1 // Auto-generated file. Do not edit!
2 // Template: src/qs8-igemm/c4-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
17
xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c4__neon_mull_ld1r(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_3x8c4__neon_mull_ld1r(
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, 4 * 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 vacc0x01 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
56 int32x4_t vacc0x23 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
57 int32x4_t vacc0x45 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
58 int32x4_t vacc0x67 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
59 int32x4_t vacc1x01 = vacc0x01;
60 int32x4_t vacc1x23 = vacc0x23;
61 int32x4_t vacc1x45 = vacc0x45;
62 int32x4_t vacc1x67 = vacc0x67;
63 int32x4_t vacc2x01 = vacc0x01;
64 int32x4_t vacc2x23 = vacc0x23;
65 int32x4_t vacc2x45 = vacc0x45;
66 int32x4_t vacc2x67 = vacc0x67;
67
68 size_t p = ks;
69 do {
70 const int8_t* restrict a0 = a[0];
71 if XNN_UNPREDICTABLE(a0 != zero) {
72 a0 = (const int8_t*) ((uintptr_t) a0 + a_offset);
73 }
74 const int8_t* restrict a1 = a[1];
75 if XNN_UNPREDICTABLE(a1 != zero) {
76 a1 = (const int8_t*) ((uintptr_t) a1 + a_offset);
77 }
78 const int8_t* restrict a2 = a[2];
79 if XNN_UNPREDICTABLE(a2 != zero) {
80 a2 = (const int8_t*) ((uintptr_t) a2 + a_offset);
81 }
82 a += 3;
83
84 size_t k = kc;
85
86
87 while (k >= 8 * sizeof(int8_t)) {
88 const int32x2_t va00 = vld1_dup_s32((const void*)a0);
89 const int32x2_t va01 = vld1_dup_s32((const void*)(a0 + 4)); a0 += 8;
90 const int32x2_t va10 = vld1_dup_s32((const void*)a1);
91 const int32x2_t va11 = vld1_dup_s32((const void*)(a1 + 4)); a1 += 8;
92 const int32x2_t va20 = vld1_dup_s32((const void*)a2);
93 const int32x2_t va21 = vld1_dup_s32((const void*)(a2 + 4)); a2 += 8;
94
95 const int8x8_t vb01c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
96 const int8x8_t vb23c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
97 const int8x8_t vb45c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
98 const int8x8_t vb67c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
99 const int8x8_t vb01c1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
100 const int8x8_t vb23c1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
101 const int8x8_t vb45c1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
102 const int8x8_t vb67c1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
103
104 const int8x8_t va0c0 = vreinterpret_s8_s32(va00);
105 const int8x8_t va1c0 = vreinterpret_s8_s32(va10);
106 const int8x8_t va2c0 = vreinterpret_s8_s32(va20);
107
108 const int16x8_t vprod0x01c0 = vmull_s8(vb01c0, va0c0);
109 const int16x8_t vprod1x01c0 = vmull_s8(vb01c0, va1c0);
110 const int16x8_t vprod2x01c0 = vmull_s8(vb01c0, va2c0);
111 vacc0x01 = vpadalq_s16(vacc0x01, vprod0x01c0);
