Skip to content

GitLab

Commit a3b4c5cb authored by wangshaojie6's avatar wangshaojie6
Browse files

merge develop branch and add gridwise pipeline v3

parents 48918ab9 1677cf70
Hide whitespace changes
Inline Side-by-side
Showing with 491 additions and 904 deletions
profiler/include/profile_gemm_reduce_impl.hpp
View file @ a3b4c5cb
......@@ -7,7 +7,7 @@
#include "tensor_layout.hpp"
#include "device_tensor.hpp"
#include "element_wise_operation.hpp"
#include "element_wise_reduce_operation.hpp"
#include "reduction_operator.hpp"
#include "device_gemm_reduce.hpp"
#include "reference_gemm.hpp"
......@@ -16,12 +16,22 @@ namespace tensor_operation {
namespace device {
namespace device_gemm_instance {
using F32 = float;
using F16 = ck::half_t;
using DPtrsGlobal = ck::Tuple<F32*, F32*>;
using Div = ck::tensor_operation::element_wise::UnaryIdentic<F32, F32, true>;
using Identity = ck::tensor_operation::element_wise::UnaryIdentic<F32, F32, false>;
using Square = ck::tensor_operation::element_wise::UnarySquare<F32, F32, false>;
using DInElementOps = ck::Tuple<Identity, Square>;
using DOutElementOps = ck::Tuple<Div, Div>;
using DeviceGemmReduceNoOpPtr = ck::tensor_operation::device::DeviceGemmReducePtr<
DPtrsGlobal,
ck::tensor_operation::element_wise::PassThrough,
ck::tensor_operation::element_wise::PassThrough,
ck::tensor_operation::element_wise::PassThrough,
ck::tensor_operation::element_wise::ReduceSum,
ck::tensor_operation::element_wise::ReduceSquareSum>;
DInElementOps,
DOutElementOps>;
void add_device_gemm_reduce_xdl_cshuffle_f16_f16_f16_f32_f32_mk_kn_mn_instances(
std::vector<DeviceGemmReduceNoOpPtr>&);
......@@ -53,7 +63,7 @@ template <typename ADataType,
bool profile_gemm_reduce_impl(int do_verification,
int init_method,
bool do_log,
int nrepeat,
bool time_kernel,
int M,
int N,
int K,
......@@ -113,11 +123,18 @@ bool profile_gemm_reduce_impl(int do_verification,
b_k_n.GenerateTensorValue(GeneratorTensor_3<BDataType>{-0.5, 0.5}, num_thread);
}
using AElementOp = ck::tensor_operation::element_wise::PassThrough;
using BElementOp = ck::tensor_operation::element_wise::PassThrough;
using CElementOp = ck::tensor_operation::element_wise::PassThrough;
using D0ReduceOp = ck::tensor_operation::element_wise::ReduceSum;
using D1ReduceOp = ck::tensor_operation::element_wise::ReduceSquareSum;
using AElementOp = ck::tensor_operation::element_wise::PassThrough;
using BElementOp = ck::tensor_operation::element_wise::PassThrough;
using CElementOp = ck::tensor_operation::element_wise::PassThrough;
using D0ReduceOp = ck::reduce::Add<float>;
using D1ReduceOp = ck::reduce::Add<float>;
using UnaryDivElementOp = ck::tensor_operation::element_wise::UnaryIdentic<float, float, true>;
using UnaryIdenticElementOp =
ck::tensor_operation::element_wise::UnaryIdentic<float, float, false>;
using UnarySquareElementOp =
ck::tensor_operation::element_wise::UnarySquare<float, float, false>;
using DxsInElementOps = ck::Tuple<UnaryIdenticElementOp, UnarySquareElementOp>;
using DxsOutElementOps = ck::Tuple<UnaryDivElementOp, UnaryDivElementOp>;
const auto a_element_op = AElementOp{};
const auto b_element_op = BElementOp{};
......@@ -125,10 +142,18 @@ bool profile_gemm_reduce_impl(int do_verification,
const auto d0_reduce_op = D0ReduceOp{};
const auto d1_reduce_op = D1ReduceOp{};
auto dxs_in_element_op = DxsInElementOps{};
auto dxs_out_element_op = DxsOutElementOps{M, M};
if(do_verification)
{
using ReferenceGemmInstance = ck::tensor_operation::host::
ReferenceGemm<ADataType, BDataType, CDataType, AElementOp, BElementOp, CElementOp>;
using ReferenceGemmInstance = ck::tensor_operation::host::ReferenceGemm<ADataType,
BDataType,
CDataType,
DDataType,
AElementOp,
BElementOp,
CElementOp>;
auto ref_gemm = ReferenceGemmInstance{};
auto ref_invoker = ref_gemm.MakeInvoker();
......@@ -140,17 +165,25 @@ bool profile_gemm_reduce_impl(int do_verification,
for(int m = 0; m < M; ++m)
{
float d0_acc = d0_reduce_op.GetReduceZeroValue();
float d1_acc = d1_reduce_op.GetReduceZeroValue();
float d0_acc = d0_reduce_op.GetIdentityValue();
float d1_acc = d1_reduce_op.GetIdentityValue();
for(int n = 0; n < N; ++n)
{
d0_reduce_op.Reduce(d0_acc, c_m_n_host_result(m, n));
d1_reduce_op.Reduce(d1_acc, c_m_n_host_result(m, n));
float c_val = ck::type_convert<float>(c_m_n_host_result(m, n));
float d0_val = 0;
float d1_val = 0;
dxs_in_element_op(ck::Number<0>{})(d0_val, c_val);
dxs_in_element_op(ck::Number<1>{})(d1_val, c_val);
d0_reduce_op(d0_acc, d0_val);
d1_reduce_op(d1_acc, d1_val);
}
d0_m_host_result(m) = d0_acc;
d1_m_host_result(m) = d1_acc;
dxs_out_element_op(ck::Number<0>{})(d0_acc, d0_acc);
dxs_out_element_op(ck::Number<1>{})(d1_acc, d1_acc);
d0_m_host_result(m) = ck::type_convert<DDataType>(d0_acc);
d1_m_host_result(m) = ck::type_convert<DDataType>(d1_acc);
}
}
......@@ -160,6 +193,9 @@ bool profile_gemm_reduce_impl(int do_verification,
DeviceMem d0_device_buf(sizeof(DDataType) * d0_m_device_result.mDesc.GetElementSpace());
DeviceMem d1_device_buf(sizeof(DDataType) * d1_m_device_result.mDesc.GetElementSpace());
auto dxs_global = ck::make_tuple(static_cast<DDataType*>(d0_device_buf.GetDeviceBuffer()),
static_cast<DDataType*>(d1_device_buf.GetDeviceBuffer()));
a_device_buf.ToDevice(a_m_k.mData.data());
b_device_buf.ToDevice(b_k_n.mData.data());
......@@ -221,8 +257,7 @@ bool profile_gemm_reduce_impl(int do_verification,
gemm_ptr->MakeArgumentPointer(static_cast<ADataType*>(a_device_buf.GetDeviceBuffer()),
static_cast<BDataType*>(b_device_buf.GetDeviceBuffer()),
static_cast<CDataType*>(c_device_buf.GetDeviceBuffer()),
static_cast<DDataType*>(d0_device_buf.GetDeviceBuffer()),
static_cast<DDataType*>(d1_device_buf.GetDeviceBuffer()),
dxs_global,
M,
N,
K,
......@@ -232,43 +267,25 @@ bool profile_gemm_reduce_impl(int do_verification,
a_element_op,
b_element_op,
c_element_op,
d0_reduce_op,
d1_reduce_op);
dxs_in_element_op,
dxs_out_element_op);
auto invoker_ptr = gemm_ptr->MakeInvokerPointer();
if(gemm_ptr->IsSupportedArgument(argument_ptr.get()))
{
// warm up
invoker_ptr->Run(argument_ptr.get());
// timing
float total_time = 0;
for(int i = 0; i < nrepeat; ++i)
{
// init DO, D1 to 0
d0_device_buf.SetZero();
d1_device_buf.SetZero();
KernelTimer timer;
timer.Start();
invoker_ptr->Run(argument_ptr.get());
timer.End();
total_time += timer.GetElapsedTime();
}
// init DO, D1 to 0
d0_device_buf.SetZero();
d1_device_buf.SetZero();
float ave_time = total_time / nrepeat;
float ave_time =
invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, time_kernel});
std::string gemm_name = gemm_ptr->GetTypeString();
std::size_t flop = std::size_t(2) * M * N * K;
std::size_t num_btype = sizeof(ADataType) * M * K + sizeof(BDataType) * K * M +
std::size_t num_btype = sizeof(ADataType) * M * K + sizeof(BDataType) * K * N +
sizeof(CDataType) * M * N + sizeof(CDataType) * N;
float tflops = static_cast<float>(flop) / 1.E9 / ave_time;
......
profiler/include/profile_grouped_gemm_impl.hpp
View file @ a3b4c5cb
......@@ -43,19 +43,20 @@ namespace profiler {
template <typename ADataType,
typename BDataType,
typename CDataType,
typename AccDataType,
typename ALayout,
typename BLayout,
typename CLayout>
void profile_grouped_gemm_impl(int do_verification,
int init_method,
bool do_log,
int nrepeat,
std::vector<int> Ms,
std::vector<int> Ns,
std::vector<int> Ks,
std::vector<int> StrideAs,
std::vector<int> StrideBs,
std::vector<int> StrideCs)
bool time_kernel,
const std::vector<int>& Ms,
const std::vector<int>& Ns,
const std::vector<int>& Ks,
const std::vector<int>& StrideAs,
const std::vector<int>& StrideBs,
const std::vector<int>& StrideCs)
{
auto f_host_tensor_descriptor =
[](std::size_t row, std::size_t col, std::size_t stride, auto layout) {
......@@ -71,7 +72,7 @@ void profile_grouped_gemm_impl(int do_verification,
}
};
int group_count = Ms.size();
std::size_t group_count = Ms.size();
if(!(group_count == Ns.size() && group_count == Ks.size() && group_count == StrideAs.size() &&
group_count == StrideBs.size() && group_count == StrideCs.size()))
......@@ -83,7 +84,7 @@ void profile_grouped_gemm_impl(int do_verification,
std::vector<Tensor<BDataType>> b_k_n;
std::vector<Tensor<CDataType>> c_m_n_device_results;
for(int i = 0; i < Ms.size(); i++)
for(std::size_t i = 0; i < group_count; i++)
{
a_m_k.push_back(
Tensor<ADataType>(f_host_tensor_descriptor(Ms[i], Ks[i], StrideAs[i], ALayout{})));
......@@ -144,7 +145,7 @@ void profile_grouped_gemm_impl(int do_verification,
gemm_shapes.reserve(group_count);
for(int i = 0; i < group_count; i++)
for(std::size_t i = 0; i < group_count; i++)
{
a_device_buf.emplace_back(
std::make_unique<DeviceMem>(sizeof(ADataType) * a_m_k[i].mDesc.GetElementSpace()));
......@@ -231,10 +232,11 @@ void profile_grouped_gemm_impl(int do_verification,
{
std::string gemm_name = gemm_ptr->GetTypeString();
float ave_time = invoker_ptr->Run(argument_ptr.get(), nrepeat);
float ave_time =
invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, time_kernel});
std::size_t flop = 0, num_btype = 0;
for(int i = 0; i < gemm_shapes.size(); i++)
for(std::size_t i = 0; i < gemm_shapes.size(); i++)
{
flop += std::size_t(2) * Ms[i] * Ns[i] * Ks[i];
......@@ -258,7 +260,7 @@ void profile_grouped_gemm_impl(int do_verification,
if(do_verification)
{
for(int i = 0; i < gemm_shapes.size(); i++)
for(std::size_t i = 0; i < gemm_shapes.size(); i++)
{
c_device_buf[i]->FromDevice(c_m_n_device_results[i].mData.data());
......@@ -270,6 +272,7 @@ void profile_grouped_gemm_impl(int do_verification,
ck::tensor_operation::host::ReferenceGemm<ADataType,
BDataType,
CDataType,
AccDataType,
AElementOp,
BElementOp,
CElementOp>;
......
