mirror of
https://github.com/NVIDIA/cuda-samples.git
synced 2026-11-17 17:38:18 +08:00
Compare commits
242 Commits
| Author | SHA1 | Date | |
|---|---|---|---|
|
|
b7c5481c55 | ||
|
|
aeab82ff30 | ||
|
|
a4526d5229 | ||
|
|
f69622b6b6 | ||
|
|
722c5dd069 | ||
|
|
e32d17e67e | ||
|
|
5cf83b5a98 | ||
|
|
191900a7b4 | ||
|
|
338df69d0f | ||
|
|
9b7a6f907a | ||
|
|
5bbae96fd3 | ||
|
|
ce9407c240 | ||
|
|
cdedff42d2 | ||
|
|
6c4d183ba3 | ||
|
|
af49216f0c | ||
|
|
17113ea87e | ||
|
|
70fcdd353f | ||
|
|
5b2dd19a21 | ||
|
|
37dcb9b61d | ||
|
|
1bc6774795 | ||
|
|
4f735616ba | ||
|
|
4ff552244d | ||
|
|
df6edf644e | ||
|
|
85231cd1b6 | ||
|
|
46a782fb44 | ||
|
|
d5dedb24d6 | ||
|
|
9600b04a3c | ||
|
|
4eae2ccc04 | ||
|
|
d759a16c4b | ||
|
|
b8191f860a | ||
|
|
547f65851f | ||
|
|
9df0ef6411 | ||
|
|
1411d243a6 | ||
|
|
57e870596c | ||
|
|
fd6707ccaa | ||
|
|
cf01c5bb28 | ||
|
|
1165d53ddf | ||
|
|
8c324afd63 | ||
|
|
0fae54dcd6 | ||
|
|
320c7e6392 | ||
|
|
22fafb0a7c | ||
|
|
cccbf2c284 | ||
|
|
69186dc51e | ||
|
|
d298bf3e6f | ||
|
|
27c7e19849 | ||
|
|
7c190dde53 | ||
|
|
5e9d44634b | ||
|
|
b0a2f2049f | ||
|
|
3e6d1858af | ||
|
|
12b046a80e | ||
|
|
036638b713 | ||
|
|
9d703d090b | ||
|
|
cbc075ef59 | ||
|
|
6e23a98575 | ||
|
|
5fc7a4c897 | ||
|
|
3f1c509650 | ||
|
|
fbc40da311 | ||
|
|
1b36fefcd5 | ||
|
|
13c2fd9717 | ||
|
|
775d1b4dd3 | ||
|
|
a5267b83a5 | ||
|
|
2ab16e6d15 | ||
|
|
903c840a8c | ||
|
|
24e7c5b428 | ||
|
|
b38ed29c95 | ||
|
|
4a631c9fd6 | ||
|
|
98afcd6515 | ||
|
|
eb8891b839 | ||
|
|
688fff15b2 | ||
|
|
555353b4b5 | ||
|
|
1a35902577 | ||
|
|
8013610840 | ||
|
|
2369d4c649 | ||
|
|
8433f89993 | ||
|
|
a9a4db666f | ||
|
|
87d03ef003 | ||
|
|
7c8fccc0d4 | ||
|
|
45314d7ff8 | ||
|
|
2f0a55d7dc | ||
|
|
0d61031846 | ||
|
|
225f84d433 | ||
|
|
e674cc36fe | ||
|
|
a47b422205 | ||
|
|
ce28796d6c | ||
|
|
9424c15848 | ||
|
|
9075c50a3d | ||
|
|
de5fa98e6e | ||
|
|
6c9e9d3cd2 | ||
|
|
7f5390cec3 | ||
|
|
5f6d46dfea | ||
|
|
49307463b5 | ||
|
|
859debae1b | ||
|
|
9c0d5aaae6 | ||
|
|
74a9a08887 | ||
|
|
0d08748ffa | ||
|
|
c1166d3298 | ||
|
|
27f47634a0 | ||
|
|
40d297dfe7 | ||
|
|
f19d33ddbc | ||
|
|
20b90b5063 | ||
|
|
8a9e2c830c | ||
|
|
61225e22f0 | ||
|
|
bf25251720 | ||
|
|
f47f06f077 | ||
|
|
0c78a3e0de | ||
|
|
5a1f5008ed | ||
|
|
8219570c15 | ||
|
|
5ce8b512ea | ||
|
|
fc88988f23 | ||
|
|
adacf1cffd | ||
|
|
3e631f467a | ||
|
|
1141dd7af4 | ||
|
|
da3b7a2b3c | ||
|
|
5987a9e9fa | ||
|
|
107f3f537f | ||
|
|
c1b03b9f81 | ||
|
|
ee0ee417c9 | ||
|
|
ebc1078379 | ||
|
|
2861e78272 | ||
|
|
494d598f86 | ||
|
|
c6208f5897 | ||
|
|
8f33cc6094 | ||
|
|
2ec9cf394a | ||
|
|
770e433a9e | ||
|
|
dd1a11f648 | ||
|
|
c6af90553e | ||
|
|
7989d1fcc8 | ||
|
|
611008fa86 | ||
|
|
bf628887f1 | ||
|
|
f288b2e261 | ||
|
|
330dd8472f | ||
|
|
b530f1cf42 | ||
|
|
b63fd83b6c | ||
|
|
8ee551c99a | ||
|
|
148014e709 | ||
|
|
cab7c66b4f | ||
|
|
8d400cfb7f | ||
|
|
f2645c5df8 | ||
|
|
6d6d964f97 | ||
|
|
ab68d58d59 | ||
|
|
c70d79cf3b | ||
|
|
9ac81370fa | ||
|
|
14b1bfdcc4 | ||
|
|
c14a0114d6 | ||
|
|
ee15cc0fe2 | ||
|
|
3438fd4875 | ||
|
|
b27b55ec70 | ||
|
|
49159f3739 | ||
|
|
93cafa8fe9 | ||
|
|
1680a1dc7f | ||
|
|
49daf0e4e0 | ||
|
|
a45fd3bd7c | ||
|
|
1627e96677 | ||
|
|
7e90d36120 | ||
|
|
9e50fdc01f | ||
|
|
2c0b36a967 | ||
|
|
83397dc811 | ||
|
|
640b566412 | ||
|
|
da24673a9f | ||
|
|
bded2585a4 | ||
|
|
5384563c57 | ||
|
|
01a62e2bc0 | ||
|
|
02fdb070ad | ||
|
|
278f4adbd2 | ||
|
|
d00076a7c1 | ||
|
|
4672b8ba2b | ||
|
|
0345908807 | ||
|
|
3b9c8ce2e9 | ||
|
|
e77d6eb5ab | ||
|
|
ac700327a2 | ||
|
|
56e669e2e4 | ||
|
|
17703dd426 | ||
|
|
a1b5a6f6e3 | ||
|
|
a32d5badf7 | ||
|
|
1fd22429c3 | ||
|
|
00ac0a1673 | ||
|
|
b013387a39 | ||
|
|
9d921e0fe7 | ||
|
|
7d1730f348 | ||
|
|
a4fba501a6 | ||
|
|
718fe6486d | ||
|
|
ad9908e32b | ||
|
|
952d6edf92 | ||
|
|
685709bfc7 | ||
|
|
0c92c34ca9 | ||
|
|
0d82634f70 | ||
|
|
4abbdf4e80 | ||
|
|
89789bd848 | ||
|
|
914ca00f89 | ||
|
|
8f6b189dae | ||
|
|
c8034f368a | ||
|
|
69522dd5b7 | ||
|
|
eddc6fd7e1 | ||
|
|
ceab6e8bcc | ||
|
|
2cd58fbc9a | ||
|
|
7ceb3122fc | ||
|
|
c0ab53f986 | ||
|
|
b87c243bbb | ||
|
|
e214cd29aa | ||
|
|
06d72496c2 | ||
|
|
2848d3bd21 | ||
|
|
bd0f630bf4 | ||
|
|
ab9166a6b2 | ||
|
|
62781dc15e | ||
|
|
c90a1c6981 | ||
|
|
3f97ef1288 | ||
|
|
408c9f69a8 | ||
|
|
9370f11e69 | ||
|
|
291435e0b4 | ||
|
|
5df07f114e | ||
|
|
8d901e745d | ||
|
|
990ebc01c2 | ||
|
|
541e9fc3f5 | ||
|
|
9adce9d9f2 | ||
|
|
b6f3b7add9 | ||
|
|
bcad2c9e61 | ||
|
|
e7b23470d5 | ||
|
|
310e7f2a11 | ||
|
|
7f0f63f311 | ||
|
|
acd3a015c8 | ||
|
|
a9869fd6ea | ||
|
|
3e8f91d1a1 | ||
|
|
f3b7c41ad6 | ||
|
|
29fb758e62 | ||
|
|
3bc08136ff | ||
|
|
85eefa06c4 | ||
|
|
c357dd1e6b | ||
|
|
efb46383e0 | ||
|
|
8d564d5e3a | ||
|
|
37c5bcbef4 | ||
|
|
940a4c7a91 | ||
|
|
61bd39800d | ||
|
|
8a96d2eee7 | ||
|
|
e762d58260 | ||
|
|
8fd1701744 | ||
|
|
94765c1597 | ||
|
|
c87881f02c | ||
|
|
25400b6b3c | ||
|
|
e24f62e28c | ||
|
|
22424227e7 | ||
|
|
42ff742bf5 | ||
|
|
8ccb13c6f0 |
49
.clang-format
Normal file
49
.clang-format
Normal file
@@ -0,0 +1,49 @@
|
|||||||
|
---
|
||||||
|
AccessModifierOffset: -4
|
||||||
|
AlignAfterOpenBracket: Align
|
||||||
|
AlignConsecutiveAssignments: Consecutive
|
||||||
|
AlignConsecutiveDeclarations: Consecutive
|
||||||
|
AlignConsecutiveMacros: Consecutive
|
||||||
|
AlignEscapedNewlines: Left
|
||||||
|
AlignOperands: AlignAfterOperator
|
||||||
|
AlignTrailingComments: true
|
||||||
|
AllowAllParametersOfDeclarationOnNextLine: false
|
||||||
|
BinPackArguments: false
|
||||||
|
BinPackParameters: false
|
||||||
|
BraceWrapping:
|
||||||
|
AfterClass: true
|
||||||
|
AfterControlStatement: false
|
||||||
|
AfterExternBlock: true
|
||||||
|
AfterFunction: true
|
||||||
|
AfterStruct: true
|
||||||
|
AfterUnion: true
|
||||||
|
BeforeCatch: true
|
||||||
|
BeforeElse: true
|
||||||
|
IndentBraces: false
|
||||||
|
BreakBeforeBraces: Custom
|
||||||
|
BreakBeforeConceptDeclarations: true
|
||||||
|
BreakBeforeBinaryOperators: NonAssignment
|
||||||
|
BreakBeforeTernaryOperators: true
|
||||||
|
BreakConstructorInitializers: BeforeComma
|
||||||
|
BreakInheritanceList: BeforeComma
|
||||||
|
ColumnLimit: 120
|
||||||
|
DerivePointerAlignment: false
|
||||||
|
FixNamespaceComments: true
|
||||||
|
IncludeCategories:
|
||||||
|
- Regex: '^<.*>'
|
||||||
|
Priority: 1
|
||||||
|
- Regex: '^".*"'
|
||||||
|
Priority: 2
|
||||||
|
SortIncludes: true
|
||||||
|
IncludeBlocks: Regroup
|
||||||
|
IndentWidth: 4
|
||||||
|
MaxEmptyLinesToKeep: 2
|
||||||
|
PointerAlignment: Right
|
||||||
|
SortUsingDeclarations: true
|
||||||
|
SpaceAfterCStyleCast: false
|
||||||
|
SpaceBeforeAssignmentOperators: true
|
||||||
|
SpaceBeforeParens: ControlStatements
|
||||||
|
Standard: c++17
|
||||||
|
TabWidth: 4
|
||||||
|
UseTab: Never
|
||||||
|
...
|
||||||
8
.gitignore
vendored
8
.gitignore
vendored
@@ -1,3 +1,11 @@
|
|||||||
build
|
build
|
||||||
|
build-*/
|
||||||
|
test-results*/
|
||||||
.vs
|
.vs
|
||||||
.clangd
|
.clangd
|
||||||
|
test
|
||||||
|
settings.json
|
||||||
|
launch.json
|
||||||
|
__pycache__/
|
||||||
|
*.py[co]
|
||||||
|
.pytest_cache/
|
||||||
|
|||||||
116
.pre-commit-config.yaml
Normal file
116
.pre-commit-config.yaml
Normal file
@@ -0,0 +1,116 @@
|
|||||||
|
# Copyright (c) 2024, NVIDIA CORPORATION.
|
||||||
|
ci:
|
||||||
|
autofix_commit_msg: |
|
||||||
|
[pre-commit.ci] auto code formatting
|
||||||
|
autofix_prs: false
|
||||||
|
autoupdate_branch: ''
|
||||||
|
autoupdate_commit_msg: '[pre-commit.ci] pre-commit autoupdate'
|
||||||
|
autoupdate_schedule: quarterly
|
||||||
|
skip: []
|
||||||
|
submodules: false
|
||||||
|
|
||||||
|
repos:
|
||||||
|
- repo: https://github.com/pre-commit/pre-commit-hooks
|
||||||
|
rev: v5.0.0
|
||||||
|
hooks:
|
||||||
|
- id: end-of-file-fixer
|
||||||
|
exclude: |
|
||||||
|
(?x)^(
|
||||||
|
.*\.raw$|
|
||||||
|
.*\.bin$|
|
||||||
|
.*\.dat$|
|
||||||
|
.*\.nv12$|
|
||||||
|
data/.*|
|
||||||
|
Common/.*
|
||||||
|
)
|
||||||
|
files: |
|
||||||
|
(?x)^(
|
||||||
|
.*\.txt$|
|
||||||
|
.*\.md$|
|
||||||
|
.*\.cpp$|
|
||||||
|
.*\.cxx$|
|
||||||
|
.*\.hpp$|
|
||||||
|
.*\.h$|
|
||||||
|
.*\.cu$|
|
||||||
|
.*\.cuh$|
|
||||||
|
.*\.py$|
|
||||||
|
.*\.json$
|
||||||
|
)
|
||||||
|
- id: mixed-line-ending
|
||||||
|
exclude: |
|
||||||
|
(?x)^(
|
||||||
|
.*\.raw$|
|
||||||
|
.*\.bin$|
|
||||||
|
.*\.dat$|
|
||||||
|
.*\.nv12$|
|
||||||
|
data/.*|
|
||||||
|
Common/.*
|
||||||
|
)
|
||||||
|
files: |
|
||||||
|
(?x)^(
|
||||||
|
.*\.txt$|
|
||||||
|
.*\.md$|
|
||||||
|
.*\.cpp$|
|
||||||
|
.*\.cxx$|
|
||||||
|
.*\.hpp$|
|
||||||
|
.*\.h$|
|
||||||
|
.*\.cu$|
|
||||||
|
.*\.cuh$|
|
||||||
|
.*\.py$|
|
||||||
|
.*\.json$
|
||||||
|
)
|
||||||
|
- id: trailing-whitespace
|
||||||
|
exclude: |
|
||||||
|
(?x)^(
|
||||||
|
.*\.raw$|
|
||||||
|
.*\.bin$|
|
||||||
|
.*\.dat$|
|
||||||
|
.*\.nv12$|
|
||||||
|
data/.*|
|
||||||
|
Common/.*
|
||||||
|
)
|
||||||
|
files: |
|
||||||
|
(?x)^(
|
||||||
|
.*\.txt$|
|
||||||
|
.*\.md$|
|
||||||
|
.*\.cpp$|
|
||||||
|
.*\.cxx$|
|
||||||
|
.*\.hpp$|
|
||||||
|
.*\.h$|
|
||||||
|
.*\.cu$|
|
||||||
|
.*\.cuh$|
|
||||||
|
.*\.py$|
|
||||||
|
.*\.json$
|
||||||
|
)
|
||||||
|
- repo: https://github.com/pre-commit/mirrors-clang-format
|
||||||
|
rev: v19.1.6
|
||||||
|
hooks:
|
||||||
|
- id: clang-format
|
||||||
|
types_or: [file]
|
||||||
|
files: |
|
||||||
|
(?x)^(
|
||||||
|
^.*\.c$|
|
||||||
|
^.*\.cpp$|
|
||||||
|
^.*\.cu$|
|
||||||
|
^.*\.cuh$|
|
||||||
|
^.*\.cxx$|
|
||||||
|
^.*\.h$|
|
||||||
|
^.*\.hpp$|
|
||||||
|
^.*\.inl$|
|
||||||
|
^.*\.mm$
|
||||||
|
)
|
||||||
|
exclude: |
|
||||||
|
(?x)^(
|
||||||
|
Common/.*
|
||||||
|
)
|
||||||
|
args: ["-fallback-style=none", "-style=file", "-i"]
|
||||||
|
- repo: https://github.com/astral-sh/ruff-pre-commit
|
||||||
|
rev: v0.8.4
|
||||||
|
hooks:
|
||||||
|
- id: ruff
|
||||||
|
name: ruff lint
|
||||||
|
args: [--fix]
|
||||||
|
files: ^python/.*\.py$
|
||||||
|
- id: ruff-format
|
||||||
|
name: ruff format
|
||||||
|
files: ^python/.*\.py$
|
||||||
113
CHANGELOG.md
113
CHANGELOG.md
@@ -1,5 +1,117 @@
|
|||||||
## Changelog
|
## Changelog
|
||||||
|
|
||||||
|
### CUDA 13.3
|
||||||
|
* Added **CUDA Tile C++** samples under `cpp/9_CUDA_Tile`.
|
||||||
|
* Added a set of **CCCL 3.3 feature samples** under `cpp/4_CUDA_Libraries/`, each built against CCCL fetched via CPM (pinned to v3.3.3, with an optional `CCCL_SOURCE_DIR` override):
|
||||||
|
* `cubDeviceFind` - `cub::DeviceFind::FindIf`, `LowerBound`, and `UpperBound` device-wide search algorithms.
|
||||||
|
* `cubDeviceSegmentedScan` - `cub::DeviceSegmentedScan::ExclusiveSegmentedSum` and `InclusiveSegmentedScan` with a custom binary operator.
|
||||||
|
* `cubDeviceTransform` - N-to-M `cub::DeviceTransform::Transform` where the op returns a `cuda::std::tuple`.
|
||||||
|
* `libcuxxRandom` - `cuda::pcg64` and `cuda::std::philox4x32` engines driving the uniform, normal, Poisson, and Bernoulli distributions from `<cuda/std/random>`.
|
||||||
|
* `libcuxxMdspan` - DLPack <-> `cuda::std::mdspan` bridging via `cuda::to_device_mdspan` / `cuda::to_dlpack_tensor`, plus `cuda::shared_memory_mdspan` for multi-dimensional views of shared memory.
|
||||||
|
* Added **cuda.compute 1.0 Python samples** under `python/2_CoreConcepts/`:
|
||||||
|
* `cudaComputeLambdas` - Python lambdas / regular callables driving `reduce_into`, `unary_transform`, and `inclusive_scan` in `cuda.compute` (from the `cuda-cccl` package).
|
||||||
|
* `binarySearch` - parallel `cuda.compute.upper_bound` / `lower_bound`, verified against `numpy.searchsorted`.
|
||||||
|
|
||||||
|
|
||||||
|
### CUDA 13.2 (update)
|
||||||
|
* Added **CUDA Python samples** under `python/`. These scripts use [CUDA Python](https://nvidia.github.io/cuda-python/) (including `cuda.core`) and are organized like the C++ tree: `1_GettingStarted`, `2_CoreConcepts`, `3_FrameworkInterop`, and `4_DistributedComputing`, plus shared helpers in `python/Utilities`. Each sample includes a `README.md` and `requirements.txt`. They are **not** built by the root CMake project; install dependencies with `pip install -r requirements.txt` in the sample directory, then run the corresponding `.py` file as documented in that sample’s README.
|
||||||
|
* Renamed top-level `Samples` directory to `cpp` to accommodate Python samples alongside existing C++ samples; updated path references in `CMakeLists.txt`, `README.md`, and `Common` headers accordingly.
|
||||||
|
|
||||||
|
### CUDA 13.2
|
||||||
|
* Added the MSVC compile flag `-Xcompiler=/Zc:preprocessor` in CMakeLists.txt to comply with CUDA13.2 CCCL. Previously, using the traditional preprocessor triggered the warning “MSVC/cl.exe with traditional preprocessor is used…”, which now leads to a build error.
|
||||||
|
|
||||||
|
### CUDA 13.1
|
||||||
|
* Minor bug fixes and enhancements, no structural or functional changes
|
||||||
|
|
||||||
|
### CUDA 13.0
|
||||||
|
* Updated the samples using the cudaDeviceProp fields which are deprecated and removed in CUDA 13.0, replacing the fields with the equivalents in "cudaDeviceGetAttribute":
|
||||||
|
* Deprecated "cudaDeviceProp" fields
|
||||||
|
`int clockRate; // - Replaced with "cudaDevAttrClockRate"`
|
||||||
|
`int deviceOverlap; // - Replaced with "cudaDevAttrGpuOverlap */`
|
||||||
|
`int kernelExecTimeoutEnabled; // - Replaced with "cudaDevAttrKernelExecTimeout`
|
||||||
|
`int computeMode; // - Replaced with "cudaDevAttrComputeMode" */`
|
||||||
|
`int memoryClockRate; // - Replaced with "cudaDevAttrMemoryClockRate"`
|
||||||
|
`int cooperativeMultiDeviceLaunch; // - Deprecated, cudaLaunchCooperativeKernelMultiDevice is deprecated.`
|
||||||
|
* `0_Introduction`
|
||||||
|
* `UnifiedMemoryStreams`
|
||||||
|
* `simpleHyperQ`
|
||||||
|
* `simpleIPC`
|
||||||
|
* `simpleMultiCopy`
|
||||||
|
* `systemWideAtomics`
|
||||||
|
* `1_Utilitie`
|
||||||
|
* `deviceQuery`
|
||||||
|
* `2_Concepts_and_Techniques`
|
||||||
|
* `streamOrderedAllocationIPC`
|
||||||
|
* `4_CUDA_Libraries`
|
||||||
|
* `simpleCUBLASXT`
|
||||||
|
* `5_Domain_Specific`
|
||||||
|
* `simpleVulkan`
|
||||||
|
* `vulkanImageCUDA`
|
||||||
|
* Updated the samples using the CUDA driver API "cuCtxCreate" with adding the parameter "CUctxCreateParams" as "cuCtxCreate" is updated to "cuCtxCreate_v4" by default in CUDA 13.0:
|
||||||
|
* `Common`
|
||||||
|
* `nvrtc_helper.h`
|
||||||
|
* `0_Introduction`
|
||||||
|
* `UnifiedMemoryStreams`
|
||||||
|
* `matrixMulDrv`
|
||||||
|
* `simpleTextureDrv`
|
||||||
|
* `vectorAddDrv`
|
||||||
|
* `vectorAddMMAP`
|
||||||
|
* `2_Concepts_and_Techniques`
|
||||||
|
* `EGLStream_CUDA_CrossGPU`
|
||||||
|
* `EGLStream_CUDA_Interop`
|
||||||
|
* `threadMigration`
|
||||||
|
* `3_CUDA_Features`
|
||||||
|
* `graphMemoryFootprint`
|
||||||
|
* `memMapIPCDrv`
|
||||||
|
* `4_CUDA_Libraries`
|
||||||
|
* `jitLto`
|
||||||
|
* `7_libNVVM`
|
||||||
|
* `cuda-c-linking`
|
||||||
|
* `device-side-launch`
|
||||||
|
* `simple`
|
||||||
|
* `uvmlite`
|
||||||
|
* `8_Platform_Specific/Tegra`
|
||||||
|
* `EGLSync_CUDAEvent_Interop`
|
||||||
|
* Updated the sample using CUDA API "cudaGraphAddNode"/"cudaStreamGetCaptureInfo" with adding "cudaGraphEdgeData" pointer parameter as they are updated to "cudaGraphAddNode_v2"/"cudaStreamGetCaptureInfo_v3" by default in CUDA 13.0:
|
||||||
|
* `3_CUDA_Features`
|
||||||
|
* `graphConditionalNodes`
|
||||||
|
* Updated the samples using CUDA API "cudaMemAdvise"/"cudaMemPrefetchAsync" with changing the parameter "int device" to "cudaMemLocation location" as they are updated to "cudaMemAdvise_v2"/"cudaMemPrefetchAsyn_v2" by default in CUDA 13.0.
|
||||||
|
* `4_CUDA_Libraries`
|
||||||
|
* `conjugateGradientMultiDeviceCG`
|
||||||
|
* `6_Performance`
|
||||||
|
* `UnifiedMemoryPerf`
|
||||||
|
* Replaced "thrust::identity<uint>()" with "cuda::std::identity()" as it is deprecated in CUDA 13.0.
|
||||||
|
* `2_Concepts_and_Techniques`
|
||||||
|
* `segmentationTreeThrust`
|
||||||
|
* Updated the the headers file and samples for CUFFT error codes update.
|
||||||
|
* Deprecated CUFFT errors:
|
||||||
|
* `CUFFT_INCOMPLETE_PARAMETER_LIST`
|
||||||
|
* `CUFFT_PARSE_ERROR`
|
||||||
|
* `CUFFT_LICENSE_ERROR`
|
||||||
|
* New added CUFFT errors:
|
||||||
|
* `CUFFT_MISSING_DEPENDENCY`
|
||||||
|
* `CUFFT_NVRTC_FAILURE`
|
||||||
|
* `CUFFT_NVJITLINK_FAILURE`
|
||||||
|
* `CUFFT_NVSHMEM_FAILURE`
|
||||||
|
* Header files and samples that are related with this change:
|
||||||
|
* `Common/helper_cuda.h`
|
||||||
|
* `4_CUDA_Libraries`
|
||||||
|
* `simpleCUFFT`
|
||||||
|
* `simpleCUFFT_2d_MGPU`
|
||||||
|
* `simpleCUFFT_MGPU`
|
||||||
|
* `simpleCUFFT_callback`
|
||||||
|
* Updated toolchain for cross-compilation for Tegra QNX platforms.
|
||||||
|
|
||||||
|
### CUDA 12.9
|
||||||
|
* Updated toolchain for cross-compilation for Tegra Linux platforms.
|
||||||
|
* Added `run_tests.py` utility to exercise all samples. See README.md for details
|
||||||
|
* Repository has been updated with consistent code formatting across all samples
|
||||||
|
* Many small code tweaks and bug fixes (see commit history for details)
|
||||||
|
* Removed the following outdated samples:
|
||||||
|
* `1_Utilities`
|
||||||
|
* `bandwidthTest` - this sample was out of date and did not produce accurate results. For bandwidth
|
||||||
|
testing of NVIDIA GPU platforms, please refer to [NVBandwidth](https://github.com/NVIDIA/nvbandwidth)
|
||||||
|
|
||||||
### CUDA 12.8
|
### CUDA 12.8
|
||||||
* Updated build system across the repository to CMake. Removed Visual Studio project files and Makefiles.
|
* Updated build system across the repository to CMake. Removed Visual Studio project files and Makefiles.
|
||||||
* Removed the following outdated samples:
|
* Removed the following outdated samples:
|
||||||
@@ -36,6 +148,7 @@
|
|||||||
* `cuDLALayerwiseStatsHybrid`
|
* `cuDLALayerwiseStatsHybrid`
|
||||||
* `cuDLALayerwiseStatsStandalone`
|
* `cuDLALayerwiseStatsStandalone`
|
||||||
* `cuDLAStandaloneMode`
|
* `cuDLAStandaloneMode`
|
||||||
|
* `cudaNvSciBufMultiplanar`
|
||||||
* `cudaNvSciNvMedia`
|
* `cudaNvSciNvMedia`
|
||||||
* `fluidsGLES`
|
* `fluidsGLES`
|
||||||
* `nbody_opengles`
|
* `nbody_opengles`
|
||||||
|
|||||||
@@ -1,7 +1,5 @@
|
|||||||
cmake_minimum_required(VERSION 3.20)
|
cmake_minimum_required(VERSION 3.20)
|
||||||
|
|
||||||
list(APPEND CMAKE_MODULE_PATH "${CMAKE_CURRENT_SOURCE_DIR}/../../../cmake/Modules")
|
|
||||||
|
|
||||||
project(cuda-samples LANGUAGES C CXX CUDA)
|
project(cuda-samples LANGUAGES C CXX CUDA)
|
||||||
|
|
||||||
find_package(CUDAToolkit REQUIRED)
|
find_package(CUDAToolkit REQUIRED)
|
||||||
@@ -14,12 +12,22 @@ set(CMAKE_CXX_STANDARD_REQUIRED ON)
|
|||||||
set(CMAKE_CUDA_STANDARD 17)
|
set(CMAKE_CUDA_STANDARD 17)
|
||||||
set(CMAKE_CUDA_STANDARD_REQUIRED ON)
|
set(CMAKE_CUDA_STANDARD_REQUIRED ON)
|
||||||
|
|
||||||
set(CMAKE_CUDA_ARCHITECTURES 50 52 60 61 70 72 75 80 86 87 89 90 100 101 120)
|
set(CMAKE_CUDA_ARCHITECTURES 75 80 86 87 89 90 100 110 120)
|
||||||
set(CMAKE_CUDA_FLAGS "${CMAKE_CUDA_FLAGS} -Wno-deprecated-gpu-targets")
|
set(CMAKE_CUDA_FLAGS "${CMAKE_CUDA_FLAGS} -Wno-deprecated-gpu-targets")
|
||||||
if(CMAKE_BUILD_TYPE STREQUAL "Debug")
|
if(ENABLE_CUDA_DEBUG)
|
||||||
# set(CMAKE_CUDA_FLAGS "${CMAKE_CUDA_FLAGS} -G") # enable cuda-gdb (expensive)
|
set(CMAKE_CUDA_FLAGS "${CMAKE_CUDA_FLAGS} -G") # enable cuda-gdb (may significantly affect performance on some targets)
|
||||||
|
else()
|
||||||
|
set(CMAKE_CUDA_FLAGS "${CMAKE_CUDA_FLAGS} -lineinfo") # add line information to all builds for debug tools (exclusive to -G option)
|
||||||
endif()
|
endif()
|
||||||
|
|
||||||
set(CMAKE_CUDA_FLAGS "${CMAKE_CUDA_FLAGS} --extended-lambda")
|
set(CMAKE_CUDA_FLAGS "${CMAKE_CUDA_FLAGS} --extended-lambda")
|
||||||
|
|
||||||
add_subdirectory(Samples)
|
# Add MSVC-specific flags for standard-conforming preprocessor (required for CCCL)
|
||||||
|
if(MSVC)
|
||||||
|
add_compile_options($<$<COMPILE_LANGUAGE:CUDA>:-Xcompiler=/Zc:preprocessor>)
|
||||||
|
endif()
|
||||||
|
|
||||||
|
# Include installation configuration before processing samples
|
||||||
|
include(cmake/InstallSamples.cmake)
|
||||||
|
|
||||||
|
add_subdirectory(cpp)
|
||||||
|
|||||||
103
CONTRIBUTING.md
Normal file
103
CONTRIBUTING.md
Normal file
@@ -0,0 +1,103 @@
|
|||||||
|
|
||||||
|
# Contributing to the CUDA Samples
|
||||||
|
|
||||||
|
Thank you for your interest in contributing to the CUDA Samples!
|
||||||
|
|
||||||
|
|
||||||
|
## Getting Started
|
||||||
|
|
||||||
|
1. **Fork & Clone the Repository**:
|
||||||
|
|
||||||
|
Fork the reporistory and clone the fork. For more information, check [GitHub's documentation on forking](https://docs.github.com/en/github/getting-started-with-github/fork-a-repo) and [cloning a repository](https://docs.github.com/en/github/creating-cloning-and-archiving-repositories/cloning-a-repository).
|
||||||
|
|
||||||
|
## Making Changes
|
||||||
|
|
||||||
|
1. **Create a New Branch**:
|
||||||
|
|
||||||
|
```bash
|
||||||
|
git checkout -b your-feature-branch
|
||||||
|
```
|
||||||
|
|
||||||
|
2. **Make Changes**.
|
||||||
|
|
||||||
|
3. **Build and Test**:
|
||||||
|
|
||||||
|
Ensure changes don't break existing functionality by building and running tests.
|
||||||
|
|
||||||
|
For more details on building and testing, refer to the [Building and Testing](#building-and-testing) section below.
|
||||||
|
|
||||||
|
4. **Commit Changes**:
|
||||||
|
|
||||||
|
```bash
|
||||||
|
git commit -m "Brief description of the change"
|
||||||
|
```
|
||||||
|
|
||||||
|
## Building and Testing
|
||||||
|
|
||||||
|
For information on building a running tests on the samples, please refer to the main [README](README.md)
|
||||||
|
|
||||||
|
## Creating a Pull Request
|
||||||
|
|
||||||
|
1. Push changes to your fork
|
||||||
|
2. Create a pull request targeting the `master` branch of the original CUDA Samples repository. Refer to [GitHub's documentation](https://docs.github.com/en/github/collaborating-with-pull-requests/proposing-changes-to-your-work-with-pull-requests/about-pull-requests) for more information on creating a pull request.
|
||||||
|
3. Describe the purpose and context of the changes in the pull request description.
|
||||||
|
|
||||||
|
## Code Formatting (pre-commit hooks)
|
||||||
|
|
||||||
|
The CUDA Samples repository uses [pre-commit](https://pre-commit.com/) to execute all code linters and formatters. These
|
||||||
|
tools ensure a consistent coding style throughout the project. Using pre-commit ensures that linter
|
||||||
|
versions and options are aligned for all developers. Additionally, there is a CI check in place to
|
||||||
|
enforce that committed code follows our standards.
|
||||||
|
|
||||||
|
The linters used by the CUDA Samples are listed in `.pre-commit-config.yaml`.
|
||||||
|
For example, C++ and CUDA code is formatted with [`clang-format`](https://clang.llvm.org/docs/ClangFormat.html).
|
||||||
|
|
||||||
|
To use `pre-commit`, install via `conda` or `pip`:
|
||||||
|
|
||||||
|
```bash
|
||||||
|
conda config --add channels conda-forge
|
||||||
|
conda install pre-commit
|
||||||
|
```
|
||||||
|
|
||||||
|
```bash
|
||||||
|
pip install pre-commit
|
||||||
|
```
|
||||||
|
|
||||||
|
Then run pre-commit hooks before committing code:
|
||||||
|
|
||||||
|
```bash
|
||||||
|
pre-commit run
|
||||||
|
```
|
||||||
|
|
||||||
|
By default, pre-commit runs on staged files (only changes and additions that will be committed).
|
||||||
|
To run pre-commit checks on all files, execute:
|
||||||
|
|
||||||
|
```bash
|
||||||
|
pre-commit run --all-files
|
||||||
|
```
|
||||||
|
|
||||||
|
Optionally, you may set up the pre-commit hooks to run automatically when you make a git commit. This can be done by running:
|
||||||
|
|
||||||
|
```bash
|
||||||
|
pre-commit install
|
||||||
|
```
|
||||||
|
|
||||||
|
Now code linters and formatters will be run each time you commit changes.
|
||||||
|
|
||||||
|
You can skip these checks with `git commit --no-verify` or with the short version `git commit -n`, althoguh please note
|
||||||
|
that this may result in pull requests being rejected if subsequent checks fail.
|
||||||
|
|
||||||
|
## Review Process
|
||||||
|
|
||||||
|
Once submitted, maintainers will be automatically assigned to review the pull request. They might suggest changes or improvements. Constructive feedback is a part of the collaborative process, aimed at ensuring the highest quality code.
|
||||||
|
|
||||||
|
For constructive feedback and effective communication during reviews, we recommend following [Conventional Comments](https://conventionalcomments.org/).
|
||||||
|
|
||||||
|
Further recommended reading for successful PR reviews:
|
||||||
|
|
||||||
|
- [How to Do Code Reviews Like a Human (Part One)](https://mtlynch.io/human-code-reviews-1/)
|
||||||
|
- [How to Do Code Reviews Like a Human (Part Two)](https://mtlynch.io/human-code-reviews-2/)
|
||||||
|
|
||||||
|
## Thank You
|
||||||
|
|
||||||
|
Your contributions enhance the CUDA Samples for the entire community. We appreciate your effort and collaboration!
|
||||||
@@ -138,26 +138,29 @@ static const char *_cudaGetErrorEnum(cufftResult error) {
|
|||||||
case CUFFT_UNALIGNED_DATA:
|
case CUFFT_UNALIGNED_DATA:
|
||||||
return "CUFFT_UNALIGNED_DATA";
|
return "CUFFT_UNALIGNED_DATA";
|
||||||
|
|
||||||
case CUFFT_INCOMPLETE_PARAMETER_LIST:
|
|
||||||
return "CUFFT_INCOMPLETE_PARAMETER_LIST";
|
|
||||||
|
|
||||||
case CUFFT_INVALID_DEVICE:
|
case CUFFT_INVALID_DEVICE:
|
||||||
return "CUFFT_INVALID_DEVICE";
|
return "CUFFT_INVALID_DEVICE";
|
||||||
|
|
||||||
case CUFFT_PARSE_ERROR:
|
|
||||||
return "CUFFT_PARSE_ERROR";
|
|
||||||
|
|
||||||
case CUFFT_NO_WORKSPACE:
|
case CUFFT_NO_WORKSPACE:
|
||||||
return "CUFFT_NO_WORKSPACE";
|
return "CUFFT_NO_WORKSPACE";
|
||||||
|
|
||||||
case CUFFT_NOT_IMPLEMENTED:
|
case CUFFT_NOT_IMPLEMENTED:
|
||||||
return "CUFFT_NOT_IMPLEMENTED";
|
return "CUFFT_NOT_IMPLEMENTED";
|
||||||
|
|
||||||
case CUFFT_LICENSE_ERROR:
|
|
||||||
return "CUFFT_LICENSE_ERROR";
|
|
||||||
|
|
||||||
case CUFFT_NOT_SUPPORTED:
|
case CUFFT_NOT_SUPPORTED:
|
||||||
return "CUFFT_NOT_SUPPORTED";
|
return "CUFFT_NOT_SUPPORTED";
|
||||||
|
|
||||||
|
case CUFFT_MISSING_DEPENDENCY:
|
||||||
|
return "CUFFT_MISSING_DEPENDENCY";
|
||||||
|
|
||||||
|
case CUFFT_NVRTC_FAILURE:
|
||||||
|
return "CUFFT_NVRTC_FAILURE";
|
||||||
|
|
||||||
|
case CUFFT_NVJITLINK_FAILURE:
|
||||||
|
return "CUFFT_NVJITLINK_FAILURE";
|
||||||
|
|
||||||
|
case CUFFT_NVSHMEM_FAILURE:
|
||||||
|
return "CUFFT_NVSHMEM_FAILURE";
|
||||||
}
|
}
|
||||||
|
|
||||||
return "<unknown>";
|
return "<unknown>";
|
||||||
@@ -670,7 +673,10 @@ inline int _ConvertSMVer2Cores(int major, int minor) {
|
|||||||
{0x90, 128},
|
{0x90, 128},
|
||||||
{0xa0, 128},
|
{0xa0, 128},
|
||||||
{0xa1, 128},
|
{0xa1, 128},
|
||||||
|
{0xa3, 128},
|
||||||
|
{0xb0, 128},
|
||||||
{0xc0, 128},
|
{0xc0, 128},
|
||||||
|
{0xc1, 128},
|
||||||
{-1, -1}};
|
{-1, -1}};
|
||||||
|
|
||||||
int index = 0;
|
int index = 0;
|
||||||
@@ -722,7 +728,10 @@ inline const char* _ConvertSMVer2ArchName(int major, int minor) {
|
|||||||
{0x90, "Hopper"},
|
{0x90, "Hopper"},
|
||||||
{0xa0, "Blackwell"},
|
{0xa0, "Blackwell"},
|
||||||
{0xa1, "Blackwell"},
|
{0xa1, "Blackwell"},
|
||||||
|
{0xa3, "Blackwell"},
|
||||||
|
{0xb0, "Blackwell"},
|
||||||
{0xc0, "Blackwell"},
|
{0xc0, "Blackwell"},
|
||||||
|
{0xc1, "Blackwell"},
|
||||||
{-1, "Graphics Device"}};
|
{-1, "Graphics Device"}};
|
||||||
|
|
||||||
int index = 0;
|
int index = 0;
|
||||||
|
|||||||
@@ -118,7 +118,10 @@ inline int _ConvertSMVer2CoresDRV(int major, int minor) {
|
|||||||
{0x90, 128},
|
{0x90, 128},
|
||||||
{0xa0, 128},
|
{0xa0, 128},
|
||||||
{0xa1, 128},
|
{0xa1, 128},
|
||||||
|
{0xa3, 128},
|
||||||
|
{0xb0, 128},
|
||||||
{0xc0, 128},
|
{0xc0, 128},
|
||||||
|
{0xc1, 128},
|
||||||
{-1, -1}};
|
{-1, -1}};
|
||||||
|
|
||||||
int index = 0;
|
int index = 0;
|
||||||
@@ -241,7 +244,7 @@ inline int gpuGetMaxGflopsDeviceIdDRV() {
|
|||||||
}
|
}
|
||||||
|
|
||||||
unsigned long long compute_perf =
|
unsigned long long compute_perf =
|
||||||
(unsigned long long)(multiProcessorCount * sm_per_multiproc *
|
((unsigned long long)multiProcessorCount * sm_per_multiproc *
|
||||||
clockRate);
|
clockRate);
|
||||||
|
|
||||||
if (compute_perf > max_compute_perf) {
|
if (compute_perf > max_compute_perf) {
|
||||||
|
|||||||
@@ -185,25 +185,30 @@ int ipcCreateSocket(ipcHandle *&handle, const char *name,
|
|||||||
return -1;
|
return -1;
|
||||||
}
|
}
|
||||||
|
|
||||||
unlink(name);
|
char path_name[50];
|
||||||
bzero(&servaddr, sizeof(servaddr));
|
|
||||||
|
// Create unique name for the socket with path if SOCK_FOLDER is set.
|
||||||
|
sprintf(path_name, "%s/%u", getSocketFolder().c_str(), getpid());
|
||||||
|
|
||||||
|
unlink(path_name);
|
||||||
|
memset(&servaddr, 0, sizeof(servaddr));
|
||||||
servaddr.sun_family = AF_UNIX;
|
servaddr.sun_family = AF_UNIX;
|
||||||
|
|
||||||
size_t len = strlen(name);
|
size_t len = strlen(path_name);
|
||||||
if (len > (sizeof(servaddr.sun_path) - 1)) {
|
if (len > (sizeof(servaddr.sun_path) - 1)) {
|
||||||
perror("IPC failure: Cannot bind provided name to socket. Name too large");
|
perror("IPC failure: Cannot bind provided name to socket. Name too large");
|
||||||
return -1;
|
return -1;
|
||||||
}
|
}
|
||||||
|
|
||||||
strncpy(servaddr.sun_path, name, len);
|
strncpy(servaddr.sun_path, path_name, len);
|
||||||
|
|
||||||
if (bind(server_fd, (struct sockaddr *)&servaddr, SUN_LEN(&servaddr)) < 0) {
|
if (bind(server_fd, (struct sockaddr *)&servaddr, SUN_LEN(&servaddr)) < 0) {
|
||||||
perror("IPC failure: Binding socket failed");
|
perror("IPC failure: Binding socket failed");
|
||||||
return -1;
|
return -1;
|
||||||
}
|
}
|
||||||
|
|
||||||
handle->socketName = new char[strlen(name) + 1];
|
handle->socketName = new char[strlen(path_name) + 1];
|
||||||
strcpy(handle->socketName, name);
|
strcpy(handle->socketName, path_name);
|
||||||
handle->socket = server_fd;
|
handle->socket = server_fd;
|
||||||
return 0;
|
return 0;
|
||||||
}
|
}
|
||||||
@@ -219,13 +224,13 @@ int ipcOpenSocket(ipcHandle *&handle) {
|
|||||||
perror("IPC failure:Socket creation error");
|
perror("IPC failure:Socket creation error");
|
||||||
return -1;
|
return -1;
|
||||||
}
|
}
|
||||||
|
|
||||||
bzero(&cliaddr, sizeof(cliaddr));
|
memset(&cliaddr, 0, sizeof(cliaddr));
|
||||||
cliaddr.sun_family = AF_UNIX;
|
cliaddr.sun_family = AF_UNIX;
|
||||||
char temp[10];
|
char temp[50];
|
||||||
|
|
||||||
// Create unique name for the socket.
|
// Create unique name for the socket with path if SOCK_FOLDER is set.
|
||||||
sprintf(temp, "%u", getpid());
|
sprintf(temp, "%s/%u", getSocketFolder().c_str(), getpid());
|
||||||
|
|
||||||
strcpy(cliaddr.sun_path, temp);
|
strcpy(cliaddr.sun_path, temp);
|
||||||
if (bind(sock, (struct sockaddr *)&cliaddr, sizeof(cliaddr)) < 0) {
|
if (bind(sock, (struct sockaddr *)&cliaddr, sizeof(cliaddr)) < 0) {
|
||||||
@@ -326,7 +331,7 @@ int ipcSendDataToServer(ipcHandle *handle, const char *serverName,
|
|||||||
ssize_t sendResult;
|
ssize_t sendResult;
|
||||||
struct sockaddr_un serveraddr;
|
struct sockaddr_un serveraddr;
|
||||||
|
|
||||||
bzero(&serveraddr, sizeof(serveraddr));
|
memset(&serveraddr, 0, sizeof(serveraddr));
|
||||||
serveraddr.sun_family = AF_UNIX;
|
serveraddr.sun_family = AF_UNIX;
|
||||||
strncpy(serveraddr.sun_path, serverName, sizeof(serveraddr.sun_path) - 1);
|
strncpy(serveraddr.sun_path, serverName, sizeof(serveraddr.sun_path) - 1);
|
||||||
|
|
||||||
@@ -359,10 +364,10 @@ int ipcSendShareableHandle(ipcHandle *handle,
|
|||||||
socklen_t len = sizeof(cliaddr);
|
socklen_t len = sizeof(cliaddr);
|
||||||
|
|
||||||
// Construct client address to send this SHareable handle to
|
// Construct client address to send this SHareable handle to
|
||||||
bzero(&cliaddr, sizeof(cliaddr));
|
memset(&cliaddr, 0, sizeof(cliaddr));
|
||||||
cliaddr.sun_family = AF_UNIX;
|
cliaddr.sun_family = AF_UNIX;
|
||||||
char temp[10];
|
char temp[20];
|
||||||
sprintf(temp, "%u", process);
|
sprintf(temp, "%s/%u", getSocketFolder().c_str(), process);
|
||||||
strcpy(cliaddr.sun_path, temp);
|
strcpy(cliaddr.sun_path, temp);
|
||||||
len = sizeof(cliaddr);
|
len = sizeof(cliaddr);
|
||||||
|
|
||||||
|
|||||||
@@ -54,6 +54,27 @@
|
|||||||
#endif
|
#endif
|
||||||
#include <vector>
|
#include <vector>
|
||||||
|
|
||||||
|
// The Unix domain sockets creating folder on QNX has been restricted to qnx6-mounted directories since QNX SDP 8.0.3.
|
||||||
|
#if defined(__QNX__)
|
||||||
|
#include <string>
|
||||||
|
inline std::string getSocketFolder() {
|
||||||
|
return "/storage";
|
||||||
|
}
|
||||||
|
// Simple filesystem compatibility for GCC 8.x
|
||||||
|
#elif defined(__GNUC__) && __GNUC__ < 9
|
||||||
|
#include <cstdlib>
|
||||||
|
#include <string>
|
||||||
|
inline std::string getSocketFolder() {
|
||||||
|
const char* tmpdir = std::getenv("TMPDIR");
|
||||||
|
return tmpdir ? std::string(tmpdir) : "/tmp";
|
||||||
|
}
|
||||||
|
#else
|
||||||
|
#include <filesystem>
|
||||||
|
inline std::string getSocketFolder() {
|
||||||
|
return std::filesystem::temp_directory_path().string();
|
||||||
|
}
|
||||||
|
#endif
|
||||||
|
|
||||||
typedef struct sharedMemoryInfo_st {
|
typedef struct sharedMemoryInfo_st {
|
||||||
void *addr;
|
void *addr;
|
||||||
size_t size;
|
size_t size;
|
||||||
|
|||||||
@@ -400,9 +400,9 @@ int getInputDir(std::string &input_dir, const char *executable_path) {
|
|||||||
std::string pathname = "";
|
std::string pathname = "";
|
||||||
const char *searchPath[] = {
|
const char *searchPath[] = {
|
||||||
"./images",
|
"./images",
|
||||||
"../../../../Samples/4_CUDA_Libraries/<executable_name>/images",
|
"../../../../cpp/4_CUDA_Libraries/<executable_name>/images",
|
||||||
"../../../Samples/4_CUDA_Libraries/<executable_name>/images",
|
"../../../cpp/4_CUDA_Libraries/<executable_name>/images",
|
||||||
"../../Samples/4_CUDA_Libraries/<executable_name>/images"};
|
"../../cpp/4_CUDA_Libraries/<executable_name>/images"};
|
||||||
|
|
||||||
for (unsigned int i = 0; i < sizeof(searchPath) / sizeof(char *); ++i) {
|
for (unsigned int i = 0; i < sizeof(searchPath) / sizeof(char *); ++i) {
|
||||||
std::string pathname(searchPath[i]);
|
std::string pathname(searchPath[i]);
|
||||||
|
|||||||
@@ -288,73 +288,79 @@ inline char *sdkFindFilePath(const char *filename,
|
|||||||
"./", // same dir
|
"./", // same dir
|
||||||
"./data/", // same dir
|
"./data/", // same dir
|
||||||
|
|
||||||
"../../../../Samples/<executable_name>/", // up 4 in tree
|
"../../../../cpp/<executable_name>/", // up 4 in tree
|
||||||
"../../../Samples/<executable_name>/", // up 3 in tree
|
"../../../cpp/<executable_name>/", // up 3 in tree
|
||||||
"../../Samples/<executable_name>/", // up 2 in tree
|
"../../cpp/<executable_name>/", // up 2 in tree
|
||||||
|
|
||||||
"../../../../Samples/<executable_name>/data/", // up 4 in tree
|
"../../../../cpp/<executable_name>/data/", // up 4 in tree
|
||||||
"../../../Samples/<executable_name>/data/", // up 3 in tree
|
"../../../cpp/<executable_name>/data/", // up 3 in tree
|
||||||
"../../Samples/<executable_name>/data/", // up 2 in tree
|
"../../cpp/<executable_name>/data/", // up 2 in tree
|
||||||
|
|
||||||
"../../../../Samples/0_Introduction/<executable_name>/", // up 4 in tree
|
"../../../../cpp/0_Introduction/<executable_name>/", // up 4 in tree
|
||||||
"../../../Samples/0_Introduction/<executable_name>/", // up 3 in tree
|
"../../../cpp/0_Introduction/<executable_name>/", // up 3 in tree
|
||||||
"../../Samples/0_Introduction/<executable_name>/", // up 2 in tree
|
"../../cpp/0_Introduction/<executable_name>/", // up 2 in tree
|
||||||
|
|
||||||
"../../../../Samples/1_Utilities/<executable_name>/", // up 4 in tree
|
"../../../../cpp/1_Utilities/<executable_name>/", // up 4 in tree
|
||||||
"../../../Samples/1_Utilities/<executable_name>/", // up 3 in tree
|
"../../../cpp/1_Utilities/<executable_name>/", // up 3 in tree
|
||||||
"../../Samples/1_Utilities/<executable_name>/", // up 2 in tree
|
"../../cpp/1_Utilities/<executable_name>/", // up 2 in tree
|
||||||
|
|
||||||
"../../../../Samples/2_Concepts_and_Techniques/<executable_name>/", // up 4 in tree
|
"../../../../cpp/2_Concepts_and_Techniques/<executable_name>/", // up 4 in tree
|
||||||
"../../../Samples/2_Concepts_and_Techniques/<executable_name>/", // up 3 in tree
|
"../../../cpp/2_Concepts_and_Techniques/<executable_name>/", // up 3 in tree
|
||||||
"../../Samples/2_Concepts_and_Techniques/<executable_name>/", // up 2 in tree
|
"../../cpp/2_Concepts_and_Techniques/<executable_name>/", // up 2 in tree
|
||||||
|
|
||||||
"../../../../Samples/3_CUDA_Features/<executable_name>/", // up 4 in tree
|
"../../../../cpp/3_CUDA_Features/<executable_name>/", // up 4 in tree
|
||||||
"../../../Samples/3_CUDA_Features/<executable_name>/", // up 3 in tree
|
"../../../cpp/3_CUDA_Features/<executable_name>/", // up 3 in tree
|
||||||
"../../Samples/3_CUDA_Features/<executable_name>/", // up 2 in tree
|
"../../cpp/3_CUDA_Features/<executable_name>/", // up 2 in tree
|
||||||
|
|
||||||
"../../../../Samples/4_CUDA_Libraries/<executable_name>/", // up 4 in tree
|
"../../../../cpp/4_CUDA_Libraries/<executable_name>/", // up 4 in tree
|
||||||
"../../../Samples/4_CUDA_Libraries/<executable_name>/", // up 3 in tree
|
"../../../cpp/4_CUDA_Libraries/<executable_name>/", // up 3 in tree
|
||||||
"../../Samples/4_CUDA_Libraries/<executable_name>/", // up 2 in tree
|
"../../cpp/4_CUDA_Libraries/<executable_name>/", // up 2 in tree
|
||||||
|
|
||||||
"../../../../Samples/5_Domain_Specific/<executable_name>/", // up 4 in tree
|
"../../../../cpp/5_Domain_Specific/<executable_name>/", // up 4 in tree
|
||||||
"../../../Samples/5_Domain_Specific/<executable_name>/", // up 3 in tree
|
"../../../cpp/5_Domain_Specific/<executable_name>/", // up 3 in tree
|
||||||
"../../Samples/5_Domain_Specific/<executable_name>/", // up 2 in tree
|
"../../cpp/5_Domain_Specific/<executable_name>/", // up 2 in tree
|
||||||
|
|
||||||
"../../../../Samples/6_Performance/<executable_name>/", // up 4 in tree
|
"../../../../cpp/6_Performance/<executable_name>/", // up 4 in tree
|
||||||
"../../../Samples/6_Performance/<executable_name>/", // up 3 in tree
|
"../../../cpp/6_Performance/<executable_name>/", // up 3 in tree
|
||||||
"../../Samples/6_Performance/<executable_name>/", // up 2 in tree
|
"../../cpp/6_Performance/<executable_name>/", // up 2 in tree
|
||||||
|
|
||||||
"../../../../Samples/0_Introduction/<executable_name>/data/", // up 4 in tree
|
"../../../../cpp/0_Introduction/<executable_name>/data/", // up 4 in tree
|
||||||
"../../../Samples/0_Introduction/<executable_name>/data/", // up 3 in tree
|
"../../../cpp/0_Introduction/<executable_name>/data/", // up 3 in tree
|
||||||
"../../Samples/0_Introduction/<executable_name>/data/", // up 2 in tree
|
"../../cpp/0_Introduction/<executable_name>/data/", // up 2 in tree
|
||||||
|
|
||||||
"../../../../Samples/1_Utilities/<executable_name>/data/", // up 4 in tree
|
"../../../../cpp/1_Utilities/<executable_name>/data/", // up 4 in tree
|
||||||
"../../../Samples/1_Utilities/<executable_name>/data/", // up 3 in tree
|
"../../../cpp/1_Utilities/<executable_name>/data/", // up 3 in tree
|
||||||
"../../Samples/1_Utilities/<executable_name>/data/", // up 2 in tree
|
"../../cpp/1_Utilities/<executable_name>/data/", // up 2 in tree
|
||||||
|
|
||||||
"../../../../Samples/2_Concepts_and_Techniques/<executable_name>/data/", // up 4 in tree
|
"../../../../cpp/2_Concepts_and_Techniques/<executable_name>/data/", // up 4 in tree
|
||||||
"../../../Samples/2_Concepts_and_Techniques/<executable_name>/data/", // up 3 in tree
|
"../../../cpp/2_Concepts_and_Techniques/<executable_name>/data/", // up 3 in tree
|
||||||
"../../Samples/2_Concepts_and_Techniques/<executable_name>/data/", // up 2 in tree
|
"../../cpp/2_Concepts_and_Techniques/<executable_name>/data/", // up 2 in tree
|
||||||
|
|
||||||
"../../../../Samples/3_CUDA_Features/<executable_name>/data/", // up 4 in tree
|
"../../../../cpp/3_CUDA_Features/<executable_name>/data/", // up 4 in tree
|
||||||
"../../../Samples/3_CUDA_Features/<executable_name>/data/", // up 3 in tree
|
"../../../cpp/3_CUDA_Features/<executable_name>/data/", // up 3 in tree
|
||||||
"../../Samples/3_CUDA_Features/<executable_name>/data/", // up 2 in tree
|
"../../cpp/3_CUDA_Features/<executable_name>/data/", // up 2 in tree
|
||||||
|
|
||||||
"../../../../Samples/4_CUDA_Libraries/<executable_name>/data/", // up 4 in tree
|
"../../../../cpp/4_CUDA_Libraries/<executable_name>/data/", // up 4 in tree
|
||||||
"../../../Samples/4_CUDA_Libraries/<executable_name>/data/", // up 3 in tree
|
"../../../cpp/4_CUDA_Libraries/<executable_name>/data/", // up 3 in tree
|
||||||
"../../Samples/4_CUDA_Libraries/<executable_name>/data/", // up 2 in tree
|
"../../cpp/4_CUDA_Libraries/<executable_name>/data/", // up 2 in tree
|
||||||
|
|
||||||
"../../../../Samples/5_Domain_Specific/<executable_name>/data/", // up 4 in tree
|
"../../../../cpp/5_Domain_Specific/<executable_name>/data/", // up 4 in tree
|
||||||
"../../../Samples/5_Domain_Specific/<executable_name>/data/", // up 3 in tree
|
"../../../cpp/5_Domain_Specific/<executable_name>/data/", // up 3 in tree
|
||||||
"../../Samples/5_Domain_Specific/<executable_name>/data/", // up 2 in tree
|
"../../cpp/5_Domain_Specific/<executable_name>/data/", // up 2 in tree
|
||||||
|
|
||||||
"../../../../Samples/6_Performance/<executable_name>/data/", // up 4 in tree
|
"../../../../cpp/6_Performance/<executable_name>/data/", // up 4 in tree
|
||||||
"../../../Samples/6_Performance/<executable_name>/data/", // up 3 in tree
|
"../../../cpp/6_Performance/<executable_name>/data/", // up 3 in tree
|
||||||
"../../Samples/6_Performance/<executable_name>/data/", // up 2 in tree
|
"../../cpp/6_Performance/<executable_name>/data/", // up 2 in tree
|
||||||
|
|
||||||
"../../../../Common/data/", // up 4 in tree
|
"../../../../Common/data/", // up 4 in tree
|
||||||
"../../../Common/data/", // up 3 in tree
|
"../../../Common/data/", // up 3 in tree
|
||||||
"../../Common/data/" // up 2 in tree
|
"../../Common/data/", // up 2 in tree
|
||||||
|
|
||||||
|
"../../../../cpp/9_CUDA_Tile/<executable_name>/", // up 4 in tree
|
||||||
|
"../../../cpp/9_CUDA_Tile/<executable_name>/", // up 3 in tree
|
||||||
|
"../../cpp/9_CUDA_Tile/<executable_name>/", // up 2 in tree
|
||||||
|
"../cpp/9_CUDA_Tile/<executable_name>/", // up 1 in tree
|
||||||
|
"./cpp/9_CUDA_Tile/<executable_name>/" // up 0 in tree
|
||||||
};
|
};
|
||||||
|
|
||||||
// Extract the executable name
|
// Extract the executable name
|
||||||
|
|||||||
@@ -258,7 +258,7 @@ namespace nv
|
|||||||
s[2] = &r3[0];
|
s[2] = &r3[0];
|
||||||
s[3] = &r4[0];
|
s[3] = &r4[0];
|
||||||
|
|
||||||
register int i,j,p,jj;
|
int i,j,p,jj;
|
||||||
|
|
||||||
for (i=0; i<4; i++)
|
for (i=0; i<4; i++)
|
||||||
{
|
{
|
||||||
|
|||||||
@@ -187,6 +187,7 @@ CUmodule loadCUBIN(char *cubin, int argc, char **argv) {
|
|||||||
CUcontext context;
|
CUcontext context;
|
||||||
int major = 0, minor = 0;
|
int major = 0, minor = 0;
|
||||||
char deviceName[256];
|
char deviceName[256];
|
||||||
|
CUctxCreateParams ctxCreateParams = {};
|
||||||
|
|
||||||
// Picks the best CUDA device available
|
// Picks the best CUDA device available
|
||||||
CUdevice cuDevice = findCudaDeviceDRV(argc, (const char **)argv);
|
CUdevice cuDevice = findCudaDeviceDRV(argc, (const char **)argv);
|
||||||
@@ -200,7 +201,7 @@ CUmodule loadCUBIN(char *cubin, int argc, char **argv) {
|
|||||||
printf("> GPU Device has SM %d.%d compute capability\n", major, minor);
|
printf("> GPU Device has SM %d.%d compute capability\n", major, minor);
|
||||||
|
|
||||||
checkCudaErrors(cuInit(0));
|
checkCudaErrors(cuInit(0));
|
||||||
checkCudaErrors(cuCtxCreate(&context, 0, cuDevice));
|
checkCudaErrors(cuCtxCreate(&context, &ctxCreateParams, 0, cuDevice));
|
||||||
|
|
||||||
checkCudaErrors(cuModuleLoadData(&module, cubin));
|
checkCudaErrors(cuModuleLoadData(&module, cubin));
|
||||||
free(cubin);
|
free(cubin);
|
||||||
|
|||||||
395
README.md
395
README.md
@@ -1,6 +1,6 @@
|
|||||||
# CUDA Samples
|
# CUDA Samples
|
||||||
|
|
||||||
Samples for CUDA Developers which demonstrates features in CUDA Toolkit. This version supports [CUDA Toolkit 12.6](https://developer.nvidia.com/cuda-downloads).
