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_ld2r(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_ld2r(
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 int16x4x2_t va00 = vld2_dup_s16((const void*)a0);
67 const int16x4x2_t va01 = vld2_dup_s16((const void*)(a0 + 4)); a0 += 8;
68 const int16x4x2_t va10 = vld2_dup_s16((const void*)a1);
69 const int16x4x2_t va11 = vld2_dup_s16((const void*)(a1 + 4)); a1 += 8;
70 const int16x4x2_t va20 = vld2_dup_s16((const void*)a2);
71 const int16x4x2_t va21 = vld2_dup_s16((const void*)(a2 + 4)); a2 += 8;
72
73 const int8x8_t vb0123c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
74 const int8x8_t vb4567c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
75 const int8x8_t vb0123c1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
76 const int8x8_t vb4567c1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
77 const int8x8_t vb0123c2 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
78 const int8x8_t vb4567c2 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
79 const int8x8_t vb0123c3 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
80 const int8x8_t vb4567c3 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
81
82 const int8x8_t va0c0 = vreinterpret_s8_s16(va00.val[0]);
83 const int8x8_t va1c0 = vreinterpret_s8_s16(va10.val[0]);
84 const int8x8_t va2c0 = vreinterpret_s8_s16(va20.val[0]);
85
86 const int16x8_t vprod0x0123c0 = vmull_s8(vb0123c0, va0c0);
87 const int16x8_t vprod1x0123c0 = vmull_s8(vb0123c0, va1c0);
88 const int16x8_t vprod2x0123c0 = vmull_s8(vb0123c0, va2c0);
89 vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c0);
90 vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c0);
91 vacc2x0123 = vpadalq_s16(vacc2x0123, vprod2x0123c0);
92 const int16x8_t vprod0x4567c0 = vmull_s8(vb4567c0, va0c0);
93 const int16x8_t vprod1x4567c0 = vmull_s8(vb4567c0, va1c0);
94 const int16x8_t vprod2x4567c0 = vmull_s8(vb4567c0, va2c0);
95 vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c0);
96 vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c0);
97 vacc2x4567 = vpadalq_s16(vacc2x4567, vprod2x4567c0);
98 const int8x8_t va0c1 = vreinterpret_s8_s16(va00.val[1]);
99 const int8x8_t va1c1 = vreinterpret_s8_s16(va10.val[1]);
100 const int8x8_t va2c1 = vreinterpret_s8_s16(va20.val[1]);
101
102 const int16x8_t vprod0x0123c1 = vmull_s8(vb0123c1, va0c1);
103 const int16x8_t vprod1x0123c1 = vmull_s8(vb0123c1, va1c1);
104 const int16x8_t vprod2x0123c1 = vmull_s8(vb0123c1, va2c1);
105 vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c1);
106 vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c1);
107 vacc2x0123 = vpadalq_s16(vacc2x0123, vprod2x0123c1);
108 const int16x8_t vprod0x4567c1 = vmull_s8(vb4567c1, va0c1);
109 const int16x8_t vprod1x4567c1 = vmull_s8(vb4567c1, va1c1);
110 const int16x8_t vprod2x4567c1 = vmull_s8(vb4567c1, va2c1);
111 vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c1);
112 vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c1);
113 vacc2x4567 = vpadalq_s16(vacc2x4567, vprod2x4567c1);
114 const int8x8_t va0c2 = vreinterpret_s8_s16(va01.val[0]);
115 const int8x8_t va1c2 = vreinterpret_s8_s16(va11.val[0]);
116 const int8x8_t va2c2 = vreinterpret_s8_s16(va21.val[0]);
117
118 const int16x8_t vprod0x0123c2 = vmull_s8(vb0123c2, va0c2);
119 const int16x8_t vprod1x0123c2 = vmull_s8(vb0123c2, va1c2);
120 const int16x8_t vprod2x0123c2 = vmull_s8(vb0123c2, va2c2);
121 vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c2);
122 vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c2);
123 vacc2x0123 = vpadalq_s16(vacc2x0123, vprod2x0123c2);
124 const int16x8_t vprod0x4567c2 = vmull_s8(vb4567c2, va0c2);
125 const int16x8_t vprod1x4567c2 = vmull_s8(vb4567c2, va1c2);
126 const int16x8_t vprod2x4567c2 = vmull_s8(vb4567c2, va2c2);
127 vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c2);
128 vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c2);
129 vacc2x4567 = vpadalq_s16(vacc2x4567, vprod2x4567c2);
130 const int8x8_t va0c3 = vreinterpret_s8_s16(va01.val[1]);
131 const int8x8_t va1c3 = vreinterpret_s8_s16(va11.val[1]);
132 const int8x8_t va2c3 = vreinterpret_s8_s16(va21.val[1]);
133
134 const int16x8_t vprod0x0123c3 = vmull_s8(vb0123c3, va0c3);
135 const int16x8_t vprod1x0123c3 = vmull_s8(vb0123c3, va1c3);
136 const int16x8_t vprod2x0123c3 = vmull_s8(vb0123c3, va2c3);
137 vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c3);
138 vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c3);
139 vacc2x0123 = vpadalq_s16(vacc2x0123, vprod2x0123c3);
140 const int16x8_t vprod0x4567c3 = vmull_s8(vb4567c3, va0c3);
141 const int16x8_t vprod1x4567c3 = vmull_s8(vb4567c3, va1c3);
142 const int16x8_t vprod2x4567c3 = vmull_s8(vb4567c3, va2c3);
143 vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c3);
144 vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c3);
145 vacc2x4567 = vpadalq_s16(vacc2x4567, vprod2x4567c3);
146
147 k -= 8 * sizeof(int8_t);
148 }
149
150 if XNN_UNLIKELY(k != 0) {
151 const int8x8_t va0 = vld1_s8(a0); a0 = (const int8_t*) ((uintptr_t) a0 + k);
152 const int8x8_t va1 = vld1_s8(a1); a1 = (const int8_t*) ((uintptr_t) a1 + k);
153 const int8x8_t va2 = vld1_s8(a2); a2 = (const int8_t*) ((uintptr_t) a2 + k);
154
155 const int8x8_t vb0123c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
156 const int8x8_t vb4567c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
157
158 const int8x8_t va0c0 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0), 0));
159 const int16x8_t vprod0x0123c0 = vmull_s8(vb0123c0, va0c0);
160 vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c0);
161 const int16x8_t vprod0x4567c0 = vmull_s8(vb4567c0, va0c0);
162 vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c0);
163 const int8x8_t va1c0 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va1), 0));
164 const int16x8_t vprod1x0123c0 = vmull_s8(vb0123c0, va1c0);
165 vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c0);
166 const int16x8_t vprod1x4567c0 = vmull_s8(vb4567c0, va1c0);
167 vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c0);
168 const int8x8_t va2c0 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va2), 0));
169 const int16x8_t vprod2x0123c0 = vmull_s8(vb0123c0, va2c0);
170 vacc2x0123 = vpadalq_s16(vacc2x0123, vprod2x0123c0);
171 const int16x8_t vprod2x4567c0 = vmull_s8(vb4567c0, va2c0);
172 vacc2x4567 = vpadalq_s16(vacc2x4567, vprod2x4567c0);
173
174 if (k > 2 * sizeof(int8_t)) {
175 const int8x8_t vb0123c1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
176 const int8x8_t vb4567c1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
177
178 const int8x8_t va0c1 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0), 1));
179 const int16x8_t vprod0x0123c1 = vmull_s8(vb0123c1, va0c1);
180 vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c1);
181 const int16x8_t vprod0x4567c1 = vmull_s8(vb4567c1, va0c1);
