1 // Auto-generated file. Do not edit!
2 // Template: src/qu8-gemm/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/gemm.h>
15 #include <xnnpack/math.h>
16
17
xnn_qu8_gemm_minmax_rndnu_ukernel_4x8c4__neondot(size_t mr,size_t nc,size_t kc,const uint8_t * restrict a,size_t a_stride,const void * restrict w,uint8_t * restrict c,size_t cm_stride,size_t cn_stride,const union xnn_qu8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS (1)])18 void xnn_qu8_gemm_minmax_rndnu_ukernel_4x8c4__neondot(
19 size_t mr,
20 size_t nc,
21 size_t kc,
22 const uint8_t* restrict a,
23 size_t a_stride,
24 const void* restrict w,
25 uint8_t* restrict c,
26 size_t cm_stride,
27 size_t cn_stride,
28 const union xnn_qu8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
29 {
30 assert(mr != 0);
31 assert(mr <= 4);
32 assert(nc != 0);
33 assert(kc != 0);
34 assert(kc % sizeof(uint8_t) == 0);
35 assert(a != NULL);
36 assert(w != NULL);
37 assert(c != NULL);
38
39 kc = round_up_po2(kc, 4 * sizeof(uint8_t));
40 const uint8_t* a0 = a;
41 uint8_t* c0 = c;
42 const uint8_t* a1 = (const uint8_t*) ((uintptr_t) a0 + a_stride);
43 uint8_t* c1 = (uint8_t*) ((uintptr_t) c0 + cm_stride);
44 if XNN_UNPREDICTABLE(mr < 2) {
45 a1 = a0;
46 c1 = c0;
47 }
48 const uint8_t* a2 = (const uint8_t*) ((uintptr_t) a1 + a_stride);
49 uint8_t* c2 = (uint8_t*) ((uintptr_t) c1 + cm_stride);
50 if XNN_UNPREDICTABLE(mr <= 2) {
51 a2 = a1;
52 c2 = c1;
53 }
54 const uint8_t* a3 = (const uint8_t*) ((uintptr_t) a2 + a_stride);
55 uint8_t* c3 = (uint8_t*) ((uintptr_t) c2 + cm_stride);
56 if XNN_UNPREDICTABLE(mr != 4) {
57 a3 = a2;
58 c3 = c2;
59 }
60
61 const uint8x8_t va_zero_point = vld1_dup_u8(¶ms->rndnu_neon.kernel_zero_point[0]);
62
63 // Loop over groups of 8 columns.
64 do {
65 // Initialize accumulators with bias. 8 bias values are loaded from the
66 // weight matrix, at the start of the group of 8 columns.
67 uint32x4_t vpacc0x0123 = vld1q_u32(w); w = (const void*) ((const uint32_t*) w + 4);
68 uint32x4_t vpacc0x4567 = vld1q_u32(w); w = (const void*) ((const uint32_t*) w + 4);
69 uint32x4_t vpacc1x0123 = vpacc0x0123;
70 uint32x4_t vpacc1x4567 = vpacc0x4567;
71 uint32x4_t vpacc2x0123 = vpacc0x0123;
72 uint32x4_t vpacc2x4567 = vpacc0x4567;
73 uint32x4_t vpacc3x0123 = vpacc0x0123;
74 uint32x4_t vpacc3x4567 = vpacc0x4567;
75 uint32x2_t vnacc0 = vmov_n_u32(0);
76 uint32x2_t vnacc1 = vmov_n_u32(0);
77 uint32x2_t vnacc2 = vmov_n_u32(0);
78 uint32x2_t vnacc3 = vmov_n_u32(0);
79
80 // Inner accumulation loop along the 8 columns.
81 size_t k = kc;
82 // 2x partial unrolled loop to load 8 bytes at a time.
83 while (k >= 8 * sizeof(uint8_t)) {
84 // Load a 4x8 block of activations.
