xref: /aosp_15_r20/external/XNNPACK/src/qs8-gemm/gen/3x8c4-minmax-rndnu-neon-mull-ld1r.c (revision 4bdc94577ba0e567308109d787f7fec7b531ce36)
1 // Auto-generated file. Do not edit!
2 //   Template: src/qs8-gemm/c4-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 
17 
xnn_qs8_gemm_minmax_rndnu_ukernel_3x8c4__neon_mull_ld1r(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_3x8c4__neon_mull_ld1r(
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 <= 3);
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   kc = round_up_po2(kc, 4 * sizeof(int8_t));
40   const int8_t* a0 = a;
41   int8_t* c0 = c;
42   const int8_t* a1 = (const int8_t*) ((uintptr_t) a0 + a_stride);
43   int8_t* c1 = (int8_t*) ((uintptr_t) c0 + cm_stride);
44   if XNN_UNPREDICTABLE(mr < 2) {
45     a1 = a0;
46     c1 = c0;
47   }
48   const int8_t* a2 = (const int8_t*) ((uintptr_t) a1 + a_stride);
49   int8_t* c2 = (int8_t*) ((uintptr_t) c1 + cm_stride);
50   if XNN_UNPREDICTABLE(mr <= 2) {
51     a2 = a1;
52     c2 = c1;
53   }
54 
55   do {
56     int32x4_t vacc0x01 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
57     int32x4_t vacc0x23 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
58     int32x4_t vacc0x45 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
59     int32x4_t vacc0x67 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
60     int32x4_t vacc1x01 = vacc0x01;
61     int32x4_t vacc1x23 = vacc0x23;
62     int32x4_t vacc1x45 = vacc0x45;
63     int32x4_t vacc1x67 = vacc0x67;
64     int32x4_t vacc2x01 = vacc0x01;
65     int32x4_t vacc2x23 = vacc0x23;
66     int32x4_t vacc2x45 = vacc0x45;
67     int32x4_t vacc2x67 = vacc0x67;
68 
69     size_t k = kc;
70 
71 
72     while (k >= 8 * sizeof(int8_t)) {
73       const int32x2_t va00 = vld1_dup_s32((const void*)a0);
74       const int32x2_t va01 = vld1_dup_s32((const void*)(a0 + 4)); a0 += 8;
75       const int32x2_t va10 = vld1_dup_s32((const void*)a1);
76       const int32x2_t va11 = vld1_dup_s32((const void*)(a1 + 4)); a1 += 8;
77       const int32x2_t va20 = vld1_dup_s32((const void*)a2);
78       const int32x2_t va21 = vld1_dup_s32((const void*)(a2 + 4)); a2 += 8;
79 
80       const int8x8_t vb01c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
81       const int8x8_t vb23c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
82       const int8x8_t vb45c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
83       const int8x8_t vb67c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
84       const int8x8_t vb01c1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
85       const int8x8_t vb23c1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
86       const int8x8_t vb45c1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
87       const int8x8_t vb67c1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
88 
89       const int8x8_t va0c0 = vreinterpret_s8_s32(va00);
90       const int8x8_t va1c0 = vreinterpret_s8_s32(va10);
91       const int8x8_t va2c0 = vreinterpret_s8_s32(va20);
92 
93       const int16x8_t vprod0x01c0 = vmull_s8(vb01c0, va0c0);
94       const int16x8_t vprod1x01c0 = vmull_s8(vb01c0, va1c0);
95       const int16x8_t vprod2x01c0 = vmull_s8(vb01c0, va2c0);
