Skip to content

GitLab

Commit b134b7d6 authored by carlushuang's avatar carlushuang
Browse files

Merge remote-tracking branch 'origin/develop' into cpu_avx2

parents 090ba885 9f71ff48
Hide whitespace changes
Inline Side-by-side
Showing with 504 additions and 519 deletions
test/gemm/gemm_fp32.cpp
View file @ b134b7d6
......@@ -15,7 +15,7 @@
#include "host_gemm.hpp"
#include "device_tensor.hpp"
#include "device_gemm_xdl.hpp"
#include "device_gemm_xdl_c_shuffle.hpp"
#include "device_gemm_xdl_cshuffle.hpp"
#include "element_wise_operation.hpp"
#include "reference_gemm.hpp"
#include "gemm_specialization.hpp"
......
test/gemm/gemm_int8.cpp
View file @ b134b7d6
......@@ -15,7 +15,7 @@
#include "host_gemm.hpp"
#include "device_tensor.hpp"
#include "device_gemm_xdl.hpp"
#include "device_gemm_xdl_c_shuffle.hpp"
#include "device_gemm_xdl_cshuffle.hpp"
#include "element_wise_operation.hpp"
#include "reference_gemm.hpp"
#include "gemm_specialization.hpp"
......
test/gemm/gemm_util.hpp
View file @ b134b7d6
#ifndef GEMM_UTILS_HPP
#define GEMM_UTILS_HPP
#include "check_err.hpp"
#include "config.hpp"
#include "device.hpp"
#include "host_tensor.hpp"
#include "host_tensor_generator.hpp"
#include "reference_gemm.hpp"
#include "tensor_layout.hpp"
namespace ck {
namespace gemm_util {
struct GemmParams
{
GemmParams()
: M(1024), N(1024), K(1024), StrideA(1024), StrideB(1024), StrideC(1024), alpha(1), beta(0)
{
}
ck::index_t M;
ck::index_t N;
ck::index_t K;
ck::index_t StrideA;
ck::index_t StrideB;
ck::index_t StrideC;
float alpha;
float beta;
};
template <typename GemmInstance,
typename ADataType,
typename BDataType,
typename CDataType,
typename AElementwiseOperation,
typename BElementwiseOperation,
typename CElementwiseOperation>
void RunHostGEMM(const Tensor<ADataType>& A,
const Tensor<BDataType>& B,
Tensor<CDataType>& C,
AElementwiseOperation a_element_op,
BElementwiseOperation b_element_op,
CElementwiseOperation c_element_op)
{
auto ref_gemm = GemmInstance{};
auto ref_invoker = ref_gemm.MakeInvoker();
auto ref_argument = ref_gemm.MakeArgument(A, B, C, a_element_op, b_element_op, c_element_op);
ref_invoker.Run(ref_argument);
}
template <typename DeviceGemmPtr_,
typename ADataType,
typename BDataType,
typename CDataType,
typename AElementwiseOperation,
typename BElementwiseOperation,
typename CElementwiseOperation>
void RunDeviceGEMM(DeviceGemmPtr_& gemmPtr,
const ck::gemm_util::GemmParams& params,
const Tensor<ADataType>& A,
const Tensor<BDataType>& B,
Tensor<CDataType>& C,
AElementwiseOperation a_element_op,
BElementwiseOperation b_element_op,
CElementwiseOperation c_element_op)
{
DeviceMem a_m_k_device_buf(sizeof(ADataType) * A.mDesc.GetElementSpace());
DeviceMem b_k_n_device_buf(sizeof(BDataType) * B.mDesc.GetElementSpace());
DeviceMem c_m_n_device_buf(sizeof(CDataType) * C.mDesc.GetElementSpace());
a_m_k_device_buf.ToDevice(A.mData.data());
b_k_n_device_buf.ToDevice(B.mData.data());
auto invoker_ptr = gemmPtr->MakeInvokerPointer();
auto argument_ptr =
gemmPtr->MakeArgumentPointer(static_cast<ADataType*>(a_m_k_device_buf.GetDeviceBuffer()),
static_cast<BDataType*>(b_k_n_device_buf.GetDeviceBuffer()),
static_cast<CDataType*>(c_m_n_device_buf.GetDeviceBuffer()),
params.M,
params.N,
params.K,
params.StrideA,
params.StrideB,
params.StrideC,
a_element_op,
b_element_op,
c_element_op);
if(!gemmPtr->IsSupportedArgument(argument_ptr.get()))
{
throw std::runtime_error(
"wrong! device_gemm with the specified compilation parameters does "
"not support this GEMM problem");
}
invoker_ptr->Run(argument_ptr.get());
c_m_n_device_buf.FromDevice(C.mData.data());
}
template <typename DeviceGemmPtr_,
typename ADataType,
typename BDataType,
typename CDataType,
typename ALayout,
typename BLayout,
typename CLayout,
typename AElementwiseOperation,
typename BElementwiseOperation,
typename CElementwiseOperation>
struct TestGemm
{
auto PrepareGemmTensor(const ck::gemm_util::GemmParams& params)
{
auto f_host_tensor_descriptor =
[](std::size_t row, std::size_t col, std::size_t stride, auto layout) {
if(std::is_same<decltype(layout), ck::tensor_layout::gemm::RowMajor>::value)
{
return HostTensorDescriptor(std::vector<std::size_t>({row, col}),
std::vector<std::size_t>({stride, 1}));
}
else
{
return HostTensorDescriptor(std::vector<std::size_t>({row, col}),
std::vector<std::size_t>({1, stride}));
}
};
Tensor<ADataType> a_m_k(
f_host_tensor_descriptor(params.M, params.K, params.StrideA, ALayout{}));
