Skip to content

Implement cache of device properties #2492

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
AuroraPerego merged 21 commits into from
May 27, 2025
Merged

Implement cache of device properties #2492

Show file tree
Hide file tree
Changes from 5 commits
Commits
Show all changes
21 commits
Select commit Hold shift + click to select a range
253efb4
Cache device properties after first API calls
sbaldu Apr 3, 2025
bc93536
Make caching of device properties thread-safe
sbaldu Apr 6, 2025
5daa9b3
Add missing include
sbaldu Apr 7, 2025
06243e1
Explicitly specify lock template parameter
sbaldu Apr 7, 2025
9d5e977
Fix typo in sycl implementation
sbaldu Apr 7, 2025
e49e604
Move device properties inside QueueRegistry
sbaldu Apr 9, 2025
ef3a99a
Fix deadlock condition in cuda/hip and make sintax lighter
sbaldu Apr 11, 2025
c130965
Add static_cast of warpSize
sbaldu Apr 11, 2025
ca3def8
Use static variables and call_once instead of locks
sbaldu Apr 25, 2025
189054e
Use call_once for cpu backends as well
sbaldu Apr 25, 2025
1f94564
Sett all properties on first once_call
sbaldu May 4, 2025
9bb4507
Use consistent syntax in method return type
sbaldu May 5, 2025
3731eb4
Fix address sanitizer error for DevCpu
sbaldu May 5, 2025
6de4bc7
Define functions to be used inside call_once lambda
sbaldu May 6, 2025
30e3663
Encaplusate set of dev props inside QueueRegistry/DevSyclImpl
sbaldu May 19, 2025
1126366
Fix typos
sbaldu May 20, 2025
026b155
Template `setDeviceProperties` and specialize for cpu
sbaldu May 20, 2025
2927629
Use const& for sycl device
sbaldu May 20, 2025
7770e22
Pass device to `setDeviceProperties`
sbaldu May 20, 2025
b137daa
Rename `QueueRegistry` as `DevGenericImpl`
sbaldu May 21, 2025
7da31c8
Make `setDeviceProperties` a static method of the device
sbaldu May 21, 2025
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
47 changes: 40 additions & 7 deletions include/alpaka/dev/DevCpu.hpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -6,6 +6,7 @@
#pragma once

#include "alpaka/dev/Traits.hpp"
#include "alpaka/dev/common/DeviceProperties.hpp"
#include "alpaka/dev/common/QueueRegistry.hpp"
#include "alpaka/dev/cpu/SysInfo.hpp"
#include "alpaka/mem/buf/Traits.hpp"
Expand Down Expand Up @@ -57,7 +58,10 @@ namespace alpaka
friend struct trait::GetDevByIdx<PlatformCpu>;

protected:
DevCpu() : m_spDevCpuImpl(std::make_shared<cpu::detail::DevCpuImpl>())
DevCpu()
: m_spDevCpuImpl(std::make_shared<cpu::detail::DevCpuImpl>())
, m_deviceProperties(std::make_shared<alpaka::DeviceProperties>())
, m_mutex(std::make_shared<std::mutex>())
{
}

Expand Down Expand Up @@ -89,8 +93,16 @@ namespace alpaka
return 0;
}

friend struct trait::GetName<DevCpu>;
friend struct trait::GetMemBytes<DevCpu>;
friend struct trait::GetFreeMemBytes<DevCpu>;
friend struct trait::GetWarpSizes<DevCpu>;
friend struct trait::GetPreferredWarpSize<DevCpu>;

private:
std::shared_ptr<cpu::detail::DevCpuImpl> m_spDevCpuImpl;
std::shared_ptr<alpaka::DeviceProperties> m_deviceProperties;
std::shared_ptr<std::mutex> m_mutex;
Copy link
Contributor

@fwyzard fwyzard Apr 7, 2025

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Can you move these into the cpu::detail::DevCpuImpl class ?

Copy link
Contributor Author

@sbaldu sbaldu Apr 9, 2025

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Yes, done

};

namespace trait
Expand All @@ -99,29 +111,50 @@ namespace alpaka
template<>
struct GetName<DevCpu>
{
ALPAKA_FN_HOST static auto getName(DevCpu const& /* dev */) -> std::string
ALPAKA_FN_HOST static auto getName(DevCpu const& dev) -> std::string
{
return cpu::detail::getCpuName();
{
std::lock_guard<std::mutex> lock(*dev.m_mutex);
if(!dev.m_deviceProperties->name.has_value())
{
dev.m_deviceProperties->name = cpu::detail::getCpuName();
}
}
return dev.m_deviceProperties->name.value();
}
};