profiler/include/profile_reduce_impl.hpp
View file @ a3b4c5cb
......@@ -5,74 +5,77 @@
#include "device_reduce_instance.hpp"
#include "reduction_enums.hpp"
#include "host_reduction.hpp"
#include "host_common_util.hpp"
#include "host_tensor_generator.hpp"
namespace ck {
namespace tensor_operation {
namespace device {
namespace device_reduce_instance {
template <int Rank, int NumReduceDim, int ReduceOpId, int NanOpt, int IndicesOpt>
template <int Rank, int NumReduceDim, int ReduceOpId, bool PropagateNan, bool UseIndex>
struct ReduceDescription
{
static constexpr int Rank_ = Rank;
static constexpr int NumReduceDim_ = NumReduceDim;
static constexpr int ReduceOpId_ = ReduceOpId;
static constexpr int NanOpt_ = NanOpt;
static constexpr int IndicesOpt_ = IndicesOpt;
static constexpr int PropagateNan_ = PropagateNan;
static constexpr int UseIndex_ = UseIndex;
};
using reduce_description_instances = std::tuple<ReduceDescription<4, 3, 0, 0, 0>, // for ADD
ReduceDescription<4, 4, 0, 0, 0>,
ReduceDescription<4, 1, 0, 0, 0>,
ReduceDescription<2, 1, 0, 0, 0>,
ReduceDescription<4, 3, 5, 0, 0>, // for AVG
ReduceDescription<4, 4, 5, 0, 0>,
ReduceDescription<4, 1, 5, 0, 0>,
ReduceDescription<2, 1, 5, 0, 0>,
ReduceDescription<4, 3, 7, 0, 0>, // for NORM2
ReduceDescription<4, 4, 7, 0, 0>,
ReduceDescription<4, 1, 7, 0, 0>,
ReduceDescription<2, 1, 7, 0, 0>,
ReduceDescription<4, 3, 2, 0, 0>, // for MIN
ReduceDescription<4, 4, 2, 0, 0>,
ReduceDescription<4, 1, 2, 0, 0>,
ReduceDescription<2, 1, 2, 0, 0>,
ReduceDescription<4, 3, 3, 0, 0>, // for MAX
ReduceDescription<4, 4, 3, 0, 0>,
ReduceDescription<4, 1, 3, 0, 0>,
ReduceDescription<2, 1, 3, 0, 0>,
ReduceDescription<4, 3, 4, 0, 0>, // for AMAX
ReduceDescription<4, 4, 4, 0, 0>,
ReduceDescription<4, 1, 4, 0, 0>,
ReduceDescription<2, 1, 4, 0, 0>,
ReduceDescription<4, 3, 2, 0, 1>, // for MIN
ReduceDescription<4, 4, 2, 0, 1>,
ReduceDescription<4, 1, 2, 0, 1>,
ReduceDescription<2, 1, 2, 0, 1>,
ReduceDescription<4, 3, 3, 0, 1>, // for MAX
ReduceDescription<4, 4, 3, 0, 1>,
ReduceDescription<4, 1, 3, 0, 1>,
ReduceDescription<2, 1, 3, 0, 1>,
ReduceDescription<4, 3, 4, 0, 1>, // for AMAX
ReduceDescription<4, 4, 4, 0, 1>,
ReduceDescription<4, 1, 4, 0, 1>,
ReduceDescription<2, 1, 4, 0, 1>>;
using reduce_description_instances =
std::tuple<ReduceDescription<4, 3, 0, false, false>, // for ADD
ReduceDescription<4, 4, 0, false, false>,
ReduceDescription<4, 1, 0, false, false>,
ReduceDescription<2, 1, 0, false, false>,
ReduceDescription<4, 3, 5, false, false>, // for AVG
ReduceDescription<4, 4, 5, false, false>,
ReduceDescription<4, 1, 5, false, false>,
ReduceDescription<2, 1, 5, false, false>,
ReduceDescription<4, 3, 7, false, false>, // for NORM2
ReduceDescription<4, 4, 7, false, false>,
ReduceDescription<4, 1, 7, false, false>,
ReduceDescription<2, 1, 7, false, false>,
ReduceDescription<4, 3, 2, false, false>, // for MIN
ReduceDescription<4, 4, 2, false, false>,
ReduceDescription<4, 1, 2, false, false>,
ReduceDescription<2, 1, 2, false, false>,
ReduceDescription<4, 3, 3, false, false>, // for MAX
ReduceDescription<4, 4, 3, false, false>,
ReduceDescription<4, 1, 3, false, false>,
ReduceDescription<2, 1, 3, false, false>,
ReduceDescription<4, 3, 4, false, false>, // for AMAX
ReduceDescription<4, 4, 4, false, false>,
ReduceDescription<4, 1, 4, false, false>,
ReduceDescription<2, 1, 4, false, false>,
ReduceDescription<4, 3, 2, false, true>, // for MIN
ReduceDescription<4, 4, 2, false, true>,
ReduceDescription<4, 1, 2, false, true>,
ReduceDescription<2, 1, 2, false, true>,
ReduceDescription<4, 3, 3, false, true>, // for MAX
ReduceDescription<4, 4, 3, false, true>,
ReduceDescription<4, 1, 3, false, true>,
ReduceDescription<2, 1, 3, false, true>,
ReduceDescription<4, 3, 4, false, true>, // for AMAX
ReduceDescription<4, 4, 4, false, true>,
ReduceDescription<4, 1, 4, false, true>,
ReduceDescription<2, 1, 4, false, true>>;
template <typename DescriptionType>
bool description_match(const DescriptionType& description,
int Rank,
const std::vector<int>& reduceDims,
ReduceTensorOp ReduceOpId,
NanPropagation NanOpt,
ReduceTensorIndices IndicesOpt)
bool PropagateNan,
bool UseIndex)
{
if(description.Rank_ != Rank || description.ReduceOpId_ != static_cast<int>(ReduceOpId) ||
description.NanOpt_ != static_cast<int>(NanOpt) ||
description.IndicesOpt_ != static_cast<int>(IndicesOpt))
description.PropagateNan_ != static_cast<int>(PropagateNan) ||
description.UseIndex_ != static_cast<int>(UseIndex))
return (false);
if(DescriptionType::NumReduceDim_ != reduceDims.size())
......@@ -116,48 +119,18 @@ static inline std::vector<int> get_invariant_dims(const std::vector<int>& reduce
return invariantDims;
};
template <typename T>
static void dumpBufferToFile(const char* fileName, T* data, size_t dataNumItems)
{
std::ofstream outFile(fileName, std::ios::binary);
if(outFile)
{
outFile.write(reinterpret_cast<char*>(data), dataNumItems * sizeof(T));
outFile.close();
std::cout << "Write output to file " << fileName << std::endl;
}
else
{
std::cout << "Could not open file " << fileName << " for writing" << std::endl;
}
};
// map the data type used by the GPU kernels to the corresponding type used by the host codes
template <typename InType>
struct type_mapping
{
using OutType = InType;
};
template <>
struct type_mapping<ck::half_t>
{
using OutType = half_float::half;
};
template <typename InDataType,
typename AccDataType,
typename OutDataType,
int Rank,
int NumReduceDim,
ReduceTensorOp ReduceOpId,
NanPropagation NanOpt,
ReduceTensorIndices IndicesOpt>
void profile_reduce_impl_impl(bool do_verification,
bool PropagateNan,
bool UseIndex>
bool profile_reduce_impl_impl(bool do_verification,
int init_method,
bool do_log,
bool do_dumpout,
int nrepeat,
bool time_kernel,
const std::vector<size_t>& inLengths,
const std::vector<int>& reduceDims,
float alpha,
......@@ -165,16 +138,13 @@ void profile_reduce_impl_impl(bool do_verification,
{
using namespace ck::tensor_operation::device;
using namespace ck::tensor_operation::device::device_reduce_instance;
using namespace ck::host_reduce;
using ck::host_common::dumpBufferToFile;
constexpr bool op_support_indices =
(ReduceOpId == ReduceTensorOp::MIN || ReduceOpId == ReduceTensorOp::MAX ||
ReduceOpId == ReduceTensorOp::AMAX);
constexpr bool NeedIndices =
(op_support_indices && (IndicesOpt != ReduceTensorIndices::NO_INDICES));
constexpr bool PropagateNan = (NanOpt == NanPropagation::PROPAGATE_NAN);
constexpr bool OutputIndex = (op_support_indices && UseIndex);
constexpr bool out_support_atomic_add = std::is_same<OutDataType, float>::value;
constexpr bool op_support_atomic_add =
......@@ -195,8 +165,7 @@ void profile_reduce_impl_impl(bool do_verification,
(op_support_indices && !std::is_same<AccDataType, float>::value);
// 1) The indices can only be used when the reduction operation is indexable
constexpr bool invalid_reduce_3 =
(!op_support_indices && IndicesOpt != ReduceTensorIndices::NO_INDICES);
constexpr bool invalid_reduce_3 = (!op_support_indices && UseIndex);
// 1) If InDataType is int8_t, must use int8_t as AccDataType for indexable reduction operations
// 2) If InDataType is int8_t, must use int32_t as AccDataType for non-indexable reduction
......@@ -219,6 +188,8 @@ void profile_reduce_impl_impl(bool do_verification,
constexpr bool invalid_reduce = (invalid_reduce_1 || invalid_reduce_2 || invalid_reduce_3 ||
invalid_reduce_4 || invalid_reduce_5 || invalid_reduce_6);
bool pass = true;
if constexpr(!invalid_reduce)
{
Tensor<InDataType> in(inLengths);
......@@ -282,42 +253,26 @@ void profile_reduce_impl_impl(bool do_verification,
if(beta != 0.0f)
out_dev.ToDevice(out.mData.data());
size_t indicesSizeInBytes = NeedIndices ? out.mDesc.GetElementSize() * sizeof(int) : 0;
size_t indicesSizeInBytes = OutputIndex ? out.mDesc.GetElementSize() * sizeof(int) : 0;
DeviceMem out_indices_dev(indicesSizeInBytes);
float best_avg_time = 0;
float best_gb_per_sec = 0;
using InElementwiseOperation_0 =
using InElementwiseOperation =
typename reduce_unary_operator<AccDataType, ReduceOpId, true, true>::
InElementwiseOperation;
using AccElementwiseOperation_0 =
using AccElementwiseOperation =
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 ReduceOperation = typename reduce_binary_operator<AccDataType, ReduceOpId>::opType;
using DeviceReduceInstPtr0 =
DeviceReducePtr<InElementwiseOperation_0, AccElementwiseOperation_0>;
using DeviceReduceInstPtr1 =
DeviceReducePtr<InElementwiseOperation_1, AccElementwiseOperation_1>;
using DeviceReduceInstPtr2 =
DeviceReducePtr<InElementwiseOperation_2, AccElementwiseOperation_2>;
DeviceReducePtr<InElementwiseOperation, AccElementwiseOperation>;
std::vector<DeviceReduceInstPtr0> reduce0_ptrs;
std::vector<DeviceReduceInstPtr1> reduce1_ptrs;
std::vector<DeviceReduceInstPtr2> reduce2_ptrs;
add_device_reduce_instance_threadwise<InDataType,
AccDataType,
......@@ -325,8 +280,8 @@ void profile_reduce_impl_impl(bool do_verification,
Rank,
NumReduceDim,
ReduceOpId,
NanOpt,
IndicesOpt>(reduce0_ptrs);
PropagateNan,
UseIndex>(reduce0_ptrs);
add_device_reduce_instance_blockwise<InDataType,
AccDataType,
......@@ -334,8 +289,8 @@ void profile_reduce_impl_impl(bool do_verification,
Rank,
NumReduceDim,
ReduceOpId,
NanOpt,
IndicesOpt>(reduce0_ptrs);
PropagateNan,
UseIndex>(reduce0_ptrs);
if constexpr(use_atomic_add)
{
......@@ -345,35 +300,11 @@ void profile_reduce_impl_impl(bool do_verification,
Rank,
NumReduceDim,
ReduceOpId,
NanOpt,
IndicesOpt>(reduce0_ptrs);
PropagateNan,
UseIndex>(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())
if(reduce0_ptrs.empty())
{
throw std::runtime_error("Wrong! No device REDUCE instance found");
};
......@@ -383,31 +314,34 @@ void profile_reduce_impl_impl(bool do_verification,
ReductionHost<InDataType,
AccDataType,
OutDataType,
ReduceOpId,
ReduceOperation,
InElementwiseOperation,
AccElementwiseOperation,
Rank,
NumReduceDim,
PropagateNan,
NeedIndices>
OutputIndex>
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);
std::vector<ck::index_t> i_inLengths;
std::vector<ck::index_t> i_inStrides;
std::vector<ck::index_t> i_outLengths;
