xref: /aosp_15_r20/external/XNNPACK/src/qs8-igemm/gen/1x16-minmax-rndnu-neon-mull-addw-dup.c (revision 4bdc94577ba0e567308109d787f7fec7b531ce36) 
1 // Auto-generated file. Do not edit!
2 //   Template: src/qs8-igemm/neon-mull-addw-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/common.h>
15 #include <xnnpack/gemm.h>
16 
17 
xnn_qs8_igemm_minmax_rndnu_ukernel_1x16__neon_mull_addw_dup(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_1x16__neon_mull_addw_dup(
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 <= 1);
34   assert(nc != 0);
35   assert(kc != 0);
36   assert(ks != 0);
37   assert(ks % (1 * 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 
45   do {
46     int32x4_t vacc0x0123 = vld1q_s32(w); w = (const void*) ((uintptr_t) w + 4 * sizeof(int32_t));
47     int32x4_t vacc0x4567 = vld1q_s32(w); w = (const void*) ((uintptr_t) w + 4 * sizeof(int32_t));
48     int32x4_t vacc0x89AB = vld1q_s32(w); w = (const void*) ((uintptr_t) w + 4 * sizeof(int32_t));
49     int32x4_t vacc0xCDEF = vld1q_s32(w); w = (const void*) ((uintptr_t) w + 4 * sizeof(int32_t));
50 
51     size_t p = ks;
52     do {
53       const int8_t* restrict a0 = a[0];
54       if XNN_UNPREDICTABLE(a0 != zero) {
55         a0 = (const int8_t*) ((uintptr_t) a0 + a_offset);
56       }
57       a += 1;
58 
59       size_t k = kc;
60       while (k >= 8 * sizeof(int8_t)) {
61         const int8x8_t va0 = vld1_s8(a0); a0 += 8;
62 
63         const int8x8_t vb01234567c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
64 
65         const int16x8_t vprod0x01234567c0 = vmull_s8(vb01234567c0, vdup_lane_s8(va0, 0));
66         vacc0x0123 = vaddw_s16(vacc0x0123, vget_low_s16(vprod0x01234567c0));
67         vacc0x4567 = vaddw_s16(vacc0x4567, vget_high_s16(vprod0x01234567c0));
68         const int8x8_t vb89ABCDEFc0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
69 
70         const int16x8_t vprod0x89ABCDEFc0 = vmull_s8(vb89ABCDEFc0, vdup_lane_s8(va0, 0));
71         vacc0x89AB = vaddw_s16(vacc0x89AB, vget_low_s16(vprod0x89ABCDEFc0));
72         vacc0xCDEF = vaddw_s16(vacc0xCDEF, vget_high_s16(vprod0x89ABCDEFc0));
73         const int8x8_t vb01234567c1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
74 
75         const int16x8_t vprod0x01234567c1 = vmull_s8(vb01234567c1, vdup_lane_s8(va0, 1));
76         vacc0x0123 = vaddw_s16(vacc0x0123, vget_low_s16(vprod0x01234567c1));
77         vacc0x4567 = vaddw_s16(vacc0x4567, vget_high_s16(vprod0x01234567c1));
78         const int8x8_t vb89ABCDEFc1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
79 
80         const int16x8_t vprod0x89ABCDEFc1 = vmull_s8(vb89ABCDEFc1, vdup_lane_s8(va0, 1));
81         vacc0x89AB = vaddw_s16(vacc0x89AB, vget_low_s16(vprod0x89ABCDEFc1));
82         vacc0xCDEF = vaddw_s16(vacc0xCDEF, vget_high_s16(vprod0x89ABCDEFc1));
83         const int8x8_t vb01234567c2 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
84 
85         const int16x8_t vprod0x01234567c2 = vmull_s8(vb01234567c2, vdup_lane_s8(va0, 2));
86         vacc0x0123 = vaddw_s16(vacc0x0123, vget_low_s16(vprod0x01234567c2));
87         vacc0x4567 = vaddw_s16(vacc0x4567, vget_high_s16(vprod0x01234567c2));
88         const int8x8_t vb89ABCDEFc2 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
89 
