1 // Auto-generated file. Do not edit!
2 // Template: src/qs8-igemm/c8-neon-mull.c.in
3 // Generator: tools/xngen
4 //
5 // Copyright 2021 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
xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c8__neon_mull(size_t mr,size_t nc,size_t kc,size_t ks,const int8_t ** restrict a,const void * restrict w,int8_t * restrict c,size_t cm_stride,size_t cn_stride,size_t a_offset,const int8_t * zero,const union xnn_qs8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS (1)])18 void xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c8__neon_mull(
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 int8_t* restrict c,
26 size_t cm_stride,
27 size_t cn_stride,
28 size_t a_offset,
29 const int8_t* zero,
30 const union xnn_qs8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
31 {
32 assert(mr != 0);
33 assert(mr <= 3);
34 assert(nc != 0);
35 assert(kc != 0);
36 assert(ks != 0);
37 assert(ks % (3 * sizeof(void*)) == 0);
38 assert(a_offset % sizeof(int8_t) == 0);
39 assert(a != NULL);
40 assert(w != NULL);
41 assert(c != NULL);
42
43 kc = round_up_po2(kc, 8 * sizeof(int8_t));
44 int8_t* c0 = c;
45 int8_t* c1 = (int8_t*) ((uintptr_t) c0 + cm_stride);
46 if XNN_UNPREDICTABLE(mr < 2) {
47 c1 = c0;
48 }
49 int8_t* c2 = (int8_t*) ((uintptr_t) c1 + cm_stride);
50 if XNN_UNPREDICTABLE(mr <= 2) {
51 c2 = c1;
52 }
53
54 do {
55 int32x4_t vacc0x0 = vld1q_lane_s32(w, vmovq_n_s32(0), 0); w = (const void*) ((uintptr_t) w + sizeof(int32_t));
56 int32x4_t vacc0x1 = vld1q_lane_s32(w, vmovq_n_s32(0), 0); w = (const void*) ((uintptr_t) w + sizeof(int32_t));
57 int32x4_t vacc0x2 = vld1q_lane_s32(w, vmovq_n_s32(0), 0); w = (const void*) ((uintptr_t) w + sizeof(int32_t));
58 int32x4_t vacc0x3 = vld1q_lane_s32(w, vmovq_n_s32(0), 0); w = (const void*) ((uintptr_t) w + sizeof(int32_t));
59 int32x4_t vacc0x4 = vld1q_lane_s32(w, vmovq_n_s32(0), 0); w = (const void*) ((uintptr_t) w + sizeof(int32_t));
60 int32x4_t vacc0x5 = vld1q_lane_s32(w, vmovq_n_s32(0), 0); w = (const void*) ((uintptr_t) w + sizeof(int32_t));
61 int32x4_t vacc0x6 = vld1q_lane_s32(w, vmovq_n_s32(0), 0); w = (const void*) ((uintptr_t) w + sizeof(int32_t));
62 int32x4_t vacc0x7 = vld1q_lane_s32(w, vmovq_n_s32(0), 0); w = (const void*) ((uintptr_t) w + sizeof(int32_t));
63 int32x4_t vacc1x0 = vacc0x0;
64 int32x4_t vacc1x1 = vacc0x1;
65 int32x4_t vacc1x2 = vacc0x2;
66 int32x4_t vacc1x3 = vacc0x3;
67 int32x4_t vacc1x4 = vacc0x4;
68 int32x4_t vacc1x5 = vacc0x5;
69 int32x4_t vacc1x6 = vacc0x6;
70 int32x4_t vacc1x7 = vacc0x7;
71 int32x4_t vacc2x0 = vacc0x0;
72 int32x4_t vacc2x1 = vacc0x1;
73 int32x4_t vacc2x2 = vacc0x2;
74 int32x4_t vacc2x3 = vacc0x3;
75 int32x4_t vacc2x4 = vacc0x4;
76 int32x4_t vacc2x5 = vacc0x5;
77 int32x4_t vacc2x6 = vacc0x6;
78 int32x4_t vacc2x7 = vacc0x7;
79
80 size_t p = ks;
81 do {
82 const int8_t* restrict a0 = a[0];
83 if XNN_UNPREDICTABLE(a0 != zero) {
84 a0 = (const int8_t*) ((uintptr_t) a0 + a_offset);
85 }
86 const int8_t* restrict a1 = a[1];
87 if XNN_UNPREDICTABLE(a1 != zero) {
88 a1 = (const int8_t*) ((uintptr_t) a1 + a_offset);
89 }
90 const int8_t* restrict a2 = a[2];
91 if XNN_UNPREDICTABLE(a2 != zero) {
92 a2 = (const int8_t*) ((uintptr_t) a2 + a_offset);
93 }
94 a += 3;
95
96 size_t k = kc;
97
98 // Handle 8 bytes at a time using MUL.
