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_1x8c2s4__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_1x8c2s4__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
50 size_t p = ks;
51 do {
52 const int8_t* restrict a0 = a[0];
53 if XNN_UNPREDICTABLE(a0 != zero) {
54 a0 = (const int8_t*) ((uintptr_t) a0 + a_offset);
55 }
56 a += 1;
57
58 size_t k = kc;
59 do {
60 int8x8_t va0x0 = vld1_s8(a0); a0 += 8;
61
62 const int8x8_t vb0123c0x0 = vld1_s8(w); w = (const int8_t*) w + 8;
63 const int8x8_t vb4567c0x0 = vld1_s8(w); w = (const int8_t*) w + 8;
64 const int8x8_t vb0123c1x0 = vld1_s8(w); w = (const int8_t*) w + 8;
65 const int8x8_t vb4567c1x0 = vld1_s8(w); w = (const int8_t*) w + 8;
66 const int8x8_t vb0123c2x0 = vld1_s8(w); w = (const int8_t*) w + 8;
67 const int8x8_t vb4567c2x0 = vld1_s8(w); w = (const int8_t*) w + 8;
68 const int8x8_t vb0123c3x0 = vld1_s8(w); w = (const int8_t*) w + 8;
69 const int8x8_t vb4567c3x0 = vld1_s8(w); w = (const int8_t*) w + 8;
70
71 int16x8_t vprod0x0123c0 = vmull_s8(vb0123c0x0, va0x0);
72 vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c0);
73 int16x8_t vprod0x4567c0 = vmull_s8(vb4567c0x0, va0x0);
74 vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c0);
75 va0x0 = vext_s8(va0x0, va0x0, 2);
76 int16x8_t vprod0x0123c1 = vmull_s8(vb0123c1x0, va0x0);
77 vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c1);
78 int16x8_t vprod0x4567c1 = vmull_s8(vb4567c1x0, va0x0);
79 vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c1);
80 va0x0 = vext_s8(va0x0, va0x0, 2);
81 int16x8_t vprod0x0123c2 = vmull_s8(vb0123c2x0, va0x0);
82 vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c2);
83 int16x8_t vprod0x4567c2 = vmull_s8(vb4567c2x0, va0x0);
84 vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c2);
85 va0x0 = vext_s8(va0x0, va0x0, 2);
86 int16x8_t vprod0x0123c3 = vmull_s8(vb0123c3x0, va0x0);
87 vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c3);
88 int16x8_t vprod0x4567c3 = vmull_s8(vb4567c3x0, va0x0);
89 vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c3);
90
91 k -= 8 * sizeof(int8_t);
92 } while (k != 0);
93
94 p -= 1 * sizeof(void*);
95 } while (p != 0);
96
97 const int32x4_t vright_pre_shift = vld1q_dup_s32(¶ms->rndnu_neon.right_pre_shift);
98 const int32x4_t vmultiplier = vld1q_dup_s32(¶ms->rndnu_neon.multiplier);
99 const int32x4_t vright_post_shift = vld1q_dup_s32(¶ms->rndnu_neon.right_post_shift);
100
101 vacc0x0123 = vqshlq_s32(vacc0x0123, vright_pre_shift);
102 vacc0x4567 = vqshlq_s32(vacc0x4567, vright_pre_shift);
103
104 vacc0x0123 = vqdmulhq_s32(vacc0x0123, vmultiplier);
105 vacc0x4567 = vqdmulhq_s32(vacc0x4567, vmultiplier);
106
107 vacc0x0123 = vrshlq_s32(vacc0x0123, vright_post_shift);
108 vacc0x4567 = vrshlq_s32(vacc0x4567, vright_post_shift);
109
110 const int16x8_t voutput_zero_point = vld1q_dup_s16(¶ms->rndnu_neon.output_zero_point);
111 #if XNN_ARCH_ARM64
112 int16x8_t vacc0x01234567 = vqmovn_high_s32(vqmovn_s32(vacc0x0123), vacc0x4567);
113
114 vacc0x01234567 = vqaddq_s16(vacc0x01234567, voutput_zero_point);
115
116 int8x8_t vout0x01234567 = vqmovn_s16(vacc0x01234567);
117 #else
118 int16x8_t vacc0x01234567 = vcombine_s16(vqmovn_s32(vacc0x0123), vqmovn_s32(vacc0x4567));
119
120 vacc0x01234567 = vqaddq_s16(vacc0x01234567, voutput_zero_point);
121
122 int8x8_t vout0x01234567 = vqmovn_s16(vacc0x01234567);
123 #endif
124
125 const int8x8_t voutput_min = vld1_dup_s8(¶ms->rndnu_neon.output_min);
126 vout0x01234567 = vmax_s8(vout0x01234567, voutput_min);
127
128 const int8x8_t voutput_max = vld1_dup_s8(¶ms->rndnu_neon.output_max);
129 vout0x01234567 = vmin_s8(vout0x01234567, voutput_max);
130
131 if (nc >= 8) {
132 vst1_s8(c0 + 0, vout0x01234567);
133
134 c0 = (int8_t*) ((uintptr_t) c0 + cn_stride);
135
136 a = (const int8_t**restrict) ((uintptr_t) a - ks);
137
138 nc -= 8;
139 } else {
140 if (nc & 4) {
141 vst1_lane_u32((void*) c0, vreinterpret_u32_s8(vout0x01234567), 0); c0 += 4;
142 vout0x01234567 = vext_s8(vout0x01234567, vout0x01234567, 4);
143 }
144 if (nc & 2) {
145 vst1_lane_u16((void*) c0, vreinterpret_u16_s8(vout0x01234567), 0); c0 += 2;
146 vout0x01234567 = vext_s8(vout0x01234567, vout0x01234567, 2);
147 }
148 if (nc & 1) {
149 vst1_lane_s8(c0, vout0x01234567, 0);
150 }
151
152 nc = 0;
153 }
154 } while (nc != 0);
155 }
156