xref: /aosp_15_r20/external/XNNPACK/src/qs8-igemm/gen/1x16c2s4-minmax-rndnu-neon-mull.c (revision 4bdc94577ba0e567308109d787f7fec7b531ce36)
1 // Auto-generated file. Do not edit!
2 //   Template: src/qs8-igemm/c2-neon-mull-shuffle.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_1x16c2s4__neon_mull(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_1x16c2s4__neon_mull(
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   kc = round_up_po2(kc, 8 * sizeof(int8_t));
46   do {
47     int32x4_t vacc0x0123 = vld1q_s32(w); w = (const int32_t*) w + 4;
48     int32x4_t vacc0x4567 = vld1q_s32(w); w = (const int32_t*) w + 4;
49     int32x4_t vacc0x89AB = vld1q_s32(w); w = (const int32_t*) w + 4;
50     int32x4_t vacc0xCDEF = vld1q_s32(w); w = (const int32_t*) w + 4;
51 
52     size_t p = ks;
53     do {
54       const int8_t* restrict a0 = a[0];
55       if XNN_UNPREDICTABLE(a0 != zero) {
56         a0 = (const int8_t*) ((uintptr_t) a0 + a_offset);
57       }
58       a += 1;
59 
60       size_t k = kc;
61       do {
62         int8x8_t va0x0 = vld1_s8(a0); a0 += 8;
63 
64         const int8x8_t vb0123c0x0 = vld1_s8(w); w = (const int8_t*) w + 8;
65         const int8x8_t vb4567c0x0 = vld1_s8(w); w = (const int8_t*) w + 8;
66         const int8x8_t vb89ABc0x0 = vld1_s8(w); w = (const int8_t*) w + 8;
67         const int8x8_t vbCDEFc0x0 = vld1_s8(w); w = (const int8_t*) w + 8;
68         const int8x8_t vb0123c1x0 = vld1_s8(w); w = (const int8_t*) w + 8;
69         const int8x8_t vb4567c1x0 = vld1_s8(w); w = (const int8_t*) w + 8;
70         const int8x8_t vb89ABc1x0 = vld1_s8(w); w = (const int8_t*) w + 8;
71         const int8x8_t vbCDEFc1x0 = vld1_s8(w); w = (const int8_t*) w + 8;
72         const int8x8_t vb0123c2x0 = vld1_s8(w); w = (const int8_t*) w + 8;
73         const int8x8_t vb4567c2x0 = vld1_s8(w); w = (const int8_t*) w + 8;
74         const int8x8_t vb89ABc2x0 = vld1_s8(w); w = (const int8_t*) w + 8;
75         const int8x8_t vbCDEFc2x0 = vld1_s8(w); w = (const int8_t*) w + 8;
76         const int8x8_t vb0123c3x0 = vld1_s8(w); w = (const int8_t*) w + 8;
77         const int8x8_t vb4567c3x0 = vld1_s8(w); w = (const int8_t*) w + 8;
78         const int8x8_t vb89ABc3x0 = vld1_s8(w); w = (const int8_t*) w + 8;
79         const int8x8_t vbCDEFc3x0 = vld1_s8(w); w = (const int8_t*) w + 8;
80 
81         int16x8_t vprod0x0123c0 = vmull_s8(vb0123c0x0, va0x0);
82         vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c0);
83         int16x8_t vprod0x4567c0 = vmull_s8(vb4567c0x0, va0x0);
84         vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c0);
85         int16x8_t vprod0x89ABc0 = vmull_s8(vb89ABc0x0, va0x0);
86         vacc0x89AB = vpadalq_s16(vacc0x89AB, vprod0x89ABc0);
87         int16x8_t vprod0xCDEFc0 = vmull_s8(vbCDEFc0x0, va0x0);
88         vacc0xCDEF = vpadalq_s16(vacc0xCDEF, vprod0xCDEFc0);
89         va0x0 = vext_s8(va0x0, va0x0, 2);
90         int16x8_t vprod0x0123c1 = vmull_s8(vb0123c1x0, va0x0);
91         vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c1);
92         int16x8_t vprod0x4567c1 = vmull_s8(vb4567c1x0, va0x0);
93         vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c1);
94         int16x8_t vprod0x89ABc1 = vmull_s8(vb89ABc1x0, va0x0);
95         vacc0x89AB = vpadalq_s16(vacc0x89AB, vprod0x89ABc1);
96         int16x8_t vprod0xCDEFc1 = vmull_s8(vbCDEFc1x0, va0x0);
97         vacc0xCDEF = vpadalq_s16(vacc0xCDEF, vprod0xCDEFc1);
98         va0x0 = vext_s8(va0x0, va0x0, 2);
99         int16x8_t vprod0x0123c2 = vmull_s8(vb0123c2x0, va0x0);
100         vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c2);
101         int16x8_t vprod0x4567c2 = vmull_s8(vb4567c2x0, va0x0);
102         vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c2);
103         int16x8_t vprod0x89ABc2 = vmull_s8(vb89ABc2x0, va0x0);
104         vacc0x89AB = vpadalq_s16(vacc0x89AB, vprod0x89ABc2);
105         int16x8_t vprod0xCDEFc2 = vmull_s8(vbCDEFc2x0, va0x0);
106         vacc0xCDEF = vpadalq_s16(vacc0xCDEF, vprod0xCDEFc2);
107         va0x0 = vext_s8(va0x0, va0x0, 2);
