From c65cee1db07508c27a780ece0558c10684b4acb1 Mon Sep 17 00:00:00 2001
From: Frederik Hennig <frederik.hennig@fau.de>
Date: Mon, 17 Mar 2025 14:45:45 +0100
Subject: [PATCH] Introduce MemTag. Use it in SparseIndexList. Adapt codegen to
use entities and views.
- Introduce the MemTag protocol (from my master's thesis)
- Implement CUDA and HIP unified memory allocator
- Adapt SparseIndexList to use memtags and select allocators according to them
- Adapt codegen to use the entity/view pattern for all fields such that the block data ID of the SparseIndexList is now fully obscured
- Make FreeSlipPipe test case run on GPU
---
lib/CMakeLists.txt | 3 +
lib/walberla/experimental/Memory.hpp | 4 +
.../experimental/lbm/IrregularFreeSlip.hpp | 340 +++++++++---------
.../experimental/memory/MemoryTags.hpp | 68 ++++
.../memory/UnifiedMemoryAllocator.hpp | 49 +++
.../experimental/sweep/SparseIndexList.hpp | 149 ++++----
src/walberla/codegen/api.py | 56 ++-
src/walberla/codegen/boundaries/hbb.py | 12 +-
src/walberla/codegen/boundaries/linkwise.py | 8 +-
src/walberla/codegen/sweep.py | 139 ++++---
tests/BasicLbmScenarios/SimDomain.hpp | 9 +
.../TestBasicLbmScenarios.cpp | 31 +-
12 files changed, 550 insertions(+), 318 deletions(-)
create mode 100644 lib/walberla/experimental/Memory.hpp
create mode 100644 lib/walberla/experimental/memory/MemoryTags.hpp
create mode 100644 lib/walberla/experimental/memory/UnifiedMemoryAllocator.hpp
diff --git a/lib/CMakeLists.txt b/lib/CMakeLists.txt
index 15a074a..7619333 100644
--- a/lib/CMakeLists.txt
+++ b/lib/CMakeLists.txt
@@ -3,9 +3,12 @@ add_library( walberla::experimental ALIAS walberla_experimental )
target_sources( walberla_experimental
INTERFACE
+ walberla/experimental/memory/UnifiedMemoryAllocator.hpp
walberla/experimental/sweep/SparseIndexList.hpp
walberla/experimental/lbm/GenericHbbBoundary.hpp
walberla/experimental/lbm/IrregularFreeSlip.hpp
)
+target_link_libraries( walberla_experimental INTERFACE walberla_core walberla_gpu )
+
target_include_directories( walberla_experimental INTERFACE ${CMAKE_CURRENT_SOURCE_DIR} )
diff --git a/lib/walberla/experimental/Memory.hpp b/lib/walberla/experimental/Memory.hpp
new file mode 100644
index 0000000..877249e
--- /dev/null
+++ b/lib/walberla/experimental/Memory.hpp
@@ -0,0 +1,4 @@
+#pragma once
+
+#include "./memory/MemoryTags.hpp"
+#include "./memory/UnifiedMemoryAllocator.hpp"
diff --git a/lib/walberla/experimental/lbm/IrregularFreeSlip.hpp b/lib/walberla/experimental/lbm/IrregularFreeSlip.hpp
index 3332d8a..0b9097f 100644
--- a/lib/walberla/experimental/lbm/IrregularFreeSlip.hpp
+++ b/lib/walberla/experimental/lbm/IrregularFreeSlip.hpp
@@ -2,182 +2,186 @@
#include "core/DataTypes.h"
#include "core/cell/Cell.h"
+
#include "domain_decomposition/BlockDataID.h"
#include "domain_decomposition/IBlock.h"
-#include "field/GhostLayerField.h"
+
#include "field/FlagField.h"
-#include "walberla/experimental/sweep/SparseIndexList.hpp"
+#include "field/GhostLayerField.h"
+
#include "stencil/Directions.h"
+
+#include <memory>
#include <tuple>
+#include <type_traits>
+
+#include "walberla/experimental/Memory.hpp"
+#include "walberla/experimental/sweep/SparseIndexList.hpp"
namespace walberla::experimental::lbm
{
- struct IrregularFreeSlipLinkInfo
- {
- int32_t x;
- int32_t y;
- int32_t z;
- int32_t dir;
- int32_t source_offset_x;
- int32_t source_offset_y;
- int32_t source_offset_z;
- int32_t source_dir;
- IrregularFreeSlipLinkInfo(
- walberla::Cell fluidCell, walberla::stencil::Direction linkDir,
