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_1x16c4__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_1x16c4__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 int32x4_t vacc0x89AB = vld1q_s32(w); w = (const void*) ((const int32_t*) w + 4);
50 int32x4_t vacc0xCDEF = vld1q_s32(w); w = (const void*) ((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 // Inner accumulation loop along the 16 columns.
61 size_t k = kc;
62 // 2x partial unrolled loop to load 8 bytes at a time.
63 while (k >= 8 * sizeof(int8_t)) {
64 // Load a 1x8 block of activations.
65 const int8x8_t va0x01234567 = vld1_s8(a0); a0 += 8;
66
67 // Load a 8x16 block of weights.
68 const int8x16_t vb0123x0123 = vld1q_s8(w); w = (const void*) ((const int8_t*) w + 16);
69 const int8x16_t vb0123x4567 = vld1q_s8(w); w = (const void*) ((const int8_t*) w + 16);
70 const int8x16_t vb0123x89AB = vld1q_s8(w); w = (const void*) ((const int8_t*) w + 16);
71 const int8x16_t vb0123xCDEF = vld1q_s8(w); w = (const void*) ((const int8_t*) w + 16);
72 const int8x16_t vb4567x0123 = vld1q_s8(w); w = (const void*) ((const int8_t*) w + 16);
73 const int8x16_t vb4567x4567 = vld1q_s8(w); w = (const void*) ((const int8_t*) w + 16);
74 const int8x16_t vb4567x89AB = vld1q_s8(w); w = (const void*) ((const int8_t*) w + 16);
75 const int8x16_t vb4567xCDEF = vld1q_s8(w); w = (const void*) ((const int8_t*) w + 16);
76
77 // Multiply-accumulate: 1x8 * 8x16 --> 1x16.
78 vacc0x0123 = vdotq_lane_s32(vacc0x0123, vb0123x0123, va0x01234567, 0);
79 vacc0x4567 = vdotq_lane_s32(vacc0x4567, vb0123x4567, va0x01234567, 0);
80 vacc0x89AB = vdotq_lane_s32(vacc0x89AB, vb0123x89AB, va0x01234567, 0);
81 vacc0xCDEF = vdotq_lane_s32(vacc0xCDEF, vb0123xCDEF, va0x01234567, 0);
82 vacc0x0123 = vdotq_lane_s32(vacc0x0123, vb4567x0123, va0x01234567, 1);
83 vacc0x4567 = vdotq_lane_s32(vacc0x4567, vb4567x4567, va0x01234567, 1);
84 vacc0x89AB = vdotq_lane_s32(vacc0x89AB, vb4567x89AB, va0x01234567, 1);
85 vacc0xCDEF = vdotq_lane_s32(vacc0xCDEF, vb4567xCDEF, va0x01234567, 1);
86
87 k -= 8 * sizeof(int8_t);
88 }
89 // Handle up to 4 final positions of `k`
90 if XNN_UNLIKELY(k != 0) {
91 // Load a 1x4 block of activations.
92 const int8x8_t va0x01234567 = vld1_s8(a0);
93
94 // Load a 4x16 block of weights.
95 const int8x16_t vb0123x0123 = vld1q_s8(w); w = (const void*) ((const int8_t*) w + 16);
96 const int8x16_t vb0123x4567 = vld1q_s8(w); w = (const void*) ((const int8_t*) w + 16);
97 const int8x16_t vb0123x89AB = vld1q_s8(w); w = (const void*) ((const int8_t*) w + 16);
98 const int8x16_t vb0123xCDEF = vld1q_s8(w); w = (const void*) ((const int8_t*) w + 16);
99
100 // Multiply-accumulate: 1x4 * 4x16 --> 1x16.
