Skip to content

GitLab

Unverified Commit 4e2be38c authored by gilbertlee-amd's avatar gilbertlee-amd Committed by GitHub
Browse files

Update rocm-rel-6.4 to use TransferBench v1.62.00 (#183) 

* updating metadata

* v1.58.00 Fixing DMA copy-on-engine (#152)

* Leo's review

* Update use-transferbench.rst

* Update Doxyfile

* Refining API library

* Update TransferBench.hpp

* Update TransferBench.hpp

* Update TransferBench.hpp

* Bump rocm-docs-core from 1.9.2 to 1.11.0 in /docs/sphinx (#153)

Bumps [rocm-docs-core](https://github.com/ROCm/rocm-docs-core) from 1.9.2 to 1.11.0.
- [Release notes](https://github.com/ROCm/rocm-docs-core/releases)
- [Changelog](https://github.com/ROCm/rocm-docs-core/blob/develop/CHANGELOG.md)
- [Commits](https://github.com/ROCm/rocm-docs-core/compare/v1.9.2...v1.11.0

)

---
updated-dependencies:
- dependency-name: rocm-docs-core
  dependency-type: direct:production
  update-type: version-update:semver-minor
...
Signed-off-by: default avatardependabot[bot] <support@github.com>
Co-authored-by: default avatardependabot[bot] <49699333+dependabot[bot]@users.noreply.github.com>

* Bump rocm-docs-core from 1.11.0 to 1.12.0 in /docs/sphinx (#155)

Bumps [rocm-docs-core](https://github.com/ROCm/rocm-docs-core) from 1.11.0 to 1.12.0.
- [Release notes](https://github.com/ROCm/rocm-docs-core/releases)
- [Changelog](https://github.com/ROCm/rocm-docs-core/blob/develop/CHANGELOG.md)
- [Commits](https://github.com/ROCm/rocm-docs-core/compare/v1.11.0...v1.12.0

)

---
updated-dependencies:
- dependency-name: rocm-docs-core
  dependency-type: direct:production
  update-type: version-update:semver-minor
...
Signed-off-by: default avatardependabot[bot] <support@github.com>
Co-authored-by: default avatardependabot[bot] <49699333+dependabot[bot]@users.noreply.github.com>

* External CI: enable CI triggers

* Apply suggestions from code review
Co-authored-by: default avatarMustafa Abduljabbar <mustafa.abduljabbar@amd.com>

* Update LICENSE.md

* Bump rocm-docs-core from 1.12.0 to 1.13.0 in /docs/sphinx (#160)

Bumps [rocm-docs-core](https://github.com/ROCm/rocm-docs-core) from 1.12.0 to 1.13.0.
- [Release notes](https://github.com/ROCm/rocm-docs-core/releases)
- [Changelog](https://github.com/ROCm/rocm-docs-core/blob/develop/CHANGELOG.md)
- [Commits](https://github.com/ROCm/rocm-docs-core/compare/v1.12.0...v1.13.0

)

---
updated-dependencies:
- dependency-name: rocm-docs-core
  dependency-type: direct:production
  update-type: version-update:semver-minor
...
Signed-off-by: default avatardependabot[bot] <support@github.com>
Co-authored-by: default avatardependabot[bot] <49699333+dependabot[bot]@users.noreply.github.com>

* TransferBench V1.59 (#162)

Adding NIC execution capabilities, various bug fixes introduced by header-only-library refactor
---------
Co-authored-by: default avatarMustafa Abduljabbar <mustafa.abduljabbar@amd.com>

* Adding ability to specify A2A_MODE=numSrcs:numDsts (#164)
Co-authored-by: default avatarMustafa Abduljabbar <mustafa.abduljabbar@amd.com>

* Fixing specific DMA engine transfers, enabling GFX_SINGLE_TEAM=1 by default (#166)

* Bump rocm-docs-core from 1.13.0 to 1.15.0 in /docs/sphinx (#165)

Bumps [rocm-docs-core](https://github.com/ROCm/rocm-docs-core) from 1.13.0 to 1.15.0.
- [Release notes](https://github.com/ROCm/rocm-docs-core/releases)
- [Changelog](https://github.com/ROCm/rocm-docs-core/blob/develop/CHANGELOG.md)
- [Commits](https://github.com/ROCm/rocm-docs-core/compare/v1.13.0...v1.15.0

)

---
updated-dependencies:
- dependency-name: rocm-docs-core
  dependency-type: direct:production
  update-type: version-update:semver-minor
...
Signed-off-by: default avatardependabot[bot] <support@github.com>
Co-authored-by: default avatardependabot[bot] <49699333+dependabot[bot]@users.noreply.github.com>

* Bump rocm-docs-core from 1.15.0 to 1.17.0 in /docs/sphinx (#171)

Bumps [rocm-docs-core](https://github.com/ROCm/rocm-docs-core) from 1.15.0 to 1.17.0.
- [Release notes](https://github.com/ROCm/rocm-docs-core/releases)
- [Changelog](https://github.com/ROCm/rocm-docs-core/blob/develop/CHANGELOG.md)
- [Commits](https://github.com/ROCm/rocm-docs-core/compare/v1.15.0...v1.17.0

)

---
updated-dependencies:
- dependency-name: rocm-docs-core
  dependency-type: direct:production
  update-type: version-update:semver-minor
...
Signed-off-by: default avatardependabot[bot] <support@github.com>
Co-authored-by: default avatardependabot[bot] <49699333+dependabot[bot]@users.noreply.github.com>

* TransferBench v1.61 (#174)
Co-authored-by: default avatarMustafa Abduljabbar <mustafa.abduljabbar@amd.com>

* Bump rocm-docs-core from 1.17.0 to 1.18.1 in /docs/sphinx (#176)

Bumps [rocm-docs-core](https://github.com/ROCm/rocm-docs-core) from 1.17.0 to 1.18.1.
- [Release notes](https://github.com/ROCm/rocm-docs-core/releases)
- [Changelog](https://github.com/ROCm/rocm-docs-core/blob/develop/CHANGELOG.md)
- [Commits](https://github.com/ROCm/rocm-docs-core/compare/v1.17.0...v1.18.1

)

---
updated-dependencies:
- dependency-name: rocm-docs-core
  dependency-type: direct:production
  update-type: version-update:semver-minor
...
Signed-off-by: default avatardependabot[bot] <support@github.com>
Co-authored-by: default avatardependabot[bot] <49699333+dependabot[bot]@users.noreply.github.com>

* Bump rocm-docs-core from 1.18.1 to 1.18.2 in /docs/sphinx (#177)

Bumps [rocm-docs-core](https://github.com/ROCm/rocm-docs-core) from 1.18.1 to 1.18.2.
- [Release notes](https://github.com/ROCm/rocm-docs-core/releases)
- [Changelog](https://github.com/ROCm/rocm-docs-core/blob/develop/CHANGELOG.md)
- [Commits](https://github.com/ROCm/rocm-docs-core/compare/v1.18.1...v1.18.2

)

---
updated-dependencies:
- dependency-name: rocm-docs-core
  dependency-version: 1.18.2
  dependency-type: direct:production
  update-type: version-update:semver-patch
...
Signed-off-by: default avatardependabot[bot] <support@github.com>
Co-authored-by: default avatardependabot[bot] <49699333+dependabot[bot]@users.noreply.github.com>

* TransferBench v1.62.00 (#181)

* Adding non-temporal loads and stores via GFX_TEMPORAL
* Adding additional summary details to a2a preset
* Add SHOW_MIN_ONLY for a2asweep preset
* Adding new P CPU memory type which is indexed by closest GPU

* Bump rocm-docs-core from 1.18.2 to 1.20.0 in /docs/sphinx (#180)

Bumps [rocm-docs-core](https://github.com/ROCm/rocm-docs-core) from 1.18.2 to 1.20.0.
- [Release notes](https://github.com/ROCm/rocm-docs-core/releases)
- [Changelog](https://github.com/ROCm/rocm-docs-core/blob/develop/CHANGELOG.md)
- [Commits](https://github.com/ROCm/rocm-docs-core/compare/v1.18.2...v1.20.0

)

---
updated-dependencies:
- dependency-name: rocm-docs-core
  dependency-version: 1.20.0
  dependency-type: direct:production
  update-type: version-update:semver-minor
...
Signed-off-by: default avatardependabot[bot] <support@github.com>
Co-authored-by: default avatardependabot[bot] <49699333+dependabot[bot]@users.noreply.github.com>

