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/intrinsics-polyfill.h>
16 #include <xnnpack/math.h>
17
18
xnn_qu8_igemm_minmax_fp32_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)])19 void xnn_qu8_igemm_minmax_fp32_ukernel_2x16c4__neondot(
20 size_t mr,
21 size_t nc,
22 size_t kc,
23 size_t ks,
24 const uint8_t** restrict a,
25 const void* restrict w,
26 uint8_t* restrict c,
27 size_t cm_stride,
28 size_t cn_stride,
29 size_t a_offset,
30 const uint8_t* zero,
31 const union xnn_qu8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
32 {
33 assert(mr != 0);
34 assert(mr <= 2);
35 assert(nc != 0);
36 assert(kc != 0);
37 assert(ks != 0);
38 assert(ks % (2 * sizeof(void*)) == 0);
39 assert(a_offset % sizeof(uint8_t) == 0);
40 assert(a != NULL);
41 assert(w != NULL);
42 assert(c != NULL);
43
44 kc = round_up_po2(kc, 4 * sizeof(uint8_t));
45 uint8_t* c0 = c;
46 uint8_t* c1 = (uint8_t*) ((uintptr_t) c0 + cm_stride);
47 if XNN_UNPREDICTABLE(mr != 2) {
48 c1 = c0;
49 }
50
51 const uint8x8_t va_zero_point = vld1_dup_u8(¶ms->fp32_neonv8.kernel_zero_point[0]);
52
53 do {
54 // Initialize accumulators with bias. 16 bias values are loaded from the
55 // weight matrix, at the start of the group of 16 columns.
56 uint32x4_t vpacc0x0123 = vld1q_u32(w); w = (const void*) ((const uint32_t*) w + 4);
57 uint32x4_t vpacc0x4567 = vld1q_u32(w); w = (const void*) ((const uint32_t*) w + 4);
58 uint32x4_t vpacc0x89AB = vld1q_u32(w); w = (const void*) ((const uint32_t*) w + 4);
59 uint32x4_t vpacc0xCDEF = vld1q_u32(w); w = (const void*) ((const uint32_t*) w + 4);
60 uint32x4_t vpacc1x0123 = vpacc0x0123;
61 uint32x4_t vpacc1x4567 = vpacc0x4567;
62 uint32x4_t vpacc1x89AB = vpacc0x89AB;
63 uint32x4_t vpacc1xCDEF = vpacc0xCDEF;
64 uint32x2_t vnacc0 = vmov_n_u32(0);
65 uint32x2_t vnacc1 = vmov_n_u32(0);
66
67 size_t p = ks;
68 do {
69 const uint8_t* restrict a0 = a[0];
70 if XNN_UNPREDICTABLE(a0 != zero) {
71 a0 = (const uint8_t*) ((uintptr_t) a0 + a_offset);
72 }
73 const uint8_t* restrict a1 = a[1];
74 if XNN_UNPREDICTABLE(a1 != zero) {
75 a1 = (const uint8_t*) ((uintptr_t) a1 + a_offset);
76 }
77 a += 2;
78
79 // Inner accumulation loop along the 16 columns.
80 size_t k = kc;
81 // 2x partial unrolled loop to load 8 bytes at a time.
82 while (k >= 8 * sizeof(uint8_t)) {
83 // Load a 2x8 block of activations.
84 const uint8x8_t va0x01234567 = vld1_u8(a0); a0 += 8;
85 const uint8x8_t va1x01234567 = vld1_u8(a1); a1 += 8;
86
87 // Load a 8x16 block of weights.
88 const uint8x16_t vb0123x0123 = vld1q_u8(w); w = (const void*) ((const uint8_t*) w + 16);
89 const uint8x16_t vb0123x4567 = vld1q_u8(w); w = (const void*) ((const uint8_t*) w + 16);
90 const uint8x16_t vb0123x89AB = vld1q_u8(w); w = (const void*) ((const uint8_t*) w + 16);
91 const uint8x16_t vb0123xCDEF = vld1q_u8(w); w = (const void*) ((const uint8_t*) w + 16);
92 const uint8x16_t vb4567x0123 = vld1q_u8(w); w = (const void*) ((const uint8_t*) w + 16);
93 const uint8x16_t vb4567x4567 = vld1q_u8(w); w = (const void*) ((const uint8_t*) w + 16);
94 const uint8x16_t vb4567x89AB = vld1q_u8(w); w = (const void*) ((const uint8_t*) w + 16);
95 const uint8x16_t vb4567xCDEF = vld1q_u8(w); w = (const void*) ((const uint8_t*) w + 16);
96
97 // Multiply-accumulate: 2x8 * 8x16 --> 2x16.
