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_3x8c4s2__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_3x8c4s2__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 <= 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 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 int8_t* c2 = (int8_t*) ((uintptr_t) c1 + cm_stride);
49 if XNN_UNPREDICTABLE(mr <= 2) {
50 c2 = c1;
51 }
52
53 kc = round_up_po2(kc, 8 * sizeof(int8_t));
54 do {
55 int32x4_t vacc0x01 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const int32_t*) w + 2;
56 int32x4_t vacc0x23 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const int32_t*) w + 2;
57 int32x4_t vacc0x45 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const int32_t*) w + 2;
58 int32x4_t vacc0x67 = 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 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 do {
86 int8x8_t va0x0 = vld1_s8(a0); a0 += 8;
87 int8x8_t va1x0 = vld1_s8(a1); a1 += 8;
88 int8x8_t va2x0 = vld1_s8(a2); a2 += 8;
89
90 const int8x8_t vb01c0x0 = vld1_s8(w); w = (const int8_t*) w + 8;
91 const int8x8_t vb23c0x0 = vld1_s8(w); w = (const int8_t*) w + 8;
92 const int8x8_t vb45c0x0 = vld1_s8(w); w = (const int8_t*) w + 8;
93 const int8x8_t vb67c0x0 = vld1_s8(w); w = (const int8_t*) w + 8;
94 const int8x8_t vb01c1x0 = vld1_s8(w); w = (const int8_t*) w + 8;
95 const int8x8_t vb23c1x0 = vld1_s8(w); w = (const int8_t*) w + 8;
96 const int8x8_t vb45c1x0 = vld1_s8(w); w = (const int8_t*) w + 8;
97 const int8x8_t vb67c1x0 = vld1_s8(w); w = (const int8_t*) w + 8;
98
99 int16x8_t vprod0x01c0 = vmull_s8(vb01c0x0, va0x0);
100 int16x8_t vprod1x01c0 = vmull_s8(vb01c0x0, va1x0);
101 int16x8_t vprod2x01c0 = vmull_s8(vb01c0x0, va2x0);
102 vacc0x01 = vpadalq_s16(vacc0x01, vprod0x01c0);
103 vacc1x01 = vpadalq_s16(vacc1x01, vprod1x01c0);
104 vacc2x01 = vpadalq_s16(vacc2x01, vprod2x01c0);
105 int16x8_t vprod0x23c0 = vmull_s8(vb23c0x0, va0x0);
106 int16x8_t vprod1x23c0 = vmull_s8(vb23c0x0, va1x0);
107 int16x8_t vprod2x23c0 = vmull_s8(vb23c0x0, va2x0);
108 vacc0x23 = vpadalq_s16(vacc0x23, vprod0x23c0);
109 vacc1x23 = vpadalq_s16(vacc1x23, vprod1x23c0);
110 vacc2x23 = vpadalq_s16(vacc2x23, vprod2x23c0);
111 int16x8_t vprod0x45c0 = vmull_s8(vb45c0x0, va0x0);
112 int16x8_t vprod1x45c0 = vmull_s8(vb45c0x0, va1x0);
113 int16x8_t vprod2x45c0 = vmull_s8(vb45c0x0, va2x0);
114 vacc0x45 = vpadalq_s16(vacc0x45, vprod0x45c0);
115 vacc1x45 = vpadalq_s16(vacc1x45, vprod1x45c0);
116 vacc2x45 = vpadalq_s16(vacc2x45, vprod2x45c0);
117 int16x8_t vprod0x67c0 = vmull_s8(vb67c0x0, va0x0);
118 int16x8_t vprod1x67c0 = vmull_s8(vb67c0x0, va1x0);
119 int16x8_t vprod2x67c0 = vmull_s8(vb67c0x0, va2x0);
120 vacc0x67 = vpadalq_s16(vacc0x67, vprod0x67c0);
