softmax4d.cpp 5.27 KB
Newer Older
Adam Osewski's avatar
Adam Osewski committed
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
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.

#include <functional>
#include <numeric>
#include <iomanip>
#include <iostream>
#include <vector>

#include "ck/ck.hpp"
#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
#include "ck/tensor_operation/gpu/device/device_softmax.hpp"
#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"

#include "ck/library/tensor_operation_instance/gpu/softmax.hpp"

using InDataType  = ck::half_t;
using OutDataType = ck::half_t;
using AccDataType = float;
using PassThrough = ck::tensor_operation::element_wise::PassThrough;

constexpr int Rank         = 4;
constexpr int NumReduceDim = 2;

struct SimpleDeviceMem
{
    SimpleDeviceMem() = delete;

    SimpleDeviceMem(std::size_t mem_size) : p_mem_{}
    {
        (void)hipMalloc(static_cast<void**>(&p_mem_), mem_size);
    }

    void* GetDeviceBuffer() { return p_mem_; }

    ~SimpleDeviceMem() { (void)hipFree(p_mem_); }

    void* p_mem_;
};

int main(int argc, char* argv[])
{
    std::vector<ck::index_t> in_lengths{2, 8, 128, 1024};
    std::vector<ck::index_t> in_strides{8 * 128 * 1024, 128 * 1024, 1024, 1};
    std::vector<ck::index_t> reduce_dims{2, 3};

    ck::index_t num_elements =
        std::accumulate(in_lengths.begin(), in_lengths.end(), 1, std::multiplies<ck::index_t>());

50
51
    double alpha{2.0};
    double beta{2.0};
Adam Osewski's avatar
Adam Osewski committed
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

    SimpleDeviceMem in(sizeof(InDataType) * num_elements);
    SimpleDeviceMem out(sizeof(OutDataType) * num_elements);

    using DeviceOp = ck::tensor_operation::device::
        DeviceSoftmax<InDataType, AccDataType, OutDataType, PassThrough, PassThrough, Rank>;
    // get device op instances
    const auto op_ptrs = ck::tensor_operation::device::instance::DeviceOperationInstanceFactory<
        DeviceOp>::GetInstances();

    std::cout << "found " << op_ptrs.size() << " instances" << std::endl;

    std::string best_op_name;
    bool found            = false;
    int best_op_id        = -1;
    float best_ave_time   = std::numeric_limits<float>::max();
    float best_gb_per_sec = 0;

    // profile device operation instances
    std::cout << "Run all instances and do timing" << std::endl;

    for(int i = 0; i < op_ptrs.size(); ++i)
    {
        auto& op_ptr = op_ptrs[i];

        if(op_ptr->GetRank() != Rank || op_ptr->GetNumReduceDim() != NumReduceDim)
        {
            continue;
        }

        auto argument_ptr   = op_ptr->MakeArgumentPointer(in_lengths,
                                                        in_strides,
                                                        reduce_dims,
85
86
                                                        alpha,
                                                        beta,
Adam Osewski's avatar
Adam Osewski committed
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
                                                        in.GetDeviceBuffer(),
                                                        out.GetDeviceBuffer(),
                                                        PassThrough{},
                                                        PassThrough{});
        auto invoker_ptr    = op_ptr->MakeInvokerPointer();
        std::string op_name = op_ptr->GetTypeString();

        if(op_ptr->IsSupportedArgument(argument_ptr.get()))
        {
            float ave_time = invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, true});

            std::size_t num_bytes = num_elements * sizeof(InDataType) +
                                    (beta == 0.0f ? 1 : 2) * num_elements * sizeof(OutDataType);

            float gb_per_sec = num_bytes / 1.E6 / ave_time;

            std::cout << "Perf: " << std::setw(10) << ave_time << " ms, " << gb_per_sec << " GB/s, "
                      << op_name << std::endl;

            if(ave_time < best_ave_time)
            {
                found           = true;
                best_op_id      = i;
                best_op_name    = op_name;
                best_ave_time   = ave_time;
                best_gb_per_sec = gb_per_sec;
            }
        }
        else
        {
            std::cout << op_name << " does not support this problem" << std::endl;
        }
    }

    std::cout << "Best Perf: " << best_ave_time << " ms, " << best_gb_per_sec << " GB/s, "
              << best_op_name << std::endl;

    // run the best intance
    {
        auto& op_ptr = op_ptrs[best_op_id];
        std::cout << "Run the best instance without timing: " << op_ptr->GetTypeString()
                  << std::endl;
        auto argument_ptr = op_ptr->MakeArgumentPointer(in_lengths,
                                                        in_strides,
                                                        reduce_dims,
132
133
                                                        alpha,
                                                        beta,
Adam Osewski's avatar
Adam Osewski committed
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
                                                        in.GetDeviceBuffer(),
                                                        out.GetDeviceBuffer(),
                                                        PassThrough{},
                                                        PassThrough{});

        auto invoker_ptr = op_ptr->MakeInvokerPointer();

        if(op_ptr->IsSupportedArgument(argument_ptr.get()))
        {
            invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, false});
        }

        std::cout << "Done" << std::endl;
    }

    return 0;
150
}