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_1x16c2s4__neon_mlal(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_1x16c2s4__neon_mlal(
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 <= 1);
34 assert(nc != 0);
35 assert(kc != 0);
36 assert(ks != 0);
37 assert(ks % (1 * 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
45 kc = round_up_po2(kc, 8 * sizeof(int8_t));
46 do {
47 int32x4_t vacc0x0123 = vld1q_s32(w); w = (const int32_t*) w + 4;
48 int32x4_t vacc0x4567 = vld1q_s32(w); w = (const int32_t*) w + 4;
49 int32x4_t vacc0x89AB = vld1q_s32(w); w = (const int32_t*) w + 4;
50 int32x4_t vacc0xCDEF = vld1q_s32(w); w = (const int32_t*) w + 4;
51
52 size_t p = ks;
53 do {
54 const int8_t* restrict a0 = a[0];
55 if XNN_UNPREDICTABLE(a0 != zero) {
56 a0 = (const int8_t*) ((uintptr_t) a0 + a_offset);
57 }
58 a += 1;
59
60 size_t k = kc;
61 while (k >= 16 * sizeof(int8_t)) {
62 int8x8_t va0x0 = vld1_s8(a0); a0 += 8;
63 int8x8_t va0x1 = vld1_s8(a0); a0 += 8;
64
65 const int8x8_t vb0123c0x0 = vld1_s8(w); w = (const int8_t*) w + 8;
66 const int8x8_t vb4567c0x0 = vld1_s8(w); w = (const int8_t*) w + 8;
67 const int8x8_t vb89ABc0x0 = vld1_s8(w); w = (const int8_t*) w + 8;
68 const int8x8_t vbCDEFc0x0 = vld1_s8(w); w = (const int8_t*) w + 8;
69 const int8x8_t vb0123c1x0 = vld1_s8(w); w = (const int8_t*) w + 8;
70 const int8x8_t vb4567c1x0 = vld1_s8(w); w = (const int8_t*) w + 8;
71 const int8x8_t vb89ABc1x0 = vld1_s8(w); w = (const int8_t*) w + 8;
72 const int8x8_t vbCDEFc1x0 = vld1_s8(w); w = (const int8_t*) w + 8;
73 const int8x8_t vb0123c2x0 = vld1_s8(w); w = (const int8_t*) w + 8;
74 const int8x8_t vb4567c2x0 = vld1_s8(w); w = (const int8_t*) w + 8;
75 const int8x8_t vb89ABc2x0 = vld1_s8(w); w = (const int8_t*) w + 8;
76 const int8x8_t vbCDEFc2x0 = vld1_s8(w); w = (const int8_t*) w + 8;
77 const int8x8_t vb0123c3x0 = vld1_s8(w); w = (const int8_t*) w + 8;
78 const int8x8_t vb4567c3x0 = vld1_s8(w); w = (const int8_t*) w + 8;
79 const int8x8_t vb89ABc3x0 = vld1_s8(w); w = (const int8_t*) w + 8;
80 const int8x8_t vbCDEFc3x0 = vld1_s8(w); w = (const int8_t*) w + 8;
81
82 int16x8_t vprod0x0123c0 = vmull_s8(vb0123c0x0, va0x0);
83 const int8x8_t vb0123c0x1 = vld1_s8(w); w = (const int8_t*) w + 8;
84 vprod0x0123c0 = vmlal_s8(vprod0x0123c0, vb0123c0x1, va0x1);
85 vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c0);
86 int16x8_t vprod0x4567c0 = vmull_s8(vb4567c0x0, va0x0);
87 const int8x8_t vb4567c0x1 = vld1_s8(w); w = (const int8_t*) w + 8;
88 vprod0x4567c0 = vmlal_s8(vprod0x4567c0, vb4567c0x1, va0x1);
