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_2x16c4__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_2x16c4__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. 16 bias values are loaded from the
54 // weight matrix, at the start of the group of 16 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 vpacc0x89AB = vld1q_u32(w); w = (const void*) ((const uint32_t*) w + 4);
58 uint32x4_t vpacc0xCDEF = vld1q_u32(w); w = (const void*) ((const uint32_t*) w + 4);
59 uint32x4_t vpacc1x0123 = vpacc0x0123;
60 uint32x4_t vpacc1x4567 = vpacc0x4567;
61 uint32x4_t vpacc1x89AB = vpacc0x89AB;
62 uint32x4_t vpacc1xCDEF = vpacc0xCDEF;
63 uint32x2_t vnacc0 = vmov_n_u32(0);
64 uint32x2_t vnacc1 = vmov_n_u32(0);
65
66 size_t p = ks;
67 do {
68 const uint8_t* restrict a0 = a[0];
69 if XNN_UNPREDICTABLE(a0 != zero) {
70 a0 = (const uint8_t*) ((uintptr_t) a0 + a_offset);
71 }
72 const uint8_t* restrict a1 = a[1];
73 if XNN_UNPREDICTABLE(a1 != zero) {
74 a1 = (const uint8_t*) ((uintptr_t) a1 + a_offset);
75 }
76 a += 2;
77
78 // Inner accumulation loop along the 16 columns.
79 size_t k = kc;
80 // 2x partial unrolled loop to load 8 bytes at a time.
81 while (k >= 8 * sizeof(uint8_t)) {
82 // Load a 2x8 block of activations.
83 const uint8x8_t va0x01234567 = vld1_u8(a0); a0 += 8;
84 const uint8x8_t va1x01234567 = vld1_u8(a1); a1 += 8;
85
86 // Load a 8x16 block of weights.
87 const uint8x16_t vb0123x0123 = vld1q_u8(w); w = (const void*) ((const uint8_t*) w + 16);
88 const uint8x16_t vb0123x4567 = vld1q_u8(w); w = (const void*) ((const uint8_t*) w + 16);
89 const uint8x16_t vb0123x89AB = vld1q_u8(w); w = (const void*) ((const uint8_t*) w + 16);
90 const uint8x16_t vb0123xCDEF = vld1q_u8(w); w = (const void*) ((const uint8_t*) w + 16);
91 const uint8x16_t vb4567x0123 = vld1q_u8(w); w = (const void*) ((const uint8_t*) w + 16);
92 const uint8x16_t vb4567x4567 = vld1q_u8(w); w = (const void*) ((const uint8_t*) w + 16);
93 const uint8x16_t vb4567x89AB = vld1q_u8(w); w = (const void*) ((const uint8_t*) w + 16);
94 const uint8x16_t vb4567xCDEF = vld1q_u8(w); w = (const void*) ((const uint8_t*) w + 16);
95
96 // Multiply-accumulate: 2x8 * 8x16 --> 2x16.
97 vnacc0 = vdot_u32(vnacc0, va_zero_point, va0x01234567);
98 vpacc0x0123 = vdotq_lane_u32(vpacc0x0123, vb0123x0123, va0x01234567, 0);
99 vpacc0x4567 = vdotq_lane_u32(vpacc0x4567, vb0123x4567, va0x01234567, 0);
100 vpacc0x89AB = vdotq_lane_u32(vpacc0x89AB, vb0123x89AB, va0x01234567, 0);
101 vpacc0xCDEF = vdotq_lane_u32(vpacc0xCDEF, vb0123xCDEF, va0x01234567, 0);
102 vpacc0x0123 = vdotq_lane_u32(vpacc0x0123, vb4567x0123, va0x01234567, 1);
103 vpacc0x4567 = vdotq_lane_u32(vpacc0x4567, vb4567x4567, va0x01234567, 1);
104 vpacc0x89AB = vdotq_lane_u32(vpacc0x89AB, vb4567x89AB, va0x01234567, 1);
105 vpacc0xCDEF = vdotq_lane_u32(vpacc0xCDEF, vb4567xCDEF, va0x01234567, 1);
106 vnacc1 = vdot_u32(vnacc1, va_zero_point, va1x01234567);
107 vpacc1x0123 = vdotq_lane_u32(vpacc1x0123, vb0123x0123, va1x01234567, 0);
108 vpacc1x4567 = vdotq_lane_u32(vpacc1x4567, vb0123x4567, va1x01234567, 0);
109 vpacc1x89AB = vdotq_lane_u32(vpacc1x89AB, vb0123x89AB, va1x01234567, 0);
110 vpacc1xCDEF = vdotq_lane_u32(vpacc1xCDEF, vb0123xCDEF, va1x01234567, 0);
111 vpacc1x0123 = vdotq_lane_u32(vpacc1x0123, vb4567x0123, va1x01234567, 1);
112 vpacc1x4567 = vdotq_lane_u32(vpacc1x4567, vb4567x4567, va1x01234567, 1);
113 vpacc1x89AB = vdotq_lane_u32(vpacc1x89AB, vb4567x89AB, va1x01234567, 1);
114 vpacc1xCDEF = vdotq_lane_u32(vpacc1xCDEF, vb4567xCDEF, va1x01234567, 1);
115
116 k -= 8 * sizeof(uint8_t);
117 }
118 // Handle up to 4 final positions of `k`
119 if XNN_UNLIKELY(k != 0) {
120 // Load a 2x4 block of activations.