---------
Signed-off-by: default avatardependabot[bot] <support@github.com>
Co-authored-by: default avatarsrawat <120587655+SwRaw@users.noreply.github.com>
Co-authored-by: default avatardependabot[bot] <49699333+dependabot[bot]@users.noreply.github.com>
Co-authored-by: default avatarDaniel Su <danielsu@amd.com>
Co-authored-by: default avatarMustafa Abduljabbar <mustafa.abduljabbar@amd.com>
parent 3ea2f226
Show whitespace changes
Inline Side-by-side
Showing with 939 additions and 77 deletions
.azuredevops/rocm-ci.yml 0 → 100644
View file @ 4e2be38c
resources:
repositories:
- repository: pipelines_repo
type: github
endpoint: ROCm
name: ROCm/ROCm
variables:
- group: common
- template: /.azuredevops/variables-global.yml@pipelines_repo
trigger:
batch: true
branches:
include:
- develop
- mainline
paths:
exclude:
- .github
- docs
- '.*.yaml'
- '.*.yml'
- '*.md'
- LICENSE
pr:
autoCancel: true
branches:
include:
- develop
- mainline
paths:
exclude:
- .github
- docs
- '.*.yaml'
- '.*.yml'
- '*.md'
- LICENSE
drafts: false
jobs:
- template: ${{ variables.CI_COMPONENT_PATH }}/TransferBench.yml@pipelines_repo
CHANGELOG.md
View file @ 4e2be38c
......@@ -3,6 +3,72 @@
Documentation for TransferBench is available at
[https://rocm.docs.amd.com/projects/TransferBench](https://rocm.docs.amd.com/projects/TransferBench).
## v1.62.00
### Added
- Adding GFX_TEMPORAL to allow for use for use of non-temporal loads/stores
- (0 = none [default], 1 = load, 2 = store, 3 = both)
- Addding "P" memory type which maps to CPU memory but is indexed by closest GPU
- For example, P4 refers to CPU memory on NUMA node closest to GPU 4
### Modified
- Adding some additional summary details to a2a preset
## v1.61.00
### Added
- Added a2a_n preset which conducts alltoall GPU-to-GPU tranfers over nearest NIC executors
- Re-implemented GFX_BLOCK_ORDER which allows for control over how threadblocks of multiple transfers are ordered
- 0 = sequential, 1 = interleaved, 2 = random
- Added a2asweep preset which tries various CU/unroll options for GFX-executed all-to-all
- Rewrite main GID index detection logic
- Show the GID index and description in the topology table. It is helpful for debugging purposes
- Added GFX_WORD_SIZE to allow for different packed float sizes to use for GFX kernel. Must be either 4 (default), 2 or 1
### Fixed
- Avoid build errors for CMake and Makefile if infiniband/verbs.h header is not present and disable NIC executor in such case
- Have a priority list of which GID entry to go for instead of hardcoding choices based on underdocumented user input (such as RoCE version and IP address family)
- Use link-local when it is the only choice (i.e. when routing information is not available beyond local link)
## v1.60.00
### Modified
- Reverted GFX_SINGLE_TEAM default back to 1
### Fixed
- Fixed bug where peer memory access was not enabled for DMA transfers, which would break specific DMA engine transfers
## v1.59.01
### Added
- The a2a preset A2A_MODE variable has been enhanced to allow for customizing the number of srcs/dsts to use
This is specified by setting A2A_MODE to numSrcs:numDsts. Extra destinations past 1 will be "local" writes (i.e. if one sets A2A_MODE=1:3, then transfers will follow this pattern: Fx Gx FyFxFx)
to simulate similar conditions normally used during collective algorithms such as ring-based AllReduce
## v1.59.00
### Added
- Adding in support for NIC executor, which allows for RDMA copies on NICs that support IBVerbs
By default, NIC executor will be enabled if IBVerbs is found in the dynamic linker cache
- NIC executor can be indexed in two methods
- "I" Ix.y will use NIC x as the source and NIC y as the destination.
E.g. (G0 I0.5 G4)
- "N" Nx.y will use NIC closest to GPU x as source, and NIC closest to GPU y as destination
E.g. (G0 N0.4 N4)
- The closest NIC can be overridden by the environment variable CLOSEST_NIC, which should be a comma-separated
list of NIC indices to use for the corresponding GPU
- This feature can be explicitly disabled at compile time by specifying DISABLE_NIC_EXEC=1
### Modified
- Changing default data size to 256M from 64M
- Adding NUM_QUEUE_PAIRS which enables NIC traffic in A2A. Each GPU will talk to the next GPU via the closest NIC
- Sweep preset now saves last sweep run configuration to /tmp/lastSweep.cfg and can be changed via SWEEP_FILE
### Fixed
- Fixed bug with reporting when using subiterations
- Fixed bug with per-Transfer data size specification
- Fixed bug when using XCC prefered table
## v1.58.00
### Fixed
- Fixed broken specific DMA-engine copies
## v1.57.01
### Added
- Re-added "scaling" GPU GFX preset benchmark, which tests copies from GPU to other devices using varying
......
CMakeLists.txt
View file @ 4e2be38c
# Copyright (c) 2023-2024 Advanced Micro Devices, Inc. All rights reserved.
# Copyright (c) 2023-2025 Advanced Micro Devices, Inc. All rights reserved.
if (DEFINED ENV{ROCM_PATH})
set(ROCM_PATH "$ENV{ROCM_PATH}" CACHE STRING "ROCm install directory")
......@@ -7,7 +7,7 @@ else()
endif()
cmake_minimum_required(VERSION 3.5)
project(TransferBench VERSION 1.57.0 LANGUAGES CXX)
project(TransferBench VERSION 1.62.00 LANGUAGES CXX)
# Default GPU architectures to build
#==================================================================================================
......@@ -56,6 +56,23 @@ set( CMAKE_CXX_FLAGS "${flags_str} ${CMAKE_CXX_FLAGS}")
set( CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -O3 -L${ROCM_PATH}/lib")
include_directories(${ROCM_PATH}/include)
find_library(IBVERBS_LIBRARY ibverbs)
find_path(IBVERBS_INCLUDE_DIR infiniband/verbs.h)
if (DEFINED ENV{DISABLE_NIC_EXEC})
message(STATUS "Disabling NIC Executor support")
elseif(IBVERBS_LIBRARY AND IBVERBS_INCLUDE_DIR)
message(STATUS "Found ibverbs: ${IBVERBS_LIBRARY}. Building with NIC executor support. Can set DISABLE_NIC_EXEC=1 to disable")
add_definitions(-DNIC_EXEC_ENABLED)
link_libraries(ibverbs)
else()
if (NOT IBVERBS_LIBRARY)
message(WARNING "IBVerbs library not found")
elseif (NOT IBVERBS_INCLUDE_DIR)
message(WARNING "infiniband/verbs.h not found")
endif()
message(WARNING "Building without NIC executor support. To use the TransferBench RDMA executor, check if your system has NICs, the NIC drivers are installed, and libibverbs-dev is installed")
endif()
link_libraries(numa hsa-runtime64 pthread)
set (CMAKE_RUNTIME_OUTPUT_DIRECTORY .)
add_executable(TransferBench src/client/Client.cpp)
......
LICENSE.md
View file @ 4e2be38c
Copyright (c) 2019-2024 Advanced Micro Devices, Inc. All rights reserved.
Copyright (c) 2019-2025 Advanced Micro Devices, Inc. All rights reserved.
Permission is hereby granted, free of charge, to any person obtaining a copy
of this software and associated documentation files (the "Software"), to deal
......
Makefile
View file @ 4e2be38c
#
# Copyright (c) 2019-2024 Advanced Micro Devices, Inc. All rights reserved.
# Copyright (c) 2019-2025 Advanced Micro Devices, Inc. All rights reserved.
#
# Configuration options
......@@ -12,8 +12,10 @@ NVCC=$(CUDA_PATH)/bin/nvcc
# Compile TransferBenchCuda if nvcc detected
ifeq ("$(shell test -e $(NVCC) && echo found)", "found")
EXE=TransferBenchCuda
CXX=$(NVCC)
else
EXE=TransferBench
CXX=$(HIPCC)
endif
CXXFLAGS = -I$(ROCM_PATH)/include -lnuma -L$(ROCM_PATH)/lib -lhsa-runtime64
......@@ -21,13 +23,40 @@ NVFLAGS = -x cu -lnuma -arch=native
COMMON_FLAGS = -O3 -I./src/header -I./src/client -I./src/client/Presets
LDFLAGS += -lpthread
# Compile RDMA executor if
# 1) DISABLE_NIC_EXEC is not set to 1
# 2) IBVerbs is found in the Dynamic Linker cache
# 3) infiniband/verbs.h is found in the default include path
NIC_ENABLED = 0
ifneq ($(DISABLE_NIC_EXEC),1)
ifeq ("$(shell ldconfig -p | grep -c ibverbs)", "0")
$(info lib IBVerbs not found)
else ifeq ("$(shell echo '#include <infiniband/verbs.h>' | $(CXX) -E - 2>/dev/null | grep -c 'infiniband/verbs.h')", "0")
$(info infiniband/verbs.h not found)
