whisper.cpp/CMakeLists.txt

578 lines
19 KiB
CMake
Raw Normal View History

whisper : Metal and ggml-alloc support (#1270) * metal : init * whisper : factor out graph builds * whisper : allocate encoder and decoder using ggml-alloc * whisper : ggml-alloc is now supported * whisper : CoreML support ggml-alloc * build : fix ggml-alloc * ios : update submodule * extra : update sync-ggml.sh script to also sync ggml-alloc * ci : see if this is causing the crash * whisper : refactor ggml-alloc init * whisper.android : try to fix build * whisper : initial Metal version * ci : try to debug vmem issue * metal : decoder works on GPU! * metal : add multi-decoder support * ggml : fix ggml_nbytes (probably temp solution) * metal : run "cross" step on the GPU * whisper : remove ggml_repeat in the encoder * whisper : offload the Encoder to Metal * ggml : use simpler ggml_bytes() implementation * ggml-alloc : try to make CI happy by reducing vram to 128GB * whisper : add whisper_allocr to wrap ggml_allocr * whisper : factor out alloc init in a function * cmake : update to support Metal build * whisper : add <functional> header * objc : fix build (no Metal yet) * ios : add Metal support * swiftui : fix build * metal : speed-up KQ multiplication * metal : sync latest llama.cpp kernels * readme : add Metal info * ios : update submodule * coreml : add code to toggle Core ML config (CPU, ANE, GPU) * bench : fix timings by running a pre-heat * bench : start benching the decoder * whisper : add ggml_mul_mat_pad * bench : fix uninitialized vars * whisper : add comment for disabling mul-mat padding * whisper : add description of ggml_mul_mat_pad * whisper : clean-up ggml_mul_mat_pad * metal : remove the "concurrent" flag * bench : variable n_past * ios : update SPM package
2023-09-15 11:18:18 +02:00
cmake_minimum_required (VERSION 3.5)
2022-12-17 18:52:42 +01:00
2024-01-05 16:11:27 +01:00
project(whisper.cpp VERSION 1.5.4)
2022-10-07 22:53:12 +02:00
# Add path to modules
list(APPEND CMAKE_MODULE_PATH "${CMAKE_CURRENT_SOURCE_DIR}/cmake/")
set(CMAKE_RUNTIME_OUTPUT_DIRECTORY ${CMAKE_BINARY_DIR}/bin)
2022-10-07 22:53:12 +02:00
if(CMAKE_SOURCE_DIR STREQUAL CMAKE_CURRENT_SOURCE_DIR)
set(WHISPER_STANDALONE ON)
include(GitVars)
include(BuildTypes)
2022-11-07 19:50:24 +01:00
# configure project version
if (EXISTS "${CMAKE_SOURCE_DIR}/bindings/ios/Makefile-tmpl")
configure_file(${CMAKE_SOURCE_DIR}/bindings/ios/Makefile-tmpl ${CMAKE_SOURCE_DIR}/bindings/ios/Makefile @ONLY)
endif()
configure_file(${CMAKE_SOURCE_DIR}/bindings/javascript/package-tmpl.json ${CMAKE_SOURCE_DIR}/bindings/javascript/package.json @ONLY)
2022-10-07 22:53:12 +02:00
else()
set(WHISPER_STANDALONE OFF)
endif()
if (EMSCRIPTEN)
set(BUILD_SHARED_LIBS_DEFAULT OFF)
option(WHISPER_WASM_SINGLE_FILE "whisper: embed WASM inside the generated whisper.js" ON)
else()
if (MINGW)
set(BUILD_SHARED_LIBS_DEFAULT OFF)
else()
set(BUILD_SHARED_LIBS_DEFAULT ON)
endif()
endif()
2022-10-07 22:53:12 +02:00
# options
whisper : Metal and ggml-alloc support (#1270) * metal : init * whisper : factor out graph builds * whisper : allocate encoder and decoder using ggml-alloc * whisper : ggml-alloc is now supported * whisper : CoreML support ggml-alloc * build : fix ggml-alloc * ios : update submodule * extra : update sync-ggml.sh script to also sync ggml-alloc * ci : see if this is causing the crash * whisper : refactor ggml-alloc init * whisper.android : try to fix build * whisper : initial Metal version * ci : try to debug vmem issue * metal : decoder works on GPU! * metal : add multi-decoder support * ggml : fix ggml_nbytes (probably temp solution) * metal : run "cross" step on the GPU * whisper : remove ggml_repeat in the encoder * whisper : offload the Encoder to Metal * ggml : use simpler ggml_bytes() implementation * ggml-alloc : try to make CI happy by reducing vram to 128GB * whisper : add whisper_allocr to wrap ggml_allocr * whisper : factor out alloc init in a function * cmake : update to support Metal build * whisper : add <functional> header * objc : fix build (no Metal yet) * ios : add Metal support * swiftui : fix build * metal : speed-up KQ multiplication * metal : sync latest llama.cpp kernels * readme : add Metal info * ios : update submodule * coreml : add code to toggle Core ML config (CPU, ANE, GPU) * bench : fix timings by running a pre-heat * bench : start benching the decoder * whisper : add ggml_mul_mat_pad * bench : fix uninitialized vars * whisper : add comment for disabling mul-mat padding * whisper : add description of ggml_mul_mat_pad * whisper : clean-up ggml_mul_mat_pad * metal : remove the "concurrent" flag * bench : variable n_past * ios : update SPM package
2023-09-15 11:18:18 +02:00
if (APPLE)
set(WHISPER_METAL_DEFAULT ON)
else()
set(WHISPER_METAL_DEFAULT OFF)
endif()
option(BUILD_SHARED_LIBS "whisper: build shared libs" ${BUILD_SHARED_LIBS_DEFAULT})
option(WHISPER_ALL_WARNINGS "whisper: enable all compiler warnings" ON)