std::vector<ck::index_t> i_outStrides;
i_inLengths.assign(inLengths.begin(), inLengths.end());
i_inStrides.assign(inStrides.begin(), inStrides.end());
i_outLengths.assign(outLengths.begin(), outLengths.end());
i_outStrides.assign(outStrides.begin(), outStrides.end());
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));
InElementwiseOperation in_elementwise_op(static_cast<int32_t>(reduce_total_length));
AccElementwiseOperation acc_elementwise_op(static_cast<int32_t>(reduce_total_length));
auto argument_ptr = reduce_ptr->MakeArgumentPointer(i_inLengths,
i_inStrides,
......@@ -417,11 +351,11 @@ void profile_reduce_impl_impl(bool do_verification,
alpha,
beta,
in_dev.GetDeviceBuffer(),
nullptr,
out_dev.GetDeviceBuffer(),
out_indices_dev.GetDeviceBuffer(),
ws_dev.GetDeviceBuffer(),
in_elementwise_op_0,
acc_elementwise_op_0);
in_elementwise_op,
acc_elementwise_op);
if(!reduce_ptr->IsSupportedArgument(argument_ptr.get()))
continue;
......@@ -430,7 +364,8 @@ void profile_reduce_impl_impl(bool do_verification,
auto invoker_ptr = reduce_ptr->MakeInvokerPointer();
float avg_time = invoker_ptr->Run(argument_ptr.get(), nrepeat);
float avg_time =
invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, time_kernel});
std::size_t num_bytes =
invariant_total_length * reduce_total_length * sizeof(InDataType) +
......@@ -438,8 +373,9 @@ void profile_reduce_impl_impl(bool do_verification,
float gb_per_sec = num_bytes / 1.E6 / avg_time;
std::cout << "Perf: " << avg_time << " ms, " << gb_per_sec << " GB/s, " << reduce_name
<< std::endl;
if(time_kernel)
std::cout << "Perf: " << avg_time << " ms, " << gb_per_sec << " GB/s, "
<< reduce_name << std::endl;
if(gb_per_sec > best_gb_per_sec)
{
......@@ -449,22 +385,24 @@ void profile_reduce_impl_impl(bool do_verification,
if(do_verification)
{
bool single_pass;
out_dev.FromDevice(out.mData.data());
ck::utils::check_err(out.mData, out_ref.mData);
single_pass = ck::utils::check_err(out.mData, out_ref.mData);
if(NeedIndices)
if(OutputIndex)
{
out_indices_dev.FromDevice(out_indices.mData.data());
ck::utils::check_err(out_indices.mData, out_indices_ref.mData);
;
single_pass = single_pass &&
ck::utils::check_err(out_indices.mData, out_indices_ref.mData);
};
if(do_log)
if(!single_pass)
{
LogRangeAsType<float>(std::cout << "out_host : ", out_ref.mData, ",")
<< std::endl;
LogRangeAsType<float>(std::cout << "out_device: ", out.mData, ",") << std::endl;
};
std::cout << "Fail Info: " << reduce_ptr->GetTypeString() << std::endl;
}
pass = pass && single_pass;
};
if(do_dumpout)
......@@ -473,7 +411,7 @@ void profile_reduce_impl_impl(bool do_verification,
dumpBufferToFile("dump_out.bin", out.mData.data(), out.mDesc.GetElementSize());
dumpBufferToFile(
"dump_out_host.bin", out_ref.mData.data(), out_ref.mDesc.GetElementSize());
if(NeedIndices)
if(OutputIndex)
{
dumpBufferToFile("dump_indices.bin",
out_indices.mData.data(),
......@@ -485,156 +423,34 @@ void profile_reduce_impl_impl(bool do_verification,
};
};
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();
float avg_time = invoker_ptr->Run(argument_ptr.get(), nrepeat);
std::size_t num_bytes =
invariant_total_length * reduce_total_length * sizeof(InDataType) +
invariant_total_length * sizeof(OutDataType);
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();
float avg_time_2 = invoker2_ptr->Run(argument2_ptr.get(), nrepeat);
std::size_t num_bytes_2 =
static_cast<size_t>(inLengths2[0]) * inLengths2[1] * sizeof(AccDataType);
float gb_per_sec = (num_bytes + num_bytes_2) / 1.E6 / (avg_time + avg_time_2);
std::cout << "Perf: " << (avg_time + avg_time_2) << " ms, " << gb_per_sec
<< " GB/s, " << reduce_name << " => " << reduce2_name << std::endl;
if(gb_per_sec > best_gb_per_sec)
{
best_avg_time = avg_time + avg_time_2;
best_gb_per_sec = gb_per_sec;
}
if(do_verification)
{
out_dev.FromDevice(out.mData.data());
ck::utils::check_err(out.mData, out_ref.mData);
if(NeedIndices)
{
out_indices_dev.FromDevice(out_indices.mData.data());
ck::utils::check_err(out_indices.mData, out_indices_ref.mData);
;
};
if(do_log)
{
LogRangeAsType<float>(std::cout << "out_host : ", out_ref.mData, ",")
<< std::endl;
LogRangeAsType<float>(std::cout << "out_device: ", out.mData, ",")
<< std::endl;
}
}
if(do_dumpout)
{
dumpBufferToFile("dump_in.bin", in.mData.data(), in.mDesc.GetElementSize());
dumpBufferToFile("dump_out.bin", out.mData.data(), out.mDesc.GetElementSize());
dumpBufferToFile(
"dump_out_host.bin", out_ref.mData.data(), out_ref.mDesc.GetElementSize());
if(NeedIndices)
{
dumpBufferToFile("dump_indices.bin",
out_indices.mData.data(),
out_indices.mDesc.GetElementSize());
dumpBufferToFile("dump_indices_host.bin",
out_indices_ref.mData.data(),
out_indices_ref.mDesc.GetElementSize());
};
};
};
};
std::cout << "Best Perf: " << best_avg_time << " ms, " << best_gb_per_sec << " GB/s"
<< std::endl;
if(time_kernel)
std::cout << "Best Perf: " << best_avg_time << " ms, " << best_gb_per_sec << " GB/s"
<< std::endl;
}
else
{
std::cout << "The requested reduction operation is not supported, please check !!!"
<< std::endl;
};
return pass;
};
template <typename InDataType, typename AccDataType, typename OutDataType>
void profile_reduce_impl(bool do_verification,
bool profile_reduce_impl(bool do_verification,
int init_method,
bool do_log,
bool do_dumpout,
int nrepeat,
bool time_kernel,
const std::vector<size_t>& inLengths,
const std::vector<int>& reduceDims,
ReduceTensorOp ReduceOpId,
NanPropagation NanOpt,
ReduceTensorIndices IndicesOpt,
bool PropagateNan,
bool UseIndex,
float alpha,
float beta)
{
bool matched = false;
bool pass = true;
using tuple_of_description_instances =
tensor_operation::device::device_reduce_instance::reduce_description_instances;
......@@ -648,29 +464,30 @@ void profile_reduce_impl(bool do_verification,
using descType = remove_cvref_t<decltype(std::get<i>(tuple_object))>;
if(!description_match(
descType{}, inLengths.size(), reduceDims, ReduceOpId, NanOpt, IndicesOpt))
descType{}, inLengths.size(), reduceDims, ReduceOpId, PropagateNan, UseIndex))
return;
profile_reduce_impl_impl<InDataType,
AccDataType,
OutDataType,
descType::Rank_,
descType::NumReduceDim_,
static_cast<ReduceTensorOp>(descType::ReduceOpId_),
static_cast<NanPropagation>(descType::NanOpt_),
static_cast<ReduceTensorIndices>(descType::IndicesOpt_)>(
do_verification,
init_method,
do_log,
do_dumpout,
nrepeat,
inLengths,
reduceDims,
alpha,
beta);
pass = pass &&
profile_reduce_impl_impl<InDataType,
AccDataType,
OutDataType,
descType::Rank_,
descType::NumReduceDim_,
static_cast<ReduceTensorOp>(descType::ReduceOpId_),
static_cast<bool>(descType::PropagateNan_),
static_cast<bool>(descType::UseIndex_)>(do_verification,
init_method,
do_dumpout,
time_kernel,
inLengths,
reduceDims,
alpha,
beta);
matched = true;
});
return pass;
};
} // namespace profiler
......
profiler/src/profile_batched_gemm.cpp
View file @ a3b4c5cb
......@@ -48,8 +48,8 @@ int profile_batched_gemm(int argc, char* argv[])
printf(" 3: A[g, k, m] * B[g, n, k] = C[g, m, n])\n");
printf("arg4: verification (0: no; 1: yes)\n");
printf("arg5: initialization (0: no init; 1: integer value; 2: decimal value)\n");
printf("arg8: print tensor value (0: no; 1: yes)\n");
printf("arg7: run kernel # of times (>1)\n");
printf("arg6: print tensor value (0: no; 1: yes)\n");
printf("arg7: time kernel (0=n0, 1=yes)\n");
printf("arg8 to 14: M, N, K, StrideA, StrideB, StrideC, BatchCount\n");
exit(1);
}
......@@ -59,7 +59,7 @@ int profile_batched_gemm(int argc, char* argv[])
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 int nrepeat = std::stoi(argv[7]);
const bool time_kernel = std::stoi(argv[7]);
const int M = std::stoi(argv[8]);
const int N = std::stoi(argv[9]);
......@@ -82,7 +82,7 @@ int profile_batched_gemm(int argc, char* argv[])
do_verification,
init_method,
do_log,
nrepeat,
time_kernel,
M,
N,
K,
......@@ -102,7 +102,7 @@ int profile_batched_gemm(int argc, char* argv[])
do_verification,
init_method,
do_log,
nrepeat,
time_kernel,
M,
N,
K,
......@@ -122,7 +122,7 @@ int profile_batched_gemm(int argc, char* argv[])
do_verification,
init_method,
do_log,
nrepeat,
time_kernel,
M,
N,
K,
......@@ -142,7 +142,7 @@ int profile_batched_gemm(int argc, char* argv[])
do_verification,
init_method,
do_log,
nrepeat,
time_kernel,
M,
N,
K,
......@@ -162,7 +162,7 @@ int profile_batched_gemm(int argc, char* argv[])
do_verification,
init_method,
do_log,
nrepeat,
time_kernel,
M,
N,
K,
......@@ -182,7 +182,7 @@ int profile_batched_gemm(int argc, char* argv[])
do_verification,
init_method,
do_log,
nrepeat,
time_kernel,
M,
N,
K,
......@@ -202,7 +202,7 @@ int profile_batched_gemm(int argc, char* argv[])
do_verification,
init_method,
do_log,
nrepeat,
time_kernel,
M,
N,
K,
......@@ -222,7 +222,7 @@ int profile_batched_gemm(int argc, char* argv[])
do_verification,
init_method,
do_log,
nrepeat,
time_kernel,
M,
N,
K,
......@@ -242,7 +242,7 @@ int profile_batched_gemm(int argc, char* argv[])
do_verification,
init_method,
do_log,
nrepeat,
time_kernel,
M,
N,
K,
......@@ -262,7 +262,7 @@ int profile_batched_gemm(int argc, char* argv[])
do_verification,
init_method,
do_log,
nrepeat,
time_kernel,
M,
N,
K,
......@@ -282,7 +282,7 @@ int profile_batched_gemm(int argc, char* argv[])
do_verification,
init_method,
do_log,
nrepeat,
time_kernel,
M,
N,
K,
......@@ -302,7 +302,7 @@ int profile_batched_gemm(int argc, char* argv[])
do_verification,
init_method,
do_log,
nrepeat,
time_kernel,
M,
N,
K,
......@@ -322,7 +322,7 @@ int profile_batched_gemm(int argc, char* argv[])
do_verification,
init_method,
do_log,
nrepeat,
time_kernel,
M,
N,
K,
......@@ -342,7 +342,7 @@ int profile_batched_gemm(int argc, char* argv[])
do_verification,
init_method,
do_log,
nrepeat,
time_kernel,
M,
N,
K,
......@@ -362,7 +362,7 @@ int profile_batched_gemm(int argc, char* argv[])
do_verification,
init_method,
do_log,
nrepeat,
time_kernel,
M,
N,
K,
......@@ -382,7 +382,7 @@ int profile_batched_gemm(int argc, char* argv[])
do_verification,
init_method,
do_log,
nrepeat,
time_kernel,
M,
N,
K,
......@@ -396,5 +396,5 @@ int profile_batched_gemm(int argc, char* argv[])