Tensor<BDataType> b_k_n(
f_host_tensor_descriptor(params.K, params.N, params.StrideB, BLayout{}));
Tensor<CDataType> c_m_n_host_result(
f_host_tensor_descriptor(params.M, params.N, params.StrideC, CLayout{}));
Tensor<CDataType> c_m_n_device_result(
f_host_tensor_descriptor(params.M, params.N, params.StrideC, CLayout{}));
auto f_generate_tensor_value = [](auto desc, auto type) {
using dataType = decltype(type);
if(std::is_same<dataType, int8_t>::value)
{
desc.GenerateTensorValue(GeneratorTensor_2<int8_t>{-5, 5});
}
else
{
desc.GenerateTensorValue(GeneratorTensor_3<dataType>{-0.5, 0.5});
}
};
f_generate_tensor_value(a_m_k, ADataType{});
f_generate_tensor_value(b_k_n, BDataType{});
return std::make_tuple(a_m_k, b_k_n, c_m_n_host_result, c_m_n_device_result);
}
auto operator()(DeviceGemmPtr_& gemmPtr)
{
std::cout << "ALayout = " << ALayout{}.name << ", BLayout = " << BLayout{}.name
<< ", CLayout = " << CLayout{}.name << std::endl;
std::cout << gemmPtr->GetTypeString() << std::endl;
// Arrange
ck::gemm_util::GemmParams params;
params.M = 1024;
params.N = 1024;
params.K = 1024;
params.StrideA = 1024;
params.StrideB = 1024;
params.StrideC = 1024;
auto host_tensors = PrepareGemmTensor(params);
const Tensor<ADataType>& a = std::get<0>(host_tensors);
const Tensor<BDataType>& b = std::get<1>(host_tensors);
Tensor<CDataType>& c_host = std::get<2>(host_tensors);
Tensor<CDataType>& c_device = std::get<3>(host_tensors);
auto a_element_op = AElementwiseOperation{};
auto b_element_op = BElementwiseOperation{};
auto c_element_op = CElementwiseOperation{};
using ReferenceGemmInstance =
ck::tensor_operation::host::ReferenceGemm<ADataType,
BDataType,
CDataType,
AElementwiseOperation,
BElementwiseOperation,
CElementwiseOperation>;
ck::gemm_util::RunHostGEMM<ReferenceGemmInstance>(
a, b, c_host, a_element_op, b_element_op, c_element_op);
// Act
ck::gemm_util::RunDeviceGEMM(
gemmPtr, params, a, b, c_device, a_element_op, b_element_op, c_element_op);
// Assert
bool res = false;
if(std::is_same<CDataType, float>::value)
{
res = ck::utils::check_err(c_device.mData, c_host.mData);
std::cout << (res ? "SUCCESS" : "FAILURE") << std::endl;
}
else if(std::is_same<CDataType, ck::half_t>::value)
{
res = ck::utils::check_err(c_device.mData, c_host.mData);
std::cout << (res ? "SUCCESS" : "FAILURE") << std::endl;
}
else if(std::is_same<CDataType, int8_t>::value)
{
res = ck::utils::check_err(c_device.mData, c_host.mData);
std::cout << (res ? "SUCCESS" : "FAILURE") << std::endl;
}
return res;
}
};
template <typename DeviceGemmPtr_,
typename ALayout,
typename BLayout,
typename CLayout,
typename AElementwiseOperation,
typename BElementwiseOperation,
typename CElementwiseOperation>
struct TestGemmBF16
{
using BF16 = ck::bhalf_t;
auto PrepareGemmTensorBF16(const ck::gemm_util::GemmParams& params)
{
auto f_host_tensor_descriptor =
[](std::size_t row, std::size_t col, std::size_t stride, auto layout) {
if(std::is_same<decltype(layout), ck::tensor_layout::gemm::RowMajor>::value)
{
return HostTensorDescriptor(std::vector<std::size_t>({row, col}),
std::vector<std::size_t>({stride, 1}));
}
else
{
return HostTensorDescriptor(std::vector<std::size_t>({row, col}),
std::vector<std::size_t>({1, stride}));
}
};
// use fp32 host kernel to verify bf16 device kernel
Tensor<BF16> a_m_k_bf16(
f_host_tensor_descriptor(params.M, params.K, params.StrideA, ALayout{}));
Tensor<BF16> b_k_n_bf16(
f_host_tensor_descriptor(params.K, params.N, params.StrideB, BLayout{}));
Tensor<BF16> c_m_n_device_bf16(
f_host_tensor_descriptor(params.M, params.N, params.StrideC, CLayout{}));
Tensor<float> a_m_k_fp32(
f_host_tensor_descriptor(params.M, params.K, params.StrideA, ALayout{}));
Tensor<float> b_k_n_fp32(
f_host_tensor_descriptor(params.K, params.N, params.StrideB, BLayout{}));
Tensor<float> c_m_n_host_fp32(
f_host_tensor_descriptor(params.M, params.N, params.StrideC, CLayout{}));
Tensor<float> c_m_n_device_fp32(
f_host_tensor_descriptor(params.M, params.N, params.StrideC, CLayout{}));
a_m_k_bf16.GenerateTensorValue(GeneratorTensor_3<BF16>{-0.5, 0.5});
b_k_n_bf16.GenerateTensorValue(GeneratorTensor_3<BF16>{-0.5, 0.5});
bf16_to_f32_(a_m_k_bf16, a_m_k_fp32);
bf16_to_f32_(b_k_n_bf16, b_k_n_fp32);
return std::make_tuple(a_m_k_bf16,
b_k_n_bf16,
c_m_n_device_bf16,
a_m_k_fp32,
b_k_n_fp32,
c_m_n_host_fp32,
c_m_n_device_fp32);
}
auto operator()(DeviceGemmPtr_& gemmPtr)
{
// Arrange
ck::gemm_util::GemmParams params;
params.M = 1024;
params.N = 1024;
