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_1x32c4__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_1x32c4__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 <= 1);
34 assert(nc != 0);
35 assert(kc != 0);
36 assert(ks != 0);
37 assert(ks % (1 * 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
46 const uint8x8_t va_zero_point = vld1_dup_u8(¶ms->rndnu_neon.kernel_zero_point[0]);
47
48 do {
49 // Initialize accumulators with bias. 32 bias values are loaded from the
50 // weight matrix, at the start of the group of 32 columns.
51 uint32x4_t vpacc0x0123 = vld1q_u32(w); w = (const void*) ((const uint32_t*) w + 4);
52 uint32x4_t vpacc0x4567 = vld1q_u32(w); w = (const void*) ((const uint32_t*) w + 4);
53 uint32x4_t vpacc0x89AB = vld1q_u32(w); w = (const void*) ((const uint32_t*) w + 4);
54 uint32x4_t vpacc0xCDEF = vld1q_u32(w); w = (const void*) ((const uint32_t*) w + 4);
55 uint32x4_t vpacc0xGHIJ = vld1q_u32(w); w = (const void*) ((const uint32_t*) w + 4);
56 uint32x4_t vpacc0xKLMN = vld1q_u32(w); w = (const void*) ((const uint32_t*) w + 4);
57 uint32x4_t vpacc0xOPQR = vld1q_u32(w); w = (const void*) ((const uint32_t*) w + 4);
58 uint32x4_t vpacc0xSTUV = vld1q_u32(w); w = (const void*) ((const uint32_t*) w + 4);
59 uint32x2_t vnacc0 = vmov_n_u32(0);
60
61 size_t p = ks;
62 do {
63 const uint8_t* restrict a0 = a[0];
64 if XNN_UNPREDICTABLE(a0 != zero) {
65 a0 = (const uint8_t*) ((uintptr_t) a0 + a_offset);
66 }
67 a += 1;
68
69 // Inner accumulation loop along the 32 columns.
70 size_t k = kc;
71 // 2x partial unrolled loop to load 8 bytes at a time.
72 while (k >= 8 * sizeof(uint8_t)) {
73 // Load a 1x8 block of activations.
74 const uint8x8_t va0x01234567 = vld1_u8(a0); a0 += 8;
75
76 // Load a 8x32 block of weights.
77 const uint8x16_t vb0123x0123 = vld1q_u8(w); w = (const void*) ((const uint8_t*) w + 16);
78 const uint8x16_t vb0123x4567 = vld1q_u8(w); w = (const void*) ((const uint8_t*) w + 16);
79 const uint8x16_t vb0123x89AB = vld1q_u8(w); w = (const void*) ((const uint8_t*) w + 16);
80 const uint8x16_t vb0123xCDEF = vld1q_u8(w); w = (const void*) ((const uint8_t*) w + 16);
81 const uint8x16_t vb0123xGHIJ = vld1q_u8(w); w = (const void*) ((const uint8_t*) w + 16);
82 const uint8x16_t vb0123xKLMN = vld1q_u8(w); w = (const void*) ((const uint8_t*) w + 16);
83 const uint8x16_t vb0123xOPQR = vld1q_u8(w); w = (const void*) ((const uint8_t*) w + 16);
84 const uint8x16_t vb0123xSTUV = 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 const uint8x16_t vb4567x89AB = vld1q_u8(w); w = (const void*) ((const uint8_t*) w + 16);
88 const uint8x16_t vb4567xCDEF = vld1q_u8(w); w = (const void*) ((const uint8_t*) w + 16);
89 const uint8x16_t vb4567xGHIJ = vld1q_u8(w); w = (const void*) ((const uint8_t*) w + 16);
90 const uint8x16_t vb4567xKLMN = vld1q_u8(w); w = (const void*) ((const uint8_t*) w + 16);
91 const uint8x16_t vb4567xOPQR = vld1q_u8(w); w = (const void*) ((const uint8_t*) w + 16);
92 const uint8x16_t vb4567xSTUV = vld1q_u8(w); w = (const void*) ((const uint8_t*) w + 16);
93
94 // Multiply-accumulate: 1x8 * 8x32 --> 1x32.
