1 // Auto-generated file. Do not edit!
2 // Template: src/qs8-igemm/c2-neon-mull-dup.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/gemm.h>
15 #include <xnnpack/math.h>
16
xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c2__neon_mull_ld2r(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)])17 void xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c2__neon_mull_ld2r(
18 size_t mr,
19 size_t nc,
20 size_t kc,
21 size_t ks,
22 const int8_t** restrict a,
23 const void* restrict w,
24 int8_t* restrict c,
25 size_t cm_stride,
26 size_t cn_stride,
27 size_t a_offset,
28 const int8_t* zero,
29 const union xnn_qs8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
30 {
31 assert(mr != 0);
32 assert(mr <= 3);
33 assert(nc != 0);
34 assert(kc != 0);
35 assert(ks != 0);
36 assert(ks % (3 * sizeof(void*)) == 0);
37 assert(a_offset % sizeof(int8_t) == 0);
38 assert(a != NULL);
39 assert(w != NULL);
40 assert(c != NULL);
41
42 kc = round_up_po2(kc, 2 * sizeof(int8_t));
43 int8_t* c0 = c;
44 int8_t* c1 = (int8_t*) ((uintptr_t) c0 + cm_stride);
45 if XNN_UNPREDICTABLE(mr < 2) {
46 c1 = c0;
47 }
48 int8_t* c2 = (int8_t*) ((uintptr_t) c1 + cm_stride);
49 if XNN_UNPREDICTABLE(mr <= 2) {
50 c2 = c1;
51 }
52
53 do {
54 int32x4_t vacc0x0123 = vld1q_s32(w); w = (const void*) ((uintptr_t) w + 4 * sizeof(int32_t));
55 int32x4_t vacc0x4567 = vld1q_s32(w); w = (const void*) ((uintptr_t) w + 4 * sizeof(int32_t));
56 int32x4_t vacc1x0123 = vacc0x0123;
57 int32x4_t vacc1x4567 = vacc0x4567;
58 int32x4_t vacc2x0123 = vacc0x0123;
59 int32x4_t vacc2x4567 = vacc0x4567;
60
61 size_t p = ks;
62 do {
63 const int8_t* restrict a0 = a[0];
64 if XNN_UNPREDICTABLE(a0 != zero) {
65 a0 = (const int8_t*) ((uintptr_t) a0 + a_offset);
66 }
67 const int8_t* restrict a1 = a[1];
68 if XNN_UNPREDICTABLE(a1 != zero) {
69 a1 = (const int8_t*) ((uintptr_t) a1 + a_offset);
70 }
71 const int8_t* restrict a2 = a[2];
72 if XNN_UNPREDICTABLE(a2 != zero) {
73 a2 = (const int8_t*) ((uintptr_t) a2 + a_offset);
74 }
75 a += 3;
76
77 size_t k = kc;
78
79
80 while (k >= 8 * sizeof(int8_t)) {
81 const int16x4x2_t va00 = vld2_dup_s16((const void*)a0);
82 const int16x4x2_t va01 = vld2_dup_s16((const void*)(a0 + 4)); a0 += 8;
83 const int16x4x2_t va10 = vld2_dup_s16((const void*)a1);
84 const int16x4x2_t va11 = vld2_dup_s16((const void*)(a1 + 4)); a1 += 8;
85 const int16x4x2_t va20 = vld2_dup_s16((const void*)a2);
86 const int16x4x2_t va21 = vld2_dup_s16((const void*)(a2 + 4)); a2 += 8;
87
88 const int8x8_t vb0123c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
89 const int8x8_t vb4567c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
90 const int8x8_t vb0123c1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
91 const int8x8_t vb4567c1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
92 const int8x8_t vb0123c2 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
93 const int8x8_t vb4567c2 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
94 const int8x8_t vb0123c3 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
95 const int8x8_t vb4567c3 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
