Use pthread key on Mac

modules/Nncase.Modules.CPU/CodeGen/CPU/CSourceBuiltn.cs

@@ -55,14 +55,14 @@ public static string TopoAwareRuntimeDef(CpuTargetOptions options, ulong dataAli
 
     public static string CMakeDef(string name)
     {
-        var cmakePath = CMakePath(Path.Combine(Path.GetDirectoryName(typeof(CSourceBuiltn).Assembly.Location)!, "Runtime", "cmake", "cpu_runtime.cmake"));
+        var cmakePath = CMakePath(Path.Combine(Path.GetDirectoryName(typeof(CSourceBuiltn).Assembly.Location)!, "Runtime", "cmake", "ntt_module.cmake"));
         var content = RazorTemplateEngine.RenderAsync("~/CodeGen/CPU/Templates/CMakeLists.txt.cshtml", new { CMakePath = cmakePath }).Result;
         return content;
     }
 
     public static string MakeMain(TIR.PrimFunction primFunction, ulong dataAlign, ulong dataUsage, ulong rdataPoolSize, IEnumerable<TIR.Buffer> rdataBuffers, CpuTargetOptions options)
     {
-        var content = RazorTemplateEngine.RenderAsync("~/CodeGen/CPU/Templates/main.cpp.cshtml", new KernelMainModel(primFunction, rdataBuffers.ToArray(), options, dataAlign, dataUsage, rdataPoolSize)).Result;
+        var content = RazorTemplateEngine.RenderAsync("~/CodeGen/CPU/Templates/thread_main.cpp.cshtml", new KernelMainModel(primFunction, rdataBuffers.ToArray(), options, dataAlign, dataUsage, rdataPoolSize)).Result;
         return content;
     }
 

modules/Nncase.Modules.CPU/CodeGen/CPU/CSourceCompiler.cs

@@ -169,9 +169,7 @@ private void ArchSpecific()
     private string ArgumentsSpecific(string sourcePath, string outPath)
     {
         var archConfig = RuntimeInformation.IsOSPlatform(OSPlatform.Windows) ?
-        "-DCMAKE_C_COMPILER=clang-cl -DCMAKE_CXX_COMPILER=clang-cl -DBUILD_SHARED=ON" :
-        RuntimeInformation.IsOSPlatform(OSPlatform.Linux) ? "-DBUILD_SHARED=ON" :
-        string.Empty;
+        "-DCMAKE_C_COMPILER=clang-cl -DCMAKE_CXX_COMPILER=clang-cl" : string.Empty;
 
 #if DEBUG
         var config = "Debug";

modules/Nncase.Modules.CPU/CodeGen/CPU/LinkableModule.cs

@@ -56,7 +56,7 @@ public ILinkedModule Link(ILinkContext linkContext)
                 Directory.CreateDirectory(dumpPath);
             }
 
-            using (var fs = File.Open(Path.Join(dumpPath, "main.cpp"), FileMode.Create))
+            using (var fs = File.Open(Path.Join(dumpPath, "thread_main.cpp"), FileMode.Create))
             {
                 using (var writer = new StreamWriter(fs))
                 {
@@ -112,7 +112,7 @@ public ILinkedModule Link(ILinkContext linkContext)
     private string CompileCSource(string sourcePath)
     {
         var compiler = new CSourceCompiler();
-        var binDir = Path.Join(sourcePath, "build", "nncase_cpu_module");
+        var binDir = Path.Join(sourcePath, "build", "nncase_ntt_module");
         return compiler.Compile(sourcePath, binDir);
     }
 }

modules/Nncase.Modules.CPU/CodeGen/CPU/Templates/CMakeLists.txt.cshtml

@@ -12,41 +12,4 @@ endif()
 
 include(@Html.Raw(Model.CMakePath))
 
