File size: 3,810 Bytes
4a28d4d
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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

#include <cuda_runtime.h>
#include <cuda_runtime_api.h>

#include "src/turbomind/core/allocator.h"
#include "src/turbomind/core/check.h"

namespace turbomind::core {

AllocatorImpl::~AllocatorImpl() = default;

Stream AllocatorImpl::stream() const noexcept
{
    return Stream{};
}

class CudaMemPoolAllocator: public AllocatorImpl {
public:
    CudaMemPoolAllocator(Stream stream, bool use_default_pool):
        pool_{}, stream_{stream}, device_{kDEVICE}, use_default_pool_{use_default_pool}
    {
        check_cuda_error(cudaGetDevice(&device_.id));
        if (use_default_pool_) {
            check_cuda_error(cudaDeviceGetDefaultMemPool(&pool_, device_.id));
        }
        else {
            cudaMemPoolProps props{};
            props.allocType     = cudaMemAllocationTypePinned;
            props.handleTypes   = cudaMemHandleTypeNone;
            props.location.type = cudaMemLocationTypeDevice;
            props.location.id   = device_.id;
            check_cuda_error(cudaMemPoolCreate(&pool_, &props));
            cuuint64_t thres = (cuuint64_t)-1;
            check_cuda_error(cudaMemPoolSetAttribute(pool_, cudaMemPoolAttrReleaseThreshold, &thres));
        }
    }

    ~CudaMemPoolAllocator() override
    {
        if (!use_default_pool_) {
            check_cuda_error(cudaMemPoolDestroy(pool_));
        }
        pool_ = {};
    }

    void* allocate(ssize_t size) override
    {
        void* ptr{};
        check_cuda_error(cudaMallocFromPoolAsync(&ptr, size, pool_, stream_.handle()));
        return ptr;
    }

    void deallocate(void* p, ssize_t) override
    {
        check_cuda_error(cudaFreeAsync(p, stream_.handle()));
    }

    Device device() const noexcept override
    {
        return device_;
    }

    Stream stream() const noexcept override
    {
        return stream_;
    }

    void trim(size_t bytes_to_keep)
    {
        check_cuda_error(cudaMemPoolTrimTo(pool_, bytes_to_keep));
    }

private:
    cudaMemPool_t pool_;
    Stream        stream_;
    Device        device_;
    bool          use_default_pool_;
};

class CudaAllocator: public AllocatorImpl {
public:
    void* allocate(ssize_t size) override
    {
        void* ptr{};
        check_cuda_error(cudaMalloc(&ptr, size));
        return ptr;
    }

    void deallocate(void* p, ssize_t) override
    {
        check_cuda_error(cudaFree(p));
    }

    Device device() const noexcept override
    {
        return kDEVICE;
    }
};

class CudaHostAllocator: public AllocatorImpl {
public:
    void* allocate(ssize_t size) override
    {
        void* ptr{};
        check_cuda_error(cudaHostAlloc(&ptr, size, cudaHostAllocDefault));
        return ptr;
    }

    void deallocate(void* p, ssize_t) override
    {
        check_cuda_error(cudaFreeHost(p));
    }

    Device device() const noexcept override
    {
        return kCPUpinned;
    }
};

class HostAllocator: public AllocatorImpl {
public:
    void* allocate(ssize_t size) override
    {
        return ::operator new(size);
    }

    void deallocate(void* p, ssize_t) override
    {
        ::operator delete(p);
    }

    Device device() const noexcept override
    {
        return kCPU;
    }
};

Allocator::Allocator(DeviceType type)
{
    impl_ = [&]() -> shared_ptr<AllocatorImpl> {
        switch (type) {
            case kCPU:
                return std::make_shared<HostAllocator>();
            case kDEVICE:
                return std::make_shared<CudaAllocator>();
            case kCPUpinned:
                return std::make_shared<CudaHostAllocator>();
        }
        return {};
    }();
    TM_CHECK_NOTNULL(impl_);
}

Allocator::Allocator(Stream stream, bool use_default_pool)
{
    impl_ = std::make_shared<CudaMemPoolAllocator>(std::move(stream), use_default_pool);
}

}  // namespace turbomind::core