1 // Auto-generated file. Do not edit!
2 // Template: src/qs8-gemm/c4-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
17
xnn_qs8_gemm_minmax_rndnu_ukernel_3x8c4__neon_mull_dup(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)])18 void xnn_qs8_gemm_minmax_rndnu_ukernel_3x8c4__neon_mull_dup(
19 size_t mr,
20 size_t nc,
21 size_t kc,
22 const int8_t* restrict a,
23 size_t a_stride,
24 const void* restrict w,
25 int8_t* restrict c,
26 size_t cm_stride,
27 size_t cn_stride,
28 const union xnn_qs8_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(int8_t) == 0);
35 assert(a != NULL);
36 assert(w != NULL);
37 assert(c != NULL);
38
39 kc = round_up_po2(kc, 4 * sizeof(int8_t));
40 const int8_t* a0 = a;
41 int8_t* c0 = c;
42 const int8_t* a1 = (const int8_t*) ((uintptr_t) a0 + a_stride);
43 int8_t* c1 = (int8_t*) ((uintptr_t) c0 + cm_stride);
44 if XNN_UNPREDICTABLE(mr < 2) {
45 a1 = a0;
46 c1 = c0;
47 }
48 const int8_t* a2 = (const int8_t*) ((uintptr_t) a1 + a_stride);
49 int8_t* c2 = (int8_t*) ((uintptr_t) c1 + cm_stride);
50 if XNN_UNPREDICTABLE(mr <= 2) {
51 a2 = a1;
52 c2 = c1;
53 }
54
55 do {
56 int32x4_t vacc0x01 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
57 int32x4_t vacc0x23 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
58 int32x4_t vacc0x45 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
59 int32x4_t vacc0x67 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
60 int32x4_t vacc1x01 = vacc0x01;
61 int32x4_t vacc1x23 = vacc0x23;
62 int32x4_t vacc1x45 = vacc0x45;
63 int32x4_t vacc1x67 = vacc0x67;
64 int32x4_t vacc2x01 = vacc0x01;
65 int32x4_t vacc2x23 = vacc0x23;
66 int32x4_t vacc2x45 = vacc0x45;
67 int32x4_t vacc2x67 = vacc0x67;
68
69 size_t k = kc;
70
71
72 while (k >= 8 * sizeof(int8_t)) {
73 const int8x8_t va0 = vld1_s8(a0); a0 += 8;
74 const int8x8_t va1 = vld1_s8(a1); a1 += 8;
75 const int8x8_t va2 = vld1_s8(a2); a2 += 8;
76
77 const int8x8_t vb01c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
78 const int8x8_t vb23c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
79 const int8x8_t vb45c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
80 const int8x8_t vb67c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
81 const int8x8_t vb01c1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
82 const int8x8_t vb23c1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
83 const int8x8_t vb45c1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
84 const int8x8_t vb67c1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
85
86 const int8x8_t va0c0 = vreinterpret_s8_s32(vdup_lane_s32(vreinterpret_s32_s8(va0), 0));
87 const int8x8_t va1c0 = vreinterpret_s8_s32(vdup_lane_s32(vreinterpret_s32_s8(va1), 0));
88 const int8x8_t va2c0 = vreinterpret_s8_s32(vdup_lane_s32(vreinterpret_s32_s8(va2), 0));
89
90 const int16x8_t vprod0x01c0 = vmull_s8(vb01c0, va0c0);
91 const int16x8_t vprod1x01c0 = vmull_s8(vb01c0, va1c0);
92 const int16x8_t vprod2x01c0 = vmull_s8(vb01c0, va2c0);
