Skip to content

GitLab

Commit f9c478e2 authored by ltqin's avatar ltqin
Browse files

Merge branch 'develop' into bmatrix_skip_lds

parents 7d85d04a 91d8b7d6
Hide whitespace changes
Inline Side-by-side
Showing with 260 additions and 1078 deletions
test/gemm_split_k/gemm_split_k.cpp
View file @ f9c478e2
...@@ -45,7 +45,7 @@ static bool check_out(const Tensor<T>& ref, const Tensor<T>& result) ...@@ -45,7 +45,7 @@ static bool check_out(const Tensor<T>& ref, const Tensor<T>& result)
{ {
float max_diff = 1e-6; float max_diff = 1e-6;
for(int i = 0; i < ref.mData.size(); ++i) for(std::size_t i = 0; i < ref.mData.size(); ++i)
{ {
float diff = std::abs(double(ref.mData[i]) - double(result.mData[i])); float diff = std::abs(double(ref.mData[i]) - double(result.mData[i]));
if(max_diff < diff) if(max_diff < diff)
...@@ -187,9 +187,10 @@ int test_gemm(const gemmArgs& args) ...@@ -187,9 +187,10 @@ int test_gemm(const gemmArgs& args)
if(gemm_ptr->IsSupportedArgument(argument_ptr.get())) 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()); c_device_buf.FromDevice(c_m_n_device_result.mData.data());
if(!check_out(c_m_n_host_result, c_m_n_device_result)) if(!check_out(c_m_n_host_result, c_m_n_device_result))
{ {
success = false; success = false;
......
test/grouped_gemm/grouped_gemm_fp16.cpp
View file @ f9c478e2
...@@ -104,7 +104,7 @@ bool TestGroupedGemm(DeviceGroupedGemmPtr_& groupedGemmPtr) ...@@ -104,7 +104,7 @@ bool TestGroupedGemm(DeviceGroupedGemmPtr_& groupedGemmPtr)
b_tensors_device.reserve(group_count); b_tensors_device.reserve(group_count);
c_tensors_device.reserve(group_count); c_tensors_device.reserve(group_count);
for(int i = 0; i < gemm_shapes.size(); i++) for(std::size_t i = 0; i < gemm_shapes.size(); i++)
{ {
a_tensors.emplace_back(Tensor<ADataType>(f_host_tensor_descriptor( a_tensors.emplace_back(Tensor<ADataType>(f_host_tensor_descriptor(
gemm_shapes[i].M, gemm_shapes[i].K, gemm_shapes[i].StrideA, ALayout{}))); gemm_shapes[i].M, gemm_shapes[i].K, gemm_shapes[i].StrideA, ALayout{})));
...@@ -119,7 +119,7 @@ bool TestGroupedGemm(DeviceGroupedGemmPtr_& groupedGemmPtr) ...@@ -119,7 +119,7 @@ bool TestGroupedGemm(DeviceGroupedGemmPtr_& groupedGemmPtr)
b_tensors[i].GenerateTensorValue(GeneratorTensor_2<BDataType>{-5, 5}); b_tensors[i].GenerateTensorValue(GeneratorTensor_2<BDataType>{-5, 5});
} }
for(int i = 0; i < gemm_shapes.size(); i++) for(std::size_t i = 0; i < gemm_shapes.size(); i++)
{ {
a_tensors_device.emplace_back( a_tensors_device.emplace_back(
std::make_unique<DeviceMem>(sizeof(ADataType) * a_tensors[i].mDesc.GetElementSize())); std::make_unique<DeviceMem>(sizeof(ADataType) * a_tensors[i].mDesc.GetElementSize()));
...@@ -147,12 +147,17 @@ bool TestGroupedGemm(DeviceGroupedGemmPtr_& groupedGemmPtr) ...@@ -147,12 +147,17 @@ bool TestGroupedGemm(DeviceGroupedGemmPtr_& groupedGemmPtr)
invoker_ptr->Run(argument_ptr.get()); invoker_ptr->Run(argument_ptr.get());
for(int i = 0; i < gemm_shapes.size(); i++) for(std::size_t i = 0; i < gemm_shapes.size(); i++)
{ {
c_tensors_device[i]->FromDevice(c_device_tensors[i].mData.data()); c_tensors_device[i]->FromDevice(c_device_tensors[i].mData.data());
using ReferenceGemmInstance = ck::tensor_operation::host:: using ReferenceGemmInstance = ck::tensor_operation::host::ReferenceGemm<ADataType,
ReferenceGemm<ADataType, BDataType, CDataType, PassThrough, PassThrough, PassThrough>; BDataType,
CDataType,
AccDataType,
PassThrough,
PassThrough,
PassThrough>;
auto ref_gemm = ReferenceGemmInstance{}; auto ref_gemm = ReferenceGemmInstance{};
auto ref_invoker = ref_gemm.MakeInvoker(); auto ref_invoker = ref_gemm.MakeInvoker();
......
