xref: /aosp_15_r20/external/XNNPACK/src/qs8-igemm/gen/1x8c4-minmax-rndnu-neon-mull-dup.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_1x8c4__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)])18 void xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c4__neon_mull_dup(
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   kc = round_up_po2(kc, 4 * sizeof(int8_t));
44   int8_t* c0 = c;
45 
46   do {
47     int32x4_t vacc0x01 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
48     int32x4_t vacc0x23 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
49     int32x4_t vacc0x45 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
50     int32x4_t vacc0x67 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
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 
62 
63       while (k >= 8 * sizeof(int8_t)) {
64         const int8x8_t va0 = vld1_s8(a0); a0 += 8;
65 
66         const int8x8_t vb01c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
67         const int8x8_t vb23c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
68         const int8x8_t vb45c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
69         const int8x8_t vb67c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
70         const int8x8_t vb01c1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
71         const int8x8_t vb23c1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
72         const int8x8_t vb45c1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
73         const int8x8_t vb67c1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
74 
75         const int8x8_t va0c0 = vreinterpret_s8_s32(vdup_lane_s32(vreinterpret_s32_s8(va0), 0));
76 
77         const int16x8_t vprod0x01c0 = vmull_s8(vb01c0, va0c0);
78         vacc0x01 = vpadalq_s16(vacc0x01, vprod0x01c0);
79         const int16x8_t vprod0x23c0 = vmull_s8(vb23c0, va0c0);
80         vacc0x23 = vpadalq_s16(vacc0x23, vprod0x23c0);
81         const int16x8_t vprod0x45c0 = vmull_s8(vb45c0, va0c0);
82         vacc0x45 = vpadalq_s16(vacc0x45, vprod0x45c0);
83         const int16x8_t vprod0x67c0 = vmull_s8(vb67c0, va0c0);
84         vacc0x67 = vpadalq_s16(vacc0x67, vprod0x67c0);
85         const int8x8_t va0c1 = vreinterpret_s8_s32(vdup_lane_s32(vreinterpret_s32_s8(va0), 1));
86 
87         const int16x8_t vprod0x01c1 = vmull_s8(vb01c1, va0c1);
88         vacc0x01 = vpadalq_s16(vacc0x01, vprod0x01c1);
89         const int16x8_t vprod0x23c1 = vmull_s8(vb23c1, va0c1);
90         vacc0x23 = vpadalq_s16(vacc0x23, vprod0x23c1);
91         const int16x8_t vprod0x45c1 = vmull_s8(vb45c1, va0c1);
92         vacc0x45 = vpadalq_s16(vacc0x45, vprod0x45c1);
93         const int16x8_t vprod0x67c1 = vmull_s8(vb67c1, va0c1);
94         vacc0x67 = vpadalq_s16(vacc0x67, vprod0x67c1);
95 
96         k -= 8 * sizeof(int8_t);
97       }
98 
99       if XNN_UNLIKELY(k != 0) {
100         const int8x8_t va0 = vld1_s8(a0); a0 = (const int8_t*) ((uintptr_t) a0 + k);
101 
102         const int8x8_t vb01c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
103         const int8x8_t vb23c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
104         const int8x8_t vb45c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
105         const int8x8_t vb67c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
106 
107         const int8x8_t va0c0 = vreinterpret_s8_s32(vdup_lane_s32(vreinterpret_s32_s8(va0), 0));
108         const int16x8_t vprod0x01c0 = vmull_s8(vb01c0, va0c0);
109         vacc0x01 = vpadalq_s16(vacc0x01, vprod0x01c0);
