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_1x16c2s4__neon_mlal(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_1x16c2s4__neon_mlal(
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 <= 1);
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
42 kc = round_up_po2(kc, 8 * sizeof(int8_t));
43 do {
44 int32x4_t vacc0x0123 = vld1q_s32(w); w = (const int32_t*) w + 4;
45 int32x4_t vacc0x4567 = vld1q_s32(w); w = (const int32_t*) w + 4;
46 int32x4_t vacc0x89AB = vld1q_s32(w); w = (const int32_t*) w + 4;
47 int32x4_t vacc0xCDEF = vld1q_s32(w); w = (const int32_t*) w + 4;
48
49 size_t k = kc;
50 while (k >= 16 * sizeof(int8_t)) {
51 int8x8_t va0x0 = vld1_s8(a0); a0 += 8;
52 int8x8_t va0x1 = vld1_s8(a0); a0 += 8;
53
54 const int8x8_t vb0123c0x0 = vld1_s8(w); w = (const int8_t*) w + 8;
55 const int8x8_t vb4567c0x0 = vld1_s8(w); w = (const int8_t*) w + 8;
56 const int8x8_t vb89ABc0x0 = vld1_s8(w); w = (const int8_t*) w + 8;
57 const int8x8_t vbCDEFc0x0 = vld1_s8(w); w = (const int8_t*) w + 8;
58 const int8x8_t vb0123c1x0 = vld1_s8(w); w = (const int8_t*) w + 8;
59 const int8x8_t vb4567c1x0 = vld1_s8(w); w = (const int8_t*) w + 8;
60 const int8x8_t vb89ABc1x0 = vld1_s8(w); w = (const int8_t*) w + 8;
61 const int8x8_t vbCDEFc1x0 = vld1_s8(w); w = (const int8_t*) w + 8;
62 const int8x8_t vb0123c2x0 = vld1_s8(w); w = (const int8_t*) w + 8;
63 const int8x8_t vb4567c2x0 = vld1_s8(w); w = (const int8_t*) w + 8;
64 const int8x8_t vb89ABc2x0 = vld1_s8(w); w = (const int8_t*) w + 8;
65 const int8x8_t vbCDEFc2x0 = vld1_s8(w); w = (const int8_t*) w + 8;
66 const int8x8_t vb0123c3x0 = vld1_s8(w); w = (const int8_t*) w + 8;
67 const int8x8_t vb4567c3x0 = vld1_s8(w); w = (const int8_t*) w + 8;
68 const int8x8_t vb89ABc3x0 = vld1_s8(w); w = (const int8_t*) w + 8;
69 const int8x8_t vbCDEFc3x0 = vld1_s8(w); w = (const int8_t*) w + 8;
70
71 int16x8_t vprod0x0123c0 = vmull_s8(vb0123c0x0, va0x0);
72 const int8x8_t vb0123c0x1 = vld1_s8(w); w = (const int8_t*) w + 8;
73 vprod0x0123c0 = vmlal_s8(vprod0x0123c0, vb0123c0x1, va0x1);
74 vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c0);
75 int16x8_t vprod0x4567c0 = vmull_s8(vb4567c0x0, va0x0);
76 const int8x8_t vb4567c0x1 = vld1_s8(w); w = (const int8_t*) w + 8;
77 vprod0x4567c0 = vmlal_s8(vprod0x4567c0, vb4567c0x1, va0x1);
78 vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c0);
79 int16x8_t vprod0x89ABc0 = vmull_s8(vb89ABc0x0, va0x0);
80 const int8x8_t vb89ABc0x1 = vld1_s8(w); w = (const int8_t*) w + 8;
