Debian Codegraph @

src/qs8-qc8w-igemm/gen/qs8-qc8w-igemm-1x16-minmax-fp32-neonv8-mlal-lane-prfm.c

raw
1 // Auto-generated file. Do not edit!
2 // Template: src/qs8-igemm/neon-mlal-lane.c.in
3 // Generator: tools/xngen
4 //
5 // Copyright 2020 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
11 #include <assert.h>
12
13 #include <arm_neon.h>
14
15 #include "xnnpack/common.h"
16 #include "xnnpack/igemm.h"
17 #include "xnnpack/intrinsics-polyfill.h"
18 #include "xnnpack/prefetch.h"
19
20
21 void xnn_qs8_qc8w_igemm_minmax_fp32_ukernel_1x16__neonv8_mlal_lane_prfm(
22 size_t mr,
23 size_t nc,
24 size_t kc,
25 size_t ks,
26 const int8_t** restrict a,
27 const void* restrict w,
28 int8_t* restrict c,
29 size_t cm_stride,
30 size_t cn_stride,
31 size_t a_offset,
32 const int8_t* zero,
33 const union xnn_qs8_qc8w_conv_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
34 {
35 assert(mr != 0);
36 assert(mr <= 1);
37 assert(nc != 0);
38 assert(kc != 0);
39 assert(ks != 0);
40 assert(ks % (1 * sizeof(void*)) == 0);
41 assert(a_offset % sizeof(int8_t) == 0);
42 assert(a != NULL);
43 assert(w != NULL);
44 assert(c != NULL);
45
46 int8_t* c0 = c;
47
48 do {
49 int32x4_t vacc0x0123 = vld1q_s32(w); w = (const void*) ((const int32_t*) w + 4);
50 int32x4_t vacc0x4567 = vld1q_s32(w); w = (const void*) ((const int32_t*) w + 4);
51 int32x4_t vacc0x89AB = vld1q_s32(w); w = (const void*) ((const int32_t*) w + 4);
52 int32x4_t vacc0xCDEF = vld1q_s32(w); w = (const void*) ((const int32_t*) w + 4);
53
54 size_t p = ks;
55 do {
56 const int8_t* restrict a0 = a[0];
57 if XNN_UNPREDICTABLE(a0 != zero) {
58 a0 = (const int8_t*) ((uintptr_t) a0 + a_offset);
59 }
60 a += 1;
61
62 size_t k = kc;
63 while (k >= 8 * sizeof(int8_t)) {
64 const int8x8_t va0 = vld1_s8(a0); a0 += 8;
65 const int16x8_t vxa0 = vmovl_s8(va0);
66
67 const int8x8_t vb01234567c0 = vld1_s8(w); w = (const int8_t*) w + 8;
68 const int16x8_t vxb01234567c0 = vmovl_s8(vb01234567c0);
69
70 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c0), vget_low_s16(vxa0), 0);
71 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c0), vget_low_s16(vxa0), 0);
72 const int8x8_t vb89ABCDEFc0 = vld1_s8(w); w = (const int8_t*) w + 8;
73 const int16x8_t vxb89ABCDEFc0 = vmovl_s8(vb89ABCDEFc0);
74
75 vacc0x89AB = vmlal_lane_s16(vacc0x89AB, vget_low_s16(vxb89ABCDEFc0), vget_low_s16(vxa0), 0);
76 vacc0xCDEF = vmlal_lane_s16(vacc0xCDEF, vget_high_s16(vxb89ABCDEFc0), vget_low_s16(vxa0), 0);
77 const int8x8_t vb01234567c1 = vld1_s8(w); w = (const int8_t*) w + 8;
78 const int16x8_t vxb01234567c1 = vmovl_s8(vb01234567c1);
79
80 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c1), vget_low_s16(vxa0), 1);
81 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c1), vget_low_s16(vxa0), 1);
82 const int8x8_t vb89ABCDEFc1 = vld1_s8(w); w = (const int8_t*) w + 8;
83 const int16x8_t vxb89ABCDEFc1 = vmovl_s8(vb89ABCDEFc1);
84
85 vacc0x89AB = vmlal_lane_s16(vacc0x89AB, vget_low_s16(vxb89ABCDEFc1), vget_low_s16(vxa0), 1);
86 vacc0xCDEF = vmlal_lane_s16(vacc0xCDEF, vget_high_s16(vxb89ABCDEFc1), vget_low_s16(vxa0), 1);