|
Samples for CUDA Developers which demonstrates features in CUDA Toolkit. This version supports [CUDA Toolkit 13.3](https://developer.nvidia.com/cuda-downloads).
|
||||||
|
|
||||||
## Release Notes
|
## Release Notes
|
||||||
|
|
||||||
@@ -14,7 +14,7 @@ This section describes the release notes for the CUDA Samples on GitHub only.
|
|||||||
|
|
||||||
### Prerequisites
|
### Prerequisites
|
||||||
|
|
||||||
Download and install the [CUDA Toolkit 12.8](https://developer.nvidia.com/cuda-downloads) for your corresponding platform.
|
Download and install the [CUDA Toolkit](https://developer.nvidia.com/cuda-downloads) for your corresponding platform.
|
||||||
For system requirements and installation instructions of cuda toolkit, please refer to the [Linux Installation Guide](http://docs.nvidia.com/cuda/cuda-installation-guide-linux/), and the [Windows Installation Guide](http://docs.nvidia.com/cuda/cuda-installation-guide-microsoft-windows/index.html).
|
For system requirements and installation instructions of cuda toolkit, please refer to the [Linux Installation Guide](http://docs.nvidia.com/cuda/cuda-installation-guide-linux/), and the [Windows Installation Guide](http://docs.nvidia.com/cuda/cuda-installation-guide-microsoft-windows/index.html).
|
||||||
|
|
||||||
### Getting the CUDA Samples
|
### Getting the CUDA Samples
|
||||||
@@ -72,6 +72,17 @@ Open the generated solution file CUDA_Samples.sln in Visual Studio. Build the sa
|
|||||||
|
|
||||||
Run the samples from the output directories specified in Visual Studio.
|
Run the samples from the output directories specified in Visual Studio.
|
||||||
|
|
||||||
|
### Enabling On-GPU Debugging
|
||||||
|
|
||||||
|
NVIDIA GPUs support on-GPU debugging through cuda-gdb. Enabling this may significantly affect application performance as certain compiler optimizations are disabled
|
||||||
|
in this configuration, hence it's not on by default. Enablement of on-device debugging is controlled via the `-G` switch to nvcc.
|
||||||
|
|
||||||
|
To enable cuda-gdb for samples builds, define the `ENABLE_CUDA_DEBUG` flag on the CMake command line. For example:
|
||||||
|
|
||||||
|
```
|
||||||
|
cmake -DENABLE_CUDA_DEBUG=True ...
|
||||||
|
```
|
||||||
|
|
||||||
### Platform-Specific Samples
|
### Platform-Specific Samples
|
||||||
|
|
||||||
Some CUDA samples are specific to certain platforms, and require passing flags into CMake to enable. In particular, we define the following platform-specific flags:
|
Some CUDA samples are specific to certain platforms, and require passing flags into CMake to enable. In particular, we define the following platform-specific flags:
|
||||||
@@ -94,9 +105,9 @@ Navigate to the root of the cloned repository and create a build directory:
|
|||||||
```
|
```
|
||||||
mkdir build && cd build
|
mkdir build && cd build
|
||||||
```
|
```
|
||||||
Configure the project with CMake, specifying the Tegra toolchain file:
|
Configure the project with CMake, specifying the Tegra toolchain file. And you can use -DTARGET_FS to point to the target file system root path for necessary include and library files:
|
||||||
```
|
```
|
||||||
cmake .. -DCMAKE_TOOLCHAIN_FILE=/path/to/tegra/toolchain.cmake
|
cmake .. -DCMAKE_TOOLCHAIN_FILE=../cmake/toolchains/toolchain-aarch64-linux.cmake -DTARGET_FS=/path/to/target/system/file/system
|
||||||
```
|
```
|
||||||
Build the samples:
|
Build the samples:
|
||||||
```
|
```
|
||||||
@@ -104,56 +115,370 @@ make -j$(nproc)
|
|||||||
```
|
```
|
||||||
Transfer the built binaries to the Tegra device and execute them there.
|
Transfer the built binaries to the Tegra device and execute them there.
|
||||||
|
|
||||||
### Building for Automotive Linux Platforms
|
|
||||||
|
|
||||||
These platforms require additional information to be passed to CMake on the command line to ensure proper resolution of all necessary include and library files.
|
### Cross Building for Automotive Linux Platforms from the DriveOS Docker containers
|
||||||
Instead of being in the default location, `/usr/local/cuda/include` or `/usr/local/cuda/lib64`, you must point to architecture-specific paths:
|
|
||||||
|
|
||||||
`/usr/local/cuda/<ARCH>/targets/aarch64-linux/lib`
|
To build CUDA samples to the target platform from the DriveOS Docker containers, use the following instructions.
|
||||||
and
|
|
||||||
`/usr/local/cuda-12.8/<ARCH>/include`
|
|
||||||
|
|
||||||
An example build might look like this:
|
Mount the target Root Filesystem (RFS) in the container so that the CUDA cmake process has the correct paths to CUDA and other system libraries required to build the samples.
|
||||||
|
|
||||||
|
Create a temporary directory, `<temp>` is any temporary directory of your choosing, for example, you can use `/drive/temp`:
|
||||||
|
|
||||||
|
```
|
||||||
|
$ mkdir /drive/<temp>
|
||||||
|
```
|
||||||
|
|
||||||
|
Mount the filesystem by running the following command:
|
||||||
|
|
||||||
|
```
|
||||||
|
$ mount /drive/drive-linux/filesystem/targetfs-images/dev_nsr_desktop_ubuntu-24.04_thor_rfs.img /drive/temp
|
||||||
|
```
|
||||||
|
|
||||||
|
Configure the project by running the following cmake command:
|
||||||
|
|
||||||
|
```
|
||||||
|
$ mkdir build && cd build
|
||||||
|
$ cmake .. -DBUILD_TEGRA=True \
|
||||||
|
-DCMAKE_CUDA_COMPILER=/usr/local/cuda/bin/nvcc \
|
||||||
|
-DCMAKE_TOOLCHAIN_FILE=../cmake/toolchains/toolchain-aarch64-linux.cmake \
|
||||||
|
-DTARGET_FS=/drive/temp \
|
||||||
|
-DCMAKE_LIBRARY_PATH=/drive/temp/usr/local/cuda-13.1/thor/lib64/ \
|
||||||
|
-DCMAKE_INCLUDE_PATH=/drive/temp/usr/local/cuda-13.1/thor/include/
|
||||||
|
```
|
||||||
|
|
||||||
|
Please note that the following libraries are not pre-installed in the DriveOS dev-nsr target filesystem:
|
||||||
|
* libdrm-dev
|
||||||
|
* Vulkan
|
||||||
|
|
||||||
|
This causes the cmake command to throw errors related to the missing files, and as a result, the related samples will not build in later steps. This issue will be addressed in a future DriveOS release.
|
||||||
|
|
||||||
|
To build the samples with ignore the error mentioned above, you can use `--ignore-errors`/`--keep-going` or comment out the comment out the corresponding `add_subdirectory` command in the CMakeLists.txt in the parent folder for the samples requiring Vulkan and libdrm_dev:
|
||||||
|
|
||||||
|
```
|
||||||
|
$ make -j$(nproc) --ignore-errors # or --keep-going
|
||||||
|
```
|
||||||
|
|
||||||
|
```
|
||||||
|
# In cpp/5_Domain_Specific/CMakeList.txt
|
||||||
|
# add_subdirectory(simpleGL)
|
||||||
|
# add_subdirectory(simpleVulkan)
|
||||||
|
# add_subdirectory(simpleVulkanMMAP)
|
||||||
|
|
||||||
|
# In cpp/8_Platform_Specific/Tegra/CMakeList.txt
|
||||||
|
# add_subdirectory(simpleGLES_EGLOutput)
|
||||||
|
```
|
||||||
|
|
||||||
|
### QNX
|
||||||
|
|
||||||
|
Cross-compilation for QNX with CMake is supported in the CUDA 13.0 samples release and newer. An example build for
|
||||||
|
the Tegra Thor QNX platform might look like this:
|
||||||
|
|
||||||
```
|
```
|
||||||
$ mkdir build
|
$ mkdir build
|
||||||
$ cd build
|
$ cd build
|
||||||
|
|
||||||
$ cmake -DCMAKE_CUDA_COMPILER=/usr/local/cuda/bin/nvcc -DCMAKE_LIBRARY_PATH=/usr/local/cuda/orin/lib64/ -DCMAKE_INCLUDE_PATH=/usr/local/cuda/orin/include -DBUILD_TEGRA=True ..
|
QNX_HOST=/path/to/qnx/host \
|
||||||
|
QNX_TARGET=/path/to/qnx/target \
|
||||||
|
cmake .. \
|
||||||
|
-DBUILD_TEGRA=True \
|
||||||
|
-DCMAKE_CUDA_COMPILER=/usr/local/cuda-13.3/bin/nvcc \
|
||||||
|
-DCMAKE_TOOLCHAIN_FILE=../cmake/toolchains/toolchain-aarch64-qnx.cmake \
|
||||||
|
-DCMAKE_LIBRARY_PATH=/usr/local/cuda-13.3/thor/targets/aarch64-qnx/lib/stubs/ \
|
||||||
|
-DCMAKE_INCLUDE_PATH=/usr/local/cuda-13.3/thor/targets/aarch64-qnx/include/
|
||||||
```
|
```
|
||||||
|
|
||||||
### QNX
|
### Forward Compatibility
|
||||||
|
|
||||||
Note that in the current branch sample cross-compilation for QNX is not fully validated. This placeholder will be updated in the
|
To build samples with new CUDA Toolkit(CUDA 13.0 or later) and UMD(Version 580 or later) and old KMD(Version 550 or earlier),you need to set the `CMAKE_PREFIX_PATH` for using new driver library, the command might like this:
|
||||||
near future with QNX cross-compilation instructions. In the meantime, if you want to cross-compile for QNX please check out one
|
|
||||||
of the previous tags prior to the CMake build system transition in 12.8.
|
```
|
||||||
|
cmake -DCMAKE_PREFIX_PATH=/usr/local/cuda/lib64/stubs/ ..
|
||||||
|
```
|
||||||
|
|
||||||
|
## CUDA Python samples
|
||||||
|
|
||||||
|
The repository includes **Python** examples under the [`python/`](./python) directory. **These samples are cuda.core–focused:** they use [CUDA Python](https://nvidia.github.io/cuda-python/), with [`cuda.core`](https://nvidia.github.io/cuda-python/cuda-core/latest/) for devices, programs, launches, and memory, alongside NumPy, CuPy, or framework interop where each sample notes.
|
||||||
|
|
||||||
|
**Layout (same themes as the C++ samples):**
|
||||||
|
|
||||||
|
| Directory | Contents |
|
||||||
|
|-----------|----------|
|
||||||
|
| `python/1_GettingStarted/` | Introductory scripts (e.g. `vectorAdd`, `deviceQuery`, `systemInfo`, image blur with unified memory, NumPy vs CuPy). |
|
||||||
|
| `python/2_CoreConcepts/` | Algorithms and techniques (e.g. reductions, histograms, FFT, stream overlap, `memoryResources`, `cudaGraphs`, `jitLtoLinking`, `tmaTensorMap`). |
|
||||||
|
| `python/3_FrameworkInterop/` | Integration with PyTorch and TensorFlow. |
|
||||||
|
| `python/4_DistributedComputing/` | Multi-GPU, peer-to-peer, and IPC patterns (`ipcMemoryPool`). |
|
||||||
|
| `python/Utilities/` | Shared helpers imported by some samples. |
|
||||||
|
|
||||||
|
**How to run:** The top-level CMake build does **not** compile these samples. For each sample, use a Python 3.10+ environment with a matching [CUDA Toolkit](https://developer.nvidia.com/cuda-downloads) (the samples target CUDA 13.x and document exact package pins in `requirements.txt`):
|
||||||
|
|
||||||
|
```bash
|
||||||
|
cd python/<category>/<sampleName>
|
||||||
|
pip install -r requirements.txt
|
||||||
|
python <sampleScript>.py
|
||||||
|
```
|
||||||
|
|
||||||
|
Use each sample’s `README.md` for prerequisites, CLI options, and expected output.
|
||||||
|
|
||||||
|
## Install Samples
|
||||||
|
|
||||||
|
### Installation Path Structure
|
||||||
|
|
||||||
|
The installation system automatically organizes samples into a structured directory layout based on:
|
||||||
|
- **Target Architecture**: ${CMAKE_SYSTEM_PROCESSOR}, e.g. `x64`, `aarch64`, `amd64`, etc.
|
||||||
|
- **Target OS**: `linux`, `windows`, `darwin`, `qnx`
|
||||||
|
- **Build Type**: `release`, `debug`, etc.
|
||||||
|
|
||||||
|
The default installation path is: `build/bin/${TARGET_ARCH}/${TARGET_OS}/${BUILD_TYPE}`
|
||||||
|
|
||||||
|
**Examples:**
|
||||||
|
- Linux x86_64 Release: `build/bin/x64/linux/release`
|
||||||
|
- Linux aarch64 Release: `build/bin/aarch64/linux/release`
|
||||||
|
- Windows amd64 Release: `build/bin/amd64/windows/release`
|
||||||
|
|
||||||
|
### Customizing Installation Paths
|
||||||
|
|
||||||
|
You can customize the installation location using CMake variables during the configuration step:
|
||||||
|
|
||||||
|
- `CMAKE_INSTALL_PREFIX`: Changes the root installation directory (default: `build/bin`)
|
||||||
|
```
|
||||||
|
cmake -DCMAKE_INSTALL_PREFIX=/custom/path ..
|
||||||
|
```
|
||||||
|
This will install to: `/custom/path/${TARGET_ARCH}/${TARGET_OS}/${BUILD_TYPE}`
|
||||||
|
|
||||||
|
- `CUDA_SAMPLES_INSTALL_DIR`: Specifies the exact final installation directory (overrides the structured path)
|
||||||
|
```
|
||||||
|
cmake -DCUDA_SAMPLES_INSTALL_DIR=/exact/install/path ..
|
||||||
|
```
|
||||||
|
|
||||||
|
### Install Samples on Linux
|
||||||
|
|
||||||
|
**Prerequisites:** You must first configure the project with CMake as described in the [Building CUDA Samples - Linux](#linux) or [Building]section.
|
||||||
|
|
||||||
|
After configuring and building, install the samples:
|
||||||
|
|
||||||
|
```
|
||||||
|
cd build/
|
||||||
|
make install
|
||||||
|
```
|
||||||
|
|
||||||
|
### Install Samples on Windows
|
||||||
|
|
||||||
|
**Prerequisites:** You must first configure the project with CMake as described in the [Building CUDA Samples - Windows](#windows) section.
|
||||||
|
|
||||||
|
#### Using Command Line
|
||||||
|
|
||||||
|
After configuring with CMake, build and install from the `x64 Native Tools Command Prompt for VS`:
|
||||||
|
|
||||||
|
```cmd
|
||||||
|
cd build
|
||||||
|
cmake --build . --config Release
|
||||||
|
cmake --install . --config Release
|
||||||
|
```
|
||||||
|
|
||||||
|
**Note:** Replace `Release` with `Debug` if you want to install debug builds. For multi-configuration generators (like Visual Studio), the `--config` flag determines which build type to install.
|
||||||
|
|
||||||
|
#### Using Visual Studio IDE
|
||||||
|
|
||||||
|
Alternatively, open the generated solution file `CUDA_Samples.sln` in Visual Studio:
|
||||||
|
1. Select the desired configuration (`Release` or `Debug`)
|
||||||
|
2. Build the solution (F7 or Build > Build Solution)
|
||||||
|
3. Right-click on the `INSTALL` target under `CMakePredefinedTargets` in Solution Explorer
|
||||||
|
4. Select "Build"
|
||||||
|
|
||||||
|
## Running All Samples as Tests
|
||||||
|
|
||||||
|
It's important to note that the CUDA samples are _not_ intended as a validation suite for CUDA. They do not cover corner cases, they do not completely cover the
|
||||||
|
runtime and driver APIs, are not intended for performance benchmarking, etc. That said, it can sometimes be useful to run all of the samples as a quick sanity check and
|
||||||
|
we provide a script to do so, `run_tests.py`.
|
||||||
|
|
||||||
|
This Python3 script finds all executables in a subdirectory you choose, matching application names with command line arguments specified in `test_args.json`. It accepts
|
||||||
|
the following command line arguments:
|
||||||
|
|
||||||
|
| Switch | Purpose | Example |
|
||||||
|
| ---------- | -------------------------------------------------------------------------------------------------------------- | ----------------------- |
|
||||||
|
| --dir | Specify the root directory to search for executables (recursively) | --dir ./build/cpp |
|
||||||
|
| --config | JSON configuration file for executable arguments | --config test_args.json |
|
||||||
|
| --output | Output directory for test results (stdout saved to .txt files - directory will be created if it doesn't exist) | --output ./test |
|
||||||
|
| --args | Global arguments to pass to all executables (not currently used) | --args arg_1 arg_2 ... |
|
||||||
|
| --parallel | Number of applications to execute in parallel. | --parallel 8 |
|
||||||
|
|
||||||
|
|
||||||
|
Application configurations are loaded from `test_args.json` and matched against executable names (discarding the `.exe` extension on Windows).
|
||||||
|
|
||||||
|
The script returns 0 on success, or the first non-zero error code encountered during testing on failure. It will also print a condensed list of samples that failed, if any.
|
||||||
|
|
||||||
|
There are three primary modes of configuration:
|
||||||
|
|
||||||
|
**Skip**
|
||||||
|
|
||||||
|
An executable configured with "skip" will not be executed. These generally rely on having attached graphical displays and are not suited to this kind of automation.
|
||||||
|
|
||||||
|
Configuration example:
|
||||||
|
```json
|
||||||
|
"fluidsGL": {
|
||||||
|
"skip": true
|
||||||
|
}
|
||||||
|
```
|
||||||
|
|
||||||
|
You will see:
|
||||||
|
```
|
||||||
|
Skipping fluidsGL (marked as skip in config)
|
||||||
|
```
|
||||||
|
|
||||||
|
**Single Run**
|
||||||
|
|
||||||
|
For executables to run one time only with arguments, specify each argument as a list entry. Each entry in the JSON file will be appended to the command line, separated
|
||||||
|
by a space.
|
||||||
|
|
||||||
|
All applications execute from their current directory, so all paths are relative to the application's location.
|
||||||
|
|
||||||
|
Note that if an application needs no arguments, this entry is optional. An executable found without a matching entry in the JSON will just run as `./application` from its
|
||||||
|
current directory.
|
||||||
|
|
||||||
|
Configuration example:
|
||||||
|
```json
|
||||||
|
"ptxgen": {
|
||||||
|
"args": [
|
||||||
|
"test.ll",
|
||||||
|
"-arch=compute_75"
|
||||||
|
]
|
||||||
|
}
|
||||||
|
```
|
||||||
|
|
||||||
|
You will see:
|
||||||
|
```
|
||||||
|
Running ptxgen
|
||||||
|
Command: ./ptxgen test.ll -arch=compute_75
|
||||||
|
Test completed with return code 0
|
||||||
|
```
|
||||||
|
|
||||||
|
**Multiple Runs**
|
||||||
|
|
||||||
|
For executables to run multiple times with different command line arguments, specify any number of sets of args within a "runs" list.
|
||||||
|
|
||||||
|
As with single runs, all applications execute from their current directory, so all paths are relative to the application's location.
|
||||||
|
|
||||||
|
Configuration example:
|
||||||
|
```json
|
||||||
|
"recursiveGaussian": {
|
||||||
|
"runs": [
|
||||||
|
{
|
||||||
|
"args": [
|
||||||
|
"-sigma=10",
|
||||||
|
"-file=data/ref_10.ppm"
|
||||||
|
]
|
||||||
|
},
|
||||||
|
{
|
||||||
|
"args": [
|
||||||
|
"-sigma=14",
|
||||||
|
"-file=data/ref_14.ppm"
|
||||||
|
]
|
||||||
|
},
|
||||||
|
{
|
||||||
|
"args": [
|
||||||
|
"-sigma=18",
|
||||||
|
"-file=data/ref_18.ppm"
|
||||||
|
]
|
||||||
|
},
|
||||||
|
{
|
||||||
|
"args": [
|
||||||
|
"-sigma=22",
|
||||||
|
"-file=data/ref_22.ppm"
|
||||||
|
]
|
||||||
|
}
|
||||||
|
]
|
||||||
|
}
|
||||||
|
```
|
||||||
|
|
||||||
|
You will see:
|
||||||
|
```
|
||||||
|
Running recursiveGaussian (run 1/4)
|
||||||
|
Command: ./recursiveGaussian -sigma=10 -file=data/ref_10.ppm
|
||||||
|
Test completed with return code 0
|
||||||
|
Running recursiveGaussian (run 2/4)
|
||||||
|
Command: ./recursiveGaussian -sigma=14 -file=data/ref_14.ppm
|
||||||
|
Test completed with return code 0
|
||||||
|
Running recursiveGaussian (run 3/4)
|
||||||
|
Command: ./recursiveGaussian -sigma=18 -file=data/ref_18.ppm
|
||||||
|
Test completed with return code 0
|
||||||
|
Running recursiveGaussian (run 4/4)
|
||||||
|
Command: ./recursiveGaussian -sigma=22 -file=data/ref_22.ppm
|
||||||
|
Test completed with return code 0
|
||||||
|
```
|
||||||
|
|
||||||
|
### Example Usage
|
||||||
|
|
||||||
|
Here is an example set of commands to build and test all of the samples.
|
||||||
|
|
||||||
|
First, build:
|
||||||
|
```bash
|
||||||
|
mkdir build
|
||||||
|
cd build
|
||||||
|
cmake ..
|
||||||
|
make -j$(nproc)
|
||||||
|
```
|
||||||
|
|
||||||
|
Now, return to the samples root directory and run the test script:
|
||||||
|
```bash
|
||||||
|
cd ..
|
||||||
|
python3 run_tests.py --output ./test --dir ./build/cpp --config test_args.json
|
||||||
|
```
|
||||||
|
|
||||||
|
If all applications run successfully, you will see something similar to this (the specific number of samples will depend on your build type
|
||||||
|
and system configuration):
|
||||||
|
|
||||||
|
```
|
||||||
|
Test Summary:
|
||||||
|
Ran 199 test runs for 180 executables.
|
||||||
|
All test runs passed!
|
||||||
|
```
|
||||||
|
|
||||||
|
If some samples fail, you will see something like this:
|
||||||
|
|
||||||
|
```
|
||||||
|
Test Summary:
|
||||||
|
Ran 199 test runs for 180 executables.
|
||||||
|
Failed runs (2):
|
||||||
|
bicubicTexture (run 1/5): Failed (code 1)
|
||||||
|
Mandelbrot (run 1/2): Failed (code 1)
|
||||||
|
```
|
||||||
|
|
||||||
|
You can inspect the stdout logs in the output directory (generally `APM_<application_name>.txt` or `APM_<application_name>.run<n>.txt`) to help
|
||||||
|
determine what may have gone wrong from the output logs. Please file issues against the samples repository if you believe a sample is failing
|
||||||
|
incorrectly on your system.
|
||||||
|
|
||||||
## Samples list
|
## Samples list
|
||||||
|
|
||||||
### [0. Introduction](./Samples/0_Introduction/README.md)
|
### [0. Introduction](./cpp/0_Introduction/README.md)
|
||||||
Basic CUDA samples for beginners that illustrate key concepts with using CUDA and CUDA runtime APIs.
|
Basic CUDA samples for beginners that illustrate key concepts with using CUDA and CUDA runtime APIs.
|
||||||
|
|
||||||
### [1. Utilities](./Samples/1_Utilities/README.md)
|
### [1. Utilities](./cpp/1_Utilities/README.md)
|
||||||
Utility samples that demonstrate how to query device capabilities and measure GPU/CPU bandwidth.
|
Utility samples that demonstrate how to query device capabilities and measure GPU/CPU bandwidth.
|
||||||
|
|
||||||
### [2. Concepts and Techniques](./Samples/2_Concepts_and_Techniques/README.md)
|
### [2. Concepts and Techniques](./cpp/2_Concepts_and_Techniques/README.md)
|
||||||
Samples that demonstrate CUDA related concepts and common problem solving techniques.
|
Samples that demonstrate CUDA related concepts and common problem solving techniques.
|
||||||
|
|
||||||
### [3. CUDA Features](./Samples/3_CUDA_Features/README.md)
|
### [3. CUDA Features](./cpp/3_CUDA_Features/README.md)
|
||||||
Samples that demonstrate CUDA Features (Cooperative Groups, CUDA Dynamic Parallelism, CUDA Graphs etc).
|
Samples that demonstrate CUDA Features (Cooperative Groups, CUDA Dynamic Parallelism, CUDA Graphs etc).
|
||||||
|
|
||||||
### [4. CUDA Libraries](./Samples/4_CUDA_Libraries/README.md)
|
### [4. CUDA Libraries](./cpp/4_CUDA_Libraries/README.md)
|
||||||
Samples that demonstrate how to use CUDA platform libraries (NPP, NVJPEG, NVGRAPH cuBLAS, cuFFT, cuSPARSE, cuSOLVER and cuRAND).
|
Samples that demonstrate how to use CUDA platform libraries (NPP, NVJPEG, NVGRAPH cuBLAS, cuFFT, cuSPARSE, cuSOLVER and cuRAND).
|
||||||
|
|
||||||
### [5. Domain Specific](./Samples/5_Domain_Specific/README.md)
|
### [5. Domain Specific](./cpp/5_Domain_Specific/README.md)
|
||||||
Samples that are specific to domain (Graphics, Finance, Image Processing).
|
Samples that are specific to domain (Graphics, Finance, Image Processing).
|
||||||
|
|
||||||
### [6. Performance](./Samples/6_Performance/README.md)
|
### [6. Performance](./cpp/6_Performance/README.md)
|
||||||
Samples that demonstrate performance optimization.
|
Samples that demonstrate performance optimization.
|
||||||
|
|
||||||
### [7. libNVVM](./Samples/7_libNVVM/README.md)
|
### [7. libNVVM](./cpp/7_libNVVM/README.md)
|
||||||
Samples that demonstrate the use of libNVVVM and NVVM IR.
|
Samples that demonstrate the use of libNVVVM and NVVM IR.
|
||||||
|
|
||||||
|
### [8. Platform Specific](./cpp/8_Platform_Specific/Tegra/README.md)
|
||||||
|
Samples that are specific to certain platforms (Tegra, cuDLA, NvMedia, NvSci, OpenGL ES).
|
||||||
|
|
||||||
|
### [9. CUDA Tile](./cpp/9_CUDA_Tile/README.md)
|
||||||
|
Samples that demonstrate how to use CUDA Tile C++.
|
||||||
|
|
||||||
## Dependencies
|
## Dependencies
|
||||||
|
|
||||||
Some CUDA Samples rely on third-party applications and/or libraries, or features provided by the CUDA Toolkit and Driver, to either build or execute. These dependencies are listed below.
|
Some CUDA Samples rely on third-party applications and/or libraries, or features provided by the CUDA Toolkit and Driver, to either build or execute. These dependencies are listed below.
|
||||||
@@ -170,7 +495,7 @@ These third-party dependencies are required by some CUDA samples. If available,
|
|||||||
|
|
||||||
FreeImage is an open source imaging library. FreeImage can usually be installed on Linux using your distribution's package manager system. FreeImage can also be downloaded from the FreeImage website.
|
FreeImage is an open source imaging library. FreeImage can usually be installed on Linux using your distribution's package manager system. FreeImage can also be downloaded from the FreeImage website.
|
||||||
|
|
||||||
To set up FreeImage on a Windows system, extract the FreeImage DLL distribution into the folder `../../../Common/FreeImage/Dist/x64` such that it contains the .h and .lib files. Copy the .dll file to the Release/ Debug/ execution folder or pass the FreeImage folder when cmake configuring with the `-DFREEIMAGE_INCLUDE_DIR` and `-DFREEIMAGE_LIBRARY` options.
|
To set up FreeImage on a Windows system, extract the FreeImage DLL distribution into the folder `./Common/FreeImage/Dist/x64` such that it contains the .h and .lib files. Copy the .dll file to the Release/ Debug/ execution folder or pass the FreeImage folder when cmake configuring with the `-DFreeImage_INCLUDE_DIR` and `-DFreeImage_LIBRARY` options.
|
||||||
|
|
||||||
#### Message Passing Interface
|
#### Message Passing Interface
|
||||||
|
|
||||||
@@ -196,18 +521,27 @@ OpenGL is a graphics library used for 2D and 3D rendering. On systems which supp
|
|||||||
|
|
||||||
OpenGL ES is an embedded systems graphics library used for 2D and 3D rendering. On systems which support OpenGL ES, NVIDIA's OpenGL ES implementation is provided with the CUDA Driver.
|
OpenGL ES is an embedded systems graphics library used for 2D and 3D rendering. On systems which support OpenGL ES, NVIDIA's OpenGL ES implementation is provided with the CUDA Driver.
|
||||||
|
|
||||||
|
#### Freeglut
|
||||||
|
|
||||||
|
Freeglut is an open-source software library that serves as a replacement for the original OpenGL Utility Toolkit (GLUT). Its primary purpose is to make it easier for developers to create and manage windows containing OpenGL contexts, as well as handle input from devices like the mouse, keyboard, and joystick, across a wide range of platforms. To set up Freeglut on a Windowson on ARM system, you need to download the source from [Freeglut website](https://freeglut.sourceforge.net/), build freeglut on your system, and copy the freeglut.lib into the folder `./Common/lib/x64` and copy the freeglut.dll file into the `./bin/win64/${BUILD_TYPE}` execution folder.
|
||||||
|
|
||||||
#### Vulkan
|
#### Vulkan
|
||||||
|
|
||||||
Vulkan is a low-overhead, cross-platform 3D graphics and compute API. Vulkan targets high-performance realtime 3D graphics applications such as video games and interactive media across all platforms. On systems which support Vulkan, NVIDIA's Vulkan implementation is provided with the CUDA Driver. For building and running Vulkan applications one needs to install the [Vulkan SDK](https://www.lunarg.com/vulkan-sdk/).
|
Vulkan is a low-overhead, cross-platform 3D graphics and compute API. Vulkan targets high-performance realtime 3D graphics applications such as video games and interactive media across all platforms. On systems which support Vulkan, NVIDIA's Vulkan implementation is provided with the CUDA Driver. For building and running Vulkan applications one needs to install the [Vulkan SDK](https://www.lunarg.com/vulkan-sdk/).
|
||||||
|
|
||||||
|
#### GLEW
|
||||||
|
|
||||||
|
GLEW (OpenGL Extension Wrangler Library) is a cross-platform, open-source C/C++ library designed to simplify the process of using modern OpenGL features and extensions. Its main function is to dynamically load OpenGL function pointers at runtime, allowing developers to access both core OpenGL functions and additional features provided by hardware vendors, known as extensions. To set up GLEW on a Windows on ARM system, you need to download the source from [GLEW website](https://glew.sourceforge.net/), build GLEW on your system, and copy the glew32.lib into the folder `./Common/lib/x64` and the glew32.dll into the `./bin/win64/${BUILD_TYPE}` execution folder.
|
||||||
|
|
||||||
#### GLFW
|
#### GLFW
|
||||||
|
|
||||||
GLFW is a lightweight, open-source library designed for managing OpenGL, OpenGL ES, and Vulkan contexts. It simplifies the process of creating and managing windows, handling user input (keyboard, mouse, and joystick), and working with multiple monitors in a cross-platform manner.
|
GLFW is a lightweight, open-source library designed for managing OpenGL, OpenGL ES, and Vulkan contexts. It simplifies the process of creating and managing windows, handling user input (keyboard, mouse, and joystick), and working with multiple monitors in a cross-platform manner.
|
||||||
|
|
||||||
To set up GLFW on a Windows system, Download the pre-built binaries from [GLFW website](https://www.glfw.org/download.html) and extract the zip file into the folder, pass the GLFW include header as `-DGLFW_INCLUDE_DIR` for cmake configuring and follow the Build_instructions.txt in the sample folder to set up the t.
|
To set up GLFW on a Windows system, Download the pre-built binaries from [GLFW website](https://www.glfw.org/download.html) and extract the zip file into the folder, pass the GLFW include header folder as `-DGLFW_INCLUDE_DIR` and lib folder as `-DGLFW_LIB_DIR` for cmake configuring.
|
||||||
|
|
||||||
#### OpenMP
|
#### OpenMP
|
||||||
|
|
||||||
OpenMP is an API for multiprocessing programming. OpenMP can be installed using your Linux distribution's package manager system. It usually comes preinstalled with GCC. It can also be found at the [OpenMP website](http://openmp.org/).
|
OpenMP is an API for multiprocessing programming. OpenMP can be installed using your Linux distribution's package manager system. It usually comes preinstalled with GCC. It can also be found at the [OpenMP website](http://openmp.org/). For compilers such as clang, make sure OpenMP is enabled when building LLVM by including `openmp` in `LLVM_ENABLE_PROJECTS`. If you use clang (from an installed prefix or directly from an LLVM build tree) with OpenMP enabled, set CMAKE_CXX_COMPILER and CMAKE_CUDA_HOST_COMPILER to that clang++ and let CMake detect OpenMP; extra OpenMP_* CMake variables are usually not needed. When using clang++ directly from an LLVM build tree, you may need to copy the generated `omp.h` into the `include/` directory under the path reported by `clang++ --print-resource-dir` before building the samples. When using an installed clang with OpenMP, if you see libomp.so: cannot open shared object file at runtime, add the directory that contains libomp.so to LD_LIBRARY_PATH (or configure it via ld.so.conf.d) so the dynamic linker can locate the OpenMP runtime.
|
||||||
|
|
||||||
#### Screen
|
#### Screen
|
||||||
|
|
||||||
@@ -297,6 +631,11 @@ NVJPEG library provides high-performance, GPU accelerated JPEG decoding function
|
|||||||
|
|
||||||
NVRTC (CUDA RunTime Compilation) is a runtime compilation library for CUDA C++.
|
NVRTC (CUDA RunTime Compilation) is a runtime compilation library for CUDA C++.
|
||||||
|
|
||||||
|
#### NVJITLINK
|
||||||
|
|
||||||
|
NVJITLINK is a CUDA runtime library that links multiple GPU device code objects at runtime and enables just‑in‑time link‑time optimization (JIT LTO).
|
||||||
|
It can take device code from tools such as NVCC and NVRTC, link and optimize it on the fly, and produce the final GPU binary.For more details, see: https://docs.nvidia.com/cuda/nvjitlink/index.html
|
||||||
|
|
||||||
#### Stream Priorities
|
#### Stream Priorities
|
||||||
|
|
||||||
Stream Priorities allows the creation of streams with specified priorities. Stream Priorities is only available on GPUs with SM architecture of 3.5 or above.
|
Stream Priorities allows the creation of streams with specified priorities. Stream Priorities is only available on GPUs with SM architecture of 3.5 or above.
|
||||||
@@ -329,7 +668,7 @@ Answers to frequently asked questions about CUDA can be found at http://develope
|
|||||||
|
|
||||||
## References
|
## References
|
||||||
|
|
||||||
* [CUDA Programming Guide](http://docs.nvidia.com/cuda/cuda-c-programming-guide/index.html)
|
* [CUDA Programming Guide](https://docs.nvidia.com/cuda/cuda-programming-guide/index.html)
|
||||||
* [Accelerated Computing Blog](https://developer.nvidia.com/blog/?tags=accelerated-computing)
|
* [Accelerated Computing Blog](https://developer.nvidia.com/blog/?tags=accelerated-computing)
|
||||||
|
|
||||||
## Attributions
|
## Attributions
|
||||||
|
|||||||
@@ -1,62 +0,0 @@
|
|||||||
cmake_minimum_required(VERSION 3.20)
|
|
||||||
|
|
||||||
list(APPEND CMAKE_MODULE_PATH "${CMAKE_CURRENT_SOURCE_DIR}/../../../cmake/Modules")
|
|
||||||
|
|
||||||
project(simpleCallback LANGUAGES C CXX CUDA)
|
|
||||||
|
|
||||||
find_package(CUDAToolkit REQUIRED)
|
|
||||||
|
|
||||||
set(CMAKE_POSITION_INDEPENDENT_CODE ON)
|
|
||||||
|
|
||||||
set(CMAKE_CUDA_ARCHITECTURES 50 52 60 61 70 72 75 80 86 87 89 90 100 101 120)
|
|
||||||
set(CMAKE_CUDA_FLAGS "${CMAKE_CUDA_FLAGS} -Wno-deprecated-gpu-targets")
|
|
||||||
if(CMAKE_BUILD_TYPE STREQUAL "Debug")
|
|
||||||
# set(CMAKE_CUDA_FLAGS "${CMAKE_CUDA_FLAGS} -G") # enable cuda-gdb (expensive)
|
|
||||||
endif()
|
|
||||||
|
|
||||||
|
|
||||||
add_subdirectory(UnifiedMemoryStreams)
|
|
||||||
add_subdirectory(asyncAPI)
|
|
||||||
add_subdirectory(clock)
|
|
||||||
add_subdirectory(clock_nvrtc)
|
|
||||||
add_subdirectory(cudaOpenMP)
|
|
||||||
add_subdirectory(fp16ScalarProduct)
|
|
||||||
add_subdirectory(matrixMul)
|
|
||||||
add_subdirectory(matrixMulDrv)
|
|
||||||
add_subdirectory(matrixMulDynlinkJIT)
|
|
||||||
add_subdirectory(matrixMul_nvrtc)
|
|
||||||
add_subdirectory(mergeSort)
|
|
||||||
add_subdirectory(simpleAWBarrier)
|
|
||||||
add_subdirectory(simpleAssert)
|
|
||||||
add_subdirectory(simpleAssert_nvrtc)
|
|
||||||
add_subdirectory(simpleAtomicIntrinsics)
|
|
||||||
add_subdirectory(simpleAtomicIntrinsics_nvrtc)
|
|
||||||
add_subdirectory(simpleAttributes)
|
|
||||||
add_subdirectory(simpleCUDA2GL)
|
|
||||||
add_subdirectory(simpleCallback)
|
|
||||||
add_subdirectory(simpleCooperativeGroups)
|
|
||||||
add_subdirectory(simpleCubemapTexture)
|
|
||||||
add_subdirectory(simpleDrvRuntime)
|
|
||||||
add_subdirectory(simpleHyperQ)
|
|
||||||
add_subdirectory(simpleIPC)
|
|
||||||
add_subdirectory(simpleLayeredTexture)
|
|
||||||
add_subdirectory(simpleMPI)
|
|
||||||
add_subdirectory(simpleMultiCopy)
|
|
||||||
add_subdirectory(simpleMultiGPU)
|
|
||||||
add_subdirectory(simpleOccupancy)
|
|
||||||
add_subdirectory(simpleP2P)
|
|
||||||
add_subdirectory(simplePitchLinearTexture)
|
|
||||||
add_subdirectory(simplePrintf)
|
|
||||||
add_subdirectory(simpleStreams)
|
|
||||||
add_subdirectory(simpleSurfaceWrite)
|
|
||||||
add_subdirectory(simpleTemplates)
|
|
||||||
add_subdirectory(simpleTexture)
|
|
||||||
add_subdirectory(simpleTexture3D)
|
|
||||||
add_subdirectory(simpleTextureDrv)
|
|
||||||
add_subdirectory(simpleVoteIntrinsics)
|
|
||||||
add_subdirectory(simpleZeroCopy)
|
|
||||||
add_subdirectory(systemWideAtomics)
|
|
||||||
add_subdirectory(vectorAdd)
|
|
||||||
add_subdirectory(vectorAddDrv)
|
|
||||||
add_subdirectory(vectorAddMMAP)
|
|
||||||
add_subdirectory(vectorAdd_nvrtc)
|
|
||||||
@@ -1,10 +0,0 @@
|
|||||||
{
|
|
||||||
"configurations": [
|
|
||||||
{
|
|
||||||
"name": "CUDA C++: Launch",
|
|
||||||
"type": "cuda-gdb",
|
|
||||||
"request": "launch",
|
|
||||||
"program": "${workspaceFolder}/UnifiedMemoryStreams"
|
|
||||||
}
|
|
||||||
]
|
|
||||||
}
|
|
||||||
@@ -1,15 +0,0 @@
|
|||||||
{
|
|
||||||
"version": "2.0.0",
|
|
||||||
"tasks": [
|
|
||||||
{
|
|
||||||
"label": "sample",
|
|
||||||
"type": "shell",
|
|
||||||
"command": "make dbg=1",
|
|
||||||
"problemMatcher": ["$nvcc"],
|
|
||||||
"group": {
|
|
||||||
"kind": "build",
|
|
||||||
"isDefault": true
|
|
||||||
}
|
|
||||||
}
|
|
||||||
]
|
|
||||||
}
|
|
||||||
@@ -1,39 +0,0 @@
|
|||||||
cmake_minimum_required(VERSION 3.20)
|
|
||||||
|
|
||||||
list(APPEND CMAKE_MODULE_PATH "${CMAKE_CURRENT_SOURCE_DIR}/../../../cmake/Modules")
|
|
||||||
|
|
||||||
project(UnifiedMemoryStreams LANGUAGES C CXX CUDA)
|
|
||||||
|
|
||||||
find_package(CUDAToolkit REQUIRED)
|
|
||||||
|
|
||||||
set(CMAKE_POSITION_INDEPENDENT_CODE ON)
|
|
||||||
|
|
||||||
set(CMAKE_CUDA_ARCHITECTURES 50 52 60 61 70 72 75 80 86 87 89 90 100 101 120)
|
|
||||||
set(CMAKE_CUDA_FLAGS "${CMAKE_CUDA_FLAGS} -Wno-deprecated-gpu-targets")
|
|
||||||
if(CMAKE_BUILD_TYPE STREQUAL "Debug")
|
|
||||||
# set(CMAKE_CUDA_FLAGS "${CMAKE_CUDA_FLAGS} -G") # enable cuda-gdb (expensive)
|
|
||||||
endif()
|
|
||||||
|
|
||||||
# Include directories and libraries
|
|
||||||
include_directories(../../../Common)
|
|
||||||
|
|
||||||
# Source file
|
|
||||||
find_package(OpenMP REQUIRED)
|
|
||||||
|
|
||||||
if(${OpenMP_FOUND})
|
|
||||||
# Add target for UnifiedMemoryStreams
|
|
||||||
add_executable(UnifiedMemoryStreams UnifiedMemoryStreams.cu)
|
|
||||||
|
|
||||||
target_compile_options(UnifiedMemoryStreams PRIVATE $<$<COMPILE_LANGUAGE:CUDA>:--extended-lambda>)
|
|
||||||
|
|
||||||
target_compile_features(UnifiedMemoryStreams PRIVATE cxx_std_17 cuda_std_17)
|
|
||||||
|
|
||||||
set_target_properties(UnifiedMemoryStreams PROPERTIES CUDA_SEPARABLE_COMPILATION ON)
|
|
||||||
|
|
||||||
target_link_libraries(UnifiedMemoryStreams PUBLIC
|
|
||||||
CUDA::cublas
|
|
||||||
OpenMP::OpenMP_CXX
|
|
||||||
)
|
|
||||||
else()
|
|
||||||
message(STATUS "OpenMP not found - will not build sample 'UnifiedMemoryStreams'")
|
|
||||||
endif()
|
|
||||||
@@ -1,34 +0,0 @@
|
|||||||
# UnifiedMemoryStreams - Unified Memory Streams
|
|
||||||
|
|
||||||
## Description
|
|
||||||
|
|
||||||
This sample demonstrates the use of OpenMP and streams with Unified Memory on a single GPU.
|
|
||||||
|
|
||||||
## Key Concepts
|
|
||||||
|
|
||||||
CUDA Systems Integration, OpenMP, CUBLAS, Multithreading, Unified Memory, CUDA Streams and Events
|
|
||||||
|
|
||||||
## Supported SM Architectures
|
|
||||||
|
|
||||||
## Supported OSes
|
|
||||||
|
|
||||||
Linux, Windows
|
|
||||||
|
|
||||||
## Supported CPU Architecture
|
|
||||||
|
|
||||||
x86_64, armv7l
|
|
||||||
|
|
||||||
## CUDA APIs involved
|
|
||||||
|
|
||||||
### [CUDA Runtime API](http://docs.nvidia.com/cuda/cuda-runtime-api/index.html)
|
|
||||||
cudaStreamDestroy, cudaFree, cudaMallocManaged, cudaStreamAttachMemAsync, cudaSetDevice, cudaDeviceSynchronize, cudaStreamSynchronize, cudaStreamCreate, cudaGetDeviceProperties
|
|
||||||
|
|
||||||
## Dependencies needed to build/run
|
|
||||||
[OpenMP](../../../README.md#openmp), [UVM](../../../README.md#uvm), [CUBLAS](../../../README.md#cublas)
|
|
||||||
|
|
||||||
## Prerequisites
|
|
||||||
|
|
||||||
Download and install the [CUDA Toolkit 12.5](https://developer.nvidia.com/cuda-downloads) for your corresponding platform.
|
|
||||||
Make sure the dependencies mentioned in [Dependencies]() section above are installed.
|
|
||||||
|
|
||||||
## References (for more details)
|
|
||||||
@@ -1,341 +0,0 @@
|
|||||||
/* Copyright (c) 2022, NVIDIA CORPORATION. All rights reserved.
|
|
||||||
*
|
|
||||||
* Redistribution and use in source and binary forms, with or without
|
|
||||||
* modification, are permitted provided that the following conditions
|
|
||||||
* are met:
|
|
||||||
* * Redistributions of source code must retain the above copyright
|
|
||||||
* notice, this list of conditions and the following disclaimer.
|
|
||||||
* * Redistributions in binary form must reproduce the above copyright
|
|
||||||
* notice, this list of conditions and the following disclaimer in the
|
|
||||||
* documentation and/or other materials provided with the distribution.
|
|
||||||
* * Neither the name of NVIDIA CORPORATION nor the names of its
|
|
||||||
* contributors may be used to endorse or promote products derived
|
|
||||||
* from this software without specific prior written permission.
|
|
||||||
*
|
|
||||||
* THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS ``AS IS'' AND ANY
|
|
||||||
* EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
|
|
||||||
* IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR
|
|
||||||
* PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER OR
|
|
||||||
* CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL,
|
|
||||||
* EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO,
|
|
||||||
* PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR
|
|
||||||
* PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY
|
|
||||||
* OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
|
|
||||||
* (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
|
|
||||||
* OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
|
|
||||||
*/
|
|
||||||
|
|
||||||
/*
|
|
||||||
* This sample implements a simple task consumer using threads and streams
|
|
||||||
* with all data in Unified Memory, and tasks consumed by both host and device
|
|
||||||
*/
|
|
||||||
|
|
||||||
// system includes
|
|
||||||
#include <cstdio>
|
|
||||||
#include <ctime>
|
|
||||||
#include <vector>
|
|
||||||
#include <algorithm>
|
|
||||||
#ifdef USE_PTHREADS
|
|
||||||
#include <pthread.h>
|
|
||||||
#else
|
|
||||||
#include <omp.h>
|
|
||||||
#endif
|
|
||||||
#include <stdlib.h>
|
|
||||||
|
|
||||||
// cuBLAS
|
|
||||||
#include <cublas_v2.h>
|
|
||||||
|
|
||||||
// utilities
|
|
||||||
#include <helper_cuda.h>
|
|
||||||
|
|
||||||
#if defined(WIN32) || defined(_WIN32) || defined(WIN64) || defined(_WIN64)
|
|
||||||
// SRAND48 and DRAND48 don't exist on windows, but these are the equivalent
|
|
||||||
// functions
|
|
||||||
void srand48(long seed) { srand((unsigned int)seed); }
|
|
||||||
double drand48() { return double(rand()) / RAND_MAX; }
|
|
||||||
#endif
|
|
||||||
|
|
||||||
const char *sSDKname = "UnifiedMemoryStreams";
|
|
||||||
|
|
||||||
// simple task
|
|
||||||
template <typename T>
|
|
||||||
struct Task {
|
|
||||||
unsigned int size, id;
|
|
||||||
T *data;
|
|
||||||
T *result;
|
|
||||||
T *vector;
|
|
||||||
|
|
||||||
Task() : size(0), id(0), data(NULL), result(NULL), vector(NULL){};
|
|
||||||
Task(unsigned int s) : size(s), id(0), data(NULL), result(NULL) {
|
|
||||||
// allocate unified memory -- the operation performed in this example will
|
|
||||||
// be a DGEMV
|
|
||||||
checkCudaErrors(cudaMallocManaged(&data, sizeof(T) * size * size));
|
|
||||||
checkCudaErrors(cudaMallocManaged(&result, sizeof(T) * size));
|
|
||||||
checkCudaErrors(cudaMallocManaged(&vector, sizeof(T) * size));
|
|
||||||
checkCudaErrors(cudaDeviceSynchronize());
|
|
||||||
}
|
|
||||||
|
|
||||||
~Task() {
|
|
||||||
// ensure all memory is deallocated
|
|
||||||
checkCudaErrors(cudaDeviceSynchronize());
|
|
||||||
checkCudaErrors(cudaFree(data));
|
|
||||||
checkCudaErrors(cudaFree(result));
|
|
||||||
checkCudaErrors(cudaFree(vector));
|
|
||||||
}
|
|
||||||
|
|
||||||
void allocate(const unsigned int s, const unsigned int unique_id) {
|
|
||||||
// allocate unified memory outside of constructor
|
|
||||||
id = unique_id;
|
|
||||||
size = s;
|
|
||||||
checkCudaErrors(cudaMallocManaged(&data, sizeof(T) * size * size));
|
|
||||||
checkCudaErrors(cudaMallocManaged(&result, sizeof(T) * size));
|
|
||||||
checkCudaErrors(cudaMallocManaged(&vector, sizeof(T) * size));
|
|
||||||
checkCudaErrors(cudaDeviceSynchronize());
|
|
||||||
|
|
||||||
// populate data with random elements
|
|
||||||
for (unsigned int i = 0; i < size * size; i++) {
|
|
||||||
data[i] = drand48();
|
|
||||||
}
|
|
||||||
|
|
||||||
for (unsigned int i = 0; i < size; i++) {
|
|
||||||
result[i] = 0.;
|
|
||||||
vector[i] = drand48();
|
|
||||||
}
|
|
||||||
}
|
|
||||||
};
|
|
||||||
|
|
||||||
#ifdef USE_PTHREADS
|
|
||||||
struct threadData_t {
|
|
||||||
int tid;
|
|
||||||
Task<double> *TaskListPtr;
|
|
||||||
cudaStream_t *streams;
|
|
||||||
cublasHandle_t *handles;
|
|
||||||
int taskSize;
|
|
||||||
};
|
|
||||||
|
|
||||||
typedef struct threadData_t threadData;
|
|
||||||
#endif
|
|
||||||
|
|
||||||
// simple host dgemv: assume data is in row-major format and square
|
|
||||||
template <typename T>
|
|
||||||
void gemv(int m, int n, T alpha, T *A, T *x, T beta, T *result) {
|
|
||||||
// rows
|
|
||||||
for (int i = 0; i < n; i++) {
|
|
||||||
result[i] *= beta;
|
|
||||||
|
|
||||||
for (int j = 0; j < n; j++) {
|
|
||||||
result[i] += A[i * n + j] * x[j];
|
|
||||||
}
|
|
||||||
}
|
|
||||||
}
|
|
||||||
|
|
||||||
// execute a single task on either host or device depending on size
|
|
||||||
#ifdef USE_PTHREADS
|
|
||||||
void *execute(void *inpArgs) {
|
|
||||||
threadData *dataPtr = (threadData *)inpArgs;
|
|
||||||
cudaStream_t *stream = dataPtr->streams;
|
|
||||||
cublasHandle_t *handle = dataPtr->handles;
|
|
||||||
int tid = dataPtr->tid;
|
|
||||||
|
|
||||||
for (int i = 0; i < dataPtr->taskSize; i++) {
|
|
||||||
Task<double> &t = dataPtr->TaskListPtr[i];
|
|
||||||
|
|
||||||
if (t.size < 100) {
|
|
||||||
// perform on host
|
|
||||||
printf("Task [%d], thread [%d] executing on host (%d)\n", t.id, tid,
|
|
||||||
t.size);
|
|
||||||
|
|
||||||
// attach managed memory to a (dummy) stream to allow host access while
|
|
||||||
// the device is running
|
|
||||||
checkCudaErrors(
|
|
||||||
cudaStreamAttachMemAsync(stream[0], t.data, 0, cudaMemAttachHost));
|
|
||||||
checkCudaErrors(
|
|
||||||
cudaStreamAttachMemAsync(stream[0], t.vector, 0, cudaMemAttachHost));
|
|
||||||
checkCudaErrors(
|
|
||||||
cudaStreamAttachMemAsync(stream[0], t.result, 0, cudaMemAttachHost));
|
|
||||||
// necessary to ensure Async cudaStreamAttachMemAsync calls have finished
|
|
||||||
checkCudaErrors(cudaStreamSynchronize(stream[0]));
|
|
||||||
// call the host operation
|
|
||||||
gemv(t.size, t.size, 1.0, t.data, t.vector, 0.0, t.result);
|
|
||||||
} else {
|
|
||||||
// perform on device
|
|
||||||
printf("Task [%d], thread [%d] executing on device (%d)\n", t.id, tid,
|
|
||||||
t.size);
|
|
||||||
double one = 1.0;
|
|
||||||
double zero = 0.0;
|
|
||||||
|
|
||||||
// attach managed memory to my stream
|
|
||||||
checkCudaErrors(cublasSetStream(handle[tid + 1], stream[tid + 1]));
|
|
||||||
checkCudaErrors(cudaStreamAttachMemAsync(stream[tid + 1], t.data, 0,
|
|
||||||
cudaMemAttachSingle));
|
|
||||||
checkCudaErrors(cudaStreamAttachMemAsync(stream[tid + 1], t.vector, 0,
|
|
||||||
cudaMemAttachSingle));
|
|
||||||
checkCudaErrors(cudaStreamAttachMemAsync(stream[tid + 1], t.result, 0,
|
|
||||||
cudaMemAttachSingle));
|
|
||||||
// call the device operation
|
|
||||||
checkCudaErrors(cublasDgemv(handle[tid + 1], CUBLAS_OP_N, t.size, t.size,
|
|
||||||
&one, t.data, t.size, t.vector, 1, &zero,
|
|
||||||
t.result, 1));
|
|
||||||
}
|
|
||||||
}
|
|
||||||
|
|
||||||
pthread_exit(NULL);
|
|
||||||
}
|
|
||||||
#else
|
|
||||||
template <typename T>
|
|
||||||
void execute(Task<T> &t, cublasHandle_t *handle, cudaStream_t *stream,
|
|
||||||
int tid) {
|
|
||||||
if (t.size < 100) {
|
|
||||||
// perform on host
|
|
||||||
printf("Task [%d], thread [%d] executing on host (%d)\n", t.id, tid,
|
|
||||||
t.size);
|
|
||||||
|
|
||||||
// attach managed memory to a (dummy) stream to allow host access while the
|
|
||||||
// device is running
|
|
||||||
checkCudaErrors(
|
|
||||||
cudaStreamAttachMemAsync(stream[0], t.data, 0, cudaMemAttachHost));
|
|
||||||
checkCudaErrors(
|
|
||||||
cudaStreamAttachMemAsync(stream[0], t.vector, 0, cudaMemAttachHost));
|
|
||||||
checkCudaErrors(
|
|
||||||
cudaStreamAttachMemAsync(stream[0], t.result, 0, cudaMemAttachHost));
|
|
||||||
// necessary to ensure Async cudaStreamAttachMemAsync calls have finished
|
|
||||||
checkCudaErrors(cudaStreamSynchronize(stream[0]));
|
|
||||||
// call the host operation
|
|
||||||
gemv(t.size, t.size, 1.0, t.data, t.vector, 0.0, t.result);
|
|
||||||
} else {
|
|
||||||
// perform on device
|
|
||||||
printf("Task [%d], thread [%d] executing on device (%d)\n", t.id, tid,
|
|
||||||
t.size);
|
|
||||||
double one = 1.0;
|
|
||||||
double zero = 0.0;
|
|
||||||
|
|
||||||
// attach managed memory to my stream
|
|
||||||
checkCudaErrors(cublasSetStream(handle[tid + 1], stream[tid + 1]));
|
|
||||||
checkCudaErrors(cudaStreamAttachMemAsync(stream[tid + 1], t.data, 0,
|
|
||||||
cudaMemAttachSingle));
|
|
||||||
checkCudaErrors(cudaStreamAttachMemAsync(stream[tid + 1], t.vector, 0,
|
|
||||||
cudaMemAttachSingle));
|
|
||||||
checkCudaErrors(cudaStreamAttachMemAsync(stream[tid + 1], t.result, 0,
|
|
||||||
cudaMemAttachSingle));
|
|
||||||
// call the device operation
|
|
||||||
checkCudaErrors(cublasDgemv(handle[tid + 1], CUBLAS_OP_N, t.size, t.size,
|
|
||||||
&one, t.data, t.size, t.vector, 1, &zero,
|
|
||||||
t.result, 1));
|
|
||||||
}
|
|
||||||
}
|
|
||||||
#endif
|
|
||||||
|
|
||||||
// populate a list of tasks with random sizes
|
|
||||||
template <typename T>
|
|
||||||
void initialise_tasks(std::vector<Task<T> > &TaskList) {
|
|
||||||
for (unsigned int i = 0; i < TaskList.size(); i++) {
|
|
||||||
// generate random size
|
|
||||||
int size;
|
|
||||||
size = std::max((int)(drand48() * 1000.0), 64);
|
|
||||||
TaskList[i].allocate(size, i);
|
|
||||||
}
|
|
||||||
}
|
|
||||||
|
|
||||||
int main(int argc, char **argv) {
|
|
||||||
// set device
|
|
||||||
cudaDeviceProp device_prop;
|
|
||||||
int dev_id = findCudaDevice(argc, (const char **)argv);
|
|
||||||
checkCudaErrors(cudaGetDeviceProperties(&device_prop, dev_id));
|
|
||||||
|
|
||||||
if (!device_prop.managedMemory) {
|
|
||||||
// This samples requires being run on a device that supports Unified Memory
|
|
||||||
fprintf(stderr, "Unified Memory not supported on this device\n");
|
|
||||||
|
|
||||||
exit(EXIT_WAIVED);
|
|
||||||
}
|
|
||||||
|
|
||||||
if (device_prop.computeMode == cudaComputeModeProhibited) {
|
|
||||||
// This sample requires being run with a default or process exclusive mode
|
|
||||||
fprintf(stderr,
|
|
||||||
"This sample requires a device in either default or process "
|
|
||||||
"exclusive mode\n");
|
|
||||||
|
|
||||||
exit(EXIT_WAIVED);
|
|
||||||
}
|
|
||||||
|
|
||||||
// randomise task sizes
|
|
||||||
int seed = (int)time(NULL);
|
|
||||||
srand48(seed);
|
|
||||||
|
|
||||||
// set number of threads
|
|
||||||
const int nthreads = 4;
|
|
||||||
|
|
||||||
// number of streams = number of threads
|
|
||||||
cudaStream_t *streams = new cudaStream_t[nthreads + 1];
|
|
||||||
cublasHandle_t *handles = new cublasHandle_t[nthreads + 1];
|
|
||||||
|
|
||||||
for (int i = 0; i < nthreads + 1; i++) {
|
|
||||||
checkCudaErrors(cudaStreamCreate(&streams[i]));
|
|
||||||
checkCudaErrors(cublasCreate(&handles[i]));
|
|
||||||
}
|
|
||||||
|
|
||||||
// create list of N tasks
|
|
||||||
unsigned int N = 40;
|
|
||||||
std::vector<Task<double> > TaskList(N);
|
|
||||||
initialise_tasks(TaskList);
|
|
||||||
|
|
||||||
printf("Executing tasks on host / device\n");
|
|
||||||
|
|
||||||
// run through all tasks using threads and streams
|
|
||||||
#ifdef USE_PTHREADS
|
|
||||||
pthread_t threads[nthreads];
|
|
||||||
threadData *InputToThreads = new threadData[nthreads];
|
|
||||||
|
|
||||||
for (int i = 0; i < nthreads; i++) {
|
|
||||||
checkCudaErrors(cudaSetDevice(dev_id));
|
|
||||||
InputToThreads[i].tid = i;
|
|
||||||
InputToThreads[i].streams = streams;
|
|
||||||
InputToThreads[i].handles = handles;
|
|
||||||
|
|
||||||
if ((TaskList.size() / nthreads) == 0) {
|
|
||||||
InputToThreads[i].taskSize = (TaskList.size() / nthreads);
|
|
||||||
InputToThreads[i].TaskListPtr =
|
|
||||||
&TaskList[i * (TaskList.size() / nthreads)];
|
|
||||||
} else {
|
|
||||||
if (i == nthreads - 1) {
|
|
||||||
InputToThreads[i].taskSize =
|
|
||||||
(TaskList.size() / nthreads) + (TaskList.size() % nthreads);
|
|
||||||
InputToThreads[i].TaskListPtr =
|
|
||||||
&TaskList[i * (TaskList.size() / nthreads) +
|
|
||||||
(TaskList.size() % nthreads)];
|
|
||||||
} else {
|
|
||||||
InputToThreads[i].taskSize = (TaskList.size() / nthreads);
|
|
||||||
InputToThreads[i].TaskListPtr =
|
|
||||||
&TaskList[i * (TaskList.size() / nthreads)];
|
|
||||||
}
|
|
||||||
}
|
|
||||||
|
|
||||||
pthread_create(&threads[i], NULL, &execute, &InputToThreads[i]);
|
|
||||||
}
|
|
||||||
for (int i = 0; i < nthreads; i++) {
|
|
||||||
pthread_join(threads[i], NULL);
|
|
||||||
}
|
|
||||||
#else
|
|
||||||
omp_set_num_threads(nthreads);
|
|
||||||
#pragma omp parallel for schedule(dynamic)
|
|
||||||
for (int i = 0; i < TaskList.size(); i++) {
|
|
||||||
checkCudaErrors(cudaSetDevice(dev_id));
|
|
||||||
int tid = omp_get_thread_num();
|
|
||||||
execute(TaskList[i], handles, streams, tid);
|
|
||||||
}
|
|
||||||
#endif
|
|
||||||
|
|
||||||
cudaDeviceSynchronize();
|
|
||||||
|
|
||||||
// Destroy CUDA Streams, cuBlas handles
|
|
||||||
for (int i = 0; i < nthreads + 1; i++) {
|
|
||||||
cudaStreamDestroy(streams[i]);
|
|
||||||
cublasDestroy(handles[i]);
|
|
||||||
}
|
|
||||||
|
|
||||||
// Free TaskList
|
|
||||||
std::vector<Task<double> >().swap(TaskList);
|
|
||||||
|
|
||||||
printf("All Done!\n");
|
|
||||||
exit(EXIT_SUCCESS);
|
|
||||||
}
|
|
||||||
@@ -1,10 +0,0 @@
|
|||||||
{
|
|
||||||
"configurations": [
|
|
||||||
{
|
|
||||||
"name": "CUDA C++: Launch",
|
|
||||||
"type": "cuda-gdb",
|
|
||||||
"request": "launch",
|
|
||||||
"program": "${workspaceFolder}/asyncAPI"
|
|
||||||
}
|
|
||||||
]
|
|
||||||
}
|
|
||||||
@@ -1,15 +0,0 @@
|
|||||||
{
|
|
||||||
"version": "2.0.0",
|
|
||||||
"tasks": [
|
|
||||||
{
|
|
||||||
"label": "sample",
|
|
||||||
"type": "shell",
|
|
||||||
"command": "make dbg=1",
|
|
||||||
"problemMatcher": ["$nvcc"],
|
|
||||||
"group": {
|
|
||||||
"kind": "build",
|
|
||||||
"isDefault": true
|
|
||||||
}
|
|
||||||
}
|
|
||||||
]
|
|
||||||
}
|
|
||||||
@@ -1,28 +0,0 @@
|
|||||||
cmake_minimum_required(VERSION 3.20)
|
|
||||||
|
|
||||||
list(APPEND CMAKE_MODULE_PATH "${CMAKE_CURRENT_SOURCE_DIR}/../../../cmake/Modules")
|
|
||||||
|
|
||||||
project(asyncAPI LANGUAGES C CXX CUDA)
|
|
||||||
|
|
||||||
find_package(CUDAToolkit REQUIRED)
|
|
||||||
|
|
||||||
set(CMAKE_POSITION_INDEPENDENT_CODE ON)
|
|
||||||
|
|
||||||
set(CMAKE_CUDA_ARCHITECTURES 50 52 60 61 70 72 75 80 86 87 89 90 100 101 120)
|
|
||||||
set(CMAKE_CUDA_FLAGS "${CMAKE_CUDA_FLAGS} -Wno-deprecated-gpu-targets")
|
|
||||||
if(CMAKE_BUILD_TYPE STREQUAL "Debug")
|
|
||||||
# set(CMAKE_CUDA_FLAGS "${CMAKE_CUDA_FLAGS} -G") # enable cuda-gdb (expensive)
|
|
||||||
endif()
|
|
||||||
|
|
||||||
# Include directories and libraries
|
|
||||||
include_directories(../../../Common)
|
|
||||||
|
|
||||||
# Source file
|
|
||||||
# Add target for asyncAPI
|
|
||||||
add_executable(asyncAPI asyncAPI.cu)
|
|
||||||
|
|
||||||
target_compile_options(asyncAPI PRIVATE $<$<COMPILE_LANGUAGE:CUDA>:--extended-lambda>)
|
|
||||||
|
|
||||||
target_compile_features(asyncAPI PRIVATE cxx_std_17 cuda_std_17)
|
|
||||||
|
|
||||||
set_target_properties(asyncAPI PROPERTIES CUDA_SEPARABLE_COMPILATION ON)
|
|
||||||
@@ -1,32 +0,0 @@
|
|||||||
# asyncAPI - asyncAPI
|
|
||||||
|
|
||||||
## Description
|
|
||||||
|
|
||||||
This sample illustrates the usage of CUDA events for both GPU timing and overlapping CPU and GPU execution. Events are inserted into a stream of CUDA calls. Since CUDA stream calls are asynchronous, the CPU can perform computations while GPU is executing (including DMA memcopies between the host and device). CPU can query CUDA events to determine whether GPU has completed tasks.