121 vacc1x67 = vpadalq_s16(vacc1x67, vprod1x67c0);
122 vacc2x67 = vpadalq_s16(vacc2x67, vprod2x67c0);
123 va0x0 = vext_s8(va0x0, va0x0, 4);
124 va1x0 = vext_s8(va1x0, va1x0, 4);
125 va2x0 = vext_s8(va2x0, va2x0, 4);
126 int16x8_t vprod0x01c1 = vmull_s8(vb01c1x0, va0x0);
127 int16x8_t vprod1x01c1 = vmull_s8(vb01c1x0, va1x0);
128 int16x8_t vprod2x01c1 = vmull_s8(vb01c1x0, va2x0);
129 vacc0x01 = vpadalq_s16(vacc0x01, vprod0x01c1);
130 vacc1x01 = vpadalq_s16(vacc1x01, vprod1x01c1);
131 vacc2x01 = vpadalq_s16(vacc2x01, vprod2x01c1);
132 int16x8_t vprod0x23c1 = vmull_s8(vb23c1x0, va0x0);
133 int16x8_t vprod1x23c1 = vmull_s8(vb23c1x0, va1x0);
134 int16x8_t vprod2x23c1 = vmull_s8(vb23c1x0, va2x0);
135 vacc0x23 = vpadalq_s16(vacc0x23, vprod0x23c1);
136 vacc1x23 = vpadalq_s16(vacc1x23, vprod1x23c1);
137 vacc2x23 = vpadalq_s16(vacc2x23, vprod2x23c1);
138 int16x8_t vprod0x45c1 = vmull_s8(vb45c1x0, va0x0);
139 int16x8_t vprod1x45c1 = vmull_s8(vb45c1x0, va1x0);
140 int16x8_t vprod2x45c1 = vmull_s8(vb45c1x0, va2x0);
141 vacc0x45 = vpadalq_s16(vacc0x45, vprod0x45c1);
142 vacc1x45 = vpadalq_s16(vacc1x45, vprod1x45c1);
143 vacc2x45 = vpadalq_s16(vacc2x45, vprod2x45c1);
144 int16x8_t vprod0x67c1 = vmull_s8(vb67c1x0, va0x0);
145 int16x8_t vprod1x67c1 = vmull_s8(vb67c1x0, va1x0);
146 int16x8_t vprod2x67c1 = vmull_s8(vb67c1x0, va2x0);
147 vacc0x67 = vpadalq_s16(vacc0x67, vprod0x67c1);
148 vacc1x67 = vpadalq_s16(vacc1x67, vprod1x67c1);
149 vacc2x67 = vpadalq_s16(vacc2x67, vprod2x67c1);
150
151 k -= 8 * sizeof(int8_t);
152 } while (k != 0);
153
154 p -= 3 * sizeof(void*);
155 } while (p != 0);
156
157 #if XNN_ARCH_ARM64
158 int32x4_t vacc0x0123 = vpaddq_s32(vacc0x01, vacc0x23);
159 int32x4_t vacc0x4567 = vpaddq_s32(vacc0x45, vacc0x67);
160 int32x4_t vacc1x0123 = vpaddq_s32(vacc1x01, vacc1x23);
161 int32x4_t vacc1x4567 = vpaddq_s32(vacc1x45, vacc1x67);
162 int32x4_t vacc2x0123 = vpaddq_s32(vacc2x01, vacc2x23);
163 int32x4_t vacc2x4567 = vpaddq_s32(vacc2x45, vacc2x67);
164 #else
165 const int32x2_t vsum0x01 = vpadd_s32(vget_low_s32(vacc0x01), vget_high_s32(vacc0x01));
166 const int32x2_t vsum0x23 = vpadd_s32(vget_low_s32(vacc0x23), vget_high_s32(vacc0x23));
167 int32x4_t vacc0x0123 = vcombine_s32(vsum0x01, vsum0x23);
168 const int32x2_t vsum0x45 = vpadd_s32(vget_low_s32(vacc0x45), vget_high_s32(vacc0x45));
169 const int32x2_t vsum0x67 = vpadd_s32(vget_low_s32(vacc0x67), vget_high_s32(vacc0x67));
170 int32x4_t vacc0x4567 = vcombine_s32(vsum0x45, vsum0x67);
171 const int32x2_t vsum1x01 = vpadd_s32(vget_low_s32(vacc1x01), vget_high_s32(vacc1x01));
172 const int32x2_t vsum1x23 = vpadd_s32(vget_low_s32(vacc1x23), vget_high_s32(vacc1x23));
173 int32x4_t vacc1x0123 = vcombine_s32(vsum1x01, vsum1x23);
174 const int32x2_t vsum1x45 = vpadd_s32(vget_low_s32(vacc1x45), vget_high_s32(vacc1x45));
175 const int32x2_t vsum1x67 = vpadd_s32(vget_low_s32(vacc1x67), vget_high_s32(vacc1x67));
176 int32x4_t vacc1x4567 = vcombine_s32(vsum1x45, vsum1x67);
177 const int32x2_t vsum2x01 = vpadd_s32(vget_low_s32(vacc2x01), vget_high_s32(vacc2x01));
178 const int32x2_t vsum2x23 = vpadd_s32(vget_low_s32(vacc2x23), vget_high_s32(vacc2x23));
179 int32x4_t vacc2x0123 = vcombine_s32(vsum2x01, vsum2x23);
180 const int32x2_t vsum2x45 = vpadd_s32(vget_low_s32(vacc2x45), vget_high_s32(vacc2x45));
181 const int32x2_t vsum2x67 = vpadd_s32(vget_low_s32(vacc2x67), vget_high_s32(vacc2x67));
182 int32x4_t vacc2x4567 = vcombine_s32(vsum2x45, vsum2x67);
183 #endif
184
185 const int32x4_t vright_pre_shift = vld1q_dup_s32(¶ms->rndnu_neon.right_pre_shift);
186 const int32x4_t vmultiplier = vld1q_dup_s32(¶ms->rndnu_neon.multiplier);
187 const int32x4_t vright_post_shift = vld1q_dup_s32(¶ms->rndnu_neon.right_post_shift);
188
189 vacc0x0123 = vqshlq_s32(vacc0x0123, vright_pre_shift);
190 vacc0x4567 = vqshlq_s32(vacc0x4567, vright_pre_shift);
191 vacc1x0123 = vqshlq_s32(vacc1x0123, vright_pre_shift);
192 vacc1x4567 = vqshlq_s32(vacc1x4567, vright_pre_shift);
193 vacc2x0123 = vqshlq_s32(vacc2x0123, vright_pre_shift);
194 vacc2x4567 = vqshlq_s32(vacc2x4567, vright_pre_shift);
195
196 vacc0x0123 = vqdmulhq_s32(vacc0x0123, vmultiplier);
197 vacc0x4567 = vqdmulhq_s32(vacc0x4567, vmultiplier);
198 vacc1x0123 = vqdmulhq_s32(vacc1x0123, vmultiplier);
199 vacc1x4567 = vqdmulhq_s32(vacc1x4567, vmultiplier);
200 vacc2x0123 = vqdmulhq_s32(vacc2x0123, vmultiplier);
201 vacc2x4567 = vqdmulhq_s32(vacc2x4567, vmultiplier);
202
203 vacc0x0123 = vrshlq_s32(vacc0x0123, vright_post_shift);
204 vacc0x4567 = vrshlq_s32(vacc0x4567, vright_post_shift);
205 vacc1x0123 = vrshlq_s32(vacc1x0123, vright_post_shift);
206 vacc1x4567 = vrshlq_s32(vacc1x4567, vright_post_shift);
207 vacc2x0123 = vrshlq_s32(vacc2x0123, vright_post_shift);
208 vacc2x4567 = vrshlq_s32(vacc2x4567, vright_post_shift);
209
210 const int16x8_t voutput_zero_point = vld1q_dup_s16(¶ms->rndnu_neon.output_zero_point);
211 #if XNN_ARCH_ARM64
212 int16x8_t vacc0x01234567 = vqmovn_high_s32(vqmovn_s32(vacc0x0123), vacc0x4567);
213 int16x8_t vacc1x01234567 = vqmovn_high_s32(vqmovn_s32(vacc1x0123), vacc1x4567);
214 int16x8_t vacc2x01234567 = vqmovn_high_s32(vqmovn_s32(vacc2x0123), vacc2x4567);
215
216 vacc0x01234567 = vqaddq_s16(vacc0x01234567, voutput_zero_point);
217 vacc1x01234567 = vqaddq_s16(vacc1x01234567, voutput_zero_point);
218 vacc2x01234567 = vqaddq_s16(vacc2x01234567, voutput_zero_point);
219
220 int8x16_t vout0x01234567_1x01234567 = vqmovn_high_s16(vqmovn_s16(vacc0x01234567), vacc1x01234567);
221 int8x8_t vout2x01234567 = vqmovn_s16(vacc2x01234567);
222 #else
223 int16x8_t vacc0x01234567 = vcombine_s16(vqmovn_s32(vacc0x0123), vqmovn_s32(vacc0x4567));
224 int16x8_t vacc1x01234567 = vcombine_s16(vqmovn_s32(vacc1x0123), vqmovn_s32(vacc1x4567));
225 int16x8_t vacc2x01234567 = vcombine_s16(vqmovn_s32(vacc2x0123), vqmovn_s32(vacc2x4567));
226
227 vacc0x01234567 = vqaddq_s16(vacc0x01234567, voutput_zero_point);
228 vacc1x01234567 = vqaddq_s16(vacc1x01234567, voutput_zero_point);
229 vacc2x01234567 = vqaddq_s16(vacc2x01234567, voutput_zero_point);
230
231 int8x16_t vout0x01234567_1x01234567 = vcombine_s8(vqmovn_s16(vacc0x01234567), vqmovn_s16(vacc1x01234567));
232 int8x8_t vout2x01234567 = vqmovn_s16(vacc2x01234567);
233 #endif
234
235 const int8x16_t voutput_min = vld1q_dup_s8(¶ms->rndnu_neon.output_min);
236 vout0x01234567_1x01234567 = vmaxq_s8(vout0x01234567_1x01234567, voutput_min);
237 vout2x01234567 = vmax_s8(vout2x01234567, vget_low_s8(voutput_min));
238
239 const int8x16_t voutput_max = vld1q_dup_s8(¶ms->rndnu_neon.output_max);
240 vout0x01234567_1x01234567 = vminq_s8(vout0x01234567_1x01234567, voutput_max);
241 vout2x01234567 = vmin_s8(vout2x01234567, vget_low_s8(voutput_max));
242
243 if (nc >= 8) {
244 vst1_s8(c2 + 0, vout2x01234567);
245 vst1_s8(c1 + 0, vget_high_s8(vout0x01234567_1x01234567));
246 vst1_s8(c0 + 0, vget_low_s8(vout0x01234567_1x01234567));
247
248 c2 = (int8_t*) ((uintptr_t) c2 + cn_stride);
249 c1 = (int8_t*) ((uintptr_t) c1 + cn_stride);
250 c0 = (int8_t*) ((uintptr_t) c0 + cn_stride);
251
252 a = (const int8_t**restrict) ((uintptr_t) a - ks);
253
254 nc -= 8;
255 } else {
256 if (nc & 4) {
257 vst1_lane_u32((void*) c2, vreinterpret_u32_s8(vout2x01234567), 0); c2 += 4;
258 vst1q_lane_u32((void*) c1, vreinterpretq_u32_s8(vout0x01234567_1x01234567), 2); c1 += 4;
259 vst1q_lane_u32((void*) c0, vreinterpretq_u32_s8(vout0x01234567_1x01234567), 0); c0 += 4;
260 vout2x01234567 = vext_s8(vout2x01234567, vout2x01234567, 4);
261 vout0x01234567_1x01234567 = vextq_s8(vout0x01234567_1x01234567, vout0x01234567_1x01234567, 4);
262 }
263 if (nc & 2) {
264 vst1_lane_u16((void*) c2, vreinterpret_u16_s8(vout2x01234567), 0); c2 += 2;
265 vst1q_lane_u16((void*) c1, vreinterpretq_u16_s8(vout0x01234567_1x01234567), 4); c1 += 2;
266 vst1q_lane_u16((void*) c0, vreinterpretq_u16_s8(vout0x01234567_1x01234567), 0); c0 += 2;
267 vout2x01234567 = vext_s8(vout2x01234567, vout2x01234567, 2);
268 vout0x01234567_1x01234567 = vextq_s8(vout0x01234567_1x01234567, vout0x01234567_1x01234567, 2);
269 }
270 if (nc & 1) {
271 vst1_lane_s8(c2, vout2x01234567, 0);
272 vst1q_lane_s8(c1, vout0x01234567_1x01234567, 8);
273 vst1q_lane_s8(c0, vout0x01234567_1x01234567, 0);
274 }
275
276 nc = 0;
277 }
278 } while (nc != 0);
279 }
280