81 vprod0x89ABc0 = vmlal_s8(vprod0x89ABc0, vb89ABc0x1, va0x1);
82 vacc0x89AB = vpadalq_s16(vacc0x89AB, vprod0x89ABc0);
83 int16x8_t vprod0xCDEFc0 = vmull_s8(vbCDEFc0x0, va0x0);
84 const int8x8_t vbCDEFc0x1 = vld1_s8(w); w = (const int8_t*) w + 8;
85 vprod0xCDEFc0 = vmlal_s8(vprod0xCDEFc0, vbCDEFc0x1, va0x1);
86 vacc0xCDEF = vpadalq_s16(vacc0xCDEF, vprod0xCDEFc0);
87 va0x0 = vext_s8(va0x0, va0x0, 2);
88 va0x1 = vext_s8(va0x1, va0x1, 2);
89 int16x8_t vprod0x0123c1 = vmull_s8(vb0123c1x0, va0x0);
90 const int8x8_t vb0123c1x1 = vld1_s8(w); w = (const int8_t*) w + 8;
91 vprod0x0123c1 = vmlal_s8(vprod0x0123c1, vb0123c1x1, va0x1);
92 vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c1);
93 int16x8_t vprod0x4567c1 = vmull_s8(vb4567c1x0, va0x0);
94 const int8x8_t vb4567c1x1 = vld1_s8(w); w = (const int8_t*) w + 8;
95 vprod0x4567c1 = vmlal_s8(vprod0x4567c1, vb4567c1x1, va0x1);
96 vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c1);
97 int16x8_t vprod0x89ABc1 = vmull_s8(vb89ABc1x0, va0x0);
98 const int8x8_t vb89ABc1x1 = vld1_s8(w); w = (const int8_t*) w + 8;
99 vprod0x89ABc1 = vmlal_s8(vprod0x89ABc1, vb89ABc1x1, va0x1);
100 vacc0x89AB = vpadalq_s16(vacc0x89AB, vprod0x89ABc1);
101 int16x8_t vprod0xCDEFc1 = vmull_s8(vbCDEFc1x0, va0x0);
102 const int8x8_t vbCDEFc1x1 = vld1_s8(w); w = (const int8_t*) w + 8;
103 vprod0xCDEFc1 = vmlal_s8(vprod0xCDEFc1, vbCDEFc1x1, va0x1);
104 vacc0xCDEF = vpadalq_s16(vacc0xCDEF, vprod0xCDEFc1);
105 va0x0 = vext_s8(va0x0, va0x0, 2);
106 va0x1 = vext_s8(va0x1, va0x1, 2);
107 int16x8_t vprod0x0123c2 = vmull_s8(vb0123c2x0, va0x0);
108 const int8x8_t vb0123c2x1 = vld1_s8(w); w = (const int8_t*) w + 8;
109 vprod0x0123c2 = vmlal_s8(vprod0x0123c2, vb0123c2x1, va0x1);
110 vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c2);
111 int16x8_t vprod0x4567c2 = vmull_s8(vb4567c2x0, va0x0);
112 const int8x8_t vb4567c2x1 = vld1_s8(w); w = (const int8_t*) w + 8;
113 vprod0x4567c2 = vmlal_s8(vprod0x4567c2, vb4567c2x1, va0x1);
114 vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c2);
115 int16x8_t vprod0x89ABc2 = vmull_s8(vb89ABc2x0, va0x0);
116 const int8x8_t vb89ABc2x1 = vld1_s8(w); w = (const int8_t*) w + 8;
117 vprod0x89ABc2 = vmlal_s8(vprod0x89ABc2, vb89ABc2x1, va0x1);
118 vacc0x89AB = vpadalq_s16(vacc0x89AB, vprod0x89ABc2);
119 int16x8_t vprod0xCDEFc2 = vmull_s8(vbCDEFc2x0, va0x0);
120 const int8x8_t vbCDEFc2x1 = vld1_s8(w); w = (const int8_t*) w + 8;
121 vprod0xCDEFc2 = vmlal_s8(vprod0xCDEFc2, vbCDEFc2x1, va0x1);
122 vacc0xCDEF = vpadalq_s16(vacc0xCDEF, vprod0xCDEFc2);
