1 // Auto-generated file. Do not edit!
2 // Template: src/qs8-gemm/c4-neon-mull-shuffle.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_4x8c4s2__neon_mull(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_4x8c4s2__neon_mull(
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 <= 4);
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 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 const int8_t* a3 = (const int8_t*) ((uintptr_t) a2 + a_stride);
54 int8_t* c3 = (int8_t*) ((uintptr_t) c2 + cm_stride);
55 if XNN_UNPREDICTABLE(mr != 4) {
56 a3 = a2;
57 c3 = c2;
58 }
59
60 kc = round_up_po2(kc, 8 * sizeof(int8_t));
61 do {
62 int32x4_t vacc0x01 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const int32_t*) w + 2;
63 int32x4_t vacc0x23 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const int32_t*) w + 2;
64 int32x4_t vacc0x45 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const int32_t*) w + 2;
65 int32x4_t vacc0x67 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const int32_t*) w + 2;
66 int32x4_t vacc1x01 = vacc0x01;
67 int32x4_t vacc1x23 = vacc0x23;
68 int32x4_t vacc1x45 = vacc0x45;
69 int32x4_t vacc1x67 = vacc0x67;
70 int32x4_t vacc2x01 = vacc0x01;
71 int32x4_t vacc2x23 = vacc0x23;
72 int32x4_t vacc2x45 = vacc0x45;
73 int32x4_t vacc2x67 = vacc0x67;
74 int32x4_t vacc3x01 = vacc0x01;
75 int32x4_t vacc3x23 = vacc0x23;
76 int32x4_t vacc3x45 = vacc0x45;
77 int32x4_t vacc3x67 = vacc0x67;
78
79 size_t k = kc;
80 do {
81 int8x8_t va0x0 = vld1_s8(a0); a0 += 8;
82 int8x8_t va1x0 = vld1_s8(a1); a1 += 8;
83 int8x8_t va2x0 = vld1_s8(a2); a2 += 8;
84 int8x8_t va3x0 = vld1_s8(a3); a3 += 8;
85
86 const int8x8_t vb01c0x0 = vld1_s8(w); w = (const int8_t*) w + 8;
87 const int8x8_t vb23c0x0 = vld1_s8(w); w = (const int8_t*) w + 8;
88 const int8x8_t vb45c0x0 = vld1_s8(w); w = (const int8_t*) w + 8;
89 const int8x8_t vb67c0x0 = vld1_s8(w); w = (const int8_t*) w + 8;
90 const int8x8_t vb01c1x0 = vld1_s8(w); w = (const int8_t*) w + 8;
91 const int8x8_t vb23c1x0 = vld1_s8(w); w = (const int8_t*) w + 8;
92 const int8x8_t vb45c1x0 = vld1_s8(w); w = (const int8_t*) w + 8;
93 const int8x8_t vb67c1x0 = vld1_s8(w); w = (const int8_t*) w + 8;
94
95 int16x8_t vprod0x01c0 = vmull_s8(vb01c0x0, va0x0);
96 int16x8_t vprod1x01c0 = vmull_s8(vb01c0x0, va1x0);
97 int16x8_t vprod2x01c0 = vmull_s8(vb01c0x0, va2x0);
98 int16x8_t vprod3x01c0 = vmull_s8(vb01c0x0, va3x0);
99 vacc0x01 = vpadalq_s16(vacc0x01, vprod0x01c0);
100 vacc1x01 = vpadalq_s16(vacc1x01, vprod1x01c0);
101 vacc2x01 = vpadalq_s16(vacc2x01, vprod2x01c0);
102 vacc3x01 = vpadalq_s16(vacc3x01, vprod3x01c0);
103 int16x8_t vprod0x23c0 = vmull_s8(vb23c0x0, va0x0);
104 int16x8_t vprod1x23c0 = vmull_s8(vb23c0x0, va1x0);
105 int16x8_t vprod2x23c0 = vmull_s8(vb23c0x0, va2x0);