option(WHISPER_ALL_WARNINGS_3RD_PARTY "whisper: enable all compiler warnings in 3rd party libs" OFF)
2022-10-07 22:53:12 +02:00
option(WHISPER_SANITIZE_THREAD "whisper: enable thread sanitizer" OFF)
option(WHISPER_SANITIZE_ADDRESS "whisper: enable address sanitizer" OFF)
option(WHISPER_SANITIZE_UNDEFINED "whisper: enable undefined sanitizer" OFF)
2022-10-07 22:53:12 +02:00
option(WHISPER_BUILD_TESTS "whisper: build tests" ${WHISPER_STANDALONE})
option(WHISPER_BUILD_EXAMPLES "whisper: build examples" ${WHISPER_STANDALONE})
2022-10-07 22:53:12 +02:00
option(WHISPER_SDL2 "whisper: support for libSDL2" OFF)
2022-10-07 22:53:12 +02:00
option(WHISPER_NO_AVX "whisper: disable AVX" OFF)
option(WHISPER_NO_AVX2 "whisper: disable AVX2" OFF)
option(WHISPER_NO_FMA "whisper: disable FMA" OFF)
option(WHISPER_NO_F16C "whisper: disable F16c" OFF)
option(WHISPER_OPENVINO "whisper: support for OpenVINO" OFF)
if (APPLE)
option(WHISPER_NO_ACCELERATE "whisper: disable Accelerate framework" OFF)
whisper : Metal and ggml-alloc support (#1270) * metal : init * whisper : factor out graph builds * whisper : allocate encoder and decoder using ggml-alloc * whisper : ggml-alloc is now supported * whisper : CoreML support ggml-alloc * build : fix ggml-alloc * ios : update submodule * extra : update sync-ggml.sh script to also sync ggml-alloc * ci : see if this is causing the crash * whisper : refactor ggml-alloc init * whisper.android : try to fix build * whisper : initial Metal version * ci : try to debug vmem issue * metal : decoder works on GPU! * metal : add multi-decoder support * ggml : fix ggml_nbytes (probably temp solution) * metal : run "cross" step on the GPU * whisper : remove ggml_repeat in the encoder * whisper : offload the Encoder to Metal * ggml : use simpler ggml_bytes() implementation * ggml-alloc : try to make CI happy by reducing vram to 128GB * whisper : add whisper_allocr to wrap ggml_allocr * whisper : factor out alloc init in a function * cmake : update to support Metal build * whisper : add <functional> header * objc : fix build (no Metal yet) * ios : add Metal support * swiftui : fix build * metal : speed-up KQ multiplication * metal : sync latest llama.cpp kernels * readme : add Metal info * ios : update submodule * coreml : add code to toggle Core ML config (CPU, ANE, GPU) * bench : fix timings by running a pre-heat * bench : start benching the decoder * whisper : add ggml_mul_mat_pad * bench : fix uninitialized vars * whisper : add comment for disabling mul-mat padding * whisper : add description of ggml_mul_mat_pad * whisper : clean-up ggml_mul_mat_pad * metal : remove the "concurrent" flag * bench : variable n_past * ios : update SPM package
2023-09-15 11:18:18 +02:00
option(WHISPER_METAL "whisper: use Metal" ${WHISPER_METAL_DEFAULT})
option(WHISPER_METAL_NDEBUG "whisper: disable Metal debugging" OFF)
option(WHISPER_COREML "whisper: enable Core ML framework" OFF)
option(WHISPER_COREML_ALLOW_FALLBACK "whisper: allow non-CoreML fallback" OFF)
else()
option(WHISPER_BLAS "whisper: use BLAS libraries" OFF)
option(WHISPER_BLAS_VENDOR "whisper: BLAS library vendor" Generic)
option(WHISPER_OPENBLAS "whisper: prefer OpenBLAS" OFF)
option(WHISPER_CUBLAS "whisper: support for cuBLAS" OFF)
option(WHISPER_HIPBLAS "whisper: support for hipBLAS" OFF)
option(WHISPER_CLBLAST "whisper: use CLBlast" OFF)
endif()
option(WHISPER_PERF "whisper: enable perf timings" OFF)
2022-10-18 17:51:59 +02:00
2022-10-07 22:53:12 +02:00
# sanitizers
2022-10-11 19:57:52 +02:00
if (NOT MSVC)
if (WHISPER_SANITIZE_THREAD)
set(CMAKE_C_FLAGS "${CMAKE_C_FLAGS} -fsanitize=thread")
2022-10-11 19:57:52 +02:00
set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -fsanitize=thread")
endif()
2022-10-07 22:53:12 +02:00
2022-10-11 19:57:52 +02:00
if (WHISPER_SANITIZE_ADDRESS)
set(CMAKE_C_FLAGS "${CMAKE_C_FLAGS} -fsanitize=address -fno-omit-frame-pointer")
set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -fsanitize=address -fno-omit-frame-pointer")
endif()
2022-10-07 22:53:12 +02:00
2022-10-11 19:57:52 +02:00
if (WHISPER_SANITIZE_UNDEFINED)
set(CMAKE_C_FLAGS "${CMAKE_C_FLAGS} -fsanitize=undefined")
set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -fsanitize=undefined")
endif()
2022-10-07 22:53:12 +02:00
endif()
#set(CMAKE_C_FLAGS "${CMAKE_C_FLAGS} -ffast-math")
#set(CMAKE_C_FLAGS "${CMAKE_C_FLAGS} -march=native")
# dependencies
find_package(Threads REQUIRED)
# on APPLE
if (APPLE)
# include Accelerate framework
if (NOT WHISPER_NO_ACCELERATE)
find_library(ACCELERATE_FRAMEWORK Accelerate)
if (ACCELERATE_FRAMEWORK)
message(STATUS "Accelerate framework found")
2022-10-18 17:51:59 +02:00
set(WHISPER_EXTRA_LIBS ${WHISPER_EXTRA_LIBS} ${ACCELERATE_FRAMEWORK})
set(WHISPER_EXTRA_FLAGS ${WHISPER_EXTRA_FLAGS} -DGGML_USE_ACCELERATE)
else()
message(FATAL_ERROR "Accelerate framework not found")
endif()
endif()
