Added function export API

CMakeLists.txt

@@ -1,10 +1,6 @@
 cmake_minimum_required(VERSION 3.14)
 project(s2cpp LANGUAGES C CXX)
 
-if (WIN32)
-    add_definitions(-DWIN32_LEAN_AND_MEAN)
-endif()
-
 set(CMAKE_CXX_STANDARD 17)
 set(CMAKE_CXX_STANDARD_REQUIRED ON)
 set(CMAKE_POSITION_INDEPENDENT_CODE ON)
@@ -15,6 +11,7 @@ set(CMAKE_POSITION_INDEPENDENT_CODE ON)
 option(S2_VULKAN  "Build with Vulkan backend"  OFF)
 option(S2_CUDA    "Build with CUDA backend"    OFF)
 option(S2_METAL   "Build with Metal backend"   OFF)
+option(S2_BUILD_SHARED_LIBRARIES   "Compile S2 as a dynamic link library(dll) and a shared(.lib) library as well."   OFF)
 
 # ---------------------------------------------------------------------------
 # GGML (git submodule — not modified)
@@ -38,6 +35,7 @@ add_subdirectory(ggml)
 # s2 executable
 # ---------------------------------------------------------------------------
 set(S2_SOURCES
+    src/s2_export_api.cpp
     src/s2_audio.cpp
     src/s2_tokenizer.cpp
     src/s2_sampler.cpp
@@ -46,36 +44,98 @@ set(S2_SOURCES
     src/s2_prompt.cpp
     src/s2_generate.cpp
     src/s2_pipeline.cpp
-    src/s2_server.cpp
     src/main.cpp
 )
 
+if(S2_BUILD_SHARED_LIBRARIES)
 add_executable(s2 ${S2_SOURCES})
+add_library(s2_dll SHARED ${S2_SOURCES})
+add_library(s2_lib STATIC ${S2_SOURCES})
 
 target_include_directories(s2 PRIVATE
     ${CMAKE_CURRENT_SOURCE_DIR}/include
     ${CMAKE_CURRENT_SOURCE_DIR}/third_party
     ${CMAKE_CURRENT_SOURCE_DIR}/ggml/include
-    ${CMAKE_CURRENT_SOURCE_DIR}/ggml/src
+    ${CMAKE_CURRENT_SOURCE_DIR}/ggml/src          # for ggml-common.h etc.
+)
+target_include_directories(s2_dll PRIVATE
+    ${CMAKE_CURRENT_SOURCE_DIR}/include
+    ${CMAKE_CURRENT_SOURCE_DIR}/third_party
+    ${CMAKE_CURRENT_SOURCE_DIR}/ggml/include
+    ${CMAKE_CURRENT_SOURCE_DIR}/ggml/src          # for ggml-common.h etc.
+)
+target_include_directories(s2_lib PRIVATE
+    ${CMAKE_CURRENT_SOURCE_DIR}/include
+    ${CMAKE_CURRENT_SOURCE_DIR}/third_party
+    ${CMAKE_CURRENT_SOURCE_DIR}/ggml/include
+    ${CMAKE_CURRENT_SOURCE_DIR}/ggml/src          # for ggml-common.h etc.
 )
 
 target_link_libraries(s2 PRIVATE ggml)
+target_link_libraries(s2_dll PRIVATE ggml)
+target_link_libraries(s2_lib PRIVATE ggml)
+
 
+target_compile_definitions(s2_dll PRIVATE S2_LIBRARY)
+target_compile_definitions(s2_lib PRIVATE S2_LIBRARY)
 if(S2_VULKAN)
     target_compile_definitions(s2 PRIVATE GGML_USE_VULKAN)
+	target_compile_definitions(s2_dll PRIVATE GGML_USE_VULKAN)
+	target_compile_definitions(s2_lib PRIVATE GGML_USE_VULKAN)
 endif()
 if(S2_CUDA)
     target_compile_definitions(s2 PRIVATE GGML_USE_CUDA)
+	target_compile_definitions(s2_dll PRIVATE GGML_USE_CUDA)
+	target_compile_definitions(s2_lib PRIVATE GGML_USE_CUDA)
 endif()
 