90         const int16x8_t vprod0x89ABCDEFc2 = vmull_s8(vb89ABCDEFc2, vdup_lane_s8(va0, 2));
91         vacc0x89AB = vaddw_s16(vacc0x89AB, vget_low_s16(vprod0x89ABCDEFc2));
92         vacc0xCDEF = vaddw_s16(vacc0xCDEF, vget_high_s16(vprod0x89ABCDEFc2));
93         const int8x8_t vb01234567c3 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
94 
95         const int16x8_t vprod0x01234567c3 = vmull_s8(vb01234567c3, vdup_lane_s8(va0, 3));
96         vacc0x0123 = vaddw_s16(vacc0x0123, vget_low_s16(vprod0x01234567c3));
97         vacc0x4567 = vaddw_s16(vacc0x4567, vget_high_s16(vprod0x01234567c3));
98         const int8x8_t vb89ABCDEFc3 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
99 
100         const int16x8_t vprod0x89ABCDEFc3 = vmull_s8(vb89ABCDEFc3, vdup_lane_s8(va0, 3));
101         vacc0x89AB = vaddw_s16(vacc0x89AB, vget_low_s16(vprod0x89ABCDEFc3));
102         vacc0xCDEF = vaddw_s16(vacc0xCDEF, vget_high_s16(vprod0x89ABCDEFc3));
103         const int8x8_t vb01234567c4 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
104 
105         const int16x8_t vprod0x01234567c4 = vmull_s8(vb01234567c4, vdup_lane_s8(va0, 4));
106         vacc0x0123 = vaddw_s16(vacc0x0123, vget_low_s16(vprod0x01234567c4));
107         vacc0x4567 = vaddw_s16(vacc0x4567, vget_high_s16(vprod0x01234567c4));
108         const int8x8_t vb89ABCDEFc4 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
109 
110         const int16x8_t vprod0x89ABCDEFc4 = vmull_s8(vb89ABCDEFc4, vdup_lane_s8(va0, 4));
111         vacc0x89AB = vaddw_s16(vacc0x89AB, vget_low_s16(vprod0x89ABCDEFc4));
112         vacc0xCDEF = vaddw_s16(vacc0xCDEF, vget_high_s16(vprod0x89ABCDEFc4));
113         const int8x8_t vb01234567c5 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
114 
115         const int16x8_t vprod0x01234567c5 = vmull_s8(vb01234567c5, vdup_lane_s8(va0, 5));
116         vacc0x0123 = vaddw_s16(vacc0x0123, vget_low_s16(vprod0x01234567c5));
117         vacc0x4567 = vaddw_s16(vacc0x4567, vget_high_s16(vprod0x01234567c5));
118         const int8x8_t vb89ABCDEFc5 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
119 
120         const int16x8_t vprod0x89ABCDEFc5 = vmull_s8(vb89ABCDEFc5, vdup_lane_s8(va0, 5));
121         vacc0x89AB = vaddw_s16(vacc0x89AB, vget_low_s16(vprod0x89ABCDEFc5));
122         vacc0xCDEF = vaddw_s16(vacc0xCDEF, vget_high_s16(vprod0x89ABCDEFc5));
123         const int8x8_t vb01234567c6 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
124 
125         const int16x8_t vprod0x01234567c6 = vmull_s8(vb01234567c6, vdup_lane_s8(va0, 6));
126         vacc0x0123 = vaddw_s16(vacc0x0123, vget_low_s16(vprod0x01234567c6));
127         vacc0x4567 = vaddw_s16(vacc0x4567, vget_high_s16(vprod0x01234567c6));
128         const int8x8_t vb89ABCDEFc6 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
129 
130         const int16x8_t vprod0x89ABCDEFc6 = vmull_s8(vb89ABCDEFc6, vdup_lane_s8(va0, 6));
131         vacc0x89AB = vaddw_s16(vacc0x89AB, vget_low_s16(vprod0x89ABCDEFc6));
132         vacc0xCDEF = vaddw_s16(vacc0xCDEF, vget_high_s16(vprod0x89ABCDEFc6));
133         const int8x8_t vb01234567c7 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
134 
135         const int16x8_t vprod0x01234567c7 = vmull_s8(vb01234567c7, vdup_lane_s8(va0, 7));
