1 // Auto-generated file. Do not edit!
2 // Template: src/qs8-igemm/c4-neon-mull-shuffle.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_2x16c4s2__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_2x16c4s2__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 <= 2);
34 assert(nc != 0);
35 assert(kc != 0);
36 assert(ks != 0);
37 assert(ks % (2 * 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 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
49 kc = round_up_po2(kc, 8 * sizeof(int8_t));
50 do {
51 int32x4_t vacc0x01 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const int32_t*) w + 2;
52 int32x4_t vacc0x23 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const int32_t*) w + 2;
53 int32x4_t vacc0x45 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const int32_t*) w + 2;
54 int32x4_t vacc0x67 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const int32_t*) w + 2;
55 int32x4_t vacc0x89 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const int32_t*) w + 2;
56 int32x4_t vacc0xAB = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const int32_t*) w + 2;
57 int32x4_t vacc0xCD = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const int32_t*) w + 2;
58 int32x4_t vacc0xEF = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const int32_t*) w + 2;
59 int32x4_t vacc1x01 = vacc0x01;
60 int32x4_t vacc1x23 = vacc0x23;
61 int32x4_t vacc1x45 = vacc0x45;
62 int32x4_t vacc1x67 = vacc0x67;
63 int32x4_t vacc1x89 = vacc0x89;
64 int32x4_t vacc1xAB = vacc0xAB;
65 int32x4_t vacc1xCD = vacc0xCD;
66 int32x4_t vacc1xEF = vacc0xEF;
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 a += 2;
79
80 size_t k = kc;
81 do {
82 int8x8_t va0x0 = vld1_s8(a0); a0 += 8;
83 int8x8_t va1x0 = vld1_s8(a1); a1 += 8;
84
85 const int8x8_t vb01c0x0 = vld1_s8(w); w = (const int8_t*) w + 8;
86 const int8x8_t vb23c0x0 = vld1_s8(w); w = (const int8_t*) w + 8;
87 const int8x8_t vb45c0x0 = vld1_s8(w); w = (const int8_t*) w + 8;
88 const int8x8_t vb67c0x0 = vld1_s8(w); w = (const int8_t*) w + 8;
89 const int8x8_t vb89c0x0 = vld1_s8(w); w = (const int8_t*) w + 8;
90 const int8x8_t vbABc0x0 = vld1_s8(w); w = (const int8_t*) w + 8;
91 const int8x8_t vbCDc0x0 = vld1_s8(w); w = (const int8_t*) w + 8;
92 const int8x8_t vbEFc0x0 = vld1_s8(w); w = (const int8_t*) w + 8;
93 const int8x8_t vb01c1x0 = vld1_s8(w); w = (const int8_t*) w + 8;
94 const int8x8_t vb23c1x0 = vld1_s8(w); w = (const int8_t*) w + 8;
95 const int8x8_t vb45c1x0 = vld1_s8(w); w = (const int8_t*) w + 8;
96 const int8x8_t vb67c1x0 = vld1_s8(w); w = (const int8_t*) w + 8;
97 const int8x8_t vb89c1x0 = vld1_s8(w); w = (const int8_t*) w + 8;
98 const int8x8_t vbABc1x0 = vld1_s8(w); w = (const int8_t*) w + 8;
99 const int8x8_t vbCDc1x0 = vld1_s8(w); w = (const int8_t*) w + 8;
100 const int8x8_t vbEFc1x0 = vld1_s8(w); w = (const int8_t*) w + 8;
101
