refactor(benchmark): hide CudaTimer as internal detail and remove dead code

- Move CudaTimer into detail namespace as implementation detail
  - Only measureGpuTime() needs it, not part of public API
  - Remove unused startEvent()/stopEvent() methods
- Delete SgemmReferenceCalculator class (unused for 78 lines)
  - Was designed for reference computation but never used
  - SGEMMBenchmark and tests use inline cuBLAS calls instead

Author: LessUp

src/utils/benchmark_core.cuh

@@ -4,14 +4,13 @@
 #include <cuda_runtime.h>
 
 // ============================================================================
-// CUDA 事件计时器
+// CUDA 性能测量器
 // ============================================================================
 
+namespace detail {
+
 /**
- * RAII 包装的 CUDA 事件计时器
- *
- * 提供精确的 GPU 时间测量，使用 CUDA 事件机制。
- * 适用于内核执行时间、数据传输时间等测量。
+ * RAII 包装的 CUDA 事件计时器（内部实现）
  */
 class CudaTimer {
   public:
@@ -45,18 +44,12 @@ class CudaTimer {
         return ms;
     }
 
-    // 获取开始和结束事件（用于手动记录）
-    cudaEvent_t startEvent() { return start_; }
-    cudaEvent_t stopEvent() { return stop_; }
-
   private:
     cudaEvent_t start_;
     cudaEvent_t stop_;
 };
 
-// ============================================================================
-// 性能测量器
-// ============================================================================
+} // namespace detail
 
 /**
  * 通用的 GPU 操作性能测量器
@@ -75,7 +68,7 @@ float measureGpuTime(RunFunc func, int warmup_runs = 5, int benchmark_runs = 20)
     CUDA_CHECK(cudaDeviceSynchronize());
 
     // 计时运行
-    CudaTimer timer;
+    detail::CudaTimer timer;
     timer.start();
     for (int i = 0; i < benchmark_runs; ++i) {
         func();

src/utils/benchmark_cublas.cuh

@@ -73,81 +73,3 @@ class CublasSgemm {
   private:
     cublasHandle_t handle_;
 };
-
-// ============================================================================
-// SGEMM 参考计算器
-// ============================================================================
-
-/**
- * 计算参考 SGEMM 结果并验证
- *
- * 提供完整的参考计算流程：
- * - 设备内存分配
- * - 随机矩阵初始化
- * - cuBLAS 参考计算
- * - 结果验证
- */
-class SgemmReferenceCalculator {
-  public:
-    SgemmReferenceCalculator() = default;
-
-    /**
-     * 初始化参考数据
-     *
-     * @param M, K, N 矩阵维度
-     * @param seed 随机种子（可重现）
-     */
-    void initialize(int M, int K, int N, unsigned int seed_A = 42, unsigned int seed_B = 123) {
-        M_ = M;
-        K_ = K;
-        N_ = N;
-
-        h_A_.resize(M * K);
-        h_B_.resize(K * N);
-
-        initRandomMatrix(h_A_.data(), M, K, -1.0f, 1.0f, seed_A);
-        initRandomMatrix(h_B_.data(), K, N, -1.0f, 1.0f, seed_B);
-
-        d_A_ = std::make_unique<DeviceMemory<float>>(M * K);
-        d_B_ = std::make_unique<DeviceMemory<float>>(K * N);
-        d_C_ref_ = std::make_unique<DeviceMemory<float>>(M * N);
-
-        d_A_->copyFromHost(h_A_.data(), M * K);
-        d_B_->copyFromHost(h_B_.data(), K * N);
-    }
-
-    /**
-     * 计算参考结果
-     */
-    void computeReference() {
-        cublas_.sgemm(d_A_->get(), d_B_->get(), d_C_ref_->get(), M_, K_, N_);
-    }
-
-    /**
-     * 获取主机端参考结果
-     */
-    std::vector<float> getReferenceResult() {
-        std::vector<float> h_ref(M_ * N_);
-        d_C_ref_->copyToHost(h_ref.data(), M_ * N_);
-        return h_ref;
-    }
-
-    // 访问器
-    const std::vector<float> &hostA() const { return h_A_; }
-    const std::vector<float> &hostB() const { return h_B_; }
-    float *deviceA() { return d_A_->get(); }
-    float *deviceB() { return d_B_->get(); }
-    float *deviceCRef() { return d_C_ref_->get(); }
-    CublasSgemm &cublas() { return cublas_; }
-    int M() const { return M_; }
-    int K() const { return K_; }
-    int N() const { return N_; }
-
-  private:
-    int M_ = 0, K_ = 0, N_ = 0;
-    std::vector<float> h_A_, h_B_;
-    std::unique_ptr<DeviceMemory<float>> d_A_;
-    std::unique_ptr<DeviceMemory<float>> d_B_;
-    std::unique_ptr<DeviceMemory<float>> d_C_ref_;
-    CublasSgemm cublas_;
-};