Error result in test-vector-hsa

[MaoxinYee]

./test-vector-hsa
ggml_hsa_precompiled_kernel_dir: no pregenerated kernel directory defined.
ggml_hsa_cached_kernel_dir: cached kernels in /home/ymx/.cache/ggml
Creating arrays of 32 elements and doing A + B.
Using HSA backend
ggml_gallocr_needs_realloc: graph has different number of nodes
ggml_gallocr_alloc_graph: reallocating buffers automatically
ggml_gallocr_reserve_n: reallocating HSA0 buffer from size 0.00 MiB to 0.00 MiB
A = [ 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 ]
B = [ 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 ]
A + B = [ 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 ]

[MaoxinYee]

I am not familiar with the HSA interface, so I have written a separate dispatch_xrt, and execution can proceed normally by using this function. However, I am still unclear about how to integrate the current GGML with XRT—what I am doing now is directly initiating an XRT for execution within the dispatch module. Additionally, the current device identification function recognizes my aie2p as aie2. I will attempt to modify only this part to see if the original HSA mechanism can execute correctly.

Registering xclbin: /home/ymx/.cache/ggml/aie2p/add-32i32-32i32-32i32.xclbin
Getting hardware context.
Getting handle to kernel:MLIR_AIE
Running Kernel.
A = [ 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 ]
B = [ 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 ]
A + B = [ 12 14 16 18 20 22 24 26 28 30 32 34 36 38 40 42 44 46 48 50 52 54 56 58 60 62 64 66 68 70 72 74 ]

[MaoxinYee]

This project is of great significance, and many people have expressed their hope for its support in the issues section of llama.cpp. Additionally, I have tried the current Windows lemonade solution with OGA Hybrid, but the result was quite unsatisfactory. Therefore, I hope to carry out the transformation of llama.cpp. I wish the author could continue to update the project or provide me with some guidance, as I am eager to push this project forward.

[MaoxinYee]

Your HSA code is working. Just change the dev_info.name to "aie2p", the result will be right.

A = [ 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 ]
B = [ 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 ]
A + B = [ 12 14 16 18 20 22 24 26 28 30 32 34 36 38 40 42 44 46 48 50 52 54 56 58 60 62 64 66 68 70 72 74 ]