params.K = 1024;
params.StrideA = 1024;
params.StrideB = 1024;
params.StrideC = 1024;
auto host_tensors = PrepareGemmTensorBF16(params);
const Tensor<BF16>& a_bf16 = std::get<0>(host_tensors);
const Tensor<BF16>& b_bf16 = std::get<1>(host_tensors);
Tensor<BF16>& c_device_bf16 = std::get<2>(host_tensors);
Tensor<float>& a_fp32 = std::get<3>(host_tensors);
Tensor<float>& b_fp32 = std::get<4>(host_tensors);
Tensor<float>& c_host_fp32 = std::get<5>(host_tensors);
Tensor<float>& c_device_fp32 = std::get<6>(host_tensors);
auto a_element_op = AElementwiseOperation{};
auto b_element_op = BElementwiseOperation{};
auto c_element_op = CElementwiseOperation{};
// use fp32 host kernel to verify bf16 device kernel
using ReferenceGemmInstance =
ck::tensor_operation::host::ReferenceGemm<float,
float,
float,
AElementwiseOperation,
BElementwiseOperation,
CElementwiseOperation>;
ck::gemm_util::RunHostGEMM<ReferenceGemmInstance>(
a_fp32, b_fp32, c_host_fp32, a_element_op, b_element_op, c_element_op);
// Act
ck::gemm_util::RunDeviceGEMM(gemmPtr,
params,
a_bf16,
b_bf16,
c_device_bf16,
a_element_op,
b_element_op,
c_element_op);
bf16_to_f32_(c_device_bf16, c_device_fp32);
// Assert
bool res = ck::utils::check_err(
c_device_fp32.mData, c_host_fp32.mData, "Error: incorrect results!", 1e-2f, 1e-3f);
std::cout << (res ? "SUCCESS" : "FAILURE") << std::endl;
return res;
};
};
} // namespace gemm_util
} // namespace ck
#endif
#ifndef GEMM_UTILS_HPP
#define GEMM_UTILS_HPP
#include "check_err.hpp"
#include "config.hpp"
#include "device.hpp"
#include "host_tensor.hpp"
#include "host_tensor_generator.hpp"
#include "reference_gemm.hpp"
#include "tensor_layout.hpp"
namespace ck {
namespace gemm_util {
struct GemmParams
{
GemmParams()
: M(1024), N(1024), K(1024), StrideA(1024), StrideB(1024), StrideC(1024), alpha(1), beta(0)
{
}
ck::index_t M;
ck::index_t N;
ck::index_t K;
ck::index_t StrideA;
ck::index_t StrideB;
ck::index_t StrideC;
float alpha;
float beta;
};
template <typename GemmInstance,
typename ADataType,
typename BDataType,
typename CDataType,
typename AElementwiseOperation,
typename BElementwiseOperation,
typename CElementwiseOperation>
void RunHostGEMM(const Tensor<ADataType>& A,
const Tensor<BDataType>& B,
Tensor<CDataType>& C,
AElementwiseOperation a_element_op,
BElementwiseOperation b_element_op,
CElementwiseOperation c_element_op)
{
auto ref_gemm = GemmInstance{};
auto ref_invoker = ref_gemm.MakeInvoker();
auto ref_argument = ref_gemm.MakeArgument(A, B, C, a_element_op, b_element_op, c_element_op);
ref_invoker.Run(ref_argument);
}
template <typename DeviceGemmPtr_,
typename ADataType,
typename BDataType,
typename CDataType,
typename AElementwiseOperation,
typename BElementwiseOperation,
typename CElementwiseOperation>
void RunDeviceGEMM(DeviceGemmPtr_& gemmPtr,
const ck::gemm_util::GemmParams& params,
const Tensor<ADataType>& A,
const Tensor<BDataType>& B,
Tensor<CDataType>& C,
AElementwiseOperation a_element_op,
BElementwiseOperation b_element_op,
CElementwiseOperation c_element_op)
{
DeviceMem a_m_k_device_buf(sizeof(ADataType) * A.mDesc.GetElementSpace());
DeviceMem b_k_n_device_buf(sizeof(BDataType) * B.mDesc.GetElementSpace());
DeviceMem c_m_n_device_buf(sizeof(CDataType) * C.mDesc.GetElementSpace());
a_m_k_device_buf.ToDevice(A.mData.data());
b_k_n_device_buf.ToDevice(B.mData.data());
auto invoker_ptr = gemmPtr->MakeInvokerPointer();
auto argument_ptr =
gemmPtr->MakeArgumentPointer(static_cast<ADataType*>(a_m_k_device_buf.GetDeviceBuffer()),
static_cast<BDataType*>(b_k_n_device_buf.GetDeviceBuffer()),
static_cast<CDataType*>(c_m_n_device_buf.GetDeviceBuffer()),
params.M,
params.N,
params.K,
params.StrideA,
params.StrideB,
params.StrideC,
a_element_op,
b_element_op,
c_element_op);
if(!gemmPtr->IsSupportedArgument(argument_ptr.get()))
{
throw std::runtime_error(
"wrong! device_gemm with the specified compilation parameters does "
"not support this GEMM problem");
}
invoker_ptr->Run(argument_ptr.get());
c_m_n_device_buf.FromDevice(C.mData.data());
}
template <typename DeviceGemmPtr_,
typename ADataType,
typename BDataType,
typename CDataType,
typename ALayout,
typename BLayout,
typename CLayout,
typename AElementwiseOperation,
typename BElementwiseOperation,
typename CElementwiseOperation>
struct TestGemm
{
auto PrepareGemmTensor(const ck::gemm_util::GemmParams& params)
{
auto f_host_tensor_descriptor =
[](std::size_t row, std::size_t col, std::size_t stride, auto layout) {
