token scratch

CHANGELOG.md

@@ -5,6 +5,7 @@
 ## Current develop
 
 ### Added (new features/APIs/variables/...)
+- [[PR 1394]](https://github.com/parthenon-hpc-lab/parthenon/pull/1394) Add token scratch utility
 - [[PR 1403]](https://github.com/parthenon-hpc-lab/parthenon/pull/1403) Add interface for Fourier transforms on uniform meshes via heFFTe
 - [[PR 1408]](https://github.com/parthenon-hpc-lab/parthenon/pull/1408) Add GetAsUnresolvedString() method to ParameterInput
 - [[PR 1050]](https://github.com/parthenon-hpc-lab/parthenon/pull/1050) Add support for OpenPMD/ADIOS2 output (incl slices and coarsened dumps)

doc/plan_histories/token_scratch.md

@@ -0,0 +1,213 @@
+# Token-Based Scratch Memory System
+
+## Overview
+
+The token scratch system provides efficient, thread-safe scratch memory management for Kokkos parallel kernels using `Kokkos::Experimental::UniqueToken`. Each parallel thread acquires a unique token ID and gets access to a pre-allocated scratch buffer that can be carved into multiple typed views.
+
+## Key Features
+
+- **Thread-safe**: Uses Kokkos UniqueToken to avoid race conditions
+- **Type-safe**: Automatic type deduction and alignment handling
+- **Zero runtime overhead**: All inline functions, compile-time resolution
+- **Flexible allocation**: Supports 1D, 2D, and 3D views of arbitrary types
+- **RAII-based**: Automatic token acquisition and release
+- **Debug validation**: Bounds checking in debug builds
+
+## Components
+
+### 1. `TokenScratchPool<MemorySpace, ExecutionSpace, TokenScope>`
+
+The main pool manager that allocates scratch memory for all tokens.
+
+**Template Parameters:**
+- `MemorySpace` - Kokkos memory space (e.g., `HostSpace`, device memory space)
+- `ExecutionSpace` - Kokkos execution space
+- `TokenScope` (optional) - `UniqueTokenScope::Global` (default) or `UniqueTokenScope::Instance`
+
+**Constructor:**
+```cpp
+explicit TokenScratchPool(size_t bytes_per_token)
+```
+
+**Methods:**
+- `acquire()` - Returns a `ScratchAllocator` for the current thread
+- `size()` - Number of tokens in the pool
+- `bytes_per_token()` - Bytes allocated per token
+- `total_bytes()` - Total pool size
+
+### 2. `ScratchAllocator<MemorySpace, ExecutionSpace>`
+
+Per-token allocator that carves typed views from the token's buffer.
+
+**Methods:**
+- `allocate_view<T>(n)` - Allocate 1D view with n elements
+- `allocate_view<T>(n1, n2)` - Allocate 2D view
+- `allocate_view<T>(n1, n2, n3)` - Allocate 3D view
+- `remaining()` - Get remaining capacity
+- `reset()` - Reset allocator to beginning (for reuse)
+
+### 3. `TypedScratchBundle<MemorySpace, ExecutionSpace, ArraySpecs...>`
+
+(Optional) Pre-configured bundle for compile-time known layouts.
+
+## Usage Examples
+
+### Basic Usage
+
+```cpp
+using ExecSpace = Kokkos::DefaultExecutionSpace;
+using MemSpace = ExecSpace::memory_space;
+
+// Create pool with 64KB per token
+TokenScratchPool<MemSpace, ExecSpace> pool(64 * 1024);
+
+Kokkos::parallel_for("my_kernel", N, KOKKOS_LAMBDA(int i) {
+  // Acquire scratch allocator
+  auto scratch = pool.acquire();
+
+  // Allocate multiple typed views
+  auto doubles = scratch.template allocate_view<double>(100);
+  auto ints = scratch.template allocate_view<int>(50);
+  auto flags = scratch.template allocate_view<bool>(200);
+
+  // Use the views...
+  for (int j = 0; j < 100; ++j) {
+    doubles(j) = compute(i, j);
+  }
+
+  // Token automatically released when scratch goes out of scope
+});
+```
+
+### Multi-dimensional Views
+
+```cpp
+TokenScratchPool<MemSpace, ExecSpace> pool(128 * 1024);
+
+Kokkos::parallel_for("multidim", N, KOKKOS_LAMBDA(int i) {
+  auto scratch = pool.acquire();
+
+  // 2D and 3D scratch arrays
+  auto work_2d = scratch.template allocate_view<double>(ni, nj);
+  auto work_3d = scratch.template allocate_view<double>(ni, nj, nk);
+
+  // Access as: work_2d(i, j), work_3d(i, j, k)
