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/intrinsics-polyfill.h>
16 #include <xnnpack/math.h>
17
18
xnn_qc8_igemm_minmax_fp32_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_qc8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS (1)])19 void xnn_qc8_igemm_minmax_fp32_ukernel_1x8c4__neondot(
20 size_t mr,
21 size_t nc,
22 size_t kc,
23 size_t ks,
24 const int8_t** restrict a,
25 const void* restrict w,
26 int8_t* restrict c,
27 size_t cm_stride,
28 size_t cn_stride,
29 size_t a_offset,
30 const int8_t* zero,
31 const union xnn_qc8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
32 {
33 assert(mr != 0);
34 assert(mr <= 1);
35 assert(nc != 0);
36 assert(kc != 0);
37 assert(ks != 0);
38 assert(ks % (1 * sizeof(void*)) == 0);
39 assert(a_offset % sizeof(int8_t) == 0);
40 assert(a != NULL);
41 assert(w != NULL);
42 assert(c != NULL);
43
44 kc = round_up_po2(kc, 4 * sizeof(int8_t));
45 int8_t* c0 = c;
46
47 do {
48 int32x4_t vacc0x0123 = vld1q_s32(w); w = (const void*) ((const int32_t*) w + 4);
49 int32x4_t vacc0x4567 = vld1q_s32(w); w = (const void*) ((const int32_t*) w + 4);
50
51 size_t p = ks;
52 do {
53 const int8_t* restrict a0 = a[0];
54 if XNN_UNPREDICTABLE(a0 != zero) {
55 a0 = (const int8_t*) ((uintptr_t) a0 + a_offset);
56 }
57 a += 1;
58
59 // Inner accumulation loop along the 8 columns.
60 size_t k = kc;
61 // 2x partial unrolled loop to load 8 bytes at a time.
62 while (k >= 8 * sizeof(int8_t)) {
63 // Load a 1x8 block of activations.
64 const int8x8_t va0x01234567 = vld1_s8(a0); a0 += 8;
65
66 // Load a 8x8 block of weights.
67 const int8x16_t vb0123x0123 = vld1q_s8(w); w = (const void*) ((const int8_t*) w + 16);
68 const int8x16_t vb0123x4567 = vld1q_s8(w); w = (const void*) ((const int8_t*) w + 16);
69 const int8x16_t vb4567x0123 = vld1q_s8(w); w = (const void*) ((const int8_t*) w + 16);
70 const int8x16_t vb4567x4567 = vld1q_s8(w); w = (const void*) ((const int8_t*) w + 16);
71
72 // Multiply-accumulate: 1x8 * 8x8 --> 1x8.
73 vacc0x0123 = vdotq_lane_s32(vacc0x0123, vb0123x0123, va0x01234567, 0);
74 vacc0x4567 = vdotq_lane_s32(vacc0x4567, vb0123x4567, va0x01234567, 0);
75 vacc0x0123 = vdotq_lane_s32(vacc0x0123, vb4567x0123, va0x01234567, 1);
76 vacc0x4567 = vdotq_lane_s32(vacc0x4567, vb4567x4567, va0x01234567, 1);
77
78 k -= 8 * sizeof(int8_t);
79 }
80 // Handle up to 4 final positions of `k`
81 if XNN_UNLIKELY(k != 0) {
82 // Load a 1x4 block of activations.
83 const int8x8_t va0x01234567 = vld1_s8(a0);
84
85 // Load a 4x8 block of weights.
86 const int8x16_t vb0123x0123 = vld1q_s8(w); w = (const void*) ((const int8_t*) w + 16);
87 const int8x16_t vb0123x4567 = vld1q_s8(w); w = (const void*) ((const int8_t*) w + 16);
88
89 // Multiply-accumulate: 1x4 * 4x8 --> 1x8.
90 vacc0x0123 = vdotq_lane_s32(vacc0x0123, vb0123x0123, va0x01234567, 0);
91 vacc0x4567 = vdotq_lane_s32(vacc0x4567, vb0123x4567, va0x01234567, 0);
92 }
93 p -= 1 * sizeof(void*);
94 } while (p != 0);
95
96 float32x4_t vfpacc0x0123 = vcvtq_f32_s32(vacc0x0123);
97 float32x4_t vfpacc0x4567 = vcvtq_f32_s32(vacc0x4567);
98
99 const float32x4_t vscale0123 = vld1q_f32((const float*) w); w = (const void*) ((const float*) w + 4);
100 vfpacc0x0123 = vmulq_f32(vfpacc0x0123, vscale0123);
101 const float32x4_t vscale4567 = vld1q_f32((const float*) w); w = (const void*) ((const float*) w + 4);
102 vfpacc0x4567 = vmulq_f32(vfpacc0x4567, vscale4567);
103
104 vacc0x0123 = vcvtnq_s32_f32(vfpacc0x0123);
105 vacc0x4567 = vcvtnq_s32_f32(vfpacc0x4567);
106
107 const int16x8_t voutput_zero_point = vld1q_dup_s16(¶ms->fp32_neonv8.output_zero_point);
108 #if XNN_ARCH_ARM64
109 const int16x8_t vacc0x01234567 = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc0x0123), vacc0x4567), voutput_zero_point);
110
111 int8x8_t vout0x01234567 = vqmovn_s16(vacc0x01234567);
112 #else
113 const int16x8_t vacc0x01234567 = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc0x0123), vqmovn_s32(vacc0x4567)), voutput_zero_point);
114
115 int8x8_t vout0x01234567 = vqmovn_s16(vacc0x01234567);
116 #endif
117 const int8x8_t voutput_min = vld1_dup_s8(¶ms->fp32_neonv8.output_min);
118 const int8x8_t voutput_max = vld1_dup_s8(¶ms->fp32_neonv8.output_max);
119
120 vout0x01234567 = vmax_s8(vout0x01234567, voutput_min);
121
122 vout0x01234567 = vmin_s8(vout0x01234567, voutput_max);
123
124 if (nc >= 8) {
125 vst1_s8(c0 + 0, vout0x01234567);
126
127 c0 = (int8_t*) ((uintptr_t) c0 + cn_stride);
128
129 a = (const int8_t**restrict) ((uintptr_t) a - ks);
130
131 nc -= 8;
132 } else {
133 if (nc & 4) {
134 vst1_lane_u32((void*) c0, vreinterpret_u32_s8(vout0x01234567), 0); c0 += 4;
135 vout0x01234567 = vext_s8(vout0x01234567, vout0x01234567, 4);
136 }
137 if (nc & 2) {
138 vst1_lane_u16((void*) c0, vreinterpret_u16_s8(vout0x01234567), 0); c0 += 2;
139 vout0x01234567 = vext_s8(vout0x01234567, vout0x01234567, 2);
140 }
141 if (nc & 1) {
142 vst1_lane_s8(c0, vout0x01234567, 0);
143 }
144
145 nc = 0;
146 }
147 } while (nc != 0);
148 }
149