1 // Auto-generated file. Do not edit!
2 // Template: src/qs8-igemm/c2-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_2x16c2s4__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_2x16c2s4__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 vacc0x0123 = vld1q_s32(w); w = (const int32_t*) w + 4;
52 int32x4_t vacc0x4567 = vld1q_s32(w); w = (const int32_t*) w + 4;
53 int32x4_t vacc0x89AB = vld1q_s32(w); w = (const int32_t*) w + 4;
54 int32x4_t vacc0xCDEF = vld1q_s32(w); w = (const int32_t*) w + 4;
55 int32x4_t vacc1x0123 = vacc0x0123;
56 int32x4_t vacc1x4567 = vacc0x4567;
57 int32x4_t vacc1x89AB = vacc0x89AB;
58 int32x4_t vacc1xCDEF = vacc0xCDEF;
59
60 size_t p = ks;
61 do {
62 const int8_t* restrict a0 = a[0];
63 if XNN_UNPREDICTABLE(a0 != zero) {
64 a0 = (const int8_t*) ((uintptr_t) a0 + a_offset);
65 }
66 const int8_t* restrict a1 = a[1];
67 if XNN_UNPREDICTABLE(a1 != zero) {
68 a1 = (const int8_t*) ((uintptr_t) a1 + a_offset);
69 }
70 a += 2;
71
72 size_t k = kc;
73 do {
74 int8x8_t va0x0 = vld1_s8(a0); a0 += 8;
75 int8x8_t va1x0 = vld1_s8(a1); a1 += 8;
76
77 const int8x8_t vb0123c0x0 = vld1_s8(w); w = (const int8_t*) w + 8;
78 const int8x8_t vb4567c0x0 = vld1_s8(w); w = (const int8_t*) w + 8;
79 const int8x8_t vb89ABc0x0 = vld1_s8(w); w = (const int8_t*) w + 8;
80 const int8x8_t vbCDEFc0x0 = vld1_s8(w); w = (const int8_t*) w + 8;
81 const int8x8_t vb0123c1x0 = vld1_s8(w); w = (const int8_t*) w + 8;
82 const int8x8_t vb4567c1x0 = vld1_s8(w); w = (const int8_t*) w + 8;
83 const int8x8_t vb89ABc1x0 = vld1_s8(w); w = (const int8_t*) w + 8;
84 const int8x8_t vbCDEFc1x0 = vld1_s8(w); w = (const int8_t*) w + 8;
85 const int8x8_t vb0123c2x0 = vld1_s8(w); w = (const int8_t*) w + 8;
86 const int8x8_t vb4567c2x0 = vld1_s8(w); w = (const int8_t*) w + 8;
87 const int8x8_t vb89ABc2x0 = vld1_s8(w); w = (const int8_t*) w + 8;
88 const int8x8_t vbCDEFc2x0 = vld1_s8(w); w = (const int8_t*) w + 8;
89 const int8x8_t vb0123c3x0 = vld1_s8(w); w = (const int8_t*) w + 8;
90 const int8x8_t vb4567c3x0 = vld1_s8(w); w = (const int8_t*) w + 8;
91 const int8x8_t vb89ABc3x0 = vld1_s8(w); w = (const int8_t*) w + 8;
92 const int8x8_t vbCDEFc3x0 = vld1_s8(w); w = (const int8_t*) w + 8;
93
94 int16x8_t vprod0x0123c0 = vmull_s8(vb0123c0x0, va0x0);
95 int16x8_t vprod1x0123c0 = vmull_s8(vb0123c0x0, va1x0);
96 vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c0);
97 vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c0);
98 int16x8_t vprod0x4567c0 = vmull_s8(vb4567c0x0, va0x0);
99 int16x8_t vprod1x4567c0 = vmull_s8(vb4567c0x0, va1x0);
100 vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c0);
101 vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c0);
102 int16x8_t vprod0x89ABc0 = vmull_s8(vb89ABc0x0, va0x0);
103 int16x8_t vprod1x89ABc0 = vmull_s8(vb89ABc0x0, va1x0);
104 vacc0x89AB = vpadalq_s16(vacc0x89AB, vprod0x89ABc0);
105 vacc1x89AB = vpadalq_s16(vacc1x89AB, vprod1x89ABc0);
106 int16x8_t vprod0xCDEFc0 = vmull_s8(vbCDEFc0x0, va0x0);