whisper : Metal and ggml-alloc support (#1270) * metal : init * whisper : factor out graph builds * whisper : allocate encoder and decoder using ggml-alloc * whisper : ggml-alloc is now supported * whisper : CoreML support ggml-alloc * build : fix ggml-alloc * ios : update submodule * extra : update sync-ggml.sh script to also sync ggml-alloc * ci : see if this is causing the crash * whisper : refactor ggml-alloc init * whisper.android : try to fix build * whisper : initial Metal version * ci : try to debug vmem issue * metal : decoder works on GPU! * metal : add multi-decoder support * ggml : fix ggml_nbytes (probably temp solution) * metal : run "cross" step on the GPU * whisper : remove ggml_repeat in the encoder * whisper : offload the Encoder to Metal * ggml : use simpler ggml_bytes() implementation * ggml-alloc : try to make CI happy by reducing vram to 128GB * whisper : add whisper_allocr to wrap ggml_allocr * whisper : factor out alloc init in a function * cmake : update to support Metal build * whisper : add <functional> header * objc : fix build (no Metal yet) * ios : add Metal support * swiftui : fix build * metal : speed-up KQ multiplication * metal : sync latest llama.cpp kernels * readme : add Metal info * ios : update submodule * coreml : add code to toggle Core ML config (CPU, ANE, GPU) * bench : fix timings by running a pre-heat * bench : start benching the decoder * whisper : add ggml_mul_mat_pad * bench : fix uninitialized vars * whisper : add comment for disabling mul-mat padding * whisper : add description of ggml_mul_mat_pad * whisper : clean-up ggml_mul_mat_pad * metal : remove the "concurrent" flag * bench : variable n_past * ios : update SPM package
2023-09-15 11:18:18 +02:00
if (WHISPER_METAL)
find_library(FOUNDATION_LIBRARY Foundation REQUIRED)
find_library(METAL_FRAMEWORK Metal REQUIRED)
find_library(METALKIT_FRAMEWORK MetalKit REQUIRED)
if (METAL_FRAMEWORK)
message(STATUS "Metal framework found")
set(WHISPER_EXTRA_LIBS ${WHISPER_EXTRA_LIBS}
${FOUNDATION_LIBRARY}
${METAL_FRAMEWORK}
${METALKIT_FRAMEWORK}
)
set(WHISPER_EXTRA_FLAGS ${WHISPER_EXTRA_FLAGS} -DGGML_USE_METAL)
if (WHISPER_METAL_NDEBUG)
set(WHISPER_EXTRA_FLAGS ${WHISPER_EXTRA_FLAGS} -DGGML_METAL_NDEBUG)
endif()
else()
message(FATAL_ERROR "Metal framework not found")
whisper : Metal and ggml-alloc support (#1270) * metal : init * whisper : factor out graph builds * whisper : allocate encoder and decoder using ggml-alloc * whisper : ggml-alloc is now supported * whisper : CoreML support ggml-alloc * build : fix ggml-alloc * ios : update submodule * extra : update sync-ggml.sh script to also sync ggml-alloc * ci : see if this is causing the crash * whisper : refactor ggml-alloc init * whisper.android : try to fix build * whisper : initial Metal version * ci : try to debug vmem issue * metal : decoder works on GPU! * metal : add multi-decoder support * ggml : fix ggml_nbytes (probably temp solution) * metal : run "cross" step on the GPU * whisper : remove ggml_repeat in the encoder * whisper : offload the Encoder to Metal * ggml : use simpler ggml_bytes() implementation * ggml-alloc : try to make CI happy by reducing vram to 128GB * whisper : add whisper_allocr to wrap ggml_allocr * whisper : factor out alloc init in a function * cmake : update to support Metal build * whisper : add <functional> header * objc : fix build (no Metal yet) * ios : add Metal support * swiftui : fix build * metal : speed-up KQ multiplication * metal : sync latest llama.cpp kernels * readme : add Metal info * ios : update submodule * coreml : add code to toggle Core ML config (CPU, ANE, GPU) * bench : fix timings by running a pre-heat * bench : start benching the decoder * whisper : add ggml_mul_mat_pad * bench : fix uninitialized vars * whisper : add comment for disabling mul-mat padding * whisper : add description of ggml_mul_mat_pad * whisper : clean-up ggml_mul_mat_pad * metal : remove the "concurrent" flag * bench : variable n_past * ios : update SPM package
2023-09-15 11:18:18 +02:00
endif()
set(GGML_SOURCES_METAL ggml-metal.m ggml-metal.h)
# copy ggml-metal.metal to bin directory
configure_file(ggml-metal.metal bin/ggml-metal.metal COPYONLY)
endif()
if (WHISPER_COREML)
find_library(FOUNDATION_FRAMEWORK Foundation)
find_library(COREML_FRAMEWORK CoreML)
if (COREML_FRAMEWORK)
message(STATUS "CoreML framework found")
set(WHISPER_EXTRA_FLAGS ${WHISPER_EXTRA_FLAGS} -DWHISPER_USE_COREML)
else()
message(FATAL_ERROR "CoreML framework not found")
endif()
if (WHISPER_COREML_ALLOW_FALLBACK)
set(WHISPER_EXTRA_FLAGS ${WHISPER_EXTRA_FLAGS} -DWHISPER_COREML_ALLOW_FALLBACK)
endif()
2022-10-18 17:51:59 +02:00
endif()
endif()
if (WHISPER_OPENBLAS)
set(WHISPER_BLAS_VENDOR "OpenBLAS")
set(WHISPER_BLAS ON)
endif()
if (WHISPER_BLAS)
if (WIN32)
if(DEFINED ENV{OPENBLAS_PATH})
set(BLAS_LIBRARIES $ENV{OPENBLAS_PATH}/lib/libopenblas.dll.a)
message(STATUS "Libraries ${BLAS_LIBRARIES}")
set(WHISPER_EXTRA_FLAGS ${WHISPER_EXTRA_FLAGS} -DGGML_USE_OPENBLAS)
include_directories($ENV{OPENBLAS_PATH}/include)
set(WHISPER_EXTRA_LIBS ${WHISPER_EXTRA_LIBS} ${BLAS_LIBRARIES})
else ()
message(FATAL_ERROR "BLAS library was not found. Environment variable OPENBLAS_PATH not defined.")