102 int16x8_t vprod0x01c0 = vmull_s8(vb01c0x0, va0x0);
103 int16x8_t vprod1x01c0 = vmull_s8(vb01c0x0, va1x0);
104 vacc0x01 = vpadalq_s16(vacc0x01, vprod0x01c0);
105 vacc1x01 = vpadalq_s16(vacc1x01, vprod1x01c0);
106 int16x8_t vprod0x23c0 = vmull_s8(vb23c0x0, va0x0);
107 int16x8_t vprod1x23c0 = vmull_s8(vb23c0x0, va1x0);
108 vacc0x23 = vpadalq_s16(vacc0x23, vprod0x23c0);
109 vacc1x23 = vpadalq_s16(vacc1x23, vprod1x23c0);
110 int16x8_t vprod0x45c0 = vmull_s8(vb45c0x0, va0x0);
111 int16x8_t vprod1x45c0 = vmull_s8(vb45c0x0, va1x0);
112 vacc0x45 = vpadalq_s16(vacc0x45, vprod0x45c0);
113 vacc1x45 = vpadalq_s16(vacc1x45, vprod1x45c0);
114 int16x8_t vprod0x67c0 = vmull_s8(vb67c0x0, va0x0);
115 int16x8_t vprod1x67c0 = vmull_s8(vb67c0x0, va1x0);
116 vacc0x67 = vpadalq_s16(vacc0x67, vprod0x67c0);
117 vacc1x67 = vpadalq_s16(vacc1x67, vprod1x67c0);
118 int16x8_t vprod0x89c0 = vmull_s8(vb89c0x0, va0x0);
119 int16x8_t vprod1x89c0 = vmull_s8(vb89c0x0, va1x0);
120 vacc0x89 = vpadalq_s16(vacc0x89, vprod0x89c0);
121 vacc1x89 = vpadalq_s16(vacc1x89, vprod1x89c0);
122 int16x8_t vprod0xABc0 = vmull_s8(vbABc0x0, va0x0);
123 int16x8_t vprod1xABc0 = vmull_s8(vbABc0x0, va1x0);
124 vacc0xAB = vpadalq_s16(vacc0xAB, vprod0xABc0);
125 vacc1xAB = vpadalq_s16(vacc1xAB, vprod1xABc0);
126 int16x8_t vprod0xCDc0 = vmull_s8(vbCDc0x0, va0x0);
127 int16x8_t vprod1xCDc0 = vmull_s8(vbCDc0x0, va1x0);
128 vacc0xCD = vpadalq_s16(vacc0xCD, vprod0xCDc0);
129 vacc1xCD = vpadalq_s16(vacc1xCD, vprod1xCDc0);
130 int16x8_t vprod0xEFc0 = vmull_s8(vbEFc0x0, va0x0);
131 int16x8_t vprod1xEFc0 = vmull_s8(vbEFc0x0, va1x0);
132 vacc0xEF = vpadalq_s16(vacc0xEF, vprod0xEFc0);
133 vacc1xEF = vpadalq_s16(vacc1xEF, vprod1xEFc0);
134 va0x0 = vext_s8(va0x0, va0x0, 4);
135 va1x0 = vext_s8(va1x0, va1x0, 4);
136 int16x8_t vprod0x01c1 = vmull_s8(vb01c1x0, va0x0);
137 int16x8_t vprod1x01c1 = vmull_s8(vb01c1x0, va1x0);
138 vacc0x01 = vpadalq_s16(vacc0x01, vprod0x01c1);
139 vacc1x01 = vpadalq_s16(vacc1x01, vprod1x01c1);
140 int16x8_t vprod0x23c1 = vmull_s8(vb23c1x0, va0x0);
141 int16x8_t vprod1x23c1 = vmull_s8(vb23c1x0, va1x0);
142 vacc0x23 = vpadalq_s16(vacc0x23, vprod0x23c1);
143 vacc1x23 = vpadalq_s16(vacc1x23, vprod1x23c1);
144 int16x8_t vprod0x45c1 = vmull_s8(vb45c1x0, va0x0);
145 int16x8_t vprod1x45c1 = vmull_s8(vb45c1x0, va1x0);
146 vacc0x45 = vpadalq_s16(vacc0x45, vprod0x45c1);
147 vacc1x45 = vpadalq_s16(vacc1x45, vprod1x45c1);
148 int16x8_t vprod0x67c1 = vmull_s8(vb67c1x0, va0x0);
149 int16x8_t vprod1x67c1 = vmull_s8(vb67c1x0, va1x0);
150 vacc0x67 = vpadalq_s16(vacc0x67, vprod0x67c1);
151 vacc1x67 = vpadalq_s16(vacc1x67, vprod1x67c1);
152 int16x8_t vprod0x89c1 = vmull_s8(vb89c1x0, va0x0);
153 int16x8_t vprod1x89c1 = vmull_s8(vb89c1x0, va1x0);
154 vacc0x89 = vpadalq_s16(vacc0x89, vprod0x89c1);
155 vacc1x89 = vpadalq_s16(vacc1x89, vprod1x89c1);
156 int16x8_t vprod0xABc1 = vmull_s8(vbABc1x0, va0x0);