test/reduce/reduce_no_index.cpp
View file @ f9c478e2
#include "getopt.h" #include "getopt.h"
#include "check_err.hpp" #include "host_common_util.hpp"
#include "device_reduce_instance.hpp" #include "profile_reduce_impl.hpp"
#include "reduction_enums.hpp"
#include "host_tensor.hpp"
#include "host_tensor_generator.hpp"
#include "host_reduction.hpp"
#include "reduce_util.hpp"
using namespace ck; 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'}, static struct option long_options[] = {{"inLengths", required_argument, nullptr, 'D'},
{"reduceDimensions", required_argument, nullptr, 'R'}, {"reduceDimensions", required_argument, nullptr, 'R'},
{"scales", required_argument, nullptr, 'S'}, {"scales", required_argument, nullptr, 'S'},
...@@ -387,48 +13,6 @@ static struct option long_options[] = {{"inLengths", required_argument, nullptr, ...@@ -387,48 +13,6 @@ static struct option long_options[] = {{"inLengths", required_argument, nullptr,
class SimpleAppArgs 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: private:
int option_index = 0; int option_index = 0;
...@@ -460,7 +44,9 @@ class SimpleAppArgs ...@@ -460,7 +44,9 @@ class SimpleAppArgs
int processArgs(int argc, char* argv[]) int processArgs(int argc, char* argv[])
{ {
unsigned int ch; using ck::host_common::getTypeValuesFromString;
int ch;
while(1) while(1)
{ {
...@@ -514,7 +100,7 @@ class SimpleAppArgs ...@@ -514,7 +100,7 @@ class SimpleAppArgs
(reduceDims.size() != 1 && reduceDims.size() != 3 && reduceDims.size() != 4)) (reduceDims.size() != 1 && reduceDims.size() != 3 && reduceDims.size() != 4))
return (-1); 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 (-1);
return (0); return (0);
...@@ -525,87 +111,92 @@ bool test_reduce_no_index(int data_type, ...@@ -525,87 +111,92 @@ bool test_reduce_no_index(int data_type,
int init_method, int init_method,
std::vector<int> reduceDims, std::vector<int> reduceDims,
std::vector<size_t> inLengths, std::vector<size_t> inLengths,
ReduceTensorOp reduceOpId,
bool propagateNan,
float alpha, float alpha,
float beta) float beta)
{ {
using ck::profiler::profile_reduce_impl;
bool result = true; bool result = true;
if(data_type == 0) if(data_type == 0)
{ {
switch(reduceDims.size()) result = profile_reduce_impl<float, float, float>(true,
{ init_method,
case 1: false,
result = test_reduce_no_index_impl<float, float, float, Rank, 1>( false,
init_method, inLengths, reduceDims, alpha, beta); inLengths,
break; reduceDims,
case 3: reduceOpId,
result = test_reduce_no_index_impl<float, float, float, Rank, 3>( propagateNan,
init_method, inLengths, reduceDims, alpha, beta); false,
break; alpha,
case 4: beta);
result = test_reduce_no_index_impl<float, float, float, Rank, 4>(
init_method, inLengths, reduceDims, alpha, beta);
break;
};
} }
else if(data_type == 1) else if(data_type == 1)
{ {
switch(reduceDims.size()) result = profile_reduce_impl<ck::half_t, float, ck::half_t>(true,
{ init_method,
case 1: false,
result = test_reduce_no_index_impl<ck::half_t, float, ck::half_t, Rank, 1>( false,