87 const int8x8_t vb01234567c2 = vld1_s8(w); w = (const int8_t*) w + 8;
88 const int16x8_t vxb01234567c2 = vmovl_s8(vb01234567c2);
89
90 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c2), vget_low_s16(vxa0), 2);
91 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c2), vget_low_s16(vxa0), 2);
92 const int8x8_t vb89ABCDEFc2 = vld1_s8(w); w = (const int8_t*) w + 8;
93 const int16x8_t vxb89ABCDEFc2 = vmovl_s8(vb89ABCDEFc2);
94
95 vacc0x89AB = vmlal_lane_s16(vacc0x89AB, vget_low_s16(vxb89ABCDEFc2), vget_low_s16(vxa0), 2);
96 vacc0xCDEF = vmlal_lane_s16(vacc0xCDEF, vget_high_s16(vxb89ABCDEFc2), vget_low_s16(vxa0), 2);
97 const int8x8_t vb01234567c3 = vld1_s8(w); w = (const int8_t*) w + 8;
98 const int16x8_t vxb01234567c3 = vmovl_s8(vb01234567c3);
99
100 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c3), vget_low_s16(vxa0), 3);
101 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c3), vget_low_s16(vxa0), 3);
102 const int8x8_t vb89ABCDEFc3 = vld1_s8(w); w = (const int8_t*) w + 8;
103 const int16x8_t vxb89ABCDEFc3 = vmovl_s8(vb89ABCDEFc3);
104
105 vacc0x89AB = vmlal_lane_s16(vacc0x89AB, vget_low_s16(vxb89ABCDEFc3), vget_low_s16(vxa0), 3);
106 vacc0xCDEF = vmlal_lane_s16(vacc0xCDEF, vget_high_s16(vxb89ABCDEFc3), vget_low_s16(vxa0), 3);
107
108 xnn_prefetch_to_l1((const int8_t*) w + 448);
109 xnn_prefetch_to_l1((const int8_t*) w + 512);
110
111 const int8x8_t vb01234567c4 = vld1_s8(w); w = (const int8_t*) w + 8;
112 const int16x8_t vxb01234567c4 = vmovl_s8(vb01234567c4);
113
114 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c4), vget_high_s16(vxa0), 0);
115 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c4), vget_high_s16(vxa0), 0);
116 const int8x8_t vb89ABCDEFc4 = vld1_s8(w); w = (const int8_t*) w + 8;
117 const int16x8_t vxb89ABCDEFc4 = vmovl_s8(vb89ABCDEFc4);
118
119 vacc0x89AB = vmlal_lane_s16(vacc0x89AB, vget_low_s16(vxb89ABCDEFc4), vget_high_s16(vxa0), 0);
120 vacc0xCDEF = vmlal_lane_s16(vacc0xCDEF, vget_high_s16(vxb89ABCDEFc4), vget_high_s16(vxa0), 0);
121 const int8x8_t vb01234567c5 = vld1_s8(w); w = (const int8_t*) w + 8;
122 const int16x8_t vxb01234567c5 = vmovl_s8(vb01234567c5);
123
124 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c5), vget_high_s16(vxa0), 1);
125 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c5), vget_high_s16(vxa0), 1);
126 const int8x8_t vb89ABCDEFc5 = vld1_s8(w); w = (const int8_t*) w + 8;
127 const int16x8_t vxb89ABCDEFc5 = vmovl_s8(vb89ABCDEFc5);
128
129 vacc0x89AB = vmlal_lane_s16(vacc0x89AB, vget_low_s16(vxb89ABCDEFc5), vget_high_s16(vxa0), 1);
130 vacc0xCDEF = vmlal_lane_s16(vacc0xCDEF, vget_high_s16(vxb89ABCDEFc5), vget_high_s16(vxa0), 1);
131 const int8x8_t vb01234567c6 = vld1_s8(w); w = (const int8_t*) w + 8;
132 const int16x8_t vxb01234567c6 = vmovl_s8(vb01234567c6);
133
134 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c6), vget_high_s16(vxa0), 2);
135 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c6), vget_high_s16(vxa0), 2);
136 const int8x8_t vb89ABCDEFc6 = vld1_s8(w); w = (const int8_t*) w + 8;
