sycl : add Kronecker product FWHT support for sizes 384, 640, 768, 1280

[philip-jingxin]

Overview

Extends FWHT with Kronecker product to support non-power-of-two common embedding dimensions 384, 640, 768, 1280

Additional information

Changes

• Implemented kronecker_kernel to handle base blocks of $m=12$ and $m=20$ using the Paley I matrix construction.
• Extended operator tests in tests/test-backend-ops.cpp to cover non-power-of-two sizes for Hadamard transformation.

Testing & Verification

Environment

• Device: Intel(R) UHD Graphics 770
• Driver Version: 26.05.037020
• Command: ./build_hadamard/bin/test-backend-ops test -o MUL_MAT_HADAMARD -b SYCL0

Correctness Test

Passed 16/16 verification tests, including non-power-of-two sizes (m= 384, 640, 768, 1280) across various configurations:

Backend 1/2: SYCL0
  Device description: Intel(R) UHD Graphics 770
  MUL_MAT_HADAMARD(m=384,  n=1,  k=384):  OK
  MUL_MAT_HADAMARD(m=384,  n=32, k=384):  OK
  MUL_MAT_HADAMARD(m=384,  n=4,  k=384):  OK
  MUL_MAT_HADAMARD(m=640,  n=1,  k=640):  OK
  MUL_MAT_HADAMARD(m=640,  n=32, k=640):  OK
  MUL_MAT_HADAMARD(m=768,  n=1,  k=768):  OK
  MUL_MAT_HADAMARD(m=1280, n=1,  k=1280): OK
  16/16 tests passed

Performance Benchmarks

Measured on Intel UHD Graphics 770:

Size	Batch	Latency (us/run)	Throughput
384	1	4.20 us	70.19 GFLOPS
384	32	5.33 us	1.77 TFLOPS
640	1	16.74 us	48.95 GFLOPS
640	32	19.57 us	1.34 TFLOPS
768	1	17.23 us	68.47 GFLOPS
1280	1	72.43 us	45.24 GFLOPS

Requirements

• I have read and agree with the contributing guidelines
• AI usage disclosure: YES - Gemini was used to debug and ensure the correctness of the implementation.

[philip-jingxin]

This pr extends the FWHT to support a wider range of common embeddings. If you guys find this useful, I'll extends this to more generalized dimensions. Thanks @ggerganov

[arthw]

It's good job!
The UT cases are passed on B60 and iGPU in Ultra 7 265K.

Thank you!

[Titaniumtown]

@philip-jingxin Thank you! Very reasonable changes.

[Titaniumtown]

I too hastilly merged this PR, I believe as it breaks github actions. @philip-jingxin can you create a follow-up PR fixing the build issues?

[philip-jingxin]

I too hastilly merged this PR, I believe as it breaks github actions. @philip-jingxin can you create a follow-up PR fixing the build issues?

@Titaniumtown My bad! I've created a follow-up PR fixing the -Werror,-Wunused-variable warning and reintroducing the feature. Here you go: #28254

Thanks!