|
|
||||||
|
|
||||||
## Key Concepts
|
|
||||||
|
|
||||||
Asynchronous Data Transfers, CUDA Streams and Events
|
|
||||||
|
|
||||||
## Supported SM Architectures
|
|
||||||
|
|
||||||
[SM 5.0 ](https://developer.nvidia.com/cuda-gpus) [SM 5.2 ](https://developer.nvidia.com/cuda-gpus) [SM 5.3 ](https://developer.nvidia.com/cuda-gpus) [SM 6.0 ](https://developer.nvidia.com/cuda-gpus) [SM 6.1 ](https://developer.nvidia.com/cuda-gpus) [SM 7.0 ](https://developer.nvidia.com/cuda-gpus) [SM 7.2 ](https://developer.nvidia.com/cuda-gpus) [SM 7.5 ](https://developer.nvidia.com/cuda-gpus) [SM 8.0 ](https://developer.nvidia.com/cuda-gpus) [SM 8.6 ](https://developer.nvidia.com/cuda-gpus) [SM 8.7 ](https://developer.nvidia.com/cuda-gpus) [SM 8.9 ](https://developer.nvidia.com/cuda-gpus) [SM 9.0 ](https://developer.nvidia.com/cuda-gpus)
|
|
||||||
|
|
||||||
## Supported OSes
|
|
||||||
|
|
||||||
Linux, Windows
|
|
||||||
|
|
||||||
## Supported CPU Architecture
|
|
||||||
|
|
||||||
x86_64, armv7l
|
|
||||||
|
|
||||||
## CUDA APIs involved
|
|
||||||
|
|
||||||
### [CUDA Runtime API](http://docs.nvidia.com/cuda/cuda-runtime-api/index.html)
|
|
||||||
cudaProfilerStop, cudaMalloc, cudaMemcpyAsync, cudaFree, cudaMallocHost, cudaProfilerStart, cudaDeviceSynchronize, cudaEventRecord, cudaFreeHost, cudaMemset, cudaEventDestroy, cudaEventQuery, cudaEventElapsedTime, cudaGetDeviceProperties, cudaEventCreate
|
|
||||||
|
|
||||||
## Prerequisites
|
|
||||||
|
|
||||||
Download and install the [CUDA Toolkit 12.5](https://developer.nvidia.com/cuda-downloads) for your corresponding platform.
|
|
||||||
|
|
||||||
## References (for more details)
|
|
||||||
@@ -1,142 +0,0 @@
|
|||||||
/* Copyright (c) 2022, NVIDIA CORPORATION. All rights reserved.
|
|
||||||
*
|
|
||||||
* Redistribution and use in source and binary forms, with or without
|
|
||||||
* modification, are permitted provided that the following conditions
|
|
||||||
* are met:
|
|
||||||
* * Redistributions of source code must retain the above copyright
|
|
||||||
* notice, this list of conditions and the following disclaimer.
|
|
||||||
* * Redistributions in binary form must reproduce the above copyright
|
|
||||||
* notice, this list of conditions and the following disclaimer in the
|
|
||||||
* documentation and/or other materials provided with the distribution.
|
|
||||||
* * Neither the name of NVIDIA CORPORATION nor the names of its
|
|
||||||
* contributors may be used to endorse or promote products derived
|
|
||||||
* from this software without specific prior written permission.
|
|
||||||
*
|
|
||||||
* THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS ``AS IS'' AND ANY
|
|
||||||
* EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
|
|
||||||
* IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR
|
|
||||||
* PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER OR
|
|
||||||
* CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL,
|
|
||||||
* EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO,
|
|
||||||
* PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR
|
|
||||||
* PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY
|
|
||||||
* OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
|
|
||||||
* (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
|
|
||||||
* OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
|
|
||||||
*/
|
|
||||||
|
|
||||||
/*
|
|
||||||
* This sample illustrates the usage of CUDA events for both GPU timing and
|
|
||||||
* overlapping CPU and GPU execution. Events are inserted into a stream
|
|
||||||
* of CUDA calls. Since CUDA stream calls are asynchronous, the CPU can
|
|
||||||
* perform computations while GPU is executing (including DMA memcopies
|
|
||||||
* between the host and device). CPU can query CUDA events to determine
|
|
||||||
* whether GPU has completed tasks.
|
|
||||||
*/
|
|
||||||
|
|
||||||
// includes, system
|
|
||||||
#include <stdio.h>
|
|
||||||
|
|
||||||
// includes CUDA Runtime
|
|
||||||
#include <cuda_runtime.h>
|
|
||||||
#include <cuda_profiler_api.h>
|
|
||||||
|
|
||||||
// includes, project
|
|
||||||
#include <helper_cuda.h>
|
|
||||||
#include <helper_functions.h> // helper utility functions
|
|
||||||
|
|
||||||
__global__ void increment_kernel(int *g_data, int inc_value) {
|
|
||||||
int idx = blockIdx.x * blockDim.x + threadIdx.x;
|
|
||||||
g_data[idx] = g_data[idx] + inc_value;
|
|
||||||
}
|
|
||||||
|
|
||||||
bool correct_output(int *data, const int n, const int x) {
|
|
||||||
for (int i = 0; i < n; i++)
|
|
||||||
if (data[i] != x) {
|
|
||||||
printf("Error! data[%d] = %d, ref = %d\n", i, data[i], x);
|
|
||||||
return false;
|
|
||||||
}
|
|
||||||
|
|
||||||
return true;
|
|
||||||
}
|
|
||||||
|
|
||||||
int main(int argc, char *argv[]) {
|
|
||||||
int devID;
|
|
||||||
cudaDeviceProp deviceProps;
|
|
||||||
|
|
||||||
printf("[%s] - Starting...\n", argv[0]);
|
|
||||||
|
|
||||||
// This will pick the best possible CUDA capable device
|
|
||||||
devID = findCudaDevice(argc, (const char **)argv);
|
|
||||||
|
|
||||||
// get device name
|
|
||||||
checkCudaErrors(cudaGetDeviceProperties(&deviceProps, devID));
|
|
||||||
printf("CUDA device [%s]\n", deviceProps.name);
|
|
||||||
|
|
||||||
int n = 16 * 1024 * 1024;
|
|
||||||
int nbytes = n * sizeof(int);
|
|
||||||
int value = 26;
|
|
||||||
|
|
||||||
// allocate host memory
|
|
||||||
int *a = 0;
|
|
||||||
checkCudaErrors(cudaMallocHost((void **)&a, nbytes));
|
|
||||||
memset(a, 0, nbytes);
|
|
||||||
|
|
||||||
// allocate device memory
|
|
||||||
int *d_a = 0;
|
|
||||||
checkCudaErrors(cudaMalloc((void **)&d_a, nbytes));
|
|
||||||
checkCudaErrors(cudaMemset(d_a, 255, nbytes));
|
|
||||||
|
|
||||||
// set kernel launch configuration
|
|
||||||
dim3 threads = dim3(512, 1);
|
|
||||||
dim3 blocks = dim3(n / threads.x, 1);
|
|
||||||
|
|
||||||
// create cuda event handles
|
|
||||||
cudaEvent_t start, stop;
|
|
||||||
checkCudaErrors(cudaEventCreate(&start));
|
|
||||||
checkCudaErrors(cudaEventCreate(&stop));
|
|
||||||
|
|
||||||
StopWatchInterface *timer = NULL;
|
|
||||||
sdkCreateTimer(&timer);
|
|
||||||
sdkResetTimer(&timer);
|
|
||||||
|
|
||||||
checkCudaErrors(cudaDeviceSynchronize());
|
|
||||||
float gpu_time = 0.0f;
|
|
||||||
|
|
||||||
// asynchronously issue work to the GPU (all to stream 0)
|
|
||||||
checkCudaErrors(cudaProfilerStart());
|
|
||||||
sdkStartTimer(&timer);
|
|
||||||
cudaEventRecord(start, 0);
|
|
||||||
cudaMemcpyAsync(d_a, a, nbytes, cudaMemcpyHostToDevice, 0);
|
|
||||||
increment_kernel<<<blocks, threads, 0, 0>>>(d_a, value);
|
|
||||||
cudaMemcpyAsync(a, d_a, nbytes, cudaMemcpyDeviceToHost, 0);
|
|
||||||
cudaEventRecord(stop, 0);
|
|
||||||
sdkStopTimer(&timer);
|
|
||||||
checkCudaErrors(cudaProfilerStop());
|
|
||||||
|
|
||||||
// have CPU do some work while waiting for stage 1 to finish
|
|
||||||
unsigned long int counter = 0;
|
|
||||||
|
|
||||||
while (cudaEventQuery(stop) == cudaErrorNotReady) {
|
|
||||||
counter++;
|
|
||||||
}
|
|
||||||
|
|
||||||
checkCudaErrors(cudaEventElapsedTime(&gpu_time, start, stop));
|
|
||||||
|
|
||||||
// print the cpu and gpu times
|
|
||||||
printf("time spent executing by the GPU: %.2f\n", gpu_time);
|
|
||||||
printf("time spent by CPU in CUDA calls: %.2f\n", sdkGetTimerValue(&timer));
|
|
||||||
printf("CPU executed %lu iterations while waiting for GPU to finish\n",
|
|
||||||
counter);
|
|
||||||
|
|
||||||
// check the output for correctness
|
|
||||||
bool bFinalResults = correct_output(a, n, value);
|
|
||||||
|
|
||||||
// release resources
|
|
||||||
checkCudaErrors(cudaEventDestroy(start));
|
|
||||||
checkCudaErrors(cudaEventDestroy(stop));
|
|
||||||
checkCudaErrors(cudaFreeHost(a));
|
|
||||||
checkCudaErrors(cudaFree(d_a));
|
|
||||||
|
|
||||||
exit(bFinalResults ? EXIT_SUCCESS : EXIT_FAILURE);
|
|
||||||
}
|
|
||||||
10
Samples/0_Introduction/clock/.vscode/launch.json
vendored
10
Samples/0_Introduction/clock/.vscode/launch.json
vendored
@@ -1,10 +0,0 @@
|
|||||||
{
|
|
||||||
"configurations": [
|
|
||||||
{
|
|
||||||
"name": "CUDA C++: Launch",
|
|
||||||
"type": "cuda-gdb",
|
|
||||||
"request": "launch",
|
|
||||||
"program": "${workspaceFolder}/clock"
|
|
||||||
}
|
|
||||||
]
|
|
||||||
}
|
|
||||||
15
Samples/0_Introduction/clock/.vscode/tasks.json
vendored
15
Samples/0_Introduction/clock/.vscode/tasks.json
vendored
@@ -1,15 +0,0 @@
|
|||||||
{
|
|
||||||
"version": "2.0.0",
|
|
||||||
"tasks": [
|
|
||||||
{
|
|
||||||
"label": "sample",
|
|
||||||
"type": "shell",
|
|
||||||
"command": "make dbg=1",
|
|
||||||
"problemMatcher": ["$nvcc"],
|
|
||||||
"group": {
|
|
||||||
"kind": "build",
|
|
||||||
"isDefault": true
|
|
||||||
}
|
|
||||||
}
|
|
||||||
]
|
|
||||||
}
|
|
||||||
@@ -1,28 +0,0 @@
|
|||||||
cmake_minimum_required(VERSION 3.20)
|
|
||||||
|
|
||||||
list(APPEND CMAKE_MODULE_PATH "${CMAKE_CURRENT_SOURCE_DIR}/../../../cmake/Modules")
|
|
||||||
|
|
||||||
project(clock LANGUAGES C CXX CUDA)
|
|
||||||
|
|
||||||
find_package(CUDAToolkit REQUIRED)
|
|
||||||
|
|
||||||
set(CMAKE_POSITION_INDEPENDENT_CODE ON)
|
|
||||||
|
|
||||||
set(CMAKE_CUDA_ARCHITECTURES 50 52 60 61 70 72 75 80 86 87 89 90 100 101 120)
|
|
||||||
set(CMAKE_CUDA_FLAGS "${CMAKE_CUDA_FLAGS} -Wno-deprecated-gpu-targets")
|
|
||||||
if(CMAKE_BUILD_TYPE STREQUAL "Debug")
|
|
||||||
# set(CMAKE_CUDA_FLAGS "${CMAKE_CUDA_FLAGS} -G") # enable cuda-gdb (expensive)
|
|
||||||
endif()
|
|
||||||
|
|
||||||
# Include directories and libraries
|
|
||||||
include_directories(../../../Common)
|
|
||||||
|
|
||||||
# Source file
|
|
||||||
# Add target for asyncAPI
|
|
||||||
add_executable(clock clock.cu)
|
|
||||||
|
|
||||||
target_compile_options(clock PRIVATE $<$<COMPILE_LANGUAGE:CUDA>:--extended-lambda>)
|
|
||||||
|
|
||||||
target_compile_features(clock PRIVATE cxx_std_17 cuda_std_17)
|
|
||||||
|
|
||||||
set_target_properties(clock PROPERTIES CUDA_SEPARABLE_COMPILATION ON)
|
|
||||||
@@ -1,32 +0,0 @@
|
|||||||
# clock - Clock
|
|
||||||
|
|
||||||
## Description
|
|
||||||
|
|
||||||
This example shows how to use the clock function to measure the performance of block of threads of a kernel accurately.
|
|
||||||
|
|
||||||
## Key Concepts
|
|
||||||
|
|
||||||
Performance Strategies
|
|
||||||
|
|
||||||
## Supported SM Architectures
|
|
||||||
|
|
||||||
[SM 5.0 ](https://developer.nvidia.com/cuda-gpus) [SM 5.2 ](https://developer.nvidia.com/cuda-gpus) [SM 5.3 ](https://developer.nvidia.com/cuda-gpus) [SM 6.0 ](https://developer.nvidia.com/cuda-gpus) [SM 6.1 ](https://developer.nvidia.com/cuda-gpus) [SM 7.0 ](https://developer.nvidia.com/cuda-gpus) [SM 7.2 ](https://developer.nvidia.com/cuda-gpus) [SM 7.5 ](https://developer.nvidia.com/cuda-gpus) [SM 8.0 ](https://developer.nvidia.com/cuda-gpus) [SM 8.6 ](https://developer.nvidia.com/cuda-gpus) [SM 8.7 ](https://developer.nvidia.com/cuda-gpus) [SM 8.9 ](https://developer.nvidia.com/cuda-gpus) [SM 9.0 ](https://developer.nvidia.com/cuda-gpus)
|
|
||||||
|
|
||||||
## Supported OSes
|
|
||||||
|
|
||||||
Linux, Windows
|
|
||||||
|
|
||||||
## Supported CPU Architecture
|
|
||||||
|
|
||||||
x86_64, armv7l
|
|
||||||
|
|
||||||
## CUDA APIs involved
|
|
||||||
|
|
||||||
### [CUDA Runtime API](http://docs.nvidia.com/cuda/cuda-runtime-api/index.html)
|
|
||||||
cudaMalloc, cudaMemcpy, cudaFree
|
|
||||||
|
|
||||||
## Prerequisites
|
|
||||||
|
|
||||||
Download and install the [CUDA Toolkit 12.5](https://developer.nvidia.com/cuda-downloads) for your corresponding platform.
|
|
||||||
|
|
||||||
## References (for more details)
|
|
||||||
@@ -1,153 +0,0 @@
|
|||||||
/* Copyright (c) 2022, NVIDIA CORPORATION. All rights reserved.
|
|
||||||
*
|
|
||||||
* Redistribution and use in source and binary forms, with or without
|
|
||||||
* modification, are permitted provided that the following conditions
|
|
||||||
* are met:
|
|
||||||
* * Redistributions of source code must retain the above copyright
|
|
||||||
* notice, this list of conditions and the following disclaimer.
|
|
||||||
* * Redistributions in binary form must reproduce the above copyright
|
|
||||||
* notice, this list of conditions and the following disclaimer in the
|
|
||||||
* documentation and/or other materials provided with the distribution.
|
|
||||||
* * Neither the name of NVIDIA CORPORATION nor the names of its
|
|
||||||
* contributors may be used to endorse or promote products derived
|
|
||||||
* from this software without specific prior written permission.
|
|
||||||
*
|
|
||||||
* THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS ``AS IS'' AND ANY
|
|
||||||
* EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
|
|
||||||
* IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR
|
|
||||||
* PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER OR
|
|
||||||
* CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL,
|
|
||||||
* EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO,
|
|
||||||
* PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR
|
|
||||||
* PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY
|
|
||||||
* OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
|
|
||||||
* (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
|
|
||||||
* OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
|
|
||||||
*/
|
|
||||||
|
|
||||||
/*
|
|
||||||
* This example shows how to use the clock function to measure the performance
|
|
||||||
* of block of threads of a kernel accurately. Blocks are executed in parallel
|
|
||||||
* and out of order. Since there's no synchronization mechanism between blocks,
|
|
||||||
* we measure the clock once for each block. The clock samples are written to
|
|
||||||
* device memory.
|
|
||||||
*/
|
|
||||||
|
|
||||||
// System includes
|
|
||||||
#include <assert.h>
|
|
||||||
#include <stdint.h>
|
|
||||||
#include <stdio.h>
|
|
||||||
|
|
||||||
// CUDA runtime
|
|
||||||
#include <cuda_runtime.h>
|
|
||||||
|
|
||||||
// helper functions and utilities to work with CUDA
|
|
||||||
#include <helper_cuda.h>
|
|
||||||
#include <helper_functions.h>
|
|
||||||
|
|
||||||
// This kernel computes a standard parallel reduction and evaluates the
|
|
||||||
// time it takes to do that for each block. The timing results are stored
|
|
||||||
// in device memory.
|
|
||||||
__global__ static void timedReduction(const float *input, float *output,
|
|
||||||
clock_t *timer) {
|
|
||||||
// __shared__ float shared[2 * blockDim.x];
|
|
||||||
extern __shared__ float shared[];
|
|
||||||
|
|
||||||
const int tid = threadIdx.x;
|
|
||||||
const int bid = blockIdx.x;
|
|
||||||
|
|
||||||
if (tid == 0) timer[bid] = clock();
|
|
||||||
|
|
||||||
// Copy input.
|
|
||||||
shared[tid] = input[tid];
|
|
||||||
shared[tid + blockDim.x] = input[tid + blockDim.x];
|
|
||||||
|
|
||||||
// Perform reduction to find minimum.
|
|
||||||
for (int d = blockDim.x; d > 0; d /= 2) {
|
|
||||||
__syncthreads();
|
|
||||||
|
|
||||||
if (tid < d) {
|
|
||||||
float f0 = shared[tid];
|
|
||||||
float f1 = shared[tid + d];
|
|
||||||
|
|
||||||
if (f1 < f0) {
|
|
||||||
shared[tid] = f1;
|
|
||||||
}
|
|
||||||
}
|
|
||||||
}
|
|
||||||
|
|
||||||
// Write result.
|
|
||||||
if (tid == 0) output[bid] = shared[0];
|
|
||||||
|
|
||||||
__syncthreads();
|
|
||||||
|
|
||||||
if (tid == 0) timer[bid + gridDim.x] = clock();
|
|
||||||
}
|
|
||||||
|
|
||||||
#define NUM_BLOCKS 64
|
|
||||||
#define NUM_THREADS 256
|
|
||||||
|
|
||||||
// It's interesting to change the number of blocks and the number of threads to
|
|
||||||
// understand how to keep the hardware busy.
|
|
||||||
//
|
|
||||||
// Here are some numbers I get on my G80:
|
|
||||||
// blocks - clocks
|
|
||||||
// 1 - 3096
|
|
||||||
// 8 - 3232
|
|
||||||
// 16 - 3364
|
|
||||||
// 32 - 4615
|
|
||||||
// 64 - 9981
|
|
||||||
//
|
|
||||||
// With less than 16 blocks some of the multiprocessors of the device are idle.
|
|
||||||
// With more than 16 you are using all the multiprocessors, but there's only one
|
|
||||||
// block per multiprocessor and that doesn't allow you to hide the latency of
|
|
||||||
// the memory. With more than 32 the speed scales linearly.
|
|
||||||
|
|
||||||
// Start the main CUDA Sample here
|
|
||||||
int main(int argc, char **argv) {
|
|
||||||
printf("CUDA Clock sample\n");
|
|
||||||
|
|
||||||
// This will pick the best possible CUDA capable device
|
|
||||||
int dev = findCudaDevice(argc, (const char **)argv);
|
|
||||||
|
|
||||||
float *dinput = NULL;
|
|
||||||
float *doutput = NULL;
|
|
||||||
clock_t *dtimer = NULL;
|
|
||||||
|
|
||||||
clock_t timer[NUM_BLOCKS * 2];
|
|
||||||
float input[NUM_THREADS * 2];
|
|
||||||
|
|
||||||
for (int i = 0; i < NUM_THREADS * 2; i++) {
|
|
||||||
input[i] = (float)i;
|
|
||||||
}
|
|
||||||
|
|
||||||
checkCudaErrors(
|
|
||||||
cudaMalloc((void **)&dinput, sizeof(float) * NUM_THREADS * 2));
|
|
||||||
checkCudaErrors(cudaMalloc((void **)&doutput, sizeof(float) * NUM_BLOCKS));
|
|
||||||
checkCudaErrors(
|
|
||||||
cudaMalloc((void **)&dtimer, sizeof(clock_t) * NUM_BLOCKS * 2));
|
|
||||||
|
|
||||||
checkCudaErrors(cudaMemcpy(dinput, input, sizeof(float) * NUM_THREADS * 2,
|
|
||||||
cudaMemcpyHostToDevice));
|
|
||||||
|
|
||||||
timedReduction<<<NUM_BLOCKS, NUM_THREADS, sizeof(float) * 2 * NUM_THREADS>>>(
|
|
||||||
dinput, doutput, dtimer);
|
|
||||||
|
|
||||||
checkCudaErrors(cudaMemcpy(timer, dtimer, sizeof(clock_t) * NUM_BLOCKS * 2,
|
|
||||||
cudaMemcpyDeviceToHost));
|
|
||||||
|
|
||||||
checkCudaErrors(cudaFree(dinput));
|
|
||||||
checkCudaErrors(cudaFree(doutput));
|
|
||||||
checkCudaErrors(cudaFree(dtimer));
|
|
||||||
|
|
||||||
long double avgElapsedClocks = 0;
|
|
||||||
|
|
||||||
for (int i = 0; i < NUM_BLOCKS; i++) {
|
|
||||||
avgElapsedClocks += (long double)(timer[i + NUM_BLOCKS] - timer[i]);
|
|
||||||
}
|
|
||||||
|
|
||||||
avgElapsedClocks = avgElapsedClocks / NUM_BLOCKS;
|
|
||||||
printf("Average clocks/block = %Lf\n", avgElapsedClocks);
|
|
||||||
|
|
||||||
return EXIT_SUCCESS;
|
|
||||||
}
|
|
||||||
@@ -1,10 +0,0 @@
|
|||||||
{
|
|
||||||
"configurations": [
|
|
||||||
{
|
|
||||||
"name": "CUDA C++: Launch",
|
|
||||||
"type": "cuda-gdb",
|
|
||||||
"request": "launch",
|
|
||||||
"program": "${workspaceFolder}/clock_nvrtc"
|
|
||||||
}
|
|
||||||
]
|
|
||||||
}
|
|
||||||
@@ -1,15 +0,0 @@
|
|||||||
{
|
|
||||||
"version": "2.0.0",
|
|
||||||
"tasks": [
|
|
||||||
{
|
|
||||||
"label": "sample",
|
|
||||||
"type": "shell",
|
|
||||||
"command": "make dbg=1",
|
|
||||||
"problemMatcher": ["$nvcc"],
|
|
||||||
"group": {
|
|
||||||
"kind": "build",
|
|
||||||
"isDefault": true
|
|
||||||
}
|
|
||||||
}
|
|
||||||
]
|
|
||||||
}
|
|
||||||
@@ -1,37 +0,0 @@
|
|||||||
cmake_minimum_required(VERSION 3.20)
|
|
||||||
|
|
||||||
list(APPEND CMAKE_MODULE_PATH "${CMAKE_CURRENT_SOURCE_DIR}/../../../cmake/Modules")
|
|
||||||
|
|
||||||
project(clock_nvrtc LANGUAGES C CXX CUDA)
|
|
||||||
|
|
||||||
find_package(CUDAToolkit REQUIRED)
|
|
||||||
|
|
||||||
set(CMAKE_POSITION_INDEPENDENT_CODE ON)
|
|
||||||
|
|
||||||
set(CMAKE_CUDA_ARCHITECTURES 50 52 60 61 70 72 75 80 86 87 89 90 100 101 120)
|
|
||||||
set(CMAKE_CUDA_FLAGS "${CMAKE_CUDA_FLAGS} -Wno-deprecated-gpu-targets")
|
|
||||||
if(CMAKE_BUILD_TYPE STREQUAL "Debug")
|
|
||||||
# set(CMAKE_CUDA_FLAGS "${CMAKE_CUDA_FLAGS} -G") # enable cuda-gdb (expensive)
|
|
||||||
endif()
|
|
||||||
|
|
||||||
# Include directories and libraries
|
|
||||||
include_directories(../../../Common)
|
|
||||||
|
|
||||||
# Source file
|
|
||||||
# Add sample target executable
|
|
||||||
add_executable(clock_nvrtc clock.cpp)
|
|
||||||
|
|
||||||
target_compile_options(clock_nvrtc PRIVATE $<$<COMPILE_LANGUAGE:CUDA>:--extended-lambda>)
|
|
||||||
|
|
||||||
target_compile_features(clock_nvrtc PRIVATE cxx_std_17 cuda_std_17)
|
|
||||||
|
|
||||||
target_link_libraries(clock_nvrtc PRIVATE
|
|
||||||
CUDA::nvrtc
|
|
||||||
CUDA::cuda_driver
|
|
||||||
)
|
|
||||||
|
|
||||||
# Copy clock_kernel.cu to the output directory
|
|
||||||
add_custom_command(TARGET clock_nvrtc POST_BUILD
|
|
||||||
COMMAND ${CMAKE_COMMAND} -E copy_if_different
|
|
||||||
${CMAKE_CURRENT_SOURCE_DIR}/clock_kernel.cu ${CMAKE_CURRENT_BINARY_DIR}
|
|
||||||
)
|
|
||||||
@@ -1,39 +0,0 @@
|
|||||||
# clock_nvrtc - Clock libNVRTC
|
|
||||||
|
|
||||||
## Description
|
|
||||||
|
|
||||||
This example shows how to use the clock function using libNVRTC to measure the performance of block of threads of a kernel accurately.
|
|
||||||
|
|
||||||
## Key Concepts
|
|
||||||
|
|
||||||
Performance Strategies, Runtime Compilation
|
|
||||||
|
|
||||||
## Supported SM Architectures
|
|
||||||
|
|
||||||
[SM 5.0 ](https://developer.nvidia.com/cuda-gpus) [SM 5.2 ](https://developer.nvidia.com/cuda-gpus) [SM 5.3 ](https://developer.nvidia.com/cuda-gpus) [SM 6.0 ](https://developer.nvidia.com/cuda-gpus) [SM 6.1 ](https://developer.nvidia.com/cuda-gpus) [SM 7.0 ](https://developer.nvidia.com/cuda-gpus) [SM 7.2 ](https://developer.nvidia.com/cuda-gpus) [SM 7.5 ](https://developer.nvidia.com/cuda-gpus) [SM 8.0 ](https://developer.nvidia.com/cuda-gpus) [SM 8.6 ](https://developer.nvidia.com/cuda-gpus) [SM 8.7 ](https://developer.nvidia.com/cuda-gpus) [SM 8.9 ](https://developer.nvidia.com/cuda-gpus) [SM 9.0 ](https://developer.nvidia.com/cuda-gpus)
|
|
||||||
|
|
||||||
## Supported OSes
|
|
||||||
|
|
||||||
Linux, Windows, QNX
|
|
||||||
|
|
||||||
## Supported CPU Architecture
|
|
||||||
|
|
||||||
x86_64, aarch64
|
|
||||||
|
|
||||||
## CUDA APIs involved
|
|
||||||
|
|
||||||
### [CUDA Driver API](http://docs.nvidia.com/cuda/cuda-driver-api/index.html)
|
|
||||||
cuMemcpyDtoH, cuLaunchKernel, cuMemcpyHtoD, cuCtxSynchronize, cuMemAlloc, cuMemFree, cuModuleGetFunction
|
|
||||||
|
|
||||||
### [CUDA Runtime API](http://docs.nvidia.com/cuda/cuda-runtime-api/index.html)
|
|
||||||
cudaBlockSize, cudaGridSize
|
|
||||||
|
|
||||||
## Dependencies needed to build/run
|
|
||||||
[NVRTC](../../../README.md#nvrtc)
|
|
||||||
|
|
||||||
## Prerequisites
|
|
||||||
|
|
||||||
Download and install the [CUDA Toolkit 12.5](https://developer.nvidia.com/cuda-downloads) for your corresponding platform.
|
|
||||||
Make sure the dependencies mentioned in [Dependencies]() section above are installed.
|
|
||||||
|
|
||||||
## References (for more details)
|
|
||||||
@@ -1,134 +0,0 @@
|
|||||||
/* Copyright (c) 2022, NVIDIA CORPORATION. All rights reserved.
|
|
||||||
*
|
|
||||||
* Redistribution and use in source and binary forms, with or without
|
|
||||||
* modification, are permitted provided that the following conditions
|
|
||||||
* are met:
|
|
||||||
* * Redistributions of source code must retain the above copyright
|
|
||||||
* notice, this list of conditions and the following disclaimer.
|
|
||||||
* * Redistributions in binary form must reproduce the above copyright
|
|
||||||
* notice, this list of conditions and the following disclaimer in the
|
|
||||||
* documentation and/or other materials provided with the distribution.
|
|
||||||
* * Neither the name of NVIDIA CORPORATION nor the names of its
|
|
||||||
* contributors may be used to endorse or promote products derived
|
|
||||||
* from this software without specific prior written permission.
|
|
||||||
*
|
|
||||||
* THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS ``AS IS'' AND ANY
|
|
||||||
* EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
|
|
||||||
* IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR
|
|
||||||
* PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER OR
|
|
||||||
* CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL,
|
|
||||||
* EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO,
|
|
||||||
* PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR
|
|
||||||
* PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY
|
|
||||||
* OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
|
|
||||||
* (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
|
|
||||||
* OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
|
|
||||||
*/
|
|
||||||
|
|
||||||
/*
|
|
||||||
* This example shows how to use the clock function to measure the performance
|
|
||||||
* of block of threads of a kernel accurately. Blocks are executed in parallel
|
|
||||||
* and out of order. Since there's no synchronization mechanism between blocks,
|
|
||||||
* we measure the clock once for each block. The clock samples are written to
|
|
||||||
* device memory.
|
|
||||||
*/
|
|
||||||
|
|
||||||
// System includes
|
|
||||||
#include <stdio.h>
|
|
||||||
#include <stdint.h>
|
|
||||||
#include <assert.h>
|
|
||||||
|
|
||||||
#include <cuda_runtime.h>
|
|
||||||
#include <nvrtc_helper.h>
|
|
||||||
|
|
||||||
// helper functions and utilities to work with CUDA
|
|
||||||
#include <helper_functions.h>
|
|
||||||
|
|
||||||
#define NUM_BLOCKS 64
|
|
||||||
|
|
||||||
#define NUM_THREADS 256
|
|
||||||
|
|
||||||
// It's interesting to change the number of blocks and the number of threads to
|
|
||||||
// understand how to keep the hardware busy.
|
|
||||||
//
|
|
||||||
|
|
||||||
// Here are some numbers I get on my G80:
|
|
||||||
// blocks - clocks
|
|
||||||
// 1 - 3096
|
|
||||||
// 8 - 3232
|
|
||||||
// 16 - 3364
|
|
||||||
// 32 - 4615
|
|
||||||
// 64 - 9981
|
|
||||||
|
|
||||||
//
|
|
||||||
// With less than 16 blocks some of the multiprocessors of the device are idle.
|
|
||||||
// With
|
|
||||||
// more than 16 you are using all the multiprocessors, but there's only one
|
|
||||||
// block per
|
|
||||||
// multiprocessor and that doesn't allow you to hide the latency of the memory.
|
|
||||||
// With
|
|
||||||
// more than 32 the speed scales linearly.
|
|
||||||
|
|
||||||
// Start the main CUDA Sample here
|
|
||||||
|
|
||||||
int main(int argc, char **argv) {
|
|
||||||
printf("CUDA Clock sample\n");
|
|
||||||
|
|
||||||
typedef long clock_t;
|
|
||||||
|
|
||||||
clock_t timer[NUM_BLOCKS * 2];
|
|
||||||
|
|
||||||
float input[NUM_THREADS * 2];
|
|
||||||
|
|
||||||
for (int i = 0; i < NUM_THREADS * 2; i++) {
|
|
||||||
input[i] = (float)i;
|
|
||||||
}
|
|
||||||
|
|
||||||
char *cubin, *kernel_file;
|
|
||||||
size_t cubinSize;
|
|
||||||
|
|
||||||
kernel_file = sdkFindFilePath("clock_kernel.cu", argv[0]);
|
|
||||||
compileFileToCUBIN(kernel_file, argc, argv, &cubin, &cubinSize, 0);
|
|
||||||
|
|
||||||
CUmodule module = loadCUBIN(cubin, argc, argv);
|
|
||||||
CUfunction kernel_addr;
|
|
||||||
|
|
||||||
checkCudaErrors(cuModuleGetFunction(&kernel_addr, module, "timedReduction"));
|
|
||||||
|
|
||||||
dim3 cudaBlockSize(NUM_THREADS, 1, 1);
|
|
||||||
dim3 cudaGridSize(NUM_BLOCKS, 1, 1);
|
|
||||||
|
|
||||||
CUdeviceptr dinput, doutput, dtimer;
|
|
||||||
checkCudaErrors(cuMemAlloc(&dinput, sizeof(float) * NUM_THREADS * 2));
|
|
||||||
checkCudaErrors(cuMemAlloc(&doutput, sizeof(float) * NUM_BLOCKS));
|
|
||||||
checkCudaErrors(cuMemAlloc(&dtimer, sizeof(clock_t) * NUM_BLOCKS * 2));
|
|
||||||
checkCudaErrors(cuMemcpyHtoD(dinput, input, sizeof(float) * NUM_THREADS * 2));
|
|
||||||
|
|
||||||
void *arr[] = {(void *)&dinput, (void *)&doutput, (void *)&dtimer};
|
|
||||||
|
|
||||||
checkCudaErrors(cuLaunchKernel(
|
|
||||||
kernel_addr, cudaGridSize.x, cudaGridSize.y,
|
|
||||||
cudaGridSize.z, /* grid dim */
|
|
||||||
cudaBlockSize.x, cudaBlockSize.y, cudaBlockSize.z, /* block dim */
|
|
||||||
sizeof(float) * 2 * NUM_THREADS, 0, /* shared mem, stream */
|
|
||||||
&arr[0], /* arguments */
|
|
||||||
0));
|
|
||||||
|
|
||||||
checkCudaErrors(cuCtxSynchronize());
|
|
||||||
checkCudaErrors(
|
|
||||||
cuMemcpyDtoH(timer, dtimer, sizeof(clock_t) * NUM_BLOCKS * 2));
|
|
||||||
checkCudaErrors(cuMemFree(dinput));
|
|
||||||
checkCudaErrors(cuMemFree(doutput));
|
|
||||||
checkCudaErrors(cuMemFree(dtimer));
|
|
||||||
|
|
||||||
long double avgElapsedClocks = 0;
|
|
||||||
|
|
||||||
for (int i = 0; i < NUM_BLOCKS; i++) {
|
|
||||||
avgElapsedClocks += (long double)(timer[i + NUM_BLOCKS] - timer[i]);
|
|
||||||
}
|
|
||||||
|
|
||||||
avgElapsedClocks = avgElapsedClocks / NUM_BLOCKS;
|
|
||||||
printf("Average clocks/block = %Lf\n", avgElapsedClocks);
|
|
||||||
|
|
||||||
return EXIT_SUCCESS;
|
|
||||||
}
|
|
||||||
@@ -1,74 +0,0 @@
|
|||||||
/* Copyright (c) 2022, NVIDIA CORPORATION. All rights reserved.
|
|
||||||
*
|
|
||||||
* Redistribution and use in source and binary forms, with or without
|
|
||||||
* modification, are permitted provided that the following conditions
|
|
||||||
* are met:
|
|
||||||
* * Redistributions of source code must retain the above copyright
|
|
||||||
* notice, this list of conditions and the following disclaimer.
|
|
||||||
* * Redistributions in binary form must reproduce the above copyright
|
|
||||||
* notice, this list of conditions and the following disclaimer in the
|
|
||||||
* documentation and/or other materials provided with the distribution.
|
|
||||||
* * Neither the name of NVIDIA CORPORATION nor the names of its
|
|
||||||
* contributors may be used to endorse or promote products derived
|
|
||||||
* from this software without specific prior written permission.
|
|
||||||
*
|
|
||||||
* THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS ``AS IS'' AND ANY
|
|
||||||
* EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
|
|
||||||
* IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR
|
|
||||||
* PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER OR
|
|
||||||
* CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL,
|
|
||||||
* EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO,
|
|
||||||
* PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR
|
|
||||||
* PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY
|
|
||||||
* OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
|
|
||||||
* (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
|
|
||||||
* OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
|
|
||||||
*/
|
|
||||||
|
|
||||||
/*
|
|
||||||
* This example shows how to use the clock function to measure the performance
|
|
||||||
* of block of threads of a kernel accurately. Blocks are executed in parallel
|
|
||||||
* and out of order. Since there's no synchronization mechanism between blocks,
|
|
||||||
* we measure the clock once for each block. The clock samples are written to
|
|
||||||
* device memory.
|
|
||||||
*/
|
|
||||||
|
|
||||||
// This kernel computes a standard parallel reduction and evaluates the
|
|
||||||
// time it takes to do that for each block. The timing results are stored
|
|
||||||
// in device memory.
|
|
||||||
|
|
||||||
extern "C" __global__ void timedReduction(const float *input, float *output,
|
|
||||||
clock_t *timer) {
|
|
||||||
// __shared__ float shared[2 * blockDim.x];
|
|
||||||
extern __shared__ float shared[];
|
|
||||||
|
|
||||||
const int tid = threadIdx.x;
|
|
||||||
const int bid = blockIdx.x;
|
|
||||||
|
|
||||||
if (tid == 0) timer[bid] = clock();
|
|
||||||
|
|
||||||
// Copy input.
|
|
||||||
shared[tid] = input[tid];
|
|
||||||
shared[tid + blockDim.x] = input[tid + blockDim.x];
|
|
||||||
|
|
||||||
// Perform reduction to find minimum.
|
|
||||||
for (int d = blockDim.x; d > 0; d /= 2) {
|
|
||||||
__syncthreads();
|
|
||||||
|
|
||||||
if (tid < d) {
|
|
||||||
float f0 = shared[tid];
|
|
||||||
float f1 = shared[tid + d];
|
|
||||||
|
|
||||||
if (f1 < f0) {
|
|
||||||
shared[tid] = f1;
|
|
||||||
}
|
|
||||||
}
|
|
||||||
}
|
|
||||||
|
|
||||||
// Write result.
|
|
||||||
if (tid == 0) output[bid] = shared[0];
|
|
||||||
|
|
||||||
__syncthreads();
|
|
||||||
|
|
||||||
if (tid == 0) timer[bid + gridDim.x] = clock();
|
|
||||||
}
|
|
||||||
@@ -1,10 +0,0 @@
|
|||||||
{
|
|
||||||
"configurations": [
|
|
||||||
{
|
|
||||||
"name": "CUDA C++: Launch",
|
|
||||||
"type": "cuda-gdb",
|
|
||||||
"request": "launch",
|
|
||||||
"program": "${workspaceFolder}/cudaOpenMP"
|
|
||||||
}
|
|
||||||
]
|
|
||||||
}
|
|
||||||
@@ -1,15 +0,0 @@
|
|||||||
{
|
|
||||||
"version": "2.0.0",
|
|
||||||
"tasks": [
|
|
||||||
{
|
|
||||||
"label": "sample",
|
|
||||||
"type": "shell",
|
|
||||||
"command": "make dbg=1",
|
|
||||||
"problemMatcher": ["$nvcc"],
|
|
||||||
"group": {
|
|
||||||
"kind": "build",
|
|
||||||
"isDefault": true
|
|
||||||
}
|
|
||||||
}
|
|
||||||
]
|
|
||||||
}
|
|
||||||
@@ -1,37 +0,0 @@
|
|||||||
cmake_minimum_required(VERSION 3.20)
|
|
||||||
|
|
||||||
list(APPEND CMAKE_MODULE_PATH "${CMAKE_CURRENT_SOURCE_DIR}/../../../cmake/Modules")
|
|
||||||
|
|
||||||
project(cudaOpenMP LANGUAGES C CXX CUDA)
|
|
||||||
|
|
||||||
find_package(CUDAToolkit REQUIRED)
|
|
||||||
|
|
||||||
set(CMAKE_POSITION_INDEPENDENT_CODE ON)
|
|
||||||
|
|
||||||
set(CMAKE_CUDA_ARCHITECTURES 50 52 60 61 70 72 75 80 86 87 89 90 100 101 120)
|
|
||||||
set(CMAKE_CUDA_FLAGS "${CMAKE_CUDA_FLAGS} -Wno-deprecated-gpu-targets")
|
|
||||||
if(CMAKE_BUILD_TYPE STREQUAL "Debug")
|
|
||||||
# set(CMAKE_CUDA_FLAGS "${CMAKE_CUDA_FLAGS} -G") # enable cuda-gdb (expensive)
|
|
||||||
endif()
|
|
||||||
|
|
||||||
# Include directories and libraries
|
|
||||||
include_directories(../../../Common)
|
|
||||||
|
|
||||||
find_package(OpenMP)
|
|
||||||
|
|
||||||
# Source file
|
|
||||||
if(OpenMP_CXX_FOUND)
|
|
||||||
# Add target for asyncAPI
|
|
||||||
add_executable(cudaOpenMP cudaOpenMP.cu)
|
|
||||||
|
|
||||||
target_compile_options(cudaOpenMP PRIVATE $<$<COMPILE_LANGUAGE:CUDA>:--extended-lambda>)
|
|
||||||
|
|
||||||
target_compile_features(cudaOpenMP PRIVATE cxx_std_17 cuda_std_17)
|
|
||||||
|
|
||||||
set_target_properties(cudaOpenMP PROPERTIES CUDA_SEPARABLE_COMPILATION ON)
|
|
||||||
target_link_libraries(cudaOpenMP PUBLIC
|
|
||||||
OpenMP::OpenMP_CXX
|
|
||||||
)
|
|
||||||
else()
|
|
||||||
message(STATUS "OpenMP not found - will not build sample 'cudaOpenMP'")
|
|
||||||
endif()
|
|
||||||
@@ -1,36 +0,0 @@
|
|||||||
# cudaOpenMP - cudaOpenMP
|
|
||||||
|
|
||||||
## Description
|
|
||||||
|
|
||||||
This sample demonstrates how to use OpenMP API to write an application for multiple GPUs.
|
|
||||||
|
|
||||||
## Key Concepts
|
|
||||||
|
|
||||||
CUDA Systems Integration, OpenMP, Multithreading
|
|
||||||
|
|
||||||
## Supported SM Architectures
|
|
||||||
|
|
||||||
[SM 5.0 ](https://developer.nvidia.com/cuda-gpus) [SM 5.2 ](https://developer.nvidia.com/cuda-gpus) [SM 5.3 ](https://developer.nvidia.com/cuda-gpus) [SM 6.0 ](https://developer.nvidia.com/cuda-gpus) [SM 6.1 ](https://developer.nvidia.com/cuda-gpus) [SM 7.0 ](https://developer.nvidia.com/cuda-gpus) [SM 7.2 ](https://developer.nvidia.com/cuda-gpus) [SM 7.5 ](https://developer.nvidia.com/cuda-gpus) [SM 8.0 ](https://developer.nvidia.com/cuda-gpus) [SM 8.6 ](https://developer.nvidia.com/cuda-gpus) [SM 8.7 ](https://developer.nvidia.com/cuda-gpus) [SM 8.9 ](https://developer.nvidia.com/cuda-gpus) [SM 9.0 ](https://developer.nvidia.com/cuda-gpus)
|
|
||||||
|
|
||||||
## Supported OSes
|
|
||||||
|
|
||||||
Linux, Windows
|
|
||||||
|
|
||||||
## Supported CPU Architecture
|
|
||||||
|
|
||||||
x86_64, armv7l
|
|
||||||
|
|
||||||
## CUDA APIs involved
|
|
||||||
|
|
||||||
### [CUDA Runtime API](http://docs.nvidia.com/cuda/cuda-runtime-api/index.html)
|
|
||||||
cudaMemcpy, cudaGetErrorString, cudaFree, cudaGetLastError, cudaSetDevice, cudaGetDeviceCount, cudaGetDevice, cudaMemset, cudaMalloc, cudaGetDeviceProperties
|
|
||||||
|
|
||||||
## Dependencies needed to build/run
|
|
||||||
[OpenMP](../../../README.md#openmp)
|
|
||||||
|
|
||||||
## Prerequisites
|
|
||||||
|
|
||||||
Download and install the [CUDA Toolkit 12.5](https://developer.nvidia.com/cuda-downloads) for your corresponding platform.
|
|
||||||
Make sure the dependencies mentioned in [Dependencies]() section above are installed.
|
|
||||||
|
|
||||||
## References (for more details)
|
|
||||||
@@ -1,159 +0,0 @@
|
|||||||
/* Copyright (c) 2022, NVIDIA CORPORATION. All rights reserved.
|
|
||||||
*
|
|
||||||
* Redistribution and use in source and binary forms, with or without
|
|
||||||
* modification, are permitted provided that the following conditions
|
|
||||||
* are met:
|
|
||||||
* * Redistributions of source code must retain the above copyright
|
|
||||||
* notice, this list of conditions and the following disclaimer.
|
|
||||||
* * Redistributions in binary form must reproduce the above copyright
|
|
||||||
* notice, this list of conditions and the following disclaimer in the
|
|
||||||
* documentation and/or other materials provided with the distribution.
|
|
||||||
* * Neither the name of NVIDIA CORPORATION nor the names of its
|
|
||||||
* contributors may be used to endorse or promote products derived
|
|
||||||
* from this software without specific prior written permission.
|
|
||||||
*
|
|
||||||
* THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS ``AS IS'' AND ANY
|
|
||||||
* EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
|
|
||||||
* IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR
|
|
||||||
* PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER OR
|
|
||||||
* CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL,
|
|
||||||
* EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO,
|
|
||||||
* PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR
|
|
||||||
* PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY
|
|
||||||
* OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
|
|
||||||
* (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
|
|
||||||
* OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
|
|
||||||
*/
|
|
||||||
|
|
||||||
/*
|
|
||||||
* Multi-GPU sample using OpenMP for threading on the CPU side
|
|
||||||
* needs a compiler that supports OpenMP 2.0
|
|
||||||
*/
|
|
||||||
|
|
||||||
#include <helper_cuda.h>
|
|
||||||
#include <omp.h>
|
|
||||||
#include <stdio.h> // stdio functions are used since C++ streams aren't necessarily thread safe
|
|
||||||
|
|
||||||
using namespace std;
|
|
||||||
|
|
||||||
// a simple kernel that simply increments each array element by b
|
|
||||||
__global__ void kernelAddConstant(int *g_a, const int b) {
|
|
||||||
int idx = blockIdx.x * blockDim.x + threadIdx.x;
|
|
||||||
g_a[idx] += b;
|
|
||||||
}
|
|
||||||
|
|
||||||
// a predicate that checks whether each array element is set to its index plus b
|
|
||||||
int correctResult(int *data, const int n, const int b) {
|
|
||||||
for (int i = 0; i < n; i++)
|
|
||||||
if (data[i] != i + b) return 0;
|
|
||||||
|
|
||||||
return 1;
|
|
||||||
}
|
|
||||||
|
|
||||||
int main(int argc, char *argv[]) {
|
|
||||||
int num_gpus = 0; // number of CUDA GPUs
|
|
||||||
|
|
||||||
printf("%s Starting...\n\n", argv[0]);
|
|
||||||
|
|
||||||
/////////////////////////////////////////////////////////////////
|
|
||||||
// determine the number of CUDA capable GPUs
|
|
||||||
//
|
|
||||||
cudaGetDeviceCount(&num_gpus);
|
|
||||||
|
|
||||||
if (num_gpus < 1) {
|
|
||||||
printf("no CUDA capable devices were detected\n");
|
|
||||||
return 1;
|
|
||||||
}
|
|
||||||
|
|
||||||
/////////////////////////////////////////////////////////////////
|
|
||||||
// display CPU and GPU configuration
|
|
||||||
//
|
|
||||||
printf("number of host CPUs:\t%d\n", omp_get_num_procs());
|
|
||||||
printf("number of CUDA devices:\t%d\n", num_gpus);
|
|
||||||
|
|
||||||
for (int i = 0; i < num_gpus; i++) {
|
|
||||||
cudaDeviceProp dprop;
|
|
||||||
cudaGetDeviceProperties(&dprop, i);
|
|
||||||
printf(" %d: %s\n", i, dprop.name);
|
|
||||||
}
|
|
||||||
|
|
||||||
printf("---------------------------\n");
|
|
||||||
|
|
||||||
/////////////////////////////////////////////////////////////////
|
|
||||||
// initialize data
|
|
||||||
//
|
|
||||||
unsigned int n = num_gpus * 8192;
|
|
||||||
unsigned int nbytes = n * sizeof(int);
|
|
||||||
int *a = 0; // pointer to data on the CPU
|
|
||||||
int b = 3; // value by which the array is incremented
|
|
||||||
a = (int *)malloc(nbytes);
|
|
||||||
|
|
||||||
if (0 == a) {
|
|
||||||
printf("couldn't allocate CPU memory\n");
|
|
||||||
return 1;
|
|
||||||
}
|
|
||||||
|
|
||||||
for (unsigned int i = 0; i < n; i++) a[i] = i;
|
|
||||||
|
|
||||||
////////////////////////////////////////////////////////////////
|
|
||||||
// run as many CPU threads as there are CUDA devices
|
|
||||||
// each CPU thread controls a different device, processing its
|
|
||||||
// portion of the data. It's possible to use more CPU threads
|
|
||||||
// than there are CUDA devices, in which case several CPU
|
|
||||||
// threads will be allocating resources and launching kernels
|
|
||||||
// on the same device. For example, try omp_set_num_threads(2*num_gpus);
|
|
||||||
// Recall that all variables declared inside an "omp parallel" scope are
|
|
||||||
// local to each CPU thread
|
|
||||||
//
|
|
||||||
omp_set_num_threads(
|
|
||||||
num_gpus); // create as many CPU threads as there are CUDA devices
|
|
||||||
// omp_set_num_threads(2*num_gpus);// create twice as many CPU threads as there
|
|
||||||
// are CUDA devices
|
|
||||||
#pragma omp parallel
|
|
||||||
{
|
|
||||||
unsigned int cpu_thread_id = omp_get_thread_num();
|
|
||||||
unsigned int num_cpu_threads = omp_get_num_threads();
|
|
||||||
|
|
||||||
// set and check the CUDA device for this CPU thread
|
|
||||||
int gpu_id = -1;
|
|
||||||
checkCudaErrors(cudaSetDevice(
|
|
||||||
cpu_thread_id %
|
|
||||||
num_gpus)); // "% num_gpus" allows more CPU threads than GPU devices
|
|
||||||
checkCudaErrors(cudaGetDevice(&gpu_id));
|
|
||||||
printf("CPU thread %d (of %d) uses CUDA device %d\n", cpu_thread_id,
|
|
||||||
num_cpu_threads, gpu_id);
|
|
||||||
|
|
||||||
int *d_a =
|
|
||||||
0; // pointer to memory on the device associated with this CPU thread
|
|
||||||
int *sub_a =
|
|
||||||
a +
|
|
||||||
cpu_thread_id * n /
|
|
||||||
num_cpu_threads; // pointer to this CPU thread's portion of data
|
|
||||||
unsigned int nbytes_per_kernel = nbytes / num_cpu_threads;
|
|
||||||
dim3 gpu_threads(128); // 128 threads per block
|
|
||||||
dim3 gpu_blocks(n / (gpu_threads.x * num_cpu_threads));
|
|
||||||
|
|
||||||
checkCudaErrors(cudaMalloc((void **)&d_a, nbytes_per_kernel));
|
|
||||||
checkCudaErrors(cudaMemset(d_a, 0, nbytes_per_kernel));
|
|
||||||
checkCudaErrors(
|
|
||||||
cudaMemcpy(d_a, sub_a, nbytes_per_kernel, cudaMemcpyHostToDevice));
|
|
||||||
kernelAddConstant<<<gpu_blocks, gpu_threads>>>(d_a, b);
|
|
||||||
|
|
||||||
checkCudaErrors(
|
|
||||||
cudaMemcpy(sub_a, d_a, nbytes_per_kernel, cudaMemcpyDeviceToHost));
|
|
||||||
checkCudaErrors(cudaFree(d_a));
|
|
||||||
}
|
|
||||||
printf("---------------------------\n");
|
|
||||||
|
|
||||||
if (cudaSuccess != cudaGetLastError())
|
|
||||||
printf("%s\n", cudaGetErrorString(cudaGetLastError()));
|
|
||||||
|
|
||||||
////////////////////////////////////////////////////////////////
|
|
||||||
// check the result
|
|
||||||
//
|
|
||||||
bool bResult = correctResult(a, n, b);
|
|
||||||
|
|
||||||
if (a) free(a); // free CPU memory
|
|
||||||
|
|
||||||
exit(bResult ? EXIT_SUCCESS : EXIT_FAILURE);
|
|
||||||
}
|
|
||||||
@@ -1,10 +0,0 @@
|
|||||||
{
|
|
||||||
"configurations": [
|
|
||||||
{
|
|
||||||
"name": "CUDA C++: Launch",
|
|
||||||
"type": "cuda-gdb",
|
|
||||||
"request": "launch",
|
|
||||||
"program": "${workspaceFolder}/fp16ScalarProduct"
|
|
||||||
}
|
|
||||||
]
|
|
||||||
}
|
|
||||||
@@ -1,15 +0,0 @@
|
|||||||
{
|
|
||||||
"version": "2.0.0",
|
|
||||||
"tasks": [
|
|
||||||
{
|
|
||||||
"label": "sample",
|
|
||||||
"type": "shell",
|
|
||||||
"command": "make dbg=1",
|
|
||||||
"problemMatcher": ["$nvcc"],
|
|
||||||
"group": {
|
|
||||||
"kind": "build",
|
|
||||||
"isDefault": true
|
|
||||||
}
|
|
||||||
}
|
|
||||||
]
|
|
||||||
}
|
|
||||||
@@ -1,27 +0,0 @@
|
|||||||
cmake_minimum_required(VERSION 3.20)
|
|
||||||
|
|
||||||
list(APPEND CMAKE_MODULE_PATH "${CMAKE_CURRENT_SOURCE_DIR}/../../../cmake/Modules")
|
|
||||||
|
|
||||||
project(fp16ScalarProduct LANGUAGES C CXX CUDA)
|
|
||||||
|
|
||||||
find_package(CUDAToolkit REQUIRED)
|
|
||||||
|
|
||||||
set(CMAKE_POSITION_INDEPENDENT_CODE ON)
|
|
||||||
|
|
||||||
set(CMAKE_CUDA_ARCHITECTURES 60 61 70 72 75 80 86 87 89 90 100 101 120)
|
|
||||||
if(CMAKE_BUILD_TYPE STREQUAL "Debug")
|
|
||||||
# set(CMAKE_CUDA_FLAGS "${CMAKE_CUDA_FLAGS} -G") # enable cuda-gdb (expensive)
|
|
||||||
endif()
|
|
||||||
|
|
||||||
# Include directories and libraries
|
|
||||||
include_directories(../../../Common)
|
|
||||||
|
|
||||||
# Source file
|
|
||||||
# Add target for asyncAPI
|
|
||||||
add_executable(fp16ScalarProduct fp16ScalarProduct.cu)
|
|
||||||
|
|
||||||
target_compile_options(fp16ScalarProduct PRIVATE $<$<COMPILE_LANGUAGE:CUDA>:--extended-lambda>)
|
|
||||||
|
|
||||||
target_compile_features(fp16ScalarProduct PRIVATE cxx_std_17 cuda_std_17)
|
|
||||||
|
|
||||||
set_target_properties(fp16ScalarProduct PROPERTIES CUDA_SEPARABLE_COMPILATION ON)
|
|
||||||
@@ -1,36 +0,0 @@
|
|||||||
# fp16ScalarProduct - FP16 Scalar Product
|
|
||||||
|
|
||||||
## Description
|
|
||||||
|
|
||||||
Calculates scalar product of two vectors of FP16 numbers.
