xref: /aosp_15_r20/external/XNNPACK/src/xnnpack/transpose.h (revision 4bdc94577ba0e567308109d787f7fec7b531ce36)
1*4bdc9457SAndroid Build Coastguard Worker // Copyright 2021 Google LLC
2*4bdc9457SAndroid Build Coastguard Worker //
3*4bdc9457SAndroid Build Coastguard Worker // This source code is licensed under the BSD-style license found in the
4*4bdc9457SAndroid Build Coastguard Worker // LICENSE file in the root directory of this source tree.
5*4bdc9457SAndroid Build Coastguard Worker 
6*4bdc9457SAndroid Build Coastguard Worker #pragma once
7*4bdc9457SAndroid Build Coastguard Worker 
8*4bdc9457SAndroid Build Coastguard Worker #include <stddef.h>
9*4bdc9457SAndroid Build Coastguard Worker #include <stdint.h>
10*4bdc9457SAndroid Build Coastguard Worker 
11*4bdc9457SAndroid Build Coastguard Worker #include <xnnpack/common.h>
12*4bdc9457SAndroid Build Coastguard Worker 
13*4bdc9457SAndroid Build Coastguard Worker #ifdef __cplusplus
14*4bdc9457SAndroid Build Coastguard Worker extern "C" {
15*4bdc9457SAndroid Build Coastguard Worker #endif
16*4bdc9457SAndroid Build Coastguard Worker 
17*4bdc9457SAndroid Build Coastguard Worker #define DECLARE_XX_TRANSPOSEV_UKERNEL_FUNCTION(fn_name)   \
18*4bdc9457SAndroid Build Coastguard Worker   XNN_INTERNAL void fn_name(const void* input,            \
19*4bdc9457SAndroid Build Coastguard Worker                             void* output,                 \
20*4bdc9457SAndroid Build Coastguard Worker                             size_t input_row_stride,      \
21*4bdc9457SAndroid Build Coastguard Worker                             size_t output_row_stride,     \
22*4bdc9457SAndroid Build Coastguard Worker                             size_t input_element_stride,  \
23*4bdc9457SAndroid Build Coastguard Worker                             size_t output_element_stride, \
24*4bdc9457SAndroid Build Coastguard Worker                             size_t element_size,          \
25*4bdc9457SAndroid Build Coastguard Worker                             size_t block_width,           \
26*4bdc9457SAndroid Build Coastguard Worker                             size_t block_height);
27*4bdc9457SAndroid Build Coastguard Worker 
28*4bdc9457SAndroid Build Coastguard Worker DECLARE_XX_TRANSPOSEV_UKERNEL_FUNCTION(xnn_xx_transposev_ukernel__1x1_memcpy)
29*4bdc9457SAndroid Build Coastguard Worker 
30*4bdc9457SAndroid Build Coastguard Worker #define DECLARE_X64_TRANSPOSEC_UKERNEL_FUNCTION(fn_name) \
31*4bdc9457SAndroid Build Coastguard Worker   XNN_INTERNAL void fn_name(const uint64_t* input,      \
32*4bdc9457SAndroid Build Coastguard Worker                             uint64_t* output,           \
33*4bdc9457SAndroid Build Coastguard Worker                             size_t input_stride,        \
34*4bdc9457SAndroid Build Coastguard Worker                             size_t output_stride,       \
35*4bdc9457SAndroid Build Coastguard Worker                             size_t block_width,         \
