1 // Auto-generated file. Do not edit!
2 // Template: src/qs8-igemm/c2-neon-mull-dup.c.in
3 // Generator: tools/xngen
4 //
5 // Copyright 2021 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
xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c2__neon_mull_dup(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)])17 void xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c2__neon_mull_dup(
18 size_t mr,
19 size_t nc,
20 size_t kc,
21 size_t ks,
22 const int8_t** restrict a,
23 const void* restrict w,
24 int8_t* restrict c,
25 size_t cm_stride,
26 size_t cn_stride,
27 size_t a_offset,
28 const int8_t* zero,
29 const union xnn_qs8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
30 {
31 assert(mr != 0);
32 assert(mr <= 2);
33 assert(nc != 0);
34 assert(kc != 0);
35 assert(ks != 0);
36 assert(ks % (2 * sizeof(void*)) == 0);
37 assert(a_offset % sizeof(int8_t) == 0);
38 assert(a != NULL);
39 assert(w != NULL);
40 assert(c != NULL);
41
42 kc = round_up_po2(kc, 2 * sizeof(int8_t));
43 int8_t* c0 = c;
44 int8_t* c1 = (int8_t*) ((uintptr_t) c0 + cm_stride);
45 if XNN_UNPREDICTABLE(mr != 2) {
46 c1 = c0;
47 }
48
49 do {
50 int32x4_t vacc0x0123 = vld1q_s32(w); w = (const void*) ((uintptr_t) w + 4 * sizeof(int32_t));
51 int32x4_t vacc0x4567 = vld1q_s32(w); w = (const void*) ((uintptr_t) w + 4 * sizeof(int32_t));
52 int32x4_t vacc1x0123 = vacc0x0123;
53 int32x4_t vacc1x4567 = vacc0x4567;
54
55 size_t p = ks;
56 do {
57 const int8_t* restrict a0 = a[0];
58 if XNN_UNPREDICTABLE(a0 != zero) {
59 a0 = (const int8_t*) ((uintptr_t) a0 + a_offset);
60 }
61 const int8_t* restrict a1 = a[1];
62 if XNN_UNPREDICTABLE(a1 != zero) {
63 a1 = (const int8_t*) ((uintptr_t) a1 + a_offset);
64 }
65 a += 2;
66
67 size_t k = kc;
68
69
70 while (k >= 8 * sizeof(int8_t)) {
71 const int8x8_t va0 = vld1_s8(a0); a0 += 8;
72 const int8x8_t va1 = vld1_s8(a1); a1 += 8;
73
74 const int8x8_t vb0123c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
75 const int8x8_t vb4567c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
76 const int8x8_t vb0123c1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
77 const int8x8_t vb4567c1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
78 const int8x8_t vb0123c2 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
79 const int8x8_t vb4567c2 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
80 const int8x8_t vb0123c3 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
81 const int8x8_t vb4567c3 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
82
83 const int8x8_t va0c0 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0), 0));
84 const int8x8_t va1c0 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va1), 0));
85
86 const int16x8_t vprod0x0123c0 = vmull_s8(vb0123c0, va0c0);
87 const int16x8_t vprod1x0123c0 = vmull_s8(vb0123c0, va1c0);
88 vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c0);
89 vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c0);
90 const int16x8_t vprod0x4567c0 = vmull_s8(vb4567c0, va0c0);
91 const int16x8_t vprod1x4567c0 = vmull_s8(vb4567c0, va1c0);
92 vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c0);
93 vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c0);
94 const int8x8_t va0c1 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0), 1));
95 const int8x8_t va1c1 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va1), 1));
96
97 const int16x8_t vprod0x0123c1 = vmull_s8(vb0123c1, va0c1);
98 const int16x8_t vprod1x0123c1 = vmull_s8(vb0123c1, va1c1);
99 vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c1);
100 vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c1);
101 const int16x8_t vprod0x4567c1 = vmull_s8(vb4567c1, va0c1);
102 const int16x8_t vprod1x4567c1 = vmull_s8(vb4567c1, va1c1);
103 vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c1);
104 vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c1);