init_method, inLengths, reduceDims, alpha, beta); inLengths,
break; reduceDims,
case 3: reduceOpId,
result = test_reduce_no_index_impl<ck::half_t, float, ck::half_t, Rank, 3>( propagateNan,
init_method, inLengths, reduceDims, alpha, beta); false,
break; alpha,
case 4: beta);
result = test_reduce_no_index_impl<ck::half_t, float, ck::half_t, Rank, 4>(
init_method, inLengths, reduceDims, alpha, beta);
break;
};
} }
else if(data_type == 3) else if(data_type == 3)
{ {
switch(reduceDims.size()) result = profile_reduce_impl<int8_t, int32_t, int8_t>(true,
{ init_method,
case 1: false,
result = test_reduce_no_index_impl<int8_t, int32_t, int8_t, Rank, 1>( false,
init_method, inLengths, reduceDims, alpha, beta); inLengths,
break; reduceDims,
case 3: reduceOpId,
result = test_reduce_no_index_impl<int8_t, int32_t, int8_t, Rank, 3>( propagateNan,
init_method, inLengths, reduceDims, alpha, beta); false,
break; alpha,
case 4: beta);
result = test_reduce_no_index_impl<int8_t, int32_t, int8_t, Rank, 4>(
init_method, inLengths, reduceDims, alpha, beta);
break;
};
} }
else if(data_type == 5) else if(data_type == 5)
{ {
switch(reduceDims.size()) result = profile_reduce_impl<ck::bhalf_t, float, ck::bhalf_t>(true,
{ init_method,
case 1: false,
result = test_reduce_no_index_impl<ck::bhalf_t, float, ck::bhalf_t, Rank, 1>( false,
init_method, inLengths, reduceDims, alpha, beta); inLengths,
break; reduceDims,
case 3: reduceOpId,
result = test_reduce_no_index_impl<ck::bhalf_t, float, ck::bhalf_t, Rank, 3>( propagateNan,
init_method, inLengths, reduceDims, alpha, beta); false,
break; alpha,
case 4: beta);
result = test_reduce_no_index_impl<ck::bhalf_t, float, ck::bhalf_t, Rank, 4>( }
init_method, inLengths, reduceDims, alpha, beta); else if(data_type == 6)
break; {
}; result = profile_reduce_impl<double, double, double>(true,
init_method,
false,
false,
inLengths,
reduceDims,
reduceOpId,
propagateNan,
false,
alpha,
beta);
} }
return (result); return (result);
}; };
constexpr ReduceTensorOp reduceOpId = ReduceTensorOp::AVG;
constexpr bool propagateNan = false;
int main(int argc, char* argv[]) int main(int argc, char* argv[])
{ {
SimpleAppArgs args; SimpleAppArgs args;
...@@ -621,8 +212,14 @@ int main(int argc, char* argv[]) ...@@ -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}}; {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) for(auto& reduceDims : v_reduceDims)
result = result && test_reduce_no_index( result = result && test_reduce_no_index(data_type,
data_type, init_method, reduceDims, inLengths, 1.0f, 0.0f); init_method,
reduceDims,
inLengths,
reduceOpId,
propagateNan,
1.0f,
0.0f);
} }
else else
{ {
...@@ -636,6 +233,8 @@ int main(int argc, char* argv[]) ...@@ -636,6 +233,8 @@ int main(int argc, char* argv[])
args.init_method, args.init_method,
args.reduceDims, args.reduceDims,
args.inLengths, args.inLengths,
reduceOpId,
propagateNan,
args.scales[0], args.scales[0],
args.scales[1]); args.scales[1]);
} }
......
