Debian Codegraph @

src/qs8-qc8w-igemm/gen/qs8-qc8w-igemm-1x16-minmax-fp32-neonv8-mlal-lane.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
19
20 void xnn_qs8_qc8w_igemm_minmax_fp32_ukernel_1x16__neonv8_mlal_lane(
21 size_t mr,
22 size_t nc,
23 size_t kc,
24 size_t ks,
25 const int8_t** restrict a,
26 const void* restrict w,
27 int8_t* restrict c,
28 size_t cm_stride,
29 size_t cn_stride,
30 size_t a_offset,
31 const int8_t* zero,
32 const union xnn_qs8_qc8w_conv_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
33 {
34 assert(mr != 0);
35 assert(mr <= 1);
36 assert(nc != 0);
37 assert(kc != 0);
38 assert(ks != 0);
39 assert(ks % (1 * sizeof(void*)) == 0);
40 assert(a_offset % sizeof(int8_t) == 0);
41 assert(a != NULL);
42 assert(w != NULL);
43 assert(c != NULL);
44
45 int8_t* c0 = c;
46
47 do {
48 int32x4_t vacc0x0123 = vld1q_s32(w); w = (const void*) ((const int32_t*) w + 4);
49 int32x4_t vacc0x4567 = vld1q_s32(w); w = (const void*) ((const int32_t*) w + 4);
50 int32x4_t vacc0x89AB = vld1q_s32(w); w = (const void*) ((const int32_t*) w + 4);
51 int32x4_t vacc0xCDEF = vld1q_s32(w); w = (const void*) ((const int32_t*) w + 4);
52
53 size_t p = ks;
54 do {
55 const int8_t* restrict a0 = a[0];
56 if XNN_UNPREDICTABLE(a0 != zero) {
57 a0 = (const int8_t*) ((uintptr_t) a0 + a_offset);
58 }
59 a += 1;
60
61 size_t k = kc;
62 while (k >= 8 * sizeof(int8_t)) {
63 const int8x8_t va0 = vld1_s8(a0); a0 += 8;
64 const int16x8_t vxa0 = vmovl_s8(va0);
65
66 const int8x8_t vb01234567c0 = vld1_s8(w); w = (const int8_t*) w + 8;
67 const int16x8_t vxb01234567c0 = vmovl_s8(vb01234567c0);
68
69 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c0), vget_low_s16(vxa0), 0);
70 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c0), vget_low_s16(vxa0), 0);
71 const int8x8_t vb89ABCDEFc0 = vld1_s8(w); w = (const int8_t*) w + 8;
72 const int16x8_t vxb89ABCDEFc0 = vmovl_s8(vb89ABCDEFc0);
73
74 vacc0x89AB = vmlal_lane_s16(vacc0x89AB, vget_low_s16(vxb89ABCDEFc0), vget_low_s16(vxa0), 0);
75 vacc0xCDEF = vmlal_lane_s16(vacc0xCDEF, vget_high_s16(vxb89ABCDEFc0), vget_low_s16(vxa0), 0);
76 const int8x8_t vb01234567c1 = vld1_s8(w); w = (const int8_t*) w + 8;
77 const int16x8_t vxb01234567c1 = vmovl_s8(vb01234567c1);
78
79 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c1), vget_low_s16(vxa0), 1);
80 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c1), vget_low_s16(vxa0), 1);
81 const int8x8_t vb89ABCDEFc1 = vld1_s8(w); w = (const int8_t*) w + 8;
82 const int16x8_t vxb89ABCDEFc1 = vmovl_s8(vb89ABCDEFc1);
83
84 vacc0x89AB = vmlal_lane_s16(vacc0x89AB, vget_low_s16(vxb89ABCDEFc1), vget_low_s16(vxa0), 1);
85 vacc0xCDEF = vmlal_lane_s16(vacc0xCDEF, vget_high_s16(vxb89ABCDEFc1), vget_low_s16(vxa0), 1);
86 const int8x8_t vb01234567c2 = vld1_s8(w); w = (const int8_t*) w + 8;
87 const int16x8_t vxb01234567c2 = vmovl_s8(vb01234567c2);
88
89 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c2), vget_low_s16(vxa0), 2);
