Debian Codegraph @

src/qd8-f16-qc8w-igemm/gen/qd8-f16-qc8w-igemm-6x16c4-minmax-neondotfp16arith.c

raw
1 // Auto-generated file. Do not edit!
2 // Template: src/qs8-igemm/c4-neondot.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/igemm.h"
15 #include "xnnpack/math.h"
16
17
18 void xnn_qd8_f16_qc8w_igemm_minmax_ukernel_6x16c4__neondotfp16arith(
19 size_t mr,
20 size_t nc,
21 size_t kc,
22 size_t ks,
23 const int8_t** restrict a,
24 const void* restrict w,
25 xnn_float16* restrict c,
26 size_t cm_stride,
27 size_t cn_stride,
28 size_t a_offset,
29 const int8_t* zero,
30 const int8_t* zero_data,
31 const union xnn_f16_minmax_params params[restrict XNN_MIN_ELEMENTS(1)],
32 const struct xnn_qd8_quantization_params quantization_params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
33 {
34 assert(mr != 0);
35 assert(mr <= 6);
36 assert(nc != 0);
37 assert(kc != 0);
38 assert(ks != 0);
39 assert(ks % (6 * sizeof(void*)) == 0);
40 assert(a_offset % sizeof(int8_t) == 0);
41 assert(a != NULL);
42 assert(w != NULL);
43 assert(c != NULL);
44
45 kc = round_up_po2(kc, 4 * sizeof(int8_t));
46 uint16_t* c0 = (uint16_t*) c;
47 uint16_t* c1 = (uint16_t*) ((uintptr_t) c0 + cm_stride);
48 if XNN_UNPREDICTABLE(mr < 2) {
49 c1 = c0;
50 }
51 uint16_t* c2 = (uint16_t*) ((uintptr_t) c1 + cm_stride);
52 if XNN_UNPREDICTABLE(mr <= 2) {
53 c2 = c1;
54 }
55 uint16_t* c3 = (uint16_t*) ((uintptr_t) c2 + cm_stride);
56 if XNN_UNPREDICTABLE(mr < 4) {
57 c3 = c2;
58 }
59 uint16_t* c4 = (uint16_t*) ((uintptr_t) c3 + cm_stride);
60 if XNN_UNPREDICTABLE(mr <= 4) {
61 c4 = c3;
62 }
63 uint16_t* c5 = (uint16_t*) ((uintptr_t) c4 + cm_stride);
64 if XNN_UNPREDICTABLE(mr != 6) {
65 c5 = c4;
66 }
67
68 do {
69 const int32x4_t vinput_zero_point = vld1q_dup_s32(&quantization_params->zero_point);
70 const int32x4_t vksum0123 = vld1q_s32(w); w = (const int32_t*) w + 4;
71 int32x4_t vacc0x0123 = vmulq_s32(vksum0123, vinput_zero_point);
72 const int32x4_t vksum4567 = vld1q_s32(w); w = (const int32_t*) w + 4;
73 int32x4_t vacc0x4567 = vmulq_s32(vksum4567, vinput_zero_point);
74 const int32x4_t vksum89AB = vld1q_s32(w); w = (const int32_t*) w + 4;
75 int32x4_t vacc0x89AB = vmulq_s32(vksum89AB, vinput_zero_point);
76 const int32x4_t vksumCDEF = vld1q_s32(w); w = (const int32_t*) w + 4;
77 int32x4_t vacc0xCDEF = vmulq_s32(vksumCDEF, vinput_zero_point);
78 int32x4_t vacc1x0123 = vacc0x0123;
79 int32x4_t vacc1x4567 = vacc0x4567;
80 int32x4_t vacc1x89AB = vacc0x89AB;
81 int32x4_t vacc1xCDEF = vacc0xCDEF;
82 int32x4_t vacc2x0123 = vacc0x0123;
83 int32x4_t vacc2x4567 = vacc0x4567;
84 int32x4_t vacc2x89AB = vacc0x89AB;
85 int32x4_t vacc2xCDEF = vacc0xCDEF;
86 int32x4_t vacc3x0123 = vacc0x0123;
87 int32x4_t vacc3x4567 = vacc0x4567;
88 int32x4_t vacc3x89AB = vacc0x89AB;
89 int32x4_t vacc3xCDEF = vacc0xCDEF;
90 int32x4_t vacc4x0123 = vacc0x0123;
91 int32x4_t vacc4x4567 = vacc0x4567;
92 int32x4_t vacc4x89AB = vacc0x89AB;
93 int32x4_t vacc4xCDEF = vacc0xCDEF;
94 int32x4_t vacc5x0123 = vacc0x0123;
95 int32x4_t vacc5x4567 = vacc0x4567;
96 int32x4_t vacc5x89AB = vacc0x89AB;
97 int32x4_t vacc5xCDEF = vacc0xCDEF;
98
99 size_t p = ks;
100 do {
101 const int8_t* restrict a0 = a[0];
102 if XNN_UNPREDICTABLE(a0 != zero) {
103 a0 = (const int8_t*) ((uintptr_t) a0 + a_offset);
104 } else {
105 a0 = zero_data;
106 }
107 const int8_t* restrict a1 = a[1];
108 if XNN_UNPREDICTABLE(a1 != zero) {
109 a1 = (const int8_t*) ((uintptr_t) a1 + a_offset);
110 } else {
111 a1 = zero_data;
112 }
113 const int8_t* restrict a2 = a[2];
114 if XNN_UNPREDICTABLE(a2 != zero) {
115 a2 = (const int8_t*) ((uintptr_t) a2 + a_offset);
116 } else {
117 a2 = zero_data;
118 }
119 const int8_t* restrict a3 = a[3];
120 if XNN_UNPREDICTABLE(a3 != zero) {
121 a3 = (const int8_t*) ((uintptr_t) a3 + a_offset);
122 } else {
123 a3 = zero_data;
124 }
125 const int8_t* restrict a4 = a[4];
126 if XNN_UNPREDICTABLE(a4 != zero) {
127 a4 = (const int8_t*) ((uintptr_t) a4 + a_offset);
128 } else {
129 a4 = zero_data;
130 }
131 const int8_t* restrict a5 = a[5];
132 if XNN_UNPREDICTABLE(a5 != zero) {
133 a5 = (const int8_t*) ((uintptr_t) a5 + a_offset);
134 } else {
135 a5 = zero_data;
136 }
137 a += 6;
138
139 // Inner accumulation loop along the 16 columns.