+});
+```
+
+## Design Considerations
+
+### Memory Layout
+
+The pool allocates a 2D view: `View<char**>` with shape `[num_tokens, bytes_per_token]`.
+
+- Each row is owned by one token
+- Allocations within a token are sequential with proper alignment
+- No fragmentation within a token's buffer
+
+### Token Scope
+
+Two scope options (specified as a template parameter):
+
+1. **`UniqueTokenScope::Global`** (default)
+   - Tokens shared across all kernel invocations
+   - More token reuse, potentially fewer tokens needed
+   - May see contention under high concurrency
+   ```cpp
+   // Explicit Global scope (or omit for default)
+   TokenScratchPool<MemSpace, ExecSpace, 
+                    Kokkos::Experimental::UniqueTokenScope::Global> 
+       pool(scratch_bytes);
+   ```
+
+2. **`UniqueTokenScope::Instance`**
+   - Tokens private to each kernel instance
+   - Better for very high thread counts
+   - May allocate more tokens
+   ```cpp
+   // Use Instance scope for high concurrency
+   TokenScratchPool<MemSpace, ExecSpace,
+                    Kokkos::Experimental::UniqueTokenScope::Instance>
+       pool(scratch_bytes);
+   ```
+
+### Sizing Guidelines
+
+Choose `bytes_per_token` based on:
+
+1. **Per-thread scratch needs**: Sum of all allocations in your kernel
+2. **Alignment overhead**: Add ~20% for alignment padding
+3. **Safety margin**: Add 10-20% buffer for debug assertions
+
+Example calculation:
+```cpp
+// Kernel needs:
+// - 100 doubles = 800 bytes
+// - 50 ints = 200 bytes  
+// - 200 bools = 200 bytes
+// Total: 1200 bytes
+// With alignment (20%): 1440 bytes
+// With safety margin (20%): 1728 bytes → round to 2KB
+constexpr size_t scratch_bytes = 2 * 1024;
+```
+
+### Performance Notes
+
+- **Zero overhead in release builds**: All functions inline, no virtual dispatch
+- **Debug validation**: Bounds checking only in debug builds (`#ifndef NDEBUG`)
+- **No dynamic allocation**: All memory pre-allocated at pool creation
+- **Cache-friendly**: Each token's buffer is contiguous
+
+## Integration with the rest of Parthenon
+
+```cpp
+// In package initialization or task setup
+namespace Package {
+
+void MyTask(MeshData<Real>* md) {
+  auto pmb = md->GetBlockData(0)->GetBlockPointer();
+  const auto &ib = pmb->cellbounds.GetBoundsI(IndexDomain::interior);
+  const auto &jb = pmb->cellbounds.GetBoundsJ(IndexDomain::interior);
+  const auto &kb = pmb->cellbounds.GetBoundsK(IndexDomain::interior);
+
+  // Size scratch based on block dimensions
+  const size_t ni = ib.e - ib.s + 1;
+  const size_t nj = jb.e - jb.s + 1;
+  const size_t nk = kb.e - kb.s + 1;
+
+  const size_t scratch_per_k = (ni * nj) * sizeof(Real) * n_vars;
+  const size_t total_scratch = scratch_per_k * 1.3; // 30% padding
+
+  TokenScratchPool<DeviceSpace, ExecSpace> pool(total_scratch);
+
+  // Use in block loop...
+}
+
+} // namespace Package
+```
+
+## Error Handling
+
+- **Overflow detection**: In debug builds, `Kokkos::abort()` if allocation exceeds capacity
+- **Release builds**: No checks for performance; ensure proper sizing
+- **Diagnostic methods**: Use `remaining()` and `current_offset()` to debug sizing
+
+## Thread Safety
+
+- **Acquire/release**: Managed by Kokkos UniqueToken (thread-safe)
+- **Within token**: No synchronization needed - each thread owns its token
+- **Across tokens**: No shared state between tokens
+
+## Future Extensions
+
+Possible enhancements:
+
+1. **Named allocations**: Track allocation names for debugging
+2. **High-water mark tracking**: Monitor maximum usage per token
+3. **Hierarchical scratch**: Separate team-shared and thread-private levels
+4. **Template bundle interface**: Compile-time view configuration
+
+## References
+
+- [Kokkos UniqueToken Documentation](https://kokkos.org/kokkos-core-wiki/API/core/UniqueToken.html)
+- Parthenon scratch memory patterns

