Skip to content

GitLab

"build_tools/vscode:/vscode.git/clone" did not exist on "ed175137faec87ed308216754a115d3db96c1852"
Commit 68886f7d authored by raman jana's avatar raman jana
Browse files

merging with latest develop branch

parents a9ee2960 1677cf70
Hide whitespace changes
Inline Side-by-side
Showing with 469 additions and 993 deletions
test/gemm/gemm_xdl_fp64.cpp 0 → 100644
View file @ 68886f7d
#include <algorithm>
#include <cstdlib>
#include <half.hpp>
#include <iostream>
#include <numeric>
#include <tuple>
#include <vector>
#include "gemm_util.hpp"
#include "config.hpp"
#include "print.hpp"
#include "device.hpp"
#include "host_tensor.hpp"
#include "host_tensor_generator.hpp"
#include "host_gemm.hpp"
#include "device_tensor.hpp"
#include "device_gemm_xdl.hpp"
#include "element_wise_operation.hpp"
#include "reference_gemm.hpp"
#include "gemm_specialization.hpp"
using PassThrough = ck::tensor_operation::element_wise::PassThrough;
using DeviceGemmNoOpPtr =
ck::tensor_operation::device::DeviceGemmPtr<ck::tensor_operation::element_wise::PassThrough,
ck::tensor_operation::element_wise::PassThrough,
ck::tensor_operation::element_wise::PassThrough>;
namespace ck {
namespace tensor_operation {
namespace device {
namespace device_gemm_instance {
void add_device_gemm_xdl_f64_f64_f64_km_kn_mn_instances(std::vector<DeviceGemmNoOpPtr>&);
void add_device_gemm_xdl_f64_f64_f64_km_nk_mn_instances(std::vector<DeviceGemmNoOpPtr>&);
void add_device_gemm_xdl_f64_f64_f64_mk_nk_mn_instances(std::vector<DeviceGemmNoOpPtr>&);
void add_device_gemm_xdl_f64_f64_f64_mk_kn_mn_instances(std::vector<DeviceGemmNoOpPtr>&);
} // namespace device_gemm_instance
} // namespace device
} // namespace tensor_operation
} // namespace ck
inline std::string get_device_name()
{
hipDeviceProp_t props{};
int device;
auto status = hipGetDevice(&device);
if(status != hipSuccess)
{
return std::string();
}
status = hipGetDeviceProperties(&props, device);
if(status != hipSuccess)
{
return std::string();
}
const std::string name(props.gcnArchName);
return name;
}
int main()
{
if(get_device_name().find("gfx90a") == std::string::npos)
{
std::cout << "TestGemm ..... SUCCESS" << std::endl;
return 0;
}
using ADataType = double;
using BDataType = double;
using CDataType = double;
using AccDataType = double;
using RowMajor = ck::tensor_layout::gemm::RowMajor;
using ColumnMajor = ck::tensor_layout::gemm::ColumnMajor;
bool res = true;
std::vector<DeviceGemmNoOpPtr> gemmPtrs;
ck::tensor_operation::device::device_gemm_instance::
add_device_gemm_xdl_f64_f64_f64_km_kn_mn_instances(gemmPtrs);
for(auto& gemmPtr : gemmPtrs)
{
res &= ck::gemm_util::TestGemm<DeviceGemmNoOpPtr,
ADataType,
BDataType,
CDataType,
AccDataType,
ColumnMajor,
RowMajor,
RowMajor,
PassThrough,
PassThrough,
PassThrough>{}(gemmPtr);
}
gemmPtrs.clear();
ck::tensor_operation::device::device_gemm_instance::
add_device_gemm_xdl_f64_f64_f64_km_nk_mn_instances(gemmPtrs);
for(auto& gemmPtr : gemmPtrs)
{
res &= ck::gemm_util::TestGemm<DeviceGemmNoOpPtr,
ADataType,
BDataType,
CDataType,
AccDataType,
ColumnMajor,
ColumnMajor,
RowMajor,
PassThrough,
PassThrough,
PassThrough>{}(gemmPtr);
}
gemmPtrs.clear();
ck::tensor_operation::device::device_gemm_instance::
add_device_gemm_xdl_f64_f64_f64_mk_kn_mn_instances(gemmPtrs);
for(auto& gemmPtr : gemmPtrs)
{
res &= ck::gemm_util::TestGemm<DeviceGemmNoOpPtr,
ADataType,
BDataType,
CDataType,
AccDataType,
RowMajor,
RowMajor,
RowMajor,
PassThrough,
PassThrough,
PassThrough>{}(gemmPtr);
}
gemmPtrs.clear();
ck::tensor_operation::device::device_gemm_instance::
add_device_gemm_xdl_f64_f64_f64_mk_nk_mn_instances(gemmPtrs);
for(auto& gemmPtr : gemmPtrs)
{
res &= ck::gemm_util::TestGemm<DeviceGemmNoOpPtr,
ADataType,
BDataType,
CDataType,
AccDataType,
RowMajor,
ColumnMajor,
RowMajor,
PassThrough,
PassThrough,
PassThrough>{}(gemmPtr);
}
std::cout << "TestGemm ..... " << (res ? "SUCCESS" : "FAILURE") << std::endl;
return res ? 0 : 1;
}
test/gemm/gemm_xdl_int8.cpp 0 → 100644
View file @ 68886f7d
#include <algorithm>
#include <cstdlib>
#include <half.hpp>
#include <iostream>
#include <numeric>
#include <tuple>
#include <vector>
#include "gemm_util.hpp"
#include "config.hpp"
#include "print.hpp"
#include "device.hpp"
#include "host_tensor.hpp"
#include "host_tensor_generator.hpp"
#include "host_gemm.hpp"
#include "device_tensor.hpp"
#include "device_gemm_xdl.hpp"
#include "device_gemm_xdl_cshuffle.hpp"
#include "element_wise_operation.hpp"
#include "reference_gemm.hpp"
#include "gemm_specialization.hpp"
using PassThrough = ck::tensor_operation::element_wise::PassThrough;
using DeviceGemmNoOpPtr =
ck::tensor_operation::device::DeviceGemmPtr<ck::tensor_operation::element_wise::PassThrough,
ck::tensor_operation::element_wise::PassThrough,
ck::tensor_operation::element_wise::PassThrough>;
namespace ck {