136         vacc0x0123 = vaddw_s16(vacc0x0123, vget_low_s16(vprod0x01234567c7));
137         vacc0x4567 = vaddw_s16(vacc0x4567, vget_high_s16(vprod0x01234567c7));
138         const int8x8_t vb89ABCDEFc7 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
139 
140         const int16x8_t vprod0x89ABCDEFc7 = vmull_s8(vb89ABCDEFc7, vdup_lane_s8(va0, 7));
141         vacc0x89AB = vaddw_s16(vacc0x89AB, vget_low_s16(vprod0x89ABCDEFc7));
142         vacc0xCDEF = vaddw_s16(vacc0xCDEF, vget_high_s16(vprod0x89ABCDEFc7));
143 
144         k -= 8 * sizeof(int8_t);
145       }
146       if XNN_UNLIKELY(k != 0) {
147         const int8x8_t va0 = vld1_s8(a0); a0 = (const int8_t*) ((uintptr_t) a0 + k);
148 
149         const int8x8_t vb01234567c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
150         const int8x8_t vb89ABCDEFc0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
151 
152         const int16x8_t vprod0x01234567c0 = vmull_s8(vb01234567c0, vdup_lane_s8(va0, 0));
153         vacc0x0123 = vaddw_s16(vacc0x0123, vget_low_s16(vprod0x01234567c0));
154         vacc0x4567 = vaddw_s16(vacc0x4567, vget_high_s16(vprod0x01234567c0));
155         const int16x8_t vprod0x89ABCDEFc0 = vmull_s8(vb89ABCDEFc0, vdup_lane_s8(va0, 0));
156         vacc0x89AB = vaddw_s16(vacc0x89AB, vget_low_s16(vprod0x89ABCDEFc0));
157         vacc0xCDEF = vaddw_s16(vacc0xCDEF, vget_high_s16(vprod0x89ABCDEFc0));
158 
159         if (k >= 2 * sizeof(int8_t)) {
160           const int8x8_t vb01234567c1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
161           const int8x8_t vb89ABCDEFc1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
162 
163           const int16x8_t vprod0x01234567c1 = vmull_s8(vb01234567c1, vdup_lane_s8(va0, 1));
164           vacc0x0123 = vaddw_s16(vacc0x0123, vget_low_s16(vprod0x01234567c1));
165           vacc0x4567 = vaddw_s16(vacc0x4567, vget_high_s16(vprod0x01234567c1));
166           const int16x8_t vprod0x89ABCDEFc1 = vmull_s8(vb89ABCDEFc1, vdup_lane_s8(va0, 1));
167           vacc0x89AB = vaddw_s16(vacc0x89AB, vget_low_s16(vprod0x89ABCDEFc1));
168           vacc0xCDEF = vaddw_s16(vacc0xCDEF, vget_high_s16(vprod0x89ABCDEFc1));
169 
170           if (k > 2 * sizeof(int8_t)) {
171             const int8x8_t vb01234567c2 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
172             const int8x8_t vb89ABCDEFc2 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
173 
174             const int16x8_t vprod0x01234567c2 = vmull_s8(vb01234567c2, vdup_lane_s8(va0, 2));
175             vacc0x0123 = vaddw_s16(vacc0x0123, vget_low_s16(vprod0x01234567c2));
176             vacc0x4567 = vaddw_s16(vacc0x4567, vget_high_s16(vprod0x01234567c2));
177             const int16x8_t vprod0x89ABCDEFc2 = vmull_s8(vb89ABCDEFc2, vdup_lane_s8(va0, 2));
178             vacc0x89AB = vaddw_s16(vacc0x89AB, vget_low_s16(vprod0x89ABCDEFc2));
179             vacc0xCDEF = vaddw_s16(vacc0xCDEF, vget_high_s16(vprod0x89ABCDEFc2));
180 
181             if (k >= 4 * sizeof(int8_t)) {
182               const int8x8_t vb01234567c3 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
183               const int8x8_t vb89ABCDEFc3 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
184 
185               const int16x8_t vprod0x01234567c3 = vmull_s8(vb01234567c3, vdup_lane_s8(va0, 3));