121 const uint8x8_t va0x01234567 = vreinterpret_u8_u32(vld1_lane_u32((const void*) a0, vmov_n_u32(0), 0)); a0 += 4;
122 const uint8x8_t va1x01234567 = vreinterpret_u8_u32(vld1_lane_u32((const void*) a1, vmov_n_u32(0), 0)); a1 += 4;
123
124 // Load a 4x16 block of weights.
125 const uint8x16_t vb0123x0123 = vld1q_u8(w); w = (const void*) ((const uint8_t*) w + 16);
126 const uint8x16_t vb0123x4567 = vld1q_u8(w); w = (const void*) ((const uint8_t*) w + 16);
127 const uint8x16_t vb0123x89AB = vld1q_u8(w); w = (const void*) ((const uint8_t*) w + 16);
128 const uint8x16_t vb0123xCDEF = vld1q_u8(w); w = (const void*) ((const uint8_t*) w + 16);
129
130 // Multiply-accumulate: 2x4 * 4x16 --> 2x16.
131 vnacc0 = vdot_u32(vnacc0, va_zero_point, va0x01234567);
132 vpacc0x0123 = vdotq_lane_u32(vpacc0x0123, vb0123x0123, va0x01234567, 0);
133 vpacc0x4567 = vdotq_lane_u32(vpacc0x4567, vb0123x4567, va0x01234567, 0);
134 vpacc0x89AB = vdotq_lane_u32(vpacc0x89AB, vb0123x89AB, va0x01234567, 0);
135 vpacc0xCDEF = vdotq_lane_u32(vpacc0xCDEF, vb0123xCDEF, va0x01234567, 0);
136 vnacc1 = vdot_u32(vnacc1, va_zero_point, va1x01234567);
137 vpacc1x0123 = vdotq_lane_u32(vpacc1x0123, vb0123x0123, va1x01234567, 0);
138 vpacc1x4567 = vdotq_lane_u32(vpacc1x4567, vb0123x4567, va1x01234567, 0);
139 vpacc1x89AB = vdotq_lane_u32(vpacc1x89AB, vb0123x89AB, va1x01234567, 0);
140 vpacc1xCDEF = vdotq_lane_u32(vpacc1xCDEF, vb0123xCDEF, va1x01234567, 0);
141 }
142 p -= 2 * sizeof(void*);
143 } while (p != 0);
144
145 // Subtract zero point from accumulators.
