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_3x8c4__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_3x8c4__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 <= 3);
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
55 const uint8x8_t va_zero_point = vld1_dup_u8(¶ms->rndnu_neon.kernel_zero_point[0]);
56
57 // Loop over groups of 8 columns.
58 do {
59 // Initialize accumulators with bias. 8 bias values are loaded from the
60 // weight matrix, at the start of the group of 8 columns.
61 uint32x4_t vpacc0x0123 = vld1q_u32(w); w = (const void*) ((const uint32_t*) w + 4);
62 uint32x4_t vpacc0x4567 = vld1q_u32(w); w = (const void*) ((const uint32_t*) w + 4);
63 uint32x4_t vpacc1x0123 = vpacc0x0123;
64 uint32x4_t vpacc1x4567 = vpacc0x4567;
65 uint32x4_t vpacc2x0123 = vpacc0x0123;
66 uint32x4_t vpacc2x4567 = vpacc0x4567;
67 uint32x2_t vnacc0 = vmov_n_u32(0);
68 uint32x2_t vnacc1 = vmov_n_u32(0);
69 uint32x2_t vnacc2 = vmov_n_u32(0);
70
71 // Inner accumulation loop along the 8 columns.
72 size_t k = kc;
73 // 2x partial unrolled loop to load 8 bytes at a time.
74 while (k >= 8 * sizeof(uint8_t)) {
75 // Load a 3x8 block of activations.
76 const uint8x8_t va0x01234567 = vld1_u8(a0); a0 += 8;
77 const uint8x8_t va1x01234567 = vld1_u8(a1); a1 += 8;
78 const uint8x8_t va2x01234567 = vld1_u8(a2); a2 += 8;
79
80 // Load a 8x8 block of weights.
81 const uint8x16_t vb0123x0123 = vld1q_u8(w); w = (const void*) ((const uint8_t*) w + 16);
82 const uint8x16_t vb0123x4567 = vld1q_u8(w); w = (const void*) ((const uint8_t*) w + 16);
83 const uint8x16_t vb4567x0123 = vld1q_u8(w); w = (const void*) ((const uint8_t*) w + 16);
84 const uint8x16_t vb4567x4567 = vld1q_u8(w); w = (const void*) ((const uint8_t*) w + 16);
85
86 // Multiply-accumulate: 3x8 * 8x8 --> 3x8.
87 vnacc0 = vdot_u32(vnacc0, va_zero_point, va0x01234567);
88 vpacc0x0123 = vdotq_lane_u32(vpacc0x0123, vb0123x0123, va0x01234567, 0);
89 vpacc0x4567 = vdotq_lane_u32(vpacc0x4567, vb0123x4567, va0x01234567, 0);
90 vpacc0x0123 = vdotq_lane_u32(vpacc0x0123, vb4567x0123, va0x01234567, 1);
91 vpacc0x4567 = vdotq_lane_u32(vpacc0x4567, vb4567x4567, va0x01234567, 1);
92 vnacc1 = vdot_u32(vnacc1, va_zero_point, va1x01234567);
93 vpacc1x0123 = vdotq_lane_u32(vpacc1x0123, vb0123x0123, va1x01234567, 0);
94 vpacc1x4567 = vdotq_lane_u32(vpacc1x4567, vb0123x4567, va1x01234567, 0);
95 vpacc1x0123 = vdotq_lane_u32(vpacc1x0123, vb4567x0123, va1x01234567, 1);
96 vpacc1x4567 = vdotq_lane_u32(vpacc1x4567, vb4567x4567, va1x01234567, 1);
97 vnacc2 = vdot_u32(vnacc2, va_zero_point, va2x01234567);
98 vpacc2x0123 = vdotq_lane_u32(vpacc2x0123, vb0123x0123, va2x01234567, 0);
99 vpacc2x4567 = vdotq_lane_u32(vpacc2x4567, vb0123x4567, va2x01234567, 0);
100 vpacc2x0123 = vdotq_lane_u32(vpacc2x0123, vb4567x0123, va2x01234567, 1);
101 vpacc2x4567 = vdotq_lane_u32(vpacc2x4567, vb4567x4567, va2x01234567, 1);
102
103 k -= 8 * sizeof(uint8_t);
104 }
105 // Handle up to 4 final positions of `k`
106 if XNN_UNLIKELY(k != 0) {
107 // Load a 3x4 block of activations.