96
97 const int8x8_t va0c0 = vreinterpret_s8_s16(va00.val[0]);
98 const int8x8_t va1c0 = vreinterpret_s8_s16(va10.val[0]);
99 const int8x8_t va2c0 = vreinterpret_s8_s16(va20.val[0]);
100
101 const int16x8_t vprod0x0123c0 = vmull_s8(vb0123c0, va0c0);
102 const int16x8_t vprod1x0123c0 = vmull_s8(vb0123c0, va1c0);
103 const int16x8_t vprod2x0123c0 = vmull_s8(vb0123c0, va2c0);
104 vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c0);
105 vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c0);
106 vacc2x0123 = vpadalq_s16(vacc2x0123, vprod2x0123c0);
107 const int16x8_t vprod0x4567c0 = vmull_s8(vb4567c0, va0c0);
108 const int16x8_t vprod1x4567c0 = vmull_s8(vb4567c0, va1c0);
109 const int16x8_t vprod2x4567c0 = vmull_s8(vb4567c0, va2c0);
110 vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c0);
111 vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c0);
112 vacc2x4567 = vpadalq_s16(vacc2x4567, vprod2x4567c0);
113 const int8x8_t va0c1 = vreinterpret_s8_s16(va00.val[1]);
114 const int8x8_t va1c1 = vreinterpret_s8_s16(va10.val[1]);
115 const int8x8_t va2c1 = vreinterpret_s8_s16(va20.val[1]);
116
117 const int16x8_t vprod0x0123c1 = vmull_s8(vb0123c1, va0c1);
118 const int16x8_t vprod1x0123c1 = vmull_s8(vb0123c1, va1c1);
119 const int16x8_t vprod2x0123c1 = vmull_s8(vb0123c1, va2c1);
120 vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c1);
121 vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c1);
122 vacc2x0123 = vpadalq_s16(vacc2x0123, vprod2x0123c1);
123 const int16x8_t vprod0x4567c1 = vmull_s8(vb4567c1, va0c1);
124 const int16x8_t vprod1x4567c1 = vmull_s8(vb4567c1, va1c1);
125 const int16x8_t vprod2x4567c1 = vmull_s8(vb4567c1, va2c1);
126 vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c1);
127 vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c1);
128 vacc2x4567 = vpadalq_s16(vacc2x4567, vprod2x4567c1);
129 const int8x8_t va0c2 = vreinterpret_s8_s16(va01.val[0]);
130 const int8x8_t va1c2 = vreinterpret_s8_s16(va11.val[0]);
131 const int8x8_t va2c2 = vreinterpret_s8_s16(va21.val[0]);
132
133 const int16x8_t vprod0x0123c2 = vmull_s8(vb0123c2, va0c2);
134 const int16x8_t vprod1x0123c2 = vmull_s8(vb0123c2, va1c2);
135 const int16x8_t vprod2x0123c2 = vmull_s8(vb0123c2, va2c2);
136 vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c2);
137 vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c2);
138 vacc2x0123 = vpadalq_s16(vacc2x0123, vprod2x0123c2);
139 const int16x8_t vprod0x4567c2 = vmull_s8(vb4567c2, va0c2);
140 const int16x8_t vprod1x4567c2 = vmull_s8(vb4567c2, va1c2);
141 const int16x8_t vprod2x4567c2 = vmull_s8(vb4567c2, va2c2);
142 vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c2);
143 vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c2);
144 vacc2x4567 = vpadalq_s16(vacc2x4567, vprod2x4567c2);
145 const int8x8_t va0c3 = vreinterpret_s8_s16(va01.val[1]);
146 const int8x8_t va1c3 = vreinterpret_s8_s16(va11.val[1]);
147 const int8x8_t va2c3 = vreinterpret_s8_s16(va21.val[1]);
148
149 const int16x8_t vprod0x0123c3 = vmull_s8(vb0123c3, va0c3);
150 const int16x8_t vprod1x0123c3 = vmull_s8(vb0123c3, va1c3);
