1 // Auto-generated file. Do not edit!
2 // Template: src/qs8-gemm/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_gemm_minmax_rndnu_ukernel_3x8c2__neon_mull_ld1r(size_t mr,size_t nc,size_t kc,const int8_t * restrict a,size_t a_stride,const void * restrict w,int8_t * restrict c,size_t cm_stride,size_t cn_stride,const union xnn_qs8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS (1)])17 void xnn_qs8_gemm_minmax_rndnu_ukernel_3x8c2__neon_mull_ld1r(
18 size_t mr,
19 size_t nc,
20 size_t kc,
21 const int8_t* restrict a,
22 size_t a_stride,
23 const void* restrict w,
24 int8_t* restrict c,
25 size_t cm_stride,
26 size_t cn_stride,
27 const union xnn_qs8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
28 {
29 assert(mr != 0);
30 assert(mr <= 3);
31 assert(nc != 0);
32 assert(kc != 0);
33 assert(kc % sizeof(int8_t) == 0);
34 assert(a != NULL);
35 assert(w != NULL);
36 assert(c != NULL);
37
38 kc = round_up_po2(kc, 2 * sizeof(int8_t));
39 const int8_t* a0 = a;
40 int8_t* c0 = c;
41 const int8_t* a1 = (const int8_t*) ((uintptr_t) a0 + a_stride);
42 int8_t* c1 = (int8_t*) ((uintptr_t) c0 + cm_stride);
43 if XNN_UNPREDICTABLE(mr < 2) {
44 a1 = a0;
45 c1 = c0;
46 }
47 const int8_t* a2 = (const int8_t*) ((uintptr_t) a1 + a_stride);
48 int8_t* c2 = (int8_t*) ((uintptr_t) c1 + cm_stride);
49 if XNN_UNPREDICTABLE(mr <= 2) {
50 a2 = a1;
51 c2 = c1;
52 }
53
54 do {
55 int32x4_t vacc0x0123 = vld1q_s32(w); w = (const void*) ((uintptr_t) w + 4 * sizeof(int32_t));
56 int32x4_t vacc0x4567 = vld1q_s32(w); w = (const void*) ((uintptr_t) w + 4 * sizeof(int32_t));
57 int32x4_t vacc1x0123 = vacc0x0123;
58 int32x4_t vacc1x4567 = vacc0x4567;
59 int32x4_t vacc2x0123 = vacc0x0123;
60 int32x4_t vacc2x4567 = vacc0x4567;
61
62 size_t k = kc;
63
64
65 while (k >= 8 * sizeof(int8_t)) {
66 const int16x4_t va00 = vld1_dup_s16((const void*)a0);
67 const int16x4_t va01 = vld1_dup_s16((const void*)(a0 + 2));
68 const int16x4_t va02 = vld1_dup_s16((const void*)(a0 + 4));
69 const int16x4_t va03 = vld1_dup_s16((const void*)(a0 + 6)); a0 += 8;
70 const int16x4_t va10 = vld1_dup_s16((const void*)a1);
71 const int16x4_t va11 = vld1_dup_s16((const void*)(a1 + 2));
72 const int16x4_t va12 = vld1_dup_s16((const void*)(a1 + 4));
73 const int16x4_t va13 = vld1_dup_s16((const void*)(a1 + 6)); a1 += 8;
74 const int16x4_t va20 = vld1_dup_s16((const void*)a2);
75 const int16x4_t va21 = vld1_dup_s16((const void*)(a2 + 2));
76 const int16x4_t va22 = vld1_dup_s16((const void*)(a2 + 4));
77 const int16x4_t va23 = vld1_dup_s16((const void*)(a2 + 6)); a2 += 8;
78
79 const int8x8_t vb0123c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
80 const int8x8_t vb4567c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
81 const int8x8_t vb0123c1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
82 const int8x8_t vb4567c1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
83 const int8x8_t vb0123c2 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
84 const int8x8_t vb4567c2 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
85 const int8x8_t vb0123c3 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
