1 // Auto-generated file. Do not edit!
2 // Template: src/qs8-gemm/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_gemm_minmax_rndnu_ukernel_4x8c2s4__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_4x8c2s4__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 <= 4);
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 const int8_t* a2 = (const int8_t*) ((uintptr_t) a1 + a_stride);
48 int8_t* c2 = (int8_t*) ((uintptr_t) c1 + cm_stride);
49 if XNN_UNPREDICTABLE(mr <= 2) {
50 a2 = a1;
51 c2 = c1;
52 }
53 const int8_t* a3 = (const int8_t*) ((uintptr_t) a2 + a_stride);
54 int8_t* c3 = (int8_t*) ((uintptr_t) c2 + cm_stride);
55 if XNN_UNPREDICTABLE(mr != 4) {
56 a3 = a2;
57 c3 = c2;
58 }
59
60 kc = round_up_po2(kc, 8 * sizeof(int8_t));
61 do {
62 int32x4_t vacc0x0123 = vld1q_s32(w); w = (const int32_t*) w + 4;
63 int32x4_t vacc0x4567 = vld1q_s32(w); w = (const int32_t*) w + 4;
64 int32x4_t vacc1x0123 = vacc0x0123;
65 int32x4_t vacc1x4567 = vacc0x4567;
66 int32x4_t vacc2x0123 = vacc0x0123;
67 int32x4_t vacc2x4567 = vacc0x4567;
68 int32x4_t vacc3x0123 = vacc0x0123;
69 int32x4_t vacc3x4567 = vacc0x4567;
70
71 size_t k = kc;
72 do {
73 int8x8_t va0x0 = vld1_s8(a0); a0 += 8;
74 int8x8_t va1x0 = vld1_s8(a1); a1 += 8;
75 int8x8_t va2x0 = vld1_s8(a2); a2 += 8;
76 int8x8_t va3x0 = vld1_s8(a3); a3 += 8;
77
78 const int8x8_t vb0123c0x0 = vld1_s8(w); w = (const int8_t*) w + 8;
79 const int8x8_t vb4567c0x0 = vld1_s8(w); w = (const int8_t*) w + 8;
80 const int8x8_t vb0123c1x0 = vld1_s8(w); w = (const int8_t*) w + 8;
81 const int8x8_t vb4567c1x0 = vld1_s8(w); w = (const int8_t*) w + 8;
82 const int8x8_t vb0123c2x0 = vld1_s8(w); w = (const int8_t*) w + 8;
83 const int8x8_t vb4567c2x0 = vld1_s8(w); w = (const int8_t*) w + 8;
84 const int8x8_t vb0123c3x0 = vld1_s8(w); w = (const int8_t*) w + 8;
85 const int8x8_t vb4567c3x0 = vld1_s8(w); w = (const int8_t*) w + 8;
86
87 int16x8_t vprod0x0123c0 = vmull_s8(vb0123c0x0, va0x0);
88 int16x8_t vprod1x0123c0 = vmull_s8(vb0123c0x0, va1x0);
89 int16x8_t vprod2x0123c0 = vmull_s8(vb0123c0x0, va2x0);
90 int16x8_t vprod3x0123c0 = vmull_s8(vb0123c0x0, va3x0);
91 vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c0);
92 vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c0);
93 vacc2x0123 = vpadalq_s16(vacc2x0123, vprod2x0123c0);
94 vacc3x0123 = vpadalq_s16(vacc3x0123, vprod3x0123c0);
95 int16x8_t vprod0x4567c0 = vmull_s8(vb4567c0x0, va0x0);
96 int16x8_t vprod1x4567c0 = vmull_s8(vb4567c0x0, va1x0);
97 int16x8_t vprod2x4567c0 = vmull_s8(vb4567c0x0, va2x0);
98 int16x8_t vprod3x4567c0 = vmull_s8(vb4567c0x0, va3x0);
99 vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c0);
100 vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c0);
101 vacc2x4567 = vpadalq_s16(vacc2x4567, vprod2x4567c0);
102 vacc3x4567 = vpadalq_s16(vacc3x4567, vprod3x4567c0);
103 va0x0 = vext_s8(va0x0, va0x0, 2);
104 va1x0 = vext_s8(va1x0, va1x0, 2);
105 va2x0 = vext_s8(va2x0, va2x0, 2);
106 va3x0 = vext_s8(va3x0, va3x0, 2);
107 int16x8_t vprod0x0123c1 = vmull_s8(vb0123c1x0, va0x0);
108 int16x8_t vprod1x0123c1 = vmull_s8(vb0123c1x0, va1x0);
109 int16x8_t vprod2x0123c1 = vmull_s8(vb0123c1x0, va2x0);
110 int16x8_t vprod3x0123c1 = vmull_s8(vb0123c1x0, va3x0);
111 vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c1);
112 vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c1);
113 vacc2x0123 = vpadalq_s16(vacc2x0123, vprod2x0123c1);
114 vacc3x0123 = vpadalq_s16(vacc3x0123, vprod3x0123c1);
115 int16x8_t vprod0x4567c1 = vmull_s8(vb4567c1x0, va0x0);
116 int16x8_t vprod1x4567c1 = vmull_s8(vb4567c1x0, va1x0);
117 int16x8_t vprod2x4567c1 = vmull_s8(vb4567c1x0, va2x0);
118 int16x8_t vprod3x4567c1 = vmull_s8(vb4567c1x0, va3x0);
119 vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c1);
120 vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c1);
121 vacc2x4567 = vpadalq_s16(vacc2x4567, vprod2x4567c1);
122 vacc3x4567 = vpadalq_s16(vacc3x4567, vprod3x4567c1);
123 va0x0 = vext_s8(va0x0, va0x0, 2);
124 va1x0 = vext_s8(va1x0, va1x0, 2);
125 va2x0 = vext_s8(va2x0, va2x0, 2);
126 va3x0 = vext_s8(va3x0, va3x0, 2);
127 int16x8_t vprod0x0123c2 = vmull_s8(vb0123c2x0, va0x0);
128 int16x8_t vprod1x0123c2 = vmull_s8(vb0123c2x0, va1x0);
129 int16x8_t vprod2x0123c2 = vmull_s8(vb0123c2x0, va2x0);
130 int16x8_t vprod3x0123c2 = vmull_s8(vb0123c2x0, va3x0);
131 vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c2);
132 vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c2);
133 vacc2x0123 = vpadalq_s16(vacc2x0123, vprod2x0123c2);
134 vacc3x0123 = vpadalq_s16(vacc3x0123, vprod3x0123c2);
135 int16x8_t vprod0x4567c2 = vmull_s8(vb4567c2x0, va0x0);
136 int16x8_t vprod1x4567c2 = vmull_s8(vb4567c2x0, va1x0);
137 int16x8_t vprod2x4567c2 = vmull_s8(vb4567c2x0, va2x0);
138 int16x8_t vprod3x4567c2 = vmull_s8(vb4567c2x0, va3x0);
139 vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c2);
140 vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c2);
141 vacc2x4567 = vpadalq_s16(vacc2x4567, vprod2x4567c2);
142 vacc3x4567 = vpadalq_s16(vacc3x4567, vprod3x4567c2);
143 va0x0 = vext_s8(va0x0, va0x0, 2);
144 va1x0 = vext_s8(va1x0, va1x0, 2);
145 va2x0 = vext_s8(va2x0, va2x0, 2);
146 va3x0 = vext_s8(va3x0, va3x0, 2);
147 int16x8_t vprod0x0123c3 = vmull_s8(vb0123c3x0, va0x0);
148 int16x8_t vprod1x0123c3 = vmull_s8(vb0123c3x0, va1x0);
149 int16x8_t vprod2x0123c3 = vmull_s8(vb0123c3x0, va2x0);
150 int16x8_t vprod3x0123c3 = vmull_s8(vb0123c3x0, va3x0);
151 vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c3);
152 vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c3);
153 vacc2x0123 = vpadalq_s16(vacc2x0123, vprod2x0123c3);
154 vacc3x0123 = vpadalq_s16(vacc3x0123, vprod3x0123c3);
155 int16x8_t vprod0x4567c3 = vmull_s8(vb4567c3x0, va0x0);
156 int16x8_t vprod1x4567c3 = vmull_s8(vb4567c3x0, va1x0);
157 int16x8_t vprod2x4567c3 = vmull_s8(vb4567c3x0, va2x0);
158 int16x8_t vprod3x4567c3 = vmull_s8(vb4567c3x0, va3x0);
159 vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c3);
160 vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c3);
161 vacc2x4567 = vpadalq_s16(vacc2x4567, vprod2x4567c3);
162 vacc3x4567 = vpadalq_s16(vacc3x4567, vprod3x4567c3);
163
164 k -= 8 * sizeof(int8_t);
165 } while (k != 0);
166