105 const int8x8_t va0c2 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0), 2));
106 const int8x8_t va1c2 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va1), 2));
107
108 const int16x8_t vprod0x0123c2 = vmull_s8(vb0123c2, va0c2);
109 const int16x8_t vprod1x0123c2 = vmull_s8(vb0123c2, va1c2);
110 vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c2);
111 vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c2);
112 const int16x8_t vprod0x4567c2 = vmull_s8(vb4567c2, va0c2);
113 const int16x8_t vprod1x4567c2 = vmull_s8(vb4567c2, va1c2);
114 vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c2);
115 vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c2);
116 const int8x8_t va0c3 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0), 3));
117 const int8x8_t va1c3 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va1), 3));
118
119 const int16x8_t vprod0x0123c3 = vmull_s8(vb0123c3, va0c3);
120 const int16x8_t vprod1x0123c3 = vmull_s8(vb0123c3, va1c3);
121 vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c3);
122 vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c3);
123 const int16x8_t vprod0x4567c3 = vmull_s8(vb4567c3, va0c3);
124 const int16x8_t vprod1x4567c3 = vmull_s8(vb4567c3, va1c3);
125 vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c3);
126 vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c3);
127
128 k -= 8 * sizeof(int8_t);
129 }
130
131 if XNN_UNLIKELY(k != 0) {
132 const int8x8_t va0 = vld1_s8(a0); a0 = (const int8_t*) ((uintptr_t) a0 + k);
133 const int8x8_t va1 = vld1_s8(a1); a1 = (const int8_t*) ((uintptr_t) a1 + k);
134
135 const int8x8_t vb0123c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
136 const int8x8_t vb4567c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
137
138 const int8x8_t va0c0 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0), 0));
139 const int16x8_t vprod0x0123c0 = vmull_s8(vb0123c0, va0c0);
140 vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c0);
141 const int16x8_t vprod0x4567c0 = vmull_s8(vb4567c0, va0c0);
142 vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c0);
143 const int8x8_t va1c0 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va1), 0));
144 const int16x8_t vprod1x0123c0 = vmull_s8(vb0123c0, va1c0);
145 vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c0);
146 const int16x8_t vprod1x4567c0 = vmull_s8(vb4567c0, va1c0);
147 vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c0);
148
149 if (k > 2 * sizeof(int8_t)) {
150 const int8x8_t vb0123c1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
151 const int8x8_t vb4567c1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
152
153 const int8x8_t va0c1 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0), 1));
154 const int16x8_t vprod0x0123c1 = vmull_s8(vb0123c1, va0c1);
155 vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c1);
156 const int16x8_t vprod0x4567c1 = vmull_s8(vb4567c1, va0c1);
157 vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c1);
158 const int8x8_t va1c1 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va1), 1));
159 const int16x8_t vprod1x0123c1 = vmull_s8(vb0123c1, va1c1);
160 vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c1);
161 const int16x8_t vprod1x4567c1 = vmull_s8(vb4567c1, va1c1);
162 vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c1);
163
164 if (k > 4 * sizeof(int8_t)) {
165 const int8x8_t vb0123c2 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
166 const int8x8_t vb4567c2 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
167
168 const int8x8_t va0c2 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0), 2));
169 const int16x8_t vprod0x0123c2 = vmull_s8(vb0123c2, va0c2);
170 vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c2);
171 const int16x8_t vprod0x4567c2 = vmull_s8(vb4567c2, va0c2);
172 vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c2);