|
|
||||||
|
|
||||||
## Key Concepts
|
|
||||||
|
|
||||||
CUDA Runtime API
|
|
||||||
|
|
||||||
## Supported SM Architectures
|
|
||||||
|
|
||||||
[SM 6.0 ](https://developer.nvidia.com/cuda-gpus) [SM 6.1 ](https://developer.nvidia.com/cuda-gpus) [SM 7.0 ](https://developer.nvidia.com/cuda-gpus) [SM 7.2 ](https://developer.nvidia.com/cuda-gpus) [SM 7.5 ](https://developer.nvidia.com/cuda-gpus) [SM 8.0 ](https://developer.nvidia.com/cuda-gpus) [SM 8.6 ](https://developer.nvidia.com/cuda-gpus) [SM 8.7 ](https://developer.nvidia.com/cuda-gpus) [SM 8.9 ](https://developer.nvidia.com/cuda-gpus) [SM 9.0 ](https://developer.nvidia.com/cuda-gpus)
|
|
||||||
|
|
||||||
## Supported OSes
|
|
||||||
|
|
||||||
Linux, Windows
|
|
||||||
|
|
||||||
## Supported CPU Architecture
|
|
||||||
|
|
||||||
x86_64, armv7l
|
|
||||||
|
|
||||||
## CUDA APIs involved
|
|
||||||
|
|
||||||
### [CUDA Runtime API](http://docs.nvidia.com/cuda/cuda-runtime-api/index.html)
|
|
||||||
cudaMemcpy, cudaFree, cudaMallocHost, cudaFreeHost, cudaMalloc, cudaGetDeviceProperties
|
|
||||||
|
|
||||||
## Dependencies needed to build/run
|
|
||||||
[FP16](../../../README.md#fp16)
|
|
||||||
|
|
||||||
## Prerequisites
|
|
||||||
|
|
||||||
Download and install the [CUDA Toolkit 12.5](https://developer.nvidia.com/cuda-downloads) for your corresponding platform.
|
|
||||||
Make sure the dependencies mentioned in [Dependencies]() section above are installed.
|
|
||||||
|
|
||||||
## References (for more details)
|
|
||||||
@@ -1,215 +0,0 @@
|
|||||||
/* Copyright (c) 2022, NVIDIA CORPORATION. All rights reserved.
|
|
||||||
*
|
|
||||||
* Redistribution and use in source and binary forms, with or without
|
|
||||||
* modification, are permitted provided that the following conditions
|
|
||||||
* are met:
|
|
||||||
* * Redistributions of source code must retain the above copyright
|
|
||||||
* notice, this list of conditions and the following disclaimer.
|
|
||||||
* * Redistributions in binary form must reproduce the above copyright
|
|
||||||
* notice, this list of conditions and the following disclaimer in the
|
|
||||||
* documentation and/or other materials provided with the distribution.
|
|
||||||
* * Neither the name of NVIDIA CORPORATION nor the names of its
|
|
||||||
* contributors may be used to endorse or promote products derived
|
|
||||||
* from this software without specific prior written permission.
|
|
||||||
*
|
|
||||||
* THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS ``AS IS'' AND ANY
|
|
||||||
* EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
|
|
||||||
* IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR
|
|
||||||
* PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER OR
|
|
||||||
* CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL,
|
|
||||||
* EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO,
|
|
||||||
* PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR
|
|
||||||
* PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY
|
|
||||||
* OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
|
|
||||||
* (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
|
|
||||||
* OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
|
|
||||||
*/
|
|
||||||
|
|
||||||
#include "cuda_fp16.h"
|
|
||||||
#include "helper_cuda.h"
|
|
||||||
|
|
||||||
#include <cstdio>
|
|
||||||
#include <cstdlib>
|
|
||||||
#include <ctime>
|
|
||||||
|
|
||||||
#define NUM_OF_BLOCKS 128
|
|
||||||
#define NUM_OF_THREADS 128
|
|
||||||
|
|
||||||
__forceinline__ __device__ void reduceInShared_intrinsics(half2 *const v) {
|
|
||||||
if (threadIdx.x < 64)
|
|
||||||
v[threadIdx.x] = __hadd2(v[threadIdx.x], v[threadIdx.x + 64]);
|
|
||||||
__syncthreads();
|
|
||||||
if (threadIdx.x < 32)
|
|
||||||
v[threadIdx.x] = __hadd2(v[threadIdx.x], v[threadIdx.x + 32]);
|
|
||||||
__syncthreads();
|
|
||||||
if (threadIdx.x < 16)
|
|
||||||
v[threadIdx.x] = __hadd2(v[threadIdx.x], v[threadIdx.x + 16]);
|
|
||||||
__syncthreads();
|
|
||||||
if (threadIdx.x < 8)
|
|
||||||
v[threadIdx.x] = __hadd2(v[threadIdx.x], v[threadIdx.x + 8]);
|
|
||||||
__syncthreads();
|
|
||||||
if (threadIdx.x < 4)
|
|
||||||
v[threadIdx.x] = __hadd2(v[threadIdx.x], v[threadIdx.x + 4]);
|
|
||||||
__syncthreads();
|
|
||||||
if (threadIdx.x < 2)
|
|
||||||
v[threadIdx.x] = __hadd2(v[threadIdx.x], v[threadIdx.x + 2]);
|
|
||||||
__syncthreads();
|
|
||||||
if (threadIdx.x < 1)
|
|
||||||
v[threadIdx.x] = __hadd2(v[threadIdx.x], v[threadIdx.x + 1]);
|
|
||||||
__syncthreads();
|
|
||||||
}
|
|
||||||
|
|
||||||
__forceinline__ __device__ void reduceInShared_native(half2 *const v) {
|
|
||||||
if (threadIdx.x < 64) v[threadIdx.x] = v[threadIdx.x] + v[threadIdx.x + 64];
|
|
||||||
__syncthreads();
|
|
||||||
if (threadIdx.x < 32) v[threadIdx.x] = v[threadIdx.x] + v[threadIdx.x + 32];
|
|
||||||
__syncthreads();
|
|
||||||
if (threadIdx.x < 16) v[threadIdx.x] = v[threadIdx.x] + v[threadIdx.x + 16];
|
|
||||||
__syncthreads();
|
|
||||||
if (threadIdx.x < 8) v[threadIdx.x] = v[threadIdx.x] + v[threadIdx.x + 8];
|
|
||||||
__syncthreads();
|
|
||||||
if (threadIdx.x < 4) v[threadIdx.x] = v[threadIdx.x] + v[threadIdx.x + 4];
|
|
||||||
__syncthreads();
|
|
||||||
if (threadIdx.x < 2) v[threadIdx.x] = v[threadIdx.x] + v[threadIdx.x + 2];
|
|
||||||
__syncthreads();
|
|
||||||
if (threadIdx.x < 1) v[threadIdx.x] = v[threadIdx.x] + v[threadIdx.x + 1];
|
|
||||||
__syncthreads();
|
|
||||||
}
|
|
||||||
|
|
||||||
__global__ void scalarProductKernel_intrinsics(half2 const *const a,
|
|
||||||
half2 const *const b,
|
|
||||||
float *const results,
|
|
||||||
size_t const size) {
|
|
||||||
const int stride = gridDim.x * blockDim.x;
|
|
||||||
__shared__ half2 shArray[NUM_OF_THREADS];
|
|
||||||
|
|
||||||
shArray[threadIdx.x] = __float2half2_rn(0.f);
|
|
||||||
half2 value = __float2half2_rn(0.f);
|
|
||||||
|
|
||||||
for (int i = threadIdx.x + blockDim.x + blockIdx.x; i < size; i += stride) {
|
|
||||||
value = __hfma2(a[i], b[i], value);
|
|
||||||
}
|
|
||||||
|
|
||||||
shArray[threadIdx.x] = value;
|
|
||||||
__syncthreads();
|
|
||||||
reduceInShared_intrinsics(shArray);
|
|
||||||
|
|
||||||
if (threadIdx.x == 0) {
|
|
||||||
half2 result = shArray[0];
|
|
||||||
float f_result = __low2float(result) + __high2float(result);
|
|
||||||
results[blockIdx.x] = f_result;
|
|
||||||
}
|
|
||||||
}
|
|
||||||
|
|
||||||
__global__ void scalarProductKernel_native(half2 const *const a,
|
|
||||||
half2 const *const b,
|
|
||||||
float *const results,
|
|
||||||
size_t const size) {
|
|
||||||
const int stride = gridDim.x * blockDim.x;
|
|
||||||
__shared__ half2 shArray[NUM_OF_THREADS];
|
|
||||||
|
|
||||||
half2 value(0.f, 0.f);
|
|
||||||
shArray[threadIdx.x] = value;
|
|
||||||
|
|
||||||
for (int i = threadIdx.x + blockDim.x + blockIdx.x; i < size; i += stride) {
|
|
||||||
value = a[i] * b[i] + value;
|
|
||||||
}
|
|
||||||
|
|
||||||
shArray[threadIdx.x] = value;
|
|
||||||
__syncthreads();
|
|
||||||
reduceInShared_native(shArray);
|
|
||||||
|
|
||||||
if (threadIdx.x == 0) {
|
|
||||||
half2 result = shArray[0];
|
|
||||||
float f_result = (float)result.y + (float)result.x;
|
|
||||||
results[blockIdx.x] = f_result;
|
|
||||||
}
|
|
||||||
}
|
|
||||||
|
|
||||||
void generateInput(half2 *a, size_t size) {
|
|
||||||
for (size_t i = 0; i < size; ++i) {
|
|
||||||
half2 temp;
|
|
||||||
temp.x = static_cast<float>(rand() % 4);
|
|
||||||
temp.y = static_cast<float>(rand() % 2);
|
|
||||||
a[i] = temp;
|
|
||||||
}
|
|
||||||
}
|
|
||||||
|
|
||||||
int main(int argc, char *argv[]) {
|
|
||||||
srand((unsigned int)time(NULL));
|
|
||||||
size_t size = NUM_OF_BLOCKS * NUM_OF_THREADS * 16;
|
|
||||||
|
|
||||||
half2 *vec[2];
|
|
||||||
half2 *devVec[2];
|
|
||||||
|
|
||||||
float *results;
|
|
||||||
float *devResults;
|
|
||||||
|
|
||||||
int devID = findCudaDevice(argc, (const char **)argv);
|
|
||||||
|
|
||||||
cudaDeviceProp devProp;
|
|
||||||
checkCudaErrors(cudaGetDeviceProperties(&devProp, devID));
|
|
||||||
|
|
||||||
if (devProp.major < 5 || (devProp.major == 5 && devProp.minor < 3)) {
|
|
||||||
printf(
|
|
||||||
"ERROR: fp16ScalarProduct requires GPU devices with compute SM 5.3 or "
|
|
||||||
"higher.\n");
|
|
||||||
return EXIT_WAIVED;
|
|
||||||
}
|
|
||||||
|
|
||||||
for (int i = 0; i < 2; ++i) {
|
|
||||||
checkCudaErrors(cudaMallocHost((void **)&vec[i], size * sizeof *vec[i]));
|
|
||||||
checkCudaErrors(cudaMalloc((void **)&devVec[i], size * sizeof *devVec[i]));
|
|
||||||
}
|
|
||||||
|
|
||||||
checkCudaErrors(
|
|
||||||
cudaMallocHost((void **)&results, NUM_OF_BLOCKS * sizeof *results));
|
|
||||||
checkCudaErrors(
|
|
||||||
cudaMalloc((void **)&devResults, NUM_OF_BLOCKS * sizeof *devResults));
|
|
||||||
|
|
||||||
for (int i = 0; i < 2; ++i) {
|
|
||||||
generateInput(vec[i], size);
|
|
||||||
checkCudaErrors(cudaMemcpy(devVec[i], vec[i], size * sizeof *vec[i],
|
|
||||||
cudaMemcpyHostToDevice));
|
|
||||||
}
|
|
||||||
|
|
||||||
scalarProductKernel_native<<<NUM_OF_BLOCKS, NUM_OF_THREADS>>>(
|
|
||||||
devVec[0], devVec[1], devResults, size);
|
|
||||||
|
|
||||||
checkCudaErrors(cudaMemcpy(results, devResults,
|
|
||||||
NUM_OF_BLOCKS * sizeof *results,
|
|
||||||
cudaMemcpyDeviceToHost));
|
|
||||||
|
|
||||||
float result_native = 0;
|
|
||||||
for (int i = 0; i < NUM_OF_BLOCKS; ++i) {
|
|
||||||
result_native += results[i];
|
|
||||||
}
|
|
||||||
printf("Result native operators\t: %f \n", result_native);
|
|
||||||
|
|
||||||
scalarProductKernel_intrinsics<<<NUM_OF_BLOCKS, NUM_OF_THREADS>>>(
|
|
||||||
devVec[0], devVec[1], devResults, size);
|
|
||||||
|
|
||||||
checkCudaErrors(cudaMemcpy(results, devResults,
|
|
||||||
NUM_OF_BLOCKS * sizeof *results,
|
|
||||||
cudaMemcpyDeviceToHost));
|
|
||||||
|
|
||||||
float result_intrinsics = 0;
|
|
||||||
for (int i = 0; i < NUM_OF_BLOCKS; ++i) {
|
|
||||||
result_intrinsics += results[i];
|
|
||||||
}
|
|
||||||
printf("Result intrinsics\t: %f \n", result_intrinsics);
|
|
||||||
|
|
||||||
printf("&&&& fp16ScalarProduct %s\n",
|
|
||||||
(fabs(result_intrinsics - result_native) < 0.00001) ? "PASSED"
|
|
||||||
: "FAILED");
|
|
||||||
|
|
||||||
for (int i = 0; i < 2; ++i) {
|
|
||||||
checkCudaErrors(cudaFree(devVec[i]));
|
|
||||||
checkCudaErrors(cudaFreeHost(vec[i]));
|
|
||||||
}
|
|
||||||
|
|
||||||
checkCudaErrors(cudaFree(devResults));
|
|
||||||
checkCudaErrors(cudaFreeHost(results));
|
|
||||||
|
|
||||||
return EXIT_SUCCESS;
|
|
||||||
}
|
|
||||||
@@ -1,10 +0,0 @@
|
|||||||
{
|
|
||||||
"configurations": [
|
|
||||||
{
|
|
||||||
"name": "CUDA C++: Launch",
|
|
||||||
"type": "cuda-gdb",
|
|
||||||
"request": "launch",
|
|
||||||
"program": "${workspaceFolder}/matrixMul"
|
|
||||||
}
|
|
||||||
]
|
|
||||||
}
|
|
||||||
@@ -1,15 +0,0 @@
|
|||||||
{
|
|
||||||
"version": "2.0.0",
|
|
||||||
"tasks": [
|
|
||||||
{
|
|
||||||
"label": "sample",
|
|
||||||
"type": "shell",
|
|
||||||
"command": "make dbg=1",
|
|
||||||
"problemMatcher": ["$nvcc"],
|
|
||||||
"group": {
|
|
||||||
"kind": "build",
|
|
||||||
"isDefault": true
|
|
||||||
}
|
|
||||||
}
|
|
||||||
]
|
|
||||||
}
|
|
||||||
@@ -1,28 +0,0 @@
|
|||||||
cmake_minimum_required(VERSION 3.20)
|
|
||||||
|
|
||||||
list(APPEND CMAKE_MODULE_PATH "${CMAKE_CURRENT_SOURCE_DIR}/../../../cmake/Modules")
|
|
||||||
|
|
||||||
project(matrixMul LANGUAGES C CXX CUDA)
|
|
||||||
|
|
||||||
find_package(CUDAToolkit REQUIRED)
|
|
||||||
|
|
||||||
set(CMAKE_POSITION_INDEPENDENT_CODE ON)
|
|
||||||
|
|
||||||
set(CMAKE_CUDA_ARCHITECTURES 50 52 60 61 70 72 75 80 86 87 89 90 100 101 120)
|
|
||||||
set(CMAKE_CUDA_FLAGS "${CMAKE_CUDA_FLAGS} -Wno-deprecated-gpu-targets")
|
|
||||||
if(CMAKE_BUILD_TYPE STREQUAL "Debug")
|
|
||||||
# set(CMAKE_CUDA_FLAGS "${CMAKE_CUDA_FLAGS} -G") # enable cuda-gdb (expensive)
|
|
||||||
endif()
|
|
||||||
|
|
||||||
# Include directories and libraries
|
|
||||||
include_directories(../../../Common)
|
|
||||||
|
|
||||||
# Source file
|
|
||||||
# Add target for asyncAPI
|
|
||||||
add_executable(matrixMul matrixMul.cu)
|
|
||||||
|
|
||||||
target_compile_options(matrixMul PRIVATE $<$<COMPILE_LANGUAGE:CUDA>:--extended-lambda>)
|
|
||||||
|
|
||||||
target_compile_features(matrixMul PRIVATE cxx_std_17 cuda_std_17)
|
|
||||||
|
|
||||||
set_target_properties(matrixMul PROPERTIES CUDA_SEPARABLE_COMPILATION ON)
|
|
||||||
@@ -1,32 +0,0 @@
|
|||||||
# matrixMul - Matrix Multiplication (CUDA Runtime API Version)
|
|
||||||
|
|
||||||
## Description
|
|
||||||
|
|
||||||
This sample implements matrix multiplication and is exactly the same as Chapter 6 of the programming guide. It has been written for clarity of exposition to illustrate various CUDA programming principles, not with the goal of providing the most performant generic kernel for matrix multiplication. To illustrate GPU performance for matrix multiply, this sample also shows how to use the new CUDA 4.0 interface for CUBLAS to demonstrate high-performance performance for matrix multiplication.
|
|
||||||
|
|
||||||
## Key Concepts
|
|
||||||
|
|
||||||
CUDA Runtime API, Linear Algebra
|
|
||||||
|
|
||||||
## Supported SM Architectures
|
|
||||||
|
|
||||||
[SM 5.0 ](https://developer.nvidia.com/cuda-gpus) [SM 5.2 ](https://developer.nvidia.com/cuda-gpus) [SM 5.3 ](https://developer.nvidia.com/cuda-gpus) [SM 6.0 ](https://developer.nvidia.com/cuda-gpus) [SM 6.1 ](https://developer.nvidia.com/cuda-gpus) [SM 7.0 ](https://developer.nvidia.com/cuda-gpus) [SM 7.2 ](https://developer.nvidia.com/cuda-gpus) [SM 7.5 ](https://developer.nvidia.com/cuda-gpus) [SM 8.0 ](https://developer.nvidia.com/cuda-gpus) [SM 8.6 ](https://developer.nvidia.com/cuda-gpus) [SM 8.7 ](https://developer.nvidia.com/cuda-gpus) [SM 8.9 ](https://developer.nvidia.com/cuda-gpus) [SM 9.0 ](https://developer.nvidia.com/cuda-gpus)
|
|
||||||
|
|
||||||
## Supported OSes
|
|
||||||
|
|
||||||
Linux, Windows
|
|
||||||
|
|
||||||
## Supported CPU Architecture
|
|
||||||
|
|
||||||
x86_64, armv7l, aarch64
|
|
||||||
|
|
||||||
## CUDA APIs involved
|
|
||||||
|
|
||||||
### [CUDA Runtime API](http://docs.nvidia.com/cuda/cuda-runtime-api/index.html)
|
|
||||||
cudaStreamCreateWithFlags, cudaProfilerStop, cudaMalloc, cudaFree, cudaMallocHost, cudaProfilerStart, cudaEventSynchronize, cudaEventRecord, cudaFreeHost, cudaStreamSynchronize, cudaEventDestroy, cudaEventElapsedTime, cudaMemcpyAsync, cudaEventCreate
|
|
||||||
|
|
||||||
## Prerequisites
|
|
||||||
|
|
||||||
Download and install the [CUDA Toolkit 12.5](https://developer.nvidia.com/cuda-downloads) for your corresponding platform.
|
|
||||||
|
|
||||||
## References (for more details)
|
|
||||||
@@ -1,353 +0,0 @@
|
|||||||
/* Copyright (c) 2022, NVIDIA CORPORATION. All rights reserved.
|
|
||||||
*
|
|
||||||
* Redistribution and use in source and binary forms, with or without
|
|
||||||
* modification, are permitted provided that the following conditions
|
|
||||||
* are met:
|
|
||||||
* * Redistributions of source code must retain the above copyright
|
|
||||||
* notice, this list of conditions and the following disclaimer.
|
|
||||||
* * Redistributions in binary form must reproduce the above copyright
|
|
||||||
* notice, this list of conditions and the following disclaimer in the
|
|
||||||
* documentation and/or other materials provided with the distribution.
|
|
||||||
* * Neither the name of NVIDIA CORPORATION nor the names of its
|
|
||||||
* contributors may be used to endorse or promote products derived
|
|
||||||
* from this software without specific prior written permission.
|
|
||||||
*
|
|
||||||
* THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS ``AS IS'' AND ANY
|
|
||||||
* EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
|
|
||||||
* IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR
|
|
||||||
* PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER OR
|
|
||||||
* CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL,
|
|
||||||
* EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO,
|
|
||||||
* PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR
|
|
||||||
* PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY
|
|
||||||
* OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
|
|
||||||
* (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
|
|
||||||
* OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
|
|
||||||
*/
|
|
||||||
|
|
||||||
/**
|
|
||||||
* Matrix multiplication: C = A * B.
|
|
||||||
* Host code.
|
|
||||||
*
|
|
||||||
* This sample implements matrix multiplication which makes use of shared memory
|
|
||||||
* to ensure data reuse, the matrix multiplication is done using tiling approach.
|
|
||||||
* It has been written for clarity of exposition to illustrate various CUDA programming
|
|
||||||
* principles, not with the goal of providing the most performant generic kernel for matrix multiplication.
|
|
||||||
* See also:
|
|
||||||
* V. Volkov and J. Demmel, "Benchmarking GPUs to tune dense linear algebra,"
|
|
||||||
* in Proc. 2008 ACM/IEEE Conf. on Supercomputing (SC '08),
|
|
||||||
* Piscataway, NJ: IEEE Press, 2008, pp. Art. 31:1-11.
|
|
||||||
*/
|
|
||||||
|
|
||||||
// System includes
|
|
||||||
#include <stdio.h>
|
|
||||||
#include <assert.h>
|
|
||||||
|
|
||||||
// CUDA runtime
|
|
||||||
#include <cuda_runtime.h>
|
|
||||||
#include <cuda_profiler_api.h>
|
|
||||||
|
|
||||||
// Helper functions and utilities to work with CUDA
|
|
||||||
#include <helper_functions.h>
|
|
||||||
#include <helper_cuda.h>
|
|
||||||
|
|
||||||
/**
|
|
||||||
* Matrix multiplication (CUDA Kernel) on the device: C = A * B
|
|
||||||
* wA is A's width and wB is B's width
|
|
||||||
*/
|
|
||||||
template <int BLOCK_SIZE> __global__ void MatrixMulCUDA(float *C, float *A,
|
|
||||||
float *B, int wA,
|
|
||||||
int wB) {
|
|
||||||
// Block index
|
|
||||||
int bx = blockIdx.x;
|
|
||||||
int by = blockIdx.y;
|
|
||||||
|
|
||||||
// Thread index
|
|
||||||
int tx = threadIdx.x;
|
|
||||||
int ty = threadIdx.y;
|
|
||||||
|
|
||||||
// Index of the first sub-matrix of A processed by the block
|
|
||||||
int aBegin = wA * BLOCK_SIZE * by;
|
|
||||||
|
|
||||||
// Index of the last sub-matrix of A processed by the block
|
|
||||||
int aEnd = aBegin + wA - 1;
|
|
||||||
|
|
||||||
// Step size used to iterate through the sub-matrices of A
|
|
||||||
int aStep = BLOCK_SIZE;
|
|
||||||
|
|
||||||
// Index of the first sub-matrix of B processed by the block
|
|
||||||
int bBegin = BLOCK_SIZE * bx;
|
|
||||||
|
|
||||||
// Step size used to iterate through the sub-matrices of B
|
|
||||||
int bStep = BLOCK_SIZE * wB;
|
|
||||||
|
|
||||||
// Csub is used to store the element of the block sub-matrix
|
|
||||||
// that is computed by the thread
|
|
||||||
float Csub = 0;
|
|
||||||
|
|
||||||
// Loop over all the sub-matrices of A and B
|
|
||||||
// required to compute the block sub-matrix
|
|
||||||
for (int a = aBegin, b = bBegin;
|
|
||||||
a <= aEnd;
|
|
||||||
a += aStep, b += bStep) {
|
|
||||||
// Declaration of the shared memory array As used to
|
|
||||||
// store the sub-matrix of A
|
|
||||||
__shared__ float As[BLOCK_SIZE][BLOCK_SIZE];
|
|
||||||
|
|
||||||
// Declaration of the shared memory array Bs used to
|
|
||||||
// store the sub-matrix of B
|
|
||||||
__shared__ float Bs[BLOCK_SIZE][BLOCK_SIZE];
|
|
||||||
|
|
||||||
// Load the matrices from device memory
|
|
||||||
// to shared memory; each thread loads
|
|
||||||
// one element of each matrix
|
|
||||||
As[ty][tx] = A[a + wA * ty + tx];
|
|
||||||
Bs[ty][tx] = B[b + wB * ty + tx];
|
|
||||||
|
|
||||||
// Synchronize to make sure the matrices are loaded
|
|
||||||
__syncthreads();
|
|
||||||
|
|
||||||
// Multiply the two matrices together;
|
|
||||||
// each thread computes one element
|
|
||||||
// of the block sub-matrix
|
|
||||||
#pragma unroll
|
|
||||||
|
|
||||||
for (int k = 0; k < BLOCK_SIZE; ++k) {
|
|
||||||
Csub += As[ty][k] * Bs[k][tx];
|
|
||||||
}
|
|
||||||
|
|
||||||
// Synchronize to make sure that the preceding
|
|
||||||
// computation is done before loading two new
|
|
||||||
// sub-matrices of A and B in the next iteration
|
|
||||||
__syncthreads();
|
|
||||||
}
|
|
||||||
|
|
||||||
// Write the block sub-matrix to device memory;
|
|
||||||
// each thread writes one element
|
|
||||||
int c = wB * BLOCK_SIZE * by + BLOCK_SIZE * bx;
|
|
||||||
C[c + wB * ty + tx] = Csub;
|
|
||||||
}
|
|
||||||
|
|
||||||
void ConstantInit(float *data, int size, float val) {
|
|
||||||
for (int i = 0; i < size; ++i) {
|
|
||||||
data[i] = val;
|
|
||||||
}
|
|
||||||
}
|
|
||||||
|
|
||||||
/**
|
|
||||||
* Run a simple test of matrix multiplication using CUDA
|
|
||||||
*/
|
|
||||||
int MatrixMultiply(int argc, char **argv,
|
|
||||||
int block_size, const dim3 &dimsA,
|
|
||||||
const dim3 &dimsB) {
|
|
||||||
// Allocate host memory for matrices A and B
|
|
||||||
unsigned int size_A = dimsA.x * dimsA.y;
|
|
||||||
unsigned int mem_size_A = sizeof(float) * size_A;
|
|
||||||
float *h_A;
|
|
||||||
checkCudaErrors(cudaMallocHost(&h_A, mem_size_A));
|
|
||||||
unsigned int size_B = dimsB.x * dimsB.y;
|
|
||||||
unsigned int mem_size_B = sizeof(float) * size_B;
|
|
||||||
float *h_B;
|
|
||||||
checkCudaErrors(cudaMallocHost(&h_B, mem_size_B));
|
|
||||||
cudaStream_t stream;
|
|
||||||
|
|
||||||
// Initialize host memory
|
|
||||||
const float valB = 0.01f;
|
|
||||||
ConstantInit(h_A, size_A, 1.0f);
|
|
||||||
ConstantInit(h_B, size_B, valB);
|
|
||||||
|
|
||||||
// Allocate device memory
|
|
||||||
float *d_A, *d_B, *d_C;
|
|
||||||
|
|
||||||
// Allocate host matrix C
|
|
||||||
dim3 dimsC(dimsB.x, dimsA.y, 1);
|
|
||||||
unsigned int mem_size_C = dimsC.x * dimsC.y * sizeof(float);
|
|
||||||
float *h_C;
|
|
||||||
checkCudaErrors(cudaMallocHost(&h_C, mem_size_C));
|
|
||||||
|
|
||||||
if (h_C == NULL) {
|
|
||||||
fprintf(stderr, "Failed to allocate host matrix C!\n");
|
|
||||||
exit(EXIT_FAILURE);
|
|
||||||
}
|
|
||||||
|
|
||||||
checkCudaErrors(cudaMalloc(reinterpret_cast<void **>(&d_A), mem_size_A));
|
|
||||||
checkCudaErrors(cudaMalloc(reinterpret_cast<void **>(&d_B), mem_size_B));
|
|
||||||
checkCudaErrors(cudaMalloc(reinterpret_cast<void **>(&d_C), mem_size_C));
|
|
||||||
// Allocate CUDA events that we'll use for timing
|
|
||||||
cudaEvent_t start, stop;
|
|
||||||
checkCudaErrors(cudaEventCreate(&start));
|
|
||||||
checkCudaErrors(cudaEventCreate(&stop));
|
|
||||||
|
|
||||||
checkCudaErrors(cudaStreamCreateWithFlags(&stream, cudaStreamNonBlocking));
|
|
||||||
|
|
||||||
// copy host memory to device
|
|
||||||
checkCudaErrors(
|
|
||||||
cudaMemcpyAsync(d_A, h_A, mem_size_A, cudaMemcpyHostToDevice, stream));
|
|
||||||
checkCudaErrors(
|
|
||||||
cudaMemcpyAsync(d_B, h_B, mem_size_B, cudaMemcpyHostToDevice, stream));
|
|
||||||
|
|
||||||
// Setup execution parameters
|
|
||||||
dim3 threads(block_size, block_size);
|
|
||||||
dim3 grid(dimsB.x / threads.x, dimsA.y / threads.y);
|
|
||||||
|
|
||||||
// Create and start timer
|
|
||||||
printf("Computing result using CUDA Kernel...\n");
|
|
||||||
|
|
||||||
// Performs warmup operation using matrixMul CUDA kernel
|
|
||||||
if (block_size == 16) {
|
|
||||||
MatrixMulCUDA<16>
|
|
||||||
<<<grid, threads, 0, stream>>>(d_C, d_A, d_B, dimsA.x, dimsB.x);
|
|
||||||
} else {
|
|
||||||
MatrixMulCUDA<32>
|
|
||||||
<<<grid, threads, 0, stream>>>(d_C, d_A, d_B, dimsA.x, dimsB.x);
|
|
||||||
}
|
|
||||||
|
|
||||||
printf("done\n");
|
|
||||||
checkCudaErrors(cudaStreamSynchronize(stream));
|
|
||||||
|
|
||||||
// Record the start event
|
|
||||||
checkCudaErrors(cudaEventRecord(start, stream));
|
|
||||||
|
|
||||||
// Execute the kernel
|
|
||||||
int nIter = 300;
|
|
||||||
|
|
||||||
for (int j = 0; j < nIter; j++) {
|
|
||||||
if (block_size == 16) {
|
|
||||||
MatrixMulCUDA<16>
|
|
||||||
<<<grid, threads, 0, stream>>>(d_C, d_A, d_B, dimsA.x, dimsB.x);
|
|
||||||
} else {
|
|
||||||
MatrixMulCUDA<32>
|
|
||||||
<<<grid, threads, 0, stream>>>(d_C, d_A, d_B, dimsA.x, dimsB.x);
|
|
||||||
}
|
|
||||||
}
|
|
||||||
|
|
||||||
// Record the stop event
|
|
||||||
checkCudaErrors(cudaEventRecord(stop, stream));
|
|
||||||
|
|
||||||
// Wait for the stop event to complete
|
|
||||||
checkCudaErrors(cudaEventSynchronize(stop));
|
|
||||||
|
|
||||||
float msecTotal = 0.0f;
|
|
||||||
checkCudaErrors(cudaEventElapsedTime(&msecTotal, start, stop));
|
|
||||||
|
|
||||||
// Compute and print the performance
|
|
||||||
float msecPerMatrixMul = msecTotal / nIter;
|
|
||||||
double flopsPerMatrixMul = 2.0 * static_cast<double>(dimsA.x) *
|
|
||||||
static_cast<double>(dimsA.y) *
|
|
||||||
static_cast<double>(dimsB.x);
|
|
||||||
double gigaFlops =
|
|
||||||
(flopsPerMatrixMul * 1.0e-9f) / (msecPerMatrixMul / 1000.0f);
|
|
||||||
printf(
|
|
||||||
"Performance= %.2f GFlop/s, Time= %.3f msec, Size= %.0f Ops,"
|
|
||||||
" WorkgroupSize= %u threads/block\n",
|
|
||||||
gigaFlops, msecPerMatrixMul, flopsPerMatrixMul, threads.x * threads.y);
|
|
||||||
|
|
||||||
// Copy result from device to host
|
|
||||||
checkCudaErrors(
|
|
||||||
cudaMemcpyAsync(h_C, d_C, mem_size_C, cudaMemcpyDeviceToHost, stream));
|
|
||||||
checkCudaErrors(cudaStreamSynchronize(stream));
|
|
||||||
|
|
||||||
printf("Checking computed result for correctness: ");
|
|
||||||
bool correct = true;
|
|
||||||
|
|
||||||
// test relative error by the formula
|
|
||||||
// |<x, y>_cpu - <x,y>_gpu|/<|x|, |y|> < eps
|
|
||||||
double eps = 1.e-6; // machine zero
|
|
||||||
|
|
||||||
for (int i = 0; i < static_cast<int>(dimsC.x * dimsC.y); i++) {
|
|
||||||
double abs_err = fabs(h_C[i] - (dimsA.x * valB));
|
|
||||||
double dot_length = dimsA.x;
|
|
||||||
double abs_val = fabs(h_C[i]);
|
|
||||||
double rel_err = abs_err / abs_val / dot_length;
|
|
||||||
|
|
||||||
if (rel_err > eps) {
|
|
||||||
printf("Error! Matrix[%05d]=%.8f, ref=%.8f error term is > %E\n",
|
|
||||||
i, h_C[i], dimsA.x * valB, eps);
|
|
||||||
correct = false;
|
|
||||||
}
|
|
||||||
}
|
|
||||||
|
|
||||||
printf("%s\n", correct ? "Result = PASS" : "Result = FAIL");
|
|
||||||
|
|
||||||
// Clean up memory
|
|
||||||
checkCudaErrors(cudaFreeHost(h_A));
|
|
||||||
checkCudaErrors(cudaFreeHost(h_B));
|
|
||||||
checkCudaErrors(cudaFreeHost(h_C));
|
|
||||||
checkCudaErrors(cudaFree(d_A));
|
|
||||||
checkCudaErrors(cudaFree(d_B));
|
|
||||||
checkCudaErrors(cudaFree(d_C));
|
|
||||||
checkCudaErrors(cudaEventDestroy(start));
|
|
||||||
checkCudaErrors(cudaEventDestroy(stop));
|
|
||||||
printf(
|
|
||||||
"\nNOTE: The CUDA Samples are not meant for performance "
|
|
||||||
"measurements. Results may vary when GPU Boost is enabled.\n");
|
|
||||||
|
|
||||||
if (correct) {
|
|
||||||
return EXIT_SUCCESS;
|
|
||||||
} else {
|
|
||||||
return EXIT_FAILURE;
|
|
||||||
}
|
|
||||||
}
|
|
||||||
|
|
||||||
|
|
||||||
/**
|
|
||||||
* Program main
|
|
||||||
*/
|
|
||||||
int main(int argc, char **argv) {
|
|
||||||
printf("[Matrix Multiply Using CUDA] - Starting...\n");
|
|
||||||
|
|
||||||
if (checkCmdLineFlag(argc, (const char **)argv, "help") ||
|
|
||||||
checkCmdLineFlag(argc, (const char **)argv, "?")) {
|
|
||||||
printf("Usage -device=n (n >= 0 for deviceID)\n");
|
|
||||||
printf(" -wA=WidthA -hA=HeightA (Width x Height of Matrix A)\n");
|
|
||||||
printf(" -wB=WidthB -hB=HeightB (Width x Height of Matrix B)\n");
|
|
||||||
printf(" Note: Outer matrix dimensions of A & B matrices" \
|
|
||||||
" must be equal.\n");
|
|
||||||
|
|
||||||
exit(EXIT_SUCCESS);
|
|
||||||
}
|
|
||||||
|
|
||||||
// This will pick the best possible CUDA capable device, otherwise
|
|
||||||
// override the device ID based on input provided at the command line
|
|
||||||
int dev = findCudaDevice(argc, (const char **)argv);
|
|
||||||
|
|
||||||
int block_size = 32;
|
|
||||||
|
|
||||||
dim3 dimsA(5 * 2 * block_size, 5 * 2 * block_size, 1);
|
|
||||||
dim3 dimsB(5 * 4 * block_size, 5 * 2 * block_size, 1);
|
|
||||||
|
|
||||||
// width of Matrix A
|
|
||||||
if (checkCmdLineFlag(argc, (const char **)argv, "wA")) {
|
|
||||||
dimsA.x = getCmdLineArgumentInt(argc, (const char **)argv, "wA");
|
|
||||||
}
|
|
||||||
|
|
||||||
// height of Matrix A
|
|
||||||
if (checkCmdLineFlag(argc, (const char **)argv, "hA")) {
|
|
||||||
dimsA.y = getCmdLineArgumentInt(argc, (const char **)argv, "hA");
|
|
||||||
}
|
|
||||||
|
|
||||||
// width of Matrix B
|
|
||||||
if (checkCmdLineFlag(argc, (const char **)argv, "wB")) {
|
|
||||||
dimsB.x = getCmdLineArgumentInt(argc, (const char **)argv, "wB");
|
|
||||||
}
|
|
||||||
|
|
||||||
// height of Matrix B
|
|
||||||
if (checkCmdLineFlag(argc, (const char **)argv, "hB")) {
|
|
||||||
dimsB.y = getCmdLineArgumentInt(argc, (const char **)argv, "hB");
|
|
||||||
}
|
|
||||||
|
|
||||||
if (dimsA.x != dimsB.y) {
|
|
||||||
printf("Error: outer matrix dimensions must be equal. (%d != %d)\n",
|
|
||||||
dimsA.x, dimsB.y);
|
|
||||||
exit(EXIT_FAILURE);
|
|
||||||
}
|
|
||||||
|
|
||||||
printf("MatrixA(%d,%d), MatrixB(%d,%d)\n", dimsA.x, dimsA.y,
|
|
||||||
dimsB.x, dimsB.y);
|
|
||||||
|
|
||||||
checkCudaErrors(cudaProfilerStart());
|
|
||||||
int matrix_result = MatrixMultiply(argc, argv, block_size, dimsA, dimsB);
|
|
||||||
checkCudaErrors(cudaProfilerStop());
|
|
||||||
|
|
||||||
exit(matrix_result);
|
|
||||||
}
|
|
||||||
@@ -1,10 +0,0 @@
|
|||||||
{
|
|
||||||
"configurations": [
|
|
||||||
{
|
|
||||||
"name": "CUDA C++: Launch",
|
|
||||||
"type": "cuda-gdb",
|
|
||||||
"request": "launch",
|
|
||||||
"program": "${workspaceFolder}/matrixMulDrv"
|
|
||||||
}
|
|
||||||
]
|
|
||||||
}
|
|
||||||
@@ -1,15 +0,0 @@
|
|||||||
{
|
|
||||||
"version": "2.0.0",
|
|
||||||
"tasks": [
|
|
||||||
{
|
|
||||||
"label": "sample",
|
|
||||||
"type": "shell",
|
|
||||||
"command": "make dbg=1",
|
|
||||||
"problemMatcher": ["$nvcc"],
|
|
||||||
"group": {
|
|
||||||
"kind": "build",
|
|
||||||
"isDefault": true
|
|
||||||
}
|
|
||||||
}
|
|
||||||
]
|
|
||||||
}
|
|
||||||
@@ -1,52 +0,0 @@
|
|||||||
cmake_minimum_required(VERSION 3.20)
|
|
||||||
|
|
||||||
list(APPEND CMAKE_MODULE_PATH "${CMAKE_CURRENT_SOURCE_DIR}/../../../cmake/Modules")
|
|
||||||
|
|
||||||
project(matrixMulDrv LANGUAGES C CXX CUDA)
|
|
||||||
|
|
||||||
find_package(CUDAToolkit REQUIRED)
|
|
||||||
|
|
||||||
set(CMAKE_POSITION_INDEPENDENT_CODE ON)
|
|
||||||
|
|
||||||
set(CMAKE_CUDA_ARCHITECTURES 50 52 60 61 70 72 75 80 86 87 89 90 100 101 120)
|
|
||||||
set(CMAKE_CUDA_FLAGS "${CMAKE_CUDA_FLAGS} -Wno-deprecated-gpu-targets")
|
|
||||||
if(CMAKE_BUILD_TYPE STREQUAL "Debug")
|
|
||||||
# set(CMAKE_CUDA_FLAGS "${CMAKE_CUDA_FLAGS} -G") # enable cuda-gdb (expensive)
|
|
||||||
endif()
|
|
||||||
|
|
||||||
# Include directories and libraries
|
|
||||||
include_directories(../../../Common)
|
|
||||||
|
|
||||||
# Source file
|
|
||||||
# Add target for matrixMulDrv
|
|
||||||
add_executable(matrixMulDrv matrixMulDrv.cpp)
|
|
||||||
|
|
||||||
target_compile_options(matrixMulDrv PRIVATE $<$<COMPILE_LANGUAGE:CUDA>:--extended-lambda>)
|
|
||||||
|
|
||||||
target_compile_features(matrixMulDrv PRIVATE cxx_std_17 cuda_std_17)
|
|
||||||
|
|
||||||
set_target_properties(matrixMulDrv PROPERTIES CUDA_SEPARABLE_COMPILATION ON)
|
|
||||||
|
|
||||||
target_include_directories(matrixMulDrv PRIVATE
|
|
||||||
${CUDAToolkit_INCLUDE_DIRS}
|
|
||||||
)
|
|
||||||
|
|
||||||
target_link_libraries(matrixMulDrv PUBLIC
|
|
||||||
CUDA::cuda_driver
|
|
||||||
)
|
|
||||||
|
|
||||||
set(CUDA_FATBIN_FILE "${CMAKE_CURRENT_BINARY_DIR}/matrixMul_kernel64.fatbin")
|
|
||||||
set(CUDA_KERNEL_SOURCE "${CMAKE_CURRENT_SOURCE_DIR}/matrixMul_kernel.cu")
|
|
||||||
|
|
||||||
add_custom_command(
|
|
||||||
OUTPUT ${CUDA_FATBIN_FILE}
|
|
||||||
COMMAND ${CMAKE_CUDA_COMPILER} ${INCLUDES} ${ALL_CCFLAGS} -Wno-deprecated-gpu-targets ${GENCODE_FLAGS} -o ${CUDA_FATBIN_FILE} -fatbin ${CUDA_KERNEL_SOURCE}
|
|
||||||
DEPENDS ${CUDA_KERNEL_SOURCE}
|
|
||||||
COMMENT "Building CUDA fatbin: ${CUDA_FATBIN_FILE}"
|
|
||||||
)
|
|
||||||
|
|
||||||
# Create a dummy target for fatbin generation
|
|
||||||
add_custom_target(generate_fatbin_matmulDrv ALL DEPENDS ${CUDA_FATBIN_FILE})
|
|
||||||
|
|
||||||
# Ensure matrixMulDrv depends on the fatbin
|
|
||||||
add_dependencies(matrixMulDrv generate_fatbin_matmulDrv)
|
|
||||||
@@ -1,32 +0,0 @@
|
|||||||
# matrixMulDrv - Matrix Multiplication (CUDA Driver API Version)
|
|
||||||
|
|
||||||
## Description
|
|
||||||
|
|
||||||
This sample implements matrix multiplication and uses the new CUDA 4.0 kernel launch Driver API. It has been written for clarity of exposition to illustrate various CUDA programming principles, not with the goal of providing the most performant generic kernel for matrix multiplication. CUBLAS provides high-performance matrix multiplication.
|
|
||||||
|
|
||||||
## Key Concepts
|
|
||||||
|
|
||||||
CUDA Driver API, Matrix Multiply
|
|
||||||
|
|
||||||
## Supported SM Architectures
|
|
||||||
|
|
||||||
[SM 5.0 ](https://developer.nvidia.com/cuda-gpus) [SM 5.2 ](https://developer.nvidia.com/cuda-gpus) [SM 5.3 ](https://developer.nvidia.com/cuda-gpus) [SM 6.0 ](https://developer.nvidia.com/cuda-gpus) [SM 6.1 ](https://developer.nvidia.com/cuda-gpus) [SM 7.0 ](https://developer.nvidia.com/cuda-gpus) [SM 7.2 ](https://developer.nvidia.com/cuda-gpus) [SM 7.5 ](https://developer.nvidia.com/cuda-gpus) [SM 8.0 ](https://developer.nvidia.com/cuda-gpus) [SM 8.6 ](https://developer.nvidia.com/cuda-gpus) [SM 8.7 ](https://developer.nvidia.com/cuda-gpus) [SM 8.9 ](https://developer.nvidia.com/cuda-gpus) [SM 9.0 ](https://developer.nvidia.com/cuda-gpus)
|
|
||||||
|
|
||||||
## Supported OSes
|
|
||||||
|
|
||||||
Linux, Windows
|
|
||||||
|
|
||||||
## Supported CPU Architecture
|
|
||||||
|
|
||||||
x86_64, armv7l, aarch64
|
|
||||||
|
|
||||||
## CUDA APIs involved
|
|
||||||
|
|
||||||
### [CUDA Driver API](http://docs.nvidia.com/cuda/cuda-driver-api/index.html)
|
|
||||||
cuMemcpyDtoH, cuLaunchKernel, cuMemcpyHtoD, cuDeviceGetName, cuDeviceTotalMem, cuDeviceGetAttribute, cuModuleLoadData, cuOccupancyMaxPotentialBlockSize, cuMemAlloc, cuMemFree, cuCtxDestroy, cuModuleGetFunction, cuCtxCreate
|
|
||||||
|
|
||||||
## Prerequisites
|
|
||||||
|
|
||||||
Download and install the [CUDA Toolkit 12.5](https://developer.nvidia.com/cuda-downloads) for your corresponding platform.
|
|
||||||
|
|
||||||
## References (for more details)
|
|
||||||
@@ -1,40 +0,0 @@
|
|||||||
/* Copyright (c) 2022, NVIDIA CORPORATION. All rights reserved.
|
|
||||||
*
|
|
||||||
* Redistribution and use in source and binary forms, with or without
|
|
||||||
* modification, are permitted provided that the following conditions
|
|
||||||
* are met:
|
|
||||||
* * Redistributions of source code must retain the above copyright
|
|
||||||
* notice, this list of conditions and the following disclaimer.
|
|
||||||
* * Redistributions in binary form must reproduce the above copyright
|
|
||||||
* notice, this list of conditions and the following disclaimer in the
|
|
||||||
* documentation and/or other materials provided with the distribution.
|
|
||||||
* * Neither the name of NVIDIA CORPORATION nor the names of its
|
|
||||||
* contributors may be used to endorse or promote products derived
|
|
||||||
* from this software without specific prior written permission.
|
|
||||||
*
|
|
||||||
* THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS ``AS IS'' AND ANY
|
|
||||||
* EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
|
|
||||||
* IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR
|
|
||||||
* PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER OR
|
|
||||||
* CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL,
|
|
||||||
* EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO,
|
|
||||||
* PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR
|
|
||||||
* PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY
|
|
||||||
* OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
|
|
||||||
* (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
|
|
||||||
* OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
|
|
||||||
*/
|
|
||||||
|
|
||||||
#ifndef _MATRIXMUL_H_
|
|
||||||
#define _MATRIXMUL_H_
|
|
||||||
|
|
||||||
// Matrix dimensions
|
|
||||||
// (chosen as multiples of the thread block size for simplicity)
|
|
||||||
#define WA (4 * block_size) // Matrix A width
|
|
||||||
#define HA (6 * block_size) // Matrix A height
|
|
||||||
#define WB (4 * block_size) // Matrix B width
|
|
||||||
#define HB WA // Matrix B height
|
|
||||||
#define WC WB // Matrix C width
|
|
||||||
#define HC HA // Matrix C height
|
|
||||||
|
|
||||||
#endif // _MATRIXMUL_H_
|
|
||||||
@@ -1,320 +0,0 @@
|
|||||||
/* Copyright (c) 2022, NVIDIA CORPORATION. All rights reserved.
|
|
||||||
*
|
|
||||||
* Redistribution and use in source and binary forms, with or without
|
|
||||||
* modification, are permitted provided that the following conditions
|
|
||||||
* are met:
|
|
||||||
* * Redistributions of source code must retain the above copyright
|
|
||||||
* notice, this list of conditions and the following disclaimer.
|
|
||||||
* * Redistributions in binary form must reproduce the above copyright
|
|
||||||
* notice, this list of conditions and the following disclaimer in the
|
|
||||||
* documentation and/or other materials provided with the distribution.
|
|
||||||
* * Neither the name of NVIDIA CORPORATION nor the names of its
|
|
||||||
* contributors may be used to endorse or promote products derived
|
|
||||||
* from this software without specific prior written permission.
|
|
||||||
*
|
|
||||||
* THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS ``AS IS'' AND ANY
|
|
||||||
* EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
|
|
||||||
* IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR
|
|
||||||
* PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER OR
|
|
||||||
* CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL,
|
|
||||||
* EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO,
|
|
||||||
* PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR
|
|
||||||
* PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY
|
|
||||||
* OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
|
|
||||||
* (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
|
|
||||||
* OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
|
|
||||||
*/
|
|
||||||
|
|
||||||
/* Matrix multiplication: C = A * B.
|
|
||||||
* Host code.
|
|
||||||
*
|
|
||||||
* This sample implements matrix multiplication using the CUDA driver API.
|
|
||||||
* It has been written for clarity of exposition to illustrate various CUDA
|
|
||||||
* programming principles, not with the goal of providing the most
|
|
||||||
* performant generic kernel for matrix multiplication.
|
|
||||||
*
|
|
||||||
* CUBLAS provides high-performance matrix multiplication.
|
|
||||||
* See also:
|
|
||||||
* V. Volkov and J. Demmel, "Benchmarking GPUs to tune dense linear algebra,"
|
|
||||||
* in Proc. 2008 ACM/IEEE Conf. on Supercomputing (SC '08),
|
|
||||||
* Piscataway, NJ: IEEE Press, 2008, pp. Art. 31:1-11.
|
|
||||||
*
|
|
||||||
* Volkov, V. 2010. Better performance at lower occupancy,
|
|
||||||
* GPU Technology Conference 2~010 (GTC 2010).
|
|
||||||
*
|
|
||||||
*/
|
|
||||||
|
|
||||||
// includes, system
|
|
||||||
#include <builtin_types.h>
|
|
||||||
#include <math.h>
|
|
||||||
#include <stdlib.h>
|
|
||||||
#include <stdio.h>
|
|
||||||
#include <string.h>
|
|
||||||
#include <iostream>
|
|
||||||
#include <cstring>
|
|
||||||
|
|
||||||
// includes, project, CUDA
|
|
||||||
#include <cuda.h>
|
|
||||||
#include <helper_cuda_drvapi.h>
|
|
||||||
#include <helper_image.h>
|
|
||||||
#include <helper_string.h>
|
|
||||||
#include <helper_timer.h>
|
|
||||||
|
|
||||||
#include <cstring>
|
|
||||||
#include <iostream>
|
|
||||||
#include <string>
|
|
||||||
#include "matrixMul.h"
|
|
||||||
|
|
||||||
|
|
||||||
////////////////////////////////////////////////////////////////////////////////
|
|
||||||
// declaration, forward
|
|
||||||
void runTest(int argc, char **argv);
|
|
||||||
void randomInit(float *, int);
|
|
||||||
|
|
||||||
extern "C" void computeGold(float *, const float *, const float *, unsigned int,
|
|
||||||
unsigned int, unsigned int);
|
|
||||||
|
|
||||||
static int initCUDA(int argc, char **argv, CUfunction *pMatrixMul,
|
|
||||||
int *blk_size);
|
|
||||||
|
|
||||||
#ifndef FATBIN_FILE
|
|
||||||
#define FATBIN_FILE "matrixMul_kernel64.fatbin"
|
|
||||||
#endif
|
|
||||||
|
|
||||||
////////////////////////////////////////////////////////////////////////////////
|
|
||||||
// Globals
|
|
||||||
////////////////////////////////////////////////////////////////////////////////
|
|
||||||
CUdevice cuDevice;
|
|
||||||
CUcontext cuContext;
|
|
||||||
CUmodule cuModule;
|
|
||||||
size_t totalGlobalMem;
|
|
||||||
|
|
||||||
const char *sSDKsample = "matrixMulDrv (Driver API)";
|
|
||||||
|
|
||||||
void constantInit(float *data, int size, float val) {
|
|
||||||
for (int i = 0; i < size; ++i) {
|
|
||||||
data[i] = val;
|
|
||||||
}
|
|
||||||
}
|
|
||||||
|
|
||||||
////////////////////////////////////////////////////////////////////////////////
|
|
||||||
// Program main
|
|
||||||
////////////////////////////////////////////////////////////////////////////////
|
|
||||||
int main(int argc, char **argv) {
|
|
||||||
printf("[ %s ]\n", sSDKsample);
|
|
||||||
|
|
||||||
runTest(argc, argv);
|
|
||||||
}
|
|
||||||
|
|
||||||
////////////////////////////////////////////////////////////////////////////////
|
|
||||||
//! Run a simple test for CUDA
|
|
||||||
////////////////////////////////////////////////////////////////////////////////
|
|
||||||
void runTest(int argc, char **argv) {
|
|
||||||
// initialize CUDA
|
|
||||||
CUfunction matrixMul = NULL;
|
|
||||||
int block_size = 0;
|
|
||||||
|
|
||||||
initCUDA(argc, argv, &matrixMul, &block_size);
|
|
||||||
|
|
||||||
// set seed for rand()
|
|
||||||
srand(2006);
|
|
||||||
|
|
||||||
// allocate host memory for matrices A and B
|
|
||||||
unsigned int size_A = WA * HA;
|
|
||||||
unsigned int mem_size_A = sizeof(float) * size_A;
|
|
||||||
float *h_A = reinterpret_cast<float *>(malloc(mem_size_A));
|
|
||||||
unsigned int size_B = WB * HB;
|
|
||||||
unsigned int mem_size_B = sizeof(float) * size_B;
|
|
||||||
float *h_B = reinterpret_cast<float *>(malloc(mem_size_B));
|
|
||||||
|
|
||||||
// initialize host memory
|
|
||||||
const float valB = 0.01f;
|
|
||||||
constantInit(h_A, size_A, 1.0f);
|
|
||||||
constantInit(h_B, size_B, valB);
|
|
||||||
|
|
||||||
// allocate device memory
|
|
||||||
CUdeviceptr d_A;
|
|
||||||
checkCudaErrors(cuMemAlloc(&d_A, mem_size_A));
|
|
||||||
CUdeviceptr d_B;
|
|
||||||
checkCudaErrors(cuMemAlloc(&d_B, mem_size_B));
|
|
||||||
|
|
||||||
// copy host memory to device
|
|
||||||
checkCudaErrors(cuMemcpyHtoD(d_A, h_A, mem_size_A));
|
|
||||||
checkCudaErrors(cuMemcpyHtoD(d_B, h_B, mem_size_B));
|
|
||||||
|
|
||||||
// allocate device memory for result
|
|
||||||
size_t size_C = WC * HC;
|
|
||||||
size_t mem_size_C = sizeof(float) * size_C;
|
|
||||||
|
|
||||||
CUdeviceptr d_C;
|
|
||||||
checkCudaErrors(cuMemAlloc(&d_C, mem_size_C));
|
|
||||||
|
|
||||||
// allocate mem for the result on host side
|
|
||||||
float *h_C = reinterpret_cast<float *>(malloc(mem_size_C));
|
|
||||||
|
|
||||||
// create and start timer
|
|
||||||
StopWatchInterface *timer = NULL;
|
|
||||||
sdkCreateTimer(&timer);
|
|
||||||
|
|
||||||
// start the timer
|
|
||||||
sdkStartTimer(&timer);
|
|
||||||
|
|
||||||
// There are two ways to launch CUDA kernels via the Driver API.