98 vnacc0 = vdot_u32(vnacc0, va_zero_point, va0x01234567);
99 vpacc0x0123 = vdotq_lane_u32(vpacc0x0123, vb0123x0123, va0x01234567, 0);
100 vpacc0x4567 = vdotq_lane_u32(vpacc0x4567, vb0123x4567, va0x01234567, 0);
101 vpacc0x89AB = vdotq_lane_u32(vpacc0x89AB, vb0123x89AB, va0x01234567, 0);
102 vpacc0xCDEF = vdotq_lane_u32(vpacc0xCDEF, vb0123xCDEF, va0x01234567, 0);
103 vpacc0x0123 = vdotq_lane_u32(vpacc0x0123, vb4567x0123, va0x01234567, 1);
104 vpacc0x4567 = vdotq_lane_u32(vpacc0x4567, vb4567x4567, va0x01234567, 1);
105 vpacc0x89AB = vdotq_lane_u32(vpacc0x89AB, vb4567x89AB, va0x01234567, 1);
106 vpacc0xCDEF = vdotq_lane_u32(vpacc0xCDEF, vb4567xCDEF, va0x01234567, 1);
107 vnacc1 = vdot_u32(vnacc1, va_zero_point, va1x01234567);
108 vpacc1x0123 = vdotq_lane_u32(vpacc1x0123, vb0123x0123, va1x01234567, 0);
109 vpacc1x4567 = vdotq_lane_u32(vpacc1x4567, vb0123x4567, va1x01234567, 0);
110 vpacc1x89AB = vdotq_lane_u32(vpacc1x89AB, vb0123x89AB, va1x01234567, 0);
111 vpacc1xCDEF = vdotq_lane_u32(vpacc1xCDEF, vb0123xCDEF, va1x01234567, 0);
112 vpacc1x0123 = vdotq_lane_u32(vpacc1x0123, vb4567x0123, va1x01234567, 1);
113 vpacc1x4567 = vdotq_lane_u32(vpacc1x4567, vb4567x4567, va1x01234567, 1);
114 vpacc1x89AB = vdotq_lane_u32(vpacc1x89AB, vb4567x89AB, va1x01234567, 1);
115 vpacc1xCDEF = vdotq_lane_u32(vpacc1xCDEF, vb4567xCDEF, va1x01234567, 1);
116
117 k -= 8 * sizeof(uint8_t);
118 }
119 // Handle up to 4 final positions of `k`
120 if XNN_UNLIKELY(k != 0) {
121 // Load a 2x4 block of activations.
122 const uint8x8_t va0x01234567 = vreinterpret_u8_u32(vld1_lane_u32((const void*) a0, vmov_n_u32(0), 0)); a0 += 4;
123 const uint8x8_t va1x01234567 = vreinterpret_u8_u32(vld1_lane_u32((const void*) a1, vmov_n_u32(0), 0)); a1 += 4;
124
125 // Load a 4x16 block of weights.
126 const uint8x16_t vb0123x0123 = vld1q_u8(w); w = (const void*) ((const uint8_t*) w + 16);
127 const uint8x16_t vb0123x4567 = vld1q_u8(w); w = (const void*) ((const uint8_t*) w + 16);
128 const uint8x16_t vb0123x89AB = vld1q_u8(w); w = (const void*) ((const uint8_t*) w + 16);
129 const uint8x16_t vb0123xCDEF = vld1q_u8(w); w = (const void*) ((const uint8_t*) w + 16);
130
131 // Multiply-accumulate: 2x4 * 4x16 --> 2x16.
132 vnacc0 = vdot_u32(vnacc0, va_zero_point, va0x01234567);
133 vpacc0x0123 = vdotq_lane_u32(vpacc0x0123, vb0123x0123, va0x01234567, 0);
134 vpacc0x4567 = vdotq_lane_u32(vpacc0x4567, vb0123x4567, va0x01234567, 0);
135 vpacc0x89AB = vdotq_lane_u32(vpacc0x89AB, vb0123x89AB, va0x01234567, 0);
136 vpacc0xCDEF = vdotq_lane_u32(vpacc0xCDEF, vb0123xCDEF, va0x01234567, 0);
137 vnacc1 = vdot_u32(vnacc1, va_zero_point, va1x01234567);
138 vpacc1x0123 = vdotq_lane_u32(vpacc1x0123, vb0123x0123, va1x01234567, 0);
139 vpacc1x4567 = vdotq_lane_u32(vpacc1x4567, vb0123x4567, va1x01234567, 0);
140 vpacc1x89AB = vdotq_lane_u32(vpacc1x89AB, vb0123x89AB, va1x01234567, 0);
141 vpacc1xCDEF = vdotq_lane_u32(vpacc1xCDEF, vb0123xCDEF, va1x01234567, 0);
142 }
143 p -= 2 * sizeof(void*);
144 } while (p != 0);
145
146 // Subtract zero point from accumulators.