test/reduce/reduce_util.hpp deleted 100644 → 0
View file @ 7d85d04a
#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(int 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 @ f9c478e2
#include "getopt.h" #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; #include "host_common_util.hpp"
#include "profile_reduce_impl.hpp"
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!");
};
if(!single_result) using namespace ck;
{
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
static struct option long_options[] = {{"inLengths", required_argument, nullptr, 'D'}, static struct option long_options[] = {{"inLengths", required_argument, nullptr, 'D'},
{"reduceDimensions", required_argument, nullptr, 'R'}, {"reduceDimensions", required_argument, nullptr, 'R'},
...@@ -390,48 +13,6 @@ static struct option long_options[] = {{"inLengths", required_argument, nullptr, ...@@ -390,48 +13,6 @@ static struct option long_options[] = {{"inLengths", required_argument, nullptr,
class SimpleAppArgs 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: private:
int option_index = 0; int option_index = 0;
...@@ -463,7 +44,9 @@ class SimpleAppArgs ...@@ -463,7 +44,9 @@ class SimpleAppArgs
int processArgs(int argc, char* argv[]) int processArgs(int argc, char* argv[])
{ {
unsigned int ch; using ck::host_common::getTypeValuesFromString;
int ch;
while(1) while(1)
{ {
...@@ -517,7 +100,7 @@ class SimpleAppArgs ...@@ -517,7 +100,7 @@ class SimpleAppArgs
(reduceDims.size() != 1 && reduceDims.size() != 3 && reduceDims.size() != 4)) (reduceDims.size() != 1 && reduceDims.size() != 3 && reduceDims.size() != 4))
return (-1); 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 (-1);
return (0); return (0);
...@@ -528,87 +111,92 @@ bool test_reduce_with_index(int data_type, ...@@ -528,87 +111,92 @@ bool test_reduce_with_index(int data_type,
int init_method, int init_method,
std::vector<int> reduceDims, std::vector<int> reduceDims,
std::vector<size_t> inLengths, std::vector<size_t> inLengths,
ReduceTensorOp reduceOpId,
bool propagateNan,
float alpha, float alpha,
float beta) float beta)
{ {
using ck::profiler::profile_reduce_impl;
bool result = true; bool result = true;
if(data_type == 0) if(data_type == 0)
{ {
switch(reduceDims.size()) result = profile_reduce_impl<float, float, float>(true,
{ init_method,
case 1: false,
result = test_reduce_with_index_impl<float, float, float, Rank, 1>( false,
init_method, inLengths, reduceDims, alpha, beta); inLengths,
break; reduceDims,
case 3: reduceOpId,
result = test_reduce_with_index_impl<float, float, float, Rank, 3>( propagateNan,
init_method, inLengths, reduceDims, alpha, beta); true,
break; alpha,
case 4: beta);
result = test_reduce_with_index_impl<float, float, float, Rank, 4>(
init_method, inLengths, reduceDims, alpha, beta);
break;
};
} }
else if(data_type == 1) else if(data_type == 1)
{ {
switch(reduceDims.size()) result = profile_reduce_impl<ck::half_t, ck::half_t, ck::half_t>(true,
{ init_method,
case 1: false,
result = test_reduce_with_index_impl<ck::half_t, ck::half_t, ck::half_t, Rank, 1>( false,
init_method, inLengths, reduceDims, alpha, beta); inLengths,
break; reduceDims,
case 3: reduceOpId,
result = test_reduce_with_index_impl<ck::half_t, ck::half_t, ck::half_t, Rank, 3>( propagateNan,
init_method, inLengths, reduceDims, alpha, beta); true,
break; alpha,
case 4: beta);
result = test_reduce_with_index_impl<ck::half_t, ck::half_t, ck::half_t, Rank, 4>(
init_method, inLengths, reduceDims, alpha, beta);
break;
};
} }
else if(data_type == 3) else if(data_type == 3)
{ {
switch(reduceDims.size()) result = profile_reduce_impl<int8_t, int8_t, int8_t>(true,
{ init_method,
case 1: false,
result = test_reduce_with_index_impl<int8_t, int8_t, int8_t, Rank, 1>( false,
init_method, inLengths, reduceDims, alpha, beta); inLengths,
break; reduceDims,
case 3: reduceOpId,
result = test_reduce_with_index_impl<int8_t, int8_t, int8_t, Rank, 3>( propagateNan,
init_method, inLengths, reduceDims, alpha, beta); true,
break; alpha,
case 4: beta);
result = test_reduce_with_index_impl<int8_t, int8_t, int8_t, Rank, 4>(
init_method, inLengths, reduceDims, alpha, beta);
break;
};
} }
else if(data_type == 5) else if(data_type == 5)
{ {
switch(reduceDims.size()) result = profile_reduce_impl<ck::bhalf_t, float, ck::bhalf_t>(true,
{ init_method,
case 1: false,
result = test_reduce_with_index_impl<ck::bhalf_t, float, ck::bhalf_t, Rank, 1>( false,
init_method, inLengths, reduceDims, alpha, beta); inLengths,
break; reduceDims,
case 3: reduceOpId,
result = test_reduce_with_index_impl<ck::bhalf_t, float, ck::bhalf_t, Rank, 3>( propagateNan,
init_method, inLengths, reduceDims, alpha, beta); true,
break; alpha,
case 4: beta);
result = test_reduce_with_index_impl<ck::bhalf_t, float, ck::bhalf_t, Rank, 4>( }
init_method, inLengths, reduceDims, alpha, beta); else if(data_type == 6)
break; {
}; result = profile_reduce_impl<double, double, double>(true,
init_method,
false,
false,
inLengths,
reduceDims,
reduceOpId,
propagateNan,
true,
alpha,
beta);
} }
return (result); return (result);
}; };
constexpr ReduceTensorOp reduceOpId = ReduceTensorOp::AMAX;
constexpr bool propagateNan = false;
int main(int argc, char* argv[]) int main(int argc, char* argv[])
{ {
SimpleAppArgs args; SimpleAppArgs args;
...@@ -624,8 +212,14 @@ int main(int argc, char* argv[]) ...@@ -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}}; {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) for(auto& reduceDims : v_reduceDims)
result = result && test_reduce_with_index( result = result && test_reduce_with_index(data_type,
data_type, init_method, reduceDims, inLengths, 1.0f, 0.0f); init_method,
reduceDims,
inLengths,
reduceOpId,
propagateNan,
1.0f,
0.0f);
} }
else else
{ {
...@@ -639,6 +233,8 @@ int main(int argc, char* argv[]) ...@@ -639,6 +233,8 @@ int main(int argc, char* argv[])
args.init_method, args.init_method,
args.reduceDims, args.reduceDims,
args.inLengths, args.inLengths,
reduceOpId,
propagateNan,
args.scales[0], args.scales[0],
args.scales[1]); args.scales[1]);
} }
......
