feat: package llama.cpp OpenCL backend for Qualcomm midstream

.github/local/Makefile.local

@@ -2,38 +2,46 @@ SRC_DIR        := src
 
 debuild: build
 
-.PHONY: pre_build_dep
-pre_build_dep:
-	sudo apt-get update
-	sudo apt-get install -y --no-install-recommends docker.io
-
 .PHONY: main_build
-main_build: build-qcom
+main_build: build-midstream
+
+# ===== Radxa Q6A/Q8B (OpenCL GPU, mainline kernel) =====
+PKG_MIDSTREAM   := llama.cpp-qcom-midstream
+BUILD_MIDSTREAM := $(SRC_DIR)/build-midstream-cross
+INSTALL_MIDSTREAM := $(SRC_DIR)/install-midstream-cross
+OPENCL_TARBALL  := qcom-adreno_1.838.3_qcm6490.tar.gz
+OPENCL_URL      := https://softwarecenter.qualcomm.com/nexus/generic/software/chip/qualcomm_linux-spf-1-0/qualcomm-linux-spf-1-0_test_device_public/r1.0_00114.0/le-qclinux-1-0-r1/apps_proc/prebuilt_HY22/qcom-adreno/251215/$(OPENCL_TARBALL)
+OPENCL_DIR      := _opencl
 
-# ===== Qualcomm Snapdragon (Hexagon DSP) =====
-PKG_QCOM       := llama.cpp-qcom
-BUILD_QCOM     := $(SRC_DIR)/build-qcom
-INSTALL_QCOM   := $(SRC_DIR)/install-qcom
-DOCKER_QCOM    := ghcr.io/snapdragon-toolchain/arm64-linux:v0.1
+$(OPENCL_TARBALL):
+	wget -q --tries=5 --timeout=60 --retry-connrefused \
+		"$(OPENCL_URL)" -O $@.tmp
+	mv $@.tmp $@
 
-.PHONY: build-qcom
-build-qcom:
-	cp $(SRC_DIR)/docs/backend/snapdragon/CMakeUserPresets.json $(SRC_DIR)/
-	sed -i 's/"version": 5/"version": 4/' $(SRC_DIR)/CMakeUserPresets.json
-	# Create container and copy source (docker cp avoids Docker-in-Docker volume mount issue)
-	docker create --name qcom-build --platform linux/amd64 $(DOCKER_QCOM) \
-		/bin/bash -c ' \
-			set -euo pipefail; \
-			cd /workspace; \
-			cmake --preset arm64-linux-snapdragon-release -B build-qcom; \
-			cmake --build build-qcom -j $$(nproc); \
-			cmake --install build-qcom --prefix install-qcom \
-		'
-	docker cp $(SRC_DIR)/. qcom-build:/workspace/
-	docker start -a qcom-build
-	rm -rf $(INSTALL_QCOM)
-	docker cp qcom-build:/workspace/install-qcom/. $(INSTALL_QCOM)/
-	docker rm qcom-build
-	rm -rf $(PKG_QCOM)
-	mkdir -p $(PKG_QCOM)/usr
-	cp -a $(INSTALL_QCOM)/* $(PKG_QCOM)/usr/
+.PHONY: build-midstream
+build-midstream: $(OPENCL_TARBALL)
+	rm -rf $(OPENCL_DIR)
+	mkdir -p $(OPENCL_DIR)
+	tar xzf "$(OPENCL_TARBALL)" -C $(OPENCL_DIR)
+	ln -sf usr/lib/libOpenCL.so.1 $(OPENCL_DIR)/libOpenCL.so
+	cmake -S $(SRC_DIR) -B $(BUILD_MIDSTREAM) -G Ninja \
+		-DCMAKE_BUILD_TYPE=Release \
+		-DCMAKE_SYSTEM_NAME=Linux \
+		-DCMAKE_SYSTEM_PROCESSOR=aarch64 \
+		-DCMAKE_C_COMPILER=aarch64-linux-gnu-gcc \
+		-DCMAKE_CXX_COMPILER=aarch64-linux-gnu-g++ \
+		-DBUILD_SHARED_LIBS=OFF \
+		-DGGML_NATIVE=OFF \
+		-DGGML_OPENCL=ON \
+		-DGGML_OPENCL_USE_ADRENO_KERNELS=ON \
+		-DGGML_OPENCL_EMBED_KERNELS=ON \
+		-DGGML_HEXAGON=OFF \
+		-DGGML_OPENCL_TARGET_VERSION=200 \
+		-DOpenCL_LIBRARY=$(abspath $(OPENCL_DIR))/libOpenCL.so \
+		-DOpenCL_INCLUDE_DIR=/usr/include
+	cmake --build $(BUILD_MIDSTREAM) -j $$(nproc)
+	rm -rf $(INSTALL_MIDSTREAM)
+	cmake --install $(BUILD_MIDSTREAM) --prefix $(abspath $(INSTALL_MIDSTREAM))
+	rm -rf $(PKG_MIDSTREAM)
+	mkdir -p $(PKG_MIDSTREAM)/usr
+	cp -a $(INSTALL_MIDSTREAM)/* $(PKG_MIDSTREAM)/usr/

