Transform Operations

This directory contains usage samples of tensor transform operators in PyPTO, including Assemble, Gather, Concat, and View operations.

Overview

The transform operator samples cover the following content:

• Assemble: Place a small tensor at a specified offset position within a large tensor.
• Gather: Collect elements from a source tensor based on an index tensor.
• Concat: Concatenate multiple tensors along a specified dimension.
• View: Create a view of a tensor, supporting specified shapes and offsets without data copying.

Code File Description

• transform_ops.py: Comprehensive sample containing all transform operators.
  • test_assemble_basic: Basic Assemble operation.
  • test_gather_basic: Basic Gather operation.
  • test_concat_basic: Basic Concat operation.
  • test_view_basic: Basic View operation.
• add_scalar_loop_view_assemble.py: Sample demonstrating scalar addition combined with View and Assemble in a loop tiling pattern.

How To Run

Environment Preparation

# Configure CANN environment variables
# After installation, configure the environment variables. Execute the following command based on the actual path of set_env.sh.
# The above environment variable configuration only takes effect in the current window. You can write the above commands into the environment variable configuration file (such as .bashrc) as needed.

# Default path installation, using root user as a sample (for non-root users, replace /usr/local with ${HOME})
source /usr/local/Ascend/ascend-toolkit/set_env.sh

# Set device ID
export TILE_FWK_DEVICE_ID=0

Run the Samples

# Run all transform-related samples
python3 transform_ops.py

# List all available transform test cases
python3 transform_ops.py --list

# Run a specific test case
python3 transform_ops.py assemble::test_assemble_basic

Core API Description

1. Assemble

Commonly used to assemble computed Tile results back into a global tensor.

# Place small_tensor at the [0, 0] offset in large_tensor
pypto.assemble(small_tensor, offsets=[0, 0], large_tensor)

2. Gather

Select data from an input tensor based on indices.

# Gather along dimension 0
pypto.gather(input_tensor, dim=0, index_tensor)

3. Concat

Concatenate multiple tensors in order.

# Concatenate two tensors along dimension 1
pypto.concat([tensor1, tensor2], dim=1)

4. View

Create a reference pointing to a partial region of the original tensor. It is key to implementing Tiling loops.

# Create a 4x4 view with an offset of [0, 4]
view = pypto.view(tensor, shape=[4, 4], offsets=[0, 4])

Precautions

• Assemble and View: These two operators are typically used together to implement manual Tiling loops.
• Dimension Alignment: When performing Concat operations, the shapes of all dimensions except the concatenation dimension must be consistent.
• Data Copying: The view operation does not involve data copying, while gather and concat typically produce new data copies.