186               vacc0x0123 = vaddw_s16(vacc0x0123, vget_low_s16(vprod0x01234567c3));
187               vacc0x4567 = vaddw_s16(vacc0x4567, vget_high_s16(vprod0x01234567c3));
188               const int16x8_t vprod0x89ABCDEFc3 = vmull_s8(vb89ABCDEFc3, vdup_lane_s8(va0, 3));
189               vacc0x89AB = vaddw_s16(vacc0x89AB, vget_low_s16(vprod0x89ABCDEFc3));
190               vacc0xCDEF = vaddw_s16(vacc0xCDEF, vget_high_s16(vprod0x89ABCDEFc3));
191 
192               if (k > 4 * sizeof(int8_t)) {
193                 const int8x8_t vb01234567c4 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
194                 const int8x8_t vb89ABCDEFc4 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
195 
196                 const int16x8_t vprod0x01234567c4 = vmull_s8(vb01234567c4, vdup_lane_s8(va0, 4));
197                 vacc0x0123 = vaddw_s16(vacc0x0123, vget_low_s16(vprod0x01234567c4));
198                 vacc0x4567 = vaddw_s16(vacc0x4567, vget_high_s16(vprod0x01234567c4));
199                 const int16x8_t vprod0x89ABCDEFc4 = vmull_s8(vb89ABCDEFc4, vdup_lane_s8(va0, 4));
200                 vacc0x89AB = vaddw_s16(vacc0x89AB, vget_low_s16(vprod0x89ABCDEFc4));
201                 vacc0xCDEF = vaddw_s16(vacc0xCDEF, vget_high_s16(vprod0x89ABCDEFc4));
202 
203                 if (k >= 6 * sizeof(int8_t)) {
204                   const int8x8_t vb01234567c5 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
205                   const int8x8_t vb89ABCDEFc5 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
206 
207                   const int16x8_t vprod0x01234567c5 = vmull_s8(vb01234567c5, vdup_lane_s8(va0, 5));
208                   vacc0x0123 = vaddw_s16(vacc0x0123, vget_low_s16(vprod0x01234567c5));
209                   vacc0x4567 = vaddw_s16(vacc0x4567, vget_high_s16(vprod0x01234567c5));
210                   const int16x8_t vprod0x89ABCDEFc5 = vmull_s8(vb89ABCDEFc5, vdup_lane_s8(va0, 5));
211                   vacc0x89AB = vaddw_s16(vacc0x89AB, vget_low_s16(vprod0x89ABCDEFc5));
212                   vacc0xCDEF = vaddw_s16(vacc0xCDEF, vget_high_s16(vprod0x89ABCDEFc5));
213 
214                   if (k > 6 * sizeof(int8_t)) {
215                     const int8x8_t vb01234567c6 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
216                     const int8x8_t vb89ABCDEFc6 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
217 
218                     const int16x8_t vprod0x01234567c6 = vmull_s8(vb01234567c6, vdup_lane_s8(va0, 6));
219                     vacc0x0123 = vaddw_s16(vacc0x0123, vget_low_s16(vprod0x01234567c6));
220                     vacc0x4567 = vaddw_s16(vacc0x4567, vget_high_s16(vprod0x01234567c6));
221                     const int16x8_t vprod0x89ABCDEFc6 = vmull_s8(vb89ABCDEFc6, vdup_lane_s8(va0, 6));
222                     vacc0x89AB = vaddw_s16(vacc0x89AB, vget_low_s16(vprod0x89ABCDEFc6));
223                     vacc0xCDEF = vaddw_s16(vacc0xCDEF, vget_high_s16(vprod0x89ABCDEFc6));
224                   }
225                 }
226               }
227             }
228           }
229         }
230       }
231       p -= 1 * sizeof(void*);
232     } while (p != 0);
233 
234     const int32x4_t vright_pre_shift = vld1q_dup_s32(&params->rndnu_neon.right_pre_shift);