93 vacc0x01 = vpadalq_s16(vacc0x01, vprod0x01c0);
94 vacc1x01 = vpadalq_s16(vacc1x01, vprod1x01c0);
95 vacc2x01 = vpadalq_s16(vacc2x01, vprod2x01c0);
96 const int16x8_t vprod0x23c0 = vmull_s8(vb23c0, va0c0);
97 const int16x8_t vprod1x23c0 = vmull_s8(vb23c0, va1c0);
98 const int16x8_t vprod2x23c0 = vmull_s8(vb23c0, va2c0);
99 vacc0x23 = vpadalq_s16(vacc0x23, vprod0x23c0);
100 vacc1x23 = vpadalq_s16(vacc1x23, vprod1x23c0);
101 vacc2x23 = vpadalq_s16(vacc2x23, vprod2x23c0);
102 const int16x8_t vprod0x45c0 = vmull_s8(vb45c0, va0c0);
103 const int16x8_t vprod1x45c0 = vmull_s8(vb45c0, va1c0);
104 const int16x8_t vprod2x45c0 = vmull_s8(vb45c0, va2c0);
105 vacc0x45 = vpadalq_s16(vacc0x45, vprod0x45c0);
106 vacc1x45 = vpadalq_s16(vacc1x45, vprod1x45c0);
107 vacc2x45 = vpadalq_s16(vacc2x45, vprod2x45c0);
108 const int16x8_t vprod0x67c0 = vmull_s8(vb67c0, va0c0);
109 const int16x8_t vprod1x67c0 = vmull_s8(vb67c0, va1c0);
110 const int16x8_t vprod2x67c0 = vmull_s8(vb67c0, va2c0);
111 vacc0x67 = vpadalq_s16(vacc0x67, vprod0x67c0);
112 vacc1x67 = vpadalq_s16(vacc1x67, vprod1x67c0);
113 vacc2x67 = vpadalq_s16(vacc2x67, vprod2x67c0);
114 const int8x8_t va0c1 = vreinterpret_s8_s32(vdup_lane_s32(vreinterpret_s32_s8(va0), 1));
115 const int8x8_t va1c1 = vreinterpret_s8_s32(vdup_lane_s32(vreinterpret_s32_s8(va1), 1));
116 const int8x8_t va2c1 = vreinterpret_s8_s32(vdup_lane_s32(vreinterpret_s32_s8(va2), 1));
117
118 const int16x8_t vprod0x01c1 = vmull_s8(vb01c1, va0c1);
119 const int16x8_t vprod1x01c1 = vmull_s8(vb01c1, va1c1);
120 const int16x8_t vprod2x01c1 = vmull_s8(vb01c1, va2c1);
121 vacc0x01 = vpadalq_s16(vacc0x01, vprod0x01c1);
122 vacc1x01 = vpadalq_s16(vacc1x01, vprod1x01c1);
123 vacc2x01 = vpadalq_s16(vacc2x01, vprod2x01c1);
124 const int16x8_t vprod0x23c1 = vmull_s8(vb23c1, va0c1);
125 const int16x8_t vprod1x23c1 = vmull_s8(vb23c1, va1c1);
126 const int16x8_t vprod2x23c1 = vmull_s8(vb23c1, va2c1);
127 vacc0x23 = vpadalq_s16(vacc0x23, vprod0x23c1);
128 vacc1x23 = vpadalq_s16(vacc1x23, vprod1x23c1);
129 vacc2x23 = vpadalq_s16(vacc2x23, vprod2x23c1);
130 const int16x8_t vprod0x45c1 = vmull_s8(vb45c1, va0c1);
131 const int16x8_t vprod1x45c1 = vmull_s8(vb45c1, va1c1);
132 const int16x8_t vprod2x45c1 = vmull_s8(vb45c1, va2c1);
133 vacc0x45 = vpadalq_s16(vacc0x45, vprod0x45c1);
134 vacc1x45 = vpadalq_s16(vacc1x45, vprod1x45c1);
135 vacc2x45 = vpadalq_s16(vacc2x45, vprod2x45c1);
136 const int16x8_t vprod0x67c1 = vmull_s8(vb67c1, va0c1);
137 const int16x8_t vprod1x67c1 = vmull_s8(vb67c1, va1c1);
138 const int16x8_t vprod2x67c1 = vmull_s8(vb67c1, va2c1);
139 vacc0x67 = vpadalq_s16(vacc0x67, vprod0x67c1);
140 vacc1x67 = vpadalq_s16(vacc1x67, vprod1x67c1);
141 vacc2x67 = vpadalq_s16(vacc2x67, vprod2x67c1);
142
143 k -= 8 * sizeof(int8_t);
144 }
145
146 if XNN_UNLIKELY(k != 0) {
147 const int8x8_t va0 = vld1_s8(a0); a0 = (const int8_t*) ((uintptr_t) a0 + k);