endif ()
else ()
set(BLA_STATIC 1)
set(BLA_VENDOR ${WHISPER_BLAS_VENDOR})
set(BLA_SIZEOF_INTEGER 8)
set(BLA_PREFER_PKGCONFIG 1)
find_package(BLAS)
if(BLAS_FOUND)
message(STATUS "BLAS compatible library found")
message(STATUS "Libraries ${BLAS_LIBRARIES}")
find_path(BLAS_INCLUDE_DIRS cblas.h /usr/include/openblas /usr/local/include/openblas $ENV{BLAS_HOME}/include)
set(WHISPER_EXTRA_FLAGS ${WHISPER_EXTRA_FLAGS} -DGGML_USE_OPENBLAS)
include_directories(${BLAS_INCLUDE_DIRS})
set(WHISPER_EXTRA_LIBS ${WHISPER_EXTRA_LIBS} ${BLAS_LIBRARIES})
else()
message(FATAL_ERROR "BLAS library was not found")
endif()
endif ()
endif ()
if (WHISPER_CUBLAS)
cmake_minimum_required(VERSION 3.17)
find_package(CUDAToolkit)
if (CUDAToolkit_FOUND)
message(STATUS "cuBLAS found")
enable_language(CUDA)
whisper : Metal and ggml-alloc support (#1270) * metal : init * whisper : factor out graph builds * whisper : allocate encoder and decoder using ggml-alloc * whisper : ggml-alloc is now supported * whisper : CoreML support ggml-alloc * build : fix ggml-alloc * ios : update submodule * extra : update sync-ggml.sh script to also sync ggml-alloc * ci : see if this is causing the crash * whisper : refactor ggml-alloc init * whisper.android : try to fix build * whisper : initial Metal version * ci : try to debug vmem issue * metal : decoder works on GPU! * metal : add multi-decoder support * ggml : fix ggml_nbytes (probably temp solution) * metal : run "cross" step on the GPU * whisper : remove ggml_repeat in the encoder * whisper : offload the Encoder to Metal * ggml : use simpler ggml_bytes() implementation * ggml-alloc : try to make CI happy by reducing vram to 128GB * whisper : add whisper_allocr to wrap ggml_allocr * whisper : factor out alloc init in a function * cmake : update to support Metal build * whisper : add <functional> header * objc : fix build (no Metal yet) * ios : add Metal support * swiftui : fix build * metal : speed-up KQ multiplication * metal : sync latest llama.cpp kernels * readme : add Metal info * ios : update submodule * coreml : add code to toggle Core ML config (CPU, ANE, GPU) * bench : fix timings by running a pre-heat * bench : start benching the decoder * whisper : add ggml_mul_mat_pad * bench : fix uninitialized vars * whisper : add comment for disabling mul-mat padding * whisper : add description of ggml_mul_mat_pad * whisper : clean-up ggml_mul_mat_pad * metal : remove the "concurrent" flag * bench : variable n_past * ios : update SPM package
2023-09-15 11:18:18 +02:00
set(GGML_SOURCES_CUDA ggml-cuda.cu ggml-cuda.h)
add_compile_definitions(GGML_USE_CUBLAS)
if (WHISPER_STATIC)
if (WIN32)
# As of 12.3.1 CUDA Tookit for Windows does not offer a static cublas library
set(WHISPER_EXTRA_LIBS ${WHISPER_EXTRA_LIBS} CUDA::cudart_static CUDA::cublas CUDA::cublasLt)
else ()
set(WHISPER_EXTRA_LIBS ${WHISPER_EXTRA_LIBS} CUDA::cudart_static CUDA::cublas_static CUDA::cublasLt_static)
endif()
else()
set(WHISPER_EXTRA_LIBS ${WHISPER_EXTRA_LIBS} CUDA::cudart CUDA::cublas CUDA::cublasLt)
endif()
set(WHISPER_EXTRA_LIBS ${WHISPER_EXTRA_LIBS} CUDA::cuda_driver)
else()
message(FATAL_ERROR "cuBLAS not found")
endif()
endif()
if (WHISPER_HIPBLAS)
list(APPEND CMAKE_PREFIX_PATH /opt/rocm)
if (NOT ${CMAKE_C_COMPILER_ID} MATCHES "Clang")
message(WARNING "Only LLVM is supported for HIP, hint: CC=/opt/rocm/llvm/bin/clang")
endif()
if (NOT ${CMAKE_CXX_COMPILER_ID} MATCHES "Clang")
message(WARNING "Only LLVM is supported for HIP, hint: CXX=/opt/rocm/llvm/bin/clang++")
endif()
find_package(hip)
find_package(hipblas)
find_package(rocblas)
if (${hipblas_FOUND} AND ${hip_FOUND})
message(STATUS "HIP and hipBLAS found")
add_compile_definitions(GGML_USE_HIPBLAS GGML_USE_CUBLAS)
add_library(ggml-rocm OBJECT ggml-cuda.cu ggml-cuda.h)
set_property(TARGET ggml-rocm PROPERTY POSITION_INDEPENDENT_CODE ON)
set_source_files_properties(ggml-cuda.cu PROPERTIES LANGUAGE CXX)
target_link_libraries(ggml-rocm PRIVATE hip::device PUBLIC hip::host roc::rocblas roc::hipblas)
if (WHISPER_STATIC)
message(FATAL_ERROR "Static linking not supported for HIP/ROCm")
endif()
set(WHISPER_EXTRA_LIBS ${WHISPER_EXTRA_LIBS} ggml-rocm)
else()
message(FATAL_ERROR "hipBLAS or HIP not found. Try setting CMAKE_PREFIX_PATH=/opt/rocm")
endif()
endif()
if (WHISPER_CLBLAST)
find_package(CLBlast)
if (CLBlast_FOUND)
message(STATUS "CLBlast found")
