Skip to content

Commit

Permalink
gfx906/MI50/MI60/Vega VII rocBLAS fix
Browse files Browse the repository at this point in the history
DTensile_CODE_OBJECT_VERSION=V5 caused the
pytorch benchmark on gfx906 to fail on error:

:0:rocm_sdk_builder_gfx906/src_projects/clr/hipamd/src/hip_global.cpp:114 : 113508223548 us: [pid:7768  tid:0x7f19583a7040] Cannot find Symbol with name: Cijk_Alik_Bljk_SB_MT32x32x16_SE_APM1_AF0EM1_AF1EM1_AMAS0_ASAE01_ASCE01_ASEM1_BL1_BS1_DTLA0_DTLB0_EPS0_FL1_GLVWA4_GLVWB4_GRVW4_GSU1_GSUASB_ISA906_IU1_K1_KLA_LPA1_LPB1_LDL1_LRVW1_MDA2_MMFGLC_NLCA1_NLCB1_ONLL1_PK0_PGR0_PLR1_SIA1_SU32_SUM0_SUS256_SVW4_SNLL0_TT2_4_USFGRO1_VAW1_VS1_VW1_VWB1_WG16_8_1_WGM8

fixes: #180

Signed-off-by: Mika Laitio <[email protected]>
  • Loading branch information
lamikr committed Dec 17, 2024
1 parent 8c244d8 commit 84faa05
Showing 1 changed file with 14 additions and 17 deletions.
31 changes: 14 additions & 17 deletions binfo/core/023_02_rocBLAS.binfo
Original file line number Diff line number Diff line change
Expand Up @@ -6,15 +6,14 @@ BINFO_APP_UPSTREAM_REPO_URL=https://github.com/ROCm/rocBLAS.git
#use default git tag
#BINFO_APP_UPSTREAM_REPO_VERSION_TAG=rocm-5.7.1

# execution of tensile is very slow and errorprone. It should work with these settings
# with special tensile version where I have added some error checking to places which could crash.
# I also did not manage build with hip-clang but building with hcc worked.
# (in the end of tensile commands happens som errors)
# If tensile is disabled from the build, rocBLAS libraries were missing some functionality needed other libs/tensorflow build afterwards
# rocBLAS can be build either with or without Tensile.
# If rocBLAS is however build without Tensile, then libraries would miss some features needed by other apps like Tensorflow.
# Tensile specific parameters are however very error-prone causing easily build failures and other problems,
# so be careufull if changing these paramters.
# I have had also problem of buildng by using the hip-clang , while hcc has worked ok.
# Some functionality in client/test/benchmars apps seems to also require linking againsts blis instead of rocBLAS.

CFG_TEMP1=-DAMDGPU_TARGETS="${SEMICOLON_SEPARATED_GPU_TARGET_LIST_DEFAULT}"
#CFG_TEMP2=-DTensile_ARCHITECTURE="${SPACE_SEPARATED_GPU_TARGET_LIST_DEFAULT}"
#CFG_TEMP2=-DTensile_ARCHITECTURE="navi21;rembrandt"