147 vnacc0 = vpadd_u32(vnacc0, vnacc0);
148 const uint32x4_t vnacc0x0123 = vcombine_u32(vnacc0, vnacc0);
149 int32x4_t vacc0x0123 = vreinterpretq_s32_u32(vsubq_u32(vpacc0x0123, vnacc0x0123));
150 int32x4_t vacc0x4567 = vreinterpretq_s32_u32(vsubq_u32(vpacc0x4567, vnacc0x0123));
151 int32x4_t vacc0x89AB = vreinterpretq_s32_u32(vsubq_u32(vpacc0x89AB, vnacc0x0123));
152 int32x4_t vacc0xCDEF = vreinterpretq_s32_u32(vsubq_u32(vpacc0xCDEF, vnacc0x0123));
153 vnacc1 = vpadd_u32(vnacc1, vnacc1);
154 const uint32x4_t vnacc1x0123 = vcombine_u32(vnacc1, vnacc1);
155 int32x4_t vacc1x0123 = vreinterpretq_s32_u32(vsubq_u32(vpacc1x0123, vnacc1x0123));
156 int32x4_t vacc1x4567 = vreinterpretq_s32_u32(vsubq_u32(vpacc1x4567, vnacc1x0123));
157 int32x4_t vacc1x89AB = vreinterpretq_s32_u32(vsubq_u32(vpacc1x89AB, vnacc1x0123));
158 int32x4_t vacc1xCDEF = vreinterpretq_s32_u32(vsubq_u32(vpacc1xCDEF, vnacc1x0123));
159
160 float32x4_t vfpacc0x0123 = vcvtq_f32_s32(vacc0x0123);
161 float32x4_t vfpacc0x4567 = vcvtq_f32_s32(vacc0x4567);
162 float32x4_t vfpacc0x89AB = vcvtq_f32_s32(vacc0x89AB);
163 float32x4_t vfpacc0xCDEF = vcvtq_f32_s32(vacc0xCDEF);
164 float32x4_t vfpacc1x0123 = vcvtq_f32_s32(vacc1x0123);
165 float32x4_t vfpacc1x4567 = vcvtq_f32_s32(vacc1x4567);
166 float32x4_t vfpacc1x89AB = vcvtq_f32_s32(vacc1x89AB);
167 float32x4_t vfpacc1xCDEF = vcvtq_f32_s32(vacc1xCDEF);
168
169 const float32x4_t vscale = vld1q_dup_f32(¶ms->fp32_neonv8.scale);
170 vfpacc0x0123 = vmulq_f32(vfpacc0x0123, vscale);
171 vfpacc0x4567 = vmulq_f32(vfpacc0x4567, vscale);
172 vfpacc0x89AB = vmulq_f32(vfpacc0x89AB, vscale);
173 vfpacc0xCDEF = vmulq_f32(vfpacc0xCDEF, vscale);
174 vfpacc1x0123 = vmulq_f32(vfpacc1x0123, vscale);
175 vfpacc1x4567 = vmulq_f32(vfpacc1x4567, vscale);
176 vfpacc1x89AB = vmulq_f32(vfpacc1x89AB, vscale);
177 vfpacc1xCDEF = vmulq_f32(vfpacc1xCDEF, vscale);
178
179 vacc0x0123 = vcvtnq_s32_f32(vfpacc0x0123);
180 vacc0x4567 = vcvtnq_s32_f32(vfpacc0x4567);
181 vacc0x89AB = vcvtnq_s32_f32(vfpacc0x89AB);
182 vacc0xCDEF = vcvtnq_s32_f32(vfpacc0xCDEF);
183 vacc1x0123 = vcvtnq_s32_f32(vfpacc1x0123);
184 vacc1x4567 = vcvtnq_s32_f32(vfpacc1x4567);
185 vacc1x89AB = vcvtnq_s32_f32(vfpacc1x89AB);
186 vacc1xCDEF = vcvtnq_s32_f32(vfpacc1xCDEF);
187
188 const int16x8_t voutput_zero_point = vld1q_dup_s16(¶ms->fp32_neonv8.output_zero_point);
189 #if XNN_ARCH_ARM64
190 const int16x8_t vacc0x01234567 = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc0x0123), vacc0x4567), voutput_zero_point);
191 const int16x8_t vacc0x89ABCDEF = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc0x89AB), vacc0xCDEF), voutput_zero_point);
192 const int16x8_t vacc1x01234567 = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc1x0123), vacc1x4567), voutput_zero_point);
193 const int16x8_t vacc1x89ABCDEF = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc1x89AB), vacc1xCDEF), voutput_zero_point);
194
195 uint8x16_t vout0x0123456789ABCDEF = vqmovun_high_s16(vqmovun_s16(vacc0x01234567), vacc0x89ABCDEF);
