1 // Auto-generated file. Do not edit!
2 // Template: src/qs8-gemm/c2-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_2x16c2s4__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_2x16c2s4__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 <= 2);
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
48 kc = round_up_po2(kc, 8 * sizeof(int8_t));
49 do {
50 int32x4_t vacc0x0123 = vld1q_s32(w); w = (const int32_t*) w + 4;
51 int32x4_t vacc0x4567 = vld1q_s32(w); w = (const int32_t*) w + 4;
52 int32x4_t vacc0x89AB = vld1q_s32(w); w = (const int32_t*) w + 4;
53 int32x4_t vacc0xCDEF = vld1q_s32(w); w = (const int32_t*) w + 4;
54 int32x4_t vacc1x0123 = vacc0x0123;
55 int32x4_t vacc1x4567 = vacc0x4567;
56 int32x4_t vacc1x89AB = vacc0x89AB;
57 int32x4_t vacc1xCDEF = vacc0xCDEF;
58
59 size_t k = kc;
60 do {
61 int8x8_t va0x0 = vld1_s8(a0); a0 += 8;
62 int8x8_t va1x0 = vld1_s8(a1); a1 += 8;
63
64 const int8x8_t vb0123c0x0 = vld1_s8(w); w = (const int8_t*) w + 8;
65 const int8x8_t vb4567c0x0 = vld1_s8(w); w = (const int8_t*) w + 8;
66 const int8x8_t vb89ABc0x0 = vld1_s8(w); w = (const int8_t*) w + 8;
67 const int8x8_t vbCDEFc0x0 = vld1_s8(w); w = (const int8_t*) w + 8;
68 const int8x8_t vb0123c1x0 = vld1_s8(w); w = (const int8_t*) w + 8;
69 const int8x8_t vb4567c1x0 = vld1_s8(w); w = (const int8_t*) w + 8;
70 const int8x8_t vb89ABc1x0 = vld1_s8(w); w = (const int8_t*) w + 8;
71 const int8x8_t vbCDEFc1x0 = vld1_s8(w); w = (const int8_t*) w + 8;
72 const int8x8_t vb0123c2x0 = vld1_s8(w); w = (const int8_t*) w + 8;
73 const int8x8_t vb4567c2x0 = vld1_s8(w); w = (const int8_t*) w + 8;
74 const int8x8_t vb89ABc2x0 = vld1_s8(w); w = (const int8_t*) w + 8;
75 const int8x8_t vbCDEFc2x0 = vld1_s8(w); w = (const int8_t*) w + 8;
76 const int8x8_t vb0123c3x0 = vld1_s8(w); w = (const int8_t*) w + 8;
77 const int8x8_t vb4567c3x0 = vld1_s8(w); w = (const int8_t*) w + 8;
78 const int8x8_t vb89ABc3x0 = vld1_s8(w); w = (const int8_t*) w + 8;
79 const int8x8_t vbCDEFc3x0 = vld1_s8(w); w = (const int8_t*) w + 8;
80
81 int16x8_t vprod0x0123c0 = vmull_s8(vb0123c0x0, va0x0);
82 int16x8_t vprod1x0123c0 = vmull_s8(vb0123c0x0, va1x0);
83 vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c0);
84 vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c0);
85 int16x8_t vprod0x4567c0 = vmull_s8(vb4567c0x0, va0x0);
86 int16x8_t vprod1x4567c0 = vmull_s8(vb4567c0x0, va1x0);
87 vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c0);
88 vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c0);
89 int16x8_t vprod0x89ABc0 = vmull_s8(vb89ABc0x0, va0x0);
90 int16x8_t vprod1x89ABc0 = vmull_s8(vb89ABc0x0, va1x0);
91 vacc0x89AB = vpadalq_s16(vacc0x89AB, vprod0x89ABc0);
92 vacc1x89AB = vpadalq_s16(vacc1x89AB, vprod1x89ABc0);
93 int16x8_t vprod0xCDEFc0 = vmull_s8(vbCDEFc0x0, va0x0);
94 int16x8_t vprod1xCDEFc0 = vmull_s8(vbCDEFc0x0, va1x0);
95 vacc0xCDEF = vpadalq_s16(vacc0xCDEF, vprod0xCDEFc0);