96       vacc0x01 = vpadalq_s16(vacc0x01, vprod0x01c0);
97       vacc1x01 = vpadalq_s16(vacc1x01, vprod1x01c0);
98       vacc2x01 = vpadalq_s16(vacc2x01, vprod2x01c0);
99       const int16x8_t vprod0x23c0 = vmull_s8(vb23c0, va0c0);
100       const int16x8_t vprod1x23c0 = vmull_s8(vb23c0, va1c0);
101       const int16x8_t vprod2x23c0 = vmull_s8(vb23c0, va2c0);
102       vacc0x23 = vpadalq_s16(vacc0x23, vprod0x23c0);
103       vacc1x23 = vpadalq_s16(vacc1x23, vprod1x23c0);
104       vacc2x23 = vpadalq_s16(vacc2x23, vprod2x23c0);
105       const int16x8_t vprod0x45c0 = vmull_s8(vb45c0, va0c0);
106       const int16x8_t vprod1x45c0 = vmull_s8(vb45c0, va1c0);
107       const int16x8_t vprod2x45c0 = vmull_s8(vb45c0, va2c0);
108       vacc0x45 = vpadalq_s16(vacc0x45, vprod0x45c0);
109       vacc1x45 = vpadalq_s16(vacc1x45, vprod1x45c0);
110       vacc2x45 = vpadalq_s16(vacc2x45, vprod2x45c0);
111       const int16x8_t vprod0x67c0 = vmull_s8(vb67c0, va0c0);
112       const int16x8_t vprod1x67c0 = vmull_s8(vb67c0, va1c0);
113       const int16x8_t vprod2x67c0 = vmull_s8(vb67c0, va2c0);
114       vacc0x67 = vpadalq_s16(vacc0x67, vprod0x67c0);
115       vacc1x67 = vpadalq_s16(vacc1x67, vprod1x67c0);
116       vacc2x67 = vpadalq_s16(vacc2x67, vprod2x67c0);
117       const int8x8_t va0c1 = vreinterpret_s8_s32(va01);
118       const int8x8_t va1c1 = vreinterpret_s8_s32(va11);
119       const int8x8_t va2c1 = vreinterpret_s8_s32(va21);
120 
121       const int16x8_t vprod0x01c1 = vmull_s8(vb01c1, va0c1);
122       const int16x8_t vprod1x01c1 = vmull_s8(vb01c1, va1c1);
123       const int16x8_t vprod2x01c1 = vmull_s8(vb01c1, va2c1);
124       vacc0x01 = vpadalq_s16(vacc0x01, vprod0x01c1);
125       vacc1x01 = vpadalq_s16(vacc1x01, vprod1x01c1);
126       vacc2x01 = vpadalq_s16(vacc2x01, vprod2x01c1);
127       const int16x8_t vprod0x23c1 = vmull_s8(vb23c1, va0c1);
128       const int16x8_t vprod1x23c1 = vmull_s8(vb23c1, va1c1);
129       const int16x8_t vprod2x23c1 = vmull_s8(vb23c1, va2c1);
130       vacc0x23 = vpadalq_s16(vacc0x23, vprod0x23c1);
131       vacc1x23 = vpadalq_s16(vacc1x23, vprod1x23c1);
132       vacc2x23 = vpadalq_s16(vacc2x23, vprod2x23c1);
133       const int16x8_t vprod0x45c1 = vmull_s8(vb45c1, va0c1);
134       const int16x8_t vprod1x45c1 = vmull_s8(vb45c1, va1c1);
135       const int16x8_t vprod2x45c1 = vmull_s8(vb45c1, va2c1);
136       vacc0x45 = vpadalq_s16(vacc0x45, vprod0x45c1);
137       vacc1x45 = vpadalq_s16(vacc1x45, vprod1x45c1);
138       vacc2x45 = vpadalq_s16(vacc2x45, vprod2x45c1);
139       const int16x8_t vprod0x67c1 = vmull_s8(vb67c1, va0c1);
140       const int16x8_t vprod1x67c1 = vmull_s8(vb67c1, va1c1);
141       const int16x8_t vprod2x67c1 = vmull_s8(vb67c1, va2c1);
142       vacc0x67 = vpadalq_s16(vacc0x67, vprod0x67c1);
143       vacc1x67 = vpadalq_s16(vacc1x67, vprod1x67c1);
144       vacc2x67 = vpadalq_s16(vacc2x67, vprod2x67c1);
145 
146       k -= 8 * sizeof(int8_t);
147     }
148 
149     if XNN_UNLIKELY(k != 0) {
150       const int8x8_t va0 = vld1_s8(a0); a0 = (const int8_t*) ((uintptr_t) a0 + k);
151       const int8x8_t va1 = vld1_s8(a1); a1 = (const int8_t*) ((uintptr_t) a1 + k);
