composable_kernel/example/ck_tile/16_batched_gemm/batched_gemm.hpp at develop · ROCm/composable_kernel · GitHub

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
// Copyright (c) Advanced Micro Devices, Inc., or its affiliates.
// SPDX-License-Identifier: MIT

#pragma once

#include <string>

#include "ck_tile/core.hpp"
#include "ck_tile/host/kernel_launch.hpp"
#include "ck_tile/ops/gemm/kernel/batched_gemm_kernel.hpp"
#include "ck_tile/ops/elementwise/unary_element_wise_operation.hpp"
#include "ck_tile/utility/json_dump.hpp"

struct GemmConfigMemory
{
    // Memory friendly for Interwave scheduler
    static constexpr ck_tile::index_t M_Tile = 128;
    static constexpr ck_tile::index_t N_Tile = 32;
    static constexpr ck_tile::index_t K_Tile = 64;

    static constexpr ck_tile::index_t M_Warp = 4;
    static constexpr ck_tile::index_t N_Warp = 1;
    static constexpr ck_tile::index_t K_Warp = 1;

    static constexpr ck_tile::index_t M_Warp_Tile = 32;
    static constexpr ck_tile::index_t N_Warp_Tile = 32;
    static constexpr ck_tile::index_t K_Warp_Tile = 8;

    static constexpr bool DoubleSmemBuffer          = false;
    static constexpr ck_tile::GemmPipeline Pipeline = ck_tile::GemmPipeline::MEMORY;
    static constexpr auto Scheduler                 = ck_tile::GemmPipelineScheduler::Interwave;
};

struct GemmConfigV3
{
    // Compute friendly for Intrawave scheduler
    static constexpr ck_tile::index_t M_Tile = 256;
    static constexpr ck_tile::index_t N_Tile = 256;
    static constexpr ck_tile::index_t K_Tile = 64;

    static constexpr ck_tile::index_t M_Warp = 2;
    static constexpr ck_tile::index_t N_Warp = 2;
    static constexpr ck_tile::index_t K_Warp = 1;

    static constexpr ck_tile::index_t M_Warp_Tile = 32;
    static constexpr ck_tile::index_t N_Warp_Tile = 32;
    static constexpr ck_tile::index_t K_Warp_Tile = 16;

    static constexpr bool DoubleSmemBuffer          = false;
    static constexpr ck_tile::GemmPipeline Pipeline = ck_tile::GemmPipeline::COMPUTE_V3;
    static constexpr auto Scheduler                 = ck_tile::GemmPipelineScheduler::Intrawave;
};

struct GemmConfigV4
{
    // Compute friendly for Intrawave scheduler
    // Using the ping pong reader in the lds level
    static constexpr ck_tile::index_t M_Tile = 256;
    static constexpr ck_tile::index_t N_Tile = 256;
    static constexpr ck_tile::index_t K_Tile = 32;

    static constexpr ck_tile::index_t M_Warp = 2;
    static constexpr ck_tile::index_t N_Warp = 2;
    static constexpr ck_tile::index_t K_Warp = 1;

    static constexpr ck_tile::index_t M_Warp_Tile = 32;
    static constexpr ck_tile::index_t N_Warp_Tile = 32;
    static constexpr ck_tile::index_t K_Warp_Tile = 16;

    static constexpr bool DoubleSmemBuffer          = true;
    static constexpr ck_tile::GemmPipeline Pipeline = ck_tile::GemmPipeline::COMPUTE_V4;
    static constexpr auto Scheduler                 = ck_tile::GemmPipelineScheduler::Intrawave;
};

struct GemmConfigV3_Wmma
{
    // Compute friendly for Intrawave scheduler
    static constexpr ck_tile::index_t M_Tile = 128;
    static constexpr ck_tile::index_t N_Tile = 128;
    static constexpr ck_tile::index_t K_Tile = 64;