148 const int8x8_t va1 = vld1_s8(a1); a1 = (const int8_t*) ((uintptr_t) a1 + k);
149 const int8x8_t va2 = vld1_s8(a2); a2 = (const int8_t*) ((uintptr_t) a2 + k);
150
151 const int8x8_t vb01c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
152 const int8x8_t vb23c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
153 const int8x8_t vb45c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
154 const int8x8_t vb67c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
155
156 const int8x8_t va0c0 = vreinterpret_s8_s32(vdup_lane_s32(vreinterpret_s32_s8(va0), 0));
157 const int16x8_t vprod0x01c0 = vmull_s8(vb01c0, va0c0);
158 vacc0x01 = vpadalq_s16(vacc0x01, vprod0x01c0);
159 const int16x8_t vprod0x23c0 = vmull_s8(vb23c0, va0c0);
160 vacc0x23 = vpadalq_s16(vacc0x23, vprod0x23c0);
161 const int16x8_t vprod0x45c0 = vmull_s8(vb45c0, va0c0);
162 vacc0x45 = vpadalq_s16(vacc0x45, vprod0x45c0);
163 const int16x8_t vprod0x67c0 = vmull_s8(vb67c0, va0c0);
164 vacc0x67 = vpadalq_s16(vacc0x67, vprod0x67c0);
165 const int8x8_t va1c0 = vreinterpret_s8_s32(vdup_lane_s32(vreinterpret_s32_s8(va1), 0));
166 const int16x8_t vprod1x01c0 = vmull_s8(vb01c0, va1c0);
167 vacc1x01 = vpadalq_s16(vacc1x01, vprod1x01c0);
168 const int16x8_t vprod1x23c0 = vmull_s8(vb23c0, va1c0);
169 vacc1x23 = vpadalq_s16(vacc1x23, vprod1x23c0);
170 const int16x8_t vprod1x45c0 = vmull_s8(vb45c0, va1c0);
171 vacc1x45 = vpadalq_s16(vacc1x45, vprod1x45c0);
172 const int16x8_t vprod1x67c0 = vmull_s8(vb67c0, va1c0);
173 vacc1x67 = vpadalq_s16(vacc1x67, vprod1x67c0);
174 const int8x8_t va2c0 = vreinterpret_s8_s32(vdup_lane_s32(vreinterpret_s32_s8(va2), 0));
175 const int16x8_t vprod2x01c0 = vmull_s8(vb01c0, va2c0);
176 vacc2x01 = vpadalq_s16(vacc2x01, vprod2x01c0);
177 const int16x8_t vprod2x23c0 = vmull_s8(vb23c0, va2c0);
178 vacc2x23 = vpadalq_s16(vacc2x23, vprod2x23c0);
179 const int16x8_t vprod2x45c0 = vmull_s8(vb45c0, va2c0);
180 vacc2x45 = vpadalq_s16(vacc2x45, vprod2x45c0);
181 const int16x8_t vprod2x67c0 = vmull_s8(vb67c0, va2c0);
182 vacc2x67 = vpadalq_s16(vacc2x67, vprod2x67c0);
183 }
184
185 #if XNN_ARCH_ARM64
186 int32x4_t vacc0x0123 = vpaddq_s32(vacc0x01, vacc0x23);
187 int32x4_t vacc0x4567 = vpaddq_s32(vacc0x45, vacc0x67);
188 int32x4_t vacc1x0123 = vpaddq_s32(vacc1x01, vacc1x23);
189 int32x4_t vacc1x4567 = vpaddq_s32(vacc1x45, vacc1x67);
190 int32x4_t vacc2x0123 = vpaddq_s32(vacc2x01, vacc2x23);
191 int32x4_t vacc2x4567 = vpaddq_s32(vacc2x45, vacc2x67);
192 #else
193 const int32x2_t vsum0x01 = vpadd_s32(vget_low_s32(vacc0x01), vget_high_s32(vacc0x01));
194 const int32x2_t vsum0x23 = vpadd_s32(vget_low_s32(vacc0x23), vget_high_s32(vacc0x23));
195 int32x4_t vacc0x0123 = vcombine_s32(vsum0x01, vsum0x23);
196 const int32x2_t vsum0x45 = vpadd_s32(vget_low_s32(vacc0x45), vget_high_s32(vacc0x45));
197 const int32x2_t vsum0x67 = vpadd_s32(vget_low_s32(vacc0x67), vget_high_s32(vacc0x67));
198 int32x4_t vacc0x4567 = vcombine_s32(vsum0x45, vsum0x67);
