xref: /aosp_15_r20/external/XNNPACK/src/qs8-gemm/gen/3x8c2-minmax-rndnu-neon-mull-dup.c (revision 4bdc94577ba0e567308109d787f7fec7b531ce36) 
1 // Auto-generated file. Do not edit!
2 //   Template: src/qs8-gemm/c2-neon-mull-dup.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 
xnn_qs8_gemm_minmax_rndnu_ukernel_3x8c2__neon_mull_dup(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)])17 void xnn_qs8_gemm_minmax_rndnu_ukernel_3x8c2__neon_mull_dup(
18     size_t mr,
19     size_t nc,
20     size_t kc,
21     const int8_t* restrict a,
22     size_t a_stride,
23     const void* restrict w,
24     int8_t* restrict c,
25     size_t cm_stride,
26     size_t cn_stride,
27     const union xnn_qs8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
28 {
29   assert(mr != 0);
30   assert(mr <= 3);
31   assert(nc != 0);
32   assert(kc != 0);
33   assert(kc % sizeof(int8_t) == 0);
34   assert(a != NULL);
35   assert(w != NULL);
36   assert(c != NULL);
37 
38   kc = round_up_po2(kc, 2 * sizeof(int8_t));
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 
54   do {
55     int32x4_t vacc0x0123 = vld1q_s32(w); w = (const void*) ((uintptr_t) w + 4 * sizeof(int32_t));
56     int32x4_t vacc0x4567 = vld1q_s32(w); w = (const void*) ((uintptr_t) w + 4 * sizeof(int32_t));
57     int32x4_t vacc1x0123 = vacc0x0123;
58     int32x4_t vacc1x4567 = vacc0x4567;
59     int32x4_t vacc2x0123 = vacc0x0123;
60     int32x4_t vacc2x4567 = vacc0x4567;
61 
62     size_t k = kc;
63 
64 
65     while (k >= 8 * sizeof(int8_t)) {
66       const int8x8_t va0 = vld1_s8(a0); a0 += 8;
67       const int8x8_t va1 = vld1_s8(a1); a1 += 8;
68       const int8x8_t va2 = vld1_s8(a2); a2 += 8;
69 
70       const int8x8_t vb0123c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
71       const int8x8_t vb4567c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
72       const int8x8_t vb0123c1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
73       const int8x8_t vb4567c1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
74       const int8x8_t vb0123c2 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
75       const int8x8_t vb4567c2 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
76       const int8x8_t vb0123c3 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
77       const int8x8_t vb4567c3 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
78 
79       const int8x8_t va0c0 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0), 0));
80       const int8x8_t va1c0 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va1), 0));
81       const int8x8_t va2c0 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va2), 0));
82 
83       const int16x8_t vprod0x0123c0 = vmull_s8(vb0123c0, va0c0);
84       const int16x8_t vprod1x0123c0 = vmull_s8(vb0123c0, va1c0);
85       const int16x8_t vprod2x0123c0 = vmull_s8(vb0123c0, va2c0);
86       vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c0);
87       vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c0);
88       vacc2x0123 = vpadalq_s16(vacc2x0123, vprod2x0123c0);
89       const int16x8_t vprod0x4567c0 = vmull_s8(vb4567c0, va0c0);
90       const int16x8_t vprod1x4567c0 = vmull_s8(vb4567c0, va1c0);
91       const int16x8_t vprod2x4567c0 = vmull_s8(vb4567c0, va2c0);
92       vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c0);
93       vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c0);
94       vacc2x4567 = vpadalq_s16(vacc2x4567, vprod2x4567c0);
95       const int8x8_t va0c1 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0), 1));
96       const int8x8_t va1c1 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va1), 1));
97       const int8x8_t va2c1 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va2), 1));
98 
99       const int16x8_t vprod0x0123c1 = vmull_s8(vb0123c1, va0c1);
100       const int16x8_t vprod1x0123c1 = vmull_s8(vb0123c1, va1c1);
101       const int16x8_t vprod2x0123c1 = vmull_s8(vb0123c1, va2c1);
102       vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c1);
103       vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c1);
104       vacc2x0123 = vpadalq_s16(vacc2x0123, vprod2x0123c1);