namespace tensor_operation {
namespace device {
namespace device_gemm_instance {
void add_device_gemm_xdl_c_shuffle_i8_i8_i8_km_kn_mn_instances(std::vector<DeviceGemmNoOpPtr>&);
void add_device_gemm_xdl_c_shuffle_i8_i8_i8_km_nk_mn_instances(std::vector<DeviceGemmNoOpPtr>&);
void add_device_gemm_xdl_c_shuffle_i8_i8_i8_mk_nk_mn_instances(std::vector<DeviceGemmNoOpPtr>&);
void add_device_gemm_xdl_c_shuffle_i8_i8_i8_mk_kn_mn_instances(std::vector<DeviceGemmNoOpPtr>&);
} // namespace device_gemm_instance
} // namespace device
} // namespace tensor_operation
} // namespace ck
int main()
{
using ADataType = int8_t;
using BDataType = int8_t;
using CDataType = int8_t;
using AccDataType = int32_t;
using RowMajor = ck::tensor_layout::gemm::RowMajor;
using ColumnMajor = ck::tensor_layout::gemm::ColumnMajor;
std::vector<DeviceGemmNoOpPtr> gemmPtrs;
bool res = true;
ck::tensor_operation::device::device_gemm_instance::
add_device_gemm_xdl_c_shuffle_i8_i8_i8_km_kn_mn_instances(gemmPtrs);
for(auto& gemmPtr : gemmPtrs)
{
res &= ck::gemm_util::TestGemm<DeviceGemmNoOpPtr,
ADataType,
BDataType,
CDataType,
AccDataType,
ColumnMajor,
RowMajor,
RowMajor,
PassThrough,
PassThrough,
PassThrough>{}(gemmPtr);
}
gemmPtrs.clear();
ck::tensor_operation::device::device_gemm_instance::
add_device_gemm_xdl_c_shuffle_i8_i8_i8_km_nk_mn_instances(gemmPtrs);
for(auto& gemmPtr : gemmPtrs)
{
res &= ck::gemm_util::TestGemm<DeviceGemmNoOpPtr,
ADataType,
BDataType,
CDataType,
AccDataType,
ColumnMajor,
ColumnMajor,
RowMajor,
PassThrough,
PassThrough,
PassThrough>{}(gemmPtr);
}
gemmPtrs.clear();
ck::tensor_operation::device::device_gemm_instance::
add_device_gemm_xdl_c_shuffle_i8_i8_i8_mk_kn_mn_instances(gemmPtrs);
for(auto& gemmPtr : gemmPtrs)
{
res &= ck::gemm_util::TestGemm<DeviceGemmNoOpPtr,
ADataType,
BDataType,
CDataType,
AccDataType,
RowMajor,
RowMajor,
RowMajor,
PassThrough,
PassThrough,
PassThrough>{}(gemmPtr);
}
gemmPtrs.clear();
ck::tensor_operation::device::device_gemm_instance::
add_device_gemm_xdl_c_shuffle_i8_i8_i8_mk_nk_mn_instances(gemmPtrs);
for(auto& gemmPtr : gemmPtrs)
{
res &= ck::gemm_util::TestGemm<DeviceGemmNoOpPtr,
ADataType,
BDataType,
CDataType,
AccDataType,
RowMajor,
ColumnMajor,
RowMajor,
PassThrough,
PassThrough,
PassThrough>{}(gemmPtr);
}
std::cout << "TestGemm ..... " << (res ? "SUCCESS" : "FAILURE") << std::endl;
return res ? 0 : 1;
}
test/gemm_reduce/gemm_reduce_fp16.cpp
View file @ 68886f7d
......@@ -16,22 +16,22 @@ int main()
pass = pass &&
ck::profiler::
profile_gemm_reduce_impl<ck::half_t, ck::half_t, ck::half_t, float, Row, Row, Row>(
true, 1, false, 1, M, N, K, K, N, N);
true, 1, false, false, M, N, K, K, N, N);
pass = pass &&
ck::profiler::
profile_gemm_reduce_impl<ck::half_t, ck::half_t, ck::half_t, float, Row, Col, Row>(
true, 1, false, 1, M, N, K, K, K, N);
true, 1, false, false, M, N, K, K, K, N);
pass = pass &&
ck::profiler::
profile_gemm_reduce_impl<ck::half_t, ck::half_t, ck::half_t, float, Col, Row, Row>(
true, 1, false, 1, M, N, K, M, N, N);
true, 1, false, false, M, N, K, M, N, N);
pass = pass &&
ck::profiler::
profile_gemm_reduce_impl<ck::half_t, ck::half_t, ck::half_t, float, Col, Col, Row>(
true, 1, false, 1, M, N, K, M, K, N);
true, 1, false, false, M, N, K, M, K, N);
if(pass)
{
......
test/gemm_split_k/gemm_split_k.cpp
View file @ 68886f7d
......@@ -187,9 +187,10 @@ int test_gemm(const gemmArgs& args)
if(gemm_ptr->IsSupportedArgument(argument_ptr.get()))
{
invoker_ptr->Run(argument_ptr.get(), 0);
invoker_ptr->Run(argument_ptr.get());
c_device_buf.FromDevice(c_m_n_device_result.mData.data());
if(!check_out(c_m_n_host_result, c_m_n_device_result))
{
success = false;
......
test/grouped_gemm/grouped_gemm_fp16.cpp
View file @ 68886f7d
......@@ -141,18 +141,28 @@ bool TestGroupedGemm(DeviceGroupedGemmPtr_& groupedGemmPtr)
auto c_element_op = PassThrough{};
// do GEMM
auto invoker_ptr = groupedGemmPtr->MakeInvokerPointer();
auto invoker_ptr = groupedGemmPtr->MakeInvokerPointer();
auto argument_ptr = groupedGemmPtr->MakeArgumentPointer(
p_a, p_b, p_c, gemm_shapes, a_element_op, b_element_op, c_element_op);
DeviceMem gemm_desc_workspace(groupedGemmPtr->GetWorkSpaceSize(argument_ptr.get()));
groupedGemmPtr->SetWorkSpacePointer(argument_ptr.get(), gemm_desc_workspace.GetDeviceBuffer());
invoker_ptr->Run(argument_ptr.get());
for(std::size_t i = 0; i < gemm_shapes.size(); i++)
{
c_tensors_device[i]->FromDevice(c_device_tensors[i].mData.data());
using ReferenceGemmInstance = ck::tensor_operation::host::
ReferenceGemm<ADataType, BDataType, CDataType, PassThrough, PassThrough, PassThrough>;
using ReferenceGemmInstance = ck::tensor_operation::host::ReferenceGemm<ADataType,
BDataType,
CDataType,
AccDataType,
PassThrough,
PassThrough,
PassThrough>;
auto ref_gemm = ReferenceGemmInstance{};
auto ref_invoker = ref_gemm.MakeInvoker();
......