else
LDFLAGS += -libverbs -DNIC_EXEC_ENABLED
NVFLAGS += -libverbs -DNIC_EXEC_ENABLED
NIC_ENABLED = 1
endif
ifeq ($(NIC_ENABLED), 0)
$(info To use the TransferBench RDMA executor, check if your system has NICs, the NIC drivers are installed, and libibverbs-dev is installed)
endif
endif
all: $(EXE)
TransferBench: ./src/client/Client.cpp $(shell find -regex ".*\.\hpp")
TransferBench: ./src/client/Client.cpp $(shell find -regex ".*\.\hpp") NicStatus
$(HIPCC) $(CXXFLAGS) $(COMMON_FLAGS) $< -o $@ $(LDFLAGS)
TransferBenchCuda: ./src/client/Client.cpp $(shell find -regex ".*\.\hpp")
TransferBenchCuda: ./src/client/Client.cpp $(shell find -regex ".*\.\hpp") NicStatus
$(NVCC) $(NVFLAGS) $(COMMON_FLAGS) $< -o $@ $(LDFLAGS)
clean:
rm -f *.o ./TransferBench ./TransferBenchCuda
NicStatus:
ifeq ($(NIC_ENABLED), 1)
$(info Building with NIC executor support. Can set DISABLE_NIC_EXEC=1 to disable)
else
$(info Building without NIC executor support)
endif
docs/api.rst deleted 100644 → 0
View file @ 3ea2f226
-----
API
-----
.. doxygenindex::
docs/doxygen/Doxyfile
View file @ 4e2be38c
......@@ -775,7 +775,7 @@ WARN_LOGFILE =
# spaces. See also FILE_PATTERNS and EXTENSION_MAPPING
# Note: If this tag is empty the current directory is searched.
INPUT = ../../src/ \
INPUT = ../../src/header \
../../src/include
# This tag can be used to specify the character encoding of the source files
......
docs/how to/use-transferbench.rst
View file @ 4e2be38c
.. meta::
:description: TransferBench is a utility to benchmark simultaneous transfers between user-specified devices (CPUs or GPUs)
:keywords: Using TransferBench, TransferBench Usage, TransferBench How To, API, ROCm, documentation, HIP
:keywords: TransferBench usage, TransferBench how to, TransferBench user guide, TransferBench user manual
.. _using-transferbench:
......
docs/index.rst
View file @ 4e2be38c
.. meta::
:description: TransferBench is a utility to benchmark simultaneous transfers between user-specified devices (CPUs or GPUs)
:keywords: TransferBench, API, ROCm, documentation, HIP
:keywords: Benchmarking utility, Memory transfers, Device transfers
****************************
TransferBench documentation
......
docs/install/install.rst
View file @ 4e2be38c
.. meta::
:description: TransferBench is a utility to benchmark simultaneous transfers between user-specified devices (CPUs or GPUs)
:keywords: Build TransferBench, Install TransferBench, API, ROCm, HIP
:keywords: Build TransferBench, Install TransferBench
.. _install-transferbench:
......
docs/reference/api.rst
View file @ 4e2be38c
.. meta::
:description: TransferBench is a utility to benchmark simultaneous transfers between user-specified devices (CPUs or GPUs)
:keywords: TransferBench API, TransferBench library, documentation, HIP
:keywords: TransferBench library, TransferBench functions, Transferbench API, Transferbench interface
.. _transferbench-api:
......
docs/sphinx/requirements.in
View file @ 4e2be38c
rocm-docs-core==1.9.2
rocm-docs-core==1.20.0
docs/sphinx/requirements.txt
View file @ 4e2be38c
......@@ -8,6 +8,13 @@ accessible-pygments==0.0.5
# via pydata-sphinx-theme
alabaster==0.7.16
# via sphinx
asttokens==3.0.0
# via stack-data
attrs==25.1.0
# via
# jsonschema
# jupyter-cache
# referencing
babel==2.15.0
# via
# pydata-sphinx-theme
......@@ -25,9 +32,17 @@ cffi==1.16.0
charset-normalizer==3.3.2
# via requests
click==8.1.7
# via sphinx-external-toc
# via
# jupyter-cache
# sphinx-external-toc
comm==0.2.2
# via ipykernel
cryptography==42.0.7
# via pyjwt
debugpy==1.8.12
# via ipykernel
decorator==5.1.1
# via ipython
deprecated==1.2.14
# via pygithub
docutils==0.21.2
......@@ -36,36 +51,104 @@ docutils==0.21.2
# myst-parser
# pydata-sphinx-theme
# sphinx
exceptiongroup==1.2.2
# via ipython
executing==2.2.0
# via stack-data
fastjsonschema==2.19.1
# via rocm-docs-core
# via
# nbformat
# rocm-docs-core
gitdb==4.0.11
# via gitpython
gitpython==3.1.43
# via rocm-docs-core
greenlet==3.1.1
# via sqlalchemy
idna==3.7
# via requests
imagesize==1.4.1
# via sphinx
importlib-metadata==8.6.1
# via
# jupyter-cache
# myst-nb
ipykernel==6.29.5
# via myst-nb
ipython==8.31.0
# via
# ipykernel
# myst-nb
jedi==0.19.2
# via ipython
jinja2==3.1.4
# via
# myst-parser
# sphinx
jsonschema==4.23.0
# via nbformat
jsonschema-specifications==2024.10.1
# via jsonschema
jupyter-cache==1.0.1
# via myst-nb
jupyter-client==8.6.3
# via
# ipykernel
# nbclient
jupyter-core==5.7.2
# via
# ipykernel
# jupyter-client
# nbclient
# nbformat
markdown-it-py==3.0.0
# via
# mdit-py-plugins
# myst-parser
markupsafe==2.1.5
# via jinja2
matplotlib-inline==0.1.7
# via
# ipykernel
# ipython
mdit-py-plugins==0.4.1
# via myst-parser
mdurl==0.1.2
# via markdown-it-py
myst-parser==3.0.1
myst-nb==1.1.2
# via rocm-docs-core
myst-parser==3.0.1
# via myst-nb
nbclient==0.10.2
# via
# jupyter-cache
# myst-nb
nbformat==5.10.4
# via
# jupyter-cache
# myst-nb
# nbclient
nest-asyncio==1.6.0
# via ipykernel
packaging==24.0
# via
# ipykernel
# pydata-sphinx-theme
# sphinx
parso==0.8.4
# via jedi
pexpect==4.9.0
# via ipython
platformdirs==4.3.6
# via jupyter-core
prompt-toolkit==3.0.50
# via ipython
psutil==6.1.1
# via ipykernel
ptyprocess==0.7.0
# via pexpect
pure-eval==0.2.3
# via stack-data
pycparser==2.22
# via cffi
pydata-sphinx-theme==0.15.2
......@@ -77,23 +160,42 @@ pygithub==2.3.0
pygments==2.18.0
# via
# accessible-pygments
# ipython
# pydata-sphinx-theme
# sphinx
pyjwt[crypto]==2.8.0
# via pygithub
pynacl==1.5.0
# via pygithub
python-dateutil==2.9.0.post0
# via jupyter-client
pyyaml==6.0.1
# via
# jupyter-cache
# myst-nb
# myst-parser
# rocm-docs-core
# sphinx-external-toc
pyzmq==26.2.1
# via
# ipykernel
# jupyter-client
referencing==0.36.2
# via
# jsonschema
# jsonschema-specifications
requests==2.32.2
# via
# pygithub
# sphinx
rocm-docs-core==1.9.2
rocm-docs-core==1.20.0
# via -r requirements.in
rpds-py==0.22.3
# via
# jsonschema
# referencing
six==1.17.0
# via python-dateutil
smmap==5.0.1
# via gitdb
snowballstemmer==2.2.0
......@@ -103,6 +205,7 @@ soupsieve==2.5
sphinx==7.3.7
# via
# breathe
# myst-nb
# myst-parser
# pydata-sphinx-theme
# rocm-docs-core
......@@ -133,15 +236,43 @@ sphinxcontrib-qthelp==1.0.7
# via sphinx
sphinxcontrib-serializinghtml==1.1.10
# via sphinx
sqlalchemy==2.0.37
# via jupyter-cache
stack-data==0.6.3
# via ipython
tabulate==0.9.0
# via jupyter-cache
tomli==2.0.1
# via sphinx
tornado==6.4.2
# via
# ipykernel
# jupyter-client
traitlets==5.14.3
# via
# comm
# ipykernel
# ipython
# jupyter-client
# jupyter-core
# matplotlib-inline
# nbclient
# nbformat
typing-extensions==4.11.0
# via
# ipython
# myst-nb
# pydata-sphinx-theme
# pygithub
# referencing
# sqlalchemy
urllib3==2.2.1
# via
# pygithub
# requests
wcwidth==0.2.13
# via prompt-toolkit
wrapt==1.16.0
# via deprecated
zipp==3.21.0
# via importlib-metadata
examples/example.cfg
View file @ 4e2be38c
......@@ -13,6 +13,7 @@
# 1) CPU CPU thread
# 2) GPU GPU threadblock/Compute Unit (CU)
# 3) DMA N/A. (May only be used for copies (single SRC/DST)
# 4) NIC Queue Pair
# Each single line in the configuration file defines a set of Transfers (a Test) to run in parallel
......@@ -37,6 +38,8 @@
# - C: CPU-executed (Indexed from 0 to # NUMA nodes - 1)
# - G: GPU-executed (Indexed from 0 to # GPUs - 1)
# - D: DMA-executor (Indexed from 0 to # GPUs - 1)
# - I#.#: NIC executor (Indexed from 0 to # NICs - 1)
# - N#.#: Nearest NIC executor (Indexed from 0 to # GPUs - 1)
# dstMemL : Destination memory locations (Where the data is to be written to)
# bytesL : Number of bytes to copy (0 means use command-line specified size)
# Must be a multiple of 4 and may be suffixed with ('K','M', or 'G')
......@@ -50,13 +53,17 @@