106 int16x8_t vprod3x23c0 = vmull_s8(vb23c0x0, va3x0);
107 vacc0x23 = vpadalq_s16(vacc0x23, vprod0x23c0);
108 vacc1x23 = vpadalq_s16(vacc1x23, vprod1x23c0);
109 vacc2x23 = vpadalq_s16(vacc2x23, vprod2x23c0);
110 vacc3x23 = vpadalq_s16(vacc3x23, vprod3x23c0);
111 int16x8_t vprod0x45c0 = vmull_s8(vb45c0x0, va0x0);
112 int16x8_t vprod1x45c0 = vmull_s8(vb45c0x0, va1x0);
113 int16x8_t vprod2x45c0 = vmull_s8(vb45c0x0, va2x0);
114 int16x8_t vprod3x45c0 = vmull_s8(vb45c0x0, va3x0);
115 vacc0x45 = vpadalq_s16(vacc0x45, vprod0x45c0);
116 vacc1x45 = vpadalq_s16(vacc1x45, vprod1x45c0);
117 vacc2x45 = vpadalq_s16(vacc2x45, vprod2x45c0);
118 vacc3x45 = vpadalq_s16(vacc3x45, vprod3x45c0);
119 int16x8_t vprod0x67c0 = vmull_s8(vb67c0x0, va0x0);
120 int16x8_t vprod1x67c0 = vmull_s8(vb67c0x0, va1x0);
121 int16x8_t vprod2x67c0 = vmull_s8(vb67c0x0, va2x0);
122 int16x8_t vprod3x67c0 = vmull_s8(vb67c0x0, va3x0);
123 vacc0x67 = vpadalq_s16(vacc0x67, vprod0x67c0);
124 vacc1x67 = vpadalq_s16(vacc1x67, vprod1x67c0);
125 vacc2x67 = vpadalq_s16(vacc2x67, vprod2x67c0);
126 vacc3x67 = vpadalq_s16(vacc3x67, vprod3x67c0);
127 va0x0 = vext_s8(va0x0, va0x0, 4);
128 va1x0 = vext_s8(va1x0, va1x0, 4);
129 va2x0 = vext_s8(va2x0, va2x0, 4);
130 va3x0 = vext_s8(va3x0, va3x0, 4);
131 int16x8_t vprod0x01c1 = vmull_s8(vb01c1x0, va0x0);
132 int16x8_t vprod1x01c1 = vmull_s8(vb01c1x0, va1x0);
133 int16x8_t vprod2x01c1 = vmull_s8(vb01c1x0, va2x0);
134 int16x8_t vprod3x01c1 = vmull_s8(vb01c1x0, va3x0);
135 vacc0x01 = vpadalq_s16(vacc0x01, vprod0x01c1);
136 vacc1x01 = vpadalq_s16(vacc1x01, vprod1x01c1);
137 vacc2x01 = vpadalq_s16(vacc2x01, vprod2x01c1);
138 vacc3x01 = vpadalq_s16(vacc3x01, vprod3x01c1);
139 int16x8_t vprod0x23c1 = vmull_s8(vb23c1x0, va0x0);
140 int16x8_t vprod1x23c1 = vmull_s8(vb23c1x0, va1x0);
141 int16x8_t vprod2x23c1 = vmull_s8(vb23c1x0, va2x0);
142 int16x8_t vprod3x23c1 = vmull_s8(vb23c1x0, va3x0);
143 vacc0x23 = vpadalq_s16(vacc0x23, vprod0x23c1);
144 vacc1x23 = vpadalq_s16(vacc1x23, vprod1x23c1);
145 vacc2x23 = vpadalq_s16(vacc2x23, vprod2x23c1);
146 vacc3x23 = vpadalq_s16(vacc3x23, vprod3x23c1);
147 int16x8_t vprod0x45c1 = vmull_s8(vb45c1x0, va0x0);
148 int16x8_t vprod1x45c1 = vmull_s8(vb45c1x0, va1x0);
149 int16x8_t vprod2x45c1 = vmull_s8(vb45c1x0, va2x0);
150 int16x8_t vprod3x45c1 = vmull_s8(vb45c1x0, va3x0);
151 vacc0x45 = vpadalq_s16(vacc0x45, vprod0x45c1);
152 vacc1x45 = vpadalq_s16(vacc1x45, vprod1x45c1);
153 vacc2x45 = vpadalq_s16(vacc2x45, vprod2x45c1);
154 vacc3x45 = vpadalq_s16(vacc3x45, vprod3x45c1);
155 int16x8_t vprod0x67c1 = vmull_s8(vb67c1x0, va0x0);
156 int16x8_t vprod1x67c1 = vmull_s8(vb67c1x0, va1x0);
157 int16x8_t vprod2x67c1 = vmull_s8(vb67c1x0, va2x0);
158 int16x8_t vprod3x67c1 = vmull_s8(vb67c1x0, va3x0);
