xref: /aosp_15_r20/external/XNNPACK/src/qs8-gemm/gen/1x16c4s2-minmax-rndnu-neon-mlal.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_1x16c4s2__neon_mlal(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_1x16c4s2__neon_mlal(
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 <= 1);
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 
42   kc = round_up_po2(kc, 8 * sizeof(int8_t));
43   do {
44     int32x4_t vacc0x01 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const int32_t*) w + 2;
45     int32x4_t vacc0x23 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const int32_t*) w + 2;
46     int32x4_t vacc0x45 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const int32_t*) w + 2;
47     int32x4_t vacc0x67 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const int32_t*) w + 2;
48     int32x4_t vacc0x89 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const int32_t*) w + 2;
49     int32x4_t vacc0xAB = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const int32_t*) w + 2;
50     int32x4_t vacc0xCD = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const int32_t*) w + 2;
51     int32x4_t vacc0xEF = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const int32_t*) w + 2;
52 
53     size_t k = kc;
54     while (k >= 16 * sizeof(int8_t)) {
55       int8x8_t va0x0 = vld1_s8(a0); a0 += 8;
56       int8x8_t va0x1 = vld1_s8(a0); a0 += 8;
57 
58       const int8x8_t vb01c0x0 = vld1_s8(w); w = (const int8_t*) w + 8;
59       const int8x8_t vb23c0x0 = vld1_s8(w); w = (const int8_t*) w + 8;
60       const int8x8_t vb45c0x0 = vld1_s8(w); w = (const int8_t*) w + 8;
61       const int8x8_t vb67c0x0 = vld1_s8(w); w = (const int8_t*) w + 8;
62       const int8x8_t vb89c0x0 = vld1_s8(w); w = (const int8_t*) w + 8;
63       const int8x8_t vbABc0x0 = vld1_s8(w); w = (const int8_t*) w + 8;
64       const int8x8_t vbCDc0x0 = vld1_s8(w); w = (const int8_t*) w + 8;
65       const int8x8_t vbEFc0x0 = vld1_s8(w); w = (const int8_t*) w + 8;
66       const int8x8_t vb01c1x0 = vld1_s8(w); w = (const int8_t*) w + 8;
67       const int8x8_t vb23c1x0 = vld1_s8(w); w = (const int8_t*) w + 8;
68       const int8x8_t vb45c1x0 = vld1_s8(w); w = (const int8_t*) w + 8;
69       const int8x8_t vb67c1x0 = vld1_s8(w); w = (const int8_t*) w + 8;
70       const int8x8_t vb89c1x0 = vld1_s8(w); w = (const int8_t*) w + 8;
71       const int8x8_t vbABc1x0 = vld1_s8(w); w = (const int8_t*) w + 8;
72       const int8x8_t vbCDc1x0 = vld1_s8(w); w = (const int8_t*) w + 8;
73       const int8x8_t vbEFc1x0 = vld1_s8(w); w = (const int8_t*) w + 8;
74 
75       int16x8_t vprod0x01c0 = vmull_s8(vb01c0x0, va0x0);
76       const int8x8_t vb01c0x1 = vld1_s8(w); w = (const int8_t*) w + 8;
77       vprod0x01c0 = vmlal_s8(vprod0x01c0, vb01c0x1, va0x1);
78       vacc0x01 = vpadalq_s16(vacc0x01, vprod0x01c0);
79       int16x8_t vprod0x23c0 = vmull_s8(vb23c0x0, va0x0);
80       const int8x8_t vb23c0x1 = vld1_s8(w); w = (const int8_t*) w + 8;
