Fixed static initialization fiascoyouki/unordered_map

8 files changed, 31 insertions, 24 deletions

diff --git a/src/CodeCache.h b/src/CodeCache.h
index 3c8e7a6..b8ee15c 100644
--- a/src/CodeCache.h
+++ b/src/CodeCache.h
@@ -30,6 +30,7 @@ class CodeCache {
     if (it != values_.end()) {
       return it->second;
     } else {
+      //std::cerr << "create" << std::endl;
       auto fn = generatorFunction();
       values_[key] = fn;
       return fn;
diff --git a/src/GenerateKernel.h b/src/GenerateKernel.h
index cd136e1..3832f25 100644
--- a/src/GenerateKernel.h
+++ b/src/GenerateKernel.h
@@ -159,12 +159,18 @@ class CodeGenBase {
   int vectorWidth_; ///< Vector width in bits.
   int VLEN_; ///< Vector width in elements.
 
-  asmjit::JitRuntime &runtime() {
-    return codeCache_.second;
+  static std::pair<CodeCache<std::tuple<bool, int, int, int, int, int, int, int>,
+    jit_micro_kernel_fp>, asmjit::JitRuntime> &runtime() {
+    static thread_local std::pair<CodeCache<std::tuple<bool, int, int, int, int, int, int, int>,
+      jit_micro_kernel_fp>, asmjit::JitRuntime> codeCache_;
+    return codeCache_;
   }
 
-  std::pair<CodeCache<std::tuple<bool, int, int, int, int, int, int, int>,
-    jit_micro_kernel_fp>, asmjit::JitRuntime> codeCache_;
 };
 
+//template <typename TA, typename TB, typename TC, typename accT>
+//thread_local std::pair<CodeCache<std::tuple<bool, int, int, int, int, int, int, int>,
+//  typename CodeGenBase<TA, TB, TC, accT>::jit_micro_kernel_fp>, asmjit::JitRuntime> 
+//  CodeGenBase<TA, TB, TC, accT>::codeCache_;
+
 } // namespace fbgemm
diff --git a/src/GenerateKernelU8S8S32ACC16.cc b/src/GenerateKernelU8S8S32ACC16.cc
index a2fdda3..cc77eab 100644
--- a/src/GenerateKernelU8S8S32ACC16.cc
+++ b/src/GenerateKernelU8S8S32ACC16.cc
@@ -168,9 +168,9 @@ CodeGenBase<uint8_t, int8_t, int32_t, int16_t>::getOrCreate<inst_set_t::avx2>(
       nRegBlockSize,
       nRegBlockSizeMin);
 
-  return codeCache_.first.getOrCreate(kernelSig, [&]() -> jit_micro_kernel_fp {
+  return runtime().first.getOrCreate(kernelSig, [&]() -> jit_micro_kernel_fp {
     asmjit::CodeHolder code;
-    code.init(runtime().codeInfo());
+    code.init(runtime().second.codeInfo());
     x86::Assembler assembler(&code);
     x86::Emitter *a = assembler.as<x86::Emitter>();
 
@@ -343,7 +343,7 @@ CodeGenBase<uint8_t, int8_t, int32_t, int16_t>::getOrCreate<inst_set_t::avx2>(
     asmjit::Error err;
     {
       //std::unique_lock<std::mutex> lock(rtMutex_);
-      err = runtime().add(&fn, &code);
+      err = runtime().second.add(&fn, &code);
     }
     if (err) {
       std::cout << "Error: in fn add" << std::endl;
diff --git a/src/GenerateKernelU8S8S32ACC16Avx512.cc b/src/GenerateKernelU8S8S32ACC16Avx512.cc
index c8d96b8..06169e9 100644
--- a/src/GenerateKernelU8S8S32ACC16Avx512.cc
+++ b/src/GenerateKernelU8S8S32ACC16Avx512.cc
@@ -175,9 +175,9 @@ CodeGenBase<uint8_t, int8_t, int32_t, int16_t>::getOrCreate<inst_set_t::avx512>(
       nRegBlockSize,
       nRegBlockSizeMin);
 
-  return codeCache_.first.getOrCreate(kernelSig, [&]() -> jit_micro_kernel_fp {
+  return runtime().first.getOrCreate(kernelSig, [&]() -> jit_micro_kernel_fp {
     asmjit::CodeHolder code;
-    code.init(runtime().codeInfo());
+    code.init(runtime().second.codeInfo());
     x86::Assembler assembler(&code);
     x86::Emitter *a = assembler.as<x86::Emitter>();
 
@@ -411,7 +411,7 @@ CodeGenBase<uint8_t, int8_t, int32_t, int16_t>::getOrCreate<inst_set_t::avx512>(
     asmjit::Error err;
     {
       //std::unique_lock<std::mutex> lock(rtMutex_);
-      err = runtime().add(&fn, &code);
+      err = runtime().second.add(&fn, &code);
     }
     if (err) {
       std::cout << "Error: in fn add" << std::endl;
diff --git a/src/GenerateKernelU8S8S32ACC32.cc b/src/GenerateKernelU8S8S32ACC32.cc
index e217c1c..984bcd4 100644
--- a/src/GenerateKernelU8S8S32ACC32.cc
+++ b/src/GenerateKernelU8S8S32ACC32.cc
@@ -173,9 +173,9 @@ CodeGenBase<uint8_t, int8_t, int32_t, int32_t>::getOrCreate<inst_set_t::avx2>(
       nRegBlockSize,
       nRegBlockSizeMin);
 