182 vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c1);
183 const int8x8_t va1c1 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va1), 1));
184 const int16x8_t vprod1x0123c1 = vmull_s8(vb0123c1, va1c1);
185 vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c1);
186 const int16x8_t vprod1x4567c1 = vmull_s8(vb4567c1, va1c1);
187 vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c1);
188 const int8x8_t va2c1 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va2), 1));
189 const int16x8_t vprod2x0123c1 = vmull_s8(vb0123c1, va2c1);
190 vacc2x0123 = vpadalq_s16(vacc2x0123, vprod2x0123c1);
191 const int16x8_t vprod2x4567c1 = vmull_s8(vb4567c1, va2c1);
192 vacc2x4567 = vpadalq_s16(vacc2x4567, vprod2x4567c1);
193
194 if (k > 4 * sizeof(int8_t)) {
195 const int8x8_t vb0123c2 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
196 const int8x8_t vb4567c2 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
197
198 const int8x8_t va0c2 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0), 2));
199 const int16x8_t vprod0x0123c2 = vmull_s8(vb0123c2, va0c2);
200 vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c2);
201 const int16x8_t vprod0x4567c2 = vmull_s8(vb4567c2, va0c2);
202 vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c2);
203 const int8x8_t va1c2 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va1), 2));
204 const int16x8_t vprod1x0123c2 = vmull_s8(vb0123c2, va1c2);
205 vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c2);
206 const int16x8_t vprod1x4567c2 = vmull_s8(vb4567c2, va1c2);
207 vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c2);
208 const int8x8_t va2c2 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va2), 2));
209 const int16x8_t vprod2x0123c2 = vmull_s8(vb0123c2, va2c2);
210 vacc2x0123 = vpadalq_s16(vacc2x0123, vprod2x0123c2);
211 const int16x8_t vprod2x4567c2 = vmull_s8(vb4567c2, va2c2);
212 vacc2x4567 = vpadalq_s16(vacc2x4567, vprod2x4567c2);
213 }
214 }
215 }
216
217 const int32x4_t vright_pre_shift = vld1q_dup_s32(¶ms->rndnu_neon.right_pre_shift);
218 const int32x4_t vmultiplier = vld1q_dup_s32(¶ms->rndnu_neon.multiplier);
219 const int32x4_t vright_post_shift = vld1q_dup_s32(¶ms->rndnu_neon.right_post_shift);
220
221 vacc0x0123 = vqshlq_s32(vacc0x0123, vright_pre_shift);
222 vacc0x4567 = vqshlq_s32(vacc0x4567, vright_pre_shift);
223 vacc1x0123 = vqshlq_s32(vacc1x0123, vright_pre_shift);
224 vacc1x4567 = vqshlq_s32(vacc1x4567, vright_pre_shift);
225 vacc2x0123 = vqshlq_s32(vacc2x0123, vright_pre_shift);
226 vacc2x4567 = vqshlq_s32(vacc2x4567, vright_pre_shift);
227
228 vacc0x0123 = vqdmulhq_s32(vacc0x0123, vmultiplier);
229 vacc0x4567 = vqdmulhq_s32(vacc0x4567, vmultiplier);
230 vacc1x0123 = vqdmulhq_s32(vacc1x0123, vmultiplier);
231 vacc1x4567 = vqdmulhq_s32(vacc1x4567, vmultiplier);
232 vacc2x0123 = vqdmulhq_s32(vacc2x0123, vmultiplier);
233 vacc2x4567 = vqdmulhq_s32(vacc2x4567, vmultiplier);
234
235 vacc0x0123 = vrshlq_s32(vacc0x0123, vright_post_shift);
236 vacc0x4567 = vrshlq_s32(vacc0x4567, vright_post_shift);
237 vacc1x0123 = vrshlq_s32(vacc1x0123, vright_post_shift);
238 vacc1x4567 = vrshlq_s32(vacc1x4567, vright_post_shift);
239 vacc2x0123 = vrshlq_s32(vacc2x0123, vright_post_shift);
240 vacc2x4567 = vrshlq_s32(vacc2x4567, vright_post_shift);
241
242 const int16x8_t voutput_zero_point = vld1q_dup_s16(¶ms->rndnu_neon.output_zero_point);
243 #if XNN_ARCH_ARM64
244 int16x8_t vacc0x01234567 = vqmovn_high_s32(vqmovn_s32(vacc0x0123), vacc0x4567);
