xref: /aosp_15_r20/external/XNNPACK/src/qs8-igemm/gen/2x8c4-minmax-rndnu-neon-mull-ld1r.c (revision 4bdc94577ba0e567308109d787f7fec7b531ce36)
1 // Auto-generated file. Do not edit!
2 //   Template: src/qs8-igemm/c4-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 
17 
xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c4__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)])18 void xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c4__neon_mull_ld1r(
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 <= 2);
34   assert(nc != 0);
35   assert(kc != 0);
36   assert(ks != 0);
37   assert(ks % (2 * 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   kc = round_up_po2(kc, 4 * sizeof(int8_t));
44   int8_t* c0 = c;
45   int8_t* c1 = (int8_t*) ((uintptr_t) c0 + cm_stride);
46   if XNN_UNPREDICTABLE(mr != 2) {
47     c1 = c0;
48   }
49 
50   do {
51     int32x4_t vacc0x01 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
52     int32x4_t vacc0x23 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
53     int32x4_t vacc0x45 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
54     int32x4_t vacc0x67 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
55     int32x4_t vacc1x01 = vacc0x01;
56     int32x4_t vacc1x23 = vacc0x23;
57     int32x4_t vacc1x45 = vacc0x45;
58     int32x4_t vacc1x67 = vacc0x67;
59 
60     size_t p = ks;
61     do {
62       const int8_t* restrict a0 = a[0];
63       if XNN_UNPREDICTABLE(a0 != zero) {
64         a0 = (const int8_t*) ((uintptr_t) a0 + a_offset);
65       }
66       const int8_t* restrict a1 = a[1];
67       if XNN_UNPREDICTABLE(a1 != zero) {
68         a1 = (const int8_t*) ((uintptr_t) a1 + a_offset);
69       }
70       a += 2;
71 
72       size_t k = kc;
73 
74 
75       while (k >= 8 * sizeof(int8_t)) {
76         const int32x2_t va00 = vld1_dup_s32((const void*)a0);
77         const int32x2_t va01 = vld1_dup_s32((const void*)(a0 + 4)); a0 += 8;
78         const int32x2_t va10 = vld1_dup_s32((const void*)a1);
79         const int32x2_t va11 = vld1_dup_s32((const void*)(a1 + 4)); a1 += 8;
80 
81         const int8x8_t vb01c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
82         const int8x8_t vb23c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
83         const int8x8_t vb45c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
84         const int8x8_t vb67c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
85         const int8x8_t vb01c1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
86         const int8x8_t vb23c1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
87         const int8x8_t vb45c1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
88         const int8x8_t vb67c1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
89 
90         const int8x8_t va0c0 = vreinterpret_s8_s32(va00);
91         const int8x8_t va1c0 = vreinterpret_s8_s32(va10);
92 
93         const int16x8_t vprod0x01c0 = vmull_s8(vb01c0, va0c0);
94         const int16x8_t vprod1x01c0 = vmull_s8(vb01c0, va1c0);