test/reference_conv_fwd/CMakeLists.txt
View file @ f9c478e2
add_gtest_executable(test_reference_conv_fwd reference_conv_fwd.cpp) add_gtest_executable(test_reference_conv_fwd reference_conv_fwd.cpp)
target_link_libraries(test_reference_conv_fwd PRIVATE host_tensor conv_fwd_util) target_link_libraries(test_reference_conv_fwd PRIVATE host_tensor conv_util)
test/reference_conv_fwd/reference_conv_fwd.cpp
View file @ f9c478e2
...@@ -8,7 +8,7 @@ ...@@ -8,7 +8,7 @@
#include "check_err.hpp" #include "check_err.hpp"
#include "config.hpp" #include "config.hpp"
#include "conv_fwd_util.hpp" #include "conv_util.hpp"
#include "element_wise_operation.hpp" #include "element_wise_operation.hpp"
#include "fill.hpp" #include "fill.hpp"
#include "host_tensor.hpp" #include "host_tensor.hpp"
...@@ -34,21 +34,21 @@ run_reference_convolution_forward(const ck::utils::conv::ConvParams& params, ...@@ -34,21 +34,21 @@ run_reference_convolution_forward(const ck::utils::conv::ConvParams& params,
const FillInputOp& fill_input_op = FillInputOp{}, const FillInputOp& fill_input_op = FillInputOp{},
const FillWeightsOp& fill_weights_op = FillWeightsOp{0.5f}) const FillWeightsOp& fill_weights_op = FillWeightsOp{0.5f})
{ {
std::vector<std::size_t> input_dims{static_cast<std::size_t>(params.N), std::vector<std::size_t> input_dims{static_cast<std::size_t>(params.N_),
static_cast<std::size_t>(params.C)}; static_cast<std::size_t>(params.C_)};
input_dims.insert(std::end(input_dims), input_dims.insert(std::end(input_dims),
std::begin(params.input_spatial_lengths), std::begin(params.input_spatial_lengths_),
std::end(params.input_spatial_lengths)); std::end(params.input_spatial_lengths_));
std::vector<std::size_t> filter_dims{static_cast<std::size_t>(params.K), std::vector<std::size_t> filter_dims{static_cast<std::size_t>(params.K_),
static_cast<std::size_t>(params.C)}; static_cast<std::size_t>(params.C_)};
filter_dims.insert(std::end(filter_dims), filter_dims.insert(std::end(filter_dims),
std::begin(params.filter_spatial_lengths), std::begin(params.filter_spatial_lengths_),
std::end(params.filter_spatial_lengths)); std::end(params.filter_spatial_lengths_));
const std::vector<ck::index_t>& output_spatial_lengths = params.GetOutputSpatialLengths(); const std::vector<ck::index_t>& output_spatial_lengths = params.GetOutputSpatialLengths();
std::vector<std::size_t> output_dims{static_cast<std::size_t>(params.N), std::vector<std::size_t> output_dims{static_cast<std::size_t>(params.N_),
static_cast<std::size_t>(params.K)}; static_cast<std::size_t>(params.K_)};
output_dims.insert(std::end(output_dims), output_dims.insert(std::end(output_dims),
std::begin(output_spatial_lengths), std::begin(output_spatial_lengths),
std::end(output_spatial_lengths)); std::end(output_spatial_lengths));
...@@ -74,10 +74,10 @@ run_reference_convolution_forward(const ck::utils::conv::ConvParams& params, ...@@ -74,10 +74,10 @@ run_reference_convolution_forward(const ck::utils::conv::ConvParams& params,
auto ref_argument = ref_conv.MakeArgument(input, auto ref_argument = ref_conv.MakeArgument(input,
weights, weights,
host_output, host_output,
params.conv_filter_strides, params.conv_filter_strides_,
params.conv_filter_dilations, params.conv_filter_dilations_,
