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_2x16c4__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_2x16c4__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 <= 2);
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
49 const uint8x8_t va_zero_point = vld1_dup_u8(¶ms->rndnu_neon.kernel_zero_point[0]);
50
51 // Loop over groups of 16 columns.
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 // Inner accumulation loop along the 16 columns.
67 size_t k = kc;
68 // 2x partial unrolled loop to load 8 bytes at a time.
69 while (k >= 8 * sizeof(uint8_t)) {
70 // Load a 2x8 block of activations.
71 const uint8x8_t va0x01234567 = vld1_u8(a0); a0 += 8;
72 const uint8x8_t va1x01234567 = vld1_u8(a1); a1 += 8;
73
74 // Load a 8x16 block of weights.
75 const uint8x16_t vb0123x0123 = vld1q_u8(w); w = (const void*) ((const uint8_t*) w + 16);
76 const uint8x16_t vb0123x4567 = vld1q_u8(w); w = (const void*) ((const uint8_t*) w + 16);
77 const uint8x16_t vb0123x89AB = vld1q_u8(w); w = (const void*) ((const uint8_t*) w + 16);
78 const uint8x16_t vb0123xCDEF = vld1q_u8(w); w = (const void*) ((const uint8_t*) w + 16);
79 const uint8x16_t vb4567x0123 = vld1q_u8(w); w = (const void*) ((const uint8_t*) w + 16);
80 const uint8x16_t vb4567x4567 = vld1q_u8(w); w = (const void*) ((const uint8_t*) w + 16);
81 const uint8x16_t vb4567x89AB = vld1q_u8(w); w = (const void*) ((const uint8_t*) w + 16);
82 const uint8x16_t vb4567xCDEF = vld1q_u8(w); w = (const void*) ((const uint8_t*) w + 16);
83
84 // Multiply-accumulate: 2x8 * 8x16 --> 2x16.
85 vnacc0 = vdot_u32(vnacc0, va_zero_point, va0x01234567);
86 vpacc0x0123 = vdotq_lane_u32(vpacc0x0123, vb0123x0123, va0x01234567, 0);
87 vpacc0x4567 = vdotq_lane_u32(vpacc0x4567, vb0123x4567, va0x01234567, 0);
88 vpacc0x89AB = vdotq_lane_u32(vpacc0x89AB, vb0123x89AB, va0x01234567, 0);
89 vpacc0xCDEF = vdotq_lane_u32(vpacc0xCDEF, vb0123xCDEF, va0x01234567, 0);
90 vpacc0x0123 = vdotq_lane_u32(vpacc0x0123, vb4567x0123, va0x01234567, 1);
91 vpacc0x4567 = vdotq_lane_u32(vpacc0x4567, vb4567x4567, va0x01234567, 1);
92 vpacc0x89AB = vdotq_lane_u32(vpacc0x89AB, vb4567x89AB, va0x01234567, 1);
93 vpacc0xCDEF = vdotq_lane_u32(vpacc0xCDEF, vb4567xCDEF, 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 vpacc1x89AB = vdotq_lane_u32(vpacc1x89AB, vb0123x89AB, va1x01234567, 0);
98 vpacc1xCDEF = vdotq_lane_u32(vpacc1xCDEF, vb0123xCDEF, va1x01234567, 0);
99 vpacc1x0123 = vdotq_lane_u32(vpacc1x0123, vb4567x0123, va1x01234567, 1);
100 vpacc1x4567 = vdotq_lane_u32(vpacc1x4567, vb4567x4567, va1x01234567, 1);
101 vpacc1x89AB = vdotq_lane_u32(vpacc1x89AB, vb4567x89AB, va1x01234567, 1);
102 vpacc1xCDEF = vdotq_lane_u32(vpacc1xCDEF, vb4567xCDEF, va1x01234567, 1);
103
104 k -= 8 * sizeof(uint8_t);
105 }
106 // Handle up to 4 final positions of `k`
107 if XNN_UNLIKELY(k != 0) {
108 // Load a 2x4 block of activations.
