Debian Codegraph @

src/qs8-qc8w-igemm/gen/qs8-qc8w-igemm-1x8-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_1x8__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
52 size_t p = ks;
53 do {
54 const int8_t* restrict a0 = a[0];
55 if XNN_UNPREDICTABLE(a0 != zero) {
56 a0 = (const int8_t*) ((uintptr_t) a0 + a_offset);
57 }
58 a += 1;
59
60 size_t k = kc;
61 while (k >= 8 * sizeof(int8_t)) {
62 const int8x8_t va0 = vld1_s8(a0); a0 += 8;
63 const int16x8_t vxa0 = vmovl_s8(va0);
64
65 const int8x8_t vb01234567c0 = vld1_s8(w); w = (const int8_t*) w + 8;
66 const int16x8_t vxb01234567c0 = vmovl_s8(vb01234567c0);
67
68 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c0), vget_low_s16(vxa0), 0);
69 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c0), vget_low_s16(vxa0), 0);
70 const int8x8_t vb01234567c1 = vld1_s8(w); w = (const int8_t*) w + 8;
71 const int16x8_t vxb01234567c1 = vmovl_s8(vb01234567c1);
72
73 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c1), vget_low_s16(vxa0), 1);
74 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c1), vget_low_s16(vxa0), 1);
75 const int8x8_t vb01234567c2 = vld1_s8(w); w = (const int8_t*) w + 8;
76 const int16x8_t vxb01234567c2 = vmovl_s8(vb01234567c2);
77
78 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c2), vget_low_s16(vxa0), 2);
79 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c2), vget_low_s16(vxa0), 2);
80 const int8x8_t vb01234567c3 = vld1_s8(w); w = (const int8_t*) w + 8;
81 const int16x8_t vxb01234567c3 = vmovl_s8(vb01234567c3);
82
83 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c3), vget_low_s16(vxa0), 3);
84 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c3), vget_low_s16(vxa0), 3);
85
86 xnn_prefetch_to_l1((const int8_t*) w + 448);
87
88 const int8x8_t vb01234567c4 = vld1_s8(w); w = (const int8_t*) w + 8;
89 const int16x8_t vxb01234567c4 = vmovl_s8(vb01234567c4);
90
91 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c4), vget_high_s16(vxa0), 0);
92 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c4), vget_high_s16(vxa0), 0);
93 const int8x8_t vb01234567c5 = vld1_s8(w); w = (const int8_t*) w + 8;
94 const int16x8_t vxb01234567c5 = vmovl_s8(vb01234567c5);
95
96 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c5), vget_high_s16(vxa0), 1);
97 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c5), vget_high_s16(vxa0), 1);
98 const int8x8_t vb01234567c6 = vld1_s8(w); w = (const int8_t*) w + 8;
99 const int16x8_t vxb01234567c6 = vmovl_s8(vb01234567c6);
100
101 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c6), vget_high_s16(vxa0), 2);
102 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c6), vget_high_s16(vxa0), 2);
103 const int8x8_t vb01234567c7 = vld1_s8(w); w = (const int8_t*) w + 8;
104 const int16x8_t vxb01234567c7 = vmovl_s8(vb01234567c7);
105
106 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c7), vget_high_s16(vxa0), 3);
107 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c7), vget_high_s16(vxa0), 3);
108
109 k -= 8 * sizeof(int8_t);
110 }
111 if XNN_UNLIKELY(k != 0) {
112 const int8x8_t va0 = vld1_s8(a0); a0 = (const int8_t*) ((uintptr_t) a0 + k);
113 const int16x8_t vxa0 = vmovl_s8(va0);
114
115 const int8x8_t vb01234567c0 = vld1_s8(w); w = (const int8_t*) w + 8;
116 const int16x8_t vxb01234567c0 = vmovl_s8(vb01234567c0);
117
118 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c0), vget_low_s16(vxa0), 0);
119 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c0), vget_low_s16(vxa0), 0);
120
121 if (k >= 2 * sizeof(int8_t)) {
122 const int8x8_t vb01234567c1 = vld1_s8(w); w = (const int8_t*) w + 8;
123 const int16x8_t vxb01234567c1 = vmovl_s8(vb01234567c1);
124
125 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c1), vget_low_s16(vxa0), 1);
126 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c1), vget_low_s16(vxa0), 1);
127
128 if (k > 2 * sizeof(int8_t)) {