36*4bdc9457SAndroid Build Coastguard Worker                             size_t block_height);
37*4bdc9457SAndroid Build Coastguard Worker 
38*4bdc9457SAndroid Build Coastguard Worker DECLARE_X64_TRANSPOSEC_UKERNEL_FUNCTION(xnn_x64_transposec_ukernel__1x2_scalar_float)
39*4bdc9457SAndroid Build Coastguard Worker DECLARE_X64_TRANSPOSEC_UKERNEL_FUNCTION(xnn_x64_transposec_ukernel__1x2_scalar_int)
40*4bdc9457SAndroid Build Coastguard Worker DECLARE_X64_TRANSPOSEC_UKERNEL_FUNCTION(xnn_x64_transposec_ukernel__2x1_scalar_float)
41*4bdc9457SAndroid Build Coastguard Worker DECLARE_X64_TRANSPOSEC_UKERNEL_FUNCTION(xnn_x64_transposec_ukernel__2x1_scalar_int)
42*4bdc9457SAndroid Build Coastguard Worker DECLARE_X64_TRANSPOSEC_UKERNEL_FUNCTION(xnn_x64_transposec_ukernel__2x2_multi_mov_sse2)
43*4bdc9457SAndroid Build Coastguard Worker DECLARE_X64_TRANSPOSEC_UKERNEL_FUNCTION(xnn_x64_transposec_ukernel__2x2_multi_multi_sse2)
44*4bdc9457SAndroid Build Coastguard Worker DECLARE_X64_TRANSPOSEC_UKERNEL_FUNCTION(xnn_x64_transposec_ukernel__2x2_multi_switch_sse2)
45*4bdc9457SAndroid Build Coastguard Worker DECLARE_X64_TRANSPOSEC_UKERNEL_FUNCTION(xnn_x64_transposec_ukernel__2x2_reuse_mov_sse2)
46*4bdc9457SAndroid Build Coastguard Worker DECLARE_X64_TRANSPOSEC_UKERNEL_FUNCTION(xnn_x64_transposec_ukernel__2x2_reuse_multi_sse2)
47*4bdc9457SAndroid Build Coastguard Worker DECLARE_X64_TRANSPOSEC_UKERNEL_FUNCTION(xnn_x64_transposec_ukernel__2x2_reuse_switch_sse2)
48*4bdc9457SAndroid Build Coastguard Worker DECLARE_X64_TRANSPOSEC_UKERNEL_FUNCTION(xnn_x64_transposec_ukernel__2x2_scalar_float)
49*4bdc9457SAndroid Build Coastguard Worker DECLARE_X64_TRANSPOSEC_UKERNEL_FUNCTION(xnn_x64_transposec_ukernel__2x2_scalar_int)
50*4bdc9457SAndroid Build Coastguard Worker DECLARE_X64_TRANSPOSEC_UKERNEL_FUNCTION(xnn_x64_transposec_ukernel__4x1_scalar_float)
51*4bdc9457SAndroid Build Coastguard Worker DECLARE_X64_TRANSPOSEC_UKERNEL_FUNCTION(xnn_x64_transposec_ukernel__4x1_scalar_int)
52*4bdc9457SAndroid Build Coastguard Worker DECLARE_X64_TRANSPOSEC_UKERNEL_FUNCTION(xnn_x64_transposec_ukernel__4x2_scalar_float)
53*4bdc9457SAndroid Build Coastguard Worker DECLARE_X64_TRANSPOSEC_UKERNEL_FUNCTION(xnn_x64_transposec_ukernel__4x2_scalar_int)
54*4bdc9457SAndroid Build Coastguard Worker 
55*4bdc9457SAndroid Build Coastguard Worker #define DECLARE_X32_TRANSPOSEC_UKERNEL_FUNCTION(fn_name) \
56*4bdc9457SAndroid Build Coastguard Worker   XNN_INTERNAL void fn_name(const uint32_t* input,      \
57*4bdc9457SAndroid Build Coastguard Worker                             uint32_t* output,           \
58*4bdc9457SAndroid Build Coastguard Worker                             size_t input_stride,        \
59*4bdc9457SAndroid Build Coastguard Worker                             size_t output_stride,       \
60*4bdc9457SAndroid Build Coastguard Worker                             size_t block_width,         \
61*4bdc9457SAndroid Build Coastguard Worker                             size_t block_height);
