Publish baldeagleos-repo candidate 0000000325-20260922-184323

1 files changed, 83 insertions, 40 deletions

diff --git a/sci-misc/stable-diffusion-cpp/stable-diffusion-cpp-9999.ebuild b/sci-misc/stable-diffusion-cpp/stable-diffusion-cpp-9999.ebuild
index e301a0baf3bd..62ff471d23d0 100644
--- a/sci-misc/stable-diffusion-cpp/stable-diffusion-cpp-9999.ebuild
+++ b/sci-misc/stable-diffusion-cpp/stable-diffusion-cpp-9999.ebuild
@@ -3,9 +3,12 @@
 
 EAPI=8
 
-ROCM_VERSION="6.3"
+ROCM_VERSION="6.1"
 
-inherit cmake cuda rocm linux-info
+inherit cmake cuda rocm linux-info toolchain-funcs
+
+MY_PN="${PN/-/.}"
+MY_P="${MY_PN}-${PV}"
 
 DESCRIPTION="Diffusion model(SD,Flux,Wan,Qwen Image,Z-Image,...) inference in pure C/C++"
 HOMEPAGE="https://github.com/leejet/stable-diffusion.cpp"
@@ -16,55 +19,77 @@ if [[ ${PV} == *9999* ]]; then
 else
 	MY_PV="b${PV#0_pre}"
 	SRC_URI="https://github.com/leejet/stable-diffusion.cpp/archive/refs/tags/${MY_PV}.tar.gz -> ${P}.tar.gz"
-	S="${WORKDIR}/stable-diffusion-cpp-${MY_PV}"
+	S="${WORKDIR}/${MY_P}"
 	KEYWORDS="~amd64"
 fi
 
 LICENSE="MIT"
 SLOT="0"
-CPU_FLAGS_X86=( avx avx2 f16c )
 
-IUSE="openblas blis rocm cuda opencl vulkan flexiblas webm webp wmma"
+# GGML CPU flags
+X86_CPU_FLAGS=(
+	amx_bf16
+	amx_int8
+	amx_tile
+	avx
+	avx2
+	avx512_bf16
+	avx512_vnni
+	avx512f
+	avx512vbmi
+	avx_vnni
+	bmi2
+	f16c
+	fma3
+	sse4_2
+)
+CPU_FLAGS=( "${X86_CPU_FLAGS[@]/#/cpu_flags_x86_}" )
+GGML_IUSE="${CPU_FLAGS[*]} blis cuda flexiblas openblas opencl +openmp rocm vulkan"
+unset X86_CPU_FLAGS CPU_FLAGS
 
+IUSE="${GGML_IUSE} webm webp"
+unset GGML_IUSE
 REQUIRED_USE="
 	?? (
 		openblas
 		blis
 		flexiblas
 	)
+	rocm? ( ${ROCM_REQUIRED_USE} )
 	webm? (
 		webp
 	)
-	wmma? (
-		rocm
-	)
 "
 
-CDEPEND="
+COMMON_DEPEND="
 	openblas? ( sci-libs/openblas:= )
+	openmp? ( llvm-runtimes/openmp:= )
 	blis? ( sci-libs/blis:= )
 	flexiblas? ( sci-libs/flexiblas:= )
 	rocm? (
 		>=dev-util/hip-${ROCM_VERSION}:=
 		>=sci-libs/hipBLAS-${ROCM_VERSION}:=
-		wmma? (
-			>=sci-libs/rocWMMA-${ROCM_VERSION}:=
-		)
 	)
 	cuda? ( dev-util/nvidia-cuda-toolkit:= )
 	webp? ( media-libs/libwebp )
 	webm? ( media-libs/libwebm )
 "
-DEPEND="${CDEPEND}
+DEPEND="${COMMON_DEPEND}
 	opencl? ( dev-util/opencl-headers )
-	vulkan? ( dev-util/vulkan-headers )
+	vulkan? (
+		dev-util/spirv-headers
+		dev-util/vulkan-headers
+	)
 "
-RDEPEND="${CDEPEND}
-	dev-python/numpy
+RDEPEND="${COMMON_DEPEND}
 	opencl? ( dev-libs/opencl-icd-loader )
 	vulkan? ( media-libs/vulkan-loader )
 "
-BDEPEND="media-libs/shaderc"
+BDEPEND="vulkan? ( media-libs/shaderc )"
+
+pkg_pretend() {
+	[[ ${MERGE_TYPE} != binary ]] && use openmp && tc-check-openmp
+}
 
 pkg_setup() {
 	if use rocm; then
@@ -75,6 +100,8 @@ pkg_setup() {
 			fi
 		fi
 	fi
+
+	[[ ${MERGE_TYPE} != binary ]] && use openmp && tc-check-openmp
 }
 