167 const int32x4_t vright_pre_shift = vld1q_dup_s32(¶ms->rndnu_neon.right_pre_shift);
168 const int32x4_t vmultiplier = vld1q_dup_s32(¶ms->rndnu_neon.multiplier);
169 const int32x4_t vright_post_shift = vld1q_dup_s32(¶ms->rndnu_neon.right_post_shift);
170
171 vacc0x0123 = vqshlq_s32(vacc0x0123, vright_pre_shift);
172 vacc0x4567 = vqshlq_s32(vacc0x4567, vright_pre_shift);
173 vacc1x0123 = vqshlq_s32(vacc1x0123, vright_pre_shift);
174 vacc1x4567 = vqshlq_s32(vacc1x4567, vright_pre_shift);
175 vacc2x0123 = vqshlq_s32(vacc2x0123, vright_pre_shift);
176 vacc2x4567 = vqshlq_s32(vacc2x4567, vright_pre_shift);
177 vacc3x0123 = vqshlq_s32(vacc3x0123, vright_pre_shift);
178 vacc3x4567 = vqshlq_s32(vacc3x4567, vright_pre_shift);
179
180 vacc0x0123 = vqdmulhq_s32(vacc0x0123, vmultiplier);
181 vacc0x4567 = vqdmulhq_s32(vacc0x4567, vmultiplier);
182 vacc1x0123 = vqdmulhq_s32(vacc1x0123, vmultiplier);
183 vacc1x4567 = vqdmulhq_s32(vacc1x4567, vmultiplier);
184 vacc2x0123 = vqdmulhq_s32(vacc2x0123, vmultiplier);
185 vacc2x4567 = vqdmulhq_s32(vacc2x4567, vmultiplier);
186 vacc3x0123 = vqdmulhq_s32(vacc3x0123, vmultiplier);
187 vacc3x4567 = vqdmulhq_s32(vacc3x4567, vmultiplier);
188
189 vacc0x0123 = vrshlq_s32(vacc0x0123, vright_post_shift);
190 vacc0x4567 = vrshlq_s32(vacc0x4567, vright_post_shift);
191 vacc1x0123 = vrshlq_s32(vacc1x0123, vright_post_shift);
192 vacc1x4567 = vrshlq_s32(vacc1x4567, vright_post_shift);
193 vacc2x0123 = vrshlq_s32(vacc2x0123, vright_post_shift);
194 vacc2x4567 = vrshlq_s32(vacc2x4567, vright_post_shift);
195 vacc3x0123 = vrshlq_s32(vacc3x0123, vright_post_shift);
196 vacc3x4567 = vrshlq_s32(vacc3x4567, vright_post_shift);
197
198 const int16x8_t voutput_zero_point = vld1q_dup_s16(¶ms->rndnu_neon.output_zero_point);
199 #if XNN_ARCH_ARM64
200 int16x8_t vacc0x01234567 = vqmovn_high_s32(vqmovn_s32(vacc0x0123), vacc0x4567);
201 int16x8_t vacc1x01234567 = vqmovn_high_s32(vqmovn_s32(vacc1x0123), vacc1x4567);
202 int16x8_t vacc2x01234567 = vqmovn_high_s32(vqmovn_s32(vacc2x0123), vacc2x4567);
203 int16x8_t vacc3x01234567 = vqmovn_high_s32(vqmovn_s32(vacc3x0123), vacc3x4567);
204
205 vacc0x01234567 = vqaddq_s16(vacc0x01234567, voutput_zero_point);
206 vacc1x01234567 = vqaddq_s16(vacc1x01234567, voutput_zero_point);
207 vacc2x01234567 = vqaddq_s16(vacc2x01234567, voutput_zero_point);
208 vacc3x01234567 = vqaddq_s16(vacc3x01234567, voutput_zero_point);
209
210 int8x16_t vout0x01234567_1x01234567 = vqmovn_high_s16(vqmovn_s16(vacc0x01234567), vacc1x01234567);
211 int8x16_t vout2x01234567_3x01234567 = vqmovn_high_s16(vqmovn_s16(vacc2x01234567), vacc3x01234567);
212 #else
213 int16x8_t vacc0x01234567 = vcombine_s16(vqmovn_s32(vacc0x0123), vqmovn_s32(vacc0x4567));
214 int16x8_t vacc1x01234567 = vcombine_s16(vqmovn_s32(vacc1x0123), vqmovn_s32(vacc1x4567));
215 int16x8_t vacc2x01234567 = vcombine_s16(vqmovn_s32(vacc2x0123), vqmovn_s32(vacc2x4567));
216 int16x8_t vacc3x01234567 = vcombine_s16(vqmovn_s32(vacc3x0123), vqmovn_s32(vacc3x4567));
217
218 vacc0x01234567 = vqaddq_s16(vacc0x01234567, voutput_zero_point);
219 vacc1x01234567 = vqaddq_s16(vacc1x01234567, voutput_zero_point);
