xref: /aosp_15_r20/external/XNNPACK/src/qs8-gemm/gen/1x16-minmax-rndnu-neon-mull-addw-dup.c (revision 4bdc94577ba0e567308109d787f7fec7b531ce36) 
1 // Auto-generated file. Do not edit!
2 //   Template: src/qs8-gemm/neon-mull-addw-dup.c.in
3 //   Generator: tools/xngen
4 //
5 // Copyright 2020 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/common.h>
15 #include <xnnpack/gemm.h>
16 
17 
xnn_qs8_gemm_minmax_rndnu_ukernel_1x16__neon_mull_addw_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)])18 void xnn_qs8_gemm_minmax_rndnu_ukernel_1x16__neon_mull_addw_dup(
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   do {
43     int32x4_t vacc0x0123 = vld1q_s32(w); w = (const void*) ((uintptr_t) w + 4 * sizeof(int32_t));
44     int32x4_t vacc0x4567 = vld1q_s32(w); w = (const void*) ((uintptr_t) w + 4 * sizeof(int32_t));
45     int32x4_t vacc0x89AB = vld1q_s32(w); w = (const void*) ((uintptr_t) w + 4 * sizeof(int32_t));
46     int32x4_t vacc0xCDEF = vld1q_s32(w); w = (const void*) ((uintptr_t) w + 4 * sizeof(int32_t));
47 
48     size_t k = kc;
49     while (k >= 8 * sizeof(int8_t)) {
50       const int8x8_t va0 = vld1_s8(a0); a0 += 8;
51 
52       const int8x8_t vb01234567c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
53 
54       const int16x8_t vprod0x01234567c0 = vmull_s8(vb01234567c0, vdup_lane_s8(va0, 0));
55       vacc0x0123 = vaddw_s16(vacc0x0123, vget_low_s16(vprod0x01234567c0));
56       vacc0x4567 = vaddw_s16(vacc0x4567, vget_high_s16(vprod0x01234567c0));
57       const int8x8_t vb89ABCDEFc0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
58 
59       const int16x8_t vprod0x89ABCDEFc0 = vmull_s8(vb89ABCDEFc0, vdup_lane_s8(va0, 0));
60       vacc0x89AB = vaddw_s16(vacc0x89AB, vget_low_s16(vprod0x89ABCDEFc0));
61       vacc0xCDEF = vaddw_s16(vacc0xCDEF, vget_high_s16(vprod0x89ABCDEFc0));
62       const int8x8_t vb01234567c1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
63 
64       const int16x8_t vprod0x01234567c1 = vmull_s8(vb01234567c1, vdup_lane_s8(va0, 1));
65       vacc0x0123 = vaddw_s16(vacc0x0123, vget_low_s16(vprod0x01234567c1));
66       vacc0x4567 = vaddw_s16(vacc0x4567, vget_high_s16(vprod0x01234567c1));
67       const int8x8_t vb89ABCDEFc1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
68 
69       const int16x8_t vprod0x89ABCDEFc1 = vmull_s8(vb89ABCDEFc1, vdup_lane_s8(va0, 1));
70       vacc0x89AB = vaddw_s16(vacc0x89AB, vget_low_s16(vprod0x89ABCDEFc1));
71       vacc0xCDEF = vaddw_s16(vacc0xCDEF, vget_high_s16(vprod0x89ABCDEFc1));
72       const int8x8_t vb01234567c2 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
73 
74       const int16x8_t vprod0x01234567c2 = vmull_s8(vb01234567c2, vdup_lane_s8(va0, 2));
75       vacc0x0123 = vaddw_s16(vacc0x0123, vget_low_s16(vprod0x01234567c2));
76       vacc0x4567 = vaddw_s16(vacc0x4567, vget_high_s16(vprod0x01234567c2));
77       const int8x8_t vb89ABCDEFc2 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
78 
79       const int16x8_t vprod0x89ABCDEFc2 = vmull_s8(vb89ABCDEFc2, vdup_lane_s8(va0, 2));
80       vacc0x89AB = vaddw_s16(vacc0x89AB, vget_low_s16(vprod0x89ABCDEFc2));
81       vacc0xCDEF = vaddw_s16(vacc0xCDEF, vget_high_s16(vprod0x89ABCDEFc2));
