
437 vout6x4567 = vfmaq_f32(vout6x4567, vf6x4567, vfilter_output_scale4567);
438 float32x4_t vf6x89AB = vcvtq_f32_s32(vacc6x89AB);
439 vout6x89AB = vfmaq_f32(vout6x89AB, vf6x89AB, vfilter_output_scale89AB);
440 float32x4_t vf6xCDEF = vcvtq_f32_s32(vacc6xCDEF);
441 vout6xCDEF = vfmaq_f32(vout6xCDEF, vf6xCDEF, vfilter_output_scaleCDEF);
442 }
443
444 const float32x4_t vinput_scale01 = vreinterpretq_f32_s32(vld1q_s32(&quantization_params[0].zero_point));
445 vout0x0123 = vmulq_lane_f32(vout0x0123, vget_low_f32(vinput_scale01), 1);
446 vout1x0123 = vmulq_lane_f32(vout1x0123, vget_high_f32(vinput_scale01), 1);
447 vout0x4567 = vmulq_lane_f32(vout0x4567, vget_low_f32(vinput_scale01), 1);
448 vout1x4567 = vmulq_lane_f32(vout1x4567, vget_high_f32(vinput_scale01), 1);
449 vout0x89AB = vmulq_lane_f32(vout0x89AB, vget_low_f32(vinput_scale01), 1);
450 vout1x89AB = vmulq_lane_f32(vout1x89AB, vget_high_f32(vinput_scale01), 1);
451 vout0xCDEF = vmulq_lane_f32(vout0xCDEF, vget_low_f32(vinput_scale01), 1);
452 vout1xCDEF = vmulq_lane_f32(vout1xCDEF, vget_high_f32(vinput_scale01), 1);
453 const float32x4_t vinput_scale23 = vreinterpretq_f32_s32(vld1q_s32(&quantization_params[2].zero_point));
454 vout2x0123 = vmulq_lane_f32(vout2x0123, vget_low_f32(vinput_scale23), 1);
455 vout3x0123 = vmulq_lane_f32(vout3x0123, vget_high_f32(vinput_scale23), 1);
456 vout2x4567 = vmulq_lane_f32(vout2x4567, vget_low_f32(vinput_scale23), 1);
457 vout3x4567 = vmulq_lane_f32(vout3x4567, vget_high_f32(vinput_scale23), 1);
458 vout2x89AB = vmulq_lane_f32(vout2x89AB, vget_low_f32(vinput_scale23), 1);
459 vout3x89AB = vmulq_lane_f32(vout3x89AB, vget_high_f32(vinput_scale23), 1);
460 vout2xCDEF = vmulq_lane_f32(vout2xCDEF, vget_low_f32(vinput_scale23), 1);
461 vout3xCDEF = vmulq_lane_f32(vout3xCDEF, vget_high_f32(vinput_scale23), 1);
462 const float32x4_t vinput_scale45 = vreinterpretq_f32_s32(vld1q_s32(&quantization_params[4].zero_point));
463 vout4x0123 = vmulq_lane_f32(vout4x0123, vget_low_f32(vinput_scale45), 1);
464 vout5x0123 = vmulq_lane_f32(vout5x0123, vget_high_f32(vinput_scale45), 1);
465 vout4x4567 = vmulq_lane_f32(vout4x4567, vget_low_f32(vinput_scale45), 1);
466 vout5x4567 = vmulq_lane_f32(vout5x4567, vget_high_f32(vinput_scale45), 1);
467 vout4x89AB = vmulq_lane_f32(vout4x89AB, vget_low_f32(vinput_scale45), 1);
468 vout5x89AB = vmulq_lane_f32(vout5x89AB, vget_high_f32(vinput_scale45), 1);
469 vout4xCDEF = vmulq_lane_f32(vout4xCDEF, vget_low_f32(vinput_scale45), 1);
470 vout5xCDEF = vmulq_lane_f32(vout5xCDEF, vget_high_f32(vinput_scale45), 1);
471 const float32x4_t vinput_scale6 = vld1q_dup_f32(&quantization_params[6].inv_scale);
472 vout6x0123 = vmulq_f32(vout6x0123, vinput_scale6);
