add hipBlas support

[Cyberhan123]

This feature is ready to be merged. @leejet
Update: Thanks to leejet's efforts, we can use the master branch directly. I think I need to wait for the following PR to be merged:
ggerganov/ggml#683
ggerganov/ggml#682

[Cyberhan123]

I'm a little confused now:
ggml has removed n_dims from the main branch, but we still have it in our branch, so what should I do because I want to modify ggml.

@leejet @FSSRepo

edit: I mean do we have to follow the master branch and make adjustments?

[Cyberhan123]

This is the performance of AMD 7900XTX on windows

ggml_init_cublas: GGML_CUDA_FORCE_MMQ:   no
ggml_init_cublas: CUDA_USE_TENSOR_CORES: yes
ggml_init_cublas: found 1 ROCm devices:
  Device 0: AMD Radeon RX 7900 XTX, compute capability 11.0
[INFO]  stable-diffusion.cpp:5386 - loading model from './models/miniSD.ckpt'
[INFO]  model.cpp:641  - load ./models/miniSD.ckpt using checkpoint format
[INFO]  stable-diffusion.cpp:5412 - Stable Diffusion 1.x
[INFO]  stable-diffusion.cpp:5418 - Stable Diffusion weight type: f32
[INFO]  stable-diffusion.cpp:5577 - total memory buffer size = 2731.37MB (clip 470.66MB, unet 2165.24MB, vae 95.47MB)
[INFO]  stable-diffusion.cpp:5579 - loading model from './models/miniSD.ckpt' completed, taking 10.19s
[INFO]  stable-diffusion.cpp:5593 - running in eps-prediction mode
[INFO]  stable-diffusion.cpp:6486 - apply_loras completed, taking 0.00s
[INFO]  stable-diffusion.cpp:6525 - get_learned_condition completed, taking 59 ms
[INFO]  stable-diffusion.cpp:6535 - sampling using Euler A method
[INFO]  stable-diffusion.cpp:6539 - generating image: 1/1 - seed 42
  |==================================================| 20/20 - 8.84it/s
[INFO]  stable-diffusion.cpp:6551 - sampling completed, taking 2.35s
[INFO]  stable-diffusion.cpp:6559 - generating 1 latent images completed, taking 2.39s
[INFO]  stable-diffusion.cpp:6561 - decoding 1 latents
[INFO]  stable-diffusion.cpp:6571 - latent 1 decoded, taking 0.13s
[INFO]  stable-diffusion.cpp:6575 - decode_first_stage completed, taking 0.13s
[INFO]  stable-diffusion.cpp:6592 - txt2img completed in 2.58s
[INFO]  main.cpp:538  - save result image to 'output.png'

[leejet]

I'm a little confused now: ggml has removed n_dims from the main branch, but we still have it in our branch, so what should I do because I want to modify ggml.

@leejet @FSSRepo

edit: I mean do we have to follow the master branch and make adjustments?

I'll attempt to adapt to the latest version of ggml.

[ggerganov]

There is ggml_n_dims that can be used to compute the n_dims of a tensor. It's a drop-in replacement in 99% of the cases

[Cyberhan123]

@leejet I think everything is ready, but the worst part is that I don't have an AMD GPU at the moment. I have to wait until I get home this weekend to test it.

[Cyberhan123]

By the way, I have replaced ggml with the latest upstream version. I also hope that friends who use cuda can also test whether there are any problems.

[leejet]

Thank you for your contribution. I'll find time to review and merge this PR.

[Cyberhan123]

Thank you for your contribution. I'll find time to review and merge this PR.

Please wait while I test it, I don't have time now

[Cyberhan123]

Now it works fine, when I generate a smaller image (256x256), twenty sampling steps only takes 1.9 seconds

 ./bin/sd -m "../models/miniSD.ckpt" -p "a cat" -H 256 -W 256              
ggml_init_cublas: GGML_CUDA_FORCE_MMQ:   no
ggml_init_cublas: CUDA_USE_TENSOR_CORES: yes
ggml_init_cublas: found 1 ROCm devices:
  Device 0: AMD Radeon RX 7900 XTX, compute capability 11.0, VMM: no
[INFO ] stable-diffusion.cpp:137  - loading model from '../models/miniSD.ckpt'
[INFO ] model.cpp:642  - load ../models/miniSD.ckpt using checkpoint format
[INFO ] stable-diffusion.cpp:163  - Stable Diffusion 1.x 
[INFO ] stable-diffusion.cpp:169  - Stable Diffusion weight type: f32
[INFO ] stable-diffusion.cpp:268  - total memory buffer size = 2749.37MB (clip 479.66MB, unet 2165.24MB, vae 104.47MB)
[INFO ] stable-diffusion.cpp:270  - loading model from '../models/miniSD.ckpt' completed, taking 9.26s
[INFO ] stable-diffusion.cpp:284  - running in eps-prediction mode
[INFO ] stable-diffusion.cpp:1182 - apply_loras completed, taking 0.00s
[INFO ] stable-diffusion.cpp:1221 - get_learned_condition completed, taking 46 ms
[INFO ] stable-diffusion.cpp:1231 - sampling using Euler A method
[INFO ] stable-diffusion.cpp:1235 - generating image: 1/1 - seed 42
  |==================================================| 20/20 - 12.18it/s
[INFO ] stable-diffusion.cpp:1247 - sampling completed, taking 1.77s
[INFO ] stable-diffusion.cpp:1255 - generating 1 latent images completed, taking 1.80s
[INFO ] stable-diffusion.cpp:1257 - decoding 1 latents
[INFO ] stable-diffusion.cpp:1267 - latent 1 decoded, taking 0.14s
[INFO ] stable-diffusion.cpp:1271 - decode_first_stage completed, taking 0.14s
[INFO ] stable-diffusion.cpp:1290 - txt2img completed in 1.99s
save result image to 'output.png'

There are more features that I haven't tested yet, due to well-known network issues :)
Although I haven't tested other models, I believe they can work properly, so I think this one can be merged.