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/math.h>
16
17
xnn_qu8_igemm_minmax_rndnu_ukernel_2x8c4__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)])18 void xnn_qu8_igemm_minmax_rndnu_ukernel_2x8c4__neondot(
19 size_t mr,
20 size_t nc,
21 size_t kc,
22 size_t ks,
23 const uint8_t** restrict a,
24 const void* restrict w,
25 uint8_t* restrict c,
26 size_t cm_stride,
27 size_t cn_stride,
28 size_t a_offset,
29 const uint8_t* zero,
30 const union xnn_qu8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
31 {
32 assert(mr != 0);
33 assert(mr <= 2);
34 assert(nc != 0);
35 assert(kc != 0);
36 assert(ks != 0);
37 assert(ks % (2 * sizeof(void*)) == 0);
38 assert(a_offset % sizeof(uint8_t) == 0);
39 assert(a != NULL);
40 assert(w != NULL);
41 assert(c != NULL);
42
43 kc = round_up_po2(kc, 4 * sizeof(uint8_t));
44 uint8_t* c0 = c;
45 uint8_t* c1 = (uint8_t*) ((uintptr_t) c0 + cm_stride);
46 if XNN_UNPREDICTABLE(mr != 2) {
47 c1 = c0;
48 }
49
50 const uint8x8_t va_zero_point = vld1_dup_u8(¶ms->rndnu_neon.kernel_zero_point[0]);
51
52 do {
53 // Initialize accumulators with bias. 8 bias values are loaded from the
54 // weight matrix, at the start of the group of 8 columns.
55 uint32x4_t vpacc0x0123 = vld1q_u32(w); w = (const void*) ((const uint32_t*) w + 4);
56 uint32x4_t vpacc0x4567 = vld1q_u32(w); w = (const void*) ((const uint32_t*) w + 4);
57 uint32x4_t vpacc1x0123 = vpacc0x0123;
58 uint32x4_t vpacc1x4567 = vpacc0x4567;
59 uint32x2_t vnacc0 = vmov_n_u32(0);
60 uint32x2_t vnacc1 = vmov_n_u32(0);
61
62 size_t p = ks;
63 do {
64 const uint8_t* restrict a0 = a[0];
65 if XNN_UNPREDICTABLE(a0 != zero) {
66 a0 = (const uint8_t*) ((uintptr_t) a0 + a_offset);
67 }
68 const uint8_t* restrict a1 = a[1];
69 if XNN_UNPREDICTABLE(a1 != zero) {
70 a1 = (const uint8_t*) ((uintptr_t) a1 + a_offset);
71 }
72 a += 2;
73
74 // Inner accumulation loop along the 8 columns.
75 size_t k = kc;
76 // 2x partial unrolled loop to load 8 bytes at a time.
77 while (k >= 8 * sizeof(uint8_t)) {
78 // Load a 2x8 block of activations.
79 const uint8x8_t va0x01234567 = vld1_u8(a0); a0 += 8;
80 const uint8x8_t va1x01234567 = vld1_u8(a1); a1 += 8;
81
82 // Load a 8x8 block of weights.
83 const uint8x16_t vb0123x0123 = vld1q_u8(w); w = (const void*) ((const uint8_t*) w + 16);
84 const uint8x16_t vb0123x4567 = vld1q_u8(w); w = (const void*) ((const uint8_t*) w + 16);
85 const uint8x16_t vb4567x0123 = vld1q_u8(w); w = (const void*) ((const uint8_t*) w + 16);
86 const uint8x16_t vb4567x4567 = vld1q_u8(w); w = (const void*) ((const uint8_t*) w + 16);
87
88 // Multiply-accumulate: 2x8 * 8x8 --> 2x8.