101 vacc0x0123 = vdotq_lane_s32(vacc0x0123, vb0123x0123, va0x01234567, 0);
102 vacc0x4567 = vdotq_lane_s32(vacc0x4567, vb0123x4567, va0x01234567, 0);
103 vacc0x89AB = vdotq_lane_s32(vacc0x89AB, vb0123x89AB, va0x01234567, 0);
104 vacc0xCDEF = vdotq_lane_s32(vacc0xCDEF, vb0123xCDEF, va0x01234567, 0);
105 }
106 p -= 1 * sizeof(void*);
107 } while (p != 0);
108
109 const int32x4_t vright_pre_shift = vld1q_dup_s32(¶ms->rndnu_neon.right_pre_shift);
110 const int32x4_t vmultiplier = vld1q_dup_s32(¶ms->rndnu_neon.multiplier);
111 const int32x4_t vright_post_shift = vld1q_dup_s32(¶ms->rndnu_neon.right_post_shift);
112
113 vacc0x0123 = vshlq_s32(vacc0x0123, vright_pre_shift);
114 vacc0x4567 = vshlq_s32(vacc0x4567, vright_pre_shift);
115 vacc0x89AB = vshlq_s32(vacc0x89AB, vright_pre_shift);
116 vacc0xCDEF = vshlq_s32(vacc0xCDEF, vright_pre_shift);
117
118 vacc0x0123 = vqdmulhq_s32(vacc0x0123, vmultiplier);
119 vacc0x4567 = vqdmulhq_s32(vacc0x4567, vmultiplier);
120 vacc0x89AB = vqdmulhq_s32(vacc0x89AB, vmultiplier);
121 vacc0xCDEF = vqdmulhq_s32(vacc0xCDEF, vmultiplier);
122
123 vacc0x0123 = vrshlq_s32(vacc0x0123, vright_post_shift);
124 vacc0x4567 = vrshlq_s32(vacc0x4567, vright_post_shift);
125 vacc0x89AB = vrshlq_s32(vacc0x89AB, vright_post_shift);
126 vacc0xCDEF = vrshlq_s32(vacc0xCDEF, vright_post_shift);
127
128 const int16x8_t voutput_zero_point = vld1q_dup_s16(¶ms->rndnu_neon.output_zero_point);
129 #if XNN_ARCH_ARM64
130 const int16x8_t vacc0x01234567 = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc0x0123), vacc0x4567), voutput_zero_point);
131 const int16x8_t vacc0x89ABCDEF = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc0x89AB), vacc0xCDEF), voutput_zero_point);
132
133 int8x16_t vout0x0123456789ABCDEF = vqmovn_high_s16(vqmovn_s16(vacc0x01234567), vacc0x89ABCDEF);
134 #else
135 const int16x8_t vacc0x01234567 = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc0x0123), vqmovn_s32(vacc0x4567)), voutput_zero_point);
136 const int16x8_t vacc0x89ABCDEF = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc0x89AB), vqmovn_s32(vacc0xCDEF)), voutput_zero_point);
137
138 int8x16_t vout0x0123456789ABCDEF = vcombine_s8(vqmovn_s16(vacc0x01234567), vqmovn_s16(vacc0x89ABCDEF));
139 #endif
140 const int8x16_t voutput_min = vld1q_dup_s8(¶ms->rndnu_neon.output_min);
141 const int8x16_t voutput_max = vld1q_dup_s8(¶ms->rndnu_neon.output_max);
142
143 vout0x0123456789ABCDEF = vmaxq_s8(vout0x0123456789ABCDEF, voutput_min);
144
145 vout0x0123456789ABCDEF = vminq_s8(vout0x0123456789ABCDEF, voutput_max);
146
147 if (nc >= 16) {
148 vst1q_s8(c0 + 0, vout0x0123456789ABCDEF);
149
150 c0 = (int8_t*) ((uintptr_t) c0 + cn_stride);
151
152 a = (const int8_t**restrict) ((uintptr_t) a - ks);
153
154 nc -= 16;
155 } else {
156 int8x8_t vout0x01234567 = vget_low_s8(vout0x0123456789ABCDEF);
157 if (nc & 8) {
158 vst1_s8(c0, vout0x01234567); c0 += 8;
159 vout0x01234567 = vget_high_s8(vout0x0123456789ABCDEF);
160 }
161 if (nc & 4) {
162 vst1_lane_u32((void*) c0, vreinterpret_u32_s8(vout0x01234567), 0); c0 += 4;
163 vout0x01234567 = vext_s8(vout0x01234567, vout0x01234567, 4);
164 }
165 if (nc & 2) {
166 vst1_lane_u16((void*) c0, vreinterpret_u16_s8(vout0x01234567), 0); c0 += 2;
167 vout0x01234567 = vext_s8(vout0x01234567, vout0x01234567, 2);
168 }
169 if (nc & 1) {
170 vst1_lane_s8(c0, vout0x01234567, 0);
171 }
172
173 nc = 0;
174 }
175 } while (nc != 0);
176 }
177