1 // Auto-generated file. Do not edit!
2 // Template: src/qs8-gemm/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_gemm_minmax_rndnu_ukernel_2x16c4s2__neon_mull(size_t mr,size_t nc,size_t kc,const int8_t * restrict a,size_t a_stride,const void * restrict w,int8_t * restrict c,size_t cm_stride,size_t cn_stride,const union xnn_qs8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS (1)])18 void xnn_qs8_gemm_minmax_rndnu_ukernel_2x16c4s2__neon_mull(
19 size_t mr,
20 size_t nc,
21 size_t kc,
22 const int8_t* restrict a,
23 size_t a_stride,
24 const void* restrict w,
25 int8_t* restrict c,
26 size_t cm_stride,
27 size_t cn_stride,
28 const union xnn_qs8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
29 {
30 assert(mr != 0);
31 assert(mr <= 2);
32 assert(nc != 0);
33 assert(kc != 0);
34 assert(kc % sizeof(int8_t) == 0);
35 assert(a != NULL);
36 assert(w != NULL);
37 assert(c != NULL);
38
39 const int8_t* a0 = a;
40 int8_t* c0 = c;
41 const int8_t* a1 = (const int8_t*) ((uintptr_t) a0 + a_stride);
42 int8_t* c1 = (int8_t*) ((uintptr_t) c0 + cm_stride);
43 if XNN_UNPREDICTABLE(mr != 2) {
44 a1 = a0;
45 c1 = c0;
46 }
47
48 kc = round_up_po2(kc, 8 * sizeof(int8_t));
49 do {
50 int32x4_t vacc0x01 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const int32_t*) w + 2;
51 int32x4_t vacc0x23 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const int32_t*) w + 2;
52 int32x4_t vacc0x45 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const int32_t*) w + 2;
53 int32x4_t vacc0x67 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const int32_t*) w + 2;
54 int32x4_t vacc0x89 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const int32_t*) w + 2;
55 int32x4_t vacc0xAB = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const int32_t*) w + 2;
56 int32x4_t vacc0xCD = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const int32_t*) w + 2;
57 int32x4_t vacc0xEF = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const int32_t*) w + 2;
58 int32x4_t vacc1x01 = vacc0x01;
59 int32x4_t vacc1x23 = vacc0x23;
60 int32x4_t vacc1x45 = vacc0x45;
61 int32x4_t vacc1x67 = vacc0x67;
62 int32x4_t vacc1x89 = vacc0x89;
63 int32x4_t vacc1xAB = vacc0xAB;
64 int32x4_t vacc1xCD = vacc0xCD;
65 int32x4_t vacc1xEF = vacc0xEF;
66
67 size_t k = kc;
68 do {
69 int8x8_t va0x0 = vld1_s8(a0); a0 += 8;
70 int8x8_t va1x0 = vld1_s8(a1); a1 += 8;
71
72 const int8x8_t vb01c0x0 = vld1_s8(w); w = (const int8_t*) w + 8;
73 const int8x8_t vb23c0x0 = vld1_s8(w); w = (const int8_t*) w + 8;
