[webgpu-native] Add transpose op #21986
Conversation
|
|
||
| class TransposeProgram final : public Program<TransposeProgram> { | ||
| public: | ||
| TransposeProgram(const std::string& kernel_name, const gsl::span<const size_t>& permutations) |
There was a problem hiding this comment.
name can remove because it should always be "Transpose"
| .TypeConstraint("T", DataTypeImpl::AllFixedSizeTensorTypes()), | ||
| Transpose); | ||
|
|
||
| const std::string permFunctionBody(const std::string& input_name, const std::string& output_name, const gsl::span<const size_t>& perm) { |
There was a problem hiding this comment.
use a std::ostringstream& ss as the first parameter.
body in this context (at least in MainFunctionBody) means the part between brackets. maybe just use Function.
may be changed to "AppendPermFunction"
There was a problem hiding this comment.
ORT prefer PascalCase (first letter upper case) for function name
There was a problem hiding this comment.
usually can safely replace const std::string& with std::string_view in function parameters
| if (!status.IsOK()) | ||
| return status; |
| Status Transpose::ComputeInternal(ComputeContext& context) const { | ||
| const auto* input_tensor = context.Input(0); | ||
| const TensorShape& input_shape = input_tensor->Shape(); | ||
| int32_t rank = gsl::narrow_cast<int32_t>(input_shape.NumDimensions()); |
There was a problem hiding this comment.
may need a discussion: which is better? we can choose one to use then keep consistent
- gsl::narrow_cast<>
- SafeInt<>
There was a problem hiding this comment.
done gsl::narrow_cast<>
| TensorShape output_shape(output_dims); | ||
| auto* output_tensor = context.Output(0, output_shape); | ||
|
|
||
| SafeInt<uint32_t> vec_size = input_tensor->Shape().Size(); |
There was a problem hiding this comment.
should not name vec_size since the input/output is not vec.
|
|
||
| SafeInt<uint32_t> vec_size = input_tensor->Shape().Size(); | ||
| TransposeProgram program{"Transpose", *p_perm}; | ||
| program |
There was a problem hiding this comment.
perm should either be a part of cache hint or uniform. currently it seems different perm may use the same shader...
axinging
force-pushed
the
transpose_webgpunative
branch
2 times, most recently
from
f1449e1
to
ea145ce
Compare
axinging
force-pushed
the
transpose_webgpunative
branch
from
ea145ce
to
ed4134f
Compare
| Status Transpose::ComputeInternal(ComputeContext& context) const { | ||
| const auto* input_tensor = context.Input(0); | ||
| const TensorShape& input_shape = input_tensor->Shape(); | ||
| int32_t rank = gsl::narrow_cast<int32_t>(input_shape.NumDimensions()); |
There was a problem hiding this comment.
done gsl::narrow_cast<>
|
|
||
| SafeInt<uint32_t> vec_size = input_tensor->Shape().Size(); | ||
| TransposeProgram program{"Transpose", *p_perm}; | ||
| program |
| .TypeConstraint("T", DataTypeImpl::AllFixedSizeTensorTypes()), | ||
| Transpose); | ||
|
|
||
| const std::string permFunctionBody(const std::string& input_name, const std::string& output_name, const gsl::span<const size_t>& perm) { |
|
|
||
| class TransposeProgram final : public Program<TransposeProgram> { | ||
| public: | ||
| TransposeProgram(const std::string& kernel_name, const gsl::span<const size_t>& permutations) |
| if (!status.IsOK()) | ||
| return status; |
| TensorShape output_shape(output_dims); | ||
| auto* output_tensor = context.Output(0, output_shape); | ||
|
|
||
| SafeInt<uint32_t> vec_size = input_tensor->Shape().Size(); |
Description
Motivation and Context