82       const int8x8_t vb01234567c3 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
83 
84       const int16x8_t vprod0x01234567c3 = vmull_s8(vb01234567c3, vdup_lane_s8(va0, 3));
85       vacc0x0123 = vaddw_s16(vacc0x0123, vget_low_s16(vprod0x01234567c3));
86       vacc0x4567 = vaddw_s16(vacc0x4567, vget_high_s16(vprod0x01234567c3));
87       const int8x8_t vb89ABCDEFc3 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
88 
89       const int16x8_t vprod0x89ABCDEFc3 = vmull_s8(vb89ABCDEFc3, vdup_lane_s8(va0, 3));
90       vacc0x89AB = vaddw_s16(vacc0x89AB, vget_low_s16(vprod0x89ABCDEFc3));
91       vacc0xCDEF = vaddw_s16(vacc0xCDEF, vget_high_s16(vprod0x89ABCDEFc3));
92       const int8x8_t vb01234567c4 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
93 
94       const int16x8_t vprod0x01234567c4 = vmull_s8(vb01234567c4, vdup_lane_s8(va0, 4));
95       vacc0x0123 = vaddw_s16(vacc0x0123, vget_low_s16(vprod0x01234567c4));
96       vacc0x4567 = vaddw_s16(vacc0x4567, vget_high_s16(vprod0x01234567c4));
97       const int8x8_t vb89ABCDEFc4 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
98 
99       const int16x8_t vprod0x89ABCDEFc4 = vmull_s8(vb89ABCDEFc4, vdup_lane_s8(va0, 4));
100       vacc0x89AB = vaddw_s16(vacc0x89AB, vget_low_s16(vprod0x89ABCDEFc4));
101       vacc0xCDEF = vaddw_s16(vacc0xCDEF, vget_high_s16(vprod0x89ABCDEFc4));
102       const int8x8_t vb01234567c5 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
103 
104       const int16x8_t vprod0x01234567c5 = vmull_s8(vb01234567c5, vdup_lane_s8(va0, 5));
105       vacc0x0123 = vaddw_s16(vacc0x0123, vget_low_s16(vprod0x01234567c5));
106       vacc0x4567 = vaddw_s16(vacc0x4567, vget_high_s16(vprod0x01234567c5));
107       const int8x8_t vb89ABCDEFc5 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
108 
109       const int16x8_t vprod0x89ABCDEFc5 = vmull_s8(vb89ABCDEFc5, vdup_lane_s8(va0, 5));
110       vacc0x89AB = vaddw_s16(vacc0x89AB, vget_low_s16(vprod0x89ABCDEFc5));
111       vacc0xCDEF = vaddw_s16(vacc0xCDEF, vget_high_s16(vprod0x89ABCDEFc5));
112       const int8x8_t vb01234567c6 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
113 
114       const int16x8_t vprod0x01234567c6 = vmull_s8(vb01234567c6, vdup_lane_s8(va0, 6));
115       vacc0x0123 = vaddw_s16(vacc0x0123, vget_low_s16(vprod0x01234567c6));
116       vacc0x4567 = vaddw_s16(vacc0x4567, vget_high_s16(vprod0x01234567c6));
117       const int8x8_t vb89ABCDEFc6 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
118 
119       const int16x8_t vprod0x89ABCDEFc6 = vmull_s8(vb89ABCDEFc6, vdup_lane_s8(va0, 6));
120       vacc0x89AB = vaddw_s16(vacc0x89AB, vget_low_s16(vprod0x89ABCDEFc6));
121       vacc0xCDEF = vaddw_s16(vacc0xCDEF, vget_high_s16(vprod0x89ABCDEFc6));
122       const int8x8_t vb01234567c7 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
123 
124       const int16x8_t vprod0x01234567c7 = vmull_s8(vb01234567c7, vdup_lane_s8(va0, 7));
125       vacc0x0123 = vaddw_s16(vacc0x0123, vget_low_s16(vprod0x01234567c7));
126       vacc0x4567 = vaddw_s16(vacc0x4567, vget_high_s16(vprod0x01234567c7));
