1 // Auto-generated file. Do not edit!
2 // Template: src/qs8-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_qs8_igemm_minmax_rndnu_ukernel_4x8c4__neondot(size_t mr,size_t nc,size_t kc,size_t ks,const int8_t ** restrict a,const void * restrict w,int8_t * restrict c,size_t cm_stride,size_t cn_stride,size_t a_offset,const int8_t * zero,const union xnn_qs8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS (1)])18 void xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c4__neondot(
19 size_t mr,
20 size_t nc,
21 size_t kc,
22 size_t ks,
23 const int8_t** restrict a,
24 const void* restrict w,
25 int8_t* restrict c,
26 size_t cm_stride,
27 size_t cn_stride,
28 size_t a_offset,
29 const int8_t* zero,
30 const union xnn_qs8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
31 {
32 assert(mr != 0);
33 assert(mr <= 4);
34 assert(nc != 0);
35 assert(kc != 0);
36 assert(ks != 0);
37 assert(ks % (4 * sizeof(void*)) == 0);
38 assert(a_offset % sizeof(int8_t) == 0);
39 assert(a != NULL);
40 assert(w != NULL);
41 assert(c != NULL);
42
43 kc = round_up_po2(kc, 4 * sizeof(int8_t));
44 int8_t* c0 = c;
45 int8_t* c1 = (int8_t*) ((uintptr_t) c0 + cm_stride);
46 if XNN_UNPREDICTABLE(mr < 2) {
47 c1 = c0;
48 }
49 int8_t* c2 = (int8_t*) ((uintptr_t) c1 + cm_stride);
50 if XNN_UNPREDICTABLE(mr <= 2) {
51 c2 = c1;
52 }
53 int8_t* c3 = (int8_t*) ((uintptr_t) c2 + cm_stride);
54 if XNN_UNPREDICTABLE(mr != 4) {
55 c3 = c2;
56 }
57
58 do {
59 int32x4_t vacc0x0123 = vld1q_s32(w); w = (const void*) ((const int32_t*) w + 4);
60 int32x4_t vacc0x4567 = vld1q_s32(w); w = (const void*) ((const int32_t*) w + 4);
61 int32x4_t vacc1x0123 = vacc0x0123;
62 int32x4_t vacc1x4567 = vacc0x4567;
63 int32x4_t vacc2x0123 = vacc0x0123;
64 int32x4_t vacc2x4567 = vacc0x4567;
65 int32x4_t vacc3x0123 = vacc0x0123;
66 int32x4_t vacc3x4567 = vacc0x4567;
67
68 size_t p = ks;
69 do {
70 const int8_t* restrict a0 = a[0];
71 if XNN_UNPREDICTABLE(a0 != zero) {
72 a0 = (const int8_t*) ((uintptr_t) a0 + a_offset);
73 }
74 const int8_t* restrict a1 = a[1];
75 if XNN_UNPREDICTABLE(a1 != zero) {
76 a1 = (const int8_t*) ((uintptr_t) a1 + a_offset);
77 }
78 const int8_t* restrict a2 = a[2];
79 if XNN_UNPREDICTABLE(a2 != zero) {
80 a2 = (const int8_t*) ((uintptr_t) a2 + a_offset);
81 }
82 const int8_t* restrict a3 = a[3];
83 if XNN_UNPREDICTABLE(a3 != zero) {
84 a3 = (const int8_t*) ((uintptr_t) a3 + a_offset);
85 }
86 a += 4;
87
88 // Inner accumulation loop along the 8 columns.
89 size_t k = kc;
90 // 2x partial unrolled loop to load 8 bytes at a time.
91 while (k >= 8 * sizeof(int8_t)) {
92 // Load a 4x8 block of activations.
93 const int8x8_t va0x01234567 = vld1_s8(a0); a0 += 8;
94 const int8x8_t va1x01234567 = vld1_s8(a1); a1 += 8;
95 const int8x8_t va2x01234567 = vld1_s8(a2); a2 += 8;
96 const int8x8_t va3x01234567 = vld1_s8(a3); a3 += 8;
97
98 // Load a 8x8 block of weights.