throw std::runtime_error("wrong! this GEMM data_type & layout is not implemented");
}
return 1;
return 0;
}
profiler/src/profile_batched_gemm_reduce.cpp
View file @ a3b4c5cb
......@@ -33,8 +33,8 @@ int profile_batched_gemm_reduce(int argc, char* argv[])
printf(" 3: A[k, m] * B[n, k] = C[m, n])\n");
printf("arg4: verification (0: no; 1: yes)\n");
printf("arg5: initialization (0: no init; 1: integer value; 2: decimal value)\n");
printf("arg8: print tensor value (0: no; 1: yes)\n");
printf("arg7: run kernel # of times (>1)\n");
printf("arg6: print tensor value (0: no; 1: yes)\n");
printf("arg7: time kernel (0=n0, 1=yes)\n");
printf("arg8 to 14: M, N, K, StrideA, StrideB, StrideC, BatchCount\n");
printf("arg15: split k into mulitiple batch\n");
exit(1);
......@@ -45,7 +45,7 @@ int profile_batched_gemm_reduce(int argc, char* argv[])
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 int nrepeat = std::stoi(argv[7]);
const bool time_kernel = std::stoi(argv[7]);
const int M = std::stoi(argv[8]);
const int N = std::stoi(argv[9]);
......@@ -69,7 +69,7 @@ int profile_batched_gemm_reduce(int argc, char* argv[])
do_verification,
init_method,
do_log,
nrepeat,
time_kernel,
M,
N,
K,
......@@ -91,7 +91,7 @@ int profile_batched_gemm_reduce(int argc, char* argv[])
do_verification,
init_method,
do_log,
nrepeat,
time_kernel,
M,
N,
K,
......@@ -113,7 +113,7 @@ int profile_batched_gemm_reduce(int argc, char* argv[])
do_verification,
init_method,
do_log,
nrepeat,
time_kernel,
M,
N,
K,
......@@ -135,7 +135,7 @@ int profile_batched_gemm_reduce(int argc, char* argv[])
do_verification,
init_method,
do_log,
nrepeat,
time_kernel,
M,
N,
K,
......@@ -149,5 +149,5 @@ int profile_batched_gemm_reduce(int argc, char* argv[])
throw std::runtime_error("wrong! this data_type & layout is not implemented");
}
return 1;
return 0;
}
profiler/src/profile_conv_bwd_data.cpp deleted 100644 → 0
View file @ 48918ab9
#include <iostream>
#include <numeric>
#include <initializer_list>
#include <cstdlib>
#include <stdlib.h>
#include <half.hpp>
#include "profile_conv_bwd_data_impl.hpp"
enum struct ConvDataType
{
F32_F32_F32, // 0
F16_F16_F16, // 1
BF16_BF16_BF16, // 2
INT8_INT8_INT8, // 3
};
enum struct ConvInputLayout
{
NCHW, // 0
NHWC, // 1
};
enum struct ConvWeightLayout
{
KCYX, // 0
KYXC, // 1
};
enum struct ConvOutputLayout
{
NKHW, // 0
NHWK, // 1
};
int profile_conv_bwd_data(int argc, char* argv[])
{
if(argc != 25)
{
printf("arg1: tensor operation (conv_bwd: BackwardConvolution)\n");
printf("arg2: data type (0: fp32; 1: fp16)\n");
printf("arg3: input tensor layout (0: NCHW; 1: NHWC)\n");
printf("arg4: weight tensor layout (0: KCYX; 1: KYXC)\n");
printf("arg5: output tensor layout (0: NKHW; 1: NHWK)\n");
printf("arg6: verification (0: no; 1: yes)\n");
printf("arg7: initialization (0: no init; 1: integer value; 2: decimal value)\n");
printf("arg8: print tensor value (0: no; 1: yes)\n");
printf("arg9: run kernel # of times (>1)\n");
printf("arg10 to 24: N, K, C, Y, X, Hi, Wi, Sy, Sx, Dy, Dx, LeftPy, LeftPx, RightPy, "
"RightPx\n");
exit(1);
}
const auto data_type = static_cast<ConvDataType>(std::stoi(argv[2]));
const auto in_layout = static_cast<ConvInputLayout>(std::stoi(argv[3]));
const auto wei_layout = static_cast<ConvWeightLayout>(std::stoi(argv[4]));
const auto out_layout = static_cast<ConvOutputLayout>(std::stoi(argv[5]));
const bool do_verification = std::stoi(argv[6]);
const int init_method = std::stoi(argv[7]);
const bool do_log = std::stoi(argv[8]);
const int nrepeat = std::stoi(argv[9]);
const ck::index_t N = std::stoi(argv[10]);
const ck::index_t K = std::stoi(argv[11]);
const ck::index_t C = std::stoi(argv[12]);
const ck::index_t Y = std::stoi(argv[13]);
const ck::index_t X = std::stoi(argv[14]);
const ck::index_t Hi = std::stoi(argv[15]);
const ck::index_t Wi = std::stoi(argv[16]);
const ck::index_t conv_stride_h = std::stoi(argv[17]);
const ck::index_t conv_stride_w = std::stoi(argv[18]);
const ck::index_t conv_dilation_h = std::stoi(argv[19]);
const ck::index_t conv_dilation_w = std::stoi(argv[20]);
const ck::index_t in_left_pad_h = std::stoi(argv[21]);
const ck::index_t in_left_pad_w = std::stoi(argv[22]);
const ck::index_t in_right_pad_h = std::stoi(argv[23]);
const ck::index_t in_right_pad_w = std::stoi(argv[24]);
const ck::index_t YEff = (Y - 1) * conv_dilation_h + 1;
const ck::index_t XEff = (X - 1) * conv_dilation_w + 1;
const ck::index_t Ho = (Hi + in_left_pad_h + in_right_pad_h - YEff) / conv_stride_h + 1;
const ck::index_t Wo = (Wi + in_left_pad_w + in_right_pad_w - XEff) / conv_stride_w + 1;
if(data_type == ConvDataType::F32_F32_F32 && in_layout == ConvInputLayout::NHWC &&
wei_layout == ConvWeightLayout::KYXC && out_layout == ConvOutputLayout::NHWK)
{
ck::profiler::profile_conv_bwd_data_impl<2,
float,
float,
float,
float,
ck::tensor_layout::convolution::NHWC,
ck::tensor_layout::convolution::KYXC,
ck::tensor_layout::convolution::NHWK>(
do_verification,
init_method,
do_log,
nrepeat,
N,
K,
C,
std::vector<ck::index_t>{Hi, Wi},
std::vector<ck::index_t>{Y, X},
std::vector<ck::index_t>{Ho, Wo},
std::vector<ck::index_t>{conv_stride_h, conv_stride_w},
std::vector<ck::index_t>{conv_dilation_h, conv_dilation_w},
std::vector<ck::index_t>{in_left_pad_h, in_left_pad_w},
std::vector<ck::index_t>{in_right_pad_h, in_right_pad_w});
}
else if(data_type == ConvDataType::F16_F16_F16 && in_layout == ConvInputLayout::NHWC &&
wei_layout == ConvWeightLayout::KYXC && out_layout == ConvOutputLayout::NHWK)
{
ck::profiler::profile_conv_bwd_data_impl<2,
ck::half_t,
ck::half_t,
ck::half_t,
float,
ck::tensor_layout::convolution::NHWC,
ck::tensor_layout::convolution::KYXC,
ck::tensor_layout::convolution::NHWK>(
do_verification,
init_method,
do_log,
nrepeat,
N,
K,
C,
std::vector<ck::index_t>{Hi, Wi},
std::vector<ck::index_t>{Y, X},
std::vector<ck::index_t>{Ho, Wo},
std::vector<ck::index_t>{conv_stride_h, conv_stride_w},
std::vector<ck::index_t>{conv_dilation_h, conv_dilation_w},
std::vector<ck::index_t>{in_left_pad_h, in_left_pad_w},
std::vector<ck::index_t>{in_right_pad_h, in_right_pad_w});
}
else if(data_type == ConvDataType::BF16_BF16_BF16 && in_layout == ConvInputLayout::NHWC &&
wei_layout == ConvWeightLayout::KYXC && out_layout == ConvOutputLayout::NHWK)
{
ck::profiler::profile_conv_bwd_data_impl<2,
uint16_t,
uint16_t,
uint16_t,
float,
ck::tensor_layout::convolution::NHWC,
ck::tensor_layout::convolution::KYXC,
ck::tensor_layout::convolution::NHWK>(
do_verification,
init_method,
do_log,
nrepeat,
N,
K,
C,
std::vector<ck::index_t>{Hi, Wi},
std::vector<ck::index_t>{Y, X},
std::vector<ck::index_t>{Ho, Wo},
std::vector<ck::index_t>{conv_stride_h, conv_stride_w},
std::vector<ck::index_t>{conv_dilation_h, conv_dilation_w},
std::vector<ck::index_t>{in_left_pad_h, in_left_pad_w},
std::vector<ck::index_t>{in_right_pad_h, in_right_pad_w});
}
else if(data_type == ConvDataType::INT8_INT8_INT8 && in_layout == ConvInputLayout::NHWC &&
wei_layout == ConvWeightLayout::KYXC && out_layout == ConvOutputLayout::NHWK)
{
ck::profiler::profile_conv_bwd_data_impl<2,
int8_t,
int8_t,
int8_t,
int32_t,
ck::tensor_layout::convolution::NHWC,
ck::tensor_layout::convolution::KYXC,
ck::tensor_layout::convolution::NHWK>(
do_verification,
init_method,
do_log,
nrepeat,
N,
K,
C,
std::vector<ck::index_t>{Hi, Wi},
std::vector<ck::index_t>{Y, X},
std::vector<ck::index_t>{Ho, Wo},
std::vector<ck::index_t>{conv_stride_h, conv_stride_w},
std::vector<ck::index_t>{conv_dilation_h, conv_dilation_w},
std::vector<ck::index_t>{in_left_pad_h, in_left_pad_w},
std::vector<ck::index_t>{in_right_pad_h, in_right_pad_w});
}
else
{
throw std::runtime_error("wrong! this Conv data_type & layout is not implemented");
}
return 1;
}
profiler/src/profile_conv_bwd_weight.cpp
View file @ a3b4c5cb
......@@ -58,7 +58,7 @@ int profile_conv_bwd_weight(int argc, char* argv[])
const bool do_verification = std::stoi(argv[6]);
const int init_method = std::stoi(argv[7]);
const bool do_log = std::stoi(argv[8]);
const int nrepeat = std::stoi(argv[9]);
const bool time_kernel = std::stoi(argv[9]);
const ck::index_t N = std::stoi(argv[10]);
const ck::index_t K = std::stoi(argv[11]);
......@@ -98,7 +98,7 @@ int profile_conv_bwd_weight(int argc, char* argv[])
do_verification,
init_method,
do_log,
nrepeat,
time_kernel,
N,
K,
C,
......@@ -124,7 +124,7 @@ int profile_conv_bwd_weight(int argc, char* argv[])
do_verification,
init_method,
do_log,
nrepeat,
time_kernel,
N,
K,
C,
......@@ -142,5 +142,5 @@ int profile_conv_bwd_weight(int argc, char* argv[])
throw std::runtime_error("wrong! this Conv data_type & layout is not implemented");
}
return 1;
return 0;
}
profiler/src/profile_conv_fwd_bias_relu.cpp
View file @ a3b4c5cb
......@@ -42,7 +42,7 @@ int profile_conv_fwd_bias_relu(int argc, char* argv[])
printf("arg6: verification (0: no; 1: yes)\n");
printf("arg7: initialization (0: no init; 1: integer value; 2: decimal value)\n");
printf("arg8: print tensor value (0: no; 1: yes)\n");
printf("arg9: run kernel # of times (>1)\n");
printf("arg9: time kernel (0=n0, 1=yes)\n");
printf("arg10 to 24: N, K, C, Y, X, Hi, Wi, Sy, Sx, Dy, Dx, LeftPy, LeftPx, RightPy, "
"RightPx\n");
exit(1);
......@@ -55,7 +55,7 @@ int profile_conv_fwd_bias_relu(int argc, char* argv[])
const bool do_verification = std::stoi(argv[6]);
const int init_method = std::stoi(argv[7]);
const bool do_log = std::stoi(argv[8]);
const int nrepeat = std::stoi(argv[9]);
const bool time_kernel = std::stoi(argv[9]);
const ck::index_t N = std::stoi(argv[10]);
const ck::index_t K = std::stoi(argv[11]);
......@@ -93,7 +93,7 @@ int profile_conv_fwd_bias_relu(int argc, char* argv[])
do_verification,
init_method,
do_log,
nrepeat,
time_kernel,
N,
K,
C,
......@@ -110,5 +110,5 @@ int profile_conv_fwd_bias_relu(int argc, char* argv[])
throw std::runtime_error("wrong! data_type & layout for this operator is not implemented");
}
return 1;
return 0;
}
profiler/src/profile_conv_fwd_bias_relu_add.cpp
View file @ a3b4c5cb
......@@ -43,7 +43,7 @@ int profile_conv_fwd_bias_relu_add(int argc, char* argv[])
printf("arg6: verification (0: no; 1: yes)\n");