whisper : Metal and ggml-alloc support (#1270) * metal : init * whisper : factor out graph builds * whisper : allocate encoder and decoder using ggml-alloc * whisper : ggml-alloc is now supported * whisper : CoreML support ggml-alloc * build : fix ggml-alloc * ios : update submodule * extra : update sync-ggml.sh script to also sync ggml-alloc * ci : see if this is causing the crash * whisper : refactor ggml-alloc init * whisper.android : try to fix build * whisper : initial Metal version * ci : try to debug vmem issue * metal : decoder works on GPU! * metal : add multi-decoder support * ggml : fix ggml_nbytes (probably temp solution) * metal : run "cross" step on the GPU * whisper : remove ggml_repeat in the encoder * whisper : offload the Encoder to Metal * ggml : use simpler ggml_bytes() implementation * ggml-alloc : try to make CI happy by reducing vram to 128GB * whisper : add whisper_allocr to wrap ggml_allocr * whisper : factor out alloc init in a function * cmake : update to support Metal build * whisper : add <functional> header * objc : fix build (no Metal yet) * ios : add Metal support * swiftui : fix build * metal : speed-up KQ multiplication * metal : sync latest llama.cpp kernels * readme : add Metal info * ios : update submodule * coreml : add code to toggle Core ML config (CPU, ANE, GPU) * bench : fix timings by running a pre-heat * bench : start benching the decoder * whisper : add ggml_mul_mat_pad * bench : fix uninitialized vars * whisper : add comment for disabling mul-mat padding * whisper : add description of ggml_mul_mat_pad * whisper : clean-up ggml_mul_mat_pad * metal : remove the "concurrent" flag * bench : variable n_past * ios : update SPM package
2023-09-15 11:18:18 +02:00
set(GGML_SOURCES_OPENCL ggml-opencl.cpp ggml-opencl.h)
add_compile_definitions(GGML_USE_CLBLAST)
set(WHISPER_EXTRA_LIBS ${WHISPER_EXTRA_LIBS} clblast)
else()
message(FATAL_ERROR "CLBlast not found")
endif()
endif()
if( WHISPER_OPENVINO )
find_package(OpenVINO REQUIRED COMPONENTS Runtime)
endif()
2022-10-08 08:00:59 +02:00
# compiler flags
2022-10-07 22:53:12 +02:00
if (NOT CMAKE_BUILD_TYPE AND NOT CMAKE_CONFIGURATION_TYPES)
set(CMAKE_BUILD_TYPE Release CACHE STRING "Build type" FORCE)
set_property(CACHE CMAKE_BUILD_TYPE PROPERTY STRINGS "Debug" "Release" "RelWithDebInfo")
endif ()
if (WHISPER_ALL_WARNINGS)
2022-10-11 19:57:52 +02:00
if (NOT MSVC)
2022-10-07 22:53:12 +02:00
set(CMAKE_C_FLAGS "${CMAKE_C_FLAGS} \
-Wall \
-Wextra \
-Wpedantic \
-Wshadow \
-Wcast-qual \
-Wstrict-prototypes \
-Wpointer-arith \
-Wno-unused-function \
2022-10-07 22:53:12 +02:00
")
set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} \
-Wall \
-Wextra \
-Wpedantic \
-Wcast-qual \
")
2022-10-07 22:53:12 +02:00
else()
2022-10-11 19:57:52 +02:00
# todo : msvc
2022-10-07 22:53:12 +02:00
endif()
endif()
2022-10-11 19:57:52 +02:00
if (NOT MSVC)
set(CMAKE_C_FLAGS "${CMAKE_C_FLAGS} -Werror=vla")
#set(CMAKE_C_FLAGS "${CMAKE_C_FLAGS} -fno-math-errno -ffinite-math-only -funsafe-math-optimizations")
endif()
2022-10-07 22:53:12 +02:00
message(STATUS "CMAKE_SYSTEM_PROCESSOR: ${CMAKE_SYSTEM_PROCESSOR}")
if (${CMAKE_SYSTEM_PROCESSOR} MATCHES "arm" OR ${CMAKE_SYSTEM_PROCESSOR} MATCHES "aarch64")
message(STATUS "ARM detected")
elseif(${CMAKE_SYSTEM_PROCESSOR} MATCHES "ppc64le")
message(STATUS "PowerPC detected")
2022-10-07 22:53:12 +02:00
else()
message(STATUS "x86 detected")
2022-10-11 19:57:52 +02:00
if (MSVC)
set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} /utf-8")
set(CMAKE_CXX_FLAGS_RELEASE "${CMAKE_CXX_FLAGS_RELEASE} /utf-8")
set(CMAKE_C_FLAGS "${CMAKE_C_FLAGS} /utf-8")
if(NOT WHISPER_NO_AVX2)
set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} /arch:AVX2")
set(CMAKE_CXX_FLAGS_RELEASE "${CMAKE_CXX_FLAGS_RELEASE} /arch:AVX2")
set(CMAKE_C_FLAGS "${CMAKE_C_FLAGS} /arch:AVX2")
else()
if(NOT WHISPER_NO_AVX)
set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} /arch:AVX")
set(CMAKE_CXX_FLAGS_RELEASE "${CMAKE_CXX_FLAGS_RELEASE} /arch:AVX")
set(CMAKE_C_FLAGS "${CMAKE_C_FLAGS} /arch:AVX")
endif()
endif()
2022-10-11 19:57:52 +02:00
else()
2022-10-22 12:45:43 +02:00
if (EMSCRIPTEN)
set(CMAKE_C_FLAGS "${CMAKE_C_FLAGS} -pthread -s TOTAL_STACK=5242880")
set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -pthread -s TOTAL_STACK=5242880")
2022-10-22 12:45:43 +02:00
else()
2023-08-19 20:39:03 +02:00
if(NOT WHISPER_NO_AVX)
set(CMAKE_C_FLAGS "${CMAKE_C_FLAGS} -mavx")
endif()
2023-08-19 20:39:03 +02:00
if(NOT WHISPER_NO_AVX2)
set(CMAKE_C_FLAGS "${CMAKE_C_FLAGS} -mavx2")
endif()
2023-08-19 20:39:03 +02:00
if(NOT WHISPER_NO_FMA)
set(CMAKE_C_FLAGS "${CMAKE_C_FLAGS} -mfma")
endif()
2023-08-19 20:39:03 +02:00
if(NOT WHISPER_NO_F16C)
set(CMAKE_C_FLAGS "${CMAKE_C_FLAGS} -mf16c")
endif()
2022-10-22 12:45:43 +02:00
endif()
2022-10-11 19:57:52 +02:00
endif()
2022-10-07 22:53:12 +02:00
endif()
#
# POSIX conformance
#
# clock_gettime came in POSIX.1b (1993)
# CLOCK_MONOTONIC came in POSIX.1-2001 / SUSv3 as optional
# posix_memalign came in POSIX.1-2001 / SUSv3
# M_PI is an XSI extension since POSIX.1-2001 / SUSv3, came in XPG1 (1985)
add_compile_definitions(_XOPEN_SOURCE=600)
# Somehow in OpenBSD whenever POSIX conformance is specified
# some string functions rely on locale_t availability,
# which was introduced in POSIX.1-2008, forcing us to go higher
2023-09-08 13:50:03 +02:00
if (CMAKE_SYSTEM_NAME MATCHES "OpenBSD")
remove_definitions(-D_XOPEN_SOURCE=600)
add_compile_definitions(_XOPEN_SOURCE=700)
2023-09-08 13:50:03 +02:00
endif()
# Data types, macros and functions related to controlling CPU affinity
# are available on Linux through GNU extensions in libc
2023-09-08 13:50:03 +02:00
if (CMAKE_SYSTEM_NAME MATCHES "Linux")
add_compile_definitions(_GNU_SOURCE)
2023-09-08 13:50:03 +02:00
endif()
