1 // Auto-generated file. Do not edit!
2 // Template: src/qu8-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_qu8_igemm_minmax_fp32_ukernel_1x16c4__neondot(size_t mr,size_t nc,size_t kc,size_t ks,const uint8_t ** restrict a,const void * restrict w,uint8_t * restrict c,size_t cm_stride,size_t cn_stride,size_t a_offset,const uint8_t * zero,const union xnn_qu8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS (1)])19 void xnn_qu8_igemm_minmax_fp32_ukernel_1x16c4__neondot(
20 size_t mr,
21 size_t nc,
22 size_t kc,
23 size_t ks,
24 const uint8_t** restrict a,
25 const void* restrict w,
26 uint8_t* restrict c,
27 size_t cm_stride,
28 size_t cn_stride,
29 size_t a_offset,
30 const uint8_t* zero,
31 const union xnn_qu8_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(uint8_t) == 0);
40 assert(a != NULL);
41 assert(w != NULL);
42 assert(c != NULL);
43
44 kc = round_up_po2(kc, 4 * sizeof(uint8_t));
45 uint8_t* c0 = c;
46
47 const uint8x8_t va_zero_point = vld1_dup_u8(¶ms->fp32_neonv8.kernel_zero_point[0]);
48
49 do {
50 // Initialize accumulators with bias. 16 bias values are loaded from the
51 // weight matrix, at the start of the group of 16 columns.
52 uint32x4_t vpacc0x0123 = vld1q_u32(w); w = (const void*) ((const uint32_t*) w + 4);
53 uint32x4_t vpacc0x4567 = vld1q_u32(w); w = (const void*) ((const uint32_t*) w + 4);
54 uint32x4_t vpacc0x89AB = vld1q_u32(w); w = (const void*) ((const uint32_t*) w + 4);
55 uint32x4_t vpacc0xCDEF = vld1q_u32(w); w = (const void*) ((const uint32_t*) w + 4);
56 uint32x2_t vnacc0 = vmov_n_u32(0);
57
58 size_t p = ks;
59 do {
60 const uint8_t* restrict a0 = a[0];
61 if XNN_UNPREDICTABLE(a0 != zero) {
62 a0 = (const uint8_t*) ((uintptr_t) a0 + a_offset);
63 }
64 a += 1;
65
66 // Inner accumulation loop along the 16 columns.
67 size_t k = kc;
68 // 2x partial unrolled loop to load 8 bytes at a time.
69 while (k >= 8 * sizeof(uint8_t)) {
70 // Load a 1x8 block of activations.
71 const uint8x8_t va0x01234567 = vld1_u8(a0); a0 += 8;
72
73 // Load a 8x16 block of weights.
74 const uint8x16_t vb0123x0123 = vld1q_u8(w); w = (const void*) ((const uint8_t*) w + 16);
75 const uint8x16_t vb0123x4567 = vld1q_u8(w); w = (const void*) ((const uint8_t*) w + 16);
76 const uint8x16_t vb0123x89AB = vld1q_u8(w); w = (const void*) ((const uint8_t*) w + 16);
77 const uint8x16_t vb0123xCDEF = vld1q_u8(w); w = (const void*) ((const uint8_t*) w + 16);
78 const uint8x16_t vb4567x0123 = vld1q_u8(w); w = (const void*) ((const uint8_t*) w + 16);
79 const uint8x16_t vb4567x4567 = vld1q_u8(w); w = (const void*) ((const uint8_t*) w + 16);
80 const uint8x16_t vb4567x89AB = vld1q_u8(w); w = (const void*) ((const uint8_t*) w + 16);
81 const uint8x16_t vb4567xCDEF = vld1q_u8(w); w = (const void*) ((const uint8_t*) w + 16);
82
83 // Multiply-accumulate: 1x8 * 8x16 --> 1x16.