140 size_t k = kc;
141 // 2x partial unrolled loop to load 8 bytes at a time.
142 while (k >= 8 * sizeof(int8_t)) {
143 // Load a 6x8 block of activations.
144 const int8x8_t va0x01234567 = vld1_s8(a0); a0 += 8;
145 const int8x8_t va1x01234567 = vld1_s8(a1); a1 += 8;
146 const int8x8_t va2x01234567 = vld1_s8(a2); a2 += 8;
147 const int8x8_t va3x01234567 = vld1_s8(a3); a3 += 8;
148 const int8x8_t va4x01234567 = vld1_s8(a4); a4 += 8;
149 const int8x8_t va5x01234567 = vld1_s8(a5); a5 += 8;
150
151 // Load a 8x16 block of weights.
152 const int8x16_t vb0123x0123 = vld1q_s8(w); w = (const int8_t*) w + 16;
153 const int8x16_t vb0123x4567 = vld1q_s8(w); w = (const int8_t*) w + 16;
154 const int8x16_t vb0123x89AB = vld1q_s8(w); w = (const int8_t*) w + 16;
155 const int8x16_t vb0123xCDEF = vld1q_s8(w); w = (const int8_t*) w + 16;
156 const int8x16_t vb4567x0123 = vld1q_s8(w); w = (const int8_t*) w + 16;
157 const int8x16_t vb4567x4567 = vld1q_s8(w); w = (const int8_t*) w + 16;
158 const int8x16_t vb4567x89AB = vld1q_s8(w); w = (const int8_t*) w + 16;
159 const int8x16_t vb4567xCDEF = vld1q_s8(w); w = (const int8_t*) w + 16;
160
161 // Multiply-accumulate: 6x8 * 8x16 --> 6x16.
162 vacc0x0123 = vdotq_lane_s32(vacc0x0123, vb0123x0123, va0x01234567, 0);
163 vacc0x4567 = vdotq_lane_s32(vacc0x4567, vb0123x4567, va0x01234567, 0);
164 vacc0x89AB = vdotq_lane_s32(vacc0x89AB, vb0123x89AB, va0x01234567, 0);
165 vacc0xCDEF = vdotq_lane_s32(vacc0xCDEF, vb0123xCDEF, va0x01234567, 0);
166 vacc1x0123 = vdotq_lane_s32(vacc1x0123, vb0123x0123, va1x01234567, 0);
167 vacc1x4567 = vdotq_lane_s32(vacc1x4567, vb0123x4567, va1x01234567, 0);
168 vacc1x89AB = vdotq_lane_s32(vacc1x89AB, vb0123x89AB, va1x01234567, 0);
169 vacc1xCDEF = vdotq_lane_s32(vacc1xCDEF, vb0123xCDEF, va1x01234567, 0);
170 vacc2x0123 = vdotq_lane_s32(vacc2x0123, vb0123x0123, va2x01234567, 0);
171 vacc2x4567 = vdotq_lane_s32(vacc2x4567, vb0123x4567, va2x01234567, 0);
172 vacc2x89AB = vdotq_lane_s32(vacc2x89AB, vb0123x89AB, va2x01234567, 0);
173 vacc2xCDEF = vdotq_lane_s32(vacc2xCDEF, vb0123xCDEF, va2x01234567, 0);
174 vacc3x0123 = vdotq_lane_s32(vacc3x0123, vb0123x0123, va3x01234567, 0);
175 vacc3x4567 = vdotq_lane_s32(vacc3x4567, vb0123x4567, va3x01234567, 0);
176 vacc3x89AB = vdotq_lane_s32(vacc3x89AB, vb0123x89AB, va3x01234567, 0);
177 vacc3xCDEF = vdotq_lane_s32(vacc3xCDEF, vb0123xCDEF, va3x01234567, 0);
178 vacc4x0123 = vdotq_lane_s32(vacc4x0123, vb0123x0123, va4x01234567, 0);
179 vacc4x4567 = vdotq_lane_s32(vacc4x4567, vb0123x4567, va4x01234567, 0);
180 vacc4x89AB = vdotq_lane_s32(vacc4x89AB, vb0123x89AB, va4x01234567, 0);
181 vacc4xCDEF = vdotq_lane_s32(vacc4xCDEF, vb0123xCDEF, va4x01234567, 0);
182 vacc5x0123 = vdotq_lane_s32(vacc5x0123, vb0123x0123, va5x01234567, 0);
183 vacc5x4567 = vdotq_lane_s32(vacc5x4567, vb0123x4567, va5x01234567, 0);