89 vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c0);
90 int16x8_t vprod0x89ABc0 = vmull_s8(vb89ABc0x0, va0x0);
91 const int8x8_t vb89ABc0x1 = vld1_s8(w); w = (const int8_t*) w + 8;
92 vprod0x89ABc0 = vmlal_s8(vprod0x89ABc0, vb89ABc0x1, va0x1);
93 vacc0x89AB = vpadalq_s16(vacc0x89AB, vprod0x89ABc0);
94 int16x8_t vprod0xCDEFc0 = vmull_s8(vbCDEFc0x0, va0x0);
95 const int8x8_t vbCDEFc0x1 = vld1_s8(w); w = (const int8_t*) w + 8;
96 vprod0xCDEFc0 = vmlal_s8(vprod0xCDEFc0, vbCDEFc0x1, va0x1);
97 vacc0xCDEF = vpadalq_s16(vacc0xCDEF, vprod0xCDEFc0);
98 va0x0 = vext_s8(va0x0, va0x0, 2);
99 va0x1 = vext_s8(va0x1, va0x1, 2);
100 int16x8_t vprod0x0123c1 = vmull_s8(vb0123c1x0, va0x0);
101 const int8x8_t vb0123c1x1 = vld1_s8(w); w = (const int8_t*) w + 8;
102 vprod0x0123c1 = vmlal_s8(vprod0x0123c1, vb0123c1x1, va0x1);
103 vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c1);
104 int16x8_t vprod0x4567c1 = vmull_s8(vb4567c1x0, va0x0);
105 const int8x8_t vb4567c1x1 = vld1_s8(w); w = (const int8_t*) w + 8;
106 vprod0x4567c1 = vmlal_s8(vprod0x4567c1, vb4567c1x1, va0x1);
107 vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c1);
108 int16x8_t vprod0x89ABc1 = vmull_s8(vb89ABc1x0, va0x0);
109 const int8x8_t vb89ABc1x1 = vld1_s8(w); w = (const int8_t*) w + 8;
110 vprod0x89ABc1 = vmlal_s8(vprod0x89ABc1, vb89ABc1x1, va0x1);
111 vacc0x89AB = vpadalq_s16(vacc0x89AB, vprod0x89ABc1);
112 int16x8_t vprod0xCDEFc1 = vmull_s8(vbCDEFc1x0, va0x0);
113 const int8x8_t vbCDEFc1x1 = vld1_s8(w); w = (const int8_t*) w + 8;
114 vprod0xCDEFc1 = vmlal_s8(vprod0xCDEFc1, vbCDEFc1x1, va0x1);
115 vacc0xCDEF = vpadalq_s16(vacc0xCDEF, vprod0xCDEFc1);
116 va0x0 = vext_s8(va0x0, va0x0, 2);
117 va0x1 = vext_s8(va0x1, va0x1, 2);
118 int16x8_t vprod0x0123c2 = vmull_s8(vb0123c2x0, va0x0);
119 const int8x8_t vb0123c2x1 = vld1_s8(w); w = (const int8_t*) w + 8;
120 vprod0x0123c2 = vmlal_s8(vprod0x0123c2, vb0123c2x1, va0x1);
121 vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c2);
122 int16x8_t vprod0x4567c2 = vmull_s8(vb4567c2x0, va0x0);
123 const int8x8_t vb4567c2x1 = vld1_s8(w); w = (const int8_t*) w + 8;
124 vprod0x4567c2 = vmlal_s8(vprod0x4567c2, vb4567c2x1, va0x1);
125 vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c2);
126 int16x8_t vprod0x89ABc2 = vmull_s8(vb89ABc2x0, va0x0);
127 const int8x8_t vb89ABc2x1 = vld1_s8(w); w = (const int8_t*) w + 8;
128 vprod0x89ABc2 = vmlal_s8(vprod0x89ABc2, vb89ABc2x1, va0x1);
129 vacc0x89AB = vpadalq_s16(vacc0x89AB, vprod0x89ABc2);
130 int16x8_t vprod0xCDEFc2 = vmull_s8(vbCDEFc2x0, va0x0);
131 const int8x8_t vbCDEFc2x1 = vld1_s8(w); w = (const int8_t*) w + 8;
