Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

New fill kernel and validation tolerances #251

Merged
merged 4 commits into from
Aug 1, 2024
Merged
Show file tree
Hide file tree
Changes from 1 commit
Commits
File filter

Filter by extension

Filter by extension

Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
96 changes: 96 additions & 0 deletions library/src/data_types.cpp
Original file line number Diff line number Diff line change
Expand Up @@ -254,6 +254,102 @@ namespace hiptensor
return;
}
}

std::string computeTypeToString(hiptensorComputeType_t computeType)
{
if(computeType == HIPTENSOR_COMPUTE_16BF)
{
return "HIPTENSOR_COMPUTE_16BF";
}
else if(computeType == HIPTENSOR_COMPUTE_16F)
{
return "HIPTENSOR_COMPUTE_16F";
}
else if(computeType == HIPTENSOR_COMPUTE_32F)
{
return "HIPTENSOR_COMPUTE_32F";
}
else if(computeType == HIPTENSOR_COMPUTE_64F)
{
return "HIPTENSOR_COMPUTE_64F";
}
else if(computeType == HIPTENSOR_COMPUTE_8I)
{
return "HIPTENSOR_COMPUTE_8I";
}
else if(computeType == HIPTENSOR_COMPUTE_8U)
{
return "HIPTENSOR_COMPUTE_8U";
}
else if(computeType == HIPTENSOR_COMPUTE_32I)
{
return "HIPTENSOR_COMPUTE_32I";
}
else if(computeType == HIPTENSOR_COMPUTE_32U)
{
return "HIPTENSOR_COMPUTE_32U";
}
else if(computeType == HIPTENSOR_COMPUTE_C32F)
{
return "HIPTENSOR_COMPUTE_C32F";
}
else if(computeType == HIPTENSOR_COMPUTE_C64F)
{
return "HIPTENSOR_COMPUTE_C64F";
}
else
{
return "HIPTENSOR_COMPUTE_NONE";
}
}

std::string hipTypeToString(hipDataType hipType)
{
if(hipType == HIP_R_16BF)
{
return "HIP_R_16BF";
}
else if(hipType == HIP_R_16F)
{
return "HIP_R_16F";
}
else if(hipType == HIP_R_32F)
{
return "HIP_R_32F";
}
else if(hipType == HIP_R_64F)
{
return "HIP_R_64F";
}
else if(hipType == HIP_R_8I)
{
return "HIP_R_8I";
}
else if(hipType == HIP_R_8U)
{
return "HIP_R_8U";
}
else if(hipType == HIP_R_32I)
{
return "HIP_R_32I";
}
else if(hipType == HIP_R_32U)
{
return "HIP_R_32U";
}
else if(hipType == HIP_C_32F)
{
return "HIP_C_32F";
}
else if(hipType == HIP_C_64F)
{
return "HIP_C_64F";
}
else
{
return "HIP_TYPE_NONE";
}
}
} // namespace hiptensor

bool operator==(hipDataType hipType, hiptensorComputeType_t computeType)
Expand Down
3 changes: 3 additions & 0 deletions library/src/include/data_types.hpp
Original file line number Diff line number Diff line change
Expand Up @@ -107,6 +107,9 @@ namespace hiptensor
T readVal(void const* value, hiptensorComputeType_t id);

void writeVal(void const* addr, hiptensorComputeType_t id, ScalarData value);

std::string computeTypeToString(hiptensorComputeType_t computeType);
std::string hipTypeToString(hipDataType hipType);
} // namespace hiptensor

bool operator==(hipDataType hipType, hiptensorComputeType_t computeType);
Expand Down
34 changes: 30 additions & 4 deletions test/01_contraction/contraction_test.cpp
Original file line number Diff line number Diff line change
Expand Up @@ -656,13 +656,36 @@ namespace hiptensor
auto reference = resource->allocDevice(sizeD);
resource->copyData(reference, resource->hostD(), sizeD);