95 vnacc0 = vdot_u32(vnacc0, va_zero_point, va0x01234567);
96 vpacc0x0123 = vdotq_lane_u32(vpacc0x0123, vb0123x0123, va0x01234567, 0);
97 vpacc0x4567 = vdotq_lane_u32(vpacc0x4567, vb0123x4567, va0x01234567, 0);
98 vpacc0x89AB = vdotq_lane_u32(vpacc0x89AB, vb0123x89AB, va0x01234567, 0);
99 vpacc0xCDEF = vdotq_lane_u32(vpacc0xCDEF, vb0123xCDEF, va0x01234567, 0);
100 vpacc0xGHIJ = vdotq_lane_u32(vpacc0xGHIJ, vb0123xGHIJ, va0x01234567, 0);
101 vpacc0xKLMN = vdotq_lane_u32(vpacc0xKLMN, vb0123xKLMN, va0x01234567, 0);
102 vpacc0xOPQR = vdotq_lane_u32(vpacc0xOPQR, vb0123xOPQR, va0x01234567, 0);
103 vpacc0xSTUV = vdotq_lane_u32(vpacc0xSTUV, vb0123xSTUV, va0x01234567, 0);
104 vpacc0x0123 = vdotq_lane_u32(vpacc0x0123, vb4567x0123, va0x01234567, 1);
105 vpacc0x4567 = vdotq_lane_u32(vpacc0x4567, vb4567x4567, va0x01234567, 1);
106 vpacc0x89AB = vdotq_lane_u32(vpacc0x89AB, vb4567x89AB, va0x01234567, 1);
107 vpacc0xCDEF = vdotq_lane_u32(vpacc0xCDEF, vb4567xCDEF, va0x01234567, 1);
108 vpacc0xGHIJ = vdotq_lane_u32(vpacc0xGHIJ, vb4567xGHIJ, va0x01234567, 1);
109 vpacc0xKLMN = vdotq_lane_u32(vpacc0xKLMN, vb4567xKLMN, va0x01234567, 1);
110 vpacc0xOPQR = vdotq_lane_u32(vpacc0xOPQR, vb4567xOPQR, va0x01234567, 1);
111 vpacc0xSTUV = vdotq_lane_u32(vpacc0xSTUV, vb4567xSTUV, va0x01234567, 1);
112
113 k -= 8 * sizeof(uint8_t);
114 }
115 // Handle up to 4 final positions of `k`
116 if XNN_UNLIKELY(k != 0) {
117 // Load a 1x4 block of activations.
118 const uint8x8_t va0x01234567 = vreinterpret_u8_u32(vld1_lane_u32((const void*) a0, vmov_n_u32(0), 0)); a0 += 4;
119
120 // Load a 4x32 block of weights.
121 const uint8x16_t vb0123x0123 = vld1q_u8(w); w = (const void*) ((const uint8_t*) w + 16);
122 const uint8x16_t vb0123x4567 = vld1q_u8(w); w = (const void*) ((const uint8_t*) w + 16);
123 const uint8x16_t vb0123x89AB = vld1q_u8(w); w = (const void*) ((const uint8_t*) w + 16);
124 const uint8x16_t vb0123xCDEF = vld1q_u8(w); w = (const void*) ((const uint8_t*) w + 16);
125 const uint8x16_t vb0123xGHIJ = vld1q_u8(w); w = (const void*) ((const uint8_t*) w + 16);
126 const uint8x16_t vb0123xKLMN = vld1q_u8(w); w = (const void*) ((const uint8_t*) w + 16);
127 const uint8x16_t vb0123xOPQR = vld1q_u8(w); w = (const void*) ((const uint8_t*) w + 16);
128 const uint8x16_t vb0123xSTUV = vld1q_u8(w); w = (const void*) ((const uint8_t*) w + 16);
129
130 // Multiply-accumulate: 1x4 * 4x32 --> 1x32.
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 vpacc0xGHIJ = vdotq_lane_u32(vpacc0xGHIJ, vb0123xGHIJ, va0x01234567, 0);
137 vpacc0xKLMN = vdotq_lane_u32(vpacc0xKLMN, vb0123xKLMN, va0x01234567, 0);
138 vpacc0xOPQR = vdotq_lane_u32(vpacc0xOPQR, vb0123xOPQR, va0x01234567, 0);
139 vpacc0xSTUV = vdotq_lane_u32(vpacc0xSTUV, vb0123xSTUV, va0x01234567, 0);
140 }
141 p -= 1 * sizeof(void*);
142 } while (p != 0);
143
144 // Subtract zero point from accumulators.
