xref: /aosp_15_r20/external/XNNPACK/src/qs8-igemm/gen/4x8c2s4-minmax-rndnu-neon-mull.c (revision 4bdc94577ba0e567308109d787f7fec7b531ce36)
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_4x8c2s4__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_4x8c2s4__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 <= 4);
34   assert(nc != 0);
35   assert(kc != 0);
36   assert(ks != 0);
37   assert(ks % (4 * 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   int8_t* c3 = (int8_t*) ((uintptr_t) c2 + cm_stride);
53   if XNN_UNPREDICTABLE(mr != 4) {
54     c3 = c2;
55   }
56 
57   kc = round_up_po2(kc, 8 * sizeof(int8_t));
58   do {
59     int32x4_t vacc0x0123 = vld1q_s32(w); w = (const int32_t*) w + 4;
60     int32x4_t vacc0x4567 = vld1q_s32(w); w = (const int32_t*) w + 4;
61     int32x4_t vacc1x0123 = vacc0x0123;
62     int32x4_t vacc1x4567 = vacc0x4567;
63     int32x4_t vacc2x0123 = vacc0x0123;
64     int32x4_t vacc2x4567 = vacc0x4567;
65     int32x4_t vacc3x0123 = vacc0x0123;
66     int32x4_t vacc3x4567 = vacc0x4567;
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       const int8_t* restrict a3 = a[3];
83       if XNN_UNPREDICTABLE(a3 != zero) {
84         a3 = (const int8_t*) ((uintptr_t) a3 + a_offset);
85       }
86       a += 4;
87 
88       size_t k = kc;
89       do {
90         int8x8_t va0x0 = vld1_s8(a0); a0 += 8;
91         int8x8_t va1x0 = vld1_s8(a1); a1 += 8;
92         int8x8_t va2x0 = vld1_s8(a2); a2 += 8;
93         int8x8_t va3x0 = vld1_s8(a3); a3 += 8;
94 
95         const int8x8_t vb0123c0x0 = vld1_s8(w); w = (const int8_t*) w + 8;
96         const int8x8_t vb4567c0x0 = vld1_s8(w); w = (const int8_t*) w + 8;
97         const int8x8_t vb0123c1x0 = vld1_s8(w); w = (const int8_t*) w + 8;
98         const int8x8_t vb4567c1x0 = vld1_s8(w); w = (const int8_t*) w + 8;
99         const int8x8_t vb0123c2x0 = vld1_s8(w); w = (const int8_t*) w + 8;
100         const int8x8_t vb4567c2x0 = vld1_s8(w); w = (const int8_t*) w + 8;
101         const int8x8_t vb0123c3x0 = vld1_s8(w); w = (const int8_t*) w + 8;
102         const int8x8_t vb4567c3x0 = vld1_s8(w); w = (const int8_t*) w + 8;
103 
104         int16x8_t vprod0x0123c0 = vmull_s8(vb0123c0x0, va0x0);
105         int16x8_t vprod1x0123c0 = vmull_s8(vb0123c0x0, va1x0);
106         int16x8_t vprod2x0123c0 = vmull_s8(vb0123c0x0, va2x0);
107         int16x8_t vprod3x0123c0 = vmull_s8(vb0123c0x0, va3x0);
108         vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c0);
109         vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c0);
110         vacc2x0123 = vpadalq_s16(vacc2x0123, vprod2x0123c0);
111         vacc3x0123 = vpadalq_s16(vacc3x0123, vprod3x0123c0);
112         int16x8_t vprod0x4567c0 = vmull_s8(vb4567c0x0, va0x0);
113         int16x8_t vprod1x4567c0 = vmull_s8(vb4567c0x0, va1x0);
114         int16x8_t vprod2x4567c0 = vmull_s8(vb4567c0x0, va2x0);
115         int16x8_t vprod3x4567c0 = vmull_s8(vb4567c0x0, va3x0);
116         vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c0);
117         vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c0);
118         vacc2x4567 = vpadalq_s16(vacc2x4567, vprod2x4567c0);
119         vacc3x4567 = vpadalq_s16(vacc3x4567, vprod3x4567c0);
120         va0x0 = vext_s8(va0x0, va0x0, 2);
121         va1x0 = vext_s8(va1x0, va1x0, 2);
122         va2x0 = vext_s8(va2x0, va2x0, 2);
123         va3x0 = vext_s8(va3x0, va3x0, 2);
124         int16x8_t vprod0x0123c1 = vmull_s8(vb0123c1x0, va0x0);
125         int16x8_t vprod1x0123c1 = vmull_s8(vb0123c1x0, va1x0);
126         int16x8_t vprod2x0123c1 = vmull_s8(vb0123c1x0, va2x0);
127         int16x8_t vprod3x0123c1 = vmull_s8(vb0123c1x0, va3x0);
128         vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c1);
129         vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c1);
130         vacc2x0123 = vpadalq_s16(vacc2x0123, vprod2x0123c1);
131         vacc3x0123 = vpadalq_s16(vacc3x0123, vprod3x0123c1);
132         int16x8_t vprod0x4567c1 = vmull_s8(vb4567c1x0, va0x0);
133         int16x8_t vprod1x4567c1 = vmull_s8(vb4567c1x0, va1x0);
134         int16x8_t vprod2x4567c1 = vmull_s8(vb4567c1x0, va2x0);
135         int16x8_t vprod3x4567c1 = vmull_s8(vb4567c1x0, va3x0);
136         vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c1);
137         vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c1);
138         vacc2x4567 = vpadalq_s16(vacc2x4567, vprod2x4567c1);
139         vacc3x4567 = vpadalq_s16(vacc3x4567, vprod3x4567c1);
140         va0x0 = vext_s8(va0x0, va0x0, 2);
141         va1x0 = vext_s8(va1x0, va1x0, 2);
142         va2x0 = vext_s8(va2x0, va2x0, 2);
143         va3x0 = vext_s8(va3x0, va3x0, 2);
144         int16x8_t vprod0x0123c2 = vmull_s8(vb0123c2x0, va0x0);
145         int16x8_t vprod1x0123c2 = vmull_s8(vb0123c2x0, va1x0);
146         int16x8_t vprod2x0123c2 = vmull_s8(vb0123c2x0, va2x0);
147         int16x8_t vprod3x0123c2 = vmull_s8(vb0123c2x0, va3x0);
148         vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c2);
149         vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c2);
150         vacc2x0123 = vpadalq_s16(vacc2x0123, vprod2x0123c2);
151         vacc3x0123 = vpadalq_s16(vacc3x0123, vprod3x0123c2);
152         int16x8_t vprod0x4567c2 = vmull_s8(vb4567c2x0, va0x0);
153         int16x8_t vprod1x4567c2 = vmull_s8(vb4567c2x0, va1x0);
154         int16x8_t vprod2x4567c2 = vmull_s8(vb4567c2x0, va2x0);
155         int16x8_t vprod3x4567c2 = vmull_s8(vb4567c2x0, va3x0);
156         vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c2);
157         vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c2);
158         vacc2x4567 = vpadalq_s16(vacc2x4567, vprod2x4567c2);
159         vacc3x4567 = vpadalq_s16(vacc3x4567, vprod3x4567c2);
160         va0x0 = vext_s8(va0x0, va0x0, 2);
161         va1x0 = vext_s8(va1x0, va1x0, 2);
162         va2x0 = vext_s8(va2x0, va2x0, 2);
163         va3x0 = vext_s8(va3x0, va3x0, 2);
164         int16x8_t vprod0x0123c3 = vmull_s8(vb0123c3x0, va0x0);
165         int16x8_t vprod1x0123c3 = vmull_s8(vb0123c3x0, va1x0);
166         int16x8_t vprod2x0123c3 = vmull_s8(vb0123c3x0, va2x0);
167         int16x8_t vprod3x0123c3 = vmull_s8(vb0123c3x0, va3x0);
168         vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c3);
169         vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c3);
170         vacc2x0123 = vpadalq_s16(vacc2x0123, vprod2x0123c3);
171         vacc3x0123 = vpadalq_s16(vacc3x0123, vprod3x0123c3);
172         int16x8_t vprod0x4567c3 = vmull_s8(vb4567c3x0, va0x0);
173         int16x8_t vprod1x4567c3 = vmull_s8(vb4567c3x0, va1x0);
174         int16x8_t vprod2x4567c3 = vmull_s8(vb4567c3x0, va2x0);
175         int16x8_t vprod3x4567c3 = vmull_s8(vb4567c3x0, va3x0);
176         vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c3);
