mindspore-Ecosystem
/
mindspore
21
0
Fork 27
Code Issues Pull Requests 6 Projects Releases Wiki Activity

fix Lerp low precision in fp16 cpu backend .

This commit is contained in:
z00512249 2022-05-30 10:16:40 +08:00
parent 5f2388d6d1
commit 17cea6d8a3
2 changed files with 13 additions and 31 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

42
mindspore/ccsrc/plugin/device/cpu/kernel/lerp_cpu_kernel.cc
View File

@ -39,41 +39,33 @@ bool LerpCpuKernelMod::Init(const BaseOperatorPtr &base_operator, const std::vec
int LerpCpuKernelMod::Resize(const BaseOperatorPtr &base_operator, const std::vector<KernelTensorPtr> &inputs,
const std::vector<KernelTensorPtr> &outputs,
const std::map<uint32_t, tensor::TensorPtr> &) {
ResetResource();
if (auto ret = KernelMod::Resize(base_operator, inputs, outputs); ret != KRET_OK) {
return ret;
}
auto start_shape = inputs.at(kIndex0)->GetShapeVector();
auto end_shape = inputs.at(kIndex1)->GetShapeVector();
auto weight_shape = inputs.at(kIndex2)->GetShapeVector();
auto output_shape = outputs.at(kIndex0)->GetShapeVector();
start_shape_.clear();
(void)std::transform(start_shape.begin(), start_shape.end(), std::back_inserter(start_shape_), LongToSize);
end_shape_.clear();
auto end_shape = inputs.at(kIndex1)->GetShapeVector();
(void)std::transform(end_shape.begin(), end_shape.end(), std::back_inserter(end_shape_), LongToSize);
weight_shape_.clear();
auto weight_shape = inputs.at(kIndex2)->GetShapeVector();
output_shape_.clear();
(void)std::transform(weight_shape.begin(), weight_shape.end(), std::back_inserter(weight_shape_), LongToSize);
auto output_shape = outputs.at(kIndex0)->GetShapeVector();
(void)std::transform(output_shape.begin(), output_shape.end(), std::back_inserter(output_shape_), LongToSize);
output_size_ = std::accumulate(output_shape_.begin(), output_shape_.end(), 1, std::multiplies<size_t>());
return KRET_OK;
}
void LerpCpuKernelMod::ResetResource() noexcept {
output_size_ = 0;
start_shape_.clear();
end_shape_.clear();
weight_shape_.clear();
output_shape_.clear();
input_size_list_.clear();
output_size_list_.clear();
workspace_size_list_.clear();
}
template <typename T>
bool LerpCpuKernelMod::LaunchKernel(const std::vector<kernel::AddressPtr> &inputs, const std::vector<AddressPtr> &,
const std::vector<kernel::AddressPtr> &outputs) {
auto input_start = GetDeviceAddress<T>(inputs, kIndex0);
auto input_end = GetDeviceAddress<T>(inputs, kIndex1);
auto input_weight = GetDeviceAddress<T>(inputs, kIndex2);
auto output = GetDeviceAddress<T>(outputs, kIndex0);
if (start_shape_ == end_shape_ && start_shape_ == weight_shape_) {
auto *input_start = reinterpret_cast<T *>(inputs.at(kIndex0)->addr);
auto *input_end = reinterpret_cast<T *>(inputs.at(kIndex1)->addr);
auto *input_weight = reinterpret_cast<T *>(inputs.at(kIndex2)->addr);
T *output = reinterpret_cast<T *>(outputs.at(kIndex0)->addr);
auto task = [&input_start, &input_end, &input_weight, &output](size_t start, size_t end) {
for (size_t i = start; i < end; i++) {
T start_value = input_start[i];
@ -85,10 +77,6 @@ bool LerpCpuKernelMod::LaunchKernel(const std::vector<kernel::AddressPtr> &input
ParallelLaunchAutoSearch(task, output_size_, this, &parallel_search_info_, pool_);
} else {
MultipleBroadcastIterator multi_broadcast_iterator({start_shape_, end_shape_, weight_shape_}, output_shape_);
auto *input_start = reinterpret_cast<T *>(inputs.at(kIndex0)->addr);
auto *input_end = reinterpret_cast<T *>(inputs.at(kIndex1)->addr);
auto *input_weight = reinterpret_cast<T *>(inputs.at(kIndex2)->addr);
T *output = reinterpret_cast<T *>(outputs.at(kIndex0)->addr);
auto task = [&input_start, &input_end, &input_weight, &output, &multi_broadcast_iterator](size_t start,
size_t end) {
auto iter = multi_broadcast_iterator;
@ -108,12 +96,8 @@ bool LerpCpuKernelMod::LaunchKernel(const std::vector<kernel::AddressPtr> &input
const std::vector<std::pair<KernelAttr, LerpCpuKernelMod::KernelRunFunc>> &LerpCpuKernelMod::GetFuncList() const {
static const std::vector<std::pair<KernelAttr, LerpCpuKernelMod::KernelRunFunc>> func_list = {
{KernelAttr()
.AddInputAttr(kNumberTypeFloat16)
.AddInputAttr(kNumberTypeFloat16)
.AddInputAttr(kNumberTypeFloat16)
.AddOutputAttr(kNumberTypeFloat16),
&LerpCpuKernelMod::LaunchKernel<float16>},
// Lerp support fp16 && fp32, but precision is too low in fp16.
// So we register fp32 and make use of ms framework to cast fp16 to fp32.
{KernelAttr()
.AddInputAttr(kNumberTypeFloat32)
.AddInputAttr(kNumberTypeFloat32)

2
mindspore/ccsrc/plugin/device/cpu/kernel/lerp_cpu_kernel.h
View File

@ -44,8 +44,6 @@ class LerpCpuKernelMod : public NativeCpuKernelMod, public MatchKernelHelper<Ler
int Resize(const BaseOperatorPtr &base_operator, const std::vector<KernelTensorPtr> &inputs,
const std::vector<KernelTensorPtr> &outputs, const std::map<uint32_t, tensor::TensorPtr> &) override;
void ResetResource() noexcept;
const std::vector<std::pair<KernelAttr, KernelRunFunc>> &GetFuncList() const override;
protected: