Skip to content

feat(hip-kernel-provider): add rocKE conv engine with ML heuristic #8982

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Open
cderb wants to merge 10 commits into
base: develop
Choose a base branch
from
Open

feat(hip-kernel-provider): add rocKE conv engine with ML heuristic #8982

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
10 commits
Select commit Hold shift + click to select a range
641ee83
[hip-kernel-provider] Add rocKE conv engine with ML heuristic and JIT…
cderb Jun 27, 2026
bf92761
[rocKE] conv sweep: C++ fork+exec isolation; replace Python sweep path
cderb Jun 30, 2026
5a810ee
Merge remote-tracking branch 'origin/develop' into users/cderb/rocke-…
cderb Jun 30, 2026
23e789f
style(hip-kernel-provider): apply black + clang-format; add gfx90a di…
cderb Jun 30, 2026
d6eee6a
chore(hip-kernel-provider): fix pre-commit end-of-file and cmake-lint
cderb Jun 30, 2026
80f384f
fix(hip-kernel-provider): review fixes for rocKE conv engine and heur…
cderb Jun 30, 2026
83cbc44
refactor(hip-kernel-provider): optimize rocKE conv engine and heurist…
cderb Jun 30, 2026
46a90ac
style(hip-kernel-provider): apply black + clang-format to optimizatio…
cderb Jun 30, 2026
5ad3cd2
Merge remote-tracking branch 'origin/develop' into users/cderb/rocke-…
cderb Jul 1, 2026
7ad9c09
fix(hip-kernel-provider): heap overflow in ConvMLHeuristic + stale _v…
cderb Jul 1, 2026
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
33 changes: 33 additions & 0 deletions dnn-providers/hip-kernel-provider/CMakeLists.txt
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -43,6 +43,7 @@ option(HIPKERNELPROVIDER_ENABLE_TESTS "Build tests" ON)
option(HIPKERNELPROVIDER_ENABLE_ROCKE "Build rocKE engine and smoke tests" OFF)
option(ENABLE_ASM_SDPA_ENGINE "Build the ASM SDPA engine" ON)
option(ENABLE_HIP_MLOPS_ENGINE "Build the HIP kernel engine" ON)
option(ENABLE_ROCKE_CONV_ENGINE "Build the rocKE implicit-GEMM conv engine" OFF)

option(BUILD_ADDRESS_SANITIZER "Build with Address Sanitizer enabled" OFF)
option(ENABLE_CLANG_FORMAT "Enable Clang-Format checks" ON)
Expand Down Expand Up @@ -231,3 +232,35 @@ if(ENABLE_ASM_SDPA_ENGINE)
DESTINATION ${HIPDNN_RELATIVE_INSTALL_PLUGIN_ENGINE_DIR}/hip_kernel_provider/asm_kernels
FILES_MATCHING PATTERN "*.co")
endif()

if(ENABLE_ROCKE_CONV_ENGINE)
# Decompress .lgbm.gz model files and install alongside feature_spec.json.
# The plain .lgbm is not committed; gunzip once at install time.
set(_rocke_models_src
"${CMAKE_CURRENT_SOURCE_DIR}/rocke/platform/Python/rocke/heuristics/models")
set(_rocke_models_dst
"${HIPDNN_RELATIVE_INSTALL_PLUGIN_ENGINE_DIR}/hip_kernel_provider/rocke_models")
install(CODE "
file(GLOB _arch_dirs LIST_DIRECTORIES true \"${_rocke_models_src}/grouped_conv_forward_fp16_*\")
foreach(_src_dir \${_arch_dirs})
get_filename_component(_arch_name \${_src_dir} NAME)
set(_dst_dir \"\$ENV{DESTDIR}\${CMAKE_INSTALL_PREFIX}/${_rocke_models_dst}/\${_arch_name}\")
file(MAKE_DIRECTORY \"\${_dst_dir}\")
set(_gz \"\${_src_dir}/model_tflops.lgbm.gz\")
if(EXISTS \"\${_gz}\")
message(STATUS \"Decompressing \${_gz}\")
execute_process(
COMMAND gunzip -c \"\${_gz}\"
OUTPUT_FILE \"\${_dst_dir}/model_tflops.lgbm\"
RESULT_VARIABLE _ret)
if(NOT _ret EQUAL 0)
message(WARNING \"gunzip failed for \${_gz} (exit \${_ret})\")
endif()
endif()
set(_spec \"\${_src_dir}/feature_spec.json\")
if(EXISTS \"\${_spec}\")
file(COPY \"\${_spec}\" DESTINATION \"\${_dst_dir}\")
endif()
endforeach()
")
endif()
6 changes: 6 additions & 0 deletions dnn-providers/hip-kernel-provider/rocke/platform/CMakeLists.txt
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -149,6 +149,7 @@ set(ROCKE_PYTHON_INSTALL_DIR "${_ROCKE_DEFAULT_PYTHON_INSTALL_DIR}"
# pybind bindings under Cpp/bindings are a separate target and are excluded.
file(GLOB_RECURSE ROCKE_CORE_SOURCES CONFIGURE_DEPENDS "${CMAKE_CURRENT_SOURCE_DIR}/Cpp/*.cpp")
list(FILTER ROCKE_CORE_SOURCES EXCLUDE REGEX "/Cpp/bindings/")
list(FILTER ROCKE_CORE_SOURCES EXCLUDE REGEX "/Cpp/tools/")

add_library(rocke_core STATIC ${ROCKE_CORE_SOURCES})
target_include_directories(rocke_core PUBLIC
Expand Down Expand Up @@ -191,6 +192,11 @@ if(ROCKE_BUILD_PYBIND)
add_subdirectory(Cpp/bindings)
endif()

option(ROCKE_BUILD_CONV_SWEEP "Build the rocke_conv_sweep candidate sweep tool" OFF)
if(ROCKE_BUILD_CONV_SWEEP)
add_subdirectory(Cpp/tools/conv_sweep)
endif()

# --- Installable engine + test artifacts -------------------------------------
install(TARGETS rocke_core
EXPORT ckcTargets
Expand Down
Loading
Loading