159 vacc0x67 = vpadalq_s16(vacc0x67, vprod0x67c1);
160 vacc1x67 = vpadalq_s16(vacc1x67, vprod1x67c1);
161 vacc2x67 = vpadalq_s16(vacc2x67, vprod2x67c1);
162 vacc3x67 = vpadalq_s16(vacc3x67, vprod3x67c1);
163
164 k -= 8 * sizeof(int8_t);
165 } while (k != 0);
166
167 #if XNN_ARCH_ARM64
168 int32x4_t vacc0x0123 = vpaddq_s32(vacc0x01, vacc0x23);
169 int32x4_t vacc0x4567 = vpaddq_s32(vacc0x45, vacc0x67);
170 int32x4_t vacc1x0123 = vpaddq_s32(vacc1x01, vacc1x23);
171 int32x4_t vacc1x4567 = vpaddq_s32(vacc1x45, vacc1x67);
172 int32x4_t vacc2x0123 = vpaddq_s32(vacc2x01, vacc2x23);
173 int32x4_t vacc2x4567 = vpaddq_s32(vacc2x45, vacc2x67);
174 int32x4_t vacc3x0123 = vpaddq_s32(vacc3x01, vacc3x23);
175 int32x4_t vacc3x4567 = vpaddq_s32(vacc3x45, vacc3x67);
176 #else
177 const int32x2_t vsum0x01 = vpadd_s32(vget_low_s32(vacc0x01), vget_high_s32(vacc0x01));
178 const int32x2_t vsum0x23 = vpadd_s32(vget_low_s32(vacc0x23), vget_high_s32(vacc0x23));
179 int32x4_t vacc0x0123 = vcombine_s32(vsum0x01, vsum0x23);
180 const int32x2_t vsum0x45 = vpadd_s32(vget_low_s32(vacc0x45), vget_high_s32(vacc0x45));
181 const int32x2_t vsum0x67 = vpadd_s32(vget_low_s32(vacc0x67), vget_high_s32(vacc0x67));
182 int32x4_t vacc0x4567 = vcombine_s32(vsum0x45, vsum0x67);
183 const int32x2_t vsum1x01 = vpadd_s32(vget_low_s32(vacc1x01), vget_high_s32(vacc1x01));
184 const int32x2_t vsum1x23 = vpadd_s32(vget_low_s32(vacc1x23), vget_high_s32(vacc1x23));
185 int32x4_t vacc1x0123 = vcombine_s32(vsum1x01, vsum1x23);
186 const int32x2_t vsum1x45 = vpadd_s32(vget_low_s32(vacc1x45), vget_high_s32(vacc1x45));
187 const int32x2_t vsum1x67 = vpadd_s32(vget_low_s32(vacc1x67), vget_high_s32(vacc1x67));
188 int32x4_t vacc1x4567 = vcombine_s32(vsum1x45, vsum1x67);
189 const int32x2_t vsum2x01 = vpadd_s32(vget_low_s32(vacc2x01), vget_high_s32(vacc2x01));
190 const int32x2_t vsum2x23 = vpadd_s32(vget_low_s32(vacc2x23), vget_high_s32(vacc2x23));
191 int32x4_t vacc2x0123 = vcombine_s32(vsum2x01, vsum2x23);
192 const int32x2_t vsum2x45 = vpadd_s32(vget_low_s32(vacc2x45), vget_high_s32(vacc2x45));
193 const int32x2_t vsum2x67 = vpadd_s32(vget_low_s32(vacc2x67), vget_high_s32(vacc2x67));
194 int32x4_t vacc2x4567 = vcombine_s32(vsum2x45, vsum2x67);
195 const int32x2_t vsum3x01 = vpadd_s32(vget_low_s32(vacc3x01), vget_high_s32(vacc3x01));
196 const int32x2_t vsum3x23 = vpadd_s32(vget_low_s32(vacc3x23), vget_high_s32(vacc3x23));
197 int32x4_t vacc3x0123 = vcombine_s32(vsum3x01, vsum3x23);
198 const int32x2_t vsum3x45 = vpadd_s32(vget_low_s32(vacc3x45), vget_high_s32(vacc3x45));
199 const int32x2_t vsum3x67 = vpadd_s32(vget_low_s32(vacc3x67), vget_high_s32(vacc3x67));
200 int32x4_t vacc3x4567 = vcombine_s32(vsum3x45, vsum3x67);
201 #endif
202
