vulkan: perf_logger improvements

[jeffbolznv]

• Move perf_logger from device to ctx.
• Add an env var to control the frequency we dump the stats. If you set a very large value, it just dumps when the ctx is destroyed.
• Add a fusion info string to the tracking, only log one item per fused op.
• Fix MUL_MAT_ID flops calculation.

[0cc4m]

Please resolve the conflict.

[jeffbolznv]

Rebased.

[0cc4m]

There's another conflict, but more importantly I'm getting a segfault with Qwen3-Next-80B-A3B-Instruct-Q4_0:

Core was generated by `build_vk_debug/bin/llama-bench -m models/Qwen3-Next-80B-A3B-Instruct-Q4_0.gguf -fa 1 --mmap 0'.
Program terminated with signal SIGABRT, Aborted.

#0  0x00007f8f3c49890c in ?? () from /usr/lib/libc.so.6
#1  0x00007f8f3c43e3a0 in raise () from /usr/lib/libc.so.6
#2  0x00007f8f3c42557a in abort () from /usr/lib/libc.so.6
#3  0x00007f8f3c89a41f in std::__glibcxx_assert_fail (file=<optimized out>, line=<optimized out>, function=<optimized out>, condition=<optimized out>) at /usr/src/debug/gcc/gcc/libstdc++-v3/src/c++11/assert_fail.cc:41
#4  0x00007f8f4075dc96 in std::vector<ggml_tensor*, std::allocator<ggml_tensor*> >::operator[] (this=0x55fdfca3d2d0, __n=3) at /usr/include/c++/15.2.1/bits/stl_vector.h:1263
#5  0x00007f8f3ce523cc in ggml_backend_vk_graph_compute (backend=0x55fdfca30890, cgraph=0x55fdfca378a8) at ggml/src/ggml-vulkan/ggml-vulkan.cpp:13173
#6  0x00007f8f4036bc25 in ggml_backend_graph_compute_async (backend=0x55fdfca30890, cgraph=0x55fdfca378a8) at ggml/src/ggml-backend.cpp:359
#7  0x00007f8f40370aa4 in ggml_backend_sched_compute_splits (sched=0x55fdfc98b840) at ggml/src/ggml-backend.cpp:1575
#8  0x00007f8f403718c3 in ggml_backend_sched_graph_compute_async (sched=0x55fdfc98b840, graph=0x7f8f28d70030) at ggml/src/ggml-backend.cpp:1784
#9  0x00007f8f4079ba93 in llama_context::graph_compute (this=0x55fdff05c900, gf=0x7f8f28d70030, batched=true) at src/llama-context.cpp:1488
#10 0x00007f8f407986e9 in llama_context::process_ubatch (this=0x55fdff05c900, ubatch=..., gtype=LLM_GRAPH_TYPE_DECODER, mctx=0x55fdfca2d350, ret=@0x7ffc0c26da58: GGML_STATUS_SUCCESS) at src/llama-context.cpp:809
#11 0x00007f8f40799bef in llama_context::decode (this=0x55fdff05c900, batch_inp=...) at src/llama-context.cpp:1113
#12 0x00007f8f407a091d in llama_decode (ctx=0x55fdff05c900, batch=...) at src/llama-context.cpp:2780
#13 0x000055fde84babcb in test_prompt (ctx=0x55fdff05c900, n_prompt=512, n_batch=2048, n_threads=16) at tools/llama-bench/llama-bench.cpp:1945
#14 0x000055fde84bb882 in main (argc=7, argv=0x7ffc0c26ea38) at tools/llama-bench/llama-bench.cpp:2125

[jeffbolznv]

I'm not able to reproduce the crash locally, and I don't see the merge conflict in the UI. The crash sounds a lot like what I fixed in e3f771b, but I can't correlate the line number in your call stack to anything. Are you sure you were on the latest commit?

[0cc4m]

You are right, I didn't notice that it failed updating the PR-branch and I tested only the first commit. My bad, sorry.