235     const int32x4_t vmultiplier = vld1q_dup_s32(&params->rndnu_neon.multiplier);
236     const int32x4_t vright_post_shift = vld1q_dup_s32(&params->rndnu_neon.right_post_shift);
237 
238     vacc0x0123 = vqshlq_s32(vacc0x0123, vright_pre_shift);
239     vacc0x4567 = vqshlq_s32(vacc0x4567, vright_pre_shift);
240     vacc0x89AB = vqshlq_s32(vacc0x89AB, vright_pre_shift);
241     vacc0xCDEF = vqshlq_s32(vacc0xCDEF, vright_pre_shift);
242 
243     vacc0x0123 = vqdmulhq_s32(vacc0x0123, vmultiplier);
244     vacc0x4567 = vqdmulhq_s32(vacc0x4567, vmultiplier);
245     vacc0x89AB = vqdmulhq_s32(vacc0x89AB, vmultiplier);
246     vacc0xCDEF = vqdmulhq_s32(vacc0xCDEF, vmultiplier);
247 
248     vacc0x0123 = vrshlq_s32(vacc0x0123, vright_post_shift);
249     vacc0x4567 = vrshlq_s32(vacc0x4567, vright_post_shift);
250     vacc0x89AB = vrshlq_s32(vacc0x89AB, vright_post_shift);
251     vacc0xCDEF = vrshlq_s32(vacc0xCDEF, vright_post_shift);
252 
253     const int16x8_t voutput_zero_point = vld1q_dup_s16(&params->rndnu_neon.output_zero_point);
254 #if XNN_ARCH_ARM64
255     const int16x8_t vacc0x01234567 = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc0x0123), vacc0x4567), voutput_zero_point);
256     const int16x8_t vacc0x89ABCDEF = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc0x89AB), vacc0xCDEF), voutput_zero_point);
257 
258     int8x16_t vout0x0123456789ABCDEF = vqmovn_high_s16(vqmovn_s16(vacc0x01234567), vacc0x89ABCDEF);
259 #else
260     const int16x8_t vacc0x01234567 = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc0x0123), vqmovn_s32(vacc0x4567)), voutput_zero_point);
261     const int16x8_t vacc0x89ABCDEF = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc0x89AB), vqmovn_s32(vacc0xCDEF)), voutput_zero_point);
262 
263     int8x16_t vout0x0123456789ABCDEF = vcombine_s8(vqmovn_s16(vacc0x01234567), vqmovn_s16(vacc0x89ABCDEF));
264 #endif
265     const int8x16_t voutput_min = vld1q_dup_s8(&params->rndnu_neon.output_min);
266     const int8x16_t voutput_max = vld1q_dup_s8(&params->rndnu_neon.output_max);
267 
268     vout0x0123456789ABCDEF = vmaxq_s8(vout0x0123456789ABCDEF, voutput_min);
269 
270     vout0x0123456789ABCDEF = vminq_s8(vout0x0123456789ABCDEF, voutput_max);
271 
272     if (nc >= 16) {
273       vst1q_s8(c0 + 0, vout0x0123456789ABCDEF);
274 
275       c0 = (int8_t*) ((uintptr_t) c0 + cn_stride);
276 
277       a = (const int8_t**restrict) ((uintptr_t) a - ks);
278 
279       nc -= 16;
280     } else {
281       int8x8_t vout0x01234567 = vget_low_s8(vout0x0123456789ABCDEF);
282       if (nc & 8) {
283         vst1_s8(c0, vout0x01234567); c0 += 8;
284         vout0x01234567 = vget_high_s8(vout0x0123456789ABCDEF);
285       }
286       if (nc & 4) {
287         vst1_lane_u32((void*) c0, vreinterpret_u32_s8(vout0x01234567), 0); c0 += 4;
288         vout0x01234567 = vext_s8(vout0x01234567, vout0x01234567, 4);
289       }
290       if (nc & 2) {
291         vst1_lane_u16((void*) c0, vreinterpret_u16_s8(vout0x01234567), 0); c0 += 2;
292         vout0x01234567 = vext_s8(vout0x01234567, vout0x01234567, 2);
293       }
294       if (nc & 1) {
295         vst1_lane_s8(c0, vout0x01234567, 0);
296       }
297 
298       nc = 0;
299     }
300   } while (nc != 0);
301 }
302