112 vacc1x01 = vpadalq_s16(vacc1x01, vprod1x01c0);
113 vacc2x01 = vpadalq_s16(vacc2x01, vprod2x01c0);
114 const int16x8_t vprod0x23c0 = vmull_s8(vb23c0, va0c0);
115 const int16x8_t vprod1x23c0 = vmull_s8(vb23c0, va1c0);
116 const int16x8_t vprod2x23c0 = vmull_s8(vb23c0, va2c0);
117 vacc0x23 = vpadalq_s16(vacc0x23, vprod0x23c0);
118 vacc1x23 = vpadalq_s16(vacc1x23, vprod1x23c0);
119 vacc2x23 = vpadalq_s16(vacc2x23, vprod2x23c0);
120 const int16x8_t vprod0x45c0 = vmull_s8(vb45c0, va0c0);
121 const int16x8_t vprod1x45c0 = vmull_s8(vb45c0, va1c0);
122 const int16x8_t vprod2x45c0 = vmull_s8(vb45c0, va2c0);
123 vacc0x45 = vpadalq_s16(vacc0x45, vprod0x45c0);
124 vacc1x45 = vpadalq_s16(vacc1x45, vprod1x45c0);
125 vacc2x45 = vpadalq_s16(vacc2x45, vprod2x45c0);
126 const int16x8_t vprod0x67c0 = vmull_s8(vb67c0, va0c0);
127 const int16x8_t vprod1x67c0 = vmull_s8(vb67c0, va1c0);
128 const int16x8_t vprod2x67c0 = vmull_s8(vb67c0, va2c0);
129 vacc0x67 = vpadalq_s16(vacc0x67, vprod0x67c0);
130 vacc1x67 = vpadalq_s16(vacc1x67, vprod1x67c0);
131 vacc2x67 = vpadalq_s16(vacc2x67, vprod2x67c0);
132 const int8x8_t va0c1 = vreinterpret_s8_s32(va01);
133 const int8x8_t va1c1 = vreinterpret_s8_s32(va11);
134 const int8x8_t va2c1 = vreinterpret_s8_s32(va21);
135
136 const int16x8_t vprod0x01c1 = vmull_s8(vb01c1, va0c1);
137 const int16x8_t vprod1x01c1 = vmull_s8(vb01c1, va1c1);
138 const int16x8_t vprod2x01c1 = vmull_s8(vb01c1, va2c1);
139 vacc0x01 = vpadalq_s16(vacc0x01, vprod0x01c1);
140 vacc1x01 = vpadalq_s16(vacc1x01, vprod1x01c1);
141 vacc2x01 = vpadalq_s16(vacc2x01, vprod2x01c1);
142 const int16x8_t vprod0x23c1 = vmull_s8(vb23c1, va0c1);
143 const int16x8_t vprod1x23c1 = vmull_s8(vb23c1, va1c1);
144 const int16x8_t vprod2x23c1 = vmull_s8(vb23c1, va2c1);
145 vacc0x23 = vpadalq_s16(vacc0x23, vprod0x23c1);
146 vacc1x23 = vpadalq_s16(vacc1x23, vprod1x23c1);
147 vacc2x23 = vpadalq_s16(vacc2x23, vprod2x23c1);
148 const int16x8_t vprod0x45c1 = vmull_s8(vb45c1, va0c1);
149 const int16x8_t vprod1x45c1 = vmull_s8(vb45c1, va1c1);
150 const int16x8_t vprod2x45c1 = vmull_s8(vb45c1, va2c1);
151 vacc0x45 = vpadalq_s16(vacc0x45, vprod0x45c1);
152 vacc1x45 = vpadalq_s16(vacc1x45, vprod1x45c1);
153 vacc2x45 = vpadalq_s16(vacc2x45, vprod2x45c1);
154 const int16x8_t vprod0x67c1 = vmull_s8(vb67c1, va0c1);
155 const int16x8_t vprod1x67c1 = vmull_s8(vb67c1, va1c1);
156 const int16x8_t vprod2x67c1 = vmull_s8(vb67c1, va2c1);
157 vacc0x67 = vpadalq_s16(vacc0x67, vprod0x67c1);
158 vacc1x67 = vpadalq_s16(vacc1x67, vprod1x67c1);
159 vacc2x67 = vpadalq_s16(vacc2x67, vprod2x67c1);
160
161 k -= 8 * sizeof(int8_t);
162 }
163
164 if XNN_UNLIKELY(k != 0) {
165 const int8x8_t va0 = vld1_s8(a0); a0 = (const int8_t*) ((uintptr_t) a0 + k);
166 const int8x8_t va1 = vld1_s8(a1); a1 = (const int8_t*) ((uintptr_t) a1 + k);
167 const int8x8_t va2 = vld1_s8(a2); a2 = (const int8_t*) ((uintptr_t) a2 + k);