145 vnacc0 = vpadd_u32(vnacc0, vnacc0);
146 const uint32x4_t vnacc0x0123 = vcombine_u32(vnacc0, vnacc0);
147 int32x4_t vacc0x0123 = vreinterpretq_s32_u32(vsubq_u32(vpacc0x0123, vnacc0x0123));
148 int32x4_t vacc0x4567 = vreinterpretq_s32_u32(vsubq_u32(vpacc0x4567, vnacc0x0123));
149 int32x4_t vacc0x89AB = vreinterpretq_s32_u32(vsubq_u32(vpacc0x89AB, vnacc0x0123));
150 int32x4_t vacc0xCDEF = vreinterpretq_s32_u32(vsubq_u32(vpacc0xCDEF, vnacc0x0123));
151 int32x4_t vacc0xGHIJ = vreinterpretq_s32_u32(vsubq_u32(vpacc0xGHIJ, vnacc0x0123));
152 int32x4_t vacc0xKLMN = vreinterpretq_s32_u32(vsubq_u32(vpacc0xKLMN, vnacc0x0123));
153 int32x4_t vacc0xOPQR = vreinterpretq_s32_u32(vsubq_u32(vpacc0xOPQR, vnacc0x0123));
154 int32x4_t vacc0xSTUV = vreinterpretq_s32_u32(vsubq_u32(vpacc0xSTUV, vnacc0x0123));
155
156 const int32x4_t vright_pre_shift = vld1q_dup_s32(¶ms->rndnu_neon.right_pre_shift);
157 const int32x4_t vmultiplier = vld1q_dup_s32(¶ms->rndnu_neon.multiplier);
158 const int32x4_t vright_post_shift = vld1q_dup_s32(¶ms->rndnu_neon.right_post_shift);
159
160 vacc0x0123 = vshlq_s32(vacc0x0123, vright_pre_shift);
161 vacc0x4567 = vshlq_s32(vacc0x4567, vright_pre_shift);
162 vacc0x89AB = vshlq_s32(vacc0x89AB, vright_pre_shift);
163 vacc0xCDEF = vshlq_s32(vacc0xCDEF, vright_pre_shift);
164 vacc0xGHIJ = vshlq_s32(vacc0xGHIJ, vright_pre_shift);
165 vacc0xKLMN = vshlq_s32(vacc0xKLMN, vright_pre_shift);
166 vacc0xOPQR = vshlq_s32(vacc0xOPQR, vright_pre_shift);
167 vacc0xSTUV = vshlq_s32(vacc0xSTUV, vright_pre_shift);
168
169 vacc0x0123 = vqdmulhq_s32(vacc0x0123, vmultiplier);
170 vacc0x4567 = vqdmulhq_s32(vacc0x4567, vmultiplier);
171 vacc0x89AB = vqdmulhq_s32(vacc0x89AB, vmultiplier);
172 vacc0xCDEF = vqdmulhq_s32(vacc0xCDEF, vmultiplier);
173 vacc0xGHIJ = vqdmulhq_s32(vacc0xGHIJ, vmultiplier);
174 vacc0xKLMN = vqdmulhq_s32(vacc0xKLMN, vmultiplier);
175 vacc0xOPQR = vqdmulhq_s32(vacc0xOPQR, vmultiplier);
176 vacc0xSTUV = vqdmulhq_s32(vacc0xSTUV, vmultiplier);
177
178 vacc0x0123 = vrshlq_s32(vacc0x0123, vright_post_shift);
179 vacc0x4567 = vrshlq_s32(vacc0x4567, vright_post_shift);
180 vacc0x89AB = vrshlq_s32(vacc0x89AB, vright_post_shift);
181 vacc0xCDEF = vrshlq_s32(vacc0xCDEF, vright_post_shift);
182 vacc0xGHIJ = vrshlq_s32(vacc0xGHIJ, vright_post_shift);
183 vacc0xKLMN = vrshlq_s32(vacc0xKLMN, vright_post_shift);
184 vacc0xOPQR = vrshlq_s32(vacc0xOPQR, vright_post_shift);
185 vacc0xSTUV = vrshlq_s32(vacc0xSTUV, vright_post_shift);
186
187 const int16x8_t voutput_zero_point = vld1q_dup_s16(¶ms->rndnu_neon.output_zero_point);
188 #if XNN_ARCH_ARM64
189 const int16x8_t vacc0x01234567 = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc0x0123), vacc0x4567), voutput_zero_point);
190 const int16x8_t vacc0x89ABCDEF = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc0x89AB), vacc0xCDEF), voutput_zero_point);
