config.hpp 4.15 KB
Newer Older
1
2
3
#ifndef CK_CONFIG_AMD_HPP
#define CK_CONFIG_AMD_HPP

Chao Liu's avatar
Chao Liu committed
4
#ifndef MIOPEN_DONT_USE_HIP_RUNTIME_HEADERS
5
6
#include "hip/hip_runtime.h"
#include "hip/hip_fp16.h"
Chao Liu's avatar
Chao Liu committed
7
#endif
8

Chao Liu's avatar
tidy  
Chao Liu committed
9
// "Constant" address space for kernel parameter
10
#define CONSTANT __attribute__((address_space(4)))
11

12
13
14
15
// GPU target
// should enable one and only one GPU target
#if !(defined(CK_AMD_GPU_GFX803) || defined(CK_AMD_GPU_GFX900) || defined(CK_AMD_GPU_GFX906) || \
      defined(CK_AMD_GPU_GFX908) || defined(CK_AMD_GPU_GFX90A) || defined(CK_AMD_GPU_GFX1030))
16
#error Need to define (only) one GPU target
Chao Liu's avatar
Chao Liu committed
17
18
19
#endif

// launch bounds
zjing14's avatar
zjing14 committed
20
#define CK_USE_LAUNCH_BOUNDS 1
Chao Liu's avatar
Chao Liu committed
21
22
23

#ifdef CK_USE_LAUNCH_BOUNDS
#define CK_MAX_THREAD_PER_BLOCK 256
Chao Liu's avatar
Chao Liu committed
24
#define CK_MIN_BLOCK_PER_CU 2
Chao Liu's avatar
Chao Liu committed
25
26
27
#endif

// buffer resourse
28
29
#if defined(CK_AMD_GPU_GFX803) || defined(CK_AMD_GPU_GFX900) || defined(CK_AMD_GPU_GFX906) || \
    defined(CK_AMD_GPU_GFX908) || defined(CK_AMD_GPU_GFX90A)
Chao Liu's avatar
Chao Liu committed
30
31
32
33
34
#define CK_BUFFER_RESOURCE_3RD_DWORD 0x00020000
#elif defined(CK_AMD_GPU_GFX1030)
#define CK_BUFFER_RESOURCE_3RD_DWORD 0x31014000
#endif

35
36
37
38
39
40
41
42
43
44
// FMA instruction
#if defined(CK_AMD_GPU_GFX803) || defined(CK_AMD_GPU_GFX900)
#define CK_USE_AMD_V_MAC_F32
#elif defined(CK_AMD_GPU_GFX906) || defined(CK_AMD_GPU_GFX908) || defined(CK_AMD_GPU_GFX90a) || \
    defined(CK_AMD_GPU_GFX1030)
#define CK_USE_AMD_V_FMAC_F32
#define CK_USE_AMD_V_DOT2_F32_F16
#define CK_USE_AMD_V_DOT4_I32_I8
#endif

Chao Liu's avatar
Chao Liu committed
45
46
47
// multi index
#define CK_USE_DYNAMICALLY_INDEXED_MULTI_INDEX 0

48
49
50
51
52
// AMD inline asm
#ifndef CK_USE_AMD_INLINE_ASM
#define CK_USE_AMD_INLINE_ASM 1
#endif

53
54
55
// AMD inner product (DLOP)
#ifndef CK_USE_AMD_INNER_PRODUCT_INLINE_ASM
#define CK_USE_AMD_INNER_PRODUCT_INLINE_ASM 1
Chao Liu's avatar
Chao Liu committed
56
57
#endif

58
59
60
61
62
// AMD buffer addressing
#ifndef CK_USE_AMD_BUFFER_ADDRESSING
#define CK_USE_AMD_BUFFER_ADDRESSING 1
#endif

Chao Liu's avatar
Chao Liu committed
63
// only gfx908 support native floating point atomic add
Chao Liu's avatar
Chao Liu committed
64
65
#ifndef CK_USE_AMD_BUFFER_ATOMIC_FADD
#define CK_USE_AMD_BUFFER_ATOMIC_FADD 0
66
67
#endif

68
69
// AMD XDLOPS
#ifndef CK_USE_AMD_XDLOPS
70
#define CK_USE_AMD_XDLOPS 0
71
72
#endif

