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_4x8c4__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_4x8c4__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 <= 4);
34 assert(nc != 0);
35 assert(kc != 0);
36 assert(ks != 0);
37 assert(ks % (4 * 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 uint8_t* c2 = (uint8_t*) ((uintptr_t) c1 + cm_stride);
50 if XNN_UNPREDICTABLE(mr <= 2) {
51 c2 = c1;
52 }
53 uint8_t* c3 = (uint8_t*) ((uintptr_t) c2 + cm_stride);
54 if XNN_UNPREDICTABLE(mr != 4) {
55 c3 = c2;
56 }
57
58 const uint8x8_t va_zero_point = vld1_dup_u8(¶ms->rndnu_neon.kernel_zero_point[0]);
59
60 do {
61 // Initialize accumulators with bias. 8 bias values are loaded from the
62 // weight matrix, at the start of the group of 8 columns.
63 uint32x4_t vpacc0x0123 = vld1q_u32(w); w = (const void*) ((const uint32_t*) w + 4);
64 uint32x4_t vpacc0x4567 = vld1q_u32(w); w = (const void*) ((const uint32_t*) w + 4);
65 uint32x4_t vpacc1x0123 = vpacc0x0123;
66 uint32x4_t vpacc1x4567 = vpacc0x4567;
67 uint32x4_t vpacc2x0123 = vpacc0x0123;
68 uint32x4_t vpacc2x4567 = vpacc0x4567;
69 uint32x4_t vpacc3x0123 = vpacc0x0123;
70 uint32x4_t vpacc3x4567 = vpacc0x4567;
71 uint32x2_t vnacc0 = vmov_n_u32(0);
72 uint32x2_t vnacc1 = vmov_n_u32(0);
73 uint32x2_t vnacc2 = vmov_n_u32(0);
74 uint32x2_t vnacc3 = vmov_n_u32(0);
75
76 size_t p = ks;
77 do {
78 const uint8_t* restrict a0 = a[0];
79 if XNN_UNPREDICTABLE(a0 != zero) {
80 a0 = (const uint8_t*) ((uintptr_t) a0 + a_offset);
81 }
82 const uint8_t* restrict a1 = a[1];
83 if XNN_UNPREDICTABLE(a1 != zero) {
84 a1 = (const uint8_t*) ((uintptr_t) a1 + a_offset);
85 }
86 const uint8_t* restrict a2 = a[2];
87 if XNN_UNPREDICTABLE(a2 != zero) {
88 a2 = (const uint8_t*) ((uintptr_t) a2 + a_offset);
89 }
90 const uint8_t* restrict a3 = a[3];
91 if XNN_UNPREDICTABLE(a3 != zero) {
92 a3 = (const uint8_t*) ((uintptr_t) a3 + a_offset);
93 }
94 a += 4;
95
96 // Inner accumulation loop along the 8 columns.
97 size_t k = kc;
98 // 2x partial unrolled loop to load 8 bytes at a time.
99 while (k >= 8 * sizeof(uint8_t)) {
100 // Load a 4x8 block of activations.
101 const uint8x8_t va0x01234567 = vld1_u8(a0); a0 += 8;
102 const uint8x8_t va1x01234567 = vld1_u8(a1); a1 += 8;
103 const uint8x8_t va2x01234567 = vld1_u8(a2); a2 += 8;
104 const uint8x8_t va3x01234567 = vld1_u8(a3); a3 += 8;
105
106 // Load a 8x8 block of weights.
107 const uint8x16_t vb0123x0123 = vld1q_u8(w); w = (const void*) ((const uint8_t*) w + 16);
108 const uint8x16_t vb0123x4567 = vld1q_u8(w); w = (const void*) ((const uint8_t*) w + 16);
109 const uint8x16_t vb4567x0123 = vld1q_u8(w); w = (const void*) ((const uint8_t*) w + 16);
110 const uint8x16_t vb4567x4567 = vld1q_u8(w); w = (const void*) ((const uint8_t*) w + 16);
111
112 // Multiply-accumulate: 4x8 * 8x8 --> 4x8.