168
169 const int8x8_t vb01c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
170 const int8x8_t vb23c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
171 const int8x8_t vb45c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
172 const int8x8_t vb67c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
173
174 const int8x8_t va0c0 = vreinterpret_s8_s32(vdup_lane_s32(vreinterpret_s32_s8(va0), 0));
175 const int16x8_t vprod0x01c0 = vmull_s8(vb01c0, va0c0);
176 vacc0x01 = vpadalq_s16(vacc0x01, vprod0x01c0);
177 const int16x8_t vprod0x23c0 = vmull_s8(vb23c0, va0c0);
178 vacc0x23 = vpadalq_s16(vacc0x23, vprod0x23c0);
179 const int16x8_t vprod0x45c0 = vmull_s8(vb45c0, va0c0);
180 vacc0x45 = vpadalq_s16(vacc0x45, vprod0x45c0);
181 const int16x8_t vprod0x67c0 = vmull_s8(vb67c0, va0c0);
182 vacc0x67 = vpadalq_s16(vacc0x67, vprod0x67c0);
183 const int8x8_t va1c0 = vreinterpret_s8_s32(vdup_lane_s32(vreinterpret_s32_s8(va1), 0));
184 const int16x8_t vprod1x01c0 = vmull_s8(vb01c0, va1c0);
185 vacc1x01 = vpadalq_s16(vacc1x01, vprod1x01c0);
186 const int16x8_t vprod1x23c0 = vmull_s8(vb23c0, va1c0);
187 vacc1x23 = vpadalq_s16(vacc1x23, vprod1x23c0);
188 const int16x8_t vprod1x45c0 = vmull_s8(vb45c0, va1c0);
189 vacc1x45 = vpadalq_s16(vacc1x45, vprod1x45c0);
190 const int16x8_t vprod1x67c0 = vmull_s8(vb67c0, va1c0);
191 vacc1x67 = vpadalq_s16(vacc1x67, vprod1x67c0);
192 const int8x8_t va2c0 = vreinterpret_s8_s32(vdup_lane_s32(vreinterpret_s32_s8(va2), 0));
193 const int16x8_t vprod2x01c0 = vmull_s8(vb01c0, va2c0);
194 vacc2x01 = vpadalq_s16(vacc2x01, vprod2x01c0);
195 const int16x8_t vprod2x23c0 = vmull_s8(vb23c0, va2c0);
196 vacc2x23 = vpadalq_s16(vacc2x23, vprod2x23c0);
197 const int16x8_t vprod2x45c0 = vmull_s8(vb45c0, va2c0);
198 vacc2x45 = vpadalq_s16(vacc2x45, vprod2x45c0);
199 const int16x8_t vprod2x67c0 = vmull_s8(vb67c0, va2c0);
200 vacc2x67 = vpadalq_s16(vacc2x67, vprod2x67c0);
201 }
202 p -= 3 * sizeof(void*);
203 } while (p != 0);
204
205 #if XNN_ARCH_ARM64
206 int32x4_t vacc0x0123 = vpaddq_s32(vacc0x01, vacc0x23);
207 int32x4_t vacc0x4567 = vpaddq_s32(vacc0x45, vacc0x67);
208 int32x4_t vacc1x0123 = vpaddq_s32(vacc1x01, vacc1x23);
209 int32x4_t vacc1x4567 = vpaddq_s32(vacc1x45, vacc1x67);
210 int32x4_t vacc2x0123 = vpaddq_s32(vacc2x01, vacc2x23);
211 int32x4_t vacc2x4567 = vpaddq_s32(vacc2x45, vacc2x67);
212 #else
213 const int32x2_t vsum0x01 = vpadd_s32(vget_low_s32(vacc0x01), vget_high_s32(vacc0x01));
214 const int32x2_t vsum0x23 = vpadd_s32(vget_low_s32(vacc0x23), vget_high_s32(vacc0x23));
215 int32x4_t vacc0x0123 = vcombine_s32(vsum0x01, vsum0x23);
216 const int32x2_t vsum0x45 = vpadd_s32(vget_low_s32(vacc0x45), vget_high_s32(vacc0x45));
217 const int32x2_t vsum0x67 = vpadd_s32(vget_low_s32(vacc0x67), vget_high_s32(vacc0x67));
218 int32x4_t vacc0x4567 = vcombine_s32(vsum0x45, vsum0x67);
