1 // Auto-generated file. Do not edit!
2 // Template: src/qs8-igemm/c4-neondot.c.in
3 // Generator: tools/xngen
4 //
5 // Copyright 2020 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/igemm.h>
15 #include <xnnpack/math.h>
16
17
xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c4__neondot(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__neondot(
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 vacc0x0123 = vld1q_s32(w); w = (const void*) ((const int32_t*) w + 4);
48 int32x4_t vacc0x4567 = vld1q_s32(w); w = (const void*) ((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 // Inner accumulation loop along the 8 columns.
59 size_t k = kc;
60 // 2x partial unrolled loop to load 8 bytes at a time.
61 while (k >= 8 * sizeof(int8_t)) {
62 // Load a 1x8 block of activations.
63 const int8x8_t va0x01234567 = vld1_s8(a0); a0 += 8;
64
65 // Load a 8x8 block of weights.
66 const int8x16_t vb0123x0123 = vld1q_s8(w); w = (const void*) ((const int8_t*) w + 16);
67 const int8x16_t vb0123x4567 = vld1q_s8(w); w = (const void*) ((const int8_t*) w + 16);
68 const int8x16_t vb4567x0123 = vld1q_s8(w); w = (const void*) ((const int8_t*) w + 16);
69 const int8x16_t vb4567x4567 = vld1q_s8(w); w = (const void*) ((const int8_t*) w + 16);
70
71 // Multiply-accumulate: 1x8 * 8x8 --> 1x8.
72 vacc0x0123 = vdotq_lane_s32(vacc0x0123, vb0123x0123, va0x01234567, 0);
73 vacc0x4567 = vdotq_lane_s32(vacc0x4567, vb0123x4567, va0x01234567, 0);
74 vacc0x0123 = vdotq_lane_s32(vacc0x0123, vb4567x0123, va0x01234567, 1);
75 vacc0x4567 = vdotq_lane_s32(vacc0x4567, vb4567x4567, va0x01234567, 1);
76
77 k -= 8 * sizeof(int8_t);
78 }
79 // Handle up to 4 final positions of `k`
80 if XNN_UNLIKELY(k != 0) {
81 // Load a 1x4 block of activations.
82 const int8x8_t va0x01234567 = vld1_s8(a0);
83
84 // Load a 4x8 block of weights.
85 const int8x16_t vb0123x0123 = vld1q_s8(w); w = (const void*) ((const int8_t*) w + 16);
86 const int8x16_t vb0123x4567 = vld1q_s8(w); w = (const void*) ((const int8_t*) w + 16);
87
88 // Multiply-accumulate: 1x4 * 4x8 --> 1x8.
89 vacc0x0123 = vdotq_lane_s32(vacc0x0123, vb0123x0123, va0x01234567, 0);
90 vacc0x4567 = vdotq_lane_s32(vacc0x4567, vb0123x4567, va0x01234567, 0);
91 }
92 p -= 1 * sizeof(void*);
93 } while (p != 0);
94
95 const int32x4_t vright_pre_shift = vld1q_dup_s32(¶ms->rndnu_neon.right_pre_shift);
96 const int32x4_t vmultiplier = vld1q_dup_s32(¶ms->rndnu_neon.multiplier);
97 const int32x4_t vright_post_shift = vld1q_dup_s32(¶ms->rndnu_neon.right_post_shift);
98
99 vacc0x0123 = vshlq_s32(vacc0x0123, vright_pre_shift);
100 vacc0x4567 = vshlq_s32(vacc0x4567, vright_pre_shift);
101
102 vacc0x0123 = vqdmulhq_s32(vacc0x0123, vmultiplier);
103 vacc0x4567 = vqdmulhq_s32(vacc0x4567, vmultiplier);
104
105 vacc0x0123 = vrshlq_s32(vacc0x0123, vright_post_shift);
106 vacc0x4567 = vrshlq_s32(vacc0x4567, vright_post_shift);
107
108 const int16x8_t voutput_zero_point = vld1q_dup_s16(¶ms->rndnu_neon.output_zero_point);
109 #if XNN_ARCH_ARM64
110 const int16x8_t vacc0x01234567 = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc0x0123), vacc0x4567), voutput_zero_point);
111
112 int8x8_t vout0x01234567 = vqmovn_s16(vacc0x01234567);
113 #else
114 const int16x8_t vacc0x01234567 = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc0x0123), vqmovn_s32(vacc0x4567)), voutput_zero_point);
115
116 int8x8_t vout0x01234567 = vqmovn_s16(vacc0x01234567);
117 #endif
118 const int8x8_t voutput_min = vld1_dup_s8(¶ms->rndnu_neon.output_min);
119 const int8x8_t voutput_max = vld1_dup_s8(¶ms->rndnu_neon.output_max);
120
121 vout0x01234567 = vmax_s8(vout0x01234567, voutput_min);
122
123 vout0x01234567 = vmin_s8(vout0x01234567, voutput_max);
124
125 if (nc >= 8) {
126 vst1_s8(c0 + 0, vout0x01234567);
127
128 c0 = (int8_t*) ((uintptr_t) c0 + cn_stride);
129
130 a = (const int8_t**restrict) ((uintptr_t) a - ks);
131
132 nc -= 8;
133 } else {
134 if (nc & 4) {
135 vst1_lane_u32((void*) c0, vreinterpret_u32_s8(vout0x01234567), 0); c0 += 4;
136 vout0x01234567 = vext_s8(vout0x01234567, vout0x01234567, 4);
137 }
138 if (nc & 2) {
139 vst1_lane_u16((void*) c0, vreinterpret_u16_s8(vout0x01234567), 0); c0 += 2;
140 vout0x01234567 = vext_s8(vout0x01234567, vout0x01234567, 2);
141 }
142 if (nc & 1) {
143 vst1_lane_s8(c0, vout0x01234567, 0);
144 }
145
146 nc = 0;
147 }
148 } while (nc != 0);
149 }
150