62*4bdc9457SAndroid Build Coastguard Worker 
63*4bdc9457SAndroid Build Coastguard Worker DECLARE_X32_TRANSPOSEC_UKERNEL_FUNCTION(xnn_x32_transposec_ukernel__1x2_scalar_float)
64*4bdc9457SAndroid Build Coastguard Worker DECLARE_X32_TRANSPOSEC_UKERNEL_FUNCTION(xnn_x32_transposec_ukernel__1x2_scalar_int)
65*4bdc9457SAndroid Build Coastguard Worker DECLARE_X32_TRANSPOSEC_UKERNEL_FUNCTION(xnn_x32_transposec_ukernel__1x4_scalar_float)
66*4bdc9457SAndroid Build Coastguard Worker DECLARE_X32_TRANSPOSEC_UKERNEL_FUNCTION(xnn_x32_transposec_ukernel__1x4_scalar_int)
67*4bdc9457SAndroid Build Coastguard Worker DECLARE_X32_TRANSPOSEC_UKERNEL_FUNCTION(xnn_x32_transposec_ukernel__2x1_scalar_float)
68*4bdc9457SAndroid Build Coastguard Worker DECLARE_X32_TRANSPOSEC_UKERNEL_FUNCTION(xnn_x32_transposec_ukernel__2x1_scalar_int)
69*4bdc9457SAndroid Build Coastguard Worker DECLARE_X32_TRANSPOSEC_UKERNEL_FUNCTION(xnn_x32_transposec_ukernel__2x2_multi_dec_zip_neon)
70*4bdc9457SAndroid Build Coastguard Worker DECLARE_X32_TRANSPOSEC_UKERNEL_FUNCTION(xnn_x32_transposec_ukernel__2x2_multi_mov_zip_neon)
71*4bdc9457SAndroid Build Coastguard Worker DECLARE_X32_TRANSPOSEC_UKERNEL_FUNCTION(xnn_x32_transposec_ukernel__2x2_multi_multi_zip_neon)
72*4bdc9457SAndroid Build Coastguard Worker DECLARE_X32_TRANSPOSEC_UKERNEL_FUNCTION(xnn_x32_transposec_ukernel__2x2_multi_switch_zip_neon)
73*4bdc9457SAndroid Build Coastguard Worker DECLARE_X32_TRANSPOSEC_UKERNEL_FUNCTION(xnn_x32_transposec_ukernel__2x2_reuse_dec_zip_neon)
74*4bdc9457SAndroid Build Coastguard Worker DECLARE_X32_TRANSPOSEC_UKERNEL_FUNCTION(xnn_x32_transposec_ukernel__2x2_reuse_mov_zip_neon)
75*4bdc9457SAndroid Build Coastguard Worker DECLARE_X32_TRANSPOSEC_UKERNEL_FUNCTION(xnn_x32_transposec_ukernel__2x2_reuse_multi_zip_neon)
76*4bdc9457SAndroid Build Coastguard Worker DECLARE_X32_TRANSPOSEC_UKERNEL_FUNCTION(xnn_x32_transposec_ukernel__2x2_reuse_switch_zip_neon)
77*4bdc9457SAndroid Build Coastguard Worker DECLARE_X32_TRANSPOSEC_UKERNEL_FUNCTION(xnn_x32_transposec_ukernel__2x2_scalar_float)
78*4bdc9457SAndroid Build Coastguard Worker DECLARE_X32_TRANSPOSEC_UKERNEL_FUNCTION(xnn_x32_transposec_ukernel__2x2_scalar_int)
79*4bdc9457SAndroid Build Coastguard Worker DECLARE_X32_TRANSPOSEC_UKERNEL_FUNCTION(xnn_x32_transposec_ukernel__2x4_scalar_float)
80*4bdc9457SAndroid Build Coastguard Worker DECLARE_X32_TRANSPOSEC_UKERNEL_FUNCTION(xnn_x32_transposec_ukernel__2x4_scalar_int)
81*4bdc9457SAndroid Build Coastguard Worker DECLARE_X32_TRANSPOSEC_UKERNEL_FUNCTION(xnn_x32_transposec_ukernel__4x1_scalar_float)
82*4bdc9457SAndroid Build Coastguard Worker DECLARE_X32_TRANSPOSEC_UKERNEL_FUNCTION(xnn_x32_transposec_ukernel__4x1_scalar_int)
83*4bdc9457SAndroid Build Coastguard Worker DECLARE_X32_TRANSPOSEC_UKERNEL_FUNCTION(xnn_x32_transposec_ukernel__4x2_scalar_float)
84*4bdc9457SAndroid Build Coastguard Worker DECLARE_X32_TRANSPOSEC_UKERNEL_FUNCTION(xnn_x32_transposec_ukernel__4x2_scalar_int)