# - G: Global device memory (on GPU device indexed from 0 to [# GPUs - 1])
# - F: Fine-grain device memory (on GPU device indexed from 0 to [# GPUs - 1])
# - N: Null memory (index ignored)
# - P: Pinned host memory (on NUMA node, but indexed by closest GPU [#GPUs -1])
# Examples:
# 1 4 (G0->G0->G1) Uses 4 CUs on GPU0 to copy from GPU0 to GPU1
# 1 4 (C1->G2->G0) Uses 4 CUs on GPU2 to copy from CPU1 to GPU0
# 2 4 G0->G0->G1 G1->G1->G0 Copes from GPU0 to GPU1, and GPU1 to GPU0, each with 4 SEs
# -2 (G0 G0 G1 4 1M) (G1 G1 G0 2 2M) Copies 1Mb from GPU0 to GPU1 with 4 SEs, and 2Mb from GPU1 to GPU0 with 2 SEs
# 1 2 (F0->I0.2->F1) Uses 2 QPs to transfer data from GPU0 via NIC0 to GPU1 via NIC2
# 1 1 (F0->N0.1->F1) Uses 1 QP to transfer data from GPU0 via GPU0's closest NIC to GPU1 via GPU1's closest NIC
# -2 (G0->N0.1->G1 2 128M) (G1->N1.0->G0 1 256M) Uses Nearest NIC executor to copy 128Mb from GPU0 to GPU1 with 2 QPs,
# and 256Mb from GPU1 to GPU0 with 1 QP
# Round brackets and arrows' ->' may be included for human clarity, but will be ignored and are unnecessary
# Lines starting with # will be ignored. Lines starting with ## will be echoed to output
......
src/client/Client.cpp
View file @ 4e2be38c
......@@ -121,13 +121,23 @@ int main(int argc, char **argv) {
}
}
// Track which transfers have already numBytes specified
std::vector<bool> bytesSpecified(transfers.size());
int hasUnspecified = false;
for (int i = 0; i < transfers.size(); i++) {
bytesSpecified[i] = (transfers[i].numBytes != 0);
if (transfers[i].numBytes == 0) hasUnspecified = true;
}
// Run the specified numbers of bytes otherwise generate a range of values
for (size_t bytes = (1<<10); bytes <= (1<<29); bytes *= 2) {
size_t deltaBytes = std::max(1UL, bytes / ev.samplingFactor);
size_t currBytes = (numBytesPerTransfer == 0) ? bytes : numBytesPerTransfer;
do {
for (auto& t : transfers)
t.numBytes = currBytes;
for (int i = 0; i < transfers.size(); i++) {
if (!bytesSpecified[i])
transfers[i].numBytes = currBytes;
}
if (maxVarCount == 0) {
if (TransferBench::RunTransfers(cfgOptions, transfers, results)) {
......@@ -162,17 +172,21 @@ int main(int argc, char **argv) {
PrintResults(ev, ++testNum, bestTransfers, bestResults);
PrintErrors(bestResults.errResults);
}
if (numBytesPerTransfer != 0) break;
if (numBytesPerTransfer != 0 || !hasUnspecified) break;
currBytes += deltaBytes;
} while (currBytes < bytes * 2);
if (numBytesPerTransfer != 0) break;
if (numBytesPerTransfer != 0 || !hasUnspecified) break;
}
}
}
void DisplayUsage(char const* cmdName)
{
printf("TransferBench v%s.%s\n", TransferBench::VERSION, CLIENT_VERSION);
std::string nicSupport = "";
#if NIC_EXEC_ENABLED
nicSupport = " (with NIC support)";
#endif
printf("TransferBench v%s.%s%s\n", TransferBench::VERSION, CLIENT_VERSION, nicSupport.c_str());
printf("========================================\n");
if (numa_available() == -1) {
......@@ -218,7 +232,7 @@ void PrintResults(EnvVars const& ev, int const testNum,
ExeType const exeType = exeDevice.exeType;
int32_t const exeIndex = exeDevice.exeIndex;
printf(" Executor: %3s %02d %c %7.3f GB/s %c %8.3f ms %c %12lu bytes %c %-7.3f GB/s (sum)\n",
printf(" Executor: %3s %02d %c %8.3f GB/s %c %8.3f ms %c %12lu bytes %c %-7.3f GB/s (sum)\n",
ExeTypeName[exeType], exeIndex, sep, exeResult.avgBandwidthGbPerSec, sep,
exeResult.avgDurationMsec, sep, exeResult.numBytes, sep, exeResult.sumBandwidthGbPerSec);
......@@ -230,14 +244,15 @@ void PrintResults(EnvVars const& ev, int const testNum,
char exeSubIndexStr[32] = "";
if (t.exeSubIndex != -1)
sprintf(exeSubIndexStr, ".%d", t.exeSubIndex);
printf(" Transfer %02d %c %7.3f GB/s %c %8.3f ms %c %12lu bytes %c %s -> %s%02d%s:%03d -> %s\n",
printf(" Transfer %02d %c %8.3f GB/s %c %8.3f ms %c %12lu bytes %c %s -> %c%03d%s:%03d -> %s\n",
idx, sep,
r.avgBandwidthGbPerSec, sep,
r.avgDurationMsec, sep,
r.numBytes, sep,
MemDevicesToStr(t.srcs).c_str(), ExeTypeName[exeType], exeIndex,
exeSubIndexStr, t.numSubExecs, MemDevicesToStr(t.dsts).c_str());
MemDevicesToStr(t.srcs).c_str(),
TransferBench::ExeTypeStr[t.exeDevice.exeType], t.exeDevice.exeIndex,
exeSubIndexStr, t.numSubExecs,
MemDevicesToStr(t.dsts).c_str());
// Show per-iteration timing information
if (ev.showIterations) {
......@@ -269,7 +284,7 @@ void PrintResults(EnvVars const& ev, int const testNum,
for (auto& time : times) {
double iterDurationMsec = time.first;
double iterBandwidthGbs = (t.numBytes / 1.0E9) / iterDurationMsec * 1000.0f;
printf(" Iter %03d %c %7.3f GB/s %c %8.3f ms %c", time.second, sep, iterBandwidthGbs, sep, iterDurationMsec, sep);
printf(" Iter %03d %c %8.3f GB/s %c %8.3f ms %c", time.second, sep, iterBandwidthGbs, sep, iterDurationMsec, sep);
std::set<int> usedXccs;
if (time.second - 1 < r.perIterCUs.size()) {
......@@ -285,11 +300,11 @@ void PrintResults(EnvVars const& ev, int const testNum,
printf(" %02d", x);
printf("\n");
}
printf(" StandardDev %c %7.3f GB/s %c %8.3f ms %c\n", sep, stdDevBw, sep, stdDevTime, sep);
printf(" StandardDev %c %8.3f GB/s %c %8.3f ms %c\n", sep, stdDevBw, sep, stdDevTime, sep);
}
}
}
printf(" Aggregate (CPU) %c %7.3f GB/s %c %8.3f ms %c %12lu bytes %c Overhead: %.3f ms\n",
printf(" Aggregate (CPU) %c %8.3f GB/s %c %8.3f ms %c %12lu bytes %c Overhead: %.3f ms\n",
sep, results.avgTotalBandwidthGbPerSec,
sep, results.avgTotalDurationMsec,
sep, results.totalBytesTransferred,
......
src/client/Client.hpp
View file @ 4e2be38c
......@@ -23,14 +23,14 @@ THE SOFTWARE.
#pragma once
// TransferBench client version
#define CLIENT_VERSION "01"
#define CLIENT_VERSION "00"
#include "TransferBench.hpp"
#include "EnvVars.hpp"
size_t const DEFAULT_BYTES_PER_TRANSFER = (1<<26);
size_t const DEFAULT_BYTES_PER_TRANSFER = (1<<28);
char const ExeTypeName[4][4] = {"CPU", "GPU", "DMA", "IBV"};
char const ExeTypeName[5][4] = {"CPU", "GPU", "DMA", "NIC", "NIC"};
// Display detected hardware
void DisplayTopology(bool outputToCsv);
......
src/client/EnvVars.hpp
View file @ 4e2be38c
/*
Copyright (c) 2021-2024 Advanced Micro Devices, Inc. All rights reserved.
Copyright (c) 2021-2025 Advanced Micro Devices, Inc. All rights reserved.
Permission is hereby granted, free of charge, to any person obtaining a copy
of this software and associated documentation files (the "Software"), to deal
......@@ -84,14 +84,17 @@ public:
int useHsaDma; // Use hsa_amd_async_copy instead of hipMemcpy for non-targetted DMA executions
// GFX options
int gfxBlockOrder; // How threadblocks for multiple Transfers are ordered 0=sequential 1=interleaved
int gfxBlockSize; // Size of each threadblock (must be multiple of 64)
vector<uint32_t> cuMask; // Bit-vector representing the CU mask
vector<vector<int>> prefXccTable; // Specifies XCC to use for given exe->dst pair
int gfxTemporal; // Non-temporal load/store mode (0=none, 1=load, 2=store, 3=both)
int gfxUnroll; // GFX-kernel unroll factor
int useHipEvents; // Use HIP events for timing GFX/DMA Executor
int useSingleStream; // Use a single stream per GPU GFX executor instead of stream per Transfer
int gfxSingleTeam; // Team all subExecutors across the data array
int gfxWaveOrder; // GFX-kernel wavefront ordering
int gfxWordSize; // GFX-kernel packed data size (4=DWORDx4, 2=DWORDx2, 1=DWORDx1)
// Client options
int hideEnv; // Skip printing environment variable
......@@ -100,6 +103,14 @@ public:
int outputToCsv; // Output in CSV format
int samplingFactor; // Affects how many different values of N are generated (when N set to 0)
// NIC options
int ibGidIndex; // GID Index for RoCE NICs
int roceVersion; // RoCE version number
int ipAddressFamily; // IP Address Famliy
uint8_t ibPort; // NIC port number to be used
int nicRelaxedOrder; // Use relaxed ordering for RDMA
std::string closestNicStr; // Holds the user-specified list of closest NICs
// Developer features
int gpuMaxHwQueues; // Tracks GPU_MAX_HW_QUEUES environment variable
......@@ -127,10 +138,13 @@ public:
alwaysValidate = GetEnvVar("ALWAYS_VALIDATE" , 0);
blockBytes = GetEnvVar("BLOCK_BYTES" , 256);
byteOffset = GetEnvVar("BYTE_OFFSET" , 0);
gfxBlockOrder = GetEnvVar("GFX_BLOCK_ORDER" , 0);