84 vnacc0 = vdot_u32(vnacc0, va_zero_point, va0x01234567);
85 vpacc0x0123 = vdotq_lane_u32(vpacc0x0123, vb0123x0123, va0x01234567, 0);
86 vpacc0x4567 = vdotq_lane_u32(vpacc0x4567, vb0123x4567, va0x01234567, 0);
87 vpacc0x89AB = vdotq_lane_u32(vpacc0x89AB, vb0123x89AB, va0x01234567, 0);
88 vpacc0xCDEF = vdotq_lane_u32(vpacc0xCDEF, vb0123xCDEF, va0x01234567, 0);
89 vpacc0x0123 = vdotq_lane_u32(vpacc0x0123, vb4567x0123, va0x01234567, 1);
90 vpacc0x4567 = vdotq_lane_u32(vpacc0x4567, vb4567x4567, va0x01234567, 1);
91 vpacc0x89AB = vdotq_lane_u32(vpacc0x89AB, vb4567x89AB, va0x01234567, 1);
92 vpacc0xCDEF = vdotq_lane_u32(vpacc0xCDEF, vb4567xCDEF, va0x01234567, 1);
93
94 k -= 8 * sizeof(uint8_t);
95 }
96 // Handle up to 4 final positions of `k`
97 if XNN_UNLIKELY(k != 0) {
98 // Load a 1x4 block of activations.
99 const uint8x8_t va0x01234567 = vreinterpret_u8_u32(vld1_lane_u32((const void*) a0, vmov_n_u32(0), 0)); a0 += 4;
100
101 // Load a 4x16 block of weights.
102 const uint8x16_t vb0123x0123 = vld1q_u8(w); w = (const void*) ((const uint8_t*) w + 16);
103 const uint8x16_t vb0123x4567 = vld1q_u8(w); w = (const void*) ((const uint8_t*) w + 16);
104 const uint8x16_t vb0123x89AB = vld1q_u8(w); w = (const void*) ((const uint8_t*) w + 16);
105 const uint8x16_t vb0123xCDEF = vld1q_u8(w); w = (const void*) ((const uint8_t*) w + 16);
106
107 // Multiply-accumulate: 1x4 * 4x16 --> 1x16.
108 vnacc0 = vdot_u32(vnacc0, va_zero_point, va0x01234567);
109 vpacc0x0123 = vdotq_lane_u32(vpacc0x0123, vb0123x0123, va0x01234567, 0);
110 vpacc0x4567 = vdotq_lane_u32(vpacc0x4567, vb0123x4567, va0x01234567, 0);
111 vpacc0x89AB = vdotq_lane_u32(vpacc0x89AB, vb0123x89AB, va0x01234567, 0);
112 vpacc0xCDEF = vdotq_lane_u32(vpacc0xCDEF, vb0123xCDEF, va0x01234567, 0);
113 }
114 p -= 1 * sizeof(void*);
115 } while (p != 0);
116
117 // Subtract zero point from accumulators.