- walberla::Cell sourceCell, walberla::stencil::Direction sourceDir)
- : x{fluidCell.x()}, y{fluidCell.y()}, z{fluidCell.z()},
- dir{int32_t(linkDir)}, source_offset_x{sourceCell.x() - fluidCell.x()},
- source_offset_y{sourceCell.y() - fluidCell.y()}, source_offset_z{sourceCell.z() - fluidCell.z()},
- source_dir{int32_t(sourceDir)} {}
-
- bool operator==(const IrregularFreeSlipLinkInfo &other) const
- {
- return std::tie(x, y, z, dir, source_offset_x, source_offset_y, source_offset_z, source_dir) ==
- std::tie(other.x, other.y, other.z, other.dir, other.source_offset_x,
- other.source_offset_y, other.source_offset_z, other.source_dir);
- }
- };
-
- namespace detail
- {
- template <typename Stencil, typename FlagField_T>
- class FreeSlipLinksFromFlagField
- {
- public:
- using IndexList = sweep::SparseIndexList<IrregularFreeSlipLinkInfo>;
- using flag_t = typename FlagField_T::flag_t;
-
- FreeSlipLinksFromFlagField(
- StructuredBlockForest &sbfs,
- BlockDataID flagFieldID,
- field::FlagUID boundaryFlagUID,
- field::FlagUID fluidFlagUID) : sbfs_{sbfs}, flagFieldID_(flagFieldID), boundaryFlagUID_(boundaryFlagUID), fluidFlagUID_(fluidFlagUID) {}
-
- IndexList collectLinks()
- {
- IndexList indexList{sbfs_};
-
- for (auto &block : sbfs_)
- {
- FlagField_T *flagField = block.getData<FlagField_T>(flagFieldID_);
- flag_t boundaryFlag{flagField->getFlag(boundaryFlagUID_)};
- flag_t fluidFlag{flagField->getFlag(fluidFlagUID_)};
-
- auto &idxVector = indexList.get(block);
-
- for (auto it = flagField->beginXYZ(); it != flagField->end(); ++it)
- {
- Cell c{it.cell()};
- if (!isFlagSet(it, fluidFlag))
- {
- continue;
- }
-
- for (auto dIt = Stencil::beginNoCenter(); dIt != Stencil::end(); ++dIt)
- {
- stencil::Direction dir{*dIt};
- Cell neighbor{c + dir};
- if (flagField->isFlagSet(neighbor, boundaryFlag))
- {
- idxVector.emplace_back(createLink(flagField, c, dir));
- }
- }
- }
- }
-
- return indexList;
- }
+struct IrregularFreeSlipLinkInfo
+{
+ int32_t x;
+ int32_t y;
+ int32_t z;
+ int32_t dir;
+ int32_t source_offset_x;
+ int32_t source_offset_y;
+ int32_t source_offset_z;
+ int32_t source_dir;
+ IrregularFreeSlipLinkInfo(walberla::Cell fluidCell, walberla::stencil::Direction linkDir, walberla::Cell sourceCell,
+ walberla::stencil::Direction sourceDir)
+ : x{ fluidCell.x() }, y{ fluidCell.y() }, z{ fluidCell.z() }, dir{ int32_t(linkDir) },
+ source_offset_x{ sourceCell.x() - fluidCell.x() }, source_offset_y{ sourceCell.y() - fluidCell.y() },
+ source_offset_z{ sourceCell.z() - fluidCell.z() }, source_dir{ int32_t(sourceDir) }
+ {}
+
+ bool operator==(const IrregularFreeSlipLinkInfo& other) const
+ {
+ return std::tie(x, y, z, dir, source_offset_x, source_offset_y, source_offset_z, source_dir) ==
+ std::tie(other.x, other.y, other.z, other.dir, other.source_offset_x, other.source_offset_y,
+ other.source_offset_z, other.source_dir);
+ }
+};
+
+namespace detail
+{
+template< typename Stencil, typename FlagField_T, typename IndexList_T >
+class FreeSlipLinksFromFlagField
+{
+ public:
+ using flag_t = typename FlagField_T::flag_t;
+
+ FreeSlipLinksFromFlagField(StructuredBlockForest& sbfs, BlockDataID flagFieldID, field::FlagUID boundaryFlagUID,
+ field::FlagUID fluidFlagUID)
+ : sbfs_{ sbfs }, flagFieldID_(flagFieldID), boundaryFlagUID_(boundaryFlagUID), fluidFlagUID_(fluidFlagUID)
+ {}
+
+ IndexList_T collectLinks()
+ {
+ IndexList_T indexList{ sbfs_ };
+
+ for (auto& block : sbfs_)
+ {
+ FlagField_T* flagField = block.getData< FlagField_T >(flagFieldID_);