123 va0x0 = vext_s8(va0x0, va0x0, 2);
124 va0x1 = vext_s8(va0x1, va0x1, 2);
125 int16x8_t vprod0x0123c3 = vmull_s8(vb0123c3x0, va0x0);
126 const int8x8_t vb0123c3x1 = vld1_s8(w); w = (const int8_t*) w + 8;
127 vprod0x0123c3 = vmlal_s8(vprod0x0123c3, vb0123c3x1, va0x1);
128 vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c3);
129 int16x8_t vprod0x4567c3 = vmull_s8(vb4567c3x0, va0x0);
130 const int8x8_t vb4567c3x1 = vld1_s8(w); w = (const int8_t*) w + 8;
131 vprod0x4567c3 = vmlal_s8(vprod0x4567c3, vb4567c3x1, va0x1);
132 vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c3);
133 int16x8_t vprod0x89ABc3 = vmull_s8(vb89ABc3x0, va0x0);
134 const int8x8_t vb89ABc3x1 = vld1_s8(w); w = (const int8_t*) w + 8;
135 vprod0x89ABc3 = vmlal_s8(vprod0x89ABc3, vb89ABc3x1, va0x1);
136 vacc0x89AB = vpadalq_s16(vacc0x89AB, vprod0x89ABc3);
137 int16x8_t vprod0xCDEFc3 = vmull_s8(vbCDEFc3x0, va0x0);
138 const int8x8_t vbCDEFc3x1 = vld1_s8(w); w = (const int8_t*) w + 8;
139 vprod0xCDEFc3 = vmlal_s8(vprod0xCDEFc3, vbCDEFc3x1, va0x1);
140 vacc0xCDEF = vpadalq_s16(vacc0xCDEF, vprod0xCDEFc3);
141
142 k -= 16 * sizeof(int8_t);
143 }
144 if (k != 0) {
145 int8x8_t va0x0 = vld1_s8(a0); a0 += 8;
146
147 const int8x8_t vb0123c0x0 = vld1_s8(w); w = (const int8_t*) w + 8;
148 const int8x8_t vb4567c0x0 = vld1_s8(w); w = (const int8_t*) w + 8;
149 const int8x8_t vb89ABc0x0 = vld1_s8(w); w = (const int8_t*) w + 8;
150 const int8x8_t vbCDEFc0x0 = vld1_s8(w); w = (const int8_t*) w + 8;
151 const int8x8_t vb0123c1x0 = vld1_s8(w); w = (const int8_t*) w + 8;
152 const int8x8_t vb4567c1x0 = vld1_s8(w); w = (const int8_t*) w + 8;
153 const int8x8_t vb89ABc1x0 = vld1_s8(w); w = (const int8_t*) w + 8;
154 const int8x8_t vbCDEFc1x0 = vld1_s8(w); w = (const int8_t*) w + 8;
155 const int8x8_t vb0123c2x0 = vld1_s8(w); w = (const int8_t*) w + 8;
156 const int8x8_t vb4567c2x0 = vld1_s8(w); w = (const int8_t*) w + 8;
157 const int8x8_t vb89ABc2x0 = vld1_s8(w); w = (const int8_t*) w + 8;
158 const int8x8_t vbCDEFc2x0 = vld1_s8(w); w = (const int8_t*) w + 8;
159 const int8x8_t vb0123c3x0 = vld1_s8(w); w = (const int8_t*) w + 8;
160 const int8x8_t vb4567c3x0 = vld1_s8(w); w = (const int8_t*) w + 8;
161 const int8x8_t vb89ABc3x0 = vld1_s8(w); w = (const int8_t*) w + 8;
162 const int8x8_t vbCDEFc3x0 = vld1_s8(w); w = (const int8_t*) w + 8;
163
164 int16x8_t vprod0x0123c0 = vmull_s8(vb0123c0x0, va0x0);
165 vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c0);
166 int16x8_t vprod0x4567c0 = vmull_s8(vb4567c0x0, va0x0);