# RLIMIT_MEMLOCK came in BSD, is not specified in POSIX.1,
# and on macOS its availability depends on enabling Darwin extensions
# similarly on DragonFly, enabling BSD extensions is necessary
2023-09-08 13:50:03 +02:00
if (CMAKE_SYSTEM_NAME MATCHES "Darwin")
add_compile_definitions(_DARWIN_C_SOURCE)
2023-09-08 13:50:03 +02:00
endif()
if (CMAKE_SYSTEM_NAME MATCHES "DragonFly")
add_compile_definitions(_DARWIN_C_SOURCE)
2023-09-08 13:50:03 +02:00
endif()
# alloca is a non-standard interface that is not visible on BSDs when
# POSIX conformance is specified, but not all of them provide a clean way
# to enable it in such cases
2023-09-08 13:50:03 +02:00
if (CMAKE_SYSTEM_NAME MATCHES "FreeBSD")
add_compile_definitions(__BSD_VISIBLE)
2023-09-08 13:50:03 +02:00
endif()
if (CMAKE_SYSTEM_NAME MATCHES "NetBSD")
add_compile_definitions(_NETBSD_SOURCE)
2023-09-08 13:50:03 +02:00
endif()
if (CMAKE_SYSTEM_NAME MATCHES "OpenBSD")
add_compile_definitions(_BSD_SOURCE)
2023-09-08 13:50:03 +02:00
endif()
if (WHISPER_PERF)
set(WHISPER_EXTRA_FLAGS ${WHISPER_EXTRA_FLAGS} -DGGML_PERF)
endif()
#
# whisper.coreml - Core ML support
#
if (WHISPER_COREML)
set(TARGET whisper.coreml)
add_library(${TARGET}
coreml/whisper-encoder.h
coreml/whisper-encoder.mm
coreml/whisper-encoder-impl.h
coreml/whisper-encoder-impl.m
)
include(DefaultTargetOptions)
target_include_directories(${TARGET} PUBLIC
.
)
target_link_libraries(${TARGET} PRIVATE ${FOUNDATION_FRAMEWORK} ${COREML_FRAMEWORK})
set_target_properties(${TARGET} PROPERTIES
COMPILE_FLAGS "-fobjc-arc"
)
endif()
if (WHISPER_OPENVINO)
set(TARGET whisper.openvino)
add_library(${TARGET} OBJECT
openvino/whisper-openvino-encoder.h
openvino/whisper-openvino-encoder.cpp
)
target_include_directories(${TARGET} PUBLIC
.
)
set_property(TARGET ${TARGET} PROPERTY POSITION_INDEPENDENT_CODE ON)
set(WHISPER_EXTRA_FLAGS ${WHISPER_EXTRA_FLAGS} -DWHISPER_USE_OPENVINO)
target_link_libraries(${TARGET} PRIVATE openvino::runtime)
endif()
#
2022-10-08 08:00:59 +02:00
# whisper - this is the main library of the project
#
2022-10-07 22:53:12 +02:00
set(TARGET whisper)
2022-10-08 08:00:59 +02:00
add_library(${TARGET}
ggml.h
2022-10-07 23:21:16 +02:00
ggml.c
whisper : Metal and ggml-alloc support (#1270) * metal : init * whisper : factor out graph builds * whisper : allocate encoder and decoder using ggml-alloc * whisper : ggml-alloc is now supported * whisper : CoreML support ggml-alloc * build : fix ggml-alloc * ios : update submodule * extra : update sync-ggml.sh script to also sync ggml-alloc * ci : see if this is causing the crash * whisper : refactor ggml-alloc init * whisper.android : try to fix build * whisper : initial Metal version * ci : try to debug vmem issue * metal : decoder works on GPU! * metal : add multi-decoder support * ggml : fix ggml_nbytes (probably temp solution) * metal : run "cross" step on the GPU * whisper : remove ggml_repeat in the encoder * whisper : offload the Encoder to Metal * ggml : use simpler ggml_bytes() implementation * ggml-alloc : try to make CI happy by reducing vram to 128GB * whisper : add whisper_allocr to wrap ggml_allocr * whisper : factor out alloc init in a function * cmake : update to support Metal build * whisper : add <functional> header * objc : fix build (no Metal yet) * ios : add Metal support * swiftui : fix build * metal : speed-up KQ multiplication * metal : sync latest llama.cpp kernels * readme : add Metal info * ios : update submodule * coreml : add code to toggle Core ML config (CPU, ANE, GPU) * bench : fix timings by running a pre-heat * bench : start benching the decoder * whisper : add ggml_mul_mat_pad * bench : fix uninitialized vars * whisper : add comment for disabling mul-mat padding * whisper : add description of ggml_mul_mat_pad * whisper : clean-up ggml_mul_mat_pad * metal : remove the "concurrent" flag * bench : variable n_past * ios : update SPM package
2023-09-15 11:18:18 +02:00
ggml-alloc.h
ggml-alloc.c
ggml-backend.h
ggml-backend.c
ggml-quants.h
ggml-quants.c
whisper : Metal and ggml-alloc support (#1270) * metal : init * whisper : factor out graph builds * whisper : allocate encoder and decoder using ggml-alloc * whisper : ggml-alloc is now supported * whisper : CoreML support ggml-alloc * build : fix ggml-alloc * ios : update submodule * extra : update sync-ggml.sh script to also sync ggml-alloc * ci : see if this is causing the crash * whisper : refactor ggml-alloc init * whisper.android : try to fix build * whisper : initial Metal version * ci : try to debug vmem issue * metal : decoder works on GPU! * metal : add multi-decoder support * ggml : fix ggml_nbytes (probably temp solution) * metal : run "cross" step on the GPU * whisper : remove ggml_repeat in the encoder * whisper : offload the Encoder to Metal * ggml : use simpler ggml_bytes() implementation * ggml-alloc : try to make CI happy by reducing vram to 128GB * whisper : add whisper_allocr to wrap ggml_allocr * whisper : factor out alloc init in a function * cmake : update to support Metal build * whisper : add <functional> header * objc : fix build (no Metal yet) * ios : add Metal support * swiftui : fix build * metal : speed-up KQ multiplication * metal : sync latest llama.cpp kernels * readme : add Metal info * ios : update submodule * coreml : add code to toggle Core ML config (CPU, ANE, GPU) * bench : fix timings by running a pre-heat * bench : start benching the decoder * whisper : add ggml_mul_mat_pad * bench : fix uninitialized vars * whisper : add comment for disabling mul-mat padding * whisper : add description of ggml_mul_mat_pad * whisper : clean-up ggml_mul_mat_pad * metal : remove the "concurrent" flag * bench : variable n_past * ios : update SPM package