196 uint8x16_t vout1x0123456789ABCDEF = vqmovun_high_s16(vqmovun_s16(vacc1x01234567), vacc1x89ABCDEF);
197 #else
198 const int16x8_t vacc0x01234567 = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc0x0123), vqmovn_s32(vacc0x4567)), voutput_zero_point);
199 const int16x8_t vacc0x89ABCDEF = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc0x89AB), vqmovn_s32(vacc0xCDEF)), voutput_zero_point);
200 const int16x8_t vacc1x01234567 = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc1x0123), vqmovn_s32(vacc1x4567)), voutput_zero_point);
201 const int16x8_t vacc1x89ABCDEF = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc1x89AB), vqmovn_s32(vacc1xCDEF)), voutput_zero_point);
202
203 uint8x16_t vout0x0123456789ABCDEF = vcombine_u8(vqmovun_s16(vacc0x01234567), vqmovun_s16(vacc0x89ABCDEF));
204 uint8x16_t vout1x0123456789ABCDEF = vcombine_u8(vqmovun_s16(vacc1x01234567), vqmovun_s16(vacc1x89ABCDEF));
205 #endif
206 const uint8x16_t voutput_min = vld1q_dup_u8(¶ms->fp32_neonv8.output_min);
207 const uint8x16_t voutput_max = vld1q_dup_u8(¶ms->fp32_neonv8.output_max);
208
209 vout0x0123456789ABCDEF = vmaxq_u8(vout0x0123456789ABCDEF, voutput_min);
210 vout1x0123456789ABCDEF = vmaxq_u8(vout1x0123456789ABCDEF, voutput_min);
211
212 vout0x0123456789ABCDEF = vminq_u8(vout0x0123456789ABCDEF, voutput_max);
213 vout1x0123456789ABCDEF = vminq_u8(vout1x0123456789ABCDEF, voutput_max);
214
215 if (nc >= 16) {
216 vst1q_u8(c1 + 0, vout1x0123456789ABCDEF);
217 vst1q_u8(c0 + 0, vout0x0123456789ABCDEF);
218
219 c1 = (uint8_t*) ((uintptr_t) c1 + cn_stride);
220 c0 = (uint8_t*) ((uintptr_t) c0 + cn_stride);
221
222 a = (const uint8_t**restrict) ((uintptr_t) a - ks);
223
224 nc -= 16;
225 } else {
226 uint8x16_t vout0x01234567_1x01234567 = vcombine_u8(vget_low_u8(vout0x0123456789ABCDEF), vget_low_u8(vout1x0123456789ABCDEF));
227 if (nc & 8) {
228 vst1_u8(c1, vget_high_u8(vout0x01234567_1x01234567)); c1 += 8;
229 vst1_u8(c0, vget_low_u8(vout0x01234567_1x01234567)); c0 += 8;
230 vout0x01234567_1x01234567 = vcombine_u8(vget_high_u8(vout0x0123456789ABCDEF), vget_high_u8(vout1x0123456789ABCDEF));
231 }
232 if (nc & 4) {
233 vst1q_lane_u32((void*) c1, vreinterpretq_u32_u8(vout0x01234567_1x01234567), 2); c1 += 4;
234 vst1q_lane_u32((void*) c0, vreinterpretq_u32_u8(vout0x01234567_1x01234567), 0); c0 += 4;
235 vout0x01234567_1x01234567 = vextq_u8(vout0x01234567_1x01234567, vout0x01234567_1x01234567, 4);
236 }
237 if (nc & 2) {
238 vst1q_lane_u16((void*) c1, vreinterpretq_u16_u8(vout0x01234567_1x01234567), 4); c1 += 2;
239 vst1q_lane_u16((void*) c0, vreinterpretq_u16_u8(vout0x01234567_1x01234567), 0); c0 += 2;
240 vout0x01234567_1x01234567 = vextq_u8(vout0x01234567_1x01234567, vout0x01234567_1x01234567, 2);
241 }
242 if (nc & 1) {
243 vst1q_lane_u8(c1, vout0x01234567_1x01234567, 8);
244 vst1q_lane_u8(c0, vout0x01234567_1x01234567, 0);
245 }
246
247 nc = 0;
248 }
249 } while (nc != 0);
250 }
251