xref: /aosp_15_r20/external/XNNPACK/src/qs8-igemm/gen/2x8c2-minmax-rndnu-neon-mull-ld2r.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_2x8c2__neon_mull_ld2r(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_2x8c2__neon_mull_ld2r(
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 <= 2);
33   assert(nc != 0);
34   assert(kc != 0);
35   assert(ks != 0);
36   assert(ks % (2 * 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   int8_t* c1 = (int8_t*) ((uintptr_t) c0 + cm_stride);
45   if XNN_UNPREDICTABLE(mr != 2) {
46     c1 = c0;
47   }
48 
49   do {
50     int32x4_t vacc0x0123 = vld1q_s32(w); w = (const void*) ((uintptr_t) w + 4 * sizeof(int32_t));
51     int32x4_t vacc0x4567 = vld1q_s32(w); w = (const void*) ((uintptr_t) w + 4 * sizeof(int32_t));
52     int32x4_t vacc1x0123 = vacc0x0123;
53     int32x4_t vacc1x4567 = vacc0x4567;
54 
55     size_t p = ks;
56     do {
57       const int8_t* restrict a0 = a[0];
58       if XNN_UNPREDICTABLE(a0 != zero) {
59         a0 = (const int8_t*) ((uintptr_t) a0 + a_offset);
60       }
61       const int8_t* restrict a1 = a[1];
62       if XNN_UNPREDICTABLE(a1 != zero) {
63         a1 = (const int8_t*) ((uintptr_t) a1 + a_offset);
64       }
65       a += 2;
66 
67       size_t k = kc;
68 
69 
70       while (k >= 8 * sizeof(int8_t)) {
71         const int16x4x2_t va00 = vld2_dup_s16((const void*)a0);
72         const int16x4x2_t va01 = vld2_dup_s16((const void*)(a0 + 4)); a0 += 8;
73         const int16x4x2_t va10 = vld2_dup_s16((const void*)a1);
74         const int16x4x2_t va11 = vld2_dup_s16((const void*)(a1 + 4)); a1 += 8;
75 
76         const int8x8_t vb0123c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
77         const int8x8_t vb4567c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
78         const int8x8_t vb0123c1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
79         const int8x8_t vb4567c1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
80         const int8x8_t vb0123c2 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
81         const int8x8_t vb4567c2 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
82         const int8x8_t vb0123c3 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
83         const int8x8_t vb4567c3 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
84 
85         const int8x8_t va0c0 = vreinterpret_s8_s16(va00.val[0]);
86         const int8x8_t va1c0 = vreinterpret_s8_s16(va10.val[0]);
87 
88         const int16x8_t vprod0x0123c0 = vmull_s8(vb0123c0, va0c0);
89         const int16x8_t vprod1x0123c0 = vmull_s8(vb0123c0, va1c0);
90         vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c0);
91         vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c0);
92         const int16x8_t vprod0x4567c0 = vmull_s8(vb4567c0, va0c0);
93         const int16x8_t vprod1x4567c0 = vmull_s8(vb4567c0, va1c0);
94         vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c0);
95         vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c0);
96         const int8x8_t va0c1 = vreinterpret_s8_s16(va00.val[1]);
97         const int8x8_t va1c1 = vreinterpret_s8_s16(va10.val[1]);
98 
99         const int16x8_t vprod0x0123c1 = vmull_s8(vb0123c1, va0c1);
100         const int16x8_t vprod1x0123c1 = vmull_s8(vb0123c1, va1c1);
101         vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c1);
102         vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c1);
103         const int16x8_t vprod0x4567c1 = vmull_s8(vb4567c1, va0c1);
104         const int16x8_t vprod1x4567c1 = vmull_s8(vb4567c1, va1c1);
105         vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c1);
106         vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c1);
107         const int8x8_t va0c2 = vreinterpret_s8_s16(va01.val[0]);
108         const int8x8_t va1c2 = vreinterpret_s8_s16(va11.val[0]);
109 
110         const int16x8_t vprod0x0123c2 = vmull_s8(vb0123c2, va0c2);
111         const int16x8_t vprod1x0123c2 = vmull_s8(vb0123c2, va1c2);
112         vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c2);