99 while (k != 0) {
100 const int8x8_t va0 = vld1_s8(a0); a0 += 8;
101 const int8x8_t va1 = vld1_s8(a1); a1 += 8;
102 const int8x8_t va2 = vld1_s8(a2); a2 += 8;
103
104 const int8x8_t vb0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
105 const int16x8_t vprod0x0 = vmull_s8(vb0, va0);
106 const int16x8_t vprod1x0 = vmull_s8(vb0, va1);
107 const int16x8_t vprod2x0 = vmull_s8(vb0, va2);
108 vacc0x0 = vpadalq_s16(vacc0x0, vprod0x0);
109 vacc1x0 = vpadalq_s16(vacc1x0, vprod1x0);
110 vacc2x0 = vpadalq_s16(vacc2x0, vprod2x0);
111 const int8x8_t vb1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
112 const int16x8_t vprod0x1 = vmull_s8(vb1, va0);
113 const int16x8_t vprod1x1 = vmull_s8(vb1, va1);
114 const int16x8_t vprod2x1 = vmull_s8(vb1, va2);
115 vacc0x1 = vpadalq_s16(vacc0x1, vprod0x1);
116 vacc1x1 = vpadalq_s16(vacc1x1, vprod1x1);
117 vacc2x1 = vpadalq_s16(vacc2x1, vprod2x1);
118 const int8x8_t vb2 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
119 const int16x8_t vprod0x2 = vmull_s8(vb2, va0);
120 const int16x8_t vprod1x2 = vmull_s8(vb2, va1);
121 const int16x8_t vprod2x2 = vmull_s8(vb2, va2);
122 vacc0x2 = vpadalq_s16(vacc0x2, vprod0x2);
123 vacc1x2 = vpadalq_s16(vacc1x2, vprod1x2);
124 vacc2x2 = vpadalq_s16(vacc2x2, vprod2x2);
125 const int8x8_t vb3 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
126 const int16x8_t vprod0x3 = vmull_s8(vb3, va0);
127 const int16x8_t vprod1x3 = vmull_s8(vb3, va1);
128 const int16x8_t vprod2x3 = vmull_s8(vb3, va2);
129 vacc0x3 = vpadalq_s16(vacc0x3, vprod0x3);
130 vacc1x3 = vpadalq_s16(vacc1x3, vprod1x3);
131 vacc2x3 = vpadalq_s16(vacc2x3, vprod2x3);
132 const int8x8_t vb4 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
133 const int16x8_t vprod0x4 = vmull_s8(vb4, va0);
134 const int16x8_t vprod1x4 = vmull_s8(vb4, va1);
135 const int16x8_t vprod2x4 = vmull_s8(vb4, va2);
136 vacc0x4 = vpadalq_s16(vacc0x4, vprod0x4);
137 vacc1x4 = vpadalq_s16(vacc1x4, vprod1x4);
138 vacc2x4 = vpadalq_s16(vacc2x4, vprod2x4);
139 const int8x8_t vb5 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
140 const int16x8_t vprod0x5 = vmull_s8(vb5, va0);
141 const int16x8_t vprod1x5 = vmull_s8(vb5, va1);
142 const int16x8_t vprod2x5 = vmull_s8(vb5, va2);
143 vacc0x5 = vpadalq_s16(vacc0x5, vprod0x5);
144 vacc1x5 = vpadalq_s16(vacc1x5, vprod1x5);
145 vacc2x5 = vpadalq_s16(vacc2x5, vprod2x5);
146 const int8x8_t vb6 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
