blob: 440a7564c4043657aaf9dfc3fab932b3d963e24f [file] [log] [blame]
XNNPACK Teamb455b122019-09-27 18:10:33 -07001// Copyright (c) Facebook, Inc. and its affiliates.
2// All rights reserved.
3//
4// Copyright 2019 Google LLC
5//
6// This source code is licensed under the BSD-style license found in the
7// LICENSE file in the root directory of this source tree.
8
9#pragma once
10
11#include <stddef.h>
12#include <stdint.h>
13
14#include <xnnpack/params.h>
15#include <xnnpack/common.h>
16
17#ifdef __cplusplus
18extern "C" {
19#endif
20
21#define DECLARE_F32_DWCONV_UNIPASS_UKERNEL_FUNCTION(fn_name) \
22 XNN_INTERNAL void fn_name( \
23 size_t channels, \
24 size_t output_width, \
25 const float** input, \
26 const float* weights, \
27 float* output, \
28 size_t input_stride, \
29 size_t output_increment, \
30 const union xnn_f32_output_params* params);
31
XNNPACK Teamb455b122019-09-27 18:10:33 -070032DECLARE_F32_DWCONV_UNIPASS_UKERNEL_FUNCTION(xnn_f32_dwconv_ukernel_up1x4__scalar)
Marat Dukhan5098c3e2019-11-07 12:01:19 -080033DECLARE_F32_DWCONV_UNIPASS_UKERNEL_FUNCTION(xnn_f32_dwconv_ukernel_up1x4__scalar_acc2)
34DECLARE_F32_DWCONV_UNIPASS_UKERNEL_FUNCTION(xnn_f32_dwconv_ukernel_up2x4__scalar)
35DECLARE_F32_DWCONV_UNIPASS_UKERNEL_FUNCTION(xnn_f32_dwconv_ukernel_up2x4__scalar_acc2)
XNNPACK Teamb455b122019-09-27 18:10:33 -070036DECLARE_F32_DWCONV_UNIPASS_UKERNEL_FUNCTION(xnn_f32_dwconv_ukernel_up4x4__psimd)
Marat Dukhan5098c3e2019-11-07 12:01:19 -080037DECLARE_F32_DWCONV_UNIPASS_UKERNEL_FUNCTION(xnn_f32_dwconv_ukernel_up4x4__psimd_acc2)
38DECLARE_F32_DWCONV_UNIPASS_UKERNEL_FUNCTION(xnn_f32_dwconv_ukernel_up8x4__psimd)
39DECLARE_F32_DWCONV_UNIPASS_UKERNEL_FUNCTION(xnn_f32_dwconv_ukernel_up8x4__psimd_acc2)
XNNPACK Teamb455b122019-09-27 18:10:33 -070040DECLARE_F32_DWCONV_UNIPASS_UKERNEL_FUNCTION(xnn_f32_dwconv_ukernel_up4x4__sse)
Marat Dukhan5098c3e2019-11-07 12:01:19 -080041DECLARE_F32_DWCONV_UNIPASS_UKERNEL_FUNCTION(xnn_f32_dwconv_ukernel_up4x4__sse_acc2)
42DECLARE_F32_DWCONV_UNIPASS_UKERNEL_FUNCTION(xnn_f32_dwconv_ukernel_up8x4__sse)
43DECLARE_F32_DWCONV_UNIPASS_UKERNEL_FUNCTION(xnn_f32_dwconv_ukernel_up8x4__sse_acc2)
44
45DECLARE_F32_DWCONV_UNIPASS_UKERNEL_FUNCTION(xnn_f32_dwconv_ukernel_up1x9__scalar)
46DECLARE_F32_DWCONV_UNIPASS_UKERNEL_FUNCTION(xnn_f32_dwconv_ukernel_up1x9__scalar_acc2)
47DECLARE_F32_DWCONV_UNIPASS_UKERNEL_FUNCTION(xnn_f32_dwconv_ukernel_up2x9__scalar)
48DECLARE_F32_DWCONV_UNIPASS_UKERNEL_FUNCTION(xnn_f32_dwconv_ukernel_up2x9__scalar_acc2)