params.input_left_pads, params.input_left_pads_,
params.input_right_pads, params.input_right_pads_,
InElementOp{}, InElementOp{},
WeiElementOp{}, WeiElementOp{},
OutElementOp{}); OutElementOp{});
...@@ -91,15 +91,15 @@ run_reference_convolution_forward(const ck::utils::conv::ConvParams& params, ...@@ -91,15 +91,15 @@ run_reference_convolution_forward(const ck::utils::conv::ConvParams& params,
TEST(ReferenceConvolutionFWD, Conv2DNHWC) TEST(ReferenceConvolutionFWD, Conv2DNHWC)
{ {
ck::utils::conv::ConvParams params; ck::utils::conv::ConvParams params;
params.N = 1; params.N_ = 1;
params.K = 1; params.K_ = 1;
params.C = 2; params.C_ = 2;
params.filter_spatial_lengths = std::vector<ck::index_t>{3, 3}; params.filter_spatial_lengths_ = std::vector<ck::index_t>{3, 3};
params.input_spatial_lengths = std::vector<ck::index_t>{6, 6}; params.input_spatial_lengths_ = std::vector<ck::index_t>{6, 6};
params.conv_filter_strides = std::vector<ck::index_t>{1, 1}; params.conv_filter_strides_ = std::vector<ck::index_t>{1, 1};
params.conv_filter_dilations = std::vector<ck::index_t>{1, 1}; params.conv_filter_dilations_ = std::vector<ck::index_t>{1, 1};
params.input_left_pads = std::vector<ck::index_t>{0, 0}; params.input_left_pads_ = std::vector<ck::index_t>{0, 0};
params.input_right_pads = std::vector<ck::index_t>{0, 0}; params.input_right_pads_ = std::vector<ck::index_t>{0, 0};
auto out_tensor = run_reference_convolution_forward<2>(params); auto out_tensor = run_reference_convolution_forward<2>(params);
std::vector<std::size_t> ref_dims{1, 1, 4, 4}; std::vector<std::size_t> ref_dims{1, 1, 4, 4};
...@@ -127,15 +127,15 @@ TEST(ReferenceConvolutionFWD, Conv2DNHWC) ...@@ -127,15 +127,15 @@ TEST(ReferenceConvolutionFWD, Conv2DNHWC)
TEST(ReferenceConvolutionFWD, Conv2DNHWCStridesDilationsPadding) TEST(ReferenceConvolutionFWD, Conv2DNHWCStridesDilationsPadding)
{ {
ck::utils::conv::ConvParams params; ck::utils::conv::ConvParams params;
params.N = 1; params.N_ = 1;
params.K = 2; params.K_ = 2;
params.C = 2; params.C_ = 2;
params.filter_spatial_lengths = std::vector<ck::index_t>{3, 3}; params.filter_spatial_lengths_ = std::vector<ck::index_t>{3, 3};
params.input_spatial_lengths = std::vector<ck::index_t>{12, 12}; params.input_spatial_lengths_ = std::vector<ck::index_t>{12, 12};
params.conv_filter_strides = std::vector<ck::index_t>{2, 2}; params.conv_filter_strides_ = std::vector<ck::index_t>{2, 2};
params.conv_filter_dilations = std::vector<ck::index_t>{2, 2}; params.conv_filter_dilations_ = std::vector<ck::index_t>{2, 2};
params.input_left_pads = std::vector<ck::index_t>{1, 1}; params.input_left_pads_ = std::vector<ck::index_t>{1, 1};
params.input_right_pads = std::vector<ck::index_t>{1, 1}; params.input_right_pads_ = std::vector<ck::index_t>{1, 1};
auto out_tensor = run_reference_convolution_forward<2>(params); auto out_tensor = run_reference_convolution_forward<2>(params);
std::vector<std::size_t> ref_dims = std::vector<std::size_t>{1, 2, 5, 5}; std::vector<std::size_t> ref_dims = std::vector<std::size_t>{1, 2, 5, 5};
...@@ -153,16 +153,16 @@ TEST(ReferenceConvolutionFWD, Conv2DNHWCStridesDilationsPadding) ...@@ -153,16 +153,16 @@ TEST(ReferenceConvolutionFWD, Conv2DNHWCStridesDilationsPadding)