199 const int32x2_t vsum1x01 = vpadd_s32(vget_low_s32(vacc1x01), vget_high_s32(vacc1x01));
200 const int32x2_t vsum1x23 = vpadd_s32(vget_low_s32(vacc1x23), vget_high_s32(vacc1x23));
201 int32x4_t vacc1x0123 = vcombine_s32(vsum1x01, vsum1x23);
202 const int32x2_t vsum1x45 = vpadd_s32(vget_low_s32(vacc1x45), vget_high_s32(vacc1x45));
203 const int32x2_t vsum1x67 = vpadd_s32(vget_low_s32(vacc1x67), vget_high_s32(vacc1x67));
204 int32x4_t vacc1x4567 = vcombine_s32(vsum1x45, vsum1x67);
205 const int32x2_t vsum2x01 = vpadd_s32(vget_low_s32(vacc2x01), vget_high_s32(vacc2x01));
206 const int32x2_t vsum2x23 = vpadd_s32(vget_low_s32(vacc2x23), vget_high_s32(vacc2x23));
207 int32x4_t vacc2x0123 = vcombine_s32(vsum2x01, vsum2x23);
208 const int32x2_t vsum2x45 = vpadd_s32(vget_low_s32(vacc2x45), vget_high_s32(vacc2x45));
209 const int32x2_t vsum2x67 = vpadd_s32(vget_low_s32(vacc2x67), vget_high_s32(vacc2x67));
210 int32x4_t vacc2x4567 = vcombine_s32(vsum2x45, vsum2x67);
211 #endif
212
213 const int32x4_t vright_pre_shift = vld1q_dup_s32(¶ms->rndnu_neon.right_pre_shift);
214 const int32x4_t vmultiplier = vld1q_dup_s32(¶ms->rndnu_neon.multiplier);
215 const int32x4_t vright_post_shift = vld1q_dup_s32(¶ms->rndnu_neon.right_post_shift);
216
217 vacc0x0123 = vqshlq_s32(vacc0x0123, vright_pre_shift);
218 vacc0x4567 = vqshlq_s32(vacc0x4567, vright_pre_shift);
219 vacc1x0123 = vqshlq_s32(vacc1x0123, vright_pre_shift);
220 vacc1x4567 = vqshlq_s32(vacc1x4567, vright_pre_shift);
221 vacc2x0123 = vqshlq_s32(vacc2x0123, vright_pre_shift);
222 vacc2x4567 = vqshlq_s32(vacc2x4567, vright_pre_shift);
223
224 vacc0x0123 = vqdmulhq_s32(vacc0x0123, vmultiplier);
225 vacc0x4567 = vqdmulhq_s32(vacc0x4567, vmultiplier);
226 vacc1x0123 = vqdmulhq_s32(vacc1x0123, vmultiplier);
227 vacc1x4567 = vqdmulhq_s32(vacc1x4567, vmultiplier);
228 vacc2x0123 = vqdmulhq_s32(vacc2x0123, vmultiplier);
229 vacc2x4567 = vqdmulhq_s32(vacc2x4567, vmultiplier);
230
231 vacc0x0123 = vrshlq_s32(vacc0x0123, vright_post_shift);
232 vacc0x4567 = vrshlq_s32(vacc0x4567, vright_post_shift);
233 vacc1x0123 = vrshlq_s32(vacc1x0123, vright_post_shift);
234 vacc1x4567 = vrshlq_s32(vacc1x4567, vright_post_shift);
235 vacc2x0123 = vrshlq_s32(vacc2x0123, vright_post_shift);
236 vacc2x4567 = vrshlq_s32(vacc2x4567, vright_post_shift);
237
238 const int16x8_t voutput_zero_point = vld1q_dup_s16(¶ms->rndnu_neon.output_zero_point);
239 #if XNN_ARCH_ARM64
240 int16x8_t vacc0x01234567 = vqmovn_high_s32(vqmovn_s32(vacc0x0123), vacc0x4567);
241 int16x8_t vacc1x01234567 = vqmovn_high_s32(vqmovn_s32(vacc1x0123), vacc1x4567);
242 int16x8_t vacc2x01234567 = vqmovn_high_s32(vqmovn_s32(vacc2x0123), vacc2x4567);
243
244 vacc0x01234567 = vqaddq_s16(vacc0x01234567, voutput_zero_point);
245 vacc1x01234567 = vqaddq_s16(vacc1x01234567, voutput_zero_point);
246 vacc2x01234567 = vqaddq_s16(vacc2x01234567, voutput_zero_point);
247
248 int8x16_t vout0x01234567_1x01234567 = vqmovn_high_s16(vqmovn_s16(vacc0x01234567), vacc1x01234567);