test/reduce/reduce_no_index.cpp
View file @ 68886f7d
#include "getopt.h"
#include "check_err.hpp"
#include "device_reduce_instance.hpp"
#include "reduction_enums.hpp"
#include "host_tensor.hpp"
#include "host_tensor_generator.hpp"
#include "host_reduction.hpp"
#include "reduce_util.hpp"
#include "host_common_util.hpp"
#include "profile_reduce_impl.hpp"
using namespace ck;
namespace {
template <index_t Rank, index_t NumReduceDim>
static inline std::vector<int> get_invariant_dims(const std::vector<int>& reduceDims)
{
assert(NumReduceDim == reduceDims.size());
int reduceFlag = 0;
// flag the bits for the reduceDims
for(int i = 0; i < NumReduceDim; i++)
{
reduceFlag |= 1 << reduceDims[i];
};
std::vector<int> invariantDims;
// collect invariant dimensions
for(int i = 0; i < Rank; i++)
if((reduceFlag & (1 << i)) == 0)
{
invariantDims.push_back(i);
};
return invariantDims;
};
constexpr int Rank = 4;
constexpr ReduceTensorOp ReduceOpId = ReduceTensorOp::AVG;
constexpr NanPropagation NanOpt = NanPropagation::PROPAGATE_NAN;
constexpr bool PropagateNan = false;
constexpr ReduceTensorIndices IndicesOpt = ReduceTensorIndices::NO_INDICES;
constexpr bool NeedIndices = false;
template <typename InDataType,
typename AccDataType,
typename OutDataType,
int Rank,
int NumReduceDim>
bool test_reduce_no_index_impl(int init_method,
const std::vector<size_t>& inLengths,
const std::vector<int>& reduceDims,
float alpha,
float beta)
{
using namespace ck::tensor_operation::device;
using namespace ck::tensor_operation::device::device_reduce_instance;
using namespace ck::host_reduce;
constexpr bool out_support_atomic_add = std::is_same<OutDataType, float>::value;
constexpr bool op_support_atomic_add = true;
constexpr bool use_atomic_add = (out_support_atomic_add && op_support_atomic_add);
Tensor<InDataType> in(inLengths);
std::vector<size_t> outLengths;
const auto invariantDims = get_invariant_dims<Rank, NumReduceDim>(reduceDims);
if(reduceDims.size() == Rank)
outLengths.push_back(1);
else
for(auto dim : invariantDims)
outLengths.push_back(inLengths[dim]);
Tensor<OutDataType> out_ref(outLengths);
Tensor<OutDataType> out(outLengths);
// only used when the OutDataType is bhalf_t
Tensor<float> out_ref_fp32(outLengths);
Tensor<float> out_fp32(outLengths);
auto inStrides = in.mDesc.GetStrides();
auto outStrides = out.mDesc.GetStrides();
size_t invariant_total_length = out.mDesc.GetElementSize();
size_t reduce_total_length = in.mDesc.GetElementSize() / invariant_total_length;
std::size_t num_thread = 1;
switch(init_method)
{
case 0: break;
case 1:
in.GenerateTensorValue(GeneratorTensor_1<InDataType>{1}, num_thread);
if(beta != 0.0f)
out_ref.GenerateTensorValue(GeneratorTensor_1<InDataType>{1}, num_thread);
break;
case 2:
in.GenerateTensorValue(GeneratorTensor_2<InDataType>{-5, 5}, num_thread);
if(beta != 0.0f)
out_ref.GenerateTensorValue(GeneratorTensor_2<InDataType>{-5, 5}, num_thread);
break;
default:
in.GenerateTensorValue(GeneratorTensor_3<InDataType>{-5.0, 5.0}, num_thread);
if(beta != 0.0f)
out_ref.GenerateTensorValue(GeneratorTensor_3<InDataType>{-5.0, 5.0}, num_thread);
}
if(beta != 0.0f)
for(size_t i = 0; i < out_ref.mDesc.GetElementSpace(); i++)
out.mData[i] = out_ref.mData[i];
// these buffers are usually provided by the user application
DeviceMem in_dev(sizeof(InDataType) * in.mDesc.GetElementSpace());
DeviceMem out_dev(sizeof(OutDataType) * out.mDesc.GetElementSpace());
in_dev.ToDevice(in.mData.data());
if(beta != 0.0f)
out_dev.ToDevice(out.mData.data());
using InElementwiseOperation_0 =
typename reduce_unary_operator<AccDataType, ReduceOpId, true, true>::InElementwiseOperation;
using AccElementwiseOperation_0 =
typename reduce_unary_operator<AccDataType, ReduceOpId, true, true>::
AccElementwiseOperation;
using InElementwiseOperation_1 =
typename reduce_unary_operator<AccDataType, ReduceOpId, true, false>::
InElementwiseOperation;
using AccElementwiseOperation_1 =
typename reduce_unary_operator<AccDataType, ReduceOpId, true, false>::
AccElementwiseOperation;
using InElementwiseOperation_2 =
typename reduce_unary_operator<AccDataType, ReduceOpId, false, true>::
InElementwiseOperation;
using AccElementwiseOperation_2 =
typename reduce_unary_operator<AccDataType, ReduceOpId, false, true>::
AccElementwiseOperation;
using DeviceReduceInstPtr0 =
DeviceReducePtr<InElementwiseOperation_0, AccElementwiseOperation_0>;
using DeviceReduceInstPtr1 =
DeviceReducePtr<InElementwiseOperation_1, AccElementwiseOperation_1>;
using DeviceReduceInstPtr2 =
DeviceReducePtr<InElementwiseOperation_2, AccElementwiseOperation_2>;
std::vector<DeviceReduceInstPtr0> reduce0_ptrs;
std::vector<DeviceReduceInstPtr1> reduce1_ptrs;
std::vector<DeviceReduceInstPtr2> reduce2_ptrs;
add_device_reduce_instance_threadwise<InDataType,
AccDataType,
OutDataType,
Rank,
NumReduceDim,
ReduceOpId,
NanOpt,
IndicesOpt>(reduce0_ptrs);
add_device_reduce_instance_blockwise<InDataType,
AccDataType,
OutDataType,
Rank,
NumReduceDim,
ReduceOpId,
NanOpt,
IndicesOpt>(reduce0_ptrs);
if constexpr(use_atomic_add)
{