90 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c2), vget_low_s16(vxa0), 2);
91 const int8x8_t vb89ABCDEFc2 = vld1_s8(w); w = (const int8_t*) w + 8;
92 const int16x8_t vxb89ABCDEFc2 = vmovl_s8(vb89ABCDEFc2);
93
94 vacc0x89AB = vmlal_lane_s16(vacc0x89AB, vget_low_s16(vxb89ABCDEFc2), vget_low_s16(vxa0), 2);
95 vacc0xCDEF = vmlal_lane_s16(vacc0xCDEF, vget_high_s16(vxb89ABCDEFc2), vget_low_s16(vxa0), 2);
96 const int8x8_t vb01234567c3 = vld1_s8(w); w = (const int8_t*) w + 8;
97 const int16x8_t vxb01234567c3 = vmovl_s8(vb01234567c3);
98
99 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c3), vget_low_s16(vxa0), 3);
100 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c3), vget_low_s16(vxa0), 3);
101 const int8x8_t vb89ABCDEFc3 = vld1_s8(w); w = (const int8_t*) w + 8;
102 const int16x8_t vxb89ABCDEFc3 = vmovl_s8(vb89ABCDEFc3);
103
104 vacc0x89AB = vmlal_lane_s16(vacc0x89AB, vget_low_s16(vxb89ABCDEFc3), vget_low_s16(vxa0), 3);
105 vacc0xCDEF = vmlal_lane_s16(vacc0xCDEF, vget_high_s16(vxb89ABCDEFc3), vget_low_s16(vxa0), 3);
106
107
108 const int8x8_t vb01234567c4 = vld1_s8(w); w = (const int8_t*) w + 8;
109 const int16x8_t vxb01234567c4 = vmovl_s8(vb01234567c4);
110
111 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c4), vget_high_s16(vxa0), 0);
112 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c4), vget_high_s16(vxa0), 0);
113 const int8x8_t vb89ABCDEFc4 = vld1_s8(w); w = (const int8_t*) w + 8;
114 const int16x8_t vxb89ABCDEFc4 = vmovl_s8(vb89ABCDEFc4);
115
116 vacc0x89AB = vmlal_lane_s16(vacc0x89AB, vget_low_s16(vxb89ABCDEFc4), vget_high_s16(vxa0), 0);
117 vacc0xCDEF = vmlal_lane_s16(vacc0xCDEF, vget_high_s16(vxb89ABCDEFc4), vget_high_s16(vxa0), 0);
118 const int8x8_t vb01234567c5 = vld1_s8(w); w = (const int8_t*) w + 8;
119 const int16x8_t vxb01234567c5 = vmovl_s8(vb01234567c5);
120
121 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c5), vget_high_s16(vxa0), 1);
122 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c5), vget_high_s16(vxa0), 1);
123 const int8x8_t vb89ABCDEFc5 = vld1_s8(w); w = (const int8_t*) w + 8;
124 const int16x8_t vxb89ABCDEFc5 = vmovl_s8(vb89ABCDEFc5);
125
126 vacc0x89AB = vmlal_lane_s16(vacc0x89AB, vget_low_s16(vxb89ABCDEFc5), vget_high_s16(vxa0), 1);
127 vacc0xCDEF = vmlal_lane_s16(vacc0xCDEF, vget_high_s16(vxb89ABCDEFc5), vget_high_s16(vxa0), 1);
128 const int8x8_t vb01234567c6 = vld1_s8(w); w = (const int8_t*) w + 8;
129 const int16x8_t vxb01234567c6 = vmovl_s8(vb01234567c6);
130
131 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c6), vget_high_s16(vxa0), 2);
132 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c6), vget_high_s16(vxa0), 2);
133 const int8x8_t vb89ABCDEFc6 = vld1_s8(w); w = (const int8_t*) w + 8;
134 const int16x8_t vxb89ABCDEFc6 = vmovl_s8(vb89ABCDEFc6);
135
136 vacc0x89AB = vmlal_lane_s16(vacc0x89AB, vget_low_s16(vxb89ABCDEFc6), vget_high_s16(vxa0), 2);
137 vacc0xCDEF = vmlal_lane_s16(vacc0xCDEF, vget_high_s16(vxb89ABCDEFc6), vget_high_s16(vxa0), 2);