184 vacc5x89AB = vdotq_lane_s32(vacc5x89AB, vb0123x89AB, va5x01234567, 0);
185 vacc5xCDEF = vdotq_lane_s32(vacc5xCDEF, vb0123xCDEF, va5x01234567, 0);
186 vacc0x0123 = vdotq_lane_s32(vacc0x0123, vb4567x0123, va0x01234567, 1);
187 vacc0x4567 = vdotq_lane_s32(vacc0x4567, vb4567x4567, va0x01234567, 1);
188 vacc0x89AB = vdotq_lane_s32(vacc0x89AB, vb4567x89AB, va0x01234567, 1);
189 vacc0xCDEF = vdotq_lane_s32(vacc0xCDEF, vb4567xCDEF, va0x01234567, 1);
190 vacc1x0123 = vdotq_lane_s32(vacc1x0123, vb4567x0123, va1x01234567, 1);
191 vacc1x4567 = vdotq_lane_s32(vacc1x4567, vb4567x4567, va1x01234567, 1);
192 vacc1x89AB = vdotq_lane_s32(vacc1x89AB, vb4567x89AB, va1x01234567, 1);
193 vacc1xCDEF = vdotq_lane_s32(vacc1xCDEF, vb4567xCDEF, va1x01234567, 1);
194 vacc2x0123 = vdotq_lane_s32(vacc2x0123, vb4567x0123, va2x01234567, 1);
195 vacc2x4567 = vdotq_lane_s32(vacc2x4567, vb4567x4567, va2x01234567, 1);
196 vacc2x89AB = vdotq_lane_s32(vacc2x89AB, vb4567x89AB, va2x01234567, 1);
197 vacc2xCDEF = vdotq_lane_s32(vacc2xCDEF, vb4567xCDEF, va2x01234567, 1);
198 vacc3x0123 = vdotq_lane_s32(vacc3x0123, vb4567x0123, va3x01234567, 1);
199 vacc3x4567 = vdotq_lane_s32(vacc3x4567, vb4567x4567, va3x01234567, 1);
200 vacc3x89AB = vdotq_lane_s32(vacc3x89AB, vb4567x89AB, va3x01234567, 1);
201 vacc3xCDEF = vdotq_lane_s32(vacc3xCDEF, vb4567xCDEF, va3x01234567, 1);
202 vacc4x0123 = vdotq_lane_s32(vacc4x0123, vb4567x0123, va4x01234567, 1);
203 vacc4x4567 = vdotq_lane_s32(vacc4x4567, vb4567x4567, va4x01234567, 1);
204 vacc4x89AB = vdotq_lane_s32(vacc4x89AB, vb4567x89AB, va4x01234567, 1);
205 vacc4xCDEF = vdotq_lane_s32(vacc4xCDEF, vb4567xCDEF, va4x01234567, 1);
206 vacc5x0123 = vdotq_lane_s32(vacc5x0123, vb4567x0123, va5x01234567, 1);
207 vacc5x4567 = vdotq_lane_s32(vacc5x4567, vb4567x4567, va5x01234567, 1);
208 vacc5x89AB = vdotq_lane_s32(vacc5x89AB, vb4567x89AB, va5x01234567, 1);
209 vacc5xCDEF = vdotq_lane_s32(vacc5xCDEF, vb4567xCDEF, va5x01234567, 1);
210
211 k -= 8 * sizeof(int8_t);
212 }
213 // Handle up to 4 final positions of `k`
214 if XNN_UNLIKELY(k != 0) {
215 // Load a 6x4 block of activations.
216 const int8x8_t va0x01234567 = vld1_s8(a0);
217 const int8x8_t va1x01234567 = vld1_s8(a1);
218 const int8x8_t va2x01234567 = vld1_s8(a2);
219 const int8x8_t va3x01234567 = vld1_s8(a3);
220 const int8x8_t va4x01234567 = vld1_s8(a4);
221 const int8x8_t va5x01234567 = vld1_s8(a5);
222
223 // Load a 4x16 block of weights.
224 const int8x16_t vb0123x0123 = vld1q_s8(w); w = (const int8_t*) w + 16;
225 const int8x16_t vb0123x4567 = vld1q_s8(w); w = (const int8_t*) w + 16;
226 const int8x16_t vb0123x89AB = vld1q_s8(w); w = (const int8_t*) w + 16;
227 const int8x16_t vb0123xCDEF = vld1q_s8(w); w = (const int8_t*) w + 16;
228
229 // Multiply-accumulate: 6x4 * 4x16 --> 6x16.
