Skip to content

Commit

Permalink
Rewrite the 3D buffer copy example using different uniformElements loops
Browse files Browse the repository at this point in the history
  • Loading branch information
@fwyzard
fwyzard committed Sep 13, 2024
1 parent b2f2f2f commit cc40198
Showing 1 changed file with 36 additions and 39 deletions.
75 changes: 36 additions & 39 deletions example/bufferCopy/src/bufferCopy.cpp
View file
Original file line number Diff line number Diff line change
@@ -1,5 +1,5 @@
/* Copyright 2023 Alexander Matthes, Benjamin Worpitz, Erik Zenker, Matthias Werner, Bernhard Manfred Gruber,
* Jan Stephan
/* Copyright 2024 Alexander Matthes, Benjamin Worpitz, Erik Zenker, Matthias Werner, Bernhard Manfred Gruber,
* Jan Stephan, Andrea Bocci
* SPDX-License-Identifier: ISC
*/

Expand All @@ -15,13 +15,12 @@ struct PrintBufferKernel
template<typename TAcc, typename MdSpan>
ALPAKA_FN_ACC auto operator()(TAcc const& acc, MdSpan data) const -> void
{
auto const idx = alpaka::getIdx<alpaka::Grid, alpaka::Threads>(acc);
auto const gridSize = alpaka::getWorkDiv<alpaka::Grid, alpaka::Threads>(acc);

for(size_t z = idx[0]; z < data.extent(0); z += gridSize[0])
for(size_t y = idx[1]; y < data.extent(1); y += gridSize[1])
for(size_t x = idx[2]; x < data.extent(2); x += gridSize[2])
printf("%zu,%zu,%zu:%u ", z, y, x, static_cast<uint32_t>(data(z, y, x)));
// Use three nested loops along the dimensions 0, 1 and 2
for(size_t z : alpaka::uniformElementsAlong<0>(acc, data.extent(0)))
for(size_t y : alpaka::uniformElementsAlong<1>(acc, data.extent(1)))
for(size_t x : alpaka::uniformElementsAlong<2>(acc, data.extent(2)))
// %zu prints garbage in some cases, while %lu seems to be working correctly
printf("%lu,%lu,%lu: %u\t", z, y, x, static_cast<uint32_t>(data(z, y, x)));
}
};

Expand All @@ -31,12 +30,10 @@ struct TestBufferKernel
template<typename TAcc, typename MdSpan>
ALPAKA_FN_ACC auto operator()(TAcc const& acc, MdSpan data) const -> void
{
auto const idx = alpaka::getIdx<alpaka::Grid, alpaka::Threads>(acc);
auto const gridSize = alpaka::getWorkDiv<alpaka::Grid, alpaka::Threads>(acc);

for(size_t z = idx[0]; z < data.extent(0); z += gridSize[0])
for(size_t y = idx[1]; y < data.extent(1); y += gridSize[1])
for(size_t x = idx[2]; x < data.extent(2); x += gridSize[2])
// Use three nested loops along the dimensions z, y and x
for(size_t z : alpaka::uniformElementsAlongZ(acc, data.extent(0)))
for(size_t y : alpaka::uniformElementsAlongY(acc, data.extent(1)))
for(size_t x : alpaka::uniformElementsAlongX(acc, data.extent(2)))
ALPAKA_ASSERT_ACC(
data(z, y, x)
== alpaka::mapIdx<1u>(
Expand All @@ -51,16 +48,10 @@ struct FillBufferKernel
template<typename TAcc, typename MdSpan>
ALPAKA_FN_ACC auto operator()(TAcc const& acc, MdSpan data) const -> void
{
using Vec = alpaka::Vec<alpaka::Dim<TAcc>, alpaka::Idx<TAcc>>;

auto const idx = alpaka::getIdx<alpaka::Grid, alpaka::Threads>(acc);
auto const gridSize = alpaka::getWorkDiv<alpaka::Grid, alpaka::Threads>(acc);

for(size_t z = idx[0]; z < data.extent(0); z += gridSize[0])
for(size_t y = idx[1]; y < data.extent(1); y += gridSize[1])
for(size_t x = idx[2]; x < data.extent(2); x += gridSize[2])
data(z, y, x)
= alpaka::mapIdx<1u>(Vec{z, y, x}, Vec{data.extent(0), data.extent(1), data.extent(2)})[0];
// Use a single 3-dimensional loop
for(auto idx : alpaka::uniformElementsND(acc, alpaka::Vec{data.extent(0), data.extent(1), data.extent(2)}))
data(idx.z(), idx.y(), idx.x()) // equivalent to data(idx[0], idx[1], idx[2])
= alpaka::mapIdx<1u>(idx, alpaka::Vec{data.extent(0), data.extent(1), data.extent(2)})[0];
}
};