138 const int8x8_t vb01234567c7 = vld1_s8(w); w = (const int8_t*) w + 8;
139 const int16x8_t vxb01234567c7 = vmovl_s8(vb01234567c7);
140
141 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c7), vget_high_s16(vxa0), 3);
142 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c7), vget_high_s16(vxa0), 3);
143 const int8x8_t vb89ABCDEFc7 = vld1_s8(w); w = (const int8_t*) w + 8;
144 const int16x8_t vxb89ABCDEFc7 = vmovl_s8(vb89ABCDEFc7);
145
146 vacc0x89AB = vmlal_lane_s16(vacc0x89AB, vget_low_s16(vxb89ABCDEFc7), vget_high_s16(vxa0), 3);
147 vacc0xCDEF = vmlal_lane_s16(vacc0xCDEF, vget_high_s16(vxb89ABCDEFc7), vget_high_s16(vxa0), 3);
148
149 k -= 8 * sizeof(int8_t);
150 }
151 if XNN_UNLIKELY(k != 0) {
152 const int8x8_t va0 = vld1_s8(a0); a0 = (const int8_t*) ((uintptr_t) a0 + k);
153 const int16x8_t vxa0 = vmovl_s8(va0);
154
155 const int8x8_t vb01234567c0 = vld1_s8(w); w = (const int8_t*) w + 8;
156 const int16x8_t vxb01234567c0 = vmovl_s8(vb01234567c0);
157 const int8x8_t vb89ABCDEFc0 = vld1_s8(w); w = (const int8_t*) w + 8;
158 const int16x8_t vxb89ABCDEFc0 = vmovl_s8(vb89ABCDEFc0);
159
160 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c0), vget_low_s16(vxa0), 0);
161 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c0), vget_low_s16(vxa0), 0);
162 vacc0x89AB = vmlal_lane_s16(vacc0x89AB, vget_low_s16(vxb89ABCDEFc0), vget_low_s16(vxa0), 0);
163 vacc0xCDEF = vmlal_lane_s16(vacc0xCDEF, vget_high_s16(vxb89ABCDEFc0), vget_low_s16(vxa0), 0);
164
165 if (k >= 2 * sizeof(int8_t)) {
166 const int8x8_t vb01234567c1 = vld1_s8(w); w = (const int8_t*) w + 8;
167 const int16x8_t vxb01234567c1 = vmovl_s8(vb01234567c1);
168 const int8x8_t vb89ABCDEFc1 = vld1_s8(w); w = (const int8_t*) w + 8;
169 const int16x8_t vxb89ABCDEFc1 = vmovl_s8(vb89ABCDEFc1);
170
171 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c1), vget_low_s16(vxa0), 1);
172 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c1), vget_low_s16(vxa0), 1);
173 vacc0x89AB = vmlal_lane_s16(vacc0x89AB, vget_low_s16(vxb89ABCDEFc1), vget_low_s16(vxa0), 1);
174 vacc0xCDEF = vmlal_lane_s16(vacc0xCDEF, vget_high_s16(vxb89ABCDEFc1), vget_low_s16(vxa0), 1);
175
176 if (k > 2 * sizeof(int8_t)) {
177 const int8x8_t vb01234567c2 = vld1_s8(w); w = (const int8_t*) w + 8;
178 const int16x8_t vxb01234567c2 = vmovl_s8(vb01234567c2);
179 const int8x8_t vb89ABCDEFc2 = vld1_s8(w); w = (const int8_t*) w + 8;
180 const int16x8_t vxb89ABCDEFc2 = vmovl_s8(vb89ABCDEFc2);
181
182 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c2), vget_low_s16(vxa0), 2);
183 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c2), vget_low_s16(vxa0), 2);
184 vacc0x89AB = vmlal_lane_s16(vacc0x89AB, vget_low_s16(vxb89ABCDEFc2), vget_low_s16(vxa0), 2);
185 vacc0xCDEF = vmlal_lane_s16(vacc0xCDEF, vget_high_s16(vxb89ABCDEFc2), vget_low_s16(vxa0), 2);
186
187 if (k >= 4 * sizeof(int8_t)) {
188 const int8x8_t vb01234567c3 = vld1_s8(w); w = (const int8_t*) w + 8;