167 vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c0);
168 int16x8_t vprod0x89ABc0 = vmull_s8(vb89ABc0x0, va0x0);
169 vacc0x89AB = vpadalq_s16(vacc0x89AB, vprod0x89ABc0);
170 int16x8_t vprod0xCDEFc0 = vmull_s8(vbCDEFc0x0, va0x0);
171 vacc0xCDEF = vpadalq_s16(vacc0xCDEF, vprod0xCDEFc0);
172 va0x0 = vext_s8(va0x0, va0x0, 2);
173 int16x8_t vprod0x0123c1 = vmull_s8(vb0123c1x0, va0x0);
174 vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c1);
175 int16x8_t vprod0x4567c1 = vmull_s8(vb4567c1x0, va0x0);
176 vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c1);
177 int16x8_t vprod0x89ABc1 = vmull_s8(vb89ABc1x0, va0x0);
178 vacc0x89AB = vpadalq_s16(vacc0x89AB, vprod0x89ABc1);
179 int16x8_t vprod0xCDEFc1 = vmull_s8(vbCDEFc1x0, va0x0);
180 vacc0xCDEF = vpadalq_s16(vacc0xCDEF, vprod0xCDEFc1);
181 va0x0 = vext_s8(va0x0, va0x0, 2);
182 int16x8_t vprod0x0123c2 = vmull_s8(vb0123c2x0, va0x0);
183 vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c2);
184 int16x8_t vprod0x4567c2 = vmull_s8(vb4567c2x0, va0x0);
185 vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c2);
186 int16x8_t vprod0x89ABc2 = vmull_s8(vb89ABc2x0, va0x0);
187 vacc0x89AB = vpadalq_s16(vacc0x89AB, vprod0x89ABc2);
188 int16x8_t vprod0xCDEFc2 = vmull_s8(vbCDEFc2x0, va0x0);
189 vacc0xCDEF = vpadalq_s16(vacc0xCDEF, vprod0xCDEFc2);
190 va0x0 = vext_s8(va0x0, va0x0, 2);
191 int16x8_t vprod0x0123c3 = vmull_s8(vb0123c3x0, va0x0);
192 vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c3);
193 int16x8_t vprod0x4567c3 = vmull_s8(vb4567c3x0, va0x0);
194 vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c3);
195 int16x8_t vprod0x89ABc3 = vmull_s8(vb89ABc3x0, va0x0);
196 vacc0x89AB = vpadalq_s16(vacc0x89AB, vprod0x89ABc3);
197 int16x8_t vprod0xCDEFc3 = vmull_s8(vbCDEFc3x0, va0x0);
198 vacc0xCDEF = vpadalq_s16(vacc0xCDEF, vprod0xCDEFc3);
199
200 }
201
202 const int32x4_t vright_pre_shift = vld1q_dup_s32(¶ms->rndnu_neon.right_pre_shift);
203 const int32x4_t vmultiplier = vld1q_dup_s32(¶ms->rndnu_neon.multiplier);
204 const int32x4_t vright_post_shift = vld1q_dup_s32(¶ms->rndnu_neon.right_post_shift);
205
206 vacc0x0123 = vqshlq_s32(vacc0x0123, vright_pre_shift);
207 vacc0x4567 = vqshlq_s32(vacc0x4567, vright_pre_shift);
208 vacc0x89AB = vqshlq_s32(vacc0x89AB, vright_pre_shift);
209 vacc0xCDEF = vqshlq_s32(vacc0xCDEF, vright_pre_shift);
210
211 vacc0x0123 = vqdmulhq_s32(vacc0x0123, vmultiplier);
212 vacc0x4567 = vqdmulhq_s32(vacc0x4567, vmultiplier);
213 vacc0x89AB = vqdmulhq_s32(vacc0x89AB, vmultiplier);