Expand All @@ -78,15 +69,15 @@ auto example(TAccTag const&) -> int
// Define the device accelerator
using Acc = alpaka::TagToAcc<TAccTag, Dim, Idx>;
std::cout << "Using alpaka accelerator: " << alpaka::getAccName<Acc>() << std::endl;
// Defines the synchronization behavior of a queue
// Defines the synchronization behavior of the device queue
//
// choose between Blocking and NonBlocking
using AccQueueProperty = alpaka::Blocking;
using DevQueue = alpaka::Queue<Acc, AccQueueProperty>;

// Define the device accelerator
// Define the host accelerator
using Host = alpaka::AccCpuSerial<Dim, Idx>;
// Defines the synchronization behavior of a queue
// Defines the synchronization behavior of the host queue
//
// choose between Blocking and NonBlocking
using HostQueueProperty = alpaka::Blocking;
Expand Down Expand Up @@ -118,14 +109,14 @@ auto example(TAccTag const&) -> int
using Data = std::uint32_t;
constexpr Idx nElementsPerDim = 2;

Vec const extents(Vec::all(static_cast<Idx>(nElementsPerDim)));
Vec const extents = Vec::all(nElementsPerDim);

// Allocate host memory buffers
//
// The `alloc` method returns a reference counted buffer handle.
// When the last such handle is destroyed, the memory is freed automatically.
using BufHost = alpaka::Buf<Host, Data, Dim, Idx>;
BufHost hostBuffer(alpaka::allocBuf<Data, Idx>(devHost, extents));
BufHost hostBuffer = alpaka::allocBuf<Data, Idx>(devHost, extents);
// You can also use already allocated memory and wrap it within a view (irrespective of the device type).
// The view does not own the underlying memory. So you have to make sure that
// the view does not outlive its underlying memory.
Expand All @@ -136,8 +127,8 @@ auto example(TAccTag const&) -> int
//
// The interface to allocate a buffer is the same on the host and on the device.
using BufAcc = alpaka::Buf<Acc, Data, Dim, Idx>;
BufAcc deviceBuffer1(alpaka::allocBuf<Data, Idx>(devAcc, extents));
BufAcc deviceBuffer2(alpaka::allocBuf<Data, Idx>(devAcc, extents));
BufAcc deviceBuffer1 = alpaka::allocBuf<Data, Idx>(devAcc, extents);
BufAcc deviceBuffer2 = alpaka::allocBuf<Data, Idx>(devAcc, extents);


// Init host buffer
Expand All @@ -152,9 +143,9 @@ auto example(TAccTag const&) -> int
// some values into the buffer memory.
// Mind, that only a host can write on host memory.
// The same holds true for device memory.
for(Idx z(0); z < extents[0]; ++z)
for(Idx y(0); y < extents[1]; ++y)
for(Idx x(0); x < extents[2]; ++x)
for(Idx z = 0; z < extents[0]; ++z)
for(Idx y = 0; y < extents[1]; ++y)
for(Idx x = 0; x < extents[2]; ++x)
hostBufferMdSpan(z, y, x) = static_cast<Data>(z * extents[1] * extents[2] + y * extents[2] + x);

// Memory views and buffers can also be initialized by executing a kernel.
Expand Down Expand Up @@ -221,19 +212,25 @@ auto example(TAccTag const&) -> int
// completely distorted.

PrintBufferKernel printBufferKernel;
alpaka::exec<Acc>(devQueue, devWorkDiv, printBufferKernel, deviceBufferMdSpan1);

// Let alpaka calculate good block and grid sizes given our full problem extent
auto const hostPrintWorkDiv
= alpaka::getValidWorkDiv(hostKernelCfg, devHost, printBufferKernel, hostViewPlainPtrMdSpan);
auto const devPrintWorkDiv = alpaka::getValidWorkDiv(devKernelCfg, devAcc, printBufferKernel, deviceBufferMdSpan1);

alpaka::exec<Acc>(devQueue, devPrintWorkDiv, printBufferKernel, deviceBufferMdSpan1);
alpaka::wait(devQueue);
std::cout << std::endl;

alpaka::exec<Acc>(devQueue, devWorkDiv, printBufferKernel, deviceBufferMdSpan2);
alpaka::exec<Acc>(devQueue, devPrintWorkDiv, printBufferKernel, deviceBufferMdSpan2);
alpaka::wait(devQueue);
std::cout << std::endl;

alpaka::exec<Host>(hostQueue, hostWorkDiv, printBufferKernel, hostBufferMdSpan);
alpaka::exec<Host>(hostQueue, hostPrintWorkDiv, printBufferKernel, hostBufferMdSpan);
alpaka::wait(hostQueue);
std::cout << std::endl;

alpaka::exec<Host>(hostQueue, hostWorkDiv, printBufferKernel, hostViewPlainPtrMdSpan);
alpaka::exec<Host>(hostQueue, hostPrintWorkDiv, printBufferKernel, hostViewPlainPtrMdSpan);
alpaka::wait(hostQueue);
std::cout << std::endl;

Expand Down

0 comments on commit cc40198

Please sign in to comment.