From d42fd59464fef92934a082bab24855fd6f6177cb Mon Sep 17 00:00:00 2001 From: rmatif <66360289+rmatif@users.noreply.github.com> Date: Mon, 30 Jun 2025 17:32:23 +0200 Subject: [PATCH] feat: add OpenCL backend support (#680) --- CMakeLists.txt | 7 +++++ README.md | 69 +++++++++++++++++++++++++++++++++++++++++++- common.hpp | 2 +- esrgan.hpp | 4 +-- ggml | 2 +- ggml_extend.hpp | 7 ++++- model.cpp | 4 +++ stable-diffusion.cpp | 8 +++++ tae.hpp | 2 +- upscaler.cpp | 4 +++ 10 files changed, 102 insertions(+), 7 deletions(-) diff --git a/CMakeLists.txt b/CMakeLists.txt index 782a893..06de0d5 100644 --- a/CMakeLists.txt +++ b/CMakeLists.txt @@ -28,6 +28,7 @@ option(SD_CUDA "sd: cuda backend" OFF) option(SD_HIPBLAS "sd: rocm backend" OFF) option(SD_METAL "sd: metal backend" OFF) option(SD_VULKAN "sd: vulkan backend" OFF) +option(SD_OPENCL "sd: opencl backend" OFF) option(SD_SYCL "sd: sycl backend" OFF) option(SD_MUSA "sd: musa backend" OFF) option(SD_FAST_SOFTMAX "sd: x1.5 faster softmax, indeterministic (sometimes, same seed don't generate same image), cuda only" OFF) @@ -52,6 +53,12 @@ if (SD_VULKAN) add_definitions(-DSD_USE_VULKAN) endif () +if (SD_OPENCL) + message("-- Use OpenCL as backend stable-diffusion") + set(GGML_OPENCL ON) + add_definitions(-DSD_USE_OPENCL) +endif () + if (SD_HIPBLAS) message("-- Use HIPBLAS as backend stable-diffusion") set(GGML_HIP ON) diff --git a/README.md b/README.md index 232e881..8ac08a0 100644 --- a/README.md +++ b/README.md @@ -22,7 +22,7 @@ Inference of Stable Diffusion and Flux in pure C/C++ - Accelerated memory-efficient CPU inference - Only requires ~2.3GB when using txt2img with fp16 precision to generate a 512x512 image, enabling Flash Attention just requires ~1.8GB. - AVX, AVX2 and AVX512 support for x86 architectures -- Full CUDA, Metal, Vulkan and SYCL backend for GPU acceleration. +- Full CUDA, Metal, Vulkan, OpenCL and SYCL backend for GPU acceleration. - Can load ckpt, safetensors and diffusers models/checkpoints. Standalone VAEs models - No need to convert to `.ggml` or `.gguf` anymore! - Flash Attention for memory usage optimization @@ -160,6 +160,73 @@ cmake .. -DSD_VULKAN=ON cmake --build . --config Release ``` +##### Using OpenCL (for Adreno GPU) + +Currently, it supports only Adreno GPUs and is primarily optimized for Q4_0 type + +To build for Windows ARM please refers to [Windows 11 Arm64 +](https://github.com/ggml-org/llama.cpp/blob/master/docs/backend/OPENCL.md#windows-11-arm64) + +Building for Android: + + Android NDK: + Download and install the Android NDK from the [official Android developer site](https://developer.android.com/ndk/downloads). + +Setup OpenCL Dependencies for NDK: + +You need to provide OpenCL headers and the ICD loader library to your NDK sysroot. + +* OpenCL Headers: + ```bash + # In a temporary working directory + git clone https://github.com/KhronosGroup/OpenCL-Headers + cd OpenCL-Headers + # Replace with your actual NDK installation path + # e.g., cp -r CL /path/to/android-ndk-r26c/toolchains/llvm/prebuilt/linux-x86_64/sysroot/usr/include + sudo cp -r CL /toolchains/llvm/prebuilt/linux-x86_64/sysroot/usr/include + cd .. + ``` + +* OpenCL ICD Loader: + ```bash + # In the same temporary working directory + git clone https://github.com/KhronosGroup/OpenCL-ICD-Loader + cd OpenCL-ICD-Loader + mkdir build_ndk && cd build_ndk + + # Replace in the CMAKE_TOOLCHAIN_FILE and OPENCL_ICD_LOADER_HEADERS_DIR + cmake .. -G Ninja -DCMAKE_BUILD_TYPE=Release \ + -DCMAKE_TOOLCHAIN_FILE=/build/cmake/android.toolchain.cmake \ + -DOPENCL_ICD_LOADER_HEADERS_DIR=/toolchains/llvm/prebuilt/linux-x86_64/sysroot/usr/include \ + -DANDROID_ABI=arm64-v8a \ + -DANDROID_PLATFORM=24 \ + -DANDROID_STL=c++_shared + + ninja + # Replace + # e.g., cp libOpenCL.so /path/to/android-ndk-r26c/toolchains/llvm/prebuilt/linux-x86_64/sysroot/usr/lib/aarch64-linux-android + sudo cp libOpenCL.so /toolchains/llvm/prebuilt/linux-x86_64/sysroot/usr/lib/aarch64-linux-android + cd ../.. + ``` + +Build `stable-diffusion.cpp` for Android with OpenCL: + +```bash +mkdir build-android && cd build-android + +# Replace with your actual NDK installation path +# e.g., -DCMAKE_TOOLCHAIN_FILE=/path/to/android-ndk-r26c/build/cmake/android.toolchain.cmake +cmake .. -G Ninja \ + -DCMAKE_TOOLCHAIN_FILE=/build/cmake/android.toolchain.cmake \ + -DANDROID_ABI=arm64-v8a \ + -DANDROID_PLATFORM=android-28 \ + -DGGML_OPENMP=OFF \ + -DSD_OPENCL=ON + +ninja +``` +*(Note: Don't forget to include `LD_LIBRARY_PATH=/vendor/lib64` in your command line before running the binary)* + ##### Using SYCL Using SYCL makes the computation run on the Intel GPU. Please make sure you have installed the related driver and [Intel® oneAPI Base toolkit](https://www.intel.com/content/www/us/en/developer/tools/oneapi/base-toolkit.html) before start. More details and steps can refer to [llama.cpp SYCL backend](https://github.com/ggerganov/llama.cpp/blob/master/docs/backend/SYCL.md#linux). diff --git a/common.hpp b/common.hpp index 337b4a0..b20c60f 100644 --- a/common.hpp +++ b/common.hpp @@ -56,7 +56,7 @@ public: // x: [N, channels, h, w] auto conv = std::dynamic_pointer_cast(blocks["conv"]); - x = ggml_upscale(ctx, x, 2); // [N, channels, h*2, w*2] + x = ggml_upscale(ctx, x, 2, GGML_SCALE_MODE_NEAREST); // [N, channels, h*2, w*2] x = conv->forward(ctx, x); // [N, out_channels, h*2, w*2] return x; } diff --git a/esrgan.hpp b/esrgan.hpp index 989d15f..5cbb4ad 100644 --- a/esrgan.hpp +++ b/esrgan.hpp @@ -130,8 +130,8 @@ public: body_feat = conv_body->forward(ctx, body_feat); feat = ggml_add(ctx, feat, body_feat); // upsample - feat = lrelu(ctx, conv_up1->forward(ctx, ggml_upscale(ctx, feat, 2))); - feat = lrelu(ctx, conv_up2->forward(ctx, ggml_upscale(ctx, feat, 2))); + feat = lrelu(ctx, conv_up1->forward(ctx, ggml_upscale(ctx, feat, 2, GGML_SCALE_MODE_NEAREST))); + feat = lrelu(ctx, conv_up2->forward(ctx, ggml_upscale(ctx, feat, 2, GGML_SCALE_MODE_NEAREST))); auto out = conv_last->forward(ctx, lrelu(ctx, conv_hr->forward(ctx, feat))); return out; } diff --git a/ggml b/ggml index ff90529..9e4bee1 160000 --- a/ggml +++ b/ggml @@ -1 +1 @@ -Subproject commit ff9052988b76e137bcf92bb335733933ca196ac0 +Subproject commit 9e4bee1c5afc2d677a5b32ecb90cbdb483e81fff diff --git a/ggml_extend.hpp b/ggml_extend.hpp index 101a5d1..e20fc4f 100644 --- a/ggml_extend.hpp +++ b/ggml_extend.hpp @@ -39,6 +39,10 @@ #include "ggml-vulkan.h" #endif +#ifdef SD_USE_OPENCL +#include "ggml-opencl.h" +#endif + #ifdef SD_USE_SYCL #include "ggml-sycl.h" #endif @@ -113,7 +117,8 @@ __STATIC_INLINE__ struct ggml_tensor* ggml_kronecker(ggml_context* ctx, struct g a->ne[0] * b->ne[0], a->ne[1] * b->ne[1], a->ne[2] * b->ne[2], - a->ne[3] * b->ne[3]), + a->ne[3] * b->ne[3], + GGML_SCALE_MODE_NEAREST), b); } diff --git a/model.cpp b/model.cpp index 24da39f..3e0bde7 100644 --- a/model.cpp +++ b/model.cpp @@ -26,6 +26,10 @@ #include "ggml-vulkan.h" #endif +#ifdef SD_USE_OPENCL +#include "ggml-opencl.h" +#endif + #define ST_HEADER_SIZE_LEN 8 uint64_t read_u64(uint8_t* buffer) { diff --git a/stable-diffusion.cpp b/stable-diffusion.cpp index 5522287..e7aa441 100644 --- a/stable-diffusion.cpp +++ b/stable-diffusion.cpp @@ -181,6 +181,14 @@ public: LOG_WARN("Failed to initialize Vulkan backend"); } #endif +#ifdef SD_USE_OPENCL + LOG_DEBUG("Using OpenCL backend"); + // ggml_log_set(ggml_log_callback_default, nullptr); // Optional ggml logs + backend = ggml_backend_opencl_init(); + if (!backend) { + LOG_WARN("Failed to initialize OpenCL backend"); + } +#endif #ifdef SD_USE_SYCL LOG_DEBUG("Using SYCL backend"); backend = ggml_backend_sycl_init(0); diff --git a/tae.hpp b/tae.hpp index c458b87..678c44c 100644 --- a/tae.hpp +++ b/tae.hpp @@ -149,7 +149,7 @@ public: if (i == 1) { h = ggml_relu_inplace(ctx, h); } else { - h = ggml_upscale(ctx, h, 2); + h = ggml_upscale(ctx, h, 2, GGML_SCALE_MODE_NEAREST); } continue; } diff --git a/upscaler.cpp b/upscaler.cpp index 0c11b66..1372134 100644 --- a/upscaler.cpp +++ b/upscaler.cpp @@ -28,6 +28,10 @@ struct UpscalerGGML { LOG_DEBUG("Using Vulkan backend"); backend = ggml_backend_vk_init(0); #endif +#ifdef SD_USE_OPENCL + LOG_DEBUG("Using OpenCL backend"); + backend = ggml_backend_opencl_init(); +#endif #ifdef SD_USE_SYCL LOG_DEBUG("Using SYCL backend"); backend = ggml_backend_sycl_init(0);