2023-09-15 11:18:18 +02:00
${GGML_SOURCES_METAL}
${GGML_SOURCES_CUDA}
${GGML_SOURCES_OPENCL}
whisper.h
2022-10-07 22:53:12 +02:00
whisper.cpp
)
include(DefaultTargetOptions)
2022-10-07 22:53:12 +02:00
target_include_directories(${TARGET} PUBLIC
.
)
if (WHISPER_COREML)
target_link_libraries(${TARGET} PRIVATE whisper.coreml)
endif()
if (WHISPER_OPENVINO)
target_link_libraries(${TARGET} PRIVATE whisper.openvino)
endif()
if (MSVC)
target_link_libraries(${TARGET} PRIVATE ${WHISPER_EXTRA_LIBS} ${CMAKE_THREAD_LIBS_INIT})
set(WHISPER_EXTRA_FLAGS ${WHISPER_EXTRA_FLAGS} -D_CRT_SECURE_NO_WARNINGS)
else()
target_link_libraries(${TARGET} PRIVATE m ${WHISPER_EXTRA_LIBS} ${CMAKE_THREAD_LIBS_INIT})
endif()
2022-10-07 22:53:12 +02:00
if (BUILD_SHARED_LIBS)
target_link_libraries(${TARGET} PUBLIC
${CMAKE_DL_LIBS}
)
target_compile_definitions(${TARGET} PUBLIC
WHISPER_SHARED
GGML_SHARED
)
target_compile_definitions(${TARGET} PRIVATE
WHISPER_BUILD
GGML_BUILD
2022-10-07 22:53:12 +02:00
)
whisper : Metal and ggml-alloc support (#1270) * metal : init * whisper : factor out graph builds * whisper : allocate encoder and decoder using ggml-alloc * whisper : ggml-alloc is now supported * whisper : CoreML support ggml-alloc * build : fix ggml-alloc * ios : update submodule * extra : update sync-ggml.sh script to also sync ggml-alloc * ci : see if this is causing the crash * whisper : refactor ggml-alloc init * whisper.android : try to fix build * whisper : initial Metal version * ci : try to debug vmem issue * metal : decoder works on GPU! * metal : add multi-decoder support * ggml : fix ggml_nbytes (probably temp solution) * metal : run "cross" step on the GPU * whisper : remove ggml_repeat in the encoder * whisper : offload the Encoder to Metal * ggml : use simpler ggml_bytes() implementation * ggml-alloc : try to make CI happy by reducing vram to 128GB * whisper : add whisper_allocr to wrap ggml_allocr * whisper : factor out alloc init in a function * cmake : update to support Metal build * whisper : add <functional> header * objc : fix build (no Metal yet) * ios : add Metal support * swiftui : fix build * metal : speed-up KQ multiplication * metal : sync latest llama.cpp kernels * readme : add Metal info * ios : update submodule * coreml : add code to toggle Core ML config (CPU, ANE, GPU) * bench : fix timings by running a pre-heat * bench : start benching the decoder * whisper : add ggml_mul_mat_pad * bench : fix uninitialized vars * whisper : add comment for disabling mul-mat padding * whisper : add description of ggml_mul_mat_pad * whisper : clean-up ggml_mul_mat_pad * metal : remove the "concurrent" flag * bench : variable n_past * ios : update SPM package
2023-09-15 11:18:18 +02:00
if (WHISPER_METAL)
# TODO: I think this should make ggml-metal.m "see" the ggml-metal.metal file from the "bin" directory
# but for some reason it does not work here like it does in llama.cpp
set_target_properties(${TARGET} PROPERTIES RESOURCE "${CMAKE_CURRENT_SOURCE_DIR}/ggml-metal.metal")
endif()
2022-10-07 22:53:12 +02:00
endif()
whisper : Metal and ggml-alloc support (#1270) * metal : init * whisper : factor out graph builds * whisper : allocate encoder and decoder using ggml-alloc * whisper : ggml-alloc is now supported * whisper : CoreML support ggml-alloc * build : fix ggml-alloc * ios : update submodule * extra : update sync-ggml.sh script to also sync ggml-alloc * ci : see if this is causing the crash * whisper : refactor ggml-alloc init * whisper.android : try to fix build * whisper : initial Metal version * ci : try to debug vmem issue * metal : decoder works on GPU! * metal : add multi-decoder support * ggml : fix ggml_nbytes (probably temp solution) * metal : run "cross" step on the GPU * whisper : remove ggml_repeat in the encoder * whisper : offload the Encoder to Metal * ggml : use simpler ggml_bytes() implementation * ggml-alloc : try to make CI happy by reducing vram to 128GB * whisper : add whisper_allocr to wrap ggml_allocr * whisper : factor out alloc init in a function * cmake : update to support Metal build * whisper : add <functional> header * objc : fix build (no Metal yet) * ios : add Metal support * swiftui : fix build * metal : speed-up KQ multiplication * metal : sync latest llama.cpp kernels * readme : add Metal info * ios : update submodule * coreml : add code to toggle Core ML config (CPU, ANE, GPU) * bench : fix timings by running a pre-heat * bench : start benching the decoder * whisper : add ggml_mul_mat_pad * bench : fix uninitialized vars * whisper : add comment for disabling mul-mat padding * whisper : add description of ggml_mul_mat_pad * whisper : clean-up ggml_mul_mat_pad * metal : remove the "concurrent" flag * bench : variable n_past * ios : update SPM package
