[DRAFT] Use TensorLayout in Tensor #15028
Conversation
| inline ttnn::Shape shape() const { return this->tensor_attributes->tensor_spec.compute_shape(); }; | ||
| inline DataType dtype() const { return this->tensor_attributes->tensor_spec.tensor_layout().get_data_type(); }; | ||
| inline Layout layout() const { return this->tensor_attributes->tensor_spec.tensor_layout().get_layout(); }; | ||
| inline Tile tile() const { return this->tensor_attributes->tensor_spec.tensor_layout().get_page_config().get_tile(); }; |
There was a problem hiding this comment.
lets discuss this. ideally we should not calculate these all the time, they are not changing dynamically
There was a problem hiding this comment.
The only calculated one out of theses getters is shape(). I'm planning to look into how often each compute_* method is called and cache some of them.
| for (const auto& output_spec : output_specs) { | ||
| output_tensors.emplace_back(create_device_tensor(output_spec, device)); |
| })); | ||
|
|
||
| }); | ||
| weight_tensor_ = ttnn::reshape(weight_tensor_, target_shape); |
There was a problem hiding this comment.
@mywoodstock please take a look
| @@ -99,8 +100,7 @@ ttnn::Tensor RepeatOperation::invoke( | |||
|
|
|||
| auto padded_to_tiled_shape = ttnn::Shape(sliced_logical_shape.view(), | |||
| tiled_output.get_padded_shape().view()); | |||
| tiled_output.set_shape(padded_to_tiled_shape); | |||
| return tiled_output; | |||
| return ttnn::reshape(tiled_output, padded_to_tiled_shape); | |||
There was a problem hiding this comment.
@ntarafdar , please take a look
| @@ -920,8 +920,8 @@ Matmul create_matmul_struct( | |||
| bool broadcast_batch = | |||
| parameters.bcast_batch.value_or(get_broadcast_batch(input_tensor_a, input_tensor_b, parameters.program_config)); | |||
| TT_FATAL(!(has_user_grid && has_program_config), "Cannot use both user core grid/coordinates and a program config"); | |||
| const auto& in0_tile = input_tensor_a.get_tile(); | |||
| const auto& in1_tile = input_tensor_b.get_tile(); | |||
| auto in0_tile = input_tensor_a.get_tile(); | |||
There was a problem hiding this comment.
what do you think about this change?
| output_tensor->tensor_attributes->dtype = local_tensor->tensor_attributes->dtype; | ||
| output_tensor->tensor_attributes->layout = local_tensor->tensor_attributes->layout; | ||
| output_tensor->tensor_attributes->tile = local_tensor->tensor_attributes->tile; | ||
| output_tensor->set_tensor_spec(local_tensor->tensor_spec()); |
There was a problem hiding this comment.
not a good public method, would like to discuss this.
| { | ||
| if(std::holds_alternative<TilePageConfig>(config_)) { | ||
| return std::get<TilePageConfig>(config_).get_tile(); | ||
| } | ||
|
|
||
| return std::nullopt; | ||
| return Tile(); |
There was a problem hiding this comment.
would like to discuss why
| @@ -18,7 +18,7 @@ size_t element_size_bytes(DataType dtype) { | |||
| case DataType::UINT8: return sizeof(uint8_t); | |||
| case DataType::BFLOAT8_B: | |||
| case DataType::BFLOAT4_B: | |||
| TT_THROW("element_size_bytes() should not be used for BFLOAT8_B and BFLOAT4_B types becaues of how they are packed"); | |||
| return sizeof(float); | |||
There was a problem hiding this comment.
would like to discuss why
| @@ -36,6 +36,12 @@ Alignment legacyShapeToAlignment(const ttnn::Shape& shape) { | |||
| values[i] = legacy_padded_shape[i] * values[i + 1]; | |||
| } | |||
|
|
|||
| for (auto& value : values) { | |||
There was a problem hiding this comment.
will this conflict with @TT-BrianLiu 's , change?
| @@ -51,6 +51,15 @@ class TensorLayout { | |||
| // H is all dimensions except W multiplied and aligned to tile and shard height | |||
| Size compute_physical_shape(const ttnn::SimpleShape& shape) const; | |||
|
|
|||
| void set_memory_config(MemoryConfig memory_config) { | |||
There was a problem hiding this comment.
wouldn't want it to be mutable like this. lets discuss
|
|
||
| namespace tt::tt_metal { | ||
|
|
||
| class TensorSpec final { |
There was a problem hiding this comment.
tbd, discuss next steps with caching some of the values
Ticket
Link to Github Issue
Problem description
Provide context for the problem.
What's changed
Describe the approach used to solve the problem.
Summarize the changes made and its impact.
Checklist