146 vnacc0 = vpadd_u32(vnacc0, vnacc0);
147 const uint32x4_t vnacc0x0123 = vcombine_u32(vnacc0, vnacc0);
148 int32x4_t vacc0x0123 = vreinterpretq_s32_u32(vsubq_u32(vpacc0x0123, vnacc0x0123));
149 int32x4_t vacc0x4567 = vreinterpretq_s32_u32(vsubq_u32(vpacc0x4567, vnacc0x0123));
150 int32x4_t vacc0x89AB = vreinterpretq_s32_u32(vsubq_u32(vpacc0x89AB, vnacc0x0123));
151 int32x4_t vacc0xCDEF = vreinterpretq_s32_u32(vsubq_u32(vpacc0xCDEF, vnacc0x0123));
152 vnacc1 = vpadd_u32(vnacc1, vnacc1);
153 const uint32x4_t vnacc1x0123 = vcombine_u32(vnacc1, vnacc1);
154 int32x4_t vacc1x0123 = vreinterpretq_s32_u32(vsubq_u32(vpacc1x0123, vnacc1x0123));
155 int32x4_t vacc1x4567 = vreinterpretq_s32_u32(vsubq_u32(vpacc1x4567, vnacc1x0123));
156 int32x4_t vacc1x89AB = vreinterpretq_s32_u32(vsubq_u32(vpacc1x89AB, vnacc1x0123));
157 int32x4_t vacc1xCDEF = vreinterpretq_s32_u32(vsubq_u32(vpacc1xCDEF, vnacc1x0123));
158
159 const int32x4_t vright_pre_shift = vld1q_dup_s32(¶ms->rndnu_neon.right_pre_shift);
160 const int32x4_t vmultiplier = vld1q_dup_s32(¶ms->rndnu_neon.multiplier);
161 const int32x4_t vright_post_shift = vld1q_dup_s32(¶ms->rndnu_neon.right_post_shift);
162
163 vacc0x0123 = vshlq_s32(vacc0x0123, vright_pre_shift);
164 vacc0x4567 = vshlq_s32(vacc0x4567, vright_pre_shift);
165 vacc0x89AB = vshlq_s32(vacc0x89AB, vright_pre_shift);
166 vacc0xCDEF = vshlq_s32(vacc0xCDEF, vright_pre_shift);
167 vacc1x0123 = vshlq_s32(vacc1x0123, vright_pre_shift);
168 vacc1x4567 = vshlq_s32(vacc1x4567, vright_pre_shift);
169 vacc1x89AB = vshlq_s32(vacc1x89AB, vright_pre_shift);
170 vacc1xCDEF = vshlq_s32(vacc1xCDEF, vright_pre_shift);
171
172 vacc0x0123 = vqdmulhq_s32(vacc0x0123, vmultiplier);
173 vacc0x4567 = vqdmulhq_s32(vacc0x4567, vmultiplier);
174 vacc0x89AB = vqdmulhq_s32(vacc0x89AB, vmultiplier);
175 vacc0xCDEF = vqdmulhq_s32(vacc0xCDEF, vmultiplier);
176 vacc1x0123 = vqdmulhq_s32(vacc1x0123, vmultiplier);
177 vacc1x4567 = vqdmulhq_s32(vacc1x4567, vmultiplier);
178 vacc1x89AB = vqdmulhq_s32(vacc1x89AB, vmultiplier);
179 vacc1xCDEF = vqdmulhq_s32(vacc1xCDEF, vmultiplier);
180
181 vacc0x0123 = vrshlq_s32(vacc0x0123, vright_post_shift);
182 vacc0x4567 = vrshlq_s32(vacc0x4567, vright_post_shift);
183 vacc0x89AB = vrshlq_s32(vacc0x89AB, vright_post_shift);
184 vacc0xCDEF = vrshlq_s32(vacc0xCDEF, vright_post_shift);
185 vacc1x0123 = vrshlq_s32(vacc1x0123, vright_post_shift);
186 vacc1x4567 = vrshlq_s32(vacc1x4567, vright_post_shift);
187 vacc1x89AB = vrshlq_s32(vacc1x89AB, vright_post_shift);
188 vacc1xCDEF = vrshlq_s32(vacc1xCDEF, vright_post_shift);
189
190 const int16x8_t voutput_zero_point = vld1q_dup_s16(¶ms->rndnu_neon.output_zero_point);
191 #if XNN_ARCH_ARM64
192 const int16x8_t vacc0x01234567 = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc0x0123), vacc0x4567), voutput_zero_point);
193 const int16x8_t vacc0x89ABCDEF = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc0x89AB), vacc0xCDEF), voutput_zero_point);
194 const int16x8_t vacc1x01234567 = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc1x0123), vacc1x4567), voutput_zero_point);