doc/sphinx/src/par_for.rst

@@ -75,6 +75,40 @@ When ommitted the ``DEFAULT_LOOP_PATTERN`` is used.
        flattened into an outer ``TeamThreadRange``. The specializations ``loop_pattern_[tpttr|tptvr|tpttrtvr]_tag`` correspond
        to ``<1,0>``, ``<0,1>``, ``<1,1>`` respectively.
 
+Token Scratch Utility
+---------------------
+
+For kernels launched with ``parthenon::par_for`` or directly with Kokkos, Parthenon provides
+``parthenon::TokenScratchPool`` in ``utils/token_scratch.hpp`` for temporary per-thread workspace.
+The pool pre-allocates a fixed number of bytes for each
+``Kokkos::Experimental::UniqueToken``, and ``acquire()`` returns a ``ScratchAllocator`` that can
+carve one or more unmanaged typed ``Kokkos::View`` objects from that token-local buffer. This is
+useful when a kernel needs short-lived scratch storage without performing repeated dynamic
+allocations inside the parallel region.
+
+.. code:: cpp
+
+   parthenon::TokenScratchPool<> pool(64 * 1024);
+
+   parthenon::par_for(
+       "token_scratch_example", 0, nblocks - 1, KOKKOS_LAMBDA(const int b) {
+         auto scratch = pool.acquire();
+         auto work = scratch.template allocate_view<double>(ni, nj);
+         // Use work(...) as temporary storage for this token.
+       });
+
+.. note::
+
+  The view created by ``allocate_view<T>`` is unmanaged and thus
+  non-blocking when destructed. The backing view, which is created by
+  the ``TokenScratchPool`` object is owning and thus **does** fence
+  upon going out of scope.
+
+.. warning::
+
+  The token scratch infrastructure is not tested for hierarchical
+  parallelism. In this case, we recommend that you use the team
+  scratch provided by Kokkos.
 
 Cmake Options
 -------------
@@ -129,4 +163,3 @@ New types can be provided by specializing the ``ProcessLoopBound`` struct in the
 These structs need to provide a ``GetNumBounds`` method to count the number of start/end bounds contained
 in the type, as well as a ``GetIndexRanges`` method to fill the ``IndexRange`` bounds used in the
 parallel dispatch.
-

src/CMakeLists.txt

@@ -321,6 +321,7 @@ add_library(parthenon
   utils/sort.hpp
   utils/string_utils.cpp
   utils/string_utils.hpp
+  utils/token_scratch.hpp
   utils/type_list.hpp
   utils/unique_id.cpp
   utils/unique_id.hpp

src/utils/cleantypes.hpp

@@ -14,6 +14,8 @@
 #ifndef UTILS_CLEANTYPES_HPP_
 #define UTILS_CLEANTYPES_HPP_
 
+// This file made with the assistance of generative AI
+
 namespace parthenon {
 namespace cleantypes {
 
@@ -44,6 +46,19 @@ struct remove_all_pointers<T *const volatile> {
   using type = typename remove_all_pointers<T>::type;
 };
 
+//! Helper to build pointer types with specified depth
+//! E.g., pointer_depth<T, 3>::type = T***
+//! Used for constructing multi-dimensional Kokkos::View data types
+template <typename T, std::size_t Rank>
+struct pointer_depth {
+  using type = typename pointer_depth<T *, Rank - 1>::type;
+};
+
+template <typename T>
+struct pointer_depth<T, 1> {
+  using type = T *;
+};
+
 } // namespace cleantypes
 } // namespace parthenon