152       const int8x8_t va2 = vld1_s8(a2); a2 = (const int8_t*) ((uintptr_t) a2 + k);
153 
154       const int8x8_t vb01c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
155       const int8x8_t vb23c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
156       const int8x8_t vb45c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
157       const int8x8_t vb67c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
158 
159       const int8x8_t va0c0 = vreinterpret_s8_s32(vdup_lane_s32(vreinterpret_s32_s8(va0), 0));
160       const int16x8_t vprod0x01c0 = vmull_s8(vb01c0, va0c0);
161       vacc0x01 = vpadalq_s16(vacc0x01, vprod0x01c0);
162       const int16x8_t vprod0x23c0 = vmull_s8(vb23c0, va0c0);
163       vacc0x23 = vpadalq_s16(vacc0x23, vprod0x23c0);
164       const int16x8_t vprod0x45c0 = vmull_s8(vb45c0, va0c0);
165       vacc0x45 = vpadalq_s16(vacc0x45, vprod0x45c0);
166       const int16x8_t vprod0x67c0 = vmull_s8(vb67c0, va0c0);
167       vacc0x67 = vpadalq_s16(vacc0x67, vprod0x67c0);
168       const int8x8_t va1c0 = vreinterpret_s8_s32(vdup_lane_s32(vreinterpret_s32_s8(va1), 0));
169       const int16x8_t vprod1x01c0 = vmull_s8(vb01c0, va1c0);
170       vacc1x01 = vpadalq_s16(vacc1x01, vprod1x01c0);
171       const int16x8_t vprod1x23c0 = vmull_s8(vb23c0, va1c0);
172       vacc1x23 = vpadalq_s16(vacc1x23, vprod1x23c0);
173       const int16x8_t vprod1x45c0 = vmull_s8(vb45c0, va1c0);
174       vacc1x45 = vpadalq_s16(vacc1x45, vprod1x45c0);
175       const int16x8_t vprod1x67c0 = vmull_s8(vb67c0, va1c0);
176       vacc1x67 = vpadalq_s16(vacc1x67, vprod1x67c0);
177       const int8x8_t va2c0 = vreinterpret_s8_s32(vdup_lane_s32(vreinterpret_s32_s8(va2), 0));
178       const int16x8_t vprod2x01c0 = vmull_s8(vb01c0, va2c0);
179       vacc2x01 = vpadalq_s16(vacc2x01, vprod2x01c0);
180       const int16x8_t vprod2x23c0 = vmull_s8(vb23c0, va2c0);
181       vacc2x23 = vpadalq_s16(vacc2x23, vprod2x23c0);
182       const int16x8_t vprod2x45c0 = vmull_s8(vb45c0, va2c0);
183       vacc2x45 = vpadalq_s16(vacc2x45, vprod2x45c0);
184       const int16x8_t vprod2x67c0 = vmull_s8(vb67c0, va2c0);
185       vacc2x67 = vpadalq_s16(vacc2x67, vprod2x67c0);
186     }
187 
188 #if XNN_ARCH_ARM64
189     int32x4_t vacc0x0123 = vpaddq_s32(vacc0x01, vacc0x23);
190     int32x4_t vacc0x4567 = vpaddq_s32(vacc0x45, vacc0x67);
191     int32x4_t vacc1x0123 = vpaddq_s32(vacc1x01, vacc1x23);
192     int32x4_t vacc1x4567 = vpaddq_s32(vacc1x45, vacc1x67);
193     int32x4_t vacc2x0123 = vpaddq_s32(vacc2x01, vacc2x23);
194     int32x4_t vacc2x4567 = vpaddq_s32(vacc2x45, vacc2x67);
195 #else
196     const int32x2_t vsum0x01 = vpadd_s32(vget_low_s32(vacc0x01), vget_high_s32(vacc0x01));
197     const int32x2_t vsum0x23 = vpadd_s32(vget_low_s32(vacc0x23), vget_high_s32(vacc0x23));
198     int32x4_t vacc0x0123 = vcombine_s32(vsum0x01, vsum0x23);
199     const int32x2_t vsum0x45 = vpadd_s32(vget_low_s32(vacc0x45), vget_high_s32(vacc0x45));
200     const int32x2_t vsum0x67 = vpadd_s32(vget_low_s32(vacc0x67), vget_high_s32(vacc0x67));
201     int32x4_t vacc0x4567 = vcombine_s32(vsum0x45, vsum0x67);