219 const int32x2_t vsum1x01 = vpadd_s32(vget_low_s32(vacc1x01), vget_high_s32(vacc1x01));
220 const int32x2_t vsum1x23 = vpadd_s32(vget_low_s32(vacc1x23), vget_high_s32(vacc1x23));
221 int32x4_t vacc1x0123 = vcombine_s32(vsum1x01, vsum1x23);
222 const int32x2_t vsum1x45 = vpadd_s32(vget_low_s32(vacc1x45), vget_high_s32(vacc1x45));
223 const int32x2_t vsum1x67 = vpadd_s32(vget_low_s32(vacc1x67), vget_high_s32(vacc1x67));
224 int32x4_t vacc1x4567 = vcombine_s32(vsum1x45, vsum1x67);
225 const int32x2_t vsum2x01 = vpadd_s32(vget_low_s32(vacc2x01), vget_high_s32(vacc2x01));
226 const int32x2_t vsum2x23 = vpadd_s32(vget_low_s32(vacc2x23), vget_high_s32(vacc2x23));
227 int32x4_t vacc2x0123 = vcombine_s32(vsum2x01, vsum2x23);
228 const int32x2_t vsum2x45 = vpadd_s32(vget_low_s32(vacc2x45), vget_high_s32(vacc2x45));
229 const int32x2_t vsum2x67 = vpadd_s32(vget_low_s32(vacc2x67), vget_high_s32(vacc2x67));
230 int32x4_t vacc2x4567 = vcombine_s32(vsum2x45, vsum2x67);
231 #endif
232
233 const int32x4_t vright_pre_shift = vld1q_dup_s32(¶ms->rndnu_neon.right_pre_shift);
234 const int32x4_t vmultiplier = vld1q_dup_s32(¶ms->rndnu_neon.multiplier);
235 const int32x4_t vright_post_shift = vld1q_dup_s32(¶ms->rndnu_neon.right_post_shift);
236
237 vacc0x0123 = vqshlq_s32(vacc0x0123, vright_pre_shift);
238 vacc0x4567 = vqshlq_s32(vacc0x4567, vright_pre_shift);
239 vacc1x0123 = vqshlq_s32(vacc1x0123, vright_pre_shift);
240 vacc1x4567 = vqshlq_s32(vacc1x4567, vright_pre_shift);
241 vacc2x0123 = vqshlq_s32(vacc2x0123, vright_pre_shift);
242 vacc2x4567 = vqshlq_s32(vacc2x4567, vright_pre_shift);
243
244 vacc0x0123 = vqdmulhq_s32(vacc0x0123, vmultiplier);
245 vacc0x4567 = vqdmulhq_s32(vacc0x4567, vmultiplier);
246 vacc1x0123 = vqdmulhq_s32(vacc1x0123, vmultiplier);
247 vacc1x4567 = vqdmulhq_s32(vacc1x4567, vmultiplier);
248 vacc2x0123 = vqdmulhq_s32(vacc2x0123, vmultiplier);
249 vacc2x4567 = vqdmulhq_s32(vacc2x4567, vmultiplier);
250
251 vacc0x0123 = vrshlq_s32(vacc0x0123, vright_post_shift);
252 vacc0x4567 = vrshlq_s32(vacc0x4567, vright_post_shift);
253 vacc1x0123 = vrshlq_s32(vacc1x0123, vright_post_shift);
254 vacc1x4567 = vrshlq_s32(vacc1x4567, vright_post_shift);
255 vacc2x0123 = vrshlq_s32(vacc2x0123, vright_post_shift);
256 vacc2x4567 = vrshlq_s32(vacc2x4567, vright_post_shift);
257
258 const int16x8_t voutput_zero_point = vld1q_dup_s16(¶ms->rndnu_neon.output_zero_point);
259 #if XNN_ARCH_ARM64
260 int16x8_t vacc0x01234567 = vqmovn_high_s32(vqmovn_s32(vacc0x0123), vacc0x4567);
261 int16x8_t vacc1x01234567 = vqmovn_high_s32(vqmovn_s32(vacc1x0123), vacc1x4567);
262 int16x8_t vacc2x01234567 = vqmovn_high_s32(vqmovn_s32(vacc2x0123), vacc2x4567);
263
264 vacc0x01234567 = vqaddq_s16(vacc0x01234567, voutput_zero_point);
265 vacc1x01234567 = vqaddq_s16(vacc1x01234567, voutput_zero_point);
266 vacc2x01234567 = vqaddq_s16(vacc2x01234567, voutput_zero_point);