TEST(ReferenceConvolutionFWD, Conv1DNWC) TEST(ReferenceConvolutionFWD, Conv1DNWC)
{ {
ck::utils::conv::ConvParams params; ck::utils::conv::ConvParams params;
params.num_dim_spatial = 1; params.num_dim_spatial_ = 1;
params.N = 1; params.N_ = 1;
params.K = 1; params.K_ = 1;
params.C = 2; params.C_ = 2;
params.filter_spatial_lengths = std::vector<ck::index_t>{3}; params.filter_spatial_lengths_ = std::vector<ck::index_t>{3};
params.input_spatial_lengths = std::vector<ck::index_t>{6}; params.input_spatial_lengths_ = std::vector<ck::index_t>{6};
params.conv_filter_strides = std::vector<ck::index_t>{1}; params.conv_filter_strides_ = std::vector<ck::index_t>{1};
params.conv_filter_dilations = std::vector<ck::index_t>{1}; params.conv_filter_dilations_ = std::vector<ck::index_t>{1};
params.input_left_pads = std::vector<ck::index_t>{0}; params.input_left_pads_ = std::vector<ck::index_t>{0};
params.input_right_pads = std::vector<ck::index_t>{0}; params.input_right_pads_ = std::vector<ck::index_t>{0};
auto out_tensor = auto out_tensor =
run_reference_convolution_forward<1, run_reference_convolution_forward<1,
...@@ -182,16 +182,16 @@ TEST(ReferenceConvolutionFWD, Conv1DNWC) ...@@ -182,16 +182,16 @@ TEST(ReferenceConvolutionFWD, Conv1DNWC)
TEST(ReferenceConvolutionFWD, Conv1DNWCStridesDilationsPadding) TEST(ReferenceConvolutionFWD, Conv1DNWCStridesDilationsPadding)
{ {
ck::utils::conv::ConvParams params; ck::utils::conv::ConvParams params;
params.num_dim_spatial = 1; params.num_dim_spatial_ = 1;
params.N = 1; params.N_ = 1;
params.K = 2; params.K_ = 2;
params.C = 2; params.C_ = 2;
params.filter_spatial_lengths = std::vector<ck::index_t>{3}; params.filter_spatial_lengths_ = std::vector<ck::index_t>{3};
params.input_spatial_lengths = std::vector<ck::index_t>{12}; params.input_spatial_lengths_ = std::vector<ck::index_t>{12};
params.conv_filter_strides = std::vector<ck::index_t>{2}; params.conv_filter_strides_ = std::vector<ck::index_t>{2};
params.conv_filter_dilations = std::vector<ck::index_t>{2}; params.conv_filter_dilations_ = std::vector<ck::index_t>{2};
params.input_left_pads = std::vector<ck::index_t>{1}; params.input_left_pads_ = std::vector<ck::index_t>{1};
params.input_right_pads = std::vector<ck::index_t>{1}; params.input_right_pads_ = std::vector<ck::index_t>{1};
auto out_tensor = auto out_tensor =
run_reference_convolution_forward<1, run_reference_convolution_forward<1,
...@@ -211,16 +211,16 @@ TEST(ReferenceConvolutionFWD, Conv1DNWCStridesDilationsPadding) ...@@ -211,16 +211,16 @@ TEST(ReferenceConvolutionFWD, Conv1DNWCStridesDilationsPadding)
TEST(ReferenceConvolutionFWD, Conv1DNWCSameOutputSize) TEST(ReferenceConvolutionFWD, Conv1DNWCSameOutputSize)
{ {
ck::utils::conv::ConvParams params; ck::utils::conv::ConvParams params;
params.num_dim_spatial = 1; params.num_dim_spatial_ = 1;
params.N = 2; params.N_ = 2;
params.K = 16; params.K_ = 16;
params.C = 4; params.C_ = 4;
params.filter_spatial_lengths = std::vector<ck::index_t>{3}; params.filter_spatial_lengths_ = std::vector<ck::index_t>{3};
params.input_spatial_lengths = std::vector<ck::index_t>{16}; params.input_spatial_lengths_ = std::vector<ck::index_t>{16};
params.conv_filter_strides = std::vector<ck::index_t>{1}; params.conv_filter_strides_ = std::vector<ck::index_t>{1};