Chao Liu's avatar
Chao Liu committed
73
74
75
76
77
// block synchronization only s_wait lgkmcnt(0), not vmcnt(0)
#ifndef CK_BLOCK_SYNC_LDS_WITHOUT_SYNC_VMEM
#define CK_BLOCK_SYNC_LDS_WITHOUT_SYNC_VMEM 1
#endif

78
// experimental implementation for buffer load/store/atomic
Chao Liu's avatar
Chao Liu committed
79
#ifndef CK_EXPERIMENTAL_USE_BUFFER_LOAD_OOB_CHECK_OFFSET_TRICK
80
#define CK_EXPERIMENTAL_USE_BUFFER_LOAD_OOB_CHECK_OFFSET_TRICK 0
Chao Liu's avatar
Chao Liu committed
81
82
83
84
85
86
#endif

#ifndef CK_EXPERIMENTAL_USE_BUFFER_STORE_OOB_CHECK_OFFSET_TRICK
#define CK_EXPERIMENTAL_USE_BUFFER_STORE_OOB_CHECK_OFFSET_TRICK 1
#endif

zjing14's avatar
zjing14 committed
87
88
#ifndef CK_EXPERIMENTAL_USE_BUFFER_ATOMIC_ADD_OOB_CHECK_OFFSET_TRICK
#define CK_EXPERIMENTAL_USE_BUFFER_ATOMIC_ADD_OOB_CHECK_OFFSET_TRICK 1
89
90
#endif

91
92
93
94
95
// experimental implementation for in-regsiter sub-dword transpose
#ifndef CK_EXPERIMENTAL_USE_IN_REGISTER_SUB_DWORD_TRANSPOSE
#define CK_EXPERIMENTAL_USE_IN_REGISTER_SUB_DWORD_TRANSPOSE 1
#endif

96
// pass tensor descriptor by value or void*
Chao Liu's avatar
Chao Liu committed
97
98
#define CK_EXPERIMENTAL_PASS_TENSOR_DESCRIPTOR_BY_VALUE 1
#define CK_EXPERIMENTAL_PASS_TENSOR_DESCRIPTOR_BY_VOID_POINTER 0
Chao Liu's avatar
Chao Liu committed
99

100
// merge transformation use magic number division
101
#define CK_EXPERIMENTAL_MERGE_USE_MAGIC_DIVISION 1
102

Chao Liu's avatar
Chao Liu committed
103
104
105
106
// hack: have underlying assumption that need to be satsified, otherwise it's a bug
// hack for forcing register to keep idx_diff_low_const in SGPR. idx_diff_low_const must be
// thread-invariant, otherwise it's a bug
// TODO: separate index calculation into "compile-time", "global", "block", "wave", "thread"
Chao Liu's avatar
rename  
Chao Liu committed
107
108
#ifndef CK_HACK_MERGE_CALCULATE_IDX_DIFF_LOW_CONST_USE_AMD_GCN_READ_FIRST_LANE
#define CK_HACK_MERGE_CALCULATE_IDX_DIFF_LOW_CONST_USE_AMD_GCN_READ_FIRST_LANE 0
Chao Liu's avatar
Chao Liu committed
109
110
111
112
113
#endif

// workaround for compiler crash when compiling recursive lambda
#ifndef CK_WORKAROUND_SWDEV_275126
#define CK_WORKAROUND_SWDEV_275126 1
Chao Liu's avatar
Chao Liu committed
114
#endif
115

116
// workaround for compiler crash when using buffer load/store for i8
117
118
119
120
121
122
123
#ifndef CK_WORKAROUND_SWDEV_XXXXXX_INT8_BUFFER_LOAD_STORE_ISSUE
#define CK_WORKAROUND_SWDEV_XXXXXX_INT8_BUFFER_LOAD_STORE_ISSUE 1
#endif

// workaround for compiler crash when using buffer load/store for i8
#ifndef CK_WORKAROUND_SWDEV_XXXXXX_INT8_DS_WRITE_ISSUE
#define CK_WORKAROUND_SWDEV_XXXXXX_INT8_DS_WRITE_ISSUE 1
124
125
#endif

126
127
namespace ck {

128
enum InMemoryDataOperationEnum_t
Chao Liu's avatar
Chao Liu committed
129
{
Chao Liu's avatar
Chao Liu committed
130
131
    Set,
    AtomicAdd
132
133
};

Chao Liu's avatar
Chao Liu committed
134
// index type
135
136
137
138
using index_t = int32_t;

} // namespace ck
#endif