85*4bdc9457SAndroid Build Coastguard Worker DECLARE_X32_TRANSPOSEC_UKERNEL_FUNCTION(xnn_x32_transposec_ukernel__4x4_aarch64_neon_tbl)
86*4bdc9457SAndroid Build Coastguard Worker DECLARE_X32_TRANSPOSEC_UKERNEL_FUNCTION(xnn_x32_transposec_ukernel__4x4_multi_dec_zip_neon)
87*4bdc9457SAndroid Build Coastguard Worker DECLARE_X32_TRANSPOSEC_UKERNEL_FUNCTION(xnn_x32_transposec_ukernel__4x4_multi_mov_sse2)
88*4bdc9457SAndroid Build Coastguard Worker DECLARE_X32_TRANSPOSEC_UKERNEL_FUNCTION(xnn_x32_transposec_ukernel__4x4_multi_mov_wasmsimd)
89*4bdc9457SAndroid Build Coastguard Worker DECLARE_X32_TRANSPOSEC_UKERNEL_FUNCTION(xnn_x32_transposec_ukernel__4x4_multi_mov_zip_neon)
90*4bdc9457SAndroid Build Coastguard Worker DECLARE_X32_TRANSPOSEC_UKERNEL_FUNCTION(xnn_x32_transposec_ukernel__4x4_multi_multi_sse2)
91*4bdc9457SAndroid Build Coastguard Worker DECLARE_X32_TRANSPOSEC_UKERNEL_FUNCTION(xnn_x32_transposec_ukernel__4x4_multi_multi_wasmsimd)
92*4bdc9457SAndroid Build Coastguard Worker DECLARE_X32_TRANSPOSEC_UKERNEL_FUNCTION(xnn_x32_transposec_ukernel__4x4_multi_multi_zip_neon)
93*4bdc9457SAndroid Build Coastguard Worker DECLARE_X32_TRANSPOSEC_UKERNEL_FUNCTION(xnn_x32_transposec_ukernel__4x4_multi_switch_sse2)
94*4bdc9457SAndroid Build Coastguard Worker DECLARE_X32_TRANSPOSEC_UKERNEL_FUNCTION(xnn_x32_transposec_ukernel__4x4_multi_switch_wasmsimd)
95*4bdc9457SAndroid Build Coastguard Worker DECLARE_X32_TRANSPOSEC_UKERNEL_FUNCTION(xnn_x32_transposec_ukernel__4x4_multi_switch_zip_neon)
96*4bdc9457SAndroid Build Coastguard Worker DECLARE_X32_TRANSPOSEC_UKERNEL_FUNCTION(xnn_x32_transposec_ukernel__4x4_reuse_dec_zip_neon)
97*4bdc9457SAndroid Build Coastguard Worker DECLARE_X32_TRANSPOSEC_UKERNEL_FUNCTION(xnn_x32_transposec_ukernel__4x4_reuse_mov_sse2)
98*4bdc9457SAndroid Build Coastguard Worker DECLARE_X32_TRANSPOSEC_UKERNEL_FUNCTION(xnn_x32_transposec_ukernel__4x4_reuse_mov_wasmsimd)
99*4bdc9457SAndroid Build Coastguard Worker DECLARE_X32_TRANSPOSEC_UKERNEL_FUNCTION(xnn_x32_transposec_ukernel__4x4_reuse_mov_zip_neon)
100*4bdc9457SAndroid Build Coastguard Worker DECLARE_X32_TRANSPOSEC_UKERNEL_FUNCTION(xnn_x32_transposec_ukernel__4x4_reuse_multi_sse2)
101*4bdc9457SAndroid Build Coastguard Worker DECLARE_X32_TRANSPOSEC_UKERNEL_FUNCTION(xnn_x32_transposec_ukernel__4x4_reuse_multi_wasmsimd)
102*4bdc9457SAndroid Build Coastguard Worker DECLARE_X32_TRANSPOSEC_UKERNEL_FUNCTION(xnn_x32_transposec_ukernel__4x4_reuse_multi_zip_neon)
103*4bdc9457SAndroid Build Coastguard Worker DECLARE_X32_TRANSPOSEC_UKERNEL_FUNCTION(xnn_x32_transposec_ukernel__4x4_reuse_switch_sse2)
104*4bdc9457SAndroid Build Coastguard Worker DECLARE_X32_TRANSPOSEC_UKERNEL_FUNCTION(xnn_x32_transposec_ukernel__4x4_reuse_switch_wasmsimd)
105*4bdc9457SAndroid Build Coastguard Worker DECLARE_X32_TRANSPOSEC_UKERNEL_FUNCTION(xnn_x32_transposec_ukernel__4x4_reuse_switch_zip_neon)
106*4bdc9457SAndroid Build Coastguard Worker DECLARE_X32_TRANSPOSEC_UKERNEL_FUNCTION(xnn_x32_transposec_ukernel__4x4_scalar_float)