if(std::is_same<decltype(layout), ck::tensor_layout::gemm::RowMajor>::value)
{
return HostTensorDescriptor(std::vector<std::size_t>({row, col}),
std::vector<std::size_t>({stride, 1}));
}
else
{
return HostTensorDescriptor(std::vector<std::size_t>({row, col}),
std::vector<std::size_t>({1, stride}));
}
};
Tensor<ADataType> a_m_k(
f_host_tensor_descriptor(params.M, params.K, params.StrideA, ALayout{}));
Tensor<BDataType> b_k_n(
f_host_tensor_descriptor(params.K, params.N, params.StrideB, BLayout{}));
Tensor<CDataType> c_m_n_host_result(
f_host_tensor_descriptor(params.M, params.N, params.StrideC, CLayout{}));
Tensor<CDataType> c_m_n_device_result(
f_host_tensor_descriptor(params.M, params.N, params.StrideC, CLayout{}));
auto f_generate_tensor_value = [](auto& tensor, auto type) {
using dataType = decltype(type);
tensor.GenerateTensorValue(GeneratorTensor_2<dataType>{-5, 5});
};
f_generate_tensor_value(a_m_k, ADataType{});
f_generate_tensor_value(b_k_n, BDataType{});
return std::make_tuple(a_m_k, b_k_n, c_m_n_host_result, c_m_n_device_result);
}
auto operator()(DeviceGemmPtr_& gemmPtr)
{
std::cout << "ALayout = " << ALayout{}.name << ", BLayout = " << BLayout{}.name
<< ", CLayout = " << CLayout{}.name << std::endl;
std::cout << gemmPtr->GetTypeString() << std::endl;
// Arrange
ck::gemm_util::GemmParams params;
params.M = 1024;
params.N = 1024;
params.K = 1024;
params.StrideA = 1024;
params.StrideB = 1024;
params.StrideC = 1024;
auto host_tensors = PrepareGemmTensor(params);
const Tensor<ADataType>& a = std::get<0>(host_tensors);
const Tensor<BDataType>& b = std::get<1>(host_tensors);
Tensor<CDataType>& c_host = std::get<2>(host_tensors);
Tensor<CDataType>& c_device = std::get<3>(host_tensors);
auto a_element_op = AElementwiseOperation{};
auto b_element_op = BElementwiseOperation{};
auto c_element_op = CElementwiseOperation{};
using ReferenceGemmInstance =
ck::tensor_operation::host::ReferenceGemm<ADataType,
BDataType,
CDataType,
AElementwiseOperation,
BElementwiseOperation,
CElementwiseOperation>;
ck::gemm_util::RunHostGEMM<ReferenceGemmInstance>(
a, b, c_host, a_element_op, b_element_op, c_element_op);
// Act
ck::gemm_util::RunDeviceGEMM(
gemmPtr, params, a, b, c_device, a_element_op, b_element_op, c_element_op);
// Assert
bool res = false;
if(std::is_same<CDataType, float>::value)
{
res = ck::utils::check_err(c_device.mData, c_host.mData);
std::cout << (res ? "SUCCESS" : "FAILURE") << std::endl;
}
else if(std::is_same<CDataType, ck::half_t>::value)
{
res = ck::utils::check_err(c_device.mData, c_host.mData);
std::cout << (res ? "SUCCESS" : "FAILURE") << std::endl;
}
else if(std::is_same<CDataType, int8_t>::value)
{
res = ck::utils::check_err(c_device.mData, c_host.mData);
std::cout << (res ? "SUCCESS" : "FAILURE") << std::endl;
}
return res;
}
};
template <typename DeviceGemmPtr_,
typename ALayout,
typename BLayout,
typename CLayout,
typename AElementwiseOperation,
typename BElementwiseOperation,
typename CElementwiseOperation>
struct TestGemmBF16
{
using BF16 = ck::bhalf_t;
auto PrepareGemmTensorBF16(const ck::gemm_util::GemmParams& params)
{
auto f_host_tensor_descriptor =
[](std::size_t row, std::size_t col, std::size_t stride, auto layout) {
if(std::is_same<decltype(layout), ck::tensor_layout::gemm::RowMajor>::value)
{
return HostTensorDescriptor(std::vector<std::size_t>({row, col}),
std::vector<std::size_t>({stride, 1}));
}
else
{
return HostTensorDescriptor(std::vector<std::size_t>({row, col}),
std::vector<std::size_t>({1, stride}));
}
};
// use fp32 host kernel to verify bf16 device kernel
Tensor<BF16> a_m_k_bf16(
f_host_tensor_descriptor(params.M, params.K, params.StrideA, ALayout{}));
Tensor<BF16> b_k_n_bf16(
f_host_tensor_descriptor(params.K, params.N, params.StrideB, BLayout{}));
Tensor<BF16> c_m_n_device_bf16(
f_host_tensor_descriptor(params.M, params.N, params.StrideC, CLayout{}));
Tensor<float> a_m_k_fp32(
f_host_tensor_descriptor(params.M, params.K, params.StrideA, ALayout{}));
Tensor<float> b_k_n_fp32(
f_host_tensor_descriptor(params.K, params.N, params.StrideB, BLayout{}));
Tensor<float> c_m_n_host_fp32(
f_host_tensor_descriptor(params.M, params.N, params.StrideC, CLayout{}));
Tensor<float> c_m_n_device_fp32(
f_host_tensor_descriptor(params.M, params.N, params.StrideC, CLayout{}));
a_m_k_bf16.GenerateTensorValue(GeneratorTensor_3<BF16>{-0.5, 0.5});
b_k_n_bf16.GenerateTensorValue(GeneratorTensor_3<BF16>{-0.5, 0.5});
bf16_to_f32_(a_m_k_bf16, a_m_k_fp32);