86 const int8x8_t vb4567c3 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
87
88 const int8x8_t va0c0 = vreinterpret_s8_s16(va00);
89 const int8x8_t va1c0 = vreinterpret_s8_s16(va10);
90 const int8x8_t va2c0 = vreinterpret_s8_s16(va20);
91
92 const int16x8_t vprod0x0123c0 = vmull_s8(vb0123c0, va0c0);
93 const int16x8_t vprod1x0123c0 = vmull_s8(vb0123c0, va1c0);
94 const int16x8_t vprod2x0123c0 = vmull_s8(vb0123c0, va2c0);
95 vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c0);
96 vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c0);
97 vacc2x0123 = vpadalq_s16(vacc2x0123, vprod2x0123c0);
98 const int16x8_t vprod0x4567c0 = vmull_s8(vb4567c0, va0c0);
99 const int16x8_t vprod1x4567c0 = vmull_s8(vb4567c0, va1c0);
100 const int16x8_t vprod2x4567c0 = vmull_s8(vb4567c0, va2c0);
101 vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c0);
102 vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c0);
103 vacc2x4567 = vpadalq_s16(vacc2x4567, vprod2x4567c0);
104 const int8x8_t va0c1 = vreinterpret_s8_s16(va01);
105 const int8x8_t va1c1 = vreinterpret_s8_s16(va11);
106 const int8x8_t va2c1 = vreinterpret_s8_s16(va21);
107
108 const int16x8_t vprod0x0123c1 = vmull_s8(vb0123c1, va0c1);
109 const int16x8_t vprod1x0123c1 = vmull_s8(vb0123c1, va1c1);
110 const int16x8_t vprod2x0123c1 = vmull_s8(vb0123c1, va2c1);
111 vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c1);
112 vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c1);
113 vacc2x0123 = vpadalq_s16(vacc2x0123, vprod2x0123c1);
114 const int16x8_t vprod0x4567c1 = vmull_s8(vb4567c1, va0c1);
115 const int16x8_t vprod1x4567c1 = vmull_s8(vb4567c1, va1c1);
116 const int16x8_t vprod2x4567c1 = vmull_s8(vb4567c1, va2c1);
117 vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c1);
118 vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c1);
119 vacc2x4567 = vpadalq_s16(vacc2x4567, vprod2x4567c1);
120 const int8x8_t va0c2 = vreinterpret_s8_s16(va02);
121 const int8x8_t va1c2 = vreinterpret_s8_s16(va12);
122 const int8x8_t va2c2 = vreinterpret_s8_s16(va22);
123
124 const int16x8_t vprod0x0123c2 = vmull_s8(vb0123c2, va0c2);
125 const int16x8_t vprod1x0123c2 = vmull_s8(vb0123c2, va1c2);
126 const int16x8_t vprod2x0123c2 = vmull_s8(vb0123c2, va2c2);
127 vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c2);
128 vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c2);
129 vacc2x0123 = vpadalq_s16(vacc2x0123, vprod2x0123c2);
130 const int16x8_t vprod0x4567c2 = vmull_s8(vb4567c2, va0c2);
131 const int16x8_t vprod1x4567c2 = vmull_s8(vb4567c2, va1c2);
132 const int16x8_t vprod2x4567c2 = vmull_s8(vb4567c2, va2c2);
133 vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c2);
134 vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c2);
135 vacc2x4567 = vpadalq_s16(vacc2x4567, vprod2x4567c2);
136 const int8x8_t va0c3 = vreinterpret_s8_s16(va03);
137 const int8x8_t va1c3 = vreinterpret_s8_s16(va13);
138 const int8x8_t va2c3 = vreinterpret_s8_s16(va23);
139
140 const int16x8_t vprod0x0123c3 = vmull_s8(vb0123c3, va0c3);
141 const int16x8_t vprod1x0123c3 = vmull_s8(vb0123c3, va1c3);
142 const int16x8_t vprod2x0123c3 = vmull_s8(vb0123c3, va2c3);
