Debian Codegraph @

src/qd8-f32-qc8w-igemm/gen/qd8-f32-qc8w-igemm-6x16c4-minmax-neondot.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_f32_qc8w_igemm_minmax_ukernel_6x16c4__neondot(
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 float* 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_f32_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 float* c0 = c;
47 float* c1 = (float*) ((uintptr_t) c0 + cm_stride);
48 if XNN_UNPREDICTABLE(mr < 2) {
49 c1 = c0;
50 }
51 float* c2 = (float*) ((uintptr_t) c1 + cm_stride);
52 if XNN_UNPREDICTABLE(mr <= 2) {
53 c2 = c1;
54 }
55 float* c3 = (float*) ((uintptr_t) c2 + cm_stride);
56 if XNN_UNPREDICTABLE(mr < 4) {
57 c3 = c2;
58 }
59 float* c4 = (float*) ((uintptr_t) c3 + cm_stride);
60 if XNN_UNPREDICTABLE(mr <= 4) {
61 c4 = c3;
62 }
63 float* c5 = (float*) ((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 const float32x4_t voutput_min = vld1q_dup_f32(&params->scalar.min);
379 vout0x0123 = vmaxq_f32(vout0x0123, voutput_min);
380 vout0x4567 = vmaxq_f32(vout0x4567, voutput_min);
381 vout0x89AB = vmaxq_f32(vout0x89AB, voutput_min);
382 vout0xCDEF = vmaxq_f32(vout0xCDEF, voutput_min);
383 vout1x0123 = vmaxq_f32(vout1x0123, voutput_min);
384 vout1x4567 = vmaxq_f32(vout1x4567, voutput_min);
385 vout1x89AB = vmaxq_f32(vout1x89AB, voutput_min);
386 vout1xCDEF = vmaxq_f32(vout1xCDEF, voutput_min);
387 vout2x0123 = vmaxq_f32(vout2x0123, voutput_min);
388 vout2x4567 = vmaxq_f32(vout2x4567, voutput_min);
389 vout2x89AB = vmaxq_f32(vout2x89AB, voutput_min);
390 vout2xCDEF = vmaxq_f32(vout2xCDEF, voutput_min);
391 vout3x0123 = vmaxq_f32(vout3x0123, voutput_min);
392 vout3x4567 = vmaxq_f32(vout3x4567, voutput_min);
393 vout3x89AB = vmaxq_f32(vout3x89AB, voutput_min);
394 vout3xCDEF = vmaxq_f32(vout3xCDEF, voutput_min);
395 vout4x0123 = vmaxq_f32(vout4x0123, voutput_min);
396 vout4x4567 = vmaxq_f32(vout4x4567, voutput_min);
397 vout4x89AB = vmaxq_f32(vout4x89AB, voutput_min);
398 vout4xCDEF = vmaxq_f32(vout4xCDEF, voutput_min);
399 vout5x0123 = vmaxq_f32(vout5x0123, voutput_min);
400 vout5x4567 = vmaxq_f32(vout5x4567, voutput_min);
401 vout5x89AB = vmaxq_f32(vout5x89AB, voutput_min);
402 vout5xCDEF = vmaxq_f32(vout5xCDEF, voutput_min);
403
404 const float32x4_t voutput_max = vld1q_dup_f32(&params->scalar.max);
405 vout0x0123 = vminq_f32(vout0x0123, voutput_max);
406 vout0x4567 = vminq_f32(vout0x4567, voutput_max);
407 vout0x89AB = vminq_f32(vout0x89AB, voutput_max);
408 vout0xCDEF = vminq_f32(vout0xCDEF, voutput_max);
409 vout1x0123 = vminq_f32(vout1x0123, voutput_max);
410 vout1x4567 = vminq_f32(vout1x4567, voutput_max);
411 vout1x89AB = vminq_f32(vout1x89AB, voutput_max);
412 vout1xCDEF = vminq_f32(vout1xCDEF, voutput_max);
413 vout2x0123 = vminq_f32(vout2x0123, voutput_max);
414 vout2x4567 = vminq_f32(vout2x4567, voutput_max);
415 vout2x89AB = vminq_f32(vout2x89AB, voutput_max);
416 vout2xCDEF = vminq_f32(vout2xCDEF, voutput_max);
417 vout3x0123 = vminq_f32(vout3x0123, voutput_max);
418 vout3x4567 = vminq_f32(vout3x4567, voutput_max);
419 vout3x89AB = vminq_f32(vout3x89AB, voutput_max);
420 vout3xCDEF = vminq_f32(vout3xCDEF, voutput_max);
421 vout4x0123 = vminq_f32(vout4x0123, voutput_max);
422 vout4x4567 = vminq_f32(vout4x4567, voutput_max);
423 vout4x89AB = vminq_f32(vout4x89AB, voutput_max);
424 vout4xCDEF = vminq_f32(vout4xCDEF, voutput_max);
425 vout5x0123 = vminq_f32(vout5x0123, voutput_max);
426 vout5x4567 = vminq_f32(vout5x4567, voutput_max);
427 vout5x89AB = vminq_f32(vout5x89AB, voutput_max);
428 vout5xCDEF = vminq_f32(vout5xCDEF, voutput_max);
429
430 if XNN_LIKELY(nc >= 16) {
431 vst1q_f32(c5, vout5x0123);
432 vst1q_f32(c5 + 4, vout5x4567);
433 vst1q_f32(c5 + 8, vout5x89AB);
434 vst1q_f32(c5 + 12, vout5xCDEF);
435 vst1q_f32(c4, vout4x0123);
436 vst1q_f32(c4 + 4, vout4x4567);
437 vst1q_f32(c4 + 8, vout4x89AB);
438 vst1q_f32(c4 + 12, vout4xCDEF);
439 vst1q_f32(c3, vout3x0123);
440 vst1q_f32(c3 + 4, vout3x4567);
441 vst1q_f32(c3 + 8, vout3x89AB);
442 vst1q_f32(c3 + 12, vout3xCDEF);
443 vst1q_f32(c2, vout2x0123);
444 vst1q_f32(c2 + 4, vout2x4567);
445 vst1q_f32(c2 + 8, vout2x89AB);
446 vst1q_f32(c2 + 12, vout2xCDEF);
447 vst1q_f32(c1, vout1x0123);
448 vst1q_f32(c1 + 4, vout1x4567);
449 vst1q_f32(c1 + 8, vout1x89AB);
450 vst1q_f32(c1 + 12, vout1xCDEF);
451 vst1q_f32(c0, vout0x0123);
452 vst1q_f32(c0 + 4, vout0x4567);
453 vst1q_f32(c0 + 8, vout0x89AB);
454 vst1q_f32(c0 + 12, vout0xCDEF);
455
456 a = (const int8_t**restrict) ((uintptr_t) a - ks);
457
458 c0 = (float*) ((uintptr_t) c0 + cn_stride);
459 c1 = (float*) ((uintptr_t) c1 + cn_stride);
460 c2 = (float*) ((uintptr_t) c2 + cn_stride);
461 c3 = (float*) ((uintptr_t) c3 + cn_stride);
462 c4 = (float*) ((uintptr_t) c4 + cn_stride);
463 c5 = (float*) ((uintptr_t) c5 + cn_stride);
464
465 nc -= 16;
466 } else {
467 if (nc & 8) {
468 vst1q_f32(c5, vout5x0123); c5 += 4;
469 vout5x0123 = vout5x89AB;
470 vst1q_f32(c4, vout4x0123); c4 += 4;
471 vout4x0123 = vout4x89AB;
472 vst1q_f32(c3, vout3x0123); c3 += 4;
473 vout3x0123 = vout3x89AB;
474 vst1q_f32(c2, vout2x0123); c2 += 4;
475 vout2x0123 = vout2x89AB;
476 vst1q_f32(c1, vout1x0123); c1 += 4;
477 vout1x0123 = vout1x89AB;
478 vst1q_f32(c0, vout0x0123); c0 += 4;
479 vout0x0123 = vout0x89AB;
480 vst1q_f32(c5, vout5x4567); c5 += 4;
481 vout5x4567 = vout5xCDEF;
482 vst1q_f32(c4, vout4x4567); c4 += 4;
483 vout4x4567 = vout4xCDEF;
484 vst1q_f32(c3, vout3x4567); c3 += 4;
485 vout3x4567 = vout3xCDEF;
486 vst1q_f32(c2, vout2x4567); c2 += 4;
487 vout2x4567 = vout2xCDEF;
488 vst1q_f32(c1, vout1x4567); c1 += 4;
489 vout1x4567 = vout1xCDEF;
490 vst1q_f32(c0, vout0x4567); c0 += 4;
491 vout0x4567 = vout0xCDEF;
492 }
493 if (nc & 4) {
494 vst1q_f32(c5, vout5x0123); c5 += 4;
495 vout5x0123 = vout5x4567;
496 vst1q_f32(c4, vout4x0123); c4 += 4;
497 vout4x0123 = vout4x4567;
498 vst1q_f32(c3, vout3x0123); c3 += 4;
499 vout3x0123 = vout3x4567;
500 vst1q_f32(c2, vout2x0123); c2 += 4;
501 vout2x0123 = vout2x4567;
502 vst1q_f32(c1, vout1x0123); c1 += 4;
503 vout1x0123 = vout1x4567;
504 vst1q_f32(c0, vout0x0123); c0 += 4;
505 vout0x0123 = vout0x4567;
506 }
507 float32x2_t vout5x01 = vget_low_f32(vout5x0123);
508 float32x2_t vout4x01 = vget_low_f32(vout4x0123);
509 float32x2_t vout3x01 = vget_low_f32(vout3x0123);
510 float32x2_t vout2x01 = vget_low_f32(vout2x0123);
511 float32x2_t vout1x01 = vget_low_f32(vout1x0123);
512 float32x2_t vout0x01 = vget_low_f32(vout0x0123);
513 if (nc & 2) {
514 vst1_f32(c5, vout5x01); c5 += 2;
515 vst1_f32(c4, vout4x01); c4 += 2;
516 vst1_f32(c3, vout3x01); c3 += 2;
517 vst1_f32(c2, vout2x01); c2 += 2;
518 vst1_f32(c1, vout1x01); c1 += 2;
519 vst1_f32(c0, vout0x01); c0 += 2;
520 vout5x01 = vget_high_f32(vout5x0123);
521 vout4x01 = vget_high_f32(vout4x0123);
522 vout3x01 = vget_high_f32(vout3x0123);
523 vout2x01 = vget_high_f32(vout2x0123);
524 vout1x01 = vget_high_f32(vout1x0123);
525 vout0x01 = vget_high_f32(vout0x0123);
526 }
527 if (nc & 1) {
528 vst1_lane_f32(c5, vout5x01, 0);
529 vst1_lane_f32(c4, vout4x01, 0);
530 vst1_lane_f32(c3, vout3x01, 0);
531 vst1_lane_f32(c2, vout2x01, 0);
532 vst1_lane_f32(c1, vout1x01, 0);
533 vst1_lane_f32(c0, vout0x01, 0);
534 }
535 nc = 0;
536 }
537 } while (nc != 0);
538 }