214 vacc0xCDEF = vqdmulhq_s32(vacc0xCDEF, vmultiplier);
215
216 vacc0x0123 = vrshlq_s32(vacc0x0123, vright_post_shift);
217 vacc0x4567 = vrshlq_s32(vacc0x4567, vright_post_shift);
218 vacc0x89AB = vrshlq_s32(vacc0x89AB, vright_post_shift);
219 vacc0xCDEF = vrshlq_s32(vacc0xCDEF, vright_post_shift);
220
221 const int16x8_t voutput_zero_point = vld1q_dup_s16(¶ms->rndnu_neon.output_zero_point);
222 #if XNN_ARCH_ARM64
223 int16x8_t vacc0x01234567 = vqmovn_high_s32(vqmovn_s32(vacc0x0123), vacc0x4567);
224 int16x8_t vacc0x89ABCDEF = vqmovn_high_s32(vqmovn_s32(vacc0x89AB), vacc0xCDEF);
225
226 vacc0x01234567 = vqaddq_s16(vacc0x01234567, voutput_zero_point);
227 vacc0x89ABCDEF = vqaddq_s16(vacc0x89ABCDEF, voutput_zero_point);
228
229 int8x16_t vout0x0123456789ABCDEF = vqmovn_high_s16(vqmovn_s16(vacc0x01234567), vacc0x89ABCDEF);
230 #else
231 int16x8_t vacc0x01234567 = vcombine_s16(vqmovn_s32(vacc0x0123), vqmovn_s32(vacc0x4567));
232 int16x8_t vacc0x89ABCDEF = vcombine_s16(vqmovn_s32(vacc0x89AB), vqmovn_s32(vacc0xCDEF));
233
234 vacc0x01234567 = vqaddq_s16(vacc0x01234567, voutput_zero_point);
235 vacc0x89ABCDEF = vqaddq_s16(vacc0x89ABCDEF, voutput_zero_point);
236
237 int8x16_t vout0x0123456789ABCDEF = vcombine_s8(vqmovn_s16(vacc0x01234567), vqmovn_s16(vacc0x89ABCDEF));
238 #endif
239
240 const int8x16_t voutput_min = vld1q_dup_s8(¶ms->rndnu_neon.output_min);
241 vout0x0123456789ABCDEF = vmaxq_s8(vout0x0123456789ABCDEF, voutput_min);
242
243 const int8x16_t voutput_max = vld1q_dup_s8(¶ms->rndnu_neon.output_max);
244 vout0x0123456789ABCDEF = vminq_s8(vout0x0123456789ABCDEF, voutput_max);
245
246 if (nc >= 16) {
247 vst1q_s8(c0 + 0, vout0x0123456789ABCDEF);
248
249 c0 = (int8_t*) ((uintptr_t) c0 + cn_stride);
250
251 a0 = (const int8_t*) ((uintptr_t) a0 - kc);
252
253 nc -= 16;
254 } else {
255 // Final case where not all of the 16 columns fit in the destination.
256 int8x8_t vout0x01234567 = vget_low_s8(vout0x0123456789ABCDEF);
257 if (nc & 8) {
258 vst1_s8(c0, vout0x01234567); c0 += 8;
259 vout0x01234567 = vget_high_s8(vout0x0123456789ABCDEF);
260 }
261 if (nc & 4) {
262 vst1_lane_u32((void*) c0, vreinterpret_u32_s8(vout0x01234567), 0); c0 += 4;
263 vout0x01234567 = vext_s8(vout0x01234567, vout0x01234567, 4);
264 }
265 if (nc & 2) {
266 vst1_lane_u16((void*) c0, vreinterpret_u16_s8(vout0x01234567), 0); c0 += 2;
267 vout0x01234567 = vext_s8(vout0x01234567, vout0x01234567, 2);
268 }
269 if (nc & 1) {
270 vst1_lane_s8(c0, vout0x01234567, 0);
271 }
272
273 nc = 0;
274 }
275 } while (nc != 0);
276 }
277