230 vacc0x0123 = vdotq_lane_s32(vacc0x0123, vb0123x0123, va0x01234567, 0);
231 vacc0x4567 = vdotq_lane_s32(vacc0x4567, vb0123x4567, va0x01234567, 0);
232 vacc0x89AB = vdotq_lane_s32(vacc0x89AB, vb0123x89AB, va0x01234567, 0);
233 vacc0xCDEF = vdotq_lane_s32(vacc0xCDEF, vb0123xCDEF, va0x01234567, 0);
234 vacc1x0123 = vdotq_lane_s32(vacc1x0123, vb0123x0123, va1x01234567, 0);
235 vacc1x4567 = vdotq_lane_s32(vacc1x4567, vb0123x4567, va1x01234567, 0);
236 vacc1x89AB = vdotq_lane_s32(vacc1x89AB, vb0123x89AB, va1x01234567, 0);
237 vacc1xCDEF = vdotq_lane_s32(vacc1xCDEF, vb0123xCDEF, va1x01234567, 0);
238 vacc2x0123 = vdotq_lane_s32(vacc2x0123, vb0123x0123, va2x01234567, 0);
239 vacc2x4567 = vdotq_lane_s32(vacc2x4567, vb0123x4567, va2x01234567, 0);
240 vacc2x89AB = vdotq_lane_s32(vacc2x89AB, vb0123x89AB, va2x01234567, 0);
241 vacc2xCDEF = vdotq_lane_s32(vacc2xCDEF, vb0123xCDEF, va2x01234567, 0);
242 vacc3x0123 = vdotq_lane_s32(vacc3x0123, vb0123x0123, va3x01234567, 0);
243 vacc3x4567 = vdotq_lane_s32(vacc3x4567, vb0123x4567, va3x01234567, 0);
244 vacc3x89AB = vdotq_lane_s32(vacc3x89AB, vb0123x89AB, va3x01234567, 0);
245 vacc3xCDEF = vdotq_lane_s32(vacc3xCDEF, vb0123xCDEF, va3x01234567, 0);
246 vacc4x0123 = vdotq_lane_s32(vacc4x0123, vb0123x0123, va4x01234567, 0);
247 vacc4x4567 = vdotq_lane_s32(vacc4x4567, vb0123x4567, va4x01234567, 0);
248 vacc4x89AB = vdotq_lane_s32(vacc4x89AB, vb0123x89AB, va4x01234567, 0);
249 vacc4xCDEF = vdotq_lane_s32(vacc4xCDEF, vb0123xCDEF, va4x01234567, 0);
250 vacc5x0123 = vdotq_lane_s32(vacc5x0123, vb0123x0123, va5x01234567, 0);
251 vacc5x4567 = vdotq_lane_s32(vacc5x4567, vb0123x4567, va5x01234567, 0);
252 vacc5x89AB = vdotq_lane_s32(vacc5x89AB, vb0123x89AB, va5x01234567, 0);
253 vacc5xCDEF = vdotq_lane_s32(vacc5xCDEF, vb0123xCDEF, va5x01234567, 0);
254 }
255 p -= 6 * sizeof(void*);
256 } while (p != 0);
257
258 float32x4_t vout0x0123 = vcvtq_f32_s32(vacc0x0123);
259 float32x4_t vout0x4567 = vcvtq_f32_s32(vacc0x4567);
260 float32x4_t vout0x89AB = vcvtq_f32_s32(vacc0x89AB);
261 float32x4_t vout0xCDEF = vcvtq_f32_s32(vacc0xCDEF);
262 float32x4_t vout1x0123 = vcvtq_f32_s32(vacc1x0123);
263 float32x4_t vout1x4567 = vcvtq_f32_s32(vacc1x4567);
264 float32x4_t vout1x89AB = vcvtq_f32_s32(vacc1x89AB);
265 float32x4_t vout1xCDEF = vcvtq_f32_s32(vacc1xCDEF);
266 float32x4_t vout2x0123 = vcvtq_f32_s32(vacc2x0123);
267 float32x4_t vout2x4567 = vcvtq_f32_s32(vacc2x4567);
268 float32x4_t vout2x89AB = vcvtq_f32_s32(vacc2x89AB);
269 float32x4_t vout2xCDEF = vcvtq_f32_s32(vacc2xCDEF);
270 float32x4_t vout3x0123 = vcvtq_f32_s32(vacc3x0123);
271 float32x4_t vout3x4567 = vcvtq_f32_s32(vacc3x4567);
272 float32x4_t vout3x89AB = vcvtq_f32_s32(vacc3x89AB);
273 float32x4_t vout3xCDEF = vcvtq_f32_s32(vacc3xCDEF);
274 float32x4_t vout4x0123 = vcvtq_f32_s32(vacc4x0123);
275 float32x4_t vout4x4567 = vcvtq_f32_s32(vacc4x4567);
276 float32x4_t vout4x89AB = vcvtq_f32_s32(vacc4x89AB);
277 float32x4_t vout4xCDEF = vcvtq_f32_s32(vacc4xCDEF);
278 float32x4_t vout5x0123 = vcvtq_f32_s32(vacc5x0123);
279 float32x4_t vout5x4567 = vcvtq_f32_s32(vacc5x4567);
280 float32x4_t vout5x89AB = vcvtq_f32_s32(vacc5x89AB);
281 float32x4_t vout5xCDEF = vcvtq_f32_s32(vacc5xCDEF);
282 const float32x4_t vinput_scale = vld1q_dup_f32(&quantization_params->inv_scale);
283 vout0x0123 = vmulq_f32(vout0x0123, vinput_scale);