109 const uint8x8_t va0x01234567 = vreinterpret_u8_u32(vld1_lane_u32((const void*) a0, vmov_n_u32(0), 0)); a0 += 4;
110 const uint8x8_t va1x01234567 = vreinterpret_u8_u32(vld1_lane_u32((const void*) a1, vmov_n_u32(0), 0)); a1 += 4;
111
112 // Load a 4x16 block of weights.
113 const uint8x16_t vb0123x0123 = vld1q_u8(w); w = (const void*) ((const uint8_t*) w + 16);
114 const uint8x16_t vb0123x4567 = vld1q_u8(w); w = (const void*) ((const uint8_t*) w + 16);
115 const uint8x16_t vb0123x89AB = vld1q_u8(w); w = (const void*) ((const uint8_t*) w + 16);
116 const uint8x16_t vb0123xCDEF = vld1q_u8(w); w = (const void*) ((const uint8_t*) w + 16);
117
118 // Multiply-accumulate: 2x4 * 4x16 --> 2x16.
119 vnacc0 = vdot_u32(vnacc0, va_zero_point, va0x01234567);
120 vpacc0x0123 = vdotq_lane_u32(vpacc0x0123, vb0123x0123, va0x01234567, 0);
121 vpacc0x4567 = vdotq_lane_u32(vpacc0x4567, vb0123x4567, va0x01234567, 0);
122 vpacc0x89AB = vdotq_lane_u32(vpacc0x89AB, vb0123x89AB, va0x01234567, 0);
123 vpacc0xCDEF = vdotq_lane_u32(vpacc0xCDEF, vb0123xCDEF, va0x01234567, 0);
124 vnacc1 = vdot_u32(vnacc1, va_zero_point, va1x01234567);
125 vpacc1x0123 = vdotq_lane_u32(vpacc1x0123, vb0123x0123, va1x01234567, 0);
126 vpacc1x4567 = vdotq_lane_u32(vpacc1x4567, vb0123x4567, va1x01234567, 0);
127 vpacc1x89AB = vdotq_lane_u32(vpacc1x89AB, vb0123x89AB, va1x01234567, 0);
128 vpacc1xCDEF = vdotq_lane_u32(vpacc1xCDEF, vb0123xCDEF, va1x01234567, 0);
129 }
130
131 // Subtract zero point from accumulators.
132 vnacc0 = vpadd_u32(vnacc0, vnacc0);
133 const uint32x4_t vnacc0x0123 = vcombine_u32(vnacc0, vnacc0);
134 int32x4_t vacc0x0123 = vreinterpretq_s32_u32(vsubq_u32(vpacc0x0123, vnacc0x0123));
135 int32x4_t vacc0x4567 = vreinterpretq_s32_u32(vsubq_u32(vpacc0x4567, vnacc0x0123));
136 int32x4_t vacc0x89AB = vreinterpretq_s32_u32(vsubq_u32(vpacc0x89AB, vnacc0x0123));
137 int32x4_t vacc0xCDEF = vreinterpretq_s32_u32(vsubq_u32(vpacc0xCDEF, vnacc0x0123));
138 vnacc1 = vpadd_u32(vnacc1, vnacc1);
139 const uint32x4_t vnacc1x0123 = vcombine_u32(vnacc1, vnacc1);
140 int32x4_t vacc1x0123 = vreinterpretq_s32_u32(vsubq_u32(vpacc1x0123, vnacc1x0123));
141 int32x4_t vacc1x4567 = vreinterpretq_s32_u32(vsubq_u32(vpacc1x4567, vnacc1x0123));
142 int32x4_t vacc1x89AB = vreinterpretq_s32_u32(vsubq_u32(vpacc1x89AB, vnacc1x0123));
143 int32x4_t vacc1xCDEF = vreinterpretq_s32_u32(vsubq_u32(vpacc1xCDEF, vnacc1x0123));
144
145 const int32x4_t vright_pre_shift = vld1q_dup_s32(¶ms->rndnu_neon.right_pre_shift);
146 const int32x4_t vmultiplier = vld1q_dup_s32(¶ms->rndnu_neon.multiplier);