107*4bdc9457SAndroid Build Coastguard Worker DECLARE_X32_TRANSPOSEC_UKERNEL_FUNCTION(xnn_x32_transposec_ukernel__4x4_scalar_int)
108*4bdc9457SAndroid Build Coastguard Worker DECLARE_X32_TRANSPOSEC_UKERNEL_FUNCTION(xnn_x32_transposec_ukernel__4x4_sse)
109*4bdc9457SAndroid Build Coastguard Worker 
110*4bdc9457SAndroid Build Coastguard Worker #define DECLARE_X24_TRANSPOSEC_UKERNEL_FUNCTION(fn_name) \
111*4bdc9457SAndroid Build Coastguard Worker   XNN_INTERNAL void fn_name(const void* input,      \
112*4bdc9457SAndroid Build Coastguard Worker                             void* output,           \
113*4bdc9457SAndroid Build Coastguard Worker                             size_t input_stride,        \
114*4bdc9457SAndroid Build Coastguard Worker                             size_t output_stride,       \
115*4bdc9457SAndroid Build Coastguard Worker                             size_t block_width,         \
116*4bdc9457SAndroid Build Coastguard Worker                             size_t block_height);
117*4bdc9457SAndroid Build Coastguard Worker 
118*4bdc9457SAndroid Build Coastguard Worker DECLARE_X24_TRANSPOSEC_UKERNEL_FUNCTION(xnn_x24_transposec_ukernel__1x2_scalar)
119*4bdc9457SAndroid Build Coastguard Worker DECLARE_X24_TRANSPOSEC_UKERNEL_FUNCTION(xnn_x24_transposec_ukernel__1x4_scalar)
120*4bdc9457SAndroid Build Coastguard Worker DECLARE_X24_TRANSPOSEC_UKERNEL_FUNCTION(xnn_x24_transposec_ukernel__2x1_scalar)
121*4bdc9457SAndroid Build Coastguard Worker DECLARE_X24_TRANSPOSEC_UKERNEL_FUNCTION(xnn_x24_transposec_ukernel__2x2_neon_tbl)
122*4bdc9457SAndroid Build Coastguard Worker DECLARE_X24_TRANSPOSEC_UKERNEL_FUNCTION(xnn_x24_transposec_ukernel__2x2_scalar)
123*4bdc9457SAndroid Build Coastguard Worker DECLARE_X24_TRANSPOSEC_UKERNEL_FUNCTION(xnn_x24_transposec_ukernel__2x4_scalar)
124*4bdc9457SAndroid Build Coastguard Worker DECLARE_X24_TRANSPOSEC_UKERNEL_FUNCTION(xnn_x24_transposec_ukernel__4x1_scalar)
125*4bdc9457SAndroid Build Coastguard Worker DECLARE_X24_TRANSPOSEC_UKERNEL_FUNCTION(xnn_x24_transposec_ukernel__4x2_scalar)
126*4bdc9457SAndroid Build Coastguard Worker DECLARE_X24_TRANSPOSEC_UKERNEL_FUNCTION(xnn_x24_transposec_ukernel__4x4_aarch64_neon_tbl)
127*4bdc9457SAndroid Build Coastguard Worker DECLARE_X24_TRANSPOSEC_UKERNEL_FUNCTION(xnn_x24_transposec_ukernel__4x4_scalar)
128*4bdc9457SAndroid Build Coastguard Worker DECLARE_X24_TRANSPOSEC_UKERNEL_FUNCTION(xnn_x24_transposec_ukernel__4x4_ssse3)
129*4bdc9457SAndroid Build Coastguard Worker 
130*4bdc9457SAndroid Build Coastguard Worker #define DECLARE_X16_TRANSPOSEC_UKERNEL_FUNCTION(fn_name) \
131*4bdc9457SAndroid Build Coastguard Worker   XNN_INTERNAL void fn_name(const uint16_t* input,      \
132*4bdc9457SAndroid Build Coastguard Worker                             uint16_t* output,           \
133*4bdc9457SAndroid Build Coastguard Worker                             size_t input_stride,        \