81       vprod0x23c0 = vmlal_s8(vprod0x23c0, vb23c0x1, va0x1);
82       vacc0x23 = vpadalq_s16(vacc0x23, vprod0x23c0);
83       int16x8_t vprod0x45c0 = vmull_s8(vb45c0x0, va0x0);
84       const int8x8_t vb45c0x1 = vld1_s8(w); w = (const int8_t*) w + 8;
85       vprod0x45c0 = vmlal_s8(vprod0x45c0, vb45c0x1, va0x1);
86       vacc0x45 = vpadalq_s16(vacc0x45, vprod0x45c0);
87       int16x8_t vprod0x67c0 = vmull_s8(vb67c0x0, va0x0);
88       const int8x8_t vb67c0x1 = vld1_s8(w); w = (const int8_t*) w + 8;
89       vprod0x67c0 = vmlal_s8(vprod0x67c0, vb67c0x1, va0x1);
90       vacc0x67 = vpadalq_s16(vacc0x67, vprod0x67c0);
91       int16x8_t vprod0x89c0 = vmull_s8(vb89c0x0, va0x0);
92       const int8x8_t vb89c0x1 = vld1_s8(w); w = (const int8_t*) w + 8;
93       vprod0x89c0 = vmlal_s8(vprod0x89c0, vb89c0x1, va0x1);
94       vacc0x89 = vpadalq_s16(vacc0x89, vprod0x89c0);
95       int16x8_t vprod0xABc0 = vmull_s8(vbABc0x0, va0x0);
96       const int8x8_t vbABc0x1 = vld1_s8(w); w = (const int8_t*) w + 8;
97       vprod0xABc0 = vmlal_s8(vprod0xABc0, vbABc0x1, va0x1);
98       vacc0xAB = vpadalq_s16(vacc0xAB, vprod0xABc0);
99       int16x8_t vprod0xCDc0 = vmull_s8(vbCDc0x0, va0x0);
100       const int8x8_t vbCDc0x1 = vld1_s8(w); w = (const int8_t*) w + 8;
101       vprod0xCDc0 = vmlal_s8(vprod0xCDc0, vbCDc0x1, va0x1);
102       vacc0xCD = vpadalq_s16(vacc0xCD, vprod0xCDc0);
103       int16x8_t vprod0xEFc0 = vmull_s8(vbEFc0x0, va0x0);
104       const int8x8_t vbEFc0x1 = vld1_s8(w); w = (const int8_t*) w + 8;
105       vprod0xEFc0 = vmlal_s8(vprod0xEFc0, vbEFc0x1, va0x1);
106       vacc0xEF = vpadalq_s16(vacc0xEF, vprod0xEFc0);
107       va0x0 = vext_s8(va0x0, va0x0, 4);
108       va0x1 = vext_s8(va0x1, va0x1, 4);
109       int16x8_t vprod0x01c1 = vmull_s8(vb01c1x0, va0x0);
110       const int8x8_t vb01c1x1 = vld1_s8(w); w = (const int8_t*) w + 8;
111       vprod0x01c1 = vmlal_s8(vprod0x01c1, vb01c1x1, va0x1);
112       vacc0x01 = vpadalq_s16(vacc0x01, vprod0x01c1);
113       int16x8_t vprod0x23c1 = vmull_s8(vb23c1x0, va0x0);
114       const int8x8_t vb23c1x1 = vld1_s8(w); w = (const int8_t*) w + 8;
115       vprod0x23c1 = vmlal_s8(vprod0x23c1, vb23c1x1, va0x1);
116       vacc0x23 = vpadalq_s16(vacc0x23, vprod0x23c1);
117       int16x8_t vprod0x45c1 = vmull_s8(vb45c1x0, va0x0);
118       const int8x8_t vb45c1x1 = vld1_s8(w); w = (const int8_t*) w + 8;
119       vprod0x45c1 = vmlal_s8(vprod0x45c1, vb45c1x1, va0x1);
120       vacc0x45 = vpadalq_s16(vacc0x45, vprod0x45c1);
121       int16x8_t vprod0x67c1 = vmull_s8(vb67c1x0, va0x0);
122       const int8x8_t vb67c1x1 = vld1_s8(w); w = (const int8_t*) w + 8;
123       vprod0x67c1 = vmlal_s8(vprod0x67c1, vb67c1x1, va0x1);
124       vacc0x67 = vpadalq_s16(vacc0x67, vprod0x67c1);
125       int16x8_t vprod0x89c1 = vmull_s8(vb89c1x0, va0x0);
126       const int8x8_t vb89c1x1 = vld1_s8(w); w = (const int8_t*) w + 8;