95         vacc0x01 = vpadalq_s16(vacc0x01, vprod0x01c0);
96         vacc1x01 = vpadalq_s16(vacc1x01, vprod1x01c0);
97         const int16x8_t vprod0x23c0 = vmull_s8(vb23c0, va0c0);
98         const int16x8_t vprod1x23c0 = vmull_s8(vb23c0, va1c0);
99         vacc0x23 = vpadalq_s16(vacc0x23, vprod0x23c0);
100         vacc1x23 = vpadalq_s16(vacc1x23, vprod1x23c0);
101         const int16x8_t vprod0x45c0 = vmull_s8(vb45c0, va0c0);
102         const int16x8_t vprod1x45c0 = vmull_s8(vb45c0, va1c0);
103         vacc0x45 = vpadalq_s16(vacc0x45, vprod0x45c0);
104         vacc1x45 = vpadalq_s16(vacc1x45, vprod1x45c0);
105         const int16x8_t vprod0x67c0 = vmull_s8(vb67c0, va0c0);
106         const int16x8_t vprod1x67c0 = vmull_s8(vb67c0, va1c0);
107         vacc0x67 = vpadalq_s16(vacc0x67, vprod0x67c0);
108         vacc1x67 = vpadalq_s16(vacc1x67, vprod1x67c0);
109         const int8x8_t va0c1 = vreinterpret_s8_s32(va01);
110         const int8x8_t va1c1 = vreinterpret_s8_s32(va11);
111 
112         const int16x8_t vprod0x01c1 = vmull_s8(vb01c1, va0c1);
113         const int16x8_t vprod1x01c1 = vmull_s8(vb01c1, va1c1);
114         vacc0x01 = vpadalq_s16(vacc0x01, vprod0x01c1);
115         vacc1x01 = vpadalq_s16(vacc1x01, vprod1x01c1);
116         const int16x8_t vprod0x23c1 = vmull_s8(vb23c1, va0c1);
117         const int16x8_t vprod1x23c1 = vmull_s8(vb23c1, va1c1);
118         vacc0x23 = vpadalq_s16(vacc0x23, vprod0x23c1);
119         vacc1x23 = vpadalq_s16(vacc1x23, vprod1x23c1);
120         const int16x8_t vprod0x45c1 = vmull_s8(vb45c1, va0c1);
121         const int16x8_t vprod1x45c1 = vmull_s8(vb45c1, va1c1);
122         vacc0x45 = vpadalq_s16(vacc0x45, vprod0x45c1);
123         vacc1x45 = vpadalq_s16(vacc1x45, vprod1x45c1);
124         const int16x8_t vprod0x67c1 = vmull_s8(vb67c1, va0c1);
125         const int16x8_t vprod1x67c1 = vmull_s8(vb67c1, va1c1);
126         vacc0x67 = vpadalq_s16(vacc0x67, vprod0x67c1);
127         vacc1x67 = vpadalq_s16(vacc1x67, vprod1x67c1);
128 
129         k -= 8 * sizeof(int8_t);
130       }
131 
132       if XNN_UNLIKELY(k != 0) {
133         const int8x8_t va0 = vld1_s8(a0); a0 = (const int8_t*) ((uintptr_t) a0 + k);
134         const int8x8_t va1 = vld1_s8(a1); a1 = (const int8_t*) ((uintptr_t) a1 + k);
135 
136         const int8x8_t vb01c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
137         const int8x8_t vb23c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
138         const int8x8_t vb45c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
139         const int8x8_t vb67c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
140 
141         const int8x8_t va0c0 = vreinterpret_s8_s32(vdup_lane_s32(vreinterpret_s32_s8(va0), 0));
142         const int16x8_t vprod0x01c0 = vmull_s8(vb01c0, va0c0);
143         vacc0x01 = vpadalq_s16(vacc0x01, vprod0x01c0);
144         const int16x8_t vprod0x23c0 = vmull_s8(vb23c0, va0c0);
145         vacc0x23 = vpadalq_s16(vacc0x23, vprod0x23c0);