173 const int8x8_t va1c2 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va1), 2));
174 const int16x8_t vprod1x0123c2 = vmull_s8(vb0123c2, va1c2);
175 vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c2);
176 const int16x8_t vprod1x4567c2 = vmull_s8(vb4567c2, va1c2);
177 vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c2);
178 }
179 }
180 }
181 p -= 2 * sizeof(void*);
182 } while (p != 0);
183
184 const int32x4_t vright_pre_shift = vld1q_dup_s32(¶ms->rndnu_neon.right_pre_shift);
185 const int32x4_t vmultiplier = vld1q_dup_s32(¶ms->rndnu_neon.multiplier);
186 const int32x4_t vright_post_shift = vld1q_dup_s32(¶ms->rndnu_neon.right_post_shift);
187
188 vacc0x0123 = vqshlq_s32(vacc0x0123, vright_pre_shift);
189 vacc0x4567 = vqshlq_s32(vacc0x4567, vright_pre_shift);
190 vacc1x0123 = vqshlq_s32(vacc1x0123, vright_pre_shift);
191 vacc1x4567 = vqshlq_s32(vacc1x4567, vright_pre_shift);
192
193 vacc0x0123 = vqdmulhq_s32(vacc0x0123, vmultiplier);
194 vacc0x4567 = vqdmulhq_s32(vacc0x4567, vmultiplier);
195 vacc1x0123 = vqdmulhq_s32(vacc1x0123, vmultiplier);
196 vacc1x4567 = vqdmulhq_s32(vacc1x4567, vmultiplier);
197
198 vacc0x0123 = vrshlq_s32(vacc0x0123, vright_post_shift);
199 vacc0x4567 = vrshlq_s32(vacc0x4567, vright_post_shift);
200 vacc1x0123 = vrshlq_s32(vacc1x0123, vright_post_shift);
201 vacc1x4567 = vrshlq_s32(vacc1x4567, vright_post_shift);
202
203 const int16x8_t voutput_zero_point = vld1q_dup_s16(¶ms->rndnu_neon.output_zero_point);
204 #if XNN_ARCH_ARM64
205 int16x8_t vacc0x01234567 = vqmovn_high_s32(vqmovn_s32(vacc0x0123), vacc0x4567);
206 int16x8_t vacc1x01234567 = vqmovn_high_s32(vqmovn_s32(vacc1x0123), vacc1x4567);
207
208 vacc0x01234567 = vqaddq_s16(vacc0x01234567, voutput_zero_point);
209 vacc1x01234567 = vqaddq_s16(vacc1x01234567, voutput_zero_point);
210
211 int8x16_t vout0x01234567_1x01234567 = vqmovn_high_s16(vqmovn_s16(vacc0x01234567), vacc1x01234567);
212 #else
213 int16x8_t vacc0x01234567 = vcombine_s16(vqmovn_s32(vacc0x0123), vqmovn_s32(vacc0x4567));
214 int16x8_t vacc1x01234567 = vcombine_s16(vqmovn_s32(vacc1x0123), vqmovn_s32(vacc1x4567));
215
216 vacc0x01234567 = vqaddq_s16(vacc0x01234567, voutput_zero_point);
217 vacc1x01234567 = vqaddq_s16(vacc1x01234567, voutput_zero_point);
218
219 int8x16_t vout0x01234567_1x01234567 = vcombine_s8(vqmovn_s16(vacc0x01234567), vqmovn_s16(vacc1x01234567));
220 #endif
221
222 const int8x16_t voutput_min = vld1q_dup_s8(¶ms->rndnu_neon.output_min);
223 vout0x01234567_1x01234567 = vmaxq_s8(vout0x01234567_1x01234567, voutput_min);
224
225 const int8x16_t voutput_max = vld1q_dup_s8(¶ms->rndnu_neon.output_max);
226 vout0x01234567_1x01234567 = vminq_s8(vout0x01234567_1x01234567, voutput_max);
227
228 if (nc >= 8) {
229 vst1_s8(c1 + 0, vget_high_s8(vout0x01234567_1x01234567));
230 vst1_s8(c0 + 0, vget_low_s8(vout0x01234567_1x01234567));
231
232 c1 = (int8_t*) ((uintptr_t) c1 + cn_stride);
233 c0 = (int8_t*) ((uintptr_t) c0 + cn_stride);
234
235 a = (const int8_t**restrict) ((uintptr_t) a - ks);
236
237 nc -= 8;
238 } else {
239 if (nc & 4) {
240 vst1q_lane_u32((void*) c1, vreinterpretq_u32_s8(vout0x01234567_1x01234567), 2); c1 += 4;
241 vst1q_lane_u32((void*) c0, vreinterpretq_u32_s8(vout0x01234567_1x01234567), 0); c0 += 4;
242 vout0x01234567_1x01234567 = vextq_s8(vout0x01234567_1x01234567, vout0x01234567_1x01234567, 4);
243 }
244 if (nc & 2) {
245 vst1q_lane_u16((void*) c1, vreinterpretq_u16_s8(vout0x01234567_1x01234567), 4); c1 += 2;
246 vst1q_lane_u16((void*) c0, vreinterpretq_u16_s8(vout0x01234567_1x01234567), 0); c0 += 2;
247 vout0x01234567_1x01234567 = vextq_s8(vout0x01234567_1x01234567, vout0x01234567_1x01234567, 2);
248 }
249 if (nc & 1) {
250 vst1q_lane_s8(c1, vout0x01234567_1x01234567, 8);
251 vst1q_lane_s8(c0, vout0x01234567_1x01234567, 0);
252 }
253
254 nc = 0;
255 }
256 } while (nc != 0);
257 }
258