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_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_qc8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS (1)])19 void xnn_qc8_igemm_minmax_fp32_ukernel_1x16c4__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 int32x4_t vacc0x89AB = vld1q_s32(w); w = (const void*) ((const int32_t*) w + 4);
51 int32x4_t vacc0xCDEF = vld1q_s32(w); w = (const void*) ((const int32_t*) w + 4);
52
53 size_t p = ks;
54 do {
55 const int8_t* restrict a0 = a[0];
56 if XNN_UNPREDICTABLE(a0 != zero) {
57 a0 = (const int8_t*) ((uintptr_t) a0 + a_offset);
58 }
59 a += 1;
60
61 // Inner accumulation loop along the 16 columns.
62 size_t k = kc;
63 // 2x partial unrolled loop to load 8 bytes at a time.
64 while (k >= 8 * sizeof(int8_t)) {
65 // Load a 1x8 block of activations.
66 const int8x8_t va0x01234567 = vld1_s8(a0); a0 += 8;
67
68 // Load a 8x16 block of weights.
69 const int8x16_t vb0123x0123 = vld1q_s8(w); w = (const void*) ((const int8_t*) w + 16);
70 const int8x16_t vb0123x4567 = vld1q_s8(w); w = (const void*) ((const int8_t*) w + 16);
71 const int8x16_t vb0123x89AB = vld1q_s8(w); w = (const void*) ((const int8_t*) w + 16);
72 const int8x16_t vb0123xCDEF = vld1q_s8(w); w = (const void*) ((const int8_t*) w + 16);
73 const int8x16_t vb4567x0123 = vld1q_s8(w); w = (const void*) ((const int8_t*) w + 16);
74 const int8x16_t vb4567x4567 = vld1q_s8(w); w = (const void*) ((const int8_t*) w + 16);
75 const int8x16_t vb4567x89AB = vld1q_s8(w); w = (const void*) ((const int8_t*) w + 16);
76 const int8x16_t vb4567xCDEF = vld1q_s8(w); w = (const void*) ((const int8_t*) w + 16);
77
78 // Multiply-accumulate: 1x8 * 8x16 --> 1x16.
79 vacc0x0123 = vdotq_lane_s32(vacc0x0123, vb0123x0123, va0x01234567, 0);
80 vacc0x4567 = vdotq_lane_s32(vacc0x4567, vb0123x4567, va0x01234567, 0);
81 vacc0x89AB = vdotq_lane_s32(vacc0x89AB, vb0123x89AB, va0x01234567, 0);
82 vacc0xCDEF = vdotq_lane_s32(vacc0xCDEF, vb0123xCDEF, va0x01234567, 0);
83 vacc0x0123 = vdotq_lane_s32(vacc0x0123, vb4567x0123, va0x01234567, 1);
84 vacc0x4567 = vdotq_lane_s32(vacc0x4567, vb4567x4567, va0x01234567, 1);
85 vacc0x89AB = vdotq_lane_s32(vacc0x89AB, vb4567x89AB, va0x01234567, 1);
86 vacc0xCDEF = vdotq_lane_s32(vacc0xCDEF, vb4567xCDEF, va0x01234567, 1);
87
88 k -= 8 * sizeof(int8_t);
89 }
90 // Handle up to 4 final positions of `k`
91 if XNN_UNLIKELY(k != 0) {
92 // Load a 1x4 block of activations.
93 const int8x8_t va0x01234567 = vld1_s8(a0);
94
95 // Load a 4x16 block of weights.
96 const int8x16_t vb0123x0123 = vld1q_s8(w); w = (const void*) ((const int8_t*) w + 16);
97 const int8x16_t vb0123x4567 = vld1q_s8(w); w = (const void*) ((const int8_t*) w + 16);
98 const int8x16_t vb0123x89AB = vld1q_s8(w); w = (const void*) ((const int8_t*) w + 16);
99 const int8x16_t vb0123xCDEF = vld1q_s8(w); w = (const void*) ((const int8_t*) w + 16);
100
101 // Multiply-accumulate: 1x4 * 4x16 --> 1x16.
