xref: /aosp_15_r20/external/XNNPACK/src/qs8-gemm/gen/3x8c4s2-minmax-rndnu-neon-mull.c (revision 4bdc94577ba0e567308109d787f7fec7b531ce36) 
1 // Auto-generated file. Do not edit!
2 //   Template: src/qs8-gemm/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_gemm_minmax_rndnu_ukernel_3x8c4s2__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_3x8c4s2__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 <= 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   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   kc = round_up_po2(kc, 8 * sizeof(int8_t));
55   do {
56     int32x4_t vacc0x01 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const int32_t*) w + 2;
57     int32x4_t vacc0x23 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const int32_t*) w + 2;
58     int32x4_t vacc0x45 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const int32_t*) w + 2;
59     int32x4_t vacc0x67 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const int32_t*) w + 2;
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     do {
71       int8x8_t va0x0 = vld1_s8(a0); a0 += 8;
72       int8x8_t va1x0 = vld1_s8(a1); a1 += 8;
73       int8x8_t va2x0 = vld1_s8(a2); a2 += 8;
74 
75       const int8x8_t vb01c0x0 = vld1_s8(w); w = (const int8_t*) w + 8;
76       const int8x8_t vb23c0x0 = vld1_s8(w); w = (const int8_t*) w + 8;
77       const int8x8_t vb45c0x0 = vld1_s8(w); w = (const int8_t*) w + 8;
78       const int8x8_t vb67c0x0 = vld1_s8(w); w = (const int8_t*) w + 8;
79       const int8x8_t vb01c1x0 = vld1_s8(w); w = (const int8_t*) w + 8;
80       const int8x8_t vb23c1x0 = vld1_s8(w); w = (const int8_t*) w + 8;
81       const int8x8_t vb45c1x0 = vld1_s8(w); w = (const int8_t*) w + 8;
82       const int8x8_t vb67c1x0 = vld1_s8(w); w = (const int8_t*) w + 8;
83 
84       int16x8_t vprod0x01c0 = vmull_s8(vb01c0x0, va0x0);
85       int16x8_t vprod1x01c0 = vmull_s8(vb01c0x0, va1x0);
86       int16x8_t vprod2x01c0 = vmull_s8(vb01c0x0, va2x0);
87       vacc0x01 = vpadalq_s16(vacc0x01, vprod0x01c0);
88       vacc1x01 = vpadalq_s16(vacc1x01, vprod1x01c0);
89       vacc2x01 = vpadalq_s16(vacc2x01, vprod2x01c0);
90       int16x8_t vprod0x23c0 = vmull_s8(vb23c0x0, va0x0);
91       int16x8_t vprod1x23c0 = vmull_s8(vb23c0x0, va1x0);
92       int16x8_t vprod2x23c0 = vmull_s8(vb23c0x0, va2x0);
93       vacc0x23 = vpadalq_s16(vacc0x23, vprod0x23c0);
94       vacc1x23 = vpadalq_s16(vacc1x23, vprod1x23c0);
95       vacc2x23 = vpadalq_s16(vacc2x23, vprod2x23c0);
96       int16x8_t vprod0x45c0 = vmull_s8(vb45c0x0, va0x0);
97       int16x8_t vprod1x45c0 = vmull_s8(vb45c0x0, va1x0);
98       int16x8_t vprod2x45c0 = vmull_s8(vb45c0x0, va2x0);
99       vacc0x45 = vpadalq_s16(vacc0x45, vprod0x45c0);
100       vacc1x45 = vpadalq_s16(vacc1x45, vprod1x45c0);
101       vacc2x45 = vpadalq_s16(vacc2x45, vprod2x45c0);
102       int16x8_t vprod0x67c0 = vmull_s8(vb67c0x0, va0x0);
103       int16x8_t vprod1x67c0 = vmull_s8(vb67c0x0, va1x0);