137 const int16x8_t vxb89ABCDEFc6 = vmovl_s8(vb89ABCDEFc6);
138
139 vacc0x89AB = vmlal_lane_s16(vacc0x89AB, vget_low_s16(vxb89ABCDEFc6), vget_high_s16(vxa0), 2);
140 vacc0xCDEF = vmlal_lane_s16(vacc0xCDEF, vget_high_s16(vxb89ABCDEFc6), vget_high_s16(vxa0), 2);
141 const int8x8_t vb01234567c7 = vld1_s8(w); w = (const int8_t*) w + 8;
142 const int16x8_t vxb01234567c7 = vmovl_s8(vb01234567c7);
143
144 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c7), vget_high_s16(vxa0), 3);
145 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c7), vget_high_s16(vxa0), 3);
146 const int8x8_t vb89ABCDEFc7 = vld1_s8(w); w = (const int8_t*) w + 8;
147 const int16x8_t vxb89ABCDEFc7 = vmovl_s8(vb89ABCDEFc7);
148
149 vacc0x89AB = vmlal_lane_s16(vacc0x89AB, vget_low_s16(vxb89ABCDEFc7), vget_high_s16(vxa0), 3);
150 vacc0xCDEF = vmlal_lane_s16(vacc0xCDEF, vget_high_s16(vxb89ABCDEFc7), vget_high_s16(vxa0), 3);
151
152 k -= 8 * sizeof(int8_t);
153 }
154 if XNN_UNLIKELY(k != 0) {
155 const int8x8_t va0 = vld1_s8(a0); a0 = (const int8_t*) ((uintptr_t) a0 + k);
156 const int16x8_t vxa0 = vmovl_s8(va0);
157
158 const int8x8_t vb01234567c0 = vld1_s8(w); w = (const int8_t*) w + 8;
159 const int16x8_t vxb01234567c0 = vmovl_s8(vb01234567c0);
160 const int8x8_t vb89ABCDEFc0 = vld1_s8(w); w = (const int8_t*) w + 8;
161 const int16x8_t vxb89ABCDEFc0 = vmovl_s8(vb89ABCDEFc0);
162
163 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c0), vget_low_s16(vxa0), 0);
164 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c0), vget_low_s16(vxa0), 0);
165 vacc0x89AB = vmlal_lane_s16(vacc0x89AB, vget_low_s16(vxb89ABCDEFc0), vget_low_s16(vxa0), 0);
166 vacc0xCDEF = vmlal_lane_s16(vacc0xCDEF, vget_high_s16(vxb89ABCDEFc0), vget_low_s16(vxa0), 0);
167
168 if (k >= 2 * sizeof(int8_t)) {
169 const int8x8_t vb01234567c1 = vld1_s8(w); w = (const int8_t*) w + 8;
170 const int16x8_t vxb01234567c1 = vmovl_s8(vb01234567c1);
171 const int8x8_t vb89ABCDEFc1 = vld1_s8(w); w = (const int8_t*) w + 8;
172 const int16x8_t vxb89ABCDEFc1 = vmovl_s8(vb89ABCDEFc1);
173
174 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c1), vget_low_s16(vxa0), 1);
175 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c1), vget_low_s16(vxa0), 1);
176 vacc0x89AB = vmlal_lane_s16(vacc0x89AB, vget_low_s16(vxb89ABCDEFc1), vget_low_s16(vxa0), 1);
177 vacc0xCDEF = vmlal_lane_s16(vacc0xCDEF, vget_high_s16(vxb89ABCDEFc1), vget_low_s16(vxa0), 1);
178
179 if (k > 2 * sizeof(int8_t)) {
180 const int8x8_t vb01234567c2 = vld1_s8(w); w = (const int8_t*) w + 8;
181 const int16x8_t vxb01234567c2 = vmovl_s8(vb01234567c2);
182 const int8x8_t vb89ABCDEFc2 = vld1_s8(w); w = (const int8_t*) w + 8;
183 const int16x8_t vxb89ABCDEFc2 = vmovl_s8(vb89ABCDEFc2);
184
185 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c2), vget_low_s16(vxa0), 2);
186 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c2), vget_low_s16(vxa0), 2);
187 vacc0x89AB = vmlal_lane_s16(vacc0x89AB, vget_low_s16(vxb89ABCDEFc2), vget_low_s16(vxa0), 2);