177         vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c3);
178         vacc2x4567 = vpadalq_s16(vacc2x4567, vprod2x4567c3);
179         vacc3x4567 = vpadalq_s16(vacc3x4567, vprod3x4567c3);
180 
181         k -= 8 * sizeof(int8_t);
182       } while (k != 0);
183 
184       p -= 4 * sizeof(void*);
185     } while (p != 0);
186 
187     const int32x4_t vright_pre_shift = vld1q_dup_s32(&params->rndnu_neon.right_pre_shift);
188     const int32x4_t vmultiplier = vld1q_dup_s32(&params->rndnu_neon.multiplier);
189     const int32x4_t vright_post_shift = vld1q_dup_s32(&params->rndnu_neon.right_post_shift);
190 
191     vacc0x0123 = vqshlq_s32(vacc0x0123, vright_pre_shift);
192     vacc0x4567 = vqshlq_s32(vacc0x4567, vright_pre_shift);
193     vacc1x0123 = vqshlq_s32(vacc1x0123, vright_pre_shift);
194     vacc1x4567 = vqshlq_s32(vacc1x4567, vright_pre_shift);
195     vacc2x0123 = vqshlq_s32(vacc2x0123, vright_pre_shift);
196     vacc2x4567 = vqshlq_s32(vacc2x4567, vright_pre_shift);
197     vacc3x0123 = vqshlq_s32(vacc3x0123, vright_pre_shift);
198     vacc3x4567 = vqshlq_s32(vacc3x4567, vright_pre_shift);
199 
200     vacc0x0123 = vqdmulhq_s32(vacc0x0123, vmultiplier);
201     vacc0x4567 = vqdmulhq_s32(vacc0x4567, vmultiplier);
202     vacc1x0123 = vqdmulhq_s32(vacc1x0123, vmultiplier);
203     vacc1x4567 = vqdmulhq_s32(vacc1x4567, vmultiplier);
204     vacc2x0123 = vqdmulhq_s32(vacc2x0123, vmultiplier);
205     vacc2x4567 = vqdmulhq_s32(vacc2x4567, vmultiplier);
206     vacc3x0123 = vqdmulhq_s32(vacc3x0123, vmultiplier);
207     vacc3x4567 = vqdmulhq_s32(vacc3x4567, vmultiplier);
208 
209     vacc0x0123 = vrshlq_s32(vacc0x0123, vright_post_shift);
210     vacc0x4567 = vrshlq_s32(vacc0x4567, vright_post_shift);
211     vacc1x0123 = vrshlq_s32(vacc1x0123, vright_post_shift);
212     vacc1x4567 = vrshlq_s32(vacc1x4567, vright_post_shift);
213     vacc2x0123 = vrshlq_s32(vacc2x0123, vright_post_shift);
214     vacc2x4567 = vrshlq_s32(vacc2x4567, vright_post_shift);
215     vacc3x0123 = vrshlq_s32(vacc3x0123, vright_post_shift);
216     vacc3x4567 = vrshlq_s32(vacc3x4567, vright_post_shift);
217 
218     const int16x8_t voutput_zero_point = vld1q_dup_s16(&params->rndnu_neon.output_zero_point);
219 #if XNN_ARCH_ARM64
220     int16x8_t vacc0x01234567 = vqmovn_high_s32(vqmovn_s32(vacc0x0123), vacc0x4567);
221     int16x8_t vacc1x01234567 = vqmovn_high_s32(vqmovn_s32(vacc1x0123), vacc1x4567);
222     int16x8_t vacc2x01234567 = vqmovn_high_s32(vqmovn_s32(vacc2x0123), vacc2x4567);
223     int16x8_t vacc3x01234567 = vqmovn_high_s32(vqmovn_s32(vacc3x0123), vacc3x4567);
224 
225     vacc0x01234567 = vqaddq_s16(vacc0x01234567, voutput_zero_point);
226     vacc1x01234567 = vqaddq_s16(vacc1x01234567, voutput_zero_point);
227     vacc2x01234567 = vqaddq_s16(vacc2x01234567, voutput_zero_point);
228     vacc3x01234567 = vqaddq_s16(vacc3x01234567, voutput_zero_point);
229 
230     int8x16_t vout0x01234567_1x01234567 = vqmovn_high_s16(vqmovn_s16(vacc0x01234567), vacc1x01234567);
231     int8x16_t vout2x01234567_3x01234567 = vqmovn_high_s16(vqmovn_s16(vacc2x01234567), vacc3x01234567);
232 #else
233     int16x8_t vacc0x01234567 = vcombine_s16(vqmovn_s32(vacc0x0123), vqmovn_s32(vacc0x4567));
234     int16x8_t vacc1x01234567 = vcombine_s16(vqmovn_s32(vacc1x0123), vqmovn_s32(vacc1x4567));
235     int16x8_t vacc2x01234567 = vcombine_s16(vqmovn_s32(vacc2x0123), vqmovn_s32(vacc2x4567));