191 const int16x8_t vacc0xGHIJKLMN = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc0xGHIJ), vacc0xKLMN), voutput_zero_point);
192 const int16x8_t vacc0xOPQRSTUV = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc0xOPQR), vacc0xSTUV), voutput_zero_point);
193
194 uint8x16_t vout0x0123456789ABCDEF = vqmovun_high_s16(vqmovun_s16(vacc0x01234567), vacc0x89ABCDEF);
195 uint8x16_t vout0xGHIJKLMNOPQRSTUV = vqmovun_high_s16(vqmovun_s16(vacc0xGHIJKLMN), vacc0xOPQRSTUV);
196 #else
197 const int16x8_t vacc0x01234567 = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc0x0123), vqmovn_s32(vacc0x4567)), voutput_zero_point);
198 const int16x8_t vacc0x89ABCDEF = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc0x89AB), vqmovn_s32(vacc0xCDEF)), voutput_zero_point);
199 const int16x8_t vacc0xGHIJKLMN = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc0xGHIJ), vqmovn_s32(vacc0xKLMN)), voutput_zero_point);
200 const int16x8_t vacc0xOPQRSTUV = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc0xOPQR), vqmovn_s32(vacc0xSTUV)), voutput_zero_point);
201
202 uint8x16_t vout0x0123456789ABCDEF = vcombine_u8(vqmovun_s16(vacc0x01234567), vqmovun_s16(vacc0x89ABCDEF));
203 uint8x16_t vout0xGHIJKLMNOPQRSTUV = vcombine_u8(vqmovun_s16(vacc0xGHIJKLMN), vqmovun_s16(vacc0xOPQRSTUV));
204 #endif
205 const uint8x16_t voutput_min = vld1q_dup_u8(¶ms->rndnu_neon.output_min);
206 const uint8x16_t voutput_max = vld1q_dup_u8(¶ms->rndnu_neon.output_max);
207
208 vout0x0123456789ABCDEF = vmaxq_u8(vout0x0123456789ABCDEF, voutput_min);
209 vout0xGHIJKLMNOPQRSTUV = vmaxq_u8(vout0xGHIJKLMNOPQRSTUV, voutput_min);
210
211 vout0x0123456789ABCDEF = vminq_u8(vout0x0123456789ABCDEF, voutput_max);
212 vout0xGHIJKLMNOPQRSTUV = vminq_u8(vout0xGHIJKLMNOPQRSTUV, voutput_max);
213
214 if (nc >= 32) {
215 vst1q_u8(c0 + 0, vout0x0123456789ABCDEF);
216 vst1q_u8(c0 + 16, vout0xGHIJKLMNOPQRSTUV);
217
218 c0 = (uint8_t*) ((uintptr_t) c0 + cn_stride);
219
220 a = (const uint8_t**restrict) ((uintptr_t) a - ks);
221
222 nc -= 32;
223 } else {
224 if (nc & 16) {
225 vst1q_u8(c0, vout0x0123456789ABCDEF); c0 += 16;
226
227 vout0x0123456789ABCDEF = vout0xGHIJKLMNOPQRSTUV;
228 }
229 uint8x8_t vout0x01234567 = vget_low_u8(vout0x0123456789ABCDEF);
230 if (nc & 8) {
231 vst1_u8(c0, vout0x01234567); c0 += 8; // This line
232 vout0x01234567 = vget_high_u8(vout0x0123456789ABCDEF);
233 }
234 if (nc & 4) {
235 vst1_lane_u32((void*) c0, vreinterpret_u32_u8(vout0x01234567), 0); c0 += 4;
236 vout0x01234567 = vext_u8(vout0x01234567, vout0x01234567, 4);
237 }
238 if (nc & 2) {
239 vst1_lane_u16((void*) c0, vreinterpret_u16_u8(vout0x01234567), 0); c0 += 2;
240 vout0x01234567 = vext_u8(vout0x01234567, vout0x01234567, 2);
241 }
242 if (nc & 1) {
243 vst1_lane_u8(c0, vout0x01234567, 0);
244 }
245
246 nc = 0;
247 }
248 } while (nc != 0);
249 }
250