146         const int16x8_t vprod0x45c0 = vmull_s8(vb45c0, va0c0);
147         vacc0x45 = vpadalq_s16(vacc0x45, vprod0x45c0);
148         const int16x8_t vprod0x67c0 = vmull_s8(vb67c0, va0c0);
149         vacc0x67 = vpadalq_s16(vacc0x67, vprod0x67c0);
150         const int8x8_t va1c0 = vreinterpret_s8_s32(vdup_lane_s32(vreinterpret_s32_s8(va1), 0));
151         const int16x8_t vprod1x01c0 = vmull_s8(vb01c0, va1c0);
152         vacc1x01 = vpadalq_s16(vacc1x01, vprod1x01c0);
153         const int16x8_t vprod1x23c0 = vmull_s8(vb23c0, va1c0);
154         vacc1x23 = vpadalq_s16(vacc1x23, vprod1x23c0);
155         const int16x8_t vprod1x45c0 = vmull_s8(vb45c0, va1c0);
156         vacc1x45 = vpadalq_s16(vacc1x45, vprod1x45c0);
157         const int16x8_t vprod1x67c0 = vmull_s8(vb67c0, va1c0);
158         vacc1x67 = vpadalq_s16(vacc1x67, vprod1x67c0);
159       }
160       p -= 2 * sizeof(void*);
161     } while (p != 0);
162 
163 #if XNN_ARCH_ARM64
164     int32x4_t vacc0x0123 = vpaddq_s32(vacc0x01, vacc0x23);
165     int32x4_t vacc0x4567 = vpaddq_s32(vacc0x45, vacc0x67);
166     int32x4_t vacc1x0123 = vpaddq_s32(vacc1x01, vacc1x23);
167     int32x4_t vacc1x4567 = vpaddq_s32(vacc1x45, vacc1x67);
168 #else
169     const int32x2_t vsum0x01 = vpadd_s32(vget_low_s32(vacc0x01), vget_high_s32(vacc0x01));
170     const int32x2_t vsum0x23 = vpadd_s32(vget_low_s32(vacc0x23), vget_high_s32(vacc0x23));
171     int32x4_t vacc0x0123 = vcombine_s32(vsum0x01, vsum0x23);
172     const int32x2_t vsum0x45 = vpadd_s32(vget_low_s32(vacc0x45), vget_high_s32(vacc0x45));
173     const int32x2_t vsum0x67 = vpadd_s32(vget_low_s32(vacc0x67), vget_high_s32(vacc0x67));
174     int32x4_t vacc0x4567 = vcombine_s32(vsum0x45, vsum0x67);
175     const int32x2_t vsum1x01 = vpadd_s32(vget_low_s32(vacc1x01), vget_high_s32(vacc1x01));
176     const int32x2_t vsum1x23 = vpadd_s32(vget_low_s32(vacc1x23), vget_high_s32(vacc1x23));
177     int32x4_t vacc1x0123 = vcombine_s32(vsum1x01, vsum1x23);
178     const int32x2_t vsum1x45 = vpadd_s32(vget_low_s32(vacc1x45), vget_high_s32(vacc1x45));
179     const int32x2_t vsum1x67 = vpadd_s32(vget_low_s32(vacc1x67), vget_high_s32(vacc1x67));
180     int32x4_t vacc1x4567 = vcombine_s32(vsum1x45, vsum1x67);
181 #endif
182 
183     const int32x4_t vright_pre_shift = vld1q_dup_s32(&params->rndnu_neon.right_pre_shift);
184     const int32x4_t vmultiplier = vld1q_dup_s32(&params->rndnu_neon.multiplier);
185     const int32x4_t vright_post_shift = vld1q_dup_s32(&params->rndnu_neon.right_post_shift);
186 
187     vacc0x0123 = vqshlq_s32(vacc0x0123, vright_pre_shift);
188     vacc0x4567 = vqshlq_s32(vacc0x4567, vright_pre_shift);
189     vacc1x0123 = vqshlq_s32(vacc1x0123, vright_pre_shift);
190     vacc1x4567 = vqshlq_s32(vacc1x4567, vright_pre_shift);
191 
192     vacc0x0123 = vqdmulhq_s32(vacc0x0123, vmultiplier);
193     vacc0x4567 = vqdmulhq_s32(vacc0x4567, vmultiplier);
194     vacc1x0123 = vqdmulhq_s32(vacc1x0123, vmultiplier);