147 const int16x8_t vprod0x6 = vmull_s8(vb6, va0);
148 const int16x8_t vprod1x6 = vmull_s8(vb6, va1);
149 const int16x8_t vprod2x6 = vmull_s8(vb6, va2);
150 vacc0x6 = vpadalq_s16(vacc0x6, vprod0x6);
151 vacc1x6 = vpadalq_s16(vacc1x6, vprod1x6);
152 vacc2x6 = vpadalq_s16(vacc2x6, vprod2x6);
153 const int8x8_t vb7 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
154 const int16x8_t vprod0x7 = vmull_s8(vb7, va0);
155 const int16x8_t vprod1x7 = vmull_s8(vb7, va1);
156 const int16x8_t vprod2x7 = vmull_s8(vb7, va2);
157 vacc0x7 = vpadalq_s16(vacc0x7, vprod0x7);
158 vacc1x7 = vpadalq_s16(vacc1x7, vprod1x7);
159 vacc2x7 = vpadalq_s16(vacc2x7, vprod2x7);
160
161 k -= 8 * sizeof(int8_t);
162 }
163
164 p -= 3 * sizeof(void*);
165 } while (p != 0);
166
167 #if XNN_ARCH_ARM64
168 const int32x4_t vsum0x01 = vpaddq_s32(vacc0x0, vacc0x1);
169 const int32x4_t vsum0x23 = vpaddq_s32(vacc0x2, vacc0x3);
170 const int32x4_t vsum0x45 = vpaddq_s32(vacc0x4, vacc0x5);
171 const int32x4_t vsum0x67 = vpaddq_s32(vacc0x6, vacc0x7);
172 const int32x4_t vsum1x01 = vpaddq_s32(vacc1x0, vacc1x1);
173 const int32x4_t vsum1x23 = vpaddq_s32(vacc1x2, vacc1x3);
174 const int32x4_t vsum1x45 = vpaddq_s32(vacc1x4, vacc1x5);
175 const int32x4_t vsum1x67 = vpaddq_s32(vacc1x6, vacc1x7);
176 const int32x4_t vsum2x01 = vpaddq_s32(vacc2x0, vacc2x1);
177 const int32x4_t vsum2x23 = vpaddq_s32(vacc2x2, vacc2x3);
178 const int32x4_t vsum2x45 = vpaddq_s32(vacc2x4, vacc2x5);
179 const int32x4_t vsum2x67 = vpaddq_s32(vacc2x6, vacc2x7);
180
181 int32x4_t vacc0x0123 = vpaddq_s32(vsum0x01, vsum0x23);
182 int32x4_t vacc0x4567 = vpaddq_s32(vsum0x45, vsum0x67);
183 int32x4_t vacc1x0123 = vpaddq_s32(vsum1x01, vsum1x23);
184 int32x4_t vacc1x4567 = vpaddq_s32(vsum1x45, vsum1x67);
185 int32x4_t vacc2x0123 = vpaddq_s32(vsum2x01, vsum2x23);
186 int32x4_t vacc2x4567 = vpaddq_s32(vsum2x45, vsum2x67);
187 #else
188 const int32x2_t vpsum0x0 = vadd_s32(vget_low_s32(vacc0x0), vget_high_s32(vacc0x0));
189 const int32x2_t vpsum0x1 = vadd_s32(vget_low_s32(vacc0x1), vget_high_s32(vacc0x1));
190 const int32x2_t vpsum0x2 = vadd_s32(vget_low_s32(vacc0x2), vget_high_s32(vacc0x2));
191 const int32x2_t vpsum0x3 = vadd_s32(vget_low_s32(vacc0x3), vget_high_s32(vacc0x3));
192 const int32x2_t vsum0x01 = vpadd_s32(vpsum0x0, vpsum0x1);
193 const int32x2_t vsum0x23 = vpadd_s32(vpsum0x2, vpsum0x3);
194 int32x4_t vacc0x0123 = vcombine_s32(vsum0x01, vsum0x23 );
195 const int32x2_t vpsum0x4 = vadd_s32(vget_low_s32(vacc0x4), vget_high_s32(vacc0x4));