127       vprod0x89c1 = vmlal_s8(vprod0x89c1, vb89c1x1, va0x1);
128       vacc0x89 = vpadalq_s16(vacc0x89, vprod0x89c1);
129       int16x8_t vprod0xABc1 = vmull_s8(vbABc1x0, va0x0);
130       const int8x8_t vbABc1x1 = vld1_s8(w); w = (const int8_t*) w + 8;
131       vprod0xABc1 = vmlal_s8(vprod0xABc1, vbABc1x1, va0x1);
132       vacc0xAB = vpadalq_s16(vacc0xAB, vprod0xABc1);
133       int16x8_t vprod0xCDc1 = vmull_s8(vbCDc1x0, va0x0);
134       const int8x8_t vbCDc1x1 = vld1_s8(w); w = (const int8_t*) w + 8;
135       vprod0xCDc1 = vmlal_s8(vprod0xCDc1, vbCDc1x1, va0x1);
136       vacc0xCD = vpadalq_s16(vacc0xCD, vprod0xCDc1);
137       int16x8_t vprod0xEFc1 = vmull_s8(vbEFc1x0, va0x0);
138       const int8x8_t vbEFc1x1 = vld1_s8(w); w = (const int8_t*) w + 8;
139       vprod0xEFc1 = vmlal_s8(vprod0xEFc1, vbEFc1x1, va0x1);
140       vacc0xEF = vpadalq_s16(vacc0xEF, vprod0xEFc1);
141 
142       k -= 16 * sizeof(int8_t);
143     }
144     if (k != 0) {
145       int8x8_t va0x0 = vld1_s8(a0); a0 += 8;
146 
147       const int8x8_t vb01c0x0 = vld1_s8(w); w = (const int8_t*) w + 8;
148       const int8x8_t vb23c0x0 = vld1_s8(w); w = (const int8_t*) w + 8;
149       const int8x8_t vb45c0x0 = vld1_s8(w); w = (const int8_t*) w + 8;
150       const int8x8_t vb67c0x0 = vld1_s8(w); w = (const int8_t*) w + 8;
151       const int8x8_t vb89c0x0 = vld1_s8(w); w = (const int8_t*) w + 8;
152       const int8x8_t vbABc0x0 = vld1_s8(w); w = (const int8_t*) w + 8;
153       const int8x8_t vbCDc0x0 = vld1_s8(w); w = (const int8_t*) w + 8;
154       const int8x8_t vbEFc0x0 = vld1_s8(w); w = (const int8_t*) w + 8;
155       const int8x8_t vb01c1x0 = vld1_s8(w); w = (const int8_t*) w + 8;
156       const int8x8_t vb23c1x0 = vld1_s8(w); w = (const int8_t*) w + 8;
157       const int8x8_t vb45c1x0 = vld1_s8(w); w = (const int8_t*) w + 8;
158       const int8x8_t vb67c1x0 = vld1_s8(w); w = (const int8_t*) w + 8;
159       const int8x8_t vb89c1x0 = vld1_s8(w); w = (const int8_t*) w + 8;
160       const int8x8_t vbABc1x0 = vld1_s8(w); w = (const int8_t*) w + 8;
161       const int8x8_t vbCDc1x0 = vld1_s8(w); w = (const int8_t*) w + 8;
162       const int8x8_t vbEFc1x0 = vld1_s8(w); w = (const int8_t*) w + 8;
163 
164       int16x8_t vprod0x01c0 = vmull_s8(vb01c0x0, va0x0);
165       vacc0x01 = vpadalq_s16(vacc0x01, vprod0x01c0);
166       int16x8_t vprod0x23c0 = vmull_s8(vb23c0x0, va0x0);
167       vacc0x23 = vpadalq_s16(vacc0x23, vprod0x23c0);
168       int16x8_t vprod0x45c0 = vmull_s8(vb45c0x0, va0x0);
169       vacc0x45 = vpadalq_s16(vacc0x45, vprod0x45c0);
170       int16x8_t vprod0x67c0 = vmull_s8(vb67c0x0, va0x0);
171       vacc0x67 = vpadalq_s16(vacc0x67, vprod0x67c0);
172       int16x8_t vprod0x89c0 = vmull_s8(vb89c0x0, va0x0);
173       vacc0x89 = vpadalq_s16(vacc0x89, vprod0x89c0);
174       int16x8_t vprod0xABc0 = vmull_s8(vbABc0x0, va0x0);
175       vacc0xAB = vpadalq_s16(vacc0xAB, vprod0xABc0);