107 int16x8_t vprod1xCDEFc0 = vmull_s8(vbCDEFc0x0, va1x0);
108 vacc0xCDEF = vpadalq_s16(vacc0xCDEF, vprod0xCDEFc0);
109 vacc1xCDEF = vpadalq_s16(vacc1xCDEF, vprod1xCDEFc0);
110 va0x0 = vext_s8(va0x0, va0x0, 2);
111 va1x0 = vext_s8(va1x0, va1x0, 2);
112 int16x8_t vprod0x0123c1 = vmull_s8(vb0123c1x0, va0x0);
113 int16x8_t vprod1x0123c1 = vmull_s8(vb0123c1x0, va1x0);
114 vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c1);
115 vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c1);
116 int16x8_t vprod0x4567c1 = vmull_s8(vb4567c1x0, va0x0);
117 int16x8_t vprod1x4567c1 = vmull_s8(vb4567c1x0, va1x0);
118 vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c1);
119 vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c1);
120 int16x8_t vprod0x89ABc1 = vmull_s8(vb89ABc1x0, va0x0);
121 int16x8_t vprod1x89ABc1 = vmull_s8(vb89ABc1x0, va1x0);
122 vacc0x89AB = vpadalq_s16(vacc0x89AB, vprod0x89ABc1);
123 vacc1x89AB = vpadalq_s16(vacc1x89AB, vprod1x89ABc1);
124 int16x8_t vprod0xCDEFc1 = vmull_s8(vbCDEFc1x0, va0x0);
125 int16x8_t vprod1xCDEFc1 = vmull_s8(vbCDEFc1x0, va1x0);
126 vacc0xCDEF = vpadalq_s16(vacc0xCDEF, vprod0xCDEFc1);
127 vacc1xCDEF = vpadalq_s16(vacc1xCDEF, vprod1xCDEFc1);
128 va0x0 = vext_s8(va0x0, va0x0, 2);
129 va1x0 = vext_s8(va1x0, va1x0, 2);
130 int16x8_t vprod0x0123c2 = vmull_s8(vb0123c2x0, va0x0);
131 int16x8_t vprod1x0123c2 = vmull_s8(vb0123c2x0, va1x0);
132 vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c2);
133 vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c2);
134 int16x8_t vprod0x4567c2 = vmull_s8(vb4567c2x0, va0x0);
135 int16x8_t vprod1x4567c2 = vmull_s8(vb4567c2x0, va1x0);
136 vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c2);
137 vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c2);
138 int16x8_t vprod0x89ABc2 = vmull_s8(vb89ABc2x0, va0x0);
139 int16x8_t vprod1x89ABc2 = vmull_s8(vb89ABc2x0, va1x0);
140 vacc0x89AB = vpadalq_s16(vacc0x89AB, vprod0x89ABc2);
141 vacc1x89AB = vpadalq_s16(vacc1x89AB, vprod1x89ABc2);
142 int16x8_t vprod0xCDEFc2 = vmull_s8(vbCDEFc2x0, va0x0);
143 int16x8_t vprod1xCDEFc2 = vmull_s8(vbCDEFc2x0, va1x0);
144 vacc0xCDEF = vpadalq_s16(vacc0xCDEF, vprod0xCDEFc2);
145 vacc1xCDEF = vpadalq_s16(vacc1xCDEF, vprod1xCDEFc2);
146 va0x0 = vext_s8(va0x0, va0x0, 2);
147 va1x0 = vext_s8(va1x0, va1x0, 2);
148 int16x8_t vprod0x0123c3 = vmull_s8(vb0123c3x0, va0x0);
149 int16x8_t vprod1x0123c3 = vmull_s8(vb0123c3x0, va1x0);
150 vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c3);
151 vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c3);
152 int16x8_t vprod0x4567c3 = vmull_s8(vb4567c3x0, va0x0);
153 int16x8_t vprod1x4567c3 = vmull_s8(vb4567c3x0, va1x0);
154 vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c3);
155 vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c3);
156 int16x8_t vprod0x89ABc3 = vmull_s8(vb89ABc3x0, va0x0);
157 int16x8_t vprod1x89ABc3 = vmull_s8(vb89ABc3x0, va1x0);