49DECLARE_F32_DWCONV_UNIPASS_UKERNEL_FUNCTION(xnn_f32_dwconv_ukernel_up4x9__neon)
50DECLARE_F32_DWCONV_UNIPASS_UKERNEL_FUNCTION(xnn_f32_dwconv_ukernel_up4x9__neon_acc2)
51DECLARE_F32_DWCONV_UNIPASS_UKERNEL_FUNCTION(xnn_f32_dwconv_ukernel_up8x9__neon)
52DECLARE_F32_DWCONV_UNIPASS_UKERNEL_FUNCTION(xnn_f32_dwconv_ukernel_up8x9__neon_acc2)
53DECLARE_F32_DWCONV_UNIPASS_UKERNEL_FUNCTION(xnn_f32_dwconv_ukernel_up4x9__neonfma)
54DECLARE_F32_DWCONV_UNIPASS_UKERNEL_FUNCTION(xnn_f32_dwconv_ukernel_up4x9__neonfma_acc2)
55DECLARE_F32_DWCONV_UNIPASS_UKERNEL_FUNCTION(xnn_f32_dwconv_ukernel_up8x9__neonfma)
56DECLARE_F32_DWCONV_UNIPASS_UKERNEL_FUNCTION(xnn_f32_dwconv_ukernel_up8x9__neonfma_acc2)
XNNPACK Teamb455b122019-09-27 18:10:33 -070057DECLARE_F32_DWCONV_UNIPASS_UKERNEL_FUNCTION(xnn_f32_dwconv_ukernel_up4x9__aarch64_neonfma)
58DECLARE_F32_DWCONV_UNIPASS_UKERNEL_FUNCTION(xnn_f32_dwconv_ukernel_up4x9__aarch64_neonfma_cortex_a55)
XNNPACK Teamb455b122019-09-27 18:10:33 -070059DECLARE_F32_DWCONV_UNIPASS_UKERNEL_FUNCTION(xnn_f32_dwconv_ukernel_up4x9__psimd)
Marat Dukhan5098c3e2019-11-07 12:01:19 -080060DECLARE_F32_DWCONV_UNIPASS_UKERNEL_FUNCTION(xnn_f32_dwconv_ukernel_up4x9__psimd_acc2)
61DECLARE_F32_DWCONV_UNIPASS_UKERNEL_FUNCTION(xnn_f32_dwconv_ukernel_up8x9__psimd)
62DECLARE_F32_DWCONV_UNIPASS_UKERNEL_FUNCTION(xnn_f32_dwconv_ukernel_up8x9__psimd_acc2)
XNNPACK Teamb455b122019-09-27 18:10:33 -070063DECLARE_F32_DWCONV_UNIPASS_UKERNEL_FUNCTION(xnn_f32_dwconv_ukernel_up4x9__sse)
Marat Dukhan5098c3e2019-11-07 12:01:19 -080064DECLARE_F32_DWCONV_UNIPASS_UKERNEL_FUNCTION(xnn_f32_dwconv_ukernel_up4x9__sse_acc2)
65DECLARE_F32_DWCONV_UNIPASS_UKERNEL_FUNCTION(xnn_f32_dwconv_ukernel_up8x9__sse)
66DECLARE_F32_DWCONV_UNIPASS_UKERNEL_FUNCTION(xnn_f32_dwconv_ukernel_up8x9__sse_acc2)
67
68DECLARE_F32_DWCONV_UNIPASS_UKERNEL_FUNCTION(xnn_f32_dwconv_ukernel_up1x25__scalar)
69DECLARE_F32_DWCONV_UNIPASS_UKERNEL_FUNCTION(xnn_f32_dwconv_ukernel_up1x25__scalar_acc2)
70DECLARE_F32_DWCONV_UNIPASS_UKERNEL_FUNCTION(xnn_f32_dwconv_ukernel_up2x25__scalar)
71DECLARE_F32_DWCONV_UNIPASS_UKERNEL_FUNCTION(xnn_f32_dwconv_ukernel_up2x25__scalar_acc2)
72DECLARE_F32_DWCONV_UNIPASS_UKERNEL_FUNCTION(xnn_f32_dwconv_ukernel_up4x25__psimd)
73DECLARE_F32_DWCONV_UNIPASS_UKERNEL_FUNCTION(xnn_f32_dwconv_ukernel_up4x25__psimd_acc2)
74DECLARE_F32_DWCONV_UNIPASS_UKERNEL_FUNCTION(xnn_f32_dwconv_ukernel_up8x25__psimd)
75DECLARE_F32_DWCONV_UNIPASS_UKERNEL_FUNCTION(xnn_f32_dwconv_ukernel_up8x25__psimd_acc2)