195     vacc1x4567 = vqdmulhq_s32(vacc1x4567, vmultiplier);
196 
197     vacc0x0123 = vrshlq_s32(vacc0x0123, vright_post_shift);
198     vacc0x4567 = vrshlq_s32(vacc0x4567, vright_post_shift);
199     vacc1x0123 = vrshlq_s32(vacc1x0123, vright_post_shift);
200     vacc1x4567 = vrshlq_s32(vacc1x4567, vright_post_shift);
201 
202     const int16x8_t voutput_zero_point = vld1q_dup_s16(&params->rndnu_neon.output_zero_point);
203 #if XNN_ARCH_ARM64
204     int16x8_t vacc0x01234567 = vqmovn_high_s32(vqmovn_s32(vacc0x0123), vacc0x4567);
205     int16x8_t vacc1x01234567 = vqmovn_high_s32(vqmovn_s32(vacc1x0123), vacc1x4567);
206 
207     vacc0x01234567 = vqaddq_s16(vacc0x01234567, voutput_zero_point);
208     vacc1x01234567 = vqaddq_s16(vacc1x01234567, voutput_zero_point);
209 
210     int8x16_t vout0x01234567_1x01234567 = vqmovn_high_s16(vqmovn_s16(vacc0x01234567), vacc1x01234567);
211 #else
212     int16x8_t vacc0x01234567 = vcombine_s16(vqmovn_s32(vacc0x0123), vqmovn_s32(vacc0x4567));
213     int16x8_t vacc1x01234567 = vcombine_s16(vqmovn_s32(vacc1x0123), vqmovn_s32(vacc1x4567));
214 
215     vacc0x01234567 = vqaddq_s16(vacc0x01234567, voutput_zero_point);
216     vacc1x01234567 = vqaddq_s16(vacc1x01234567, voutput_zero_point);
217 
218     int8x16_t vout0x01234567_1x01234567 = vcombine_s8(vqmovn_s16(vacc0x01234567), vqmovn_s16(vacc1x01234567));
219 #endif
220 
221     const int8x16_t voutput_min = vld1q_dup_s8(&params->rndnu_neon.output_min);
222     vout0x01234567_1x01234567 = vmaxq_s8(vout0x01234567_1x01234567, voutput_min);
223 
224     const int8x16_t voutput_max = vld1q_dup_s8(&params->rndnu_neon.output_max);
225     vout0x01234567_1x01234567 = vminq_s8(vout0x01234567_1x01234567, voutput_max);
226 
227     if (nc >= 8) {
228       vst1_s8(c1 + 0, vget_high_s8(vout0x01234567_1x01234567));
229       vst1_s8(c0 + 0, vget_low_s8(vout0x01234567_1x01234567));
230 
231       c1 = (int8_t*) ((uintptr_t) c1 + cn_stride);
232       c0 = (int8_t*) ((uintptr_t) c0 + cn_stride);
233 
234       a = (const int8_t**restrict) ((uintptr_t) a - ks);
235 
236       nc -= 8;
237     } else {
238       if (nc & 4) {
239         vst1q_lane_u32((void*) c1, vreinterpretq_u32_s8(vout0x01234567_1x01234567), 2); c1 += 4;
240         vst1q_lane_u32((void*) c0, vreinterpretq_u32_s8(vout0x01234567_1x01234567), 0); c0 += 4;
241         vout0x01234567_1x01234567 = vextq_s8(vout0x01234567_1x01234567, vout0x01234567_1x01234567, 4);
242       }
243       if (nc & 2) {
244         vst1q_lane_u16((void*) c1, vreinterpretq_u16_s8(vout0x01234567_1x01234567), 4); c1 += 2;
245         vst1q_lane_u16((void*) c0, vreinterpretq_u16_s8(vout0x01234567_1x01234567), 0); c0 += 2;
246         vout0x01234567_1x01234567 = vextq_s8(vout0x01234567_1x01234567, vout0x01234567_1x01234567, 2);
247       }
248       if (nc & 1) {
249         vst1q_lane_s8(c1, vout0x01234567_1x01234567, 8);
250         vst1q_lane_s8(c0, vout0x01234567_1x01234567, 0);
251       }
252 
253       nc = 0;
254     }
255   } while (nc != 0);
256 }
257