|
|
||||||
// In this CUDA Sample, we illustrate both ways to pass parameters
|
|
||||||
// and specify parameters. By default we use the simpler method.
|
|
||||||
dim3 block(block_size, block_size, 1);
|
|
||||||
dim3 grid(WC / block_size, HC / block_size, 1);
|
|
||||||
|
|
||||||
if (1) {
|
|
||||||
// This is the new CUDA 4.0 API for Kernel Parameter passing and Kernel
|
|
||||||
// Launching (simplier method)
|
|
||||||
size_t Matrix_Width_A = (size_t)WA;
|
|
||||||
size_t Matrix_Width_B = (size_t)WB;
|
|
||||||
void *args[5] = {&d_C, &d_A, &d_B, &Matrix_Width_A, &Matrix_Width_B};
|
|
||||||
// new CUDA 4.0 Driver API Kernel launch call
|
|
||||||
checkCudaErrors(cuLaunchKernel(
|
|
||||||
matrixMul, grid.x, grid.y, grid.z, block.x, block.y, block.z,
|
|
||||||
2 * block_size * block_size * sizeof(float), NULL, args, NULL));
|
|
||||||
} else {
|
|
||||||
// This is the new CUDA 4.0 API for Kernel Parameter passing and Kernel
|
|
||||||
// Launching (advanced method)
|
|
||||||
int offset = 0;
|
|
||||||
char argBuffer[256];
|
|
||||||
|
|
||||||
// pass in launch parameters (not actually de-referencing CUdeviceptr).
|
|
||||||
// CUdeviceptr is storing the value of the parameters
|
|
||||||
*(reinterpret_cast<CUdeviceptr *>(&argBuffer[offset])) = d_C;
|
|
||||||
offset += sizeof(d_C);
|
|
||||||
*(reinterpret_cast<CUdeviceptr *>(&argBuffer[offset])) = d_A;
|
|
||||||
offset += sizeof(d_A);
|
|
||||||
*(reinterpret_cast<CUdeviceptr *>(&argBuffer[offset])) = d_B;
|
|
||||||
offset += sizeof(d_B);
|
|
||||||
|
|
||||||
size_t Matrix_Width_A = (size_t)WA;
|
|
||||||
size_t Matrix_Width_B = (size_t)WB;
|
|
||||||
|
|
||||||
*(reinterpret_cast<CUdeviceptr *>(&argBuffer[offset])) = Matrix_Width_A;
|
|
||||||
offset += sizeof(Matrix_Width_A);
|
|
||||||
*(reinterpret_cast<CUdeviceptr *>(&argBuffer[offset])) = Matrix_Width_B;
|
|
||||||
offset += sizeof(Matrix_Width_B);
|
|
||||||
|
|
||||||
void *kernel_launch_config[5] = {CU_LAUNCH_PARAM_BUFFER_POINTER, argBuffer,
|
|
||||||
CU_LAUNCH_PARAM_BUFFER_SIZE, &offset,
|
|
||||||
CU_LAUNCH_PARAM_END};
|
|
||||||
|
|
||||||
// new CUDA 4.0 Driver API Kernel launch call
|
|
||||||
checkCudaErrors(cuLaunchKernel(
|
|
||||||
matrixMul, grid.x, grid.y, grid.z, block.x, block.y, block.z,
|
|
||||||
2 * block_size * block_size * sizeof(float), NULL, NULL,
|
|
||||||
reinterpret_cast<void **>(&kernel_launch_config)));
|
|
||||||
}
|
|
||||||
|
|
||||||
// copy result from device to host
|
|
||||||
checkCudaErrors(cuMemcpyDtoH(reinterpret_cast<void *>(h_C), d_C, mem_size_C));
|
|
||||||
|
|
||||||
// stop and destroy timer
|
|
||||||
sdkStopTimer(&timer);
|
|
||||||
printf("Processing time: %f (ms)\n", sdkGetTimerValue(&timer));
|
|
||||||
sdkDeleteTimer(&timer);
|
|
||||||
|
|
||||||
printf("Checking computed result for correctness: ");
|
|
||||||
bool correct = true;
|
|
||||||
|
|
||||||
for (int i = 0; i < static_cast<int>(WC * HC); i++) {
|
|
||||||
if (fabs(h_C[i] - (WA * valB)) > 1e-5) {
|
|
||||||
printf("Error! Matrix[%05d]=%.8f, ref=%.8f error term is > 1e-5\n", i,
|
|
||||||
h_C[i], WA * valB);
|
|
||||||
correct = false;
|
|
||||||
}
|
|
||||||
}
|
|
||||||
|
|
||||||
printf("%s\n", correct ? "Result = PASS" : "Result = FAIL");
|
|
||||||
|
|
||||||
printf("\nNOTE: The CUDA Samples are not meant for performance measurements. "
|
|
||||||
"Results may vary when GPU Boost is enabled.\n");
|
|
||||||
|
|
||||||
// clean up memory
|
|
||||||
free(h_A);
|
|
||||||
free(h_B);
|
|
||||||
free(h_C);
|
|
||||||
checkCudaErrors(cuMemFree(d_A));
|
|
||||||
checkCudaErrors(cuMemFree(d_B));
|
|
||||||
checkCudaErrors(cuMemFree(d_C));
|
|
||||||
checkCudaErrors(cuCtxDestroy(cuContext));
|
|
||||||
}
|
|
||||||
|
|
||||||
// Allocates a matrix with random float entries.
|
|
||||||
void randomInit(float *data, int size) {
|
|
||||||
for (int i = 0; i < size; ++i) {
|
|
||||||
data[i] = rand() / static_cast<float>(RAND_MAX);
|
|
||||||
}
|
|
||||||
}
|
|
||||||
|
|
||||||
static int initCUDA(int argc, char **argv, CUfunction *pMatrixMul,
|
|
||||||
int *blk_size) {
|
|
||||||
CUfunction cuFunction = 0;
|
|
||||||
int major = 0, minor = 0;
|
|
||||||
char deviceName[100];
|
|
||||||
|
|
||||||
cuDevice = findCudaDeviceDRV(argc, (const char **)argv);
|
|
||||||
|
|
||||||
// get compute capabilities and the devicename
|
|
||||||
checkCudaErrors(cuDeviceGetAttribute(
|
|
||||||
&major, CU_DEVICE_ATTRIBUTE_COMPUTE_CAPABILITY_MAJOR, cuDevice));
|
|
||||||
checkCudaErrors(cuDeviceGetAttribute(
|
|
||||||
&minor, CU_DEVICE_ATTRIBUTE_COMPUTE_CAPABILITY_MINOR, cuDevice));
|
|
||||||
checkCudaErrors(cuDeviceGetName(deviceName, sizeof(deviceName), cuDevice));
|
|
||||||
printf("> GPU Device has SM %d.%d compute capability\n", major, minor);
|
|
||||||
|
|
||||||
checkCudaErrors(cuDeviceTotalMem(&totalGlobalMem, cuDevice));
|
|
||||||
printf(" Total amount of global memory: %llu bytes\n",
|
|
||||||
(long long unsigned int)totalGlobalMem);
|
|
||||||
|
|
||||||
checkCudaErrors(cuCtxCreate(&cuContext, 0, cuDevice));
|
|
||||||
|
|
||||||
// first search for the module path before we load the results
|
|
||||||
std::string module_path;
|
|
||||||
std::ostringstream fatbin;
|
|
||||||
|
|
||||||
if (!findFatbinPath(FATBIN_FILE, module_path, argv, fatbin)) {
|
|
||||||
exit(EXIT_FAILURE);
|
|
||||||
} else {
|
|
||||||
printf("> initCUDA loading module: <%s>\n", module_path.c_str());
|
|
||||||
}
|
|
||||||
|
|
||||||
if (!fatbin.str().size()) {
|
|
||||||
printf("fatbin file empty. exiting..\n");
|
|
||||||
exit(EXIT_FAILURE);
|
|
||||||
}
|
|
||||||
|
|
||||||
// Create module from binary file (FATBIN)
|
|
||||||
checkCudaErrors(cuModuleLoadData(&cuModule, fatbin.str().c_str()));
|
|
||||||
|
|
||||||
// select the suitable kernel function
|
|
||||||
const char *kernels[] = {"matrixMul_bs32_64bit", "matrixMul_bs16_64bit",
|
|
||||||
"matrixMul_bs8_64bit"};
|
|
||||||
|
|
||||||
int idx = 0;
|
|
||||||
int block_size = 32;
|
|
||||||
while (idx < 3) {
|
|
||||||
int threadsPerBlock = 0;
|
|
||||||
int blocksPerGrid = 0;
|
|
||||||
|
|
||||||
checkCudaErrors(cuModuleGetFunction(&cuFunction, cuModule, kernels[idx]));
|
|
||||||
checkCudaErrors(cuOccupancyMaxPotentialBlockSize(
|
|
||||||
&blocksPerGrid, &threadsPerBlock, cuFunction, 0,
|
|
||||||
2 * block_size * block_size * sizeof(float), 0));
|
|
||||||
if (block_size * block_size <= threadsPerBlock) {
|
|
||||||
printf("> %d block size selected\n", block_size);
|
|
||||||
break;
|
|
||||||
} else {
|
|
||||||
block_size /= 2;
|
|
||||||
}
|
|
||||||
idx++;
|
|
||||||
}
|
|
||||||
|
|
||||||
*pMatrixMul = cuFunction;
|
|
||||||
*blk_size = block_size;
|
|
||||||
|
|
||||||
return 0;
|
|
||||||
}
|
|
||||||
@@ -1,127 +0,0 @@
|
|||||||
/* Copyright (c) 2022, NVIDIA CORPORATION. All rights reserved.
|
|
||||||
*
|
|
||||||
* Redistribution and use in source and binary forms, with or without
|
|
||||||
* modification, are permitted provided that the following conditions
|
|
||||||
* are met:
|
|
||||||
* * Redistributions of source code must retain the above copyright
|
|
||||||
* notice, this list of conditions and the following disclaimer.
|
|
||||||
* * Redistributions in binary form must reproduce the above copyright
|
|
||||||
* notice, this list of conditions and the following disclaimer in the
|
|
||||||
* documentation and/or other materials provided with the distribution.
|
|
||||||
* * Neither the name of NVIDIA CORPORATION nor the names of its
|
|
||||||
* contributors may be used to endorse or promote products derived
|
|
||||||
* from this software without specific prior written permission.
|
|
||||||
*
|
|
||||||
* THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS ``AS IS'' AND ANY
|
|
||||||
* EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
|
|
||||||
* IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR
|
|
||||||
* PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER OR
|
|
||||||
* CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL,
|
|
||||||
* EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO,
|
|
||||||
* PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR
|
|
||||||
* PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY
|
|
||||||
* OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
|
|
||||||
* (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
|
|
||||||
* OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
|
|
||||||
*/
|
|
||||||
|
|
||||||
/* Matrix multiplication: C = A * B.
|
|
||||||
* Device code.
|
|
||||||
*/
|
|
||||||
|
|
||||||
#ifndef _MATRIXMUL_KERNEL_H_
|
|
||||||
#define _MATRIXMUL_KERNEL_H_
|
|
||||||
|
|
||||||
#include <stdio.h>
|
|
||||||
|
|
||||||
#define AS(i, j) As[i][j]
|
|
||||||
#define BS(i, j) Bs[i][j]
|
|
||||||
|
|
||||||
////////////////////////////////////////////////////////////////////////////////
|
|
||||||
//! Matrix multiplication on the device: C = A * B
|
|
||||||
//! wA is A's width and wB is B's width
|
|
||||||
////////////////////////////////////////////////////////////////////////////////
|
|
||||||
template <int block_size, typename size_type>
|
|
||||||
__device__ void matrixMul(float *C, float *A, float *B, size_type wA,
|
|
||||||
size_type wB) {
|
|
||||||
// Block index
|
|
||||||
size_type bx = blockIdx.x;
|
|
||||||
size_type by = blockIdx.y;
|
|
||||||
|
|
||||||
// Thread index
|
|
||||||
size_type tx = threadIdx.x;
|
|
||||||
size_type ty = threadIdx.y;
|
|
||||||
|
|
||||||
// Index of the first sub-matrix of A processed by the block
|
|
||||||
size_type aBegin = wA * block_size * by;
|
|
||||||
|
|
||||||
// Index of the last sub-matrix of A processed by the block
|
|
||||||
size_type aEnd = aBegin + wA - 1;
|
|
||||||
|
|
||||||
// Step size used to iterate through the sub-matrices of A
|
|
||||||
size_type aStep = block_size;
|
|
||||||
|
|
||||||
// Index of the first sub-matrix of B processed by the block
|
|
||||||
size_type bBegin = block_size * bx;
|
|
||||||
|
|
||||||
// Step size used to iterate through the sub-matrices of B
|
|
||||||
size_type bStep = block_size * wB;
|
|
||||||
|
|
||||||
// Csub is used to store the element of the block sub-matrix
|
|
||||||
// that is computed by the thread
|
|
||||||
float Csub = 0;
|
|
||||||
|
|
||||||
// Loop over all the sub-matrices of A and B
|
|
||||||
// required to compute the block sub-matrix
|
|
||||||
for (size_type a = aBegin, b = bBegin; a <= aEnd; a += aStep, b += bStep) {
|
|
||||||
// Declaration of the shared memory array As used to
|
|
||||||
// store the sub-matrix of A
|
|
||||||
__shared__ float As[block_size][block_size];
|
|
||||||
|
|
||||||
// Declaration of the shared memory array Bs used to
|
|
||||||
// store the sub-matrix of B
|
|
||||||
__shared__ float Bs[block_size][block_size];
|
|
||||||
|
|
||||||
// Load the matrices from device memory
|
|
||||||
// to shared memory; each thread loads
|
|
||||||
// one element of each matrix
|
|
||||||
AS(ty, tx) = A[a + wA * ty + tx];
|
|
||||||
BS(ty, tx) = B[b + wB * ty + tx];
|
|
||||||
|
|
||||||
// Synchronize to make sure the matrices are loaded
|
|
||||||
__syncthreads();
|
|
||||||
|
|
||||||
// Multiply the two matrices together;
|
|
||||||
// each thread computes one element
|
|
||||||
// of the block sub-matrix
|
|
||||||
#pragma unroll
|
|
||||||
|
|
||||||
for (size_type k = 0; k < block_size; ++k) Csub += AS(ty, k) * BS(k, tx);
|
|
||||||
|
|
||||||
// Synchronize to make sure that the preceding
|
|
||||||
// computation is done before loading two new
|
|
||||||
// sub-matrices of A and B in the next iteration
|
|
||||||
__syncthreads();
|
|
||||||
}
|
|
||||||
|
|
||||||
// Write the block sub-matrix to device memory;
|
|
||||||
// each thread writes one element
|
|
||||||
size_type c = wB * block_size * by + block_size * bx;
|
|
||||||
C[c + wB * ty + tx] = Csub;
|
|
||||||
}
|
|
||||||
|
|
||||||
// C wrappers around our template kernel
|
|
||||||
extern "C" __global__ void matrixMul_bs8_64bit(float *C, float *A, float *B,
|
|
||||||
size_t wA, size_t wB) {
|
|
||||||
matrixMul<8, size_t>(C, A, B, wA, wB);
|
|
||||||
}
|
|
||||||
extern "C" __global__ void matrixMul_bs16_64bit(float *C, float *A, float *B,
|
|
||||||
size_t wA, size_t wB) {
|
|
||||||
matrixMul<16, size_t>(C, A, B, wA, wB);
|
|
||||||
}
|
|
||||||
extern "C" __global__ void matrixMul_bs32_64bit(float *C, float *A, float *B,
|
|
||||||
size_t wA, size_t wB) {
|
|
||||||
matrixMul<32, size_t>(C, A, B, wA, wB);
|
|
||||||
}
|
|
||||||
|
|
||||||
#endif // #ifndef _MATRIXMUL_KERNEL_H_
|
|
||||||
@@ -1,10 +0,0 @@
|
|||||||
{
|
|
||||||
"configurations": [
|
|
||||||
{
|
|
||||||
"name": "CUDA C++: Launch",
|
|
||||||
"type": "cuda-gdb",
|
|
||||||
"request": "launch",
|
|
||||||
"program": "${workspaceFolder}/matrixMulDynlinkJIT"
|
|
||||||
}
|
|
||||||
]
|
|
||||||
}
|
|
||||||
@@ -1,15 +0,0 @@
|
|||||||
{
|
|
||||||
"version": "2.0.0",
|
|
||||||
"tasks": [
|
|
||||||
{
|
|
||||||
"label": "sample",
|
|
||||||
"type": "shell",
|
|
||||||
"command": "make dbg=1",
|
|
||||||
"problemMatcher": ["$nvcc"],
|
|
||||||
"group": {
|
|
||||||
"kind": "build",
|
|
||||||
"isDefault": true
|
|
||||||
}
|
|
||||||
}
|
|
||||||
]
|
|
||||||
}
|
|
||||||
@@ -1,40 +0,0 @@
|
|||||||
cmake_minimum_required(VERSION 3.20)
|
|
||||||
|
|
||||||
list(APPEND CMAKE_MODULE_PATH "${CMAKE_CURRENT_SOURCE_DIR}/../../../cmake/Modules")
|
|
||||||
|
|
||||||
project(matrixMulDynlinkJIT LANGUAGES C CXX CUDA)
|
|
||||||
|
|
||||||
find_package(CUDAToolkit REQUIRED)
|
|
||||||
|
|
||||||
set(CMAKE_POSITION_INDEPENDENT_CODE ON)
|
|
||||||
|
|
||||||
set(CMAKE_CUDA_ARCHITECTURES 50 52 60 61 70 72 75 80 86 87 89 90 100 101 120)
|
|
||||||
set(CMAKE_CUDA_FLAGS "${CMAKE_CUDA_FLAGS} -Wno-deprecated-gpu-targets")
|
|
||||||
if(CMAKE_BUILD_TYPE STREQUAL "Debug")
|
|
||||||
# set(CMAKE_CUDA_FLAGS "${CMAKE_CUDA_FLAGS} -G") # enable cuda-gdb (expensive)
|
|
||||||
endif()
|
|
||||||
|
|
||||||
# Include directories and libraries
|
|
||||||
include_directories(../../../Common)
|
|
||||||
|
|
||||||
# Source file
|
|
||||||
# Add target for matrixMulDynlinkJIT
|
|
||||||
add_executable(matrixMulDynlinkJIT cuda_drvapi_dynlink.c matrixMulDynlinkJIT.cpp matrixMul_gold.cpp matrixMul_kernel_32_ptxdump.c matrixMul_kernel_64_ptxdump.c)
|
|
||||||
|
|
||||||
target_compile_options(matrixMulDynlinkJIT PRIVATE $<$<COMPILE_LANGUAGE:CUDA>:--extended-lambda>)
|
|
||||||
|
|
||||||
target_compile_features(matrixMulDynlinkJIT PRIVATE cxx_std_17 cuda_std_17)
|
|
||||||
|
|
||||||
set_target_properties(matrixMulDynlinkJIT PROPERTIES
|
|
||||||
CUDA_SEPARABLE_COMPILATION ON
|
|
||||||
POSITION_INDEPENDENT_CODE OFF
|
|
||||||
)
|
|
||||||
set(CMAKE_EXE_LINKER_FLAGS "${CMAKE_EXE_LINKER_FLAGS} -no-pie")
|
|
||||||
target_link_libraries(matrixMulDynlinkJIT PUBLIC
|
|
||||||
CUDA::cudart
|
|
||||||
CUDA::cuda_driver
|
|
||||||
)
|
|
||||||
|
|
||||||
if(UNIX)
|
|
||||||
target_link_libraries(matrixMulDynlinkJIT PUBLIC dl)
|
|
||||||
endif()
|
|
||||||
@@ -1,32 +0,0 @@
|
|||||||
# matrixMulDynlinkJIT - Matrix Multiplication (CUDA Driver API version with Dynamic Linking Version)
|
|
||||||
|
|
||||||
## Description
|
|
||||||
|
|
||||||
This sample revisits matrix multiplication using the CUDA driver API. It demonstrates how to link to CUDA driver at runtime and how to use JIT (just-in-time) compilation from PTX code. It has been written for clarity of exposition to illustrate various CUDA programming principles, not with the goal of providing the most performant generic kernel for matrix multiplication. CUBLAS provides high-performance matrix multiplication.
|
|
||||||
|
|
||||||
## Key Concepts
|
|
||||||
|
|
||||||
CUDA Driver API, CUDA Dynamically Linked Library
|
|
||||||
|
|
||||||
## Supported SM Architectures
|
|
||||||
|
|
||||||
[SM 5.0 ](https://developer.nvidia.com/cuda-gpus) [SM 5.2 ](https://developer.nvidia.com/cuda-gpus) [SM 5.3 ](https://developer.nvidia.com/cuda-gpus) [SM 6.0 ](https://developer.nvidia.com/cuda-gpus) [SM 6.1 ](https://developer.nvidia.com/cuda-gpus) [SM 7.0 ](https://developer.nvidia.com/cuda-gpus) [SM 7.2 ](https://developer.nvidia.com/cuda-gpus) [SM 7.5 ](https://developer.nvidia.com/cuda-gpus) [SM 8.0 ](https://developer.nvidia.com/cuda-gpus) [SM 8.6 ](https://developer.nvidia.com/cuda-gpus) [SM 8.7 ](https://developer.nvidia.com/cuda-gpus) [SM 8.9 ](https://developer.nvidia.com/cuda-gpus) [SM 9.0 ](https://developer.nvidia.com/cuda-gpus)
|
|
||||||
|
|
||||||
## Supported OSes
|
|
||||||
|
|
||||||
Linux, Windows
|
|
||||||
|
|
||||||
## Supported CPU Architecture
|
|
||||||
|
|
||||||
x86_64, armv7l
|
|
||||||
|
|
||||||
## CUDA APIs involved
|
|
||||||
|
|
||||||
### [CUDA Driver API](http://docs.nvidia.com/cuda/cuda-driver-api/index.html)
|
|
||||||
cuMemcpyDtoH, cuDeviceGetName, cuParamSeti, cuModuleLoadDataEx, cuModuleGetFunction, cuLaunchGrid, cuFuncSetSharedSize, cuMemFree, cuParamSetSize, cuParamSetv, cuInit, cuMemcpyHtoD, cuLaunchKernel, cuDeviceGet, cuFuncSetBlockShape, cuCtxDestroy, cuDeviceGetCount, cuDeviceComputeCapability, cuCtxSynchronize, cuMemAlloc, cuCtxCreate
|
|
||||||
|
|
||||||
## Prerequisites
|
|
||||||
|
|
||||||
Download and install the [CUDA Toolkit 12.5](https://developer.nvidia.com/cuda-downloads) for your corresponding platform.
|
|
||||||
|
|
||||||
## References (for more details)
|
|
||||||
@@ -1,647 +0,0 @@
|
|||||||
/*
|
|
||||||
* Copyright 1993-2014 NVIDIA Corporation. All rights reserved.
|
|
||||||
*
|
|
||||||
* NVIDIA Corporation and its licensors retain all intellectual property and
|
|
||||||
* proprietary rights in and to this software and related documentation.
|
|
||||||
* Any use, reproduction, disclosure, or distribution of this software
|
|
||||||
* and related documentation without an express license agreement from
|
|
||||||
* NVIDIA Corporation is strictly prohibited.
|
|
||||||
*
|
|
||||||
* Please refer to the applicable NVIDIA end user license agreement (EULA)
|
|
||||||
* associated with this source code for terms and conditions that govern
|
|
||||||
* your use of this NVIDIA software.
|
|
||||||
*
|
|
||||||
*/
|
|
||||||
|
|
||||||
// With these flags defined, this source file will dynamically
|
|
||||||
// load the corresponding functions. Disabled by default.
|
|
||||||
//#define CUDA_INIT_D3D9
|
|
||||||
//#define CUDA_INIT_D3D10
|
|
||||||
//#define CUDA_INIT_D3D11
|
|
||||||
//#define CUDA_INIT_OPENGL
|
|
||||||
|
|
||||||
#include <stdio.h>
|
|
||||||
#include "cuda_drvapi_dynlink.h"
|
|
||||||
|
|
||||||
tcuInit *_cuInit;
|
|
||||||
tcuDriverGetVersion *cuDriverGetVersion;
|
|
||||||
tcuDeviceGet *cuDeviceGet;
|
|
||||||
tcuDeviceGetCount *cuDeviceGetCount;
|
|
||||||
tcuDeviceGetName *cuDeviceGetName;
|
|
||||||
tcuDeviceComputeCapability *cuDeviceComputeCapability;
|
|
||||||
tcuDeviceTotalMem *cuDeviceTotalMem;
|
|
||||||
tcuDeviceGetProperties *cuDeviceGetProperties;
|
|
||||||
tcuDeviceGetAttribute *cuDeviceGetAttribute;
|
|
||||||
tcuGetErrorString *cuGetErrorString;
|
|
||||||
tcuCtxCreate *cuCtxCreate;
|
|
||||||
tcuCtxDestroy *cuCtxDestroy;
|
|
||||||
tcuCtxAttach *cuCtxAttach;
|
|
||||||
tcuCtxDetach *cuCtxDetach;
|
|
||||||
tcuCtxPushCurrent *cuCtxPushCurrent;
|
|
||||||
tcuCtxPopCurrent *cuCtxPopCurrent;
|
|
||||||
tcuCtxGetCurrent *cuCtxGetCurrent;
|
|
||||||
tcuCtxSetCurrent *cuCtxSetCurrent;
|
|
||||||
tcuCtxGetDevice *cuCtxGetDevice;
|
|
||||||
tcuCtxSynchronize *cuCtxSynchronize;
|
|
||||||
tcuModuleLoad *cuModuleLoad;
|
|
||||||
tcuModuleLoadData *cuModuleLoadData;
|
|
||||||
tcuModuleLoadDataEx *cuModuleLoadDataEx;
|
|
||||||
tcuModuleLoadFatBinary *cuModuleLoadFatBinary;
|
|
||||||
tcuModuleUnload *cuModuleUnload;
|
|
||||||
tcuModuleGetFunction *cuModuleGetFunction;
|
|
||||||
tcuModuleGetGlobal *cuModuleGetGlobal;
|
|
||||||
tcuModuleGetTexRef *cuModuleGetTexRef;
|
|
||||||
tcuModuleGetSurfRef *cuModuleGetSurfRef;
|
|
||||||
tcuMemGetInfo *cuMemGetInfo;
|
|
||||||
tcuMemAlloc *cuMemAlloc;
|
|
||||||
tcuMemAllocPitch *cuMemAllocPitch;
|
|
||||||
tcuMemFree *cuMemFree;
|
|
||||||
tcuMemGetAddressRange *cuMemGetAddressRange;
|
|
||||||
tcuMemAllocHost *cuMemAllocHost;
|
|
||||||
tcuMemFreeHost *cuMemFreeHost;
|
|
||||||
tcuMemHostAlloc *cuMemHostAlloc;
|
|
||||||
tcuMemHostGetFlags *cuMemHostGetFlags;
|
|
||||||
|
|
||||||
tcuMemHostGetDevicePointer *cuMemHostGetDevicePointer;
|
|
||||||
tcuDeviceGetByPCIBusId *cuDeviceGetByPCIBusId;
|
|
||||||
tcuDeviceGetPCIBusId *cuDeviceGetPCIBusId;
|
|
||||||
tcuIpcGetEventHandle *cuIpcGetEventHandle;
|
|
||||||
tcuIpcOpenEventHandle *cuIpcOpenEventHandle;
|
|
||||||
tcuIpcGetMemHandle *cuIpcGetMemHandle;
|
|
||||||
tcuIpcOpenMemHandle *cuIpcOpenMemHandle;
|
|
||||||
tcuIpcCloseMemHandle *cuIpcCloseMemHandle;
|
|
||||||
|
|
||||||
tcuMemHostRegister *cuMemHostRegister;
|
|
||||||
tcuMemHostUnregister *cuMemHostUnregister;
|
|
||||||
tcuMemcpyHtoD *cuMemcpyHtoD;
|
|
||||||
tcuMemcpyDtoH *cuMemcpyDtoH;
|
|
||||||
tcuMemcpyDtoD *cuMemcpyDtoD;
|
|
||||||
tcuMemcpyDtoA *cuMemcpyDtoA;
|
|
||||||
tcuMemcpyAtoD *cuMemcpyAtoD;
|
|
||||||
tcuMemcpyHtoA *cuMemcpyHtoA;
|
|
||||||
tcuMemcpyAtoH *cuMemcpyAtoH;
|
|
||||||
tcuMemcpyAtoA *cuMemcpyAtoA;
|
|
||||||
tcuMemcpy2D *cuMemcpy2D;
|
|
||||||
tcuMemcpy2DUnaligned *cuMemcpy2DUnaligned;
|
|
||||||
tcuMemcpy3D *cuMemcpy3D;
|
|
||||||
tcuMemcpyHtoDAsync *cuMemcpyHtoDAsync;
|
|
||||||
tcuMemcpyDtoHAsync *cuMemcpyDtoHAsync;
|
|
||||||
tcuMemcpyDtoDAsync *cuMemcpyDtoDAsync;
|
|
||||||
tcuMemcpyHtoAAsync *cuMemcpyHtoAAsync;
|
|
||||||
tcuMemcpyAtoHAsync *cuMemcpyAtoHAsync;
|
|
||||||
tcuMemcpy2DAsync *cuMemcpy2DAsync;
|
|
||||||
tcuMemcpy3DAsync *cuMemcpy3DAsync;
|
|
||||||
tcuMemcpy *cuMemcpy;
|
|
||||||
tcuMemcpyPeer *cuMemcpyPeer;
|
|
||||||
tcuMemsetD8 *cuMemsetD8;
|
|
||||||
tcuMemsetD16 *cuMemsetD16;
|
|
||||||
tcuMemsetD32 *cuMemsetD32;
|
|
||||||
tcuMemsetD2D8 *cuMemsetD2D8;
|
|
||||||
tcuMemsetD2D16 *cuMemsetD2D16;
|
|
||||||
tcuMemsetD2D32 *cuMemsetD2D32;
|
|
||||||
tcuFuncSetBlockShape *cuFuncSetBlockShape;
|
|
||||||
tcuFuncSetSharedSize *cuFuncSetSharedSize;
|
|
||||||
tcuFuncGetAttribute *cuFuncGetAttribute;
|
|
||||||
tcuFuncSetCacheConfig *cuFuncSetCacheConfig;
|
|
||||||
tcuFuncSetSharedMemConfig *cuFuncSetSharedMemConfig;
|
|
||||||
tcuLaunchKernel *cuLaunchKernel;
|
|
||||||
tcuArrayCreate *cuArrayCreate;
|
|
||||||
tcuArrayGetDescriptor *cuArrayGetDescriptor;
|
|
||||||
tcuArrayDestroy *cuArrayDestroy;
|
|
||||||
tcuArray3DCreate *cuArray3DCreate;
|
|
||||||
tcuArray3DGetDescriptor *cuArray3DGetDescriptor;
|
|
||||||
tcuTexRefCreate *cuTexRefCreate;
|
|
||||||
tcuTexRefDestroy *cuTexRefDestroy;
|
|
||||||
tcuTexRefSetArray *cuTexRefSetArray;
|
|
||||||
tcuTexRefSetAddress *cuTexRefSetAddress;
|
|
||||||
tcuTexRefSetAddress2D *cuTexRefSetAddress2D;
|
|
||||||
tcuTexRefSetFormat *cuTexRefSetFormat;
|
|
||||||
tcuTexRefSetAddressMode *cuTexRefSetAddressMode;
|
|
||||||
tcuTexRefSetFilterMode *cuTexRefSetFilterMode;
|
|
||||||
tcuTexRefSetFlags *cuTexRefSetFlags;
|
|
||||||
tcuTexRefGetAddress *cuTexRefGetAddress;
|
|
||||||
tcuTexRefGetArray *cuTexRefGetArray;
|
|
||||||
tcuTexRefGetAddressMode *cuTexRefGetAddressMode;
|
|
||||||
tcuTexRefGetFilterMode *cuTexRefGetFilterMode;
|
|
||||||
tcuTexRefGetFormat *cuTexRefGetFormat;
|
|
||||||
tcuTexRefGetFlags *cuTexRefGetFlags;
|
|
||||||
tcuSurfRefSetArray *cuSurfRefSetArray;
|
|
||||||
tcuSurfRefGetArray *cuSurfRefGetArray;
|
|
||||||
tcuParamSetSize *cuParamSetSize;
|
|
||||||
tcuParamSeti *cuParamSeti;
|
|
||||||
tcuParamSetf *cuParamSetf;
|
|
||||||
tcuParamSetv *cuParamSetv;
|
|
||||||
tcuParamSetTexRef *cuParamSetTexRef;
|
|
||||||
tcuLaunch *cuLaunch;
|
|
||||||
tcuLaunchGrid *cuLaunchGrid;
|
|
||||||
tcuLaunchGridAsync *cuLaunchGridAsync;
|
|
||||||
tcuEventCreate *cuEventCreate;
|
|
||||||
tcuEventRecord *cuEventRecord;
|
|
||||||
tcuEventQuery *cuEventQuery;
|
|
||||||
tcuEventSynchronize *cuEventSynchronize;
|
|
||||||
tcuEventDestroy *cuEventDestroy;
|
|
||||||
tcuEventElapsedTime *cuEventElapsedTime;
|
|
||||||
tcuStreamCreate *cuStreamCreate;
|
|
||||||
tcuStreamWaitEvent *cuStreamWaitEvent;
|
|
||||||
tcuStreamAddCallback *cuStreamAddCallback;
|
|
||||||
tcuStreamQuery *cuStreamQuery;
|
|
||||||
tcuStreamSynchronize *cuStreamSynchronize;
|
|
||||||
tcuStreamDestroy *cuStreamDestroy;
|
|
||||||
tcuGraphicsUnregisterResource *cuGraphicsUnregisterResource;
|
|
||||||
tcuGraphicsSubResourceGetMappedArray *cuGraphicsSubResourceGetMappedArray;
|
|
||||||
tcuGraphicsResourceGetMappedPointer *cuGraphicsResourceGetMappedPointer;
|
|
||||||
tcuGraphicsResourceSetMapFlags *cuGraphicsResourceSetMapFlags;
|
|
||||||
tcuGraphicsMapResources *cuGraphicsMapResources;
|
|
||||||
tcuGraphicsUnmapResources *cuGraphicsUnmapResources;
|
|
||||||
tcuGetExportTable *cuGetExportTable;
|
|
||||||
tcuCtxSetLimit *cuCtxSetLimit;
|
|
||||||
tcuCtxGetLimit *cuCtxGetLimit;
|
|
||||||
tcuCtxGetCacheConfig *cuCtxGetCacheConfig;
|
|
||||||
tcuCtxSetCacheConfig *cuCtxSetCacheConfig;
|
|
||||||
tcuCtxGetSharedMemConfig *cuCtxGetSharedMemConfig;
|
|
||||||
tcuCtxSetSharedMemConfig *cuCtxSetSharedMemConfig;
|
|
||||||
tcuCtxGetApiVersion *cuCtxGetApiVersion;
|
|
||||||
|
|
||||||
tcuMipmappedArrayCreate *cuMipmappedArrayCreate;
|
|
||||||
tcuMipmappedArrayGetLevel *cuMipmappedArrayGetLevel;
|
|
||||||
tcuMipmappedArrayDestroy *cuMipmappedArrayDestroy;
|
|
||||||
|
|
||||||
tcuProfilerStop *cuProfilerStop;
|
|
||||||
|
|
||||||
#ifdef CUDA_INIT_D3D9
|
|
||||||
// D3D9/CUDA interop (CUDA 1.x compatible API). These functions
|
|
||||||
// are deprecated; please use the ones below
|
|
||||||
tcuD3D9Begin *cuD3D9Begin;
|
|
||||||
tcuD3D9End *cuD3DEnd;
|
|
||||||
tcuD3D9RegisterVertexBuffer *cuD3D9RegisterVertexBuffer;
|
|
||||||
tcuD3D9MapVertexBuffer *cuD3D9MapVertexBuffer;
|
|
||||||
tcuD3D9UnmapVertexBuffer *cuD3D9UnmapVertexBuffer;
|
|
||||||
tcuD3D9UnregisterVertexBuffer *cuD3D9UnregisterVertexBuffer;
|
|
||||||
|
|
||||||
// D3D9/CUDA interop (CUDA 2.x compatible)
|
|
||||||
tcuD3D9GetDirect3DDevice *cuD3D9GetDirect3DDevice;
|
|
||||||
tcuD3D9RegisterResource *cuD3D9RegisterResource;
|
|
||||||
tcuD3D9UnregisterResource *cuD3D9UnregisterResource;
|
|
||||||
tcuD3D9MapResources *cuD3D9MapResources;
|
|
||||||
tcuD3D9UnmapResources *cuD3D9UnmapResources;
|
|
||||||
tcuD3D9ResourceSetMapFlags *cuD3D9ResourceSetMapFlags;
|
|
||||||
tcuD3D9ResourceGetSurfaceDimensions *cuD3D9ResourceGetSurfaceDimensions;
|
|
||||||
tcuD3D9ResourceGetMappedArray *cuD3D9ResourceGetMappedArray;
|
|
||||||
tcuD3D9ResourceGetMappedPointer *cuD3D9ResourceGetMappedPointer;
|
|
||||||
tcuD3D9ResourceGetMappedSize *cuD3D9ResourceGetMappedSize;
|
|
||||||
tcuD3D9ResourceGetMappedPitch *cuD3D9ResourceGetMappedPitch;
|
|
||||||
|
|
||||||
// D3D9/CUDA interop (CUDA 2.0+)
|
|
||||||
tcuD3D9GetDevice *cuD3D9GetDevice;
|
|
||||||
tcuD3D9CtxCreate *cuD3D9CtxCreate;
|
|
||||||
tcuGraphicsD3D9RegisterResource *cuGraphicsD3D9RegisterResource;
|
|
||||||
#endif
|
|
||||||
|
|
||||||
#ifdef CUDA_INIT_D3D10
|
|
||||||
// D3D10/CUDA interop (CUDA 3.0+)
|
|
||||||
tcuD3D10GetDevice *cuD3D10GetDevice;
|
|
||||||
tcuD3D10CtxCreate *cuD3D10CtxCreate;
|
|
||||||
tcuGraphicsD3D10RegisterResource *cuGraphicsD3D10RegisterResource;
|
|
||||||
#endif
|
|
||||||
|
|
||||||
|
|
||||||
#ifdef CUDA_INIT_D3D11
|
|
||||||
// D3D11/CUDA interop (CUDA 3.0+)
|
|
||||||
tcuD3D11GetDevice *cuD3D11GetDevice;
|
|
||||||
tcuD3D11CtxCreate *cuD3D11CtxCreate;
|
|
||||||
tcuGraphicsD3D11RegisterResource *cuGraphicsD3D11RegisterResource;
|
|
||||||
#endif
|
|
||||||
|
|
||||||
// GL/CUDA interop
|
|
||||||
#ifdef CUDA_INIT_OPENGL
|
|
||||||
tcuGLCtxCreate *cuGLCtxCreate;
|
|
||||||
tcuGraphicsGLRegisterBuffer *cuGraphicsGLRegisterBuffer;
|
|
||||||
tcuGraphicsGLRegisterImage *cuGraphicsGLRegisterImage;
|
|
||||||
#if defined(WIN32) || defined(_WIN32) || defined(WIN64) || defined(_WIN64)
|
|
||||||
tcuWGLGetDevice *cuWGLGetDevice;
|
|
||||||
#endif
|
|
||||||
#endif
|
|
||||||
|
|
||||||
#define STRINGIFY(X) #X
|
|
||||||
|
|
||||||
#if defined(WIN32) || defined(_WIN32) || defined(WIN64) || defined(_WIN64)
|
|
||||||
#include <Windows.h>
|
|
||||||
|
|
||||||
#ifdef UNICODE
|
|
||||||
static LPCWSTR __CudaLibName = L"nvcuda.dll";
|
|
||||||
#else
|
|
||||||
static LPCSTR __CudaLibName = "nvcuda.dll";
|
|
||||||
#endif
|
|
||||||
|
|
||||||
typedef HMODULE CUDADRIVER;
|
|
||||||
|
|
||||||
static CUresult LOAD_LIBRARY(CUDADRIVER *pInstance)
|
|
||||||
{
|
|
||||||
*pInstance = LoadLibrary(__CudaLibName);
|
|
||||||
|
|
||||||
if (*pInstance == NULL)
|
|
||||||
{
|
|
||||||
printf("LoadLibrary \"%s\" failed!\n", __CudaLibName);
|
|
||||||
return CUDA_ERROR_UNKNOWN;
|
|
||||||
}
|
|
||||||
|
|
||||||
return CUDA_SUCCESS;
|
|
||||||
}
|
|
||||||
|
|
||||||
#define GET_PROC_EX(name, alias, required) \
|
|
||||||
alias = (t##name *)GetProcAddress(CudaDrvLib, #name); \
|
|
||||||
if (alias == NULL && required) { \
|
|
||||||
printf("Failed to find required function \"%s\" in %s\n", \
|
|
||||||
#name, __CudaLibName); \
|
|
||||||
return CUDA_ERROR_UNKNOWN; \
|
|
||||||
}
|
|
||||||
|
|
||||||
#define GET_PROC_EX_V2(name, alias, required) \
|
|
||||||
alias = (t##name *)GetProcAddress(CudaDrvLib, STRINGIFY(name##_v2));\
|
|
||||||
if (alias == NULL && required) { \
|
|
||||||
printf("Failed to find required function \"%s\" in %s\n", \
|
|
||||||
STRINGIFY(name##_v2), __CudaLibName); \
|
|
||||||
return CUDA_ERROR_UNKNOWN; \
|
|
||||||
}
|
|
||||||
|
|
||||||
#define GET_PROC_EX_V3(name, alias, required) \
|
|
||||||
alias = (t##name *)GetProcAddress(CudaDrvLib, STRINGIFY(name##_v3));\
|
|
||||||
if (alias == NULL && required) { \
|
|
||||||
printf("Failed to find required function \"%s\" in %s\n", \
|
|
||||||
STRINGIFY(name##_v3), __CudaLibName); \
|
|
||||||
return CUDA_ERROR_UNKNOWN; \
|
|
||||||
}
|
|
||||||
|
|
||||||
#elif defined(__unix__) || defined (__QNX__) || defined(__APPLE__) || defined(__MACOSX)
|
|
||||||
|
|
||||||
#include <dlfcn.h>
|
|
||||||
|
|
||||||
#if defined(__APPLE__) || defined(__MACOSX)
|
|
||||||
static char __CudaLibName[] = "/usr/local/cuda/lib/libcuda.dylib";
|
|
||||||
#elif defined(__ANDROID__)
|
|
||||||
#if defined (__aarch64__)
|
|
||||||
static char __CudaLibName[] = "/system/vendor/lib64/libcuda.so";
|
|
||||||
#elif defined(__arm__)
|
|
||||||
static char __CudaLibName[] = "/system/vendor/lib/libcuda.so";
|
|
||||||
#endif
|
|
||||||
#else
|
|
||||||
static char __CudaLibName[] = "libcuda.so.1";
|
|
||||||
#endif
|
|
||||||
|
|
||||||
typedef void *CUDADRIVER;
|
|
||||||
|
|
||||||
static CUresult LOAD_LIBRARY(CUDADRIVER *pInstance)
|
|
||||||
{
|
|
||||||
*pInstance = dlopen(__CudaLibName, RTLD_NOW);
|
|
||||||
|
|
||||||
if (*pInstance == NULL)
|
|
||||||
{
|
|
||||||
printf("dlopen \"%s\" failed!\n", __CudaLibName);
|
|
||||||
return CUDA_ERROR_UNKNOWN;
|
|
||||||
}
|
|
||||||
|
|
||||||
return CUDA_SUCCESS;
|
|
||||||
}
|
|
||||||
|
|
||||||
#define GET_PROC_EX(name, alias, required) \
|
|
||||||
alias = (t##name *)dlsym(CudaDrvLib, #name); \
|
|
||||||
if (alias == NULL && required) { \
|
|
||||||
printf("Failed to find required function \"%s\" in %s\n", \
|
|
||||||
#name, __CudaLibName); \
|
|
||||||
return CUDA_ERROR_UNKNOWN; \
|
|
||||||
}
|
|
||||||
|
|
||||||
#define GET_PROC_EX_V2(name, alias, required) \
|
|
||||||
alias = (t##name *)dlsym(CudaDrvLib, STRINGIFY(name##_v2)); \
|
|
||||||
if (alias == NULL && required) { \
|
|
||||||
printf("Failed to find required function \"%s\" in %s\n", \
|
|
||||||
STRINGIFY(name##_v2), __CudaLibName); \
|
|
||||||
return CUDA_ERROR_UNKNOWN; \
|
|
||||||
}
|
|
||||||
|
|
||||||
#define GET_PROC_EX_V3(name, alias, required) \
|
|
||||||
alias = (t##name *)dlsym(CudaDrvLib, STRINGIFY(name##_v3)); \
|
|
||||||
if (alias == NULL && required) { \
|
|
||||||
printf("Failed to find required function \"%s\" in %s\n", \
|
|
||||||
STRINGIFY(name##_v3), __CudaLibName); \
|
|
||||||
return CUDA_ERROR_UNKNOWN; \
|
|
||||||
}
|
|
||||||
|
|
||||||
#else
|
|
||||||
#error unsupported platform
|
|
||||||
#endif
|
|
||||||
|
|
||||||
#define CHECKED_CALL(call) \
|
|
||||||
do { \
|
|
||||||
CUresult result = (call); \
|
|
||||||
if (CUDA_SUCCESS != result) { \
|
|
||||||
return result; \
|
|
||||||
} \
|
|
||||||
} while(0)
|
|
||||||
|
|
||||||
#define GET_PROC_REQUIRED(name) GET_PROC_EX(name,name,1)
|
|
||||||
#define GET_PROC_OPTIONAL(name) GET_PROC_EX(name,name,0)
|
|
||||||
#define GET_PROC(name) GET_PROC_REQUIRED(name)
|
|
||||||
#define GET_PROC_V2(name) GET_PROC_EX_V2(name,name,1)
|
|
||||||
#define GET_PROC_V3(name) GET_PROC_EX_V3(name,name,1)
|
|
||||||
|
|
||||||
CUresult CUDAAPI cuInit(unsigned int Flags, int cudaVersion)
|
|
||||||
{
|
|
||||||
CUDADRIVER CudaDrvLib;
|
|
||||||
int driverVer = 1000;
|
|
||||||
|
|
||||||
CHECKED_CALL(LOAD_LIBRARY(&CudaDrvLib));
|
|
||||||
|
|
||||||
// cuInit is required; alias it to _cuInit
|
|
||||||
GET_PROC_EX(cuInit, _cuInit, 1);
|
|
||||||
CHECKED_CALL(_cuInit(Flags));
|
|
||||||
|
|
||||||
// available since 2.2. if not present, version 1.0 is assumed
|
|
||||||
GET_PROC_OPTIONAL(cuDriverGetVersion);
|
|
||||||
|
|
||||||
if (cuDriverGetVersion)
|
|
||||||
{
|
|
||||||
CHECKED_CALL(cuDriverGetVersion(&driverVer));
|
|
||||||
}
|
|
||||||
|
|
||||||
// fetch all function pointers
|
|
||||||
GET_PROC(cuDeviceGet);
|
|
||||||
GET_PROC(cuDeviceGetCount);
|
|
||||||
GET_PROC(cuDeviceGetName);
|
|
||||||
GET_PROC(cuDeviceComputeCapability);
|
|
||||||
GET_PROC(cuDeviceGetProperties);
|
|
||||||
GET_PROC(cuDeviceGetAttribute);
|
|
||||||
GET_PROC(cuGetErrorString);
|
|
||||||
GET_PROC(cuCtxDestroy);
|
|
||||||
GET_PROC(cuCtxAttach);
|
|
||||||
GET_PROC(cuCtxDetach);
|
|
||||||
GET_PROC(cuCtxPushCurrent);
|
|
||||||
GET_PROC(cuCtxPopCurrent);
|
|
||||||
GET_PROC(cuCtxGetDevice);
|
|
||||||
GET_PROC(cuCtxSynchronize);
|
|
||||||
GET_PROC(cuCtxSetLimit);
|
|
||||||
GET_PROC(cuCtxGetCacheConfig);
|
|
||||||
GET_PROC(cuCtxSetCacheConfig);
|
|
||||||
GET_PROC(cuCtxGetApiVersion);
|
|
||||||
GET_PROC(cuModuleLoad);
|
|
||||||
GET_PROC(cuModuleLoadData);
|
|
||||||
GET_PROC(cuModuleUnload);
|
|
||||||
GET_PROC(cuModuleGetFunction);
|
|
||||||
GET_PROC(cuModuleGetTexRef);
|
|
||||||
GET_PROC(cuMemFreeHost);
|
|
||||||
GET_PROC(cuMemHostAlloc);
|
|
||||||
GET_PROC(cuFuncSetBlockShape);
|
|
||||||
GET_PROC(cuFuncSetSharedSize);
|
|
||||||
GET_PROC(cuFuncGetAttribute);
|
|
||||||
GET_PROC(cuArrayDestroy);
|
|
||||||
GET_PROC(cuTexRefCreate);
|
|
||||||
GET_PROC(cuTexRefDestroy);
|
|
||||||
GET_PROC(cuTexRefSetArray);
|
|
||||||
GET_PROC(cuTexRefSetFormat);
|
|
||||||
GET_PROC(cuTexRefSetAddressMode);
|
|
||||||
GET_PROC(cuTexRefSetFilterMode);
|
|
||||||
GET_PROC(cuTexRefSetFlags);
|
|
||||||
GET_PROC(cuTexRefGetArray);
|
|
||||||
GET_PROC(cuTexRefGetAddressMode);
|
|
||||||
GET_PROC(cuTexRefGetFilterMode);
|
|
||||||
GET_PROC(cuTexRefGetFormat);
|
|
||||||
GET_PROC(cuTexRefGetFlags);
|
|
||||||
GET_PROC(cuParamSetSize);
|
|
||||||
GET_PROC(cuParamSeti);
|
|
||||||
GET_PROC(cuParamSetf);
|
|
||||||
GET_PROC(cuParamSetv);
|
|
||||||
GET_PROC(cuParamSetTexRef);
|
|
||||||
GET_PROC(cuLaunch);
|
|
||||||
GET_PROC(cuLaunchGrid);
|
|
||||||
GET_PROC(cuLaunchGridAsync);
|
|
||||||
GET_PROC(cuEventCreate);
|
|
||||||
GET_PROC(cuEventRecord);
|
|
||||||
GET_PROC(cuEventQuery);
|
|
||||||
GET_PROC(cuEventSynchronize);
|
|
||||||
GET_PROC(cuEventDestroy);
|
|
||||||
GET_PROC(cuEventElapsedTime);
|
|
||||||
GET_PROC(cuStreamCreate);
|
|
||||||
GET_PROC(cuStreamWaitEvent);
|
|
||||||
GET_PROC(cuStreamAddCallback);
|
|
||||||
GET_PROC(cuStreamQuery);
|
|
||||||
GET_PROC(cuStreamSynchronize);
|
|
||||||
GET_PROC(cuStreamDestroy);
|
|
||||||
|
|
||||||
// These are CUDA 5.0 new functions
|
|
||||||
if (driverVer >= 5000)
|
|
||||||
{
|
|
||||||
GET_PROC(cuMipmappedArrayCreate);
|
|
||||||
GET_PROC(cuMipmappedArrayDestroy);
|
|
||||||
GET_PROC(cuMipmappedArrayGetLevel);
|
|
||||||
}
|
|
||||||
|
|
||||||
// These are CUDA 4.2 new functions
|
|
||||||
if (driverVer >= 4020)
|
|
||||||
{
|
|
||||||
GET_PROC(cuFuncSetSharedMemConfig);
|
|
||||||
GET_PROC(cuCtxGetSharedMemConfig);
|
|
||||||
GET_PROC(cuCtxSetSharedMemConfig);
|
|
||||||
}
|
|
||||||
|
|
||||||
// These are CUDA 4.1 new functions
|
|
||||||
if (cudaVersion >= 4010 && __CUDA_API_VERSION >= 4010)
|
|
||||||
{
|
|
||||||
GET_PROC(cuDeviceGetByPCIBusId);
|
|
||||||
GET_PROC(cuDeviceGetPCIBusId);
|
|
||||||
GET_PROC(cuIpcGetEventHandle);
|
|
||||||
GET_PROC(cuIpcOpenEventHandle);
|
|
||||||
GET_PROC(cuIpcGetMemHandle);
|
|
||||||
GET_PROC(cuIpcOpenMemHandle);
|
|
||||||
GET_PROC(cuIpcCloseMemHandle);
|
|
||||||
}
|
|
||||||
|
|
||||||
// These could be _v2 interfaces
|
|
||||||
if (cudaVersion >= 4000 && __CUDA_API_VERSION >= 4000)
|
|
||||||
{
|
|
||||||
GET_PROC_V2(cuCtxDestroy);
|
|
||||||
GET_PROC_V2(cuCtxPopCurrent);
|
|
||||||
GET_PROC_V2(cuCtxPushCurrent);
|
|
||||||
GET_PROC_V2(cuStreamDestroy);
|
|
||||||
GET_PROC_V2(cuEventDestroy);
|
|
||||||
}
|
|
||||||
|
|
||||||
if (cudaVersion >= 3020 && __CUDA_API_VERSION >= 3020)
|
|
||||||
{
|
|
||||||
GET_PROC_V2(cuDeviceTotalMem);
|
|
||||||
GET_PROC_V2(cuCtxCreate);
|
|
||||||
GET_PROC_V2(cuModuleGetGlobal);
|
|
||||||
GET_PROC_V2(cuMemGetInfo);
|
|
||||||
GET_PROC_V2(cuMemAlloc);
|
|
||||||
GET_PROC_V2(cuMemAllocPitch);
|
|
||||||
GET_PROC_V2(cuMemFree);
|
|
||||||
GET_PROC_V2(cuMemGetAddressRange);
|
|
||||||
GET_PROC_V2(cuMemAllocHost);
|
|
||||||
GET_PROC_V2(cuMemHostGetDevicePointer);
|
|
||||||
GET_PROC_V2(cuMemcpyHtoD);
|
|
||||||
GET_PROC_V2(cuMemcpyDtoH);
|
|
||||||
GET_PROC_V2(cuMemcpyDtoD);
|
|
||||||
GET_PROC_V2(cuMemcpyDtoA);
|
|
||||||
GET_PROC_V2(cuMemcpyAtoD);
|
|
||||||
GET_PROC_V2(cuMemcpyHtoA);
|
|
||||||
GET_PROC_V2(cuMemcpyAtoH);
|
|
||||||
GET_PROC_V2(cuMemcpyAtoA);
|
|
||||||
GET_PROC_V2(cuMemcpy2D);
|
|
||||||
GET_PROC_V2(cuMemcpy2DUnaligned);
|
|
||||||
GET_PROC_V2(cuMemcpy3D);
|
|
||||||
GET_PROC_V2(cuMemcpyHtoDAsync);
|
|
||||||
GET_PROC_V2(cuMemcpyDtoHAsync);
|
|
||||||
GET_PROC_V2(cuMemcpyHtoAAsync);
|
|
||||||
GET_PROC_V2(cuMemcpyAtoHAsync);
|
|
||||||
GET_PROC_V2(cuMemcpy2DAsync);
|
|
||||||
GET_PROC_V2(cuMemcpy3DAsync);
|
|
||||||
GET_PROC_V2(cuMemsetD8);
|
|
||||||
GET_PROC_V2(cuMemsetD16);
|
|
||||||
GET_PROC_V2(cuMemsetD32);
|
|
||||||
GET_PROC_V2(cuMemsetD2D8);
|
|
||||||
GET_PROC_V2(cuMemsetD2D16);
|
|
||||||
GET_PROC_V2(cuMemsetD2D32);
|
|
||||||
GET_PROC_V2(cuArrayCreate);
|
|
||||||
GET_PROC_V2(cuArrayGetDescriptor);
|
|
||||||
GET_PROC_V2(cuArray3DCreate);
|
|
||||||
GET_PROC_V2(cuArray3DGetDescriptor);
|
|
||||||
GET_PROC_V2(cuTexRefSetAddress);
|
|
||||||
GET_PROC_V2(cuTexRefGetAddress);
|
|
||||||
|
|
||||||
if (cudaVersion >= 4010 && __CUDA_API_VERSION >= 4010)
|
|
||||||
{
|
|
||||||
GET_PROC_V3(cuTexRefSetAddress2D);
|
|
||||||
}
|
|
||||||
else
|
|
||||||
{
|
|
||||||
GET_PROC_V2(cuTexRefSetAddress2D);
|
|
||||||
}
|
|
||||||
}
|
|
||||||
else
|
|
||||||
{
|
|
||||||
// versions earlier than 3020
|
|
||||||
GET_PROC(cuDeviceTotalMem);
|
|
||||||
GET_PROC(cuCtxCreate);
|
|
||||||
GET_PROC(cuModuleGetGlobal);
|
|
||||||
GET_PROC(cuMemGetInfo);
|
|
||||||
GET_PROC(cuMemAlloc);
|
|
||||||
GET_PROC(cuMemAllocPitch);
|
|
||||||
GET_PROC(cuMemFree);
|
|
||||||
GET_PROC(cuMemGetAddressRange);
|
|
||||||
GET_PROC(cuMemAllocHost);
|
|
||||||
GET_PROC(cuMemHostGetDevicePointer);
|
|
||||||
GET_PROC(cuMemcpyHtoD);
|
|
||||||
GET_PROC(cuMemcpyDtoH);
|
|
||||||
GET_PROC(cuMemcpyDtoD);
|
|
||||||
GET_PROC(cuMemcpyDtoA);
|
|
||||||
GET_PROC(cuMemcpyAtoD);
|
|
||||||
GET_PROC(cuMemcpyHtoA);
|
|
||||||
GET_PROC(cuMemcpyAtoH);
|
|
||||||
GET_PROC(cuMemcpyAtoA);
|
|
||||||
GET_PROC(cuMemcpy2D);
|
|
||||||
GET_PROC(cuMemcpy2DUnaligned);
|
|
||||||
GET_PROC(cuMemcpy3D);
|
|
||||||
GET_PROC(cuMemcpyHtoDAsync);
|
|
||||||
GET_PROC(cuMemcpyDtoHAsync);
|
|
||||||
GET_PROC(cuMemcpyHtoAAsync);
|
|
||||||
GET_PROC(cuMemcpyAtoHAsync);
|
|
||||||
GET_PROC(cuMemcpy2DAsync);
|
|
||||||
GET_PROC(cuMemcpy3DAsync);
|
|
||||||
GET_PROC(cuMemsetD8);
|
|
||||||
GET_PROC(cuMemsetD16);
|
|
||||||
GET_PROC(cuMemsetD32);
|
|
||||||
GET_PROC(cuMemsetD2D8);
|
|
||||||
GET_PROC(cuMemsetD2D16);
|
|
||||||
GET_PROC(cuMemsetD2D32);
|
|
||||||
GET_PROC(cuArrayCreate);
|
|
||||||
GET_PROC(cuArrayGetDescriptor);
|
|
||||||
GET_PROC(cuArray3DCreate);
|
|
||||||
GET_PROC(cuArray3DGetDescriptor);
|
|
||||||
GET_PROC(cuTexRefSetAddress);
|
|
||||||
GET_PROC(cuTexRefSetAddress2D);
|
|
||||||
GET_PROC(cuTexRefGetAddress);
|
|
||||||
}
|
|
||||||
|
|
||||||
// The following functions are specific to CUDA versions
|
|
||||||
if (driverVer >= 4000)
|
|
||||||
{
|
|
||||||
GET_PROC(cuCtxSetCurrent);
|
|
||||||
GET_PROC(cuCtxGetCurrent);
|
|
||||||
GET_PROC(cuMemHostRegister);
|
|
||||||
GET_PROC(cuMemHostUnregister);
|
|
||||||
GET_PROC(cuMemcpy);
|
|
||||||
GET_PROC(cuMemcpyPeer);
|
|
||||||
GET_PROC(cuLaunchKernel);
|
|
||||||
GET_PROC(cuProfilerStop);
|
|
||||||
}
|
|
||||||
|
|
||||||
if (driverVer >= 3010)
|
|
||||||
{
|
|
||||||
GET_PROC(cuModuleGetSurfRef);
|
|
||||||
GET_PROC(cuSurfRefSetArray);
|
|
||||||
GET_PROC(cuSurfRefGetArray);
|
|
||||||
GET_PROC(cuCtxSetLimit);
|
|
||||||
GET_PROC(cuCtxGetLimit);
|
|
||||||
}
|
|
||||||
|
|
||||||
if (driverVer >= 3000)
|
|
||||||
{
|
|
||||||
GET_PROC(cuMemcpyDtoDAsync);
|
|
||||||
GET_PROC(cuFuncSetCacheConfig);
|
|
||||||
#ifdef CUDA_INIT_D3D11
|
|
||||||
GET_PROC(cuD3D11GetDevice);
|
|
||||||
GET_PROC(cuD3D11CtxCreate);
|
|
||||||
GET_PROC(cuGraphicsD3D11RegisterResource);
|
|
||||||
#endif
|
|
||||||
GET_PROC(cuGraphicsUnregisterResource);
|
|
||||||
GET_PROC(cuGraphicsSubResourceGetMappedArray);
|
|
||||||
|
|
||||||
if (cudaVersion >= 3020 && __CUDA_API_VERSION >= 3020)
|
|
||||||
{
|
|
||||||
GET_PROC_V2(cuGraphicsResourceGetMappedPointer);
|
|
||||||
}
|
|
||||||
else
|
|
||||||
{
|
|
||||||
GET_PROC(cuGraphicsResourceGetMappedPointer);
|
|
||||||
}
|
|
||||||
|
|
||||||
GET_PROC(cuGraphicsResourceSetMapFlags);
|
|
||||||
GET_PROC(cuGraphicsMapResources);
|
|
||||||
GET_PROC(cuGraphicsUnmapResources);
|
|
||||||
GET_PROC(cuGetExportTable);
|
|
||||||
}
|
|
||||||
|
|
||||||
if (driverVer >= 2030)
|
|
||||||
{
|
|
||||||
GET_PROC(cuMemHostGetFlags);
|
|
||||||
#ifdef CUDA_INIT_D3D10
|
|
||||||
GET_PROC(cuD3D10GetDevice);
|
|
||||||
GET_PROC(cuD3D10CtxCreate);
|
|
||||||
GET_PROC(cuGraphicsD3D10RegisterResource);
|
|
||||||
#endif
|
|
||||||
#ifdef CUDA_INIT_OPENGL
|
|
||||||
GET_PROC(cuGraphicsGLRegisterBuffer);
|
|
||||||
GET_PROC(cuGraphicsGLRegisterImage);
|
|
||||||
#endif
|
|
||||||
}
|
|
||||||
|
|
||||||
if (driverVer >= 2010)
|
|
||||||
{
|
|
||||||
GET_PROC(cuModuleLoadDataEx);
|
|
||||||
GET_PROC(cuModuleLoadFatBinary);
|
|
||||||
#ifdef CUDA_INIT_OPENGL
|
|
||||||
GET_PROC(cuGLCtxCreate);
|
|
||||||
GET_PROC(cuGraphicsGLRegisterBuffer);
|
|
||||||
GET_PROC(cuGraphicsGLRegisterImage);
|
|
||||||
# ifdef WIN32
|
|
||||||
GET_PROC(cuWGLGetDevice);
|
|
||||||
# endif
|
|
||||||
#endif
|
|
||||||
#ifdef CUDA_INIT_D3D9
|
|
||||||
GET_PROC(cuD3D9GetDevice);
|
|
||||||
GET_PROC(cuD3D9CtxCreate);
|
|
||||||
GET_PROC(cuGraphicsD3D9RegisterResource);
|
|
||||||
#endif
|
|
||||||
}
|
|
||||||
|
|
||||||
return CUDA_SUCCESS;
|
|
||||||
}
|
|
||||||
File diff suppressed because it is too large
Load Diff
@@ -1,357 +0,0 @@
|
|||||||
/**
|
|
||||||
* Copyright 1993-2017 NVIDIA Corporation. All rights reserved.