196 const int32x2_t vpsum0x5 = vadd_s32(vget_low_s32(vacc0x5), vget_high_s32(vacc0x5));
197 const int32x2_t vpsum0x6 = vadd_s32(vget_low_s32(vacc0x6), vget_high_s32(vacc0x6));
198 const int32x2_t vpsum0x7 = vadd_s32(vget_low_s32(vacc0x7), vget_high_s32(vacc0x7));
199 const int32x2_t vsum0x45 = vpadd_s32(vpsum0x4, vpsum0x5);
200 const int32x2_t vsum0x67 = vpadd_s32(vpsum0x6, vpsum0x7);
201 int32x4_t vacc0x4567 = vcombine_s32(vsum0x45, vsum0x67 );
202 const int32x2_t vpsum1x0 = vadd_s32(vget_low_s32(vacc1x0), vget_high_s32(vacc1x0));
203 const int32x2_t vpsum1x1 = vadd_s32(vget_low_s32(vacc1x1), vget_high_s32(vacc1x1));
204 const int32x2_t vpsum1x2 = vadd_s32(vget_low_s32(vacc1x2), vget_high_s32(vacc1x2));
205 const int32x2_t vpsum1x3 = vadd_s32(vget_low_s32(vacc1x3), vget_high_s32(vacc1x3));
206 const int32x2_t vsum1x01 = vpadd_s32(vpsum1x0, vpsum1x1);
207 const int32x2_t vsum1x23 = vpadd_s32(vpsum1x2, vpsum1x3);
208 int32x4_t vacc1x0123 = vcombine_s32(vsum1x01, vsum1x23 );
209 const int32x2_t vpsum1x4 = vadd_s32(vget_low_s32(vacc1x4), vget_high_s32(vacc1x4));
210 const int32x2_t vpsum1x5 = vadd_s32(vget_low_s32(vacc1x5), vget_high_s32(vacc1x5));
211 const int32x2_t vpsum1x6 = vadd_s32(vget_low_s32(vacc1x6), vget_high_s32(vacc1x6));
212 const int32x2_t vpsum1x7 = vadd_s32(vget_low_s32(vacc1x7), vget_high_s32(vacc1x7));
213 const int32x2_t vsum1x45 = vpadd_s32(vpsum1x4, vpsum1x5);
214 const int32x2_t vsum1x67 = vpadd_s32(vpsum1x6, vpsum1x7);
215 int32x4_t vacc1x4567 = vcombine_s32(vsum1x45, vsum1x67 );
216 const int32x2_t vpsum2x0 = vadd_s32(vget_low_s32(vacc2x0), vget_high_s32(vacc2x0));
217 const int32x2_t vpsum2x1 = vadd_s32(vget_low_s32(vacc2x1), vget_high_s32(vacc2x1));
218 const int32x2_t vpsum2x2 = vadd_s32(vget_low_s32(vacc2x2), vget_high_s32(vacc2x2));
219 const int32x2_t vpsum2x3 = vadd_s32(vget_low_s32(vacc2x3), vget_high_s32(vacc2x3));
220 const int32x2_t vsum2x01 = vpadd_s32(vpsum2x0, vpsum2x1);
221 const int32x2_t vsum2x23 = vpadd_s32(vpsum2x2, vpsum2x3);
222 int32x4_t vacc2x0123 = vcombine_s32(vsum2x01, vsum2x23 );
223 const int32x2_t vpsum2x4 = vadd_s32(vget_low_s32(vacc2x4), vget_high_s32(vacc2x4));
224 const int32x2_t vpsum2x5 = vadd_s32(vget_low_s32(vacc2x5), vget_high_s32(vacc2x5));
225 const int32x2_t vpsum2x6 = vadd_s32(vget_low_s32(vacc2x6), vget_high_s32(vacc2x6));
226 const int32x2_t vpsum2x7 = vadd_s32(vget_low_s32(vacc2x7), vget_high_s32(vacc2x7));
227 const int32x2_t vsum2x45 = vpadd_s32(vpsum2x4, vpsum2x5);