473 vout6x4567 = vmulq_f32(vout6x4567, vinput_scale6);
474 vout6x89AB = vmulq_f32(vout6x89AB, vinput_scale6);
475 vout6xCDEF = vmulq_f32(vout6xCDEF, vinput_scale6);
476
477
478 const float32x4_t vbias0123 = vld1q_f32(w); w = (const float*) w + 4;
479 vout0x0123 = vaddq_f32(vbias0123, vout0x0123);
480 vout1x0123 = vaddq_f32(vbias0123, vout1x0123);
481 vout2x0123 = vaddq_f32(vbias0123, vout2x0123);
482 vout3x0123 = vaddq_f32(vbias0123, vout3x0123);
483 vout4x0123 = vaddq_f32(vbias0123, vout4x0123);
484 vout5x0123 = vaddq_f32(vbias0123, vout5x0123);
485 vout6x0123 = vaddq_f32(vbias0123, vout6x0123);
486 const float32x4_t vbias4567 = vld1q_f32(w); w = (const float*) w + 4;
487 vout0x4567 = vaddq_f32(vbias4567, vout0x4567);
488 vout1x4567 = vaddq_f32(vbias4567, vout1x4567);
489 vout2x4567 = vaddq_f32(vbias4567, vout2x4567);
490 vout3x4567 = vaddq_f32(vbias4567, vout3x4567);
491 vout4x4567 = vaddq_f32(vbias4567, vout4x4567);
492 vout5x4567 = vaddq_f32(vbias4567, vout5x4567);
493 vout6x4567 = vaddq_f32(vbias4567, vout6x4567);
494 const float32x4_t vbias89AB = vld1q_f32(w); w = (const float*) w + 4;
495 vout0x89AB = vaddq_f32(vbias89AB, vout0x89AB);
496 vout1x89AB = vaddq_f32(vbias89AB, vout1x89AB);
497 vout2x89AB = vaddq_f32(vbias89AB, vout2x89AB);
498 vout3x89AB = vaddq_f32(vbias89AB, vout3x89AB);
499 vout4x89AB = vaddq_f32(vbias89AB, vout4x89AB);
500 vout5x89AB = vaddq_f32(vbias89AB, vout5x89AB);
501 vout6x89AB = vaddq_f32(vbias89AB, vout6x89AB);
502 const float32x4_t vbiasCDEF = vld1q_f32(w); w = (const float*) w + 4;
503 vout0xCDEF = vaddq_f32(vbiasCDEF, vout0xCDEF);
504 vout1xCDEF = vaddq_f32(vbiasCDEF, vout1xCDEF);
505 vout2xCDEF = vaddq_f32(vbiasCDEF, vout2xCDEF);
506 vout3xCDEF = vaddq_f32(vbiasCDEF, vout3xCDEF);
507 vout4xCDEF = vaddq_f32(vbiasCDEF, vout4xCDEF);
508 vout5xCDEF = vaddq_f32(vbiasCDEF, vout5xCDEF);
509 vout6xCDEF = vaddq_f32(vbiasCDEF, vout6xCDEF);
510
511 float16x8_t vfp16out0x01234567 = vcombine_f16(vcvt_f16_f32(vout0x0123), vcvt_f16_f32(vout0x4567));
512 float16x8_t vfp16out0x89ABCDEF = vcombine_f16(vcvt_f16_f32(vout0x89AB), vcvt_f16_f32(vout0xCDEF));
513 float16x8_t vfp16out1x01234567 = vcombine_f16(vcvt_f16_f32(vout1x0123), vcvt_f16_f32(vout1x4567));
514 float16x8_t vfp16out1x89ABCDEF = vcombine_f16(vcvt_f16_f32(vout1x89AB), vcvt_f16_f32(vout1xCDEF));
515 float16x8_t vfp16out2x01234567 = vcombine_f16(vcvt_f16_f32(vout2x0123), vcvt_f16_f32(vout2x4567));
516 float16x8_t vfp16out2x89ABCDEF = vcombine_f16(vcvt_f16_f32(vout2x89AB), vcvt_f16_f32(vout2xCDEF));
517 float16x8_t vfp16out3x01234567 = vcombine_f16(vcvt_f16_f32(vout3x0123), vcvt_f16_f32(vout3x4567));
518 float16x8_t vfp16out3x89ABCDEF = vcombine_f16(vcvt_f16_f32(vout3x89AB), vcvt_f16_f32(vout3xCDEF));
