xref: /aosp_15_r20/external/XNNPACK/src/qs8-igemm/gen/1x8c2-minmax-rndnu-neon-mull-dup.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_1x8c2__neon_mull_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)])17 void xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c2__neon_mull_dup(
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 
49     size_t p = ks;
50     do {
51       const int8_t* restrict a0 = a[0];
52       if XNN_UNPREDICTABLE(a0 != zero) {
53         a0 = (const int8_t*) ((uintptr_t) a0 + a_offset);
54       }
55       a += 1;
56 
57       size_t k = kc;
58 
59 
60       while (k >= 8 * sizeof(int8_t)) {
61         const int8x8_t va0 = vld1_s8(a0); a0 += 8;
62 
63         const int8x8_t vb0123c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
64         const int8x8_t vb4567c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
65         const int8x8_t vb0123c1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
66         const int8x8_t vb4567c1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
67         const int8x8_t vb0123c2 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
68         const int8x8_t vb4567c2 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
69         const int8x8_t vb0123c3 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
70         const int8x8_t vb4567c3 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
71 
72         const int8x8_t va0c0 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0), 0));
73 
74         const int16x8_t vprod0x0123c0 = vmull_s8(vb0123c0, va0c0);
75         vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c0);
76         const int16x8_t vprod0x4567c0 = vmull_s8(vb4567c0, va0c0);
77         vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c0);
78         const int8x8_t va0c1 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0), 1));
79 
80         const int16x8_t vprod0x0123c1 = vmull_s8(vb0123c1, va0c1);
81         vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c1);
82         const int16x8_t vprod0x4567c1 = vmull_s8(vb4567c1, va0c1);
83         vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c1);
84         const int8x8_t va0c2 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0), 2));
85 
86         const int16x8_t vprod0x0123c2 = vmull_s8(vb0123c2, va0c2);
87         vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c2);
88         const int16x8_t vprod0x4567c2 = vmull_s8(vb4567c2, va0c2);
89         vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c2);
90         const int8x8_t va0c3 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0), 3));
91 
92         const int16x8_t vprod0x0123c3 = vmull_s8(vb0123c3, va0c3);
93         vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c3);
94         const int16x8_t vprod0x4567c3 = vmull_s8(vb4567c3, va0c3);
95         vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c3);
96 
97         k -= 8 * sizeof(int8_t);
98       }
99 
100       if XNN_UNLIKELY(k != 0) {
101         const int8x8_t va0 = vld1_s8(a0); a0 = (const int8_t*) ((uintptr_t) a0 + k);
102 
103         const int8x8_t vb0123c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
104         const int8x8_t vb4567c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
105 
106         const int8x8_t va0c0 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0), 0));
107         const int16x8_t vprod0x0123c0 = vmull_s8(vb0123c0, va0c0);
108         vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c0);
109         const int16x8_t vprod0x4567c0 = vmull_s8(vb4567c0, va0c0);
110         vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c0);
111 
112         if (k > 2 * sizeof(int8_t)) {
113           const int8x8_t vb0123c1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
114           const int8x8_t vb4567c1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
115 
116           const int8x8_t va0c1 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0), 1));
117           const int16x8_t vprod0x0123c1 = vmull_s8(vb0123c1, va0c1);
118           vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c1);
119           const int16x8_t vprod0x4567c1 = vmull_s8(vb4567c1, va0c1);
120           vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c1);
121 
122           if (k > 4 * sizeof(int8_t)) {
123             const int8x8_t vb0123c2 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
124             const int8x8_t vb4567c2 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
125 
126             const int8x8_t va0c2 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0), 2));
127             const int16x8_t vprod0x0123c2 = vmull_s8(vb0123c2, va0c2);
128             vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c2);
129             const int16x8_t vprod0x4567c2 = vmull_s8(vb4567c2, va0c2);
130             vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c2);
131           }
132         }
133       }
134       p -= 1 * sizeof(void*);
135     } while (p != 0);
136 
137     const int32x4_t vright_pre_shift = vld1q_dup_s32(&params->rndnu_neon.right_pre_shift);
138     const int32x4_t vmultiplier = vld1q_dup_s32(&params->rndnu_neon.multiplier);
139     const int32x4_t vright_post_shift = vld1q_dup_s32(&params->rndnu_neon.right_post_shift);
140 
141     vacc0x0123 = vqshlq_s32(vacc0x0123, vright_pre_shift);
142     vacc0x4567 = vqshlq_s32(vacc0x4567, vright_pre_shift);
143 
144     vacc0x0123 = vqdmulhq_s32(vacc0x0123, vmultiplier);
145     vacc0x4567 = vqdmulhq_s32(vacc0x4567, vmultiplier);
146 
147     vacc0x0123 = vrshlq_s32(vacc0x0123, vright_post_shift);
148     vacc0x4567 = vrshlq_s32(vacc0x4567, vright_post_shift);
149 
150     const int16x8_t voutput_zero_point = vld1q_dup_s16(&params->rndnu_neon.output_zero_point);
151 #if XNN_ARCH_ARM64
152     int16x8_t vacc0x01234567 = vqmovn_high_s32(vqmovn_s32(vacc0x0123), vacc0x4567);
153 
154     vacc0x01234567 = vqaddq_s16(vacc0x01234567, voutput_zero_point);
155 
156     int8x8_t vout0x01234567 = vqmovn_s16(vacc0x01234567);
157 #else
158     int16x8_t vacc0x01234567 = vcombine_s16(vqmovn_s32(vacc0x0123), vqmovn_s32(vacc0x4567));
159 
160     vacc0x01234567 = vqaddq_s16(vacc0x01234567, voutput_zero_point);
161 
162     int8x8_t vout0x01234567 = vqmovn_s16(vacc0x01234567);
163 #endif
164 
165     const int8x8_t voutput_min = vld1_dup_s8(&params->rndnu_neon.output_min);
166     vout0x01234567 = vmax_s8(vout0x01234567, voutput_min);
167 
168     const int8x8_t voutput_max = vld1_dup_s8(&params->rndnu_neon.output_max);
169     vout0x01234567 = vmin_s8(vout0x01234567, voutput_max);
170 
171     if (nc >= 8) {
172       vst1_s8(c0 + 0, vout0x01234567);
173 
174       c0 = (int8_t*) ((uintptr_t) c0 + cn_stride);
175 
176       a = (const int8_t**restrict) ((uintptr_t) a - ks);
177 
178       nc -= 8;
179     } else {
180       if (nc & 4) {
181         vst1_lane_u32((void*) c0, vreinterpret_u32_s8(vout0x01234567), 0); c0 += 4;
182         vout0x01234567 = vext_s8(vout0x01234567, vout0x01234567, 4);
183       }
184       if (nc & 2) {
185         vst1_lane_u16((void*) c0, vreinterpret_u16_s8(vout0x01234567), 0); c0 += 2;
186         vout0x01234567 = vext_s8(vout0x01234567, vout0x01234567, 2);
187       }
188       if (nc & 1) {
189         vst1_lane_s8(c0, vout0x01234567, 0);
190       }
191 
192       nc = 0;
193     }
194   } while (nc != 0);
195 }
196