202     const int32x2_t vsum1x01 = vpadd_s32(vget_low_s32(vacc1x01), vget_high_s32(vacc1x01));
203     const int32x2_t vsum1x23 = vpadd_s32(vget_low_s32(vacc1x23), vget_high_s32(vacc1x23));
204     int32x4_t vacc1x0123 = vcombine_s32(vsum1x01, vsum1x23);
205     const int32x2_t vsum1x45 = vpadd_s32(vget_low_s32(vacc1x45), vget_high_s32(vacc1x45));
206     const int32x2_t vsum1x67 = vpadd_s32(vget_low_s32(vacc1x67), vget_high_s32(vacc1x67));
207     int32x4_t vacc1x4567 = vcombine_s32(vsum1x45, vsum1x67);
208     const int32x2_t vsum2x01 = vpadd_s32(vget_low_s32(vacc2x01), vget_high_s32(vacc2x01));
209     const int32x2_t vsum2x23 = vpadd_s32(vget_low_s32(vacc2x23), vget_high_s32(vacc2x23));
210     int32x4_t vacc2x0123 = vcombine_s32(vsum2x01, vsum2x23);
211     const int32x2_t vsum2x45 = vpadd_s32(vget_low_s32(vacc2x45), vget_high_s32(vacc2x45));
212     const int32x2_t vsum2x67 = vpadd_s32(vget_low_s32(vacc2x67), vget_high_s32(vacc2x67));
213     int32x4_t vacc2x4567 = vcombine_s32(vsum2x45, vsum2x67);
214 #endif
215 
216     const int32x4_t vright_pre_shift = vld1q_dup_s32(&params->rndnu_neon.right_pre_shift);
217     const int32x4_t vmultiplier = vld1q_dup_s32(&params->rndnu_neon.multiplier);
218     const int32x4_t vright_post_shift = vld1q_dup_s32(&params->rndnu_neon.right_post_shift);
219 
220     vacc0x0123 = vqshlq_s32(vacc0x0123, vright_pre_shift);
221     vacc0x4567 = vqshlq_s32(vacc0x4567, vright_pre_shift);
222     vacc1x0123 = vqshlq_s32(vacc1x0123, vright_pre_shift);
223     vacc1x4567 = vqshlq_s32(vacc1x4567, vright_pre_shift);
224     vacc2x0123 = vqshlq_s32(vacc2x0123, vright_pre_shift);
225     vacc2x4567 = vqshlq_s32(vacc2x4567, vright_pre_shift);
226 
227     vacc0x0123 = vqdmulhq_s32(vacc0x0123, vmultiplier);
228     vacc0x4567 = vqdmulhq_s32(vacc0x4567, vmultiplier);
229     vacc1x0123 = vqdmulhq_s32(vacc1x0123, vmultiplier);
230     vacc1x4567 = vqdmulhq_s32(vacc1x4567, vmultiplier);
231     vacc2x0123 = vqdmulhq_s32(vacc2x0123, vmultiplier);
232     vacc2x4567 = vqdmulhq_s32(vacc2x4567, vmultiplier);
233 
234     vacc0x0123 = vrshlq_s32(vacc0x0123, vright_post_shift);
235     vacc0x4567 = vrshlq_s32(vacc0x4567, vright_post_shift);
236     vacc1x0123 = vrshlq_s32(vacc1x0123, vright_post_shift);
237     vacc1x4567 = vrshlq_s32(vacc1x4567, vright_post_shift);
238     vacc2x0123 = vrshlq_s32(vacc2x0123, vright_post_shift);
239     vacc2x4567 = vrshlq_s32(vacc2x4567, vright_post_shift);
240 
241     const int16x8_t voutput_zero_point = vld1q_dup_s16(&params->rndnu_neon.output_zero_point);
242 #if XNN_ARCH_ARM64
243     int16x8_t vacc0x01234567 = vqmovn_high_s32(vqmovn_s32(vacc0x0123), vacc0x4567);
244     int16x8_t vacc1x01234567 = vqmovn_high_s32(vqmovn_s32(vacc1x0123), vacc1x4567);
245     int16x8_t vacc2x01234567 = vqmovn_high_s32(vqmovn_s32(vacc2x0123), vacc2x4567);
246 
247     vacc0x01234567 = vqaddq_s16(vacc0x01234567, voutput_zero_point);
248     vacc1x01234567 = vqaddq_s16(vacc1x01234567, voutput_zero_point);
249     vacc2x01234567 = vqaddq_s16(vacc2x01234567, voutput_zero_point);
250 
251     int8x16_t vout0x01234567_1x01234567 = vqmovn_high_s16(vqmovn_s16(vacc0x01234567), vacc1x01234567);
252     int8x8_t vout2x01234567 = vqmovn_s16(vacc2x01234567);
