#7449: allow users to specify compile opt level

Add a field for users to set the compiler optimization level in
the config passed to CreateKernel

Default is still O3 for compute and Os for rest

tt_metal/api/tt-metalium/build.hpp

@@ -1,8 +1,9 @@
-// SPDX-FileCopyrightText: © 2023 Tenstorrent Inc.
+// SPDX-FileCopyrightText: © 2025 Tenstorrent Inc.
 //
 // SPDX-License-Identifier: Apache-2.0
 
 #pragma once
+#include <string_view>
 #include <thread>
 #include <string>
 #include <future>
@@ -101,14 +102,22 @@ class alignas(CACHE_LINE_ALIGNMENT) JitBuildState {
 
     string link_objs_;
 
+    // Default compiler optimization setting
+    // Used when JitBuildSettings is not provided
+    string default_compile_opt_level_;
+
+    // Default linker optimization setting
+    // Used when JitBuildSettings is not provided
+    string default_linker_opt_level_;
+
     void compile(const string& log_file, const string& out_path, const JitBuildSettings* settings) const;
     void compile_one(
         const string& log_file,
         const string& out_path,
         const JitBuildSettings* settings,
         const string& src,
         const string& obj) const;
-    void link(const string& log_file, const string& out_path) const;
+    void link(const string& log_file, const string& out_path, const JitBuildSettings* settings) const;
     void weaken(const string& log_file, const string& out_path) const;
     void copy_kernel(const string& kernel_in_path, const string& op_out_path) const;
     void extract_zone_src_locations(const string& log_file) const;
@@ -169,12 +178,19 @@ class JitBuildIdleEthernet : public JitBuildState {
 // (eg, API specified settings)
 class JitBuildSettings {
 public:
-    virtual const string& get_full_kernel_name() const = 0;
+    // Returns the full kernel name
+    virtual const std::string& get_full_kernel_name() const = 0;
+    // Returns the compiler optimization level
+    virtual std::string_view get_compiler_opt_level() const = 0;
+    // Returns the linker optimization level
+    virtual std::string_view get_linker_opt_level() const = 0;
+
+    // Called to process the user defines
     virtual void process_defines(const std::function<void(const string& define, const string& value)>) const = 0;
+    // Called to process the user compile time args
     virtual void process_compile_time_args(const std::function<void(int i, uint32_t value)>) const = 0;
 
-private:
-    bool use_multi_threaded_compile = true;
+    virtual ~JitBuildSettings() = default;
 };
 
 void jit_build(const JitBuildState& build, const JitBuildSettings* settings);

tt_metal/api/tt-metalium/kernel.hpp

@@ -1,9 +1,10 @@
-// SPDX-FileCopyrightText: © 2023 Tenstorrent Inc.
+// SPDX-FileCopyrightText: © 2025 Tenstorrent Inc.
 //
 // SPDX-License-Identifier: Apache-2.0
 
 #pragma once
 
+#include <string_view>
 #include <vector>
 #include <map>
 #include <variant>
@@ -176,7 +177,11 @@ class DataMovementKernel : public Kernel {
 
     void process_defines(const std::function<void (const string& define, const string &value)>) const override;
 
-   private:
+    std::string_view get_compiler_opt_level() const override;
+
+    std::string_view get_linker_opt_level() const override;
+
+private:
     const DataMovementConfig config_;
 
     uint8_t expected_num_binaries() const override;
@@ -204,7 +209,11 @@ class EthernetKernel : public Kernel {
 
     void process_defines(const std::function<void(const string &define, const string &value)>) const override;
 
-   private:
+    std::string_view get_compiler_opt_level() const override;
+
+    std::string_view get_linker_opt_level() const override;
+
+private:
     const EthernetConfig config_;
 
     uint8_t expected_num_binaries() const override;
@@ -233,7 +242,11 @@ class ComputeKernel : public Kernel {
 
     void process_defines(const std::function<void (const string& define, const string &value)>) const override;
 
-   private:
+    std::string_view get_compiler_opt_level() const override;
+
+    std::string_view get_linker_opt_level() const override;
+
+private:
     const ComputeConfig config_;
 
     uint8_t expected_num_binaries() const override;

tt_metal/api/tt-metalium/kernel_types.hpp

@@ -15,15 +15,14 @@ namespace tt::tt_metal {
 
 using KernelHandle = std::uint32_t;
 
