whisper : disable CUDA mel + fix FFMPEG

examples/CMakeLists.txt

@@ -11,7 +11,7 @@ if (WHISPER_SDL2)
     string(STRIP "${SDL2_LIBRARIES}" SDL2_LIBRARIES)
 
     message(STATUS "SDL2_INCLUDE_DIRS = ${SDL2_INCLUDE_DIRS}")
-    message(STATUS "SDL2_LIBRARIES = ${SDL2_LIBRARIES}")
+    message(STATUS "SDL2_LIBRARIES    = ${SDL2_LIBRARIES}")
 endif()
 
 if (WHISPER_CLBLAST)
@@ -22,10 +22,35 @@ endif()
 
 set(TARGET common)
 
+unset(COMMON_EXTRA_LIBS)
+
 if (WHISPER_FFMPEG)
+    # As of cmake 3.27, there is no official cmake support for FindFFmpeg.
+    # Consequnelty we added a FindFFmpeg.cmake script the cmake subfolder:
+    # whisper.cpp does not need the full ffmpeg libs, just AVFORMAT AVCODEC AVUTIL SWRESAMPLE
+    # libswresample  performs highly optimized audio resampling, rematrixing and sample format conversion operations
+    # libavcodec provides a generic encoding/decoding framework and contains multiple decoders and encoders for audio, video and subtitle streams, and several bitstream filters.
+    # libavformat provides a generic framework for multiplexing and demultiplexing (muxing and demuxing) audio, video and subtitle streams.
+    find_package(FFmpeg REQUIRED)
+
+    if (NOT ${FFMPEG_FOUND})
+        message(FATAL_ERROR "Cannot find ffmpeg libs/headers")
+    endif()
+
+    message(STATUS "Found ffmpeg libs:       ${FFMPEG_LIBRARIES}")
+    message(STATUS "Found ffmpeg headers in: ${FFMPEG_INCLUDE_DIRS}")
+    message(STATUS "ffmpeg definitions:      ${FFMPEG_DEFINITIONS}")
+    message(STATUS "Found avformat ${AVFORMAT_VERSION}")
+
+    include_directories(${FFMPEG_INCLUDE_DIRS})
+    add_compile_definitions(WHISPER_FFMPEG)
+
+    list(APPEND COMMON_EXTRA_LIBS ${FFMPEG_LIBRARIES})
+
     set(COMMON_SOURCES_FFMPEG ffmpeg-transcode.cpp)
 endif()
 
+
 add_library(${TARGET} STATIC
     common.h
     common.cpp
@@ -38,7 +63,7 @@ add_library(${TARGET} STATIC
 
 include(DefaultTargetOptions)
 
-target_link_libraries(${TARGET} PRIVATE whisper)
+target_link_libraries(${TARGET} PRIVATE whisper ${COMMON_EXTRA_LIBS})
 
 set_target_properties(${TARGET} PROPERTIES POSITION_INDEPENDENT_CODE ON)
 set_target_properties(${TARGET} PROPERTIES FOLDER "libs")
@@ -55,8 +80,8 @@ if (WHISPER_SDL2)
 
     include(DefaultTargetOptions)
 
-    target_include_directories(${TARGET} PUBLIC ${SDL2_INCLUDE_DIRS})
-    target_link_libraries(${TARGET} PRIVATE ${SDL2_LIBRARIES})
+    target_include_directories(${TARGET} PUBLIC  ${SDL2_INCLUDE_DIRS})
+    target_link_libraries     (${TARGET} PRIVATE ${SDL2_LIBRARIES})
 
     set_target_properties(${TARGET} PROPERTIES POSITION_INDEPENDENT_CODE ON)
     set_target_properties(${TARGET} PROPERTIES FOLDER "libs")
@@ -108,7 +133,7 @@ if (WHISPER_SDL2)
     set_target_properties(talk-llama PROPERTIES FOLDER "examples")
     add_subdirectory(lsp)
     set_target_properties(lsp PROPERTIES FOLDER "examples")
-    if (LLAMA_SYCL)
+    if (GGML_SYCL)
         add_subdirectory(sycl)
         set_target_properties(sycl PROPERTIES FOLDER "examples")
     endif()

examples/common.cpp

@@ -30,7 +30,7 @@ extern bool ffmpeg_decode_audio(const std::string & ifname, std::vector<uint8_t>
 #endif
 
 // Function to check if the next argument exists
-std::string get_next_arg(int& i, int argc, char** argv, const std::string& flag, gpt_params& params) {
+static std::string get_next_arg(int& i, int argc, char** argv, const std::string& flag, gpt_params& params) {
     if (i + 1 < argc && argv[i + 1][0] != '-') {
         return argv[++i];
     } else {
@@ -346,7 +346,7 @@ std::vector<gpt_vocab::id> gpt_tokenize(const gpt_vocab & vocab, const std::stri
     return tokens;
 }
 
-std::vector<gpt_vocab::id> parse_tokens_from_string(const std::string& input, char delimiter) {
+static std::vector<gpt_vocab::id> parse_tokens_from_string(const std::string& input, char delimiter) {
     std::vector<gpt_vocab::id> output;
     std::stringstream ss(input);
     std::string token;
@@ -358,7 +358,7 @@ std::vector<gpt_vocab::id> parse_tokens_from_string(const std::string& input, ch
     return output;
 }
 