add_device_reduce_instance_multiblock_atomic_add<InDataType,
AccDataType,
OutDataType,
Rank,
NumReduceDim,
ReduceOpId,
NanOpt,
IndicesOpt>(reduce0_ptrs);
}
else
{
add_device_reduce_instance_multiblock_partial_reduce<InDataType,
AccDataType,
OutDataType,
Rank,
NumReduceDim,
ReduceOpId,
NanOpt,
IndicesOpt>(reduce1_ptrs);
};
// used for secondary reduction
if constexpr(!use_atomic_add)
{
add_device_reduce_instance_blockwise_second_call<AccDataType,
AccDataType,
OutDataType,
Rank,
NumReduceDim,
ReduceOpId,
NanOpt,
IndicesOpt>(reduce2_ptrs);
};
if(reduce0_ptrs.empty() && reduce1_ptrs.empty())
{
throw std::runtime_error("Wrong! No device REDUCE instance found");
};
bool result = true;
ReductionHost<InDataType,
AccDataType,
OutDataType,
ReduceOpId,
Rank,
NumReduceDim,
PropagateNan,
NeedIndices>
hostReduce(in.mDesc, out_ref.mDesc, invariantDims, reduceDims);
hostReduce.Run(alpha, in.mData.data(), beta, out_ref.mData.data(), nullptr);
const auto i_inLengths = to_int_vector(inLengths);
const auto i_inStrides = to_int_vector(inStrides);
const auto i_outLengths = to_int_vector(outLengths);
const auto i_outStrides = to_int_vector(outStrides);
for(auto& reduce_ptr : reduce0_ptrs)
{
auto wsSizeInBytes = reduce_ptr->GetWorkspaceSizeInBytes(i_inLengths, reduceDims);
DeviceMem ws_dev(wsSizeInBytes);
InElementwiseOperation_0 in_elementwise_op_0(static_cast<int32_t>(reduce_total_length));
AccElementwiseOperation_0 acc_elementwise_op_0(static_cast<int32_t>(reduce_total_length));
auto argument_ptr = reduce_ptr->MakeArgumentPointer(i_inLengths,
i_inStrides,
i_outLengths,
i_outStrides,
reduceDims,
alpha,
beta,
in_dev.GetDeviceBuffer(),
out_dev.GetDeviceBuffer(),
nullptr,
ws_dev.GetDeviceBuffer(),
in_elementwise_op_0,
acc_elementwise_op_0);
if(!reduce_ptr->IsSupportedArgument(argument_ptr.get()))
continue;
auto invoker_ptr = reduce_ptr->MakeInvokerPointer();
(void)invoker_ptr->Run(argument_ptr.get());
out_dev.FromDevice(out.mData.data());
bool single_result = true;
if constexpr(std::is_same<OutDataType, ck::half_t>::value ||
std::is_same<OutDataType, ck::bhalf_t>::value)
{
reduce_util::to_f32_vector(out, out_fp32);
reduce_util::to_f32_vector(out_ref, out_ref_fp32);
single_result = ck::utils::check_err(
out_fp32.mData, out_ref_fp32.mData, "Error: incorrect data result!");
}
else
{
single_result =
ck::utils::check_err(out.mData, out_ref.mData, "Error: incorrect data result!");
};
if(!single_result)
{
std::cout << "Fail Info: " << reduce_ptr->GetTypeString() << std::endl;
result = false;
}
};
for(auto& reduce_ptr : reduce1_ptrs)
{
auto wsSizeInBytes = reduce_ptr->GetWorkspaceSizeInBytes(i_inLengths, reduceDims);
DeviceMem ws_dev(wsSizeInBytes);
InElementwiseOperation_1 in_elementwise_op_1(static_cast<int32_t>(reduce_total_length));
AccElementwiseOperation_1 acc_elementwise_op_1(static_cast<int32_t>(reduce_total_length));
auto argument_ptr = reduce_ptr->MakeArgumentPointer(i_inLengths,
i_inStrides,
i_outLengths,
i_outStrides,
reduceDims,
alpha,
beta,
in_dev.GetDeviceBuffer(),
out_dev.GetDeviceBuffer(),
nullptr,
ws_dev.GetDeviceBuffer(),
in_elementwise_op_1,
acc_elementwise_op_1);
if(!reduce_ptr->IsSupportedArgument(argument_ptr.get()))
continue;
auto invoker_ptr = reduce_ptr->MakeInvokerPointer();
(void)invoker_ptr->Run(argument_ptr.get());
std::vector<int> inLengths2 = reduce_ptr->GetWorkspace2dLengths(argument_ptr.get());
std::vector<int> inStrides2{inLengths2[1], 1};
for(auto& reduce2_ptr : reduce2_ptrs)
{
InElementwiseOperation_2 in_elementwise_op_2(static_cast<int32_t>(reduce_total_length));
AccElementwiseOperation_2 acc_elementwise_op_2(
static_cast<int32_t>(reduce_total_length));
auto argument2_ptr = reduce2_ptr->MakeArgumentPointer(inLengths2,
inStrides2,
i_outLengths,
i_outStrides,
reduceDims,
alpha,
beta,
ws_dev.GetDeviceBuffer(),
out_dev.GetDeviceBuffer(),
nullptr,
ws_dev.GetDeviceBuffer(),
in_elementwise_op_2,
acc_elementwise_op_2);
if(!reduce2_ptr->IsSupportedArgument(argument2_ptr.get()))
continue;
std::string reduce2_name = reduce2_ptr->GetTypeString();
auto invoker2_ptr = reduce2_ptr->MakeInvokerPointer();
(void)invoker2_ptr->Run(argument2_ptr.get());
out_dev.FromDevice(out.mData.data());
bool single_result = true;
if constexpr(std::is_same<OutDataType, ck::half_t>::value ||
std::is_same<OutDataType, ck::bhalf_t>::value)
{
reduce_util::to_f32_vector(out, out_fp32);
reduce_util::to_f32_vector(out_ref, out_ref_fp32);
single_result = ck::utils::check_err(
out_fp32.mData, out_ref_fp32.mData, "Error: incorrect data result!");
}
else
{
single_result =
ck::utils::check_err(out.mData, out_ref.mData, "Error: incorrect data result!");
};
if(!single_result)
{
std::cout << "Fail Info: " << reduce_ptr->GetTypeString() << " => "
<< reduce2_ptr->GetTypeString() << std::endl;
result = false;
}
};
};
return (result);
};
} // anonymous namespace
static struct option long_options[] = {{"inLengths", required_argument, nullptr, 'D'},
{"reduceDimensions", required_argument, nullptr, 'R'},
{"scales", required_argument, nullptr, 'S'},
......@@ -387,48 +13,6 @@ static struct option long_options[] = {{"inLengths", required_argument, nullptr,