-if(NOT APPLE AND BUILD_SHARED)
-    add_library(nncase_cpu_module SHARED main.cpp)
-    set_target_properties(nncase_cpu_module PROPERTIES PREFIX "" SUFFIX "")
-    set_target_properties(nncase_cpu_runtime PROPERTIES POSITION_INDEPENDENT_CODE ON)
-else()
-    add_executable(nncase_cpu_module main.cpp)
-endif()
-target_compile_features(nncase_cpu_module PUBLIC cxx_std_20)
-target_link_libraries(nncase_cpu_module PRIVATE nncase_cpu_runtime)
-target_compile_definitions(nncase_cpu_module PUBLIC -DNNCASE_CPU_MODULE=1)
-
-if (MSVC)
-    set_target_properties(nncase_cpu_module PROPERTIES LINK_FLAGS /SUBSYSTEM:CONSOLE)
-    target_link_options(nncase_cpu_module PRIVATE /NODEFAULTLIB)
-    target_link_libraries(nncase_cpu_module PRIVATE libvcruntime msvcrt ucrt "libcpmt$<$<CONFIG:Debug>:d>")
-    set_property(TARGET nncase_cpu_module PROPERTY
-        MSVC_RUNTIME_LIBRARY "MultiThreaded$<$<CONFIG:Debug>:Debug>")
-else()
-    if (APPLE)
-        if(BUILD_STANDALONE)
-            target_link_options(nncase_cpu_module PRIVATE -ld_classic -lc)
-        else()
-            target_link_options(nncase_cpu_module PRIVATE -static)
-            target_link_options(nncase_cpu_module PRIVATE -e _module_entry -bundle -ld_classic -lc)
-            target_compile_options(nncase_cpu_module PRIVATE "$<$<CONFIG:Debug>:-O1>")
-        endif(BUILD_STANDALONE)
-    else()
-        if (BUILD_SHARED)
-            target_link_options(nncase_cpu_module PRIVATE -e module_entry)
-        else()
-            if(NOT BUILD_STANDALONE)
-                target_link_options(nncase_cpu_module PRIVATE -static)
-                target_link_options(nncase_cpu_module PRIVATE -e module_entry -nostdlib)
-            endif(NOT BUILD_STANDALONE)
-        endif()
-        target_link_libraries(nncase_cpu_module PRIVATE gcc)
-    endif()
-endif()
+target_sources(nncase_ntt_module PRIVATE thread_main.cpp)

modules/Nncase.Modules.CPU/CodeGen/CPU/Templates/main.cpp.cshtml → modules/Nncase.Modules.CPU/CodeGen/CPU/Templates/thread_main.cpp.cshtml

@@ -14,7 +14,7 @@
 }
 #include "kernel.h"
 
-alignas(@(Model.Alignment)) static thread_local uint8_t local_data[@(Model.DataSize)];
+//alignas(@(Model.Alignment)) static thread_local uint8_t local_data[@(Model.DataSize)];
 
 extern "C" void thread_main(std::byte *const *inouts, const std::byte *rdata) {
   /* prepare inputs */
@@ -41,7 +41,9 @@ extern "C" void thread_main(std::byte *const *inouts, const std::byte *rdata) {
     throw new NotSupportedException($"not support multi form topology!");
   }
 
+  auto local_data = (uint8_t *)nncase::ntt::runtime::thread_alloc(@Model.DataSize, @Model.Alignment);
   @(Model.PrimFunction.Name)(@(string.Join(", ", names)), local_data);
+  nncase::ntt::runtime::thread_free(local_data);
 }
 
 #ifdef NNCASE_STANDALONE

ntt/cmake/ntt_module.cmake

@@ -0,0 +1,33 @@
+cmake_minimum_required(VERSION 3.15)
+
+include(${CMAKE_CURRENT_LIST_DIR}/compile_flags.cmake)
+
+if (BUILD_STANDALONE)
+    add_executable(nncase_ntt_module ${CMAKE_CURRENT_LIST_DIR}/../src/dummy.cpp)
+else()
+    add_library(nncase_ntt_module SHARED ${CMAKE_CURRENT_LIST_DIR}/../src/dummy.cpp)
+endif()
+
+target_compile_features(nncase_ntt_module PUBLIC cxx_std_20)
+target_include_directories(nncase_ntt_module PUBLIC ${CMAKE_CURRENT_LIST_DIR}/../include)
+set_target_properties(nncase_ntt_module PROPERTIES PREFIX "" SUFFIX "")
+set_target_properties(nncase_ntt_module PROPERTIES POSITION_INDEPENDENT_CODE ON)
+target_compile_definitions(nncase_ntt_module PUBLIC -DNNCASE_CPU_MODULE=1)
+
+target_sources(nncase_ntt_module PRIVATE ${CMAKE_CURRENT_LIST_DIR}/../src/cpu_runtime.cpp)
+
+if (BUILD_STANDALONE)
+    target_compile_definitions(nncase_ntt_module PUBLIC -DNNCASE_STANDALONE=1)
+endif()
+
+if (MSVC)
+    set_property(TARGET nncase_ntt_module PROPERTY
+        MSVC_RUNTIME_LIBRARY "MultiThreaded$<$<CONFIG:Debug>:Debug>")
+    set_target_properties(nncase_ntt_module PROPERTIES LINK_FLAGS /SUBSYSTEM:CONSOLE)
+    target_link_options(nncase_ntt_module PRIVATE /NODEFAULTLIB)
+    target_link_libraries(nncase_ntt_module PRIVATE libvcruntime msvcrt ucrt "libcpmt$<$<CONFIG:Debug>:d>")
+elseif(APPLE)
+    target_link_options(nncase_ntt_module PRIVATE -ld_classic -lc)
+else()
+    target_link_options(nncase_ntt_module PRIVATE -static)
+endif()

ntt/include/nncase/ntt/runtime.h

@@ -23,16 +23,8 @@
 #endif
 
 namespace nncase::ntt::runtime {
-enum class module_main_reason {
-    block_main,
-    thread_main,
-};
-}
+void *thread_alloc(size_t bytes, size_t alignment);
+void thread_free(void *ptr);
+} // namespace nncase::ntt::runtime
 