BINFO_APP_CMAKE_CFG="-DCMAKE_INSTALL_PREFIX=${INSTALL_DIR_PREFIX_SDK_ROOT}"
BINFO_APP_CMAKE_CFG="${BINFO_APP_CMAKE_CFG} -DCMAKE_PREFIX_PATH=${INSTALL_DIR_PREFIX_SDK_ROOT}/lib64/cmake;${INSTALL_DIR_PREFIX_SDK_ROOT}/lib/cmake"
Expand All @@ -26,33 +25,31 @@ BINFO_APP_CMAKE_CFG="${BINFO_APP_CMAKE_CFG} -DROCM_PATH=${INSTALL_DIR_PREFIX_SDK
#BINFO_APP_CMAKE_CFG="${BINFO_APP_CMAKE_CFG} -DLINK_BLIS=ON"
#BINFO_APP_CMAKE_CFG="${BINFO_APP_CMAKE_CFG} -DBLIS_DIR=${INSTALL_DIR_PREFIX_SDK_ROOT}"
BINFO_APP_CMAKE_CFG="${BINFO_APP_CMAKE_CFG} -DOPENBLAS_DIR=${INSTALL_DIR_PREFIX_SDK_ROOT}"
BINFO_APP_CMAKE_CFG="${BINFO_APP_CMAKE_CFG} -DTensile_CODE_OBJECT_VERSION=V5"
BINFO_APP_CMAKE_CFG="${BINFO_APP_CMAKE_CFG} -DTensile_LOGIC=asm_full"
BINFO_APP_CMAKE_CFG="${BINFO_APP_CMAKE_CFG} -DTensile_SEPARATE_ARCHITECTURES=ON"
BINFO_APP_CMAKE_CFG="${BINFO_APP_CMAKE_CFG} -DTensile_LAZY_LIBRARY_LOADING=ON"
#
# Warning!
# -DTensile_CODE_OBJECT_VERSION=V5
# causes pytorch benchmarks to fail on gfx906 (tested on rocm sdk builder 6.1.2)
#
BINFO_APP_CMAKE_CFG="${BINFO_APP_CMAKE_CFG} -DTensile_CODE_OBJECT_VERSION=default"
BINFO_APP_CMAKE_CFG="${BINFO_APP_CMAKE_CFG} -DTensile_MERGE_FILES=ON"
BINFO_APP_CMAKE_CFG="${BINFO_APP_CMAKE_CFG} -DBUILD_FILE_REORG_BACKWARD_COMPATIBILITY=ON"
BINFO_APP_CMAKE_CFG="${BINFO_APP_CMAKE_CFG} -DTensile_LIBRARY_FORMAT=msgpack"
BINFO_APP_CMAKE_CFG="${BINFO_APP_CMAKE_CFG} -DTensile_CPU_THREADS=${BUILD_CPU_COUNT_MODERATE}"
BINFO_APP_CMAKE_CFG="${BINFO_APP_CMAKE_CFG} -DBUILD_FILE_REORG_BACKWARD_COMPATIBILITY=ON"
BINFO_APP_CMAKE_CFG="${BINFO_APP_CMAKE_CFG} -DTensile_TEST_LOCAL_PATH=${SDK_SRC_ROOT_DIR}/Tensile"
BINFO_APP_CMAKE_CFG="${BINFO_APP_CMAKE_CFG} -DTENSILE_VERSION=4.40.0"
BINFO_APP_CMAKE_CFG="${BINFO_APP_CMAKE_CFG} -DCMAKE_C_COMPILER=${SDK_C_COMPILER_DEFAULT}"
BINFO_APP_CMAKE_CFG="${BINFO_APP_CMAKE_CFG} -DCMAKE_CXX_COMPILER=${SDK_CXX_COMPILER_DEFAULT}"
BINFO_APP_CMAKE_CFG="${BINFO_APP_CMAKE_CFG} -DRUN_HEADER_TESTING=OFF"
BINFO_APP_CMAKE_CFG="${BINFO_APP_CMAKE_CFG} -DCMAKE_BUILD_TYPE=Release"
BINFO_APP_CMAKE_CFG="${BINFO_APP_CMAKE_CFG} -DBUILD_CLIENTS_BENCHMARKS=ON"
BINFO_APP_CMAKE_CFG="${BINFO_APP_CMAKE_CFG} -DBUILD_CLIENTS_BENCHMARKS=OFF"
BINFO_APP_CMAKE_CFG="${BINFO_APP_CMAKE_CFG} -DBUILD_CLIENTS_SAMPLES=OFF"
BINFO_APP_CMAKE_CFG="${BINFO_APP_CMAKE_CFG} -DBUILD_CLIENTS_TESTS=OFF"
BINFO_APP_CMAKE_CFG="${BINFO_APP_CMAKE_CFG} ${BINFO_APP_SRC_DIR}"

##BINFO_APP_CMAKE_CFG="${BINFO_APP_CMAKE_CFG} -DHIP_CLANG_INCLUDE_PATH=${INSTALL_DIR_PREFIX_HIP_LLVM}/include"
#BINFO_APP_CMAKE_CFG="${BINFO_APP_CMAKE_CFG} -Damd_comgr_DIR=${INSTALL_DIR_PREFIX_SDK_ROOT}/lib/cmake/amd_comgr"
#BINFO_APP_CMAKE_CFG="${BINFO_APP_CMAKE_CFG} -DBUILD_WITH_TENSILE=ON"
##BINFO_APP_CMAKE_CFG="${BINFO_APP_CMAKE_CFG} -DBUILD_WITH_TENSILE_HOST=ON"
#BINFO_APP_CMAKE_CFG="${BINFO_APP_CMAKE_CFG} ${BINFO_APP_SRC_DIR}"

BINFO_APP_POST_INSTALL_CMD_ARRAY=(
"if [ ! -e ${INSTALL_DIR_PREFIX_SDK_ROOT}/include/rocblas.h ]; then ln -s ${INSTALL_DIR_PREFIX_SDK_ROOT}/include/rocblas/rocblas.h ${INSTALL_DIR_PREFIX_SDK_ROOT}/include/rocblas.h; fi"
"if [ ! -e ${INSTALL_DIR_PREFIX_SDK_ROOT}/include/internal ]; then ln -s ${INSTALL_DIR_PREFIX_SDK_ROOT}/include/rocblas/internal ${INSTALL_DIR_PREFIX_SDK_ROOT}/include/internal; fi"
)
#"unset LDFLAGS"

0 comments on commit 84faa05

Please sign in to comment.