Clang 9 - Tripple amdgcn-amd-amdpal produces incomplete binary

[Lolliedieb]

Hello,

I am used to compile my kernels for AMD ROCm platform (Vega 64) with clang-9 using the following call:

clang-9 -x cl -Xclang -finclude-default-header
-target amdgcn-amd-amdhsa -mcpu=gfx900
-Xclang -mlink-bitcode-file -Xclang ./opencl/bitcode/opencl.amdgcn.bc
-Xclang -mlink-bitcode-file -Xclang ./opencl/bitcode/ocml.amdgcn.bc
-Xclang -mlink-bitcode-file -Xclang ./opencl/bitcode/ockl.amdgcn.bc
-Xclang -mlink-bitcode-file -Xclang ./opencl/bitcode/oclc_correctly_rounded_sqrt_off.amdgcn.bc
-Xclang -mlink-bitcode-file -Xclang ./opencl/bitcode/oclc_daz_opt_off.amdgcn.bc
-Xclang -mlink-bitcode-file -Xclang ./opencl/bitcode/oclc_finite_only_off.amdgcn.bc
-Xclang -mlink-bitcode-file -Xclang ./opencl/bitcode/oclc_unsafe_math_off.amdgcn.bc
-Xclang -mlink-bitcode-file -Xclang ./opencl/bitcode/oclc_wavefrontsize64_off.amdgcn.bc
-Xclang -mlink-bitcode-file -Xclang ./opencl/bitcode/oclc_isa_version_900.amdgcn.bc
kernel.cl -o kernel.so

The kernel builds and runs without a problem on ROCm 2.10 (Ubuntu 18.4). Changing the target to amdgcn-amd-amdpal works without any warning or error message, but the resulting binary file is only approx 2/3 the size (196.4 kb vs. 128.1 kb) and the kernels do not run (tested with amdgpu-pro 19.30, also Ubuntu). Disassembling the .so file shows valid gcn asm, but it seems all metadata, kernel names and so on are missing.

I am not sure if this is a bug or if I am doing something wrong - but tried to follow the instructions of the OpenCL back end as close as possible.

Ps: The kernels do not use any inline ASM or similar - its mostly stock OpenCL 1.2.

[arsenm]

amdpal kernels aren't intended to work with rocm. These use different binary metadata formats, so they aren't expected to be identically sized. Why are you trying to use amdpal with rocm?

[Lolliedieb]

Sry for the confusion. To make it short: no, I am not trying to run the pal kernel on rocm. Using the pal tripple I want to load the kernel with amdgpu-pro 19.30, but that fails even for simple kernels while the hsa tripple kernels seem fine for me inspecting their disassembly (not able to test that yet on rocm, but will soon. Only connection to rocm is that I used the bitcode files I found there since I found no other source for them. The code itself generated seems ok as said above, but the amdgpu-pro runtime fails loading the binary and looking to kernel dissassembly the metadata with kernel information do not appear, different to pal binaries compiled with the inregrated OpenCL offline compiler. I can provide a minimal example later today. Matt Arsenault <notifications@github.com> schrieb am Fr., 10. Jan. 2020, 00:26:

…

amdpal kernels aren't intended to work with rocm. These use different binary metadata formats, so they aren't expected to be identically sized. Why are you trying to use amdpal with rocm? — You are receiving this because you authored the thread. Reply to this email directly, view it on GitHub <#79?email_source=notifications&email_token=AJS63R23Z355LQWUJSDSVZ3Q46XDRA5CNFSM4J7GBMH2YY3PNVWWK3TUL52HS4DFVREXG43VMVBW63LNMVXHJKTDN5WW2ZLOORPWSZGOEISEVDQ#issuecomment-572803726>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/AJS63R7SAKVL7K2Y7EBJPLTQ46XDRANCNFSM4J7GBMHQ> .