157 int16x8_t vprod1xABc1 = vmull_s8(vbABc1x0, va1x0);
158 vacc0xAB = vpadalq_s16(vacc0xAB, vprod0xABc1);
159 vacc1xAB = vpadalq_s16(vacc1xAB, vprod1xABc1);
160 int16x8_t vprod0xCDc1 = vmull_s8(vbCDc1x0, va0x0);
161 int16x8_t vprod1xCDc1 = vmull_s8(vbCDc1x0, va1x0);
162 vacc0xCD = vpadalq_s16(vacc0xCD, vprod0xCDc1);
163 vacc1xCD = vpadalq_s16(vacc1xCD, vprod1xCDc1);
164 int16x8_t vprod0xEFc1 = vmull_s8(vbEFc1x0, va0x0);
165 int16x8_t vprod1xEFc1 = vmull_s8(vbEFc1x0, va1x0);
166 vacc0xEF = vpadalq_s16(vacc0xEF, vprod0xEFc1);
167 vacc1xEF = vpadalq_s16(vacc1xEF, vprod1xEFc1);
168
169 k -= 8 * sizeof(int8_t);
170 } while (k != 0);
171
172 p -= 2 * sizeof(void*);
173 } while (p != 0);
174
175 #if XNN_ARCH_ARM64
176 int32x4_t vacc0x0123 = vpaddq_s32(vacc0x01, vacc0x23);
177 int32x4_t vacc0x4567 = vpaddq_s32(vacc0x45, vacc0x67);
178 int32x4_t vacc0x89AB = vpaddq_s32(vacc0x89, vacc0xAB);
179 int32x4_t vacc0xCDEF = vpaddq_s32(vacc0xCD, vacc0xEF);
180 int32x4_t vacc1x0123 = vpaddq_s32(vacc1x01, vacc1x23);
181 int32x4_t vacc1x4567 = vpaddq_s32(vacc1x45, vacc1x67);
182 int32x4_t vacc1x89AB = vpaddq_s32(vacc1x89, vacc1xAB);
183 int32x4_t vacc1xCDEF = vpaddq_s32(vacc1xCD, vacc1xEF);
184 #else
185 const int32x2_t vsum0x01 = vpadd_s32(vget_low_s32(vacc0x01), vget_high_s32(vacc0x01));
186 const int32x2_t vsum0x23 = vpadd_s32(vget_low_s32(vacc0x23), vget_high_s32(vacc0x23));
187 int32x4_t vacc0x0123 = vcombine_s32(vsum0x01, vsum0x23);
188 const int32x2_t vsum0x45 = vpadd_s32(vget_low_s32(vacc0x45), vget_high_s32(vacc0x45));
189 const int32x2_t vsum0x67 = vpadd_s32(vget_low_s32(vacc0x67), vget_high_s32(vacc0x67));
190 int32x4_t vacc0x4567 = vcombine_s32(vsum0x45, vsum0x67);
191 const int32x2_t vsum0x89 = vpadd_s32(vget_low_s32(vacc0x89), vget_high_s32(vacc0x89));
192 const int32x2_t vsum0xAB = vpadd_s32(vget_low_s32(vacc0xAB), vget_high_s32(vacc0xAB));
193 int32x4_t vacc0x89AB = vcombine_s32(vsum0x89, vsum0xAB);
194 const int32x2_t vsum0xCD = vpadd_s32(vget_low_s32(vacc0xCD), vget_high_s32(vacc0xCD));
195 const int32x2_t vsum0xEF = vpadd_s32(vget_low_s32(vacc0xEF), vget_high_s32(vacc0xEF));
196 int32x4_t vacc0xCDEF = vcombine_s32(vsum0xCD, vsum0xEF);
197 const int32x2_t vsum1x01 = vpadd_s32(vget_low_s32(vacc1x01), vget_high_s32(vacc1x01));
198 const int32x2_t vsum1x23 = vpadd_s32(vget_low_s32(vacc1x23), vget_high_s32(vacc1x23));
199 int32x4_t vacc1x0123 = vcombine_s32(vsum1x01, vsum1x23);
200 const int32x2_t vsum1x45 = vpadd_s32(vget_low_s32(vacc1x45), vget_high_s32(vacc1x45));
201 const int32x2_t vsum1x67 = vpadd_s32(vget_low_s32(vacc1x67), vget_high_s32(vacc1x67));
202 int32x4_t vacc1x4567 = vcombine_s32(vsum1x45, vsum1x67);
203 const int32x2_t vsum1x89 = vpadd_s32(vget_low_s32(vacc1x89), vget_high_s32(vacc1x89));
204 const int32x2_t vsum1xAB = vpadd_s32(vget_low_s32(vacc1xAB), vget_high_s32(vacc1xAB));
205 int32x4_t vacc1x89AB = vcombine_s32(vsum1x89, vsum1xAB);