85 const uint8x8_t va0x01234567 = vld1_u8(a0); a0 += 8;
86 const uint8x8_t va1x01234567 = vld1_u8(a1); a1 += 8;
87 const uint8x8_t va2x01234567 = vld1_u8(a2); a2 += 8;
88 const uint8x8_t va3x01234567 = vld1_u8(a3); a3 += 8;
89
90 // Load a 8x8 block of weights.
91 const uint8x16_t vb0123x0123 = vld1q_u8(w); w = (const void*) ((const uint8_t*) w + 16);
92 const uint8x16_t vb0123x4567 = vld1q_u8(w); w = (const void*) ((const uint8_t*) w + 16);
93 const uint8x16_t vb4567x0123 = vld1q_u8(w); w = (const void*) ((const uint8_t*) w + 16);
94 const uint8x16_t vb4567x4567 = vld1q_u8(w); w = (const void*) ((const uint8_t*) w + 16);
95
96 // Multiply-accumulate: 4x8 * 8x8 --> 4x8.
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 vpacc0x0123 = vdotq_lane_u32(vpacc0x0123, vb4567x0123, va0x01234567, 1);
101 vpacc0x4567 = vdotq_lane_u32(vpacc0x4567, vb4567x4567, va0x01234567, 1);
102 vnacc1 = vdot_u32(vnacc1, va_zero_point, va1x01234567);
103 vpacc1x0123 = vdotq_lane_u32(vpacc1x0123, vb0123x0123, va1x01234567, 0);
104 vpacc1x4567 = vdotq_lane_u32(vpacc1x4567, vb0123x4567, va1x01234567, 0);
105 vpacc1x0123 = vdotq_lane_u32(vpacc1x0123, vb4567x0123, va1x01234567, 1);
106 vpacc1x4567 = vdotq_lane_u32(vpacc1x4567, vb4567x4567, va1x01234567, 1);
107 vnacc2 = vdot_u32(vnacc2, va_zero_point, va2x01234567);
108 vpacc2x0123 = vdotq_lane_u32(vpacc2x0123, vb0123x0123, va2x01234567, 0);
109 vpacc2x4567 = vdotq_lane_u32(vpacc2x4567, vb0123x4567, va2x01234567, 0);
110 vpacc2x0123 = vdotq_lane_u32(vpacc2x0123, vb4567x0123, va2x01234567, 1);
111 vpacc2x4567 = vdotq_lane_u32(vpacc2x4567, vb4567x4567, va2x01234567, 1);
112 vnacc3 = vdot_u32(vnacc3, va_zero_point, va3x01234567);
113 vpacc3x0123 = vdotq_lane_u32(vpacc3x0123, vb0123x0123, va3x01234567, 0);
114 vpacc3x4567 = vdotq_lane_u32(vpacc3x4567, vb0123x4567, va3x01234567, 0);
115 vpacc3x0123 = vdotq_lane_u32(vpacc3x0123, vb4567x0123, va3x01234567, 1);
116 vpacc3x4567 = vdotq_lane_u32(vpacc3x4567, vb4567x4567, va3x01234567, 1);
117
118 k -= 8 * sizeof(uint8_t);
119 }
120 // Handle up to 4 final positions of `k`
121 if XNN_UNLIKELY(k != 0) {
122 // Load a 4x4 block of activations.
123 const uint8x8_t va0x01234567 = vreinterpret_u8_u32(vld1_lane_u32((const void*) a0, vmov_n_u32(0), 0)); a0 += 4;
124 const uint8x8_t va1x01234567 = vreinterpret_u8_u32(vld1_lane_u32((const void*) a1, vmov_n_u32(0), 0)); a1 += 4;
125 const uint8x8_t va2x01234567 = vreinterpret_u8_u32(vld1_lane_u32((const void*) a2, vmov_n_u32(0), 0)); a2 += 4;
126 const uint8x8_t va3x01234567 = vreinterpret_u8_u32(vld1_lane_u32((const void*) a3, vmov_n_u32(0), 0)); a3 += 4;
127
128 // Load a 4x8 block of weights.
