Loading...
Loading...
Found 6 Skills
AMD ROCm / HIP GPU コンピューティングリファレンス。 hipMalloc, hipLaunchKernelGGL, HIP Graph, HIPRTC, hipify, rocBLAS, MIOpen, RCCL, rocprofv3, amd-smi, rocminfo, CDNA1-4 / RDNA1-4, MI300 / MI350, gfx942, Matrix Core, wave32/64。
Diagnoses why ROCm, the HIP SDK, PyTorch, or llama.cpp is broken on an AMD GPU on Linux, Windows, or WSL2, then applies a low-risk fix with consent or hands back the exact next step. Also routes Lemonade, LM Studio, and Ollama problems to the right upstream channel. Use when the user reports that ROCm or HIP "isn't working", torch.cuda.is_available() is False, rocminfo / hipInfo can't see the GPU, or hits hipErrorNoBinaryForGpu, HSA_STATUS_ERROR_INVALID_ISA, "invalid device function", "no kernel image is available", cannot open /dev/kfd, permission denied on /dev/kfd, "ROCk module is NOT loaded", a missing libamdhip64.so / amdhip64_6.dll / hipblas.dll / vcruntime140_1.dll, an HSA_OVERRIDE_GFX_VERSION page fault, an iGPU+dGPU crash, a container that can't see the GPU, or an amdgpu-install / DKMS failure. Backed by the `rocm` CLI (`rocm examine` / `rocm diagnose` / `rocm fix`); this skill is a thin driver over those commands, not a re-implementation.
Collects and reads rocprof-compute roofline data to place an AMD GPU kernel against the hardware's peak compute and peak bandwidth ceilings. Use when the user asks whether a kernel is compute-bound or memory-bound, asks about arithmetic intensity or FLOPs per byte, asks how much performance is left on the table, or asks for a roofline. Not for CUDA tools, Windows, or per-instruction analysis.
Profiles an AMD GPU application with rocprof-compute and finds which kernel is slow and why, using performance metrics and PC sampling. Collects counters and selects kernels and dispatches, then analyzes GPU memory throughput, compute efficiency, wavefront occupancy and wavefront limits, scheduler stalls, etc. Use when the user asks to profile, benchmark, or speed up a HIP/ROCm kernel or application, or asks where a kernel's bottleneck is. Not for CUDA tools, Windows even when the user names rocprof-compute, system-wide CPU/MPI tracing, or writing kernel source.
Analyzes the AMD GPU memory hierarchy with rocprof-compute, covering the Memory Chart, vL1D and L2 caches, LDS bank conflicts, Infinity Fabric traffic, and the experimental gfx950 memory bandwidth analysis. Use when the user asks about memory bandwidth, cache hit rates, coalescing, LDS conflicts, data locality, HBM traffic, or says a kernel is memory-bound. Not for CUDA tools, Nsight, Windows, host-memory leaks, GPU out-of-memory allocation failures, or system-wide tracing.
Attributes AMD GPU kernel performance counters back to the PyTorch or Triton operators that launched them, using rocprof-compute ML API tracing. Use when the user asks which PyTorch layer or operator is slow, wants to map GPU kernels to model code, profiles a training or inference script, asks about torch.compile and Inductor kernels, or wants to select user-defined ROCTx regions from a traced Python profile. Not for CUDA tools, Windows, compiled non-Python binaries, or starting and stopping a profile at ROCTx ranges.