134*4bdc9457SAndroid Build Coastguard Worker                             size_t output_stride,       \
135*4bdc9457SAndroid Build Coastguard Worker                             size_t block_width,         \
136*4bdc9457SAndroid Build Coastguard Worker                             size_t block_height);
137*4bdc9457SAndroid Build Coastguard Worker 
138*4bdc9457SAndroid Build Coastguard Worker DECLARE_X16_TRANSPOSEC_UKERNEL_FUNCTION(xnn_x16_transposec_ukernel__1x2_scalar_int)
139*4bdc9457SAndroid Build Coastguard Worker DECLARE_X16_TRANSPOSEC_UKERNEL_FUNCTION(xnn_x16_transposec_ukernel__1x4_scalar_int)
140*4bdc9457SAndroid Build Coastguard Worker DECLARE_X16_TRANSPOSEC_UKERNEL_FUNCTION(xnn_x16_transposec_ukernel__2x1_scalar_int)
141*4bdc9457SAndroid Build Coastguard Worker DECLARE_X16_TRANSPOSEC_UKERNEL_FUNCTION(xnn_x16_transposec_ukernel__2x2_scalar_int)
142*4bdc9457SAndroid Build Coastguard Worker DECLARE_X16_TRANSPOSEC_UKERNEL_FUNCTION(xnn_x16_transposec_ukernel__2x4_scalar_int)
143*4bdc9457SAndroid Build Coastguard Worker DECLARE_X16_TRANSPOSEC_UKERNEL_FUNCTION(xnn_x16_transposec_ukernel__4x1_scalar_int)
144*4bdc9457SAndroid Build Coastguard Worker DECLARE_X16_TRANSPOSEC_UKERNEL_FUNCTION(xnn_x16_transposec_ukernel__4x2_scalar_int)
145*4bdc9457SAndroid Build Coastguard Worker DECLARE_X16_TRANSPOSEC_UKERNEL_FUNCTION(xnn_x16_transposec_ukernel__4x4_multi_dec_zip_neon)
146*4bdc9457SAndroid Build Coastguard Worker DECLARE_X16_TRANSPOSEC_UKERNEL_FUNCTION(xnn_x16_transposec_ukernel__4x4_multi_mov_zip_neon)
147*4bdc9457SAndroid Build Coastguard Worker DECLARE_X16_TRANSPOSEC_UKERNEL_FUNCTION(xnn_x16_transposec_ukernel__4x4_multi_multi_zip_neon)
148*4bdc9457SAndroid Build Coastguard Worker DECLARE_X16_TRANSPOSEC_UKERNEL_FUNCTION(xnn_x16_transposec_ukernel__4x4_multi_switch_zip_neon)
149*4bdc9457SAndroid Build Coastguard Worker DECLARE_X16_TRANSPOSEC_UKERNEL_FUNCTION(xnn_x16_transposec_ukernel__4x4_reuse_dec_zip_neon)
150*4bdc9457SAndroid Build Coastguard Worker DECLARE_X16_TRANSPOSEC_UKERNEL_FUNCTION(xnn_x16_transposec_ukernel__4x4_reuse_mov_zip_neon)
151*4bdc9457SAndroid Build Coastguard Worker DECLARE_X16_TRANSPOSEC_UKERNEL_FUNCTION(xnn_x16_transposec_ukernel__4x4_reuse_multi_zip_neon)
152*4bdc9457SAndroid Build Coastguard Worker DECLARE_X16_TRANSPOSEC_UKERNEL_FUNCTION(xnn_x16_transposec_ukernel__4x4_reuse_switch_zip_neon)
153*4bdc9457SAndroid Build Coastguard Worker DECLARE_X16_TRANSPOSEC_UKERNEL_FUNCTION(xnn_x16_transposec_ukernel__4x4_scalar_int)
154*4bdc9457SAndroid Build Coastguard Worker DECLARE_X16_TRANSPOSEC_UKERNEL_FUNCTION(xnn_x16_transposec_ukernel__4x8_sse2)
155*4bdc9457SAndroid Build Coastguard Worker DECLARE_X16_TRANSPOSEC_UKERNEL_FUNCTION(xnn_x16_transposec_ukernel__8x8_multi_dec_zip_neon)
156*4bdc9457SAndroid Build Coastguard Worker DECLARE_X16_TRANSPOSEC_UKERNEL_FUNCTION(xnn_x16_transposec_ukernel__8x8_multi_mov_sse2)
157*4bdc9457SAndroid Build Coastguard Worker DECLARE_X16_TRANSPOSEC_UKERNEL_FUNCTION(xnn_x16_transposec_ukernel__8x8_multi_mov_wasmsimd)