|
|
||||||
*
|
|
||||||
* Please refer to the NVIDIA end user license agreement (EULA) associated
|
|
||||||
* with this source code for terms and conditions that govern your use of
|
|
||||||
* this software. Any use, reproduction, disclosure, or distribution of
|
|
||||||
* this software and related documentation outside the terms of the EULA
|
|
||||||
* is strictly prohibited.
|
|
||||||
*
|
|
||||||
*/
|
|
||||||
|
|
||||||
// Helper functions for CUDA Driver API error handling (make sure that CUDA_H is
|
|
||||||
// included in your projects)
|
|
||||||
#ifndef HELPER_CUDA_DRVAPI_H
|
|
||||||
#define HELPER_CUDA_DRVAPI_H
|
|
||||||
|
|
||||||
#include <stdio.h>
|
|
||||||
#include <stdlib.h>
|
|
||||||
#include <string.h>
|
|
||||||
|
|
||||||
#include <helper_string.h>
|
|
||||||
|
|
||||||
#ifndef MAX
|
|
||||||
#define MAX(a, b) (a > b ? a : b)
|
|
||||||
#endif
|
|
||||||
|
|
||||||
#ifndef HELPER_CUDA_DRVAPI_H
|
|
||||||
inline int ftoi(float value) {
|
|
||||||
return (value >= 0 ? static_cast<int>(value + 0.5)
|
|
||||||
: static_cast<int>(value - 0.5));
|
|
||||||
}
|
|
||||||
#endif
|
|
||||||
|
|
||||||
#ifndef EXIT_WAIVED
|
|
||||||
#define EXIT_WAIVED 2
|
|
||||||
#endif
|
|
||||||
|
|
||||||
////////////////////////////////////////////////////////////////////////////////
|
|
||||||
// These are CUDA Helper functions
|
|
||||||
|
|
||||||
// add a level of protection to the CUDA SDK samples, let's force samples to
|
|
||||||
// explicitly include CUDA.H
|
|
||||||
#ifdef __cuda_cuda_h__
|
|
||||||
// This will output the proper CUDA error strings in the event that a CUDA host
|
|
||||||
// call returns an error
|
|
||||||
#ifndef checkCudaErrors
|
|
||||||
#define checkCudaErrors(err) __checkCudaErrors(err, __FILE__, __LINE__)
|
|
||||||
|
|
||||||
// These are the inline versions for all of the SDK helper functions
|
|
||||||
inline void __checkCudaErrors(CUresult err, const char *file, const int line) {
|
|
||||||
if (CUDA_SUCCESS != err) {
|
|
||||||
const char *errorStr = NULL;
|
|
||||||
cuGetErrorString(err, &errorStr);
|
|
||||||
fprintf(stderr,
|
|
||||||
"checkCudaErrors() Driver API error = %04d \"%s\" from file <%s>, "
|
|
||||||
"line %i.\n",
|
|
||||||
err, errorStr, file, line);
|
|
||||||
exit(EXIT_FAILURE);
|
|
||||||
}
|
|
||||||
}
|
|
||||||
#endif
|
|
||||||
|
|
||||||
// This function wraps the CUDA Driver API into a template function
|
|
||||||
template <class T>
|
|
||||||
inline void getCudaAttribute(T *attribute, CUdevice_attribute device_attribute,
|
|
||||||
int device) {
|
|
||||||
checkCudaErrors(cuDeviceGetAttribute(attribute, device_attribute, device));
|
|
||||||
}
|
|
||||||
#endif
|
|
||||||
|
|
||||||
// Beginning of GPU Architecture definitions
|
|
||||||
inline int _ConvertSMVer2CoresDRV(int major, int minor) {
|
|
||||||
// Defines for GPU Architecture types (using the SM version to determine the #
|
|
||||||
// of cores per SM
|
|
||||||
typedef struct {
|
|
||||||
int SM; // 0xMm (hexidecimal notation), M = SM Major version, and m = SM
|
|
||||||
// minor version
|
|
||||||
int Cores;
|
|
||||||
} sSMtoCores;
|
|
||||||
|
|
||||||
sSMtoCores nGpuArchCoresPerSM[] = {
|
|
||||||
{0x30, 192},
|
|
||||||
{0x32, 192},
|
|
||||||
{0x35, 192},
|
|
||||||
{0x37, 192},
|
|
||||||
{0x50, 128},
|
|
||||||
{0x52, 128},
|
|
||||||
{0x53, 128},
|
|
||||||
{0x60, 64},
|
|
||||||
{0x61, 128},
|
|
||||||
{0x62, 128},
|
|
||||||
{0x70, 64},
|
|
||||||
{0x72, 64},
|
|
||||||
{0x75, 64},
|
|
||||||
{0x80, 64},
|
|
||||||
{0x86, 128},
|
|
||||||
{0x87, 128},
|
|
||||||
{0x90, 128},
|
|
||||||
{-1, -1}};
|
|
||||||
|
|
||||||
int index = 0;
|
|
||||||
|
|
||||||
while (nGpuArchCoresPerSM[index].SM != -1) {
|
|
||||||
if (nGpuArchCoresPerSM[index].SM == ((major << 4) + minor)) {
|
|
||||||
return nGpuArchCoresPerSM[index].Cores;
|
|
||||||
}
|
|
||||||
|
|
||||||
index++;
|
|
||||||
}
|
|
||||||
|
|
||||||
// If we don't find the values, we default use the previous one to run
|
|
||||||
// properly
|
|
||||||
printf(
|
|
||||||
"MapSMtoCores for SM %d.%d is undefined. Default to use %d Cores/SM\n",
|
|
||||||
major, minor, nGpuArchCoresPerSM[index - 1].Cores);
|
|
||||||
return nGpuArchCoresPerSM[index - 1].Cores;
|
|
||||||
}
|
|
||||||
// end of GPU Architecture definitions
|
|
||||||
|
|
||||||
#ifdef __cuda_cuda_h__
|
|
||||||
// General GPU Device CUDA Initialization
|
|
||||||
inline int gpuDeviceInitDRV(int ARGC, const char **ARGV) {
|
|
||||||
int cuDevice = 0;
|
|
||||||
int deviceCount = 0;
|
|
||||||
checkCudaErrors(cuInit(0, __CUDA_API_VERSION));
|
|
||||||
|
|
||||||
checkCudaErrors(cuDeviceGetCount(&deviceCount));
|
|
||||||
|
|
||||||
if (deviceCount == 0) {
|
|
||||||
fprintf(stderr, "cudaDeviceInit error: no devices supporting CUDA\n");
|
|
||||||
exit(EXIT_FAILURE);
|
|
||||||
}
|
|
||||||
|
|
||||||
int dev = 0;
|
|
||||||
dev = getCmdLineArgumentInt(ARGC, (const char **)ARGV, "device=");
|
|
||||||
|
|
||||||
if (dev < 0) {
|
|
||||||
dev = 0;
|
|
||||||
}
|
|
||||||
|
|
||||||
if (dev > deviceCount - 1) {
|
|
||||||
fprintf(stderr, "\n");
|
|
||||||
fprintf(stderr, ">> %d CUDA capable GPU device(s) detected. <<\n",
|
|
||||||
deviceCount);
|
|
||||||
fprintf(stderr,
|
|
||||||
">> cudaDeviceInit (-device=%d) is not a valid GPU device. <<\n",
|
|
||||||
dev);
|
|
||||||
fprintf(stderr, "\n");
|
|
||||||
return -dev;
|
|
||||||
}
|
|
||||||
|
|
||||||
checkCudaErrors(cuDeviceGet(&cuDevice, dev));
|
|
||||||
char name[100];
|
|
||||||
checkCudaErrors(cuDeviceGetName(name, 100, cuDevice));
|
|
||||||
|
|
||||||
int computeMode;
|
|
||||||
getCudaAttribute<int>(&computeMode, CU_DEVICE_ATTRIBUTE_COMPUTE_MODE, dev);
|
|
||||||
|
|
||||||
if (computeMode == CU_COMPUTEMODE_PROHIBITED) {
|
|
||||||
fprintf(stderr,
|
|
||||||
"Error: device is running in <CU_COMPUTEMODE_PROHIBITED>, no "
|
|
||||||
"threads can use this CUDA Device.\n");
|
|
||||||
return -1;
|
|
||||||
}
|
|
||||||
|
|
||||||
if (checkCmdLineFlag(ARGC, (const char **)ARGV, "quiet") == false) {
|
|
||||||
printf("gpuDeviceInitDRV() Using CUDA Device [%d]: %s\n", dev, name);
|
|
||||||
}
|
|
||||||
|
|
||||||
return dev;
|
|
||||||
}
|
|
||||||
|
|
||||||
// This function returns the best GPU based on performance
|
|
||||||
inline int gpuGetMaxGflopsDeviceIdDRV() {
|
|
||||||
CUdevice current_device = 0;
|
|
||||||
CUdevice max_perf_device = 0;
|
|
||||||
int device_count = 0;
|
|
||||||
int sm_per_multiproc = 0;
|
|
||||||
unsigned long long max_compute_perf = 0;
|
|
||||||
int major = 0;
|
|
||||||
int minor = 0;
|
|
||||||
int multiProcessorCount;
|
|
||||||
int clockRate;
|
|
||||||
int devices_prohibited = 0;
|
|
||||||
|
|
||||||
cuInit(0, __CUDA_API_VERSION);
|
|
||||||
checkCudaErrors(cuDeviceGetCount(&device_count));
|
|
||||||
|
|
||||||
if (device_count == 0) {
|
|
||||||
fprintf(stderr,
|
|
||||||
"gpuGetMaxGflopsDeviceIdDRV error: no devices supporting CUDA\n");
|
|
||||||
exit(EXIT_FAILURE);
|
|
||||||
}
|
|
||||||
|
|
||||||
// Find the best CUDA capable GPU device
|
|
||||||
current_device = 0;
|
|
||||||
|
|
||||||
while (current_device < device_count) {
|
|
||||||
checkCudaErrors(cuDeviceGetAttribute(
|
|
||||||
&multiProcessorCount, CU_DEVICE_ATTRIBUTE_MULTIPROCESSOR_COUNT,
|
|
||||||
current_device));
|
|
||||||
checkCudaErrors(cuDeviceGetAttribute(
|
|
||||||
&clockRate, CU_DEVICE_ATTRIBUTE_CLOCK_RATE, current_device));
|
|
||||||
checkCudaErrors(cuDeviceGetAttribute(
|
|
||||||
&major, CU_DEVICE_ATTRIBUTE_COMPUTE_CAPABILITY_MAJOR, current_device));
|
|
||||||
checkCudaErrors(cuDeviceGetAttribute(
|
|
||||||
&minor, CU_DEVICE_ATTRIBUTE_COMPUTE_CAPABILITY_MINOR, current_device));
|
|
||||||
|
|
||||||
int computeMode;
|
|
||||||
getCudaAttribute<int>(&computeMode, CU_DEVICE_ATTRIBUTE_COMPUTE_MODE,
|
|
||||||
current_device);
|
|
||||||
|
|
||||||
if (computeMode != CU_COMPUTEMODE_PROHIBITED) {
|
|
||||||
if (major == 9999 && minor == 9999) {
|
|
||||||
sm_per_multiproc = 1;
|
|
||||||
} else {
|
|
||||||
sm_per_multiproc = _ConvertSMVer2CoresDRV(major, minor);
|
|
||||||
}
|
|
||||||
|
|
||||||
unsigned long long compute_perf =
|
|
||||||
(unsigned long long)(multiProcessorCount * sm_per_multiproc *
|
|
||||||
clockRate);
|
|
||||||
|
|
||||||
if (compute_perf > max_compute_perf) {
|
|
||||||
max_compute_perf = compute_perf;
|
|
||||||
max_perf_device = current_device;
|
|
||||||
}
|
|
||||||
} else {
|
|
||||||
devices_prohibited++;
|
|
||||||
}
|
|
||||||
|
|
||||||
++current_device;
|
|
||||||
}
|
|
||||||
|
|
||||||
if (devices_prohibited == device_count) {
|
|
||||||
fprintf(stderr,
|
|
||||||
"gpuGetMaxGflopsDeviceIdDRV error: all devices have compute mode "
|
|
||||||
"prohibited.\n");
|
|
||||||
exit(EXIT_FAILURE);
|
|
||||||
}
|
|
||||||
|
|
||||||
return max_perf_device;
|
|
||||||
}
|
|
||||||
|
|
||||||
// General initialization call to pick the best CUDA Device
|
|
||||||
inline CUdevice findCudaDeviceDRV(int argc, const char **argv) {
|
|
||||||
CUdevice cuDevice;
|
|
||||||
int devID = 0;
|
|
||||||
|
|
||||||
// If the command-line has a device number specified, use it
|
|
||||||
if (checkCmdLineFlag(argc, (const char **)argv, "device")) {
|
|
||||||
devID = gpuDeviceInitDRV(argc, argv);
|
|
||||||
|
|
||||||
if (devID < 0) {
|
|
||||||
printf("exiting...\n");
|
|
||||||
exit(EXIT_SUCCESS);
|
|
||||||
}
|
|
||||||
} else {
|
|
||||||
// Otherwise pick the device with highest Gflops/s
|
|
||||||
char name[100];
|
|
||||||
devID = gpuGetMaxGflopsDeviceIdDRV();
|
|
||||||
checkCudaErrors(cuDeviceGet(&cuDevice, devID));
|
|
||||||
cuDeviceGetName(name, 100, cuDevice);
|
|
||||||
printf("> Using CUDA Device [%d]: %s\n", devID, name);
|
|
||||||
}
|
|
||||||
|
|
||||||
cuDeviceGet(&cuDevice, devID);
|
|
||||||
|
|
||||||
return cuDevice;
|
|
||||||
}
|
|
||||||
|
|
||||||
inline CUdevice findIntegratedGPUDrv() {
|
|
||||||
CUdevice current_device = 0;
|
|
||||||
int device_count = 0;
|
|
||||||
int devices_prohibited = 0;
|
|
||||||
int isIntegrated;
|
|
||||||
|
|
||||||
cuInit(0, __CUDA_API_VERSION);
|
|
||||||
checkCudaErrors(cuDeviceGetCount(&device_count));
|
|
||||||
|
|
||||||
if (device_count == 0) {
|
|
||||||
fprintf(stderr, "CUDA error: no devices supporting CUDA.\n");
|
|
||||||
exit(EXIT_FAILURE);
|
|
||||||
}
|
|
||||||
|
|
||||||
// Find the integrated GPU which is compute capable
|
|
||||||
while (current_device < device_count) {
|
|
||||||
int computeMode = -1;
|
|
||||||
checkCudaErrors(cuDeviceGetAttribute(
|
|
||||||
&isIntegrated, CU_DEVICE_ATTRIBUTE_INTEGRATED, current_device));
|
|
||||||
checkCudaErrors(cuDeviceGetAttribute(
|
|
||||||
&computeMode, CU_DEVICE_ATTRIBUTE_COMPUTE_MODE, current_device));
|
|
||||||
|
|
||||||
// If GPU is integrated and is not running on Compute Mode prohibited use
|
|
||||||
// that
|
|
||||||
if (isIntegrated && (computeMode != CU_COMPUTEMODE_PROHIBITED)) {
|
|
||||||
int major = 0, minor = 0;
|
|
||||||
char deviceName[256];
|
|
||||||
checkCudaErrors(cuDeviceGetAttribute(
|
|
||||||
&major, CU_DEVICE_ATTRIBUTE_COMPUTE_CAPABILITY_MAJOR,
|
|
||||||
current_device));
|
|
||||||
checkCudaErrors(cuDeviceGetAttribute(
|
|
||||||
&minor, CU_DEVICE_ATTRIBUTE_COMPUTE_CAPABILITY_MINOR,
|
|
||||||
current_device));
|
|
||||||
checkCudaErrors(cuDeviceGetName(deviceName, 256, current_device));
|
|
||||||
printf("GPU Device %d: \"%s\" with compute capability %d.%d\n\n",
|
|
||||||
current_device, deviceName, major, minor);
|
|
||||||
|
|
||||||
return current_device;
|
|
||||||
} else {
|
|
||||||
devices_prohibited++;
|
|
||||||
}
|
|
||||||
|
|
||||||
current_device++;
|
|
||||||
}
|
|
||||||
|
|
||||||
if (devices_prohibited == device_count) {
|
|
||||||
fprintf(stderr, "CUDA error: No Integrated CUDA capable GPU found.\n");
|
|
||||||
exit(EXIT_FAILURE);
|
|
||||||
}
|
|
||||||
|
|
||||||
return -1;
|
|
||||||
}
|
|
||||||
|
|
||||||
// General check for CUDA GPU SM Capabilities
|
|
||||||
inline bool checkCudaCapabilitiesDRV(int major_version, int minor_version,
|
|
||||||
int devID) {
|
|
||||||
CUdevice cuDevice;
|
|
||||||
char name[256];
|
|
||||||
int major = 0, minor = 0;
|
|
||||||
|
|
||||||
checkCudaErrors(cuDeviceGet(&cuDevice, devID));
|
|
||||||
checkCudaErrors(cuDeviceGetName(name, 100, cuDevice));
|
|
||||||
checkCudaErrors(cuDeviceGetAttribute(
|
|
||||||
&major, CU_DEVICE_ATTRIBUTE_COMPUTE_CAPABILITY_MAJOR, cuDevice));
|
|
||||||
checkCudaErrors(cuDeviceGetAttribute(
|
|
||||||
&minor, CU_DEVICE_ATTRIBUTE_COMPUTE_CAPABILITY_MINOR, cuDevice));
|
|
||||||
|
|
||||||
if ((major > major_version) ||
|
|
||||||
(major == major_version && minor >= minor_version)) {
|
|
||||||
printf("> Device %d: <%16s >, Compute SM %d.%d detected\n", devID, name,
|
|
||||||
major, minor);
|
|
||||||
return true;
|
|
||||||
} else {
|
|
||||||
printf(
|
|
||||||
"No GPU device was found that can support CUDA compute capability "
|
|
||||||
"%d.%d.\n",
|
|
||||||
major_version, minor_version);
|
|
||||||
return false;
|
|
||||||
}
|
|
||||||
}
|
|
||||||
#endif
|
|
||||||
|
|
||||||
// end of CUDA Helper Functions
|
|
||||||
|
|
||||||
#endif // HELPER_CUDA_DRVAPI_H
|
|
||||||
|
|
||||||
@@ -1,41 +0,0 @@
|
|||||||
/* Copyright (c) 2022, NVIDIA CORPORATION. All rights reserved.
|
|
||||||
*
|
|
||||||
* Redistribution and use in source and binary forms, with or without
|
|
||||||
* modification, are permitted provided that the following conditions
|
|
||||||
* are met:
|
|
||||||
* * Redistributions of source code must retain the above copyright
|
|
||||||
* notice, this list of conditions and the following disclaimer.
|
|
||||||
* * Redistributions in binary form must reproduce the above copyright
|
|
||||||
* notice, this list of conditions and the following disclaimer in the
|
|
||||||
* documentation and/or other materials provided with the distribution.
|
|
||||||
* * Neither the name of NVIDIA CORPORATION nor the names of its
|
|
||||||
* contributors may be used to endorse or promote products derived
|
|
||||||
* from this software without specific prior written permission.
|
|
||||||
*
|
|
||||||
* THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS ``AS IS'' AND ANY
|
|
||||||
* EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
|
|
||||||
* IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR
|
|
||||||
* PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER OR
|
|
||||||
* CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL,
|
|
||||||
* EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO,
|
|
||||||
* PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR
|
|
||||||
* PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY
|
|
||||||
* OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
|
|
||||||
* (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
|
|
||||||
* OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
|
|
||||||
*/
|
|
||||||
|
|
||||||
|
|
||||||
#ifndef _MATRIXMUL_H_
|
|
||||||
#define _MATRIXMUL_H_
|
|
||||||
|
|
||||||
// Matrix dimensions
|
|
||||||
// (chosen as multiples of the thread block size for simplicity)
|
|
||||||
#define WA (4 * block_size) // Matrix A width
|
|
||||||
#define HA (6 * block_size) // Matrix A height
|
|
||||||
#define WB (4 * block_size) // Matrix B width
|
|
||||||
#define HB WA // Matrix B height
|
|
||||||
#define WC WB // Matrix C width
|
|
||||||
#define HC HA // Matrix C height
|
|
||||||
|
|
||||||
#endif // _MATRIXMUL_H_
|
|
||||||
File diff suppressed because it is too large
Load Diff
File diff suppressed because it is too large
Load Diff
@@ -1,10 +0,0 @@
|
|||||||
{
|
|
||||||
"configurations": [
|
|
||||||
{
|
|
||||||
"name": "CUDA C++: Launch",
|
|
||||||
"type": "cuda-gdb",
|
|
||||||
"request": "launch",
|
|
||||||
"program": "${workspaceFolder}/matrixMul_nvrtc"
|
|
||||||
}
|
|
||||||
]
|
|
||||||
}
|
|
||||||
@@ -1,15 +0,0 @@
|
|||||||
{
|
|
||||||
"version": "2.0.0",
|
|
||||||
"tasks": [
|
|
||||||
{
|
|
||||||
"label": "sample",
|
|
||||||
"type": "shell",
|
|
||||||
"command": "make dbg=1",
|
|
||||||
"problemMatcher": ["$nvcc"],
|
|
||||||
"group": {
|
|
||||||
"kind": "build",
|
|
||||||
"isDefault": true
|
|
||||||
}
|
|
||||||
}
|
|
||||||
]
|
|
||||||
}
|
|
||||||
@@ -1,52 +0,0 @@
|
|||||||
cmake_minimum_required(VERSION 3.20)
|
|
||||||
|
|
||||||
list(APPEND CMAKE_MODULE_PATH "${CMAKE_CURRENT_SOURCE_DIR}/../../../cmake/Modules")
|
|
||||||
|
|
||||||
project(matrixMul_nvrtc LANGUAGES C CXX CUDA)
|
|
||||||
|
|
||||||
find_package(CUDAToolkit REQUIRED)
|
|
||||||
|
|
||||||
set(CMAKE_POSITION_INDEPENDENT_CODE ON)
|
|
||||||
|
|
||||||
set(CMAKE_CUDA_ARCHITECTURES 50 52 60 61 70 72 75 80 86 87 89 90 100 101 120)
|
|
||||||
set(CMAKE_CUDA_FLAGS "${CMAKE_CUDA_FLAGS} -Wno-deprecated-gpu-targets")
|
|
||||||
if(CMAKE_BUILD_TYPE STREQUAL "Debug")
|
|
||||||
# set(CMAKE_CUDA_FLAGS "${CMAKE_CUDA_FLAGS} -G") # enable cuda-gdb (expensive)
|
|
||||||
endif()
|
|
||||||
|
|
||||||
# Include directories and libraries
|
|
||||||
include_directories(../../../Common)
|
|
||||||
|
|
||||||
# Source file
|
|
||||||
# Add sample target executable
|
|
||||||
add_executable(matrixMul_nvrtc matrixMul.cpp)
|
|
||||||
|
|
||||||
target_compile_options(matrixMul_nvrtc PRIVATE $<$<COMPILE_LANGUAGE:CUDA>:--extended-lambda>)
|
|
||||||
|
|
||||||
target_compile_features(matrixMul_nvrtc PRIVATE cxx_std_17 cuda_std_17)
|
|
||||||
|
|
||||||
target_link_libraries(matrixMul_nvrtc PRIVATE
|
|
||||||
CUDA::nvrtc
|
|
||||||
CUDA::cuda_driver
|
|
||||||
)
|
|
||||||
|
|
||||||
# Copy clock_kernel.cu to the output directory
|
|
||||||
add_custom_command(TARGET matrixMul_nvrtc POST_BUILD
|
|
||||||
COMMAND ${CMAKE_COMMAND} -E copy_if_different
|
|
||||||
${CMAKE_CURRENT_SOURCE_DIR}/matrixMul_kernel.cu ${CUDAToolkit_BIN_DIR}/../include/cooperative_groups.h ${CMAKE_CURRENT_BINARY_DIR}
|
|
||||||
)
|
|
||||||
|
|
||||||
add_custom_command(TARGET matrixMul_nvrtc POST_BUILD
|
|
||||||
COMMAND ${CMAKE_COMMAND} -E copy_directory
|
|
||||||
${CUDAToolkit_BIN_DIR}/../include/cooperative_groups ${CMAKE_CURRENT_BINARY_DIR}/cooperative_groups
|
|
||||||
)
|
|
||||||
|
|
||||||
add_custom_command(TARGET matrixMul_nvrtc POST_BUILD
|
|
||||||
COMMAND ${CMAKE_COMMAND} -E copy_directory
|
|
||||||
${CUDAToolkit_BIN_DIR}/../include/nv ${CMAKE_CURRENT_BINARY_DIR}/nv
|
|
||||||
)
|
|
||||||
|
|
||||||
add_custom_command(TARGET matrixMul_nvrtc POST_BUILD
|
|
||||||
COMMAND ${CMAKE_COMMAND} -E copy_directory
|
|
||||||
${CUDAToolkit_BIN_DIR}/../include/cuda ${CMAKE_CURRENT_BINARY_DIR}/cuda
|
|
||||||
)
|
|
||||||
@@ -1,36 +0,0 @@
|
|||||||
# matrixMul_nvrtc - Matrix Multiplication with libNVRTC
|
|
||||||
|
|
||||||
## Description
|
|
||||||
|
|
||||||
This sample implements matrix multiplication and is exactly the same as Chapter 6 of the programming guide. It has been written for clarity of exposition to illustrate various CUDA programming principles, not with the goal of providing the most performant generic kernel for matrix multiplication. To illustrate GPU performance for matrix multiply, this sample also shows how to use the new CUDA 4.0 interface for CUBLAS to demonstrate high-performance performance for matrix multiplication.
|
|
||||||
|
|
||||||
## Key Concepts
|
|
||||||
|
|
||||||
CUDA Runtime API, Linear Algebra, Runtime Compilation
|
|
||||||
|
|
||||||
## Supported SM Architectures
|
|
||||||
|
|
||||||
[SM 5.0 ](https://developer.nvidia.com/cuda-gpus) [SM 5.2 ](https://developer.nvidia.com/cuda-gpus) [SM 5.3 ](https://developer.nvidia.com/cuda-gpus) [SM 6.0 ](https://developer.nvidia.com/cuda-gpus) [SM 6.1 ](https://developer.nvidia.com/cuda-gpus) [SM 7.0 ](https://developer.nvidia.com/cuda-gpus) [SM 7.2 ](https://developer.nvidia.com/cuda-gpus) [SM 7.5 ](https://developer.nvidia.com/cuda-gpus) [SM 8.0 ](https://developer.nvidia.com/cuda-gpus) [SM 8.6 ](https://developer.nvidia.com/cuda-gpus) [SM 8.7 ](https://developer.nvidia.com/cuda-gpus) [SM 8.9 ](https://developer.nvidia.com/cuda-gpus) [SM 9.0 ](https://developer.nvidia.com/cuda-gpus)
|
|
||||||
|
|
||||||
## Supported OSes
|
|
||||||
|
|
||||||
Linux, Windows, QNX
|
|
||||||
|
|
||||||
## Supported CPU Architecture
|
|
||||||
|
|
||||||
x86_64, aarch64
|
|
||||||
|
|
||||||
## CUDA APIs involved
|
|
||||||
|
|
||||||
### [CUDA Driver API](http://docs.nvidia.com/cuda/cuda-driver-api/index.html)
|
|
||||||
cuMemcpyDtoH, cuLaunchKernel, cuMemcpyHtoD, cuCtxSynchronize, cuMemAlloc, cuMemFree, cuModuleGetFunction
|
|
||||||
|
|
||||||
## Dependencies needed to build/run
|
|
||||||
[NVRTC](../../../README.md#nvrtc)
|
|
||||||
|
|
||||||
## Prerequisites
|
|
||||||
|
|
||||||
Download and install the [CUDA Toolkit 12.5](https://developer.nvidia.com/cuda-downloads) for your corresponding platform.
|
|
||||||
Make sure the dependencies mentioned in [Dependencies]() section above are installed.
|
|
||||||
|
|
||||||
## References (for more details)
|
|
||||||
@@ -1,248 +0,0 @@
|
|||||||
/* Copyright (c) 2022, NVIDIA CORPORATION. All rights reserved.
|
|
||||||
*
|
|
||||||
* Redistribution and use in source and binary forms, with or without
|
|
||||||
* modification, are permitted provided that the following conditions
|
|
||||||
* are met:
|
|
||||||
* * Redistributions of source code must retain the above copyright
|
|
||||||
* notice, this list of conditions and the following disclaimer.
|
|
||||||
* * Redistributions in binary form must reproduce the above copyright
|
|
||||||
* notice, this list of conditions and the following disclaimer in the
|
|
||||||
* documentation and/or other materials provided with the distribution.
|
|
||||||
* * Neither the name of NVIDIA CORPORATION nor the names of its
|
|
||||||
* contributors may be used to endorse or promote products derived
|
|
||||||
* from this software without specific prior written permission.
|
|
||||||
*
|
|
||||||
* THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS ``AS IS'' AND ANY
|
|
||||||
* EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
|
|
||||||
* IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR
|
|
||||||
* PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER OR
|
|
||||||
* CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL,
|
|
||||||
* EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO,
|
|
||||||
* PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR
|
|
||||||
* PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY
|
|
||||||
* OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
|
|
||||||
* (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
|
|
||||||
* OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
|
|
||||||
*/
|
|
||||||
|
|
||||||
/**
|
|
||||||
* Matrix multiplication: C = A * B.
|
|
||||||
* Host code.
|
|
||||||
*
|
|
||||||
* This sample implements matrix multiplication as described in Chapter 3
|
|
||||||
* of the programming guide.
|
|
||||||
* It has been written for clarity of exposition to illustrate various CUDA
|
|
||||||
* programming principles, not with the goal of providing the most
|
|
||||||
* performant generic kernel for matrix multiplication.
|
|
||||||
*
|
|
||||||
* See also:
|
|
||||||
* V. Volkov and J. Demmel, "Benchmarking GPUs to tune dense linear algebra,"
|
|
||||||
* in Proc. 2008 ACM/IEEE Conf. on Supercomputing (SC '08),
|
|
||||||
* Piscataway, NJ: IEEE Press, 2008, pp. Art. 31:1-11.
|
|
||||||
*/
|
|
||||||
|
|
||||||
// System includes
|
|
||||||
#include <stdio.h>
|
|
||||||
#include <assert.h>
|
|
||||||
|
|
||||||
// CUDA runtime
|
|
||||||
#include <cuda_runtime.h>
|
|
||||||
#include "nvrtc_helper.h"
|
|
||||||
|
|
||||||
// Helper functions and utilities to work with CUDA
|
|
||||||
#include <helper_functions.h>
|
|
||||||
|
|
||||||
void constantInit(float *data, int size, float val) {
|
|
||||||
for (int i = 0; i < size; ++i) {
|
|
||||||
data[i] = val;
|
|
||||||
}
|
|
||||||
}
|
|
||||||
|
|
||||||
/**
|
|
||||||
* Run a simple test of matrix multiplication using CUDA
|
|
||||||
*/
|
|
||||||
int matrixMultiply(int argc, char **argv, int block_size, dim3 &dimsA,
|
|
||||||
dim3 &dimsB) {
|
|
||||||
// Allocate host memory for matrices A and B
|
|
||||||
unsigned int size_A = dimsA.x * dimsA.y;
|
|
||||||
unsigned int mem_size_A = sizeof(float) * size_A;
|
|
||||||
float *h_A = (float *)malloc(mem_size_A);
|
|
||||||
unsigned int size_B = dimsB.x * dimsB.y;
|
|
||||||
unsigned int mem_size_B = sizeof(float) * size_B;
|
|
||||||
float *h_B = (float *)malloc(mem_size_B);
|
|
||||||
|
|
||||||
// Initialize host memory
|
|
||||||
const float valB = 0.01f;
|
|
||||||
constantInit(h_A, size_A, 1.0f);
|
|
||||||
constantInit(h_B, size_B, valB);
|
|
||||||
|
|
||||||
// Allocate device memory
|
|
||||||
CUdeviceptr d_A, d_B, d_C;
|
|
||||||
|
|
||||||
char *cubin, *kernel_file;
|
|
||||||
size_t cubinSize;
|
|
||||||
|
|
||||||
kernel_file = sdkFindFilePath("matrixMul_kernel.cu", argv[0]);
|
|
||||||
compileFileToCUBIN(kernel_file, argc, argv, &cubin, &cubinSize, 1);
|
|
||||||
|
|
||||||
CUmodule module = loadCUBIN(cubin, argc, argv);
|
|
||||||
|
|
||||||
// Allocate host matrix C
|
|
||||||
dim3 dimsC(dimsB.x, dimsA.y, 1);
|
|
||||||
unsigned int mem_size_C = dimsC.x * dimsC.y * sizeof(float);
|
|
||||||
float *h_C = (float *)malloc(mem_size_C);
|
|
||||||
|
|
||||||
if (h_C == NULL) {
|
|
||||||
fprintf(stderr, "Failed to allocate host matrix C!\n");
|
|
||||||
exit(EXIT_FAILURE);
|
|
||||||
}
|
|
||||||
|
|
||||||
checkCudaErrors(cuMemAlloc(&d_A, mem_size_A));
|
|
||||||
checkCudaErrors(cuMemAlloc(&d_B, mem_size_B));
|
|
||||||
checkCudaErrors(cuMemAlloc(&d_C, mem_size_C));
|
|
||||||
|
|
||||||
// copy host memory to device
|
|
||||||
checkCudaErrors(cuMemcpyHtoD(d_A, h_A, mem_size_A));
|
|
||||||
checkCudaErrors(cuMemcpyHtoD(d_B, h_B, mem_size_B));
|
|
||||||
|
|
||||||
// Setup execution parameters
|
|
||||||
dim3 threads(block_size, block_size);
|
|
||||||
dim3 grid(dimsB.x / threads.x, dimsA.y / threads.y);
|
|
||||||
|
|
||||||
// Create and start timer
|
|
||||||
printf("Computing result using CUDA Kernel...\n");
|
|
||||||
|
|
||||||
CUfunction kernel_addr;
|
|
||||||
if (block_size == 16) {
|
|
||||||
checkCudaErrors(
|
|
||||||
cuModuleGetFunction(&kernel_addr, module, "matrixMulCUDA_block16"));
|
|
||||||
} else {
|
|
||||||
checkCudaErrors(
|
|
||||||
cuModuleGetFunction(&kernel_addr, module, "matrixMulCUDA_block32"));
|
|
||||||
}
|
|
||||||
|
|
||||||
void *arr[] = {(void *)&d_C, (void *)&d_A, (void *)&d_B, (void *)&dimsA.x,
|
|
||||||
(void *)&dimsB.x};
|
|
||||||
|
|
||||||
// Execute the kernel
|
|
||||||
int nIter = 300;
|
|
||||||
|
|
||||||
for (int j = 0; j < nIter; j++) {
|
|
||||||
checkCudaErrors(
|
|
||||||
cuLaunchKernel(kernel_addr, grid.x, grid.y, grid.z, /* grid dim */
|
|
||||||
threads.x, threads.y, threads.z, /* block dim */
|
|
||||||
0, 0, /* shared mem, stream */
|
|
||||||
&arr[0], /* arguments */
|
|
||||||
0));
|
|
||||||
|
|
||||||
checkCudaErrors(cuCtxSynchronize());
|
|
||||||
}
|
|
||||||
|
|
||||||
// Copy result from device to host
|
|
||||||
checkCudaErrors(cuMemcpyDtoH(h_C, d_C, mem_size_C));
|
|
||||||
|
|
||||||
printf("Checking computed result for correctness: ");
|
|
||||||
|
|
||||||
bool correct = true;
|
|
||||||
|
|
||||||
// test relative error by the formula
|
|
||||||
// |<x, y>_cpu - <x,y>_gpu|/<|x|, |y|> < eps
|
|
||||||
|
|
||||||
double eps = 1.e-6; // machine zero
|
|
||||||
|
|
||||||
for (int i = 0; i < (int)(dimsC.x * dimsC.y); i++) {
|
|
||||||
double abs_err = fabs(h_C[i] - (dimsA.x * valB));
|
|
||||||
double dot_length = dimsA.x;
|
|
||||||
double abs_val = fabs(h_C[i]);
|
|
||||||
double rel_err = abs_err / abs_val / dot_length;
|
|
||||||
|
|
||||||
if (rel_err > eps) {
|
|
||||||
printf("Error! Matrix[%05d]=%.8f, ref=%.8f error term is > %E\n", i,
|
|
||||||
h_C[i], dimsA.x * valB, eps);
|
|
||||||
correct = false;
|
|
||||||
}
|
|
||||||
}
|
|
||||||
|
|
||||||
printf("%s\n", correct ? "Result = PASS" : "Result = FAIL");
|
|
||||||
|
|
||||||
printf(
|
|
||||||
"\nNOTE: The CUDA Samples are not meant for performance measurements. "
|
|
||||||
"Results may vary when GPU Boost is enabled.\n");
|
|
||||||
|
|
||||||
// Clean up memory
|
|
||||||
free(h_A);
|
|
||||||
free(h_B);
|
|
||||||
free(h_C);
|
|
||||||
|
|
||||||
checkCudaErrors(cuMemFree(d_A));
|
|
||||||
checkCudaErrors(cuMemFree(d_B));
|
|
||||||
checkCudaErrors(cuMemFree(d_C));
|
|
||||||
|
|
||||||
if (correct) {
|
|
||||||
return EXIT_SUCCESS;
|
|
||||||
} else {
|
|
||||||
return EXIT_FAILURE;
|
|
||||||
}
|
|
||||||
}
|
|
||||||
|
|
||||||
/**
|
|
||||||
* Program main
|
|
||||||
*/
|
|
||||||
|
|
||||||
int main(int argc, char **argv) {
|
|
||||||
printf("[Matrix Multiply Using CUDA] - Starting...\n");
|
|
||||||
|
|
||||||
if (checkCmdLineFlag(argc, (const char **)argv, "help") ||
|
|
||||||
checkCmdLineFlag(argc, (const char **)argv, "?")) {
|
|
||||||
printf("Usage -device=n (n >= 0 for deviceID)\n");
|
|
||||||
printf(" -wA=WidthA -hA=HeightA (Width x Height of Matrix A)\n");
|
|
||||||
printf(" -wB=WidthB -hB=HeightB (Width x Height of Matrix B)\n");
|
|
||||||
printf(
|
|
||||||
" Note: Outer matrix dimensions of A & B matrices must be equal.\n");
|
|
||||||
|
|
||||||
exit(EXIT_SUCCESS);
|
|
||||||
}
|
|
||||||
|
|
||||||
int block_size = 32;
|
|
||||||
|
|
||||||
// original:
|
|
||||||
dim3 dimsA(5 * 2 * block_size, 5 * 2 * block_size, 1);
|
|
||||||
dim3 dimsB(5 * 4 * block_size, 5 * 2 * block_size, 1);
|
|
||||||
|
|
||||||
// reduce sizes to avoid running out of memory
|
|
||||||
// dim3 dimsA(32,32, 1);
|
|
||||||
// dim3 dimsB(32,32,1);
|
|
||||||
|
|
||||||
// width of Matrix A
|
|
||||||
if (checkCmdLineFlag(argc, (const char **)argv, "wA")) {
|
|
||||||
dimsA.x = getCmdLineArgumentInt(argc, (const char **)argv, "wA");
|
|
||||||
}
|
|
||||||
|
|
||||||
// height of Matrix A
|
|
||||||
if (checkCmdLineFlag(argc, (const char **)argv, "hA")) {
|
|
||||||
dimsA.y = getCmdLineArgumentInt(argc, (const char **)argv, "hA");
|
|
||||||
}
|
|
||||||
|
|
||||||
// width of Matrix B
|
|
||||||
if (checkCmdLineFlag(argc, (const char **)argv, "wB")) {
|
|
||||||
dimsB.x = getCmdLineArgumentInt(argc, (const char **)argv, "wB");
|
|
||||||
}
|
|
||||||
|
|
||||||
// height of Matrix B
|
|
||||||
if (checkCmdLineFlag(argc, (const char **)argv, "hB")) {
|
|
||||||
dimsB.y = getCmdLineArgumentInt(argc, (const char **)argv, "hB");
|
|
||||||
}
|
|
||||||
|
|
||||||
if (dimsA.x != dimsB.y) {
|
|
||||||
printf("Error: outer matrix dimensions must be equal. (%d != %d)\n",
|
|
||||||
dimsA.x, dimsB.y);
|
|
||||||
exit(EXIT_FAILURE);
|
|
||||||
}
|
|
||||||
|
|
||||||
printf("MatrixA(%d,%d), MatrixB(%d,%d)\n", dimsA.x, dimsA.y, dimsB.x,
|
|
||||||
dimsB.y);
|
|
||||||
|
|
||||||
int matrix_result = matrixMultiply(argc, argv, block_size, dimsA, dimsB);
|
|
||||||
|
|
||||||
exit(matrix_result);
|
|
||||||
}
|
|
||||||
@@ -1,131 +0,0 @@
|
|||||||
/* Copyright (c) 2022, NVIDIA CORPORATION. All rights reserved.
|
|
||||||
*
|
|
||||||
* Redistribution and use in source and binary forms, with or without
|
|
||||||
* modification, are permitted provided that the following conditions
|
|
||||||
* are met:
|
|
||||||
* * Redistributions of source code must retain the above copyright
|
|
||||||
* notice, this list of conditions and the following disclaimer.
|
|
||||||
* * Redistributions in binary form must reproduce the above copyright
|
|
||||||
* notice, this list of conditions and the following disclaimer in the
|
|
||||||
* documentation and/or other materials provided with the distribution.
|
|
||||||
* * Neither the name of NVIDIA CORPORATION nor the names of its
|
|
||||||
* contributors may be used to endorse or promote products derived
|
|
||||||
* from this software without specific prior written permission.
|
|
||||||
*
|
|
||||||
* THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS ``AS IS'' AND ANY
|
|
||||||
* EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
|
|
||||||
* IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR
|
|
||||||
* PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER OR
|
|
||||||
* CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL,
|
|
||||||
* EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO,
|
|
||||||
* PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR
|
|
||||||
* PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY
|
|
||||||
* OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
|
|
||||||
* (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
|
|
||||||
* OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
|
|
||||||
*/
|
|
||||||
|
|
||||||
/**
|
|
||||||
* Matrix multiplication: C = A * B.
|
|
||||||
* Host code.
|
|
||||||
*
|
|
||||||
* This sample implements matrix multiplication as described in Chapter 3
|
|
||||||
* of the programming guide.
|
|
||||||
* It has been written for clarity of exposition to illustrate various CUDA
|
|
||||||
* programming principles, not with the goal of providing the most
|
|
||||||
* performant generic kernel for matrix multiplication.
|
|
||||||
*
|
|
||||||
* See also:
|
|
||||||
* V. Volkov and J. Demmel, "Benchmarking GPUs to tune dense linear algebra,"
|
|
||||||
* in Proc. 2008 ACM/IEEE Conf. on Supercomputing (SC '08),
|
|
||||||
* Piscataway, NJ: IEEE Press, 2008, pp. Art. 31:1-11.
|
|
||||||
*/
|
|
||||||
|
|
||||||
/**
|
|
||||||
* Matrix multiplication (CUDA Kernel) on the device: C = A * B
|
|
||||||
* wA is A's width and wB is B's width
|
|
||||||
*/
|
|
||||||
|
|
||||||
#include <cooperative_groups.h>
|
|
||||||
|
|
||||||
template <int BLOCK_SIZE>
|
|
||||||
__device__ void matrixMulCUDA(float *C, float *A, float *B, int wA, int wB) {
|
|
||||||
// Handle to thread block group
|
|
||||||
cooperative_groups::thread_block cta =
|
|
||||||
cooperative_groups::this_thread_block();
|
|
||||||
// Block index
|
|
||||||
int bx = blockIdx.x;
|
|
||||||
int by = blockIdx.y;
|
|
||||||
|
|
||||||
// Thread index
|
|
||||||
int tx = threadIdx.x;
|
|
||||||
int ty = threadIdx.y;
|
|
||||||
|
|
||||||
// Index of the first sub-matrix of A processed by the block
|
|
||||||
int aBegin = wA * BLOCK_SIZE * by;
|
|
||||||
|
|
||||||
// Index of the last sub-matrix of A processed by the block
|
|
||||||
int aEnd = aBegin + wA - 1;
|
|
||||||
|
|
||||||
// Step size used to iterate through the sub-matrices of A
|
|
||||||
int aStep = BLOCK_SIZE;
|
|
||||||
|
|
||||||
// Index of the first sub-matrix of B processed by the block
|
|
||||||
int bBegin = BLOCK_SIZE * bx;
|
|
||||||
|
|
||||||
// Step size used to iterate through the sub-matrices of B
|
|
||||||
int bStep = BLOCK_SIZE * wB;
|
|
||||||
|
|
||||||
// Csub is used to store the element of the block sub-matrix
|
|
||||||
// that is computed by the thread
|
|
||||||
float Csub = 0;
|
|
||||||
|
|
||||||
// Loop over all the sub-matrices of A and B
|
|
||||||
// required to compute the block sub-matrix
|
|
||||||
for (int a = aBegin, b = bBegin; a <= aEnd; a += aStep, b += bStep) {
|
|
||||||
// Declaration of the shared memory array As used to
|
|
||||||
// store the sub-matrix of A
|
|
||||||
__shared__ float As[BLOCK_SIZE][BLOCK_SIZE];
|
|
||||||
|
|
||||||
// Declaration of the shared memory array Bs used to
|
|
||||||
// store the sub-matrix of B
|
|
||||||
__shared__ float Bs[BLOCK_SIZE][BLOCK_SIZE];
|
|
||||||
|
|
||||||
// Load the matrices from device memory
|
|
||||||
// to shared memory; each thread loads
|
|
||||||
// one element of each matrix
|
|
||||||
As[ty][tx] = A[a + wA * ty + tx];
|
|
||||||
Bs[ty][tx] = B[b + wB * ty + tx];
|
|
||||||
|
|
||||||
// Synchronize to make sure the matrices are loaded
|
|
||||||
cooperative_groups::sync(cta);
|
|
||||||
|
|
||||||
// Multiply the two matrices together;
|
|
||||||
// each thread computes one element
|
|
||||||
// of the block sub-matrix
|
|
||||||
#pragma unroll
|
|
||||||
for (int k = 0; k < BLOCK_SIZE; ++k) {
|
|
||||||
Csub += As[ty][k] * Bs[k][tx];
|
|
||||||
}
|
|
||||||
|
|
||||||
// Synchronize to make sure that the preceding
|
|
||||||
// computation is done before loading two new
|
|
||||||
// sub-matrices of A and B in the next iteration
|
|
||||||
cooperative_groups::sync(cta);
|
|
||||||
}
|
|
||||||
|
|
||||||
// Write the block sub-matrix to device memory;
|
|
||||||
// each thread writes one element
|
|
||||||
int c = wB * BLOCK_SIZE * by + BLOCK_SIZE * bx;
|
|
||||||
C[c + wB * ty + tx] = Csub;
|
|
||||||
}
|
|
||||||
|
|
||||||
extern "C" __global__ void matrixMulCUDA_block16(float *C, float *A, float *B,
|
|
||||||
int wA, int wB) {
|
|
||||||
matrixMulCUDA<16>(C, A, B, wA, wB);
|
|
||||||
}
|
|
||||||
|
|
||||||
extern "C" __global__ void matrixMulCUDA_block32(float *C, float *A, float *B,
|
|
||||||
int wA, int wB) {
|
|
||||||
matrixMulCUDA<32>(C, A, B, wA, wB);
|
|
||||||
}
|
|
||||||
@@ -1,10 +0,0 @@
|
|||||||
{
|
|
||||||
"configurations": [
|
|
||||||
{
|
|
||||||
"name": "CUDA C++: Launch",
|
|
||||||
"type": "cuda-gdb",
|
|
||||||
"request": "launch",
|
|
||||||
"program": "${workspaceFolder}/mergeSort"
|
|
||||||
}
|
|
||||||
]
|
|
||||||
}
|
|
||||||
@@ -1,15 +0,0 @@
|
|||||||
{
|
|
||||||
"version": "2.0.0",
|
|
||||||
"tasks": [
|
|
||||||
{
|
|
||||||
"label": "sample",
|
|
||||||
"type": "shell",
|
|
||||||
"command": "make dbg=1",
|
|
||||||
"problemMatcher": ["$nvcc"],
|
|
||||||
"group": {
|
|
||||||
"kind": "build",
|
|
||||||
"isDefault": true
|
|
||||||
}
|
|
||||||
}
|
|
||||||
]
|
|
||||||
}
|
|
||||||
@@ -1,31 +0,0 @@
|
|||||||
cmake_minimum_required(VERSION 3.20)
|
|
||||||
|
|
||||||
list(APPEND CMAKE_MODULE_PATH "${CMAKE_CURRENT_SOURCE_DIR}/../../../cmake/Modules")
|
|
||||||
|
|
||||||
project(mergeSort LANGUAGES C CXX CUDA)
|
|
||||||
|
|
||||||
find_package(CUDAToolkit REQUIRED)
|
|
||||||
|
|
||||||
set(CMAKE_POSITION_INDEPENDENT_CODE ON)
|
|
||||||
|
|
||||||
set(CMAKE_CUDA_ARCHITECTURES 50 52 60 61 70 72 75 80 86 87 89 90 100 101 120)
|
|
||||||
set(CMAKE_CUDA_FLAGS "${CMAKE_CUDA_FLAGS} -Wno-deprecated-gpu-targets")
|
|
||||||
if(CMAKE_BUILD_TYPE STREQUAL "Debug")
|
|
||||||
# set(CMAKE_CUDA_FLAGS "${CMAKE_CUDA_FLAGS} -G") # enable cuda-gdb (expensive)
|
|
||||||
endif()
|
|
||||||
|
|
||||||
# Include directories and libraries
|
|
||||||
include_directories(../../../Common)
|
|
||||||
|
|
||||||
# Source file
|
|
||||||
# Add target for mergeSort
|
|
||||||
add_executable(mergeSort mergeSort.cu main.cpp mergeSort_host.cpp mergeSort_validate.cpp)
|
|
||||||
|
|
||||||
target_compile_options(mergeSort PRIVATE $<$<COMPILE_LANGUAGE:CUDA>:--extended-lambda>)
|
|
||||||
|
|
||||||
target_compile_features(mergeSort PRIVATE cxx_std_17 cuda_std_17)
|
|
||||||
|
|
||||||
set_target_properties(mergeSort PROPERTIES CUDA_SEPARABLE_COMPILATION ON)
|
|
||||||
target_include_directories(mergeSort PRIVATE
|
|
||||||
${CUDAToolkit_INCLUDE_DIRS}
|
|
||||||
)
|
|
||||||
@@ -1,32 +0,0 @@
|
|||||||
# mergeSort - Merge Sort
|
|
||||||
|
|
||||||
## Description
|
|
||||||
|
|
||||||
This sample implements a merge sort (also known as Batcher's sort), algorithms belonging to the class of sorting networks. While generally subefficient on large sequences compared to algorithms with better asymptotic algorithmic complexity (i.e. merge sort or radix sort), may be the algorithms of choice for sorting batches of short- to mid-sized (key, value) array pairs. Refer to the excellent tutorial by H. W. Lang http://www.iti.fh-flensburg.de/lang/algorithmen/sortieren/networks/indexen.htm
|
|
||||||
|
|
||||||
## Key Concepts
|
|
||||||
|
|
||||||
Data-Parallel Algorithms
|
|
||||||
|
|
||||||
## Supported SM Architectures
|
|
||||||
|
|
||||||
[SM 5.0 ](https://developer.nvidia.com/cuda-gpus) [SM 5.2 ](https://developer.nvidia.com/cuda-gpus) [SM 5.3 ](https://developer.nvidia.com/cuda-gpus) [SM 6.0 ](https://developer.nvidia.com/cuda-gpus) [SM 6.1 ](https://developer.nvidia.com/cuda-gpus) [SM 7.0 ](https://developer.nvidia.com/cuda-gpus) [SM 7.2 ](https://developer.nvidia.com/cuda-gpus) [SM 7.5 ](https://developer.nvidia.com/cuda-gpus) [SM 8.0 ](https://developer.nvidia.com/cuda-gpus) [SM 8.6 ](https://developer.nvidia.com/cuda-gpus) [SM 8.7 ](https://developer.nvidia.com/cuda-gpus) [SM 8.9 ](https://developer.nvidia.com/cuda-gpus) [SM 9.0 ](https://developer.nvidia.com/cuda-gpus)
|
|
||||||
|
|
||||||
## Supported OSes
|
|
||||||
|
|
||||||
Linux, Windows
|
|
||||||
|
|
||||||
## Supported CPU Architecture
|
|
||||||
|
|
||||||
x86_64, armv7l
|
|
||||||
|
|
||||||
## CUDA APIs involved
|
|
||||||
|
|
||||||
### [CUDA Runtime API](http://docs.nvidia.com/cuda/cuda-runtime-api/index.html)
|
|
||||||
cudaMalloc, cudaDeviceSynchronize, cudaMemcpy, cudaFree
|
|
||||||
|
|
||||||
## Prerequisites
|
|
||||||
|
|
||||||
Download and install the [CUDA Toolkit 12.5](https://developer.nvidia.com/cuda-downloads) for your corresponding platform.
|
|
||||||
|
|
||||||
## References (for more details)
|
|
||||||
@@ -1,279 +0,0 @@
|
|||||||
/* Copyright (c) 2022, NVIDIA CORPORATION. All rights reserved.
|
|
||||||
*
|
|
||||||
* Redistribution and use in source and binary forms, with or without
|
|
||||||
* modification, are permitted provided that the following conditions
|
|
||||||
* are met:
|
|
||||||
* * Redistributions of source code must retain the above copyright
|
|
||||||
* notice, this list of conditions and the following disclaimer.
|
|
||||||
* * Redistributions in binary form must reproduce the above copyright
|
|
||||||
* notice, this list of conditions and the following disclaimer in the
|
|
||||||
* documentation and/or other materials provided with the distribution.
|
|
||||||
* * Neither the name of NVIDIA CORPORATION nor the names of its
|
|
||||||
* contributors may be used to endorse or promote products derived
|
|
||||||
* from this software without specific prior written permission.