2023-09-15 11:18:18 +02:00
if (GGML_SOURCES_CUDA)
message(STATUS "GGML CUDA sources found, configuring CUDA architecture")
# Only configure gmml CUDA architectures is not globally set
if (NOT DEFINED GGML_CUDA_ARCHITECTURES)
# Not overriden by user, so set defaults
set(GGML_CUDA_ARCHITECTURES 52 61 70)
endif()
message(STATUS "GGML Configuring CUDA architectures ${GGML_CUDA_ARCHITECTURES}")
set_property(TARGET whisper PROPERTY CUDA_ARCHITECTURES ${GGML_CUDA_ARCHITECTURES})
set_property(TARGET whisper PROPERTY CUDA_SELECT_NVCC_ARCH_FLAGS "Auto")
endif()
if (EMSCRIPTEN)
set_target_properties(${TARGET} PROPERTIES COMPILE_FLAGS "-msimd128")
endif()
2022-10-07 22:53:12 +02:00
target_compile_definitions(${TARGET} PUBLIC
${WHISPER_EXTRA_FLAGS}
)
set_target_properties(${TARGET} PROPERTIES PUBLIC_HEADER "ggml.h;whisper.h")
whisper : Metal and ggml-alloc support (#1270) * metal : init * whisper : factor out graph builds * whisper : allocate encoder and decoder using ggml-alloc * whisper : ggml-alloc is now supported * whisper : CoreML support ggml-alloc * build : fix ggml-alloc * ios : update submodule * extra : update sync-ggml.sh script to also sync ggml-alloc * ci : see if this is causing the crash * whisper : refactor ggml-alloc init * whisper.android : try to fix build * whisper : initial Metal version * ci : try to debug vmem issue * metal : decoder works on GPU! * metal : add multi-decoder support * ggml : fix ggml_nbytes (probably temp solution) * metal : run "cross" step on the GPU * whisper : remove ggml_repeat in the encoder * whisper : offload the Encoder to Metal * ggml : use simpler ggml_bytes() implementation * ggml-alloc : try to make CI happy by reducing vram to 128GB * whisper : add whisper_allocr to wrap ggml_allocr * whisper : factor out alloc init in a function * cmake : update to support Metal build * whisper : add <functional> header * objc : fix build (no Metal yet) * ios : add Metal support * swiftui : fix build * metal : speed-up KQ multiplication * metal : sync latest llama.cpp kernels * readme : add Metal info * ios : update submodule * coreml : add code to toggle Core ML config (CPU, ANE, GPU) * bench : fix timings by running a pre-heat * bench : start benching the decoder * whisper : add ggml_mul_mat_pad * bench : fix uninitialized vars * whisper : add comment for disabling mul-mat padding * whisper : add description of ggml_mul_mat_pad * whisper : clean-up ggml_mul_mat_pad * metal : remove the "concurrent" flag * bench : variable n_past * ios : update SPM package
2023-09-15 11:18:18 +02:00
include(GNUInstallDirs)
2022-10-07 22:53:12 +02:00
install(TARGETS ${TARGET}
whisper : Metal and ggml-alloc support (#1270) * metal : init * whisper : factor out graph builds * whisper : allocate encoder and decoder using ggml-alloc * whisper : ggml-alloc is now supported * whisper : CoreML support ggml-alloc * build : fix ggml-alloc * ios : update submodule * extra : update sync-ggml.sh script to also sync ggml-alloc * ci : see if this is causing the crash * whisper : refactor ggml-alloc init * whisper.android : try to fix build * whisper : initial Metal version * ci : try to debug vmem issue * metal : decoder works on GPU! * metal : add multi-decoder support * ggml : fix ggml_nbytes (probably temp solution) * metal : run "cross" step on the GPU * whisper : remove ggml_repeat in the encoder * whisper : offload the Encoder to Metal * ggml : use simpler ggml_bytes() implementation * ggml-alloc : try to make CI happy by reducing vram to 128GB * whisper : add whisper_allocr to wrap ggml_allocr * whisper : factor out alloc init in a function * cmake : update to support Metal build * whisper : add <functional> header * objc : fix build (no Metal yet) * ios : add Metal support * swiftui : fix build * metal : speed-up KQ multiplication * metal : sync latest llama.cpp kernels * readme : add Metal info * ios : update submodule * coreml : add code to toggle Core ML config (CPU, ANE, GPU) * bench : fix timings by running a pre-heat * bench : start benching the decoder * whisper : add ggml_mul_mat_pad * bench : fix uninitialized vars * whisper : add comment for disabling mul-mat padding * whisper : add description of ggml_mul_mat_pad * whisper : clean-up ggml_mul_mat_pad * metal : remove the "concurrent" flag * bench : variable n_past * ios : update SPM package
2023-09-15 11:18:18 +02:00
LIBRARY DESTINATION lib
ARCHIVE DESTINATION lib/static
RUNTIME DESTINATION bin
RESOURCE DESTINATION bin
PUBLIC_HEADER DESTINATION include
2022-10-07 22:53:12 +02:00
)
#
# bindings
#
add_subdirectory(bindings)
#
2022-10-08 08:00:59 +02:00
# programs, examples and tests
#
2022-10-08 08:00:59 +02:00
if (WHISPER_BUILD_TESTS AND NOT CMAKE_JS_VERSION)
enable_testing()
add_subdirectory(tests)
2022-10-07 22:53:12 +02:00
endif ()
if (WHISPER_BUILD_EXAMPLES)
add_subdirectory(examples)
endif()