105       const int16x8_t vprod0x4567c1 = vmull_s8(vb4567c1, va0c1);
106       const int16x8_t vprod1x4567c1 = vmull_s8(vb4567c1, va1c1);
107       const int16x8_t vprod2x4567c1 = vmull_s8(vb4567c1, va2c1);
108       vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c1);
109       vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c1);
110       vacc2x4567 = vpadalq_s16(vacc2x4567, vprod2x4567c1);
111       const int8x8_t va0c2 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0), 2));
112       const int8x8_t va1c2 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va1), 2));
113       const int8x8_t va2c2 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va2), 2));
114 
115       const int16x8_t vprod0x0123c2 = vmull_s8(vb0123c2, va0c2);
116       const int16x8_t vprod1x0123c2 = vmull_s8(vb0123c2, va1c2);
117       const int16x8_t vprod2x0123c2 = vmull_s8(vb0123c2, va2c2);
118       vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c2);
119       vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c2);
120       vacc2x0123 = vpadalq_s16(vacc2x0123, vprod2x0123c2);
121       const int16x8_t vprod0x4567c2 = vmull_s8(vb4567c2, va0c2);
122       const int16x8_t vprod1x4567c2 = vmull_s8(vb4567c2, va1c2);
123       const int16x8_t vprod2x4567c2 = vmull_s8(vb4567c2, va2c2);
124       vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c2);
125       vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c2);
126       vacc2x4567 = vpadalq_s16(vacc2x4567, vprod2x4567c2);
127       const int8x8_t va0c3 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0), 3));
128       const int8x8_t va1c3 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va1), 3));
129       const int8x8_t va2c3 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va2), 3));
130 
131       const int16x8_t vprod0x0123c3 = vmull_s8(vb0123c3, va0c3);
132       const int16x8_t vprod1x0123c3 = vmull_s8(vb0123c3, va1c3);
133       const int16x8_t vprod2x0123c3 = vmull_s8(vb0123c3, va2c3);
134       vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c3);
135       vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c3);
136       vacc2x0123 = vpadalq_s16(vacc2x0123, vprod2x0123c3);
137       const int16x8_t vprod0x4567c3 = vmull_s8(vb4567c3, va0c3);
138       const int16x8_t vprod1x4567c3 = vmull_s8(vb4567c3, va1c3);
139       const int16x8_t vprod2x4567c3 = vmull_s8(vb4567c3, va2c3);
140       vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c3);
141       vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c3);
142       vacc2x4567 = vpadalq_s16(vacc2x4567, vprod2x4567c3);
143 
144       k -= 8 * sizeof(int8_t);
145     }
146 
147     if XNN_UNLIKELY(k != 0) {
148       const int8x8_t va0 = vld1_s8(a0); a0 = (const int8_t*) ((uintptr_t) a0 + k);
149       const int8x8_t va1 = vld1_s8(a1); a1 = (const int8_t*) ((uintptr_t) a1 + k);
150       const int8x8_t va2 = vld1_s8(a2); a2 = (const int8_t*) ((uintptr_t) a2 + k);
151 
152       const int8x8_t vb0123c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
153       const int8x8_t vb4567c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
154 
155       const int8x8_t va0c0 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0), 0));
156       const int16x8_t vprod0x0123c0 = vmull_s8(vb0123c0, va0c0);
157       vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c0);
158       const int16x8_t vprod0x4567c0 = vmull_s8(vb4567c0, va0c0);
159       vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c0);
160       const int8x8_t va1c0 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va1), 0));
161       const int16x8_t vprod1x0123c0 = vmull_s8(vb0123c0, va1c0);
162       vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c0);
163       const int16x8_t vprod1x4567c0 = vmull_s8(vb4567c0, va1c0);
164       vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c0);
165       const int8x8_t va2c0 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va2), 0));
166       const int16x8_t vprod2x0123c0 = vmull_s8(vb0123c0, va2c0);
167       vacc2x0123 = vpadalq_s16(vacc2x0123, vprod2x0123c0);