129 const int8x8_t vb01234567c2 = vld1_s8(w); w = (const int8_t*) w + 8;
130 const int16x8_t vxb01234567c2 = vmovl_s8(vb01234567c2);
131
132 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c2), vget_low_s16(vxa0), 2);
133 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c2), vget_low_s16(vxa0), 2);
134
135 if (k >= 4 * sizeof(int8_t)) {
136 const int8x8_t vb01234567c3 = vld1_s8(w); w = (const int8_t*) w + 8;
137 const int16x8_t vxb01234567c3 = vmovl_s8(vb01234567c3);
138
139 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c3), vget_low_s16(vxa0), 3);
140 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c3), vget_low_s16(vxa0), 3);
141
142 if (k > 4 * sizeof(int8_t)) {
143 const int8x8_t vb01234567c4 = vld1_s8(w); w = (const int8_t*) w + 8;
144 const int16x8_t vxb01234567c4 = vmovl_s8(vb01234567c4);
145
146 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c4), vget_high_s16(vxa0), 0);
147 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c4), vget_high_s16(vxa0), 0);
148
149 if (k >= 6 * sizeof(int8_t)) {
150 const int8x8_t vb01234567c5 = vld1_s8(w); w = (const int8_t*) w + 8;
151 const int16x8_t vxb01234567c5 = vmovl_s8(vb01234567c5);
152
153 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c5), vget_high_s16(vxa0), 1);
154 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c5), vget_high_s16(vxa0), 1);
155
156 if (k > 6 * sizeof(int8_t)) {
157 const int8x8_t vb01234567c6 = vld1_s8(w); w = (const int8_t*) w + 8;
158 const int16x8_t vxb01234567c6 = vmovl_s8(vb01234567c6);
159
160 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c6), vget_high_s16(vxa0), 2);
161 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c6), vget_high_s16(vxa0), 2);
162 }
163 }
164 }
165 }
166 }
167 }
168 }
169 p -= 1 * sizeof(void*);
170 } while (p != 0);
171
172 // Post-accumulation work
173 float32x4_t vfpacc0x0123 = vcvtq_f32_s32(vacc0x0123);
174 float32x4_t vfpacc0x4567 = vcvtq_f32_s32(vacc0x4567);
175
176 const float32x4_t vscale0123 = vld1q_f32((const float*) w); w = (const float*) w + 4;
177 vfpacc0x0123 = vmulq_f32(vfpacc0x0123, vscale0123);
178 const float32x4_t vscale4567 = vld1q_f32((const float*) w); w = (const float*) w + 4;
179 vfpacc0x4567 = vmulq_f32(vfpacc0x4567, vscale4567);
180
181 vacc0x0123 = vcvtnq_s32_f32(vfpacc0x0123);
182 vacc0x4567 = vcvtnq_s32_f32(vfpacc0x4567);
183
184 const int16x8_t voutput_zero_point = vld1q_dup_s16(&params->fp32_neonv8.output_zero_point);
185 #if XNN_ARCH_ARM64
186 int16x8_t vacc0x01234567 = vqmovn_high_s32(vqmovn_s32(vacc0x0123), vacc0x4567);
187
188 vacc0x01234567 = vqaddq_s16(vacc0x01234567, voutput_zero_point);
189
190 int8x8_t vout0x01234567 = vqmovn_s16(vacc0x01234567);
191 #else
192 int16x8_t vacc0x01234567 = vcombine_s16(vqmovn_s32(vacc0x0123), vqmovn_s32(vacc0x4567));
193
194 vacc0x01234567 = vqaddq_s16(vacc0x01234567, voutput_zero_point);
195
196 int8x8_t vout0x01234567 = vqmovn_s16(vacc0x01234567);
197 #endif
198
199 const int8x8_t voutput_min = vld1_dup_s8(&params->fp32_neonv8.output_min);
200 vout0x01234567 = vmax_s8(vout0x01234567, voutput_min);
201
202 const int8x8_t voutput_max = vld1_dup_s8(&params->fp32_neonv8.output_max);
203 vout0x01234567 = vmin_s8(vout0x01234567, voutput_max);
204
205 if (nc >= 8) {
206 vst1_s8(c0 + 0, vout0x01234567);
207
208 c0 = (int8_t*) ((uintptr_t) c0 + cn_stride);
209
210 a = (const int8_t**restrict) ((uintptr_t) a - ks);
211
212 nc -= 8;
213 } else {
214 if (nc & 4) {
215 vst1_lane_u32((void*) c0, vreinterpret_u32_s8(vout0x01234567), 0); c0 += 4;
216 vout0x01234567 = vext_s8(vout0x01234567, vout0x01234567, 4);
217 }
218 if (nc & 2) {
219 vst1_lane_u16((void*) c0, vreinterpret_u16_s8(vout0x01234567), 0); c0 += 2;
220 vout0x01234567 = vext_s8(vout0x01234567, vout0x01234567, 2);
221 }
222 if (nc & 1) {
223 vst1_lane_s8(c0, vout0x01234567, 0);
224 }
225
226 nc = 0;
227 }
228 } while (nc != 0);
229 }