-  return codeCache_.first.getOrCreate(kernelSig, [&]() -> jit_micro_kernel_fp {
+  return runtime().first.getOrCreate(kernelSig, [&]() -> jit_micro_kernel_fp {
     asmjit::CodeHolder code;
-    code.init(runtime().codeInfo());
+    code.init(runtime().second.codeInfo());
     x86::Assembler assembler(&code);
     x86::Emitter *a = assembler.as<x86::Emitter>();
 #if defined(FBGEMM_LOG_CODE)
@@ -331,10 +331,10 @@ CodeGenBase<uint8_t, int8_t, int32_t, int32_t>::getOrCreate<inst_set_t::avx2>(
     asmjit::Error err;
     {
       //std::unique_lock<std::mutex> lock(rtMutex_);
-      err = runtime().add(&fn, &code);
+      err = runtime().second.add(&fn, &code);
     }
     if (err) {
-      std::cout << "Error: in fn add" << std::endl;
+      //std::cout << "Error: in fn add" << std::endl;
       return nullptr;
     }
 
diff --git a/src/GenerateKernelU8S8S32ACC32Avx512.cc b/src/GenerateKernelU8S8S32ACC32Avx512.cc
index fb2f8d1..6ad03ec 100644
--- a/src/GenerateKernelU8S8S32ACC32Avx512.cc
+++ b/src/GenerateKernelU8S8S32ACC32Avx512.cc
@@ -174,9 +174,9 @@ CodeGenBase<uint8_t, int8_t, int32_t, int32_t>::getOrCreate<inst_set_t::avx512>(
       nRegBlockSize,
       nRegBlockSizeMin);
 
-  return codeCache_.first.getOrCreate(kernelSig, [&]() -> jit_micro_kernel_fp {
+  return runtime().first.getOrCreate(kernelSig, [&]() -> jit_micro_kernel_fp {
     asmjit::CodeHolder code;
-    code.init(runtime().codeInfo());
+    code.init(runtime().second.codeInfo());
     x86::Assembler assembler(&code);
     x86::Emitter *a = assembler.as<x86::Emitter>();
 #if defined(FBGEMM_LOG_CODE)
@@ -423,7 +423,7 @@ CodeGenBase<uint8_t, int8_t, int32_t, int32_t>::getOrCreate<inst_set_t::avx512>(
     asmjit::Error err;
     {
       //std::unique_lock<std::mutex> lock(rtMutex_);
-      err = runtime().add(&fn, &code);
+      err = runtime().second.add(&fn, &code);
     }
     if (err) {
       std::cout << "Error: in fn add" << std::endl;
diff --git a/src/GenerateKernelU8S8S32ACC32Avx512VNNI.cc b/src/GenerateKernelU8S8S32ACC32Avx512VNNI.cc
index fe649ec..2722fc4 100644
--- a/src/GenerateKernelU8S8S32ACC32Avx512VNNI.cc
+++ b/src/GenerateKernelU8S8S32ACC32Avx512VNNI.cc
@@ -167,9 +167,9 @@ CodeGenBase<uint8_t, int8_t, int32_t, int32_t>::getOrCreate<
       nRegBlockSize,
       nRegBlockSizeMin);
 
-  return codeCache_.first.getOrCreate(kernelSig, [&]() -> jit_micro_kernel_fp {
+  return runtime().first.getOrCreate(kernelSig, [&]() -> jit_micro_kernel_fp {
     asmjit::CodeHolder code;
-    code.init(runtime().codeInfo());
+    code.init(runtime().second.codeInfo());
     x86::Assembler assembler(&code);
     x86::Emitter *a = assembler.as<x86::Emitter>();
 
@@ -416,7 +416,7 @@ CodeGenBase<uint8_t, int8_t, int32_t, int32_t>::getOrCreate<
     asmjit::Error err;
     {
       //std::unique_lock<std::mutex> lock(rtMutex_);
-      err = runtime().add(&fn, &code);
+      err = runtime().second.add(&fn, &code);
     }
     if (err) {
       std::cout << "Error: in fn add" << std::endl;
diff --git a/src/GroupwiseConv.h b/src/GroupwiseConv.h
index 8bb9beb..58ee24d 100644
--- a/src/GroupwiseConv.h
+++ b/src/GroupwiseConv.h
@@ -230,9 +230,9 @@ class GenConvKernel {
 
   static std::mutex rtMutex_; ///< Controll access to runtime;
 
-  static CodeCache<size_t /*std::tuple<bool, int, int, int>*/, jit_conv_kernel_fp>
+  static CodeCache<std::tuple<bool, int, int, int>, jit_conv_kernel_fp>
       codeCache_; ///< JIT Code Cache for reuse.
-  static CodeCache<size_t /*std::tuple<bool, int, int, int>*/, jit_rowoffset_kernel_fp>
+  static CodeCache<std::tuple<bool, int, int, int>, jit_rowoffset_kernel_fp>
       codeCacheRowOffset_; ///< JIT Code Cache for row offset kernel.
 
 private:
@@ -285,11 +285,11 @@ template <int SPATIAL_DIM, typename accT>
 std::mutex GenConvKernel<SPATIAL_DIM, accT>::rtMutex_;
 
 template <int SPATIAL_DIM, typename accT>
-CodeCache<size_t /*std::tuple<bool, int, int, int>*/, jit_conv_kernel_fp>
+CodeCache<std::tuple<bool, int, int, int>, jit_conv_kernel_fp>
     GenConvKernel<SPATIAL_DIM, accT>::codeCache_;
 
 template <int SPATIAL_DIM, typename accT>
-CodeCache<size_t /*std::tuple<bool, int, int, int>*/, jit_rowoffset_kernel_fp>
+CodeCache<std::tuple<bool, int, int, int>, jit_rowoffset_kernel_fp>
     GenConvKernel<SPATIAL_DIM, accT>::codeCacheRowOffset_;
 
 } // namespace fbgemm