74 const int8x8_t vb45c0x0 = vld1_s8(w); w = (const int8_t*) w + 8;
75 const int8x8_t vb67c0x0 = vld1_s8(w); w = (const int8_t*) w + 8;
76 const int8x8_t vb89c0x0 = vld1_s8(w); w = (const int8_t*) w + 8;
77 const int8x8_t vbABc0x0 = vld1_s8(w); w = (const int8_t*) w + 8;
78 const int8x8_t vbCDc0x0 = vld1_s8(w); w = (const int8_t*) w + 8;
79 const int8x8_t vbEFc0x0 = vld1_s8(w); w = (const int8_t*) w + 8;
80 const int8x8_t vb01c1x0 = vld1_s8(w); w = (const int8_t*) w + 8;
81 const int8x8_t vb23c1x0 = vld1_s8(w); w = (const int8_t*) w + 8;
82 const int8x8_t vb45c1x0 = vld1_s8(w); w = (const int8_t*) w + 8;
83 const int8x8_t vb67c1x0 = vld1_s8(w); w = (const int8_t*) w + 8;
84 const int8x8_t vb89c1x0 = vld1_s8(w); w = (const int8_t*) w + 8;
85 const int8x8_t vbABc1x0 = vld1_s8(w); w = (const int8_t*) w + 8;
86 const int8x8_t vbCDc1x0 = vld1_s8(w); w = (const int8_t*) w + 8;
87 const int8x8_t vbEFc1x0 = vld1_s8(w); w = (const int8_t*) w + 8;
88
89 int16x8_t vprod0x01c0 = vmull_s8(vb01c0x0, va0x0);
90 int16x8_t vprod1x01c0 = vmull_s8(vb01c0x0, va1x0);
91 vacc0x01 = vpadalq_s16(vacc0x01, vprod0x01c0);
92 vacc1x01 = vpadalq_s16(vacc1x01, vprod1x01c0);
93 int16x8_t vprod0x23c0 = vmull_s8(vb23c0x0, va0x0);
94 int16x8_t vprod1x23c0 = vmull_s8(vb23c0x0, va1x0);
95 vacc0x23 = vpadalq_s16(vacc0x23, vprod0x23c0);
96 vacc1x23 = vpadalq_s16(vacc1x23, vprod1x23c0);
97 int16x8_t vprod0x45c0 = vmull_s8(vb45c0x0, va0x0);
98 int16x8_t vprod1x45c0 = vmull_s8(vb45c0x0, va1x0);
99 vacc0x45 = vpadalq_s16(vacc0x45, vprod0x45c0);
100 vacc1x45 = vpadalq_s16(vacc1x45, vprod1x45c0);
101 int16x8_t vprod0x67c0 = vmull_s8(vb67c0x0, va0x0);
102 int16x8_t vprod1x67c0 = vmull_s8(vb67c0x0, va1x0);
103 vacc0x67 = vpadalq_s16(vacc0x67, vprod0x67c0);
104 vacc1x67 = vpadalq_s16(vacc1x67, vprod1x67c0);
105 int16x8_t vprod0x89c0 = vmull_s8(vb89c0x0, va0x0);
106 int16x8_t vprod1x89c0 = vmull_s8(vb89c0x0, va1x0);
107 vacc0x89 = vpadalq_s16(vacc0x89, vprod0x89c0);
108 vacc1x89 = vpadalq_s16(vacc1x89, vprod1x89c0);
109 int16x8_t vprod0xABc0 = vmull_s8(vbABc0x0, va0x0);
110 int16x8_t vprod1xABc0 = vmull_s8(vbABc0x0, va1x0);
111 vacc0xAB = vpadalq_s16(vacc0xAB, vprod0xABc0);
112 vacc1xAB = vpadalq_s16(vacc1xAB, vprod1xABc0);
113 int16x8_t vprod0xCDc0 = vmull_s8(vbCDc0x0, va0x0);
114 int16x8_t vprod1xCDc0 = vmull_s8(vbCDc0x0, va1x0);
115 vacc0xCD = vpadalq_s16(vacc0xCD, vprod0xCDc0);
116 vacc1xCD = vpadalq_s16(vacc1xCD, vprod1xCDc0);
117 int16x8_t vprod0xEFc0 = vmull_s8(vbEFc0x0, va0x0);