147 const int32x4_t vright_post_shift = vld1q_dup_s32(¶ms->rndnu_neon.right_post_shift);
148
149 vacc0x0123 = vshlq_s32(vacc0x0123, vright_pre_shift);
150 vacc0x4567 = vshlq_s32(vacc0x4567, vright_pre_shift);
151 vacc0x89AB = vshlq_s32(vacc0x89AB, vright_pre_shift);
152 vacc0xCDEF = vshlq_s32(vacc0xCDEF, vright_pre_shift);
153 vacc1x0123 = vshlq_s32(vacc1x0123, vright_pre_shift);
154 vacc1x4567 = vshlq_s32(vacc1x4567, vright_pre_shift);
155 vacc1x89AB = vshlq_s32(vacc1x89AB, vright_pre_shift);
156 vacc1xCDEF = vshlq_s32(vacc1xCDEF, vright_pre_shift);
157
158 vacc0x0123 = vqdmulhq_s32(vacc0x0123, vmultiplier);
159 vacc0x4567 = vqdmulhq_s32(vacc0x4567, vmultiplier);
160 vacc0x89AB = vqdmulhq_s32(vacc0x89AB, vmultiplier);
161 vacc0xCDEF = vqdmulhq_s32(vacc0xCDEF, vmultiplier);
162 vacc1x0123 = vqdmulhq_s32(vacc1x0123, vmultiplier);
163 vacc1x4567 = vqdmulhq_s32(vacc1x4567, vmultiplier);
164 vacc1x89AB = vqdmulhq_s32(vacc1x89AB, vmultiplier);
165 vacc1xCDEF = vqdmulhq_s32(vacc1xCDEF, vmultiplier);
166
167 vacc0x0123 = vrshlq_s32(vacc0x0123, vright_post_shift);
168 vacc0x4567 = vrshlq_s32(vacc0x4567, vright_post_shift);
169 vacc0x89AB = vrshlq_s32(vacc0x89AB, vright_post_shift);
170 vacc0xCDEF = vrshlq_s32(vacc0xCDEF, vright_post_shift);
171 vacc1x0123 = vrshlq_s32(vacc1x0123, vright_post_shift);
172 vacc1x4567 = vrshlq_s32(vacc1x4567, vright_post_shift);
173 vacc1x89AB = vrshlq_s32(vacc1x89AB, vright_post_shift);
174 vacc1xCDEF = vrshlq_s32(vacc1xCDEF, vright_post_shift);
175
176 const int16x8_t voutput_zero_point = vld1q_dup_s16(¶ms->rndnu_neon.output_zero_point);
177 #if XNN_ARCH_ARM64
178 const int16x8_t vacc0x01234567 = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc0x0123), vacc0x4567), voutput_zero_point);
179 const int16x8_t vacc0x89ABCDEF = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc0x89AB), vacc0xCDEF), voutput_zero_point);
180 const int16x8_t vacc1x01234567 = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc1x0123), vacc1x4567), voutput_zero_point);
181 const int16x8_t vacc1x89ABCDEF = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc1x89AB), vacc1xCDEF), voutput_zero_point);
182
183 uint8x16_t vout0x0123456789ABCDEF = vqmovun_high_s16(vqmovun_s16(vacc0x01234567), vacc0x89ABCDEF);
184 uint8x16_t vout1x0123456789ABCDEF = vqmovun_high_s16(vqmovun_s16(vacc1x01234567), vacc1x89ABCDEF);
185 #else
186 const int16x8_t vacc0x01234567 = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc0x0123), vqmovn_s32(vacc0x4567)), voutput_zero_point);
187 const int16x8_t vacc0x89ABCDEF = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc0x89AB), vqmovn_s32(vacc0xCDEF)), voutput_zero_point);