gfxBlockSize = GetEnvVar("GFX_BLOCK_SIZE" , 256);
gfxSingleTeam = GetEnvVar("GFX_SINGLE_TEAM" , 0);
gfxSingleTeam = GetEnvVar("GFX_SINGLE_TEAM" , 1);
gfxTemporal = GetEnvVar("GFX_TEMPORAL" , 0);
gfxUnroll = GetEnvVar("GFX_UNROLL" , defaultGfxUnroll);
gfxWaveOrder = GetEnvVar("GFX_WAVE_ORDER" , 0);
gfxWordSize = GetEnvVar("GFX_WORD_SIZE" , 4);
hideEnv = GetEnvVar("HIDE_ENV" , 0);
minNumVarSubExec = GetEnvVar("MIN_VAR_SUBEXEC" , 1);
maxNumVarSubExec = GetEnvVar("MAX_VAR_SUBEXEC" , 0);
......@@ -147,8 +161,16 @@ public:
validateDirect = GetEnvVar("VALIDATE_DIRECT" , 0);
validateSource = GetEnvVar("VALIDATE_SOURCE" , 0);
ibGidIndex = GetEnvVar("IB_GID_INDEX" ,-1);
ibPort = GetEnvVar("IB_PORT_NUMBER" , 1);
roceVersion = GetEnvVar("ROCE_VERSION" , 2);
ipAddressFamily = GetEnvVar("IP_ADDRESS_FAMILY" , 4);
nicRelaxedOrder = GetEnvVar("NIC_RELAX_ORDER" , 1);
closestNicStr = GetEnvVar("CLOSEST_NIC" , "");
gpuMaxHwQueues = GetEnvVar("GPU_MAX_HW_QUEUES" , 4);
// Check for fill pattern
char* pattern = getenv("FILL_PATTERN");
if (pattern != NULL) {
......@@ -270,27 +292,55 @@ public:
}
}
static std::string ToStr(std::vector<int> const& values) {
std::string result = "";
bool isFirst = true;
for (int v : values) {
if (isFirst) isFirst = false;
else result += ",";
result += std::to_string(v);
}
return result;
}
// Display info on the env vars that can be used
static void DisplayUsage()
{
printf("Environment variables:\n");
printf("======================\n");
printf(" ALWAYS_VALIDATE - Validate after each iteration instead of once after all iterations\n");
printf(" BLOCK_SIZE - # of threads per threadblock (Must be multiple of 64)\n");
printf(" BLOCK_BYTES - Controls granularity of how work is divided across subExecutors\n");
printf(" BYTE_OFFSET - Initial byte-offset for memory allocations. Must be multiple of 4\n");
#if NIC_EXEC_ENABLED
printf(" CLOSEST_NIC - Comma-separated list of per-GPU closest NIC (default=auto)\n");
#endif
printf(" CU_MASK - CU mask for streams. Can specify ranges e.g '5,10-12,14'\n");
printf(" FILL_PATTERN - Big-endian pattern for source data, specified in hex digits. Must be even # of digits\n");
printf(" GFX_BLOCK_ORDER - How blocks for transfers are ordered. 0=sequential, 1=interleaved\n");
printf(" GFX_BLOCK_SIZE - # of threads per threadblock (Must be multiple of 64)\n");
printf(" GFX_TEMPORAL - Use of non-temporal loads or stores (0=none 1=loads 2=stores 3=both)\n");
printf(" GFX_UNROLL - Unroll factor for GFX kernel (0=auto), must be less than %d\n", TransferBench::GetIntAttribute(ATR_GFX_MAX_UNROLL));
printf(" GFX_SINGLE_TEAM - Have subexecutors work together on full array instead of working on disjoint subarrays\n");
printf(" GFX_WAVE_ORDER - Stride pattern for GFX kernel (0=UWC,1=UCW,2=WUC,3=WCU,4=CUW,5=CWU)\n");
printf(" GFX_WORD_SIZE - GFX kernel packed data size (4=DWORDx4, 2=DWORDx2, 1=DWORDx1)\n");
printf(" HIDE_ENV - Hide environment variable value listing\n");
#if NIC_EXEC_ENABLED
printf(" IB_GID_INDEX - Required for RoCE NICs (default=-1/auto)\n");
printf(" IB_PORT_NUMBER - RDMA port count for RDMA NIC (default=1)\n");
printf(" IP_ADDRESS_FAMILY - IP address family (4=v4, 6=v6, default=v4)\n");
#endif
printf(" MIN_VAR_SUBEXEC - Minumum # of subexecutors to use for variable subExec Transfers\n");
printf(" MAX_VAR_SUBEXEC - Maximum # of subexecutors to use for variable subExec Transfers (0 for device limits)\n");
#if NIC_EXEC_ENABLED
printf(" NIC_RELAX_ORDER - Set to non-zero to use relaxed ordering");
#endif
printf(" NUM_ITERATIONS - # of timed iterations per test. If negative, run for this many seconds instead\n");
printf(" NUM_SUBITERATIONS - # of sub-iterations to run per iteration. Must be non-negative\n");
printf(" NUM_WARMUPS - # of untimed warmup iterations per test\n");
printf(" OUTPUT_TO_CSV - Outputs to CSV format if set\n");
#if NIC_EXEC_ENABLED
printf(" ROCE_VERSION - RoCE version (default=2)\n");
#endif
printf(" SAMPLING_FACTOR - Add this many samples (when possible) between powers of 2 when auto-generating data sizes\n");
printf(" SHOW_ITERATIONS - Show per-iteration timing info\n");
printf(" USE_HIP_EVENTS - Use HIP events for GFX executor timing\n");
......@@ -301,6 +351,7 @@ public:
printf(" VALIDATE_SOURCE - Validate GPU src memory immediately after preparation\n");
}
void Print(std::string const& name, int32_t const value, const char* format, ...) const
{
printf("%-20s%s%12d%s", name.c_str(), outputToCsv ? "," : " = ", value, outputToCsv ? "," : " : ");
......@@ -325,9 +376,12 @@ public:
void DisplayEnvVars() const
{
int numGpuDevices = TransferBench::GetNumExecutors(EXE_GPU_GFX);
std::string nicSupport = "";
#if NIC_EXEC_ENABLED
nicSupport = " (with NIC support)";
#endif
if (!outputToCsv) {
printf("TransferBench Client v%s Backend v%s\n", CLIENT_VERSION, TransferBench::VERSION);
printf("TransferBench v%s.%s%s\n", TransferBench::VERSION, CLIENT_VERSION, nicSupport.c_str());
printf("===============================================================\n");
if (!hideEnv) printf("[Common] (Suppress by setting HIDE_ENV=1)\n");
}
......@@ -341,15 +395,27 @@ public:
"Each CU gets a mulitple of %d bytes to copy", blockBytes);
Print("BYTE_OFFSET", byteOffset,
"Using byte offset of %d", byteOffset);
#if NIC_EXEC_ENABLED
Print("CLOSEST_NIC", (closestNicStr == "" ? "auto" : "user-input"),
"Per-GPU closest NIC is set as %s", (closestNicStr == "" ? "auto" : closestNicStr.c_str()));
#endif
Print("CU_MASK", getenv("CU_MASK") ? 1 : 0,
"%s", (cuMask.size() ? GetCuMaskDesc().c_str() : "All"));
Print("FILL_PATTERN", getenv("FILL_PATTERN") ? 1 : 0,
"%s", (fillPattern.size() ? getenv("FILL_PATTERN") : TransferBench::GetStrAttribute(ATR_SRC_PREP_DESCRIPTION).c_str()));
Print("GFX_BLOCK_ORDER", gfxBlockOrder,
"Thread block ordering: %s", gfxBlockOrder == 0 ? "Sequential" : "Interleaved");
Print("GFX_BLOCK_SIZE", gfxBlockSize,
"Threadblock size of %d", gfxBlockSize);
Print("GFX_SINGLE_TEAM", gfxSingleTeam,
"%s", (gfxSingleTeam ? "Combining CUs to work across entire data array" :
"Each CUs operates on its own disjoint subarray"));
Print("GFX_TEMPORAL", gfxTemporal,
"%s", (gfxTemporal == 0 ? "Not using non-temporal loads/stores" :
gfxTemporal == 1 ? "Using non-temporal loads" :
gfxTemporal == 2 ? "Using non-temporal stores" :
"Using non-temporal loads and stores"));
Print("GFX_UNROLL", gfxUnroll,
"Using GFX unroll factor of %d", gfxUnroll);
Print("GFX_WAVE_ORDER", gfxWaveOrder,
......@@ -359,11 +425,27 @@ public:
gfxWaveOrder == 3 ? "Wavefront,CU,Unroll" :
gfxWaveOrder == 4 ? "CU,Unroll,Wavefront" :
"CU,Wavefront,Unroll"));
Print("GFX_WORD_SIZE", gfxWordSize,
"Using GFX word size of %d (DWORDx%d)", gfxWordSize, gfxWordSize);
#if NIC_EXEC_ENABLED
Print("IP_ADDRESS_FAMILY", ipAddressFamily,
"IP address family is set to IPv%d", ipAddressFamily);
Print("IB_GID_INDEX", ibGidIndex,
"RoCE GID index is set to %s", (ibGidIndex < 0 ? "auto" : std::to_string(ibGidIndex).c_str()));
Print("IB_PORT_NUMBER", ibPort,
"IB port number is set to %d", ibPort);
#endif
Print("MIN_VAR_SUBEXEC", minNumVarSubExec,
"Using at least %d subexecutor(s) for variable subExec tranfers", minNumVarSubExec);
Print("MAX_VAR_SUBEXEC", maxNumVarSubExec,
"Using up to %s subexecutors for variable subExec transfers",
maxNumVarSubExec ? std::to_string(maxNumVarSubExec).c_str() : "all available");
#if NIC_EXEC_ENABLED
Print("NIC_RELAX_ORDER", nicRelaxedOrder,
"Using %s ordering for NIC RDMA", nicRelaxedOrder ? "relaxed" : "strict");
#endif
Print("NUM_ITERATIONS", numIterations,
(numIterations == 0) ? "Running infinitely" :
"Running %d %s", abs(numIterations), (numIterations > 0 ? " timed iteration(s)" : "seconds(s) per Test"));
......@@ -371,6 +453,10 @@ public:
"Running %s subiterations", (numSubIterations == 0 ? "infinite" : std::to_string(numSubIterations)).c_str());
Print("NUM_WARMUPS", numWarmups,
"Running %d warmup iteration(s) per Test", numWarmups);
#if NIC_EXEC_ENABLED