printf("arg7: initialization (0: no init; 1: integer value; 2: decimal value)\n");
printf("arg8: print tensor value (0: no; 1: yes)\n");
printf("arg9: run kernel # of times (>1)\n");
printf("arg9: time kernel (0=n0, 1=yes)\n");
printf("arg10 to 24: N, K, C, Y, X, Hi, Wi, Sy, Sx, Dy, Dx, LeftPy, LeftPx, RightPy, "
"RightPx\n");
exit(1);
......@@ -56,7 +56,7 @@ int profile_conv_fwd_bias_relu_add(int argc, char* argv[])
const bool do_verification = std::stoi(argv[6]);
const int init_method = std::stoi(argv[7]);
const bool do_log = std::stoi(argv[8]);
const int nrepeat = std::stoi(argv[9]);
const bool time_kernel = std::stoi(argv[9]);
const ck::index_t N = std::stoi(argv[10]);
const ck::index_t K = std::stoi(argv[11]);
......@@ -94,7 +94,7 @@ int profile_conv_fwd_bias_relu_add(int argc, char* argv[])
do_verification,
init_method,
do_log,
nrepeat,
time_kernel,
N,
K,
C,
......@@ -111,5 +111,5 @@ int profile_conv_fwd_bias_relu_add(int argc, char* argv[])
throw std::runtime_error("wrong! data_type & layout for this operator is not implemented");
}
return 1;
return 0;
}
profiler/src/profile_conv_fwd_bias_relu_atomic_add.cpp
View file @ a3b4c5cb
......@@ -43,7 +43,7 @@ int profile_conv_fwd_bias_relu_atomic_add(int argc, char* argv[])
printf("arg6: verification (0: no; 1: yes)\n");
printf("arg7: initialization (0: no init; 1: integer value; 2: decimal value)\n");
printf("arg8: print tensor value (0: no; 1: yes)\n");
printf("arg9: run kernel # of times (>1)\n");
printf("arg9: time kernel (0=n0, 1=yes)\n");
printf("arg10 to 24: N, K, C, Y, X, Hi, Wi, Sy, Sx, Dy, Dx, LeftPy, LeftPx, RightPy, "
"RightPx\n");
exit(1);
......@@ -56,7 +56,7 @@ int profile_conv_fwd_bias_relu_atomic_add(int argc, char* argv[])
const bool do_verification = std::stoi(argv[6]);
const int init_method = std::stoi(argv[7]);
const bool do_log = std::stoi(argv[8]);
const int nrepeat = std::stoi(argv[9]);
const bool time_kernel = std::stoi(argv[9]);
const ck::index_t N = std::stoi(argv[10]);
const ck::index_t K = std::stoi(argv[11]);
......@@ -95,7 +95,7 @@ int profile_conv_fwd_bias_relu_atomic_add(int argc, char* argv[])
do_verification,
init_method,
do_log,
nrepeat,
time_kernel,
N,
K,
C,
......@@ -112,5 +112,5 @@ int profile_conv_fwd_bias_relu_atomic_add(int argc, char* argv[])
throw std::runtime_error("wrong! data_type & layout for this operator is not implemented");
}
return 1;
return 0;
}
profiler/src/profile_convnd_bwd_data.cpp
View file @ a3b4c5cb
......@@ -39,40 +39,40 @@ ck::utils::conv::ConvParams parse_conv_params(int num_dim_spatial, char* argv[],
// (N, K, C) + num_dim_spatial * 6 (filter, input, strides, dilations, pad left, pad right)
ck::utils::conv::ConvParams params;
params.num_dim_spatial = num_dim_spatial;
params.N = std::stoi(argv[arg_idx++]);
params.K = std::stoi(argv[arg_idx++]);
params.C = std::stoi(argv[arg_idx++]);
params.num_dim_spatial_ = num_dim_spatial;
params.N_ = std::stoi(argv[arg_idx++]);
params.K_ = std::stoi(argv[arg_idx++]);
params.C_ = std::stoi(argv[arg_idx++]);
params.filter_spatial_lengths.resize(num_dim_spatial);
params.filter_spatial_lengths_.resize(num_dim_spatial);
for(int i = 0; i < num_dim_spatial; ++i)
{
params.filter_spatial_lengths[i] = std::stoi(argv[arg_idx++]);
params.filter_spatial_lengths_[i] = std::stoi(argv[arg_idx++]);
}
params.input_spatial_lengths.resize(num_dim_spatial);
params.input_spatial_lengths_.resize(num_dim_spatial);
for(int i = 0; i < num_dim_spatial; ++i)
{
params.input_spatial_lengths[i] = std::stoi(argv[arg_idx++]);
params.input_spatial_lengths_[i] = std::stoi(argv[arg_idx++]);
}
params.conv_filter_strides.resize(num_dim_spatial);
params.conv_filter_strides_.resize(num_dim_spatial);
for(int i = 0; i < num_dim_spatial; ++i)
{
params.conv_filter_strides[i] = std::stoi(argv[arg_idx++]);
params.conv_filter_strides_[i] = std::stoi(argv[arg_idx++]);
}
params.conv_filter_dilations.resize(num_dim_spatial);
params.conv_filter_dilations_.resize(num_dim_spatial);
for(int i = 0; i < num_dim_spatial; ++i)
{
params.conv_filter_dilations[i] = std::stoi(argv[arg_idx++]);
params.conv_filter_dilations_[i] = std::stoi(argv[arg_idx++]);
}
params.input_left_pads.resize(num_dim_spatial);
params.input_left_pads_.resize(num_dim_spatial);
for(int i = 0; i < num_dim_spatial; ++i)
{
params.input_left_pads[i] = std::stoi(argv[arg_idx++]);
params.input_left_pads_[i] = std::stoi(argv[arg_idx++]);
}
params.input_right_pads.resize(num_dim_spatial);
params.input_right_pads_.resize(num_dim_spatial);
for(int i = 0; i < num_dim_spatial; ++i)
{
params.input_right_pads[i] = std::stoi(argv[arg_idx++]);
params.input_right_pads_[i] = std::stoi(argv[arg_idx++]);
}
return params;
......@@ -95,7 +95,7 @@ int profile_convnd_bwd_data(int argc, char* argv[], int num_dim_spatial)
printf("arg6: verification (0: no; 1: yes)\n");
printf("arg7: initialization (0: no init; 1: integer value; 2: decimal value)\n");
printf("arg8: print tensor value (0: no; 1: yes)\n");
printf("arg9: run kernel # of times (>1)\n");
printf("arg9: time kernel (0=n0, 1=yes)\n");
printf("arg10 to 24: N, K, C, Y, X, Hi, Wi, Sy, Sx, Dy, Dx, LeftPy, LeftPx, RightPy, "
"RightPx\n");
return 1;
......@@ -108,7 +108,7 @@ int profile_convnd_bwd_data(int argc, char* argv[], int num_dim_spatial)
const bool do_verification = std::stoi(argv[6]);
const int init_method = std::stoi(argv[7]);
const bool do_log = std::stoi(argv[8]);
const int nrepeat = std::stoi(argv[9]);
const bool time_kernel = std::stoi(argv[9]);
ck::utils::conv::ConvParams params = parse_conv_params(num_dim_spatial, argv, preParams);
......@@ -132,17 +132,17 @@ int profile_convnd_bwd_data(int argc, char* argv[], int num_dim_spatial)
do_verification,
init_method,
do_log,
nrepeat,
params.N,
params.K,
params.C,
params.input_spatial_lengths,
params.filter_spatial_lengths,
time_kernel,
params.N_,
params.K_,
params.C_,
params.input_spatial_lengths_,
params.filter_spatial_lengths_,
params.GetOutputSpatialLengths(),
params.conv_filter_strides,
params.conv_filter_dilations,
params.input_left_pads,
params.input_right_pads);
params.conv_filter_strides_,
params.conv_filter_dilations_,
params.input_left_pads_,
params.input_right_pads_);
break;
case 2:
......@@ -157,17 +157,17 @@ int profile_convnd_bwd_data(int argc, char* argv[], int num_dim_spatial)
do_verification,
init_method,
do_log,
nrepeat,
params.N,
params.K,
params.C,
params.input_spatial_lengths,
params.filter_spatial_lengths,
time_kernel,
params.N_,
params.K_,
params.C_,
params.input_spatial_lengths_,
params.filter_spatial_lengths_,
params.GetOutputSpatialLengths(),
params.conv_filter_strides,
params.conv_filter_dilations,
params.input_left_pads,
params.input_right_pads);
params.conv_filter_strides_,
params.conv_filter_dilations_,
params.input_left_pads_,
params.input_right_pads_);
break;
case 3:
......@@ -182,17 +182,17 @@ int profile_convnd_bwd_data(int argc, char* argv[], int num_dim_spatial)
do_verification,
init_method,
do_log,
nrepeat,
params.N,
params.K,
params.C,
params.input_spatial_lengths,
params.filter_spatial_lengths,
time_kernel,
params.N_,
params.K_,
params.C_,
params.input_spatial_lengths_,
params.filter_spatial_lengths_,
params.GetOutputSpatialLengths(),
params.conv_filter_strides,
params.conv_filter_dilations,
params.input_left_pads,
params.input_right_pads);
params.conv_filter_strides_,
params.conv_filter_dilations_,
params.input_left_pads_,
params.input_right_pads_);
break;
default: break;
......
profiler/src/profile_convnd_fwd.cpp
View file @ a3b4c5cb
......@@ -5,7 +5,7 @@
#include <vector>
#include <half.hpp>
#include "conv_fwd_util.hpp"
#include "conv_util.hpp"
#include "element_wise_operation.hpp"
#include "fill.hpp"
#include "profile_convnd_fwd.hpp"
......@@ -119,7 +119,7 @@ template <int NDim,
void profile_convnd_instances_impl(const ck::utils::conv::ConvParams& params,
bool do_verification,
bool do_log,
int nrepeat,
bool time_kernel,
int init_method,
ConvLayouts)
{
......@@ -185,7 +185,7 @@ void profile_convnd_instances_impl(const ck::utils::conv::ConvParams& params,
reference_conv_fwd_fun);
auto best_conf = run_engine.Profile(
conv::ConvolutionFwdInstances<InDataType, WeiDataType, OutDataType>::template Get<NDim>(),
nrepeat,
time_kernel,
do_verification,
do_log);
......@@ -201,7 +201,7 @@ void profile_convnd_instances(ConvDataType data_type,
const ck::utils::conv::ConvParams& params,
bool do_verification,
bool do_log,
int nrepeat,
bool time_kernel,
int init_method)
{
switch(data_layout)
......@@ -214,7 +214,7 @@ void profile_convnd_instances(ConvDataType data_type,
params,
do_verification,
do_log,
nrepeat,
time_kernel,
init_method,
ConvolutionLayouts<NDim, ConvDataLayout::NHWC>{});
break;
......@@ -223,7 +223,7 @@ void profile_convnd_instances(ConvDataType data_type,
params,
do_verification,
do_log,
nrepeat,
time_kernel,
init_method,
ConvolutionLayouts<NDim, ConvDataLayout::NHWC>{});
break;
......@@ -232,7 +232,7 @@ void profile_convnd_instances(ConvDataType data_type,
params,
do_verification,
do_log,
nrepeat,
time_kernel,
init_method,
ConvolutionLayouts<NDim, ConvDataLayout::NHWC>{});
break;
......@@ -241,7 +241,7 @@ void profile_convnd_instances(ConvDataType data_type,
params,
do_verification,
do_log,
nrepeat,
time_kernel,
init_method,
ConvolutionLayouts<NDim, ConvDataLayout::NHWC>{});
break;
......@@ -256,7 +256,7 @@ void profile_convnd_instances(ConvDataType data_type,
params,
do_verification,
do_log,
nrepeat,
time_kernel,
init_method,
ConvolutionLayouts<NDim, ConvDataLayout::NCHW>{});
break;
......@@ -265,7 +265,7 @@ void profile_convnd_instances(ConvDataType data_type,
params,
do_verification,
do_log,
nrepeat,
time_kernel,
init_method,
ConvolutionLayouts<NDim, ConvDataLayout::NCHW>{});
break;
......@@ -274,7 +274,7 @@ void profile_convnd_instances(ConvDataType data_type,
params,
do_verification,
do_log,
nrepeat,
time_kernel,
init_method,
ConvolutionLayouts<NDim, ConvDataLayout::NCHW>{});
break;
......@@ -283,7 +283,7 @@ void profile_convnd_instances(ConvDataType data_type,
params,
do_verification,
do_log,
nrepeat,
time_kernel,
init_method,
ConvolutionLayouts<NDim, ConvDataLayout::NCHW>{});
break;