104       int16x8_t vprod2x67c0 = vmull_s8(vb67c0x0, va2x0);
105       vacc0x67 = vpadalq_s16(vacc0x67, vprod0x67c0);
106       vacc1x67 = vpadalq_s16(vacc1x67, vprod1x67c0);
107       vacc2x67 = vpadalq_s16(vacc2x67, vprod2x67c0);
108       va0x0 = vext_s8(va0x0, va0x0, 4);
109       va1x0 = vext_s8(va1x0, va1x0, 4);
110       va2x0 = vext_s8(va2x0, va2x0, 4);
111       int16x8_t vprod0x01c1 = vmull_s8(vb01c1x0, va0x0);
112       int16x8_t vprod1x01c1 = vmull_s8(vb01c1x0, va1x0);
113       int16x8_t vprod2x01c1 = vmull_s8(vb01c1x0, va2x0);
114       vacc0x01 = vpadalq_s16(vacc0x01, vprod0x01c1);
115       vacc1x01 = vpadalq_s16(vacc1x01, vprod1x01c1);
116       vacc2x01 = vpadalq_s16(vacc2x01, vprod2x01c1);
117       int16x8_t vprod0x23c1 = vmull_s8(vb23c1x0, va0x0);
118       int16x8_t vprod1x23c1 = vmull_s8(vb23c1x0, va1x0);
119       int16x8_t vprod2x23c1 = vmull_s8(vb23c1x0, va2x0);
120       vacc0x23 = vpadalq_s16(vacc0x23, vprod0x23c1);
121       vacc1x23 = vpadalq_s16(vacc1x23, vprod1x23c1);
122       vacc2x23 = vpadalq_s16(vacc2x23, vprod2x23c1);
123       int16x8_t vprod0x45c1 = vmull_s8(vb45c1x0, va0x0);
124       int16x8_t vprod1x45c1 = vmull_s8(vb45c1x0, va1x0);
125       int16x8_t vprod2x45c1 = vmull_s8(vb45c1x0, va2x0);
126       vacc0x45 = vpadalq_s16(vacc0x45, vprod0x45c1);
127       vacc1x45 = vpadalq_s16(vacc1x45, vprod1x45c1);
128       vacc2x45 = vpadalq_s16(vacc2x45, vprod2x45c1);
129       int16x8_t vprod0x67c1 = vmull_s8(vb67c1x0, va0x0);
130       int16x8_t vprod1x67c1 = vmull_s8(vb67c1x0, va1x0);
131       int16x8_t vprod2x67c1 = vmull_s8(vb67c1x0, va2x0);
132       vacc0x67 = vpadalq_s16(vacc0x67, vprod0x67c1);
133       vacc1x67 = vpadalq_s16(vacc1x67, vprod1x67c1);
134       vacc2x67 = vpadalq_s16(vacc2x67, vprod2x67c1);
135 
136       k -= 8 * sizeof(int8_t);
137     } while (k != 0);
138 
139 #if XNN_ARCH_ARM64
140     int32x4_t vacc0x0123 = vpaddq_s32(vacc0x01, vacc0x23);
141     int32x4_t vacc0x4567 = vpaddq_s32(vacc0x45, vacc0x67);
142     int32x4_t vacc1x0123 = vpaddq_s32(vacc1x01, vacc1x23);
143     int32x4_t vacc1x4567 = vpaddq_s32(vacc1x45, vacc1x67);
144     int32x4_t vacc2x0123 = vpaddq_s32(vacc2x01, vacc2x23);
145     int32x4_t vacc2x4567 = vpaddq_s32(vacc2x45, vacc2x67);
146 #else
147     const int32x2_t vsum0x01 = vpadd_s32(vget_low_s32(vacc0x01), vget_high_s32(vacc0x01));
148     const int32x2_t vsum0x23 = vpadd_s32(vget_low_s32(vacc0x23), vget_high_s32(vacc0x23));
149     int32x4_t vacc0x0123 = vcombine_s32(vsum0x01, vsum0x23);
150     const int32x2_t vsum0x45 = vpadd_s32(vget_low_s32(vacc0x45), vget_high_s32(vacc0x45));
151     const int32x2_t vsum0x67 = vpadd_s32(vget_low_s32(vacc0x67), vget_high_s32(vacc0x67));
152     int32x4_t vacc0x4567 = vcombine_s32(vsum0x45, vsum0x67);
153     const int32x2_t vsum1x01 = vpadd_s32(vget_low_s32(vacc1x01), vget_high_s32(vacc1x01));
154     const int32x2_t vsum1x23 = vpadd_s32(vget_low_s32(vacc1x23), vget_high_s32(vacc1x23));