Print("ROCE_VERSION", roceVersion,
"RoCE version is set to %d", roceVersion);
#endif
Print("SHOW_ITERATIONS", showIterations,
"%s per-iteration timing", showIterations ? "Showing" : "Hiding");
Print("USE_HIP_EVENTS", useHipEvents,
......@@ -381,7 +467,6 @@ public:
"Running in %s mode", useInteractive ? "interactive" : "non-interactive");
Print("USE_SINGLE_STREAM", useSingleStream,
"Using single stream per GFX %s", useSingleStream ? "device" : "Transfer");
if (getenv("XCC_PREF_TABLE")) {
printf("%36s: Preferred XCC Table (XCC_PREF_TABLE)\n", "");
printf("%36s: ", "");
......@@ -408,6 +493,31 @@ public:
return defaultValue;
}
static std::vector<int> GetEnvVarArray(std::string const& varname, std::vector<int> const& defaultValue)
{
if (getenv(varname.c_str())) {
char* rangeStr = getenv(varname.c_str());
std::set<int> values;
char* token = strtok(rangeStr, ",");
while (token) {
int start, end;
if (sscanf(token, "%d-%d", &start, &end) == 2) {
for (int i = start; i <= end; i++) values.insert(i);
} else if (sscanf(token, "%d", &start) == 1) {
values.insert(start);
} else {
printf("[ERROR] Unrecognized token [%s]\n", token);
exit(1);
}
token = strtok(NULL, ",");
}
std::vector<int> result;
for (auto v : values) result.push_back(v);
return result;
}
return defaultValue;
}
static std::string GetEnvVar(std::string const& varname, std::string const& defaultValue)
{
if (getenv(varname.c_str()))
......@@ -470,15 +580,39 @@ public:
cfg.dma.useHipEvents = useHipEvents;
cfg.dma.useHsaCopy = useHsaDma;
cfg.gfx.blockOrder = gfxBlockOrder;
cfg.gfx.blockSize = gfxBlockSize;
cfg.gfx.cuMask = cuMask;
cfg.gfx.prefXccTable = prefXccTable;
cfg.gfx.unrollFactor = gfxUnroll;
cfg.gfx.temporalMode = gfxTemporal;
cfg.gfx.useHipEvents = useHipEvents;
cfg.gfx.useMultiStream = !useSingleStream;
cfg.gfx.useSingleTeam = gfxSingleTeam;
cfg.gfx.waveOrder = gfxWaveOrder;
cfg.gfx.wordSize = gfxWordSize;
cfg.nic.ibGidIndex = ibGidIndex;
cfg.nic.ibPort = ibPort;
cfg.nic.ipAddressFamily = ipAddressFamily;
cfg.nic.useRelaxedOrder = nicRelaxedOrder;
cfg.nic.roceVersion = roceVersion;
std::vector<int> closestNics;
if(closestNicStr != "") {
std::stringstream ss(closestNicStr);
std::string item;
while (std::getline(ss, item, ',')) {
try {
int nic = std::stoi(item);
closestNics.push_back(nic);
} catch (const std::invalid_argument& e) {
printf("[ERROR] Invalid NIC index (%s) by user in %s\n", item.c_str(), closestNicStr.c_str());
exit(1);
}
}
cfg.nic.closestNics = closestNics;
}
return cfg;
}
};
......
src/client/Presets/AllToAll.hpp
View file @ 4e2be38c
......@@ -30,9 +30,10 @@ void AllToAllPreset(EnvVars& ev,
{
A2A_COPY = 0,
A2A_READ_ONLY = 1,
A2A_WRITE_ONLY = 2
A2A_WRITE_ONLY = 2,
A2A_CUSTOM = 3,
};
char a2aModeStr[3][20] = {"Copy", "Read-Only", "Write-Only"};
char a2aModeStr[4][20] = {"Copy", "Read-Only", "Write-Only", "Custom"};
// Force single-stream mode for all-to-all benchmark
ev.useSingleStream = 1;
......@@ -45,21 +46,39 @@ void AllToAllPreset(EnvVars& ev,
// Collect env vars for this preset
int a2aDirect = EnvVars::GetEnvVar("A2A_DIRECT" , 1);
int a2aLocal = EnvVars::GetEnvVar("A2A_LOCAL" , 0);
int a2aMode = EnvVars::GetEnvVar("A2A_MODE" , 0);
int numGpus = EnvVars::GetEnvVar("NUM_GPU_DEVICES", numDetectedGpus);
int numQueuePairs = EnvVars::GetEnvVar("NUM_QUEUE_PAIRS", 0);
int numSubExecs = EnvVars::GetEnvVar("NUM_SUB_EXEC" , 8);
int useDmaExec = EnvVars::GetEnvVar("USE_DMA_EXEC" , 0);
int useFineGrain = EnvVars::GetEnvVar("USE_FINE_GRAIN" , 1);
int useRemoteRead = EnvVars::GetEnvVar("USE_REMOTE_READ", 0);
// A2A_MODE may be 0,1,2 or else custom numSrcs:numDsts
int numSrcs, numDsts;
int a2aMode = 0;
if (getenv("A2A_MODE") && sscanf(getenv("A2A_MODE"), "%d:%d", &numSrcs, &numDsts) == 2) {
a2aMode = A2A_CUSTOM;
} else {
a2aMode = EnvVars::GetEnvVar("A2A_MODE", 0);
if (a2aMode < 0 || a2aMode > 2) {
printf("[ERROR] a2aMode must be between 0 and 2, or else numSrcs:numDsts\n");
exit(1);
}
numSrcs = (a2aMode == A2A_WRITE_ONLY ? 0 : 1);
numDsts = (a2aMode == A2A_READ_ONLY ? 0 : 1);
}
// Print off environment variables
ev.DisplayEnvVars();
if (!ev.hideEnv) {
if (!ev.outputToCsv) printf("[AllToAll Related]\n");
ev.Print("A2A_DIRECT" , a2aDirect , a2aDirect ? "Only using direct links" : "Full all-to-all");
ev.Print("A2A_LOCAL" , a2aLocal , "%s local transfers", a2aLocal ? "Include" : "Exclude");
ev.Print("A2A_MODE" , a2aMode , a2aModeStr[a2aMode]);
ev.Print("A2A_MODE" , (a2aMode == A2A_CUSTOM) ? std::to_string(numSrcs) + ":" + std::to_string(numDsts) : std::to_string(a2aMode),
(a2aMode == A2A_CUSTOM) ? (std::to_string(numSrcs) + " read(s) " +
std::to_string(numDsts) + " write(s)").c_str(): a2aModeStr[a2aMode]);
ev.Print("NUM_GPU_DEVICES", numGpus , "Using %d GPUs", numGpus);
ev.Print("NUM_QUEUE_PAIRS", numQueuePairs, "Using %d queue pairs for NIC transfers", numQueuePairs);
ev.Print("NUM_SUB_EXEC" , numSubExecs , "Using %d subexecutors/CUs per Transfer", numSubExecs);
ev.Print("USE_DMA_EXEC" , useDmaExec , "Using %s executor", useDmaExec ? "DMA" : "GFX");
ev.Print("USE_FINE_GRAIN" , useFineGrain , "Using %s-grained memory", useFineGrain ? "fine" : "coarse");
......@@ -68,19 +87,16 @@ void AllToAllPreset(EnvVars& ev,
}
// Validate env vars
if (a2aMode < 0 || a2aMode > 2) {
printf("[ERROR] a2aMode must be between 0 and 2\n");
exit(1);
}
if (numGpus < 0 || numGpus > numDetectedGpus) {
printf("[ERROR] Cannot use %d GPUs. Detected %d GPUs\n", numGpus, numDetectedGpus);
exit(1);
}
if (useDmaExec && (numSrcs != 1 || numDsts != 1)) {
printf("[ERROR] DMA execution can only be used for copies (A2A_MODE=0)\n");
exit(1);
}
// Collect the number of GPU devices to use
int const numSrcs = (a2aMode == A2A_WRITE_ONLY ? 0 : 1);
int const numDsts = (a2aMode == A2A_READ_ONLY ? 0 : 1);
MemType memType = useFineGrain ? MEM_GPU_FINE : MEM_GPU;
ExeType exeType = useDmaExec ? EXE_GPU_DMA : EXE_GPU_GFX;
......@@ -103,8 +119,11 @@ void AllToAllPreset(EnvVars& ev,
// Build Transfer and add it to list
TransferBench::Transfer transfer;
transfer.numBytes = numBytesPerTransfer;
if (numSrcs) transfer.srcs.push_back({memType, i});
for (int x = 0; x < numSrcs; x++) transfer.srcs.push_back({memType, i});
// When using multiple destinations, the additional destinations are "local"
if (numDsts) transfer.dsts.push_back({memType, j});
for (int x = 1; x < numDsts; x++) transfer.dsts.push_back({memType, i});
transfer.exeDevice = {exeType, (useRemoteRead ? j : i)};
transfer.exeSubIndex = -1;
transfer.numSubExecs = numSubExecs;
......@@ -114,6 +133,23 @@ void AllToAllPreset(EnvVars& ev,
}
}
// Create a ring using NICs
std::vector<int> nicTransferIdx(numGpus);
if (numQueuePairs > 0) {
int numNics = TransferBench::GetNumExecutors(EXE_NIC);
for (int i = 0; i < numGpus; i++) {
TransferBench::Transfer transfer;
transfer.numBytes = numBytesPerTransfer;
transfer.srcs.push_back({memType, i});
transfer.dsts.push_back({memType, (i+1) % numGpus});
transfer.exeDevice = {TransferBench::EXE_NIC_NEAREST, i};
transfer.exeSubIndex = (i+1) % numGpus;
transfer.numSubExecs = numQueuePairs;
nicTransferIdx[i] = transfers.size();
transfers.push_back(transfer);
}
}
printf("GPU-GFX All-To-All benchmark:\n");
printf("==========================\n");
printf("- Copying %lu bytes between %s pairs of GPUs using %d CUs (%lu Transfers)\n",
......@@ -133,20 +169,24 @@ void AllToAllPreset(EnvVars& ev,
// Print results
char separator = (ev.outputToCsv ? ',' : ' ');
printf("\nSummary: [%lu bytes per Transfer]\n", numBytesPerTransfer);
printf("==========================================================\n");
printf("\nSummary: [%lu bytes per Transfer] [%s:%d] [%d Read(s) %d Write(s)]\n",
numBytesPerTransfer, useDmaExec ? "DMA" : "GFX", numSubExecs, numSrcs, numDsts);
printf("===========================================================================\n");
printf("SRC\\DST ");
for (int dst = 0; dst < numGpus; dst++)
printf("%cGPU %02d ", separator, dst);
if (numQueuePairs > 0)