176       int16x8_t vprod0xCDc0 = vmull_s8(vbCDc0x0, va0x0);
177       vacc0xCD = vpadalq_s16(vacc0xCD, vprod0xCDc0);
178       int16x8_t vprod0xEFc0 = vmull_s8(vbEFc0x0, va0x0);
179       vacc0xEF = vpadalq_s16(vacc0xEF, vprod0xEFc0);
180       va0x0 = vext_s8(va0x0, va0x0, 4);
181       int16x8_t vprod0x01c1 = vmull_s8(vb01c1x0, va0x0);
182       vacc0x01 = vpadalq_s16(vacc0x01, vprod0x01c1);
183       int16x8_t vprod0x23c1 = vmull_s8(vb23c1x0, va0x0);
184       vacc0x23 = vpadalq_s16(vacc0x23, vprod0x23c1);
185       int16x8_t vprod0x45c1 = vmull_s8(vb45c1x0, va0x0);
186       vacc0x45 = vpadalq_s16(vacc0x45, vprod0x45c1);
187       int16x8_t vprod0x67c1 = vmull_s8(vb67c1x0, va0x0);
188       vacc0x67 = vpadalq_s16(vacc0x67, vprod0x67c1);
189       int16x8_t vprod0x89c1 = vmull_s8(vb89c1x0, va0x0);
190       vacc0x89 = vpadalq_s16(vacc0x89, vprod0x89c1);
191       int16x8_t vprod0xABc1 = vmull_s8(vbABc1x0, va0x0);
192       vacc0xAB = vpadalq_s16(vacc0xAB, vprod0xABc1);
193       int16x8_t vprod0xCDc1 = vmull_s8(vbCDc1x0, va0x0);
194       vacc0xCD = vpadalq_s16(vacc0xCD, vprod0xCDc1);
195       int16x8_t vprod0xEFc1 = vmull_s8(vbEFc1x0, va0x0);
196       vacc0xEF = vpadalq_s16(vacc0xEF, vprod0xEFc1);
197 
198     }
199 
200 #if XNN_ARCH_ARM64
201     int32x4_t vacc0x0123 = vpaddq_s32(vacc0x01, vacc0x23);
202     int32x4_t vacc0x4567 = vpaddq_s32(vacc0x45, vacc0x67);
203     int32x4_t vacc0x89AB = vpaddq_s32(vacc0x89, vacc0xAB);
204     int32x4_t vacc0xCDEF = vpaddq_s32(vacc0xCD, vacc0xEF);
205 #else
206     const int32x2_t vsum0x01 = vpadd_s32(vget_low_s32(vacc0x01), vget_high_s32(vacc0x01));
207     const int32x2_t vsum0x23 = vpadd_s32(vget_low_s32(vacc0x23), vget_high_s32(vacc0x23));
208     int32x4_t vacc0x0123 = vcombine_s32(vsum0x01, vsum0x23);
209     const int32x2_t vsum0x45 = vpadd_s32(vget_low_s32(vacc0x45), vget_high_s32(vacc0x45));
210     const int32x2_t vsum0x67 = vpadd_s32(vget_low_s32(vacc0x67), vget_high_s32(vacc0x67));
211     int32x4_t vacc0x4567 = vcombine_s32(vsum0x45, vsum0x67);
212     const int32x2_t vsum0x89 = vpadd_s32(vget_low_s32(vacc0x89), vget_high_s32(vacc0x89));
213     const int32x2_t vsum0xAB = vpadd_s32(vget_low_s32(vacc0xAB), vget_high_s32(vacc0xAB));
214     int32x4_t vacc0x89AB = vcombine_s32(vsum0x89, vsum0xAB);
215     const int32x2_t vsum0xCD = vpadd_s32(vget_low_s32(vacc0xCD), vget_high_s32(vacc0xCD));
216     const int32x2_t vsum0xEF = vpadd_s32(vget_low_s32(vacc0xEF), vget_high_s32(vacc0xEF));
217     int32x4_t vacc0xCDEF = vcombine_s32(vsum0xCD, vsum0xEF);
218 #endif
219 
220     const int32x4_t vright_pre_shift = vld1q_dup_s32(&params->rndnu_neon.right_pre_shift);
221     const int32x4_t vmultiplier = vld1q_dup_s32(&params->rndnu_neon.multiplier);
222     const int32x4_t vright_post_shift = vld1q_dup_s32(&params->rndnu_neon.right_post_shift);
223 
224     vacc0x0123 = vqshlq_s32(vacc0x0123, vright_pre_shift);
225     vacc0x4567 = vqshlq_s32(vacc0x4567, vright_pre_shift);