bf16_to_f32_(b_k_n_bf16, b_k_n_fp32);
return std::make_tuple(a_m_k_bf16,
b_k_n_bf16,
c_m_n_device_bf16,
a_m_k_fp32,
b_k_n_fp32,
c_m_n_host_fp32,
c_m_n_device_fp32);
}
auto operator()(DeviceGemmPtr_& gemmPtr)
{
// Arrange
ck::gemm_util::GemmParams params;
params.M = 1024;
params.N = 1024;
params.K = 1024;
params.StrideA = 1024;
params.StrideB = 1024;
params.StrideC = 1024;
auto host_tensors = PrepareGemmTensorBF16(params);
const Tensor<BF16>& a_bf16 = std::get<0>(host_tensors);
const Tensor<BF16>& b_bf16 = std::get<1>(host_tensors);
Tensor<BF16>& c_device_bf16 = std::get<2>(host_tensors);
Tensor<float>& a_fp32 = std::get<3>(host_tensors);
Tensor<float>& b_fp32 = std::get<4>(host_tensors);
Tensor<float>& c_host_fp32 = std::get<5>(host_tensors);
Tensor<float>& c_device_fp32 = std::get<6>(host_tensors);
auto a_element_op = AElementwiseOperation{};
auto b_element_op = BElementwiseOperation{};
auto c_element_op = CElementwiseOperation{};
// use fp32 host kernel to verify bf16 device kernel
using ReferenceGemmInstance =
ck::tensor_operation::host::ReferenceGemm<float,
float,
float,
AElementwiseOperation,
BElementwiseOperation,
CElementwiseOperation>;
ck::gemm_util::RunHostGEMM<ReferenceGemmInstance>(
a_fp32, b_fp32, c_host_fp32, a_element_op, b_element_op, c_element_op);
// Act
ck::gemm_util::RunDeviceGEMM(gemmPtr,
params,
a_bf16,
b_bf16,
c_device_bf16,
a_element_op,
b_element_op,
c_element_op);
bf16_to_f32_(c_device_bf16, c_device_fp32);
// Assert
bool res = ck::utils::check_err(
c_device_fp32.mData, c_host_fp32.mData, "Error: incorrect results!", 1e-2f, 1e-3f);
std::cout << (res ? "SUCCESS" : "FAILURE") << std::endl;
return res;
};
};
} // namespace gemm_util
} // namespace ck
#endif
test/gemm_reduce/gemm_reduce_fp16.cpp
View file @ b134b7d6
......@@ -16,22 +16,22 @@ int main()
pass = pass &&
ck::profiler::
profile_gemm_reduce_impl<ck::half_t, ck::half_t, ck::half_t, float, Row, Row, Row>(
true, 1, false, 1, M, N, K, K, N, N);
true, 1, false, false, M, N, K, K, N, N);
pass = pass &&
ck::profiler::
profile_gemm_reduce_impl<ck::half_t, ck::half_t, ck::half_t, float, Row, Col, Row>(
true, 1, false, 1, M, N, K, K, K, N);
true, 1, false, false, M, N, K, K, K, N);
pass = pass &&
ck::profiler::
profile_gemm_reduce_impl<ck::half_t, ck::half_t, ck::half_t, float, Col, Row, Row>(
true, 1, false, 1, M, N, K, M, N, N);
true, 1, false, false, M, N, K, M, N, N);
pass = pass &&
ck::profiler::
profile_gemm_reduce_impl<ck::half_t, ck::half_t, ck::half_t, float, Col, Col, Row>(
true, 1, false, 1, M, N, K, M, K, N);
true, 1, false, false, M, N, K, M, K, N);
if(pass)
{
......
test/gemm_split_k/gemm_split_k.cpp
View file @ b134b7d6
......@@ -45,7 +45,7 @@ static bool check_out(const Tensor<T>& ref, const Tensor<T>& result)
{
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]));
if(max_diff < diff)
......@@ -187,9 +187,10 @@ int test_gemm(const gemmArgs& args)
if(gemm_ptr->IsSupportedArgument(argument_ptr.get()))
{
invoker_ptr->Run(argument_ptr.get(), 0);
invoker_ptr->Run(argument_ptr.get());
c_device_buf.FromDevice(c_m_n_device_result.mData.data());
if(!check_out(c_m_n_host_result, c_m_n_device_result))
{
success = false;
......
test/grouped_gemm/grouped_gemm_fp16.cpp
View file @ b134b7d6
......@@ -104,7 +104,7 @@ bool TestGroupedGemm(DeviceGroupedGemmPtr_& groupedGemmPtr)
b_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(
gemm_shapes[i].M, gemm_shapes[i].K, gemm_shapes[i].StrideA, ALayout{})));
......@@ -119,7 +119,7 @@ bool TestGroupedGemm(DeviceGroupedGemmPtr_& groupedGemmPtr)
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(
std::make_unique<DeviceMem>(sizeof(ADataType) * a_tensors[i].mDesc.GetElementSize()));
......@@ -147,7 +147,7 @@ bool TestGroupedGemm(DeviceGroupedGemmPtr_& groupedGemmPtr)
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());
......
test/reduce/reduce_no_index.cpp
View file @ b134b7d6
......@@ -460,7 +460,7 @@ class SimpleAppArgs
int processArgs(int argc, char* argv[])
{
unsigned int ch;
int ch;
while(1)
{
......
test/reduce/reduce_util.hpp
View file @ b134b7d6
......@@ -9,7 +9,7 @@ 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)
for(std::size_t i = 0; i < src.mData.size(); ++i)
dst.mData[i] = type_convert<float>(src.mData[i]);
}
......
test/reduce/reduce_with_index.cpp
View file @ b134b7d6
......@@ -463,7 +463,7 @@ class SimpleAppArgs
int processArgs(int argc, char* argv[])
{
unsigned int ch;
int ch;
while(1)
{
......
