xref: /aosp_15_r20/external/XNNPACK/src/x32-transposec/gen/4x4-reuse-dec-zip-neon.c (revision 4bdc94577ba0e567308109d787f7fec7b531ce36)
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_dec_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_dec_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) - tile_hbytes;
35 
36   const uint32_t* i0 = input;
37   uint32_t* o = (uint32_t*) ((uintptr_t) output - tile_hbytes);
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     const size_t oN_offset = oN_stride + tile_hbytes;
44     size_t bh = block_height;
45     for (; bh >= 4; bh -= 4) {
46       const uint32x4_t v2_0 = vld1q_u32(i0); i0 = (uint32_t*) ((uintptr_t) i0 + input_stride);
47       const uint32x4_t v2_1 = vld1q_u32(i0); i0 = (uint32_t*) ((uintptr_t) i0 + input_stride);
48       const uint32x4_t v2_2 = vld1q_u32(i0); i0 = (uint32_t*) ((uintptr_t) i0 + input_stride);
49       const uint32x4_t v2_3 = vld1q_u32(i0); i0 = (uint32_t*) ((uintptr_t) i0 + input_stride);
50 
51       const uint32x4x2_t v1_0 = vzipq_u32(v2_0, v2_2);
52       const uint32x4x2_t v1_1 = vzipq_u32(v2_1, v2_3);
53 
54       const uint32x4x2_t v0_0 = vzipq_u32(v1_0.val[0], v1_1.val[0]);
55       const uint32x4x2_t v0_1 = vzipq_u32(v1_0.val[1], v1_1.val[1]);
56 
57       o = (uint32_t*) ((uintptr_t) o + oN_offset);
58       vst1q_u32(o, v0_1.val[1]);
59       if XNN_UNPREDICTABLE(block_width > 3) {
60         o = (uint32_t*) ((uintptr_t) o + minus_output_stride);
61       }
62       vst1q_u32(o, v0_1.val[0]);
63       if XNN_UNPREDICTABLE(block_width >= 3) {
64         o = (uint32_t*) ((uintptr_t) o + minus_output_stride);
65       }
66       vst1q_u32(o, v0_0.val[1]);
67       if XNN_UNPREDICTABLE(block_width > 1) {
68         o = (uint32_t*) ((uintptr_t) o + minus_output_stride);
69       }
70       vst1q_u32(o, v0_0.val[0]);
71     }
72     o = (uint32_t*) ((uintptr_t) o + tile_hbytes);
73 
74     if (bh != 0) {
75       const uint32x4_t v2_0 = vld1q_u32(i0);
76       const uint32_t *i1 = (const uint32_t*) ((uintptr_t) i0 + input_stride);
77       if XNN_UNPREDICTABLE(bh < 2) {
78         i1 = i0;
79       }
80       const uint32x4_t v2_1 = vld1q_u32(i1);
81       const uint32_t *i2 = (const uint32_t*) ((uintptr_t) i1 + input_stride);
82       if XNN_UNPREDICTABLE(bh <= 2) {
83         i2 = i1;
84       }
85       const uint32x4_t v2_2 = vld1q_u32(i2);
86       const uint32x4_t v2_3 = vmovq_n_u32(0);
87 
88       const uint32x4x2_t v1_0 = vzipq_u32(v2_0, v2_2);
89       const uint32x4x2_t v1_1 = vzipq_u32(v2_1, v2_3);
90 
91       const uint32x4x2_t v0_0 = vzipq_u32(v1_0.val[0], v1_1.val[0]);
92       const uint32x4x2_t v0_1 = vzipq_u32(v1_0.val[1], v1_1.val[1]);
93 
94       uint32x2_t v0_low = vget_low_u32(v0_0.val[0]);
95       uint32x2_t v1_low = vget_low_u32(v0_0.val[1]);
96       uint32x2_t v2_low = vget_low_u32(v0_1.val[0]);
97       uint32x2_t v3_low = vget_low_u32(v0_1.val[1]);
98 
99       if (bh & 2) {
100         o = (uint32_t*) ((uintptr_t) o + oN_stride);
101         vst1_u32(o, v3_low);
102         if XNN_UNPREDICTABLE(block_width > 3) {
103           o = (uint32_t*) ((uintptr_t) o + minus_output_stride);
104         }
105         vst1_u32(o, v2_low);
106         if XNN_UNPREDICTABLE(block_width >= 3) {
107           o = (uint32_t*) ((uintptr_t) o + minus_output_stride);
108         }
109         vst1_u32(o, v1_low);
110         if XNN_UNPREDICTABLE(block_width > 1) {
111           o = (uint32_t*) ((uintptr_t) o + minus_output_stride);
112         }
113         vst1_u32(o, v0_low); o += 2;
114         v0_low = vget_high_u32(v0_0.val[0]);
115         v1_low = vget_high_u32(v0_0.val[1]);
116         v2_low = vget_high_u32(v0_1.val[0]);
117         v3_low = vget_high_u32(v0_1.val[1]);
118       }
119 
120       if (bh & 1) {
121         o = (uint32_t*) ((uintptr_t) o + oN_stride);
122         vst1_lane_u32(o, v3_low, 0);
123         if XNN_UNPREDICTABLE(block_width > 3) {
124           o = (uint32_t*) ((uintptr_t) o + minus_output_stride);
125         }
126         vst1_lane_u32(o, v2_low, 0);
127         if XNN_UNPREDICTABLE(block_width >= 3) {
128           o = (uint32_t*) ((uintptr_t) o + minus_output_stride);
129         }
130         vst1_lane_u32(o, v1_low, 0);
131         if XNN_UNPREDICTABLE(block_width > 1) {
132           o = (uint32_t*) ((uintptr_t) o + minus_output_stride);
133         }
134         vst1_lane_u32(o, v0_low, 0);
135       }
136     }
137 
138     i0 = (const uint32_t*) ((uintptr_t) i0 + input_reset);
139     o = (uint32_t*) ((uintptr_t) o + output_reset);
140     block_width = doz(block_width, tile_width);
141   } while (block_width != 0);
142 }
143