[Feat]: Support for GPU Acceleration on Newer Qualcomm Devices

[shadow3aaa]

Description:
I would like to request the addition of GPU acceleration support for Qualcomm devices using the custom version of llama.cpp. This enhancement could significantly improve the performance of the PocketPal AI application on devices with Qualcomm Adreno GPUs.

Reference:
For more details, please refer to the Qualcomm developer blog post: Introducing the new OpenCL GPU backend for llama.cpp

Benefits:

• Enhanced performance on Qualcomm devices
• Better utilization of device hardware capabilities
• Potential for more responsive and efficient AI processing on mobile devices

Thank you for considering this feature request.

[shadow3aaa]

My test results (tested on Snapdragon 8 Elite platform):

• GPU Inference

OP5D0DL1:/data/local/tmp $ ./bin/llama-cli -m ggml-model-qwen1.5-7b-chat-Q4_0.gguf -b 128 -ngl 99 -c 2048 -p "Hello"

llama_perf_sampler_print:    sampling time =      24.03 ms /    46 runs   (    0.52 ms per token,  1914.51 tokens per second)
llama_perf_context_print:        load time =   11765.90 ms
llama_perf_context_print: prompt eval time =       0.00 ms /     1 tokens (    0.00 ms per token,      inf tokens per second)
llama_perf_context_print:        eval time =    5038.54 ms /    45 runs   (  111.97 ms per token,     8.93 tokens per second)
llama_perf_context_print:       total time =    5120.83 ms /    46 tokens

• CPU Inference

llama_perf_sampler_print:    sampling time =       1.34 ms /    13 runs   (    0.10 ms per token,  9708.74 tokens per second)
llama_perf_context_print:        load time =    4218.12 ms
llama_perf_context_print: prompt eval time =       0.00 ms /     1 tokens (    0.00 ms per token,      inf tokens per second)
llama_perf_context_print:        eval time =    6745.14 ms /    12 runs   (  562.10 ms per token,     1.78 tokens per second)
llama_perf_context_print:       total time =    8294.89 ms /    13 tokens

The GPU-accelerated llama.cpp inference speed is approximately 5 times faster than CPU inference on my device.