155     int32x4_t vacc1x0123 = vcombine_s32(vsum1x01, vsum1x23);
156     const int32x2_t vsum1x45 = vpadd_s32(vget_low_s32(vacc1x45), vget_high_s32(vacc1x45));
157     const int32x2_t vsum1x67 = vpadd_s32(vget_low_s32(vacc1x67), vget_high_s32(vacc1x67));
158     int32x4_t vacc1x4567 = vcombine_s32(vsum1x45, vsum1x67);
159     const int32x2_t vsum2x01 = vpadd_s32(vget_low_s32(vacc2x01), vget_high_s32(vacc2x01));
160     const int32x2_t vsum2x23 = vpadd_s32(vget_low_s32(vacc2x23), vget_high_s32(vacc2x23));
161     int32x4_t vacc2x0123 = vcombine_s32(vsum2x01, vsum2x23);
162     const int32x2_t vsum2x45 = vpadd_s32(vget_low_s32(vacc2x45), vget_high_s32(vacc2x45));
163     const int32x2_t vsum2x67 = vpadd_s32(vget_low_s32(vacc2x67), vget_high_s32(vacc2x67));
164     int32x4_t vacc2x4567 = vcombine_s32(vsum2x45, vsum2x67);
165 #endif
166 
167     const int32x4_t vright_pre_shift = vld1q_dup_s32(&params->rndnu_neon.right_pre_shift);
168     const int32x4_t vmultiplier = vld1q_dup_s32(&params->rndnu_neon.multiplier);
169     const int32x4_t vright_post_shift = vld1q_dup_s32(&params->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 
178     vacc0x0123 = vqdmulhq_s32(vacc0x0123, vmultiplier);
179     vacc0x4567 = vqdmulhq_s32(vacc0x4567, vmultiplier);
180     vacc1x0123 = vqdmulhq_s32(vacc1x0123, vmultiplier);
181     vacc1x4567 = vqdmulhq_s32(vacc1x4567, vmultiplier);
182     vacc2x0123 = vqdmulhq_s32(vacc2x0123, vmultiplier);
183     vacc2x4567 = vqdmulhq_s32(vacc2x4567, vmultiplier);
184 
185     vacc0x0123 = vrshlq_s32(vacc0x0123, vright_post_shift);
186     vacc0x4567 = vrshlq_s32(vacc0x4567, vright_post_shift);
187     vacc1x0123 = vrshlq_s32(vacc1x0123, vright_post_shift);
188     vacc1x4567 = vrshlq_s32(vacc1x4567, vright_post_shift);
189     vacc2x0123 = vrshlq_s32(vacc2x0123, vright_post_shift);
190     vacc2x4567 = vrshlq_s32(vacc2x4567, vright_post_shift);
191 
192     const int16x8_t voutput_zero_point = vld1q_dup_s16(&params->rndnu_neon.output_zero_point);
193 #if XNN_ARCH_ARM64
194     int16x8_t vacc0x01234567 = vqmovn_high_s32(vqmovn_s32(vacc0x0123), vacc0x4567);
195     int16x8_t vacc1x01234567 = vqmovn_high_s32(vqmovn_s32(vacc1x0123), vacc1x4567);
196     int16x8_t vacc2x01234567 = vqmovn_high_s32(vqmovn_s32(vacc2x0123), vacc2x4567);
197 
198     vacc0x01234567 = vqaddq_s16(vacc0x01234567, voutput_zero_point);
199     vacc1x01234567 = vqaddq_s16(vacc1x01234567, voutput_zero_point);
200     vacc2x01234567 = vqaddq_s16(vacc2x01234567, voutput_zero_point);
201 
202     int8x16_t vout0x01234567_1x01234567 = vqmovn_high_s16(vqmovn_s16(vacc0x01234567), vacc1x01234567);
203     int8x8_t vout2x01234567 = vqmovn_s16(vacc2x01234567);
204 #else
205     int16x8_t vacc0x01234567 = vcombine_s16(vqmovn_s32(vacc0x0123), vqmovn_s32(vacc0x4567));
206     int16x8_t vacc1x01234567 = vcombine_s16(vqmovn_s32(vacc1x0123), vqmovn_s32(vacc1x4567));