206 const int32x2_t vsum1xCD = vpadd_s32(vget_low_s32(vacc1xCD), vget_high_s32(vacc1xCD));
207 const int32x2_t vsum1xEF = vpadd_s32(vget_low_s32(vacc1xEF), vget_high_s32(vacc1xEF));
208 int32x4_t vacc1xCDEF = vcombine_s32(vsum1xCD, vsum1xEF);
209 #endif
210
211 const int32x4_t vright_pre_shift = vld1q_dup_s32(¶ms->rndnu_neon.right_pre_shift);
212 const int32x4_t vmultiplier = vld1q_dup_s32(¶ms->rndnu_neon.multiplier);
213 const int32x4_t vright_post_shift = vld1q_dup_s32(¶ms->rndnu_neon.right_post_shift);
214
215 vacc0x0123 = vqshlq_s32(vacc0x0123, vright_pre_shift);
216 vacc0x4567 = vqshlq_s32(vacc0x4567, vright_pre_shift);
217 vacc0x89AB = vqshlq_s32(vacc0x89AB, vright_pre_shift);
218 vacc0xCDEF = vqshlq_s32(vacc0xCDEF, vright_pre_shift);
219 vacc1x0123 = vqshlq_s32(vacc1x0123, vright_pre_shift);
220 vacc1x4567 = vqshlq_s32(vacc1x4567, vright_pre_shift);
221 vacc1x89AB = vqshlq_s32(vacc1x89AB, vright_pre_shift);
222 vacc1xCDEF = vqshlq_s32(vacc1xCDEF, vright_pre_shift);
223
224 vacc0x0123 = vqdmulhq_s32(vacc0x0123, vmultiplier);
225 vacc0x4567 = vqdmulhq_s32(vacc0x4567, vmultiplier);
226 vacc0x89AB = vqdmulhq_s32(vacc0x89AB, vmultiplier);
227 vacc0xCDEF = vqdmulhq_s32(vacc0xCDEF, vmultiplier);
228 vacc1x0123 = vqdmulhq_s32(vacc1x0123, vmultiplier);
229 vacc1x4567 = vqdmulhq_s32(vacc1x4567, vmultiplier);
230 vacc1x89AB = vqdmulhq_s32(vacc1x89AB, vmultiplier);
231 vacc1xCDEF = vqdmulhq_s32(vacc1xCDEF, vmultiplier);
232
233 vacc0x0123 = vrshlq_s32(vacc0x0123, vright_post_shift);
234 vacc0x4567 = vrshlq_s32(vacc0x4567, vright_post_shift);
235 vacc0x89AB = vrshlq_s32(vacc0x89AB, vright_post_shift);
236 vacc0xCDEF = vrshlq_s32(vacc0xCDEF, vright_post_shift);
237 vacc1x0123 = vrshlq_s32(vacc1x0123, vright_post_shift);
238 vacc1x4567 = vrshlq_s32(vacc1x4567, vright_post_shift);
239 vacc1x89AB = vrshlq_s32(vacc1x89AB, vright_post_shift);
240 vacc1xCDEF = vrshlq_s32(vacc1xCDEF, vright_post_shift);
241
242 const int16x8_t voutput_zero_point = vld1q_dup_s16(¶ms->rndnu_neon.output_zero_point);
243 #if XNN_ARCH_ARM64
244 int16x8_t vacc0x01234567 = vqmovn_high_s32(vqmovn_s32(vacc0x0123), vacc0x4567);
245 int16x8_t vacc0x89ABCDEF = vqmovn_high_s32(vqmovn_s32(vacc0x89AB), vacc0xCDEF);
246 int16x8_t vacc1x01234567 = vqmovn_high_s32(vqmovn_s32(vacc1x0123), vacc1x4567);
247 int16x8_t vacc1x89ABCDEF = vqmovn_high_s32(vqmovn_s32(vacc1x89AB), vacc1xCDEF);
248
249 vacc0x01234567 = vqaddq_s16(vacc0x01234567, voutput_zero_point);
250 vacc0x89ABCDEF = vqaddq_s16(vacc0x89ABCDEF, voutput_zero_point);
251 vacc1x01234567 = vqaddq_s16(vacc1x01234567, voutput_zero_point);
252 vacc1x89ABCDEF = vqaddq_s16(vacc1x89ABCDEF, voutput_zero_point);
253
254 int8x16_t vout0x0123456789ABCDEF = vqmovn_high_s16(vqmovn_s16(vacc0x01234567), vacc0x89ABCDEF);
255 int8x16_t vout1x0123456789ABCDEF = vqmovn_high_s16(vqmovn_s16(vacc1x01234567), vacc1x89ABCDEF);