132 vprod0xCDEFc2 = vmlal_s8(vprod0xCDEFc2, vbCDEFc2x1, va0x1);
133 vacc0xCDEF = vpadalq_s16(vacc0xCDEF, vprod0xCDEFc2);
134 va0x0 = vext_s8(va0x0, va0x0, 2);
135 va0x1 = vext_s8(va0x1, va0x1, 2);
136 int16x8_t vprod0x0123c3 = vmull_s8(vb0123c3x0, va0x0);
137 const int8x8_t vb0123c3x1 = vld1_s8(w); w = (const int8_t*) w + 8;
138 vprod0x0123c3 = vmlal_s8(vprod0x0123c3, vb0123c3x1, va0x1);
139 vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c3);
140 int16x8_t vprod0x4567c3 = vmull_s8(vb4567c3x0, va0x0);
141 const int8x8_t vb4567c3x1 = vld1_s8(w); w = (const int8_t*) w + 8;
142 vprod0x4567c3 = vmlal_s8(vprod0x4567c3, vb4567c3x1, va0x1);
143 vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c3);
144 int16x8_t vprod0x89ABc3 = vmull_s8(vb89ABc3x0, va0x0);
145 const int8x8_t vb89ABc3x1 = vld1_s8(w); w = (const int8_t*) w + 8;
146 vprod0x89ABc3 = vmlal_s8(vprod0x89ABc3, vb89ABc3x1, va0x1);
147 vacc0x89AB = vpadalq_s16(vacc0x89AB, vprod0x89ABc3);
148 int16x8_t vprod0xCDEFc3 = vmull_s8(vbCDEFc3x0, va0x0);
149 const int8x8_t vbCDEFc3x1 = vld1_s8(w); w = (const int8_t*) w + 8;
150 vprod0xCDEFc3 = vmlal_s8(vprod0xCDEFc3, vbCDEFc3x1, va0x1);
151 vacc0xCDEF = vpadalq_s16(vacc0xCDEF, vprod0xCDEFc3);
152
153 k -= 16 * sizeof(int8_t);
154 }
155 if (k != 0) {
156 int8x8_t va0x0 = vld1_s8(a0); a0 += 8;
157
158 const int8x8_t vb0123c0x0 = vld1_s8(w); w = (const int8_t*) w + 8;
159 const int8x8_t vb4567c0x0 = vld1_s8(w); w = (const int8_t*) w + 8;
160 const int8x8_t vb89ABc0x0 = vld1_s8(w); w = (const int8_t*) w + 8;
161 const int8x8_t vbCDEFc0x0 = vld1_s8(w); w = (const int8_t*) w + 8;
162 const int8x8_t vb0123c1x0 = vld1_s8(w); w = (const int8_t*) w + 8;
163 const int8x8_t vb4567c1x0 = vld1_s8(w); w = (const int8_t*) w + 8;
164 const int8x8_t vb89ABc1x0 = vld1_s8(w); w = (const int8_t*) w + 8;
165 const int8x8_t vbCDEFc1x0 = vld1_s8(w); w = (const int8_t*) w + 8;
166 const int8x8_t vb0123c2x0 = vld1_s8(w); w = (const int8_t*) w + 8;
167 const int8x8_t vb4567c2x0 = vld1_s8(w); w = (const int8_t*) w + 8;
168 const int8x8_t vb89ABc2x0 = vld1_s8(w); w = (const int8_t*) w + 8;
169 const int8x8_t vbCDEFc2x0 = vld1_s8(w); w = (const int8_t*) w + 8;
170 const int8x8_t vb0123c3x0 = vld1_s8(w); w = (const int8_t*) w + 8;
171 const int8x8_t vb4567c3x0 = vld1_s8(w); w = (const int8_t*) w + 8;
172 const int8x8_t vb89ABc3x0 = vld1_s8(w); w = (const int8_t*) w + 8;
173 const int8x8_t vbCDEFc3x0 = vld1_s8(w); w = (const int8_t*) w + 8;
174
175 int16x8_t vprod0x0123c0 = vmull_s8(vb0123c0x0, va0x0);
