[AdvancedCompiler]Contiguous（cpp wrapper） #750
Conversation
lib/contiguous.cpp
Outdated
| at::Tensor input_sizes = torch::tensor(input.sizes(), options); | ||
| at::Tensor input_strides = torch::tensor(input.strides(), options); | ||
| at::Tensor out_strides = torch::tensor(out.strides(), options); |
There was a problem hiding this comment.
Interesting, torch has a factory function for creating tensors like this.
lib/contiguous.cpp
Outdated
| namespace flag_gems { | ||
| using namespace triton_jit; | ||
|
|
||
| at::Tensor contiguous(at::Tensor &input, at::MemoryFormat memory_format) { |
There was a problem hiding this comment.
Use signature
at::Tensor contiguous(const at::Tensor & self, at::MemoryFormat memory_format=c10::MemoryFormat::Contiguous);since this is the signature translated from the schema
contiguous(Tensor(a) self, *, MemoryFormat memory_format=contiguous_format) -> Tensor(a)There was a problem hiding this comment.
Check this file torch/include/ATen/ops/contiguous_native.h
src/flag_gems/csrc/cstub.cpp
Outdated
| @@ -20,6 +20,7 @@ TORCH_LIBRARY(flag_gems, m) { | |||
| m.def( | |||
| "rotary_embedding(Tensor q, Tensor k, Tensor cos, Tensor sin, Tensor? position_ids=None, " | |||
| "bool rotary_interleaved=False) -> (Tensor, Tensor)"); // q and k may be view to other size | |||
| m.def("contiguous(Tensor self, *, MemoryFormat memory_format=contiguous_format) -> Tensor"); | |||
There was a problem hiding this comment.
Use the schema from native_functions.yaml if we are implementing the operator with the same semantics.
contiguous(Tensor(a) self, *, MemoryFormat memory_format=contiguous_format) -> 
src/flag_gems/csrc/cstub.cpp
Outdated
| @@ -20,6 +20,7 @@ TORCH_LIBRARY(flag_gems, m) { | |||
| m.def( | |||
| "rotary_embedding(Tensor q, Tensor k, Tensor cos, Tensor sin, Tensor? position_ids=None, " | |||
| "bool rotary_interleaved=False) -> (Tensor, Tensor)"); // q and k may be view to other size | |||
| m.def("contiguous(Tensor(a) self, *, MemoryFormat memory_format=contiguous_format) -> Tensor"); | |||
There was a problem hiding this comment.
Please use the same schema with the one in native_functions.yaml
contiguous(Tensor(a) self, *, MemoryFormat memory_format=contiguous_format) -> Tensor(a)(a) here means aliasing.
| namespace flag_gems { | ||
| using namespace triton_jit; | ||
|
|
||
| at::Tensor contiguous(const at::Tensor &self, at::MemoryFormat memory_format) { |
There was a problem hiding this comment.
Since our current implementation does not support ChannelLast and ChannelLast3d format, it is necessary to raise an error if the memory format passed in is not supported. We can leave it for the future.
PR Category
Operator
Type of Change
New Feature
Description
CPP wrapper packaging for Contiguous op.
Issue
Progress
Performance