220 vacc2x01234567 = vqaddq_s16(vacc2x01234567, voutput_zero_point);
221 vacc3x01234567 = vqaddq_s16(vacc3x01234567, voutput_zero_point);
222
223 int8x16_t vout0x01234567_1x01234567 = vcombine_s8(vqmovn_s16(vacc0x01234567), vqmovn_s16(vacc1x01234567));
224 int8x16_t vout2x01234567_3x01234567 = vcombine_s8(vqmovn_s16(vacc2x01234567), vqmovn_s16(vacc3x01234567));
225 #endif
226
227 const int8x16_t voutput_min = vld1q_dup_s8(¶ms->rndnu_neon.output_min);
228 vout0x01234567_1x01234567 = vmaxq_s8(vout0x01234567_1x01234567, voutput_min);
229 vout2x01234567_3x01234567 = vmaxq_s8(vout2x01234567_3x01234567, voutput_min);
230
231 const int8x16_t voutput_max = vld1q_dup_s8(¶ms->rndnu_neon.output_max);
232 vout0x01234567_1x01234567 = vminq_s8(vout0x01234567_1x01234567, voutput_max);
233 vout2x01234567_3x01234567 = vminq_s8(vout2x01234567_3x01234567, voutput_max);
234
235 if (nc >= 8) {
236 vst1_s8(c0 + 0, vget_low_s8(vout0x01234567_1x01234567));
237 vst1_s8(c1 + 0, vget_high_s8(vout0x01234567_1x01234567));
238 vst1_s8(c2 + 0, vget_low_s8(vout2x01234567_3x01234567));
239 vst1_s8(c3 + 0, vget_high_s8(vout2x01234567_3x01234567));
240
241 c0 = (int8_t*) ((uintptr_t) c0 + cn_stride);
242 c1 = (int8_t*) ((uintptr_t) c1 + cn_stride);
243 c2 = (int8_t*) ((uintptr_t) c2 + cn_stride);
244 c3 = (int8_t*) ((uintptr_t) c3 + cn_stride);
245
246 a0 = (const int8_t*) ((uintptr_t) a0 - kc);
247 a1 = (const int8_t*) ((uintptr_t) a1 - kc);
248 a2 = (const int8_t*) ((uintptr_t) a2 - kc);
249 a3 = (const int8_t*) ((uintptr_t) a3 - kc);
250
251 nc -= 8;
252 } else {
253 // Final case where not all of the 8 columns fit in the destination.
254 if (nc & 4) {
255 vst1q_lane_u32((void*) c0, vreinterpretq_u32_s8(vout0x01234567_1x01234567), 0); c0 += 4;
256 vst1q_lane_u32((void*) c1, vreinterpretq_u32_s8(vout0x01234567_1x01234567), 2); c1 += 4;
257 vst1q_lane_u32((void*) c2, vreinterpretq_u32_s8(vout2x01234567_3x01234567), 0); c2 += 4;
258 vst1q_lane_u32((void*) c3, vreinterpretq_u32_s8(vout2x01234567_3x01234567), 2); c3 += 4;
259 vout0x01234567_1x01234567 = vextq_s8(vout0x01234567_1x01234567, vout0x01234567_1x01234567, 4);
260 vout2x01234567_3x01234567 = vextq_s8(vout2x01234567_3x01234567, vout2x01234567_3x01234567, 4);
261 }
262 if (nc & 2) {
263 vst1q_lane_u16((void*) c0, vreinterpretq_u16_s8(vout0x01234567_1x01234567), 0); c0 += 2;
264 vst1q_lane_u16((void*) c1, vreinterpretq_u16_s8(vout0x01234567_1x01234567), 4); c1 += 2;
265 vst1q_lane_u16((void*) c2, vreinterpretq_u16_s8(vout2x01234567_3x01234567), 0); c2 += 2;
266 vst1q_lane_u16((void*) c3, vreinterpretq_u16_s8(vout2x01234567_3x01234567), 4); c3 += 2;
267 vout0x01234567_1x01234567 = vextq_s8(vout0x01234567_1x01234567, vout0x01234567_1x01234567, 2);
268 vout2x01234567_3x01234567 = vextq_s8(vout2x01234567_3x01234567, vout2x01234567_3x01234567, 2);
269 }
270 if (nc & 1) {
271 vst1q_lane_s8(c0, vout0x01234567_1x01234567, 0);
272 vst1q_lane_s8(c1, vout0x01234567_1x01234567, 8);
273 vst1q_lane_s8(c2, vout2x01234567_3x01234567, 0);
274 vst1q_lane_s8(c3, vout2x01234567_3x01234567, 8);
275 }
276
277 nc = 0;
278 }
279 } while (nc != 0);
280 }
281