188 vacc0xCDEF = vmlal_lane_s16(vacc0xCDEF, vget_high_s16(vxb89ABCDEFc2), vget_low_s16(vxa0), 2);
189
190 if (k >= 4 * sizeof(int8_t)) {
191 const int8x8_t vb01234567c3 = vld1_s8(w); w = (const int8_t*) w + 8;
192 const int16x8_t vxb01234567c3 = vmovl_s8(vb01234567c3);
193 const int8x8_t vb89ABCDEFc3 = vld1_s8(w); w = (const int8_t*) w + 8;
194 const int16x8_t vxb89ABCDEFc3 = vmovl_s8(vb89ABCDEFc3);
195
196 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c3), vget_low_s16(vxa0), 3);
197 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c3), vget_low_s16(vxa0), 3);
198 vacc0x89AB = vmlal_lane_s16(vacc0x89AB, vget_low_s16(vxb89ABCDEFc3), vget_low_s16(vxa0), 3);
199 vacc0xCDEF = vmlal_lane_s16(vacc0xCDEF, vget_high_s16(vxb89ABCDEFc3), vget_low_s16(vxa0), 3);
200
201 if (k > 4 * sizeof(int8_t)) {
202 const int8x8_t vb01234567c4 = vld1_s8(w); w = (const int8_t*) w + 8;
203 const int16x8_t vxb01234567c4 = vmovl_s8(vb01234567c4);
204 const int8x8_t vb89ABCDEFc4 = vld1_s8(w); w = (const int8_t*) w + 8;
205 const int16x8_t vxb89ABCDEFc4 = vmovl_s8(vb89ABCDEFc4);
206
207 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c4), vget_high_s16(vxa0), 0);
208 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c4), vget_high_s16(vxa0), 0);
209 vacc0x89AB = vmlal_lane_s16(vacc0x89AB, vget_low_s16(vxb89ABCDEFc4), vget_high_s16(vxa0), 0);
210 vacc0xCDEF = vmlal_lane_s16(vacc0xCDEF, vget_high_s16(vxb89ABCDEFc4), vget_high_s16(vxa0), 0);
211
212 if (k >= 6 * sizeof(int8_t)) {
213 const int8x8_t vb01234567c5 = vld1_s8(w); w = (const int8_t*) w + 8;
214 const int16x8_t vxb01234567c5 = vmovl_s8(vb01234567c5);
215 const int8x8_t vb89ABCDEFc5 = vld1_s8(w); w = (const int8_t*) w + 8;
216 const int16x8_t vxb89ABCDEFc5 = vmovl_s8(vb89ABCDEFc5);
217
218 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c5), vget_high_s16(vxa0), 1);
219 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c5), vget_high_s16(vxa0), 1);
220 vacc0x89AB = vmlal_lane_s16(vacc0x89AB, vget_low_s16(vxb89ABCDEFc5), vget_high_s16(vxa0), 1);
221 vacc0xCDEF = vmlal_lane_s16(vacc0xCDEF, vget_high_s16(vxb89ABCDEFc5), vget_high_s16(vxa0), 1);
222
223 if (k > 6 * sizeof(int8_t)) {
224 const int8x8_t vb01234567c6 = vld1_s8(w); w = (const int8_t*) w + 8;
225 const int16x8_t vxb01234567c6 = vmovl_s8(vb01234567c6);
226 const int8x8_t vb89ABCDEFc6 = vld1_s8(w); w = (const int8_t*) w + 8;
227 const int16x8_t vxb89ABCDEFc6 = vmovl_s8(vb89ABCDEFc6);
228
229 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c6), vget_high_s16(vxa0), 2);
230 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c6), vget_high_s16(vxa0), 2);
231 vacc0x89AB = vmlal_lane_s16(vacc0x89AB, vget_low_s16(vxb89ABCDEFc6), vget_high_s16(vxa0), 2);
232 vacc0xCDEF = vmlal_lane_s16(vacc0xCDEF, vget_high_s16(vxb89ABCDEFc6), vget_high_s16(vxa0), 2);
233 }
234 }
235 }
236 }
237 }
238 }
239 }
240 p -= 1 * sizeof(void*);
241 } while (p != 0);
242
243 // Post-accumulation work
244 float32x4_t vfpacc0x0123 = vcvtq_f32_s32(vacc0x0123);
245 float32x4_t vfpacc0x4567 = vcvtq_f32_s32(vacc0x4567);