284 vout0x4567 = vmulq_f32(vout0x4567, vinput_scale);
285 vout0x89AB = vmulq_f32(vout0x89AB, vinput_scale);
286 vout0xCDEF = vmulq_f32(vout0xCDEF, vinput_scale);
287 vout1x0123 = vmulq_f32(vout1x0123, vinput_scale);
288 vout1x4567 = vmulq_f32(vout1x4567, vinput_scale);
289 vout1x89AB = vmulq_f32(vout1x89AB, vinput_scale);
290 vout1xCDEF = vmulq_f32(vout1xCDEF, vinput_scale);
291 vout2x0123 = vmulq_f32(vout2x0123, vinput_scale);
292 vout2x4567 = vmulq_f32(vout2x4567, vinput_scale);
293 vout2x89AB = vmulq_f32(vout2x89AB, vinput_scale);
294 vout2xCDEF = vmulq_f32(vout2xCDEF, vinput_scale);
295 vout3x0123 = vmulq_f32(vout3x0123, vinput_scale);
296 vout3x4567 = vmulq_f32(vout3x4567, vinput_scale);
297 vout3x89AB = vmulq_f32(vout3x89AB, vinput_scale);
298 vout3xCDEF = vmulq_f32(vout3xCDEF, vinput_scale);
299 vout4x0123 = vmulq_f32(vout4x0123, vinput_scale);
300 vout4x4567 = vmulq_f32(vout4x4567, vinput_scale);
301 vout4x89AB = vmulq_f32(vout4x89AB, vinput_scale);
302 vout4xCDEF = vmulq_f32(vout4xCDEF, vinput_scale);
303 vout5x0123 = vmulq_f32(vout5x0123, vinput_scale);
304 vout5x4567 = vmulq_f32(vout5x4567, vinput_scale);
305 vout5x89AB = vmulq_f32(vout5x89AB, vinput_scale);
306 vout5xCDEF = vmulq_f32(vout5xCDEF, vinput_scale);
307
308 const float32x4_t vfilter_output_scale0123 = vld1q_f32(w); w = (const float*) w + 4;
309 const float32x4_t vfilter_output_scale4567 = vld1q_f32(w); w = (const float*) w + 4;
310 const float32x4_t vfilter_output_scale89AB = vld1q_f32(w); w = (const float*) w + 4;
311 const float32x4_t vfilter_output_scaleCDEF = vld1q_f32(w); w = (const float*) w + 4;
312
313 const float32x4_t vbias0123 = vld1q_f32(w); w = (const float*) w + 4;
314 #if XNN_ARCH_ARM64
315 vout0x0123 = vfmaq_f32(vbias0123, vout0x0123, vfilter_output_scale0123);
316 vout1x0123 = vfmaq_f32(vbias0123, vout1x0123, vfilter_output_scale0123);
317 vout2x0123 = vfmaq_f32(vbias0123, vout2x0123, vfilter_output_scale0123);
318 vout3x0123 = vfmaq_f32(vbias0123, vout3x0123, vfilter_output_scale0123);
319 vout4x0123 = vfmaq_f32(vbias0123, vout4x0123, vfilter_output_scale0123);
320 vout5x0123 = vfmaq_f32(vbias0123, vout5x0123, vfilter_output_scale0123);
321 #else
322 vout0x0123 = vmlaq_f32(vbias0123, vout0x0123, vfilter_output_scale0123);
323 vout1x0123 = vmlaq_f32(vbias0123, vout1x0123, vfilter_output_scale0123);
324 vout2x0123 = vmlaq_f32(vbias0123, vout2x0123, vfilter_output_scale0123);
325 vout3x0123 = vmlaq_f32(vbias0123, vout3x0123, vfilter_output_scale0123);
326 vout4x0123 = vmlaq_f32(vbias0123, vout4x0123, vfilter_output_scale0123);
327 vout5x0123 = vmlaq_f32(vbias0123, vout5x0123, vfilter_output_scale0123);
328 #endif
329 const float32x4_t vbias4567 = vld1q_f32(w); w = (const float*) w + 4;
330 #if XNN_ARCH_ARM64
331 vout0x4567 = vfmaq_f32(vbias4567, vout0x4567, vfilter_output_scale4567);
332 vout1x4567 = vfmaq_f32(vbias4567, vout1x4567, vfilter_output_scale4567);
333 vout2x4567 = vfmaq_f32(vbias4567, vout2x4567, vfilter_output_scale4567);
334 vout3x4567 = vfmaq_f32(vbias4567, vout3x4567, vfilter_output_scale4567);
335 vout4x4567 = vfmaq_f32(vbias4567, vout4x4567, vfilter_output_scale4567);
336 vout5x4567 = vfmaq_f32(vbias4567, vout5x4567, vfilter_output_scale4567);
337 #else
338 vout0x4567 = vmlaq_f32(vbias4567, vout0x4567, vfilter_output_scale4567);