99 const int8x16_t vb0123x0123 = vld1q_s8(w); w = (const void*) ((const int8_t*) w + 16);
100 const int8x16_t vb0123x4567 = vld1q_s8(w); w = (const void*) ((const int8_t*) w + 16);
101 const int8x16_t vb4567x0123 = vld1q_s8(w); w = (const void*) ((const int8_t*) w + 16);
102 const int8x16_t vb4567x4567 = vld1q_s8(w); w = (const void*) ((const int8_t*) w + 16);
103
104 // Multiply-accumulate: 4x8 * 8x8 --> 4x8.
105 vacc0x0123 = vdotq_lane_s32(vacc0x0123, vb0123x0123, va0x01234567, 0);
106 vacc0x4567 = vdotq_lane_s32(vacc0x4567, vb0123x4567, va0x01234567, 0);
107 vacc1x0123 = vdotq_lane_s32(vacc1x0123, vb0123x0123, va1x01234567, 0);
108 vacc1x4567 = vdotq_lane_s32(vacc1x4567, vb0123x4567, va1x01234567, 0);
109 vacc2x0123 = vdotq_lane_s32(vacc2x0123, vb0123x0123, va2x01234567, 0);
110 vacc2x4567 = vdotq_lane_s32(vacc2x4567, vb0123x4567, va2x01234567, 0);
111 vacc3x0123 = vdotq_lane_s32(vacc3x0123, vb0123x0123, va3x01234567, 0);
112 vacc3x4567 = vdotq_lane_s32(vacc3x4567, vb0123x4567, va3x01234567, 0);
113 vacc0x0123 = vdotq_lane_s32(vacc0x0123, vb4567x0123, va0x01234567, 1);
114 vacc0x4567 = vdotq_lane_s32(vacc0x4567, vb4567x4567, va0x01234567, 1);
115 vacc1x0123 = vdotq_lane_s32(vacc1x0123, vb4567x0123, va1x01234567, 1);
116 vacc1x4567 = vdotq_lane_s32(vacc1x4567, vb4567x4567, va1x01234567, 1);
117 vacc2x0123 = vdotq_lane_s32(vacc2x0123, vb4567x0123, va2x01234567, 1);
118 vacc2x4567 = vdotq_lane_s32(vacc2x4567, vb4567x4567, va2x01234567, 1);
119 vacc3x0123 = vdotq_lane_s32(vacc3x0123, vb4567x0123, va3x01234567, 1);
120 vacc3x4567 = vdotq_lane_s32(vacc3x4567, vb4567x4567, va3x01234567, 1);
121
122 k -= 8 * sizeof(int8_t);
123 }
124 // Handle up to 4 final positions of `k`
125 if XNN_UNLIKELY(k != 0) {
126 // Load a 4x4 block of activations.
127 const int8x8_t va0x01234567 = vld1_s8(a0);
128 const int8x8_t va1x01234567 = vld1_s8(a1);
129 const int8x8_t va2x01234567 = vld1_s8(a2);
130 const int8x8_t va3x01234567 = vld1_s8(a3);
131
132 // Load a 4x8 block of weights.
133 const int8x16_t vb0123x0123 = vld1q_s8(w); w = (const void*) ((const int8_t*) w + 16);
134 const int8x16_t vb0123x4567 = vld1q_s8(w); w = (const void*) ((const int8_t*) w + 16);
135
136 // Multiply-accumulate: 4x4 * 4x8 --> 4x8.