118 int16x8_t vprod1xEFc0 = vmull_s8(vbEFc0x0, va1x0);
119 vacc0xEF = vpadalq_s16(vacc0xEF, vprod0xEFc0);
120 vacc1xEF = vpadalq_s16(vacc1xEF, vprod1xEFc0);
121 va0x0 = vext_s8(va0x0, va0x0, 4);
122 va1x0 = vext_s8(va1x0, va1x0, 4);
123 int16x8_t vprod0x01c1 = vmull_s8(vb01c1x0, va0x0);
124 int16x8_t vprod1x01c1 = vmull_s8(vb01c1x0, va1x0);
125 vacc0x01 = vpadalq_s16(vacc0x01, vprod0x01c1);
126 vacc1x01 = vpadalq_s16(vacc1x01, vprod1x01c1);
127 int16x8_t vprod0x23c1 = vmull_s8(vb23c1x0, va0x0);
128 int16x8_t vprod1x23c1 = vmull_s8(vb23c1x0, va1x0);
129 vacc0x23 = vpadalq_s16(vacc0x23, vprod0x23c1);
130 vacc1x23 = vpadalq_s16(vacc1x23, vprod1x23c1);
131 int16x8_t vprod0x45c1 = vmull_s8(vb45c1x0, va0x0);
132 int16x8_t vprod1x45c1 = vmull_s8(vb45c1x0, va1x0);
133 vacc0x45 = vpadalq_s16(vacc0x45, vprod0x45c1);
134 vacc1x45 = vpadalq_s16(vacc1x45, vprod1x45c1);
135 int16x8_t vprod0x67c1 = vmull_s8(vb67c1x0, va0x0);
136 int16x8_t vprod1x67c1 = vmull_s8(vb67c1x0, va1x0);
137 vacc0x67 = vpadalq_s16(vacc0x67, vprod0x67c1);
138 vacc1x67 = vpadalq_s16(vacc1x67, vprod1x67c1);
139 int16x8_t vprod0x89c1 = vmull_s8(vb89c1x0, va0x0);
140 int16x8_t vprod1x89c1 = vmull_s8(vb89c1x0, va1x0);
141 vacc0x89 = vpadalq_s16(vacc0x89, vprod0x89c1);
142 vacc1x89 = vpadalq_s16(vacc1x89, vprod1x89c1);
143 int16x8_t vprod0xABc1 = vmull_s8(vbABc1x0, va0x0);
144 int16x8_t vprod1xABc1 = vmull_s8(vbABc1x0, va1x0);
145 vacc0xAB = vpadalq_s16(vacc0xAB, vprod0xABc1);
146 vacc1xAB = vpadalq_s16(vacc1xAB, vprod1xABc1);
147 int16x8_t vprod0xCDc1 = vmull_s8(vbCDc1x0, va0x0);
148 int16x8_t vprod1xCDc1 = vmull_s8(vbCDc1x0, va1x0);
149 vacc0xCD = vpadalq_s16(vacc0xCD, vprod0xCDc1);
150 vacc1xCD = vpadalq_s16(vacc1xCD, vprod1xCDc1);
151 int16x8_t vprod0xEFc1 = vmull_s8(vbEFc1x0, va0x0);
152 int16x8_t vprod1xEFc1 = vmull_s8(vbEFc1x0, va1x0);
153 vacc0xEF = vpadalq_s16(vacc0xEF, vprod0xEFc1);
154 vacc1xEF = vpadalq_s16(vacc1xEF, vprod1xEFc1);
155
156 k -= 8 * sizeof(int8_t);
157 } while (k != 0);
158
159 #if XNN_ARCH_ARM64
160 int32x4_t vacc0x0123 = vpaddq_s32(vacc0x01, vacc0x23);
161 int32x4_t vacc0x4567 = vpaddq_s32(vacc0x45, vacc0x67);
162 int32x4_t vacc0x89AB = vpaddq_s32(vacc0x89, vacc0xAB);
163 int32x4_t vacc0xCDEF = vpaddq_s32(vacc0xCD, vacc0xEF);
164 int32x4_t vacc1x0123 = vpaddq_s32(vacc1x01, vacc1x23);
165 int32x4_t vacc1x4567 = vpaddq_s32(vacc1x45, vacc1x67);
166 int32x4_t vacc1x89AB = vpaddq_s32(vacc1x89, vacc1xAB);
167 int32x4_t vacc1xCDEF = vpaddq_s32(vacc1xCD, vacc1xEF);
168 #else