339 vout1x4567 = vmlaq_f32(vbias4567, vout1x4567, vfilter_output_scale4567);
340 vout2x4567 = vmlaq_f32(vbias4567, vout2x4567, vfilter_output_scale4567);
341 vout3x4567 = vmlaq_f32(vbias4567, vout3x4567, vfilter_output_scale4567);
342 vout4x4567 = vmlaq_f32(vbias4567, vout4x4567, vfilter_output_scale4567);
343 vout5x4567 = vmlaq_f32(vbias4567, vout5x4567, vfilter_output_scale4567);
344 #endif
345 const float32x4_t vbias89AB = vld1q_f32(w); w = (const float*) w + 4;
346 #if XNN_ARCH_ARM64
347 vout0x89AB = vfmaq_f32(vbias89AB, vout0x89AB, vfilter_output_scale89AB);
348 vout1x89AB = vfmaq_f32(vbias89AB, vout1x89AB, vfilter_output_scale89AB);
349 vout2x89AB = vfmaq_f32(vbias89AB, vout2x89AB, vfilter_output_scale89AB);
350 vout3x89AB = vfmaq_f32(vbias89AB, vout3x89AB, vfilter_output_scale89AB);
351 vout4x89AB = vfmaq_f32(vbias89AB, vout4x89AB, vfilter_output_scale89AB);
352 vout5x89AB = vfmaq_f32(vbias89AB, vout5x89AB, vfilter_output_scale89AB);
353 #else
354 vout0x89AB = vmlaq_f32(vbias89AB, vout0x89AB, vfilter_output_scale89AB);
355 vout1x89AB = vmlaq_f32(vbias89AB, vout1x89AB, vfilter_output_scale89AB);
356 vout2x89AB = vmlaq_f32(vbias89AB, vout2x89AB, vfilter_output_scale89AB);
357 vout3x89AB = vmlaq_f32(vbias89AB, vout3x89AB, vfilter_output_scale89AB);
358 vout4x89AB = vmlaq_f32(vbias89AB, vout4x89AB, vfilter_output_scale89AB);
359 vout5x89AB = vmlaq_f32(vbias89AB, vout5x89AB, vfilter_output_scale89AB);
360 #endif
361 const float32x4_t vbiasCDEF = vld1q_f32(w); w = (const float*) w + 4;
362 #if XNN_ARCH_ARM64
363 vout0xCDEF = vfmaq_f32(vbiasCDEF, vout0xCDEF, vfilter_output_scaleCDEF);
364 vout1xCDEF = vfmaq_f32(vbiasCDEF, vout1xCDEF, vfilter_output_scaleCDEF);
365 vout2xCDEF = vfmaq_f32(vbiasCDEF, vout2xCDEF, vfilter_output_scaleCDEF);
366 vout3xCDEF = vfmaq_f32(vbiasCDEF, vout3xCDEF, vfilter_output_scaleCDEF);
367 vout4xCDEF = vfmaq_f32(vbiasCDEF, vout4xCDEF, vfilter_output_scaleCDEF);
368 vout5xCDEF = vfmaq_f32(vbiasCDEF, vout5xCDEF, vfilter_output_scaleCDEF);
369 #else
370 vout0xCDEF = vmlaq_f32(vbiasCDEF, vout0xCDEF, vfilter_output_scaleCDEF);
371 vout1xCDEF = vmlaq_f32(vbiasCDEF, vout1xCDEF, vfilter_output_scaleCDEF);
372 vout2xCDEF = vmlaq_f32(vbiasCDEF, vout2xCDEF, vfilter_output_scaleCDEF);
373 vout3xCDEF = vmlaq_f32(vbiasCDEF, vout3xCDEF, vfilter_output_scaleCDEF);
374 vout4xCDEF = vmlaq_f32(vbiasCDEF, vout4xCDEF, vfilter_output_scaleCDEF);
375 vout5xCDEF = vmlaq_f32(vbiasCDEF, vout5xCDEF, vfilter_output_scaleCDEF);
376 #endif
377
378 float16x8_t vfp16out0x01234567 = vcombine_f16(vcvt_f16_f32(vout0x0123), vcvt_f16_f32(vout0x4567));
379 float16x8_t vfp16out0x89ABCDEF = vcombine_f16(vcvt_f16_f32(vout0x89AB), vcvt_f16_f32(vout0xCDEF));
380 float16x8_t vfp16out1x01234567 = vcombine_f16(vcvt_f16_f32(vout1x0123), vcvt_f16_f32(vout1x4567));
381 float16x8_t vfp16out1x89ABCDEF = vcombine_f16(vcvt_f16_f32(vout1x89AB), vcvt_f16_f32(vout1xCDEF));
382 float16x8_t vfp16out2x01234567 = vcombine_f16(vcvt_f16_f32(vout2x0123), vcvt_f16_f32(vout2x4567));
383 float16x8_t vfp16out2x89ABCDEF = vcombine_f16(vcvt_f16_f32(vout2x89AB), vcvt_f16_f32(vout2xCDEF));
384 float16x8_t vfp16out3x01234567 = vcombine_f16(vcvt_f16_f32(vout3x0123), vcvt_f16_f32(vout3x4567));
385 float16x8_t vfp16out3x89ABCDEF = vcombine_f16(vcvt_f16_f32(vout3x89AB), vcvt_f16_f32(vout3xCDEF));
386 float16x8_t vfp16out4x01234567 = vcombine_f16(vcvt_f16_f32(vout4x0123), vcvt_f16_f32(vout4x4567));