96 vacc1xCDEF = vpadalq_s16(vacc1xCDEF, vprod1xCDEFc0);
97 va0x0 = vext_s8(va0x0, va0x0, 2);
98 va1x0 = vext_s8(va1x0, va1x0, 2);
99 int16x8_t vprod0x0123c1 = vmull_s8(vb0123c1x0, va0x0);
100 int16x8_t vprod1x0123c1 = vmull_s8(vb0123c1x0, va1x0);
101 vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c1);
102 vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c1);
103 int16x8_t vprod0x4567c1 = vmull_s8(vb4567c1x0, va0x0);
104 int16x8_t vprod1x4567c1 = vmull_s8(vb4567c1x0, va1x0);
105 vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c1);
106 vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c1);
107 int16x8_t vprod0x89ABc1 = vmull_s8(vb89ABc1x0, va0x0);
108 int16x8_t vprod1x89ABc1 = vmull_s8(vb89ABc1x0, va1x0);
109 vacc0x89AB = vpadalq_s16(vacc0x89AB, vprod0x89ABc1);
110 vacc1x89AB = vpadalq_s16(vacc1x89AB, vprod1x89ABc1);
111 int16x8_t vprod0xCDEFc1 = vmull_s8(vbCDEFc1x0, va0x0);
112 int16x8_t vprod1xCDEFc1 = vmull_s8(vbCDEFc1x0, va1x0);
113 vacc0xCDEF = vpadalq_s16(vacc0xCDEF, vprod0xCDEFc1);
114 vacc1xCDEF = vpadalq_s16(vacc1xCDEF, vprod1xCDEFc1);
115 va0x0 = vext_s8(va0x0, va0x0, 2);
116 va1x0 = vext_s8(va1x0, va1x0, 2);
117 int16x8_t vprod0x0123c2 = vmull_s8(vb0123c2x0, va0x0);
118 int16x8_t vprod1x0123c2 = vmull_s8(vb0123c2x0, va1x0);
119 vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c2);
120 vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c2);
121 int16x8_t vprod0x4567c2 = vmull_s8(vb4567c2x0, va0x0);
122 int16x8_t vprod1x4567c2 = vmull_s8(vb4567c2x0, va1x0);
123 vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c2);
124 vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c2);
125 int16x8_t vprod0x89ABc2 = vmull_s8(vb89ABc2x0, va0x0);
126 int16x8_t vprod1x89ABc2 = vmull_s8(vb89ABc2x0, va1x0);
127 vacc0x89AB = vpadalq_s16(vacc0x89AB, vprod0x89ABc2);
128 vacc1x89AB = vpadalq_s16(vacc1x89AB, vprod1x89ABc2);
129 int16x8_t vprod0xCDEFc2 = vmull_s8(vbCDEFc2x0, va0x0);
130 int16x8_t vprod1xCDEFc2 = vmull_s8(vbCDEFc2x0, va1x0);
131 vacc0xCDEF = vpadalq_s16(vacc0xCDEF, vprod0xCDEFc2);
132 vacc1xCDEF = vpadalq_s16(vacc1xCDEF, vprod1xCDEFc2);
133 va0x0 = vext_s8(va0x0, va0x0, 2);
134 va1x0 = vext_s8(va1x0, va1x0, 2);
135 int16x8_t vprod0x0123c3 = vmull_s8(vb0123c3x0, va0x0);
136 int16x8_t vprod1x0123c3 = vmull_s8(vb0123c3x0, va1x0);
137 vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c3);
138 vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c3);
139 int16x8_t vprod0x4567c3 = vmull_s8(vb4567c3x0, va0x0);
140 int16x8_t vprod1x4567c3 = vmull_s8(vb4567c3x0, va1x0);
141 vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c3);
142 vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c3);
143 int16x8_t vprod0x89ABc3 = vmull_s8(vb89ABc3x0, va0x0);
144 int16x8_t vprod1x89ABc3 = vmull_s8(vb89ABc3x0, va1x0);
145 vacc0x89AB = vpadalq_s16(vacc0x89AB, vprod0x89ABc3);
146 vacc1x89AB = vpadalq_s16(vacc1x89AB, vprod1x89ABc3);
147 int16x8_t vprod0xCDEFc3 = vmull_s8(vbCDEFc3x0, va0x0);