test/reference_conv_fwd/CMakeLists.txt
View file @ b134b7d6
add_test_executable(test_reference_conv_fwd reference_conv_fwd.cpp)
target_link_libraries(test_reference_conv_fwd PRIVATE host_tensor conv_fwd_util)
add_gtest_executable(test_reference_conv_fwd reference_conv_fwd.cpp)
target_link_libraries(test_reference_conv_fwd PRIVATE host_tensor conv_util)
test/reference_conv_fwd/reference_conv_fwd.cpp
View file @ b134b7d6
......@@ -4,10 +4,11 @@
#include <numeric>
#include <type_traits>
#include <vector>
#include "gtest/gtest.h"
#include "check_err.hpp"
#include "config.hpp"
#include "conv_fwd_util.hpp"
#include "conv_util.hpp"
#include "element_wise_operation.hpp"
#include "fill.hpp"
#include "host_tensor.hpp"
......@@ -33,21 +34,21 @@ run_reference_convolution_forward(const ck::utils::conv::ConvParams& params,
const FillInputOp& fill_input_op = FillInputOp{},
const FillWeightsOp& fill_weights_op = FillWeightsOp{0.5f})
{
std::vector<std::size_t> input_dims{static_cast<std::size_t>(params.N),
static_cast<std::size_t>(params.C)};
std::vector<std::size_t> input_dims{static_cast<std::size_t>(params.N_),
static_cast<std::size_t>(params.C_)};
input_dims.insert(std::end(input_dims),
std::begin(params.input_spatial_lengths),
std::end(params.input_spatial_lengths));
std::begin(params.input_spatial_lengths_),
std::end(params.input_spatial_lengths_));
std::vector<std::size_t> filter_dims{static_cast<std::size_t>(params.K),
static_cast<std::size_t>(params.C)};
std::vector<std::size_t> filter_dims{static_cast<std::size_t>(params.K_),
static_cast<std::size_t>(params.C_)};
filter_dims.insert(std::end(filter_dims),
std::begin(params.filter_spatial_lengths),
std::end(params.filter_spatial_lengths));
std::begin(params.filter_spatial_lengths_),
std::end(params.filter_spatial_lengths_));
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),
static_cast<std::size_t>(params.K)};
std::vector<std::size_t> output_dims{static_cast<std::size_t>(params.N_),
static_cast<std::size_t>(params.K_)};
output_dims.insert(std::end(output_dims),
std::begin(output_spatial_lengths),
std::end(output_spatial_lengths));
......@@ -73,32 +74,32 @@ run_reference_convolution_forward(const ck::utils::conv::ConvParams& params,
auto ref_argument = ref_conv.MakeArgument(input,
weights,
host_output,
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_,
InElementOp{},
WeiElementOp{},
OutElementOp{});
ref_invoker.Run(ref_argument);
// std::cout <<"output: " << host_output.mDesc << std::endl << host_output.mData << std::endl;
return host_output;
}
bool test_conv2d_nhwc()
} // anonymous namespace
TEST(ReferenceConvolutionFWD, Conv2DNHWC)
{
bool res{true};
ck::utils::conv::ConvParams params;
params.N = 1;
params.K = 1;
params.C = 2;
params.filter_spatial_lengths = std::vector<ck::index_t>{3, 3};
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_dilations = std::vector<ck::index_t>{1, 1};
params.input_left_pads = std::vector<ck::index_t>{0, 0};
params.input_right_pads = std::vector<ck::index_t>{0, 0};
params.N_ = 1;
params.K_ = 1;
params.C_ = 2;
params.filter_spatial_lengths_ = std::vector<ck::index_t>{3, 3};
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_dilations_ = std::vector<ck::index_t>{1, 1};
params.input_left_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);
std::vector<std::size_t> ref_dims{1, 1, 4, 4};
......@@ -118,51 +119,50 @@ bool test_conv2d_nhwc()
472.5,
490.5,
508.5};
res = res && ck::utils::check_err(out_tensor.mDesc.GetLengths(),
ref_dims,
"Error: wrong output tensor dimensions!");
res = res && ck::utils::check_err(out_tensor.mData, ref_data, "Error: incorrect results!");
EXPECT_TRUE(ck::utils::check_err(
out_tensor.mDesc.GetLengths(), ref_dims, "Error: wrong output tensor dimensions!"));
EXPECT_TRUE(ck::utils::check_err(out_tensor.mData, ref_data, "Error: incorrect results!"));
}
params.N = 1;
params.K = 2;
params.C = 2;
params.filter_spatial_lengths = std::vector<ck::index_t>{3, 3};
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_dilations = std::vector<ck::index_t>{2, 2};
params.input_left_pads = std::vector<ck::index_t>{1, 1};
params.input_right_pads = std::vector<ck::index_t>{1, 1};
TEST(ReferenceConvolutionFWD, Conv2DNHWCStridesDilationsPadding)
{
ck::utils::conv::ConvParams params;
params.N_ = 1;
params.K_ = 2;
params.C_ = 2;
params.filter_spatial_lengths_ = std::vector<ck::index_t>{3, 3};
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_dilations_ = std::vector<ck::index_t>{2, 2};
params.input_left_pads_ = std::vector<ck::index_t>{1, 1};
params.input_right_pads_ = std::vector<ck::index_t>{1, 1};
out_tensor = run_reference_convolution_forward<2>(params);
ref_dims = std::vector<std::size_t>{1, 2, 5, 5};
ref_data = std::vector<float>{
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<float> ref_data{
210., 210., 327., 327., 351., 351., 375., 375., 399., 399.,
459., 459., 706.5, 706.5, 742.5, 742.5, 778.5, 778.5, 814.5, 814.5,