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 vpacc0x0123 = vdotq_lane_u32(vpacc0x0123, vb4567x0123, va0x01234567, 1);
117 vpacc0x4567 = vdotq_lane_u32(vpacc0x4567, vb4567x4567, va0x01234567, 1);
118 vnacc1 = vdot_u32(vnacc1, va_zero_point, va1x01234567);
119 vpacc1x0123 = vdotq_lane_u32(vpacc1x0123, vb0123x0123, va1x01234567, 0);
120 vpacc1x4567 = vdotq_lane_u32(vpacc1x4567, vb0123x4567, va1x01234567, 0);
121 vpacc1x0123 = vdotq_lane_u32(vpacc1x0123, vb4567x0123, va1x01234567, 1);
122 vpacc1x4567 = vdotq_lane_u32(vpacc1x4567, vb4567x4567, va1x01234567, 1);
123 vnacc2 = vdot_u32(vnacc2, va_zero_point, va2x01234567);
124 vpacc2x0123 = vdotq_lane_u32(vpacc2x0123, vb0123x0123, va2x01234567, 0);
125 vpacc2x4567 = vdotq_lane_u32(vpacc2x4567, vb0123x4567, va2x01234567, 0);
126 vpacc2x0123 = vdotq_lane_u32(vpacc2x0123, vb4567x0123, va2x01234567, 1);
127 vpacc2x4567 = vdotq_lane_u32(vpacc2x4567, vb4567x4567, va2x01234567, 1);
128 vnacc3 = vdot_u32(vnacc3, va_zero_point, va3x01234567);
129 vpacc3x0123 = vdotq_lane_u32(vpacc3x0123, vb0123x0123, va3x01234567, 0);
130 vpacc3x4567 = vdotq_lane_u32(vpacc3x4567, vb0123x4567, va3x01234567, 0);
131 vpacc3x0123 = vdotq_lane_u32(vpacc3x0123, vb4567x0123, va3x01234567, 1);
132 vpacc3x4567 = vdotq_lane_u32(vpacc3x4567, vb4567x4567, va3x01234567, 1);
133
134 k -= 8 * sizeof(uint8_t);
135 }
136 // Handle up to 4 final positions of `k`
137 if XNN_UNLIKELY(k != 0) {
138 // Load a 4x4 block of activations.
139 const uint8x8_t va0x01234567 = vreinterpret_u8_u32(vld1_lane_u32((const void*) a0, vmov_n_u32(0), 0)); a0 += 4;
140 const uint8x8_t va1x01234567 = vreinterpret_u8_u32(vld1_lane_u32((const void*) a1, vmov_n_u32(0), 0)); a1 += 4;
141 const uint8x8_t va2x01234567 = vreinterpret_u8_u32(vld1_lane_u32((const void*) a2, vmov_n_u32(0), 0)); a2 += 4;
142 const uint8x8_t va3x01234567 = vreinterpret_u8_u32(vld1_lane_u32((const void*) a3, vmov_n_u32(0), 0)); a3 += 4;
143
144 // Load a 4x8 block of weights.
145 const uint8x16_t vb0123x0123 = vld1q_u8(w); w = (const void*) ((const uint8_t*) w + 16);
146 const uint8x16_t vb0123x4567 = vld1q_u8(w); w = (const void*) ((const uint8_t*) w + 16);
147
148 // Multiply-accumulate: 4x4 * 4x8 --> 4x8.
149 vnacc0 = vdot_u32(vnacc0, va_zero_point, va0x01234567);
150 vpacc0x0123 = vdotq_lane_u32(vpacc0x0123, vb0123x0123, va0x01234567, 0);
151 vpacc0x4567 = vdotq_lane_u32(vpacc0x4567, vb0123x4567, va0x01234567, 0);
152 vnacc1 = vdot_u32(vnacc1, va_zero_point, va1x01234567);
153 vpacc1x0123 = vdotq_lane_u32(vpacc1x0123, vb0123x0123, va1x01234567, 0);
154 vpacc1x4567 = vdotq_lane_u32(vpacc1x4567, vb0123x4567, va1x01234567, 0);
155 vnacc2 = vdot_u32(vnacc2, va_zero_point, va2x01234567);
156 vpacc2x0123 = vdotq_lane_u32(vpacc2x0123, vb0123x0123, va2x01234567, 0);
157 vpacc2x4567 = vdotq_lane_u32(vpacc2x4567, vb0123x4567, va2x01234567, 0);
158 vnacc3 = vdot_u32(vnacc3, va_zero_point, va3x01234567);
159 vpacc3x0123 = vdotq_lane_u32(vpacc3x0123, vb0123x0123, va3x01234567, 0);
160 vpacc3x4567 = vdotq_lane_u32(vpacc3x4567, vb0123x4567, va3x01234567, 0);
161 }
162 p -= 4 * sizeof(void*);
163 } while (p != 0);
164
165 // Subtract zero point from accumulators.
