Debian Codegraph @

src/x32-packx/gen/x32-packx-8x-neon-st4-u4-prfm.c

raw
1 // Auto-generated file. Do not edit!
2 // Template: src/x32-packx/neon.c.in
3 // Generator: tools/xngen
4 //
5 // Copyright 2019 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/packx.h"
16 #include "xnnpack/prefetch.h"
17
18
19 void xnn_x32_packx_ukernel_8x__neon_st4_x4_prfm(
20 size_t m,
21 size_t k,
22 const uint32_t* x,
23 size_t x_stride,
24 uint32_t* restrict y)
25 {
26 assert(m != 0);
27 assert(m <= 8);
28 assert(k != 0);
29 assert(x != NULL);
30 assert(y != NULL);
31
32 const uint32_t* x0 = x;
33 const uint32_t* x1 = (const uint32_t*) ((uintptr_t) x0 + x_stride);
34 if XNN_UNPREDICTABLE(m < 2) {
35 x1 = x0;
36 }
37 const uint32_t* x2 = (const uint32_t*) ((uintptr_t) x1 + x_stride);
38 if XNN_UNPREDICTABLE(m <= 2) {
39 x2 = x1;
40 }
41 const uint32_t* x3 = (const uint32_t*) ((uintptr_t) x2 + x_stride);
42 if XNN_UNPREDICTABLE(m < 4) {
43 x3 = x2;
44 }
45 const uint32_t* x4 = (const uint32_t*) ((uintptr_t) x3 + x_stride);
46 if XNN_UNPREDICTABLE(m <= 4) {
47 x4 = x3;
48 }
49 const uint32_t* x5 = (const uint32_t*) ((uintptr_t) x4 + x_stride);
50 if XNN_UNPREDICTABLE(m < 6) {
51 x5 = x4;
52 }
53 const uint32_t* x6 = (const uint32_t*) ((uintptr_t) x5 + x_stride);
54 if XNN_UNPREDICTABLE(m <= 6) {
55 x6 = x5;
56 }
57 const uint32_t* x7 = (const uint32_t*) ((uintptr_t) x6 + x_stride);
58 if XNN_UNPREDICTABLE(m != 8) {
59 x7 = x6;
60 }
61
62 for (; k >= 4; k -= 4) {
63 const uint32x4_t vx0123x0 = vld1q_u32(x0); x0 += 4;
64 const uint32x4_t vx0123x1 = vld1q_u32(x1); x1 += 4;
65 const uint32x4_t vx0123x2 = vld1q_u32(x2); x2 += 4;
66 const uint32x4_t vx0123x3 = vld1q_u32(x3); x3 += 4;
67 const uint32x4_t vx0123x4 = vld1q_u32(x4); x4 += 4;
68 const uint32x4_t vx0123x5 = vld1q_u32(x5); x5 += 4;
69 const uint32x4_t vx0123x6 = vld1q_u32(x6); x6 += 4;
70 const uint32x4_t vx0123x7 = vld1q_u32(x7); x7 += 4;
71 xnn_prefetch_to_l1((const int8_t*) x0 + 128);
72 xnn_prefetch_to_l1((const int8_t*) x1 + 128);
73 xnn_prefetch_to_l1((const int8_t*) x2 + 128);
74 xnn_prefetch_to_l1((const int8_t*) x3 + 128);
75 xnn_prefetch_to_l1((const int8_t*) x4 + 128);
76 xnn_prefetch_to_l1((const int8_t*) x5 + 128);
77 xnn_prefetch_to_l1((const int8_t*) x6 + 128);
78 xnn_prefetch_to_l1((const int8_t*) x7 + 128);
79 const uint32x4x2_t vz0123x0 = vzipq_u32(vx0123x0, vx0123x4);
80 const uint32x4x2_t vz0123x1 = vzipq_u32(vx0123x1, vx0123x5);
81 const uint32x4x2_t vz0123x2 = vzipq_u32(vx0123x2, vx0123x6);
82 const uint32x4x2_t vz0123x3 = vzipq_u32(vx0123x3, vx0123x7);
83
84 const uint32x4x4_t vy0123x0 = { vz0123x0.val[0], vz0123x1.val[0], vz0123x2.val[0], vz0123x3.val[0] };
85 const uint32x4x4_t vy0123x1 = { vz0123x0.val[1], vz0123x1.val[1], vz0123x2.val[1], vz0123x3.val[1] };
86 vst4q_u32(y, vy0123x0); y += 16;
87 vst4q_u32(y, vy0123x1); y += 16;
88 }
89
90 if XNN_UNLIKELY(k != 0) {
91 uint32x4_t vt0123 = vdupq_n_u32(0);
92 uint32x4_t vt4567 = vdupq_n_u32(0);
93 do {
94 vt0123 = vld1q_lane_u32(x0, vt0123, 0); x0 += 1;
95 vt0123 = vld1q_lane_u32(x1, vt0123, 1); x1 += 1;
96 vt0123 = vld1q_lane_u32(x2, vt0123, 2); x2 += 1;
97 vt0123 = vld1q_lane_u32(x3, vt0123, 3); x3 += 1;
98 vt4567 = vld1q_lane_u32(x4, vt4567, 0); x4 += 1;
99 vt4567 = vld1q_lane_u32(x5, vt4567, 1); x5 += 1;
100 vt4567 = vld1q_lane_u32(x6, vt4567, 2); x6 += 1;
101 vt4567 = vld1q_lane_u32(x7, vt4567, 3); x7 += 1;
102 xnn_prefetch_to_l1((const int8_t*) x0 + 128);
103 xnn_prefetch_to_l1((const int8_t*) x1 + 128);
104 xnn_prefetch_to_l1((const int8_t*) x2 + 128);
105 xnn_prefetch_to_l1((const int8_t*) x3 + 128);
106 xnn_prefetch_to_l1((const int8_t*) x4 + 128);
107 xnn_prefetch_to_l1((const int8_t*) x5 + 128);
108 xnn_prefetch_to_l1((const int8_t*) x6 + 128);
109 xnn_prefetch_to_l1((const int8_t*) x7 + 128);
110 vst1q_u32(y, vt0123); y += 4;
111 vst1q_u32(y, vt4567); y += 4;
112 } while (--k != 0);
113 }
114 }