158*4bdc9457SAndroid Build Coastguard Worker DECLARE_X16_TRANSPOSEC_UKERNEL_FUNCTION(xnn_x16_transposec_ukernel__8x8_multi_mov_zip_neon)
159*4bdc9457SAndroid Build Coastguard Worker DECLARE_X16_TRANSPOSEC_UKERNEL_FUNCTION(xnn_x16_transposec_ukernel__8x8_multi_switch_sse2)
160*4bdc9457SAndroid Build Coastguard Worker DECLARE_X16_TRANSPOSEC_UKERNEL_FUNCTION(xnn_x16_transposec_ukernel__8x8_multi_switch_wasmsimd)
161*4bdc9457SAndroid Build Coastguard Worker DECLARE_X16_TRANSPOSEC_UKERNEL_FUNCTION(xnn_x16_transposec_ukernel__8x8_multi_switch_zip_neon)
162*4bdc9457SAndroid Build Coastguard Worker DECLARE_X16_TRANSPOSEC_UKERNEL_FUNCTION(xnn_x16_transposec_ukernel__8x8_reuse_dec_zip_neon)
163*4bdc9457SAndroid Build Coastguard Worker DECLARE_X16_TRANSPOSEC_UKERNEL_FUNCTION(xnn_x16_transposec_ukernel__8x8_reuse_mov_sse2)
164*4bdc9457SAndroid Build Coastguard Worker DECLARE_X16_TRANSPOSEC_UKERNEL_FUNCTION(xnn_x16_transposec_ukernel__8x8_reuse_mov_wasmsimd)
165*4bdc9457SAndroid Build Coastguard Worker DECLARE_X16_TRANSPOSEC_UKERNEL_FUNCTION(xnn_x16_transposec_ukernel__8x8_reuse_mov_zip_neon)
166*4bdc9457SAndroid Build Coastguard Worker DECLARE_X16_TRANSPOSEC_UKERNEL_FUNCTION(xnn_x16_transposec_ukernel__8x8_reuse_multi_sse2)
167*4bdc9457SAndroid Build Coastguard Worker DECLARE_X16_TRANSPOSEC_UKERNEL_FUNCTION(xnn_x16_transposec_ukernel__8x8_reuse_multi_wasmsimd)
168*4bdc9457SAndroid Build Coastguard Worker DECLARE_X16_TRANSPOSEC_UKERNEL_FUNCTION(xnn_x16_transposec_ukernel__8x8_reuse_multi_zip_neon)
169*4bdc9457SAndroid Build Coastguard Worker DECLARE_X16_TRANSPOSEC_UKERNEL_FUNCTION(xnn_x16_transposec_ukernel__8x8_reuse_switch_sse2)
170*4bdc9457SAndroid Build Coastguard Worker DECLARE_X16_TRANSPOSEC_UKERNEL_FUNCTION(xnn_x16_transposec_ukernel__8x8_reuse_switch_wasmsimd)
171*4bdc9457SAndroid Build Coastguard Worker DECLARE_X16_TRANSPOSEC_UKERNEL_FUNCTION(xnn_x16_transposec_ukernel__8x8_reuse_switch_zip_neon)
172*4bdc9457SAndroid Build Coastguard Worker 
173*4bdc9457SAndroid Build Coastguard Worker #define DECLARE_X8_TRANSPOSEC_UKERNEL_FUNCTION(fn_name)  \
174*4bdc9457SAndroid Build Coastguard Worker   XNN_INTERNAL void fn_name(const uint8_t* input,       \
175*4bdc9457SAndroid Build Coastguard Worker                             uint8_t* output,            \
176*4bdc9457SAndroid Build Coastguard Worker                             size_t input_stride,        \
177*4bdc9457SAndroid Build Coastguard Worker                             size_t output_stride,       \
178*4bdc9457SAndroid Build Coastguard Worker                             size_t block_width,         \
179*4bdc9457SAndroid Build Coastguard Worker                             size_t block_height);
180*4bdc9457SAndroid Build Coastguard Worker 
181*4bdc9457SAndroid Build Coastguard Worker DECLARE_X8_TRANSPOSEC_UKERNEL_FUNCTION(xnn_x8_transposec_ukernel__1x2_scalar_int)
182*4bdc9457SAndroid Build Coastguard Worker DECLARE_X8_TRANSPOSEC_UKERNEL_FUNCTION(xnn_x8_transposec_ukernel__1x4_scalar_int)