387 float16x8_t vfp16out4x89ABCDEF = vcombine_f16(vcvt_f16_f32(vout4x89AB), vcvt_f16_f32(vout4xCDEF));
388 float16x8_t vfp16out5x01234567 = vcombine_f16(vcvt_f16_f32(vout5x0123), vcvt_f16_f32(vout5x4567));
389 float16x8_t vfp16out5x89ABCDEF = vcombine_f16(vcvt_f16_f32(vout5x89AB), vcvt_f16_f32(vout5xCDEF));
390
391 const float16x8_t voutput_min = vreinterpretq_f16_u16(vld1q_dup_u16((const uint16_t*) &params->scalar.min));
392 vfp16out0x01234567 = vmaxq_f16(vfp16out0x01234567, voutput_min);
393 vfp16out0x89ABCDEF = vmaxq_f16(vfp16out0x89ABCDEF, voutput_min);
394 vfp16out1x01234567 = vmaxq_f16(vfp16out1x01234567, voutput_min);
395 vfp16out1x89ABCDEF = vmaxq_f16(vfp16out1x89ABCDEF, voutput_min);
396 vfp16out2x01234567 = vmaxq_f16(vfp16out2x01234567, voutput_min);
397 vfp16out2x89ABCDEF = vmaxq_f16(vfp16out2x89ABCDEF, voutput_min);
398 vfp16out3x01234567 = vmaxq_f16(vfp16out3x01234567, voutput_min);
399 vfp16out3x89ABCDEF = vmaxq_f16(vfp16out3x89ABCDEF, voutput_min);
400 vfp16out4x01234567 = vmaxq_f16(vfp16out4x01234567, voutput_min);
401 vfp16out4x89ABCDEF = vmaxq_f16(vfp16out4x89ABCDEF, voutput_min);
402 vfp16out5x01234567 = vmaxq_f16(vfp16out5x01234567, voutput_min);
403 vfp16out5x89ABCDEF = vmaxq_f16(vfp16out5x89ABCDEF, voutput_min);
404 const float16x8_t voutput_max = vreinterpretq_f16_u16(vld1q_dup_u16((const uint16_t*) &params->scalar.max));
405 vfp16out0x01234567 = vminq_f16(vfp16out0x01234567, voutput_max);
406 vfp16out0x89ABCDEF = vminq_f16(vfp16out0x89ABCDEF, voutput_max);
407 vfp16out1x01234567 = vminq_f16(vfp16out1x01234567, voutput_max);
408 vfp16out1x89ABCDEF = vminq_f16(vfp16out1x89ABCDEF, voutput_max);
409 vfp16out2x01234567 = vminq_f16(vfp16out2x01234567, voutput_max);
410 vfp16out2x89ABCDEF = vminq_f16(vfp16out2x89ABCDEF, voutput_max);
411 vfp16out3x01234567 = vminq_f16(vfp16out3x01234567, voutput_max);
412 vfp16out3x89ABCDEF = vminq_f16(vfp16out3x89ABCDEF, voutput_max);
413 vfp16out4x01234567 = vminq_f16(vfp16out4x01234567, voutput_max);
414 vfp16out4x89ABCDEF = vminq_f16(vfp16out4x89ABCDEF, voutput_max);
415 vfp16out5x01234567 = vminq_f16(vfp16out5x01234567, voutput_max);
416 vfp16out5x89ABCDEF = vminq_f16(vfp16out5x89ABCDEF, voutput_max);
417 if XNN_LIKELY(nc >= 16) {
418 vst1q_u16(c5, vreinterpretq_u16_f16(vfp16out5x01234567));
419 vst1q_u16(c5 + 8, vreinterpretq_u16_f16(vfp16out5x89ABCDEF));
420 vst1q_u16(c4, vreinterpretq_u16_f16(vfp16out4x01234567));
421 vst1q_u16(c4 + 8, vreinterpretq_u16_f16(vfp16out4x89ABCDEF));
422 vst1q_u16(c3, vreinterpretq_u16_f16(vfp16out3x01234567));
423 vst1q_u16(c3 + 8, vreinterpretq_u16_f16(vfp16out3x89ABCDEF));
424 vst1q_u16(c2, vreinterpretq_u16_f16(vfp16out2x01234567));
425 vst1q_u16(c2 + 8, vreinterpretq_u16_f16(vfp16out2x89ABCDEF));
426 vst1q_u16(c1, vreinterpretq_u16_f16(vfp16out1x01234567));
427 vst1q_u16(c1 + 8, vreinterpretq_u16_f16(vfp16out1x89ABCDEF));
428 vst1q_u16(c0, vreinterpretq_u16_f16(vfp16out0x01234567));
429 vst1q_u16(c0 + 8, vreinterpretq_u16_f16(vfp16out0x89ABCDEF));
430
431 a = (const int8_t**restrict) ((uintptr_t) a - ks);
432
433 c0 = (uint16_t*) ((uintptr_t) c0 + cn_stride);
434 c1 = (uint16_t*) ((uintptr_t) c1 + cn_stride);
435 c2 = (uint16_t*) ((uintptr_t) c2 + cn_stride);
436 c3 = (uint16_t*) ((uintptr_t) c3 + cn_stride);
437 c4 = (uint16_t*) ((uintptr_t) c4 + cn_stride);