151 const int16x8_t vprod2x0123c3 = vmull_s8(vb0123c3, va2c3);
152 vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c3);
153 vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c3);
154 vacc2x0123 = vpadalq_s16(vacc2x0123, vprod2x0123c3);
155 const int16x8_t vprod0x4567c3 = vmull_s8(vb4567c3, va0c3);
156 const int16x8_t vprod1x4567c3 = vmull_s8(vb4567c3, va1c3);
157 const int16x8_t vprod2x4567c3 = vmull_s8(vb4567c3, va2c3);
158 vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c3);
159 vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c3);
160 vacc2x4567 = vpadalq_s16(vacc2x4567, vprod2x4567c3);
161
162 k -= 8 * sizeof(int8_t);
163 }
164
165 if XNN_UNLIKELY(k != 0) {
166 const int8x8_t va0 = vld1_s8(a0); a0 = (const int8_t*) ((uintptr_t) a0 + k);
167 const int8x8_t va1 = vld1_s8(a1); a1 = (const int8_t*) ((uintptr_t) a1 + k);
168 const int8x8_t va2 = vld1_s8(a2); a2 = (const int8_t*) ((uintptr_t) a2 + k);
169
170 const int8x8_t vb0123c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
171 const int8x8_t vb4567c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
172
173 const int8x8_t va0c0 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0), 0));
174 const int16x8_t vprod0x0123c0 = vmull_s8(vb0123c0, va0c0);
175 vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c0);
176 const int16x8_t vprod0x4567c0 = vmull_s8(vb4567c0, va0c0);
177 vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c0);
178 const int8x8_t va1c0 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va1), 0));
179 const int16x8_t vprod1x0123c0 = vmull_s8(vb0123c0, va1c0);
180 vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c0);
181 const int16x8_t vprod1x4567c0 = vmull_s8(vb4567c0, va1c0);
182 vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c0);
183 const int8x8_t va2c0 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va2), 0));
184 const int16x8_t vprod2x0123c0 = vmull_s8(vb0123c0, va2c0);
185 vacc2x0123 = vpadalq_s16(vacc2x0123, vprod2x0123c0);
186 const int16x8_t vprod2x4567c0 = vmull_s8(vb4567c0, va2c0);
187 vacc2x4567 = vpadalq_s16(vacc2x4567, vprod2x4567c0);
188
189 if (k > 2 * sizeof(int8_t)) {
190 const int8x8_t vb0123c1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
191 const int8x8_t vb4567c1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
192
193 const int8x8_t va0c1 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0), 1));
194 const int16x8_t vprod0x0123c1 = vmull_s8(vb0123c1, va0c1);
195 vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c1);
196 const int16x8_t vprod0x4567c1 = vmull_s8(vb4567c1, va0c1);
197 vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c1);
198 const int8x8_t va1c1 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va1), 1));
199 const int16x8_t vprod1x0123c1 = vmull_s8(vb0123c1, va1c1);
200 vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c1);
201 const int16x8_t vprod1x4567c1 = vmull_s8(vb4567c1, va1c1);
202 vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c1);
203 const int8x8_t va2c1 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va2), 1));
204 const int16x8_t vprod2x0123c1 = vmull_s8(vb0123c1, va2c1);