    static constexpr ck_tile::index_t M_Warp = 2;
    static constexpr ck_tile::index_t N_Warp = 2;
    static constexpr ck_tile::index_t K_Warp = 1;

    static constexpr ck_tile::index_t M_Warp_Tile = 16;
    static constexpr ck_tile::index_t N_Warp_Tile = 16;
    static constexpr ck_tile::index_t K_Warp_Tile = 16;

    static constexpr bool DoubleSmemBuffer          = false;
    static constexpr ck_tile::GemmPipeline Pipeline = ck_tile::GemmPipeline::COMPUTE_V3;
    static constexpr auto Scheduler                 = ck_tile::GemmPipelineScheduler::Intrawave;
};

template <ck_tile::GemmPipeline PipelineId>
struct PipelineTypeTraits;

template <>
struct PipelineTypeTraits<ck_tile::GemmPipeline::MEMORY>
{
    template <typename PipelineProblem>
    using GemmPipeline = ck_tile::GemmPipelineAgBgCrMem<PipelineProblem>;
};

template <>
struct PipelineTypeTraits<ck_tile::GemmPipeline::COMPUTE_V3>
{
    template <typename PipelineProblem>
    using GemmPipeline = ck_tile::GemmPipelineAgBgCrCompV3<PipelineProblem>;
};

template <>
struct PipelineTypeTraits<ck_tile::GemmPipeline::COMPUTE_V4>
{
    template <typename PipelineProblem>
    using GemmPipeline = ck_tile::GemmPipelineAgBgCrCompV4<PipelineProblem>;
};

template <typename DataType>
struct BatchedGemmTypeConfig;

template <>
struct BatchedGemmTypeConfig<ck_tile::half_t>
{
    using ADataType   = ck_tile::half_t;
    using BDataType   = ck_tile::half_t;
    using AccDataType = float;
    using CDataType   = ck_tile::half_t;
};

using Types = BatchedGemmTypeConfig<ck_tile::half_t>;

// Specific type aliases for easy access
using ADataType   = Types::ADataType;
using BDataType   = Types::BDataType;
using AccDataType = Types::AccDataType;
using CDataType   = Types::CDataType;

auto create_args(int argc, char* argv[])
{
    ck_tile::ArgParser arg_parser;
    arg_parser.insert("m", "512", "m dimension")
        .insert("n", "1024", "n dimension")
        .insert("k", "2048", "k dimension")
        .insert("stride_a", "0", "Tensor A stride")
        .insert("stride_b", "0", "Tensor B stride")
        .insert("stride_c", "0", "Tensor C stride")
        .insert("a_layout", "R", "A tensor data layout - Row by default")
        .insert("b_layout", "C", "B tensor data layout - Row by default")
        .insert("c_layout", "R", "C tensor data layout - Row by default")
        .insert("batch_stride_a", "1048576", "Batch A stride")
        .insert("batch_stride_b", "2097152", "Batch B stride")
        .insert("batch_stride_c", "524288", "Batch C stride")
        .insert("batch_count", "8", "Batch count")
        .insert("v", "1", "0. No validation, 1. Validation on CPU, 2. Validation on GPU")
        .insert("prec", "fp16", "data type. fp16/bf16/fp8/bf8")
        .insert("warmup", "50", "number of iterations before benchmark the kernel")
        .insert("repeat", "100", "number of iterations to benchmark the kernel")
        .insert("timer", "gpu", "gpu:gpu timer, cpu:cpu timer")
        .insert("split_k", "1", "splitK value")
        .insert("json", "0", "0: No Json, 1: Dump Results in Json format")
        .insert("jsonfile", "cktile_batched_gemm.json", "json file name to dump results");

    bool result = arg_parser.parse(argc, argv);
    return std::make_tuple(result, arg_parser);
}

// host API
float batched_gemm(const ck_tile::BatchedGemmHostArgs& args, const ck_tile::stream_config& s);