89 vnacc0 = vdot_u32(vnacc0, va_zero_point, va0x01234567);
90 vpacc0x0123 = vdotq_lane_u32(vpacc0x0123, vb0123x0123, va0x01234567, 0);
91 vpacc0x4567 = vdotq_lane_u32(vpacc0x4567, vb0123x4567, va0x01234567, 0);
92 vpacc0x0123 = vdotq_lane_u32(vpacc0x0123, vb4567x0123, va0x01234567, 1);
93 vpacc0x4567 = vdotq_lane_u32(vpacc0x4567, vb4567x4567, va0x01234567, 1);
94 vnacc1 = vdot_u32(vnacc1, va_zero_point, va1x01234567);
95 vpacc1x0123 = vdotq_lane_u32(vpacc1x0123, vb0123x0123, va1x01234567, 0);
96 vpacc1x4567 = vdotq_lane_u32(vpacc1x4567, vb0123x4567, va1x01234567, 0);
97 vpacc1x0123 = vdotq_lane_u32(vpacc1x0123, vb4567x0123, va1x01234567, 1);
98 vpacc1x4567 = vdotq_lane_u32(vpacc1x4567, vb4567x4567, va1x01234567, 1);
99
100 k -= 8 * sizeof(uint8_t);
101 }
102 // Handle up to 4 final positions of `k`
103 if XNN_UNLIKELY(k != 0) {
104 // Load a 2x4 block of activations.
105 const uint8x8_t va0x01234567 = vreinterpret_u8_u32(vld1_lane_u32((const void*) a0, vmov_n_u32(0), 0)); a0 += 4;
106 const uint8x8_t va1x01234567 = vreinterpret_u8_u32(vld1_lane_u32((const void*) a1, vmov_n_u32(0), 0)); a1 += 4;
107
108 // Load a 4x8 block of weights.
109 const uint8x16_t vb0123x0123 = vld1q_u8(w); w = (const void*) ((const uint8_t*) w + 16);
110 const uint8x16_t vb0123x4567 = vld1q_u8(w); w = (const void*) ((const uint8_t*) w + 16);
111
112 // Multiply-accumulate: 2x4 * 4x8 --> 2x8.
113 vnacc0 = vdot_u32(vnacc0, va_zero_point, va0x01234567);
114 vpacc0x0123 = vdotq_lane_u32(vpacc0x0123, vb0123x0123, va0x01234567, 0);
115 vpacc0x4567 = vdotq_lane_u32(vpacc0x4567, vb0123x4567, va0x01234567, 0);
116 vnacc1 = vdot_u32(vnacc1, va_zero_point, va1x01234567);
117 vpacc1x0123 = vdotq_lane_u32(vpacc1x0123, vb0123x0123, va1x01234567, 0);
118 vpacc1x4567 = vdotq_lane_u32(vpacc1x4567, vb0123x4567, va1x01234567, 0);
119 }
120 p -= 2 * sizeof(void*);
121 } while (p != 0);
122
123 // Subtract zero point from accumulators.
124 vnacc0 = vpadd_u32(vnacc0, vnacc0);
125 const uint32x4_t vnacc0x0123 = vcombine_u32(vnacc0, vnacc0);
126 int32x4_t vacc0x0123 = vreinterpretq_s32_u32(vsubq_u32(vpacc0x0123, vnacc0x0123));
127 int32x4_t vacc0x4567 = vreinterpretq_s32_u32(vsubq_u32(vpacc0x4567, vnacc0x0123));
128 vnacc1 = vpadd_u32(vnacc1, vnacc1);
129 const uint32x4_t vnacc1x0123 = vcombine_u32(vnacc1, vnacc1);
130 int32x4_t vacc1x0123 = vreinterpretq_s32_u32(vsubq_u32(vpacc1x0123, vnacc1x0123));
131 int32x4_t vacc1x4567 = vreinterpretq_s32_u32(vsubq_u32(vpacc1x4567, vnacc1x0123));
132
133 const int32x4_t vright_pre_shift = vld1q_dup_s32(¶ms->rndnu_neon.right_pre_shift);
134 const int32x4_t vmultiplier = vld1q_dup_s32(¶ms->rndnu_neon.multiplier);
135 const int32x4_t vright_post_shift = vld1q_dup_s32(¶ms->rndnu_neon.right_post_shift);
136
137 vacc0x0123 = vshlq_s32(vacc0x0123, vright_pre_shift);
138 vacc0x4567 = vshlq_s32(vacc0x4567, vright_pre_shift);
139 vacc1x0123 = vshlq_s32(vacc1x0123, vright_pre_shift);