......@@ -304,7 +304,7 @@ int ck::profiler::profile_convnd_fwd(int argc, char* argv[])
bool do_verification{true};
int init_method{2};
bool do_log{false};
int nrepeat{100};
bool time_kernel{false};
int num_dim_spatial{2};
ConvParams params;
......@@ -318,7 +318,7 @@ int ck::profiler::profile_convnd_fwd(int argc, char* argv[])
do_verification = std::stoi(argv[4]);
init_method = std::stoi(argv[5]);
do_log = std::stoi(argv[6]);
nrepeat = std::stoi(argv[7]);
time_kernel = std::stoi(argv[7]);
num_dim_spatial = std::stoi(argv[8]);
}
if(argc >= 10)
......@@ -332,20 +332,20 @@ int ck::profiler::profile_convnd_fwd(int argc, char* argv[])
{
case 1:
profile_convnd_instances<1>(
data_type, data_layout, params, do_verification, do_log, nrepeat, init_method);
data_type, data_layout, params, do_verification, do_log, time_kernel, init_method);
break;
case 2:
profile_convnd_instances<2>(
data_type, data_layout, params, do_verification, do_log, nrepeat, init_method);
data_type, data_layout, params, do_verification, do_log, time_kernel, init_method);
break;
case 3:
profile_convnd_instances<3>(
data_type, data_layout, params, do_verification, do_log, nrepeat, init_method);
data_type, data_layout, params, do_verification, do_log, time_kernel, init_method);
break;
default:
throw std::runtime_error("profile_conv_fwd: unsupported num_dim_spatial value: " +
std::to_string(num_dim_spatial));
}
return 1;
return 0;
}
profiler/src/profile_gemm.cpp
View file @ a3b4c5cb
......@@ -38,8 +38,8 @@ int profile_gemm(int argc, char* argv[])
printf(" 3: A[k, m] * B[n, k] = C[m, n])\n");
printf("arg4: verification (0: no; 1: yes)\n");
printf("arg5: initialization (0: no init; 1: integer value; 2: decimal value)\n");
printf("arg8: print tensor value (0: no; 1: yes)\n");
printf("arg7: run kernel # of times (>1)\n");
printf("arg6: print tensor value (0: no; 1: yes)\n");
printf("arg7: time kernel (0=n0, 1=yes)\n");
printf("arg8 to 13: M, N, K, StrideA, StrideB, StrideC\n");
printf("arg14: split k into mulitiple batch\n");
exit(1);
......@@ -50,7 +50,7 @@ int profile_gemm(int argc, char* argv[])
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 int nrepeat = std::stoi(argv[7]);
const bool time_kernel = std::stoi(argv[7]);
const int M = std::stoi(argv[8]);
const int N = std::stoi(argv[9]);
......@@ -68,13 +68,14 @@ int profile_gemm(int argc, char* argv[])
ck::profiler::profile_gemm_impl<ck::half_t,
ck::half_t,
ck::half_t,
float,
ck::tensor_layout::gemm::RowMajor,
ck::tensor_layout::gemm::RowMajor,
ck::tensor_layout::gemm::RowMajor>(
do_verification,
init_method,
do_log,
nrepeat,
time_kernel,
M,
N,
K,
......@@ -88,13 +89,14 @@ int profile_gemm(int argc, char* argv[])
ck::profiler::profile_gemm_impl<ck::half_t,
ck::half_t,
ck::half_t,
float,
ck::tensor_layout::gemm::RowMajor,
ck::tensor_layout::gemm::ColumnMajor,
ck::tensor_layout::gemm::RowMajor>(
do_verification,
init_method,
do_log,
nrepeat,
time_kernel,
M,
N,
K,
......@@ -108,13 +110,14 @@ int profile_gemm(int argc, char* argv[])
ck::profiler::profile_gemm_impl<ck::half_t,
ck::half_t,
ck::half_t,
float,
ck::tensor_layout::gemm::ColumnMajor,
ck::tensor_layout::gemm::RowMajor,
ck::tensor_layout::gemm::RowMajor>(
do_verification,
init_method,
do_log,
nrepeat,
time_kernel,
M,
N,
K,
......@@ -128,13 +131,14 @@ int profile_gemm(int argc, char* argv[])
ck::profiler::profile_gemm_impl<ck::half_t,
ck::half_t,
ck::half_t,
float,
ck::tensor_layout::gemm::ColumnMajor,
ck::tensor_layout::gemm::ColumnMajor,
ck::tensor_layout::gemm::RowMajor>(
do_verification,
init_method,
do_log,
nrepeat,
time_kernel,
M,
N,
K,
......@@ -146,6 +150,7 @@ int profile_gemm(int argc, char* argv[])
else if(data_type == GemmDataType::F32_F32_F32 && layout == GemmMatrixLayout::MK_KN_MN)
{
ck::profiler::profile_gemm_impl<float,
float,
float,
float,
ck::tensor_layout::gemm::RowMajor,
......@@ -154,7 +159,7 @@ int profile_gemm(int argc, char* argv[])
do_verification,
init_method,
do_log,
nrepeat,
time_kernel,
M,
N,
K,
......@@ -166,6 +171,7 @@ int profile_gemm(int argc, char* argv[])
else if(data_type == GemmDataType::F32_F32_F32 && layout == GemmMatrixLayout::MK_NK_MN)
{
ck::profiler::profile_gemm_impl<float,
float,
float,
float,
ck::tensor_layout::gemm::RowMajor,
......@@ -174,7 +180,7 @@ int profile_gemm(int argc, char* argv[])
do_verification,
init_method,
do_log,
nrepeat,
time_kernel,
M,
N,
K,
......@@ -186,6 +192,7 @@ int profile_gemm(int argc, char* argv[])
else if(data_type == GemmDataType::F32_F32_F32 && layout == GemmMatrixLayout::KM_KN_MN)
{
ck::profiler::profile_gemm_impl<float,
float,
float,
float,
ck::tensor_layout::gemm::ColumnMajor,
......@@ -194,7 +201,7 @@ int profile_gemm(int argc, char* argv[])
do_verification,
init_method,
do_log,
nrepeat,
time_kernel,
M,
N,
K,
......@@ -206,6 +213,7 @@ int profile_gemm(int argc, char* argv[])
else if(data_type == GemmDataType::F32_F32_F32 && layout == GemmMatrixLayout::KM_NK_MN)
{
ck::profiler::profile_gemm_impl<float,
float,
float,
float,
ck::tensor_layout::gemm::ColumnMajor,
......@@ -214,7 +222,7 @@ int profile_gemm(int argc, char* argv[])
do_verification,
init_method,
do_log,
nrepeat,
time_kernel,
M,
N,
K,
......@@ -228,13 +236,14 @@ int profile_gemm(int argc, char* argv[])
ck::profiler::profile_gemm_impl<int8_t,
int8_t,
int8_t,
int32_t,
ck::tensor_layout::gemm::RowMajor,
ck::tensor_layout::gemm::RowMajor,
ck::tensor_layout::gemm::RowMajor>(
do_verification,
init_method,
do_log,
nrepeat,
time_kernel,
M,
N,
K,
......@@ -248,13 +257,14 @@ int profile_gemm(int argc, char* argv[])
ck::profiler::profile_gemm_impl<int8_t,
int8_t,
int8_t,
int32_t,
ck::tensor_layout::gemm::RowMajor,
ck::tensor_layout::gemm::ColumnMajor,
ck::tensor_layout::gemm::RowMajor>(
do_verification,
init_method,
do_log,
nrepeat,
time_kernel,
M,
N,
K,
......@@ -268,13 +278,14 @@ int profile_gemm(int argc, char* argv[])
ck::profiler::profile_gemm_impl<int8_t,
int8_t,
int8_t,
int32_t,
ck::tensor_layout::gemm::ColumnMajor,
ck::tensor_layout::gemm::RowMajor,
ck::tensor_layout::gemm::RowMajor>(
do_verification,
init_method,
do_log,
nrepeat,
time_kernel,
M,
N,
K,
......@@ -288,13 +299,14 @@ int profile_gemm(int argc, char* argv[])
ck::profiler::profile_gemm_impl<int8_t,
int8_t,
int8_t,
int32_t,
ck::tensor_layout::gemm::ColumnMajor,
ck::tensor_layout::gemm::ColumnMajor,
ck::tensor_layout::gemm::RowMajor>(
do_verification,
init_method,
do_log,
nrepeat,
time_kernel,
M,
N,
K,
......@@ -308,13 +320,14 @@ int profile_gemm(int argc, char* argv[])
ck::profiler::profile_gemm_impl<ck::bhalf_t,
ck::bhalf_t,
ck::bhalf_t,
float,
ck::tensor_layout::gemm::RowMajor,
ck::tensor_layout::gemm::RowMajor,
ck::tensor_layout::gemm::RowMajor>(
do_verification,
init_method,
do_log,
nrepeat,
time_kernel,
M,
N,
K,
......@@ -328,13 +341,14 @@ int profile_gemm(int argc, char* argv[])
ck::profiler::profile_gemm_impl<ck::bhalf_t,
ck::bhalf_t,
ck::bhalf_t,
float,
ck::tensor_layout::gemm::RowMajor,
ck::tensor_layout::gemm::ColumnMajor,
ck::tensor_layout::gemm::RowMajor>(
do_verification,
init_method,
do_log,
nrepeat,
time_kernel,
M,
N,
K,
......@@ -348,13 +362,14 @@ int profile_gemm(int argc, char* argv[])
ck::profiler::profile_gemm_impl<ck::bhalf_t,
ck::bhalf_t,
ck::bhalf_t,
float,
ck::tensor_layout::gemm::ColumnMajor,
ck::tensor_layout::gemm::RowMajor,
ck::tensor_layout::gemm::RowMajor>(
do_verification,
init_method,
do_log,
nrepeat,
time_kernel,
M,
N,
K,
......@@ -368,13 +383,14 @@ int profile_gemm(int argc, char* argv[])
ck::profiler::profile_gemm_impl<ck::bhalf_t,
ck::bhalf_t,
ck::bhalf_t,
float,
ck::tensor_layout::gemm::ColumnMajor,
ck::tensor_layout::gemm::ColumnMajor,
ck::tensor_layout::gemm::RowMajor>(
do_verification,
init_method,
do_log,
nrepeat,
time_kernel,
M,
N,
K,
......@@ -388,5 +404,5 @@ int profile_gemm(int argc, char* argv[])
throw std::runtime_error("wrong! this GEMM data_type & layout is not implemented");
}
return 1;
return 0;
}
profiler/src/profile_gemm_bias_2d.cpp
View file @ a3b4c5cb
......@@ -36,8 +36,8 @@ int profile_gemm_bias_2d(int argc, char* argv[])
printf(" 3: A[k, m] * B[n, k] = C[m, n])\n");
printf("arg4: verification (0: no; 1: yes)\n");
printf("arg5: initialization (0: no init; 1: integer value; 2: decimal value)\n");
printf("arg8: print tensor value (0: no; 1: yes)\n");
printf("arg7: run kernel # of times (>1)\n");
printf("arg6: print tensor value (0: no; 1: yes)\n");
printf("arg7: time kernel (0=n0, 1=yes)\n");
printf("arg8 to 13: M, N, K, StrideA, StrideB, StrideC\n");
printf("arg14: alpha\n");
printf("arg15: beta\n");
......@@ -50,7 +50,7 @@ int profile_gemm_bias_2d(int argc, char* argv[])
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 int nrepeat = std::stoi(argv[7]);
const bool time_kernel = std::stoi(argv[7]);
const int M = std::stoi(argv[8]);
const int N = std::stoi(argv[9]);
......@@ -76,7 +76,7 @@ int profile_gemm_bias_2d(int argc, char* argv[])
do_verification,
init_method,
do_log,
nrepeat,
time_kernel,
M,
N,
K,
......@@ -99,7 +99,7 @@ int profile_gemm_bias_2d(int argc, char* argv[])
do_verification,
init_method,
do_log,
nrepeat,
time_kernel,
M,
N,
K,
......@@ -122,7 +122,7 @@ int profile_gemm_bias_2d(int argc, char* argv[])
do_verification,
init_method,
do_log,
nrepeat,
time_kernel,
M,
N,
K,
......@@ -145,7 +145,7 @@ int profile_gemm_bias_2d(int argc, char* argv[])
do_verification,
init_method,
do_log,
nrepeat,
time_kernel,
M,
N,
K,
......@@ -168,7 +168,7 @@ int profile_gemm_bias_2d(int argc, char* argv[])
do_verification,
init_method,
do_log,
nrepeat,
time_kernel,
M,
N,
K,
......@@ -191,7 +191,7 @@ int profile_gemm_bias_2d(int argc, char* argv[])
do_verification,
init_method,
do_log,
nrepeat,
time_kernel,
M,
N,
K,
......@@ -214,7 +214,7 @@ int profile_gemm_bias_2d(int argc, char* argv[])
do_verification,
init_method,
do_log,
nrepeat,
time_kernel,
M,
N,
K,
......@@ -237,7 +237,7 @@ int profile_gemm_bias_2d(int argc, char* argv[])
do_verification,
init_method,
do_log,
nrepeat,
time_kernel,
M,
N,
K,
......@@ -252,5 +252,5 @@ int profile_gemm_bias_2d(int argc, char* argv[])
throw std::runtime_error("wrong! this data_type & layout is not implemented");
}
return 1;
return 0;
}
