#!/usr/bin/make -f

FEATURE_CHECK := $(if $(filter nocheck,$(DEB_BUILD_OPTIONS) $(DEB_BUILD_PROFILES)),OFF,ON)
FEATURE_INSTTEST := $(if $(filter noinsttest,$(DEB_BUILD_PROFILES)),OFF,ON)

export CXX=hipcc
export CC=hipcc
export DEB_BUILD_MAINT_OPTIONS = hardening=+all optimize=-lto
export DEB_CXXFLAGS_MAINT_PREPEND = -gz
export DPKG_GENSYMBOLS_CHECK_LEVEL=4
# -Wl,--gc-sections:
# Fixes linker errors ("relocation refers to a discarded section") when using gtest.
# The linker's section garbage collection incorrectly discards needed code due to
# a toolchain mismatch (ROCm's clang vs. the system's GCC). This flag forces the
# linker to correctly re-evaluate dependencies, keeping the required sections.
export DEB_LDFLAGS_MAINT_PREPEND = -Wl,--gc-sections
export VERBOSE=1
export TENSILE_ROCM_ASSEMBLER_PATH=/usr/bin/clang++
export TENSILE_ROCM_OFFLOAD_BUNDLER_PATH=/usr/bin/clang-offload-bundler
export PATH:=$(PATH):/usr/lib/llvm-22/bin
#export AMD_LOG_LEVEL=4
#--merge-architectures --no-lazy-library-loading
DEBUG_FLAGS=-DCMAKE_FIND_DEBUG_MODE=ON
DEBUG_FLAGS=

CXXFLAGS := $(subst -fstack-protector-strong,-Xarch_host -fstack-protector-strong,$(CXXFLAGS))
CXXFLAGS := $(subst -fcf-protection,-Xarch_host -fcf-protection,$(CXXFLAGS))

GPU_TARGETS := "$(shell rocm-target-arch --sep ';')"

CMAKE_FLAGS = \
	-DROCTX_PATH=/usr/lib/$(DEB_HOST_MULTIARCH) \
	-DCMAKE_BUILD_TYPE=Release \
	-DCMAKE_SKIP_INSTALL_RPATH=ON \
	-DROCM_SYMLINK_LIBS=OFF \
	-DBUILD_CLIENTS_BENCHMARKS=$(FEATURE_INSTTEST) \
	-DBUILD_FILE_REORG_BACKWARD_COMPATIBILITY=OFF \
	-DGPU_TARGETS=$(GPU_TARGETS) \
	-DBUILD_WITH_HIPBLASLT=OFF \
	-DROCBLAS_TENSILE_LIBRARY_DIR=/usr/lib/$(DEB_HOST_MULTIARCH)/rocblas/5 \
	-DINSTALL_TEST_DATA_DIR=/usr/share/librocblas5-tests/data \
	-DTensile_COMPILER=hipcc \
	-DTensile_TEST_LOCAL_PATH="../tensile" \
	-DTensile_ROOT="$(CURDIR)/tensile/Tensile" \
	-DTensile_LOGIC=asm_full \
	-DTensile_LIBRARY_FORMAT=msgpack \
	-DTensile_LAZY_LIBRARY_LOADING=ON \
	-DTensile_SEPARATE_ARCHITECTURES=ON \
	-DBUILD_CLIENTS_TESTS=$(if $(filter ON,$(FEATURE_CHECK) $(FEATURE_INSTTEST)),ON,OFF) \
	$(DEBUG_FLAGS) \
	-DBUILD_WITH_PIP=false \
	-DOptions=--ignore-asm-cap-cache \
	-DRUN_HEADER_TESTING=OFF \
	-DBUILD_OFFLOAD_COMPRESS=ON

%:
	dh $@

override_dh_auto_configure-arch:
	dh_auto_configure -- $(CMAKE_FLAGS)

# strip cannot handle code object files
override_dh_strip-arch:
	dh_strip --exclude=.hsaco --exclude=.co

override_dh_dwz-arch:
	:

# note: openblas crashes in dot for complex values with M=128, N=800000, incx=-3, incy=-3
override_dh_auto_test-arch:
ifeq (,$(filter nocheck,$(DEB_BUILD_OPTIONS)))
	set -e \
	; if [ -r /dev/kfd ] \
	; then obj-$(DEB_HOST_GNU_TYPE)/clients/staging/rocblas-test \
	    --gtest_filter="*checkin*:-*known_bug*" \
	; else echo "W: /dev/kfd unreadable: no available AMD GPU." \
	;      echo "W: tests skipped." \
	; fi
endif

override_dh_auto_configure-indep:
	:

override_dh_auto_build-indep:
ifeq ($(FEATURE_INSTTEST),ON)
	python3 clients/common/rocblas_gentest.py -I clients/include clients/gtest/rocblas_gtest.yaml -o rocblas_gtest.data
endif

execute_after_dh_auto_build-indep: export http_proxy=127.0.0.1:9
execute_after_dh_auto_build-indep: export https_proxy=127.0.0.1:9
execute_after_dh_auto_build-indep:
ifeq (,$(filter nodoc,$(DEB_BUILD_OPTIONS)))
	rocm-docs-build
endif

override_dh_auto_test-indep:
	:

override_dh_auto_install-indep:
	:

override_dh_gencontrol:
	dh_gencontrol -- -Vrocm:GPU-Architecture="$(shell rocm-target-arch)"