class SimpleAppArgs
{
template <typename T>
static T getSingleValueFromString(const std::string& valueStr)
{
std::istringstream iss(valueStr);
T ret;
iss >> ret;
return (ret);
};
template <typename T>
static std::vector<T> getTypeValuesFromString(const char* cstr_values)
{
std::string valuesStr(cstr_values);
std::vector<T> values;
std::size_t pos = 0;
std::size_t new_pos;
new_pos = valuesStr.find(',', pos);
while(new_pos != std::string::npos)
{
const std::string sliceStr = valuesStr.substr(pos, new_pos - pos);
T val = getSingleValueFromString<T>(sliceStr);
values.push_back(val);
pos = new_pos + 1;
new_pos = valuesStr.find(',', pos);
};
std::string sliceStr = valuesStr.substr(pos);
T val = getSingleValueFromString<T>(sliceStr);
values.push_back(val);
return (values);
};
private:
int option_index = 0;
......@@ -460,6 +44,8 @@ class SimpleAppArgs
int processArgs(int argc, char* argv[])
{
using ck::host_common::getTypeValuesFromString;
int ch;
while(1)
......@@ -514,7 +100,7 @@ class SimpleAppArgs
(reduceDims.size() != 1 && reduceDims.size() != 3 && reduceDims.size() != 4))
return (-1);
if(data_type != 0 && data_type != 1 && data_type != 3 && data_type != 5)
if(data_type != 0 && data_type != 1 && data_type != 3 && data_type != 5 && data_type != 6)
return (-1);
return (0);
......@@ -525,87 +111,92 @@ bool test_reduce_no_index(int data_type,
int init_method,
std::vector<int> reduceDims,
std::vector<size_t> inLengths,
ReduceTensorOp reduceOpId,
bool propagateNan,
float alpha,
float beta)
{
using ck::profiler::profile_reduce_impl;
bool result = true;
if(data_type == 0)
{
switch(reduceDims.size())
{
case 1:
result = test_reduce_no_index_impl<float, float, float, Rank, 1>(
init_method, inLengths, reduceDims, alpha, beta);
break;
case 3:
result = test_reduce_no_index_impl<float, float, float, Rank, 3>(
init_method, inLengths, reduceDims, alpha, beta);
break;
case 4:
result = test_reduce_no_index_impl<float, float, float, Rank, 4>(
init_method, inLengths, reduceDims, alpha, beta);
break;
};
result = profile_reduce_impl<float, float, float>(true,
init_method,
false,
false,
inLengths,
reduceDims,
reduceOpId,
propagateNan,
false,
alpha,
beta);
}
else if(data_type == 1)
{
switch(reduceDims.size())
{
case 1:
result = test_reduce_no_index_impl<ck::half_t, float, ck::half_t, Rank, 1>(
init_method, inLengths, reduceDims, alpha, beta);
break;
case 3:
result = test_reduce_no_index_impl<ck::half_t, float, ck::half_t, Rank, 3>(
init_method, inLengths, reduceDims, alpha, beta);
break;
case 4:
result = test_reduce_no_index_impl<ck::half_t, float, ck::half_t, Rank, 4>(
init_method, inLengths, reduceDims, alpha, beta);
break;
};
result = profile_reduce_impl<ck::half_t, float, ck::half_t>(true,
init_method,
false,
false,
inLengths,
reduceDims,
reduceOpId,
propagateNan,
false,
alpha,
beta);
}
else if(data_type == 3)
{
switch(reduceDims.size())
{
case 1:
result = test_reduce_no_index_impl<int8_t, int32_t, int8_t, Rank, 1>(
init_method, inLengths, reduceDims, alpha, beta);
break;
case 3:
result = test_reduce_no_index_impl<int8_t, int32_t, int8_t, Rank, 3>(
init_method, inLengths, reduceDims, alpha, beta);
break;
case 4:
result = test_reduce_no_index_impl<int8_t, int32_t, int8_t, Rank, 4>(
init_method, inLengths, reduceDims, alpha, beta);
break;
};
result = profile_reduce_impl<int8_t, int32_t, int8_t>(true,
init_method,
false,
false,
inLengths,
reduceDims,
reduceOpId,
propagateNan,
false,
alpha,
beta);
}
else if(data_type == 5)
{
switch(reduceDims.size())
{
case 1:
result = test_reduce_no_index_impl<ck::bhalf_t, float, ck::bhalf_t, Rank, 1>(
init_method, inLengths, reduceDims, alpha, beta);
break;
case 3:
result = test_reduce_no_index_impl<ck::bhalf_t, float, ck::bhalf_t, Rank, 3>(
init_method, inLengths, reduceDims, alpha, beta);
break;
case 4:
result = test_reduce_no_index_impl<ck::bhalf_t, float, ck::bhalf_t, Rank, 4>(
init_method, inLengths, reduceDims, alpha, beta);
break;
};
result = profile_reduce_impl<ck::bhalf_t, float, ck::bhalf_t>(true,
init_method,
false,
false,
inLengths,
reduceDims,
reduceOpId,
propagateNan,
false,
alpha,
beta);
}
else if(data_type == 6)
{
result = profile_reduce_impl<double, double, double>(true,
init_method,
false,
false,
inLengths,
reduceDims,
reduceOpId,
propagateNan,
false,
alpha,
beta);
}
return (result);
};
constexpr ReduceTensorOp reduceOpId = ReduceTensorOp::AVG;
constexpr bool propagateNan = false;
int main(int argc, char* argv[])
{
SimpleAppArgs args;
......@@ -621,8 +212,14 @@ int main(int argc, char* argv[])
{0, 1, 2, 3}, {0, 1, 2}, {1, 2, 3}, {0, 1, 3}, {0, 2, 3}, {0}, {1}, {2}, {3}};
for(auto& reduceDims : v_reduceDims)
result = result && test_reduce_no_index(
data_type, init_method, reduceDims, inLengths, 1.0f, 0.0f);
result = result && test_reduce_no_index(data_type,
init_method,
reduceDims,
inLengths,
reduceOpId,
propagateNan,
1.0f,
0.0f);
}
else
{
......@@ -636,6 +233,8 @@ int main(int argc, char* argv[])
args.init_method,
args.reduceDims,
args.inLengths,
reduceOpId,
propagateNan,
args.scales[0],
args.scales[1]);
}
......