113         vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c2);
114         const int16x8_t vprod0x4567c2 = vmull_s8(vb4567c2, va0c2);
115         const int16x8_t vprod1x4567c2 = vmull_s8(vb4567c2, va1c2);
116         vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c2);
117         vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c2);
118         const int8x8_t va0c3 = vreinterpret_s8_s16(va01.val[1]);
119         const int8x8_t va1c3 = vreinterpret_s8_s16(va11.val[1]);
120 
121         const int16x8_t vprod0x0123c3 = vmull_s8(vb0123c3, va0c3);
122         const int16x8_t vprod1x0123c3 = vmull_s8(vb0123c3, va1c3);
123         vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c3);
124         vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c3);
125         const int16x8_t vprod0x4567c3 = vmull_s8(vb4567c3, va0c3);
126         const int16x8_t vprod1x4567c3 = vmull_s8(vb4567c3, va1c3);
127         vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c3);
128         vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c3);
129 
130         k -= 8 * sizeof(int8_t);
131       }
132 
133       if XNN_UNLIKELY(k != 0) {
134         const int8x8_t va0 = vld1_s8(a0); a0 = (const int8_t*) ((uintptr_t) a0 + k);
135         const int8x8_t va1 = vld1_s8(a1); a1 = (const int8_t*) ((uintptr_t) a1 + k);
136 
137         const int8x8_t vb0123c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
138         const int8x8_t vb4567c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
139 
140         const int8x8_t va0c0 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0), 0));
141         const int16x8_t vprod0x0123c0 = vmull_s8(vb0123c0, va0c0);
142         vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c0);
143         const int16x8_t vprod0x4567c0 = vmull_s8(vb4567c0, va0c0);
144         vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c0);
145         const int8x8_t va1c0 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va1), 0));
146         const int16x8_t vprod1x0123c0 = vmull_s8(vb0123c0, va1c0);
147         vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c0);
148         const int16x8_t vprod1x4567c0 = vmull_s8(vb4567c0, va1c0);
149         vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c0);
150 
151         if (k > 2 * sizeof(int8_t)) {
152           const int8x8_t vb0123c1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
153           const int8x8_t vb4567c1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
154 
155           const int8x8_t va0c1 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0), 1));
156           const int16x8_t vprod0x0123c1 = vmull_s8(vb0123c1, va0c1);
157           vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c1);
158           const int16x8_t vprod0x4567c1 = vmull_s8(vb4567c1, va0c1);
159           vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c1);
160           const int8x8_t va1c1 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va1), 1));
161           const int16x8_t vprod1x0123c1 = vmull_s8(vb0123c1, va1c1);
162           vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c1);
163           const int16x8_t vprod1x4567c1 = vmull_s8(vb4567c1, va1c1);
164           vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c1);
165 
166           if (k > 4 * sizeof(int8_t)) {
167             const int8x8_t vb0123c2 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
168             const int8x8_t vb4567c2 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
169 
170             const int8x8_t va0c2 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0), 2));
171             const int16x8_t vprod0x0123c2 = vmull_s8(vb0123c2, va0c2);
172             vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c2);
173             const int16x8_t vprod0x4567c2 = vmull_s8(vb4567c2, va0c2);
174             vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c2);
175             const int8x8_t va1c2 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va1), 2));
176             const int16x8_t vprod1x0123c2 = vmull_s8(vb0123c2, va1c2);
