[QDP] Add zero-copy amplitude batch encoding from float32 GPU tensors

[viiccwen]

Purpose of PR

This PR adds batch float32 amplitude encoding support in QDP core and kernels.

It extends the existing float32 GPU-pointer amplitude path from single-sample encoding to batched encoding, and refactors batch state allocation so the output precision is selected explicitly at allocation time.

What changed

Core

• Added QdpEngine::encode_batch_from_gpu_ptr_f32
• Added QdpEngine::encode_batch_from_gpu_ptr_f32_with_stream
• Added the corresponding float32 batch amplitude entry point in AmplitudeEncoder

Kernels

• Added launch_amplitude_encode_batch_f32
• Added the float32 batch amplitude CUDA kernel wiring in qdp-kernels
• Reused the existing float32 batched L2 norm reduction for batch normalization

Allocation

• Refactored GpuStateVector::new_batch to accept an explicit Precision
• Updated all batch allocation call sites to pass the correct output precision
  • amplitude batch uses Float64 or Float32 as appropriate
  • angle, basis, IQP, and the streaming pipeline continue to request Float64

Python

• Kept float32 + amplitude batch encoding unsupported in Python bindings for now
• Existing Python-facing error behavior remains unchanged and can be handled in a follow-up PR

Tests

• Added Float32 batch DLPack shape coverage
• Added core GPU-pointer tests for float32 batch amplitude success and validation cases

Related Issues or PRs

closes #1028

Changes Made

• Bug fix
• New feature
• Refactoring
• Documentation
• Test
• CI/CD pipeline
• Other

Breaking Changes

• Yes
• No

Checklist

• Added or updated unit tests for all changes
• Added or updated documentation for all changes
• Successfully built and ran all unit tests or manual tests locally
• PR title follows "MAHOUT-XXX: Brief Description" format (if related to an issue)
• Code follows ASF guidelines

[ryankert01]

Please resolve the conflict, Ty for the contribuiton.

[viiccwen]

solved!

[rich7420]

sorry for the late,
Overall lg
I'll review deeper later

[rich7420]

@viiccwen thanks for the patch!
left some comments:
I think this new f32 batch GPU‑pointer API and related pipelines are not fully covered by tests.

[rich7420]

In this file around lines 251–276 and 351–376, amplitude_encode_batch_kernel / _f32 compute input_base = sample_idx * input_len and then do reinterpret_cast<const double2*>(input_batch + input_base) + elem_pair / float2. For odd input_len and sample_idx > 0 this base pointer is only 8‑byte (double) / 4‑byte (float) aligned, not 16‑byte, so the double2/float2 loads are potentially misaligned. This alignment pattern already existed in the original f64 batch kernel and this PR copies it into the new f32 batch path; please either enforce even input_len at the Rust call‑site or rework the kernels to index from a properly aligned double2* / float2* base pointer with a scalar fallback.

[ryankert01]

Please help resolve conflicts

[Copilot]

Pull request overview

This PR adds comprehensive float32 amplitude batch encoding support to the QDP engine, enabling zero-copy GPU encoding from PyTorch float32 CUDA tensors.

Changes:

• Added f32 batch amplitude encoding kernels and GPU pointer APIs in core
• Refactored GpuStateVector::new_batch to accept precision parameter for flexible buffer allocation
• Updated all batch encoding call sites to specify output precision

Reviewed changes

Copilot reviewed 15 out of 15 changed files in this pull request and generated 1 comment.

Show a summary per file

File	Description
qdp/qdp-kernels/src/amplitude.cu	Added f32 batch amplitude kernel (amplitude_encode_batch_kernel_f32) and launcher, plus f32 batch L2 norm support
qdp/qdp-kernels/src/lib.rs	Added FFI declarations for f32 batch kernels and stubs for non-CUDA builds
qdp/qdp-core/src/lib.rs	Added encode_batch_from_gpu_ptr_f32 public API with validation and added precision() accessor
qdp/qdp-core/src/gpu/memory.rs	Refactored new_batch to accept Precision parameter for f32 or f64 buffer allocation
qdp/qdp-core/src/gpu/encodings/mod.rs	Added default encode_batch_from_gpu_ptr_f32 trait method
qdp/qdp-core/src/gpu/encodings/amplitude.rs	Implemented f32 batch encoding with norm validation
qdp/qdp-core/src/gpu/encodings/{angle,basis,iqp}.rs	Updated new_batch calls to pass Precision::Float64
qdp/qdp-core/src/encoding/mod.rs	Updated streaming pipeline to use engine.precision()
qdp/qdp-python/src/lib.rs	Added validation for f32/f64 amplitude tensors; 1D f32 supported, 2D f32 returns clear error
qdp/qdp-core/tests/gpu_ptr_encoding.rs	Comprehensive f32 batch tests covering success and error paths
qdp/qdp-core/tests/dlpack.rs	Updated to pass precision to new_batch
testing/qdp/test_bindings.py	Added tests for f32 input with both f32 and f64 engine precision
qdp/qdp-python/tests/test_dlpack_validation.py	Updated to verify 1D f32 works and 2D f32 gives clear error

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.