ck.hpp 7.65 KB
Newer Older
Chao Liu's avatar
Chao Liu committed
1
// SPDX-License-Identifier: MIT
Illia Silin's avatar
Illia Silin committed
2
// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
Chao Liu's avatar
Chao Liu committed
3

Chao Liu's avatar
Chao Liu committed
4
#pragma once
5

JD's avatar
JD committed
6
#ifndef CK_DONT_USE_HIP_RUNTIME_HEADERS
7
8
#include "hip/hip_runtime.h"
#include "hip/hip_fp16.h"
Chao Liu's avatar
Chao Liu committed
9
#endif
10

Chao Liu's avatar
Chao Liu committed
11
12
#define CK_TIME_KERNEL 1

13
14
15
// constant address space for kernel parameter
// https://llvm.org/docs/AMDGPUUsage.html#address-spaces
#define CK_CONSTANT_ADDRESS_SPACE __attribute__((address_space(4)))
Chao Liu's avatar
Chao Liu committed
16
17

// launch bounds
zjing14's avatar
zjing14 committed
18
#define CK_USE_LAUNCH_BOUNDS 1
Chao Liu's avatar
Chao Liu committed
19
20

#ifdef CK_USE_LAUNCH_BOUNDS
21
// for most kernels
Chao Liu's avatar
Chao Liu committed
22
#define CK_MAX_THREAD_PER_BLOCK 256
Chao Liu's avatar
Chao Liu committed
23
#define CK_MIN_BLOCK_PER_CU 2
24
25
26
27

// for wavelet GEMM kernel
#define CK_WAVELET_MAX_THREAD_PER_BLOCK 512
#define CK_WAVELET_MIN_BLOCK_PER_CU 2
Chao Liu's avatar
Chao Liu committed
28
29
#endif

Chao Liu's avatar
Chao Liu committed
30
// buffer resource
31
32
33
#ifndef __HIP_DEVICE_COMPILE__ // for host code
#define CK_BUFFER_RESOURCE_3RD_DWORD -1
#elif defined(__gfx803__) || defined(__gfx900__) || defined(__gfx906__) || defined(__gfx908__) || \
34
35
    defined(__gfx90a__) || defined(__gfx940__) || defined(__gfx941__) ||                          \
    defined(__gfx942__) // for GPU code
Chao Liu's avatar
Chao Liu committed
36
#define CK_BUFFER_RESOURCE_3RD_DWORD 0x00020000
37
#elif defined(__gfx1030__) // for GPU code
Chao Liu's avatar
Chao Liu committed
38
#define CK_BUFFER_RESOURCE_3RD_DWORD 0x31014000
39
#elif defined(__gfx1100__) || defined(__gfx1101__) || defined(__gfx1102__) // for GPU code
40
#define CK_BUFFER_RESOURCE_3RD_DWORD 0x31004000
Chao Liu's avatar
Chao Liu committed
41
42
#endif

43
// FMA instruction
44
45
#ifndef __HIP_DEVICE_COMPILE__                   // for host code, define nothing
#elif defined(__gfx803__) || defined(__gfx900__) // for GPU code
46
#define CK_USE_AMD_V_MAC_F32
47
#elif defined(__gfx906__) || defined(__gfx908__) || defined(__gfx90a__) || defined(__gfx1030__) || \
48
    defined(__gfx940__) || defined(__gfx941__) || defined(__gfx942__) // for GPU code
49
50
51
52
53
#define CK_USE_AMD_V_FMAC_F32
#define CK_USE_AMD_V_DOT2_F32_F16
#define CK_USE_AMD_V_DOT4_I32_I8
#endif

54
55
56
// MFMA instruction
#ifndef __HIP_DEVICE_COMPILE__ // for host code
#define CK_USE_AMD_MFMA
57
58
#elif defined(__gfx908__) || defined(__gfx90a__) || defined(__gfx940__) || defined(__gfx941__) || \
    defined(__gfx942__) // for GPU code
59
#define CK_USE_AMD_MFMA
60
61
#endif

62
#if(defined(__gfx90a__) || defined(__gfx940__) || defined(__gfx941__) || defined(__gfx942__))
63
#define CK_USE_AMD_MFMA_BF16_1K_OP
Chao Liu's avatar
Chao Liu committed
64
65
#endif

66
#if defined(__gfx940__) || defined(__gfx941__) || defined(__gfx942__)
67
68
69
#define CK_USE_AMD_MFMA_GFX940
#endif