169 const int32x2_t vsum0x01 = vpadd_s32(vget_low_s32(vacc0x01), vget_high_s32(vacc0x01));
170 const int32x2_t vsum0x23 = vpadd_s32(vget_low_s32(vacc0x23), vget_high_s32(vacc0x23));
171 int32x4_t vacc0x0123 = vcombine_s32(vsum0x01, vsum0x23);
172 const int32x2_t vsum0x45 = vpadd_s32(vget_low_s32(vacc0x45), vget_high_s32(vacc0x45));
173 const int32x2_t vsum0x67 = vpadd_s32(vget_low_s32(vacc0x67), vget_high_s32(vacc0x67));
174 int32x4_t vacc0x4567 = vcombine_s32(vsum0x45, vsum0x67);
175 const int32x2_t vsum0x89 = vpadd_s32(vget_low_s32(vacc0x89), vget_high_s32(vacc0x89));
176 const int32x2_t vsum0xAB = vpadd_s32(vget_low_s32(vacc0xAB), vget_high_s32(vacc0xAB));
177 int32x4_t vacc0x89AB = vcombine_s32(vsum0x89, vsum0xAB);
178 const int32x2_t vsum0xCD = vpadd_s32(vget_low_s32(vacc0xCD), vget_high_s32(vacc0xCD));
179 const int32x2_t vsum0xEF = vpadd_s32(vget_low_s32(vacc0xEF), vget_high_s32(vacc0xEF));
180 int32x4_t vacc0xCDEF = vcombine_s32(vsum0xCD, vsum0xEF);
181 const int32x2_t vsum1x01 = vpadd_s32(vget_low_s32(vacc1x01), vget_high_s32(vacc1x01));
182 const int32x2_t vsum1x23 = vpadd_s32(vget_low_s32(vacc1x23), vget_high_s32(vacc1x23));
183 int32x4_t vacc1x0123 = vcombine_s32(vsum1x01, vsum1x23);
184 const int32x2_t vsum1x45 = vpadd_s32(vget_low_s32(vacc1x45), vget_high_s32(vacc1x45));
185 const int32x2_t vsum1x67 = vpadd_s32(vget_low_s32(vacc1x67), vget_high_s32(vacc1x67));
186 int32x4_t vacc1x4567 = vcombine_s32(vsum1x45, vsum1x67);
187 const int32x2_t vsum1x89 = vpadd_s32(vget_low_s32(vacc1x89), vget_high_s32(vacc1x89));
188 const int32x2_t vsum1xAB = vpadd_s32(vget_low_s32(vacc1xAB), vget_high_s32(vacc1xAB));
189 int32x4_t vacc1x89AB = vcombine_s32(vsum1x89, vsum1xAB);
190 const int32x2_t vsum1xCD = vpadd_s32(vget_low_s32(vacc1xCD), vget_high_s32(vacc1xCD));
191 const int32x2_t vsum1xEF = vpadd_s32(vget_low_s32(vacc1xEF), vget_high_s32(vacc1xEF));
192 int32x4_t vacc1xCDEF = vcombine_s32(vsum1xCD, vsum1xEF);
193 #endif
194
195 const int32x4_t vright_pre_shift = vld1q_dup_s32(¶ms->rndnu_neon.right_pre_shift);
196 const int32x4_t vmultiplier = vld1q_dup_s32(¶ms->rndnu_neon.multiplier);
197 const int32x4_t vright_post_shift = vld1q_dup_s32(¶ms->rndnu_neon.right_post_shift);
198
199 vacc0x0123 = vqshlq_s32(vacc0x0123, vright_pre_shift);
200 vacc0x4567 = vqshlq_s32(vacc0x4567, vright_pre_shift);
201 vacc0x89AB = vqshlq_s32(vacc0x89AB, vright_pre_shift);
202 vacc0xCDEF = vqshlq_s32(vacc0xCDEF, vright_pre_shift);
203 vacc1x0123 = vqshlq_s32(vacc1x0123, vright_pre_shift);