printf("%cNIC(%02d QP)", separator, numQueuePairs);
printf(" %cSTotal %cActual\n", separator, separator);
double totalBandwidthGpu = 0.0;
double minExecutorBandwidth = std::numeric_limits<double>::max();
double maxExecutorBandwidth = 0.0;
std::vector<double> colTotalBandwidth(numGpus+1, 0.0);
double minActualBandwidth = std::numeric_limits<double>::max();
double maxActualBandwidth = 0.0;
std::vector<double> colTotalBandwidth(numGpus+2, 0.0);
for (int src = 0; src < numGpus; src++) {
double rowTotalBandwidth = 0;
double executorBandwidth = 0;
int transferCount = 0;
double minBandwidth = std::numeric_limits<double>::max();
printf("GPU %02d", src);
for (int dst = 0; dst < numGpus; dst++) {
if (reIndex.count(std::make_pair(src, dst))) {
......@@ -155,24 +195,38 @@ void AllToAllPreset(EnvVars& ev,
colTotalBandwidth[dst] += r.avgBandwidthGbPerSec;
rowTotalBandwidth += r.avgBandwidthGbPerSec;
totalBandwidthGpu += r.avgBandwidthGbPerSec;
executorBandwidth = std::max(executorBandwidth,
results.exeResults[transfers[transferIdx].exeDevice].avgBandwidthGbPerSec);
minBandwidth = std::min(minBandwidth, r.avgBandwidthGbPerSec);
transferCount++;
printf("%c%8.3f ", separator, r.avgBandwidthGbPerSec);
} else {
printf("%c%8s ", separator, "N/A");
}
}
printf(" %c%8.3f %c%8.3f\n", separator, rowTotalBandwidth, separator, executorBandwidth);
minExecutorBandwidth = std::min(minExecutorBandwidth, executorBandwidth);
maxExecutorBandwidth = std::max(maxExecutorBandwidth, executorBandwidth);
colTotalBandwidth[numGpus] += rowTotalBandwidth;
if (numQueuePairs > 0) {
TransferBench::TransferResult const& r = results.tfrResults[nicTransferIdx[src]];
colTotalBandwidth[numGpus] += r.avgBandwidthGbPerSec;
rowTotalBandwidth += r.avgBandwidthGbPerSec;
totalBandwidthGpu += r.avgBandwidthGbPerSec;
minBandwidth = std::min(minBandwidth, r.avgBandwidthGbPerSec);
transferCount++;
printf("%c%8.3f ", separator, r.avgBandwidthGbPerSec);
}
double actualBandwidth = minBandwidth * transferCount;
printf(" %c%8.3f %c%8.3f\n", separator, rowTotalBandwidth, separator, actualBandwidth);
minActualBandwidth = std::min(minActualBandwidth, actualBandwidth);
maxActualBandwidth = std::max(maxActualBandwidth, actualBandwidth);
colTotalBandwidth[numGpus+1] += rowTotalBandwidth;
}
printf("\nRTotal");
for (int dst = 0; dst < numGpus; dst++) {
printf("%c%8.3f ", separator, colTotalBandwidth[dst]);
}
printf(" %c%8.3f %c%8.3f %c%8.3f\n", separator, colTotalBandwidth[numGpus],
separator, minExecutorBandwidth, separator, maxExecutorBandwidth);
if (numQueuePairs > 0) {
printf("%c%8.3f ", separator, colTotalBandwidth[numGpus]);
}
printf(" %c%8.3f %c%8.3f %c%8.3f\n", separator, colTotalBandwidth[numGpus+1],
separator, minActualBandwidth, separator, maxActualBandwidth);
printf("\n");
printf("Average bandwidth (GPU Timed): %8.3f GB/s\n", totalBandwidthGpu / transfers.size());
......
src/client/Presets/AllToAllN.hpp 0 → 100644
View file @ 4e2be38c
/*
Copyright (c) 2024 Advanced Micro Devices, Inc. All rights reserved.
Permission is hereby granted, free of charge, to any person obtaining a copy
of this software and associated documentation files (the "Software"), to deal
in the Software without restriction, including without limitation the rights
to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
copies of the Software, and to permit persons to whom the Software is
furnished to do so, subject to the following conditions:
The above copyright notice and this permission notice shall be included in
all copies or substantial portions of the Software.
THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN
THE SOFTWARE.
*/
#include "EnvVars.hpp"
void AllToAllRdmaPreset(EnvVars& ev,
size_t const numBytesPerTransfer,
std::string const presetName)
{
int numDetectedGpus = TransferBench::GetNumExecutors(EXE_GPU_GFX);
// Collect env vars for this preset
int numGpus = EnvVars::GetEnvVar("NUM_GPU_DEVICES", numDetectedGpus);
int numQueuePairs = EnvVars::GetEnvVar("NUM_QUEUE_PAIRS", 1);
int useFineGrain = EnvVars::GetEnvVar("USE_FINE_GRAIN" , 1);
// Print off environment variables
ev.DisplayEnvVars();
if (!ev.hideEnv) {
if (!ev.outputToCsv) printf("[AllToAll Network Related]\n");
ev.Print("NUM_GPU_DEVICES", numGpus , "Using %d GPUs", numGpus);
ev.Print("NUM_QUEUE_PAIRS", numQueuePairs, "Using %d queue pairs for NIC transfers", numQueuePairs);
ev.Print("USE_FINE_GRAIN" , useFineGrain , "Using %s-grained memory", useFineGrain ? "fine" : "coarse");
printf("\n");
}
// Validate env vars
if (numGpus < 0 || numGpus > numDetectedGpus) {
printf("[ERROR] Cannot use %d GPUs. Detected %d GPUs\n", numGpus, numDetectedGpus);
exit(1);
}
MemType memType = useFineGrain ? MEM_GPU_FINE : MEM_GPU;
std::map<std::pair<int, int>, int> reIndex;
std::vector<Transfer> transfers;
for (int i = 0; i < numGpus; i++) {
for (int j = 0; j < numGpus; j++) {
// Build Transfer and add it to list
TransferBench::Transfer transfer;
transfer.numBytes = numBytesPerTransfer;
transfer.srcs.push_back({memType, i});
transfer.dsts.push_back({memType, j});
transfer.exeDevice = {EXE_NIC_NEAREST, i};
transfer.exeSubIndex = j;
transfer.numSubExecs = numQueuePairs;
reIndex[std::make_pair(i,j)] = transfers.size();
transfers.push_back(transfer);
}
}
printf("GPU-RDMA All-To-All benchmark:\n");
printf("==========================\n");
printf("- Copying %lu bytes between all pairs of GPUs using %d QPs per Transfer (%lu Transfers)\n",
numBytesPerTransfer, numQueuePairs, transfers.size());
if (transfers.size() == 0) return;
// Execute Transfers
TransferBench::ConfigOptions cfg = ev.ToConfigOptions();
TransferBench::TestResults results;
if (!TransferBench::RunTransfers(cfg, transfers, results)) {
for (auto const& err : results.errResults)
printf("%s\n", err.errMsg.c_str());
exit(0);
} else {
PrintResults(ev, 1, transfers, results);
}
// Print results
char separator = (ev.outputToCsv ? ',' : ' ');
printf("\nSummary: [%lu bytes per Transfer]\n", numBytesPerTransfer);
printf("==========================================================\n");
printf("SRC\\DST ");
for (int dst = 0; dst < numGpus; dst++)
printf("%cGPU %02d ", separator, dst);
printf(" %cSTotal %cActual\n", separator, separator);
double totalBandwidthGpu = 0.0;
double minActualBandwidth = std::numeric_limits<double>::max();
double maxActualBandwidth = 0.0;
std::vector<double> colTotalBandwidth(numGpus+2, 0.0);
for (int src = 0; src < numGpus; src++) {
double rowTotalBandwidth = 0;
int transferCount = 0;
double minBandwidth = std::numeric_limits<double>::max();
printf("GPU %02d", src);
for (int dst = 0; dst < numGpus; dst++) {
if (reIndex.count(std::make_pair(src, dst))) {
int const transferIdx = reIndex[std::make_pair(src,dst)];
TransferBench::TransferResult const& r = results.tfrResults[transferIdx];
colTotalBandwidth[dst] += r.avgBandwidthGbPerSec;
rowTotalBandwidth += r.avgBandwidthGbPerSec;
totalBandwidthGpu += r.avgBandwidthGbPerSec;
minBandwidth = std::min(minBandwidth, r.avgBandwidthGbPerSec);
transferCount++;
printf("%c%8.3f ", separator, r.avgBandwidthGbPerSec);
} else {
printf("%c%8s ", separator, "N/A");
}
}
double actualBandwidth = minBandwidth * transferCount;
printf(" %c%8.3f %c%8.3f\n", separator, rowTotalBandwidth, separator, actualBandwidth);
minActualBandwidth = std::min(minActualBandwidth, actualBandwidth);
maxActualBandwidth = std::max(maxActualBandwidth, actualBandwidth);
colTotalBandwidth[numGpus+1] += rowTotalBandwidth;
}
printf("\nRTotal");
for (int dst = 0; dst < numGpus; dst++) {
printf("%c%8.3f ", separator, colTotalBandwidth[dst]);
}
printf(" %c%8.3f %c%8.3f %c%8.3f\n", separator, colTotalBandwidth[numGpus+1],
separator, minActualBandwidth, separator, maxActualBandwidth);
printf("\n");
printf("Average bandwidth (Tx Thread Timed): %8.3f GB/s\n", totalBandwidthGpu / transfers.size());
printf("Aggregate bandwidth (Tx Thread Timed): %8.3f GB/s\n", totalBandwidthGpu);