168       const int16x8_t vprod2x4567c0 = vmull_s8(vb4567c0, va2c0);
169       vacc2x4567 = vpadalq_s16(vacc2x4567, vprod2x4567c0);
170 
171       if (k > 2 * sizeof(int8_t)) {
172         const int8x8_t vb0123c1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
173         const int8x8_t vb4567c1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
174 
175         const int8x8_t va0c1 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0), 1));
176         const int16x8_t vprod0x0123c1 = vmull_s8(vb0123c1, va0c1);
177         vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c1);
178         const int16x8_t vprod0x4567c1 = vmull_s8(vb4567c1, va0c1);
179         vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c1);
180         const int8x8_t va1c1 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va1), 1));
181         const int16x8_t vprod1x0123c1 = vmull_s8(vb0123c1, va1c1);
182         vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c1);
183         const int16x8_t vprod1x4567c1 = vmull_s8(vb4567c1, va1c1);
184         vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c1);
185         const int8x8_t va2c1 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va2), 1));
186         const int16x8_t vprod2x0123c1 = vmull_s8(vb0123c1, va2c1);
187         vacc2x0123 = vpadalq_s16(vacc2x0123, vprod2x0123c1);
188         const int16x8_t vprod2x4567c1 = vmull_s8(vb4567c1, va2c1);
189         vacc2x4567 = vpadalq_s16(vacc2x4567, vprod2x4567c1);
190 
191         if (k > 4 * sizeof(int8_t)) {
192           const int8x8_t vb0123c2 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
193           const int8x8_t vb4567c2 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
194 
195           const int8x8_t va0c2 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0), 2));
196           const int16x8_t vprod0x0123c2 = vmull_s8(vb0123c2, va0c2);
197           vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c2);
198           const int16x8_t vprod0x4567c2 = vmull_s8(vb4567c2, va0c2);
199           vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c2);
200           const int8x8_t va1c2 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va1), 2));
201           const int16x8_t vprod1x0123c2 = vmull_s8(vb0123c2, va1c2);
202           vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c2);
203           const int16x8_t vprod1x4567c2 = vmull_s8(vb4567c2, va1c2);
204           vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c2);
205           const int8x8_t va2c2 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va2), 2));
206           const int16x8_t vprod2x0123c2 = vmull_s8(vb0123c2, va2c2);
207           vacc2x0123 = vpadalq_s16(vacc2x0123, vprod2x0123c2);
208           const int16x8_t vprod2x4567c2 = vmull_s8(vb4567c2, va2c2);
209           vacc2x4567 = vpadalq_s16(vacc2x4567, vprod2x4567c2);
210         }
211       }
212     }
213 
214     const int32x4_t vright_pre_shift = vld1q_dup_s32(&params->rndnu_neon.right_pre_shift);
215     const int32x4_t vmultiplier = vld1q_dup_s32(&params->rndnu_neon.multiplier);
216     const int32x4_t vright_post_shift = vld1q_dup_s32(&params->rndnu_neon.right_post_shift);
217 
218     vacc0x0123 = vqshlq_s32(vacc0x0123, vright_pre_shift);
219     vacc0x4567 = vqshlq_s32(vacc0x4567, vright_pre_shift);
220     vacc1x0123 = vqshlq_s32(vacc1x0123, vright_pre_shift);
221     vacc1x4567 = vqshlq_s32(vacc1x4567, vright_pre_shift);
222     vacc2x0123 = vqshlq_s32(vacc2x0123, vright_pre_shift);
223     vacc2x4567 = vqshlq_s32(vacc2x4567, vright_pre_shift);
224 
225     vacc0x0123 = vqdmulhq_s32(vacc0x0123, vmultiplier);
226     vacc0x4567 = vqdmulhq_s32(vacc0x4567, vmultiplier);
227     vacc1x0123 = vqdmulhq_s32(vacc1x0123, vmultiplier);
228     vacc1x4567 = vqdmulhq_s32(vacc1x4567, vmultiplier);
229     vacc2x0123 = vqdmulhq_s32(vacc2x0123, vmultiplier);
230     vacc2x4567 = vqdmulhq_s32(vacc2x4567, vmultiplier);
231 
232     vacc0x0123 = vrshlq_s32(vacc0x0123, vright_post_shift);
233     vacc0x4567 = vrshlq_s32(vacc0x4567, vright_post_shift);
234     vacc1x0123 = vrshlq_s32(vacc1x0123, vright_post_shift);