-// Option that controls compiler optimization level
+// Option that controls build optimization level
 enum class KernelBuildOptLevel : uint8_t {
-    O1,     // Level 1 optimization. Same as O.
-    O2,     // Level 2 optimization. Turns on all flags specified by O1.
-    O3,     // Level 3 optimizaiton. Turns on all flags specified by O2.
+    O1,     // Turns on level 1 optimization. Same as O.
+    O2,     // Turns on level 2 optimization and also all flags specified by O1.
+    O3,     // Turns on level 3 optimization and also all flags specified by O2.
     O0,     // Reduce compilation time and make debugging produce the expected results.
-    Os,     // Optimize for size. Enables O2 optimizations except for those that increase binary size.
-    Ofast,  // Enable all O3 and non standard optimizations.
-    Og,     // Optimize for debugging.
+    Os,     // Optimize for size and also O2 optimizations except for those that increase binary size.
+    Ofast,  // Turns on level O3 and also non standard optimizations.
     Oz,     // Aggresively optimize for size rather than speed.
 };
 
@@ -36,7 +35,7 @@ struct DataMovementConfig {
     // Each unique combination of defines will produce a unique compiled instantiation
     // This file is then automatically included in the generated compiled kernel files
     std::map<std::string, std::string> defines;
-    // Kernel optimization level
+    // Set the compiler and linker optimization level
     KernelBuildOptLevel opt_level = KernelBuildOptLevel::Os;
 };
 
@@ -60,7 +59,7 @@ struct ComputeConfig {
     // Each unique combination of defines will produce a unique compiled instantiation
     // This file is then automatically included in the generated compiled kernel files
     std::map<std::string, std::string> defines;
-    // Kernel optimization level
+    // Set the compiler and linker optimization level
     KernelBuildOptLevel opt_level = KernelBuildOptLevel::O3;
 };
 
@@ -73,7 +72,7 @@ struct EthernetConfig {
     // Each unique combination of defines will produce a unique compiled instantiation
     // This file is then automatically included in the generated compiled kernel files
     std::map<std::string, std::string> defines;
-    // Kernel optimization level
+    // Set the compiler and linker optimization level
     KernelBuildOptLevel opt_level = KernelBuildOptLevel::Os;
 };
 

tt_metal/impl/kernels/kernel.cpp

@@ -1,16 +1,19 @@
-// SPDX-FileCopyrightText: © 2023 Tenstorrent Inc.
+// SPDX-FileCopyrightText: © 2025 Tenstorrent Inc.
 //
 // SPDX-License-Identifier: Apache-2.0
 
 #include <kernel.hpp>
+#include <kernel_types.hpp>
 
 #include <fmt/core.h>
 #include <fmt/ranges.h>
 
+#include <magic_enum/magic_enum.hpp>
 #include <set>
 
 #include <build.hpp>
 #include "llrt.hpp"
+#include <string_view>
 #include <tt_metal.hpp>
 #include "tt_metal/impl/debug/watcher_server.hpp"
 #include "tt_metal/kernel.hpp"
@@ -105,7 +108,7 @@ CoreType Kernel::get_kernel_core_type() const {
     return CoreType::WORKER;
 }
 
-const string &Kernel::get_full_kernel_name() const { return this->kernel_full_name_; }
+const std::string& Kernel::get_full_kernel_name() const { return this->kernel_full_name_; }
 
 void Kernel::add_defines(const std::map<std::string, std::string>& defines) {
     this->defines_.insert(defines.begin(), defines.end());
@@ -141,6 +144,24 @@ void EthernetKernel::process_defines(
     callback("NOC_MODE", std::to_string(NOC_MODE::DM_DEDICATED_NOC));
 }
 
+std::string_view DataMovementKernel::get_compiler_opt_level() const {
+    return magic_enum::enum_name(this->config_.opt_level);
+}
+
+std::string_view DataMovementKernel::get_linker_opt_level() const { return this->get_compiler_opt_level(); }
+
+std::string_view ComputeKernel::get_compiler_opt_level() const {
+    return magic_enum::enum_name(this->config_.opt_level);
+}
+
+std::string_view ComputeKernel::get_linker_opt_level() const { return this->get_compiler_opt_level(); }
+
+std::string_view EthernetKernel::get_compiler_opt_level() const {
+    return magic_enum::enum_name(this->config_.opt_level);
+}
+
+std::string_view EthernetKernel::get_linker_opt_level() const { return this->get_compiler_opt_level(); }
+
 void Kernel::process_compile_time_args(const std::function<void(int i, uint32_t value)> callback) const {
     for (int i = 0; i < this->compile_time_args_.size(); i++) {
         callback(i, this->compile_time_args_[i]);