xref: /aosp_15_r20/external/XNNPACK/src/qs8-igemm/gen/1x16c2-minmax-rndnu-neon-mull-ld1r.c (revision 4bdc94577ba0e567308109d787f7fec7b531ce36) 
1 // Auto-generated file. Do not edit!
2 //   Template: src/qs8-igemm/c2-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 
xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c2__neon_mull_ld1r(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)])17 void xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c2__neon_mull_ld1r(
18     size_t mr,
19     size_t nc,
20     size_t kc,
21     size_t ks,
22     const int8_t** restrict a,
23     const void* restrict w,
24     int8_t* restrict c,
25     size_t cm_stride,
26     size_t cn_stride,
27     size_t a_offset,
28     const int8_t* zero,
29     const union xnn_qs8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
30 {
31   assert(mr != 0);
32   assert(mr <= 1);
33   assert(nc != 0);
34   assert(kc != 0);
35   assert(ks != 0);
36   assert(ks % (1 * sizeof(void*)) == 0);
37   assert(a_offset % sizeof(int8_t) == 0);
38   assert(a != NULL);
39   assert(w != NULL);
40   assert(c != NULL);
41 
42   kc = round_up_po2(kc, 2 * sizeof(int8_t));
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 
61 
62       while (k >= 8 * sizeof(int8_t)) {
63         const int16x4_t va00 = vld1_dup_s16((const void*)a0);
64         const int16x4_t va01 = vld1_dup_s16((const void*)(a0 + 2));
65         const int16x4_t va02 = vld1_dup_s16((const void*)(a0 + 4));
66         const int16x4_t va03 = vld1_dup_s16((const void*)(a0 + 6)); a0 += 8;
67 
68         const int8x8_t vb0123c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
69         const int8x8_t vb4567c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
70         const int8x8_t vb89ABc0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
71         const int8x8_t vbCDEFc0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
72         const int8x8_t vb0123c1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
73         const int8x8_t vb4567c1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
74         const int8x8_t vb89ABc1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
75         const int8x8_t vbCDEFc1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
76         const int8x8_t vb0123c2 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
77         const int8x8_t vb4567c2 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
78         const int8x8_t vb89ABc2 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
79         const int8x8_t vbCDEFc2 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
80         const int8x8_t vb0123c3 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
81         const int8x8_t vb4567c3 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
82         const int8x8_t vb89ABc3 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
83         const int8x8_t vbCDEFc3 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
84 
85         const int8x8_t va0c0 = vreinterpret_s8_s16(va00);
86 
87         const int16x8_t vprod0x0123c0 = vmull_s8(vb0123c0, va0c0);
88         vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c0);
89         const int16x8_t vprod0x4567c0 = vmull_s8(vb4567c0, va0c0);
90         vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c0);
91         const int16x8_t vprod0x89ABc0 = vmull_s8(vb89ABc0, va0c0);
92         vacc0x89AB = vpadalq_s16(vacc0x89AB, vprod0x89ABc0);
93         const int16x8_t vprod0xCDEFc0 = vmull_s8(vbCDEFc0, va0c0);
94         vacc0xCDEF = vpadalq_s16(vacc0xCDEF, vprod0xCDEFc0);
95         const int8x8_t va0c1 = vreinterpret_s8_s16(va01);
96 
97         const int16x8_t vprod0x0123c1 = vmull_s8(vb0123c1, va0c1);
98         vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c1);
99         const int16x8_t vprod0x4567c1 = vmull_s8(vb4567c1, va0c1);
100         vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c1);
101         const int16x8_t vprod0x89ABc1 = vmull_s8(vb89ABc1, va0c1);
102         vacc0x89AB = vpadalq_s16(vacc0x89AB, vprod0x89ABc1);
103         const int16x8_t vprod0xCDEFc1 = vmull_s8(vbCDEFc1, va0c1);