.gitignore

@@ -17,9 +17,16 @@
 /*.changes
 /.pc
 
-# Output folder
-/out
+/llama.cpp-qcom-midstream
+
+# Build artifacts in src/
+/src/build-midstream-cross/
+/src/install-midstream-cross/
+
+# Cached OpenCL tarball
+/qcom-adreno_*.tar.gz
+/qcom-adreno_*.tar.gz.tmp
+/_opencl/
 
 # Logs
 /*.log
-

debian/.gitignore

@@ -8,3 +8,4 @@
 /*.substvars
 /tmp
 
+/llama.cpp-qcom-midstream

debian/changelog

@@ -1,3 +1,11 @@
+llama.cpp-qcom-midstream (0.0.7992-1) UNRELEASED; urgency=medium
+
+  * Initial packaging for Qualcomm mainline-kernel devices
+  * OpenCL GPU backend with Adreno optimized kernels
+  * Based on upstream tag b7992
+
+ -- "Radxa Computer Co., Ltd" <dev@radxa.com>  Mon, 20 Jul 2026 08:00:00 +0800
+
 llama.cpp (0.0.10034-1) unstable; urgency=medium
 
   [ Chongbiao Chen ]

debian/control

@@ -1,4 +1,4 @@
-Source: llama.cpp
+Source: llama.cpp-qcom-midstream
 Maintainer: "Radxa Computer Co., Ltd" <dev@radxa.com>
 Section: admin
 Priority: standard
@@ -9,13 +9,20 @@ Build-Depends: debhelper-compat (= 13),
                lintian,
                dh-exec,
                pandoc,
+               ca-certificates,
+               cmake,
+               crossbuild-essential-arm64,
+               ninja-build,
+               opencl-headers,
+               python3,
+               wget,
 
-Package: llama.cpp-qcom
+Package: llama.cpp-qcom-midstream
 Architecture: all
 Section: admin
 Priority: optional
 Depends: ${misc:Depends},
-         fastrpc,
-Description: llama.cpp for Qualcomm Snapdragon
- llama.cpp with Hexagon DSP backend for Qualcomm Snapdragon devices.
-
+         qcom-adreno-opencl-midstream,
+         python3,
+Description: llama.cpp for Qualcomm mainline (OpenCL GPU)
+ llama.cpp with OpenCL GPU backend for Qualcomm mainline-kernel devices.

debian/llama.cpp-qcom-midstream.install

@@ -0,0 +1 @@
+llama.cpp-qcom-midstream/usr /

debian/llama.cpp-qcom-midstream.lintian-overrides

@@ -0,0 +1,7 @@
+# Prebuilt cross-compiled binaries in Architecture: all package
+arch-independent-package-contains-binary-or-object
+missing-dependency-on-libc
+unstripped-binary-or-object
+no-manual-page
+pkg-config-unavailable-for-cross-compilation
+script-with-language-extension

pkg.conf.template

@@ -1,9 +1,18 @@
 {
   "*": {
     "Releases": [
-      "qcs9075-noble",
-      "qcs9075-noble-test"
+      "qcs6490-noble",
+      "qcs6490-noble-test",
+      "qcs6490-trixie",
+      "qcs6490-trixie-test",
+      "qcs6490-resolute",
+      "qcs6490-resolute-test",
+      "sc8280xp-noble",
+      "sc8280xp-noble-test",
+      "sc8280xp-trixie",
+      "sc8280xp-trixie-test",
+      "sc8280xp-resolute",
+      "sc8280xp-resolute-test"
     ]
   }
 }
-