108 const uint8x8_t va0x01234567 = vreinterpret_u8_u32(vld1_lane_u32((const void*) a0, vmov_n_u32(0), 0)); a0 += 4;
109 const uint8x8_t va1x01234567 = vreinterpret_u8_u32(vld1_lane_u32((const void*) a1, vmov_n_u32(0), 0)); a1 += 4;
110 const uint8x8_t va2x01234567 = vreinterpret_u8_u32(vld1_lane_u32((const void*) a2, vmov_n_u32(0), 0)); a2 += 4;
111
112 // Load a 4x8 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
116 // Multiply-accumulate: 3x4 * 4x8 --> 3x8.
117 vnacc0 = vdot_u32(vnacc0, va_zero_point, va0x01234567);
118 vpacc0x0123 = vdotq_lane_u32(vpacc0x0123, vb0123x0123, va0x01234567, 0);
119 vpacc0x4567 = vdotq_lane_u32(vpacc0x4567, vb0123x4567, va0x01234567, 0);
120 vnacc1 = vdot_u32(vnacc1, va_zero_point, va1x01234567);
121 vpacc1x0123 = vdotq_lane_u32(vpacc1x0123, vb0123x0123, va1x01234567, 0);
122 vpacc1x4567 = vdotq_lane_u32(vpacc1x4567, vb0123x4567, va1x01234567, 0);
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 }
127
128 // Subtract zero point from accumulators.
129 vnacc0 = vpadd_u32(vnacc0, vnacc0);
130 const uint32x4_t vnacc0x0123 = vcombine_u32(vnacc0, vnacc0);
131 int32x4_t vacc0x0123 = vreinterpretq_s32_u32(vsubq_u32(vpacc0x0123, vnacc0x0123));
132 int32x4_t vacc0x4567 = vreinterpretq_s32_u32(vsubq_u32(vpacc0x4567, vnacc0x0123));
133 vnacc1 = vpadd_u32(vnacc1, vnacc1);
134 const uint32x4_t vnacc1x0123 = vcombine_u32(vnacc1, vnacc1);
135 int32x4_t vacc1x0123 = vreinterpretq_s32_u32(vsubq_u32(vpacc1x0123, vnacc1x0123));
136 int32x4_t vacc1x4567 = vreinterpretq_s32_u32(vsubq_u32(vpacc1x4567, vnacc1x0123));
137 vnacc2 = vpadd_u32(vnacc2, vnacc2);
138 const uint32x4_t vnacc2x0123 = vcombine_u32(vnacc2, vnacc2);
139 int32x4_t vacc2x0123 = vreinterpretq_s32_u32(vsubq_u32(vpacc2x0123, vnacc2x0123));
140 int32x4_t vacc2x4567 = vreinterpretq_s32_u32(vsubq_u32(vpacc2x4567, vnacc2x0123));
141
142 const int32x4_t vright_pre_shift = vld1q_dup_s32(¶ms->rndnu_neon.right_pre_shift);
143 const int32x4_t vmultiplier = vld1q_dup_s32(¶ms->rndnu_neon.multiplier);
144 const int32x4_t vright_post_shift = vld1q_dup_s32(¶ms->rndnu_neon.right_post_shift);
145
146 vacc0x0123 = vshlq_s32(vacc0x0123, vright_pre_shift);
147 vacc0x4567 = vshlq_s32(vacc0x4567, vright_pre_shift);
148 vacc1x0123 = vshlq_s32(vacc1x0123, vright_pre_shift);
149 vacc1x4567 = vshlq_s32(vacc1x4567, vright_pre_shift);
150 vacc2x0123 = vshlq_s32(vacc2x0123, vright_pre_shift);
151 vacc2x4567 = vshlq_s32(vacc2x4567, vright_pre_shift);
152
153 vacc0x0123 = vqdmulhq_s32(vacc0x0123, vmultiplier);
154 vacc0x4567 = vqdmulhq_s32(vacc0x4567, vmultiplier);
155 vacc1x0123 = vqdmulhq_s32(vacc1x0123, vmultiplier);
156 vacc1x4567 = vqdmulhq_s32(vacc1x4567, vmultiplier);
157 vacc2x0123 = vqdmulhq_s32(vacc2x0123, vmultiplier);
158 vacc2x4567 = vqdmulhq_s32(vacc2x4567, vmultiplier);
159
160 vacc0x0123 = vrshlq_s32(vacc0x0123, vright_post_shift);
161 vacc0x4567 = vrshlq_s32(vacc0x4567, vright_post_shift);
162 vacc1x0123 = vrshlq_s32(vacc1x0123, vright_post_shift);
163 vacc1x4567 = vrshlq_s32(vacc1x4567, vright_post_shift);
164 vacc2x0123 = vrshlq_s32(vacc2x0123, vright_post_shift);
165 vacc2x4567 = vrshlq_s32(vacc2x4567, vright_post_shift);
166
167 const int16x8_t voutput_zero_point = vld1q_dup_s16(¶ms->rndnu_neon.output_zero_point);
168 #if XNN_ARCH_ARM64