519 float16x8_t vfp16out4x01234567 = vcombine_f16(vcvt_f16_f32(vout4x0123), vcvt_f16_f32(vout4x4567));
520 float16x8_t vfp16out4x89ABCDEF = vcombine_f16(vcvt_f16_f32(vout4x89AB), vcvt_f16_f32(vout4xCDEF));
521 float16x8_t vfp16out5x01234567 = vcombine_f16(vcvt_f16_f32(vout5x0123), vcvt_f16_f32(vout5x4567));
522 float16x8_t vfp16out5x89ABCDEF = vcombine_f16(vcvt_f16_f32(vout5x89AB), vcvt_f16_f32(vout5xCDEF));
523 float16x8_t vfp16out6x01234567 = vcombine_f16(vcvt_f16_f32(vout6x0123), vcvt_f16_f32(vout6x4567));
524 float16x8_t vfp16out6x89ABCDEF = vcombine_f16(vcvt_f16_f32(vout6x89AB), vcvt_f16_f32(vout6xCDEF));
525
526 const float16x8_t voutput_min = vreinterpretq_f16_u16(vld1q_dup_u16((const uint16_t*) &params->scalar.min));
527 vfp16out0x01234567 = vmaxq_f16(vfp16out0x01234567, voutput_min);
528 vfp16out0x89ABCDEF = vmaxq_f16(vfp16out0x89ABCDEF, voutput_min);
529 vfp16out1x01234567 = vmaxq_f16(vfp16out1x01234567, voutput_min);
530 vfp16out1x89ABCDEF = vmaxq_f16(vfp16out1x89ABCDEF, voutput_min);
531 vfp16out2x01234567 = vmaxq_f16(vfp16out2x01234567, voutput_min);
532 vfp16out2x89ABCDEF = vmaxq_f16(vfp16out2x89ABCDEF, voutput_min);
533 vfp16out3x01234567 = vmaxq_f16(vfp16out3x01234567, voutput_min);
534 vfp16out3x89ABCDEF = vmaxq_f16(vfp16out3x89ABCDEF, voutput_min);
535 vfp16out4x01234567 = vmaxq_f16(vfp16out4x01234567, voutput_min);
536 vfp16out4x89ABCDEF = vmaxq_f16(vfp16out4x89ABCDEF, voutput_min);
537 vfp16out5x01234567 = vmaxq_f16(vfp16out5x01234567, voutput_min);
538 vfp16out5x89ABCDEF = vmaxq_f16(vfp16out5x89ABCDEF, voutput_min);
539 vfp16out6x01234567 = vmaxq_f16(vfp16out6x01234567, voutput_min);
540 vfp16out6x89ABCDEF = vmaxq_f16(vfp16out6x89ABCDEF, voutput_min);
541 const float16x8_t voutput_max = vreinterpretq_f16_u16(vld1q_dup_u16((const uint16_t*) &params->scalar.max));
542 vfp16out0x01234567 = vminq_f16(vfp16out0x01234567, voutput_max);
543 vfp16out0x89ABCDEF = vminq_f16(vfp16out0x89ABCDEF, voutput_max);
544 vfp16out1x01234567 = vminq_f16(vfp16out1x01234567, voutput_max);
545 vfp16out1x89ABCDEF = vminq_f16(vfp16out1x89ABCDEF, voutput_max);
546 vfp16out2x01234567 = vminq_f16(vfp16out2x01234567, voutput_max);
547 vfp16out2x89ABCDEF = vminq_f16(vfp16out2x89ABCDEF, voutput_max);
548 vfp16out3x01234567 = vminq_f16(vfp16out3x01234567, voutput_max);
549 vfp16out3x89ABCDEF = vminq_f16(vfp16out3x89ABCDEF, voutput_max);
550 vfp16out4x01234567 = vminq_f16(vfp16out4x01234567, voutput_max);
551 vfp16out4x89ABCDEF = vminq_f16(vfp16out4x89ABCDEF, voutput_max);
552 vfp16out5x01234567 = vminq_f16(vfp16out5x01234567, voutput_max);
553 vfp16out5x89ABCDEF = vminq_f16(vfp16out5x89ABCDEF, voutput_max);
554 vfp16out6x01234567 = vminq_f16(vfp16out6x01234567, voutput_max);
