File size: 12,220 Bytes
b024d42 | 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 | // SPDX-License-Identifier: MIT
// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
#pragma once
#include <string>
#include <sstream>
#include <tuple>
#include <vector>
#include <gtest/gtest.h>
#include "ck/ck.hpp"
#include "ck/tensor_operation/gpu/device/impl/device_softmax_impl.hpp"
#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
#include "include/ck/utility/data_type.hpp"
#include "profiler/profile_softmax_impl.hpp"
namespace ck {
template <typename Range>
std::string serialize_range(const Range& range)
{
std::stringstream ss;
for(auto& r : range)
{
ss << r << ", ";
}
std::string str = ss.str();
return std::string(str.begin(), str.end() - 2);
}
template <typename Tuple>
class TestSoftmax : public ::testing::Test
{
protected:
using InDataType = std::tuple_element_t<0, Tuple>;
using AccDataType = std::tuple_element_t<1, Tuple>;
using OutDataType = std::tuple_element_t<2, Tuple>;
static constexpr index_t Rank = std::tuple_element_t<3, Tuple>{}.value;
public:
std::vector<std::vector<index_t>> in_lengths_ = {{2, 128, 1024}, {4, 16, 8448}, {128, 128, 64}};
std::vector<std::vector<AccDataType>> scales_ = {{2, 0}, {0, 2}, {2, 2}};
bool bench_ = false; // measure kernel performance
bool verify_ = true;
void SetUp() override
{
if constexpr(Rank == 4)
{
in_lengths_ = std::vector<std::vector<index_t>>{
{1, 2, 128, 1024}, {2, 4, 16, 8448}, {1, 128, 128, 64}};
}
}
void RunSingle(std::vector<index_t> in_length,
std::vector<index_t> reduce_dims,
AccDataType alpha,
AccDataType beta)
{
int init_method = 1; // integer value initialization
bool log = false;
std::vector<ck::index_t> strides; // intenionally empty, to get packed layout.
bool pass = false;
if constexpr(Rank == 3)
{
if(reduce_dims.size() == 1)
pass = ck::profiler::
profile_softmax_impl<InDataType, AccDataType, OutDataType, Rank, 1>(verify_,
init_method,
log,
bench_,
in_length,
strides,
reduce_dims,
alpha,
beta);
else if(reduce_dims.size() == 2)
pass = ck::profiler::
profile_softmax_impl<InDataType, AccDataType, OutDataType, Rank, 2>(verify_,
init_method,
log,
bench_,
in_length,
strides,
reduce_dims,
alpha,
beta);
else if(reduce_dims.size() == 3)
pass = ck::profiler::
profile_softmax_impl<InDataType, AccDataType, OutDataType, Rank, 3>(verify_,
init_method,
log,
bench_,
in_length,
strides,
reduce_dims,
alpha,
beta);
}
else if constexpr(Rank == 4)
{
if(reduce_dims.size() == 1)
pass = ck::profiler::
profile_softmax_impl<InDataType, AccDataType, OutDataType, Rank, 1>(verify_,
init_method,
log,
bench_,
in_length,
strides,
reduce_dims,
alpha,
beta);
else if(reduce_dims.size() == 2)
pass = ck::profiler::
profile_softmax_impl<InDataType, AccDataType, OutDataType, Rank, 2>(verify_,
init_method,
log,
bench_,
in_length,
strides,
reduce_dims,
alpha,
beta);
else if(reduce_dims.size() == 3)
pass = ck::profiler::
profile_softmax_impl<InDataType, AccDataType, OutDataType, Rank, 3>(verify_,
init_method,
log,
bench_,
in_length,
strides,
reduce_dims,
alpha,
beta);
else if(reduce_dims.size() == 4)
pass = ck::profiler::
profile_softmax_impl<InDataType, AccDataType, OutDataType, Rank, 4>(verify_,
init_method,
log,
bench_,
in_length,
strides,
reduce_dims,
alpha,
beta);
};
EXPECT_TRUE(pass);
}
void Run(std::vector<index_t> reduce_dims = {})
{
if(reduce_dims.empty())
{
reduce_dims.push_back(Rank - 1);
}
for(auto in_length : this->in_lengths_)
{
for(auto scale : this->scales_)
{
this->RunSingle(in_length, reduce_dims, scale[0], scale[1]);
}
}
}
};
template <index_t Rank,
index_t NumReduceDim,
index_t BlockSize,
index_t MThreadClusterSize,
index_t KThreadClusterSize,
index_t MThreadSliceSize,
index_t KThreadSliceSize,
index_t InSrcVectorDim,
index_t InSrcVectorSize,
index_t OutDstVectorSize>
struct DeviceSoftmaxInstanceWrapper
{
using F16 = half_t;
using F32 = float;
using Pass = tensor_operation::element_wise::PassThrough;
using InDataType = F16;
using AccDataType = F32;
using OutDataType = F16;
using InElementOp = Pass;
using AccElementOp = Pass;
using DeviceSoftmaxInstance = tensor_operation::device::DeviceSoftmaxImpl<InDataType,
AccDataType,
OutDataType,
InElementOp,
AccElementOp,
Rank,
NumReduceDim,
BlockSize,
MThreadClusterSize,
KThreadClusterSize,
MThreadSliceSize,
KThreadSliceSize,
InSrcVectorDim,
InSrcVectorSize,
OutDstVectorSize>;
bool IsSupported(const std::vector<index_t> in_lengths,
const std::vector<index_t> in_strides,
const std::vector<index_t> reduce_dims) const
{
auto softmax = DeviceSoftmaxInstance{};
auto argument = softmax.MakeArgument(in_lengths,
in_strides,
reduce_dims,
1, // alpha
1, // beta
nullptr, // in_dev
nullptr, // in_out
Pass{}, // in elementwise op
Pass{}); // acc elementwise op
return softmax.IsSupportedArgument(argument);
}
};
} // namespace ck
|