Enable restore_tracker usage in device kernels

Replace the STL-based implementation of clad::restore_tracker with a
device-compatible implementation. The previous implementation relied on
std::vector and std::map, which are unavailable in device code,
preventing restore_tracker from being used inside GPU kernels.

Author: Vedant2005goyal

include/clad/Differentiator/RestoreTracker.h

@@ -7,6 +7,12 @@
 #include <map>
 #include <utility>
 #include <vector>
+#ifndef Max_Records
+#define Max_Records 64
+#endif
+#ifndef Max_Bytes
+#define Max_Bytes 1024
+#endif
 
 namespace clad {
 
@@ -21,6 +27,41 @@ namespace clad {
 /// clad::tape is not viable.
 class restore_tracker {
   // m_data consists of pairs of memory addresses and bitwise values
+#ifdef __CUDACC__
+  struct MetaData {
+    char* addr;
+    size_t size;
+    size_t off;
+  };
+  MetaData m_meta[Max_Records];
+  uint8_t m_buf[Max_Bytes];
+  size_t m_cnt = 0, m_off = 0;
+
+public:
+  __host__ __device__ restore_tracker() = default;
+
+  template <typename T> __host__ __device__ void store(const T& val) {
+    for (size_t i = 0; i < m_cnt; ++i)
+      if (m_meta[i].addr == (char*)&val)
+        return;
+
+    if (m_cnt >= Max_Records || m_off + sizeof(T) > Max_Bytes) {
+      // Clad restore_tracker GPU capacity exceeded. Try again with larger value
+      return;
+    }
+
+    m_meta[m_cnt] = {(char*)&val, sizeof(T), m_off};
+    std::memcpy(m_buf + m_off, &val, sizeof(T));
+    m_off += sizeof(T);
+    m_cnt++;
+  }
+
+  __host__ __device__ void restore() {
+    for (size_t i = 0; i < m_cnt; ++i)
+      std::memcpy(m_meta[i].addr, m_buf + m_meta[i].off, m_meta[i].size);
+    m_cnt = m_off = 0;
+  }
+#else
   using RawMemory = std::vector<uint8_t>;
   using Address = char*;
   std::map<const Address, RawMemory> m_data;
@@ -47,6 +88,7 @@ class restore_tracker {
     }
     m_data.clear();
   }
+#endif
 };
 } // namespace clad
 

test/CUDA/RestoreTrackerDevice.cu

@@ -0,0 +1,30 @@
+// RUN: %cladclang_cuda -I%S/../../include --cuda-path=%cudapath \
+// RUN:     --cuda-gpu-arch=%cudaarch %cudaldflags -o RestoreTrackerDevice.out %s
+//
+// RUN: %cudarun ./RestoreTrackerDevice.out | %filecheck_exec %s
+//
+// REQUIRES: cuda-runtime
+#include "clad/Differentiator/Differentiator.h"
+#include <iostream>
+
+__global__ void test_kernel() {
+    clad::restore_tracker tracker;
+    
+    double val = 3.14;
+    tracker.store(val);
+    val = 0.0;
+    tracker.restore();
+    
+    if (val == 3.14) {
+        printf("Working on device!\n");
+    }
+}
+
+int main() {
+    
+    test_kernel<<<1, 1>>>();
+    cudaDeviceSynchronize();
+    
+    // CHECK-EXEC: Working on device!
+    return 0;
+}