747., 747., 1138.5, 1138.5, 1174.5, 1174.5, 1210.5, 1210.5, 1246.5, 1246.5,
1035., 1035., 1570.5, 1570.5, 1606.5, 1606.5, 1642.5, 1642.5, 1678.5, 1678.5,
1323., 1323., 2002.5, 2002.5, 2038.5, 2038.5, 2074.5, 2074.5, 2110.5, 2110.5};
res = res && ck::utils::check_err(out_tensor.mDesc.GetLengths(),
ref_dims,
"Error: wrong output tensor dimensions!");
res = res && ck::utils::check_err(out_tensor.mData, ref_data, "Error: incorrect results!");
return res;
EXPECT_TRUE(ck::utils::check_err(
out_tensor.mDesc.GetLengths(), ref_dims, "Error: wrong output tensor dimensions!"));
EXPECT_TRUE(ck::utils::check_err(out_tensor.mData, ref_data, "Error: incorrect results!"));
}
bool test_conv1d_nwc()
TEST(ReferenceConvolutionFWD, Conv1DNWC)
{
bool res{true};
ck::utils::conv::ConvParams params;
params.num_dim_spatial = 1;
params.N = 1;
params.K = 1;
params.C = 2;
params.filter_spatial_lengths = std::vector<ck::index_t>{3};
params.input_spatial_lengths = std::vector<ck::index_t>{6};
params.conv_filter_strides = 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_right_pads = std::vector<ck::index_t>{0};
params.num_dim_spatial_ = 1;
params.N_ = 1;
params.K_ = 1;
params.C_ = 2;
params.filter_spatial_lengths_ = std::vector<ck::index_t>{3};
params.input_spatial_lengths_ = std::vector<ck::index_t>{6};
params.conv_filter_strides_ = 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_right_pads_ = std::vector<ck::index_t>{0};
auto out_tensor =
run_reference_convolution_forward<1,
......@@ -174,46 +174,53 @@ bool test_conv1d_nwc()
ck::tensor_layout::convolution::NWK>(params);
std::vector<std::size_t> ref_dims{1, 1, 4};
std::vector<float> ref_data{7.5, 13.5, 19.5, 25.5};
res = res && ck::utils::check_err(out_tensor.mDesc.GetLengths(),
ref_dims,
"Error: wrong output tensor dimensions!");
res = res && ck::utils::check_err(out_tensor.mData, ref_data, "Error: incorrect results!");
EXPECT_TRUE(ck::utils::check_err(
out_tensor.mDesc.GetLengths(), ref_dims, "Error: wrong output tensor dimensions!"));
EXPECT_TRUE(ck::utils::check_err(out_tensor.mData, ref_data, "Error: incorrect results!"));
}
params.num_dim_spatial = 1;
params.N = 1;
params.K = 2;
params.C = 2;
params.filter_spatial_lengths = std::vector<ck::index_t>{3};
params.input_spatial_lengths = std::vector<ck::index_t>{12};
params.conv_filter_strides = 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_right_pads = std::vector<ck::index_t>{1};
TEST(ReferenceConvolutionFWD, Conv1DNWCStridesDilationsPadding)
{
ck::utils::conv::ConvParams params;
params.num_dim_spatial_ = 1;
params.N_ = 1;
params.K_ = 2;
params.C_ = 2;
params.filter_spatial_lengths_ = std::vector<ck::index_t>{3};
params.input_spatial_lengths_ = std::vector<ck::index_t>{12};
params.conv_filter_strides_ = 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_right_pads_ = std::vector<ck::index_t>{1};
out_tensor = run_reference_convolution_forward<1,
float,
float,
float,
ck::tensor_layout::convolution::NWC,
ck::tensor_layout::convolution::KXC,
ck::tensor_layout::convolution::NWK>(params);
ref_dims = std::vector<std::size_t>{1, 2, 5};
ref_data = std::vector<float>{9., 9., 19.5, 19.5, 31.5, 31.5, 43.5, 43.5, 55.5, 55.5};
res = res && ck::utils::check_err(out_tensor.mDesc.GetLengths(),
ref_dims,
"Error: wrong output tensor dimensions!");
res = res && ck::utils::check_err(out_tensor.mData, ref_data, "Error: incorrect results!");
auto out_tensor =
run_reference_convolution_forward<1,
float,
float,
float,
ck::tensor_layout::convolution::NWC,
ck::tensor_layout::convolution::KXC,
ck::tensor_layout::convolution::NWK>(params);
std::vector<std::size_t> ref_dims{1, 2, 5};
std::vector<float> ref_data{9., 9., 19.5, 19.5, 31.5, 31.5, 43.5, 43.5, 55.5, 55.5};
EXPECT_TRUE(ck::utils::check_err(
out_tensor.mDesc.GetLengths(), ref_dims, "Error: wrong output tensor dimensions!"));
EXPECT_TRUE(ck::utils::check_err(out_tensor.mData, ref_data, "Error: incorrect results!"));
}
params.num_dim_spatial = 1;
params.N = 2;
params.K = 16;
params.C = 4;
params.filter_spatial_lengths = std::vector<ck::index_t>{3};
params.input_spatial_lengths = std::vector<ck::index_t>{16};
params.conv_filter_strides = 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_right_pads = std::vector<ck::index_t>{1};
TEST(ReferenceConvolutionFWD, Conv1DNWCSameOutputSize)
{
ck::utils::conv::ConvParams params;
params.num_dim_spatial_ = 1;
params.N_ = 2;
params.K_ = 16;
params.C_ = 4;
params.filter_spatial_lengths_ = std::vector<ck::index_t>{3};
params.input_spatial_lengths_ = std::vector<ck::index_t>{16};
params.conv_filter_strides_ = 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_right_pads_ = std::vector<ck::index_t>{1};
auto out_tensor2 = run_reference_convolution_forward<1,
float,
......@@ -224,8 +231,8 @@ bool test_conv1d_nwc()
ck::tensor_layout::convolution::NWK>(
params, ck::utils::FillMonotonicSeq<float>{0.f, 0.1f});
ref_dims = std::vector<std::size_t>{2, 16, 16};
ref_data = std::vector<float>{