127       const int8x8_t vb89ABCDEFc7 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
128 
129       const int16x8_t vprod0x89ABCDEFc7 = vmull_s8(vb89ABCDEFc7, vdup_lane_s8(va0, 7));
130       vacc0x89AB = vaddw_s16(vacc0x89AB, vget_low_s16(vprod0x89ABCDEFc7));
131       vacc0xCDEF = vaddw_s16(vacc0xCDEF, vget_high_s16(vprod0x89ABCDEFc7));
132 
133       k -= 8 * sizeof(int8_t);
134     }
135     if XNN_UNLIKELY(k != 0) {
136       const int8x8_t va0 = vld1_s8(a0); a0 = (const int8_t*) ((uintptr_t) a0 + k);
137 
138       const int8x8_t vb01234567c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
139       const int8x8_t vb89ABCDEFc0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
140 
141       const int16x8_t vprod0x01234567c0 = vmull_s8(vb01234567c0, vdup_lane_s8(va0, 0));
142       vacc0x0123 = vaddw_s16(vacc0x0123, vget_low_s16(vprod0x01234567c0));
143       vacc0x4567 = vaddw_s16(vacc0x4567, vget_high_s16(vprod0x01234567c0));
144       const int16x8_t vprod0x89ABCDEFc0 = vmull_s8(vb89ABCDEFc0, vdup_lane_s8(va0, 0));
145       vacc0x89AB = vaddw_s16(vacc0x89AB, vget_low_s16(vprod0x89ABCDEFc0));
146       vacc0xCDEF = vaddw_s16(vacc0xCDEF, vget_high_s16(vprod0x89ABCDEFc0));
147 
148       if (k >= 2 * sizeof(int8_t)) {
149         const int8x8_t vb01234567c1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
150         const int8x8_t vb89ABCDEFc1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
151 
152         const int16x8_t vprod0x01234567c1 = vmull_s8(vb01234567c1, vdup_lane_s8(va0, 1));
153         vacc0x0123 = vaddw_s16(vacc0x0123, vget_low_s16(vprod0x01234567c1));
154         vacc0x4567 = vaddw_s16(vacc0x4567, vget_high_s16(vprod0x01234567c1));
155         const int16x8_t vprod0x89ABCDEFc1 = vmull_s8(vb89ABCDEFc1, vdup_lane_s8(va0, 1));
156         vacc0x89AB = vaddw_s16(vacc0x89AB, vget_low_s16(vprod0x89ABCDEFc1));
157         vacc0xCDEF = vaddw_s16(vacc0xCDEF, vget_high_s16(vprod0x89ABCDEFc1));
158 
159         if (k > 2 * sizeof(int8_t)) {
160           const int8x8_t vb01234567c2 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
161           const int8x8_t vb89ABCDEFc2 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
162 
163           const int16x8_t vprod0x01234567c2 = vmull_s8(vb01234567c2, vdup_lane_s8(va0, 2));
164           vacc0x0123 = vaddw_s16(vacc0x0123, vget_low_s16(vprod0x01234567c2));
165           vacc0x4567 = vaddw_s16(vacc0x4567, vget_high_s16(vprod0x01234567c2));
166           const int16x8_t vprod0x89ABCDEFc2 = vmull_s8(vb89ABCDEFc2, vdup_lane_s8(va0, 2));
167           vacc0x89AB = vaddw_s16(vacc0x89AB, vget_low_s16(vprod0x89ABCDEFc2));
168           vacc0xCDEF = vaddw_s16(vacc0xCDEF, vget_high_s16(vprod0x89ABCDEFc2));
169 
170           if (k >= 4 * sizeof(int8_t)) {
171             const int8x8_t vb01234567c3 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
172             const int8x8_t vb89ABCDEFc3 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
173 
174             const int16x8_t vprod0x01234567c3 = vmull_s8(vb01234567c3, vdup_lane_s8(va0, 3));
175             vacc0x0123 = vaddw_s16(vacc0x0123, vget_low_s16(vprod0x01234567c3));
176             vacc0x4567 = vaddw_s16(vacc0x4567, vget_high_s16(vprod0x01234567c3));