143 vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c3);
144 vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c3);
145 vacc2x0123 = vpadalq_s16(vacc2x0123, vprod2x0123c3);
146 const int16x8_t vprod0x4567c3 = vmull_s8(vb4567c3, va0c3);
147 const int16x8_t vprod1x4567c3 = vmull_s8(vb4567c3, va1c3);
148 const int16x8_t vprod2x4567c3 = vmull_s8(vb4567c3, va2c3);
149 vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c3);
150 vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c3);
151 vacc2x4567 = vpadalq_s16(vacc2x4567, vprod2x4567c3);
152
153 k -= 8 * sizeof(int8_t);
154 }
155
156 if XNN_UNLIKELY(k != 0) {
157 const int8x8_t va0 = vld1_s8(a0); a0 = (const int8_t*) ((uintptr_t) a0 + k);
158 const int8x8_t va1 = vld1_s8(a1); a1 = (const int8_t*) ((uintptr_t) a1 + k);
159 const int8x8_t va2 = vld1_s8(a2); a2 = (const int8_t*) ((uintptr_t) a2 + k);
160
161 const int8x8_t vb0123c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
162 const int8x8_t vb4567c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
163
164 const int8x8_t va0c0 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0), 0));
165 const int16x8_t vprod0x0123c0 = vmull_s8(vb0123c0, va0c0);
166 vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c0);
167 const int16x8_t vprod0x4567c0 = vmull_s8(vb4567c0, va0c0);
168 vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c0);
169 const int8x8_t va1c0 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va1), 0));
170 const int16x8_t vprod1x0123c0 = vmull_s8(vb0123c0, va1c0);
171 vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c0);
172 const int16x8_t vprod1x4567c0 = vmull_s8(vb4567c0, va1c0);
173 vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c0);
174 const int8x8_t va2c0 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va2), 0));
175 const int16x8_t vprod2x0123c0 = vmull_s8(vb0123c0, va2c0);
176 vacc2x0123 = vpadalq_s16(vacc2x0123, vprod2x0123c0);
177 const int16x8_t vprod2x4567c0 = vmull_s8(vb4567c0, va2c0);
178 vacc2x4567 = vpadalq_s16(vacc2x4567, vprod2x4567c0);
179
180 if (k > 2 * sizeof(int8_t)) {
181 const int8x8_t vb0123c1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
182 const int8x8_t vb4567c1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
183
184 const int8x8_t va0c1 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0), 1));
185 const int16x8_t vprod0x0123c1 = vmull_s8(vb0123c1, va0c1);
186 vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c1);
187 const int16x8_t vprod0x4567c1 = vmull_s8(vb4567c1, va0c1);
188 vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c1);
189 const int8x8_t va1c1 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va1), 1));
190 const int16x8_t vprod1x0123c1 = vmull_s8(vb0123c1, va1c1);
191 vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c1);
192 const int16x8_t vprod1x4567c1 = vmull_s8(vb4567c1, va1c1);
193 vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c1);
194 const int8x8_t va2c1 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va2), 1));
195 const int16x8_t vprod2x0123c1 = vmull_s8(vb0123c1, va2c1);
196 vacc2x0123 = vpadalq_s16(vacc2x0123, vprod2x0123c1);