246 float32x4_t vfpacc0x89AB = vcvtq_f32_s32(vacc0x89AB);
247 float32x4_t vfpacc0xCDEF = vcvtq_f32_s32(vacc0xCDEF);
248
249 const float32x4_t vscale0123 = vld1q_f32((const float*) w); w = (const float*) w + 4;
250 vfpacc0x0123 = vmulq_f32(vfpacc0x0123, vscale0123);
251 const float32x4_t vscale4567 = vld1q_f32((const float*) w); w = (const float*) w + 4;
252 vfpacc0x4567 = vmulq_f32(vfpacc0x4567, vscale4567);
253 const float32x4_t vscale89AB = vld1q_f32((const float*) w); w = (const float*) w + 4;
254 vfpacc0x89AB = vmulq_f32(vfpacc0x89AB, vscale89AB);
255 const float32x4_t vscaleCDEF = vld1q_f32((const float*) w); w = (const float*) w + 4;
256 vfpacc0xCDEF = vmulq_f32(vfpacc0xCDEF, vscaleCDEF);
257
258 vacc0x0123 = vcvtnq_s32_f32(vfpacc0x0123);
259 vacc0x4567 = vcvtnq_s32_f32(vfpacc0x4567);
260 vacc0x89AB = vcvtnq_s32_f32(vfpacc0x89AB);
261 vacc0xCDEF = vcvtnq_s32_f32(vfpacc0xCDEF);
262
263 const int16x8_t voutput_zero_point = vld1q_dup_s16(&params->fp32_neonv8.output_zero_point);
264 #if XNN_ARCH_ARM64
265 int16x8_t vacc0x01234567 = vqmovn_high_s32(vqmovn_s32(vacc0x0123), vacc0x4567);
266 int16x8_t vacc0x89ABCDEF = vqmovn_high_s32(vqmovn_s32(vacc0x89AB), vacc0xCDEF);
267
268 vacc0x01234567 = vqaddq_s16(vacc0x01234567, voutput_zero_point);
269 vacc0x89ABCDEF = vqaddq_s16(vacc0x89ABCDEF, voutput_zero_point);
270
271 int8x16_t vout0x0123456789ABCDEF = vqmovn_high_s16(vqmovn_s16(vacc0x01234567), vacc0x89ABCDEF);
272 #else
273 int16x8_t vacc0x01234567 = vcombine_s16(vqmovn_s32(vacc0x0123), vqmovn_s32(vacc0x4567));
274 int16x8_t vacc0x89ABCDEF = vcombine_s16(vqmovn_s32(vacc0x89AB), vqmovn_s32(vacc0xCDEF));
275
276 vacc0x01234567 = vqaddq_s16(vacc0x01234567, voutput_zero_point);
277 vacc0x89ABCDEF = vqaddq_s16(vacc0x89ABCDEF, voutput_zero_point);
278
279 int8x16_t vout0x0123456789ABCDEF = vcombine_s8(vqmovn_s16(vacc0x01234567), vqmovn_s16(vacc0x89ABCDEF));
280 #endif
281
282 const int8x16_t voutput_min = vld1q_dup_s8(&params->fp32_neonv8.output_min);
283 vout0x0123456789ABCDEF = vmaxq_s8(vout0x0123456789ABCDEF, voutput_min);
284
285 const int8x16_t voutput_max = vld1q_dup_s8(&params->fp32_neonv8.output_max);
286 vout0x0123456789ABCDEF = vminq_s8(vout0x0123456789ABCDEF, voutput_max);
287
288 if (nc >= 16) {
289 vst1q_s8(c0 + 0, vout0x0123456789ABCDEF);
290
291 c0 = (int8_t*) ((uintptr_t) c0 + cn_stride);
292
293 a = (const int8_t**restrict) ((uintptr_t) a - ks);
294
295 nc -= 16;
296 } else {
297 int8x8_t vout0x01234567 = vget_low_s8(vout0x0123456789ABCDEF);
298 if (nc & 8) {
299 vst1_s8(c0, vout0x01234567); c0 += 8;
300 vout0x01234567 = vget_high_s8(vout0x0123456789ABCDEF);
301 }
302 if (nc & 4) {
303 vst1_lane_u32((void*) c0, vreinterpret_u32_s8(vout0x01234567), 0); c0 += 4;
304 vout0x01234567 = vext_s8(vout0x01234567, vout0x01234567, 4);
305 }
306 if (nc & 2) {
307 vst1_lane_u16((void*) c0, vreinterpret_u16_s8(vout0x01234567), 0); c0 += 2;
308 vout0x01234567 = vext_s8(vout0x01234567, vout0x01234567, 2);
309 }
310 if (nc & 1) {
311 vst1_lane_s8(c0, vout0x01234567, 0);
312 }
313
314 nc = 0;
315 }
316 } while (nc != 0);
317 }