256 #else
257 int16x8_t vacc0x01234567 = vcombine_s16(vqmovn_s32(vacc0x0123), vqmovn_s32(vacc0x4567));
258 int16x8_t vacc0x89ABCDEF = vcombine_s16(vqmovn_s32(vacc0x89AB), vqmovn_s32(vacc0xCDEF));
259 int16x8_t vacc1x01234567 = vcombine_s16(vqmovn_s32(vacc1x0123), vqmovn_s32(vacc1x4567));
260 int16x8_t vacc1x89ABCDEF = vcombine_s16(vqmovn_s32(vacc1x89AB), vqmovn_s32(vacc1xCDEF));
261
262 vacc0x01234567 = vqaddq_s16(vacc0x01234567, voutput_zero_point);
263 vacc0x89ABCDEF = vqaddq_s16(vacc0x89ABCDEF, voutput_zero_point);
264 vacc1x01234567 = vqaddq_s16(vacc1x01234567, voutput_zero_point);
265 vacc1x89ABCDEF = vqaddq_s16(vacc1x89ABCDEF, voutput_zero_point);
266
267 int8x16_t vout0x0123456789ABCDEF = vcombine_s8(vqmovn_s16(vacc0x01234567), vqmovn_s16(vacc0x89ABCDEF));
268 int8x16_t vout1x0123456789ABCDEF = vcombine_s8(vqmovn_s16(vacc1x01234567), vqmovn_s16(vacc1x89ABCDEF));
269 #endif
270
271 const int8x16_t voutput_min = vld1q_dup_s8(¶ms->rndnu_neon.output_min);
272 vout0x0123456789ABCDEF = vmaxq_s8(vout0x0123456789ABCDEF, voutput_min);
273 vout1x0123456789ABCDEF = vmaxq_s8(vout1x0123456789ABCDEF, voutput_min);
274
275 const int8x16_t voutput_max = vld1q_dup_s8(¶ms->rndnu_neon.output_max);
276 vout0x0123456789ABCDEF = vminq_s8(vout0x0123456789ABCDEF, voutput_max);
277 vout1x0123456789ABCDEF = vminq_s8(vout1x0123456789ABCDEF, voutput_max);
278
279 if (nc >= 16) {
280 vst1q_s8(c1 + 0, vout1x0123456789ABCDEF);
281 vst1q_s8(c0 + 0, vout0x0123456789ABCDEF);
282
283 c1 = (int8_t*) ((uintptr_t) c1 + cn_stride);
284 c0 = (int8_t*) ((uintptr_t) c0 + cn_stride);
285
286 a = (const int8_t**restrict) ((uintptr_t) a - ks);
287
288 nc -= 16;
289 } else {
290 int8x16_t vout0x01234567_1x01234567 = vcombine_s8(vget_low_s8(vout0x0123456789ABCDEF), vget_low_s8(vout1x0123456789ABCDEF));
291 if (nc & 8) {
292 vst1_s8(c1, vget_high_s8(vout0x01234567_1x01234567)); c1 += 8;
293 vst1_s8(c0, vget_low_s8(vout0x01234567_1x01234567)); c0 += 8;
294 vout0x01234567_1x01234567 = vcombine_s8(vget_high_s8(vout0x0123456789ABCDEF), vget_high_s8(vout1x0123456789ABCDEF));
295 }
296 if (nc & 4) {
297 vst1q_lane_u32((void*) c1, vreinterpretq_u32_s8(vout0x01234567_1x01234567), 2); c1 += 4;
298 vst1q_lane_u32((void*) c0, vreinterpretq_u32_s8(vout0x01234567_1x01234567), 0); c0 += 4;
299 vout0x01234567_1x01234567 = vextq_s8(vout0x01234567_1x01234567, vout0x01234567_1x01234567, 4);
300 }
301 if (nc & 2) {
302 vst1q_lane_u16((void*) c1, vreinterpretq_u16_s8(vout0x01234567_1x01234567), 4); c1 += 2;
303 vst1q_lane_u16((void*) c0, vreinterpretq_u16_s8(vout0x01234567_1x01234567), 0); c0 += 2;
304 vout0x01234567_1x01234567 = vextq_s8(vout0x01234567_1x01234567, vout0x01234567_1x01234567, 2);
305 }
306 if (nc & 1) {
307 vst1q_lane_s8(c1, vout0x01234567_1x01234567, 8);
308 vst1q_lane_s8(c0, vout0x01234567_1x01234567, 0);
309 }
310
311 nc = 0;
312 }
313 } while (nc != 0);
314 }
315