177             const int16x8_t vprod0x89ABCDEFc3 = vmull_s8(vb89ABCDEFc3, vdup_lane_s8(va0, 3));
178             vacc0x89AB = vaddw_s16(vacc0x89AB, vget_low_s16(vprod0x89ABCDEFc3));
179             vacc0xCDEF = vaddw_s16(vacc0xCDEF, vget_high_s16(vprod0x89ABCDEFc3));
180 
181             if (k > 4 * sizeof(int8_t)) {
182               const int8x8_t vb01234567c4 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
183               const int8x8_t vb89ABCDEFc4 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
184 
185               const int16x8_t vprod0x01234567c4 = vmull_s8(vb01234567c4, vdup_lane_s8(va0, 4));
186               vacc0x0123 = vaddw_s16(vacc0x0123, vget_low_s16(vprod0x01234567c4));
187               vacc0x4567 = vaddw_s16(vacc0x4567, vget_high_s16(vprod0x01234567c4));
188               const int16x8_t vprod0x89ABCDEFc4 = vmull_s8(vb89ABCDEFc4, vdup_lane_s8(va0, 4));
189               vacc0x89AB = vaddw_s16(vacc0x89AB, vget_low_s16(vprod0x89ABCDEFc4));
190               vacc0xCDEF = vaddw_s16(vacc0xCDEF, vget_high_s16(vprod0x89ABCDEFc4));
191 
192               if (k >= 6 * sizeof(int8_t)) {
193                 const int8x8_t vb01234567c5 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
194                 const int8x8_t vb89ABCDEFc5 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
195 
196                 const int16x8_t vprod0x01234567c5 = vmull_s8(vb01234567c5, vdup_lane_s8(va0, 5));
197                 vacc0x0123 = vaddw_s16(vacc0x0123, vget_low_s16(vprod0x01234567c5));
198                 vacc0x4567 = vaddw_s16(vacc0x4567, vget_high_s16(vprod0x01234567c5));
199                 const int16x8_t vprod0x89ABCDEFc5 = vmull_s8(vb89ABCDEFc5, vdup_lane_s8(va0, 5));
200                 vacc0x89AB = vaddw_s16(vacc0x89AB, vget_low_s16(vprod0x89ABCDEFc5));
201                 vacc0xCDEF = vaddw_s16(vacc0xCDEF, vget_high_s16(vprod0x89ABCDEFc5));
202 
203                 if (k > 6 * sizeof(int8_t)) {
204                   const int8x8_t vb01234567c6 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
205                   const int8x8_t vb89ABCDEFc6 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
206 
207                   const int16x8_t vprod0x01234567c6 = vmull_s8(vb01234567c6, vdup_lane_s8(va0, 6));
208                   vacc0x0123 = vaddw_s16(vacc0x0123, vget_low_s16(vprod0x01234567c6));
209                   vacc0x4567 = vaddw_s16(vacc0x4567, vget_high_s16(vprod0x01234567c6));
210                   const int16x8_t vprod0x89ABCDEFc6 = vmull_s8(vb89ABCDEFc6, vdup_lane_s8(va0, 6));
211                   vacc0x89AB = vaddw_s16(vacc0x89AB, vget_low_s16(vprod0x89ABCDEFc6));
212                   vacc0xCDEF = vaddw_s16(vacc0xCDEF, vget_high_s16(vprod0x89ABCDEFc6));
213                 }
214               }
215             }
216           }
217         }
218       }
219     }
220 
221     // Post-accumulation work
222     const int32x4_t vright_pre_shift = vld1q_dup_s32(&params->rndnu_neon.right_pre_shift);
223     const int32x4_t vmultiplier = vld1q_dup_s32(&params->rndnu_neon.multiplier);
224     const int32x4_t vright_post_shift = vld1q_dup_s32(&params->rndnu_neon.right_post_shift);