226     vacc0x89AB = vqshlq_s32(vacc0x89AB, vright_pre_shift);
227     vacc0xCDEF = vqshlq_s32(vacc0xCDEF, vright_pre_shift);
228 
229     vacc0x0123 = vqdmulhq_s32(vacc0x0123, vmultiplier);
230     vacc0x4567 = vqdmulhq_s32(vacc0x4567, vmultiplier);
231     vacc0x89AB = vqdmulhq_s32(vacc0x89AB, vmultiplier);
232     vacc0xCDEF = vqdmulhq_s32(vacc0xCDEF, vmultiplier);
233 
234     vacc0x0123 = vrshlq_s32(vacc0x0123, vright_post_shift);
235     vacc0x4567 = vrshlq_s32(vacc0x4567, vright_post_shift);
236     vacc0x89AB = vrshlq_s32(vacc0x89AB, vright_post_shift);
237     vacc0xCDEF = vrshlq_s32(vacc0xCDEF, 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 vacc0x89ABCDEF = vqmovn_high_s32(vqmovn_s32(vacc0x89AB), vacc0xCDEF);
243 
244     vacc0x01234567 = vqaddq_s16(vacc0x01234567, voutput_zero_point);
245     vacc0x89ABCDEF = vqaddq_s16(vacc0x89ABCDEF, voutput_zero_point);
246 
247     int8x16_t vout0x0123456789ABCDEF = vqmovn_high_s16(vqmovn_s16(vacc0x01234567), vacc0x89ABCDEF);
248 #else
249     int16x8_t vacc0x01234567 = vcombine_s16(vqmovn_s32(vacc0x0123), vqmovn_s32(vacc0x4567));
250     int16x8_t vacc0x89ABCDEF = vcombine_s16(vqmovn_s32(vacc0x89AB), vqmovn_s32(vacc0xCDEF));
251 
252     vacc0x01234567 = vqaddq_s16(vacc0x01234567, voutput_zero_point);
253     vacc0x89ABCDEF = vqaddq_s16(vacc0x89ABCDEF, voutput_zero_point);
254 
255     int8x16_t vout0x0123456789ABCDEF = vcombine_s8(vqmovn_s16(vacc0x01234567), vqmovn_s16(vacc0x89ABCDEF));
256 #endif
257 
258     const int8x16_t voutput_min = vld1q_dup_s8(&params->rndnu_neon.output_min);
259     vout0x0123456789ABCDEF = vmaxq_s8(vout0x0123456789ABCDEF, voutput_min);
260 
261     const int8x16_t voutput_max = vld1q_dup_s8(&params->rndnu_neon.output_max);
262     vout0x0123456789ABCDEF = vminq_s8(vout0x0123456789ABCDEF, voutput_max);
263 
264     if (nc >= 16) {
265       vst1q_s8(c0 + 0, vout0x0123456789ABCDEF);
266 
267       c0 = (int8_t*) ((uintptr_t) c0 + cn_stride);
268 
269       a0 = (const int8_t*) ((uintptr_t) a0 - kc);
270 
271       nc -= 16;
272     } else {
273       // Final case where not all of the 16 columns fit in the destination.
274       int8x8_t vout0x01234567 = vget_low_s8(vout0x0123456789ABCDEF);
275       if (nc & 8) {
276         vst1_s8(c0, vout0x01234567); c0 += 8;
277         vout0x01234567 = vget_high_s8(vout0x0123456789ABCDEF);
278       }
279       if (nc & 4) {
280         vst1_lane_u32((void*) c0, vreinterpret_u32_s8(vout0x01234567), 0); c0 += 4;
281         vout0x01234567 = vext_s8(vout0x01234567, vout0x01234567, 4);
282       }
283       if (nc & 2) {
284         vst1_lane_u16((void*) c0, vreinterpret_u16_s8(vout0x01234567), 0); c0 += 2;
285         vout0x01234567 = vext_s8(vout0x01234567, vout0x01234567, 2);
286       }
287       if (nc & 1) {
288         vst1_lane_s8(c0, vout0x01234567, 0);
289       }
290 
291       nc = 0;
292     }
293   } while (nc != 0);
294 }
295