197 const int16x8_t vprod2x4567c1 = vmull_s8(vb4567c1, va2c1);
198 vacc2x4567 = vpadalq_s16(vacc2x4567, vprod2x4567c1);
199
200 if (k > 4 * sizeof(int8_t)) {
201 const int8x8_t vb0123c2 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
202 const int8x8_t vb4567c2 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
203
204 const int8x8_t va0c2 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0), 2));
205 const int16x8_t vprod0x0123c2 = vmull_s8(vb0123c2, va0c2);
206 vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c2);
207 const int16x8_t vprod0x4567c2 = vmull_s8(vb4567c2, va0c2);
208 vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c2);
209 const int8x8_t va1c2 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va1), 2));
210 const int16x8_t vprod1x0123c2 = vmull_s8(vb0123c2, va1c2);
211 vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c2);
212 const int16x8_t vprod1x4567c2 = vmull_s8(vb4567c2, va1c2);
213 vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c2);
214 const int8x8_t va2c2 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va2), 2));
215 const int16x8_t vprod2x0123c2 = vmull_s8(vb0123c2, va2c2);
216 vacc2x0123 = vpadalq_s16(vacc2x0123, vprod2x0123c2);
217 const int16x8_t vprod2x4567c2 = vmull_s8(vb4567c2, va2c2);
218 vacc2x4567 = vpadalq_s16(vacc2x4567, vprod2x4567c2);
219 }
220 }
221 }
222
223 const int32x4_t vright_pre_shift = vld1q_dup_s32(¶ms->rndnu_neon.right_pre_shift);
224 const int32x4_t vmultiplier = vld1q_dup_s32(¶ms->rndnu_neon.multiplier);
225 const int32x4_t vright_post_shift = vld1q_dup_s32(¶ms->rndnu_neon.right_post_shift);
226
227 vacc0x0123 = vqshlq_s32(vacc0x0123, vright_pre_shift);
228 vacc0x4567 = vqshlq_s32(vacc0x4567, vright_pre_shift);
229 vacc1x0123 = vqshlq_s32(vacc1x0123, vright_pre_shift);
230 vacc1x4567 = vqshlq_s32(vacc1x4567, vright_pre_shift);
231 vacc2x0123 = vqshlq_s32(vacc2x0123, vright_pre_shift);
232 vacc2x4567 = vqshlq_s32(vacc2x4567, vright_pre_shift);
233
234 vacc0x0123 = vqdmulhq_s32(vacc0x0123, vmultiplier);
235 vacc0x4567 = vqdmulhq_s32(vacc0x4567, vmultiplier);
236 vacc1x0123 = vqdmulhq_s32(vacc1x0123, vmultiplier);
237 vacc1x4567 = vqdmulhq_s32(vacc1x4567, vmultiplier);
238 vacc2x0123 = vqdmulhq_s32(vacc2x0123, vmultiplier);
239 vacc2x4567 = vqdmulhq_s32(vacc2x4567, vmultiplier);
240
241 vacc0x0123 = vrshlq_s32(vacc0x0123, vright_post_shift);
242 vacc0x4567 = vrshlq_s32(vacc0x4567, vright_post_shift);
243 vacc1x0123 = vrshlq_s32(vacc1x0123, vright_post_shift);
244 vacc1x4567 = vrshlq_s32(vacc1x4567, vright_post_shift);
245 vacc2x0123 = vrshlq_s32(vacc2x0123, vright_post_shift);
246 vacc2x4567 = vrshlq_s32(vacc2x4567, vright_post_shift);
247
248 const int16x8_t voutput_zero_point = vld1q_dup_s16(¶ms->rndnu_neon.output_zero_point);
249 #if XNN_ARCH_ARM64
250 int16x8_t vacc0x01234567 = vqmovn_high_s32(vqmovn_s32(vacc0x0123), vacc0x4567);
251 int16x8_t vacc1x01234567 = vqmovn_high_s32(vqmovn_s32(vacc1x0123), vacc1x4567);
252 int16x8_t vacc2x01234567 = vqmovn_high_s32(vqmovn_s32(vacc2x0123), vacc2x4567);
253
254 vacc0x01234567 = vqaddq_s16(vacc0x01234567, voutput_zero_point);