203 const int32x4_t vright_pre_shift = vld1q_dup_s32(¶ms->rndnu_neon.right_pre_shift);
204 const int32x4_t vmultiplier = vld1q_dup_s32(¶ms->rndnu_neon.multiplier);
205 const int32x4_t vright_post_shift = vld1q_dup_s32(¶ms->rndnu_neon.right_post_shift);
206
207 vacc0x0123 = vqshlq_s32(vacc0x0123, vright_pre_shift);
208 vacc0x4567 = vqshlq_s32(vacc0x4567, vright_pre_shift);
209 vacc1x0123 = vqshlq_s32(vacc1x0123, vright_pre_shift);
210 vacc1x4567 = vqshlq_s32(vacc1x4567, vright_pre_shift);
211 vacc2x0123 = vqshlq_s32(vacc2x0123, vright_pre_shift);
212 vacc2x4567 = vqshlq_s32(vacc2x4567, vright_pre_shift);
213 vacc3x0123 = vqshlq_s32(vacc3x0123, vright_pre_shift);
214 vacc3x4567 = vqshlq_s32(vacc3x4567, vright_pre_shift);
215
216 vacc0x0123 = vqdmulhq_s32(vacc0x0123, vmultiplier);
217 vacc0x4567 = vqdmulhq_s32(vacc0x4567, vmultiplier);
218 vacc1x0123 = vqdmulhq_s32(vacc1x0123, vmultiplier);
219 vacc1x4567 = vqdmulhq_s32(vacc1x4567, vmultiplier);
220 vacc2x0123 = vqdmulhq_s32(vacc2x0123, vmultiplier);
221 vacc2x4567 = vqdmulhq_s32(vacc2x4567, vmultiplier);
222 vacc3x0123 = vqdmulhq_s32(vacc3x0123, vmultiplier);
223 vacc3x4567 = vqdmulhq_s32(vacc3x4567, vmultiplier);
224
225 vacc0x0123 = vrshlq_s32(vacc0x0123, vright_post_shift);
226 vacc0x4567 = vrshlq_s32(vacc0x4567, vright_post_shift);
227 vacc1x0123 = vrshlq_s32(vacc1x0123, vright_post_shift);
228 vacc1x4567 = vrshlq_s32(vacc1x4567, vright_post_shift);
229 vacc2x0123 = vrshlq_s32(vacc2x0123, vright_post_shift);
230 vacc2x4567 = vrshlq_s32(vacc2x4567, vright_post_shift);
231 vacc3x0123 = vrshlq_s32(vacc3x0123, vright_post_shift);
232 vacc3x4567 = vrshlq_s32(vacc3x4567, vright_post_shift);
233
234 const int16x8_t voutput_zero_point = vld1q_dup_s16(¶ms->rndnu_neon.output_zero_point);
235 #if XNN_ARCH_ARM64
236 int16x8_t vacc0x01234567 = vqmovn_high_s32(vqmovn_s32(vacc0x0123), vacc0x4567);
237 int16x8_t vacc1x01234567 = vqmovn_high_s32(vqmovn_s32(vacc1x0123), vacc1x4567);
238 int16x8_t vacc2x01234567 = vqmovn_high_s32(vqmovn_s32(vacc2x0123), vacc2x4567);
239 int16x8_t vacc3x01234567 = vqmovn_high_s32(vqmovn_s32(vacc3x0123), vacc3x4567);
240
241 vacc0x01234567 = vqaddq_s16(vacc0x01234567, voutput_zero_point);
242 vacc1x01234567 = vqaddq_s16(vacc1x01234567, voutput_zero_point);
243 vacc2x01234567 = vqaddq_s16(vacc2x01234567, voutput_zero_point);
244 vacc3x01234567 = vqaddq_s16(vacc3x01234567, voutput_zero_point);
245
246 int8x16_t vout0x01234567_1x01234567 = vqmovn_high_s16(vqmovn_s16(vacc0x01234567), vacc1x01234567);
247 int8x16_t vout2x01234567_3x01234567 = vqmovn_high_s16(vqmovn_s16(vacc2x01234567), vacc3x01234567);
248 #else
249 int16x8_t vacc0x01234567 = vcombine_s16(vqmovn_s32(vacc0x0123), vqmovn_s32(vacc0x4567));
250 int16x8_t vacc1x01234567 = vcombine_s16(vqmovn_s32(vacc1x0123), vqmovn_s32(vacc1x4567));