std::vector<std::size_t> ref_dims{2, 16, 16};
std::vector<float> ref_data{
1.4, 1.4, 1.4, 1.4, 1.4, 1.4, 1.4, 1.4,
1.4, 1.4, 1.4, 1.4, 1.4, 1.4, 1.4, 1.4,
3.3, 3.3, 3.3, 3.3, 3.3, 3.3, 3.3, 3.3,
......@@ -290,28 +297,24 @@ bool test_conv1d_nwc()
72.9, 72.9, 72.9, 72.9, 72.9, 72.9, 72.9, 72.9,
49.4, 49.4, 49.4, 49.4, 49.4, 49.4, 49.4, 49.4,
49.4, 49.4, 49.4, 49.4, 49.4, 49.4, 49.4, 49.4};
res = res && ck::utils::check_err(out_tensor2.mDesc.GetLengths(),
ref_dims,
"Error: wrong output tensor dimensions!");
res = res && ck::utils::check_err(out_tensor2.mData, ref_data, "Error: incorrect results!");
return res;
EXPECT_TRUE(ck::utils::check_err(
out_tensor2.mDesc.GetLengths(), ref_dims, "Error: wrong output tensor dimensions!"));
EXPECT_TRUE(ck::utils::check_err(out_tensor2.mData, ref_data, "Error: incorrect results!"));
}
bool test_conv3d_ncdhw()
TEST(ReferenceConvolutionFWD, Conv3DNCDHW)
{
bool res{true};
ck::utils::conv::ConvParams params;
params.num_dim_spatial = 3;
params.N = 1;
params.K = 1;
params.C = 2;
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.conv_filter_strides = 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_right_pads = std::vector<ck::index_t>{0, 0, 0};
params.num_dim_spatial_ = 3;
params.N_ = 1;
params.K_ = 1;
params.C_ = 2;
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.conv_filter_strides_ = 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_right_pads_ = std::vector<ck::index_t>{0, 0, 0};
auto out_tensor = run_reference_convolution_forward<3,
float,
......@@ -331,32 +334,37 @@ bool test_conv3d_ncdhw()
634.5, 637.2, 639.9, 642.60004, 650.7, 653.4, 656.10004, 658.8,
699.3, 702., 704.7, 707.4, 715.5, 718.2, 720.9, 723.60004,
731.7, 734.4001, 737.10004, 739.8, 747.9001, 750.60004, 753.3, 756.};
res = res && ck::utils::check_err(out_tensor.mDesc.GetLengths(),
ref_dims,
"Error [case 1]: wrong output tensor dimensions!");
res = res &&
ck::utils::check_err(out_tensor.mData, ref_data, "Error [case 1]: incorrect results!");
EXPECT_TRUE(ck::utils::check_err(out_tensor.mDesc.GetLengths(),
ref_dims,
"Error [case 1]: wrong output tensor dimensions!"));
EXPECT_TRUE(
ck::utils::check_err(out_tensor.mData, ref_data, "Error [case 1]: incorrect results!"));
}
params.N = 1;
params.K = 2;
params.C = 2;
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.conv_filter_strides = std::vector<ck::index_t>{3, 3, 3};
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_right_pads = std::vector<ck::index_t>{0, 0, 0};
TEST(ReferenceConvolutionFWD, Conv3DNCDHWStridesDilations)
{
ck::utils::conv::ConvParams params;
params.num_dim_spatial_ = 3;
params.N_ = 1;
params.K_ = 2;
params.C_ = 2;
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.conv_filter_strides_ = std::vector<ck::index_t>{3, 3, 3};
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_right_pads_ = std::vector<ck::index_t>{0, 0, 0};
out_tensor = run_reference_convolution_forward<3,
float,
float,
float,
ck::tensor_layout::convolution::NCDHW,
ck::tensor_layout::convolution::KCZYX,
ck::tensor_layout::convolution::NKDHW>(
auto out_tensor = run_reference_convolution_forward<3,
float,
float,
float,
ck::tensor_layout::convolution::NCDHW,
ck::tensor_layout::convolution::KCZYX,
ck::tensor_layout::convolution::NKDHW>(
params, ck::utils::FillMonotonicSeq<float>{0.f, 0.1f});
ref_dims = std::vector<std::size_t>{1, 2, 4, 4, 4};
ref_data = std::vector<float>{
std::vector<std::size_t> ref_dims{1, 2, 4, 4, 4};
std::vector<float> ref_data{
2756.7002, 2764.7998, 2772.9001, 2781., 2853.9001, 2862., 2870.1, 2878.2002,
2951.1, 2959.2002, 2967.2998, 2975.4001, 3048.2998, 3056.4001, 3064.5, 3072.6,
3923.1, 3931.2, 3939.2998, 3947.4, 4020.2998, 4028.4001, 4036.5002, 4044.5999,
......@@ -373,26 +381,9 @@ bool test_conv3d_ncdhw()
5283.9004, 5292., 5300.0996, 5308.2, 5381.0996, 5389.2, 5397.3, 5405.4004,
6255.9004, 6264.0005, 6272.1, 6280.2, 6353.1, 6361.2, 6369.301, 6377.4,
6450.301, 6458.4, 6466.5, 6474.6, 6547.5, 6555.6, 6563.699, 6571.801};
res = res && ck::utils::check_err(out_tensor.mDesc.GetLengths(),
ref_dims,
"Error [case 2]: wrong output tensor dimensions!");
res =
res && ck::utils::check_err(
out_tensor.mData, ref_data, "Error [case 2]: incorrect results!", 1e-4f, 1e-6f);
return res;
}
} // anonymous namespace
int main(void)
{
bool res{true};
res = test_conv2d_nhwc();
std::cout << "test_conv2d_nhwc ..... " << (res ? "SUCCESS" : "FAILURE") << std::endl;
res = test_conv1d_nwc();
std::cout << "TestConv1DNHWC ..... " << (res ? "SUCCESS" : "FAILURE") << std::endl;
res = test_conv3d_ncdhw();
std::cout << "test_conv3d_ncdhw ..... " << (res ? "SUCCESS" : "FAILURE") << std::endl;
return res ? 0 : 1;
EXPECT_TRUE(ck::utils::check_err(out_tensor.mDesc.GetLengths(),
ref_dims,
"Error [case 2]: wrong output tensor dimensions!"));
EXPECT_TRUE(ck::utils::check_err(
out_tensor.mData, ref_data, "Error [case 2]: incorrect results!", 1e-4f, 1e-6f));
}
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