169 const int16x8_t vacc0x01234567 = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc0x0123), vacc0x4567), voutput_zero_point);
170 const int16x8_t vacc1x01234567 = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc1x0123), vacc1x4567), voutput_zero_point);
171 const int16x8_t vacc2x01234567 = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc2x0123), vacc2x4567), voutput_zero_point);
172
173 uint8x16_t vout0x01234567_1x01234567 = vqmovun_high_s16(vqmovun_s16(vacc0x01234567), vacc1x01234567);
174 uint8x8_t vout2x01234567 = vqmovun_s16(vacc2x01234567);
175 #else
176 const int16x8_t vacc0x01234567 = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc0x0123), vqmovn_s32(vacc0x4567)), voutput_zero_point);
177 const int16x8_t vacc1x01234567 = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc1x0123), vqmovn_s32(vacc1x4567)), voutput_zero_point);
178 const int16x8_t vacc2x01234567 = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc2x0123), vqmovn_s32(vacc2x4567)), voutput_zero_point);
179
180 uint8x16_t vout0x01234567_1x01234567 = vcombine_u8(vqmovun_s16(vacc0x01234567), vqmovun_s16(vacc1x01234567));
181 uint8x8_t vout2x01234567 = vqmovun_s16(vacc2x01234567);
182 #endif
183 const uint8x16_t voutput_min = vld1q_dup_u8(¶ms->rndnu_neon.output_min);
184 const uint8x16_t voutput_max = vld1q_dup_u8(¶ms->rndnu_neon.output_max);
185
186 vout0x01234567_1x01234567 = vmaxq_u8(vout0x01234567_1x01234567, voutput_min);
187 vout2x01234567 = vmax_u8(vout2x01234567, vget_low_u8(voutput_min));
188
189 vout0x01234567_1x01234567 = vminq_u8(vout0x01234567_1x01234567, voutput_max);
190 vout2x01234567 = vmin_u8(vout2x01234567, vget_low_u8(voutput_max));
191
192 if (nc >= 8) {
193 vst1_u8(c0 + 0, vget_low_u8(vout0x01234567_1x01234567));
194 vst1_u8(c1 + 0, vget_high_u8(vout0x01234567_1x01234567));
195 vst1_u8(c2 + 0, vout2x01234567);
196
197 c0 = (uint8_t*) ((uintptr_t) c0 + cn_stride);
198 c1 = (uint8_t*) ((uintptr_t) c1 + cn_stride);
199 c2 = (uint8_t*) ((uintptr_t) c2 + cn_stride);
200
201 a0 = (const uint8_t*) ((uintptr_t) a0 - kc);
202 a1 = (const uint8_t*) ((uintptr_t) a1 - kc);
203 a2 = (const uint8_t*) ((uintptr_t) a2 - kc);
204
205 nc -= 8;
206 } else {
207 if (nc & 4) {
208 vst1q_lane_u32((void*) c0, vreinterpretq_u32_u8(vout0x01234567_1x01234567), 0); c0 += 4;
209 vst1q_lane_u32((void*) c1, vreinterpretq_u32_u8(vout0x01234567_1x01234567), 2); c1 += 4;
210 vst1_lane_u32((void*) c2, vreinterpret_u32_u8(vout2x01234567), 0); c2 += 4;
211 vout0x01234567_1x01234567 = vextq_u8(vout0x01234567_1x01234567, vout0x01234567_1x01234567, 4);
212 vout2x01234567 = vext_u8(vout2x01234567, vout2x01234567, 4);
213 }
214 if (nc & 2) {
215 vst1q_lane_u16((void*) c0, vreinterpretq_u16_u8(vout0x01234567_1x01234567), 0); c0 += 2;
216 vst1q_lane_u16((void*) c1, vreinterpretq_u16_u8(vout0x01234567_1x01234567), 4); c1 += 2;
217 vst1_lane_u16((void*) c2, vreinterpret_u16_u8(vout2x01234567), 0); c2 += 2;
218 vout0x01234567_1x01234567 = vextq_u8(vout0x01234567_1x01234567, vout0x01234567_1x01234567, 2);
219 vout2x01234567 = vext_u8(vout2x01234567, vout2x01234567, 2);
220 }
221 if (nc & 1) {
222 vst1q_lane_u8(c0, vout0x01234567_1x01234567, 0);
223 vst1q_lane_u8(c1, vout0x01234567_1x01234567, 8);
224 vst1_lane_u8(c2, vout2x01234567, 0);
225 }
226
227 nc = 0;
228 }
229 } while (nc != 0);
230 }
231