[refactor] Remove the difficult-to-implement CompiledKernelData::size…

…() (taichi-dev#7540)

Issue: taichi-dev#7002 

### Brief Summary
It is challenging to implement the function
`CompiledKernelData::size()`, particularly for the
`llvm::CompiledKernelData` class, as it is not possible to obtain the
size of the `llvm::Module` without dumping it.

taichi/codegen/compiled_kernel_data.h

@@ -86,7 +86,6 @@ class CompiledKernelData {
   virtual ~CompiledKernelData() = default;
 
   virtual Arch arch() const = 0;
-  virtual std::size_t size() const = 0;
 
   Err load(std::istream &is);
   Err dump(std::ostream &os) const;

taichi/codegen/spirv/compiled_kernel_data.cpp

@@ -1,7 +1,5 @@
 #include "taichi/codegen/spirv/compiled_kernel_data.h"
 
-#include <numeric>
-
 namespace taichi::lang {
 
 static std::unique_ptr<CompiledKernelData> new_spirv_compiled_kernel_data() {
@@ -21,16 +19,6 @@ Arch CompiledKernelData::arch() const {
   return arch_;
 }
 
-std::size_t CompiledKernelData::size() const {
-  return sizeof(std::uint32_t) *
-         std::accumulate(
-             data_.src.spirv_src.begin(), data_.src.spirv_src.end(),
-             (std::size_t)0,
-             [](std::size_t val, const std::vector<std::uint32_t> &c) {
-               return val + c.size();
-             });
-}
-
 std::unique_ptr<lang::CompiledKernelData> CompiledKernelData::clone() const {
   return std::make_unique<CompiledKernelData>(arch_, data_);
 }

taichi/codegen/spirv/compiled_kernel_data.h

@@ -30,7 +30,6 @@ class CompiledKernelData : public lang::CompiledKernelData {
   CompiledKernelData(Arch arch, InternalData data);
 
   Arch arch() const override;
-  std::size_t size() const override;
   std::unique_ptr<lang::CompiledKernelData> clone() const override;
 
   const InternalData &get_internal_data() const {

taichi/compilation_manager/kernel_compilation_manager.cpp

@@ -115,18 +115,20 @@ void KernelCompilationManager::dump() {
   for (auto &[kernel_key, kernel] : caching_kernels_) {
     if (kernel.cache_mode == CacheData::MemAndDiskCache) {
       auto [iter, ok] = kernels.insert({kernel_key, std::move(kernel)});
-      data.size += ok ? iter->second.size : 0;
+      TI_ASSERT(!ok || iter->second.size == 0);
     }
   }
   // Dump cached CompiledKernelData to disk
-  for (const auto &[_, k] : kernels) {
+  for (auto &[_, k] : kernels) {
     if (k.compiled_kernel_data) {
       const auto arch = k.compiled_kernel_data->arch();
       auto cache_filename = make_filename(k.kernel_key, arch);
       if (try_lock_with_file(cache_filename)) {
         std::ofstream fs{cache_filename, std::ios::out | std::ios::binary};
         TI_ASSERT(fs.is_open());
         k.compiled_kernel_data->dump(fs);
+        k.size = fs.tellp();
+        data.size += k.size;
       }
     }
   }
@@ -235,7 +237,7 @@ const CompiledKernelData &KernelCompilationManager::compile_and_cache_kernel(
   k.kernel_key = kernel_key;
   k.created_at = k.last_used_at = std::time(nullptr);
   k.compiled_kernel_data = compile_kernel(compile_config, caps, kernel_def);
-  k.size = k.compiled_kernel_data->size();
+  k.size = 0;  // Populate `size` within the KernelCompilationManager::dump()
   k.cache_mode = cache_mode;
   const auto &kernel_data = (caching_kernels_[kernel_key] = std::move(k));
   return *kernel_data.compiled_kernel_data;

tests/cpp/codegen/compiled_kernel_data_test.cpp

@@ -24,10 +24,6 @@ class FakeCompiledKernelData : public CompiledKernelData {
     return kFakeArch;
   }
 
-  std::size_t size() const override {
-    return compiled_data_.so_bin.size();
-  }
-
   std::unique_ptr<CompiledKernelData> clone() const override {
     return std::make_unique<FakeCompiledKernelData>(*this);
   }