267
268 int8x16_t vout0x01234567_1x01234567 = vqmovn_high_s16(vqmovn_s16(vacc0x01234567), vacc1x01234567);
269 int8x8_t vout2x01234567 = vqmovn_s16(vacc2x01234567);
270 #else
271 int16x8_t vacc0x01234567 = vcombine_s16(vqmovn_s32(vacc0x0123), vqmovn_s32(vacc0x4567));
272 int16x8_t vacc1x01234567 = vcombine_s16(vqmovn_s32(vacc1x0123), vqmovn_s32(vacc1x4567));
273 int16x8_t vacc2x01234567 = vcombine_s16(vqmovn_s32(vacc2x0123), vqmovn_s32(vacc2x4567));
274
275 vacc0x01234567 = vqaddq_s16(vacc0x01234567, voutput_zero_point);
276 vacc1x01234567 = vqaddq_s16(vacc1x01234567, voutput_zero_point);
277 vacc2x01234567 = vqaddq_s16(vacc2x01234567, voutput_zero_point);
278
279 int8x16_t vout0x01234567_1x01234567 = vcombine_s8(vqmovn_s16(vacc0x01234567), vqmovn_s16(vacc1x01234567));
280 int8x8_t vout2x01234567 = vqmovn_s16(vacc2x01234567);
281 #endif
282
283 const int8x16_t voutput_min = vld1q_dup_s8(¶ms->rndnu_neon.output_min);
284 vout0x01234567_1x01234567 = vmaxq_s8(vout0x01234567_1x01234567, voutput_min);
285 vout2x01234567 = vmax_s8(vout2x01234567, vget_low_s8(voutput_min));
286
287 const int8x16_t voutput_max = vld1q_dup_s8(¶ms->rndnu_neon.output_max);
288 vout0x01234567_1x01234567 = vminq_s8(vout0x01234567_1x01234567, voutput_max);
289 vout2x01234567 = vmin_s8(vout2x01234567, vget_low_s8(voutput_max));
290
291 if (nc >= 8) {
292 vst1_s8(c2 + 0, vout2x01234567);
293 vst1_s8(c1 + 0, vget_high_s8(vout0x01234567_1x01234567));
294 vst1_s8(c0 + 0, vget_low_s8(vout0x01234567_1x01234567));
295
296 c2 = (int8_t*) ((uintptr_t) c2 + cn_stride);
297 c1 = (int8_t*) ((uintptr_t) c1 + cn_stride);
298 c0 = (int8_t*) ((uintptr_t) c0 + cn_stride);
299
300 a = (const int8_t**restrict) ((uintptr_t) a - ks);
301
302 nc -= 8;
303 } else {
304 if (nc & 4) {
305 vst1_lane_u32((void*) c2, vreinterpret_u32_s8(vout2x01234567), 0); c2 += 4;
306 vst1q_lane_u32((void*) c1, vreinterpretq_u32_s8(vout0x01234567_1x01234567), 2); c1 += 4;
307 vst1q_lane_u32((void*) c0, vreinterpretq_u32_s8(vout0x01234567_1x01234567), 0); c0 += 4;
308 vout2x01234567 = vext_s8(vout2x01234567, vout2x01234567, 4);
309 vout0x01234567_1x01234567 = vextq_s8(vout0x01234567_1x01234567, vout0x01234567_1x01234567, 4);
310 }
311 if (nc & 2) {
312 vst1_lane_u16((void*) c2, vreinterpret_u16_s8(vout2x01234567), 0); c2 += 2;
313 vst1q_lane_u16((void*) c1, vreinterpretq_u16_s8(vout0x01234567_1x01234567), 4); c1 += 2;
314 vst1q_lane_u16((void*) c0, vreinterpretq_u16_s8(vout0x01234567_1x01234567), 0); c0 += 2;
315 vout2x01234567 = vext_s8(vout2x01234567, vout2x01234567, 2);
316 vout0x01234567_1x01234567 = vextq_s8(vout0x01234567_1x01234567, vout0x01234567_1x01234567, 2);
317 }
318 if (nc & 1) {
319 vst1_lane_s8(c2, vout2x01234567, 0);
320 vst1q_lane_s8(c1, vout0x01234567_1x01234567, 8);
321 vst1q_lane_s8(c0, vout0x01234567_1x01234567, 0);
322 }
323
324 nc = 0;
325 }
326 } while (nc != 0);
327 }
328