//! The CPU device available memory get trait specialization.
template<>
struct GetMemBytes<DevCpu>
{
ALPAKA_FN_HOST static auto getMemBytes(DevCpu const& /* dev */) -> std::size_t
ALPAKA_FN_HOST static auto getMemBytes(DevCpu const& dev) -> std::size_t
{
return cpu::detail::getTotalGlobalMemSizeBytes();
{
std::lock_guard<std::mutex> lock(*dev.m_mutex);
if(!dev.m_deviceProperties->totalGlobalMem.has_value())
{
dev.m_deviceProperties->totalGlobalMem = cpu::detail::getTotalGlobalMemSizeBytes();
}
}
return dev.m_deviceProperties->totalGlobalMem.value();
}
};

//! The CPU device free memory get trait specialization.
template<>
struct GetFreeMemBytes<DevCpu>
{
ALPAKA_FN_HOST static auto getFreeMemBytes(DevCpu const& /* dev */) -> std::size_t
ALPAKA_FN_HOST static auto getFreeMemBytes(DevCpu const& dev) -> std::size_t
{
return cpu::detail::getFreeGlobalMemSizeBytes();
{
std::lock_guard<std::mutex> lock(*dev.m_mutex);
if(!dev.m_deviceProperties->freeGlobalMem.has_value())
{
dev.m_deviceProperties->freeGlobalMem = cpu::detail::getFreeGlobalMemSizeBytes();
}
}
return dev.m_deviceProperties->freeGlobalMem.value();
}
};

Expand Down
62 changes: 49 additions & 13 deletions include/alpaka/dev/DevGenericSycl.hpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -9,6 +9,7 @@
#include "alpaka/core/Common.hpp"
#include "alpaka/core/Sycl.hpp"
#include "alpaka/dev/Traits.hpp"
#include "alpaka/dev/common/DeviceProperties.hpp"
#include "alpaka/mem/buf/Traits.hpp"
#include "alpaka/platform/Traits.hpp"
#include "alpaka/queue/Properties.hpp"
Expand Down Expand Up @@ -131,6 +132,8 @@ namespace alpaka
public:
DevGenericSycl(sycl::device device, sycl::context context)
: m_impl{std::make_shared<detail::DevGenericSyclImpl>(std::move(device), std::move(context))}
, m_deviceProperties{std::make_shared<alpaka::DeviceProperties>()}
, m_mutex{std::make_shared<std::mutex>()}
{
}

Expand All @@ -150,6 +153,8 @@ namespace alpaka
}

std::shared_ptr<detail::DevGenericSyclImpl> m_impl;
std::shared_ptr<alpaka::DeviceProperties> m_deviceProperties;
std::shared_ptr<std::mutex> m_mutex;
};

namespace trait
Expand All @@ -160,8 +165,15 @@ namespace alpaka
{
static auto getName(DevGenericSycl<TTag> const& dev) -> std::string
{
auto const device = dev.getNativeHandle().first;
return device.template get_info<sycl::info::device::name>();
{
std::lock_guard<std::mutex> lock(*dev.m_mutex);
if(!dev.m_deviceProperties->name.has_value())
{
auto const device = dev.getNativeHandle().first;
dev.m_deviceProperties->name = device.template get_info<sycl::info::device::name>();
}
}
return dev.m_deviceProperties->name.value();
}
};

Expand All @@ -171,8 +183,16 @@ namespace alpaka
{
static auto getMemBytes(DevGenericSycl<TTag> const& dev) -> std::size_t
{
auto const device = dev.getNativeHandle().first;
return device.template get_info<sycl::info::device::global_mem_size>();
{
std::lock_guard<std::mutex> lock(*dev.m_mutex);
if(!dev.m_deviceProperties->totalGlobalMem.has_value())
{
auto const device = dev.getNativeHandle().first;
dev.m_deviceProperties->totalGlobalMem
= device.template get_info<sycl::info::device::global_mem_size>();
}
}
return dev.m_deviceProperties->totalGlobalMem.value();
}
};

Expand All @@ -195,15 +215,24 @@ namespace alpaka
{
static auto getWarpSizes(DevGenericSycl<TTag> const& dev) -> std::vector<std::size_t>
{
auto const device = dev.getNativeHandle().first;
std::vector<std::size_t> warp_sizes = device.template get_info<sycl::info::device::sub_group_sizes>();
// The CPU runtime supports a sub-group size of 64, but the SYCL implementation currently does not
auto find64 = std::find(warp_sizes.begin(), warp_sizes.end(), 64);
if(find64 != warp_sizes.end())
warp_sizes.erase(find64);
// Sort the warp sizes in decreasing order
std::sort(warp_sizes.begin(), warp_sizes.end(), std::greater<>{});
return warp_sizes;
{
std::lock_guard<std::mutex> lock(*dev.m_mutex);
if(!dev.m_deviceProperties->warpSizes.has_value())
{
auto const device = dev.getNativeHandle().first;
std::vector<std::size_t> warp_sizes
= device.template get_info<sycl::info::device::sub_group_sizes>();
// The CPU runtime supports a sub-group size of 64, but the SYCL implementation currently does
// not
auto find64 = std::find(warp_sizes.begin(), warp_sizes.end(), 64);
if(find64 != warp_sizes.end())
warp_sizes.erase(find64);
// Sort the warp sizes in decreasing order
std::sort(warp_sizes.begin(), warp_sizes.end(), std::greater<>{});
dev.m_deviceProperties->warpSizes = std::move(warp_sizes);
}
}
return dev.m_deviceProperties->warpSizes.value();
}
};