108         int16x8_t vprod0x0123c3 = vmull_s8(vb0123c3x0, va0x0);
109         vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c3);
110         int16x8_t vprod0x4567c3 = vmull_s8(vb4567c3x0, va0x0);
111         vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c3);
112         int16x8_t vprod0x89ABc3 = vmull_s8(vb89ABc3x0, va0x0);
113         vacc0x89AB = vpadalq_s16(vacc0x89AB, vprod0x89ABc3);
114         int16x8_t vprod0xCDEFc3 = vmull_s8(vbCDEFc3x0, va0x0);
115         vacc0xCDEF = vpadalq_s16(vacc0xCDEF, vprod0xCDEFc3);
116 
117         k -= 8 * sizeof(int8_t);
118       } while (k != 0);
119 
120       p -= 1 * sizeof(void*);
121     } while (p != 0);
122 
123     const int32x4_t vright_pre_shift = vld1q_dup_s32(&params->rndnu_neon.right_pre_shift);
124     const int32x4_t vmultiplier = vld1q_dup_s32(&params->rndnu_neon.multiplier);
125     const int32x4_t vright_post_shift = vld1q_dup_s32(&params->rndnu_neon.right_post_shift);
126 
127     vacc0x0123 = vqshlq_s32(vacc0x0123, vright_pre_shift);
128     vacc0x4567 = vqshlq_s32(vacc0x4567, vright_pre_shift);
129     vacc0x89AB = vqshlq_s32(vacc0x89AB, vright_pre_shift);
130     vacc0xCDEF = vqshlq_s32(vacc0xCDEF, vright_pre_shift);
131 
132     vacc0x0123 = vqdmulhq_s32(vacc0x0123, vmultiplier);
133     vacc0x4567 = vqdmulhq_s32(vacc0x4567, vmultiplier);
134     vacc0x89AB = vqdmulhq_s32(vacc0x89AB, vmultiplier);
135     vacc0xCDEF = vqdmulhq_s32(vacc0xCDEF, vmultiplier);
136 
137     vacc0x0123 = vrshlq_s32(vacc0x0123, vright_post_shift);
138     vacc0x4567 = vrshlq_s32(vacc0x4567, vright_post_shift);
139     vacc0x89AB = vrshlq_s32(vacc0x89AB, vright_post_shift);
140     vacc0xCDEF = vrshlq_s32(vacc0xCDEF, vright_post_shift);
141 
142     const int16x8_t voutput_zero_point = vld1q_dup_s16(&params->rndnu_neon.output_zero_point);
143 #if XNN_ARCH_ARM64
144     int16x8_t vacc0x01234567 = vqmovn_high_s32(vqmovn_s32(vacc0x0123), vacc0x4567);
145     int16x8_t vacc0x89ABCDEF = vqmovn_high_s32(vqmovn_s32(vacc0x89AB), vacc0xCDEF);
146 
147     vacc0x01234567 = vqaddq_s16(vacc0x01234567, voutput_zero_point);
148     vacc0x89ABCDEF = vqaddq_s16(vacc0x89ABCDEF, voutput_zero_point);
149 
150     int8x16_t vout0x0123456789ABCDEF = vqmovn_high_s16(vqmovn_s16(vacc0x01234567), vacc0x89ABCDEF);
151 #else
152     int16x8_t vacc0x01234567 = vcombine_s16(vqmovn_s32(vacc0x0123), vqmovn_s32(vacc0x4567));
153     int16x8_t vacc0x89ABCDEF = vcombine_s16(vqmovn_s32(vacc0x89AB), vqmovn_s32(vacc0xCDEF));
154 
155     vacc0x01234567 = vqaddq_s16(vacc0x01234567, voutput_zero_point);
156     vacc0x89ABCDEF = vqaddq_s16(vacc0x89ABCDEF, voutput_zero_point);
157 
158     int8x16_t vout0x0123456789ABCDEF = vcombine_s8(vqmovn_s16(vacc0x01234567), vqmovn_s16(vacc0x89ABCDEF));
159 #endif
160 
161     const int8x16_t voutput_min = vld1q_dup_s8(&params->rndnu_neon.output_min);
162     vout0x0123456789ABCDEF = vmaxq_s8(vout0x0123456789ABCDEF, voutput_min);
163 
164     const int8x16_t voutput_max = vld1q_dup_s8(&params->rndnu_neon.output_max);
165     vout0x0123456789ABCDEF = vminq_s8(vout0x0123456789ABCDEF, voutput_max);
166 
167     if (nc >= 16) {
168       vst1q_s8(c0 + 0, vout0x0123456789ABCDEF);
169 
170       c0 = (int8_t*) ((uintptr_t) c0 + cn_stride);
171 
172       a = (const int8_t**restrict) ((uintptr_t) a - ks);
173 
174       nc -= 16;
175     } else {
176       int8x8_t vout0x01234567 = vget_low_s8(vout0x0123456789ABCDEF);
177       if (nc & 8) {
178         vst1_s8(c0, vout0x01234567); c0 += 8;
179         vout0x01234567 = vget_high_s8(vout0x0123456789ABCDEF);
180       }
181       if (nc & 4) {
182         vst1_lane_u32((void*) c0, vreinterpret_u32_s8(vout0x01234567), 0); c0 += 4;
183         vout0x01234567 = vext_s8(vout0x01234567, vout0x01234567, 4);
184       }
185       if (nc & 2) {
186         vst1_lane_u16((void*) c0, vreinterpret_u16_s8(vout0x01234567), 0); c0 += 2;
187         vout0x01234567 = vext_s8(vout0x01234567, vout0x01234567, 2);
188       }
189       if (nc & 1) {
190         vst1_lane_s8(c0, vout0x01234567, 0);
191       }
192 
193       nc = 0;
194     }
195   } while (nc != 0);
196 }
197