Debian Codegraph @

src/qs8-qc8w-igemm/gen/qs8-qc8w-igemm-1x8-minmax-fp32-neon-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/prefetch.h"
18
19
20 void xnn_qs8_qc8w_igemm_minmax_fp32_ukernel_1x8__neon_mlal_lane_prfm(
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
51 size_t p = ks;
52 do {
53 const int8_t* restrict a0 = a[0];
54 if XNN_UNPREDICTABLE(a0 != zero) {
55 a0 = (const int8_t*) ((uintptr_t) a0 + a_offset);
56 }
57 a += 1;
58
59 size_t k = kc;
60 while (k >= 8 * sizeof(int8_t)) {
61 const int8x8_t va0 = vld1_s8(a0); a0 += 8;
62 const int16x8_t vxa0 = vmovl_s8(va0);
63
64 const int8x8_t vb01234567c0 = vld1_s8(w); w = (const int8_t*) w + 8;
65 const int16x8_t vxb01234567c0 = vmovl_s8(vb01234567c0);
66
67 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c0), vget_low_s16(vxa0), 0);
68 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c0), vget_low_s16(vxa0), 0);
69 const int8x8_t vb01234567c1 = vld1_s8(w); w = (const int8_t*) w + 8;
70 const int16x8_t vxb01234567c1 = vmovl_s8(vb01234567c1);
71
72 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c1), vget_low_s16(vxa0), 1);
73 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c1), vget_low_s16(vxa0), 1);
74 const int8x8_t vb01234567c2 = vld1_s8(w); w = (const int8_t*) w + 8;
75 const int16x8_t vxb01234567c2 = vmovl_s8(vb01234567c2);
76
77 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c2), vget_low_s16(vxa0), 2);
78 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c2), vget_low_s16(vxa0), 2);
79 const int8x8_t vb01234567c3 = vld1_s8(w); w = (const int8_t*) w + 8;
80 const int16x8_t vxb01234567c3 = vmovl_s8(vb01234567c3);
81
82 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c3), vget_low_s16(vxa0), 3);
83 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c3), vget_low_s16(vxa0), 3);
84
85 xnn_prefetch_to_l1((const int8_t*) w + 448);
86
87 const int8x8_t vb01234567c4 = vld1_s8(w); w = (const int8_t*) w + 8;
88 const int16x8_t vxb01234567c4 = vmovl_s8(vb01234567c4);
89
90 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c4), vget_high_s16(vxa0), 0);
91 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c4), vget_high_s16(vxa0), 0);
92 const int8x8_t vb01234567c5 = vld1_s8(w); w = (const int8_t*) w + 8;
93 const int16x8_t vxb01234567c5 = vmovl_s8(vb01234567c5);
94
95 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c5), vget_high_s16(vxa0), 1);
96 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c5), vget_high_s16(vxa0), 1);
97 const int8x8_t vb01234567c6 = vld1_s8(w); w = (const int8_t*) w + 8;
98 const int16x8_t vxb01234567c6 = vmovl_s8(vb01234567c6);
99
100 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c6), vget_high_s16(vxa0), 2);
101 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c6), vget_high_s16(vxa0), 2);
102 const int8x8_t vb01234567c7 = vld1_s8(w); w = (const int8_t*) w + 8;
103 const int16x8_t vxb01234567c7 = vmovl_s8(vb01234567c7);
104
105 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c7), vget_high_s16(vxa0), 3);
106 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c7), vget_high_s16(vxa0), 3);
107
108 k -= 8 * sizeof(int8_t);
109 }
110 if XNN_UNLIKELY(k != 0) {
111 const int8x8_t va0 = vld1_s8(a0); a0 = (const int8_t*) ((uintptr_t) a0 + k);
112 const int16x8_t vxa0 = vmovl_s8(va0);
113
114 const int8x8_t vb01234567c0 = vld1_s8(w); w = (const int8_t*) w + 8;
115 const int16x8_t vxb01234567c0 = vmovl_s8(vb01234567c0);
116
117 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c0), vget_low_s16(vxa0), 0);
118 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c0), vget_low_s16(vxa0), 0);
119
120 if (k >= 2 * sizeof(int8_t)) {
121 const int8x8_t vb01234567c1 = vld1_s8(w); w = (const int8_t*) w + 8;
122 const int16x8_t vxb01234567c1 = vmovl_s8(vb01234567c1);
123
124 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c1), vget_low_s16(vxa0), 1);
125 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c1), vget_low_s16(vxa0), 1);
126
127 if (k > 2 * sizeof(int8_t)) {
