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_switch_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_switch_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* o = (uint32_t*) output;
38 const size_t minus_output_stride = -output_stride;
39
40 do {
41 const size_t rem = min(block_width - 1, 3);
42 const size_t oN_stride = rem * output_stride;
43 size_t bh = block_height;
44 for (; bh >= 4; bh -= 4) {
45 const uint32x4_t v2_0 = vld1q_u32(i0); i0 = (uint32_t*) ((uintptr_t) i0 + input_stride);
46 const uint32x4_t v2_1 = vld1q_u32(i0); i0 = (uint32_t*) ((uintptr_t) i0 + input_stride);
47 const uint32x4_t v2_2 = vld1q_u32(i0); i0 = (uint32_t*) ((uintptr_t) i0 + input_stride);
48 const uint32x4_t v2_3 = vld1q_u32(i0); i0 = (uint32_t*) ((uintptr_t) i0 + input_stride);
49
50 const uint32x4x2_t v1_0 = vzipq_u32(v2_0, v2_2);
51 const uint32x4x2_t v1_1 = vzipq_u32(v2_1, v2_3);
52
53 const uint32x4x2_t v0_0 = vzipq_u32(v1_0.val[0], v1_1.val[0]);
54 const uint32x4x2_t v0_1 = vzipq_u32(v1_0.val[1], v1_1.val[1]);
55
56 uint32_t *oN = (uint32_t*) ((uintptr_t) o + oN_stride);
57 switch (rem) {
58 case 3:
59 vst1q_u32(oN, v0_1.val[1]); oN = (uint32_t*) ((uintptr_t) oN + minus_output_stride);
60 case 2:
61 vst1q_u32(oN, v0_1.val[0]); oN = (uint32_t*) ((uintptr_t) oN + minus_output_stride);
62 case 1:
63 vst1q_u32(oN, v0_0.val[1]);
64 case 0:
65 vst1q_u32(o, v0_0.val[0]); o = (uint32_t*) ((uintptr_t) o + tile_hbytes);
66 break;
67 default:
68 XNN_UNREACHABLE;
69 }
70 }
71
72 if (bh != 0) {
73 const uint32x4_t v2_0 = vld1q_u32(i0);
74 const uint32_t *i1 = (const uint32_t*) ((uintptr_t) i0 + input_stride);
75 if XNN_UNPREDICTABLE(bh < 2) {
76 i1 = i0;
77 }
78 const uint32x4_t v2_1 = vld1q_u32(i1);
79 const uint32_t *i2 = (const uint32_t*) ((uintptr_t) i1 + input_stride);
80 if XNN_UNPREDICTABLE(bh <= 2) {
81 i2 = i1;
82 }
83 const uint32x4_t v2_2 = vld1q_u32(i2);
84 const uint32x4_t v2_3 = vmovq_n_u32(0);
85
86 const uint32x4x2_t v1_0 = vzipq_u32(v2_0, v2_2);
87 const uint32x4x2_t v1_1 = vzipq_u32(v2_1, v2_3);
88
89 const uint32x4x2_t v0_0 = vzipq_u32(v1_0.val[0], v1_1.val[0]);
90 const uint32x4x2_t v0_1 = vzipq_u32(v1_0.val[1], v1_1.val[1]);
91
92 uint32x2_t v0_low = vget_low_u32(v0_0.val[0]);
93 uint32x2_t v1_low = vget_low_u32(v0_0.val[1]);
94 uint32x2_t v2_low = vget_low_u32(v0_1.val[0]);
95 uint32x2_t v3_low = vget_low_u32(v0_1.val[1]);
96
97 if (bh & 2) {
98 uint32_t* oN = (uint32_t*) ((uintptr_t) o + oN_stride);
99 switch (rem) {
100 case 3:
101 vst1_u32(oN, v3_low); oN = (uint32_t*) ((uintptr_t) oN + minus_output_stride);
102 case 2:
103 vst1_u32(oN, v2_low); oN = (uint32_t*) ((uintptr_t) oN + minus_output_stride);
104 case 1:
105 vst1_u32(oN, v1_low);
106 case 0:
107 vst1_u32(o, v0_low); o += 2;
108 break;
109 default:
110 XNN_UNREACHABLE;
111 }
112 v0_low = vget_high_u32(v0_0.val[0]);
113 v1_low = vget_high_u32(v0_0.val[1]);
114 v2_low = vget_high_u32(v0_1.val[0]);
115 v3_low = vget_high_u32(v0_1.val[1]);
116 }
117
118 if (bh & 1) {
119 uint32_t* oN = (uint32_t*) ((uintptr_t) o + oN_stride);
120 switch (rem) {
121 case 3:
122 vst1_lane_u32(oN, v3_low, 0); oN = (uint32_t*) ((uintptr_t) oN + minus_output_stride);
123 case 2:
124 vst1_lane_u32(oN, v2_low, 0); oN = (uint32_t*) ((uintptr_t) oN + minus_output_stride);
125 case 1:
126 vst1_lane_u32(oN, v1_low, 0);
127 case 0:
128 vst1_lane_u32(o, v0_low, 0);
129 break;
130 default:
131 XNN_UNREACHABLE;
132 }
133 }
134 }
135
136 i0 = (const uint32_t*) ((uintptr_t) i0 + input_reset);
137 o = (uint32_t*) ((uintptr_t) o + output_reset);
138 block_width = doz(block_width, tile_width);
139 } while (block_width != 0);
140 }
141