129 const uint8x16_t vb0123x0123 = vld1q_u8(w); w = (const void*) ((const uint8_t*) w + 16);
130 const uint8x16_t vb0123x4567 = vld1q_u8(w); w = (const void*) ((const uint8_t*) w + 16);
131
132 // Multiply-accumulate: 4x4 * 4x8 --> 4x8.
133 vnacc0 = vdot_u32(vnacc0, va_zero_point, va0x01234567);
134 vpacc0x0123 = vdotq_lane_u32(vpacc0x0123, vb0123x0123, va0x01234567, 0);
135 vpacc0x4567 = vdotq_lane_u32(vpacc0x4567, vb0123x4567, 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 vnacc2 = vdot_u32(vnacc2, va_zero_point, va2x01234567);
140 vpacc2x0123 = vdotq_lane_u32(vpacc2x0123, vb0123x0123, va2x01234567, 0);
141 vpacc2x4567 = vdotq_lane_u32(vpacc2x4567, vb0123x4567, va2x01234567, 0);
142 vnacc3 = vdot_u32(vnacc3, va_zero_point, va3x01234567);
143 vpacc3x0123 = vdotq_lane_u32(vpacc3x0123, vb0123x0123, va3x01234567, 0);
144 vpacc3x4567 = vdotq_lane_u32(vpacc3x4567, vb0123x4567, va3x01234567, 0);
145 }
146
147 // Subtract zero point from accumulators.
148 vnacc0 = vpadd_u32(vnacc0, vnacc0);
149 const uint32x4_t vnacc0x0123 = vcombine_u32(vnacc0, vnacc0);
150 int32x4_t vacc0x0123 = vreinterpretq_s32_u32(vsubq_u32(vpacc0x0123, vnacc0x0123));
151 int32x4_t vacc0x4567 = vreinterpretq_s32_u32(vsubq_u32(vpacc0x4567, 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 vnacc2 = vpadd_u32(vnacc2, vnacc2);
157 const uint32x4_t vnacc2x0123 = vcombine_u32(vnacc2, vnacc2);
158 int32x4_t vacc2x0123 = vreinterpretq_s32_u32(vsubq_u32(vpacc2x0123, vnacc2x0123));
159 int32x4_t vacc2x4567 = vreinterpretq_s32_u32(vsubq_u32(vpacc2x4567, vnacc2x0123));
160 vnacc3 = vpadd_u32(vnacc3, vnacc3);
161 const uint32x4_t vnacc3x0123 = vcombine_u32(vnacc3, vnacc3);
162 int32x4_t vacc3x0123 = vreinterpretq_s32_u32(vsubq_u32(vpacc3x0123, vnacc3x0123));
163 int32x4_t vacc3x4567 = vreinterpretq_s32_u32(vsubq_u32(vpacc3x4567, vnacc3x0123));
164
165 const int32x4_t vright_pre_shift = vld1q_dup_s32(¶ms->rndnu_neon.right_pre_shift);
166 const int32x4_t vmultiplier = vld1q_dup_s32(¶ms->rndnu_neon.multiplier);
167 const int32x4_t vright_post_shift = vld1q_dup_s32(¶ms->rndnu_neon.right_post_shift);
168
169 vacc0x0123 = vshlq_s32(vacc0x0123, vright_pre_shift);
170 vacc0x4567 = vshlq_s32(vacc0x4567, vright_pre_shift);
171 vacc1x0123 = vshlq_s32(vacc1x0123, vright_pre_shift);
172 vacc1x4567 = vshlq_s32(vacc1x4567, vright_pre_shift);
173 vacc2x0123 = vshlq_s32(vacc2x0123, vright_pre_shift);
174 vacc2x4567 = vshlq_s32(vacc2x4567, vright_pre_shift);
175 vacc3x0123 = vshlq_s32(vacc3x0123, vright_pre_shift);
176 vacc3x4567 = vshlq_s32(vacc3x4567, vright_pre_shift);
177
178 vacc0x0123 = vqdmulhq_s32(vacc0x0123, vmultiplier);