189 const int16x8_t vxb01234567c3 = vmovl_s8(vb01234567c3);
190 const int8x8_t vb89ABCDEFc3 = vld1_s8(w); w = (const int8_t*) w + 8;
191 const int16x8_t vxb89ABCDEFc3 = vmovl_s8(vb89ABCDEFc3);
192
193 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c3), vget_low_s16(vxa0), 3);
194 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c3), vget_low_s16(vxa0), 3);
195 vacc0x89AB = vmlal_lane_s16(vacc0x89AB, vget_low_s16(vxb89ABCDEFc3), vget_low_s16(vxa0), 3);
196 vacc0xCDEF = vmlal_lane_s16(vacc0xCDEF, vget_high_s16(vxb89ABCDEFc3), vget_low_s16(vxa0), 3);
197
198 if (k > 4 * sizeof(int8_t)) {
199 const int8x8_t vb01234567c4 = vld1_s8(w); w = (const int8_t*) w + 8;
200 const int16x8_t vxb01234567c4 = vmovl_s8(vb01234567c4);
201 const int8x8_t vb89ABCDEFc4 = vld1_s8(w); w = (const int8_t*) w + 8;
202 const int16x8_t vxb89ABCDEFc4 = vmovl_s8(vb89ABCDEFc4);
203
204 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c4), vget_high_s16(vxa0), 0);
205 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c4), vget_high_s16(vxa0), 0);
206 vacc0x89AB = vmlal_lane_s16(vacc0x89AB, vget_low_s16(vxb89ABCDEFc4), vget_high_s16(vxa0), 0);
207 vacc0xCDEF = vmlal_lane_s16(vacc0xCDEF, vget_high_s16(vxb89ABCDEFc4), vget_high_s16(vxa0), 0);
208
209 if (k >= 6 * sizeof(int8_t)) {
210 const int8x8_t vb01234567c5 = vld1_s8(w); w = (const int8_t*) w + 8;
211 const int16x8_t vxb01234567c5 = vmovl_s8(vb01234567c5);
212 const int8x8_t vb89ABCDEFc5 = vld1_s8(w); w = (const int8_t*) w + 8;
213 const int16x8_t vxb89ABCDEFc5 = vmovl_s8(vb89ABCDEFc5);
214
215 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c5), vget_high_s16(vxa0), 1);
216 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c5), vget_high_s16(vxa0), 1);
217 vacc0x89AB = vmlal_lane_s16(vacc0x89AB, vget_low_s16(vxb89ABCDEFc5), vget_high_s16(vxa0), 1);
218 vacc0xCDEF = vmlal_lane_s16(vacc0xCDEF, vget_high_s16(vxb89ABCDEFc5), vget_high_s16(vxa0), 1);
219
220 if (k > 6 * sizeof(int8_t)) {
221 const int8x8_t vb01234567c6 = vld1_s8(w); w = (const int8_t*) w + 8;
222 const int16x8_t vxb01234567c6 = vmovl_s8(vb01234567c6);
223 const int8x8_t vb89ABCDEFc6 = vld1_s8(w); w = (const int8_t*) w + 8;
224 const int16x8_t vxb89ABCDEFc6 = vmovl_s8(vb89ABCDEFc6);
225
226 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c6), vget_high_s16(vxa0), 2);
227 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c6), vget_high_s16(vxa0), 2);
228 vacc0x89AB = vmlal_lane_s16(vacc0x89AB, vget_low_s16(vxb89ABCDEFc6), vget_high_s16(vxa0), 2);
229 vacc0xCDEF = vmlal_lane_s16(vacc0xCDEF, vget_high_s16(vxb89ABCDEFc6), vget_high_s16(vxa0), 2);
230 }
231 }
232 }
233 }
234 }
235 }
236 }
237 p -= 1 * sizeof(void*);
238 } while (p != 0);
239
240 // Post-accumulation work
241 float32x4_t vfpacc0x0123 = vcvtq_f32_s32(vacc0x0123);
242 float32x4_t vfpacc0x4567 = vcvtq_f32_s32(vacc0x4567);
243 float32x4_t vfpacc0x89AB = vcvtq_f32_s32(vacc0x89AB);
244 float32x4_t vfpacc0xCDEF = vcvtq_f32_s32(vacc0xCDEF);