555 vfp16out6x89ABCDEF = vminq_f16(vfp16out6x89ABCDEF, voutput_max);
556 if XNN_LIKELY(nc >= 16) {
557 vst1q_u16(c0, vreinterpretq_u16_f16(vfp16out0x01234567));
558 vst1q_u16(c0 + 8, vreinterpretq_u16_f16(vfp16out0x89ABCDEF));
559 vst1q_u16(c1, vreinterpretq_u16_f16(vfp16out1x01234567));
560 vst1q_u16(c1 + 8, vreinterpretq_u16_f16(vfp16out1x89ABCDEF));
561 vst1q_u16(c2, vreinterpretq_u16_f16(vfp16out2x01234567));
562 vst1q_u16(c2 + 8, vreinterpretq_u16_f16(vfp16out2x89ABCDEF));
563 vst1q_u16(c3, vreinterpretq_u16_f16(vfp16out3x01234567));
564 vst1q_u16(c3 + 8, vreinterpretq_u16_f16(vfp16out3x89ABCDEF));
565 vst1q_u16(c4, vreinterpretq_u16_f16(vfp16out4x01234567));
566 vst1q_u16(c4 + 8, vreinterpretq_u16_f16(vfp16out4x89ABCDEF));
567 vst1q_u16(c5, vreinterpretq_u16_f16(vfp16out5x01234567));
568 vst1q_u16(c5 + 8, vreinterpretq_u16_f16(vfp16out5x89ABCDEF));
569 vst1q_u16(c6, vreinterpretq_u16_f16(vfp16out6x01234567));
570 vst1q_u16(c6 + 8, vreinterpretq_u16_f16(vfp16out6x89ABCDEF));
571
572 a0 = (const int8_t*) ((uintptr_t) a0 - kc);
573 a1 = (const int8_t*) ((uintptr_t) a1 - kc);
574 a2 = (const int8_t*) ((uintptr_t) a2 - kc);
575 a3 = (const int8_t*) ((uintptr_t) a3 - kc);
576 a4 = (const int8_t*) ((uintptr_t) a4 - kc);
577 a5 = (const int8_t*) ((uintptr_t) a5 - kc);
578 a6 = (const int8_t*) ((uintptr_t) a6 - kc);
579
580 c0 = (uint16_t*) ((uintptr_t) c0 + cn_stride);
581 c1 = (uint16_t*) ((uintptr_t) c1 + cn_stride);
582 c2 = (uint16_t*) ((uintptr_t) c2 + cn_stride);
583 c3 = (uint16_t*) ((uintptr_t) c3 + cn_stride);
584 c4 = (uint16_t*) ((uintptr_t) c4 + cn_stride);
585 c5 = (uint16_t*) ((uintptr_t) c5 + cn_stride);
586 c6 = (uint16_t*) ((uintptr_t) c6 + cn_stride);
587
588 nc -= 16;
589 } else {
590 if (nc & 8) {
591 vst1q_u16(c0, vreinterpretq_u16_f16(vfp16out0x01234567)); c0 += 8;
592 vfp16out0x01234567 = vfp16out0x89ABCDEF;
593 vst1q_u16(c1, vreinterpretq_u16_f16(vfp16out1x01234567)); c1 += 8;
594 vfp16out1x01234567 = vfp16out1x89ABCDEF;
595 vst1q_u16(c2, vreinterpretq_u16_f16(vfp16out2x01234567)); c2 += 8;
596 vfp16out2x01234567 = vfp16out2x89ABCDEF;
597 vst1q_u16(c3, vreinterpretq_u16_f16(vfp16out3x01234567)); c3 += 8;
598 vfp16out3x01234567 = vfp16out3x89ABCDEF;
599 vst1q_u16(c4, vreinterpretq_u16_f16(vfp16out4x01234567)); c4 += 8;
600 vfp16out4x01234567 = vfp16out4x89ABCDEF;
601 vst1q_u16(c5, vreinterpretq_u16_f16(vfp16out5x01234567)); c5 += 8;
602 vfp16out5x01234567 = vfp16out5x89ABCDEF;
603 vst1q_u16(c6, vreinterpretq_u16_f16(vfp16out6x01234567)); c6 += 8;
604 vfp16out6x01234567 = vfp16out6x89ABCDEF;
605 }
606 float16x4_t vfp16out0x0123 = vget_low_f16(vfp16out0x01234567);
607 float16x4_t vfp16out1x0123 = vget_low_f16(vfp16out1x01234567);
608 float16x4_t vfp16out2x0123 = vget_low_f16(vfp16out2x01234567);