+
 # Platform-specific
 if(UNIX AND NOT APPLE)
     target_link_libraries(s2 PRIVATE pthread m)
+	target_link_libraries(s2_dll PRIVATE pthread m)
+	target_link_libraries(s2_lib PRIVATE pthread m)
+endif()
+
+# Install
+install(TARGETS s2 RUNTIME DESTINATION bin)
+install(TARGETS s2_dll RUNTIME DESTINATION bin)
+install(TARGETS s2_lib RUNTIME DESTINATION bin)
+
+else()
+add_executable(s2 ${S2_SOURCES})
+
+target_include_directories(s2 PRIVATE
+    ${CMAKE_CURRENT_SOURCE_DIR}/include
+    ${CMAKE_CURRENT_SOURCE_DIR}/third_party
+    ${CMAKE_CURRENT_SOURCE_DIR}/ggml/include
+    ${CMAKE_CURRENT_SOURCE_DIR}/ggml/src          # for ggml-common.h etc.
+)
+
+target_link_libraries(s2 PRIVATE ggml)
+
+if(S2_VULKAN)
+    target_compile_definitions(s2 PRIVATE GGML_USE_VULKAN)
+endif()
+if(S2_CUDA)
+    target_compile_definitions(s2 PRIVATE GGML_USE_CUDA)
 endif()
 
-if(WIN32)
-    target_link_libraries(s2 PRIVATE ws2_32 wsock32)
+# Platform-specific
+if(UNIX AND NOT APPLE)
+    target_link_libraries(s2 PRIVATE pthread m)
 endif()
 
 # Install
 install(TARGETS s2 RUNTIME DESTINATION bin)
+
+endif()
+
+
+
+
+
+
+
+
+

include/s2_codec.h

@@ -32,7 +32,7 @@ class AudioCodec {
     // Decode VQ codes to mono float32 audio.
     // codes: (num_codebooks, n_frames) flattened row-major.
     bool decode(const int32_t * codes, int32_t n_frames, int32_t n_threads,
-                std::vector<float> & audio_out);
+                std::vector<float> & audio_out, int32_t* audio_n_frames_out);
 
     int32_t sample_rate()     const { return sample_rate_; }
     int32_t hop_length()      const { return hop_length_; }

include/s2_config.h

@@ -0,0 +1,20 @@
+#pragma once
+
+#if defined(_WIN32) || defined(__CYGWIN__)
+
+#ifdef S2_LIBRARY
+#define S2_Export __declspec(dllexport)
+static bool SuppressNonEssentialVerbosity = true;
+#else
+#define S2_Export 
+static bool SuppressNonEssentialVerbosity = false;
+#endif
+#else
+
+// On Unix-like systems, check for GCC 4+ visibility support
+#if __GNUC__ >= 4
+#define S2_Export __attribute__((visibility("default")))
+#else
+#define S2_Export
+#endif
+#endif