104         vacc0xCDEF = vpadalq_s16(vacc0xCDEF, vprod0xCDEFc1);
105         const int8x8_t va0c2 = vreinterpret_s8_s16(va02);
106 
107         const int16x8_t vprod0x0123c2 = vmull_s8(vb0123c2, va0c2);
108         vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c2);
109         const int16x8_t vprod0x4567c2 = vmull_s8(vb4567c2, va0c2);
110         vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c2);
111         const int16x8_t vprod0x89ABc2 = vmull_s8(vb89ABc2, va0c2);
112         vacc0x89AB = vpadalq_s16(vacc0x89AB, vprod0x89ABc2);
113         const int16x8_t vprod0xCDEFc2 = vmull_s8(vbCDEFc2, va0c2);
114         vacc0xCDEF = vpadalq_s16(vacc0xCDEF, vprod0xCDEFc2);
115         const int8x8_t va0c3 = vreinterpret_s8_s16(va03);
116 
117         const int16x8_t vprod0x0123c3 = vmull_s8(vb0123c3, va0c3);
118         vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c3);
119         const int16x8_t vprod0x4567c3 = vmull_s8(vb4567c3, va0c3);
120         vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c3);
121         const int16x8_t vprod0x89ABc3 = vmull_s8(vb89ABc3, va0c3);
122         vacc0x89AB = vpadalq_s16(vacc0x89AB, vprod0x89ABc3);
123         const int16x8_t vprod0xCDEFc3 = vmull_s8(vbCDEFc3, va0c3);
124         vacc0xCDEF = vpadalq_s16(vacc0xCDEF, vprod0xCDEFc3);
125 
126         k -= 8 * sizeof(int8_t);
127       }
128 
129       if XNN_UNLIKELY(k != 0) {
130         const int8x8_t va0 = vld1_s8(a0); a0 = (const int8_t*) ((uintptr_t) a0 + k);
131 
132         const int8x8_t vb0123c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
133         const int8x8_t vb4567c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
134         const int8x8_t vb89ABc0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
135         const int8x8_t vbCDEFc0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
136 
137         const int8x8_t va0c0 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0), 0));
138         const int16x8_t vprod0x0123c0 = vmull_s8(vb0123c0, va0c0);
139         vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c0);
140         const int16x8_t vprod0x4567c0 = vmull_s8(vb4567c0, va0c0);
141         vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c0);
142         const int16x8_t vprod0x89ABc0 = vmull_s8(vb89ABc0, va0c0);
143         vacc0x89AB = vpadalq_s16(vacc0x89AB, vprod0x89ABc0);
144         const int16x8_t vprod0xCDEFc0 = vmull_s8(vbCDEFc0, va0c0);
145         vacc0xCDEF = vpadalq_s16(vacc0xCDEF, vprod0xCDEFc0);
146 
147         if (k > 2 * sizeof(int8_t)) {
148           const int8x8_t vb0123c1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
149           const int8x8_t vb4567c1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
150           const int8x8_t vb89ABc1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
151           const int8x8_t vbCDEFc1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
152 
153           const int8x8_t va0c1 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0), 1));
154           const int16x8_t vprod0x0123c1 = vmull_s8(vb0123c1, va0c1);
155           vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c1);
156           const int16x8_t vprod0x4567c1 = vmull_s8(vb4567c1, va0c1);
157           vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c1);
158           const int16x8_t vprod0x89ABc1 = vmull_s8(vb89ABc1, va0c1);
159           vacc0x89AB = vpadalq_s16(vacc0x89AB, vprod0x89ABc1);
160           const int16x8_t vprod0xCDEFc1 = vmull_s8(vbCDEFc1, va0c1);
161           vacc0xCDEF = vpadalq_s16(vacc0xCDEF, vprod0xCDEFc1);
162 
163           if (k > 4 * sizeof(int8_t)) {
164             const int8x8_t vb0123c2 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
165             const int8x8_t vb4567c2 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
166             const int8x8_t vb89ABc2 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
167             const int8x8_t vbCDEFc2 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
168 
169             const int8x8_t va0c2 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0), 2));