printf("Aggregate bandwidth (CPU Timed): %8.3f GB/s\n", results.avgTotalBandwidthGbPerSec);
PrintErrors(results.errResults);
}
src/client/Presets/AllToAllSweep.hpp 0 → 100644
View file @ 4e2be38c
/*
Copyright (c) 2025 Advanced Micro Devices, Inc. All rights reserved.
Permission is hereby granted, free of charge, to any person obtaining a copy
of this software and associated documentation files (the "Software"), to deal
in the Software without restriction, including without limitation the rights
to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
copies of the Software, and to permit persons to whom the Software is
furnished to do so, subject to the following conditions:
The above copyright notice and this permission notice shall be included in
all copies or substantial portions of the Software.
THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN
THE SOFTWARE.
*/
#include "EnvVars.hpp"
void AllToAllSweepPreset(EnvVars& ev,
size_t const numBytesPerTransfer,
std::string const presetName)
{
enum
{
A2A_COPY = 0,
A2A_READ_ONLY = 1,
A2A_WRITE_ONLY = 2,
A2A_CUSTOM = 3,
};
char a2aModeStr[4][20] = {"Copy", "Read-Only", "Write-Only", "Custom"};
// Force single-stream mode for all-to-all benchmark
ev.useSingleStream = 1;
int numDetectedGpus = TransferBench::GetNumExecutors(EXE_GPU_GFX);
// Collect env vars for this preset
int a2aDirect = EnvVars::GetEnvVar("A2A_DIRECT" , 1);
int a2aLocal = EnvVars::GetEnvVar("A2A_LOCAL" , 0);
int numGpus = EnvVars::GetEnvVar("NUM_GPU_DEVICES", numDetectedGpus);
int showMinOnly = EnvVars::GetEnvVar("SHOW_MIN_ONLY", 1);
int useFineGrain = EnvVars::GetEnvVar("USE_FINE_GRAIN" , 1);
int useRemoteRead = EnvVars::GetEnvVar("USE_REMOTE_READ", 0);
int useSpray = EnvVars::GetEnvVar("USE_SPRAY", 0);
int verbose = EnvVars::GetEnvVar("VERBOSE", 0);
std::vector<int> unrollList = EnvVars::GetEnvVarArray("UNROLLS", {1,2,3,4,6,8});
std::vector<int> numCusList = EnvVars::GetEnvVarArray("NUM_CUS", {4,8,12,16,24,32});
// A2A_MODE may be 0,1,2 or else custom numSrcs:numDsts
int numSrcs, numDsts;
int a2aMode = 0;
if (getenv("A2A_MODE") && sscanf(getenv("A2A_MODE"), "%d:%d", &numSrcs, &numDsts) == 2) {
a2aMode = A2A_CUSTOM;
} else {
a2aMode = EnvVars::GetEnvVar("A2A_MODE", 0);
if (a2aMode < 0 || a2aMode > 2) {
printf("[ERROR] a2aMode must be between 0 and 2, or else numSrcs:numDsts\n");
exit(1);
}
numSrcs = (a2aMode == A2A_WRITE_ONLY ? 0 : 1);
numDsts = (a2aMode == A2A_READ_ONLY ? 0 : 1);
}
// Print off environment variables
ev.DisplayEnvVars();
if (!ev.hideEnv) {
if (!ev.outputToCsv) printf("[AllToAll Related]\n");
ev.Print("A2A_DIRECT" , a2aDirect , a2aDirect ? "Only using direct links" : "Full all-to-all");
ev.Print("A2A_LOCAL" , a2aLocal , "%s local transfers", a2aLocal ? "Include" : "Exclude");
ev.Print("A2A_MODE" , (a2aMode == A2A_CUSTOM) ? std::to_string(numSrcs) + ":" + std::to_string(numDsts) : std::to_string(a2aMode),
(a2aMode == A2A_CUSTOM) ? (std::to_string(numSrcs) + " read(s) " +
std::to_string(numDsts) + " write(s)").c_str(): a2aModeStr[a2aMode]);
ev.Print("SHOW_MIN_ONLY" , showMinOnly , showMinOnly ? "Showing only slowest GPU results" : "Showing slowest and fastest GPU results");
ev.Print("NUM_CUS" , numCusList.size(), EnvVars::ToStr(numCusList).c_str());
ev.Print("NUM_GPU_DEVICES", numGpus , "Using %d GPUs", numGpus);
ev.Print("UNROLLS" , unrollList.size(), EnvVars::ToStr(unrollList).c_str());
ev.Print("USE_FINE_GRAIN" , useFineGrain , "Using %s-grained memory", useFineGrain ? "fine" : "coarse");
ev.Print("USE_REMOTE_READ", useRemoteRead , "Using %s as executor", useRemoteRead ? "DST" : "SRC");
ev.Print("USE_SPRAY" , useSpray , "%s per CU", useSpray ? "All targets" : "One target");
ev.Print("VERBOSE" , verbose , verbose ? "Display test results" : "Display summary only");
printf("\n");
}
// Validate env vars
if (numGpus < 0 || numGpus > numDetectedGpus) {
printf("[ERROR] Cannot use %d GPUs. Detected %d GPUs\n", numGpus, numDetectedGpus);
exit(1);
}
if (useSpray && numDsts > 1) {
printf("[ERROR] Cannot use USE_SPRAY with multiple destination buffers\n");
exit(1);
}
// Collect the number of GPU devices to use
MemType memType = useFineGrain ? MEM_GPU_FINE : MEM_GPU;
ExeType exeType = EXE_GPU_GFX;
std::vector<Transfer> transfers;
int targetCount = 0;
if (!useSpray) {
// Each CU will work on just one target
for (int i = 0; i < numGpus; i++) {
targetCount = 0;
for (int j = 0; j < numGpus; j++) {
// Check whether or not to execute this pair
if (i == j) {
if (!a2aLocal) continue;
} else if (a2aDirect) {
#if !defined(__NVCC__)
uint32_t linkType, hopCount;
HIP_CALL(hipExtGetLinkTypeAndHopCount(i, j, &linkType, &hopCount));
if (hopCount != 1) continue;
#endif
}
// Build Transfer and add it to list
TransferBench::Transfer transfer;
targetCount++;
transfer.numBytes = numBytesPerTransfer;
for (int x = 0; x < numSrcs; x++) transfer.srcs.push_back({memType, i});
// When using multiple destinations, the additional destinations are "local"
if (numDsts) transfer.dsts.push_back({memType, j});
for (int x = 1; x < numDsts; x++) transfer.dsts.push_back({memType, i});
transfer.exeDevice = {exeType, (useRemoteRead ? j : i)};
transfer.exeSubIndex = -1;
transfers.push_back(transfer);
}
}
} else {
// Each CU will work on all targets
for (int i = 0; i < numGpus; i++) {
TransferBench::Transfer transfer;
transfer.numBytes = numBytesPerTransfer;
transfer.exeDevice = {exeType, i};
transfer.exeSubIndex = -1;
targetCount = 0;
for (int j = 0; j < numGpus; j++) {
// Check whether or not to transfer to this GPU
if (i == j) {
if (!a2aLocal) continue;
} else if (a2aDirect) {
#if !defined(__NVCC__)
uint32_t linkType, hopCount;
HIP_CALL(hipExtGetLinkTypeAndHopCount(i, j, &linkType, &hopCount));
if (hopCount != 1) continue;
#endif
}
targetCount++;
for (int x = 0; x < numSrcs; x++) transfer.srcs.push_back({memType, useRemoteRead ? j : i});
if (numDsts) transfer.dsts.push_back({memType, j});
for (int x = 1; x < numDsts; x++) transfer.dsts.push_back({memType, i});
}
transfers.push_back(transfer);
}
}
printf("GPU-GFX All-To-All Sweep benchmark:\n");
printf("==========================\n");
printf("- Copying %lu bytes between %s pairs of GPUs\n", numBytesPerTransfer, a2aDirect ? "directly connected" : "all");
if (transfers.size() == 0) {
printf("[WARN} No transfers requested. Try adjusting A2A_DIRECT or A2A_LOCAL\n");
return;
}
// Execute Transfers
TransferBench::ConfigOptions cfg = ev.ToConfigOptions();
// Run tests
std::map<std::pair<int, int>, TransferBench::TestResults> results;
// Display summary
printf("#CUs\\Unroll");
for (int u : unrollList) {
printf(" %d(Min) ", u);
if (!showMinOnly) printf(" %d(Max) ", u);
}
printf("\n");
for (int c : numCusList) {
printf(" %5d ", c); fflush(stdout);
for (int u : unrollList) {
ev.gfxUnroll = cfg.gfx.unrollFactor = u;
for (auto& transfer : transfers)
transfer.numSubExecs = useSpray ? (c * targetCount) : c;
double minBandwidth = std::numeric_limits<double>::max();
double maxBandwidth = std::numeric_limits<double>::min();
TransferBench::TestResults result;
if (TransferBench::RunTransfers(cfg, transfers, result)) {
for (auto const& exeResult : result.exeResults) {
minBandwidth = std::min(minBandwidth, exeResult.second.avgBandwidthGbPerSec);
maxBandwidth = std::max(maxBandwidth, exeResult.second.avgBandwidthGbPerSec);
}
if (useSpray) {
minBandwidth *= targetCount;
maxBandwidth *= targetCount;
}
results[std::make_pair(c,u)] = result;
} else {
minBandwidth = 0.0;
}
printf(" %7.2f ", minBandwidth);
if (!showMinOnly) printf(" %7.2f ", maxBandwidth);
fflush(stdout);
}
printf("\n"); fflush(stdout);
}
if (verbose) {
int testNum = 0;
for (int c : numCusList) {
for (int u : unrollList) {
printf("CUs: %d Unroll %d\n", c, u);
PrintResults(ev, ++testNum, transfers, results[std::make_pair(c,u)]);
}
}
}
}
Markdown is supported
0% or .
You are about to add 0 people to the discussion. Proceed with caution.
Finish editing this message first!
Please register or to comment