Hi, this is a PR for improving the SYCL backend compatibility on Intel GPUs. We fixed and updated some sycl kernels to make sure stable-diffusion models could generate larger images (for example, 1024 x 1024).

Changes

• update submodule ggml commit to latest 21d3a to sync SYCL kernels and avoid conflict with Vulkan backend PR at the same time.
• turn off fast-math on the CPU if using SYCL backend. Otherwise, it will affect some host calculations (for example, start_merge_step )
• update SYCL related readme and refine some messages in CMakeLists for good looking.

Results

test in Intel Data Center GPU Max 1100 with linux system.

SD series

• SD2
  ./build/bin/sd -m ../sd_models/v2-1_768-nonema-pruned.safetensors -p "a lovely cat" -o "output_sd2_hw1024.png" -H 1024 -W 1024 -v
  total time: 30.67s
  

• SDXL
  ./build/bin/sd -m ../sd_models/sdxl/sd_xl_base_1.0.safetensors --vae ../sd_models/sdxl/sdxl_vae.safetensors -H 1024 -W 1024 -p "a lovely cat" -v -o "output_sdxl_hw1024.png" --seed 16
  total time: 35.24s
  

• SD3
  ./build/bin/sd -m ../sd_models/sd3_medium_incl_clips_t5xxlfp16.safetensors -H 1024 -W 1024 -p "a lovely cat holding a sign says \"Stable Diffusion CPP\"" --cfg-scale 4.5 --sampling-method euler -v -o "output_sd3_hw1024.png"
  total time: 46.91s

FLUX

• FLUX-dev
  ./build/bin/sd --diffusion-model ../sd_models/flux/flux1-dev-q8_0.gguf --vae ../sd_models/flux/ae.safetensors --clip_l ../sd_models/flux/clip_l.safetensors --t5xxl ../sd_models/flux/t5xxl_fp16.safetensors -p "a lovely cat holding a sign says 'flux.cpp'" --cfg-scale 1.0 --sampling-method euler -v -H 1024 -W 1024

type	q8_0	q4_0	q4_k	q3_k	q2_k
total time (s)	114.84	116.15	114.77	115.16	114.08
total mem (MB)	21481.50	15807.93	15808.58	14301.56	13149.14
result (1024x1024)

• FLUX-schnell
  ./build/bin/sd --diffusion-model ../sd_models/flux/flux1-schnell-q8_0.gguf --vae ../sd_models/flux/ae.safetensors --clip_l ../sd_models/flux/clip_l.safetensors --t5xxl ../sd_models/flux/t5xxl_fp16.safetensors -p "a lovely cat holding a sign says 'flux.cpp'" --cfg-scale 1.0 --sampling-method euler -v --steps 4 -o "output_flux_schnell.png" -H 1024 -W 1024
  total time: 29.61s
  

Remain issues

• diffusion model weights seem to have no change after lora was applied, causing the unexpected generation results in PhotoMaker and other Lora applications.
  photomaker: ./build/bin/sd -m ../sd_models/sdxl/sdxlUnstableDiffusers_v11.safetensors --vae ../sd_models/sdxl/sdxl_vae.safetensors --stacked-id-embd-dir ../sd_models/photo_maker/photomaker-v1.safetensors --input-id-images-dir ./assets/photomaker_examples/scarletthead_woman/ -p "a girl img, retro futurism, retro game art style but extremely beautiful, intricate details, masterpiece, best quality, space-themed, cosmic, celestial, stars, galaxies, nebulas, planets, science fiction, highly detailed" -n "realistic, photo-realistic, worst quality, greyscale, bad anatomy, bad hands, error, text" --cfg-scale 5.0 --sampling-method euler -H 1024 -W 1024 --style-ratio 15
  result:
  
  It will take some time to debug and fix. Would put it in another PR.

cc @airMeng, @luoyu-intel, @hshen14