170             const int16x8_t vprod0x0123c2 = vmull_s8(vb0123c2, va0c2);
171             vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c2);
172             const int16x8_t vprod0x4567c2 = vmull_s8(vb4567c2, va0c2);
173             vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c2);
174             const int16x8_t vprod0x89ABc2 = vmull_s8(vb89ABc2, va0c2);
175             vacc0x89AB = vpadalq_s16(vacc0x89AB, vprod0x89ABc2);
176             const int16x8_t vprod0xCDEFc2 = vmull_s8(vbCDEFc2, va0c2);
177             vacc0xCDEF = vpadalq_s16(vacc0xCDEF, vprod0xCDEFc2);
178           }
179         }
180       }
181       p -= 1 * sizeof(void*);
182     } while (p != 0);
183 
184     const int32x4_t vright_pre_shift = vld1q_dup_s32(&params->rndnu_neon.right_pre_shift);
185     const int32x4_t vmultiplier = vld1q_dup_s32(&params->rndnu_neon.multiplier);
186     const int32x4_t vright_post_shift = vld1q_dup_s32(&params->rndnu_neon.right_post_shift);
187 
188     vacc0x0123 = vqshlq_s32(vacc0x0123, vright_pre_shift);
189     vacc0x4567 = vqshlq_s32(vacc0x4567, vright_pre_shift);
190     vacc0x89AB = vqshlq_s32(vacc0x89AB, vright_pre_shift);
191     vacc0xCDEF = vqshlq_s32(vacc0xCDEF, vright_pre_shift);
192 
193     vacc0x0123 = vqdmulhq_s32(vacc0x0123, vmultiplier);
194     vacc0x4567 = vqdmulhq_s32(vacc0x4567, vmultiplier);
195     vacc0x89AB = vqdmulhq_s32(vacc0x89AB, vmultiplier);
196     vacc0xCDEF = vqdmulhq_s32(vacc0xCDEF, vmultiplier);
197 
198     vacc0x0123 = vrshlq_s32(vacc0x0123, vright_post_shift);
199     vacc0x4567 = vrshlq_s32(vacc0x4567, vright_post_shift);
200     vacc0x89AB = vrshlq_s32(vacc0x89AB, vright_post_shift);
201     vacc0xCDEF = vrshlq_s32(vacc0xCDEF, vright_post_shift);
202 
203     const int16x8_t voutput_zero_point = vld1q_dup_s16(&params->rndnu_neon.output_zero_point);
204 #if XNN_ARCH_ARM64
205     int16x8_t vacc0x01234567 = vqmovn_high_s32(vqmovn_s32(vacc0x0123), vacc0x4567);
206     int16x8_t vacc0x89ABCDEF = vqmovn_high_s32(vqmovn_s32(vacc0x89AB), vacc0xCDEF);
207 
208     vacc0x01234567 = vqaddq_s16(vacc0x01234567, voutput_zero_point);
209     vacc0x89ABCDEF = vqaddq_s16(vacc0x89ABCDEF, voutput_zero_point);
210 
211     int8x16_t vout0x0123456789ABCDEF = vqmovn_high_s16(vqmovn_s16(vacc0x01234567), vacc0x89ABCDEF);
212 #else
213     int16x8_t vacc0x01234567 = vcombine_s16(vqmovn_s32(vacc0x0123), vqmovn_s32(vacc0x4567));
214     int16x8_t vacc0x89ABCDEF = vcombine_s16(vqmovn_s32(vacc0x89AB), vqmovn_s32(vacc0xCDEF));
215 
216     vacc0x01234567 = vqaddq_s16(vacc0x01234567, voutput_zero_point);
217     vacc0x89ABCDEF = vqaddq_s16(vacc0x89ABCDEF, voutput_zero_point);
218 
219     int8x16_t vout0x0123456789ABCDEF = vcombine_s8(vqmovn_s16(vacc0x01234567), vqmovn_s16(vacc0x89ABCDEF));
220 #endif
221 
222     const int8x16_t voutput_min = vld1q_dup_s8(&params->rndnu_neon.output_min);
223     vout0x0123456789ABCDEF = vmaxq_s8(vout0x0123456789ABCDEF, voutput_min);
224 
225     const int8x16_t voutput_max = vld1q_dup_s8(&params->rndnu_neon.output_max);
226     vout0x0123456789ABCDEF = vminq_s8(vout0x0123456789ABCDEF, voutput_max);
227 
228     if (nc >= 16) {
229       vst1q_s8(c0 + 0, vout0x0123456789ABCDEF);
230 
231       c0 = (int8_t*) ((uintptr_t) c0 + cn_stride);
232 
233       a = (const int8_t**restrict) ((uintptr_t) a - ks);
234 
235       nc -= 16;
236     } else {
237       int8x8_t vout0x01234567 = vget_low_s8(vout0x0123456789ABCDEF);
238       if (nc & 8) {
239         vst1_s8(c0, vout0x01234567); c0 += 8;
240         vout0x01234567 = vget_high_s8(vout0x0123456789ABCDEF);
241       }
242       if (nc & 4) {
243         vst1_lane_u32((void*) c0, vreinterpret_u32_s8(vout0x01234567), 0); c0 += 4;
244         vout0x01234567 = vext_s8(vout0x01234567, vout0x01234567, 4);
245       }
246       if (nc & 2) {
247         vst1_lane_u16((void*) c0, vreinterpret_u16_s8(vout0x01234567), 0); c0 += 2;
248         vout0x01234567 = vext_s8(vout0x01234567, vout0x01234567, 2);
249       }
250       if (nc & 1) {
251         vst1_lane_s8(c0, vout0x01234567, 0);
252       }
253 
254       nc = 0;
255     }
256   } while (nc != 0);
257 }
258