228 const int32x2_t vsum2x67 = vpadd_s32(vpsum2x6, vpsum2x7);
229 int32x4_t vacc2x4567 = vcombine_s32(vsum2x45, vsum2x67 );
230 #endif
231
232 const int32x4_t vright_pre_shift = vld1q_dup_s32(¶ms->rndnu_neon.right_pre_shift);
233 const int32x4_t vmultiplier = vld1q_dup_s32(¶ms->rndnu_neon.multiplier);
234 const int32x4_t vright_post_shift = vld1q_dup_s32(¶ms->rndnu_neon.right_post_shift);
235
236 vacc0x0123 = vqshlq_s32(vacc0x0123, vright_pre_shift);
237 vacc0x4567 = vqshlq_s32(vacc0x4567, vright_pre_shift);
238 vacc1x0123 = vqshlq_s32(vacc1x0123, vright_pre_shift);
239 vacc1x4567 = vqshlq_s32(vacc1x4567, vright_pre_shift);
240 vacc2x0123 = vqshlq_s32(vacc2x0123, vright_pre_shift);
241 vacc2x4567 = vqshlq_s32(vacc2x4567, vright_pre_shift);
242
243 vacc0x0123 = vqdmulhq_s32(vacc0x0123, vmultiplier);
244 vacc0x4567 = vqdmulhq_s32(vacc0x4567, vmultiplier);
245 vacc1x0123 = vqdmulhq_s32(vacc1x0123, vmultiplier);
246 vacc1x4567 = vqdmulhq_s32(vacc1x4567, vmultiplier);
247 vacc2x0123 = vqdmulhq_s32(vacc2x0123, vmultiplier);
248 vacc2x4567 = vqdmulhq_s32(vacc2x4567, vmultiplier);
249
250 vacc0x0123 = vrshlq_s32(vacc0x0123, vright_post_shift);
251 vacc0x4567 = vrshlq_s32(vacc0x4567, vright_post_shift);
252 vacc1x0123 = vrshlq_s32(vacc1x0123, vright_post_shift);
253 vacc1x4567 = vrshlq_s32(vacc1x4567, vright_post_shift);
254 vacc2x0123 = vrshlq_s32(vacc2x0123, vright_post_shift);
255 vacc2x4567 = vrshlq_s32(vacc2x4567, vright_post_shift);
256
257 const int16x8_t voutput_zero_point = vld1q_dup_s16(¶ms->rndnu_neon.output_zero_point);
258 #if XNN_ARCH_ARM64
259 int16x8_t vacc0x01234567 = vqmovn_high_s32(vqmovn_s32(vacc0x0123), vacc0x4567);
260 int16x8_t vacc1x01234567 = vqmovn_high_s32(vqmovn_s32(vacc1x0123), vacc1x4567);
261 int16x8_t vacc2x01234567 = vqmovn_high_s32(vqmovn_s32(vacc2x0123), vacc2x4567);
262
263 vacc0x01234567 = vqaddq_s16(vacc0x01234567, voutput_zero_point);
264 vacc1x01234567 = vqaddq_s16(vacc1x01234567, voutput_zero_point);
265 vacc2x01234567 = vqaddq_s16(vacc2x01234567, voutput_zero_point);
266
267 int8x16_t vout0x01234567_1x01234567 = vqmovn_high_s16(vqmovn_s16(vacc0x01234567), vacc1x01234567);
268 int8x8_t vout2x01234567 = vqmovn_s16(vacc2x01234567);
269 #else
270 int16x8_t vacc0x01234567 = vcombine_s16(vqmovn_s32(vacc0x0123), vqmovn_s32(vacc0x4567));
271 int16x8_t vacc1x01234567 = vcombine_s16(vqmovn_s32(vacc1x0123), vqmovn_s32(vacc1x4567));