176 vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c0);
177 int16x8_t vprod0x4567c0 = vmull_s8(vb4567c0x0, va0x0);
178 vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c0);
179 int16x8_t vprod0x89ABc0 = vmull_s8(vb89ABc0x0, va0x0);
180 vacc0x89AB = vpadalq_s16(vacc0x89AB, vprod0x89ABc0);
181 int16x8_t vprod0xCDEFc0 = vmull_s8(vbCDEFc0x0, va0x0);
182 vacc0xCDEF = vpadalq_s16(vacc0xCDEF, vprod0xCDEFc0);
183 va0x0 = vext_s8(va0x0, va0x0, 2);
184 int16x8_t vprod0x0123c1 = vmull_s8(vb0123c1x0, va0x0);
185 vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c1);
186 int16x8_t vprod0x4567c1 = vmull_s8(vb4567c1x0, va0x0);
187 vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c1);
188 int16x8_t vprod0x89ABc1 = vmull_s8(vb89ABc1x0, va0x0);
189 vacc0x89AB = vpadalq_s16(vacc0x89AB, vprod0x89ABc1);
190 int16x8_t vprod0xCDEFc1 = vmull_s8(vbCDEFc1x0, va0x0);
191 vacc0xCDEF = vpadalq_s16(vacc0xCDEF, vprod0xCDEFc1);
192 va0x0 = vext_s8(va0x0, va0x0, 2);
193 int16x8_t vprod0x0123c2 = vmull_s8(vb0123c2x0, va0x0);
194 vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c2);
195 int16x8_t vprod0x4567c2 = vmull_s8(vb4567c2x0, va0x0);
196 vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c2);
197 int16x8_t vprod0x89ABc2 = vmull_s8(vb89ABc2x0, va0x0);
198 vacc0x89AB = vpadalq_s16(vacc0x89AB, vprod0x89ABc2);
199 int16x8_t vprod0xCDEFc2 = vmull_s8(vbCDEFc2x0, va0x0);
200 vacc0xCDEF = vpadalq_s16(vacc0xCDEF, vprod0xCDEFc2);
201 va0x0 = vext_s8(va0x0, va0x0, 2);
202 int16x8_t vprod0x0123c3 = vmull_s8(vb0123c3x0, va0x0);
203 vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c3);
204 int16x8_t vprod0x4567c3 = vmull_s8(vb4567c3x0, va0x0);
205 vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c3);
206 int16x8_t vprod0x89ABc3 = vmull_s8(vb89ABc3x0, va0x0);
207 vacc0x89AB = vpadalq_s16(vacc0x89AB, vprod0x89ABc3);
208 int16x8_t vprod0xCDEFc3 = vmull_s8(vbCDEFc3x0, va0x0);
209 vacc0xCDEF = vpadalq_s16(vacc0xCDEF, vprod0xCDEFc3);
210
211 }
212
213 p -= 1 * sizeof(void*);
214 } while (p != 0);
215
216 const int32x4_t vright_pre_shift = vld1q_dup_s32(¶ms->rndnu_neon.right_pre_shift);
217 const int32x4_t vmultiplier = vld1q_dup_s32(¶ms->rndnu_neon.multiplier);
218 const int32x4_t vright_post_shift = vld1q_dup_s32(¶ms->rndnu_neon.right_post_shift);
219
220 vacc0x0123 = vqshlq_s32(vacc0x0123, vright_pre_shift);
221 vacc0x4567 = vqshlq_s32(vacc0x4567, vright_pre_shift);
222 vacc0x89AB = vqshlq_s32(vacc0x89AB, vright_pre_shift);
223 vacc0xCDEF = vqshlq_s32(vacc0xCDEF, vright_pre_shift);
224
225 vacc0x0123 = vqdmulhq_s32(vacc0x0123, vmultiplier);