225 
226     vacc0x0123 = vqshlq_s32(vacc0x0123, vright_pre_shift);
227     vacc0x4567 = vqshlq_s32(vacc0x4567, vright_pre_shift);
228     vacc0x89AB = vqshlq_s32(vacc0x89AB, vright_pre_shift);
229     vacc0xCDEF = vqshlq_s32(vacc0xCDEF, vright_pre_shift);
230 
231     vacc0x0123 = vqdmulhq_s32(vacc0x0123, vmultiplier);
232     vacc0x4567 = vqdmulhq_s32(vacc0x4567, vmultiplier);
233     vacc0x89AB = vqdmulhq_s32(vacc0x89AB, vmultiplier);
234     vacc0xCDEF = vqdmulhq_s32(vacc0xCDEF, vmultiplier);
235 
236     vacc0x0123 = vrshlq_s32(vacc0x0123, vright_post_shift);
237     vacc0x4567 = vrshlq_s32(vacc0x4567, vright_post_shift);
238     vacc0x89AB = vrshlq_s32(vacc0x89AB, vright_post_shift);
239     vacc0xCDEF = vrshlq_s32(vacc0xCDEF, 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     const int16x8_t vacc0x01234567 = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc0x0123), vacc0x4567), voutput_zero_point);
244     const int16x8_t vacc0x89ABCDEF = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc0x89AB), vacc0xCDEF), voutput_zero_point);
245 
246     int8x16_t vout0x0123456789ABCDEF = vqmovn_high_s16(vqmovn_s16(vacc0x01234567), vacc0x89ABCDEF);
247 #else
248     const int16x8_t vacc0x01234567 = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc0x0123), vqmovn_s32(vacc0x4567)), voutput_zero_point);
249     const int16x8_t vacc0x89ABCDEF = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc0x89AB), vqmovn_s32(vacc0xCDEF)), voutput_zero_point);
250 
251     int8x16_t vout0x0123456789ABCDEF = vcombine_s8(vqmovn_s16(vacc0x01234567), vqmovn_s16(vacc0x89ABCDEF));
252 #endif
253     const int8x16_t voutput_min = vld1q_dup_s8(&params->rndnu_neon.output_min);
254     const int8x16_t voutput_max = vld1q_dup_s8(&params->rndnu_neon.output_max);
255 
256     vout0x0123456789ABCDEF = vmaxq_s8(vout0x0123456789ABCDEF, voutput_min);
257 
258     vout0x0123456789ABCDEF = vminq_s8(vout0x0123456789ABCDEF, voutput_max);
259 
260     if (nc >= 16) {
261       // Main case where there the 16 columns fit in the destination.
262       vst1q_s8(c0 + 0, vout0x0123456789ABCDEF);
263 
264       // Advance to the next 16 columns.
265       c0 = (int8_t*) ((uintptr_t) c0 + cn_stride);
266 
267       a0 = (const int8_t*) ((uintptr_t) a0 - kc);
268 
269       nc -= 16;
270     } else {
271       // Final case where not all of the 16 columns fit in the destination.
272       int8x8_t vout0x01234567 = vget_low_s8(vout0x0123456789ABCDEF);
273       if (nc & 8) {
274         vst1_s8(c0, vout0x01234567); c0 += 8;
275         vout0x01234567 = vget_high_s8(vout0x0123456789ABCDEF);
276       }
277       if (nc & 4) {
278         vst1_lane_u32((void*) c0, vreinterpret_u32_s8(vout0x01234567), 0); c0 += 4;
279         vout0x01234567 = vext_s8(vout0x01234567, vout0x01234567, 4);
280       }
281       if (nc & 2) {
282         vst1_lane_u16((void*) c0, vreinterpret_u16_s8(vout0x01234567), 0); c0 += 2;
283         vout0x01234567 = vext_s8(vout0x01234567, vout0x01234567, 2);
284       }
285       if (nc & 1) {
286         vst1_lane_s8(c0, vout0x01234567, 0);
287       }
288 
289       nc = 0;
290     }
291   } while (nc != 0);
292 }
293