204 vacc1x4567 = vqshlq_s32(vacc1x4567, vright_pre_shift);
205 vacc1x89AB = vqshlq_s32(vacc1x89AB, vright_pre_shift);
206 vacc1xCDEF = vqshlq_s32(vacc1xCDEF, vright_pre_shift);
207
208 vacc0x0123 = vqdmulhq_s32(vacc0x0123, vmultiplier);
209 vacc0x4567 = vqdmulhq_s32(vacc0x4567, vmultiplier);
210 vacc0x89AB = vqdmulhq_s32(vacc0x89AB, vmultiplier);
211 vacc0xCDEF = vqdmulhq_s32(vacc0xCDEF, vmultiplier);
212 vacc1x0123 = vqdmulhq_s32(vacc1x0123, vmultiplier);
213 vacc1x4567 = vqdmulhq_s32(vacc1x4567, vmultiplier);
214 vacc1x89AB = vqdmulhq_s32(vacc1x89AB, vmultiplier);
215 vacc1xCDEF = vqdmulhq_s32(vacc1xCDEF, vmultiplier);
216
217 vacc0x0123 = vrshlq_s32(vacc0x0123, vright_post_shift);
218 vacc0x4567 = vrshlq_s32(vacc0x4567, vright_post_shift);
219 vacc0x89AB = vrshlq_s32(vacc0x89AB, vright_post_shift);
220 vacc0xCDEF = vrshlq_s32(vacc0xCDEF, vright_post_shift);
221 vacc1x0123 = vrshlq_s32(vacc1x0123, vright_post_shift);
222 vacc1x4567 = vrshlq_s32(vacc1x4567, vright_post_shift);
223 vacc1x89AB = vrshlq_s32(vacc1x89AB, vright_post_shift);
224 vacc1xCDEF = vrshlq_s32(vacc1xCDEF, vright_post_shift);
225
226 const int16x8_t voutput_zero_point = vld1q_dup_s16(¶ms->rndnu_neon.output_zero_point);
227 #if XNN_ARCH_ARM64
228 int16x8_t vacc0x01234567 = vqmovn_high_s32(vqmovn_s32(vacc0x0123), vacc0x4567);
229 int16x8_t vacc0x89ABCDEF = vqmovn_high_s32(vqmovn_s32(vacc0x89AB), vacc0xCDEF);
230 int16x8_t vacc1x01234567 = vqmovn_high_s32(vqmovn_s32(vacc1x0123), vacc1x4567);
231 int16x8_t vacc1x89ABCDEF = vqmovn_high_s32(vqmovn_s32(vacc1x89AB), vacc1xCDEF);
232
233 vacc0x01234567 = vqaddq_s16(vacc0x01234567, voutput_zero_point);
234 vacc0x89ABCDEF = vqaddq_s16(vacc0x89ABCDEF, voutput_zero_point);
235 vacc1x01234567 = vqaddq_s16(vacc1x01234567, voutput_zero_point);
236 vacc1x89ABCDEF = vqaddq_s16(vacc1x89ABCDEF, voutput_zero_point);
237
238 int8x16_t vout0x0123456789ABCDEF = vqmovn_high_s16(vqmovn_s16(vacc0x01234567), vacc0x89ABCDEF);
239 int8x16_t vout1x0123456789ABCDEF = vqmovn_high_s16(vqmovn_s16(vacc1x01234567), vacc1x89ABCDEF);
240 #else
241 int16x8_t vacc0x01234567 = vcombine_s16(vqmovn_s32(vacc0x0123), vqmovn_s32(vacc0x4567));
242 int16x8_t vacc0x89ABCDEF = vcombine_s16(vqmovn_s32(vacc0x89AB), vqmovn_s32(vacc0xCDEF));
243 int16x8_t vacc1x01234567 = vcombine_s16(vqmovn_s32(vacc1x0123), vqmovn_s32(vacc1x4567));
244 int16x8_t vacc1x89ABCDEF = vcombine_s16(vqmovn_s32(vacc1x89AB), vqmovn_s32(vacc1xCDEF));
245
246 vacc0x01234567 = vqaddq_s16(vacc0x01234567, voutput_zero_point);