158 vacc0x89AB = vpadalq_s16(vacc0x89AB, vprod0x89ABc3);
159 vacc1x89AB = vpadalq_s16(vacc1x89AB, vprod1x89ABc3);
160 int16x8_t vprod0xCDEFc3 = vmull_s8(vbCDEFc3x0, va0x0);
161 int16x8_t vprod1xCDEFc3 = vmull_s8(vbCDEFc3x0, va1x0);
162 vacc0xCDEF = vpadalq_s16(vacc0xCDEF, vprod0xCDEFc3);
163 vacc1xCDEF = vpadalq_s16(vacc1xCDEF, vprod1xCDEFc3);
164
165 k -= 8 * sizeof(int8_t);
166 } while (k != 0);
167
168 p -= 2 * sizeof(void*);
169 } while (p != 0);
170
171 const int32x4_t vright_pre_shift = vld1q_dup_s32(¶ms->rndnu_neon.right_pre_shift);
172 const int32x4_t vmultiplier = vld1q_dup_s32(¶ms->rndnu_neon.multiplier);
173 const int32x4_t vright_post_shift = vld1q_dup_s32(¶ms->rndnu_neon.right_post_shift);
174
175 vacc0x0123 = vqshlq_s32(vacc0x0123, vright_pre_shift);
176 vacc0x4567 = vqshlq_s32(vacc0x4567, vright_pre_shift);
177 vacc0x89AB = vqshlq_s32(vacc0x89AB, vright_pre_shift);
178 vacc0xCDEF = vqshlq_s32(vacc0xCDEF, vright_pre_shift);
179 vacc1x0123 = vqshlq_s32(vacc1x0123, vright_pre_shift);
180 vacc1x4567 = vqshlq_s32(vacc1x4567, vright_pre_shift);
181 vacc1x89AB = vqshlq_s32(vacc1x89AB, vright_pre_shift);
182 vacc1xCDEF = vqshlq_s32(vacc1xCDEF, vright_pre_shift);
183
184 vacc0x0123 = vqdmulhq_s32(vacc0x0123, vmultiplier);
185 vacc0x4567 = vqdmulhq_s32(vacc0x4567, vmultiplier);
186 vacc0x89AB = vqdmulhq_s32(vacc0x89AB, vmultiplier);
187 vacc0xCDEF = vqdmulhq_s32(vacc0xCDEF, vmultiplier);
188 vacc1x0123 = vqdmulhq_s32(vacc1x0123, vmultiplier);
189 vacc1x4567 = vqdmulhq_s32(vacc1x4567, vmultiplier);
190 vacc1x89AB = vqdmulhq_s32(vacc1x89AB, vmultiplier);
191 vacc1xCDEF = vqdmulhq_s32(vacc1xCDEF, vmultiplier);
192
193 vacc0x0123 = vrshlq_s32(vacc0x0123, vright_post_shift);
194 vacc0x4567 = vrshlq_s32(vacc0x4567, vright_post_shift);
195 vacc0x89AB = vrshlq_s32(vacc0x89AB, vright_post_shift);
196 vacc0xCDEF = vrshlq_s32(vacc0xCDEF, vright_post_shift);
197 vacc1x0123 = vrshlq_s32(vacc1x0123, vright_post_shift);
198 vacc1x4567 = vrshlq_s32(vacc1x4567, vright_post_shift);
199 vacc1x89AB = vrshlq_s32(vacc1x89AB, vright_post_shift);
200 vacc1xCDEF = vrshlq_s32(vacc1xCDEF, vright_post_shift);
201
202 const int16x8_t voutput_zero_point = vld1q_dup_s16(¶ms->rndnu_neon.output_zero_point);
203 #if XNN_ARCH_ARM64
204 int16x8_t vacc0x01234567 = vqmovn_high_s32(vqmovn_s32(vacc0x0123), vacc0x4567);
205 int16x8_t vacc0x89ABCDEF = vqmovn_high_s32(vqmovn_s32(vacc0x89AB), vacc0xCDEF);
206 int16x8_t vacc1x01234567 = vqmovn_high_s32(vqmovn_s32(vacc1x0123), vacc1x4567);
207 int16x8_t vacc1x89ABCDEF = vqmovn_high_s32(vqmovn_s32(vacc1x89AB), vacc1xCDEF);
208
209 vacc0x01234567 = vqaddq_s16(vacc0x01234567, voutput_zero_point);
210 vacc0x89ABCDEF = vqaddq_s16(vacc0x89ABCDEF, voutput_zero_point);
211 vacc1x01234567 = vqaddq_s16(vacc1x01234567, voutput_zero_point);
212 vacc1x89ABCDEF = vqaddq_s16(vacc1x89ABCDEF, voutput_zero_point);
213
214 int8x16_t vout0x0123456789ABCDEF = vqmovn_high_s16(vqmovn_s16(vacc0x01234567), vacc0x89ABCDEF);