188 const int16x8_t vacc1x01234567 = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc1x0123), vqmovn_s32(vacc1x4567)), voutput_zero_point);
189 const int16x8_t vacc1x89ABCDEF = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc1x89AB), vqmovn_s32(vacc1xCDEF)), voutput_zero_point);
190
191 uint8x16_t vout0x0123456789ABCDEF = vcombine_u8(vqmovun_s16(vacc0x01234567), vqmovun_s16(vacc0x89ABCDEF));
192 uint8x16_t vout1x0123456789ABCDEF = vcombine_u8(vqmovun_s16(vacc1x01234567), vqmovun_s16(vacc1x89ABCDEF));
193 #endif
194 const uint8x16_t voutput_min = vld1q_dup_u8(¶ms->rndnu_neon.output_min);
195 const uint8x16_t voutput_max = vld1q_dup_u8(¶ms->rndnu_neon.output_max);
196
197 vout0x0123456789ABCDEF = vmaxq_u8(vout0x0123456789ABCDEF, voutput_min);
198 vout1x0123456789ABCDEF = vmaxq_u8(vout1x0123456789ABCDEF, voutput_min);
199
200 vout0x0123456789ABCDEF = vminq_u8(vout0x0123456789ABCDEF, voutput_max);
201 vout1x0123456789ABCDEF = vminq_u8(vout1x0123456789ABCDEF, voutput_max);
202
203 if (nc >= 16) {
204 vst1q_u8(c0 + 0, vout0x0123456789ABCDEF);
205 vst1q_u8(c1 + 0, vout1x0123456789ABCDEF);
206
207 c0 = (uint8_t*) ((uintptr_t) c0 + cn_stride);
208 c1 = (uint8_t*) ((uintptr_t) c1 + cn_stride);
209
210 a0 = (const uint8_t*) ((uintptr_t) a0 - kc);
211 a1 = (const uint8_t*) ((uintptr_t) a1 - kc);
212
213 nc -= 16;
214 } else {
215 uint8x16_t vout0x01234567_1x01234567 = vcombine_u8(vget_low_u8(vout0x0123456789ABCDEF), vget_low_u8(vout1x0123456789ABCDEF));
216 if (nc & 8) {
217 vst1_u8(c0, vget_low_u8(vout0x01234567_1x01234567)); c0 += 8;
218 vst1_u8(c1, vget_high_u8(vout0x01234567_1x01234567)); c1 += 8;
219 vout0x01234567_1x01234567 = vcombine_u8(vget_high_u8(vout0x0123456789ABCDEF), vget_high_u8(vout1x0123456789ABCDEF));
220 }
221 if (nc & 4) {
222 vst1q_lane_u32((void*) c0, vreinterpretq_u32_u8(vout0x01234567_1x01234567), 0); c0 += 4;
223 vst1q_lane_u32((void*) c1, vreinterpretq_u32_u8(vout0x01234567_1x01234567), 2); c1 += 4;
224 vout0x01234567_1x01234567 = vextq_u8(vout0x01234567_1x01234567, vout0x01234567_1x01234567, 4);
225 }
226 if (nc & 2) {
227 vst1q_lane_u16((void*) c0, vreinterpretq_u16_u8(vout0x01234567_1x01234567), 0); c0 += 2;
228 vst1q_lane_u16((void*) c1, vreinterpretq_u16_u8(vout0x01234567_1x01234567), 4); c1 += 2;
229 vout0x01234567_1x01234567 = vextq_u8(vout0x01234567_1x01234567, vout0x01234567_1x01234567, 2);
230 }
231 if (nc & 1) {
232 vst1q_lane_u8(c0, vout0x01234567_1x01234567, 0);
233 vst1q_lane_u8(c1, vout0x01234567_1x01234567, 8);
234 }
235
236 nc = 0;
237 }
238 } while (nc != 0);
239 }
240