params.conv_filter_dilations = std::vector<ck::index_t>{1}; params.conv_filter_dilations_ = std::vector<ck::index_t>{1};
params.input_left_pads = std::vector<ck::index_t>{1}; params.input_left_pads_ = std::vector<ck::index_t>{1};
params.input_right_pads = std::vector<ck::index_t>{1}; params.input_right_pads_ = std::vector<ck::index_t>{1};
auto out_tensor2 = run_reference_convolution_forward<1, auto out_tensor2 = run_reference_convolution_forward<1,
float, float,
...@@ -305,16 +305,16 @@ TEST(ReferenceConvolutionFWD, Conv1DNWCSameOutputSize) ...@@ -305,16 +305,16 @@ TEST(ReferenceConvolutionFWD, Conv1DNWCSameOutputSize)
TEST(ReferenceConvolutionFWD, Conv3DNCDHW) TEST(ReferenceConvolutionFWD, Conv3DNCDHW)
{ {
ck::utils::conv::ConvParams params; ck::utils::conv::ConvParams params;
params.num_dim_spatial = 3; params.num_dim_spatial_ = 3;
params.N = 1; params.N_ = 1;
params.K = 1; params.K_ = 1;
params.C = 2; params.C_ = 2;
params.filter_spatial_lengths = std::vector<ck::index_t>{3, 3, 3}; params.filter_spatial_lengths_ = std::vector<ck::index_t>{3, 3, 3};
params.input_spatial_lengths = std::vector<ck::index_t>{6, 6, 6}; params.input_spatial_lengths_ = std::vector<ck::index_t>{6, 6, 6};
params.conv_filter_strides = std::vector<ck::index_t>{1, 1, 1}; params.conv_filter_strides_ = std::vector<ck::index_t>{1, 1, 1};
params.conv_filter_dilations = std::vector<ck::index_t>{1, 1, 1}; params.conv_filter_dilations_ = std::vector<ck::index_t>{1, 1, 1};
params.input_left_pads = std::vector<ck::index_t>{0, 0, 0}; params.input_left_pads_ = std::vector<ck::index_t>{0, 0, 0};
params.input_right_pads = std::vector<ck::index_t>{0, 0, 0}; params.input_right_pads_ = std::vector<ck::index_t>{0, 0, 0};
auto out_tensor = run_reference_convolution_forward<3, auto out_tensor = run_reference_convolution_forward<3,
float, float,
...@@ -344,16 +344,16 @@ TEST(ReferenceConvolutionFWD, Conv3DNCDHW) ...@@ -344,16 +344,16 @@ TEST(ReferenceConvolutionFWD, Conv3DNCDHW)
TEST(ReferenceConvolutionFWD, Conv3DNCDHWStridesDilations) TEST(ReferenceConvolutionFWD, Conv3DNCDHWStridesDilations)
{ {
ck::utils::conv::ConvParams params; ck::utils::conv::ConvParams params;
params.num_dim_spatial = 3; params.num_dim_spatial_ = 3;
params.N = 1; params.N_ = 1;
params.K = 2; params.K_ = 2;
params.C = 2; params.C_ = 2;
params.filter_spatial_lengths = std::vector<ck::index_t>{3, 3, 3}; params.filter_spatial_lengths_ = std::vector<ck::index_t>{3, 3, 3};
params.input_spatial_lengths = std::vector<ck::index_t>{12, 12, 12}; params.input_spatial_lengths_ = std::vector<ck::index_t>{12, 12, 12};
params.conv_filter_strides = std::vector<ck::index_t>{3, 3, 3}; params.conv_filter_strides_ = std::vector<ck::index_t>{3, 3, 3};
params.conv_filter_dilations = std::vector<ck::index_t>{1, 1, 1}; params.conv_filter_dilations_ = std::vector<ck::index_t>{1, 1, 1};
params.input_left_pads = std::vector<ck::index_t>{0, 0, 0}; params.input_left_pads_ = std::vector<ck::index_t>{0, 0, 0};
params.input_right_pads = std::vector<ck::index_t>{0, 0, 0}; params.input_right_pads_ = std::vector<ck::index_t>{0, 0, 0};
auto out_tensor = run_reference_convolution_forward<3, auto out_tensor = run_reference_convolution_forward<3,
float, float,
......
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