-std::map<std::string, std::vector<gpt_vocab::id>> extract_tests_from_file(const std::string & fpath_test){
+static std::map<std::string, std::vector<gpt_vocab::id>> extract_tests_from_file(const std::string & fpath_test){
     if (fpath_test.empty()){
         fprintf(stderr, "%s : No test file found.\n", __func__);
         return std::map<std::string, std::vector<gpt_vocab::id>>();

scripts/build-info.sh

@@ -24,7 +24,7 @@ if out=$($CC -dumpmachine); then
     build_target=$out
 fi
 
-echo "int LLAMA_BUILD_NUMBER = ${build_number};"
-echo "char const *LLAMA_COMMIT = \"${build_commit}\";"
-echo "char const *LLAMA_COMPILER = \"${build_compiler}\";"
-echo "char const *LLAMA_BUILD_TARGET = \"${build_target}\";"
+echo "int WHISPER_BUILD_NUMBER = ${build_number};"
+echo "char const *WHISPER_COMMIT = \"${build_commit}\";"
+echo "char const *WHISPER_COMPILER = \"${build_compiler}\";"
+echo "char const *WHISPER_BUILD_TARGET = \"${build_target}\";"

src/CMakeLists.txt

@@ -77,27 +77,27 @@ if (WHISPER_OPENVINO)
     set_target_properties(${TARGET} PROPERTIES FOLDER "libs")
 endif()
 
-if (GGML_CUDA)
-    cmake_minimum_required(VERSION 3.18)  # for CMAKE_CUDA_ARCHITECTURES
-
-    find_package(CUDAToolkit)
-    if (CUDAToolkit_FOUND)
-        message(STATUS "CUDA found")
-
-        if (NOT DEFINED CMAKE_CUDA_ARCHITECTURES)
-            # 52 == lowest CUDA 12 standard
-            # 60 == f16 CUDA intrinsics
-            # 61 == integer CUDA intrinsics
-            # 70 == compute capability at which unrolling a loop in mul_mat_q kernels is faster
-            set(CMAKE_CUDA_ARCHITECTURES "52;61;70") # lowest CUDA 12 standard + lowest for integer intrinsics
-        endif()
-        message(STATUS "Using CUDA architectures: ${CMAKE_CUDA_ARCHITECTURES}")
-
-        enable_language(CUDA)
-    else()
-        message(WARNING "CUDA not found")
-    endif()
-endif()
+#if (GGML_CUDA)
+#    cmake_minimum_required(VERSION 3.18)  # for CMAKE_CUDA_ARCHITECTURES
+#
+#    find_package(CUDAToolkit)
+#    if (CUDAToolkit_FOUND)
+#        message(STATUS "CUDA found")
+#
+#        if (NOT DEFINED CMAKE_CUDA_ARCHITECTURES)
+#            # 52 == lowest CUDA 12 standard
+#            # 60 == f16 CUDA intrinsics
+#            # 61 == integer CUDA intrinsics
+#            # 70 == compute capability at which unrolling a loop in mul_mat_q kernels is faster
+#            set(CMAKE_CUDA_ARCHITECTURES "52;61;70") # lowest CUDA 12 standard + lowest for integer intrinsics
+#        endif()
+#        message(STATUS "Using CUDA architectures: ${CMAKE_CUDA_ARCHITECTURES}")
+#
+#        enable_language(CUDA)
+#    else()
+#        message(WARNING "CUDA not found")
+#    endif()
+#endif()
 
 # whisper
 
@@ -107,11 +107,12 @@ add_library(whisper
             whisper-mel.hpp
             )
 
-if (GGML_CUDA)
-    target_sources(whisper PRIVATE whisper-mel-cuda.cu)
-
-    target_link_libraries(whisper PRIVATE CUDA::cufft)
-endif()
+# TODO: disabled because it relies on ggml internals that are no longer accessible (ggml-backend-impl.h, ggml-cuda/common.cuh, ..)
+#if (GGML_CUDA)
+#    target_sources(whisper PRIVATE whisper-mel-cuda.cu)
+#
+#    target_link_libraries(whisper PRIVATE CUDA::cufft)
+#endif()
 
 # Set the version numbers
 set_target_properties(whisper PROPERTIES

src/whisper-mel-cuda.cu

@@ -2,8 +2,7 @@
 #include "whisper-mel-cuda.hpp"
 #include "whisper.h"
 
-#include <ggml-cuda/common.cuh>
-#include <ggml-backend-impl.h>
+#include <ggml-backend.h>
 
 #include <cuda.h>
 #include <cuda_runtime.h>

src/whisper.cpp

@@ -3215,7 +3215,9 @@ struct mel_calc_cpu : public whisper_mel_calc {
 }
 
 static whisper_mel_calc * whisper_mel_calc_create(ggml_backend_t backend, const whisper_filters & filters) {
-#if defined(GGML_USE_CUDA) && !defined(GGML_USE_HIPBLAS)
+// TODO: disabled because it relies on ggml internals that are no longer accessible (ggml-backend-impl.h, ggml-cuda/common.cuh, ..)
+//#if defined(GGML_USE_CUDA) && !defined(GGML_USE_HIPBLAS)
+#if 0
     if (ggml_backend_is_cuda(backend)) {
         auto ret = whisper_mel_calc_create_cuda(backend, filters);
         if (ret) {