Expand All @@ -213,6 +242,13 @@ namespace alpaka
{
static auto getPreferredWarpSize(DevGenericSycl<TTag> const& dev) -> std::size_t
{
{
std::lock_guard<std::mutex> lock(*dev.m_mutex);
if(dev.m_deviceProperties->warpSizes.has_value())
{
return dev.m_deviceProperties->warpSizes.value().front();
}
}
return GetWarpSizes<DevGenericSycl<TTag>>::getWarpSizes(dev).front();
}
};
Expand Down
106 changes: 81 additions & 25 deletions include/alpaka/dev/DevUniformCudaHipRt.hpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -10,6 +10,7 @@
#include "alpaka/core/Hip.hpp"
#include "alpaka/core/Interface.hpp"
#include "alpaka/dev/Traits.hpp"
#include "alpaka/dev/common/DeviceProperties.hpp"
#include "alpaka/dev/common/QueueRegistry.hpp"
#include "alpaka/mem/buf/Traits.hpp"
#include "alpaka/platform/Traits.hpp"
Expand All @@ -20,6 +21,7 @@
#include "alpaka/wait/Traits.hpp"

#include <cstddef>
#include <mutex>
#include <string>
#include <vector>

Expand Down Expand Up @@ -62,7 +64,10 @@ namespace alpaka
using IDeviceQueue = uniform_cuda_hip::detail::QueueUniformCudaHipRtImpl<TApi>;

protected:
DevUniformCudaHipRt() : m_QueueRegistry{std::make_shared<alpaka::detail::QueueRegistry<IDeviceQueue>>()}
DevUniformCudaHipRt()
: m_QueueRegistry{std::make_shared<alpaka::detail::QueueRegistry<IDeviceQueue>>()}
, m_deviceProperties{std::make_shared<alpaka::DeviceProperties>()}
, m_mutex(std::make_shared<std::mutex>())
{
}

Expand Down Expand Up @@ -94,16 +99,26 @@ namespace alpaka
m_QueueRegistry->registerQueue(spQueue);
}

friend struct trait::GetName<DevUniformCudaHipRt<TApi>>;
friend struct trait::GetMemBytes<DevUniformCudaHipRt<TApi>>;
friend struct trait::GetFreeMemBytes<DevUniformCudaHipRt<TApi>>;
friend struct trait::GetWarpSizes<DevUniformCudaHipRt<TApi>>;
friend struct trait::GetPreferredWarpSize<DevUniformCudaHipRt<TApi>>;

private:
DevUniformCudaHipRt(int iDevice)
: m_iDevice(iDevice)
, m_QueueRegistry(std::make_shared<alpaka::detail::QueueRegistry<IDeviceQueue>>())
, m_deviceProperties(std::make_shared<alpaka::DeviceProperties>())
, m_mutex(std::make_shared<std::mutex>())
{
}

int m_iDevice;

std::shared_ptr<alpaka::detail::QueueRegistry<IDeviceQueue>> m_QueueRegistry;
std::shared_ptr<alpaka::DeviceProperties> m_deviceProperties;
std::shared_ptr<std::mutex> m_mutex;
};

namespace trait
Expand All @@ -114,12 +129,19 @@ namespace alpaka
{
ALPAKA_FN_HOST static auto getName(DevUniformCudaHipRt<TApi> const& dev) -> std::string
{
// There is cuda/hip-DeviceGetAttribute as faster alternative to cuda/hip-GetDeviceProperties to get a
// single device property but it has no option to get the name
typename TApi::DeviceProp_t devProp;
ALPAKA_UNIFORM_CUDA_HIP_RT_CHECK(TApi::getDeviceProperties(&devProp, dev.getNativeHandle()));

return std::string(devProp.name);
{
std::lock_guard<std::mutex> lock(*dev.m_mutex);
if(!dev.m_deviceProperties->name.has_value())
{
// There is cuda/hip-DeviceGetAttribute as faster alternative to cuda/hip-GetDeviceProperties
// to get a single device property but it has no option to get the name
typename TApi::DeviceProp_t devProp;
ALPAKA_UNIFORM_CUDA_HIP_RT_CHECK(TApi::getDeviceProperties(&devProp, dev.getNativeHandle()));
dev.m_deviceProperties->name = std::string(devProp.name);
}
}

return dev.m_deviceProperties->name.value();
}
};