195 const int16x8_t vacc1x89ABCDEF = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc1x89AB), vacc1xCDEF), voutput_zero_point);
196
197 uint8x16_t vout0x0123456789ABCDEF = vqmovun_high_s16(vqmovun_s16(vacc0x01234567), vacc0x89ABCDEF);
198 uint8x16_t vout1x0123456789ABCDEF = vqmovun_high_s16(vqmovun_s16(vacc1x01234567), vacc1x89ABCDEF);
199 #else
200 const int16x8_t vacc0x01234567 = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc0x0123), vqmovn_s32(vacc0x4567)), voutput_zero_point);
201 const int16x8_t vacc0x89ABCDEF = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc0x89AB), vqmovn_s32(vacc0xCDEF)), voutput_zero_point);
202 const int16x8_t vacc1x01234567 = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc1x0123), vqmovn_s32(vacc1x4567)), voutput_zero_point);
203 const int16x8_t vacc1x89ABCDEF = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc1x89AB), vqmovn_s32(vacc1xCDEF)), voutput_zero_point);
204
205 uint8x16_t vout0x0123456789ABCDEF = vcombine_u8(vqmovun_s16(vacc0x01234567), vqmovun_s16(vacc0x89ABCDEF));
206 uint8x16_t vout1x0123456789ABCDEF = vcombine_u8(vqmovun_s16(vacc1x01234567), vqmovun_s16(vacc1x89ABCDEF));
207 #endif
208 const uint8x16_t voutput_min = vld1q_dup_u8(¶ms->rndnu_neon.output_min);
209 const uint8x16_t voutput_max = vld1q_dup_u8(¶ms->rndnu_neon.output_max);
210
211 vout0x0123456789ABCDEF = vmaxq_u8(vout0x0123456789ABCDEF, voutput_min);
212 vout1x0123456789ABCDEF = vmaxq_u8(vout1x0123456789ABCDEF, voutput_min);
213
214 vout0x0123456789ABCDEF = vminq_u8(vout0x0123456789ABCDEF, voutput_max);
215 vout1x0123456789ABCDEF = vminq_u8(vout1x0123456789ABCDEF, voutput_max);
216
217 if (nc >= 16) {
218 vst1q_u8(c1 + 0, vout1x0123456789ABCDEF);
219 vst1q_u8(c0 + 0, vout0x0123456789ABCDEF);
220
221 c1 = (uint8_t*) ((uintptr_t) c1 + cn_stride);
222 c0 = (uint8_t*) ((uintptr_t) c0 + cn_stride);
223
224 a = (const uint8_t**restrict) ((uintptr_t) a - ks);
225
226 nc -= 16;
227 } else {
228 uint8x16_t vout0x01234567_1x01234567 = vcombine_u8(vget_low_u8(vout0x0123456789ABCDEF), vget_low_u8(vout1x0123456789ABCDEF));
229 if (nc & 8) {
230 vst1_u8(c1, vget_high_u8(vout0x01234567_1x01234567)); c1 += 8;
231 vst1_u8(c0, vget_low_u8(vout0x01234567_1x01234567)); c0 += 8;
232 vout0x01234567_1x01234567 = vcombine_u8(vget_high_u8(vout0x0123456789ABCDEF), vget_high_u8(vout1x0123456789ABCDEF));
233 }
234 if (nc & 4) {
235 vst1q_lane_u32((void*) c1, vreinterpretq_u32_u8(vout0x01234567_1x01234567), 2); c1 += 4;
236 vst1q_lane_u32((void*) c0, vreinterpretq_u32_u8(vout0x01234567_1x01234567), 0); c0 += 4;
237 vout0x01234567_1x01234567 = vextq_u8(vout0x01234567_1x01234567, vout0x01234567_1x01234567, 4);
238 }
239 if (nc & 2) {
240 vst1q_lane_u16((void*) c1, vreinterpretq_u16_u8(vout0x01234567_1x01234567), 4); c1 += 2;
241 vst1q_lane_u16((void*) c0, vreinterpretq_u16_u8(vout0x01234567_1x01234567), 0); c0 += 2;
242 vout0x01234567_1x01234567 = vextq_u8(vout0x01234567_1x01234567, vout0x01234567_1x01234567, 2);
243 }
244 if (nc & 1) {
245 vst1q_lane_u8(c1, vout0x01234567_1x01234567, 8);
246 vst1q_lane_u8(c0, vout0x01234567_1x01234567, 0);
247 }
248
249 nc = 0;
250 }
251 } while (nc != 0);
252 }
253