|
|
||||||
*
|
|
||||||
* THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS ``AS IS'' AND ANY
|
|
||||||
* EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
|
|
||||||
* IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR
|
|
||||||
* PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER OR
|
|
||||||
* CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL,
|
|
||||||
* EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO,
|
|
||||||
* PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR
|
|
||||||
* PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY
|
|
||||||
* OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
|
|
||||||
* (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
|
|
||||||
* OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
|
|
||||||
*/
|
|
||||||
|
|
||||||
#include <cooperative_groups.h>
|
|
||||||
|
|
||||||
namespace cg = cooperative_groups;
|
|
||||||
#include <helper_cuda.h>
|
|
||||||
#include <assert.h>
|
|
||||||
#include "mergeSort_common.h"
|
|
||||||
|
|
||||||
inline __device__ void Comparator(uint &keyA, uint &valA, uint &keyB,
|
|
||||||
uint &valB, uint arrowDir) {
|
|
||||||
uint t;
|
|
||||||
|
|
||||||
if ((keyA > keyB) == arrowDir) {
|
|
||||||
t = keyA;
|
|
||||||
keyA = keyB;
|
|
||||||
keyB = t;
|
|
||||||
t = valA;
|
|
||||||
valA = valB;
|
|
||||||
valB = t;
|
|
||||||
}
|
|
||||||
}
|
|
||||||
|
|
||||||
__global__ void bitonicSortSharedKernel(uint *d_DstKey, uint *d_DstVal,
|
|
||||||
uint *d_SrcKey, uint *d_SrcVal,
|
|
||||||
uint arrayLength, uint sortDir) {
|
|
||||||
// Handle to thread block group
|
|
||||||
cg::thread_block cta = cg::this_thread_block();
|
|
||||||
// Shared memory storage for one or more short vectors
|
|
||||||
__shared__ uint s_key[SHARED_SIZE_LIMIT];
|
|
||||||
__shared__ uint s_val[SHARED_SIZE_LIMIT];
|
|
||||||
|
|
||||||
// Offset to the beginning of subbatch and load data
|
|
||||||
d_SrcKey += blockIdx.x * SHARED_SIZE_LIMIT + threadIdx.x;
|
|
||||||
d_SrcVal += blockIdx.x * SHARED_SIZE_LIMIT + threadIdx.x;
|
|
||||||
d_DstKey += blockIdx.x * SHARED_SIZE_LIMIT + threadIdx.x;
|
|
||||||
d_DstVal += blockIdx.x * SHARED_SIZE_LIMIT + threadIdx.x;
|
|
||||||
s_key[threadIdx.x + 0] = d_SrcKey[0];
|
|
||||||
s_val[threadIdx.x + 0] = d_SrcVal[0];
|
|
||||||
s_key[threadIdx.x + (SHARED_SIZE_LIMIT / 2)] =
|
|
||||||
d_SrcKey[(SHARED_SIZE_LIMIT / 2)];
|
|
||||||
s_val[threadIdx.x + (SHARED_SIZE_LIMIT / 2)] =
|
|
||||||
d_SrcVal[(SHARED_SIZE_LIMIT / 2)];
|
|
||||||
|
|
||||||
for (uint size = 2; size < arrayLength; size <<= 1) {
|
|
||||||
// Bitonic merge
|
|
||||||
uint dir = (threadIdx.x & (size / 2)) != 0;
|
|
||||||
|
|
||||||
for (uint stride = size / 2; stride > 0; stride >>= 1) {
|
|
||||||
cg::sync(cta);
|
|
||||||
uint pos = 2 * threadIdx.x - (threadIdx.x & (stride - 1));
|
|
||||||
Comparator(s_key[pos + 0], s_val[pos + 0], s_key[pos + stride],
|
|
||||||
s_val[pos + stride], dir);
|
|
||||||
}
|
|
||||||
}
|
|
||||||
|
|
||||||
// ddd == sortDir for the last bitonic merge step
|
|
||||||
{
|
|
||||||
for (uint stride = arrayLength / 2; stride > 0; stride >>= 1) {
|
|
||||||
cg::sync(cta);
|
|
||||||
uint pos = 2 * threadIdx.x - (threadIdx.x & (stride - 1));
|
|
||||||
Comparator(s_key[pos + 0], s_val[pos + 0], s_key[pos + stride],
|
|
||||||
s_val[pos + stride], sortDir);
|
|
||||||
}
|
|
||||||
}
|
|
||||||
|
|
||||||
cg::sync(cta);
|
|
||||||
d_DstKey[0] = s_key[threadIdx.x + 0];
|
|
||||||
d_DstVal[0] = s_val[threadIdx.x + 0];
|
|
||||||
d_DstKey[(SHARED_SIZE_LIMIT / 2)] =
|
|
||||||
s_key[threadIdx.x + (SHARED_SIZE_LIMIT / 2)];
|
|
||||||
d_DstVal[(SHARED_SIZE_LIMIT / 2)] =
|
|
||||||
s_val[threadIdx.x + (SHARED_SIZE_LIMIT / 2)];
|
|
||||||
}
|
|
||||||
|
|
||||||
// Helper function (also used by odd-even merge sort)
|
|
||||||
extern "C" uint factorRadix2(uint *log2L, uint L) {
|
|
||||||
if (!L) {
|
|
||||||
*log2L = 0;
|
|
||||||
return 0;
|
|
||||||
} else {
|
|
||||||
for (*log2L = 0; (L & 1) == 0; L >>= 1, *log2L++)
|
|
||||||
;
|
|
||||||
|
|
||||||
return L;
|
|
||||||
}
|
|
||||||
}
|
|
||||||
|
|
||||||
extern "C" void bitonicSortShared(uint *d_DstKey, uint *d_DstVal,
|
|
||||||
uint *d_SrcKey, uint *d_SrcVal,
|
|
||||||
uint batchSize, uint arrayLength,
|
|
||||||
uint sortDir) {
|
|
||||||
// Nothing to sort
|
|
||||||
if (arrayLength < 2) {
|
|
||||||
return;
|
|
||||||
}
|
|
||||||
|
|
||||||
// Only power-of-two array lengths are supported by this implementation
|
|
||||||
uint log2L;
|
|
||||||
uint factorizationRemainder = factorRadix2(&log2L, arrayLength);
|
|
||||||
assert(factorizationRemainder == 1);
|
|
||||||
|
|
||||||
uint blockCount = batchSize * arrayLength / SHARED_SIZE_LIMIT;
|
|
||||||
uint threadCount = SHARED_SIZE_LIMIT / 2;
|
|
||||||
|
|
||||||
assert(arrayLength <= SHARED_SIZE_LIMIT);
|
|
||||||
assert((batchSize * arrayLength) % SHARED_SIZE_LIMIT == 0);
|
|
||||||
|
|
||||||
bitonicSortSharedKernel<<<blockCount, threadCount>>>(
|
|
||||||
d_DstKey, d_DstVal, d_SrcKey, d_SrcVal, arrayLength, sortDir);
|
|
||||||
getLastCudaError("bitonicSortSharedKernel<<<>>> failed!\n");
|
|
||||||
}
|
|
||||||
|
|
||||||
////////////////////////////////////////////////////////////////////////////////
|
|
||||||
// Merge step 3: merge elementary intervals
|
|
||||||
////////////////////////////////////////////////////////////////////////////////
|
|
||||||
static inline __host__ __device__ uint iDivUp(uint a, uint b) {
|
|
||||||
return ((a % b) == 0) ? (a / b) : (a / b + 1);
|
|
||||||
}
|
|
||||||
|
|
||||||
static inline __host__ __device__ uint getSampleCount(uint dividend) {
|
|
||||||
return iDivUp(dividend, SAMPLE_STRIDE);
|
|
||||||
}
|
|
||||||
|
|
||||||
template <uint sortDir>
|
|
||||||
static inline __device__ void ComparatorExtended(uint &keyA, uint &valA,
|
|
||||||
uint &flagA, uint &keyB,
|
|
||||||
uint &valB, uint &flagB,
|
|
||||||
uint arrowDir) {
|
|
||||||
uint t;
|
|
||||||
|
|
||||||
if ((!(flagA || flagB) && ((keyA > keyB) == arrowDir)) ||
|
|
||||||
((arrowDir == sortDir) && (flagA == 1)) ||
|
|
||||||
((arrowDir != sortDir) && (flagB == 1))) {
|
|
||||||
t = keyA;
|
|
||||||
keyA = keyB;
|
|
||||||
keyB = t;
|
|
||||||
t = valA;
|
|
||||||
valA = valB;
|
|
||||||
valB = t;
|
|
||||||
t = flagA;
|
|
||||||
flagA = flagB;
|
|
||||||
flagB = t;
|
|
||||||
}
|
|
||||||
}
|
|
||||||
|
|
||||||
template <uint sortDir>
|
|
||||||
__global__ void bitonicMergeElementaryIntervalsKernel(
|
|
||||||
uint *d_DstKey, uint *d_DstVal, uint *d_SrcKey, uint *d_SrcVal,
|
|
||||||
uint *d_LimitsA, uint *d_LimitsB, uint stride, uint N) {
|
|
||||||
// Handle to thread block group
|
|
||||||
cg::thread_block cta = cg::this_thread_block();
|
|
||||||
__shared__ uint s_key[2 * SAMPLE_STRIDE];
|
|
||||||
__shared__ uint s_val[2 * SAMPLE_STRIDE];
|
|
||||||
__shared__ uint s_inf[2 * SAMPLE_STRIDE];
|
|
||||||
|
|
||||||
const uint intervalI = blockIdx.x & ((2 * stride) / SAMPLE_STRIDE - 1);
|
|
||||||
const uint segmentBase = (blockIdx.x - intervalI) * SAMPLE_STRIDE;
|
|
||||||
d_SrcKey += segmentBase;
|
|
||||||
d_SrcVal += segmentBase;
|
|
||||||
d_DstKey += segmentBase;
|
|
||||||
d_DstVal += segmentBase;
|
|
||||||
|
|
||||||
// Set up threadblock-wide parameters
|
|
||||||
__shared__ uint startSrcA, lenSrcA, startSrcB, lenSrcB, startDst;
|
|
||||||
|
|
||||||
if (threadIdx.x == 0) {
|
|
||||||
uint segmentElementsA = stride;
|
|
||||||
uint segmentElementsB = umin(stride, N - segmentBase - stride);
|
|
||||||
uint segmentSamplesA = stride / SAMPLE_STRIDE;
|
|
||||||
uint segmentSamplesB = getSampleCount(segmentElementsB);
|
|
||||||
uint segmentSamples = segmentSamplesA + segmentSamplesB;
|
|
||||||
|
|
||||||
startSrcA = d_LimitsA[blockIdx.x];
|
|
||||||
startSrcB = d_LimitsB[blockIdx.x];
|
|
||||||
startDst = startSrcA + startSrcB;
|
|
||||||
|
|
||||||
uint endSrcA = (intervalI + 1 < segmentSamples) ? d_LimitsA[blockIdx.x + 1]
|
|
||||||
: segmentElementsA;
|
|
||||||
uint endSrcB = (intervalI + 1 < segmentSamples) ? d_LimitsB[blockIdx.x + 1]
|
|
||||||
: segmentElementsB;
|
|
||||||
lenSrcA = endSrcA - startSrcA;
|
|
||||||
lenSrcB = endSrcB - startSrcB;
|
|
||||||
}
|
|
||||||
|
|
||||||
s_inf[threadIdx.x + 0] = 1;
|
|
||||||
s_inf[threadIdx.x + SAMPLE_STRIDE] = 1;
|
|
||||||
|
|
||||||
// Load input data
|
|
||||||
cg::sync(cta);
|
|
||||||
|
|
||||||
if (threadIdx.x < lenSrcA) {
|
|
||||||
s_key[threadIdx.x] = d_SrcKey[0 + startSrcA + threadIdx.x];
|
|
||||||
s_val[threadIdx.x] = d_SrcVal[0 + startSrcA + threadIdx.x];
|
|
||||||
s_inf[threadIdx.x] = 0;
|
|
||||||
}
|
|
||||||
|
|
||||||
// Prepare for bitonic merge by inversing the ordering
|
|
||||||
if (threadIdx.x < lenSrcB) {
|
|
||||||
s_key[2 * SAMPLE_STRIDE - 1 - threadIdx.x] =
|
|
||||||
d_SrcKey[stride + startSrcB + threadIdx.x];
|
|
||||||
s_val[2 * SAMPLE_STRIDE - 1 - threadIdx.x] =
|
|
||||||
d_SrcVal[stride + startSrcB + threadIdx.x];
|
|
||||||
s_inf[2 * SAMPLE_STRIDE - 1 - threadIdx.x] = 0;
|
|
||||||
}
|
|
||||||
|
|
||||||
//"Extended" bitonic merge
|
|
||||||
for (uint stride = SAMPLE_STRIDE; stride > 0; stride >>= 1) {
|
|
||||||
cg::sync(cta);
|
|
||||||
uint pos = 2 * threadIdx.x - (threadIdx.x & (stride - 1));
|
|
||||||
ComparatorExtended<sortDir>(s_key[pos + 0], s_val[pos + 0], s_inf[pos + 0],
|
|
||||||
s_key[pos + stride], s_val[pos + stride],
|
|
||||||
s_inf[pos + stride], sortDir);
|
|
||||||
}
|
|
||||||
|
|
||||||
// Store sorted data
|
|
||||||
cg::sync(cta);
|
|
||||||
d_DstKey += startDst;
|
|
||||||
d_DstVal += startDst;
|
|
||||||
|
|
||||||
if (threadIdx.x < lenSrcA) {
|
|
||||||
d_DstKey[threadIdx.x] = s_key[threadIdx.x];
|
|
||||||
d_DstVal[threadIdx.x] = s_val[threadIdx.x];
|
|
||||||
}
|
|
||||||
|
|
||||||
if (threadIdx.x < lenSrcB) {
|
|
||||||
d_DstKey[lenSrcA + threadIdx.x] = s_key[lenSrcA + threadIdx.x];
|
|
||||||
d_DstVal[lenSrcA + threadIdx.x] = s_val[lenSrcA + threadIdx.x];
|
|
||||||
}
|
|
||||||
}
|
|
||||||
|
|
||||||
extern "C" void bitonicMergeElementaryIntervals(uint *d_DstKey, uint *d_DstVal,
|
|
||||||
uint *d_SrcKey, uint *d_SrcVal,
|
|
||||||
uint *d_LimitsA,
|
|
||||||
uint *d_LimitsB, uint stride,
|
|
||||||
uint N, uint sortDir) {
|
|
||||||
uint lastSegmentElements = N % (2 * stride);
|
|
||||||
|
|
||||||
uint mergePairs = (lastSegmentElements > stride)
|
|
||||||
? getSampleCount(N)
|
|
||||||
: (N - lastSegmentElements) / SAMPLE_STRIDE;
|
|
||||||
|
|
||||||
if (sortDir) {
|
|
||||||
bitonicMergeElementaryIntervalsKernel<1U><<<mergePairs, SAMPLE_STRIDE>>>(
|
|
||||||
d_DstKey, d_DstVal, d_SrcKey, d_SrcVal, d_LimitsA, d_LimitsB, stride,
|
|
||||||
N);
|
|
||||||
getLastCudaError("mergeElementaryIntervalsKernel<1> failed\n");
|
|
||||||
} else {
|
|
||||||
bitonicMergeElementaryIntervalsKernel<0U><<<mergePairs, SAMPLE_STRIDE>>>(
|
|
||||||
d_DstKey, d_DstVal, d_SrcKey, d_SrcVal, d_LimitsA, d_LimitsB, stride,
|
|
||||||
N);
|
|
||||||
getLastCudaError("mergeElementaryIntervalsKernel<0> failed\n");
|
|
||||||
}
|
|
||||||
}
|
|
||||||
@@ -1,121 +0,0 @@
|
|||||||
/* Copyright (c) 2022, NVIDIA CORPORATION. All rights reserved.
|
|
||||||
*
|
|
||||||
* Redistribution and use in source and binary forms, with or without
|
|
||||||
* modification, are permitted provided that the following conditions
|
|
||||||
* are met:
|
|
||||||
* * Redistributions of source code must retain the above copyright
|
|
||||||
* notice, this list of conditions and the following disclaimer.
|
|
||||||
* * Redistributions in binary form must reproduce the above copyright
|
|
||||||
* notice, this list of conditions and the following disclaimer in the
|
|
||||||
* documentation and/or other materials provided with the distribution.
|
|
||||||
* * Neither the name of NVIDIA CORPORATION nor the names of its
|
|
||||||
* contributors may be used to endorse or promote products derived
|
|
||||||
* from this software without specific prior written permission.
|
|
||||||
*
|
|
||||||
* THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS ``AS IS'' AND ANY
|
|
||||||
* EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
|
|
||||||
* IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR
|
|
||||||
* PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER OR
|
|
||||||
* CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL,
|
|
||||||
* EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO,
|
|
||||||
* PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR
|
|
||||||
* PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY
|
|
||||||
* OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
|
|
||||||
* (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
|
|
||||||
* OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
|
|
||||||
*/
|
|
||||||
|
|
||||||
#include <assert.h>
|
|
||||||
#include <stdio.h>
|
|
||||||
#include <stdlib.h>
|
|
||||||
#include <cuda_runtime.h>
|
|
||||||
#include <helper_functions.h>
|
|
||||||
#include <helper_cuda.h>
|
|
||||||
#include "mergeSort_common.h"
|
|
||||||
|
|
||||||
////////////////////////////////////////////////////////////////////////////////
|
|
||||||
// Test driver
|
|
||||||
////////////////////////////////////////////////////////////////////////////////
|
|
||||||
int main(int argc, char **argv) {
|
|
||||||
uint *h_SrcKey, *h_SrcVal, *h_DstKey, *h_DstVal;
|
|
||||||
uint *d_SrcKey, *d_SrcVal, *d_BufKey, *d_BufVal, *d_DstKey, *d_DstVal;
|
|
||||||
StopWatchInterface *hTimer = NULL;
|
|
||||||
|
|
||||||
const uint N = 4 * 1048576;
|
|
||||||
const uint DIR = 1;
|
|
||||||
const uint numValues = 65536;
|
|
||||||
|
|
||||||
printf("%s Starting...\n\n", argv[0]);
|
|
||||||
|
|
||||||
int dev = findCudaDevice(argc, (const char **)argv);
|
|
||||||
|
|
||||||
if (dev == -1) {
|
|
||||||
return EXIT_FAILURE;
|
|
||||||
}
|
|
||||||
|
|
||||||
printf("Allocating and initializing host arrays...\n\n");
|
|
||||||
sdkCreateTimer(&hTimer);
|
|
||||||
h_SrcKey = (uint *)malloc(N * sizeof(uint));
|
|
||||||
h_SrcVal = (uint *)malloc(N * sizeof(uint));
|
|
||||||
h_DstKey = (uint *)malloc(N * sizeof(uint));
|
|
||||||
h_DstVal = (uint *)malloc(N * sizeof(uint));
|
|
||||||
|
|
||||||
srand(2009);
|
|
||||||
|
|
||||||
for (uint i = 0; i < N; i++) {
|
|
||||||
h_SrcKey[i] = rand() % numValues;
|
|
||||||
}
|
|
||||||
|
|
||||||
fillValues(h_SrcVal, N);
|
|
||||||
|
|
||||||
printf("Allocating and initializing CUDA arrays...\n\n");
|
|
||||||
checkCudaErrors(cudaMalloc((void **)&d_DstKey, N * sizeof(uint)));
|
|
||||||
checkCudaErrors(cudaMalloc((void **)&d_DstVal, N * sizeof(uint)));
|
|
||||||
checkCudaErrors(cudaMalloc((void **)&d_BufKey, N * sizeof(uint)));
|
|
||||||
checkCudaErrors(cudaMalloc((void **)&d_BufVal, N * sizeof(uint)));
|
|
||||||
checkCudaErrors(cudaMalloc((void **)&d_SrcKey, N * sizeof(uint)));
|
|
||||||
checkCudaErrors(cudaMalloc((void **)&d_SrcVal, N * sizeof(uint)));
|
|
||||||
checkCudaErrors(
|
|
||||||
cudaMemcpy(d_SrcKey, h_SrcKey, N * sizeof(uint), cudaMemcpyHostToDevice));
|
|
||||||
checkCudaErrors(
|
|
||||||
cudaMemcpy(d_SrcVal, h_SrcVal, N * sizeof(uint), cudaMemcpyHostToDevice));
|
|
||||||
|
|
||||||
printf("Initializing GPU merge sort...\n");
|
|
||||||
initMergeSort();
|
|
||||||
|
|
||||||
printf("Running GPU merge sort...\n");
|
|
||||||
checkCudaErrors(cudaDeviceSynchronize());
|
|
||||||
sdkResetTimer(&hTimer);
|
|
||||||
sdkStartTimer(&hTimer);
|
|
||||||
mergeSort(d_DstKey, d_DstVal, d_BufKey, d_BufVal, d_SrcKey, d_SrcVal, N, DIR);
|
|
||||||
checkCudaErrors(cudaDeviceSynchronize());
|
|
||||||
sdkStopTimer(&hTimer);
|
|
||||||
printf("Time: %f ms\n", sdkGetTimerValue(&hTimer));
|
|
||||||
|
|
||||||
printf("Reading back GPU merge sort results...\n");
|
|
||||||
checkCudaErrors(
|
|
||||||
cudaMemcpy(h_DstKey, d_DstKey, N * sizeof(uint), cudaMemcpyDeviceToHost));
|
|
||||||
checkCudaErrors(
|
|
||||||
cudaMemcpy(h_DstVal, d_DstVal, N * sizeof(uint), cudaMemcpyDeviceToHost));
|
|
||||||
|
|
||||||
printf("Inspecting the results...\n");
|
|
||||||
uint keysFlag = validateSortedKeys(h_DstKey, h_SrcKey, 1, N, numValues, DIR);
|
|
||||||
|
|
||||||
uint valuesFlag = validateSortedValues(h_DstKey, h_DstVal, h_SrcKey, 1, N);
|
|
||||||
|
|
||||||
printf("Shutting down...\n");
|
|
||||||
closeMergeSort();
|
|
||||||
sdkDeleteTimer(&hTimer);
|
|
||||||
checkCudaErrors(cudaFree(d_SrcVal));
|
|
||||||
checkCudaErrors(cudaFree(d_SrcKey));
|
|
||||||
checkCudaErrors(cudaFree(d_BufVal));
|
|
||||||
checkCudaErrors(cudaFree(d_BufKey));
|
|
||||||
checkCudaErrors(cudaFree(d_DstVal));
|
|
||||||
checkCudaErrors(cudaFree(d_DstKey));
|
|
||||||
free(h_DstVal);
|
|
||||||
free(h_DstKey);
|
|
||||||
free(h_SrcVal);
|
|
||||||
free(h_SrcKey);
|
|
||||||
|
|
||||||
exit((keysFlag && valuesFlag) ? EXIT_SUCCESS : EXIT_FAILURE);
|
|
||||||
}
|
|
||||||
@@ -1,529 +0,0 @@
|
|||||||
/* Copyright (c) 2022, NVIDIA CORPORATION. All rights reserved.
|
|
||||||
*
|
|
||||||
* Redistribution and use in source and binary forms, with or without
|
|
||||||
* modification, are permitted provided that the following conditions
|
|
||||||
* are met:
|
|
||||||
* * Redistributions of source code must retain the above copyright
|
|
||||||
* notice, this list of conditions and the following disclaimer.
|
|
||||||
* * Redistributions in binary form must reproduce the above copyright
|
|
||||||
* notice, this list of conditions and the following disclaimer in the
|
|
||||||
* documentation and/or other materials provided with the distribution.
|
|
||||||
* * Neither the name of NVIDIA CORPORATION nor the names of its
|
|
||||||
* contributors may be used to endorse or promote products derived
|
|
||||||
* from this software without specific prior written permission.
|
|
||||||
*
|
|
||||||
* THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS ``AS IS'' AND ANY
|
|
||||||
* EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
|
|
||||||
* IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR
|
|
||||||
* PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER OR
|
|
||||||
* CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL,
|
|
||||||
* EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO,
|
|
||||||
* PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR
|
|
||||||
* PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY
|
|
||||||
* OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
|
|
||||||
* (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
|
|
||||||
* OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
|
|
||||||
*/
|
|
||||||
|
|
||||||
/*
|
|
||||||
* Based on "Designing efficient sorting algorithms for manycore GPUs"
|
|
||||||
* by Nadathur Satish, Mark Harris, and Michael Garland
|
|
||||||
* http://mgarland.org/files/papers/gpusort-ipdps09.pdf
|
|
||||||
*
|
|
||||||
* Victor Podlozhnyuk 09/24/2009
|
|
||||||
*/
|
|
||||||
|
|
||||||
#include <assert.h>
|
|
||||||
#include <cooperative_groups.h>
|
|
||||||
|
|
||||||
namespace cg = cooperative_groups;
|
|
||||||
|
|
||||||
#include <helper_cuda.h>
|
|
||||||
#include "mergeSort_common.h"
|
|
||||||
|
|
||||||
////////////////////////////////////////////////////////////////////////////////
|
|
||||||
// Helper functions
|
|
||||||
////////////////////////////////////////////////////////////////////////////////
|
|
||||||
static inline __host__ __device__ uint iDivUp(uint a, uint b) {
|
|
||||||
return ((a % b) == 0) ? (a / b) : (a / b + 1);
|
|
||||||
}
|
|
||||||
|
|
||||||
static inline __host__ __device__ uint getSampleCount(uint dividend) {
|
|
||||||
return iDivUp(dividend, SAMPLE_STRIDE);
|
|
||||||
}
|
|
||||||
|
|
||||||
#define W (sizeof(uint) * 8)
|
|
||||||
static inline __device__ uint nextPowerOfTwo(uint x) {
|
|
||||||
/*
|
|
||||||
--x;
|
|
||||||
x |= x >> 1;
|
|
||||||
x |= x >> 2;
|
|
||||||
x |= x >> 4;
|
|
||||||
x |= x >> 8;
|
|
||||||
x |= x >> 16;
|
|
||||||
return ++x;
|
|
||||||
*/
|
|
||||||
return 1U << (W - __clz(x - 1));
|
|
||||||
}
|
|
||||||
|
|
||||||
template <uint sortDir>
|
|
||||||
static inline __device__ uint binarySearchInclusive(uint val, uint *data,
|
|
||||||
uint L, uint stride) {
|
|
||||||
if (L == 0) {
|
|
||||||
return 0;
|
|
||||||
}
|
|
||||||
|
|
||||||
uint pos = 0;
|
|
||||||
|
|
||||||
for (; stride > 0; stride >>= 1) {
|
|
||||||
uint newPos = umin(pos + stride, L);
|
|
||||||
|
|
||||||
if ((sortDir && (data[newPos - 1] <= val)) ||
|
|
||||||
(!sortDir && (data[newPos - 1] >= val))) {
|
|
||||||
pos = newPos;
|
|
||||||
}
|
|
||||||
}
|
|
||||||
|
|
||||||
return pos;
|
|
||||||
}
|
|
||||||
|
|
||||||
template <uint sortDir>
|
|
||||||
static inline __device__ uint binarySearchExclusive(uint val, uint *data,
|
|
||||||
uint L, uint stride) {
|
|
||||||
if (L == 0) {
|
|
||||||
return 0;
|
|
||||||
}
|
|
||||||
|
|
||||||
uint pos = 0;
|
|
||||||
|
|
||||||
for (; stride > 0; stride >>= 1) {
|
|
||||||
uint newPos = umin(pos + stride, L);
|
|
||||||
|
|
||||||
if ((sortDir && (data[newPos - 1] < val)) ||
|
|
||||||
(!sortDir && (data[newPos - 1] > val))) {
|
|
||||||
pos = newPos;
|
|
||||||
}
|
|
||||||
}
|
|
||||||
|
|
||||||
return pos;
|
|
||||||
}
|
|
||||||
|
|
||||||
////////////////////////////////////////////////////////////////////////////////
|
|
||||||
// Bottom-level merge sort (binary search-based)
|
|
||||||
////////////////////////////////////////////////////////////////////////////////
|
|
||||||
template <uint sortDir>
|
|
||||||
__global__ void mergeSortSharedKernel(uint *d_DstKey, uint *d_DstVal,
|
|
||||||
uint *d_SrcKey, uint *d_SrcVal,
|
|
||||||
uint arrayLength) {
|
|
||||||
// Handle to thread block group
|
|
||||||
cg::thread_block cta = cg::this_thread_block();
|
|
||||||
__shared__ uint s_key[SHARED_SIZE_LIMIT];
|
|
||||||
__shared__ uint s_val[SHARED_SIZE_LIMIT];
|
|
||||||
|
|
||||||
d_SrcKey += blockIdx.x * SHARED_SIZE_LIMIT + threadIdx.x;
|
|
||||||
d_SrcVal += blockIdx.x * SHARED_SIZE_LIMIT + threadIdx.x;
|
|
||||||
d_DstKey += blockIdx.x * SHARED_SIZE_LIMIT + threadIdx.x;
|
|
||||||
d_DstVal += blockIdx.x * SHARED_SIZE_LIMIT + threadIdx.x;
|
|
||||||
s_key[threadIdx.x + 0] = d_SrcKey[0];
|
|
||||||
s_val[threadIdx.x + 0] = d_SrcVal[0];
|
|
||||||
s_key[threadIdx.x + (SHARED_SIZE_LIMIT / 2)] =
|
|
||||||
d_SrcKey[(SHARED_SIZE_LIMIT / 2)];
|
|
||||||
s_val[threadIdx.x + (SHARED_SIZE_LIMIT / 2)] =
|
|
||||||
d_SrcVal[(SHARED_SIZE_LIMIT / 2)];
|
|
||||||
|
|
||||||
for (uint stride = 1; stride < arrayLength; stride <<= 1) {
|
|
||||||
uint lPos = threadIdx.x & (stride - 1);
|
|
||||||
uint *baseKey = s_key + 2 * (threadIdx.x - lPos);
|
|
||||||
uint *baseVal = s_val + 2 * (threadIdx.x - lPos);
|
|
||||||
|
|
||||||
cg::sync(cta);
|
|
||||||
uint keyA = baseKey[lPos + 0];
|
|
||||||
uint valA = baseVal[lPos + 0];
|
|
||||||
uint keyB = baseKey[lPos + stride];
|
|
||||||
uint valB = baseVal[lPos + stride];
|
|
||||||
uint posA =
|
|
||||||
binarySearchExclusive<sortDir>(keyA, baseKey + stride, stride, stride) +
|
|
||||||
lPos;
|
|
||||||
uint posB =
|
|
||||||
binarySearchInclusive<sortDir>(keyB, baseKey + 0, stride, stride) +
|
|
||||||
lPos;
|
|
||||||
|
|
||||||
cg::sync(cta);
|
|
||||||
baseKey[posA] = keyA;
|
|
||||||
baseVal[posA] = valA;
|
|
||||||
baseKey[posB] = keyB;
|
|
||||||
baseVal[posB] = valB;
|
|
||||||
}
|
|
||||||
|
|
||||||
cg::sync(cta);
|
|
||||||
d_DstKey[0] = s_key[threadIdx.x + 0];
|
|
||||||
d_DstVal[0] = s_val[threadIdx.x + 0];
|
|
||||||
d_DstKey[(SHARED_SIZE_LIMIT / 2)] =
|
|
||||||
s_key[threadIdx.x + (SHARED_SIZE_LIMIT / 2)];
|
|
||||||
d_DstVal[(SHARED_SIZE_LIMIT / 2)] =
|
|
||||||
s_val[threadIdx.x + (SHARED_SIZE_LIMIT / 2)];
|
|
||||||
}
|
|
||||||
|
|
||||||
static void mergeSortShared(uint *d_DstKey, uint *d_DstVal, uint *d_SrcKey,
|
|
||||||
uint *d_SrcVal, uint batchSize, uint arrayLength,
|
|
||||||
uint sortDir) {
|
|
||||||
if (arrayLength < 2) {
|
|
||||||
return;
|
|
||||||
}
|
|
||||||
|
|
||||||
assert(SHARED_SIZE_LIMIT % arrayLength == 0);
|
|
||||||
assert(((batchSize * arrayLength) % SHARED_SIZE_LIMIT) == 0);
|
|
||||||
uint blockCount = batchSize * arrayLength / SHARED_SIZE_LIMIT;
|
|
||||||
uint threadCount = SHARED_SIZE_LIMIT / 2;
|
|
||||||
|
|
||||||
if (sortDir) {
|
|
||||||
mergeSortSharedKernel<1U><<<blockCount, threadCount>>>(
|
|
||||||
d_DstKey, d_DstVal, d_SrcKey, d_SrcVal, arrayLength);
|
|
||||||
getLastCudaError("mergeSortShared<1><<<>>> failed\n");
|
|
||||||
} else {
|
|
||||||
mergeSortSharedKernel<0U><<<blockCount, threadCount>>>(
|
|
||||||
d_DstKey, d_DstVal, d_SrcKey, d_SrcVal, arrayLength);
|
|
||||||
getLastCudaError("mergeSortShared<0><<<>>> failed\n");
|
|
||||||
}
|
|
||||||
}
|
|
||||||
|
|
||||||
////////////////////////////////////////////////////////////////////////////////
|
|
||||||
// Merge step 1: generate sample ranks
|
|
||||||
////////////////////////////////////////////////////////////////////////////////
|
|
||||||
template <uint sortDir>
|
|
||||||
__global__ void generateSampleRanksKernel(uint *d_RanksA, uint *d_RanksB,
|
|
||||||
uint *d_SrcKey, uint stride, uint N,
|
|
||||||
uint threadCount) {
|
|
||||||
uint pos = blockIdx.x * blockDim.x + threadIdx.x;
|
|
||||||
|
|
||||||
if (pos >= threadCount) {
|
|
||||||
return;
|
|
||||||
}
|
|
||||||
|
|
||||||
const uint i = pos & ((stride / SAMPLE_STRIDE) - 1);
|
|
||||||
const uint segmentBase = (pos - i) * (2 * SAMPLE_STRIDE);
|
|
||||||
d_SrcKey += segmentBase;
|
|
||||||
d_RanksA += segmentBase / SAMPLE_STRIDE;
|
|
||||||
d_RanksB += segmentBase / SAMPLE_STRIDE;
|
|
||||||
|
|
||||||
const uint segmentElementsA = stride;
|
|
||||||
const uint segmentElementsB = umin(stride, N - segmentBase - stride);
|
|
||||||
const uint segmentSamplesA = getSampleCount(segmentElementsA);
|
|
||||||
const uint segmentSamplesB = getSampleCount(segmentElementsB);
|
|
||||||
|
|
||||||
if (i < segmentSamplesA) {
|
|
||||||
d_RanksA[i] = i * SAMPLE_STRIDE;
|
|
||||||
d_RanksB[i] = binarySearchExclusive<sortDir>(
|
|
||||||
d_SrcKey[i * SAMPLE_STRIDE], d_SrcKey + stride, segmentElementsB,
|
|
||||||
nextPowerOfTwo(segmentElementsB));
|
|
||||||
}
|
|
||||||
|
|
||||||
if (i < segmentSamplesB) {
|
|
||||||
d_RanksB[(stride / SAMPLE_STRIDE) + i] = i * SAMPLE_STRIDE;
|
|
||||||
d_RanksA[(stride / SAMPLE_STRIDE) + i] = binarySearchInclusive<sortDir>(
|
|
||||||
d_SrcKey[stride + i * SAMPLE_STRIDE], d_SrcKey + 0, segmentElementsA,
|
|
||||||
nextPowerOfTwo(segmentElementsA));
|
|
||||||
}
|
|
||||||
}
|
|
||||||
|
|
||||||
static void generateSampleRanks(uint *d_RanksA, uint *d_RanksB, uint *d_SrcKey,
|
|
||||||
uint stride, uint N, uint sortDir) {
|
|
||||||
uint lastSegmentElements = N % (2 * stride);
|
|
||||||
uint threadCount =
|
|
||||||
(lastSegmentElements > stride)
|
|
||||||
? (N + 2 * stride - lastSegmentElements) / (2 * SAMPLE_STRIDE)
|
|
||||||
: (N - lastSegmentElements) / (2 * SAMPLE_STRIDE);
|
|
||||||
|
|
||||||
if (sortDir) {
|
|
||||||
generateSampleRanksKernel<1U><<<iDivUp(threadCount, 256), 256>>>(
|
|
||||||
d_RanksA, d_RanksB, d_SrcKey, stride, N, threadCount);
|
|
||||||
getLastCudaError("generateSampleRanksKernel<1U><<<>>> failed\n");
|
|
||||||
} else {
|
|
||||||
generateSampleRanksKernel<0U><<<iDivUp(threadCount, 256), 256>>>(
|
|
||||||
d_RanksA, d_RanksB, d_SrcKey, stride, N, threadCount);
|
|
||||||
getLastCudaError("generateSampleRanksKernel<0U><<<>>> failed\n");
|
|
||||||
}
|
|
||||||
}
|
|
||||||
|
|
||||||
////////////////////////////////////////////////////////////////////////////////
|
|
||||||
// Merge step 2: generate sample ranks and indices
|
|
||||||
////////////////////////////////////////////////////////////////////////////////
|
|
||||||
__global__ void mergeRanksAndIndicesKernel(uint *d_Limits, uint *d_Ranks,
|
|
||||||
uint stride, uint N,
|
|
||||||
uint threadCount) {
|
|
||||||
uint pos = blockIdx.x * blockDim.x + threadIdx.x;
|
|
||||||
|
|
||||||
if (pos >= threadCount) {
|
|
||||||
return;
|
|
||||||
}
|
|
||||||
|
|
||||||
const uint i = pos & ((stride / SAMPLE_STRIDE) - 1);
|
|
||||||
const uint segmentBase = (pos - i) * (2 * SAMPLE_STRIDE);
|
|
||||||
d_Ranks += (pos - i) * 2;
|
|
||||||
d_Limits += (pos - i) * 2;
|
|
||||||
|
|
||||||
const uint segmentElementsA = stride;
|
|
||||||
const uint segmentElementsB = umin(stride, N - segmentBase - stride);
|
|
||||||
const uint segmentSamplesA = getSampleCount(segmentElementsA);
|
|
||||||
const uint segmentSamplesB = getSampleCount(segmentElementsB);
|
|
||||||
|
|
||||||
if (i < segmentSamplesA) {
|
|
||||||
uint dstPos = binarySearchExclusive<1U>(
|
|
||||||
d_Ranks[i], d_Ranks + segmentSamplesA, segmentSamplesB,
|
|
||||||
nextPowerOfTwo(segmentSamplesB)) +
|
|
||||||
i;
|
|
||||||
d_Limits[dstPos] = d_Ranks[i];
|
|
||||||
}
|
|
||||||
|
|
||||||
if (i < segmentSamplesB) {
|
|
||||||
uint dstPos = binarySearchInclusive<1U>(d_Ranks[segmentSamplesA + i],
|
|
||||||
d_Ranks, segmentSamplesA,
|
|
||||||
nextPowerOfTwo(segmentSamplesA)) +
|
|
||||||
i;
|
|
||||||
d_Limits[dstPos] = d_Ranks[segmentSamplesA + i];
|
|
||||||
}
|
|
||||||
}
|
|
||||||
|
|
||||||
static void mergeRanksAndIndices(uint *d_LimitsA, uint *d_LimitsB,
|
|
||||||
uint *d_RanksA, uint *d_RanksB, uint stride,
|
|
||||||
uint N) {
|
|
||||||
uint lastSegmentElements = N % (2 * stride);
|
|
||||||
uint threadCount =
|
|
||||||
(lastSegmentElements > stride)
|
|
||||||
? (N + 2 * stride - lastSegmentElements) / (2 * SAMPLE_STRIDE)
|
|
||||||
: (N - lastSegmentElements) / (2 * SAMPLE_STRIDE);
|
|
||||||
|
|
||||||
mergeRanksAndIndicesKernel<<<iDivUp(threadCount, 256), 256>>>(
|
|
||||||
d_LimitsA, d_RanksA, stride, N, threadCount);
|
|
||||||
getLastCudaError("mergeRanksAndIndicesKernel(A)<<<>>> failed\n");
|
|
||||||
|
|
||||||
mergeRanksAndIndicesKernel<<<iDivUp(threadCount, 256), 256>>>(
|
|
||||||
d_LimitsB, d_RanksB, stride, N, threadCount);
|
|
||||||
getLastCudaError("mergeRanksAndIndicesKernel(B)<<<>>> failed\n");
|
|
||||||
}
|
|
||||||
|
|
||||||
////////////////////////////////////////////////////////////////////////////////
|
|
||||||
// Merge step 3: merge elementary intervals
|
|
||||||
////////////////////////////////////////////////////////////////////////////////
|
|
||||||
template <uint sortDir>
|
|
||||||
inline __device__ void merge(uint *dstKey, uint *dstVal, uint *srcAKey,
|
|
||||||
uint *srcAVal, uint *srcBKey, uint *srcBVal,
|
|
||||||
uint lenA, uint nPowTwoLenA, uint lenB,
|
|
||||||
uint nPowTwoLenB, cg::thread_block cta) {
|
|
||||||
uint keyA, valA, keyB, valB, dstPosA, dstPosB;
|
|
||||||
|
|
||||||
if (threadIdx.x < lenA) {
|
|
||||||
keyA = srcAKey[threadIdx.x];
|
|
||||||
valA = srcAVal[threadIdx.x];
|
|
||||||
dstPosA = binarySearchExclusive<sortDir>(keyA, srcBKey, lenB, nPowTwoLenB) +
|
|
||||||
threadIdx.x;
|
|
||||||
}
|
|
||||||
|
|
||||||
if (threadIdx.x < lenB) {
|
|
||||||
keyB = srcBKey[threadIdx.x];
|
|
||||||
valB = srcBVal[threadIdx.x];
|
|
||||||
dstPosB = binarySearchInclusive<sortDir>(keyB, srcAKey, lenA, nPowTwoLenA) +
|
|
||||||
threadIdx.x;
|
|
||||||
}
|
|
||||||
|
|
||||||
cg::sync(cta);
|
|
||||||
|
|
||||||
if (threadIdx.x < lenA) {
|
|
||||||
dstKey[dstPosA] = keyA;
|
|
||||||
dstVal[dstPosA] = valA;
|
|
||||||
}
|
|
||||||
|
|
||||||
if (threadIdx.x < lenB) {
|
|
||||||
dstKey[dstPosB] = keyB;
|
|
||||||
dstVal[dstPosB] = valB;
|
|
||||||
}
|
|
||||||
}
|
|
||||||
|
|
||||||
template <uint sortDir>
|
|
||||||
__global__ void mergeElementaryIntervalsKernel(uint *d_DstKey, uint *d_DstVal,
|
|
||||||
uint *d_SrcKey, uint *d_SrcVal,
|
|
||||||
uint *d_LimitsA, uint *d_LimitsB,
|
|
||||||
uint stride, uint N) {
|
|
||||||
// Handle to thread block group
|
|
||||||
cg::thread_block cta = cg::this_thread_block();
|
|
||||||
__shared__ uint s_key[2 * SAMPLE_STRIDE];
|
|
||||||
__shared__ uint s_val[2 * SAMPLE_STRIDE];
|
|
||||||
|
|
||||||
const uint intervalI = blockIdx.x & ((2 * stride) / SAMPLE_STRIDE - 1);
|
|
||||||
const uint segmentBase = (blockIdx.x - intervalI) * SAMPLE_STRIDE;
|
|
||||||
d_SrcKey += segmentBase;
|
|
||||||
d_SrcVal += segmentBase;
|
|
||||||
d_DstKey += segmentBase;
|
|
||||||
d_DstVal += segmentBase;
|
|
||||||
|
|
||||||
// Set up threadblock-wide parameters
|
|
||||||
__shared__ uint startSrcA, startSrcB, lenSrcA, lenSrcB, startDstA, startDstB;
|
|
||||||
|
|
||||||
if (threadIdx.x == 0) {
|
|
||||||
uint segmentElementsA = stride;
|
|
||||||
uint segmentElementsB = umin(stride, N - segmentBase - stride);
|
|
||||||
uint segmentSamplesA = getSampleCount(segmentElementsA);
|
|
||||||
uint segmentSamplesB = getSampleCount(segmentElementsB);
|
|
||||||
uint segmentSamples = segmentSamplesA + segmentSamplesB;
|
|
||||||
|
|
||||||
startSrcA = d_LimitsA[blockIdx.x];
|
|
||||||
startSrcB = d_LimitsB[blockIdx.x];
|
|
||||||
uint endSrcA = (intervalI + 1 < segmentSamples) ? d_LimitsA[blockIdx.x + 1]
|
|
||||||
: segmentElementsA;
|
|
||||||
uint endSrcB = (intervalI + 1 < segmentSamples) ? d_LimitsB[blockIdx.x + 1]
|
|
||||||
: segmentElementsB;
|
|
||||||
lenSrcA = endSrcA - startSrcA;
|
|
||||||
lenSrcB = endSrcB - startSrcB;
|
|
||||||
startDstA = startSrcA + startSrcB;
|
|
||||||
startDstB = startDstA + lenSrcA;
|
|
||||||
}
|
|
||||||
|
|
||||||
// Load main input data
|
|
||||||
cg::sync(cta);
|
|
||||||
|
|
||||||
if (threadIdx.x < lenSrcA) {
|
|
||||||
s_key[threadIdx.x + 0] = d_SrcKey[0 + startSrcA + threadIdx.x];
|
|
||||||
s_val[threadIdx.x + 0] = d_SrcVal[0 + startSrcA + threadIdx.x];
|
|
||||||
}
|
|
||||||
|
|
||||||
if (threadIdx.x < lenSrcB) {
|
|
||||||
s_key[threadIdx.x + SAMPLE_STRIDE] =
|
|
||||||
d_SrcKey[stride + startSrcB + threadIdx.x];
|
|
||||||
s_val[threadIdx.x + SAMPLE_STRIDE] =
|
|
||||||
d_SrcVal[stride + startSrcB + threadIdx.x];
|
|
||||||
}
|
|
||||||
|
|
||||||
// Merge data in shared memory
|
|
||||||
cg::sync(cta);
|
|
||||||
merge<sortDir>(s_key, s_val, s_key + 0, s_val + 0, s_key + SAMPLE_STRIDE,
|
|
||||||
s_val + SAMPLE_STRIDE, lenSrcA, SAMPLE_STRIDE, lenSrcB,
|
|
||||||
SAMPLE_STRIDE, cta);
|
|
||||||
|
|
||||||
// Store merged data
|
|
||||||
cg::sync(cta);
|
|
||||||
|
|
||||||
if (threadIdx.x < lenSrcA) {
|
|
||||||
d_DstKey[startDstA + threadIdx.x] = s_key[threadIdx.x];
|
|
||||||
d_DstVal[startDstA + threadIdx.x] = s_val[threadIdx.x];
|
|
||||||
}
|
|
||||||
|
|
||||||
if (threadIdx.x < lenSrcB) {
|
|
||||||
d_DstKey[startDstB + threadIdx.x] = s_key[lenSrcA + threadIdx.x];
|
|
||||||
d_DstVal[startDstB + threadIdx.x] = s_val[lenSrcA + threadIdx.x];
|
|
||||||
}
|
|
||||||
}
|
|
||||||
|
|
||||||
static void mergeElementaryIntervals(uint *d_DstKey, uint *d_DstVal,
|
|
||||||
uint *d_SrcKey, uint *d_SrcVal,
|
|
||||||
uint *d_LimitsA, uint *d_LimitsB,
|
|
||||||
uint stride, uint N, uint sortDir) {
|
|
||||||
uint lastSegmentElements = N % (2 * stride);
|
|
||||||
uint mergePairs = (lastSegmentElements > stride)
|
|
||||||
? getSampleCount(N)
|
|
||||||
: (N - lastSegmentElements) / SAMPLE_STRIDE;
|
|
||||||
|
|
||||||
if (sortDir) {
|
|
||||||
mergeElementaryIntervalsKernel<1U><<<mergePairs, SAMPLE_STRIDE>>>(
|
|
||||||
d_DstKey, d_DstVal, d_SrcKey, d_SrcVal, d_LimitsA, d_LimitsB, stride,
|
|
||||||
N);
|
|
||||||
getLastCudaError("mergeElementaryIntervalsKernel<1> failed\n");
|
|
||||||
} else {
|
|
||||||
mergeElementaryIntervalsKernel<0U><<<mergePairs, SAMPLE_STRIDE>>>(
|
|
||||||
d_DstKey, d_DstVal, d_SrcKey, d_SrcVal, d_LimitsA, d_LimitsB, stride,
|
|
||||||
N);
|
|
||||||
getLastCudaError("mergeElementaryIntervalsKernel<0> failed\n");
|
|
||||||
}
|
|
||||||
}
|
|
||||||
|
|
||||||
extern "C" void bitonicSortShared(uint *d_DstKey, uint *d_DstVal,
|
|
||||||
uint *d_SrcKey, uint *d_SrcVal,
|
|
||||||
uint batchSize, uint arrayLength,
|
|
||||||
uint sortDir);
|
|
||||||
|
|
||||||
extern "C" void bitonicMergeElementaryIntervals(uint *d_DstKey, uint *d_DstVal,
|
|
||||||
uint *d_SrcKey, uint *d_SrcVal,
|
|
||||||
uint *d_LimitsA,
|
|
||||||
uint *d_LimitsB, uint stride,
|
|
||||||
uint N, uint sortDir);
|
|
||||||
|
|
||||||
static uint *d_RanksA, *d_RanksB, *d_LimitsA, *d_LimitsB;
|
|
||||||
static const uint MAX_SAMPLE_COUNT = 32768;
|
|
||||||
|
|
||||||
extern "C" void initMergeSort(void) {
|
|
||||||
checkCudaErrors(
|
|
||||||
cudaMalloc((void **)&d_RanksA, MAX_SAMPLE_COUNT * sizeof(uint)));
|
|
||||||
checkCudaErrors(
|
|
||||||
cudaMalloc((void **)&d_RanksB, MAX_SAMPLE_COUNT * sizeof(uint)));
|
|
||||||
checkCudaErrors(
|
|
||||||
cudaMalloc((void **)&d_LimitsA, MAX_SAMPLE_COUNT * sizeof(uint)));
|
|
||||||
checkCudaErrors(
|
|
||||||
cudaMalloc((void **)&d_LimitsB, MAX_SAMPLE_COUNT * sizeof(uint)));
|
|
||||||
}
|
|
||||||
|
|
||||||
extern "C" void closeMergeSort(void) {
|
|
||||||
checkCudaErrors(cudaFree(d_RanksA));
|
|
||||||
checkCudaErrors(cudaFree(d_RanksB));
|
|
||||||
checkCudaErrors(cudaFree(d_LimitsB));
|
|
||||||
checkCudaErrors(cudaFree(d_LimitsA));
|
|
||||||
}
|
|
||||||
|
|
||||||
extern "C" void mergeSort(uint *d_DstKey, uint *d_DstVal, uint *d_BufKey,
|
|
||||||
uint *d_BufVal, uint *d_SrcKey, uint *d_SrcVal,
|
|
||||||
uint N, uint sortDir) {
|
|
||||||
uint stageCount = 0;
|
|
||||||
|
|
||||||
for (uint stride = SHARED_SIZE_LIMIT; stride < N; stride <<= 1, stageCount++)
|
|
||||||
;
|
|
||||||
|
|
||||||
uint *ikey, *ival, *okey, *oval;
|
|
||||||
|
|
||||||
if (stageCount & 1) {
|
|
||||||
ikey = d_BufKey;
|
|
||||||
ival = d_BufVal;
|
|
||||||
okey = d_DstKey;
|
|
||||||
oval = d_DstVal;
|
|
||||||
} else {
|
|
||||||
ikey = d_DstKey;
|
|
||||||
ival = d_DstVal;
|
|
||||||
okey = d_BufKey;
|
|
||||||
oval = d_BufVal;
|
|
||||||
}
|
|
||||||
|
|
||||||
assert(N <= (SAMPLE_STRIDE * MAX_SAMPLE_COUNT));
|
|
||||||
assert(N % SHARED_SIZE_LIMIT == 0);
|
|
||||||
mergeSortShared(ikey, ival, d_SrcKey, d_SrcVal, N / SHARED_SIZE_LIMIT,
|
|
||||||
SHARED_SIZE_LIMIT, sortDir);
|
|
||||||
|
|
||||||
for (uint stride = SHARED_SIZE_LIMIT; stride < N; stride <<= 1) {
|
|
||||||
uint lastSegmentElements = N % (2 * stride);
|
|
||||||
|
|
||||||
// Find sample ranks and prepare for limiters merge
|
|
||||||
generateSampleRanks(d_RanksA, d_RanksB, ikey, stride, N, sortDir);
|
|
||||||
|
|
||||||
// Merge ranks and indices
|
|
||||||
mergeRanksAndIndices(d_LimitsA, d_LimitsB, d_RanksA, d_RanksB, stride, N);
|
|
||||||
|
|
||||||
// Merge elementary intervals
|
|
||||||
mergeElementaryIntervals(okey, oval, ikey, ival, d_LimitsA, d_LimitsB,
|
|
||||||
stride, N, sortDir);
|
|
||||||
|
|
||||||
if (lastSegmentElements <= stride) {
|
|
||||||
// Last merge segment consists of a single array which just needs to be
|
|
||||||
// passed through
|
|
||||||
checkCudaErrors(cudaMemcpy(
|
|
||||||
okey + (N - lastSegmentElements), ikey + (N - lastSegmentElements),
|
|
||||||
lastSegmentElements * sizeof(uint), cudaMemcpyDeviceToDevice));
|
|
||||||
checkCudaErrors(cudaMemcpy(
|
|
||||||
oval + (N - lastSegmentElements), ival + (N - lastSegmentElements),
|
|
||||||
lastSegmentElements * sizeof(uint), cudaMemcpyDeviceToDevice));
|
|
||||||
}
|
|
||||||
|
|
||||||
uint *t;
|
|
||||||
t = ikey;
|
|
||||||
ikey = okey;
|
|
||||||
okey = t;
|
|
||||||
t = ival;
|
|
||||||
ival = oval;
|
|
||||||
oval = t;
|
|
||||||
}
|
|
||||||
}
|
|
||||||
@@ -1,357 +0,0 @@
|
|||||||
/* Copyright (c) 2022, NVIDIA CORPORATION. All rights reserved.
|
|
||||||
*
|
|
||||||
* Redistribution and use in source and binary forms, with or without
|
|
||||||
* modification, are permitted provided that the following conditions
|
|
||||||
* are met:
|
|
||||||
* * Redistributions of source code must retain the above copyright
|
|
||||||
* notice, this list of conditions and the following disclaimer.
|
|
||||||
* * Redistributions in binary form must reproduce the above copyright
|
|
||||||
* notice, this list of conditions and the following disclaimer in the
|
|
||||||
* documentation and/or other materials provided with the distribution.
|
|
||||||
* * Neither the name of NVIDIA CORPORATION nor the names of its
|
|
||||||
* contributors may be used to endorse or promote products derived
|
|
||||||
* from this software without specific prior written permission.
|
|
||||||
*
|
|
||||||
* THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS ``AS IS'' AND ANY
|
|
||||||
* EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
|
|
||||||
* IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR
|
|
||||||
* PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER OR
|
|
||||||
* CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL,
|
|
||||||
* EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO,
|
|
||||||
* PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR
|
|
||||||
* PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY
|
|
||||||
* OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
|
|
||||||
* (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
|
|
||||||
* OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
|
|
||||||
*/
|
|
||||||
|
|
||||||
#include <assert.h>
|
|
||||||
#include <stdio.h>
|
|
||||||
#include <stdlib.h>
|
|
||||||
#include <string.h>
|
|
||||||
#include "mergeSort_common.h"
|
|
||||||
|
|
||||||
////////////////////////////////////////////////////////////////////////////////
|
|
||||||
// Helper functions
|
|
||||||
////////////////////////////////////////////////////////////////////////////////
|
|
||||||
static void checkOrder(uint *data, uint N, uint sortDir) {
|
|
||||||
if (N <= 1) {
|
|
||||||
return;
|
|
||||||
}
|
|
||||||
|
|
||||||
for (uint i = 0; i < N - 1; i++)
|
|
||||||
if ((sortDir && (data[i] > data[i + 1])) ||
|
|
||||||
(!sortDir && (data[i] < data[i + 1]))) {
|
|
||||||
fprintf(stderr, "checkOrder() failed!!!\n");
|
|
||||||
exit(EXIT_FAILURE);
|
|
||||||
}
|
|
||||||
}
|
|
||||||
|
|
||||||
static uint umin(uint a, uint b) { return (a <= b) ? a : b; }
|
|
||||||
|
|
||||||
static uint getSampleCount(uint dividend) {
|
|
||||||
return ((dividend % SAMPLE_STRIDE) != 0) ? (dividend / SAMPLE_STRIDE + 1)
|
|
||||||
: (dividend / SAMPLE_STRIDE);
|
|
||||||
}
|
|
||||||
|
|
||||||
static uint nextPowerOfTwo(uint x) {
|
|
||||||
--x;
|
|
||||||
x |= x >> 1;
|
|
||||||
x |= x >> 2;
|
|
||||||
x |= x >> 4;
|
|
||||||
x |= x >> 8;
|
|
||||||
x |= x >> 16;
|
|
||||||
return ++x;
|
|
||||||
}
|
|
||||||
|
|
||||||
static uint binarySearchInclusive(uint val, uint *data, uint L, uint sortDir) {
|
|
||||||
if (L == 0) {
|
|
||||||
return 0;
|
|
||||||
}
|
|
||||||
|
|
||||||
uint pos = 0;
|
|
||||||
|
|
||||||
for (uint stride = nextPowerOfTwo(L); stride > 0; stride >>= 1) {
|
|
||||||
uint newPos = umin(pos + stride, L);
|
|
||||||
|
|
||||||
if ((sortDir && (data[newPos - 1] <= val)) ||
|
|
||||||
(!sortDir && (data[newPos - 1] >= val))) {
|
|
||||||
pos = newPos;
|
|
||||||
}
|
|
||||||
}
|
|
||||||
|
|
||||||
return pos;
|
|
||||||
}
|
|
||||||
|
|
||||||
static uint binarySearchExclusive(uint val, uint *data, uint L, uint sortDir) {
|
|
||||||
if (L == 0) {
|
|
||||||
return 0;
|
|
||||||
}
|
|
||||||
|
|
||||||
uint pos = 0;
|
|
||||||
|
|
||||||
for (uint stride = nextPowerOfTwo(L); stride > 0; stride >>= 1) {
|
|
||||||
uint newPos = umin(pos + stride, L);
|
|
||||||
|
|
||||||
if ((sortDir && (data[newPos - 1] < val)) ||
|
|
||||||
(!sortDir && (data[newPos - 1] > val))) {
|
|
||||||
pos = newPos;
|
|
||||||
}
|
|
||||||
}
|
|
||||||
|
|
||||||
return pos;
|
|
||||||
}
|
|
||||||
|
|
||||||
////////////////////////////////////////////////////////////////////////////////
|
|
||||||
// Merge step 1: find sample ranks in each segment
|
|
||||||
////////////////////////////////////////////////////////////////////////////////
|
|
||||||
static void generateSampleRanks(uint *ranksA, uint *ranksB, uint *srcKey,
|
|
||||||
uint stride, uint N, uint sortDir) {
|
|
||||||
uint lastSegmentElements = N % (2 * stride);
|
|
||||||
uint sampleCount =
|
|
||||||
(lastSegmentElements > stride)
|
|
||||||
? (N + 2 * stride - lastSegmentElements) / (2 * SAMPLE_STRIDE)
|
|
||||||
: (N - lastSegmentElements) / (2 * SAMPLE_STRIDE);
|
|
||||||
|
|
||||||
for (uint pos = 0; pos < sampleCount; pos++) {
|
|
||||||
const uint i = pos & ((stride / SAMPLE_STRIDE) - 1);
|
|
||||||
const uint segmentBase = (pos - i) * (2 * SAMPLE_STRIDE);
|
|
||||||
|
|
||||||
const uint lenA = stride;
|
|
||||||
const uint lenB = umin(stride, N - segmentBase - stride);
|
|
||||||
const uint nA = stride / SAMPLE_STRIDE;
|
|
||||||
const uint nB = getSampleCount(lenB);
|
|
||||||
|
|
||||||
if (i < nA) {
|
|
||||||
ranksA[(segmentBase + 0) / SAMPLE_STRIDE + i] = i * SAMPLE_STRIDE;
|
|
||||||
ranksB[(segmentBase + 0) / SAMPLE_STRIDE + i] =
|
|
||||||
binarySearchExclusive(srcKey[segmentBase + i * SAMPLE_STRIDE],
|
|
||||||
srcKey + segmentBase + stride, lenB, sortDir);
|
|
||||||
}
|
|
||||||
|
|
||||||
if (i < nB) {
|
|
||||||
ranksB[(segmentBase + stride) / SAMPLE_STRIDE + i] = i * SAMPLE_STRIDE;
|
|
||||||
ranksA[(segmentBase + stride) / SAMPLE_STRIDE + i] =
|
|
||||||
binarySearchInclusive(
|
|
||||||
srcKey[segmentBase + stride + i * SAMPLE_STRIDE],
|
|
||||||
srcKey + segmentBase, lenA, sortDir);
|
|
||||||
}
|
|
||||||
}
|
|
||||||
}
|
|
||||||
|
|
||||||
////////////////////////////////////////////////////////////////////////////////
|
|
||||||
// Merge step 2: merge ranks and indices to derive elementary intervals
|
|
||||||
////////////////////////////////////////////////////////////////////////////////
|
|
||||||
static void mergeRanksAndIndices(uint *limits, uint *ranks, uint stride,
|
|
||||||
uint N) {
|
|
||||||
uint lastSegmentElements = N % (2 * stride);
|
|
||||||
uint sampleCount =
|
|
||||||
(lastSegmentElements > stride)
|
|
||||||
? (N + 2 * stride - lastSegmentElements) / (2 * SAMPLE_STRIDE)
|
|
||||||
: (N - lastSegmentElements) / (2 * SAMPLE_STRIDE);
|
|
||||||
|
|
||||||
for (uint pos = 0; pos < sampleCount; pos++) {
|
|
||||||
const uint i = pos & ((stride / SAMPLE_STRIDE) - 1);
|
|
||||||
const uint segmentBase = (pos - i) * (2 * SAMPLE_STRIDE);
|
|
||||||
|
|
||||||
const uint lenA = stride;
|
|
||||||
const uint lenB = umin(stride, N - segmentBase - stride);
|
|
||||||
const uint nA = stride / SAMPLE_STRIDE;
|
|
||||||
const uint nB = getSampleCount(lenB);
|
|
||||||
|
|
||||||
if (i < nA) {
|
|
||||||
uint dstPosA =
|
|
||||||
binarySearchExclusive(ranks[(segmentBase + 0) / SAMPLE_STRIDE + i],
|
|
||||||
ranks + (segmentBase + stride) / SAMPLE_STRIDE,
|
|
||||||
nB, 1) +
|
|
||||||
i;
|
|
||||||
assert(dstPosA < nA + nB);
|
|
||||||
limits[(segmentBase / SAMPLE_STRIDE) + dstPosA] =
|
|
||||||
ranks[(segmentBase + 0) / SAMPLE_STRIDE + i];
|
|
||||||
}
|
|
||||||
|
|
||||||
if (i < nB) {
|
|
||||||
uint dstPosA = binarySearchInclusive(
|
|
||||||
ranks[(segmentBase + stride) / SAMPLE_STRIDE + i],
|
|
||||||
ranks + (segmentBase + 0) / SAMPLE_STRIDE, nA, 1) +
|
|
||||||
i;
|
|
||||||
assert(dstPosA < nA + nB);
|
|
||||||
limits[(segmentBase / SAMPLE_STRIDE) + dstPosA] =
|
|
||||||
ranks[(segmentBase + stride) / SAMPLE_STRIDE + i];
|
|
||||||
}
|
|
||||||
}
|
|
||||||
}
|
|
||||||
|
|
||||||
////////////////////////////////////////////////////////////////////////////////
|
|
||||||
// Merge step 3: merge elementary intervals (each interval is <= SAMPLE_STRIDE)
|
|
||||||
////////////////////////////////////////////////////////////////////////////////
|
|
||||||
static void merge(uint *dstKey, uint *dstVal, uint *srcAKey, uint *srcAVal,
|
|
||||||
uint *srcBKey, uint *srcBVal, uint lenA, uint lenB,
|
|
||||||
uint sortDir) {
|
|
||||||
checkOrder(srcAKey, lenA, sortDir);
|
|
||||||
checkOrder(srcBKey, lenB, sortDir);
|
|
||||||
|
|
||||||
for (uint i = 0; i < lenA; i++) {
|
|
||||||
uint dstPos = binarySearchExclusive(srcAKey[i], srcBKey, lenB, sortDir) + i;
|
|
||||||
assert(dstPos < lenA + lenB);
|
|
||||||
dstKey[dstPos] = srcAKey[i];
|
|
||||||
dstVal[dstPos] = srcAVal[i];
|
|
||||||
}
|
|
||||||
|
|
||||||
for (uint i = 0; i < lenB; i++) {
|
|
||||||
uint dstPos = binarySearchInclusive(srcBKey[i], srcAKey, lenA, sortDir) + i;
|
|
||||||
assert(dstPos < lenA + lenB);
|
|
||||||
dstKey[dstPos] = srcBKey[i];
|
|
||||||
dstVal[dstPos] = srcBVal[i];
|
|
||||||
}
|
|
||||||
}
|
|
||||||
|
|
||||||
static void mergeElementaryIntervals(uint *dstKey, uint *dstVal, uint *srcKey,
|
|
||||||
uint *srcVal, uint *limitsA, uint *limitsB,
|
|
||||||
uint stride, uint N, uint sortDir) {
|
|
||||||
uint lastSegmentElements = N % (2 * stride);
|
|
||||||
uint mergePairs = (lastSegmentElements > stride)
|
|
||||||
? getSampleCount(N)
|
|
||||||
: (N - lastSegmentElements) / SAMPLE_STRIDE;
|
|
||||||
|
|
||||||
for (uint pos = 0; pos < mergePairs; pos++) {
|
|
||||||
uint i = pos & ((2 * stride) / SAMPLE_STRIDE - 1);
|
|
||||||
uint segmentBase = (pos - i) * SAMPLE_STRIDE;
|
|
||||||
|
|
||||||
const uint lenA = stride;
|
|
||||||
const uint lenB = umin(stride, N - segmentBase - stride);
|
|
||||||
const uint nA = stride / SAMPLE_STRIDE;
|
|
||||||
const uint nB = getSampleCount(lenB);
|
|
||||||
const uint n = nA + nB;
|
|
||||||
|
|
||||||
const uint startPosA = limitsA[pos];
|
|
||||||
const uint endPosA = (i + 1 < n) ? limitsA[pos + 1] : lenA;
|
|
||||||
const uint startPosB = limitsB[pos];
|
|
||||||
const uint endPosB = (i + 1 < n) ? limitsB[pos + 1] : lenB;
|
|
||||||
const uint startPosDst = startPosA + startPosB;
|
|
||||||
|
|
||||||
assert(startPosA <= endPosA && endPosA <= lenA);
|
|
||||||
assert(startPosB <= endPosB && endPosB <= lenB);
|
|
||||||
assert((endPosA - startPosA) <= SAMPLE_STRIDE);
|
|
||||||
assert((endPosB - startPosB) <= SAMPLE_STRIDE);
|
|
||||||
|
|
||||||
merge(dstKey + segmentBase + startPosDst,
|
|
||||||
dstVal + segmentBase + startPosDst,
|
|
||||||
(srcKey + segmentBase + 0) + startPosA,
|
|
||||||
(srcVal + segmentBase + 0) + startPosA,
|
|
||||||
(srcKey + segmentBase + stride) + startPosB,
|
|
||||||
(srcVal + segmentBase + stride) + startPosB, endPosA - startPosA,
|
|
||||||
endPosB - startPosB, sortDir);
|
|
||||||
}
|
|
||||||
}
|
|
||||||
|
|
||||||
////////////////////////////////////////////////////////////////////////////////
|
|
||||||
// Retarded bubble sort
|
|
||||||
////////////////////////////////////////////////////////////////////////////////
|
|
||||||
static void bubbleSort(uint *key, uint *val, uint N, uint sortDir) {
|
|
||||||
if (N <= 1) {
|
|
||||||
return;
|
|
||||||
}
|
|
||||||
|
|
||||||
for (uint bottom = 0; bottom < N - 1; bottom++) {
|
|
||||||
uint savePos = bottom;
|
|
||||||
uint saveKey = key[bottom];
|
|
||||||
|
|
||||||
for (uint i = bottom + 1; i < N; i++)
|
|
||||||
if ((sortDir && (key[i] < saveKey)) || (!sortDir && (key[i] > saveKey))) {
|
|
||||||
savePos = i;
|
|
||||||
saveKey = key[i];
|
|
||||||
}
|
|
||||||
|
|
||||||
if (savePos != bottom) {
|
|
||||||
uint t;
|
|
||||||
t = key[savePos];
|
|
||||||
key[savePos] = key[bottom];
|
|
||||||
key[bottom] = t;
|
|
||||||
t = val[savePos];
|
|
||||||
val[savePos] = val[bottom];
|
|
||||||
val[bottom] = t;
|
|
||||||
}
|
|
||||||
}
|
|
||||||
}
|
|
||||||
|
|
||||||
////////////////////////////////////////////////////////////////////////////////
|
|
||||||
// Interface function
|
|
||||||
////////////////////////////////////////////////////////////////////////////////
|
|
||||||
extern "C" void mergeSortHost(uint *dstKey, uint *dstVal, uint *bufKey,
|
|
||||||
uint *bufVal, uint *srcKey, uint *srcVal, uint N,
|
|
||||||
uint sortDir) {
|
|
||||||
uint *ikey, *ival, *okey, *oval;
|
|
||||||
uint stageCount = 0;
|
|
||||||
|
|
||||||
for (uint stride = SHARED_SIZE_LIMIT; stride < N; stride <<= 1, stageCount++)
|
|
||||||
;
|
|
||||||
|
|
||||||
if (stageCount & 1) {
|
|
||||||
ikey = bufKey;
|
|
||||||
ival = bufVal;
|
|
||||||
okey = dstKey;
|
|
||||||
oval = dstVal;
|
|
||||||
} else {
|
|
||||||
ikey = dstKey;
|
|
||||||
ival = dstVal;
|
|
||||||
okey = bufKey;
|
|
||||||
oval = bufVal;
|
|
||||||
}
|
|
||||||
|
|
||||||
printf("Bottom-level sort...\n");
|
|
||||||
memcpy(ikey, srcKey, N * sizeof(uint));
|
|
||||||
memcpy(ival, srcVal, N * sizeof(uint));
|
|
||||||
|
|
||||||
for (uint pos = 0; pos < N; pos += SHARED_SIZE_LIMIT) {
|
|
||||||
bubbleSort(ikey + pos, ival + pos, umin(SHARED_SIZE_LIMIT, N - pos),
|
|
||||||
sortDir);
|
|
||||||
}
|
|
||||||
|
|
||||||
printf("Merge...\n");
|
|
||||||
uint *ranksA = (uint *)malloc(getSampleCount(N) * sizeof(uint));
|
|
||||||
uint *ranksB = (uint *)malloc(getSampleCount(N) * sizeof(uint));
|
|
||||||
uint *limitsA = (uint *)malloc(getSampleCount(N) * sizeof(uint));
|
|
||||||
uint *limitsB = (uint *)malloc(getSampleCount(N) * sizeof(uint));
|
|
||||||
memset(ranksA, 0xFF, getSampleCount(N) * sizeof(uint));
|
|
||||||
memset(ranksB, 0xFF, getSampleCount(N) * sizeof(uint));
|
|
||||||
memset(limitsA, 0xFF, getSampleCount(N) * sizeof(uint));
|
|
||||||
memset(limitsB, 0xFF, getSampleCount(N) * sizeof(uint));
|
|
||||||
|
|
||||||
for (uint stride = SHARED_SIZE_LIMIT; stride < N; stride <<= 1) {
|
|
||||||
uint lastSegmentElements = N % (2 * stride);
|
|
||||||
|
|
||||||
// Find sample ranks and prepare for limiters merge
|
|
||||||
generateSampleRanks(ranksA, ranksB, ikey, stride, N, sortDir);
|
|
||||||
|
|
||||||
// Merge ranks and indices
|
|
||||||
mergeRanksAndIndices(limitsA, ranksA, stride, N);
|
|
||||||
mergeRanksAndIndices(limitsB, ranksB, stride, N);
|
|
||||||
|
|
||||||
// Merge elementary intervals
|
|
||||||
mergeElementaryIntervals(okey, oval, ikey, ival, limitsA, limitsB, stride,
|
|
||||||
N, sortDir);
|
|
||||||
|
|
||||||
if (lastSegmentElements <= stride) {
|
|
||||||
// Last merge segment consists of a single array which just needs to be
|
|
||||||
// passed through
|
|
||||||
memcpy(okey + (N - lastSegmentElements), ikey + (N - lastSegmentElements),
|
|
||||||
lastSegmentElements * sizeof(uint));
|
|
||||||
memcpy(oval + (N - lastSegmentElements), ival + (N - lastSegmentElements),
|
|
||||||
lastSegmentElements * sizeof(uint));
|
|
||||||
}
|
|
||||||
|
|
||||||
uint *t;
|
|
||||||
t = ikey;
|
|
||||||
ikey = okey;
|
|
||||||
okey = t;
|
|
||||||
t = ival;
|
|
||||||
ival = oval;
|
|
||||||
oval = t;
|
|
||||||
}
|
|
||||||
|
|
||||||
free(limitsB);
|
|
||||||
free(limitsA);
|
|
||||||
free(ranksB);
|
|
||||||
free(ranksA);
|
|
||||||
}
|
|
||||||
@@ -1,132 +0,0 @@
|
|||||||
/* Copyright (c) 2022, NVIDIA CORPORATION. All rights reserved.
|
|
||||||
*
|
|
||||||
* Redistribution and use in source and binary forms, with or without
|
|
||||||
* modification, are permitted provided that the following conditions
|
|
||||||
* are met:
|
|
||||||
* * Redistributions of source code must retain the above copyright
|
|
||||||
* notice, this list of conditions and the following disclaimer.
|
|
||||||
* * Redistributions in binary form must reproduce the above copyright
|
|
||||||
* notice, this list of conditions and the following disclaimer in the
|
|
||||||
* documentation and/or other materials provided with the distribution.
|
|
||||||
* * Neither the name of NVIDIA CORPORATION nor the names of its
|
|
||||||
* contributors may be used to endorse or promote products derived
|
|
||||||
* from this software without specific prior written permission.
|
|
||||||
*
|
|
||||||
* THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS ``AS IS'' AND ANY
|
|
||||||
* EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
|
|
||||||
* IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR
|
|
||||||
* PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER OR
|
|
||||||
* CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL,
|
|
||||||
* EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO,
|
|
||||||
* PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR
|
|
||||||
* PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY
|
|
||||||
* OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
|
|
||||||
* (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
|
|
||||||
* OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
|
|
||||||
*/
|
|
||||||
|
|
||||||
#include <assert.h>
|
|
||||||
#include <stdio.h>
|
|
||||||
#include <stdlib.h>
|
|
||||||
#include <string.h>
|
|
||||||
#include "mergeSort_common.h"
|
|
||||||
|
|
||||||
////////////////////////////////////////////////////////////////////////////////
|
|
||||||
// Validate sorted keys array (check for integrity and proper order)
|
|
||||||
////////////////////////////////////////////////////////////////////////////////
|
|
||||||
extern "C" uint validateSortedKeys(uint *resKey, uint *srcKey, uint batchSize,
|
|
||||||
uint arrayLength, uint numValues,
|
|
||||||
uint sortDir) {
|
|
||||||
uint *srcHist;
|
|
||||||
uint *resHist;
|
|
||||||
|
|
||||||
if (arrayLength < 2) {
|
|
||||||
printf("validateSortedKeys(): arrays too short, exiting...\n");
|
|
||||||
return 1;
|
|
||||||
}
|
|
||||||
|
|
||||||
printf("...inspecting keys array: ");
|
|
||||||
srcHist = (uint *)malloc(numValues * sizeof(uint));
|
|
||||||
resHist = (uint *)malloc(numValues * sizeof(uint));
|
|
||||||
|
|
||||||
int flag = 1;
|
|
||||||
|
|
||||||
for (uint j = 0; j < batchSize;
|
|
||||||
j++, srcKey += arrayLength, resKey += arrayLength) {
|
|
||||||
// Build histograms for keys arrays
|
|
||||||
memset(srcHist, 0, numValues * sizeof(uint));
|
|
||||||
memset(resHist, 0, numValues * sizeof(uint));
|
|
||||||
|
|
||||||
for (uint i = 0; i < arrayLength; i++) {
|
|
||||||
if ((srcKey[i] < numValues) && (resKey[i] < numValues)) {
|
|
||||||
srcHist[srcKey[i]]++;
|
|
||||||
resHist[resKey[i]]++;
|
|
||||||
} else {
|
|
||||||
fprintf(
|
|
||||||
stderr,
|
|
||||||
"***Set %u source/result key arrays are not limited properly***\n",
|
|
||||||
j);
|
|
||||||
flag = 0;
|
|
||||||
goto brk;
|
|
||||||
}
|
|
||||||
}
|
|
||||||
|
|
||||||
// Compare the histograms
|
|
||||||
for (uint i = 0; i < numValues; i++)
|
|
||||||
if (srcHist[i] != resHist[i]) {
|
|
||||||
fprintf(stderr,
|
|
||||||
"***Set %u source/result keys histograms do not match***\n", j);
|
|
||||||
flag = 0;
|
|
||||||
goto brk;
|
|
||||||
}
|
|
||||||
|
|
||||||
// Finally check the ordering
|
|
||||||
for (uint i = 0; i < arrayLength - 1; i++)
|
|
||||||
if ((sortDir && (resKey[i] > resKey[i + 1])) ||
|
|
||||||
(!sortDir && (resKey[i] < resKey[i + 1]))) {
|
|
||||||
fprintf(stderr,
|
|
||||||
"***Set %u result key array is not ordered properly***\n", j);
|
|
||||||
flag = 0;
|
|
||||||
goto brk;
|
|
||||||
}
|
|
||||||
}
|
|
||||||
|
|
||||||
brk:
|
|
||||||
free(resHist);
|
|
||||||
free(srcHist);
|
|
||||||
|
|
||||||
if (flag) printf("OK\n");
|
|
||||||
|
|
||||||
return flag;
|
|
||||||
}
|
|
||||||
|
|
||||||
////////////////////////////////////////////////////////////////////////////////
|
|
||||||
// Value validation / stability check routines
|
|
||||||
////////////////////////////////////////////////////////////////////////////////
|
|
||||||
extern "C" void fillValues(uint *val, uint N) {
|
|
||||||
for (uint i = 0; i < N; i++) val[i] = i;
|
|
||||||
}
|
|
||||||
|
|
||||||
extern "C" int validateSortedValues(uint *resKey, uint *resVal, uint *srcKey,
|
|
||||||
uint batchSize, uint arrayLength) {
|
|
||||||
int correctFlag = 1, stableFlag = 1;
|
|
||||||
|
|
||||||
printf("...inspecting keys and values array: ");
|
|
||||||
|
|
||||||
for (uint i = 0; i < batchSize;
|
|
||||||
i++, resKey += arrayLength, resVal += arrayLength) {
|
|
||||||
for (uint j = 0; j < arrayLength; j++) {
|
|
||||||
if (resKey[j] != srcKey[resVal[j]]) correctFlag = 0;
|
|
||||||
|
|
||||||
if ((j < arrayLength - 1) && (resKey[j] == resKey[j + 1]) &&
|
|
||||||
(resVal[j] > resVal[j + 1]))
|
|
||||||
stableFlag = 0;
|
|
||||||
}
|
|
||||||
}
|
|
||||||
|
|
||||||
printf(correctFlag ? "OK\n" : "***corrupted!!!***\n");
|
|
||||||
printf(stableFlag ? "...stability property: stable!\n"
|
|
||||||
: "...stability property: NOT stable\n");
|
|
||||||
|
|
||||||
return correctFlag;
|
|
||||||
}
|
|
||||||
@@ -1,10 +0,0 @@
|
|||||||
{
|
|
||||||
"configurations": [
|
|
||||||
{
|
|
||||||
"name": "CUDA C++: Launch",
|
|
||||||
"type": "cuda-gdb",
|
|
||||||
"request": "launch",
|
|
||||||
"program": "${workspaceFolder}/simpleAWBarrier"
|
|
||||||
}
|
|
||||||
]
|
|
||||||
}
|
|
||||||
@@ -1,15 +0,0 @@
|
|||||||
{
|
|
||||||
"version": "2.0.0",
|
|
||||||
"tasks": [
|
|
||||||
{
|
|
||||||
"label": "sample",
|
|
||||||
"type": "shell",
|
|
||||||
"command": "make dbg=1",
|
|
||||||
"problemMatcher": ["$nvcc"],
|
|
||||||
"group": {
|
|
||||||
"kind": "build",
|
|
||||||
"isDefault": true
|
|
||||||
}
|
|
||||||
}
|
|
||||||
]
|
|
||||||
}
|
|
||||||
@@ -1,29 +0,0 @@
|
|||||||
cmake_minimum_required(VERSION 3.20)
|
|
||||||
|
|
||||||
list(APPEND CMAKE_MODULE_PATH "${CMAKE_CURRENT_SOURCE_DIR}/../../../cmake/Modules")
|
|
||||||
|
|
||||||
project(simpleAWBarrier LANGUAGES C CXX CUDA)
|
|
||||||
|
|
||||||
find_package(CUDAToolkit REQUIRED)
|
|
||||||
|
|
||||||
set(CMAKE_POSITION_INDEPENDENT_CODE ON)
|
|
||||||
|
|
||||||
set(CMAKE_CUDA_ARCHITECTURES 70 72 75 80 86 87 89 90 100 101 120)
|
|
||||||
set(CMAKE_CUDA_FLAGS "${CMAKE_CUDA_FLAGS} -Wno-deprecated-gpu-targets")
|
|
||||||
|
|
||||||
if(CMAKE_BUILD_TYPE STREQUAL "Debug")
|
|
||||||
# set(CMAKE_CUDA_FLAGS "${CMAKE_CUDA_FLAGS} -G") # enable cuda-gdb (expensive)
|
|
||||||
endif()
|
|
||||||
|
|
||||||
# Include directories and libraries
|
|
||||||
include_directories(../../../Common)
|
|
||||||
|
|
||||||
# Source file
|
|
||||||
# Add target for simpleAWBarrier
|
|
||||||
add_executable(simpleAWBarrier simpleAWBarrier.cu)
|
|
||||||
|
|
||||||
target_compile_options(simpleAWBarrier PRIVATE $<$<COMPILE_LANGUAGE:CUDA>:--extended-lambda>)
|
|
||||||
|
|
||||||
target_compile_features(simpleAWBarrier PRIVATE cxx_std_17 cuda_std_17)
|
|
||||||
|
|
||||||
set_target_properties(simpleAWBarrier PROPERTIES CUDA_SEPARABLE_COMPILATION ON)
|
|
||||||
@@ -1,36 +0,0 @@
|
|||||||
# simpleAWBarrier - Simple Arrive Wait Barrier
|
|
||||||
|
|
||||||
## Description
|
|
||||||
|
|
||||||
A simple demonstration of arrive wait barriers.
|
|
||||||
|
|
||||||
## Key Concepts
|
|
||||||
|
|
||||||
Arrive Wait Barrier
|
|
||||||
|
|
||||||
## Supported SM Architectures
|
|
||||||
|
|
||||||
[SM 7.0 ](https://developer.nvidia.com/cuda-gpus) [SM 7.2 ](https://developer.nvidia.com/cuda-gpus) [SM 7.5 ](https://developer.nvidia.com/cuda-gpus) [SM 8.0 ](https://developer.nvidia.com/cuda-gpus) [SM 8.6 ](https://developer.nvidia.com/cuda-gpus) [SM 8.7 ](https://developer.nvidia.com/cuda-gpus) [SM 8.9 ](https://developer.nvidia.com/cuda-gpus) [SM 9.0 ](https://developer.nvidia.com/cuda-gpus)
|
|
||||||
|
|
||||||
## Supported OSes
|
|
||||||
|
|
||||||
Linux, Windows, QNX
|
|
||||||
|
|
||||||
## Supported CPU Architecture
|
|
||||||
|
|
||||||
x86_64, armv7l, aarch64
|
|
||||||
|
|
||||||
## CUDA APIs involved
|
|
||||||
|
|
||||||
### [CUDA Runtime API](http://docs.nvidia.com/cuda/cuda-runtime-api/index.html)
|
|
||||||
cudaStreamCreateWithFlags, cudaFree, cudaDeviceGetAttribute, cudaMallocHost, cudaFreeHost, cudaStreamSynchronize, cudaLaunchCooperativeKernel, cudaMalloc, cudaOccupancyMaxActiveBlocksPerMultiprocessor, cudaMemcpyAsync, cudaOccupancyMaxPotentialBlockSize
|
|
||||||
|
|
||||||
## Dependencies needed to build/run
|
|
||||||
[CPP11](../../../README.md#cpp11), [MBCG](../../../README.md#mbcg)
|
|
||||||
|
|
||||||
## Prerequisites
|
|
||||||
|
|
||||||
Download and install the [CUDA Toolkit 12.5](https://developer.nvidia.com/cuda-downloads) for your corresponding platform.
|
|
||||||
Make sure the dependencies mentioned in [Dependencies]() section above are installed.
|
|
||||||
|
|
||||||
## References (for more details)
|
|
||||||
@@ -1,255 +0,0 @@
|
|||||||
/* Copyright (c) 2022, NVIDIA CORPORATION. All rights reserved.
|
|
||||||
*
|
|
||||||
* Redistribution and use in source and binary forms, with or without
|
|
||||||
* modification, are permitted provided that the following conditions
|
|
||||||
* are met:
|
|
||||||
* * Redistributions of source code must retain the above copyright
|
|
||||||
* notice, this list of conditions and the following disclaimer.
|
|
||||||
* * Redistributions in binary form must reproduce the above copyright
|
|
||||||
* notice, this list of conditions and the following disclaimer in the
|
|
||||||
* documentation and/or other materials provided with the distribution.
|
|
||||||
* * Neither the name of NVIDIA CORPORATION nor the names of its
|
|
||||||
* contributors may be used to endorse or promote products derived
|
|
||||||
* from this software without specific prior written permission.
|
|
||||||
*
|
|
||||||
* THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS ``AS IS'' AND ANY
|
|
||||||
* EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
|
|
||||||
* IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR
|
|
||||||
* PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER OR
|
|
||||||
* CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL,
|
|
||||||
* EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO,
|
|
||||||
* PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR
|
|
||||||
* PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY
|
|
||||||
* OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
|
|
||||||
* (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
|
|
||||||
* OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
|
|
||||||
*/
|
|
||||||
|
|
||||||
// Includes, system
|
|
||||||
#include <stdio.h>
|
|
||||||
|
|
||||||
// Includes CUDA
|
|
||||||
#include <cuda_runtime.h>
|
|
||||||
#include <cuda/barrier>
|
|
||||||
#include <cooperative_groups.h>
|
|
||||||
|
|
||||||
// Utilities and timing functions
|
|
||||||
#include <helper_functions.h> // includes cuda.h and cuda_runtime_api.h
|
|
||||||
|
|
||||||
// CUDA helper functions
|
|
||||||
#include <helper_cuda.h> // helper functions for CUDA error check
|
|
||||||
|
|
||||||
namespace cg = cooperative_groups;
|
|
||||||
|
|
||||||
#if __CUDA_ARCH__ >= 700
|
|
||||||
template <bool writeSquareRoot>
|
|
||||||
__device__ void reduceBlockData(
|
|
||||||
cuda::barrier<cuda::thread_scope_block> &barrier,
|
|
||||||
cg::thread_block_tile<32> &tile32, double &threadSum, double *result) {
|
|
||||||
extern __shared__ double tmp[];
|
|
||||||
|
|
||||||
#pragma unroll
|
|
||||||
for (int offset = tile32.size() / 2; offset > 0; offset /= 2) {
|
|
||||||
threadSum += tile32.shfl_down(threadSum, offset);
|
|
||||||
}
|
|
||||||
if (tile32.thread_rank() == 0) {
|
|
||||||
tmp[tile32.meta_group_rank()] = threadSum;
|
|
||||||
}
|
|
||||||
|
|
||||||
auto token = barrier.arrive();
|
|
||||||
|
|
||||||
barrier.wait(std::move(token));
|
|
||||||
|
|
||||||
// The warp 0 will perform last round of reduction
|
|
||||||
if (tile32.meta_group_rank() == 0) {
|
|
||||||
double beta = tile32.thread_rank() < tile32.meta_group_size()
|
|
||||||
? tmp[tile32.thread_rank()]
|
|
||||||
: 0.0;
|
|
||||||
|
|
||||||
#pragma unroll
|
|
||||||
for (int offset = tile32.size() / 2; offset > 0; offset /= 2) {
|
|
||||||
beta += tile32.shfl_down(beta, offset);
|
|
||||||
}
|
|
||||||
|
|
||||||
if (tile32.thread_rank() == 0) {
|
|
||||||
if (writeSquareRoot)
|
|
||||||
*result = sqrt(beta);
|
|
||||||
else
|
|
||||||
*result = beta;
|
|
||||||
}
|
|
||||||
}
|
|
||||||
}
|
|
||||||
#endif
|
|
||||||
|
|
||||||
__global__ void normVecByDotProductAWBarrier(float *vecA, float *vecB,
|
|
||||||
double *partialResults, int size) {
|
|
||||||
#if __CUDA_ARCH__ >= 700
|
|
||||||
#pragma diag_suppress static_var_with_dynamic_init
|
|
||||||
cg::thread_block cta = cg::this_thread_block();
|
|
||||||
cg::grid_group grid = cg::this_grid();
|
|
||||||
;
|
|
||||||
cg::thread_block_tile<32> tile32 = cg::tiled_partition<32>(cta);
|
|
||||||
|
|
||||||
__shared__ cuda::barrier<cuda::thread_scope_block> barrier;
|
|
||||||
|
|
||||||
if (threadIdx.x == 0) {
|
|
||||||
init(&barrier, blockDim.x);
|
|
||||||
}
|
|
||||||
|
|
||||||
cg::sync(cta);
|
|
||||||
|
|
||||||
double threadSum = 0.0;
|
|
||||||
for (int i = grid.thread_rank(); i < size; i += grid.size()) {
|
|
||||||
threadSum += (double)(vecA[i] * vecB[i]);
|
|
||||||
}
|
|
||||||
|
|
||||||
// Each thread block performs reduction of partial dotProducts and writes to
|
|
||||||
// global mem.
|
|
||||||
reduceBlockData<false>(barrier, tile32, threadSum,
|
|
||||||
&partialResults[blockIdx.x]);
|
|
||||||
|
|
||||||
cg::sync(grid);
|
|
||||||
|
|
||||||
// One block performs the final summation of partial dot products
|
|
||||||
// of all the thread blocks and writes the sqrt of final dot product.
|
|
||||||
if (blockIdx.x == 0) {
|
|
||||||
threadSum = 0.0;
|
|
||||||
for (int i = cta.thread_rank(); i < gridDim.x; i += cta.size()) {
|
|
||||||
threadSum += partialResults[i];
|
|
||||||
}
|
|
||||||
reduceBlockData<true>(barrier, tile32, threadSum, &partialResults[0]);
|
|
||||||
}
|
|
||||||
|
|
||||||
cg::sync(grid);
|
|
||||||
|
|
||||||
const double finalValue = partialResults[0];
|
|
||||||
|
|
||||||
// Perform normalization of vecA & vecB.
|
|
||||||
for (int i = grid.thread_rank(); i < size; i += grid.size()) {
|
|
||||||
vecA[i] = (float)vecA[i] / finalValue;
|
|
||||||
vecB[i] = (float)vecB[i] / finalValue;
|
|
||||||
}
|
|
||||||
#endif
|
|
||||||
}
|
|
||||||
|
|
||||||
int runNormVecByDotProductAWBarrier(int argc, char **argv, int deviceId);
|
|
||||||
|
|
||||||
////////////////////////////////////////////////////////////////////////////////
|
|
||||||
// Program main
|
|
||||||
////////////////////////////////////////////////////////////////////////////////
|
|
||||||
int main(int argc, char **argv) {
|
|
||||||
printf("%s starting...\n", argv[0]);
|
|
||||||
|
|
||||||
// This will pick the best possible CUDA capable device
|
|
||||||
int dev = findCudaDevice(argc, (const char **)argv);
|
|
||||||
|
|
||||||
int major = 0;
|
|
||||||
checkCudaErrors(
|
|
||||||
cudaDeviceGetAttribute(&major, cudaDevAttrComputeCapabilityMajor, dev));
|
|
||||||
|
|
||||||
// Arrive-Wait Barrier require a GPU of Volta (SM7X) architecture or higher.
|
|
||||||
if (major < 7) {
|
|
||||||
printf("simpleAWBarrier requires SM 7.0 or higher. Exiting...\n");
|
|
||||||
exit(EXIT_WAIVED);
|
|
||||||
}
|
|
||||||
|
|
||||||
int supportsCooperativeLaunch = 0;
|
|
||||||
checkCudaErrors(cudaDeviceGetAttribute(&supportsCooperativeLaunch,
|
|
||||||
cudaDevAttrCooperativeLaunch, dev));
|
|
||||||
|
|
||||||
if (!supportsCooperativeLaunch) {
|
|
||||||
printf(
|
|
||||||
"\nSelected GPU (%d) does not support Cooperative Kernel Launch, "
|
|
||||||
"Waiving the run\n",
|
|
||||||
dev);
|
|
||||||
exit(EXIT_WAIVED);
|
|
||||||
}
|
|
||||||
|
|
||||||
int testResult = runNormVecByDotProductAWBarrier(argc, argv, dev);
|
|
||||||
|
|
||||||
printf("%s completed, returned %s\n", argv[0], testResult ? "OK" : "ERROR!");
|
|
||||||
exit(testResult ? EXIT_SUCCESS : EXIT_FAILURE);
|
|
||||||
}
|
|
||||||
|
|
||||||
int runNormVecByDotProductAWBarrier(int argc, char **argv, int deviceId) {
|
|
||||||
float *vecA, *d_vecA;
|
|
||||||
float *vecB, *d_vecB;
|
|
||||||
double *d_partialResults;
|
|
||||||
int size = 10000000;
|
|
||||||
|
|
||||||
checkCudaErrors(cudaMallocHost(&vecA, sizeof(float) * size));
|
|
||||||
checkCudaErrors(cudaMallocHost(&vecB, sizeof(float) * size));
|
|
||||||
|
|
||||||
checkCudaErrors(cudaMalloc(&d_vecA, sizeof(float) * size));
|
|
||||||
checkCudaErrors(cudaMalloc(&d_vecB, sizeof(float) * size));
|
|
||||||
|
|
||||||
float baseVal = 2.0;
|
|
||||||
for (int i = 0; i < size; i++) {
|
|
||||||
vecA[i] = vecB[i] = baseVal;
|
|
||||||
}
|
|
||||||
|
|
||||||
cudaStream_t stream;
|
|
||||||
checkCudaErrors(cudaStreamCreateWithFlags(&stream, cudaStreamNonBlocking));
|
|
||||||
|
|
||||||
checkCudaErrors(cudaMemcpyAsync(d_vecA, vecA, sizeof(float) * size,
|
|
||||||
cudaMemcpyHostToDevice, stream));
|
|
||||||
checkCudaErrors(cudaMemcpyAsync(d_vecB, vecB, sizeof(float) * size,
|
|
||||||
cudaMemcpyHostToDevice, stream));
|
|
||||||
|
|
||||||
// Kernel configuration, where a one-dimensional
|
|
||||||
// grid and one-dimensional blocks are configured.
|
|
||||||
int minGridSize = 0, blockSize = 0;
|
|
||||||
checkCudaErrors(cudaOccupancyMaxPotentialBlockSize(
|
|
||||||
&minGridSize, &blockSize, (void *)normVecByDotProductAWBarrier, 0, size));
|
|
||||||
|
|
||||||
int smemSize = ((blockSize / 32) + 1) * sizeof(double);
|
|
||||||
|
|
||||||
int numBlocksPerSm = 0;
|
|
||||||
checkCudaErrors(cudaOccupancyMaxActiveBlocksPerMultiprocessor(
|
|
||||||
&numBlocksPerSm, normVecByDotProductAWBarrier, blockSize, smemSize));
|
|
||||||
|
|
||||||
int multiProcessorCount = 0;
|
|
||||||
checkCudaErrors(cudaDeviceGetAttribute(
|
|
||||||
&multiProcessorCount, cudaDevAttrMultiProcessorCount, deviceId));
|
|
||||||
|
|
||||||
minGridSize = multiProcessorCount * numBlocksPerSm;
|
|
||||||
checkCudaErrors(cudaMalloc(&d_partialResults, minGridSize * sizeof(double)));
|
|
||||||
|
|
||||||
printf(
|
|
||||||
"Launching normVecByDotProductAWBarrier kernel with numBlocks = %d "
|
|
||||||
"blockSize = %d\n",
|
|
||||||
minGridSize, blockSize);
|
|
||||||
|
|
||||||
dim3 dimGrid(minGridSize, 1, 1), dimBlock(blockSize, 1, 1);
|
|
||||||
|
|
||||||
void *kernelArgs[] = {(void *)&d_vecA, (void *)&d_vecB,
|
|
||||||
(void *)&d_partialResults, (void *)&size};
|
|
||||||
|
|
||||||
checkCudaErrors(
|
|
||||||
cudaLaunchCooperativeKernel((void *)normVecByDotProductAWBarrier, dimGrid,
|
|
||||||
dimBlock, kernelArgs, smemSize, stream));
|
|
||||||
|
|
||||||
checkCudaErrors(cudaMemcpyAsync(vecA, d_vecA, sizeof(float) * size,
|
|
||||||
cudaMemcpyDeviceToHost, stream));
|
|
||||||
checkCudaErrors(cudaStreamSynchronize(stream));
|
|
||||||
|
|
||||||
float expectedResult = (baseVal / sqrt(size * baseVal * baseVal));
|
|
||||||
unsigned int matches = 0;
|
|
||||||
for (int i = 0; i < size; i++) {
|
|
||||||
if ((vecA[i] - expectedResult) > 0.00001) {
|
|
||||||
printf("mismatch at i = %d\n", i);
|
|
||||||
break;
|
|
||||||
} else {
|
|
||||||
matches++;
|
|
||||||
}
|
|
||||||
}
|
|
||||||
|
|
||||||
printf("Result = %s\n", matches == size ? "PASSED" : "FAILED");
|
|
||||||
checkCudaErrors(cudaFree(d_vecA));
|
|
||||||
checkCudaErrors(cudaFree(d_vecB));
|
|
||||||
checkCudaErrors(cudaFree(d_partialResults));
|
|
||||||
|
|
||||||
checkCudaErrors(cudaFreeHost(vecA));
|
|
||||||
checkCudaErrors(cudaFreeHost(vecB));
|
|
||||||
return matches == size;
|
|
||||||
}
|
|
||||||
@@ -1,10 +0,0 @@
|
|||||||
{
|
|
||||||
"configurations": [
|
|
||||||
{
|
|
||||||
"name": "CUDA C++: Launch",
|
|
||||||
"type": "cuda-gdb",
|
|
||||||
"request": "launch",
|
|
||||||
"program": "${workspaceFolder}/simpleAssert"
|
|
||||||
}
|
|
||||||
]
|
|
||||||
}
|
|
||||||
@@ -1,15 +0,0 @@
|
|||||||
{
|
|
||||||
"version": "2.0.0",
|
|
||||||
"tasks": [
|
|
||||||
{
|
|
||||||
"label": "sample",
|
|
||||||
"type": "shell",
|
|
||||||
"command": "make dbg=1",
|
|
||||||
"problemMatcher": ["$nvcc"],
|
|
||||||
"group": {
|
|
||||||
"kind": "build",
|
|
||||||
"isDefault": true
|
|
||||||
}
|
|
||||||
}
|
|
||||||
]
|
|
||||||
}
|
|
||||||
@@ -1,31 +0,0 @@
|
|||||||
cmake_minimum_required(VERSION 3.20)
|
|
||||||
|
|
||||||
list(APPEND CMAKE_MODULE_PATH "${CMAKE_CURRENT_SOURCE_DIR}/../../../cmake/Modules")
|
|
||||||
|
|
||||||
project(simpleAssert LANGUAGES C CXX CUDA)
|
|
||||||
|
|
||||||
find_package(CUDAToolkit REQUIRED)
|
|
||||||
|
|
||||||
set(CMAKE_POSITION_INDEPENDENT_CODE ON)
|
|
||||||
|
|
||||||
set(CMAKE_CUDA_ARCHITECTURES 50 52 60 61 70 72 75 80 86 87 89 90 100 101 120)
|
|
||||||
set(CMAKE_CUDA_FLAGS "${CMAKE_CUDA_FLAGS} -Wno-deprecated-gpu-targets")
|
|
||||||
if(CMAKE_BUILD_TYPE STREQUAL "Debug")
|
|
||||||
# set(CMAKE_CUDA_FLAGS "${CMAKE_CUDA_FLAGS} -G") # enable cuda-gdb (expensive)
|
|
||||||
endif()
|
|
||||||
|
|
||||||
# Removes -DNDEBUG For Print specific logs in this sample.
|
|
||||||
string(REPLACE "-DNDEBUG" "" CMAKE_CUDA_FLAGS_RELEASE "${CMAKE_CUDA_FLAGS_RELEASE}")
|
|
||||||
|
|
||||||
# Include directories and libraries
|
|
||||||
include_directories(../../../Common)
|
|
||||||
|
|
||||||
# Source file
|
|
||||||
# Add target for simpleAssert
|
|
||||||
add_executable(simpleAssert simpleAssert.cu)
|
|
||||||
|
|
||||||
target_compile_options(simpleAssert PRIVATE $<$<COMPILE_LANGUAGE:CUDA>:--extended-lambda>)
|
|
||||||
|
|
||||||
target_compile_features(simpleAssert PRIVATE cxx_std_17 cuda_std_17)
|
|
||||||
|
|
||||||
set_target_properties(simpleAssert PROPERTIES CUDA_SEPARABLE_COMPILATION ON)
|
|
||||||
@@ -1,32 +0,0 @@
|
|||||||
# simpleAssert - simpleAssert
|
|
||||||
|
|
||||||
## Description
|
|
||||||
|
|
||||||
This CUDA Runtime API sample is a very basic sample that implements how to use the assert function in the device code. Requires Compute Capability 2.0 .
|
|
||||||
|
|
||||||
## Key Concepts
|
|
||||||
|
|
||||||
Assert
|
|
||||||
|
|
||||||
## Supported SM Architectures
|
|
||||||
|
|
||||||
[SM 5.0 ](https://developer.nvidia.com/cuda-gpus) [SM 5.2 ](https://developer.nvidia.com/cuda-gpus) [SM 5.3 ](https://developer.nvidia.com/cuda-gpus) [SM 6.0 ](https://developer.nvidia.com/cuda-gpus) [SM 6.1 ](https://developer.nvidia.com/cuda-gpus) [SM 7.0 ](https://developer.nvidia.com/cuda-gpus) [SM 7.2 ](https://developer.nvidia.com/cuda-gpus) [SM 7.5 ](https://developer.nvidia.com/cuda-gpus) [SM 8.0 ](https://developer.nvidia.com/cuda-gpus) [SM 8.6 ](https://developer.nvidia.com/cuda-gpus) [SM 8.7 ](https://developer.nvidia.com/cuda-gpus) [SM 8.9 ](https://developer.nvidia.com/cuda-gpus) [SM 9.0 ](https://developer.nvidia.com/cuda-gpus)
|
|
||||||
|
|
||||||
## Supported OSes
|
|
||||||
|
|
||||||
Linux, Windows
|
|
||||||
|
|
||||||
## Supported CPU Architecture
|
|
||||||
|
|
||||||
x86_64, armv7l
|
|
||||||
|
|
||||||
## CUDA APIs involved
|
|
||||||
|
|
||||||
### [CUDA Runtime API](http://docs.nvidia.com/cuda/cuda-runtime-api/index.html)
|
|
||||||
cudaDeviceSynchronize, cudaGetErrorString
|
|
||||||
|
|
||||||
## Prerequisites
|
|
||||||
|
|
||||||
Download and install the [CUDA Toolkit 12.5](https://developer.nvidia.com/cuda-downloads) for your corresponding platform.
|
|
||||||
|
|
||||||
## References (for more details)
|
|
||||||
@@ -1,128 +0,0 @@
|
|||||||
/* Copyright (c) 2022, NVIDIA CORPORATION. All rights reserved.
|
|
||||||
*
|
|
||||||
* Redistribution and use in source and binary forms, with or without
|
|
||||||
* modification, are permitted provided that the following conditions
|
|
||||||
* are met:
|
|
||||||
* * Redistributions of source code must retain the above copyright
|
|
||||||
* notice, this list of conditions and the following disclaimer.
|
|
||||||
* * Redistributions in binary form must reproduce the above copyright
|
|
||||||
* notice, this list of conditions and the following disclaimer in the
|
|
||||||
* documentation and/or other materials provided with the distribution.
|
|
||||||
* * Neither the name of NVIDIA CORPORATION nor the names of its
|
|
||||||
* contributors may be used to endorse or promote products derived
|
|
||||||
* from this software without specific prior written permission.
|
|
||||||
*
|
|
||||||
* THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS ``AS IS'' AND ANY
|
|
||||||
* EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
|
|
||||||
* IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR
|
|
||||||
* PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER OR
|
|
||||||
* CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL,
|
|
||||||
* EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO,
|
|
||||||
* PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR
|
|
||||||
* PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY
|
|
||||||
* OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
|
|
||||||
* (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
|
|
||||||
* OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
|
|
||||||
*/
|
|
||||||
|
|
||||||
#ifdef _WIN32
|
|
||||||
#define WINDOWS_LEAN_AND_MEAN
|
|
||||||
#define NOMINMAX
|
|
||||||
#include <windows.h>
|
|
||||||
#else
|
|
||||||
#include <sys/utsname.h>
|
|
||||||
#endif
|
|
||||||
|
|
||||||
// Includes, system
|
|
||||||
#include <stdio.h>
|
|
||||||
#include <cassert>
|
|
||||||
|
|
||||||
// Includes CUDA
|
|
||||||
#include <cuda_runtime.h>
|
|
||||||
|
|
||||||
// Utilities and timing functions
|
|
||||||
#include <helper_functions.h> // includes cuda.h and cuda_runtime_api.h
|
|
||||||
|
|
||||||
// CUDA helper functions
|
|
||||||
#include <helper_cuda.h> // helper functions for CUDA error check
|
|
||||||
|
|
||||||
const char *sampleName = "simpleAssert";
|
|
||||||
|
|
||||||
////////////////////////////////////////////////////////////////////////////////
|
|
||||||
// Auto-Verification Code
|
|
||||||
bool testResult = true;
|
|
||||||
|
|
||||||
////////////////////////////////////////////////////////////////////////////////
|
|
||||||
// Kernels
|
|
||||||
////////////////////////////////////////////////////////////////////////////////
|
|
||||||
//! Tests assert function.
|
|
||||||
//! Thread whose id > N will print assertion failed error message.
|
|
||||||
////////////////////////////////////////////////////////////////////////////////
|
|
||||||
__global__ void testKernel(int N) {
|
|
||||||
int gtid = blockIdx.x * blockDim.x + threadIdx.x;
|
|
||||||
assert(gtid < N);
|
|
||||||
}
|
|
||||||
|
|
||||||
////////////////////////////////////////////////////////////////////////////////
|
|
||||||
// Declaration, forward
|
|
||||||
void runTest(int argc, char **argv);
|
|
||||||
|
|
||||||
////////////////////////////////////////////////////////////////////////////////
|
|
||||||
// Program main
|
|
||||||
////////////////////////////////////////////////////////////////////////////////
|
|
||||||
int main(int argc, char **argv) {
|
|
||||||
printf("%s starting...\n", sampleName);
|
|
||||||
|
|
||||||
runTest(argc, argv);
|
|
||||||
|
|
||||||
printf("%s completed, returned %s\n", sampleName,
|
|
||||||
testResult ? "OK" : "ERROR!");
|
|
||||||
exit(testResult ? EXIT_SUCCESS : EXIT_FAILURE);
|
|
||||||
}
|
|
||||||
|
|
||||||
void runTest(int argc, char **argv) {
|
|
||||||
int Nblocks = 2;
|
|
||||||
int Nthreads = 32;
|
|
||||||
cudaError_t error;
|
|
||||||
|
|
||||||
#ifndef _WIN32
|
|
||||||
utsname OS_System_Type;
|
|
||||||
uname(&OS_System_Type);
|
|
||||||
|
|
||||||
printf("OS_System_Type.release = %s\n", OS_System_Type.release);
|
|
||||||
|
|
||||||
if (!strcasecmp(OS_System_Type.sysname, "Darwin")) {
|
|
||||||
printf("simpleAssert is not current supported on Mac OSX\n\n");
|
|
||||||
exit(EXIT_SUCCESS);
|
|
||||||
} else {
|
|
||||||
printf("OS Info: <%s>\n\n", OS_System_Type.version);
|
|
||||||
}
|
|
||||||
|
|
||||||
#endif
|
|
||||||
|
|
||||||
// This will pick the best possible CUDA capable device
|
|
||||||
findCudaDevice(argc, (const char **)argv);
|
|
||||||
|
|
||||||
// Kernel configuration, where a one-dimensional
|
|
||||||
// grid and one-dimensional blocks are configured.
|
|
||||||
dim3 dimGrid(Nblocks);
|
|
||||||
dim3 dimBlock(Nthreads);
|
|
||||||
|
|
||||||
printf("Launch kernel to generate assertion failures\n");
|
|
||||||
testKernel<<<dimGrid, dimBlock>>>(60);
|
|
||||||
|
|
||||||
// Synchronize (flushes assert output).
|
|
||||||
printf("\n-- Begin assert output\n\n");
|
|
||||||
error = cudaDeviceSynchronize();
|
|
||||||
printf("\n-- End assert output\n\n");
|
|
||||||
|
|
||||||
// Check for errors and failed asserts in asynchronous kernel launch.
|
|
||||||
if (error == cudaErrorAssert) {
|
|
||||||
printf(
|
|
||||||
"Device assert failed as expected, "
|
|
||||||
"CUDA error message is: %s\n\n",
|
|
||||||
cudaGetErrorString(error));
|
|
||||||
}
|
|
||||||
|
|
||||||
testResult = error == cudaErrorAssert;
|
|
||||||
}
|
|
||||||
@@ -1,10 +0,0 @@
|
|||||||
{
|
|
||||||
"configurations": [
|
|
||||||
{
|
|
||||||
"name": "CUDA C++: Launch",
|
|
||||||
"type": "cuda-gdb",
|
|
||||||
"request": "launch",
|
|
||||||
"program": "${workspaceFolder}/simpleAssert_nvrtc"
|
|
||||||
}
|
|
||||||
]
|
|
||||||
}
|
|
||||||
@@ -1,15 +0,0 @@
|
|||||||
{
|
|
||||||
"version": "2.0.0",
|
|
||||||
"tasks": [
|
|
||||||
{
|
|
||||||
"label": "sample",
|
|
||||||
"type": "shell",
|
|
||||||
"command": "make dbg=1",
|
|
||||||
"problemMatcher": ["$nvcc"],
|
|
||||||
"group": {
|
|
||||||
"kind": "build",
|
|
||||||
"isDefault": true
|
|
||||||
}
|
|
||||||
}
|
|
||||||
]
|
|
||||||
}
|
|
||||||
@@ -1,37 +0,0 @@
|
|||||||
cmake_minimum_required(VERSION 3.20)
|
|
||||||
|
|
||||||
list(APPEND CMAKE_MODULE_PATH "${CMAKE_CURRENT_SOURCE_DIR}/../../../cmake/Modules")
|
|
||||||
|
|
||||||
project(simpleAssert_nvrtc LANGUAGES C CXX CUDA)
|
|
||||||
|
|
||||||
find_package(CUDAToolkit REQUIRED)
|
|
||||||
|
|
||||||
set(CMAKE_POSITION_INDEPENDENT_CODE ON)
|
|
||||||
|
|
||||||
set(CMAKE_CUDA_ARCHITECTURES 50 52 60 61 70 72 75 80 86 87 89 90 100 101 120)
|
|
||||||
set(CMAKE_CUDA_FLAGS "${CMAKE_CUDA_FLAGS} -Wno-deprecated-gpu-targets")
|
|
||||||
if(CMAKE_BUILD_TYPE STREQUAL "Debug")
|
|
||||||
# set(CMAKE_CUDA_FLAGS "${CMAKE_CUDA_FLAGS} -G") # enable cuda-gdb (expensive)
|
|
||||||
endif()
|
|
||||||
|
|
||||||
# Include directories and libraries
|
|
||||||
include_directories(../../../Common)
|
|
||||||
|
|
||||||
# Source file
|
|
||||||
# Add sample target executable
|
|
||||||
add_executable(simpleAssert_nvrtc simpleAssert.cpp)
|
|
||||||
|
|
||||||
target_compile_options(simpleAssert_nvrtc PRIVATE $<$<COMPILE_LANGUAGE:CUDA>:--extended-lambda>)
|
|
||||||
|
|
||||||
target_compile_features(simpleAssert_nvrtc PRIVATE cxx_std_17 cuda_std_17)
|
|
||||||
|
|
||||||
target_link_libraries(simpleAssert_nvrtc PRIVATE
|
|
||||||
CUDA::nvrtc
|
|
||||||
CUDA::cuda_driver
|
|
||||||
)
|
|
||||||
|
|
||||||
# Copy clock_kernel.cu to the output directory
|
|
||||||
add_custom_command(TARGET simpleAssert_nvrtc POST_BUILD
|
|
||||||
COMMAND ${CMAKE_COMMAND} -E copy_if_different
|
|
||||||
${CMAKE_CURRENT_SOURCE_DIR}/simpleAssert_kernel.cu ${CMAKE_CURRENT_BINARY_DIR}
|
|
||||||
)
|
|
||||||
@@ -1,36 +0,0 @@
|
|||||||
# simpleAssert_nvrtc - simpleAssert with libNVRTC
|
|
||||||
|
|
||||||
## Description
|
|
||||||
|
|
||||||
This CUDA Runtime API sample is a very basic sample that implements how to use the assert function in the device code. Requires Compute Capability 2.0 .
|
|
||||||
|
|
||||||
## Key Concepts
|
|
||||||
|
|
||||||
Assert, Runtime Compilation
|
|
||||||
|
|
||||||
## Supported SM Architectures
|
|
||||||
|
|
||||||
[SM 5.0 ](https://developer.nvidia.com/cuda-gpus) [SM 5.2 ](https://developer.nvidia.com/cuda-gpus) [SM 5.3 ](https://developer.nvidia.com/cuda-gpus) [SM 6.0 ](https://developer.nvidia.com/cuda-gpus) [SM 6.1 ](https://developer.nvidia.com/cuda-gpus) [SM 7.0 ](https://developer.nvidia.com/cuda-gpus) [SM 7.2 ](https://developer.nvidia.com/cuda-gpus) [SM 7.5 ](https://developer.nvidia.com/cuda-gpus) [SM 8.0 ](https://developer.nvidia.com/cuda-gpus) [SM 8.6 ](https://developer.nvidia.com/cuda-gpus) [SM 8.7 ](https://developer.nvidia.com/cuda-gpus) [SM 8.9 ](https://developer.nvidia.com/cuda-gpus) [SM 9.0 ](https://developer.nvidia.com/cuda-gpus)
|
|
||||||
|
|
||||||
## Supported OSes
|
|
||||||
|
|
||||||
Linux, Windows, QNX
|
|
||||||
|
|
||||||
## Supported CPU Architecture
|
|
||||||
|
|
||||||
x86_64, aarch64
|
|
||||||
|
|
||||||
## CUDA APIs involved
|
|
||||||
|
|
||||||
### [CUDA Driver API](http://docs.nvidia.com/cuda/cuda-driver-api/index.html)
|
|
||||||
cuModuleGetFunction, cuLaunchKernel, cuCtxSynchronize
|
|
||||||
|
|
||||||
## Dependencies needed to build/run
|
|
||||||
[NVRTC](../../../README.md#nvrtc)
|
|
||||||
|
|
||||||
## Prerequisites
|
|
||||||
|
|
||||||
Download and install the [CUDA Toolkit 12.5](https://developer.nvidia.com/cuda-downloads) for your corresponding platform.
|
|
||||||
Make sure the dependencies mentioned in [Dependencies]() section above are installed.
|
|
||||||
|
|
||||||
## References (for more details)
|
|
||||||
@@ -1,113 +0,0 @@
|
|||||||
/* Copyright (c) 2022, NVIDIA CORPORATION. All rights reserved.
|
|
||||||
*
|
|
||||||
* Redistribution and use in source and binary forms, with or without
|
|
||||||
* modification, are permitted provided that the following conditions
|
|
||||||
* are met:
|
|
||||||
* * Redistributions of source code must retain the above copyright
|
|
||||||
* notice, this list of conditions and the following disclaimer.
|
|
||||||
* * Redistributions in binary form must reproduce the above copyright
|
|
||||||
* notice, this list of conditions and the following disclaimer in the
|
|
||||||
* documentation and/or other materials provided with the distribution.
|
|
||||||
* * Neither the name of NVIDIA CORPORATION nor the names of its
|
|
||||||
* contributors may be used to endorse or promote products derived
|
|
||||||
* from this software without specific prior written permission.
|
|
||||||
*
|
|
||||||
* THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS ``AS IS'' AND ANY
|
|
||||||
* EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
|
|
||||||
* IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR
|
|
||||||
* PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER OR
|
|
||||||
* CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL,
|
|
||||||
* EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO,
|
|
||||||
* PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR
|
|
||||||
* PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY
|
|
||||||
* OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
|
|
||||||
* (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
|
|
||||||
* OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
|
|
||||||
*/
|
|
||||||
|
|
||||||
#ifdef _WIN32
|
|
||||||
#define WINDOWS_LEAN_AND_MEAN
|
|
||||||
#define NOMINMAX
|
|
||||||
#include <windows.h>
|
|
||||||
#else
|
|
||||||
#include <sys/utsname.h>
|
|
||||||
#endif
|
|
||||||
|
|
||||||
// Includes, system
|
|
||||||
#include <stdio.h>
|
|
||||||
#include <cassert>
|
|
||||||
|
|
||||||
// Includes CUDA
|
|
||||||
#include <cuda_runtime.h>
|
|
||||||
#include "nvrtc_helper.h"
|
|
||||||
|
|
||||||
// Utilities and timing functions
|
|
||||||
#include <helper_functions.h> // includes cuda.h and cuda_runtime_api.h
|
|
||||||
|
|
||||||
const char *sampleName = "simpleAssert_nvrtc";
|
|
||||||
|
|
||||||
////////////////////////////////////////////////////////////////////////////////
|
|
||||||
// Auto-Verification Code
|
|
||||||
bool testResult = true;
|
|
||||||
|
|
||||||
////////////////////////////////////////////////////////////////////////////////
|
|
||||||
// Declaration, forward
|
|
||||||
void runTest(int argc, char **argv);
|
|
||||||
|
|
||||||
////////////////////////////////////////////////////////////////////////////////
|
|
||||||
// Program main
|
|
||||||
////////////////////////////////////////////////////////////////////////////////
|
|
||||||
|
|
||||||
int main(int argc, char **argv) {
|
|
||||||
printf("%s starting...\n", sampleName);
|
|
||||||
|
|
||||||
runTest(argc, argv);
|
|
||||||
|
|
||||||
exit(testResult ? EXIT_SUCCESS : EXIT_FAILURE);
|
|
||||||
}
|
|
||||||
|
|
||||||
void runTest(int argc, char **argv) {
|
|
||||||
int Nblocks = 2;
|
|
||||||
int Nthreads = 32;
|
|
||||||
|
|
||||||
// Kernel configuration, where a one-dimensional
|
|
||||||
// grid and one-dimensional blocks are configured.
|
|
||||||
|
|
||||||
dim3 dimGrid(Nblocks);
|
|
||||||
dim3 dimBlock(Nthreads);
|
|
||||||
|
|
||||||
printf("Launch kernel to generate assertion failures\n");
|
|
||||||
char *cubin, *kernel_file;
|
|
||||||
size_t cubinSize;
|
|
||||||
|
|
||||||
kernel_file = sdkFindFilePath("simpleAssert_kernel.cu", argv[0]);
|
|
||||||
compileFileToCUBIN(kernel_file, argc, argv, &cubin, &cubinSize, 0);
|
|
||||||
|
|
||||||
CUmodule module = loadCUBIN(cubin, argc, argv);
|
|
||||||
CUfunction kernel_addr;
|
|
||||||
|
|
||||||
checkCudaErrors(cuModuleGetFunction(&kernel_addr, module, "testKernel"));
|
|
||||||
|
|
||||||
int count = 60;
|
|
||||||
void *args[] = {(void *)&count};
|
|
||||||
|
|
||||||
checkCudaErrors(cuLaunchKernel(
|
|
||||||
kernel_addr, dimGrid.x, dimGrid.y, dimGrid.z, /* grid dim */
|
|
||||||
dimBlock.x, dimBlock.y, dimBlock.z, /* block dim */
|
|
||||||
0, 0, /* shared mem, stream */
|
|
||||||
&args[0], /* arguments */
|
|
||||||
0));
|
|
||||||
|
|
||||||
// Synchronize (flushes assert output).
|
|
||||||
printf("\n-- Begin assert output\n\n");
|
|
||||||
CUresult res = cuCtxSynchronize();
|
|
||||||
|
|
||||||
printf("\n-- End assert output\n\n");
|
|
||||||
|
|
||||||
// Check for errors and failed asserts in asynchronous kernel launch.
|
|
||||||
if (res == CUDA_ERROR_ASSERT) {
|
|
||||||
printf("Device assert failed as expected\n");
|
|
||||||
}
|
|
||||||
|
|
||||||
testResult = res == CUDA_ERROR_ASSERT;
|
|
||||||
}
|
|
||||||
@@ -1,10 +0,0 @@
|
|||||||
{
|
|
||||||
"configurations": [
|
|
||||||
{
|
|
||||||
"name": "CUDA C++: Launch",
|
|
||||||
"type": "cuda-gdb",
|
|
||||||
"request": "launch",
|
|
||||||
"program": "${workspaceFolder}/simpleAtomicIntrinsics"
|
|
||||||
}
|
|
||||||
]
|
|
||||||
}
|
|
||||||
Some files were not shown because too many files have changed in this diff Show More
Reference in New Issue
Block a user