1 // Auto-generated file. Do not edit!
2 // Template: src/x32-transposec/neon-zip.c.in
3 // Generator: tools/xngen
4 //
5 // Copyright 2021 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 #include <arm_neon.h>
11
12 #include <assert.h>
13
14 #include <xnnpack/common.h>
15 #include <xnnpack/math.h>
16 #include <xnnpack/transpose.h>
17
xnn_x32_transposec_ukernel__4x4_reuse_multi_zip_neon(const uint32_t * input,uint32_t * output,size_t input_stride,size_t output_stride,size_t block_width,size_t block_height)18 void xnn_x32_transposec_ukernel__4x4_reuse_multi_zip_neon(
19 const uint32_t* input,
20 uint32_t* output,
21 size_t input_stride,
22 size_t output_stride,
23 size_t block_width,
24 size_t block_height) XNN_OOB_READS
25 {
26 assert(output_stride >= block_height * sizeof(uint32_t));
27 assert(input_stride >= block_width * sizeof(uint32_t));
28
29 const size_t tile_height = 4;
30 const size_t tile_width = 4;
31 const size_t tile_hbytes = tile_height * sizeof(uint32_t);
32 const size_t tile_wbytes = tile_width * sizeof(uint32_t);
33 const size_t input_reset = tile_wbytes - round_down_po2(block_height, tile_height) * input_stride;
34 const size_t output_reset = tile_width * output_stride - round_down_po2(block_height, 2) * sizeof(uint32_t);
35
36 const uint32_t* i0 = input;
37 uint32_t* o0 = (uint32_t*) output;
38 uint32_t* o1 = (uint32_t*) ((uintptr_t) o0 + output_stride);
39 uint32_t* o2 = (uint32_t*) ((uintptr_t) o1 + output_stride);
40 uint32_t* o3 = (uint32_t*) ((uintptr_t) o2 + output_stride);
41
42 do {
43 if XNN_UNPREDICTABLE(block_width < 2) {
44 o1 = o0;
45 }
46 if XNN_UNPREDICTABLE(block_width <= 2) {
47 o2 = o0;
48 }
49 if XNN_UNPREDICTABLE(block_width < 4) {
50 o3 = o0;
51 }
52 size_t bh = block_height;
53 for (; bh >= 4; bh -= 4) {
54 const uint32x4_t v2_0 = vld1q_u32(i0); i0 = (uint32_t*) ((uintptr_t) i0 + input_stride);
55 const uint32x4_t v2_1 = vld1q_u32(i0); i0 = (uint32_t*) ((uintptr_t) i0 + input_stride);
56 const uint32x4_t v2_2 = vld1q_u32(i0); i0 = (uint32_t*) ((uintptr_t) i0 + input_stride);
57 const uint32x4_t v2_3 = vld1q_u32(i0); i0 = (uint32_t*) ((uintptr_t) i0 + input_stride);
58
59 const uint32x4x2_t v1_0 = vzipq_u32(v2_0, v2_2);
60 const uint32x4x2_t v1_1 = vzipq_u32(v2_1, v2_3);
61
62 const uint32x4x2_t v0_0 = vzipq_u32(v1_0.val[0], v1_1.val[0]);
63 const uint32x4x2_t v0_1 = vzipq_u32(v1_0.val[1], v1_1.val[1]);
64
65 vst1q_u32(o3, v0_1.val[1]); o3 = (uint32_t*) ((uintptr_t) o3 + tile_hbytes);
66 vst1q_u32(o2, v0_1.val[0]); o2 = (uint32_t*) ((uintptr_t) o2 + tile_hbytes);
67 vst1q_u32(o1, v0_0.val[1]); o1 = (uint32_t*) ((uintptr_t) o1 + tile_hbytes);
68 vst1q_u32(o0, v0_0.val[0]); o0 = (uint32_t*) ((uintptr_t) o0 + tile_hbytes);
69 }
70
71 if (bh != 0) {
72 const uint32x4_t v2_0 = vld1q_u32(i0);
73 const uint32_t *i1 = (const uint32_t*) ((uintptr_t) i0 + input_stride);
74 if XNN_UNPREDICTABLE(bh < 2) {
75 i1 = i0;
76 }
77 const uint32x4_t v2_1 = vld1q_u32(i1);
78 const uint32_t *i2 = (const uint32_t*) ((uintptr_t) i1 + input_stride);
79 if XNN_UNPREDICTABLE(bh <= 2) {
80 i2 = i1;
81 }
82 const uint32x4_t v2_2 = vld1q_u32(i2);
83 const uint32x4_t v2_3 = vmovq_n_u32(0);
84
85 const uint32x4x2_t v1_0 = vzipq_u32(v2_0, v2_2);
86 const uint32x4x2_t v1_1 = vzipq_u32(v2_1, v2_3);
87
88 const uint32x4x2_t v0_0 = vzipq_u32(v1_0.val[0], v1_1.val[0]);
89 const uint32x4x2_t v0_1 = vzipq_u32(v1_0.val[1], v1_1.val[1]);
90
91 uint32x2_t v0_low = vget_low_u32(v0_0.val[0]);
92 uint32x2_t v1_low = vget_low_u32(v0_0.val[1]);
93 uint32x2_t v2_low = vget_low_u32(v0_1.val[0]);
94 uint32x2_t v3_low = vget_low_u32(v0_1.val[1]);
95
96 if (bh & 2) {
97 vst1_u32(o3, v3_low); o3 += 2;
98 vst1_u32(o2, v2_low); o2 += 2;
99 vst1_u32(o1, v1_low); o1 += 2;
100 vst1_u32(o0, v0_low); o0 += 2;
101 v0_low = vget_high_u32(v0_0.val[0]);
102 v1_low = vget_high_u32(v0_0.val[1]);
103 v2_low = vget_high_u32(v0_1.val[0]);
104 v3_low = vget_high_u32(v0_1.val[1]);
105 }
106
107 if (bh & 1) {
108 vst1_lane_u32(o3, v3_low, 0);
109 vst1_lane_u32(o2, v2_low, 0);
110 vst1_lane_u32(o1, v1_low, 0);
111 vst1_lane_u32(o0, v0_low, 0);
112 }
113 }
114
115 i0 = (const uint32_t*) ((uintptr_t) i0 + input_reset);
116 o0 = (uint32_t*) ((uintptr_t) o0 + output_reset);
117 o1 = (uint32_t*) ((uintptr_t) o1 + output_reset);
118 o2 = (uint32_t*) ((uintptr_t) o2 + output_reset);
119 o3 = (uint32_t*) ((uintptr_t) o3 + output_reset);
120 block_width = doz(block_width, tile_width);
121 } while (block_width != 0);
122 }
123