-extern "C" {
-extern void thread_main(std::byte *const *inouts, const std::byte *rdata);
-
-extern NTT_RUNTIME_API void
-module_entry(nncase::ntt::runtime::module_main_reason reason, void *params);
-using module_entry_t = decltype(module_entry) *;
-}
+extern "C" void thread_main(std::byte *const *inouts, const std::byte *rdata);

ntt/include/nncase/ntt/runtime/cpu_runtime.h

@@ -17,78 +17,61 @@
 #include "../runtime.h"
 #include <cstdarg>
 
-extern "C" {
-struct nncase_runtime_cpu_mt_t {
-    float (*acosf)(float v);
-    float (*acoshf)(float v);
-    float (*asinf)(float v);
-    float (*asinhf)(float v);
-    float (*copysignf)(float mag, float sgn);
-    float (*cosf)(float v);
-    float (*coshf)(float v);
-    float (*erff)(float v);
-    float (*expf)(float v);
-    float (*fmodf)(float x, float y);
-    float (*logf)(float v);
-    float (*nearbyintf)(float v);
-    float (*powf)(float x, float y);
-    float (*roundf)(float v);
-    float (*sinf)(float v);
-    float (*sinhf)(float v);
-    float (*sqrtf)(float v);
-    float (*tanhf)(float v);
-
-    uint8_t *(*sram_address)(int bid, int tid);
-
-    void (*failfast)(const char *format, va_list args);
-
-#ifndef WIN32
-    void *(*memcpy)(void *dst, const void *src, size_t len);
-    void *(*memmove)(void *dst, const void *src, size_t len);
-    void *(*memset)(void *b, int c, size_t len);
+#ifdef __APPLE__
+#include <pthread.h>
 #endif
-};
 
-struct nncase_runtime_cpu_block_params_t {
-    const nncase_runtime_cpu_mt_t *cpu_mt;
+namespace nncase::ntt::runtime {
+struct cpu_block_entry_params_t {
     size_t tdim;
     size_t bdim;
     size_t cdim;
+    size_t bid;
+    size_t cid;
+    size_t cpu_id_offset;
+    std::byte *const *inouts;
+    const std::byte *rdata;
+#ifdef __APPLE__
+    pthread_key_t cpu_thread_context_key;
+#endif
 };
 
-struct nncase_runtime_cpu_thread_params_t {
+struct cpu_thread_context_t {
     size_t tid;
     size_t bid;
     size_t cid;
-    std::byte *const *inouts;
-    const std::byte *rdata;
+
+    static cpu_thread_context_t &current() noexcept;
 };
-}
 
-namespace nncase::ntt::runtime {
-extern const nncase_runtime_cpu_mt_t *cpu_mt;
 extern size_t tdim;
 extern size_t bdim;
 extern size_t cdim;
-
-extern thread_local size_t tid;
-extern thread_local size_t bid;
-extern thread_local size_t cid;
 } // namespace nncase::ntt::runtime
 
 namespace nncase::ntt {
 template <> struct program_id_getter<0> {
-    static size_t id() noexcept { return runtime::tid; }
+    static size_t id() noexcept {
+        return runtime::cpu_thread_context_t::current().tid;
+    }
     static size_t dim() noexcept { return runtime::tdim; }
 };
 
 template <> struct program_id_getter<1> {
-    static size_t id() noexcept { return runtime::bid; }
+    static size_t id() noexcept {
+        return runtime::cpu_thread_context_t::current().bid;
+    }
     static size_t dim() noexcept { return runtime::bdim; }
 };
 
 template <> struct program_id_getter<2> {
-    static size_t id() noexcept { return runtime::cid; }
+    static size_t id() noexcept {
+        return runtime::cpu_thread_context_t::current().cid;
+    }
     static size_t dim() noexcept { return runtime::cdim; }
 };
 } // namespace nncase::ntt
+
+extern "C" void
+block_entry(const nncase::ntt::runtime::cpu_block_entry_params_t &params);
+using block_entry_t = decltype(block_entry) *;