215 int8x16_t vout1x0123456789ABCDEF = vqmovn_high_s16(vqmovn_s16(vacc1x01234567), vacc1x89ABCDEF);
216 #else
217 int16x8_t vacc0x01234567 = vcombine_s16(vqmovn_s32(vacc0x0123), vqmovn_s32(vacc0x4567));
218 int16x8_t vacc0x89ABCDEF = vcombine_s16(vqmovn_s32(vacc0x89AB), vqmovn_s32(vacc0xCDEF));
219 int16x8_t vacc1x01234567 = vcombine_s16(vqmovn_s32(vacc1x0123), vqmovn_s32(vacc1x4567));
220 int16x8_t vacc1x89ABCDEF = vcombine_s16(vqmovn_s32(vacc1x89AB), vqmovn_s32(vacc1xCDEF));
221
222 vacc0x01234567 = vqaddq_s16(vacc0x01234567, voutput_zero_point);
223 vacc0x89ABCDEF = vqaddq_s16(vacc0x89ABCDEF, voutput_zero_point);
224 vacc1x01234567 = vqaddq_s16(vacc1x01234567, voutput_zero_point);
225 vacc1x89ABCDEF = vqaddq_s16(vacc1x89ABCDEF, voutput_zero_point);
226
227 int8x16_t vout0x0123456789ABCDEF = vcombine_s8(vqmovn_s16(vacc0x01234567), vqmovn_s16(vacc0x89ABCDEF));
228 int8x16_t vout1x0123456789ABCDEF = vcombine_s8(vqmovn_s16(vacc1x01234567), vqmovn_s16(vacc1x89ABCDEF));
229 #endif
230
231 const int8x16_t voutput_min = vld1q_dup_s8(¶ms->rndnu_neon.output_min);
232 vout0x0123456789ABCDEF = vmaxq_s8(vout0x0123456789ABCDEF, voutput_min);
233 vout1x0123456789ABCDEF = vmaxq_s8(vout1x0123456789ABCDEF, voutput_min);
234
235 const int8x16_t voutput_max = vld1q_dup_s8(¶ms->rndnu_neon.output_max);
236 vout0x0123456789ABCDEF = vminq_s8(vout0x0123456789ABCDEF, voutput_max);
237 vout1x0123456789ABCDEF = vminq_s8(vout1x0123456789ABCDEF, voutput_max);
238
239 if (nc >= 16) {
240 vst1q_s8(c1 + 0, vout1x0123456789ABCDEF);
241 vst1q_s8(c0 + 0, vout0x0123456789ABCDEF);
242
243 c1 = (int8_t*) ((uintptr_t) c1 + cn_stride);
244 c0 = (int8_t*) ((uintptr_t) c0 + cn_stride);
245
246 a = (const int8_t**restrict) ((uintptr_t) a - ks);
247
248 nc -= 16;
249 } else {
250 int8x16_t vout0x01234567_1x01234567 = vcombine_s8(vget_low_s8(vout0x0123456789ABCDEF), vget_low_s8(vout1x0123456789ABCDEF));
251 if (nc & 8) {
252 vst1_s8(c1, vget_high_s8(vout0x01234567_1x01234567)); c1 += 8;
253 vst1_s8(c0, vget_low_s8(vout0x01234567_1x01234567)); c0 += 8;
254 vout0x01234567_1x01234567 = vcombine_s8(vget_high_s8(vout0x0123456789ABCDEF), vget_high_s8(vout1x0123456789ABCDEF));
255 }
256 if (nc & 4) {
257 vst1q_lane_u32((void*) c1, vreinterpretq_u32_s8(vout0x01234567_1x01234567), 2); c1 += 4;
258 vst1q_lane_u32((void*) c0, vreinterpretq_u32_s8(vout0x01234567_1x01234567), 0); c0 += 4;
259 vout0x01234567_1x01234567 = vextq_s8(vout0x01234567_1x01234567, vout0x01234567_1x01234567, 4);
260 }
261 if (nc & 2) {
262 vst1q_lane_u16((void*) c1, vreinterpretq_u16_s8(vout0x01234567_1x01234567), 4); c1 += 2;
263 vst1q_lane_u16((void*) c0, vreinterpretq_u16_s8(vout0x01234567_1x01234567), 0); c0 += 2;
264 vout0x01234567_1x01234567 = vextq_s8(vout0x01234567_1x01234567, vout0x01234567_1x01234567, 2);
265 }
266 if (nc & 1) {
267 vst1q_lane_s8(c1, vout0x01234567_1x01234567, 8);
268 vst1q_lane_s8(c0, vout0x01234567_1x01234567, 0);
269 }
270
271 nc = 0;
272 }
273 } while (nc != 0);
274 }
275