test/reduce/reduce_util.hpp deleted 100644 → 0
View file @ a9ee2960
#ifndef REDUCE_UTILS_HPP
#define REDUCE_UTILS_HPP
#include "data_type.hpp"
namespace ck {
namespace reduce_util {
template <typename T>
void to_f32_vector(const Tensor<T>& src, Tensor<float>& dst)
{
for(std::size_t i = 0; i < src.mData.size(); ++i)
dst.mData[i] = type_convert<float>(src.mData[i]);
}
} // namespace reduce_util
} // namespace ck
#endif
test/reduce/reduce_with_index.cpp
View file @ 68886f7d
#include "getopt.h"
#include "device_reduce_instance.hpp"
#include "reduction_enums.hpp"
#include "host_tensor.hpp"
#include "host_tensor_generator.hpp"
#include "host_reduction.hpp"
#include "check_err.hpp"
#include "reduce_util.hpp"
using namespace ck;
namespace {
template <index_t Rank, index_t NumReduceDim>
static inline std::vector<int> get_invariant_dims(const std::vector<int>& reduceDims)
{
assert(NumReduceDim == reduceDims.size());
int reduceFlag = 0;
// flag the bits for the reduceDims
for(int i = 0; i < NumReduceDim; i++)
{
reduceFlag |= 1 << reduceDims[i];
};
std::vector<int> invariantDims;
// collect invariant dimensions
for(int i = 0; i < Rank; i++)
if((reduceFlag & (1 << i)) == 0)
{
invariantDims.push_back(i);
};
return invariantDims;
};
constexpr int Rank = 4;
constexpr ReduceTensorOp ReduceOpId = ReduceTensorOp::AMAX;
constexpr NanPropagation NanOpt = NanPropagation::PROPAGATE_NAN;
constexpr bool PropagateNan = false;
constexpr ReduceTensorIndices IndicesOpt = ReduceTensorIndices::FLATTENED_INDICES;
constexpr bool NeedIndices = true;
template <typename InDataType,
typename AccDataType,
typename OutDataType,
int Rank,
int NumReduceDim>
bool test_reduce_with_index_impl(int init_method,
const std::vector<size_t>& inLengths,
const std::vector<int>& reduceDims,
float alpha,
float beta)
{
using namespace ck::tensor_operation::device;
using namespace ck::tensor_operation::device::device_reduce_instance;
using namespace ck::host_reduce;
Tensor<InDataType> in(inLengths);
std::vector<size_t> outLengths;
const auto invariantDims = get_invariant_dims<Rank, NumReduceDim>(reduceDims);
if(reduceDims.size() == Rank)
outLengths.push_back(1);
else
for(auto dim : invariantDims)
outLengths.push_back(inLengths[dim]);
Tensor<OutDataType> out_ref(outLengths);
Tensor<OutDataType> out(outLengths);
Tensor<int32_t> out_indices_ref(outLengths);
Tensor<int32_t> out_indices(outLengths);
// only used when the OutDataType is bhalf_t
Tensor<float> out_ref_fp32(outLengths);
Tensor<float> out_fp32(outLengths);
auto inStrides = in.mDesc.GetStrides();
auto outStrides = out.mDesc.GetStrides();
size_t invariant_total_length = out.mDesc.GetElementSize();
size_t reduce_total_length = in.mDesc.GetElementSize() / invariant_total_length;
std::size_t num_thread = 1;
switch(init_method)
{
case 0: break;
case 1:
in.GenerateTensorValue(GeneratorTensor_1<InDataType>{1}, num_thread);
if(beta != 0.0f)
out_ref.GenerateTensorValue(GeneratorTensor_1<InDataType>{1}, num_thread);
break;
case 2:
in.GenerateTensorValue(GeneratorTensor_2<InDataType>{-5, 5}, num_thread);
if(beta != 0.0f)
out_ref.GenerateTensorValue(GeneratorTensor_2<InDataType>{-5, 5}, num_thread);
break;
default:
in.GenerateTensorValue(GeneratorTensor_3<InDataType>{-5.0, 5.0}, num_thread);
if(beta != 0.0f)
out_ref.GenerateTensorValue(GeneratorTensor_3<InDataType>{-5.0, 5.0}, num_thread);
}
if(beta != 0.0f)
for(size_t i = 0; i < out_ref.mDesc.GetElementSpace(); i++)
out.mData[i] = out_ref.mData[i];
// these buffers are usually provided by the user application
DeviceMem in_dev(sizeof(InDataType) * in.mDesc.GetElementSpace());
DeviceMem out_dev(sizeof(OutDataType) * out.mDesc.GetElementSpace());
in_dev.ToDevice(in.mData.data());
if(beta != 0.0f)
out_dev.ToDevice(out.mData.data());
size_t indicesSizeInBytes = NeedIndices ? out.mDesc.GetElementSize() * sizeof(int) : 0;
DeviceMem out_indices_dev(indicesSizeInBytes);
using InElementwiseOperation_0 =
typename reduce_unary_operator<AccDataType, ReduceOpId, true, true>::InElementwiseOperation;
using AccElementwiseOperation_0 =
typename reduce_unary_operator<AccDataType, ReduceOpId, true, true>::
AccElementwiseOperation;
using InElementwiseOperation_1 =
typename reduce_unary_operator<AccDataType, ReduceOpId, true, false>::
InElementwiseOperation;
using AccElementwiseOperation_1 =
typename reduce_unary_operator<AccDataType, ReduceOpId, true, false>::
AccElementwiseOperation;
using InElementwiseOperation_2 =
typename reduce_unary_operator<AccDataType, ReduceOpId, false, true>::
InElementwiseOperation;
using AccElementwiseOperation_2 =
typename reduce_unary_operator<AccDataType, ReduceOpId, false, true>::
AccElementwiseOperation;
using DeviceReduceInstPtr0 =
DeviceReducePtr<InElementwiseOperation_0, AccElementwiseOperation_0>;
using DeviceReduceInstPtr1 =
DeviceReducePtr<InElementwiseOperation_1, AccElementwiseOperation_1>;
using DeviceReduceInstPtr2 =
DeviceReducePtr<InElementwiseOperation_2, AccElementwiseOperation_2>;
std::vector<DeviceReduceInstPtr0> reduce0_ptrs;
std::vector<DeviceReduceInstPtr1> reduce1_ptrs;
std::vector<DeviceReduceInstPtr2> reduce2_ptrs;
add_device_reduce_instance_threadwise<InDataType,
AccDataType,
OutDataType,
Rank,
NumReduceDim,
ReduceOpId,
NanOpt,
IndicesOpt>(reduce0_ptrs);
add_device_reduce_instance_blockwise<InDataType,
AccDataType,
OutDataType,
Rank,
NumReduceDim,
ReduceOpId,
NanOpt,
IndicesOpt>(reduce0_ptrs);