205 vacc2x0123 = vpadalq_s16(vacc2x0123, vprod2x0123c1);
206 const int16x8_t vprod2x4567c1 = vmull_s8(vb4567c1, va2c1);
207 vacc2x4567 = vpadalq_s16(vacc2x4567, vprod2x4567c1);
208
209 if (k > 4 * sizeof(int8_t)) {
210 const int8x8_t vb0123c2 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
211 const int8x8_t vb4567c2 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
212
213 const int8x8_t va0c2 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0), 2));
214 const int16x8_t vprod0x0123c2 = vmull_s8(vb0123c2, va0c2);
215 vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c2);
216 const int16x8_t vprod0x4567c2 = vmull_s8(vb4567c2, va0c2);
217 vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c2);
218 const int8x8_t va1c2 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va1), 2));
219 const int16x8_t vprod1x0123c2 = vmull_s8(vb0123c2, va1c2);
220 vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c2);
221 const int16x8_t vprod1x4567c2 = vmull_s8(vb4567c2, va1c2);
222 vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c2);
223 const int8x8_t va2c2 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va2), 2));
224 const int16x8_t vprod2x0123c2 = vmull_s8(vb0123c2, va2c2);
225 vacc2x0123 = vpadalq_s16(vacc2x0123, vprod2x0123c2);
226 const int16x8_t vprod2x4567c2 = vmull_s8(vb4567c2, va2c2);
227 vacc2x4567 = vpadalq_s16(vacc2x4567, vprod2x4567c2);
228 }
229 }
230 }
231 p -= 3 * sizeof(void*);
232 } while (p != 0);
233
234 const int32x4_t vright_pre_shift = vld1q_dup_s32(¶ms->rndnu_neon.right_pre_shift);
235 const int32x4_t vmultiplier = vld1q_dup_s32(¶ms->rndnu_neon.multiplier);
236 const int32x4_t vright_post_shift = vld1q_dup_s32(¶ms->rndnu_neon.right_post_shift);
237
238 vacc0x0123 = vqshlq_s32(vacc0x0123, vright_pre_shift);
239 vacc0x4567 = vqshlq_s32(vacc0x4567, vright_pre_shift);
240 vacc1x0123 = vqshlq_s32(vacc1x0123, vright_pre_shift);
241 vacc1x4567 = vqshlq_s32(vacc1x4567, vright_pre_shift);
242 vacc2x0123 = vqshlq_s32(vacc2x0123, vright_pre_shift);
243 vacc2x4567 = vqshlq_s32(vacc2x4567, vright_pre_shift);
244
245 vacc0x0123 = vqdmulhq_s32(vacc0x0123, vmultiplier);
246 vacc0x4567 = vqdmulhq_s32(vacc0x4567, vmultiplier);
247 vacc1x0123 = vqdmulhq_s32(vacc1x0123, vmultiplier);
248 vacc1x4567 = vqdmulhq_s32(vacc1x4567, vmultiplier);
249 vacc2x0123 = vqdmulhq_s32(vacc2x0123, vmultiplier);
250 vacc2x4567 = vqdmulhq_s32(vacc2x4567, vmultiplier);
251
252 vacc0x0123 = vrshlq_s32(vacc0x0123, vright_post_shift);
253 vacc0x4567 = vrshlq_s32(vacc0x4567, vright_post_shift);
254 vacc1x0123 = vrshlq_s32(vacc1x0123, vright_post_shift);
255 vacc1x4567 = vrshlq_s32(vacc1x4567, vright_post_shift);
256 vacc2x0123 = vrshlq_s32(vacc2x0123, vright_post_shift);
257 vacc2x4567 = vrshlq_s32(vacc2x4567, vright_post_shift);
258
259 const int16x8_t voutput_zero_point = vld1q_dup_s16(¶ms->rndnu_neon.output_zero_point);
260 #if XNN_ARCH_ARM64
261 int16x8_t vacc0x01234567 = vqmovn_high_s32(vqmovn_s32(vacc0x0123), vacc0x4567);
262 int16x8_t vacc1x01234567 = vqmovn_high_s32(vqmovn_s32(vacc1x0123), vacc1x4567);
263 int16x8_t vacc2x01234567 = vqmovn_high_s32(vqmovn_s32(vacc2x0123), vacc2x4567);