118 vnacc0 = vpadd_u32(vnacc0, vnacc0);
119 const uint32x4_t vnacc0x0123 = vcombine_u32(vnacc0, vnacc0);
120 int32x4_t vacc0x0123 = vreinterpretq_s32_u32(vsubq_u32(vpacc0x0123, vnacc0x0123));
121 int32x4_t vacc0x4567 = vreinterpretq_s32_u32(vsubq_u32(vpacc0x4567, vnacc0x0123));
122 int32x4_t vacc0x89AB = vreinterpretq_s32_u32(vsubq_u32(vpacc0x89AB, vnacc0x0123));
123 int32x4_t vacc0xCDEF = vreinterpretq_s32_u32(vsubq_u32(vpacc0xCDEF, vnacc0x0123));
124
125 float32x4_t vfpacc0x0123 = vcvtq_f32_s32(vacc0x0123);
126 float32x4_t vfpacc0x4567 = vcvtq_f32_s32(vacc0x4567);
127 float32x4_t vfpacc0x89AB = vcvtq_f32_s32(vacc0x89AB);
128 float32x4_t vfpacc0xCDEF = vcvtq_f32_s32(vacc0xCDEF);
129
130 const float32x4_t vscale = vld1q_dup_f32(¶ms->fp32_neonv8.scale);
131 vfpacc0x0123 = vmulq_f32(vfpacc0x0123, vscale);
132 vfpacc0x4567 = vmulq_f32(vfpacc0x4567, vscale);
133 vfpacc0x89AB = vmulq_f32(vfpacc0x89AB, vscale);
134 vfpacc0xCDEF = vmulq_f32(vfpacc0xCDEF, vscale);
135
136 vacc0x0123 = vcvtnq_s32_f32(vfpacc0x0123);
137 vacc0x4567 = vcvtnq_s32_f32(vfpacc0x4567);
138 vacc0x89AB = vcvtnq_s32_f32(vfpacc0x89AB);
139 vacc0xCDEF = vcvtnq_s32_f32(vfpacc0xCDEF);
140
141 const int16x8_t voutput_zero_point = vld1q_dup_s16(¶ms->fp32_neonv8.output_zero_point);
142 #if XNN_ARCH_ARM64
143 const int16x8_t vacc0x01234567 = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc0x0123), vacc0x4567), voutput_zero_point);
144 const int16x8_t vacc0x89ABCDEF = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc0x89AB), vacc0xCDEF), voutput_zero_point);
145
146 uint8x16_t vout0x0123456789ABCDEF = vqmovun_high_s16(vqmovun_s16(vacc0x01234567), vacc0x89ABCDEF);
147 #else
148 const int16x8_t vacc0x01234567 = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc0x0123), vqmovn_s32(vacc0x4567)), voutput_zero_point);
149 const int16x8_t vacc0x89ABCDEF = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc0x89AB), vqmovn_s32(vacc0xCDEF)), voutput_zero_point);
150
151 uint8x16_t vout0x0123456789ABCDEF = vcombine_u8(vqmovun_s16(vacc0x01234567), vqmovun_s16(vacc0x89ABCDEF));
152 #endif
153 const uint8x16_t voutput_min = vld1q_dup_u8(¶ms->fp32_neonv8.output_min);
154 const uint8x16_t voutput_max = vld1q_dup_u8(¶ms->fp32_neonv8.output_max);
155
156 vout0x0123456789ABCDEF = vmaxq_u8(vout0x0123456789ABCDEF, voutput_min);
157
158 vout0x0123456789ABCDEF = vminq_u8(vout0x0123456789ABCDEF, voutput_max);
159
160 if (nc >= 16) {
161 vst1q_u8(c0 + 0, vout0x0123456789ABCDEF);
162
163 c0 = (uint8_t*) ((uintptr_t) c0 + cn_stride);
164
165 a = (const uint8_t**restrict) ((uintptr_t) a - ks);
166
167 nc -= 16;
168 } else {
169 uint8x8_t vout0x01234567 = vget_low_u8(vout0x0123456789ABCDEF);
170 if (nc & 8) {
171 vst1_u8(c0, vout0x01234567); c0 += 8; // This line
172 vout0x01234567 = vget_high_u8(vout0x0123456789ABCDEF);
173 }
174 if (nc & 4) {
175 vst1_lane_u32((void*) c0, vreinterpret_u32_u8(vout0x01234567), 0); c0 += 4;
176 vout0x01234567 = vext_u8(vout0x01234567, vout0x01234567, 4);
177 }
178 if (nc & 2) {
179 vst1_lane_u16((void*) c0, vreinterpret_u16_u8(vout0x01234567), 0); c0 += 2;
180 vout0x01234567 = vext_u8(vout0x01234567, vout0x01234567, 2);
181 }
182 if (nc & 1) {
183 vst1_lane_u8(c0, vout0x01234567, 0);
184 }
185
186 nc = 0;
187 }
188 } while (nc != 0);
189 }
190