profiler/src/profile_gemm_bias_relu.cpp
View file @ a3b4c5cb
......@@ -36,8 +36,8 @@ int profile_gemm_bias_relu(int argc, char* argv[])
printf(" 3: A[k, m] * B[n, k] = C[m, n])\n");
printf("arg4: verification (0: no; 1: yes)\n");
printf("arg5: initialization (0: no init; 1: integer value; 2: decimal value)\n");
printf("arg8: print tensor value (0: no; 1: yes)\n");
printf("arg7: run kernel # of times (>1)\n");
printf("arg6: print tensor value (0: no; 1: yes)\n");
printf("arg7: time kernel (0=n0, 1=yes)\n");
printf("arg8 to 13: M, N, K, StrideA, StrideB, StrideC\n");
printf("arg14: split k into mulitiple batch\n");
exit(1);
......@@ -48,7 +48,7 @@ int profile_gemm_bias_relu(int argc, char* argv[])
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 int nrepeat = std::stoi(argv[7]);
const bool time_kernel = std::stoi(argv[7]);
const int M = std::stoi(argv[8]);
const int N = std::stoi(argv[9]);
......@@ -69,7 +69,7 @@ int profile_gemm_bias_relu(int argc, char* argv[])
do_verification,
init_method,
do_log,
nrepeat,
time_kernel,
M,
N,
K,
......@@ -88,7 +88,7 @@ int profile_gemm_bias_relu(int argc, char* argv[])
do_verification,
init_method,
do_log,
nrepeat,
time_kernel,
M,
N,
K,
......@@ -107,7 +107,7 @@ int profile_gemm_bias_relu(int argc, char* argv[])
do_verification,
init_method,
do_log,
nrepeat,
time_kernel,
M,
N,
K,
......@@ -126,7 +126,7 @@ int profile_gemm_bias_relu(int argc, char* argv[])
do_verification,
init_method,
do_log,
nrepeat,
time_kernel,
M,
N,
K,
......@@ -139,5 +139,5 @@ int profile_gemm_bias_relu(int argc, char* argv[])
throw std::runtime_error("wrong! this data_type & layout is not implemented");
}
return 1;
return 0;
}
profiler/src/profile_gemm_bias_relu_add.cpp
View file @ a3b4c5cb
......@@ -36,8 +36,8 @@ int profile_gemm_bias_relu_add(int argc, char* argv[])
printf(" 3: A[k, m] * B[n, k] = C[m, n])\n");
printf("arg4: verification (0: no; 1: yes)\n");
printf("arg5: initialization (0: no init; 1: integer value; 2: decimal value)\n");
printf("arg8: print tensor value (0: no; 1: yes)\n");
printf("arg7: run kernel # of times (>1)\n");
printf("arg6: print tensor value (0: no; 1: yes)\n");
printf("arg7: time kernel (0=n0, 1=yes)\n");
printf("arg8 to 14: M, N, K, StrideA, StrideB, StrideC, StrideC1\n");
printf("arg15: split k into mulitiple batch\n");
exit(1);
......@@ -48,7 +48,7 @@ int profile_gemm_bias_relu_add(int argc, char* argv[])
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 int nrepeat = std::stoi(argv[7]);
const bool time_kernel = std::stoi(argv[7]);
const int M = std::stoi(argv[8]);
const int N = std::stoi(argv[9]);
......@@ -70,7 +70,7 @@ int profile_gemm_bias_relu_add(int argc, char* argv[])
do_verification,
init_method,
do_log,
nrepeat,
time_kernel,
M,
N,
K,
......@@ -90,7 +90,7 @@ int profile_gemm_bias_relu_add(int argc, char* argv[])
do_verification,
init_method,
do_log,
nrepeat,
time_kernel,
M,
N,
K,
......@@ -110,7 +110,7 @@ int profile_gemm_bias_relu_add(int argc, char* argv[])
do_verification,
init_method,
do_log,
nrepeat,
time_kernel,
M,
N,
K,
......@@ -130,7 +130,7 @@ int profile_gemm_bias_relu_add(int argc, char* argv[])
do_verification,
init_method,
do_log,
nrepeat,
time_kernel,
M,
N,
K,
......@@ -144,5 +144,5 @@ int profile_gemm_bias_relu_add(int argc, char* argv[])
throw std::runtime_error("wrong! this data_type & layout is not implemented");
}
return 1;
return 0;
}
profiler/src/profile_gemm_reduce.cpp
View file @ a3b4c5cb
......@@ -32,8 +32,8 @@ int profile_gemm_reduce(int argc, char* argv[])
printf(" 3: A[k, m] * B[n, k] = C[m, n])\n");
printf("arg4: verification (0: no; 1: yes)\n");
printf("arg5: initialization (0: no init; 1: integer value; 2: decimal value)\n");
printf("arg8: print tensor value (0: no; 1: yes)\n");
printf("arg7: run kernel # of times (>1)\n");
printf("arg6: print tensor value (0: no; 1: yes)\n");
printf("arg7: time kernel (0=n0, 1=yes)\n");
printf("arg8 to 13: M, N, K, StrideA, StrideB, StrideC\n");
printf("arg14: split k into mulitiple batch\n");
exit(1);
......@@ -44,7 +44,7 @@ int profile_gemm_reduce(int argc, char* argv[])
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 int nrepeat = std::stoi(argv[7]);
const bool time_kernel = std::stoi(argv[7]);
const int M = std::stoi(argv[8]);
const int N = std::stoi(argv[9]);
......@@ -66,7 +66,7 @@ int profile_gemm_reduce(int argc, char* argv[])
do_verification,
init_method,
do_log,
nrepeat,
time_kernel,
M,
N,
K,
......@@ -87,7 +87,7 @@ int profile_gemm_reduce(int argc, char* argv[])
do_verification,
init_method,
do_log,
nrepeat,
time_kernel,
M,
N,
K,
......@@ -108,7 +108,7 @@ int profile_gemm_reduce(int argc, char* argv[])
do_verification,
init_method,
do_log,
nrepeat,
time_kernel,
M,
N,
K,
......@@ -129,7 +129,7 @@ int profile_gemm_reduce(int argc, char* argv[])
do_verification,
init_method,
do_log,
nrepeat,
time_kernel,
M,
N,
K,
......@@ -142,5 +142,5 @@ int profile_gemm_reduce(int argc, char* argv[])
throw std::runtime_error("wrong! this data_type & layout is not implemented");
}
return 1;
return 0;
}
profiler/src/profile_grouped_gemm.cpp
View file @ a3b4c5cb
......@@ -54,8 +54,8 @@ int profile_grouped_gemm(int argc, char* argv[])
printf(" 3: A[k, m] * B[n, k] = C[m, n])\n");
printf("arg4: verification (0: no; 1: yes)\n");
printf("arg5: initialization (0: no init; 1: integer value; 2: decimal value)\n");
printf("arg8: print tensor value (0: no; 1: yes)\n");
printf("arg7: run kernel # of times (>1)\n");
printf("arg6: print tensor value (0: no; 1: yes)\n");
printf("arg7: time kernel (0=n0, 1=yes)\n");
printf("arg8 to 13: Ms, Ns, Ks, StrideAs, StrideBs, StrideCs (e.g., 256,256 128,128 64,64 "
"64,64 64,64 128,128)\n");
exit(1);
......@@ -66,7 +66,7 @@ int profile_grouped_gemm(int argc, char* argv[])
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 int nrepeat = std::stoi(argv[7]);
const bool time_kernel = std::stoi(argv[7]);
const auto Ms = argToIntArray(argv[8]);
const auto Ns = argToIntArray(argv[9]);
......@@ -79,6 +79,7 @@ int profile_grouped_gemm(int argc, char* argv[])
if(data_type == GemmDataType::F16_F16_F16 && layout == GemmMatrixLayout::MK_KN_MN)
{
ck::profiler::profile_grouped_gemm_impl<ck::half_t,
ck::half_t,
ck::half_t,
ck::half_t,
ck::tensor_layout::gemm::RowMajor,
......@@ -86,7 +87,7 @@ int profile_grouped_gemm(int argc, char* argv[])
ck::tensor_layout::gemm::RowMajor>(do_verification,
init_method,
do_log,
nrepeat,
time_kernel,
Ms,
Ns,
Ks,
......@@ -97,6 +98,7 @@ int profile_grouped_gemm(int argc, char* argv[])
else if(data_type == GemmDataType::F16_F16_F16 && layout == GemmMatrixLayout::MK_NK_MN)
{
ck::profiler::profile_grouped_gemm_impl<ck::half_t,
ck::half_t,
ck::half_t,
ck::half_t,
ck::tensor_layout::gemm::RowMajor,
......@@ -104,7 +106,7 @@ int profile_grouped_gemm(int argc, char* argv[])
ck::tensor_layout::gemm::RowMajor>(do_verification,
init_method,
do_log,
nrepeat,
time_kernel,
Ms,
Ns,
Ks,
......@@ -115,6 +117,7 @@ int profile_grouped_gemm(int argc, char* argv[])
else if(data_type == GemmDataType::F16_F16_F16 && layout == GemmMatrixLayout::KM_KN_MN)
{
ck::profiler::profile_grouped_gemm_impl<ck::half_t,
ck::half_t,
ck::half_t,
ck::half_t,
ck::tensor_layout::gemm::ColumnMajor,
......@@ -122,7 +125,7 @@ int profile_grouped_gemm(int argc, char* argv[])
ck::tensor_layout::gemm::RowMajor>(do_verification,
init_method,
do_log,
nrepeat,
time_kernel,
Ms,
Ns,
Ks,
......@@ -133,6 +136,7 @@ int profile_grouped_gemm(int argc, char* argv[])
else if(data_type == GemmDataType::F16_F16_F16 && layout == GemmMatrixLayout::KM_NK_MN)
{
ck::profiler::profile_grouped_gemm_impl<ck::half_t,
ck::half_t,
ck::half_t,
ck::half_t,
ck::tensor_layout::gemm::ColumnMajor,
......@@ -140,7 +144,7 @@ int profile_grouped_gemm(int argc, char* argv[])
ck::tensor_layout::gemm::RowMajor>(do_verification,
init_method,
do_log,
nrepeat,
time_kernel,
Ms,
Ns,
Ks,
......@@ -153,5 +157,5 @@ int profile_grouped_gemm(int argc, char* argv[])
throw std::runtime_error("wrong! this GEMM data_type & layout is not implemented");
}
return 1;
return 0;
}
profiler/src/profile_reduce.cpp
View file @ a3b4c5cb
#include <iostream>
#include <fstream>
#include <numeric>
#include <initializer_list>
#include <cstdlib>
#include <vector>
#include <stdexcept>
#include <sstream>
#include <getopt.h>
#include "config.hpp"
#include "print.hpp"
#include "device.hpp"
#include "host_tensor.hpp"
#include "host_tensor_generator.hpp"
#include "device_tensor.hpp"
#include "data_type_enum.hpp"
#include "reduction_enums.hpp"
#include "host_common_util.hpp"
#include "profile_reduce_impl.hpp"
using namespace std;
using ck::NanPropagation;
using ck::ReduceTensorIndices;
using ck::ReduceTensorOp;
static struct option long_options[] = {{"inLengths", required_argument, nullptr, 'D'},
......@@ -38,63 +30,9 @@ static struct option long_options[] = {{"inLengths", required_argument, nullptr,
{"bf16", no_argument, nullptr, '?'},
{"dumpout", required_argument, nullptr, 'o'},
{"verify", required_argument, nullptr, 'v'},
{"log", required_argument, nullptr, 'l'},
{"help", no_argument, nullptr, '?'},
{nullptr, 0, nullptr, 0}};
template <typename T>
static T getSingleValueFromString(const string& valueStr)
{
std::istringstream iss(valueStr);
T val;
iss >> val;
return (val);
};
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);
}
enum struct AppDataType
{
appHalf = 0,
appFloat = 1,
appInt32 = 2,
appInt8 = 3,
appInt8x4 = 4,
appBFloat16 = 5,
appDouble = 6,
};
static void check_reduce_dims(const int rank, const std::vector<int>& reduceDims)
{
for(auto dim : reduceDims)
......@@ -113,7 +51,7 @@ static void check_reduce_dims(const int rank, const std::vector<int>& reduceDims
};
};
class AppArgs
class ReduceProfilerArgs
{
private:
int option_index = 0;
......@@ -130,26 +68,23 @@ class AppArgs
std::vector<float> scales;
ReduceTensorOp reduceOp = ReduceTensorOp::ADD;
AppDataType compTypeId = AppDataType::appFloat;
AppDataType outTypeId = AppDataType::appFloat;
ReduceTensorOp reduceOp = ReduceTensorOp::ADD;
ck::DataTypeEnum compTypeId = ck::DataTypeEnum::Float;
ck::DataTypeEnum outTypeId = ck::DataTypeEnum::Float;
bool compType_assigned = false;
