convert-xtensor-to-qnn

name	convert-xtensor-to-qnn
description	Convert xtensor prototype code to production QNN direct-buffer operations (zero-copy)
allowed-tools	Read, Edit, Write, Bash

Convert xtensor to QNN tensors

Convert xtensor-based prototype code to direct QNN buffer operations for production. This eliminates the extra memory copy that xtensor introduces.

Conversion patterns

xtensor	QNN equivalent
xt::xarray<T>	Raw pointer + shape metadata
xt::view(arr, range)	Offset calculation into shared buffer
xt::zeros<T>({...})	memset(ptr, 0, size)
xt::ones<T>({...})	Loop initialization
xt::adapt(ptr, shape)	Direct pointer use (already zero-copy)
xt::eval(expr)	Materialize into pre-allocated buffer

QNN buffer access

// Read from graph output
float* output_ptr = static_cast<float*>(graph.outputPtr("tensor_name"));

// Write to graph input
float* input_ptr = static_cast<float*>(graph.inputPtr("tensor_name"));

// Using Graph API (handles dtype conversion: float32, float16, ufixed8/16, int32)
graph.write("tensor_name", float_data_ptr);
graph.read("tensor_name", float_output_ptr);

Memory layout

QNN uses row-major (C-style) layout, same as xtensor default.

For 2D tensor X[rows][cols]:

X[0,0], X[0,1], ..., X[0,cols-1], X[1,0], X[1,1], ...

Shared buffers

ARN (prefill) and AR1 (decode) graphs share input/output buffers since they never run concurrently. Modifying one graph's input buffer affects the other.

Reshape operations

Expand (2x2 → 2x3): Process last row to first (move backward), zero-pad new positions.

Shrink (2x3 → 2x2): Process first row to last (move forward), compact memory.

When to convert

Scenario	Use
Initial prototyping	xtensor
Debugging complex logic	xtensor
Production deployment	QNN tensors
Performance-critical paths	QNN tensors
One-off preprocessing	xtensor (acceptable overhead)

Workflow

1. Read the xtensor code to understand the tensor operations
2. Identify buffer pointers from graph.inputPtr() / graph.outputPtr()
3. Replace xtensor operations with direct pointer arithmetic
4. Ensure correct stride calculations for multi-dimensional access
5. Use memset for zero-initialization, memcpy for block copies
6. Verify correctness by comparing output values against the xtensor version