272 int16x8_t vacc2x01234567 = vcombine_s16(vqmovn_s32(vacc2x0123), vqmovn_s32(vacc2x4567));
273
274 vacc0x01234567 = vqaddq_s16(vacc0x01234567, voutput_zero_point);
275 vacc1x01234567 = vqaddq_s16(vacc1x01234567, voutput_zero_point);
276 vacc2x01234567 = vqaddq_s16(vacc2x01234567, voutput_zero_point);
277
278 int8x16_t vout0x01234567_1x01234567 = vcombine_s8(vqmovn_s16(vacc0x01234567), vqmovn_s16(vacc1x01234567));
279 int8x8_t vout2x01234567 = vqmovn_s16(vacc2x01234567);
280 #endif
281
282 const int8x16_t voutput_min = vld1q_dup_s8(¶ms->rndnu_neon.output_min);
283 vout0x01234567_1x01234567 = vmaxq_s8(vout0x01234567_1x01234567, voutput_min);
284 vout2x01234567 = vmax_s8(vout2x01234567, vget_low_s8(voutput_min));
285
286 const int8x16_t voutput_max = vld1q_dup_s8(¶ms->rndnu_neon.output_max);
287 vout0x01234567_1x01234567 = vminq_s8(vout0x01234567_1x01234567, voutput_max);
288 vout2x01234567 = vmin_s8(vout2x01234567, vget_low_s8(voutput_max));
289
290 if (nc >= 8) {
291 vst1_s8(c2 + 0, vout2x01234567);
292 vst1_s8(c1 + 0, vget_high_s8(vout0x01234567_1x01234567));
293 vst1_s8(c0 + 0, vget_low_s8(vout0x01234567_1x01234567));
294
295 c2 = (int8_t*) ((uintptr_t) c2 + cn_stride);
296 c1 = (int8_t*) ((uintptr_t) c1 + cn_stride);
297 c0 = (int8_t*) ((uintptr_t) c0 + cn_stride);
298
299 a = (const int8_t**restrict) ((uintptr_t) a - ks);
300
301 nc -= 8;
302 } else {
303 if (nc & 4) {
304 vst1_lane_u32((void*) c2, vreinterpret_u32_s8(vout2x01234567), 0); c2 += 4;
305 vst1q_lane_u32((void*) c1, vreinterpretq_u32_s8(vout0x01234567_1x01234567), 2); c1 += 4;
306 vst1q_lane_u32((void*) c0, vreinterpretq_u32_s8(vout0x01234567_1x01234567), 0); c0 += 4;
307 vout2x01234567 = vext_s8(vout2x01234567, vout2x01234567, 4);
308 vout0x01234567_1x01234567 = vextq_s8(vout0x01234567_1x01234567, vout0x01234567_1x01234567, 4);
309 }
310 if (nc & 2) {
311 vst1_lane_u16((void*) c2, vreinterpret_u16_s8(vout2x01234567), 0); c2 += 2;
312 vst1q_lane_u16((void*) c1, vreinterpretq_u16_s8(vout0x01234567_1x01234567), 4); c1 += 2;
313 vst1q_lane_u16((void*) c0, vreinterpretq_u16_s8(vout0x01234567_1x01234567), 0); c0 += 2;
314 vout2x01234567 = vext_s8(vout2x01234567, vout2x01234567, 2);
315 vout0x01234567_1x01234567 = vextq_s8(vout0x01234567_1x01234567, vout0x01234567_1x01234567, 2);
316 }
317 if (nc & 1) {
318 vst1_lane_s8(c2, vout2x01234567, 0);
319 vst1q_lane_s8(c1, vout0x01234567_1x01234567, 8);
320 vst1q_lane_s8(c0, vout0x01234567_1x01234567, 0);
321 }
322
323 nc = 0;
324 }
325 } while (nc != 0);
326 }
327