110         const int16x8_t vprod0x23c0 = vmull_s8(vb23c0, va0c0);
111         vacc0x23 = vpadalq_s16(vacc0x23, vprod0x23c0);
112         const int16x8_t vprod0x45c0 = vmull_s8(vb45c0, va0c0);
113         vacc0x45 = vpadalq_s16(vacc0x45, vprod0x45c0);
114         const int16x8_t vprod0x67c0 = vmull_s8(vb67c0, va0c0);
115         vacc0x67 = vpadalq_s16(vacc0x67, vprod0x67c0);
116       }
117       p -= 1 * sizeof(void*);
118     } while (p != 0);
119 
120 #if XNN_ARCH_ARM64
121     int32x4_t vacc0x0123 = vpaddq_s32(vacc0x01, vacc0x23);
122     int32x4_t vacc0x4567 = vpaddq_s32(vacc0x45, vacc0x67);
123 #else
124     const int32x2_t vsum0x01 = vpadd_s32(vget_low_s32(vacc0x01), vget_high_s32(vacc0x01));
125     const int32x2_t vsum0x23 = vpadd_s32(vget_low_s32(vacc0x23), vget_high_s32(vacc0x23));
126     int32x4_t vacc0x0123 = vcombine_s32(vsum0x01, vsum0x23);
127     const int32x2_t vsum0x45 = vpadd_s32(vget_low_s32(vacc0x45), vget_high_s32(vacc0x45));
128     const int32x2_t vsum0x67 = vpadd_s32(vget_low_s32(vacc0x67), vget_high_s32(vacc0x67));
129     int32x4_t vacc0x4567 = vcombine_s32(vsum0x45, vsum0x67);
130 #endif
131 
132     const int32x4_t vright_pre_shift = vld1q_dup_s32(&params->rndnu_neon.right_pre_shift);
133     const int32x4_t vmultiplier = vld1q_dup_s32(&params->rndnu_neon.multiplier);
134     const int32x4_t vright_post_shift = vld1q_dup_s32(&params->rndnu_neon.right_post_shift);
135 
136     vacc0x0123 = vqshlq_s32(vacc0x0123, vright_pre_shift);
137     vacc0x4567 = vqshlq_s32(vacc0x4567, vright_pre_shift);
138 
139     vacc0x0123 = vqdmulhq_s32(vacc0x0123, vmultiplier);
140     vacc0x4567 = vqdmulhq_s32(vacc0x4567, vmultiplier);
141 
142     vacc0x0123 = vrshlq_s32(vacc0x0123, vright_post_shift);
143     vacc0x4567 = vrshlq_s32(vacc0x4567, vright_post_shift);
144 
145     const int16x8_t voutput_zero_point = vld1q_dup_s16(&params->rndnu_neon.output_zero_point);
146 #if XNN_ARCH_ARM64
147     int16x8_t vacc0x01234567 = vqmovn_high_s32(vqmovn_s32(vacc0x0123), vacc0x4567);
148 
149     vacc0x01234567 = vqaddq_s16(vacc0x01234567, voutput_zero_point);
150 
151     int8x8_t vout0x01234567 = vqmovn_s16(vacc0x01234567);
152 #else
153     int16x8_t vacc0x01234567 = vcombine_s16(vqmovn_s32(vacc0x0123), vqmovn_s32(vacc0x4567));
154 
155     vacc0x01234567 = vqaddq_s16(vacc0x01234567, voutput_zero_point);
156 
157     int8x8_t vout0x01234567 = vqmovn_s16(vacc0x01234567);
158 #endif
159 
160     const int8x8_t voutput_min = vld1_dup_s8(&params->rndnu_neon.output_min);
161     vout0x01234567 = vmax_s8(vout0x01234567, voutput_min);
162 
163     const int8x8_t voutput_max = vld1_dup_s8(&params->rndnu_neon.output_max);
164     vout0x01234567 = vmin_s8(vout0x01234567, voutput_max);
165 
166     if (nc >= 8) {
167       vst1_s8(c0 + 0, vout0x01234567);
168 
169       c0 = (int8_t*) ((uintptr_t) c0 + cn_stride);
170 
171       a = (const int8_t**restrict) ((uintptr_t) a - ks);
172 
173       nc -= 8;
174     } else {
175       if (nc & 4) {
176         vst1_lane_u32((void*) c0, vreinterpret_u32_s8(vout0x01234567), 0); c0 += 4;
177         vout0x01234567 = vext_s8(vout0x01234567, vout0x01234567, 4);
178       }
179       if (nc & 2) {
180         vst1_lane_u16((void*) c0, vreinterpret_u16_s8(vout0x01234567), 0); c0 += 2;
181         vout0x01234567 = vext_s8(vout0x01234567, vout0x01234567, 2);
182       }
183       if (nc & 1) {
184         vst1_lane_s8(c0, vout0x01234567, 0);
185       }
186 
187       nc = 0;
188     }
189   } while (nc != 0);
190 }
191