166 vnacc0 = vpadd_u32(vnacc0, vnacc0);
167 const uint32x4_t vnacc0x0123 = vcombine_u32(vnacc0, vnacc0);
168 int32x4_t vacc0x0123 = vreinterpretq_s32_u32(vsubq_u32(vpacc0x0123, vnacc0x0123));
169 int32x4_t vacc0x4567 = vreinterpretq_s32_u32(vsubq_u32(vpacc0x4567, vnacc0x0123));
170 vnacc1 = vpadd_u32(vnacc1, vnacc1);
171 const uint32x4_t vnacc1x0123 = vcombine_u32(vnacc1, vnacc1);
172 int32x4_t vacc1x0123 = vreinterpretq_s32_u32(vsubq_u32(vpacc1x0123, vnacc1x0123));
173 int32x4_t vacc1x4567 = vreinterpretq_s32_u32(vsubq_u32(vpacc1x4567, vnacc1x0123));
174 vnacc2 = vpadd_u32(vnacc2, vnacc2);
175 const uint32x4_t vnacc2x0123 = vcombine_u32(vnacc2, vnacc2);
176 int32x4_t vacc2x0123 = vreinterpretq_s32_u32(vsubq_u32(vpacc2x0123, vnacc2x0123));
177 int32x4_t vacc2x4567 = vreinterpretq_s32_u32(vsubq_u32(vpacc2x4567, vnacc2x0123));
178 vnacc3 = vpadd_u32(vnacc3, vnacc3);
179 const uint32x4_t vnacc3x0123 = vcombine_u32(vnacc3, vnacc3);
180 int32x4_t vacc3x0123 = vreinterpretq_s32_u32(vsubq_u32(vpacc3x0123, vnacc3x0123));
181 int32x4_t vacc3x4567 = vreinterpretq_s32_u32(vsubq_u32(vpacc3x4567, vnacc3x0123));
182
183 const int32x4_t vright_pre_shift = vld1q_dup_s32(¶ms->rndnu_neon.right_pre_shift);
184 const int32x4_t vmultiplier = vld1q_dup_s32(¶ms->rndnu_neon.multiplier);
185 const int32x4_t vright_post_shift = vld1q_dup_s32(¶ms->rndnu_neon.right_post_shift);
186
187 vacc0x0123 = vshlq_s32(vacc0x0123, vright_pre_shift);
188 vacc0x4567 = vshlq_s32(vacc0x4567, vright_pre_shift);
189 vacc1x0123 = vshlq_s32(vacc1x0123, vright_pre_shift);
190 vacc1x4567 = vshlq_s32(vacc1x4567, vright_pre_shift);
191 vacc2x0123 = vshlq_s32(vacc2x0123, vright_pre_shift);
192 vacc2x4567 = vshlq_s32(vacc2x4567, vright_pre_shift);
193 vacc3x0123 = vshlq_s32(vacc3x0123, vright_pre_shift);
194 vacc3x4567 = vshlq_s32(vacc3x4567, vright_pre_shift);
195
196 vacc0x0123 = vqdmulhq_s32(vacc0x0123, vmultiplier);
197 vacc0x4567 = vqdmulhq_s32(vacc0x4567, vmultiplier);
198 vacc1x0123 = vqdmulhq_s32(vacc1x0123, vmultiplier);
199 vacc1x4567 = vqdmulhq_s32(vacc1x4567, vmultiplier);
200 vacc2x0123 = vqdmulhq_s32(vacc2x0123, vmultiplier);
201 vacc2x4567 = vqdmulhq_s32(vacc2x4567, vmultiplier);
202 vacc3x0123 = vqdmulhq_s32(vacc3x0123, vmultiplier);
203 vacc3x4567 = vqdmulhq_s32(vacc3x4567, vmultiplier);
204
205 vacc0x0123 = vrshlq_s32(vacc0x0123, vright_post_shift);
206 vacc0x4567 = vrshlq_s32(vacc0x4567, vright_post_shift);
207 vacc1x0123 = vrshlq_s32(vacc1x0123, vright_post_shift);
208 vacc1x4567 = vrshlq_s32(vacc1x4567, vright_post_shift);
209 vacc2x0123 = vrshlq_s32(vacc2x0123, vright_post_shift);
210 vacc2x4567 = vrshlq_s32(vacc2x4567, vright_post_shift);
211 vacc3x0123 = vrshlq_s32(vacc3x0123, vright_post_shift);
212 vacc3x4567 = vrshlq_s32(vacc3x4567, vright_post_shift);
213
214 const int16x8_t voutput_zero_point = vld1q_dup_s16(¶ms->rndnu_neon.output_zero_point);
215 #if XNN_ARCH_ARM64
216 const int16x8_t vacc0x01234567 = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc0x0123), vacc0x4567), voutput_zero_point);
217 const int16x8_t vacc1x01234567 = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc1x0123), vacc1x4567), voutput_zero_point);
218 const int16x8_t vacc2x01234567 = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc2x0123), vacc2x4567), voutput_zero_point);
219 const int16x8_t vacc3x01234567 = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc3x0123), vacc3x4567), voutput_zero_point);
220
221 uint8x16_t vout0x01234567_1x01234567 = vqmovun_high_s16(vqmovun_s16(vacc0x01234567), vacc1x01234567);