247 vacc0x89ABCDEF = vqaddq_s16(vacc0x89ABCDEF, voutput_zero_point);
248 vacc1x01234567 = vqaddq_s16(vacc1x01234567, voutput_zero_point);
249 vacc1x89ABCDEF = vqaddq_s16(vacc1x89ABCDEF, voutput_zero_point);
250
251 int8x16_t vout0x0123456789ABCDEF = vcombine_s8(vqmovn_s16(vacc0x01234567), vqmovn_s16(vacc0x89ABCDEF));
252 int8x16_t vout1x0123456789ABCDEF = vcombine_s8(vqmovn_s16(vacc1x01234567), vqmovn_s16(vacc1x89ABCDEF));
253 #endif
254
255 const int8x16_t voutput_min = vld1q_dup_s8(¶ms->rndnu_neon.output_min);
256 vout0x0123456789ABCDEF = vmaxq_s8(vout0x0123456789ABCDEF, voutput_min);
257 vout1x0123456789ABCDEF = vmaxq_s8(vout1x0123456789ABCDEF, voutput_min);
258
259 const int8x16_t voutput_max = vld1q_dup_s8(¶ms->rndnu_neon.output_max);
260 vout0x0123456789ABCDEF = vminq_s8(vout0x0123456789ABCDEF, voutput_max);
261 vout1x0123456789ABCDEF = vminq_s8(vout1x0123456789ABCDEF, voutput_max);
262
263 if (nc >= 16) {
264 vst1q_s8(c0 + 0, vout0x0123456789ABCDEF);
265 vst1q_s8(c1 + 0, vout1x0123456789ABCDEF);
266
267 c0 = (int8_t*) ((uintptr_t) c0 + cn_stride);
268 c1 = (int8_t*) ((uintptr_t) c1 + cn_stride);
269
270 a0 = (const int8_t*) ((uintptr_t) a0 - kc);
271 a1 = (const int8_t*) ((uintptr_t) a1 - kc);
272
273 nc -= 16;
274 } else {
275 // Final case where not all of the 16 columns fit in the destination.
276 int8x16_t vout0x01234567_1x01234567 = vcombine_s8(vget_low_s8(vout0x0123456789ABCDEF), vget_low_s8(vout1x0123456789ABCDEF));
277 if (nc & 8) {
278 vst1_s8(c0, vget_low_s8(vout0x01234567_1x01234567)); c0 += 8;
279 vst1_s8(c1, vget_high_s8(vout0x01234567_1x01234567)); c1 += 8;
280 vout0x01234567_1x01234567 = vcombine_s8(vget_high_s8(vout0x0123456789ABCDEF), vget_high_s8(vout1x0123456789ABCDEF));
281 }
282 if (nc & 4) {
283 vst1q_lane_u32((void*) c0, vreinterpretq_u32_s8(vout0x01234567_1x01234567), 0); c0 += 4;
284 vst1q_lane_u32((void*) c1, vreinterpretq_u32_s8(vout0x01234567_1x01234567), 2); c1 += 4;
285 vout0x01234567_1x01234567 = vextq_s8(vout0x01234567_1x01234567, vout0x01234567_1x01234567, 4);
286 }
287 if (nc & 2) {
288 vst1q_lane_u16((void*) c0, vreinterpretq_u16_s8(vout0x01234567_1x01234567), 0); c0 += 2;
289 vst1q_lane_u16((void*) c1, vreinterpretq_u16_s8(vout0x01234567_1x01234567), 4); c1 += 2;
290 vout0x01234567_1x01234567 = vextq_s8(vout0x01234567_1x01234567, vout0x01234567_1x01234567, 2);
291 }
292 if (nc & 1) {
293 vst1q_lane_s8(c0, vout0x01234567_1x01234567, 0);
294 vst1q_lane_s8(c1, vout0x01234567_1x01234567, 8);
295 }
296
297 nc = 0;
298 }
299 } while (nc != 0);
300 }
301