Module Tolk.Compiler_amd
The compiler of HIP for AMD GPUs.
comgr, ROCm's code object manager library, compiles the source. comgr compiles one source at a time in a process, so compiles run in processes of their own, which tolk starts from a program it carries: compiles from several domains run at once, and a compile that crashes ends its own process only. Each process serves compiles in turn, from any domain, and waits idle between them, so that loading comgr is paid once per process; a compile that finds none idle starts another. This process keeps as many as it ran compiles at once, a search's parallelism, each holding about 15 MB of its own memory, until it exits. A process forked from this one starts its own and never uses its parent's. On Linux the program runs from memory; elsewhere it is written to Helpers.cache_dir, under a name holding its digest. The library, libamd_comgr, is the file C.findlib finds when the compiler is made: the file COMGR_PATH names, then lib/libamd_comgr.so in the directory Setting.rocm_path names, then the system's library directories. Each process loads it. Its major version picks its numbering of languages and actions, which version 3 changed. comgr's own cache of the code it generates, which it keeps under ~/.cache/comgr or AMD_COMGR_CACHE_DIR, is disabled in those processes alone: code objects are cached in one place, Setting.ccache's table.
val hip : string -> Renderer.Compiler.thip arch is the compiler of HIP source to executable code objects for the GPU architecture arch, as in "gfx1100". A source whose first line is .text is assembly, which is assembled instead.
HIP is compiled with the device libraries, at -O3, in CU mode, as HIP 6.0 without HIP's own headers, whose declarations rendered kernels carry. The kernels' symbols are internalized before code generation, so that only kernels stay visible. Code generation runs without LLVM's loop idiom recognition, whose time from LLVM 22 grows exponentially with a chain of selections that each read the last twice: a loop it would replace is computed as written. From comgr 2.8 (ROCm 6.2), LLVM's verifier checks the IR before it is optimized and before code generation: IR that an optimization makes invalid raises Renderer.Compiler.Compile_error naming the verifier's finding instead of compiling. Earlier comgr compiles without it.
With Setting.ccache, code objects are cached in the table compile_hip_A_D, where A is arch and D the digest of comgr's library (C.identity), the program that runs it, and the options it compiles and links with: a code object is read back only for its source, compiled by the same comgr with the same options. Raises Failure naming COMGR_PATH if no comgr is found. Renderer.Compiler.compile raises Renderer.Compiler.Compile_error with comgr's status and log if comgr rejects the source, with the reason if comgr cannot be loaded, and with how the process ended and what it printed if it ends without answering, and Failure if the process cannot be started. Renderer.Compiler.disassemble prints what llvm-objdump -d prints (Helpers.amdgpu_disassemble).