Abstract:

This video documents a lengthy live-streamed session where developer René Rebe attempts to compile AMD's ROCm (Radeon Open Compute platform) GPU compute stack for AI workloads (specifically targeting Llama.cpp) on a high-core-count RISC-V system (Sophgo Milk-V Pioneer "Ripper" with 64 cores/128GB RAM) using the T2 SDE Linux distribution. The stream highlights the significant challenges involved in porting such a complex software suite to a non-mainstream architecture. Key difficulties include navigating extremely inconsistent and often broken CMake build configurations across different ROCm components (using varying variable names like hip_ARCHITECTURES, GPU_TARGETS, AMD_GPU_TARGETS), resolving numerous dependency issues related to hardcoded paths and library linking errors, and battling exceptionally long build and link times (often exceeding 40 minutes for linking single components on RISC-V) exacerbated by single-threaded bottlenecks and software bloat. Despite applying numerous patches and workarounds, and successfully building some components and even Llama.cpp with ROCm support on an x86_64 system, the effort is ultimately halted by a critical segmentation fault within the clang compiler while building the rocksolver library on the RISC-V hardware, demonstrating the arduous nature of this porting task.

Building ROCm and AI on RISC-V: A Debugging Journey

• 0:01:50 Hardware Showcase: Introduces the powerful build environment featuring multiple systems, including a 64-core RISC-V "Ripper" (Sophgo Milk-V Pioneer) with 128GB RAM, totaling over 100 CPU threads and 240GB RAM across the setup.
• 0:02:47 Goal Definition: The primary aim is to build the AMD ROCm stack on RISC-V to enable GPU-accelerated AI workloads, specifically targeting Llama file / Llama.cpp.
• 0:07:15 Initial Build Struggles: Encounters immediate problems with the ROCm build system (CMake), including extremely long configuration times (9+ minutes) and failures due to incorrect prerequisites like Rust versions.
• 0:11:01 Compiling Core Libraries: Begins compiling key ROCm components like hipblaslt and composable_kernel, anticipating very long build durations.
• 0:19:18 Performance & Bottlenecks: Notes relatively slow SSD speed on the RISC-V Ripper and highlights significant single-threaded performance bottlenecks during CMake configuration stages on the RISC-V hardware.
• 0:24:18 CMake Hell - Inconsistent Targets: Spends significant time debugging build failures caused by AMD using different CMake variable names (hip_ARCHITECTURES, GPU_TARGETS, AMD_GPU_TARGETS) across various ROCm projects to specify target GPU architectures, leading to major frustration.
• 0:30:24 Debugging Linker Paths: Battles numerous build failures related to the build system not finding required libraries or headers due to incorrect or hardcoded paths, requiring patches and workarounds. Discusses issues with symlinks added previously causing problems with path detection scripts (hipconfig).
• 0:53:52 Side Discussion: Apple Silicon: Briefly discusses building for ARM64 on Apple Silicon, past issues with Asahi Linux dual-booting, and QEMU performance degradation.
• 1:10:11 Uncovering hipconfig Bug: Discovers that the hipconfig script Error: 'rowid' incorrectly determines library paths when executed via a symlink (a previous workaround), leading to build failures.
• 1:23:00 Fixing Target Variables: Corrects CMake build scripts where the wrong variable for GPU targets was used (due to earlier assumption of consistency), preventing builds from incorrectly attempting to compile for all architectures and significantly shortening build times for those components.
• 1:44:14 Painfully Long Link Times: Experiences and laments extremely long link times (e.g., rockblast taking over 40 minutes on the 64-core RISC-V system), attributing it to software bloat and lack of parallelism in the build process.
• 2:01:44 Configuring Llama.cpp: Modifies the build configuration for Llama.cpp to enable the ROCm/HIP backend.
• 2:05:15 Patching hipconfig: Patches the hipconfig Perl script directly to correctly handle library paths, avoiding the need for problematic symlinks.
• 2:12:07 Success: Llama.cpp Builds on x86: Successfully compiles Llama.cpp with the ROCm/HIP backend enabled on the x86_64 Ryzen system.
• 2:21:16 Further Build Failures: Continued attempts to build remaining ROCm dependencies (hipfft, rocksolver) on RISC-V hit more linking errors and missing dependencies (like hipRTC), requiring more CMake patching.
• 2:32:06 Critical Failure: rocksolver Segfaults on RISC-V: The build of rocksolver on the RISC-V Ripper fails catastrophically with a segmentation fault originating from the clang compiler frontend (LLVM ERROR: Unable to execute command: ... Segmentation fault). This halts progress as rocksolver is needed for hipblas.
• 3:21:08 Session End & Next Steps: Due to the rocksolver crash, the developer concludes the stream. The plan is to rebuild the RISC-V kernel without potentially unstable experimental features (like vector extensions) and retry the build tomorrow. The RISC-V machine is powered off.