140 vacc1x4567 = vshlq_s32(vacc1x4567, vright_pre_shift);
141
142 vacc0x0123 = vqdmulhq_s32(vacc0x0123, vmultiplier);
143 vacc0x4567 = vqdmulhq_s32(vacc0x4567, vmultiplier);
144 vacc1x0123 = vqdmulhq_s32(vacc1x0123, vmultiplier);
145 vacc1x4567 = vqdmulhq_s32(vacc1x4567, vmultiplier);
146
147 vacc0x0123 = vrshlq_s32(vacc0x0123, vright_post_shift);
148 vacc0x4567 = vrshlq_s32(vacc0x4567, vright_post_shift);
149 vacc1x0123 = vrshlq_s32(vacc1x0123, vright_post_shift);
150 vacc1x4567 = vrshlq_s32(vacc1x4567, vright_post_shift);
151
152 const int16x8_t voutput_zero_point = vld1q_dup_s16(¶ms->rndnu_neon.output_zero_point);
153 #if XNN_ARCH_ARM64
154 const int16x8_t vacc0x01234567 = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc0x0123), vacc0x4567), voutput_zero_point);
155 const int16x8_t vacc1x01234567 = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc1x0123), vacc1x4567), voutput_zero_point);
156
157 uint8x16_t vout0x01234567_1x01234567 = vqmovun_high_s16(vqmovun_s16(vacc0x01234567), vacc1x01234567);
158 #else
159 const int16x8_t vacc0x01234567 = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc0x0123), vqmovn_s32(vacc0x4567)), voutput_zero_point);
160 const int16x8_t vacc1x01234567 = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc1x0123), vqmovn_s32(vacc1x4567)), voutput_zero_point);
161
162 uint8x16_t vout0x01234567_1x01234567 = vcombine_u8(vqmovun_s16(vacc0x01234567), vqmovun_s16(vacc1x01234567));
163 #endif
164 const uint8x16_t voutput_min = vld1q_dup_u8(¶ms->rndnu_neon.output_min);
165 const uint8x16_t voutput_max = vld1q_dup_u8(¶ms->rndnu_neon.output_max);
166
167 vout0x01234567_1x01234567 = vmaxq_u8(vout0x01234567_1x01234567, voutput_min);
168
169 vout0x01234567_1x01234567 = vminq_u8(vout0x01234567_1x01234567, voutput_max);
170
171 if (nc >= 8) {
172 vst1_u8(c1 + 0, vget_high_u8(vout0x01234567_1x01234567));
173 vst1_u8(c0 + 0, vget_low_u8(vout0x01234567_1x01234567));
174
175 c1 = (uint8_t*) ((uintptr_t) c1 + cn_stride);
176 c0 = (uint8_t*) ((uintptr_t) c0 + cn_stride);
177
178 a = (const uint8_t**restrict) ((uintptr_t) a - ks);
179
180 nc -= 8;
181 } else {
182 if (nc & 4) {
183 vst1q_lane_u32((void*) c1, vreinterpretq_u32_u8(vout0x01234567_1x01234567), 2); c1 += 4;
184 vst1q_lane_u32((void*) c0, vreinterpretq_u32_u8(vout0x01234567_1x01234567), 0); c0 += 4;
185 vout0x01234567_1x01234567 = vextq_u8(vout0x01234567_1x01234567, vout0x01234567_1x01234567, 4);
186 }
187 if (nc & 2) {
188 vst1q_lane_u16((void*) c1, vreinterpretq_u16_u8(vout0x01234567_1x01234567), 4); c1 += 2;
189 vst1q_lane_u16((void*) c0, vreinterpretq_u16_u8(vout0x01234567_1x01234567), 0); c0 += 2;
190 vout0x01234567_1x01234567 = vextq_u8(vout0x01234567_1x01234567, vout0x01234567_1x01234567, 2);
191 }
192 if (nc & 1) {
193 vst1q_lane_u8(c1, vout0x01234567_1x01234567, 8);
194 vst1q_lane_u8(c0, vout0x01234567_1x01234567, 0);
195 }
196
197 nc = 0;
198 }
199 } while (nc != 0);
200 }
201