bool outType_assigned = false;
NanPropagation nanOpt = NanPropagation::NOT_PROPAGATE_NAN;
ReduceTensorIndices indicesOpt = ReduceTensorIndices::NO_INDICES;
bool do_log = false;
bool do_verification = false;
bool do_dumpout = false;
int nanOpt = 0;
int indicesOpt = 0;
bool do_verification = false;
bool do_dumpout = false;
int init_method;
int nrepeat;
bool time_kernel;
bool need_indices = false;
AppArgs() = default;
~AppArgs() = default;
ReduceProfilerArgs() = default;
~ReduceProfilerArgs() = default;
void show_usage(const char* cmd)
{
......@@ -166,8 +101,11 @@ class AppArgs
std::cout << "--outType or -W, optional enum value indicating the type of the reduced "
"output, which could be float when the input data is half"
<< std::endl;
std::cout << "--nanOpt or -N, enum value indicates the selection for NanOpt" << std::endl;
std::cout << "--indicesOpt or -I, enum value indicates the selection for IndicesOpt"
std::cout
<< "--nanOpt or -N, 1/0 value indicates the selection to use or not use Nan-Propagation"
<< std::endl;
std::cout << "--indicesOpt or -I, 1/0 value indicates the selection to use or not use "
"index in reduction"
<< std::endl;
std::cout << "--scales or -S, comma separated two float values for alpha and beta"
<< std::endl;
......@@ -181,18 +119,19 @@ class AppArgs
std::cout << "--dumpout or -o, 1/0 to indicate where to save the reduction result to files "
"for further analysis"
<< std::endl;
std::cout << "--log or -l, 1/0 to indicate whether to log some information" << std::endl;
};
int processArgs(int argc, char* argv[])
{
unsigned int ch;
using ck::host_common::getTypeValuesFromString;
int ch;
optind++; // to skip the "reduce" module name
while(1)
{
ch = getopt_long(argc, argv, "D:R:O:C:W:N:I:S:v:o:l:", long_options, &option_index);
ch = getopt_long(argc, argv, "D:R:O:C:W:N:I:S:v:o:", long_options, &option_index);
if(ch == -1)
break;
switch(ch)
......@@ -219,27 +158,27 @@ class AppArgs
if(!optarg)
throw std::runtime_error("Invalid option format!");
compTypeId = static_cast<AppDataType>(std::atoi(optarg));
compTypeId = static_cast<ck::DataTypeEnum>(std::atoi(optarg));
compType_assigned = true;
break;
case 'W':
if(!optarg)
throw std::runtime_error("Invalid option format!");
outTypeId = static_cast<AppDataType>(std::atoi(optarg));
outTypeId = static_cast<ck::DataTypeEnum>(std::atoi(optarg));
outType_assigned = true;
break;
case 'N':
if(!optarg)
throw std::runtime_error("Invalid option format!");
nanOpt = static_cast<NanPropagation>(std::atoi(optarg));
nanOpt = std::atoi(optarg);
break;
case 'I':
if(!optarg)
throw std::runtime_error("Invalid option format!");
indicesOpt = static_cast<ReduceTensorIndices>(std::atoi(optarg));
indicesOpt = std::atoi(optarg);
break;
case 'S':
if(!optarg)
......@@ -262,12 +201,6 @@ class AppArgs
do_dumpout = static_cast<bool>(std::atoi(optarg));
break;
case 'l':
if(!optarg)
throw std::runtime_error("Invalid option format!");
do_log = static_cast<bool>(std::atoi(optarg));
break;
case '?':
if(std::string(long_options[option_index].name) == "half")
use_half = true;
......@@ -295,7 +228,7 @@ class AppArgs
throw std::runtime_error("Invalid cmd-line arguments, more argumetns are needed!");
init_method = std::atoi(argv[optind++]);
nrepeat = std::atoi(argv[optind]);
time_kernel = static_cast<bool>(std::atoi(argv[optind]));
if(scales.empty())
{
......@@ -306,9 +239,6 @@ class AppArgs
if(reduceOp == ReduceTensorOp::MIN || reduceOp == ReduceTensorOp::MAX ||
reduceOp == ReduceTensorOp::AMAX)
{
if(indicesOpt != ReduceTensorIndices::NO_INDICES)
need_indices = true;
// for indexable operations, no need to assign compType and outType, just let them be
// same as inType
compType_assigned = false;
......@@ -322,9 +252,10 @@ class AppArgs
int profile_reduce(int argc, char* argv[])
{
using namespace ck::profiler;
using ck::DataTypeEnum;
using ck::profiler::profile_reduce_impl;
AppArgs args;
ReduceProfilerArgs args;
if(args.processArgs(argc, argv) < 0)
return (-1);
......@@ -339,42 +270,41 @@ int profile_reduce(int argc, char* argv[])
if(args.use_half)
{
if(!args.compType_assigned)
args.compTypeId = AppDataType::appHalf;
args.compTypeId = DataTypeEnum::Half;
if(args.outType_assigned &&
(args.outTypeId != AppDataType::appHalf && args.outTypeId != AppDataType::appFloat))
args.outTypeId = AppDataType::appFloat;
(args.outTypeId != DataTypeEnum::Half && args.outTypeId != DataTypeEnum::Float))
args.outTypeId = DataTypeEnum::Float;
if(!args.outType_assigned)
args.outTypeId = AppDataType::appHalf;
args.outTypeId = DataTypeEnum::Half;
if(args.compTypeId == AppDataType::appHalf)
if(args.compTypeId == DataTypeEnum::Half)
{
profile_reduce_impl<ck::half_t, ck::half_t, ck::half_t>(args.do_verification,
args.init_method,
args.do_log,
args.do_dumpout,
args.nrepeat,
args.inLengths,
args.reduceDims,
args.reduceOp,
args.nanOpt,
args.indicesOpt,
args.scales[0],
args.scales[1]);
profile_reduce_impl<ck::half_t, ck::half_t, ck::half_t>(
args.do_verification,
args.init_method,
args.do_dumpout,
args.time_kernel,
args.inLengths,
args.reduceDims,
args.reduceOp,
static_cast<bool>(args.nanOpt),
static_cast<bool>(args.indicesOpt),
args.scales[0],
args.scales[1]);
}
else if(args.compTypeId == AppDataType::appFloat)
else if(args.compTypeId == DataTypeEnum::Float)
{
profile_reduce_impl<ck::half_t, float, ck::half_t>(args.do_verification,
args.init_method,
args.do_log,
args.do_dumpout,
args.nrepeat,
args.time_kernel,
args.inLengths,
args.reduceDims,
args.reduceOp,
args.nanOpt,
args.indicesOpt,
static_cast<bool>(args.nanOpt),
static_cast<bool>(args.indicesOpt),
args.scales[0],
args.scales[1]);
}
......@@ -385,56 +315,53 @@ int profile_reduce(int argc, char* argv[])
{
profile_reduce_impl<double, double, double>(args.do_verification,
args.init_method,
args.do_log,
args.do_dumpout,
args.nrepeat,
args.time_kernel,
args.inLengths,
args.reduceDims,
args.reduceOp,
args.nanOpt,
args.indicesOpt,
static_cast<bool>(args.nanOpt),
static_cast<bool>(args.indicesOpt),
args.scales[0],
args.scales[1]);
}
else if(args.use_int8)
{
if(!args.compType_assigned)
args.compTypeId = AppDataType::appInt8;
args.compTypeId = DataTypeEnum::Int8;
if(args.outType_assigned &&
(args.outTypeId != AppDataType::appInt8 && args.outTypeId != AppDataType::appInt32))
args.outTypeId = AppDataType::appInt32;
(args.outTypeId != DataTypeEnum::Int8 && args.outTypeId != DataTypeEnum::Int32))
args.outTypeId = DataTypeEnum::Int32;
if(!args.outType_assigned)
args.outTypeId = AppDataType::appInt8;
args.outTypeId = DataTypeEnum::Int8;
if(args.compTypeId == AppDataType::appInt8)
if(args.compTypeId == DataTypeEnum::Int8)
{
profile_reduce_impl<int8_t, int8_t, int8_t>(args.do_verification,
args.init_method,
args.do_log,
args.do_dumpout,
args.nrepeat,
args.time_kernel,
args.inLengths,
args.reduceDims,
args.reduceOp,
args.nanOpt,
args.indicesOpt,
static_cast<bool>(args.nanOpt),
static_cast<bool>(args.indicesOpt),
args.scales[0],
args.scales[1]);
}
else if(args.compTypeId == AppDataType::appInt32)
else if(args.compTypeId == DataTypeEnum::Int32)
{
profile_reduce_impl<int8_t, int32_t, int8_t>(args.do_verification,
args.init_method,
args.do_log,
args.do_dumpout,
args.nrepeat,
args.time_kernel,
args.inLengths,
args.reduceDims,
args.reduceOp,
args.nanOpt,
args.indicesOpt,
static_cast<bool>(args.nanOpt),
static_cast<bool>(args.indicesOpt),
args.scales[0],
args.scales[1]);
}
......@@ -444,54 +371,51 @@ int profile_reduce(int argc, char* argv[])
else if(args.use_bf16)
{
if(args.outType_assigned &&
(args.outTypeId != AppDataType::appBFloat16 && args.outTypeId != AppDataType::appFloat))
args.outTypeId = AppDataType::appFloat;
(args.outTypeId != DataTypeEnum::BFloat16 && args.outTypeId != DataTypeEnum::Float))
args.outTypeId = DataTypeEnum::Float;
if(!args.outType_assigned)
args.outTypeId = AppDataType::appBFloat16;
args.outTypeId = DataTypeEnum::BFloat16;
profile_reduce_impl<ck::bhalf_t, float, ck::bhalf_t>(args.do_verification,
args.init_method,
args.do_log,
args.do_dumpout,
args.nrepeat,
args.time_kernel,
args.inLengths,
args.reduceDims,
args.reduceOp,
args.nanOpt,
args.indicesOpt,
static_cast<bool>(args.nanOpt),
static_cast<bool>(args.indicesOpt),
args.scales[0],
args.scales[1]);
}
else
{
if(args.compTypeId == AppDataType::appFloat)
if(args.compTypeId == DataTypeEnum::Float)
{
profile_reduce_impl<float, float, float>(args.do_verification,
args.init_method,
args.do_log,
args.do_dumpout,
args.nrepeat,
args.time_kernel,
args.inLengths,
args.reduceDims,
args.reduceOp,
args.nanOpt,
args.indicesOpt,
static_cast<bool>(args.nanOpt),
static_cast<bool>(args.indicesOpt),
args.scales[0],
args.scales[1]);
}
else if(args.compTypeId == AppDataType::appDouble)
else if(args.compTypeId == DataTypeEnum::Double)
{
profile_reduce_impl<float, double, float>(args.do_verification,
args.init_method,
args.do_log,
args.do_dumpout,
args.nrepeat,
args.time_kernel,
args.inLengths,
args.reduceDims,
args.reduceOp,
args.nanOpt,
args.indicesOpt,
static_cast<bool>(args.nanOpt),
static_cast<bool>(args.indicesOpt),
args.scales[0],
args.scales[1]);
}
......
profiler/src/profiler.cpp
View file @ a3b4c5cb
......@@ -13,6 +13,7 @@ int profile_gemm_bias_relu_add(int, char*[]);
int profile_gemm_reduce(int, char*[]);
int profile_batched_gemm(int, char*[]);
int profile_grouped_gemm(int, char*[]);
int profile_conv_fwd(int, char*[]);
int profile_conv_fwd_bias_relu(int, char*[]);
int profile_conv_fwd_bias_relu_add(int, char*[]);
int profile_conv_fwd_bias_relu_atomic_add(int, char*[]);
......@@ -53,7 +54,7 @@ int main(int argc, char* argv[])
}
else if(strcmp(argv[1], "grouped_gemm") == 0)
{
profile_grouped_gemm(argc, argv);
return profile_grouped_gemm(argc, argv);
}
else if(strcmp(argv[1], "conv_fwd") == 0)
{
......@@ -107,7 +108,7 @@ int main(int argc, char* argv[])
" conv1d_bwd_data: BackwardConvolution data 1 dim\n"
" conv2d_bwd_data: BackwardConvolution data 2 dim\n"
" conv3d_bwd_data: BackwardConvolution data 3 dim\n"
" reduce: REDUCE\n"
" reduce: Reduce\n"
" conv2d_bwd_weight: Backward Weight Convolution 2d\n");
// clang-format on
}
......
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