148 int16x8_t vprod1xCDEFc3 = vmull_s8(vbCDEFc3x0, va1x0);
149 vacc0xCDEF = vpadalq_s16(vacc0xCDEF, vprod0xCDEFc3);
150 vacc1xCDEF = vpadalq_s16(vacc1xCDEF, vprod1xCDEFc3);
151
152 k -= 8 * sizeof(int8_t);
153 } while (k != 0);
154
155 const int32x4_t vright_pre_shift = vld1q_dup_s32(¶ms->rndnu_neon.right_pre_shift);
156 const int32x4_t vmultiplier = vld1q_dup_s32(¶ms->rndnu_neon.multiplier);
157 const int32x4_t vright_post_shift = vld1q_dup_s32(¶ms->rndnu_neon.right_post_shift);
158
159 vacc0x0123 = vqshlq_s32(vacc0x0123, vright_pre_shift);
160 vacc0x4567 = vqshlq_s32(vacc0x4567, vright_pre_shift);
161 vacc0x89AB = vqshlq_s32(vacc0x89AB, vright_pre_shift);
162 vacc0xCDEF = vqshlq_s32(vacc0xCDEF, vright_pre_shift);
163 vacc1x0123 = vqshlq_s32(vacc1x0123, vright_pre_shift);
164 vacc1x4567 = vqshlq_s32(vacc1x4567, vright_pre_shift);
165 vacc1x89AB = vqshlq_s32(vacc1x89AB, vright_pre_shift);
166 vacc1xCDEF = vqshlq_s32(vacc1xCDEF, vright_pre_shift);
167
168 vacc0x0123 = vqdmulhq_s32(vacc0x0123, vmultiplier);
169 vacc0x4567 = vqdmulhq_s32(vacc0x4567, vmultiplier);
170 vacc0x89AB = vqdmulhq_s32(vacc0x89AB, vmultiplier);
171 vacc0xCDEF = vqdmulhq_s32(vacc0xCDEF, vmultiplier);
172 vacc1x0123 = vqdmulhq_s32(vacc1x0123, vmultiplier);
173 vacc1x4567 = vqdmulhq_s32(vacc1x4567, vmultiplier);
174 vacc1x89AB = vqdmulhq_s32(vacc1x89AB, vmultiplier);
175 vacc1xCDEF = vqdmulhq_s32(vacc1xCDEF, vmultiplier);
176
177 vacc0x0123 = vrshlq_s32(vacc0x0123, vright_post_shift);
178 vacc0x4567 = vrshlq_s32(vacc0x4567, vright_post_shift);
179 vacc0x89AB = vrshlq_s32(vacc0x89AB, vright_post_shift);
180 vacc0xCDEF = vrshlq_s32(vacc0xCDEF, vright_post_shift);
181 vacc1x0123 = vrshlq_s32(vacc1x0123, vright_post_shift);
182 vacc1x4567 = vrshlq_s32(vacc1x4567, vright_post_shift);
183 vacc1x89AB = vrshlq_s32(vacc1x89AB, vright_post_shift);
184 vacc1xCDEF = vrshlq_s32(vacc1xCDEF, vright_post_shift);
185
186 const int16x8_t voutput_zero_point = vld1q_dup_s16(¶ms->rndnu_neon.output_zero_point);
187 #if XNN_ARCH_ARM64
188 int16x8_t vacc0x01234567 = vqmovn_high_s32(vqmovn_s32(vacc0x0123), vacc0x4567);
189 int16x8_t vacc0x89ABCDEF = vqmovn_high_s32(vqmovn_s32(vacc0x89AB), vacc0xCDEF);
190 int16x8_t vacc1x01234567 = vqmovn_high_s32(vqmovn_s32(vacc1x0123), vacc1x4567);
191 int16x8_t vacc1x89ABCDEF = vqmovn_high_s32(vqmovn_s32(vacc1x89AB), vacc1xCDEF);
192
193 vacc0x01234567 = vqaddq_s16(vacc0x01234567, voutput_zero_point);
194 vacc0x89ABCDEF = vqaddq_s16(vacc0x89ABCDEF, voutput_zero_point);
195 vacc1x01234567 = vqaddq_s16(vacc1x01234567, voutput_zero_point);
196 vacc1x89ABCDEF = vqaddq_s16(vacc1x89ABCDEF, voutput_zero_point);
197
198 int8x16_t vout0x0123456789ABCDEF = vqmovn_high_s16(vqmovn_s16(vacc0x01234567), vacc0x89ABCDEF);
199 int8x16_t vout1x0123456789ABCDEF = vqmovn_high_s16(vqmovn_s16(vacc1x01234567), vacc1x89ABCDEF);
200 #else
201 int16x8_t vacc0x01234567 = vcombine_s16(vqmovn_s32(vacc0x0123), vqmovn_s32(vacc0x4567));