264
265 vacc0x01234567 = vqaddq_s16(vacc0x01234567, voutput_zero_point);
266 vacc1x01234567 = vqaddq_s16(vacc1x01234567, voutput_zero_point);
267 vacc2x01234567 = vqaddq_s16(vacc2x01234567, voutput_zero_point);
268
269 int8x16_t vout0x01234567_1x01234567 = vqmovn_high_s16(vqmovn_s16(vacc0x01234567), vacc1x01234567);
270 int8x8_t vout2x01234567 = vqmovn_s16(vacc2x01234567);
271 #else
272 int16x8_t vacc0x01234567 = vcombine_s16(vqmovn_s32(vacc0x0123), vqmovn_s32(vacc0x4567));
273 int16x8_t vacc1x01234567 = vcombine_s16(vqmovn_s32(vacc1x0123), vqmovn_s32(vacc1x4567));
274 int16x8_t vacc2x01234567 = vcombine_s16(vqmovn_s32(vacc2x0123), vqmovn_s32(vacc2x4567));
275
276 vacc0x01234567 = vqaddq_s16(vacc0x01234567, voutput_zero_point);
277 vacc1x01234567 = vqaddq_s16(vacc1x01234567, voutput_zero_point);
278 vacc2x01234567 = vqaddq_s16(vacc2x01234567, voutput_zero_point);
279
280 int8x16_t vout0x01234567_1x01234567 = vcombine_s8(vqmovn_s16(vacc0x01234567), vqmovn_s16(vacc1x01234567));
281 int8x8_t vout2x01234567 = vqmovn_s16(vacc2x01234567);
282 #endif
283
284 const int8x16_t voutput_min = vld1q_dup_s8(¶ms->rndnu_neon.output_min);
285 vout0x01234567_1x01234567 = vmaxq_s8(vout0x01234567_1x01234567, voutput_min);
286 vout2x01234567 = vmax_s8(vout2x01234567, vget_low_s8(voutput_min));
287
288 const int8x16_t voutput_max = vld1q_dup_s8(¶ms->rndnu_neon.output_max);
289 vout0x01234567_1x01234567 = vminq_s8(vout0x01234567_1x01234567, voutput_max);
290 vout2x01234567 = vmin_s8(vout2x01234567, vget_low_s8(voutput_max));
291
292 if (nc >= 8) {
293 vst1_s8(c2 + 0, vout2x01234567);
294 vst1_s8(c1 + 0, vget_high_s8(vout0x01234567_1x01234567));
295 vst1_s8(c0 + 0, vget_low_s8(vout0x01234567_1x01234567));
296
297 c2 = (int8_t*) ((uintptr_t) c2 + cn_stride);
298 c1 = (int8_t*) ((uintptr_t) c1 + cn_stride);
299 c0 = (int8_t*) ((uintptr_t) c0 + cn_stride);
300
301 a = (const int8_t**restrict) ((uintptr_t) a - ks);
302
303 nc -= 8;
304 } else {
305 if (nc & 4) {
306 vst1_lane_u32((void*) c2, vreinterpret_u32_s8(vout2x01234567), 0); c2 += 4;
307 vst1q_lane_u32((void*) c1, vreinterpretq_u32_s8(vout0x01234567_1x01234567), 2); c1 += 4;
308 vst1q_lane_u32((void*) c0, vreinterpretq_u32_s8(vout0x01234567_1x01234567), 0); c0 += 4;
309 vout2x01234567 = vext_s8(vout2x01234567, vout2x01234567, 4);
310 vout0x01234567_1x01234567 = vextq_s8(vout0x01234567_1x01234567, vout0x01234567_1x01234567, 4);
311 }
312 if (nc & 2) {
313 vst1_lane_u16((void*) c2, vreinterpret_u16_s8(vout2x01234567), 0); c2 += 2;
314 vst1q_lane_u16((void*) c1, vreinterpretq_u16_s8(vout0x01234567_1x01234567), 4); c1 += 2;
315 vst1q_lane_u16((void*) c0, vreinterpretq_u16_s8(vout0x01234567_1x01234567), 0); c0 += 2;
316 vout2x01234567 = vext_s8(vout2x01234567, vout2x01234567, 2);
317 vout0x01234567_1x01234567 = vextq_s8(vout0x01234567_1x01234567, vout0x01234567_1x01234567, 2);
318 }
319 if (nc & 1) {
320 vst1_lane_s8(c2, vout2x01234567, 0);
321 vst1q_lane_s8(c1, vout0x01234567_1x01234567, 8);
322 vst1q_lane_s8(c0, vout0x01234567_1x01234567, 0);
323 }
324
325 nc = 0;
326 }
327 } while (nc != 0);
328 }
329