236     int16x8_t vacc3x01234567 = vcombine_s16(vqmovn_s32(vacc3x0123), vqmovn_s32(vacc3x4567));
237 
238     vacc0x01234567 = vqaddq_s16(vacc0x01234567, voutput_zero_point);
239     vacc1x01234567 = vqaddq_s16(vacc1x01234567, voutput_zero_point);
240     vacc2x01234567 = vqaddq_s16(vacc2x01234567, voutput_zero_point);
241     vacc3x01234567 = vqaddq_s16(vacc3x01234567, voutput_zero_point);
242 
243     int8x16_t vout0x01234567_1x01234567 = vcombine_s8(vqmovn_s16(vacc0x01234567), vqmovn_s16(vacc1x01234567));
244     int8x16_t vout2x01234567_3x01234567 = vcombine_s8(vqmovn_s16(vacc2x01234567), vqmovn_s16(vacc3x01234567));
245 #endif
246 
247     const int8x16_t voutput_min = vld1q_dup_s8(&params->rndnu_neon.output_min);
248     vout0x01234567_1x01234567 = vmaxq_s8(vout0x01234567_1x01234567, voutput_min);
249     vout2x01234567_3x01234567 = vmaxq_s8(vout2x01234567_3x01234567, voutput_min);
250 
251     const int8x16_t voutput_max = vld1q_dup_s8(&params->rndnu_neon.output_max);
252     vout0x01234567_1x01234567 = vminq_s8(vout0x01234567_1x01234567, voutput_max);
253     vout2x01234567_3x01234567 = vminq_s8(vout2x01234567_3x01234567, voutput_max);
254 
255     if (nc >= 8) {
256       vst1_s8(c3 + 0, vget_high_s8(vout2x01234567_3x01234567));
257       vst1_s8(c2 + 0, vget_low_s8(vout2x01234567_3x01234567));
258       vst1_s8(c1 + 0, vget_high_s8(vout0x01234567_1x01234567));
259       vst1_s8(c0 + 0, vget_low_s8(vout0x01234567_1x01234567));
260 
261       c3 = (int8_t*) ((uintptr_t) c3 + cn_stride);
262       c2 = (int8_t*) ((uintptr_t) c2 + cn_stride);
263       c1 = (int8_t*) ((uintptr_t) c1 + cn_stride);
264       c0 = (int8_t*) ((uintptr_t) c0 + cn_stride);
265 
266       a = (const int8_t**restrict) ((uintptr_t) a - ks);
267 
268       nc -= 8;
269     } else {
270       if (nc & 4) {
271         vst1q_lane_u32((void*) c3, vreinterpretq_u32_s8(vout2x01234567_3x01234567), 2); c3 += 4;
272         vst1q_lane_u32((void*) c2, vreinterpretq_u32_s8(vout2x01234567_3x01234567), 0); c2 += 4;
273         vst1q_lane_u32((void*) c1, vreinterpretq_u32_s8(vout0x01234567_1x01234567), 2); c1 += 4;
274         vst1q_lane_u32((void*) c0, vreinterpretq_u32_s8(vout0x01234567_1x01234567), 0); c0 += 4;
275         vout2x01234567_3x01234567 = vextq_s8(vout2x01234567_3x01234567, vout2x01234567_3x01234567, 4);
276         vout0x01234567_1x01234567 = vextq_s8(vout0x01234567_1x01234567, vout0x01234567_1x01234567, 4);
277       }
278       if (nc & 2) {
279         vst1q_lane_u16((void*) c3, vreinterpretq_u16_s8(vout2x01234567_3x01234567), 4); c3 += 2;
280         vst1q_lane_u16((void*) c2, vreinterpretq_u16_s8(vout2x01234567_3x01234567), 0); c2 += 2;
281         vst1q_lane_u16((void*) c1, vreinterpretq_u16_s8(vout0x01234567_1x01234567), 4); c1 += 2;
282         vst1q_lane_u16((void*) c0, vreinterpretq_u16_s8(vout0x01234567_1x01234567), 0); c0 += 2;
283         vout2x01234567_3x01234567 = vextq_s8(vout2x01234567_3x01234567, vout2x01234567_3x01234567, 2);
284         vout0x01234567_1x01234567 = vextq_s8(vout0x01234567_1x01234567, vout0x01234567_1x01234567, 2);
285       }
286       if (nc & 1) {
287         vst1q_lane_s8(c3, vout2x01234567_3x01234567, 8);
288         vst1q_lane_s8(c2, vout2x01234567_3x01234567, 0);
289         vst1q_lane_s8(c1, vout0x01234567_1x01234567, 8);
290         vst1q_lane_s8(c0, vout0x01234567_1x01234567, 0);
291       }
292 
293       nc = 0;
294     }
295   } while (nc != 0);
296 }
297