245 int16x8_t vacc1x01234567 = vqmovn_high_s32(vqmovn_s32(vacc1x0123), vacc1x4567);
246 int16x8_t vacc2x01234567 = vqmovn_high_s32(vqmovn_s32(vacc2x0123), vacc2x4567);
247
248 vacc0x01234567 = vqaddq_s16(vacc0x01234567, voutput_zero_point);
249 vacc1x01234567 = vqaddq_s16(vacc1x01234567, voutput_zero_point);
250 vacc2x01234567 = vqaddq_s16(vacc2x01234567, voutput_zero_point);
251
252 int8x16_t vout0x01234567_1x01234567 = vqmovn_high_s16(vqmovn_s16(vacc0x01234567), vacc1x01234567);
253 int8x8_t vout2x01234567 = vqmovn_s16(vacc2x01234567);
254 #else
255 int16x8_t vacc0x01234567 = vcombine_s16(vqmovn_s32(vacc0x0123), vqmovn_s32(vacc0x4567));
256 int16x8_t vacc1x01234567 = vcombine_s16(vqmovn_s32(vacc1x0123), vqmovn_s32(vacc1x4567));
257 int16x8_t vacc2x01234567 = vcombine_s16(vqmovn_s32(vacc2x0123), vqmovn_s32(vacc2x4567));
258
259 vacc0x01234567 = vqaddq_s16(vacc0x01234567, voutput_zero_point);
260 vacc1x01234567 = vqaddq_s16(vacc1x01234567, voutput_zero_point);
261 vacc2x01234567 = vqaddq_s16(vacc2x01234567, voutput_zero_point);
262
263 int8x16_t vout0x01234567_1x01234567 = vcombine_s8(vqmovn_s16(vacc0x01234567), vqmovn_s16(vacc1x01234567));
264 int8x8_t vout2x01234567 = vqmovn_s16(vacc2x01234567);
265 #endif
266
267 const int8x16_t voutput_min = vld1q_dup_s8(¶ms->rndnu_neon.output_min);
268 vout0x01234567_1x01234567 = vmaxq_s8(vout0x01234567_1x01234567, voutput_min);
269 vout2x01234567 = vmax_s8(vout2x01234567, vget_low_s8(voutput_min));
270
271 const int8x16_t voutput_max = vld1q_dup_s8(¶ms->rndnu_neon.output_max);
272 vout0x01234567_1x01234567 = vminq_s8(vout0x01234567_1x01234567, voutput_max);
273 vout2x01234567 = vmin_s8(vout2x01234567, vget_low_s8(voutput_max));
274
275 if (nc >= 8) {
276 vst1_s8(c0 + 0, vget_low_s8(vout0x01234567_1x01234567));
277 vst1_s8(c1 + 0, vget_high_s8(vout0x01234567_1x01234567));
278 vst1_s8(c2 + 0, vout2x01234567);
279
280 c0 = (int8_t*) ((uintptr_t) c0 + cn_stride);
281 c1 = (int8_t*) ((uintptr_t) c1 + cn_stride);
282 c2 = (int8_t*) ((uintptr_t) c2 + cn_stride);
283
284 a0 = (const int8_t*) ((uintptr_t) a0 - kc);
285 a1 = (const int8_t*) ((uintptr_t) a1 - kc);
286 a2 = (const int8_t*) ((uintptr_t) a2 - kc);
287
288 nc -= 8;
289 } else {
290 // Final case where not all of the 8 columns fit in the destination.
291 if (nc & 4) {
292 vst1q_lane_u32((void*) c0, vreinterpretq_u32_s8(vout0x01234567_1x01234567), 0); c0 += 4;
293 vst1q_lane_u32((void*) c1, vreinterpretq_u32_s8(vout0x01234567_1x01234567), 2); c1 += 4;
294 vst1_lane_u32((void*) c2, vreinterpret_u32_s8(vout2x01234567), 0); c2 += 4;
295 vout0x01234567_1x01234567 = vextq_s8(vout0x01234567_1x01234567, vout0x01234567_1x01234567, 4);
296 vout2x01234567 = vext_s8(vout2x01234567, vout2x01234567, 4);
297 }
298 if (nc & 2) {
299 vst1q_lane_u16((void*) c0, vreinterpretq_u16_s8(vout0x01234567_1x01234567), 0); c0 += 2;
300 vst1q_lane_u16((void*) c1, vreinterpretq_u16_s8(vout0x01234567_1x01234567), 4); c1 += 2;
301 vst1_lane_u16((void*) c2, vreinterpret_u16_s8(vout2x01234567), 0); c2 += 2;
302 vout0x01234567_1x01234567 = vextq_s8(vout0x01234567_1x01234567, vout0x01234567_1x01234567, 2);
303 vout2x01234567 = vext_s8(vout2x01234567, vout2x01234567, 2);
304 }
305 if (nc & 1) {
306 vst1q_lane_s8(c0, vout0x01234567_1x01234567, 0);
307 vst1q_lane_s8(c1, vout0x01234567_1x01234567, 8);
308 vst1_lane_s8(c2, vout2x01234567, 0);
309 }
310
311 nc = 0;
312 }
313 } while (nc != 0);
314 }
315