249 int8x8_t vout2x01234567 = vqmovn_s16(vacc2x01234567);
250 #else
251 int16x8_t vacc0x01234567 = vcombine_s16(vqmovn_s32(vacc0x0123), vqmovn_s32(vacc0x4567));
252 int16x8_t vacc1x01234567 = vcombine_s16(vqmovn_s32(vacc1x0123), vqmovn_s32(vacc1x4567));
253 int16x8_t vacc2x01234567 = vcombine_s16(vqmovn_s32(vacc2x0123), vqmovn_s32(vacc2x4567));
254
255 vacc0x01234567 = vqaddq_s16(vacc0x01234567, voutput_zero_point);
256 vacc1x01234567 = vqaddq_s16(vacc1x01234567, voutput_zero_point);
257 vacc2x01234567 = vqaddq_s16(vacc2x01234567, voutput_zero_point);
258
259 int8x16_t vout0x01234567_1x01234567 = vcombine_s8(vqmovn_s16(vacc0x01234567), vqmovn_s16(vacc1x01234567));
260 int8x8_t vout2x01234567 = vqmovn_s16(vacc2x01234567);
261 #endif
262
263 const int8x16_t voutput_min = vld1q_dup_s8(¶ms->rndnu_neon.output_min);
264 vout0x01234567_1x01234567 = vmaxq_s8(vout0x01234567_1x01234567, voutput_min);
265 vout2x01234567 = vmax_s8(vout2x01234567, vget_low_s8(voutput_min));
266
267 const int8x16_t voutput_max = vld1q_dup_s8(¶ms->rndnu_neon.output_max);
268 vout0x01234567_1x01234567 = vminq_s8(vout0x01234567_1x01234567, voutput_max);
269 vout2x01234567 = vmin_s8(vout2x01234567, vget_low_s8(voutput_max));
270
271 if (nc >= 8) {
272 vst1_s8(c0 + 0, vget_low_s8(vout0x01234567_1x01234567));
273 vst1_s8(c1 + 0, vget_high_s8(vout0x01234567_1x01234567));
274 vst1_s8(c2 + 0, vout2x01234567);
275
276 c0 = (int8_t*) ((uintptr_t) c0 + cn_stride);
277 c1 = (int8_t*) ((uintptr_t) c1 + cn_stride);
278 c2 = (int8_t*) ((uintptr_t) c2 + cn_stride);
279
280 a0 = (const int8_t*) ((uintptr_t) a0 - kc);
281 a1 = (const int8_t*) ((uintptr_t) a1 - kc);
282 a2 = (const int8_t*) ((uintptr_t) a2 - kc);
283
284 nc -= 8;
285 } else {
286 // Final case where not all of the 8 columns fit in the destination.
287 if (nc & 4) {
288 vst1q_lane_u32((void*) c0, vreinterpretq_u32_s8(vout0x01234567_1x01234567), 0); c0 += 4;
289 vst1q_lane_u32((void*) c1, vreinterpretq_u32_s8(vout0x01234567_1x01234567), 2); c1 += 4;
290 vst1_lane_u32((void*) c2, vreinterpret_u32_s8(vout2x01234567), 0); c2 += 4;
291 vout0x01234567_1x01234567 = vextq_s8(vout0x01234567_1x01234567, vout0x01234567_1x01234567, 4);
292 vout2x01234567 = vext_s8(vout2x01234567, vout2x01234567, 4);
293 }
294 if (nc & 2) {
295 vst1q_lane_u16((void*) c0, vreinterpretq_u16_s8(vout0x01234567_1x01234567), 0); c0 += 2;
296 vst1q_lane_u16((void*) c1, vreinterpretq_u16_s8(vout0x01234567_1x01234567), 4); c1 += 2;
297 vst1_lane_u16((void*) c2, vreinterpret_u16_s8(vout2x01234567), 0); c2 += 2;
298 vout0x01234567_1x01234567 = vextq_s8(vout0x01234567_1x01234567, vout0x01234567_1x01234567, 2);
299 vout2x01234567 = vext_s8(vout2x01234567, vout2x01234567, 2);
300 }
301 if (nc & 1) {
302 vst1q_lane_s8(c0, vout0x01234567_1x01234567, 0);
303 vst1q_lane_s8(c1, vout0x01234567_1x01234567, 8);
304 vst1_lane_s8(c2, vout2x01234567, 0);
305 }
306
307 nc = 0;
308 }
309 } while (nc != 0);
310 }
311