255 vacc1x01234567 = vqaddq_s16(vacc1x01234567, voutput_zero_point);
256 vacc2x01234567 = vqaddq_s16(vacc2x01234567, voutput_zero_point);
257
258 int8x16_t vout0x01234567_1x01234567 = vqmovn_high_s16(vqmovn_s16(vacc0x01234567), vacc1x01234567);
259 int8x8_t vout2x01234567 = vqmovn_s16(vacc2x01234567);
260 #else
261 int16x8_t vacc0x01234567 = vcombine_s16(vqmovn_s32(vacc0x0123), vqmovn_s32(vacc0x4567));
262 int16x8_t vacc1x01234567 = vcombine_s16(vqmovn_s32(vacc1x0123), vqmovn_s32(vacc1x4567));
263 int16x8_t vacc2x01234567 = vcombine_s16(vqmovn_s32(vacc2x0123), vqmovn_s32(vacc2x4567));
264
265 vacc0x01234567 = vqaddq_s16(vacc0x01234567, voutput_zero_point);
266 vacc1x01234567 = vqaddq_s16(vacc1x01234567, voutput_zero_point);
267 vacc2x01234567 = vqaddq_s16(vacc2x01234567, voutput_zero_point);
268
269 int8x16_t vout0x01234567_1x01234567 = vcombine_s8(vqmovn_s16(vacc0x01234567), vqmovn_s16(vacc1x01234567));
270 int8x8_t vout2x01234567 = vqmovn_s16(vacc2x01234567);
271 #endif
272
273 const int8x16_t voutput_min = vld1q_dup_s8(¶ms->rndnu_neon.output_min);
274 vout0x01234567_1x01234567 = vmaxq_s8(vout0x01234567_1x01234567, voutput_min);
275 vout2x01234567 = vmax_s8(vout2x01234567, vget_low_s8(voutput_min));
276
277 const int8x16_t voutput_max = vld1q_dup_s8(¶ms->rndnu_neon.output_max);
278 vout0x01234567_1x01234567 = vminq_s8(vout0x01234567_1x01234567, voutput_max);
279 vout2x01234567 = vmin_s8(vout2x01234567, vget_low_s8(voutput_max));
280
281 if (nc >= 8) {
282 vst1_s8(c0 + 0, vget_low_s8(vout0x01234567_1x01234567));
283 vst1_s8(c1 + 0, vget_high_s8(vout0x01234567_1x01234567));
284 vst1_s8(c2 + 0, vout2x01234567);
285
286 c0 = (int8_t*) ((uintptr_t) c0 + cn_stride);
287 c1 = (int8_t*) ((uintptr_t) c1 + cn_stride);
288 c2 = (int8_t*) ((uintptr_t) c2 + cn_stride);
289
290 a0 = (const int8_t*) ((uintptr_t) a0 - kc);
291 a1 = (const int8_t*) ((uintptr_t) a1 - kc);
292 a2 = (const int8_t*) ((uintptr_t) a2 - kc);
293
294 nc -= 8;
295 } else {
296 // Final case where not all of the 8 columns fit in the destination.
297 if (nc & 4) {
298 vst1q_lane_u32((void*) c0, vreinterpretq_u32_s8(vout0x01234567_1x01234567), 0); c0 += 4;
299 vst1q_lane_u32((void*) c1, vreinterpretq_u32_s8(vout0x01234567_1x01234567), 2); c1 += 4;
300 vst1_lane_u32((void*) c2, vreinterpret_u32_s8(vout2x01234567), 0); c2 += 4;
301 vout0x01234567_1x01234567 = vextq_s8(vout0x01234567_1x01234567, vout0x01234567_1x01234567, 4);
302 vout2x01234567 = vext_s8(vout2x01234567, vout2x01234567, 4);
303 }
304 if (nc & 2) {
305 vst1q_lane_u16((void*) c0, vreinterpretq_u16_s8(vout0x01234567_1x01234567), 0); c0 += 2;
306 vst1q_lane_u16((void*) c1, vreinterpretq_u16_s8(vout0x01234567_1x01234567), 4); c1 += 2;
307 vst1_lane_u16((void*) c2, vreinterpret_u16_s8(vout2x01234567), 0); c2 += 2;
308 vout0x01234567_1x01234567 = vextq_s8(vout0x01234567_1x01234567, vout0x01234567_1x01234567, 2);
309 vout2x01234567 = vext_s8(vout2x01234567, vout2x01234567, 2);
310 }
311 if (nc & 1) {
312 vst1q_lane_s8(c0, vout0x01234567_1x01234567, 0);
313 vst1q_lane_s8(c1, vout0x01234567_1x01234567, 8);
314 vst1_lane_s8(c2, vout2x01234567, 0);
315 }
316
317 nc = 0;
318 }
319 } while (nc != 0);
320 }
321