Expand All @@ -129,15 +151,24 @@ namespace alpaka
{
ALPAKA_FN_HOST static auto getMemBytes(DevUniformCudaHipRt<TApi> const& dev) -> std::size_t
{
// Set the current device to wait for.
ALPAKA_UNIFORM_CUDA_HIP_RT_CHECK(TApi::setDevice(dev.getNativeHandle()));
{
std::lock_guard<std::mutex> lock(*dev.m_mutex);
if(!dev.m_deviceProperties->freeGlobalMem.has_value())
{
// Set the current device to wait for.
ALPAKA_UNIFORM_CUDA_HIP_RT_CHECK(TApi::setDevice(dev.getNativeHandle()));

std::size_t freeInternal(0u);
std::size_t totalInternal(0u);

std::size_t freeInternal(0u);
std::size_t totalInternal(0u);
ALPAKA_UNIFORM_CUDA_HIP_RT_CHECK(TApi::memGetInfo(&freeInternal, &totalInternal));

ALPAKA_UNIFORM_CUDA_HIP_RT_CHECK(TApi::memGetInfo(&freeInternal, &totalInternal));
dev.m_deviceProperties->totalGlobalMem = totalInternal;
dev.m_deviceProperties->freeGlobalMem = freeInternal;
}
}

return totalInternal;
return dev.m_deviceProperties->totalGlobalMem.value();
}
};

Expand All @@ -147,15 +178,24 @@ namespace alpaka
{
ALPAKA_FN_HOST static auto getFreeMemBytes(DevUniformCudaHipRt<TApi> const& dev) -> std::size_t
{
// Set the current device to wait for.
ALPAKA_UNIFORM_CUDA_HIP_RT_CHECK(TApi::setDevice(dev.getNativeHandle()));
{
std::lock_guard<std::mutex> lock(*dev.m_mutex);
if(!dev.m_deviceProperties->totalGlobalMem.has_value())
{
// Set the current device to wait for.
ALPAKA_UNIFORM_CUDA_HIP_RT_CHECK(TApi::setDevice(dev.getNativeHandle()));

std::size_t freeInternal(0u);
std::size_t totalInternal(0u);
std::size_t freeInternal(0u);
std::size_t totalInternal(0u);

ALPAKA_UNIFORM_CUDA_HIP_RT_CHECK(TApi::memGetInfo(&freeInternal, &totalInternal));
ALPAKA_UNIFORM_CUDA_HIP_RT_CHECK(TApi::memGetInfo(&freeInternal, &totalInternal));

return freeInternal;
dev.m_deviceProperties->totalGlobalMem = totalInternal;
dev.m_deviceProperties->freeGlobalMem = freeInternal;
}
}

return dev.m_deviceProperties->freeGlobalMem.value();
}
};

Expand All @@ -165,7 +205,15 @@ namespace alpaka
{
ALPAKA_FN_HOST static auto getWarpSizes(DevUniformCudaHipRt<TApi> const& dev) -> std::vector<std::size_t>
{
return {GetPreferredWarpSize<DevUniformCudaHipRt<TApi>>::getPreferredWarpSize(dev)};
{
std::lock_guard<std::mutex> lock(*dev.m_mutex);
if(!dev.m_deviceProperties->warpSizes.has_value())
{
dev.m_deviceProperties->warpSizes = std::vector<std::size_t>{
GetPreferredWarpSize<DevUniformCudaHipRt<TApi>>::getPreferredWarpSize(dev)};
}
}
return dev.m_deviceProperties->warpSizes.value();
}
};

Expand All @@ -175,11 +223,19 @@ namespace alpaka
{
ALPAKA_FN_HOST static auto getPreferredWarpSize(DevUniformCudaHipRt<TApi> const& dev) -> std::size_t
{
int warpSize = 0;

ALPAKA_UNIFORM_CUDA_HIP_RT_CHECK(
TApi::deviceGetAttribute(&warpSize, TApi::deviceAttributeWarpSize, dev.getNativeHandle()));
return static_cast<std::size_t>(warpSize);
{
std::lock_guard<std::mutex> lock(*dev.m_mutex);
if(!dev.m_deviceProperties->preferredWarpSize.has_value())
{
int warpSize = 0;

ALPAKA_UNIFORM_CUDA_HIP_RT_CHECK(
TApi::deviceGetAttribute(&warpSize, TApi::deviceAttributeWarpSize, dev.getNativeHandle()));
dev.m_deviceProperties->preferredWarpSize = static_cast<std::size_t>(warpSize);
}
}

return dev.m_deviceProperties->preferredWarpSize.value();
}
};

Expand Down
Loading