609 float16x4_t vfp16out3x0123 = vget_low_f16(vfp16out3x01234567);
610 float16x4_t vfp16out4x0123 = vget_low_f16(vfp16out4x01234567);
611 float16x4_t vfp16out5x0123 = vget_low_f16(vfp16out5x01234567);
612 float16x4_t vfp16out6x0123 = vget_low_f16(vfp16out6x01234567);
613 if (nc & 4) {
614 vst1_u16(c0, vreinterpret_u16_f16(vfp16out0x0123)); c0 += 4;
615 vst1_u16(c1, vreinterpret_u16_f16(vfp16out1x0123)); c1 += 4;
616 vst1_u16(c2, vreinterpret_u16_f16(vfp16out2x0123)); c2 += 4;
617 vst1_u16(c3, vreinterpret_u16_f16(vfp16out3x0123)); c3 += 4;
618 vst1_u16(c4, vreinterpret_u16_f16(vfp16out4x0123)); c4 += 4;
619 vst1_u16(c5, vreinterpret_u16_f16(vfp16out5x0123)); c5 += 4;
620 vst1_u16(c6, vreinterpret_u16_f16(vfp16out6x0123)); c6 += 4;
621 vfp16out0x0123 = vget_high_f16(vfp16out0x01234567);
622 vfp16out1x0123 = vget_high_f16(vfp16out1x01234567);
623 vfp16out2x0123 = vget_high_f16(vfp16out2x01234567);
624 vfp16out3x0123 = vget_high_f16(vfp16out3x01234567);
625 vfp16out4x0123 = vget_high_f16(vfp16out4x01234567);
626 vfp16out5x0123 = vget_high_f16(vfp16out5x01234567);
627 vfp16out6x0123 = vget_high_f16(vfp16out6x01234567);
628 }
629 if (nc & 2) {
630 vst1_lane_u32((void*) c0, vreinterpret_u32_f16(vfp16out0x0123), 0); c0 += 2;
631 vst1_lane_u32((void*) c1, vreinterpret_u32_f16(vfp16out1x0123), 0); c1 += 2;
632 vst1_lane_u32((void*) c2, vreinterpret_u32_f16(vfp16out2x0123), 0); c2 += 2;
633 vst1_lane_u32((void*) c3, vreinterpret_u32_f16(vfp16out3x0123), 0); c3 += 2;
634 vst1_lane_u32((void*) c4, vreinterpret_u32_f16(vfp16out4x0123), 0); c4 += 2;
635 vst1_lane_u32((void*) c5, vreinterpret_u32_f16(vfp16out5x0123), 0); c5 += 2;
636 vst1_lane_u32((void*) c6, vreinterpret_u32_f16(vfp16out6x0123), 0); c6 += 2;
637 vfp16out0x0123 = vext_f16(vfp16out0x0123, vfp16out0x0123, 2);
638 vfp16out1x0123 = vext_f16(vfp16out1x0123, vfp16out1x0123, 2);
639 vfp16out2x0123 = vext_f16(vfp16out2x0123, vfp16out2x0123, 2);
640 vfp16out3x0123 = vext_f16(vfp16out3x0123, vfp16out3x0123, 2);
641 vfp16out4x0123 = vext_f16(vfp16out4x0123, vfp16out4x0123, 2);
642 vfp16out5x0123 = vext_f16(vfp16out5x0123, vfp16out5x0123, 2);
643 vfp16out6x0123 = vext_f16(vfp16out6x0123, vfp16out6x0123, 2);
644 }
645 if (nc & 1) {
646 vst1_lane_u16(c0, vreinterpret_u16_f16(vfp16out0x0123), 0);
647 vst1_lane_u16(c1, vreinterpret_u16_f16(vfp16out1x0123), 0);
648 vst1_lane_u16(c2, vreinterpret_u16_f16(vfp16out2x0123), 0);
649 vst1_lane_u16(c3, vreinterpret_u16_f16(vfp16out3x0123), 0);
650 vst1_lane_u16(c4, vreinterpret_u16_f16(vfp16out4x0123), 0);
651 vst1_lane_u16(c5, vreinterpret_u16_f16(vfp16out5x0123), 0);
652 vst1_lane_u16(c6, vreinterpret_u16_f16(vfp16out6x0123), 0);
653 }
654 nc = 0;
655 }
656 } while (nc != 0);
657 }