179 vacc0x4567 = vqdmulhq_s32(vacc0x4567, vmultiplier);
180 vacc1x0123 = vqdmulhq_s32(vacc1x0123, vmultiplier);
181 vacc1x4567 = vqdmulhq_s32(vacc1x4567, vmultiplier);
182 vacc2x0123 = vqdmulhq_s32(vacc2x0123, vmultiplier);
183 vacc2x4567 = vqdmulhq_s32(vacc2x4567, vmultiplier);
184 vacc3x0123 = vqdmulhq_s32(vacc3x0123, vmultiplier);
185 vacc3x4567 = vqdmulhq_s32(vacc3x4567, vmultiplier);
186
187 vacc0x0123 = vrshlq_s32(vacc0x0123, vright_post_shift);
188 vacc0x4567 = vrshlq_s32(vacc0x4567, vright_post_shift);
189 vacc1x0123 = vrshlq_s32(vacc1x0123, vright_post_shift);
190 vacc1x4567 = vrshlq_s32(vacc1x4567, vright_post_shift);
191 vacc2x0123 = vrshlq_s32(vacc2x0123, vright_post_shift);
192 vacc2x4567 = vrshlq_s32(vacc2x4567, vright_post_shift);
193 vacc3x0123 = vrshlq_s32(vacc3x0123, vright_post_shift);
194 vacc3x4567 = vrshlq_s32(vacc3x4567, vright_post_shift);
195
196 const int16x8_t voutput_zero_point = vld1q_dup_s16(¶ms->rndnu_neon.output_zero_point);
197 #if XNN_ARCH_ARM64
198 const int16x8_t vacc0x01234567 = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc0x0123), vacc0x4567), voutput_zero_point);
199 const int16x8_t vacc1x01234567 = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc1x0123), vacc1x4567), voutput_zero_point);
200 const int16x8_t vacc2x01234567 = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc2x0123), vacc2x4567), voutput_zero_point);
201 const int16x8_t vacc3x01234567 = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc3x0123), vacc3x4567), voutput_zero_point);
202
203 uint8x16_t vout0x01234567_1x01234567 = vqmovun_high_s16(vqmovun_s16(vacc0x01234567), vacc1x01234567);
204 uint8x16_t vout2x01234567_3x01234567 = vqmovun_high_s16(vqmovun_s16(vacc2x01234567), vacc3x01234567);
205 #else
206 const int16x8_t vacc0x01234567 = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc0x0123), vqmovn_s32(vacc0x4567)), voutput_zero_point);
207 const int16x8_t vacc1x01234567 = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc1x0123), vqmovn_s32(vacc1x4567)), voutput_zero_point);
208 const int16x8_t vacc2x01234567 = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc2x0123), vqmovn_s32(vacc2x4567)), voutput_zero_point);
209 const int16x8_t vacc3x01234567 = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc3x0123), vqmovn_s32(vacc3x4567)), voutput_zero_point);
210
211 uint8x16_t vout0x01234567_1x01234567 = vcombine_u8(vqmovun_s16(vacc0x01234567), vqmovun_s16(vacc1x01234567));
212 uint8x16_t vout2x01234567_3x01234567 = vcombine_u8(vqmovun_s16(vacc2x01234567), vqmovun_s16(vacc3x01234567));
213 #endif
214 const uint8x16_t voutput_min = vld1q_dup_u8(¶ms->rndnu_neon.output_min);
