README.md 5.06 KB
Newer Older
1
2
3
# Composable Kernel

## Methodology
Sam Wu's avatar
Sam Wu committed
4

Chao Liu's avatar
Chao Liu committed
5
Composable Kernel (CK) library aims to provide a programming model for writing performance critical kernels for machine learning workloads across multiple architectures including GPUs, CPUs, etc, through general purpose kernel languages, like HIP C++.
6

Chao Liu's avatar
Chao Liu committed
7
CK utilizes two concepts to achieve performance portability and code maintainability:
8
9
10
* A tile-based programming model
* Algorithm complexity reduction for complex ML operators, using innovative technique we call "Tensor Coordinate Transformation".

Sam Wu's avatar
Sam Wu committed
11
![ALT](/docs/data/ck_component.png "CK Components")
12
13

## Code Structure
Sam Wu's avatar
Sam Wu committed
14

15
Current CK library are structured into 4 layers:
Chao Liu's avatar
Chao Liu committed
16
* "Templated Tile Operators" layer
17
18
19
20
* "Templated Kernel and Invoker" layer
* "Instantiated Kernel and Invoker" layer
* "Client API" layer

Sam Wu's avatar
Sam Wu committed
21
22
23
24
25
26
27
28
![ALT](/docs/data/ck_layer.png "CK Layers")

## Documentation

Run the steps below to build documentation locally.

```
cd docs
Sam Wu's avatar
Sam Wu committed
29
pip3 install -r sphinx/requirements.txt
Sam Wu's avatar
Sam Wu committed
30
31
python3 -m sphinx -T -E -b html -d _build/doctrees -D language=en . _build/html
```
32
33

## Contributors
Sam Wu's avatar
Sam Wu committed
34

35
36
37
The list of developers and contributors is here: [Contributors](/CONTRIBUTORS.md)

## Citation
Sam Wu's avatar
Sam Wu committed
38

39
40
41
42
43
If you use CK, please use following citations:
* CK paper will be freely available on arXiv soon: [Realizing Tensor Operators Using Coordinate Transformations and Tile Based Programming](???)
* [CITATION.cff](/CITATION.cff)

## License
Sam Wu's avatar
Sam Wu committed
44

45
46
47
48
49
50
CK is released under the MIT license. [License File](/LICENSE)


# Build CK

## Build docker image
Sam Wu's avatar
Sam Wu committed
51

52
53
54
```bash
DOCKER_BUILDKIT=1 docker build -t ck:latest -f Dockerfile .
```
55
56
Pre-built dockers are available from this public repo: 
https://hub.docker.com/r/rocm/composable_kernel/tags
57
58

## Launch docker
Sam Wu's avatar
Sam Wu committed
59

60
61
62
63
64
65
66
```bash
docker run                                     \
-it                                            \
--privileged                                   \
--group-add sudo                               \
-w /root/workspace                             \
-v ${PATH_TO_LOCAL_WORKSPACE}:/root/workspace  \
67
ck:latest                                      \
68
69
/bin/bash
```
Chao Liu's avatar
Chao Liu committed
70

71
## Build CK
Sam Wu's avatar
Sam Wu committed
72

73
74
75
```bash
mkdir build && cd build

76
# Need to specify target ID, example below is for gfx908 and gfx90a
Sam Wu's avatar
Sam Wu committed
77

78
79
80
81
cmake                                                                                             \
-D CMAKE_PREFIX_PATH=/opt/rocm                                                                    \
-D CMAKE_CXX_COMPILER=/opt/rocm/bin/hipcc                                                         \
-D CMAKE_BUILD_TYPE=Release                                                                       \
82
-D GPU_TARGETS="gfx908;gfx90a"                                                                    \
83
84
85
..
```

86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
If GPU_TARGETS is not set on the cmake command line, CK will be built for all targets supported by the 
current compiler.


Additional cmake flags can be used to significantly speed-up the build:

INSTANCES_ONLY (by default is OFF) must be set to ON in order to build only the instances and library
while skipping all tests, examples, and profiler. This is useful for libraries that use CK as a dependency.

DTYPES (by default not set) can be set to any subset of "fp64;fp32;fp16;fp8;bf16;int8" to build instances 
of select data types only. Currently, building of int8 instances is taking a lot of time (the compiler fix is in the works).

DL_KERNELS (by default is OFF) must be set to ON in order to build the gemm_dl and batched_gemm_multi_d_dl 
instances. Those instances are only needed for the NAVI2x platforms.

101
### Build examples and tests
Sam Wu's avatar
Sam Wu committed
102

103
```bash
104
 make -j examples tests
105
106
107
 make test
```

108
109
110
Instructions for running each individual examples are under [example](/example)


111
## Build ckProfiler
Sam Wu's avatar
Sam Wu committed
112

113
114
115
```bash
 make -j ckProfiler
```
116
Instructions for running ckProfiler are under [profiler](/profiler)
JD's avatar
JD committed
117

118
## Install CK
Sam Wu's avatar
Sam Wu committed
119

120
121
122
123
124
```bash
make install
```

## Using CK as pre-built kernel library
Sam Wu's avatar
Sam Wu committed
125

126
Instructions for using CK as a pre-built kernel library are under [client_example](/client_example)
JD's avatar
JD committed
127

128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
## Contributing

When you contribute to Composable Kernel, make sure to run `clang-format` on all the changed files. We highly recommend using git hooks that are managed by the `pre-commit` framework. To install hooks, run:

```bash
sudo script/install_precommit.sh
```

This way, `pre-commit` will add the appropriate hooks to your local repository and automatically run `clang-format` (and possibly additional checks) before any commit is created.

If you need to uninstall hooks from the repository, you can do so by running the following command:

```bash
script/uninstall_precommit.sh
```

If for any reason, you need to temporarily disable precommit hooks, you can add the `--no-verify` option to the `git commit` command.

JD's avatar
JD committed
146
147
## Caveat
### Kernel Timing and Verification
Sam Wu's avatar
Sam Wu committed
148

JD's avatar
JD committed
149
150
CK's own kernel timer will warn up kernel once, and then run it multiple times
to get average kernel time. For some kernels that use atomic add, this will cause
Chao Liu's avatar
Chao Liu committed
151
output buffer to be accumulated multiple times, causing verification failure.
JD's avatar
JD committed
152
153
154
To work around it, do not use CK's own timer and do verification at the same time.
CK's own timer and verification in each example and ckProfiler can be enabled or
disabled from command line.