202 int16x8_t vacc0x89ABCDEF = vcombine_s16(vqmovn_s32(vacc0x89AB), vqmovn_s32(vacc0xCDEF));
203 int16x8_t vacc1x01234567 = vcombine_s16(vqmovn_s32(vacc1x0123), vqmovn_s32(vacc1x4567));
204 int16x8_t vacc1x89ABCDEF = vcombine_s16(vqmovn_s32(vacc1x89AB), vqmovn_s32(vacc1xCDEF));
205
206 vacc0x01234567 = vqaddq_s16(vacc0x01234567, voutput_zero_point);
207 vacc0x89ABCDEF = vqaddq_s16(vacc0x89ABCDEF, voutput_zero_point);
208 vacc1x01234567 = vqaddq_s16(vacc1x01234567, voutput_zero_point);
209 vacc1x89ABCDEF = vqaddq_s16(vacc1x89ABCDEF, voutput_zero_point);
210
211 int8x16_t vout0x0123456789ABCDEF = vcombine_s8(vqmovn_s16(vacc0x01234567), vqmovn_s16(vacc0x89ABCDEF));
212 int8x16_t vout1x0123456789ABCDEF = vcombine_s8(vqmovn_s16(vacc1x01234567), vqmovn_s16(vacc1x89ABCDEF));
213 #endif
214
215 const int8x16_t voutput_min = vld1q_dup_s8(¶ms->rndnu_neon.output_min);
216 vout0x0123456789ABCDEF = vmaxq_s8(vout0x0123456789ABCDEF, voutput_min);
217 vout1x0123456789ABCDEF = vmaxq_s8(vout1x0123456789ABCDEF, voutput_min);
218
219 const int8x16_t voutput_max = vld1q_dup_s8(¶ms->rndnu_neon.output_max);
220 vout0x0123456789ABCDEF = vminq_s8(vout0x0123456789ABCDEF, voutput_max);
221 vout1x0123456789ABCDEF = vminq_s8(vout1x0123456789ABCDEF, voutput_max);
222
223 if (nc >= 16) {
224 vst1q_s8(c0 + 0, vout0x0123456789ABCDEF);
225 vst1q_s8(c1 + 0, vout1x0123456789ABCDEF);
226
227 c0 = (int8_t*) ((uintptr_t) c0 + cn_stride);
228 c1 = (int8_t*) ((uintptr_t) c1 + cn_stride);
229
230 a0 = (const int8_t*) ((uintptr_t) a0 - kc);
231 a1 = (const int8_t*) ((uintptr_t) a1 - kc);
232
233 nc -= 16;
234 } else {
235 // Final case where not all of the 16 columns fit in the destination.
236 int8x16_t vout0x01234567_1x01234567 = vcombine_s8(vget_low_s8(vout0x0123456789ABCDEF), vget_low_s8(vout1x0123456789ABCDEF));
237 if (nc & 8) {
238 vst1_s8(c0, vget_low_s8(vout0x01234567_1x01234567)); c0 += 8;
239 vst1_s8(c1, vget_high_s8(vout0x01234567_1x01234567)); c1 += 8;
240 vout0x01234567_1x01234567 = vcombine_s8(vget_high_s8(vout0x0123456789ABCDEF), vget_high_s8(vout1x0123456789ABCDEF));
241 }
242 if (nc & 4) {
243 vst1q_lane_u32((void*) c0, vreinterpretq_u32_s8(vout0x01234567_1x01234567), 0); c0 += 4;
244 vst1q_lane_u32((void*) c1, vreinterpretq_u32_s8(vout0x01234567_1x01234567), 2); c1 += 4;
245 vout0x01234567_1x01234567 = vextq_s8(vout0x01234567_1x01234567, vout0x01234567_1x01234567, 4);
246 }
247 if (nc & 2) {
248 vst1q_lane_u16((void*) c0, vreinterpretq_u16_s8(vout0x01234567_1x01234567), 0); c0 += 2;
249 vst1q_lane_u16((void*) c1, vreinterpretq_u16_s8(vout0x01234567_1x01234567), 4); c1 += 2;
250 vout0x01234567_1x01234567 = vextq_s8(vout0x01234567_1x01234567, vout0x01234567_1x01234567, 2);
251 }
252 if (nc & 1) {
253 vst1q_lane_s8(c0, vout0x01234567_1x01234567, 0);
254 vst1q_lane_s8(c1, vout0x01234567_1x01234567, 8);
255 }
256
257 nc = 0;
258 }
259 } while (nc != 0);
260 }
261