245
246 const float32x4_t vscale0123 = vld1q_f32((const float*) w); w = (const float*) w + 4;
247 vfpacc0x0123 = vmulq_f32(vfpacc0x0123, vscale0123);
248 const float32x4_t vscale4567 = vld1q_f32((const float*) w); w = (const float*) w + 4;
249 vfpacc0x4567 = vmulq_f32(vfpacc0x4567, vscale4567);
250 const float32x4_t vscale89AB = vld1q_f32((const float*) w); w = (const float*) w + 4;
251 vfpacc0x89AB = vmulq_f32(vfpacc0x89AB, vscale89AB);
252 const float32x4_t vscaleCDEF = vld1q_f32((const float*) w); w = (const float*) w + 4;
253 vfpacc0xCDEF = vmulq_f32(vfpacc0xCDEF, vscaleCDEF);
254
255 vacc0x0123 = vcvtnq_s32_f32(vfpacc0x0123);
256 vacc0x4567 = vcvtnq_s32_f32(vfpacc0x4567);
257 vacc0x89AB = vcvtnq_s32_f32(vfpacc0x89AB);
258 vacc0xCDEF = vcvtnq_s32_f32(vfpacc0xCDEF);
259
260 const int16x8_t voutput_zero_point = vld1q_dup_s16(&params->fp32_neonv8.output_zero_point);
261 #if XNN_ARCH_ARM64
262 int16x8_t vacc0x01234567 = vqmovn_high_s32(vqmovn_s32(vacc0x0123), vacc0x4567);
263 int16x8_t vacc0x89ABCDEF = vqmovn_high_s32(vqmovn_s32(vacc0x89AB), vacc0xCDEF);
264
265 vacc0x01234567 = vqaddq_s16(vacc0x01234567, voutput_zero_point);
266 vacc0x89ABCDEF = vqaddq_s16(vacc0x89ABCDEF, voutput_zero_point);
267
268 int8x16_t vout0x0123456789ABCDEF = vqmovn_high_s16(vqmovn_s16(vacc0x01234567), vacc0x89ABCDEF);
269 #else
270 int16x8_t vacc0x01234567 = vcombine_s16(vqmovn_s32(vacc0x0123), vqmovn_s32(vacc0x4567));
271 int16x8_t vacc0x89ABCDEF = vcombine_s16(vqmovn_s32(vacc0x89AB), vqmovn_s32(vacc0xCDEF));
272
273 vacc0x01234567 = vqaddq_s16(vacc0x01234567, voutput_zero_point);
274 vacc0x89ABCDEF = vqaddq_s16(vacc0x89ABCDEF, voutput_zero_point);
275
276 int8x16_t vout0x0123456789ABCDEF = vcombine_s8(vqmovn_s16(vacc0x01234567), vqmovn_s16(vacc0x89ABCDEF));
277 #endif
278
279 const int8x16_t voutput_min = vld1q_dup_s8(&params->fp32_neonv8.output_min);
280 vout0x0123456789ABCDEF = vmaxq_s8(vout0x0123456789ABCDEF, voutput_min);
281
282 const int8x16_t voutput_max = vld1q_dup_s8(&params->fp32_neonv8.output_max);
283 vout0x0123456789ABCDEF = vminq_s8(vout0x0123456789ABCDEF, voutput_max);
284
285 if (nc >= 16) {
286 vst1q_s8(c0 + 0, vout0x0123456789ABCDEF);
287
288 c0 = (int8_t*) ((uintptr_t) c0 + cn_stride);
289
290 a = (const int8_t**restrict) ((uintptr_t) a - ks);
291
292 nc -= 16;
293 } else {
294 int8x8_t vout0x01234567 = vget_low_s8(vout0x0123456789ABCDEF);
295 if (nc & 8) {
296 vst1_s8(c0, vout0x01234567); c0 += 8;
297 vout0x01234567 = vget_high_s8(vout0x0123456789ABCDEF);
298 }
299 if (nc & 4) {
300 vst1_lane_u32((void*) c0, vreinterpret_u32_s8(vout0x01234567), 0); c0 += 4;
301 vout0x01234567 = vext_s8(vout0x01234567, vout0x01234567, 4);
302 }
303 if (nc & 2) {
304 vst1_lane_u16((void*) c0, vreinterpret_u16_s8(vout0x01234567), 0); c0 += 2;
305 vout0x01234567 = vext_s8(vout0x01234567, vout0x01234567, 2);
306 }
307 if (nc & 1) {
308 vst1_lane_s8(c0, vout0x01234567, 0);
309 }
310
311 nc = 0;
312 }
313 } while (nc != 0);
314 }