profile_grouped_conv_fwd.cpp 9.22 KB
Newer Older
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
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.

#include <iostream>
#include <numeric>
#include <initializer_list>
#include <cstdlib>

#include "profiler/include/profile_grouped_conv_fwd_impl.hpp"

namespace {

enum struct ConvLayout
{
    GNHWC_GKYXC_GNHWK, // 0
    NHWGC_KYXGC_NHWGK, // 1
};

enum struct ConvDataType
{
    F32_F32_F32,    // 0
    F16_F16_F16,    // 1
    BF16_BF16_BF16, // 2
    INT8_INT8_INT8, // 3
};

static void print_helper_msg()
{
    std::cout
        // clang-format off
        << "arg1: tensor operation (grouped_conv_fwd: Grouped Convolution Forward)\n"
        << "arg2: data type (0: Input fp32, Weight fp32, Output fp32\n"
        << "                 1: Input fp16, Weight fp16, Output fp16\n"
        << "                 2: Input bf16, Weight bf16, Output bf16\n"
        << "                 3: Input int8, Weight int8, Output int8)\n"
        << "arg3: tensor layout (0: Input[G, N, Hi, Wi, C], Weight[G, K, Y, X, C], Output[G, N, Ho, Wo, K]\n"
        << "                     1: Input[N, Hi, Wi, G, C], Weight[K, Y, X, G, C], Output[N, Ho, Wo, G, K])\n"
        << "arg4: verification (0: no, 1: yes)\n"
        << "arg5: initialization (0: no init, 1: integer value, 2: decimal value)\n"
        << "arg6: print tensor value (0: no; 1: yes)\n"
        << "arg7: time kernel (0: no, 1: yes)\n"
        << ck::utils::conv::get_conv_param_parser_helper_msg() << std::endl;
    // clang-format on
}

} // namespace

int profile_grouped_conv_fwd(int argc, char* argv[])
{
    // 8 for control, 1 for num_dim_spatial
    if(argc < 9)
    {
        print_helper_msg();
        return 1;
    }

    const auto data_type       = static_cast<ConvDataType>(std::stoi(argv[2]));
    const auto layout          = static_cast<ConvLayout>(std::stoi(argv[3]));
    const bool do_verification = std::stoi(argv[4]);
    const int init_method      = std::stoi(argv[5]);
    const bool do_log          = std::stoi(argv[6]);
    const bool time_kernel     = std::stoi(argv[7]);
    const int num_dim_spatial  = std::stoi(argv[8]);

    // 8 for control, 1 for num_dim_spatial, 4 for G/N/K/C, and 6 * num_dim_spatial
    if(argc != 8 + 1 + 4 + 6 * num_dim_spatial)
    {
        print_helper_msg();
        return 1;
    }

    const auto params = ck::utils::conv::parse_conv_param(num_dim_spatial, 9, argv);

    using F32  = float;
    using F16  = ck::half_t;
    using BF16 = ck::bhalf_t;
    using INT8 = int8_t;

    //
    using GNWC   = ck::tensor_layout::convolution::GNWC;
    using GNHWC  = ck::tensor_layout::convolution::GNHWC;
    using GNDHWC = ck::tensor_layout::convolution::GNDHWC;

    using GKXC   = ck::tensor_layout::convolution::GKXC;
    using GKYXC  = ck::tensor_layout::convolution::GKYXC;
    using GKZYXC = ck::tensor_layout::convolution::GKZYXC;

    using GNWK   = ck::tensor_layout::convolution::GNWK;
    using GNHWK  = ck::tensor_layout::convolution::GNHWK;
    using GNDHWK = ck::tensor_layout::convolution::GNDHWK;

    //
    using NWGC   = ck::tensor_layout::convolution::NWGC;
    using NHWGC  = ck::tensor_layout::convolution::NHWGC;
    using NDHWGC = ck::tensor_layout::convolution::NDHWGC;

    using KXGC   = ck::tensor_layout::convolution::KXGC;
    using KYXGC  = ck::tensor_layout::convolution::KYXGC;
    using KZYXGC = ck::tensor_layout::convolution::KZYXGC;

    using NWGK   = ck::tensor_layout::convolution::NWGK;
    using NHWGK  = ck::tensor_layout::convolution::NHWGK;
    using NDHWGK = ck::tensor_layout::convolution::NDHWGK;

    constexpr auto I1 = ck::Number<1>{};
    constexpr auto I2 = ck::Number<2>{};
    constexpr auto I3 = ck::Number<3>{};

    auto profile = [&](auto num_dim_spatial_tmp,
                       auto in_layout,
                       auto wei_layout,
                       auto out_layout,
                       auto in_type,
                       auto wei_type,
                       auto out_type) {
        constexpr ck::index_t NDimSpatial = num_dim_spatial_tmp.value;

        using InLayout  = decltype(in_layout);
        using WeiLayout = decltype(wei_layout);
        using OutLayout = decltype(out_layout);

        using InDataType  = decltype(in_type);
        using WeiDataType = decltype(wei_type);
        using OutDataType = decltype(out_type);

        bool pass = ck::profiler::profile_grouped_conv_fwd_impl<NDimSpatial,
                                                                InLayout,
                                                                WeiLayout,
                                                                OutLayout,
                                                                InDataType,
                                                                WeiDataType,
                                                                OutDataType>(
            do_verification, init_method, do_log, time_kernel, params);

        return pass ? 0 : 1;
    };

    // GNHWC_GKYXC_GNHWK
    if(num_dim_spatial == 1 && layout == ConvLayout::GNHWC_GKYXC_GNHWK)
    {
        if(data_type == ConvDataType::F32_F32_F32)
        {
            return profile(I1, GNWC{}, GKXC{}, GNWK{}, F32{}, F32{}, F32{});
        }
        else if(data_type == ConvDataType::F16_F16_F16)
        {
            return profile(I1, GNWC{}, GKXC{}, GNWK{}, F16{}, F16{}, F16{});
        }
        else if(data_type == ConvDataType::BF16_BF16_BF16)
        {
            return profile(I1, GNWC{}, GKXC{}, GNWK{}, BF16{}, BF16{}, BF16{});
        }
        else if(data_type == ConvDataType::INT8_INT8_INT8)
        {
            return profile(I1, GNWC{}, GKXC{}, GNWK{}, INT8{}, INT8{}, INT8{});
        }
    }
    else if(num_dim_spatial == 2 && layout == ConvLayout::GNHWC_GKYXC_GNHWK)
    {
        if(data_type == ConvDataType::F32_F32_F32)
        {
            return profile(I2, GNHWC{}, GKYXC{}, GNHWK{}, F32{}, F32{}, F32{});
        }
        else if(data_type == ConvDataType::F16_F16_F16)
        {
            return profile(I2, GNHWC{}, GKYXC{}, GNHWK{}, F16{}, F16{}, F16{});
        }
        else if(data_type == ConvDataType::BF16_BF16_BF16)
        {
            return profile(I2, GNHWC{}, GKYXC{}, GNHWK{}, BF16{}, BF16{}, BF16{});
        }
        else if(data_type == ConvDataType::INT8_INT8_INT8)
        {
            return profile(I2, GNHWC{}, GKYXC{}, GNHWK{}, INT8{}, INT8{}, INT8{});
        }
    }
    else if(num_dim_spatial == 3 && layout == ConvLayout::GNHWC_GKYXC_GNHWK)
    {
        if(data_type == ConvDataType::F32_F32_F32)
        {
            return profile(I3, GNDHWC{}, GKZYXC{}, GNDHWK{}, F32{}, F32{}, F32{});
        }
        else if(data_type == ConvDataType::F16_F16_F16)
        {
            return profile(I3, GNDHWC{}, GKZYXC{}, GNDHWK{}, F16{}, F16{}, F16{});
        }
        else if(data_type == ConvDataType::BF16_BF16_BF16)
        {
            return profile(I3, GNDHWC{}, GKZYXC{}, GNDHWK{}, BF16{}, BF16{}, BF16{});
        }
        else if(data_type == ConvDataType::INT8_INT8_INT8)
        {
            return profile(I3, GNDHWC{}, GKZYXC{}, GNDHWK{}, INT8{}, INT8{}, INT8{});
        }
    }
    // NHWGC_KYXGC_NHWGK
    else if(num_dim_spatial == 1 && layout == ConvLayout::NHWGC_KYXGC_NHWGK)
    {
        if(data_type == ConvDataType::F32_F32_F32)
        {
            return profile(I1, NWGC{}, KXGC{}, NWGK{}, F32{}, F32{}, F32{});
        }
        else if(data_type == ConvDataType::F16_F16_F16)
        {
            return profile(I1, NWGC{}, KXGC{}, NWGK{}, F16{}, F16{}, F16{});
        }
        else if(data_type == ConvDataType::BF16_BF16_BF16)
        {
            return profile(I1, NWGC{}, KXGC{}, NWGK{}, BF16{}, BF16{}, BF16{});
        }
        else if(data_type == ConvDataType::INT8_INT8_INT8)
        {
            return profile(I1, NWGC{}, KXGC{}, NWGK{}, INT8{}, INT8{}, INT8{});
        }
    }
    else if(num_dim_spatial == 2 && layout == ConvLayout::NHWGC_KYXGC_NHWGK)
    {
        if(data_type == ConvDataType::F32_F32_F32)
        {
            return profile(I2, NHWGC{}, KYXGC{}, NHWGK{}, F32{}, F32{}, F32{});
        }
        else if(data_type == ConvDataType::F16_F16_F16)
        {
            return profile(I2, NHWGC{}, KYXGC{}, NHWGK{}, F16{}, F16{}, F16{});
        }
        else if(data_type == ConvDataType::BF16_BF16_BF16)
        {
            return profile(I2, NHWGC{}, KYXGC{}, NHWGK{}, BF16{}, BF16{}, BF16{});
        }
        else if(data_type == ConvDataType::INT8_INT8_INT8)
        {
            return profile(I2, NHWGC{}, KYXGC{}, NHWGK{}, INT8{}, INT8{}, INT8{});
        }
    }
    else if(num_dim_spatial == 3 && layout == ConvLayout::NHWGC_KYXGC_NHWGK)
    {
        if(data_type == ConvDataType::F32_F32_F32)
        {
            return profile(I3, NDHWGC{}, KZYXGC{}, NDHWGK{}, F32{}, F32{}, F32{});
        }
        else if(data_type == ConvDataType::F16_F16_F16)
        {
            return profile(I3, NDHWGC{}, KZYXGC{}, NDHWGK{}, F16{}, F16{}, F16{});
        }
        else if(data_type == ConvDataType::BF16_BF16_BF16)
        {
            return profile(I3, NDHWGC{}, KZYXGC{}, NDHWGK{}, BF16{}, BF16{}, BF16{});
        }
        else if(data_type == ConvDataType::INT8_INT8_INT8)
        {
            return profile(I3, NDHWGC{}, KZYXGC{}, NDHWGK{}, INT8{}, INT8{}, INT8{});
        }
    }

    std::cout << "this data_type & layout is not implemented" << std::endl;

    return 1;
}