// Compute tolerance based on compute type
auto dimension = a_ms_ks.mLengths.size() / 2;
auto nelems_k = std::accumulate(a_ms_ks.mLengths.begin() + dimension,
a_ms_ks.mLengths.end(),
size_t{1},
std::multiplies<size_t>());

auto eps = getEpsilon(computeType);
CongMa13 marked this conversation as resolved.
Show resolved Hide resolved
double tolerance = 2 * nelems_k * eps;

// use the same default tolerance value as CK
if (computeType == HIPTENSOR_COMPUTE_16BF || DDataType == HIP_R_16BF)
{
const double epsilon = std::pow(2, -7);
tolerance += epsilon * 2;
}
else if (computeType == HIPTENSOR_COMPUTE_16F || DDataType == HIP_R_16F)
{
const double epsilon = std::pow(2, -10);
tolerance += epsilon * 2;
}

if(DDataType == HIP_R_16F)
{
std::tie(mValidationResult, mMaxRelativeError)
= compareEqualLaunchKernel<_Float16>((_Float16*)resource->deviceD().get(),
(_Float16*)reference.get(),
elementsCD,
computeType);
computeType,
tolerance);
}
else if(DDataType == HIP_R_16BF)
{
Expand All @@ -671,23 +694,26 @@ namespace hiptensor
(hip_bfloat16*)resource->deviceD().get(),
(hip_bfloat16*)reference.get(),
elementsCD,
computeType);
computeType,
tolerance);
}
else if(DDataType == HIP_R_32F || DDataType == HIP_C_32F)
{
std::tie(mValidationResult, mMaxRelativeError)
= compareEqualLaunchKernel<float>((float*)resource->deviceD().get(),
(float*)reference.get(),
elementsCD,
computeType);
computeType,
tolerance);
}
else if(DDataType == HIP_R_64F || DDataType == HIP_C_64F)
{
std::tie(mValidationResult, mMaxRelativeError)
= compareEqualLaunchKernel<double>((double*)resource->deviceD().get(),
(double*)reference.get(),
elementsCD,
computeType);
computeType,
tolerance);
}

EXPECT_TRUE(mValidationResult) << "Max relative error: " << mMaxRelativeError;
Expand Down
24 changes: 19 additions & 5 deletions test/device/common.hpp
Original file line number Diff line number Diff line change
Expand Up @@ -65,28 +65,42 @@ __global__ static void
}
}

__device__ inline unsigned pcg_hash(unsigned input)
{
unsigned state = input * 747796405u + 2891336453u;
unsigned word = ((state >> ((state >> 28u) + 4u)) ^ state) * 277803737u;
return (word >> 22u) ^ word;
}

// gen random float in range [-range, range)
template <unsigned range = 1>
__device__ inline float gen_random_float(unsigned input)
{
return (static_cast<float>(pcg_hash(input)) / static_cast<float>(UINT_MAX) - 0.5f)
* static_cast<float>(range) * 2;
}

// fill kernel for 'elementSize' elements
template <typename DataType>
__global__ void fillKernel(DataType* data, uint32_t elementSize, uint32_t seed)
__global__ void fillKernel(DataType* data, uint32_t elementSize)
{
uint32_t index = (blockIdx.x * blockDim.x + threadIdx.x);

if(index < elementSize)
{
// Input values scaled by 10, Doing UnarySquare Operation for tensors(16F) may cause overflow.
if constexpr(std::is_same_v<DataType, hipFloatComplex>)
{
auto value = (float(index / float(RAND_MAX) - 0.5) * 10) / elementSize;
auto value = gen_random_float(index);
data[index] = make_hipFloatComplex(value, value);
}
else if constexpr(std::is_same_v<DataType, hipDoubleComplex>)
{
auto value = (double(index / double(RAND_MAX) - 0.5) * 10) / elementSize;
auto value = static_cast<double>(gen_random_float(index));
data[index] = make_hipDoubleComplex(value, value);
}
else
{
auto value = (DataType(index / double(RAND_MAX) - 0.5) * 10) / elementSize;
auto value = gen_random_float(index);
data[index] = static_cast<DataType>(value);
}
}
Expand Down
52 changes: 45 additions & 7 deletions test/utils.hpp
Original file line number Diff line number Diff line change
Expand Up @@ -174,8 +174,7 @@ __host__ static inline void fillLaunchKernel(DataType* data, uint32_t elementSiz
0,
0,
data,
elementSize,
static_cast<uint32_t>(std::time(nullptr)));
elementSize);
}