251 int16x8_t vacc2x01234567 = vcombine_s16(vqmovn_s32(vacc2x0123), vqmovn_s32(vacc2x4567));
252 int16x8_t vacc3x01234567 = vcombine_s16(vqmovn_s32(vacc3x0123), vqmovn_s32(vacc3x4567));
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 vacc3x01234567 = vqaddq_s16(vacc3x01234567, voutput_zero_point);
258
259 int8x16_t vout0x01234567_1x01234567 = vcombine_s8(vqmovn_s16(vacc0x01234567), vqmovn_s16(vacc1x01234567));
260 int8x16_t vout2x01234567_3x01234567 = vcombine_s8(vqmovn_s16(vacc2x01234567), vqmovn_s16(vacc3x01234567));
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_3x01234567 = vmaxq_s8(vout2x01234567_3x01234567, 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_3x01234567 = vminq_s8(vout2x01234567_3x01234567, 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, vget_low_s8(vout2x01234567_3x01234567));
275 vst1_s8(c3 + 0, vget_high_s8(vout2x01234567_3x01234567));
276
277 c0 = (int8_t*) ((uintptr_t) c0 + cn_stride);
278 c1 = (int8_t*) ((uintptr_t) c1 + cn_stride);
279 c2 = (int8_t*) ((uintptr_t) c2 + cn_stride);
280 c3 = (int8_t*) ((uintptr_t) c3 + cn_stride);
281
282 a0 = (const int8_t*) ((uintptr_t) a0 - kc);
283 a1 = (const int8_t*) ((uintptr_t) a1 - kc);
284 a2 = (const int8_t*) ((uintptr_t) a2 - kc);
285 a3 = (const int8_t*) ((uintptr_t) a3 - kc);
286
287 nc -= 8;
288 } else {
289 // Final case where not all of the 8 columns fit in the destination.
290 if (nc & 4) {
291 vst1q_lane_u32((void*) c0, vreinterpretq_u32_s8(vout0x01234567_1x01234567), 0); c0 += 4;
292 vst1q_lane_u32((void*) c1, vreinterpretq_u32_s8(vout0x01234567_1x01234567), 2); c1 += 4;
293 vst1q_lane_u32((void*) c2, vreinterpretq_u32_s8(vout2x01234567_3x01234567), 0); c2 += 4;
294 vst1q_lane_u32((void*) c3, vreinterpretq_u32_s8(vout2x01234567_3x01234567), 2); c3 += 4;
295 vout0x01234567_1x01234567 = vextq_s8(vout0x01234567_1x01234567, vout0x01234567_1x01234567, 4);
296 vout2x01234567_3x01234567 = vextq_s8(vout2x01234567_3x01234567, vout2x01234567_3x01234567, 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 vst1q_lane_u16((void*) c2, vreinterpretq_u16_s8(vout2x01234567_3x01234567), 0); c2 += 2;
302 vst1q_lane_u16((void*) c3, vreinterpretq_u16_s8(vout2x01234567_3x01234567), 4); c3 += 2;
303 vout0x01234567_1x01234567 = vextq_s8(vout0x01234567_1x01234567, vout0x01234567_1x01234567, 2);
304 vout2x01234567_3x01234567 = vextq_s8(vout2x01234567_3x01234567, vout2x01234567_3x01234567, 2);
305 }
306 if (nc & 1) {
307 vst1q_lane_s8(c0, vout0x01234567_1x01234567, 0);
308 vst1q_lane_s8(c1, vout0x01234567_1x01234567, 8);
309 vst1q_lane_s8(c2, vout2x01234567_3x01234567, 0);
310 vst1q_lane_s8(c3, vout2x01234567_3x01234567, 8);
311 }
312
313 nc = 0;
314 }
315 } while (nc != 0);
316 }
317