 src_prepare() {
@@ -84,23 +111,48 @@ src_prepare() {
 
 src_configure() {
 	local mycmakeargs=(
-		-DCMAKE_SKIP_BUILD_RPATH=ON
-		-DSD_BUILD_SHARED_LIBS=OFF
-		-DSD_SERVER_BUILD_FRONTEND=OFF # requires pnpm and network access
-		-DGGML_NATIVE=0	# don't set march
-		-DGGML_RPC=ON
-		-DGENTOO_REMOVE_CMAKE_BLAS_HACK=ON
-		-DSD_CUDA=$(usex cuda)
-		-DSD_OPENCL=$(usex opencl)
+		-DSD_BUILD_EXAMPLES=ON
 		-DSD_WEBP=$(usex webp)
 		-DSD_USE_SYSTEM_WEBP=$(usex webp)
 		-DSD_WEBM=$(usex webm)
 		-DSD_USE_SYSTEM_WEBM=$(usex webm)
+
+		# GGML features controls
+		-DSD_CUDA=$(usex cuda)
 		-DSD_VULKAN=$(usex vulkan)
+		-DSD_OPENCL=$(usex opencl)
+		-DSD_RPC=ON
+
+		-DSD_BUILD_SHARED_LIBS=OFF
+		-DSD_SERVER_BUILD_FRONTEND=OFF # requires pnpm and network access
+
+		# avoid clashing with sci-ml/ggml
+		-DCMAKE_INSTALL_INCLUDEDIR="include/${MY_PN}"
+		-DCMAKE_INSTALL_LIBDIR="$(get_libdir)/${MY_PN}"
+		-DCMAKE_INSTALL_RPATH="\$ORIGIN/../$(get_libdir)/${MY_PN};\$ORIGIN"
+	)
 
-		# avoid clashing with whisper.cpp
-		-DCMAKE_INSTALL_LIBDIR="${EPREFIX}/usr/$(get_libdir)/stable-diffusion.cpp"
-		-DCMAKE_INSTALL_RPATH="${EPREFIX}/usr/$(get_libdir)/stable-diffusion.cpp"
+	# GGML backends
+	mycmakeargs+=(
+		-DGGML_NATIVE=OFF	# don't set march
+
+		# CPU Flags
+		-DGGML_SSE42=$(usex cpu_flags_x86_sse4_2)
+		-DGGML_AVX=$(usex cpu_flags_x86_avx)
+		-DGGML_AVX_VNNI=$(usex cpu_flags_x86_avx_vnni)
+		-DGGML_AVX2=$(usex cpu_flags_x86_avx2)
+		-DGGML_BMI2=$(usex cpu_flags_x86_bmi2)
+		-DGGML_AVX512=$(usex cpu_flags_x86_avx512f)
+		-DGGML_AVX512_VBMI=$(usex cpu_flags_x86_avx512vbmi)
+		-DGGML_AVX512_VNNI=$(usex cpu_flags_x86_avx512_vnni)
+		-DGGML_AVX512_BF16=$(usex cpu_flags_x86_avx512_bf16)
+		-DGGML_FMA=$(usex cpu_flags_x86_fma3)
+		-DGGML_F16C=$(usex cpu_flags_x86_f16c)
+		-DGGML_AMX_TILE=$(usex cpu_flags_x86_amx_tile)
+		-DGGML_AMX_INT8=$(usex cpu_flags_x86_amx_int8)
+		-DGGML_AMX_BF16=$(usex cpu_flags_x86_amx_bf16)
+
+		-DGGML_OPENMP=$(usex openmp)
 	)
 
 	if use openblas ; then
@@ -129,21 +181,12 @@ src_configure() {
 	fi
 
 	if use rocm; then
-		rocm_use_hipcc
+		export HIPCXX="$(hipconfig -l)/clang" HIP_PATH="$(hipconfig -p)"
 		mycmakeargs+=(
-			-DSD_HIPBLAS=ON -DAMDGPU_TARGETS=$(get_amdgpu_flags) -DGPU_TARGETS=$(get_amdgpu_flags)
-			-DGGML_HIP_ROCWMMA_FATTN=$(usex wmma)
+			-DSD_HIPBLAS=ON
+			-DGPU_TARGETS=$(get_amdgpu_flags)
 		)
 	fi
 
 	cmake_src_configure
 }
-
-src_install() {
-	cmake_src_install
-
-	# avoid clashing with whisper.cpp
-	rm -rf "${ED}/usr/include"
-
-	find "${ED}" -name "*.a" -delete || die
-}