207     int16x8_t vacc2x01234567 = vcombine_s16(vqmovn_s32(vacc2x0123), vqmovn_s32(vacc2x4567));
208 
209     vacc0x01234567 = vqaddq_s16(vacc0x01234567, voutput_zero_point);
210     vacc1x01234567 = vqaddq_s16(vacc1x01234567, voutput_zero_point);
211     vacc2x01234567 = vqaddq_s16(vacc2x01234567, voutput_zero_point);
212 
213     int8x16_t vout0x01234567_1x01234567 = vcombine_s8(vqmovn_s16(vacc0x01234567), vqmovn_s16(vacc1x01234567));
214     int8x8_t vout2x01234567 = vqmovn_s16(vacc2x01234567);
215 #endif
216 
217     const int8x16_t voutput_min = vld1q_dup_s8(&params->rndnu_neon.output_min);
218     vout0x01234567_1x01234567 = vmaxq_s8(vout0x01234567_1x01234567, voutput_min);
219     vout2x01234567 = vmax_s8(vout2x01234567, vget_low_s8(voutput_min));
220 
221     const int8x16_t voutput_max = vld1q_dup_s8(&params->rndnu_neon.output_max);
222     vout0x01234567_1x01234567 = vminq_s8(vout0x01234567_1x01234567, voutput_max);
223     vout2x01234567 = vmin_s8(vout2x01234567, vget_low_s8(voutput_max));
224 
225     if (nc >= 8) {
226       vst1_s8(c0 + 0, vget_low_s8(vout0x01234567_1x01234567));
227       vst1_s8(c1 + 0, vget_high_s8(vout0x01234567_1x01234567));
228       vst1_s8(c2 + 0, vout2x01234567);
229 
230       c0 = (int8_t*) ((uintptr_t) c0 + cn_stride);
231       c1 = (int8_t*) ((uintptr_t) c1 + cn_stride);
232       c2 = (int8_t*) ((uintptr_t) c2 + cn_stride);
233 
234       a0 = (const int8_t*) ((uintptr_t) a0 - kc);
235       a1 = (const int8_t*) ((uintptr_t) a1 - kc);
236       a2 = (const int8_t*) ((uintptr_t) a2 - kc);
237 
238       nc -= 8;
239     } else {
240       // Final case where not all of the 8 columns fit in the destination.
241       if (nc & 4) {
242         vst1q_lane_u32((void*) c0, vreinterpretq_u32_s8(vout0x01234567_1x01234567), 0); c0 += 4;
243         vst1q_lane_u32((void*) c1, vreinterpretq_u32_s8(vout0x01234567_1x01234567), 2); c1 += 4;
244         vst1_lane_u32((void*) c2, vreinterpret_u32_s8(vout2x01234567), 0); c2 += 4;
245         vout0x01234567_1x01234567 = vextq_s8(vout0x01234567_1x01234567, vout0x01234567_1x01234567, 4);
246         vout2x01234567 = vext_s8(vout2x01234567, vout2x01234567, 4);
247       }
248       if (nc & 2) {
249         vst1q_lane_u16((void*) c0, vreinterpretq_u16_s8(vout0x01234567_1x01234567), 0); c0 += 2;
250         vst1q_lane_u16((void*) c1, vreinterpretq_u16_s8(vout0x01234567_1x01234567), 4); c1 += 2;
251         vst1_lane_u16((void*) c2, vreinterpret_u16_s8(vout2x01234567), 0); c2 += 2;
252         vout0x01234567_1x01234567 = vextq_s8(vout0x01234567_1x01234567, vout0x01234567_1x01234567, 2);
253         vout2x01234567 = vext_s8(vout2x01234567, vout2x01234567, 2);
254       }
255       if (nc & 1) {
256         vst1q_lane_s8(c0, vout0x01234567_1x01234567, 0);
257         vst1q_lane_s8(c1, vout0x01234567_1x01234567, 8);
258         vst1_lane_s8(c2, vout2x01234567, 0);
259       }
260 
261       nc = 0;
262     }
263   } while (nc != 0);
264 }
265