70
71
72
// WMMA instruction
#ifndef __HIP_DEVICE_COMPILE__ // for host code
#define CK_USE_AMD_WMMA
73
#elif defined(__gfx1100__) || defined(__gfx1101__) || defined(__gfx1102__) // for GPU code
74
75
76
#define CK_USE_AMD_WMMA
#endif

77
// buffer load
Jianfeng Yan's avatar
Jianfeng Yan committed
78
#define CK_USE_AMD_BUFFER_LOAD 1
79

80
// buffer store
Jianfeng Yan's avatar
Jianfeng Yan committed
81
82
#define CK_USE_AMD_BUFFER_STORE 1

83
84
// buffer atomic add: integer
#define CK_USE_AMD_BUFFER_ATOMIC_ADD_INTEGER 1
85

86
87
88
// buffer atomic add: floating point
#ifndef __HIP_DEVICE_COMPILE__ // for host code
#define CK_USE_AMD_BUFFER_ATOMIC_ADD_FLOAT 1
89
90
#elif defined(__gfx908__) || defined(__gfx90a__) || defined(__gfx940__) || defined(__gfx941__) || \
    defined(__gfx942__) // for GPU code
91
92
93
#define CK_USE_AMD_BUFFER_ATOMIC_ADD_FLOAT 1
#else // for GPU code
#define CK_USE_AMD_BUFFER_ATOMIC_ADD_FLOAT 0
94
95
#endif

96
97
#if(defined(__gfx90a__) || defined(__gfx940__) || defined(__gfx941__) || \
    defined(__gfx942__)) // for GPU code
rocking5566's avatar
rocking5566 committed
98
99
100
101
102
#define CK_USE_AMD_BUFFER_ATOMIC_MAX_FLOAT64 1
#else
#define CK_USE_AMD_BUFFER_ATOMIC_MAX_FLOAT64 0
#endif

103
104
105
106
107
108
// inline asm
#define CK_USE_AMD_INLINE_ASM 1

// inner product (DLOP)
#define CK_USE_AMD_INNER_PRODUCT_INLINE_ASM 1

Chao Liu's avatar
Chao Liu committed
109
// block synchronization only s_wait lgkmcnt(0), not vmcnt(0)
110
#define CK_EXPERIMENTAL_BLOCK_SYNC_LDS_WITHOUT_SYNC_VMEM 1
Chao Liu's avatar
Chao Liu committed
111

112
113
// experimental feature: multi index implemented as array
#define CK_EXPERIMENTAL_USE_DYNAMICALLY_INDEXED_MULTI_INDEX 0
Chao Liu's avatar
Chao Liu committed
114

115
116
// experimental feature: static tensor descriptor
#define CK_EXPERIMENTAL_STATIC_TENSOR_DESCRIPTOR 0
Chao Liu's avatar
Chao Liu committed
117

rocking5566's avatar
rocking5566 committed
118
// experimental feature: buffer load/store/atomic-add/ OOB trick
119
// This (ifndef) is a hack to use customized behavior for buffer load rather than using default
Po Yen Chen's avatar
Po Yen Chen committed
120
121
// setting. Don't use this hack unless absolutely necessary!
// FIXME: make the behavior of buffer load a configurable (template) parameter for each usage
122
#ifndef CK_EXPERIMENTAL_USE_BUFFER_LOAD_OOB_CHECK_OFFSET_TRICK
123
#define CK_EXPERIMENTAL_USE_BUFFER_LOAD_OOB_CHECK_OFFSET_TRICK 0
124
#endif
125
#define CK_EXPERIMENTAL_USE_BUFFER_STORE_OOB_CHECK_OFFSET_TRICK 1
zjing14's avatar
zjing14 committed
126
#define CK_EXPERIMENTAL_USE_BUFFER_ATOMIC_ADD_OOB_CHECK_OFFSET_TRICK 1
rocking5566's avatar
rocking5566 committed
127
#define CK_EXPERIMENTAL_USE_BUFFER_ATOMIC_MAX_OOB_CHECK_OFFSET_TRICK 1
128

129
// experimental feature: in-regsiter sub-dword transpose
130
#define CK_EXPERIMENTAL_USE_IN_REGISTER_SUB_DWORD_TRANSPOSE 1
Chao Liu's avatar
Chao Liu committed
131

132
// experimental feature: merge transformation use magic number division
133
#define CK_EXPERIMENTAL_MERGE_USE_MAGIC_DIVISION 1
134