183*4bdc9457SAndroid Build Coastguard Worker DECLARE_X8_TRANSPOSEC_UKERNEL_FUNCTION(xnn_x8_transposec_ukernel__2x1_scalar_int)
184*4bdc9457SAndroid Build Coastguard Worker DECLARE_X8_TRANSPOSEC_UKERNEL_FUNCTION(xnn_x8_transposec_ukernel__2x2_scalar_int)
185*4bdc9457SAndroid Build Coastguard Worker DECLARE_X8_TRANSPOSEC_UKERNEL_FUNCTION(xnn_x8_transposec_ukernel__2x4_scalar_int)
186*4bdc9457SAndroid Build Coastguard Worker DECLARE_X8_TRANSPOSEC_UKERNEL_FUNCTION(xnn_x8_transposec_ukernel__4x1_scalar_int)
187*4bdc9457SAndroid Build Coastguard Worker DECLARE_X8_TRANSPOSEC_UKERNEL_FUNCTION(xnn_x8_transposec_ukernel__4x2_scalar_int)
188*4bdc9457SAndroid Build Coastguard Worker DECLARE_X8_TRANSPOSEC_UKERNEL_FUNCTION(xnn_x8_transposec_ukernel__4x4_scalar_int)
189*4bdc9457SAndroid Build Coastguard Worker DECLARE_X8_TRANSPOSEC_UKERNEL_FUNCTION(xnn_x8_transposec_ukernel__8x8_multi_dec_zip_neon)
190*4bdc9457SAndroid Build Coastguard Worker DECLARE_X8_TRANSPOSEC_UKERNEL_FUNCTION(xnn_x8_transposec_ukernel__8x8_multi_mov_zip_neon)
191*4bdc9457SAndroid Build Coastguard Worker DECLARE_X8_TRANSPOSEC_UKERNEL_FUNCTION(xnn_x8_transposec_ukernel__8x8_multi_switch_zip_neon)
192*4bdc9457SAndroid Build Coastguard Worker DECLARE_X8_TRANSPOSEC_UKERNEL_FUNCTION(xnn_x8_transposec_ukernel__8x8_reuse_dec_zip_neon)
193*4bdc9457SAndroid Build Coastguard Worker DECLARE_X8_TRANSPOSEC_UKERNEL_FUNCTION(xnn_x8_transposec_ukernel__8x8_reuse_mov_zip_neon)
194*4bdc9457SAndroid Build Coastguard Worker DECLARE_X8_TRANSPOSEC_UKERNEL_FUNCTION(xnn_x8_transposec_ukernel__8x8_reuse_multi_zip_neon)
195*4bdc9457SAndroid Build Coastguard Worker DECLARE_X8_TRANSPOSEC_UKERNEL_FUNCTION(xnn_x8_transposec_ukernel__8x8_reuse_switch_zip_neon)
196*4bdc9457SAndroid Build Coastguard Worker DECLARE_X8_TRANSPOSEC_UKERNEL_FUNCTION(xnn_x8_transposec_ukernel__16x16_reuse_dec_zip_neon)
197*4bdc9457SAndroid Build Coastguard Worker DECLARE_X8_TRANSPOSEC_UKERNEL_FUNCTION(xnn_x8_transposec_ukernel__16x16_reuse_mov_sse2)
198*4bdc9457SAndroid Build Coastguard Worker DECLARE_X8_TRANSPOSEC_UKERNEL_FUNCTION(xnn_x8_transposec_ukernel__16x16_reuse_mov_wasmsimd)
199*4bdc9457SAndroid Build Coastguard Worker DECLARE_X8_TRANSPOSEC_UKERNEL_FUNCTION(xnn_x8_transposec_ukernel__16x16_reuse_mov_zip_neon)
200*4bdc9457SAndroid Build Coastguard Worker DECLARE_X8_TRANSPOSEC_UKERNEL_FUNCTION(xnn_x8_transposec_ukernel__16x16_reuse_switch_sse2)
201*4bdc9457SAndroid Build Coastguard Worker DECLARE_X8_TRANSPOSEC_UKERNEL_FUNCTION(xnn_x8_transposec_ukernel__16x16_reuse_switch_wasmsimd)
202*4bdc9457SAndroid Build Coastguard Worker DECLARE_X8_TRANSPOSEC_UKERNEL_FUNCTION(xnn_x8_transposec_ukernel__16x16_reuse_switch_zip_neon)
203*4bdc9457SAndroid Build Coastguard Worker 
204*4bdc9457SAndroid Build Coastguard Worker #ifdef __cplusplus
205*4bdc9457SAndroid Build Coastguard Worker }  // extern "C"
206*4bdc9457SAndroid Build Coastguard Worker #endif
207