137 vacc0x0123 = vdotq_lane_s32(vacc0x0123, vb0123x0123, va0x01234567, 0);
138 vacc0x4567 = vdotq_lane_s32(vacc0x4567, vb0123x4567, va0x01234567, 0);
139 vacc1x0123 = vdotq_lane_s32(vacc1x0123, vb0123x0123, va1x01234567, 0);
140 vacc1x4567 = vdotq_lane_s32(vacc1x4567, vb0123x4567, va1x01234567, 0);
141 vacc2x0123 = vdotq_lane_s32(vacc2x0123, vb0123x0123, va2x01234567, 0);
142 vacc2x4567 = vdotq_lane_s32(vacc2x4567, vb0123x4567, va2x01234567, 0);
143 vacc3x0123 = vdotq_lane_s32(vacc3x0123, vb0123x0123, va3x01234567, 0);
144 vacc3x4567 = vdotq_lane_s32(vacc3x4567, vb0123x4567, va3x01234567, 0);
145 }
146 p -= 4 * sizeof(void*);
147 } while (p != 0);
148
149 const int32x4_t vright_pre_shift = vld1q_dup_s32(¶ms->rndnu_neon.right_pre_shift);
150 const int32x4_t vmultiplier = vld1q_dup_s32(¶ms->rndnu_neon.multiplier);
151 const int32x4_t vright_post_shift = vld1q_dup_s32(¶ms->rndnu_neon.right_post_shift);
152
153 vacc0x0123 = vshlq_s32(vacc0x0123, vright_pre_shift);
154 vacc0x4567 = vshlq_s32(vacc0x4567, vright_pre_shift);
155 vacc1x0123 = vshlq_s32(vacc1x0123, vright_pre_shift);
156 vacc1x4567 = vshlq_s32(vacc1x4567, vright_pre_shift);
157 vacc2x0123 = vshlq_s32(vacc2x0123, vright_pre_shift);
158 vacc2x4567 = vshlq_s32(vacc2x4567, vright_pre_shift);
159 vacc3x0123 = vshlq_s32(vacc3x0123, vright_pre_shift);
160 vacc3x4567 = vshlq_s32(vacc3x4567, vright_pre_shift);
161
162 vacc0x0123 = vqdmulhq_s32(vacc0x0123, vmultiplier);
163 vacc0x4567 = vqdmulhq_s32(vacc0x4567, vmultiplier);
164 vacc1x0123 = vqdmulhq_s32(vacc1x0123, vmultiplier);
165 vacc1x4567 = vqdmulhq_s32(vacc1x4567, vmultiplier);
166 vacc2x0123 = vqdmulhq_s32(vacc2x0123, vmultiplier);
167 vacc2x4567 = vqdmulhq_s32(vacc2x4567, vmultiplier);
168 vacc3x0123 = vqdmulhq_s32(vacc3x0123, vmultiplier);
169 vacc3x4567 = vqdmulhq_s32(vacc3x4567, vmultiplier);
170
171 vacc0x0123 = vrshlq_s32(vacc0x0123, vright_post_shift);
172 vacc0x4567 = vrshlq_s32(vacc0x4567, vright_post_shift);
173 vacc1x0123 = vrshlq_s32(vacc1x0123, vright_post_shift);
174 vacc1x4567 = vrshlq_s32(vacc1x4567, vright_post_shift);
175 vacc2x0123 = vrshlq_s32(vacc2x0123, vright_post_shift);
176 vacc2x4567 = vrshlq_s32(vacc2x4567, vright_post_shift);
177 vacc3x0123 = vrshlq_s32(vacc3x0123, vright_post_shift);
178 vacc3x4567 = vrshlq_s32(vacc3x4567, vright_post_shift);
179
180 const int16x8_t voutput_zero_point = vld1q_dup_s16(¶ms->rndnu_neon.output_zero_point);
181 #if XNN_ARCH_ARM64
182 const int16x8_t vacc0x01234567 = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc0x0123), vacc0x4567), voutput_zero_point);
183 const int16x8_t vacc1x01234567 = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc1x0123), vacc1x4567), voutput_zero_point);
184 const int16x8_t vacc2x01234567 = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc2x0123), vacc2x4567), voutput_zero_point);
185 const int16x8_t vacc3x01234567 = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc3x0123), vacc3x4567), voutput_zero_point);
186
187 int8x16_t vout0x01234567_1x01234567 = vqmovn_high_s16(vqmovn_s16(vacc0x01234567), vacc1x01234567);
188 int8x16_t vout2x01234567_3x01234567 = vqmovn_high_s16(vqmovn_s16(vacc2x01234567), vacc3x01234567);
189 #else
190 const int16x8_t vacc0x01234567 = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc0x0123), vqmovn_s32(vacc0x4567)), voutput_zero_point);