135
136
// experimental feature: use __builtin_memcpy instead of pointer cast to access a vector from
// pointer of scalar
137
138
#define CK_EXPERIMENTAL_USE_MEMCPY_FOR_VECTOR_ACCESS 0

139
// experimental feature: use __builtin_memcpy instead of union to do bit_cast
140
#define CK_EXPERIMENTAL_USE_MEMCPY_FOR_BIT_CAST 1
141

142
// experimental feature: optimize for inter-wave scheduling policy
143
#define CK_EXPERIMENTAL_INTER_WAVE_SCHEDULING 1
144
#define CK_EXPERIMENTAL_INTER_WAVE_SCHEDULING_MAC_CLUSTERS 1
145
146
147
148
149
150
// this will let make_default_loop_scheduler() return interwave scheduling flag by default
#define CK_EXPERIMENTAL_DEFAULT_TO_INTER_WAVE_SCHEDULING 0
// experimental feature: add instances using interwave scheduling
#define CK_EXPERIMENTAL_INTER_WAVE_INSTANCES 1
// experimental feature: add instances using pipeline v2
#define CK_EXPERIMENTAL_PIPELINE_V2_INSTANCES 1
151

Chao Liu's avatar
Chao Liu committed
152
153
154
155
// 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
156
#define CK_HACK_MERGE_CALCULATE_IDX_DIFF_LOW_CONST_USE_AMD_GCN_READ_FIRST_LANE 0
Chao Liu's avatar
Chao Liu committed
157

158
// workaround: compiler crash when compiling recursive lambda
Chao Liu's avatar
Chao Liu committed
159
#define CK_WORKAROUND_SWDEV_275126 1
160

161
// workaround: compiler crash when using buffer load/store for i8
162
163
#define CK_WORKAROUND_SWDEV_XXXXXX_INT8_BUFFER_LOAD_STORE_ISSUE 1

164
// workaround: compiler gnerating inefficient ds_write instructions
165
#define CK_WORKAROUND_SWDEV_XXXXXX_INT8_DS_WRITE_ISSUE 1
Chao Liu's avatar
Chao Liu committed
166

167
// workaround: verifaction failure, due to compiler regression, for conv bwd-data fp16 using some
168
// tuning parameter
169
#define CK_WORKAROUND_SWDEV_325164 0
170

171
172
173
// workaround: compiler not emiting reciprocal instruction frm __frcp_rn()
#define CK_WORKAROUND_SWDEV_383542 1

ltqin's avatar
ltqin committed
174
175
// workaround: compiler issue on gfx908
#define CK_WORKAROUND_SWDEV_388832 1
176
177
178
179

// workaround: Grouped Conv2d_bwd_data fails for already implemented instance
#define CK_WORKAROUND_SWDEV_3318619 0

180
181
182
// flag to enable (1) or disable (0) the debugging output in some kernels
#define DEBUG_LOG 0

183
184
185
// denorm test fix, required to work around dissue
#ifndef CK_WORKAROUND_DENORM_FIX
#define CK_WORKAROUND_DENORM_FIX 0
186
187
188
189
#elif
// enable only on MI200
#define CK_WORKAROUND_DENORM_FIX = CK_WORKAROUND_DENORM_FIX && defined(__gfx90a__)
#endif // CK_WORKAROUND_DENORM_FIX
190

191
192
namespace ck {

193
enum struct InMemoryDataOperationEnum
Chao Liu's avatar
Chao Liu committed
194
{
Chao Liu's avatar
Chao Liu committed
195
    Set,
196
    AtomicAdd,
rocking5566's avatar
rocking5566 committed
197
    AtomicMax,
198
199
200
    Add
};

Chao Liu's avatar
Chao Liu committed
201
// FIXME: use regular Sequence and remove this
rocking5566's avatar
rocking5566 committed
202
203
204
205
206
207
208
209
210
211
212
213
214
template <InMemoryDataOperationEnum... Is>
struct InMemoryDataOperationEnumSequence
{
    static constexpr int mSize = sizeof...(Is);

    __host__ __device__ static constexpr InMemoryDataOperationEnum At(int I)
    {
        // the last dummy element is to prevent compiler complain about empty array, when mSize = 0
        const InMemoryDataOperationEnum mData[mSize + 1] = {Is..., InMemoryDataOperationEnum::Set};
        return mData[I];
    }
};

Chao Liu's avatar
Chao Liu committed
215
// index type
Jianfeng Yan's avatar
Jianfeng Yan committed
216
217
using index_t      = int32_t;
using long_index_t = int64_t;
218
219

} // namespace ck