include/s2_export_api.h

@@ -0,0 +1,44 @@
+#pragma once
+
+#include "s2_audio.h"
+#include "s2_codec.h"
+#include "s2_generate.h"
+#include "s2_model.h"
+#include "s2_tokenizer.h"
+#include "s2_pipeline.h"
+#include "s2_prompt.h"
+#include "s2_config.h"
+
+extern "C" 
+{
+	S2_Export s2::Pipeline* AllocS2Pipeline();
+	S2_Export void ReleaseS2Pipeline(s2::Pipeline* Pipeline);
+	S2_Export void SyncS2TokenizerConfigFromS2Model(s2::SlowARModel* Model, s2::Tokenizer* Tokenizer);
+	S2_Export int InitializeS2Pipeline(s2::Pipeline* Pipeline, s2::Tokenizer* Tokenizer, s2::SlowARModel* Model, s2::AudioCodec* AudioCodec);
+
+	S2_Export s2::GenerateParams* AllocS2GenerateParams();
+	S2_Export void ReleaseS2GenerateParams(s2::GenerateParams* GenerateParams);
+	S2_Export int InitializeS2GenerateParams(s2::GenerateParams* GenerateParams, int32_t max_new_tokens = -1, float temperature = -1, float top_p = -1, int32_t top_k = -1, int32_t min_tokens_before_end = -1, int32_t n_threads = -1, int verbose = -1);
+
+	S2_Export s2::SlowARModel* AllocS2Model();
+	S2_Export void ReleaseS2Model(s2::SlowARModel* Model);
+	S2_Export int InitializeS2Model(s2::SlowARModel* Model, const char* gguf_path, int32_t gpu_device, int32_t backend_type);
+
+	S2_Export s2::Tokenizer* AllocS2Tokenizer();
+	S2_Export void ReleaseS2Tokenizer(s2::Tokenizer* Tokenizer);
+	S2_Export int InitializeS2Tokenizer(s2::Tokenizer* Tokenizer, const char* path);
+
+	S2_Export s2::AudioCodec* AllocS2AudioCodec();
+	S2_Export void ReleaseS2AudioCodec(s2::AudioCodec* AudioCodec);
+	S2_Export int InitializeS2AudioCodec(s2::AudioCodec* AudioCodec, const char* gguf_path, int32_t gpu_device, int32_t backend_type);
+
+	S2_Export std::vector<int32_t>* AllocS2AudioPromptCodes();
+	S2_Export void ReleaseS2AudioPromptCodes(std::vector<int32_t>* AudioPromptCodes);
+	S2_Export int InitializeAudioPromptCodes(s2::Pipeline* Pipeline, int32_t ThreadCount, const char* ReferenceAudioPath, std::vector<int32_t>* AudioPromptCodes, int* TPrompt);
+
+	S2_Export std::vector<float>* AllocS2AudioBuffer(int InitialSize);
+	S2_Export void ReleaseS2AudioBuffer(std::vector<float>* AudioBuffer);
+	S2_Export float* GetS2AudioBufferDataPointer(std::vector<float>* AudioBuffer);
+
+	S2_Export int S2Synthesize(s2::Pipeline* Pipeline, const s2::GenerateParams* GenerateParams, std::vector<float>* AudioBuffer, std::vector<int32_t>* ReferenceAudioPromptCodes, int32_t* ReferenceAudioTPrompt, const char* ReferenceAudioPath, const char* ReferenceAudioTranscript, const char* TextToInfer, const char* OutputAudioPath, int32_t* AudioBufferOutputLength);
+}

include/s2_pipeline.h

@@ -36,9 +36,9 @@ class Pipeline {
     bool synthesize(const PipelineParams & params);
 
     bool synthesize_to_memory(const PipelineParams & params, void** ref_audio_buffer, size_t* ref_audio_size, void** wav_buffer, size_t* wav_size);
-    bool synthesize_raw(const PipelineParams & params, AudioData & ref_audio, std::vector<float> & audio_out);
+    bool synthesize_raw(const PipelineParams & params, AudioData & ref_audio, std::vector<float> & audio_out, int32_t* audio_out_length);
 
-private:
+//private:
     Tokenizer   tokenizer_;
     SlowARModel model_;
     AudioCodec  codec_;

src/s2_codec.cpp

@@ -1028,7 +1028,7 @@ bool AudioCodec::encode(const float * audio, int32_t n_samples, int32_t n_thread
 // ---------------------------------------------------------------------------
 
 bool AudioCodec::decode(const int32_t * codes, int32_t n_frames, int32_t n_threads,
-                         std::vector<float> & audio_out) {
+                         std::vector<float> & audio_out, int32_t* audio_n_frames_out) {
     if (n_frames <= 0) return false;
 
     // Step 1: dequantize VQ codes to stage vector (n_frames, quantizer_input_dim)
@@ -1132,10 +1132,12 @@ bool AudioCodec::decode(const int32_t * codes, int32_t n_frames, int32_t n_threa
 
         // audio_t is (1, T) or (C, T) — we expect (1, T), take total elements
         const int32_t n_samples = static_cast<int32_t>(ggml_nelements(audio_t));
-        audio_out.resize(n_samples);
+        if(audio_out.size() < n_samples) { audio_out.resize(n_samples); } //Allows for pre-allocated audio buffers that don't need to resized each time.
         ggml_backend_tensor_get(audio_t, audio_out.data(), 0, n_samples * sizeof(float));
         ggml_gallocr_free(allocr);
         ggml_free(ctx);
+
+        *audio_n_frames_out = n_samples;
     }
     return true;
 }