add_device_reduce_instance_multiblock_partial_reduce<InDataType,
AccDataType,
OutDataType,
Rank,
NumReduceDim,
ReduceOpId,
NanOpt,
IndicesOpt>(reduce1_ptrs);
add_device_reduce_instance_blockwise_second_call<AccDataType,
AccDataType,
OutDataType,
Rank,
NumReduceDim,
ReduceOpId,
NanOpt,
IndicesOpt>(reduce2_ptrs);
if(reduce0_ptrs.empty() && reduce1_ptrs.empty())
{
throw std::runtime_error("Wrong! No device REDUCE instance found");
};
bool result = true;
ReductionHost<InDataType,
AccDataType,
OutDataType,
ReduceOpId,
Rank,
NumReduceDim,
PropagateNan,
NeedIndices>
hostReduce(in.mDesc, out_ref.mDesc, invariantDims, reduceDims);
hostReduce.Run(
alpha, in.mData.data(), beta, out_ref.mData.data(), out_indices_ref.mData.data());
const auto i_inLengths = to_int_vector(inLengths);
const auto i_inStrides = to_int_vector(inStrides);
const auto i_outLengths = to_int_vector(outLengths);
const auto i_outStrides = to_int_vector(outStrides);
for(auto& reduce_ptr : reduce0_ptrs)
{
auto wsSizeInBytes = reduce_ptr->GetWorkspaceSizeInBytes(i_inLengths, reduceDims);
DeviceMem ws_dev(wsSizeInBytes);
InElementwiseOperation_0 in_elementwise_op_0(static_cast<int32_t>(reduce_total_length));
AccElementwiseOperation_0 acc_elementwise_op_0(static_cast<int32_t>(reduce_total_length));
auto argument_ptr = reduce_ptr->MakeArgumentPointer(i_inLengths,
i_inStrides,
i_outLengths,
i_outStrides,
reduceDims,
alpha,
beta,
in_dev.GetDeviceBuffer(),
out_dev.GetDeviceBuffer(),
out_indices_dev.GetDeviceBuffer(),
ws_dev.GetDeviceBuffer(),
in_elementwise_op_0,
acc_elementwise_op_0);
if(!reduce_ptr->IsSupportedArgument(argument_ptr.get()))
continue;
auto invoker_ptr = reduce_ptr->MakeInvokerPointer();
(void)invoker_ptr->Run(argument_ptr.get());
out_dev.FromDevice(out.mData.data());
bool single_result = true;
if constexpr(std::is_same<OutDataType, ck::half_t>::value ||
std::is_same<OutDataType, ck::bhalf_t>::value)
{
reduce_util::to_f32_vector(out, out_fp32);
reduce_util::to_f32_vector(out_ref, out_ref_fp32);
single_result = ck::utils::check_err(
out_fp32.mData, out_ref_fp32.mData, "Error: incorrect data result!");
}
else
{
single_result =
ck::utils::check_err(out.mData, out_ref.mData, "Error: incorrect data result!");
};
if(NeedIndices)
{
out_indices_dev.FromDevice(out_indices.mData.data());
single_result = single_result && ck::utils::check_err(out_indices_ref.mData,
out_indices.mData,
"Error: incorrect index result!");
};
#include "host_common_util.hpp"
#include "profile_reduce_impl.hpp"
if(!single_result)
{
std::cout << "Fail Info: " << reduce_ptr->GetTypeString() << std::endl;
result = false;
}
};
for(auto& reduce_ptr : reduce1_ptrs)
{
auto wsSizeInBytes = reduce_ptr->GetWorkspaceSizeInBytes(i_inLengths, reduceDims);
DeviceMem ws_dev(wsSizeInBytes);
InElementwiseOperation_1 in_elementwise_op_1(static_cast<int32_t>(reduce_total_length));
AccElementwiseOperation_1 acc_elementwise_op_1(static_cast<int32_t>(reduce_total_length));
auto argument_ptr = reduce_ptr->MakeArgumentPointer(i_inLengths,
i_inStrides,
i_outLengths,
i_outStrides,
reduceDims,
alpha,
beta,
in_dev.GetDeviceBuffer(),
out_dev.GetDeviceBuffer(),
out_indices_dev.GetDeviceBuffer(),
ws_dev.GetDeviceBuffer(),
in_elementwise_op_1,
acc_elementwise_op_1);
if(!reduce_ptr->IsSupportedArgument(argument_ptr.get()))
continue;
std::string reduce_name = reduce_ptr->GetTypeString();
auto invoker_ptr = reduce_ptr->MakeInvokerPointer();
(void)invoker_ptr->Run(argument_ptr.get());
std::vector<int> inLengths2 = reduce_ptr->GetWorkspace2dLengths(argument_ptr.get());
std::vector<int> inStrides2{inLengths2[1], 1};
for(auto& reduce2_ptr : reduce2_ptrs)
{
InElementwiseOperation_2 in_elementwise_op_2(static_cast<int32_t>(reduce_total_length));
AccElementwiseOperation_2 acc_elementwise_op_2(
static_cast<int32_t>(reduce_total_length));
auto argument2_ptr = reduce2_ptr->MakeArgumentPointer(inLengths2,
inStrides2,
i_outLengths,
i_outStrides,
reduceDims,
alpha,
beta,
ws_dev.GetDeviceBuffer(),
out_dev.GetDeviceBuffer(),
out_indices_dev.GetDeviceBuffer(),
ws_dev.GetDeviceBuffer(),
in_elementwise_op_2,
acc_elementwise_op_2);
if(!reduce2_ptr->IsSupportedArgument(argument2_ptr.get()))
continue;
std::string reduce2_name = reduce2_ptr->GetTypeString();
auto invoker2_ptr = reduce2_ptr->MakeInvokerPointer();
(void)invoker2_ptr->Run(argument2_ptr.get());
out_dev.FromDevice(out.mData.data());
bool single_result = true;
if constexpr(std::is_same<OutDataType, ck::half_t>::value ||
std::is_same<OutDataType, ck::bhalf_t>::value)
{
reduce_util::to_f32_vector(out, out_fp32);
reduce_util::to_f32_vector(out_ref, out_ref_fp32);
single_result = ck::utils::check_err(
out_fp32.mData, out_ref_fp32.mData, "Error: incorrect data result!");
}
else
{
single_result =
ck::utils::check_err(out.mData, out_ref.mData, "Error: incorrect data result!");
};
if(NeedIndices)
{
out_indices_dev.FromDevice(out_indices.mData.data());
single_result =
single_result && ck::utils::check_err(out_indices_ref.mData,
out_indices.mData,
"Error: incorrect index result!");
};
if(!single_result)
{
std::cout << "Fail Info: " << reduce_ptr->GetTypeString() << " => "
<< reduce2_ptr->GetTypeString() << std::endl;
result = false;
}
};
};
return (result);
};
} // anonymous namespace
using namespace ck;
static struct option long_options[] = {{"inLengths", required_argument, nullptr, 'D'},
{"reduceDimensions", required_argument, nullptr, 'R'},