215 const uint8x16_t voutput_max = vld1q_dup_u8(¶ms->rndnu_neon.output_max);
216
217 vout0x01234567_1x01234567 = vmaxq_u8(vout0x01234567_1x01234567, voutput_min);
218 vout2x01234567_3x01234567 = vmaxq_u8(vout2x01234567_3x01234567, voutput_min);
219
220 vout0x01234567_1x01234567 = vminq_u8(vout0x01234567_1x01234567, voutput_max);
221 vout2x01234567_3x01234567 = vminq_u8(vout2x01234567_3x01234567, voutput_max);
222
223 if (nc >= 8) {
224 vst1_u8(c0 + 0, vget_low_u8(vout0x01234567_1x01234567));
225 vst1_u8(c1 + 0, vget_high_u8(vout0x01234567_1x01234567));
226 vst1_u8(c2 + 0, vget_low_u8(vout2x01234567_3x01234567));
227 vst1_u8(c3 + 0, vget_high_u8(vout2x01234567_3x01234567));
228
229 c0 = (uint8_t*) ((uintptr_t) c0 + cn_stride);
230 c1 = (uint8_t*) ((uintptr_t) c1 + cn_stride);
231 c2 = (uint8_t*) ((uintptr_t) c2 + cn_stride);
232 c3 = (uint8_t*) ((uintptr_t) c3 + cn_stride);
233
234 a0 = (const uint8_t*) ((uintptr_t) a0 - kc);
235 a1 = (const uint8_t*) ((uintptr_t) a1 - kc);
236 a2 = (const uint8_t*) ((uintptr_t) a2 - kc);
237 a3 = (const uint8_t*) ((uintptr_t) a3 - kc);
238
239 nc -= 8;
240 } else {
241 if (nc & 4) {
242 vst1q_lane_u32((void*) c0, vreinterpretq_u32_u8(vout0x01234567_1x01234567), 0); c0 += 4;
243 vst1q_lane_u32((void*) c1, vreinterpretq_u32_u8(vout0x01234567_1x01234567), 2); c1 += 4;
244 vst1q_lane_u32((void*) c2, vreinterpretq_u32_u8(vout2x01234567_3x01234567), 0); c2 += 4;
245 vst1q_lane_u32((void*) c3, vreinterpretq_u32_u8(vout2x01234567_3x01234567), 2); c3 += 4;
246 vout0x01234567_1x01234567 = vextq_u8(vout0x01234567_1x01234567, vout0x01234567_1x01234567, 4);
247 vout2x01234567_3x01234567 = vextq_u8(vout2x01234567_3x01234567, vout2x01234567_3x01234567, 4);
248 }
249 if (nc & 2) {
250 vst1q_lane_u16((void*) c0, vreinterpretq_u16_u8(vout0x01234567_1x01234567), 0); c0 += 2;
251 vst1q_lane_u16((void*) c1, vreinterpretq_u16_u8(vout0x01234567_1x01234567), 4); c1 += 2;
252 vst1q_lane_u16((void*) c2, vreinterpretq_u16_u8(vout2x01234567_3x01234567), 0); c2 += 2;
253 vst1q_lane_u16((void*) c3, vreinterpretq_u16_u8(vout2x01234567_3x01234567), 4); c3 += 2;
254 vout0x01234567_1x01234567 = vextq_u8(vout0x01234567_1x01234567, vout0x01234567_1x01234567, 2);
255 vout2x01234567_3x01234567 = vextq_u8(vout2x01234567_3x01234567, vout2x01234567_3x01234567, 2);
256 }
257 if (nc & 1) {
258 vst1q_lane_u8(c0, vout0x01234567_1x01234567, 0);
259 vst1q_lane_u8(c1, vout0x01234567_1x01234567, 8);
260 vst1q_lane_u8(c2, vout2x01234567_3x01234567, 0);
261 vst1q_lane_u8(c3, vout2x01234567_3x01234567, 8);
262 }
263
264 nc = 0;
265 }
266 } while (nc != 0);
267 }
268