76DECLARE_F32_DWCONV_UNIPASS_UKERNEL_FUNCTION(xnn_f32_dwconv_ukernel_up4x25__sse)
77DECLARE_F32_DWCONV_UNIPASS_UKERNEL_FUNCTION(xnn_f32_dwconv_ukernel_up4x25__sse_acc2)
78DECLARE_F32_DWCONV_UNIPASS_UKERNEL_FUNCTION(xnn_f32_dwconv_ukernel_up8x25__sse)
79DECLARE_F32_DWCONV_UNIPASS_UKERNEL_FUNCTION(xnn_f32_dwconv_ukernel_up8x25__sse_acc2)
XNNPACK Teamb455b122019-09-27 18:10:33 -070080
81
82#define DECLARE_Q8_DWCONV_UNIPASS_UKERNEL_FUNCTION(fn_name) \
83 XNN_INTERNAL void fn_name( \
84 size_t channels, \
85 size_t output_width, \
86 const uint8_t** input, \
87 const void* weights, \
88 uint8_t* output, \
89 size_t input_stride, \
90 size_t output_increment, \
91 const union xnn_q8_gemm_params* params);
92
93DECLARE_Q8_DWCONV_UNIPASS_UKERNEL_FUNCTION(xnn_q8_dwconv_ukernel_up1x9__scalar)
94DECLARE_Q8_DWCONV_UNIPASS_UKERNEL_FUNCTION(xnn_q8_dwconv_ukernel_up8x9__aarch32_neon)
95DECLARE_Q8_DWCONV_UNIPASS_UKERNEL_FUNCTION(xnn_q8_dwconv_ukernel_up8x9__neon)
96DECLARE_Q8_DWCONV_UNIPASS_UKERNEL_FUNCTION(xnn_q8_dwconv_ukernel_up8x9__sse2)
97
98
99#define DECLARE_F32_DWCONV_SPCHW_UKERNEL_FUNCTION(fn_name) \
100 XNN_INTERNAL void fn_name( \
101 size_t m, \
102 size_t n, \
103 const float* input, \
104 const float* weights, \
105 float* output, \
106 size_t input_tuple_stride, \
107 size_t output_tuple_stride, \
108 size_t input_height_stride, \
109 size_t output_height_stride, \
110 const union xnn_f32_spchw_params* params);
111
Erich Elsen0cc2c532019-10-15 04:44:18 -0700112DECLARE_F32_DWCONV_SPCHW_UKERNEL_FUNCTION(xnn_f32_dwconv_spchw_ukernel_3x3p1__scalar)
Erich Elsenac4de802019-10-16 04:35:30 -0700113DECLARE_F32_DWCONV_SPCHW_UKERNEL_FUNCTION(xnn_f32_dwconv_spchw_ukernel_3x3s2p1__scalar)
XNNPACK Teamb455b122019-09-27 18:10:33 -0700114DECLARE_F32_DWCONV_SPCHW_UKERNEL_FUNCTION(xnn_f32_dwconv_spchw_ukernel_3x3p1__neonfma)
115DECLARE_F32_DWCONV_SPCHW_UKERNEL_FUNCTION(xnn_f32_dwconv_spchw_ukernel_5x5p2__neonfma)
116DECLARE_F32_DWCONV_SPCHW_UKERNEL_FUNCTION(xnn_f32_dwconv_spchw_ukernel_3x3p1__sse)
117DECLARE_F32_DWCONV_SPCHW_UKERNEL_FUNCTION(xnn_f32_dwconv_spchw_ukernel_3x3s2p1__neonfma)
118DECLARE_F32_DWCONV_SPCHW_UKERNEL_FUNCTION(xnn_f32_dwconv_spchw_ukernel_5x5s2p2__neonfma)
119DECLARE_F32_DWCONV_SPCHW_UKERNEL_FUNCTION(xnn_f32_dwconv_spchw_ukernel_3x3s2p1__sse)
120
121
122#ifdef __cplusplus
Marat Dukhan80fc9322019-09-29 21:06:36 -0700123} // extern "C"
XNNPACK Teamb455b122019-09-27 18:10:33 -0700124#endif