226 vacc0x4567 = vqdmulhq_s32(vacc0x4567, vmultiplier);
227 vacc0x89AB = vqdmulhq_s32(vacc0x89AB, vmultiplier);
228 vacc0xCDEF = vqdmulhq_s32(vacc0xCDEF, vmultiplier);
229
230 vacc0x0123 = vrshlq_s32(vacc0x0123, vright_post_shift);
231 vacc0x4567 = vrshlq_s32(vacc0x4567, vright_post_shift);
232 vacc0x89AB = vrshlq_s32(vacc0x89AB, vright_post_shift);
233 vacc0xCDEF = vrshlq_s32(vacc0xCDEF, vright_post_shift);
234
235 const int16x8_t voutput_zero_point = vld1q_dup_s16(¶ms->rndnu_neon.output_zero_point);
236 #if XNN_ARCH_ARM64
237 int16x8_t vacc0x01234567 = vqmovn_high_s32(vqmovn_s32(vacc0x0123), vacc0x4567);
238 int16x8_t vacc0x89ABCDEF = vqmovn_high_s32(vqmovn_s32(vacc0x89AB), vacc0xCDEF);
239
240 vacc0x01234567 = vqaddq_s16(vacc0x01234567, voutput_zero_point);
241 vacc0x89ABCDEF = vqaddq_s16(vacc0x89ABCDEF, voutput_zero_point);
242
243 int8x16_t vout0x0123456789ABCDEF = vqmovn_high_s16(vqmovn_s16(vacc0x01234567), vacc0x89ABCDEF);
244 #else
245 int16x8_t vacc0x01234567 = vcombine_s16(vqmovn_s32(vacc0x0123), vqmovn_s32(vacc0x4567));
246 int16x8_t vacc0x89ABCDEF = vcombine_s16(vqmovn_s32(vacc0x89AB), vqmovn_s32(vacc0xCDEF));
247
248 vacc0x01234567 = vqaddq_s16(vacc0x01234567, voutput_zero_point);
249 vacc0x89ABCDEF = vqaddq_s16(vacc0x89ABCDEF, voutput_zero_point);
250
251 int8x16_t vout0x0123456789ABCDEF = vcombine_s8(vqmovn_s16(vacc0x01234567), vqmovn_s16(vacc0x89ABCDEF));
252 #endif
253
254 const int8x16_t voutput_min = vld1q_dup_s8(¶ms->rndnu_neon.output_min);
255 vout0x0123456789ABCDEF = vmaxq_s8(vout0x0123456789ABCDEF, voutput_min);
256
257 const int8x16_t voutput_max = vld1q_dup_s8(¶ms->rndnu_neon.output_max);
258 vout0x0123456789ABCDEF = vminq_s8(vout0x0123456789ABCDEF, voutput_max);
259
260 if (nc >= 16) {
261 vst1q_s8(c0 + 0, vout0x0123456789ABCDEF);
262
263 c0 = (int8_t*) ((uintptr_t) c0 + cn_stride);
264
265 a = (const int8_t**restrict) ((uintptr_t) a - ks);
266
267 nc -= 16;
268 } else {
269 int8x8_t vout0x01234567 = vget_low_s8(vout0x0123456789ABCDEF);
270 if (nc & 8) {
271 vst1_s8(c0, vout0x01234567); c0 += 8;
272 vout0x01234567 = vget_high_s8(vout0x0123456789ABCDEF);
273 }
274 if (nc & 4) {
275 vst1_lane_u32((void*) c0, vreinterpret_u32_s8(vout0x01234567), 0); c0 += 4;
276 vout0x01234567 = vext_s8(vout0x01234567, vout0x01234567, 4);
277 }
278 if (nc & 2) {
279 vst1_lane_u16((void*) c0, vreinterpret_u16_s8(vout0x01234567), 0); c0 += 2;
280 vout0x01234567 = vext_s8(vout0x01234567, vout0x01234567, 2);
281 }
282 if (nc & 1) {
283 vst1_lane_s8(c0, vout0x01234567, 0);
284 }
285
286 nc = 0;
287 }
288 } while (nc != 0);
289 }
290