102 vacc0x0123 = vdotq_lane_s32(vacc0x0123, vb0123x0123, va0x01234567, 0);
103 vacc0x4567 = vdotq_lane_s32(vacc0x4567, vb0123x4567, va0x01234567, 0);
104 vacc0x89AB = vdotq_lane_s32(vacc0x89AB, vb0123x89AB, va0x01234567, 0);
105 vacc0xCDEF = vdotq_lane_s32(vacc0xCDEF, vb0123xCDEF, va0x01234567, 0);
106 }
107 p -= 1 * sizeof(void*);
108 } while (p != 0);
109
110 float32x4_t vfpacc0x0123 = vcvtq_f32_s32(vacc0x0123);
111 float32x4_t vfpacc0x4567 = vcvtq_f32_s32(vacc0x4567);
112 float32x4_t vfpacc0x89AB = vcvtq_f32_s32(vacc0x89AB);
113 float32x4_t vfpacc0xCDEF = vcvtq_f32_s32(vacc0xCDEF);
114
115 const float32x4_t vscale0123 = vld1q_f32((const float*) w); w = (const void*) ((const float*) w + 4);
116 vfpacc0x0123 = vmulq_f32(vfpacc0x0123, vscale0123);
117 const float32x4_t vscale4567 = vld1q_f32((const float*) w); w = (const void*) ((const float*) w + 4);
118 vfpacc0x4567 = vmulq_f32(vfpacc0x4567, vscale4567);
119 const float32x4_t vscale89AB = vld1q_f32((const float*) w); w = (const void*) ((const float*) w + 4);
120 vfpacc0x89AB = vmulq_f32(vfpacc0x89AB, vscale89AB);
121 const float32x4_t vscaleCDEF = vld1q_f32((const float*) w); w = (const void*) ((const float*) w + 4);
122 vfpacc0xCDEF = vmulq_f32(vfpacc0xCDEF, vscaleCDEF);
123
124 vacc0x0123 = vcvtnq_s32_f32(vfpacc0x0123);
125 vacc0x4567 = vcvtnq_s32_f32(vfpacc0x4567);
126 vacc0x89AB = vcvtnq_s32_f32(vfpacc0x89AB);
127 vacc0xCDEF = vcvtnq_s32_f32(vfpacc0xCDEF);
128
129 const int16x8_t voutput_zero_point = vld1q_dup_s16(¶ms->fp32_neonv8.output_zero_point);
130 #if XNN_ARCH_ARM64
131 const int16x8_t vacc0x01234567 = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc0x0123), vacc0x4567), voutput_zero_point);
132 const int16x8_t vacc0x89ABCDEF = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc0x89AB), vacc0xCDEF), voutput_zero_point);
133
134 int8x16_t vout0x0123456789ABCDEF = vqmovn_high_s16(vqmovn_s16(vacc0x01234567), vacc0x89ABCDEF);
135 #else
136 const int16x8_t vacc0x01234567 = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc0x0123), vqmovn_s32(vacc0x4567)), voutput_zero_point);
137 const int16x8_t vacc0x89ABCDEF = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc0x89AB), vqmovn_s32(vacc0xCDEF)), voutput_zero_point);
138
139 int8x16_t vout0x0123456789ABCDEF = vcombine_s8(vqmovn_s16(vacc0x01234567), vqmovn_s16(vacc0x89ABCDEF));
140 #endif
141 const int8x16_t voutput_min = vld1q_dup_s8(¶ms->fp32_neonv8.output_min);
142 const int8x16_t voutput_max = vld1q_dup_s8(¶ms->fp32_neonv8.output_max);
143
144 vout0x0123456789ABCDEF = vmaxq_s8(vout0x0123456789ABCDEF, voutput_min);
145
146 vout0x0123456789ABCDEF = vminq_s8(vout0x0123456789ABCDEF, voutput_max);
147
148 if (nc >= 16) {
149 vst1q_s8(c0 + 0, vout0x0123456789ABCDEF);
150
151 c0 = (int8_t*) ((uintptr_t) c0 + cn_stride);
152
153 a = (const int8_t**restrict) ((uintptr_t) a - ks);
154
155 nc -= 16;
156 } else {
157 int8x8_t vout0x01234567 = vget_low_s8(vout0x0123456789ABCDEF);
158 if (nc & 8) {
159 vst1_s8(c0, vout0x01234567); c0 += 8;
160 vout0x01234567 = vget_high_s8(vout0x0123456789ABCDEF);
161 }
162 if (nc & 4) {
163 vst1_lane_u32((void*) c0, vreinterpret_u32_s8(vout0x01234567), 0); c0 += 4;
164 vout0x01234567 = vext_s8(vout0x01234567, vout0x01234567, 4);
165 }
166 if (nc & 2) {
167 vst1_lane_u16((void*) c0, vreinterpret_u16_s8(vout0x01234567), 0); c0 += 2;
168 vout0x01234567 = vext_s8(vout0x01234567, vout0x01234567, 2);
169 }
170 if (nc & 1) {
171 vst1_lane_s8(c0, vout0x01234567, 0);
172 }
173
174 nc = 0;
175 }
176 } while (nc != 0);
177 }
178