CHIP-SPV/chipStar

PoCL - Device library link step failed

Closed this issue · 6 comments

pvelesko commented

The run: https://github.com/CHIP-SPV/chipStar/actions/runs/8030196087

These don't fail for me locally

list(APPEND CPU_POCL_FAILED_TESTS "Unit_deviceFunctions_CompileTest___dsqrt_rd_double") # Failed
list(APPEND CPU_POCL_FAILED_TESTS "Unit_deviceFunctions_CompileTest___dsqrt_rn_double") # Failed
list(APPEND CPU_POCL_FAILED_TESTS "Unit_deviceFunctions_CompileTest___dsqrt_ru_double") # Failed
list(APPEND CPU_POCL_FAILED_TESTS "Unit_deviceFunctions_CompileTest___dsqrt_rz_double") # Failedlist(APPEND CPU_POCL_FAILED_TESTS "fp16_math") # Failed
list(APPEND CPU_POCL_FAILED_TESTS "fp16_half2_math") # Failed
list(APPEND CPU_POCL_FAILED_TESTS "Unit_hipGraphAddMemcpyNodeToSymbol_MemcpyToSymbolNodeWithKernel") # Failed

CHIP error [TID 36443] [1708774709.417000043] : hipErrorNotInitialized (Device library link step failed.) in /home/runner/work/chipStar/chipStar/src/backend/OpenCL/CHIPBackendOpenCL.cc:840:compile

CHIP error [TID 36443] [1708774709.417225324] : Caught Error: hipErrorNotInitialized
pjaaskel commented

Do you mean there could be a bug in PoCL or the CI?

pvelesko commented

from UnitTests.cmake

# The following tests fail for LLVM 15 Debug & Release : Cannot find symbol _Z4sqrtDh in kernel library
list(APPEND CPU_POCL_FAILED_TESTS "Unit_deviceFunctions_CompileTest___dsqrt_rd_double") # Failed
list(APPEND CPU_POCL_FAILED_TESTS "Unit_deviceFunctions_CompileTest___dsqrt_rn_double") # Failed
list(APPEND CPU_POCL_FAILED_TESTS "Unit_deviceFunctions_CompileTest___dsqrt_ru_double") # Failed
list(APPEND CPU_POCL_FAILED_TESTS "Unit_deviceFunctions_CompileTest___dsqrt_rz_double") # Failed

# Fails for LLVM 15 Debug: SPIR-V Parser: Failed to find size for type id 83
list(APPEND CPU_POCL_FAILED_TESTS "Unit_deviceFunctions_CompileTest_rnorm_double") # Failed
franz commented

PoCL CPU device has FP16 support only when compiled with LLVM 16 and higher (and that support is quite incomplete).

pjaaskel commented

Do we need to support LLVM 15 still?

pvelesko commented

Do we need to support LLVM 15 still?

I don't think so

pjaaskel commented

OK. So this issue is either invalid or should be a PR to PoCL to add the missing fp16 bits, right?