......@@ -390,48 +13,6 @@ static struct option long_options[] = {{"inLengths", required_argument, nullptr,
class SimpleAppArgs
{
template <typename T>
static T getSingleValueFromString(const std::string& valueStr)
{
std::istringstream iss(valueStr);
T ret;
iss >> ret;
return (ret);
};
template <typename T>
static std::vector<T> getTypeValuesFromString(const char* cstr_values)
{
std::string valuesStr(cstr_values);
std::vector<T> values;
std::size_t pos = 0;
std::size_t new_pos;
new_pos = valuesStr.find(',', pos);
while(new_pos != std::string::npos)
{
const std::string sliceStr = valuesStr.substr(pos, new_pos - pos);
T val = getSingleValueFromString<T>(sliceStr);
values.push_back(val);
pos = new_pos + 1;
new_pos = valuesStr.find(',', pos);
};
std::string sliceStr = valuesStr.substr(pos);
T val = getSingleValueFromString<T>(sliceStr);
values.push_back(val);
return (values);
};
private:
int option_index = 0;
......@@ -463,6 +44,8 @@ class SimpleAppArgs
int processArgs(int argc, char* argv[])
{
using ck::host_common::getTypeValuesFromString;
int ch;
while(1)
......@@ -517,7 +100,7 @@ class SimpleAppArgs
(reduceDims.size() != 1 && reduceDims.size() != 3 && reduceDims.size() != 4))
return (-1);
if(data_type != 0 && data_type != 1 && data_type != 3 && data_type != 5)
if(data_type != 0 && data_type != 1 && data_type != 3 && data_type != 5 && data_type != 6)
return (-1);
return (0);
......@@ -528,87 +111,92 @@ bool test_reduce_with_index(int data_type,
int init_method,
std::vector<int> reduceDims,
std::vector<size_t> inLengths,
ReduceTensorOp reduceOpId,
bool propagateNan,
float alpha,
float beta)
{
using ck::profiler::profile_reduce_impl;
bool result = true;
if(data_type == 0)
{
switch(reduceDims.size())
{
case 1:
result = test_reduce_with_index_impl<float, float, float, Rank, 1>(
init_method, inLengths, reduceDims, alpha, beta);
break;
case 3:
result = test_reduce_with_index_impl<float, float, float, Rank, 3>(
init_method, inLengths, reduceDims, alpha, beta);
break;
case 4:
result = test_reduce_with_index_impl<float, float, float, Rank, 4>(
init_method, inLengths, reduceDims, alpha, beta);
break;
};
result = profile_reduce_impl<float, float, float>(true,
init_method,
false,
false,
inLengths,
reduceDims,
reduceOpId,
propagateNan,
true,
alpha,
beta);
}
else if(data_type == 1)
{
switch(reduceDims.size())
{
case 1:
result = test_reduce_with_index_impl<ck::half_t, ck::half_t, ck::half_t, Rank, 1>(
init_method, inLengths, reduceDims, alpha, beta);
break;
case 3:
result = test_reduce_with_index_impl<ck::half_t, ck::half_t, ck::half_t, Rank, 3>(
init_method, inLengths, reduceDims, alpha, beta);
break;
case 4:
result = test_reduce_with_index_impl<ck::half_t, ck::half_t, ck::half_t, Rank, 4>(
init_method, inLengths, reduceDims, alpha, beta);
break;
};
result = profile_reduce_impl<ck::half_t, ck::half_t, ck::half_t>(true,
init_method,
false,
false,
inLengths,
reduceDims,
reduceOpId,
propagateNan,
true,
alpha,
beta);
}
else if(data_type == 3)
{
switch(reduceDims.size())
{
case 1:
result = test_reduce_with_index_impl<int8_t, int8_t, int8_t, Rank, 1>(
init_method, inLengths, reduceDims, alpha, beta);
break;
case 3:
result = test_reduce_with_index_impl<int8_t, int8_t, int8_t, Rank, 3>(
init_method, inLengths, reduceDims, alpha, beta);
break;
case 4:
result = test_reduce_with_index_impl<int8_t, int8_t, int8_t, Rank, 4>(
init_method, inLengths, reduceDims, alpha, beta);
break;
};
result = profile_reduce_impl<int8_t, int8_t, int8_t>(true,
init_method,
false,
false,
inLengths,
reduceDims,
reduceOpId,
propagateNan,
true,
alpha,
beta);
}
else if(data_type == 5)
{
switch(reduceDims.size())
{
case 1:
result = test_reduce_with_index_impl<ck::bhalf_t, float, ck::bhalf_t, Rank, 1>(
init_method, inLengths, reduceDims, alpha, beta);
break;
case 3:
result = test_reduce_with_index_impl<ck::bhalf_t, float, ck::bhalf_t, Rank, 3>(
init_method, inLengths, reduceDims, alpha, beta);
break;
case 4:
result = test_reduce_with_index_impl<ck::bhalf_t, float, ck::bhalf_t, Rank, 4>(
init_method, inLengths, reduceDims, alpha, beta);
break;
};
result = profile_reduce_impl<ck::bhalf_t, float, ck::bhalf_t>(true,
init_method,
false,
false,
inLengths,
reduceDims,
reduceOpId,
propagateNan,
true,
alpha,
beta);
}
else if(data_type == 6)
{
result = profile_reduce_impl<double, double, double>(true,
init_method,
false,
false,
inLengths,
reduceDims,
reduceOpId,
propagateNan,
true,
alpha,
beta);
}
return (result);
};
constexpr ReduceTensorOp reduceOpId = ReduceTensorOp::AMAX;
constexpr bool propagateNan = false;
int main(int argc, char* argv[])
{
SimpleAppArgs args;
......@@ -624,8 +212,14 @@ int main(int argc, char* argv[])
{0, 1, 2, 3}, {0, 1, 2}, {1, 2, 3}, {0, 1, 3}, {0, 2, 3}, {0}, {1}, {2}, {3}};
for(auto& reduceDims : v_reduceDims)
result = result && test_reduce_with_index(
data_type, init_method, reduceDims, inLengths, 1.0f, 0.0f);
result = result && test_reduce_with_index(data_type,
init_method,
reduceDims,
inLengths,
reduceOpId,
propagateNan,
1.0f,
0.0f);
}
else
{
......@@ -639,6 +233,8 @@ int main(int argc, char* argv[])
args.init_method,
args.reduceDims,
args.inLengths,
reduceOpId,
propagateNan,
args.scales[0],
args.scales[1]);
}
......
Markdown is supported
0% or .
You are about to add 0 people to the discussion. Proceed with caution.
Finish editing this message first!
Please register or to comment