253 #else
254     int16x8_t vacc0x01234567 = vcombine_s16(vqmovn_s32(vacc0x0123), vqmovn_s32(vacc0x4567));
255     int16x8_t vacc1x01234567 = vcombine_s16(vqmovn_s32(vacc1x0123), vqmovn_s32(vacc1x4567));
256     int16x8_t vacc2x01234567 = vcombine_s16(vqmovn_s32(vacc2x0123), vqmovn_s32(vacc2x4567));
257 
258     vacc0x01234567 = vqaddq_s16(vacc0x01234567, voutput_zero_point);
259     vacc1x01234567 = vqaddq_s16(vacc1x01234567, voutput_zero_point);
260     vacc2x01234567 = vqaddq_s16(vacc2x01234567, voutput_zero_point);
261 
262     int8x16_t vout0x01234567_1x01234567 = vcombine_s8(vqmovn_s16(vacc0x01234567), vqmovn_s16(vacc1x01234567));
263     int8x8_t vout2x01234567 = vqmovn_s16(vacc2x01234567);
264 #endif
265 
266     const int8x16_t voutput_min = vld1q_dup_s8(&params->rndnu_neon.output_min);
267     vout0x01234567_1x01234567 = vmaxq_s8(vout0x01234567_1x01234567, voutput_min);
268     vout2x01234567 = vmax_s8(vout2x01234567, vget_low_s8(voutput_min));
269 
270     const int8x16_t voutput_max = vld1q_dup_s8(&params->rndnu_neon.output_max);
271     vout0x01234567_1x01234567 = vminq_s8(vout0x01234567_1x01234567, voutput_max);
272     vout2x01234567 = vmin_s8(vout2x01234567, vget_low_s8(voutput_max));
273 
274     if (nc >= 8) {
275       vst1_s8(c0 + 0, vget_low_s8(vout0x01234567_1x01234567));
276       vst1_s8(c1 + 0, vget_high_s8(vout0x01234567_1x01234567));
277       vst1_s8(c2 + 0, vout2x01234567);
278 
279       c0 = (int8_t*) ((uintptr_t) c0 + cn_stride);
280       c1 = (int8_t*) ((uintptr_t) c1 + cn_stride);
281       c2 = (int8_t*) ((uintptr_t) c2 + cn_stride);
282 
283       a0 = (const int8_t*) ((uintptr_t) a0 - kc);
284       a1 = (const int8_t*) ((uintptr_t) a1 - kc);
285       a2 = (const int8_t*) ((uintptr_t) a2 - kc);
286 
287       nc -= 8;
288     } else {
289       // Final case where not all of the 8 columns fit in the destination.
290       if (nc & 4) {
291         vst1q_lane_u32((void*) c0, vreinterpretq_u32_s8(vout0x01234567_1x01234567), 0); c0 += 4;
292         vst1q_lane_u32((void*) c1, vreinterpretq_u32_s8(vout0x01234567_1x01234567), 2); c1 += 4;
293         vst1_lane_u32((void*) c2, vreinterpret_u32_s8(vout2x01234567), 0); c2 += 4;
294         vout0x01234567_1x01234567 = vextq_s8(vout0x01234567_1x01234567, vout0x01234567_1x01234567, 4);
295         vout2x01234567 = vext_s8(vout2x01234567, vout2x01234567, 4);
296       }
297       if (nc & 2) {
298         vst1q_lane_u16((void*) c0, vreinterpretq_u16_s8(vout0x01234567_1x01234567), 0); c0 += 2;
299         vst1q_lane_u16((void*) c1, vreinterpretq_u16_s8(vout0x01234567_1x01234567), 4); c1 += 2;
300         vst1_lane_u16((void*) c2, vreinterpret_u16_s8(vout2x01234567), 0); c2 += 2;
301         vout0x01234567_1x01234567 = vextq_s8(vout0x01234567_1x01234567, vout0x01234567_1x01234567, 2);
302         vout2x01234567 = vext_s8(vout2x01234567, vout2x01234567, 2);
303       }
304       if (nc & 1) {
305         vst1q_lane_s8(c0, vout0x01234567_1x01234567, 0);
306         vst1q_lane_s8(c1, vout0x01234567_1x01234567, 8);
307         vst1_lane_s8(c2, vout2x01234567, 0);
308       }
309 
310       nc = 0;
311     }
312   } while (nc != 0);
313 }
314