128 const int8x8_t vb01234567c2 = vld1_s8(w); w = (const int8_t*) w + 8;
129 const int16x8_t vxb01234567c2 = vmovl_s8(vb01234567c2);
130
131 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c2), vget_low_s16(vxa0), 2);
132 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c2), vget_low_s16(vxa0), 2);
133
134 if (k >= 4 * sizeof(int8_t)) {
135 const int8x8_t vb01234567c3 = vld1_s8(w); w = (const int8_t*) w + 8;
136 const int16x8_t vxb01234567c3 = vmovl_s8(vb01234567c3);
137
138 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c3), vget_low_s16(vxa0), 3);
139 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c3), vget_low_s16(vxa0), 3);
140
141 if (k > 4 * sizeof(int8_t)) {
142 const int8x8_t vb01234567c4 = vld1_s8(w); w = (const int8_t*) w + 8;
143 const int16x8_t vxb01234567c4 = vmovl_s8(vb01234567c4);
144
145 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c4), vget_high_s16(vxa0), 0);
146 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c4), vget_high_s16(vxa0), 0);
147
148 if (k >= 6 * sizeof(int8_t)) {
149 const int8x8_t vb01234567c5 = vld1_s8(w); w = (const int8_t*) w + 8;
150 const int16x8_t vxb01234567c5 = vmovl_s8(vb01234567c5);
151
152 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c5), vget_high_s16(vxa0), 1);
153 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c5), vget_high_s16(vxa0), 1);
154
155 if (k > 6 * sizeof(int8_t)) {
156 const int8x8_t vb01234567c6 = vld1_s8(w); w = (const int8_t*) w + 8;
157 const int16x8_t vxb01234567c6 = vmovl_s8(vb01234567c6);
158
159 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c6), vget_high_s16(vxa0), 2);
160 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c6), vget_high_s16(vxa0), 2);
161 }
162 }
163 }
164 }
165 }
166 }
167 }
168 p -= 1 * sizeof(void*);
169 } while (p != 0);
170
171 // Post-accumulation work
172 float32x4_t vfpacc0x0123 = vcvtq_f32_s32(vacc0x0123);
173 float32x4_t vfpacc0x4567 = vcvtq_f32_s32(vacc0x4567);
174
175 const float32x4_t vscale0123 = vld1q_f32((const float*) w); w = (const float*) w + 4;
176 vfpacc0x0123 = vmulq_f32(vfpacc0x0123, vscale0123);
177 const float32x4_t vscale4567 = vld1q_f32((const float*) w); w = (const float*) w + 4;
178 vfpacc0x4567 = vmulq_f32(vfpacc0x4567, vscale4567);
179
180 const float32x4_t vmagic_bias = vld1q_dup_f32(&params->fp32_neon.magic_bias);
181 vacc0x0123 = vreinterpretq_s32_f32(vaddq_f32(vfpacc0x0123, vmagic_bias));
182 vacc0x4567 = vreinterpretq_s32_f32(vaddq_f32(vfpacc0x4567, vmagic_bias));
183
184 const int32x4_t vmagic_bias_less_output_zero_point = vld1q_dup_s32(&params->fp32_neon.magic_bias_less_output_zero_point);
185 vacc0x0123 = vqsubq_s32(vacc0x0123, vmagic_bias_less_output_zero_point);
186 vacc0x4567 = vqsubq_s32(vacc0x4567, vmagic_bias_less_output_zero_point);
187
188 #if XNN_ARCH_ARM64
189 int16x8_t vacc0x01234567 = vqmovn_high_s32(vqmovn_s32(vacc0x0123), vacc0x4567);
190
191
192 int8x8_t vout0x01234567 = vqmovn_s16(vacc0x01234567);
193 #else
194 int16x8_t vacc0x01234567 = vcombine_s16(vqmovn_s32(vacc0x0123), vqmovn_s32(vacc0x4567));
195
196
197 int8x8_t vout0x01234567 = vqmovn_s16(vacc0x01234567);
198 #endif
199
200 const int8x8_t voutput_min = vld1_dup_s8(&params->fp32_neon.output_min);
201 vout0x01234567 = vmax_s8(vout0x01234567, voutput_min);
202
203 const int8x8_t voutput_max = vld1_dup_s8(&params->fp32_neon.output_max);
204 vout0x01234567 = vmin_s8(vout0x01234567, voutput_max);
205
206 if (nc >= 8) {
207 vst1_s8(c0 + 0, vout0x01234567);
208
209 c0 = (int8_t*) ((uintptr_t) c0 + cn_stride);
210
211 a = (const int8_t**restrict) ((uintptr_t) a - ks);
212
213 nc -= 8;
214 } else {
215 if (nc & 4) {
216 vst1_lane_u32((void*) c0, vreinterpret_u32_s8(vout0x01234567), 0); c0 += 4;
217 vout0x01234567 = vext_s8(vout0x01234567, vout0x01234567, 4);
218 }
219 if (nc & 2) {
220 vst1_lane_u16((void*) c0, vreinterpret_u16_s8(vout0x01234567), 0); c0 += 2;
221 vout0x01234567 = vext_s8(vout0x01234567, vout0x01234567, 2);
222 }
223 if (nc & 1) {
224 vst1_lane_s8(c0, vout0x01234567, 0);
225 }
226
227 nc = 0;
228 }
229 } while (nc != 0);
230 }