222 uint8x16_t vout2x01234567_3x01234567 = vqmovun_high_s16(vqmovun_s16(vacc2x01234567), vacc3x01234567);
223 #else
224 const int16x8_t vacc0x01234567 = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc0x0123), vqmovn_s32(vacc0x4567)), voutput_zero_point);
225 const int16x8_t vacc1x01234567 = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc1x0123), vqmovn_s32(vacc1x4567)), voutput_zero_point);
226 const int16x8_t vacc2x01234567 = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc2x0123), vqmovn_s32(vacc2x4567)), voutput_zero_point);
227 const int16x8_t vacc3x01234567 = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc3x0123), vqmovn_s32(vacc3x4567)), voutput_zero_point);
228
229 uint8x16_t vout0x01234567_1x01234567 = vcombine_u8(vqmovun_s16(vacc0x01234567), vqmovun_s16(vacc1x01234567));
230 uint8x16_t vout2x01234567_3x01234567 = vcombine_u8(vqmovun_s16(vacc2x01234567), vqmovun_s16(vacc3x01234567));
231 #endif
232 const uint8x16_t voutput_min = vld1q_dup_u8(¶ms->rndnu_neon.output_min);
233 const uint8x16_t voutput_max = vld1q_dup_u8(¶ms->rndnu_neon.output_max);
234
235 vout0x01234567_1x01234567 = vmaxq_u8(vout0x01234567_1x01234567, voutput_min);
236 vout2x01234567_3x01234567 = vmaxq_u8(vout2x01234567_3x01234567, voutput_min);
237
238 vout0x01234567_1x01234567 = vminq_u8(vout0x01234567_1x01234567, voutput_max);
239 vout2x01234567_3x01234567 = vminq_u8(vout2x01234567_3x01234567, voutput_max);
240
241 if (nc >= 8) {
242 vst1_u8(c3 + 0, vget_high_u8(vout2x01234567_3x01234567));
243 vst1_u8(c2 + 0, vget_low_u8(vout2x01234567_3x01234567));
244 vst1_u8(c1 + 0, vget_high_u8(vout0x01234567_1x01234567));
245 vst1_u8(c0 + 0, vget_low_u8(vout0x01234567_1x01234567));
246
247 c3 = (uint8_t*) ((uintptr_t) c3 + cn_stride);
248 c2 = (uint8_t*) ((uintptr_t) c2 + cn_stride);
249 c1 = (uint8_t*) ((uintptr_t) c1 + cn_stride);
250 c0 = (uint8_t*) ((uintptr_t) c0 + cn_stride);
251
252 a = (const uint8_t**restrict) ((uintptr_t) a - ks);
253
254 nc -= 8;
255 } else {
256 if (nc & 4) {
257 vst1q_lane_u32((void*) c3, vreinterpretq_u32_u8(vout2x01234567_3x01234567), 2); c3 += 4;
258 vst1q_lane_u32((void*) c2, vreinterpretq_u32_u8(vout2x01234567_3x01234567), 0); c2 += 4;
259 vst1q_lane_u32((void*) c1, vreinterpretq_u32_u8(vout0x01234567_1x01234567), 2); c1 += 4;
260 vst1q_lane_u32((void*) c0, vreinterpretq_u32_u8(vout0x01234567_1x01234567), 0); c0 += 4;
261 vout2x01234567_3x01234567 = vextq_u8(vout2x01234567_3x01234567, vout2x01234567_3x01234567, 4);
262 vout0x01234567_1x01234567 = vextq_u8(vout0x01234567_1x01234567, vout0x01234567_1x01234567, 4);
263 }
264 if (nc & 2) {
265 vst1q_lane_u16((void*) c3, vreinterpretq_u16_u8(vout2x01234567_3x01234567), 4); c3 += 2;
266 vst1q_lane_u16((void*) c2, vreinterpretq_u16_u8(vout2x01234567_3x01234567), 0); c2 += 2;
267 vst1q_lane_u16((void*) c1, vreinterpretq_u16_u8(vout0x01234567_1x01234567), 4); c1 += 2;
268 vst1q_lane_u16((void*) c0, vreinterpretq_u16_u8(vout0x01234567_1x01234567), 0); c0 += 2;
269 vout2x01234567_3x01234567 = vextq_u8(vout2x01234567_3x01234567, vout2x01234567_3x01234567, 2);
270 vout0x01234567_1x01234567 = vextq_u8(vout0x01234567_1x01234567, vout0x01234567_1x01234567, 2);
271 }
272 if (nc & 1) {
273 vst1q_lane_u8(c3, vout2x01234567_3x01234567, 8);
274 vst1q_lane_u8(c2, vout2x01234567_3x01234567, 0);
275 vst1q_lane_u8(c1, vout0x01234567_1x01234567, 8);
276 vst1q_lane_u8(c0, vout0x01234567_1x01234567, 0);
277 }
278
279 nc = 0;
280 }
281 } while (nc != 0);
282 }
283