191 const int16x8_t vacc1x01234567 = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc1x0123), vqmovn_s32(vacc1x4567)), voutput_zero_point);
192 const int16x8_t vacc2x01234567 = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc2x0123), vqmovn_s32(vacc2x4567)), voutput_zero_point);
193 const int16x8_t vacc3x01234567 = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc3x0123), vqmovn_s32(vacc3x4567)), voutput_zero_point);
194
195 int8x16_t vout0x01234567_1x01234567 = vcombine_s8(vqmovn_s16(vacc0x01234567), vqmovn_s16(vacc1x01234567));
196 int8x16_t vout2x01234567_3x01234567 = vcombine_s8(vqmovn_s16(vacc2x01234567), vqmovn_s16(vacc3x01234567));
197 #endif
198 const int8x16_t voutput_min = vld1q_dup_s8(¶ms->rndnu_neon.output_min);
199 const int8x16_t voutput_max = vld1q_dup_s8(¶ms->rndnu_neon.output_max);
200
201 vout2x01234567_3x01234567 = vmaxq_s8(vout2x01234567_3x01234567, voutput_min);
202 vout0x01234567_1x01234567 = vmaxq_s8(vout0x01234567_1x01234567, voutput_min);
203
204 vout2x01234567_3x01234567 = vminq_s8(vout2x01234567_3x01234567, voutput_max);
205 vout0x01234567_1x01234567 = vminq_s8(vout0x01234567_1x01234567, voutput_max);
206
207 if (nc >= 8) {
208 vst1_s8(c3 + 0, vget_high_s8(vout2x01234567_3x01234567));
209 vst1_s8(c2 + 0, vget_low_s8(vout2x01234567_3x01234567));
210 vst1_s8(c1 + 0, vget_high_s8(vout0x01234567_1x01234567));
211 vst1_s8(c0 + 0, vget_low_s8(vout0x01234567_1x01234567));
212
213 c3 = (int8_t*) ((uintptr_t) c3 + cn_stride);
214 c2 = (int8_t*) ((uintptr_t) c2 + cn_stride);
215 c1 = (int8_t*) ((uintptr_t) c1 + cn_stride);
216 c0 = (int8_t*) ((uintptr_t) c0 + cn_stride);
217
218 a = (const int8_t**restrict) ((uintptr_t) a - ks);
219
220 nc -= 8;
221 } else {
222 if (nc & 4) {
223 vst1q_lane_u32((void*) c3, vreinterpretq_u32_s8(vout2x01234567_3x01234567), 2); c3 += 4;
224 vst1q_lane_u32((void*) c2, vreinterpretq_u32_s8(vout2x01234567_3x01234567), 0); c2 += 4;
225 vst1q_lane_u32((void*) c1, vreinterpretq_u32_s8(vout0x01234567_1x01234567), 2); c1 += 4;
226 vst1q_lane_u32((void*) c0, vreinterpretq_u32_s8(vout0x01234567_1x01234567), 0); c0 += 4;
227 vout2x01234567_3x01234567 = vextq_s8(vout2x01234567_3x01234567, vout2x01234567_3x01234567, 4);
228 vout0x01234567_1x01234567 = vextq_s8(vout0x01234567_1x01234567, vout0x01234567_1x01234567, 4);
229 }
230 if (nc & 2) {
231 vst1q_lane_u16((void*) c3, vreinterpretq_u16_s8(vout2x01234567_3x01234567), 4); c3 += 2;
232 vst1q_lane_u16((void*) c2, vreinterpretq_u16_s8(vout2x01234567_3x01234567), 0); c2 += 2;
233 vst1q_lane_u16((void*) c1, vreinterpretq_u16_s8(vout0x01234567_1x01234567), 4); c1 += 2;
234 vst1q_lane_u16((void*) c0, vreinterpretq_u16_s8(vout0x01234567_1x01234567), 0); c0 += 2;
235 vout2x01234567_3x01234567 = vextq_s8(vout2x01234567_3x01234567, vout2x01234567_3x01234567, 2);
236 vout0x01234567_1x01234567 = vextq_s8(vout0x01234567_1x01234567, vout0x01234567_1x01234567, 2);
237 }
238 if (nc & 1) {
239 vst1q_lane_s8(c3, vout2x01234567_3x01234567, 8);
240 vst1q_lane_s8(c2, vout2x01234567_3x01234567, 0);
241 vst1q_lane_s8(c1, vout0x01234567_1x01234567, 8);
242 vst1q_lane_s8(c0, vout0x01234567_1x01234567, 0);
243 }
244
245 nc = 0;
246 }
247 } while (nc != 0);
248 }
249