177             vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c2);
178             const int16x8_t vprod1x4567c2 = vmull_s8(vb4567c2, va1c2);
179             vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c2);
180           }
181         }
182       }
183       p -= 2 * sizeof(void*);
184     } while (p != 0);
185 
186     const int32x4_t vright_pre_shift = vld1q_dup_s32(&params->rndnu_neon.right_pre_shift);
187     const int32x4_t vmultiplier = vld1q_dup_s32(&params->rndnu_neon.multiplier);
188     const int32x4_t vright_post_shift = vld1q_dup_s32(&params->rndnu_neon.right_post_shift);
189 
190     vacc0x0123 = vqshlq_s32(vacc0x0123, vright_pre_shift);
191     vacc0x4567 = vqshlq_s32(vacc0x4567, vright_pre_shift);
192     vacc1x0123 = vqshlq_s32(vacc1x0123, vright_pre_shift);
193     vacc1x4567 = vqshlq_s32(vacc1x4567, vright_pre_shift);
194 
195     vacc0x0123 = vqdmulhq_s32(vacc0x0123, vmultiplier);
196     vacc0x4567 = vqdmulhq_s32(vacc0x4567, vmultiplier);
197     vacc1x0123 = vqdmulhq_s32(vacc1x0123, vmultiplier);
198     vacc1x4567 = vqdmulhq_s32(vacc1x4567, vmultiplier);
199 
200     vacc0x0123 = vrshlq_s32(vacc0x0123, vright_post_shift);
201     vacc0x4567 = vrshlq_s32(vacc0x4567, vright_post_shift);
202     vacc1x0123 = vrshlq_s32(vacc1x0123, vright_post_shift);
203     vacc1x4567 = vrshlq_s32(vacc1x4567, vright_post_shift);
204 
205     const int16x8_t voutput_zero_point = vld1q_dup_s16(&params->rndnu_neon.output_zero_point);
206 #if XNN_ARCH_ARM64
207     int16x8_t vacc0x01234567 = vqmovn_high_s32(vqmovn_s32(vacc0x0123), vacc0x4567);
208     int16x8_t vacc1x01234567 = vqmovn_high_s32(vqmovn_s32(vacc1x0123), vacc1x4567);
209 
210     vacc0x01234567 = vqaddq_s16(vacc0x01234567, voutput_zero_point);
211     vacc1x01234567 = vqaddq_s16(vacc1x01234567, voutput_zero_point);
212 
213     int8x16_t vout0x01234567_1x01234567 = vqmovn_high_s16(vqmovn_s16(vacc0x01234567), vacc1x01234567);
214 #else
215     int16x8_t vacc0x01234567 = vcombine_s16(vqmovn_s32(vacc0x0123), vqmovn_s32(vacc0x4567));
216     int16x8_t vacc1x01234567 = vcombine_s16(vqmovn_s32(vacc1x0123), vqmovn_s32(vacc1x4567));
217 
218     vacc0x01234567 = vqaddq_s16(vacc0x01234567, voutput_zero_point);
219     vacc1x01234567 = vqaddq_s16(vacc1x01234567, voutput_zero_point);
220 
221     int8x16_t vout0x01234567_1x01234567 = vcombine_s8(vqmovn_s16(vacc0x01234567), vqmovn_s16(vacc1x01234567));
222 #endif
223 
224     const int8x16_t voutput_min = vld1q_dup_s8(&params->rndnu_neon.output_min);
225     vout0x01234567_1x01234567 = vmaxq_s8(vout0x01234567_1x01234567, voutput_min);
226 
227     const int8x16_t voutput_max = vld1q_dup_s8(&params->rndnu_neon.output_max);
228     vout0x01234567_1x01234567 = vminq_s8(vout0x01234567_1x01234567, voutput_max);
229 
230     if (nc >= 8) {
231       vst1_s8(c1 + 0, vget_high_s8(vout0x01234567_1x01234567));
232       vst1_s8(c0 + 0, vget_low_s8(vout0x01234567_1x01234567));
233 
234       c1 = (int8_t*) ((uintptr_t) c1 + cn_stride);
235       c0 = (int8_t*) ((uintptr_t) c0 + cn_stride);
236 
237       a = (const int8_t**restrict) ((uintptr_t) a - ks);
238 
239       nc -= 8;
240     } else {
241       if (nc & 4) {
242         vst1q_lane_u32((void*) c1, vreinterpretq_u32_s8(vout0x01234567_1x01234567), 2); c1 += 4;
243         vst1q_lane_u32((void*) c0, vreinterpretq_u32_s8(vout0x01234567_1x01234567), 0); c0 += 4;
244         vout0x01234567_1x01234567 = vextq_s8(vout0x01234567_1x01234567, vout0x01234567_1x01234567, 4);
245       }
246       if (nc & 2) {
247         vst1q_lane_u16((void*) c1, vreinterpretq_u16_s8(vout0x01234567_1x01234567), 4); c1 += 2;
248         vst1q_lane_u16((void*) c0, vreinterpretq_u16_s8(vout0x01234567_1x01234567), 0); c0 += 2;
249         vout0x01234567_1x01234567 = vextq_s8(vout0x01234567_1x01234567, vout0x01234567_1x01234567, 2);
250       }
251       if (nc & 1) {
252         vst1q_lane_s8(c1, vout0x01234567_1x01234567, 8);
253         vst1q_lane_s8(c0, vout0x01234567_1x01234567, 0);
254       }
255 
256       nc = 0;
257     }
258   } while (nc != 0);
259 }
260