235     vacc1x4567 = vrshlq_s32(vacc1x4567, vright_post_shift);
236     vacc2x0123 = vrshlq_s32(vacc2x0123, vright_post_shift);
237     vacc2x4567 = vrshlq_s32(vacc2x4567, vright_post_shift);
238 
239     const int16x8_t voutput_zero_point = vld1q_dup_s16(&params->rndnu_neon.output_zero_point);
240 #if XNN_ARCH_ARM64
241     int16x8_t vacc0x01234567 = vqmovn_high_s32(vqmovn_s32(vacc0x0123), vacc0x4567);
242     int16x8_t vacc1x01234567 = vqmovn_high_s32(vqmovn_s32(vacc1x0123), vacc1x4567);
243     int16x8_t vacc2x01234567 = vqmovn_high_s32(vqmovn_s32(vacc2x0123), vacc2x4567);
244 
245     vacc0x01234567 = vqaddq_s16(vacc0x01234567, voutput_zero_point);
246     vacc1x01234567 = vqaddq_s16(vacc1x01234567, voutput_zero_point);
247     vacc2x01234567 = vqaddq_s16(vacc2x01234567, voutput_zero_point);
248 
249     int8x16_t vout0x01234567_1x01234567 = vqmovn_high_s16(vqmovn_s16(vacc0x01234567), vacc1x01234567);
250     int8x8_t vout2x01234567 = vqmovn_s16(vacc2x01234567);
251 #else
252     int16x8_t vacc0x01234567 = vcombine_s16(vqmovn_s32(vacc0x0123), vqmovn_s32(vacc0x4567));
253     int16x8_t vacc1x01234567 = vcombine_s16(vqmovn_s32(vacc1x0123), vqmovn_s32(vacc1x4567));
254     int16x8_t vacc2x01234567 = vcombine_s16(vqmovn_s32(vacc2x0123), vqmovn_s32(vacc2x4567));
255 
256     vacc0x01234567 = vqaddq_s16(vacc0x01234567, voutput_zero_point);
257     vacc1x01234567 = vqaddq_s16(vacc1x01234567, voutput_zero_point);
258     vacc2x01234567 = vqaddq_s16(vacc2x01234567, voutput_zero_point);
259 
260     int8x16_t vout0x01234567_1x01234567 = vcombine_s8(vqmovn_s16(vacc0x01234567), vqmovn_s16(vacc1x01234567));
261     int8x8_t vout2x01234567 = vqmovn_s16(vacc2x01234567);
262 #endif
263 
264     const int8x16_t voutput_min = vld1q_dup_s8(&params->rndnu_neon.output_min);
265     vout0x01234567_1x01234567 = vmaxq_s8(vout0x01234567_1x01234567, voutput_min);
266     vout2x01234567 = vmax_s8(vout2x01234567, vget_low_s8(voutput_min));
267 
268     const int8x16_t voutput_max = vld1q_dup_s8(&params->rndnu_neon.output_max);
269     vout0x01234567_1x01234567 = vminq_s8(vout0x01234567_1x01234567, voutput_max);
270     vout2x01234567 = vmin_s8(vout2x01234567, vget_low_s8(voutput_max));
271 
272     if (nc >= 8) {
273       vst1_s8(c0 + 0, vget_low_s8(vout0x01234567_1x01234567));
274       vst1_s8(c1 + 0, vget_high_s8(vout0x01234567_1x01234567));
275       vst1_s8(c2 + 0, vout2x01234567);
276 
277       c0 = (int8_t*) ((uintptr_t) c0 + cn_stride);
278       c1 = (int8_t*) ((uintptr_t) c1 + cn_stride);
279       c2 = (int8_t*) ((uintptr_t) c2 + cn_stride);
280 
281       a0 = (const int8_t*) ((uintptr_t) a0 - kc);
282       a1 = (const int8_t*) ((uintptr_t) a1 - kc);
283       a2 = (const int8_t*) ((uintptr_t) a2 - kc);
284 
285       nc -= 8;
286     } else {
287       // Final case where not all of the 8 columns fit in the destination.
288       if (nc & 4) {
289         vst1q_lane_u32((void*) c0, vreinterpretq_u32_s8(vout0x01234567_1x01234567), 0); c0 += 4;
290         vst1q_lane_u32((void*) c1, vreinterpretq_u32_s8(vout0x01234567_1x01234567), 2); c1 += 4;
291         vst1_lane_u32((void*) c2, vreinterpret_u32_s8(vout2x01234567), 0); c2 += 4;
292         vout0x01234567_1x01234567 = vextq_s8(vout0x01234567_1x01234567, vout0x01234567_1x01234567, 4);
293         vout2x01234567 = vext_s8(vout2x01234567, vout2x01234567, 4);
294       }
295       if (nc & 2) {
296         vst1q_lane_u16((void*) c0, vreinterpretq_u16_s8(vout0x01234567_1x01234567), 0); c0 += 2;
297         vst1q_lane_u16((void*) c1, vreinterpretq_u16_s8(vout0x01234567_1x01234567), 4); c1 += 2;
298         vst1_lane_u16((void*) c2, vreinterpret_u16_s8(vout2x01234567), 0); c2 += 2;
299         vout0x01234567_1x01234567 = vextq_s8(vout0x01234567_1x01234567, vout0x01234567_1x01234567, 2);
300         vout2x01234567 = vext_s8(vout2x01234567, vout2x01234567, 2);
301       }
302       if (nc & 1) {
303         vst1q_lane_s8(c0, vout0x01234567_1x01234567, 0);
304         vst1q_lane_s8(c1, vout0x01234567_1x01234567, 8);
305         vst1_lane_s8(c2, vout2x01234567, 0);
306       }
307 
308       nc = 0;
309     }
310   } while (nc != 0);
311 }
312