// fill kernel wrapper for 'elementSize' elements with a specific value
Expand All @@ -194,7 +193,7 @@ std::pair<bool, double> compareEqual(DDataType const* deviceD,
DDataType const* hostD,
std::size_t elementsD,
hiptensorComputeType_t computeType,
double tolerance = 100.0)
double tolerance = 0.0)
{
bool retval = true;
double max_relative_error = 0.0;
Expand Down Expand Up @@ -245,7 +244,25 @@ std::pair<bool, double> compareEqual(DDataType const* deviceD,
}
}

auto eps = getEpsilon(computeType);
if(tolerance == 0.0)
{
// use the same default tolerance value as CK
if (computeType == HIPTENSOR_COMPUTE_16BF || std::is_same_v<DDataType, hiptensor::bfloat16_t>)
{
const double epsilon = std::pow(2, -7);
tolerance = epsilon * 2;
}
else if (computeType == HIPTENSOR_COMPUTE_16F || std::is_same_v<DDataType, hiptensor::float16_t>)
{
const double epsilon = std::pow(2, -10);
tolerance = epsilon * 2;
}
else
{
tolerance = 1e-5;
}
}

if(isInf)
{
retval = false;
Expand All @@ -256,7 +273,7 @@ std::pair<bool, double> compareEqual(DDataType const* deviceD,
retval = false;
max_relative_error = std::numeric_limits<DDataType>::signaling_NaN();
}
else if(max_relative_error > (eps * tolerance))
else if(max_relative_error > tolerance)
{
retval = false;
}
Expand All @@ -269,13 +286,14 @@ std::pair<bool, double> compareEqualLaunchKernel(DDataType* deviceD,
DDataType* hostD,
std::size_t elementsD,
hiptensorComputeType_t computeType,
double tolerance = 100.0)
double tolerance = 0.0)
{
auto blockDim = dim3(1024, 1, 1);
auto gridDim = dim3(ceilDiv(elementsD, blockDim.x), 1, 1);

double* d_relativeError;
double maxRelativeError;

CHECK_HIP_ERROR(hipMalloc(&d_relativeError, elementsD * sizeof(double)));

hipEvent_t syncEvent;
Expand Down Expand Up @@ -332,12 +350,32 @@ std::pair<bool, double> compareEqualLaunchKernel(DDataType* deviceD,
= [](DDataType const& val) { return static_cast<double>(static_cast<float>(val)); };

auto eps = getEpsilon(computeType);

if(tolerance == 0.0)
{
// use the same default tolerance value as CK
if (computeType == HIPTENSOR_COMPUTE_16BF || std::is_same_v<DDataType, hiptensor::bfloat16_t>)
{
const double epsilon = std::pow(2, -7);
tolerance = epsilon * 2;
}
else if (computeType == HIPTENSOR_COMPUTE_16F || std::is_same_v<DDataType, hiptensor::float16_t>)
{
const double epsilon = std::pow(2, -10);
tolerance = epsilon * 2;
}
else
{
tolerance = 1e-5;
}
}

if(isNaN)
{
retval = false;
maxRelativeError = std::numeric_limits<DDataType>::signaling_NaN();
}
else if(maxRelativeError > (eps * tolerance))
else if(maxRelativeError > (tolerance))
{
retval = false;
}
Expand Down
Loading