aboutsummaryrefslogtreecommitdiff
path: root/src/backends/cl/ClContextControl.cpp
blob: 7ab825f59e4c21b6e208924a1fd65fbc2967db59 (plain)
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
//
// Copyright © 2017 Arm Ltd. All rights reserved.
// SPDX-License-Identifier: MIT
//

#include "ClContextControl.hpp"

#include <armnn/Exceptions.hpp>

#include <LeakChecking.hpp>

#include <armnn/utility/Assert.hpp>
#include <armnn/utility/IgnoreUnused.hpp>

#include <arm_compute/core/CL/CLKernelLibrary.h>
#include <arm_compute/runtime/CL/CLScheduler.h>

#include <fmt/format.h>

namespace cl
{
class Context;
class CommandQueue;
class Device;
}

namespace armnn
{

ClContextControl::ClContextControl(arm_compute::CLTuner *tuner,
                                   bool profilingEnabled)
    : m_Tuner(tuner)
    , m_ProfilingEnabled(profilingEnabled)
{
    // Ignore m_ProfilingEnabled if unused to avoid compiling problems when ArmCompute is disabled.
    IgnoreUnused(m_ProfilingEnabled);

    try
    {
        std::vector<cl::Platform> platforms;
        cl::Platform::get(&platforms);

        // Selects default platform for the first element.
        cl::Platform::setDefault(platforms[0]);

        std::vector<cl::Device> devices;
        platforms[0].getDevices(CL_DEVICE_TYPE_GPU, &devices);

        // Selects default device for the first element.
        cl::Device::setDefault(devices[0]);
    }
    catch (const cl::Error& clError)
    {
        throw ClRuntimeUnavailableException(fmt::format(
            "Could not initialize the CL runtime. Error description: {0}. CL error code: {1}",
            clError.what(), clError.err()));
    }

    // Removes the use of global CL context.
    cl::Context::setDefault(cl::Context{});
    ARMNN_ASSERT(cl::Context::getDefault()() == NULL);

    // Removes the use of global CL command queue.
    cl::CommandQueue::setDefault(cl::CommandQueue{});
    ARMNN_ASSERT(cl::CommandQueue::getDefault()() == NULL);

    // Always load the OpenCL runtime.
    LoadOpenClRuntime();
}

ClContextControl::~ClContextControl()
{
    // Load the OpencCL runtime without the tuned parameters to free the memory for them.
    try
    {
        UnloadOpenClRuntime();
    }
    catch (const cl::Error& clError)
    {
        // This should not happen, it is ignored if it does.

        // Coverity fix: BOOST_LOG_TRIVIAL (previously used here to report the error) may throw an
        // exception of type std::length_error.
        // Using stderr instead in this context as there is no point in nesting try-catch blocks here.
        std::cerr << "A CL error occurred unloading the runtime tuner parameters: "
                  << clError.what() << ". CL error code is: " << clError.err() << std::endl;
    }
}

void ClContextControl::LoadOpenClRuntime()
{
    DoLoadOpenClRuntime(true);
}

void ClContextControl::UnloadOpenClRuntime()
{
    DoLoadOpenClRuntime(false);
}

void ClContextControl::DoLoadOpenClRuntime(bool updateTunedParameters)
{
    cl::Device device = cl::Device::getDefault();
    cl::Context context;
    cl::CommandQueue commandQueue;

    if (arm_compute::CLScheduler::get().is_initialised() && arm_compute::CLScheduler::get().context()() != NULL)
    {
        // Wait for all queued CL requests to finish before reinitialising it.
        arm_compute::CLScheduler::get().sync();
    }

    try
    {
        arm_compute::CLKernelLibrary::get().clear_programs_cache();
        // Initialise the scheduler with a dummy context to release the LLVM data (which only happens when there are no
        // context references); it is initialised again, with a proper context, later.
        arm_compute::CLScheduler::get().init(context, commandQueue, device);
        arm_compute::CLKernelLibrary::get().init(".", context, device);

        {
            //
            // Here we replace the context with a new one in which
            // the memory leak checks show it as an extra allocation but
            // because of the scope of the leak checks, it doesn't count
            // the disposal of the original object. On the other hand it
            // does count the creation of this context which it flags
            // as a memory leak. By adding the following line we prevent
            // this to happen.
            //
            ARMNN_DISABLE_LEAK_CHECKING_IN_SCOPE();
            context = cl::Context(device);
        }

        // NOTE: In this specific case profiling has to be enabled on the command queue
        // in order for the CLTuner to work.
        bool profilingNeededForClTuner = updateTunedParameters && m_Tuner &&
            m_Tuner->tune_new_kernels();

        if (m_ProfilingEnabled || profilingNeededForClTuner)
        {
            // Create a new queue with profiling enabled.
            commandQueue = cl::CommandQueue(context, device, CL_QUEUE_PROFILING_ENABLE);
        }
        else
        {
            // Use default queue.
            commandQueue = cl::CommandQueue(context, device);
        }
    }
    catch (const cl::Error& clError)
    {
        throw ClRuntimeUnavailableException(fmt::format(
            "Could not initialize the CL runtime. Error description: {0}. CL error code: {1}",
            clError.what(), clError.err()));
    }

    // Note the first argument (path to cl source code) will be ignored as they should be embedded in the armcompute.
    arm_compute::CLKernelLibrary::get().init(".", context, device);
    arm_compute::CLScheduler::get().init(context, commandQueue, device, m_Tuner);
}

void ClContextControl::ClearClCache()
{
    DoLoadOpenClRuntime(true);
}

armnn::IGpuAccTunedParameters* IGpuAccTunedParameters::CreateRaw(armnn::IGpuAccTunedParameters::Mode mode,
                                                                 armnn::IGpuAccTunedParameters::TuningLevel tuningLevel)
{
    return new ClTunedParameters(mode, tuningLevel);
}

armnn::IGpuAccTunedParametersPtr IGpuAccTunedParameters::Create(armnn::IGpuAccTunedParameters::Mode mode,
                                                                armnn::IGpuAccTunedParameters::TuningLevel tuningLevel)
{
    return IGpuAccTunedParametersPtr(CreateRaw(mode, tuningLevel), &IGpuAccTunedParameters::Destroy);
}

void IGpuAccTunedParameters::Destroy(IGpuAccTunedParameters* params)
{
    delete params;
}

ClTunedParameters::ClTunedParameters(armnn::IGpuAccTunedParameters::Mode mode,
                                     armnn::IGpuAccTunedParameters::TuningLevel tuningLevel)
    : m_Mode(mode)
    , m_TuningLevel(tuningLevel)
    , m_Tuner(mode == ClTunedParameters::Mode::UpdateTunedParameters)
{
}

void ClTunedParameters::Load(const char* filename)
{
    try
    {
        m_Tuner.load_from_file(filename);
    }
    catch (const std::exception& e)
    {
        throw armnn::Exception(std::string("Failed to load tuned parameters file '") + filename + "': " +
                               e.what());
    }
}

void ClTunedParameters::Save(const char* filename) const
{
    try
    {
        m_Tuner.save_to_file(filename);
    }
    catch (const std::exception& e)
    {
        throw armnn::Exception(std::string("Failed to save tuned parameters file to '") + filename + "': " +
                               e.what());
    }
}

} // namespace armnn