438 c5 = (uint16_t*) ((uintptr_t) c5 + cn_stride);
439
440 nc -= 16;
441 } else {
442 if (nc & 8) {
443 vst1q_u16(c5, vreinterpretq_u16_f16(vfp16out5x01234567)); c5 += 8;
444 vfp16out5x01234567 = vfp16out5x89ABCDEF;
445 vst1q_u16(c4, vreinterpretq_u16_f16(vfp16out4x01234567)); c4 += 8;
446 vfp16out4x01234567 = vfp16out4x89ABCDEF;
447 vst1q_u16(c3, vreinterpretq_u16_f16(vfp16out3x01234567)); c3 += 8;
448 vfp16out3x01234567 = vfp16out3x89ABCDEF;
449 vst1q_u16(c2, vreinterpretq_u16_f16(vfp16out2x01234567)); c2 += 8;
450 vfp16out2x01234567 = vfp16out2x89ABCDEF;
451 vst1q_u16(c1, vreinterpretq_u16_f16(vfp16out1x01234567)); c1 += 8;
452 vfp16out1x01234567 = vfp16out1x89ABCDEF;
453 vst1q_u16(c0, vreinterpretq_u16_f16(vfp16out0x01234567)); c0 += 8;
454 vfp16out0x01234567 = vfp16out0x89ABCDEF;
455 }
456 float16x4_t vfp16out5x0123 = vget_low_f16(vfp16out5x01234567);
457 float16x4_t vfp16out4x0123 = vget_low_f16(vfp16out4x01234567);
458 float16x4_t vfp16out3x0123 = vget_low_f16(vfp16out3x01234567);
459 float16x4_t vfp16out2x0123 = vget_low_f16(vfp16out2x01234567);
460 float16x4_t vfp16out1x0123 = vget_low_f16(vfp16out1x01234567);
461 float16x4_t vfp16out0x0123 = vget_low_f16(vfp16out0x01234567);
462 if (nc & 4) {
463 vst1_u16(c5, vreinterpret_u16_f16(vfp16out5x0123)); c5 += 4;
464 vst1_u16(c4, vreinterpret_u16_f16(vfp16out4x0123)); c4 += 4;
465 vst1_u16(c3, vreinterpret_u16_f16(vfp16out3x0123)); c3 += 4;
466 vst1_u16(c2, vreinterpret_u16_f16(vfp16out2x0123)); c2 += 4;
467 vst1_u16(c1, vreinterpret_u16_f16(vfp16out1x0123)); c1 += 4;
468 vst1_u16(c0, vreinterpret_u16_f16(vfp16out0x0123)); c0 += 4;
469 vfp16out5x0123 = vget_high_f16(vfp16out5x01234567);
470 vfp16out4x0123 = vget_high_f16(vfp16out4x01234567);
471 vfp16out3x0123 = vget_high_f16(vfp16out3x01234567);
472 vfp16out2x0123 = vget_high_f16(vfp16out2x01234567);
473 vfp16out1x0123 = vget_high_f16(vfp16out1x01234567);
474 vfp16out0x0123 = vget_high_f16(vfp16out0x01234567);
475 }
476 if (nc & 2) {
477 vst1_lane_u32((void*) c5, vreinterpret_u32_f16(vfp16out5x0123), 0); c5 += 2;
478 vst1_lane_u32((void*) c4, vreinterpret_u32_f16(vfp16out4x0123), 0); c4 += 2;
479 vst1_lane_u32((void*) c3, vreinterpret_u32_f16(vfp16out3x0123), 0); c3 += 2;
480 vst1_lane_u32((void*) c2, vreinterpret_u32_f16(vfp16out2x0123), 0); c2 += 2;
481 vst1_lane_u32((void*) c1, vreinterpret_u32_f16(vfp16out1x0123), 0); c1 += 2;
482 vst1_lane_u32((void*) c0, vreinterpret_u32_f16(vfp16out0x0123), 0); c0 += 2;
483 vfp16out5x0123 = vext_f16(vfp16out5x0123, vfp16out5x0123, 2);
484 vfp16out4x0123 = vext_f16(vfp16out4x0123, vfp16out4x0123, 2);
485 vfp16out3x0123 = vext_f16(vfp16out3x0123, vfp16out3x0123, 2);
486 vfp16out2x0123 = vext_f16(vfp16out2x0123, vfp16out2x0123, 2);
487 vfp16out1x0123 = vext_f16(vfp16out1x0123, vfp16out1x0123, 2);
488 vfp16out0x0123 = vext_f16(vfp16out0x0123, vfp16out0x0123, 2);
489 }
490 if (nc & 1) {
491 vst1_lane_u16(c5, vreinterpret_u16_f16(vfp16out5x0123), 0);
492 vst1_lane_u16(c4, vreinterpret_u16_f16(vfp16out4x0123), 0);
493 vst1_lane_u16(c3, vreinterpret_u16_f16(vfp16out3x0123), 0);
494 vst1_lane_u16(c2, vreinterpret_u16_f16(vfp16out2x0123), 0);
495 vst1_lane_u16(c1, vreinterpret_u16_f16(vfp16out1x0123), 0);
496 vst1_lane_u16(c0, vreinterpret_u16_f16(vfp16out0x0123), 0);
497 }
498 nc = 0;
499 }
500 } while (nc != 0);
501 }