ATLAS Offline Software
Loading...
Searching...
No Matches
AlpakaExampleAlg.cxx
Go to the documentation of this file.
1//
2// Copyright (C) 2002-2026 CERN for the benefit of the ATLAS collaboration
3//
4
5#include "AlpakaExampleAlg.h"
6
7// Alpaka include(s).
8#include <alpaka/alpaka.hpp>
9#include <alpaka/example/ExampleDefaultAcc.hpp>
10
11// System include(s).
12#include <cstddef>
13#include <cstdint>
14
15namespace {
16
18struct VectorTransformKernel {
19 template < typename TAcc, typename TElem, typename TIdx >
20 ALPAKA_FN_ACC auto operator()(
21 TAcc const& acc,
22 TElem const* const in,
23 TElem* const out,
24 TIdx const numElements ) const -> void {
25 for( auto i : alpaka::uniformElements( acc, numElements ) ) {
26 out[i] = in[i] * 2 + 1;
27 }
28 }
29};
30
31} // namespace
32
34 using Dim = alpaka::DimInt< 1u >;
35 using Idx = std::size_t;
36 using Acc = alpaka::ExampleDefaultAcc< Dim, Idx >;
37
38 ATH_MSG_INFO( "Using Alpaka accelerator: " << alpaka::getAccName< Acc >() );
39 return StatusCode::SUCCESS;
40}
41
42StatusCode AlpakaExampleAlg::execute( const EventContext& /*ctx*/ ) const {
43 using Dim = alpaka::DimInt< 1u >;
44 using Idx = std::size_t;
45 using Acc = alpaka::ExampleDefaultAcc< Dim, Idx >;
46 using DevAcc = alpaka::Dev< Acc >;
47 using QueueAcc = alpaka::Queue< Acc, alpaka::Blocking >;
48 using Data = std::uint32_t;
49
50 const Idx numElements = m_nElements;
51 const alpaka::Vec< Dim, Idx > extent( numElements );
52
53 // Select device and queue.
54 auto const platform = alpaka::Platform< Acc >{};
55 auto const devAcc = alpaka::getDevByIdx( platform, 0 );
56 QueueAcc queue( devAcc );
57
58 // Select host device.
59 using DevHost = alpaka::DevCpu;
60 auto const platformHost = alpaka::PlatformCpu{};
61 auto const devHost = alpaka::getDevByIdx( platformHost, 0 );
62
63 // Allocate host buffers.
64 using BufHost = alpaka::Buf< DevHost, Data, Dim, Idx >;
65 BufHost bufHostIn( alpaka::allocBuf< Data, Idx >( devHost, extent ) );
66 BufHost bufHostOut( alpaka::allocBuf< Data, Idx >( devHost, extent ) );
67
68 // Initialize host input.
69 for( Idx i = 0; i < numElements; ++i ) {
70 bufHostIn[i] = static_cast< Data >( i );
71 bufHostOut[i] = 0;
72 }
73
74 // Allocate device buffers.
75 using BufAcc = alpaka::Buf< DevAcc, Data, Dim, Idx >;
76 BufAcc bufAccIn( alpaka::allocBuf< Data, Idx >( devAcc, extent ) );
77 BufAcc bufAccOut( alpaka::allocBuf< Data, Idx >( devAcc, extent ) );
78
79 // Copy Host -> Device.
80 alpaka::memcpy( queue, bufAccIn, bufHostIn );
81 alpaka::memcpy( queue, bufAccOut, bufHostOut );
82
83 // Run kernel on queue.
84 VectorTransformKernel kernel;
85 const Idx elementsPerThread = 1u;
86 alpaka::KernelCfg< Acc > const kernelCfg = { extent, elementsPerThread };
87 auto const workDiv = alpaka::getValidWorkDiv(
88 kernelCfg, devAcc, kernel,
89 alpaka::getPtrNative( bufAccIn ),
90 alpaka::getPtrNative( bufAccOut ),
91 numElements );
92
93 auto const taskKernel = alpaka::createTaskKernel< Acc >(
94 workDiv, kernel,
95 std::data( bufAccIn ),
96 std::data( bufAccOut ),
97 numElements );
98
99 alpaka::enqueue( queue, taskKernel );
100
101 // Copy Device -> Host.
102 alpaka::memcpy( queue, bufHostOut, bufAccOut );
103 alpaka::wait( queue );
104
105 // Validate results.
106 bool success = true;
107 ATH_MSG_INFO( "Alpaka kernel execution with " << numElements << " elements:" );
108 for( Idx i = 0; i < numElements; ++i ) {
109 Data const expected = static_cast< Data >( i * 2 + 1 );
110 Data const actual = bufHostOut[i];
111 ATH_MSG_INFO( " in[" << i << "] = " << bufHostIn[i]
112 << " -> out[" << i << "] = " << actual
113 << " (expected " << expected << ")" );
114 if( actual != expected ) {
115 success = false;
116 }
117 }
118
119 ATH_MSG_INFO( "Result " << ( success ? "VALIDATED" : "FAILED" ) );
120 return success ? StatusCode::SUCCESS : StatusCode::FAILURE;
121}
122
#define ATH_MSG_INFO(x,...)
@ Data
Definition BaseObject.h:11
void operator()(T1)
virtual StatusCode initialize() override
The function setting up the algorithm.
virtual StatusCode execute(const EventContext &ctx) const override
The function executing this algorithm.
Gaudi::Property< unsigned int > m_nElements
The number of synthetic elements to process per event.