ROCm/composable_kernel
1
0
Fork 0
mirror of https://github.com/ROCm/composable_kernel.git synced 2026-05-20 04:49:54 +00:00
Code Issues Packages Projects Releases Wiki Activity

fixed multiple definition issue of bfp16/fp32 conversion function when building ckProfiler (#51)

Browse Source 
* fixed bfloat16 issues

* refactor type_convert

Co-authored-by: Chao Liu <chao.liu2@amd.com>

[ROCm/composable_kernel commit: 0a66c54e95]
This commit is contained in:
zjing14
2021-11-16 15:44:17 -06:00
committed by GitHub
parent ea6fa92eea
commit 1e7102575b
18 changed files with 157 additions and 337 deletions
Download Patch File Download Diff File Expand all files Collapse all files

58
host/host_tensor/include/host_tensor.hpp
View File

@@ -299,53 +299,41 @@ HostTensorDescriptor::HostTensorDescriptor(std::vector<X> lens, std::vector<Y> s
void ostream_HostTensorDescriptor(const HostTensorDescriptor& desc, std::ostream& os = std::cout);
float bf16_to_f32_(ushort src_val);
template <typename T>
void check_error(const Tensor<T>& ref, const Tensor<T>& result)
{
float error = 0;
float max_diff = -1;
float ref_value = 0, result_value = 0;
for(int i = 0; i < ref.mData.size(); ++i)
if constexpr(std::is_same<ushort, T>::value)
{
error += 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)
for(int i = 0; i < ref.mData.size(); ++i)
{
max_diff = diff;
ref_value = ref.mData[i];
result_value = result.mData[i];
error += std::abs(bf16_to_f32_(ref.mData[i]) - bf16_to_f32_(result.mData[i]));
float diff = std::abs(bf16_to_f32_(ref.mData[i]) - bf16_to_f32_(result.mData[i]));
if(max_diff < diff)
{
max_diff = diff;
ref_value = bf16_to_f32_(ref.mData[i]);
result_value = bf16_to_f32_(result.mData[i]);
}
}
}
std::cout << "error: " << error << std::endl;
std::cout << "max_diff: " << max_diff << ", " << ref_value << ", " << result_value << std::endl;
}
__host__ __device__ float bf16_to_f32(ushort src_val)
{
union
else
{
uint32_t int32;
float fp32;
} u = {uint32_t(src_val) << 16};
return u.fp32;
}
template <>
void check_error<ushort>(const Tensor<ushort>& ref, const Tensor<ushort>& result)
{
float error = 0;
float max_diff = -1;
float ref_value = 0, result_value = 0;
for(int i = 0; i < ref.mData.size(); ++i)
{
error += std::abs(bf16_to_f32(ref.mData[i]) - bf16_to_f32(result.mData[i]));
float diff = std::abs(bf16_to_f32(ref.mData[i]) - bf16_to_f32(result.mData[i]));
if(max_diff < diff)
for(int i = 0; i < ref.mData.size(); ++i)
{
max_diff = diff;
ref_value = bf16_to_f32(ref.mData[i]);
result_value = bf16_to_f32(result.mData[i]);
error += 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)
{
max_diff = diff;
ref_value = ref.mData[i];
result_value = result.mData[i];
}
}
}