+ flag_t boundaryFlag{ flagField->getFlag(boundaryFlagUID_) };
+ flag_t fluidFlag{ flagField->getFlag(fluidFlagUID_) };
- private:
- IrregularFreeSlipLinkInfo createLink(FlagField_T *flagField, const Cell &fluidCell, const stencil::Direction dir)
+ auto& idxVector = indexList.getVector(block);
+
+ for (auto it = flagField->beginXYZ(); it != flagField->end(); ++it)
+ {
+ Cell c{ it.cell() };
+ if (!isFlagSet(it, fluidFlag)) { continue; }
+
+ for (auto dIt = Stencil::beginNoCenter(); dIt != Stencil::end(); ++dIt)
{
- const flag_t fluidFlag{flagField->getFlag(fluidFlagUID_)};
- const Cell wallCell{fluidCell + dir};
-
- // inverse direction of 'dir' as lattice vector
-
- const cell_idx_t ix = stencil::cx[stencil::inverseDir[dir]];
- const cell_idx_t iy = stencil::cy[stencil::inverseDir[dir]];
- const cell_idx_t iz = stencil::cz[stencil::inverseDir[dir]];
-
- stencil::Direction sourceDir = stencil::inverseDir[dir]; // compute reflected (mirrored) of inverse direction of 'dir'
-
- cell_idx_t wnx = 0; // compute "normal" vector of free slip wall
- cell_idx_t wny = 0;
- cell_idx_t wnz = 0;
-
- if (flagField->isFlagSet(wallCell.x() + ix, wallCell.y(), wallCell.z(), fluidFlag))
- {
- wnx = ix;
- sourceDir = stencil::mirrorX[sourceDir];
- }
- if (flagField->isFlagSet(wallCell.x(), wallCell.y() + iy, wallCell.z(), fluidFlag))
- {
- wny = iy;
- sourceDir = stencil::mirrorY[sourceDir];
- }
- if (flagField->isFlagSet(wallCell.x(), wallCell.y(), wallCell.z() + iz, fluidFlag))
- {
- wnz = iz;
- sourceDir = stencil::mirrorZ[sourceDir];
- }
-
- // concave corner (neighbors are non-fluid)
- if (wnx == 0 && wny == 0 && wnz == 0)
- {
- wnx = ix;
- wny = iy;
- wnz = iz;
- sourceDir = dir;
- }
-
- const Cell sourceCell{
- wallCell.x() + wnx,
- wallCell.y() + wny,
- wallCell.z() + wnz};
-
- return {
- fluidCell,
- dir,
- sourceCell,
- sourceDir};
+ stencil::Direction dir{ *dIt };
+ Cell neighbor{ c + dir };
+ if (flagField->isFlagSet(neighbor, boundaryFlag))
+ {
+ idxVector.emplace_back(createLink(flagField, c, dir));
+ }
}
-
- StructuredBlockForest &sbfs_;
- const BlockDataID flagFieldID_;
- const field::FlagUID boundaryFlagUID_;
- const field::FlagUID fluidFlagUID_;
- };
- }
-
- template <typename Impl>
- class IrregularFreeSlipFactory
- {
- public:
- IrregularFreeSlipFactory(const shared_ptr<StructuredBlockForest> blocks, BlockDataID pdfFieldID)
- : blocks_{blocks}, pdfFieldID_{pdfFieldID} {}
-
- template <typename FlagField_T>
- auto fromFlagField(BlockDataID flagFieldID, field::FlagUID boundaryFlagUID, field::FlagUID fluidFlagUID)
- {
- detail::FreeSlipLinksFromFlagField<typename Impl::Stencil, FlagField_T> linksFromFlagField{*blocks_, flagFieldID, boundaryFlagUID, fluidFlagUID};
- auto indexVector = linksFromFlagField.collectLinks();
- return impl().irregularFromIndexVector(indexVector);
- }
-
- protected:
- shared_ptr<StructuredBlockForest> blocks_;
- BlockDataID pdfFieldID_;
-
- private:
- Impl &impl()
- {
- return *static_cast<Impl *>(this);
- }
-
- const Impl &impl() const
- {
- return *static_cast<const Impl *>(this);
- }
- };
-}
+ }
+ }
+
+ return indexList;
+ }
+
+ private:
+ IrregularFreeSlipLinkInfo createLink(FlagField_T* flagField, const Cell& fluidCell, const stencil::Direction dir)
+ {
+ const flag_t fluidFlag{ flagField->getFlag(fluidFlagUID_) };
+ const Cell wallCell{ fluidCell + dir };
+
+ // inverse direction of 'dir' as lattice vector
+
+ const cell_idx_t ix = stencil::cx[stencil::inverseDir[dir]];
+ const cell_idx_t iy = stencil::cy[stencil::inverseDir[dir]];
+ const cell_idx_t iz = stencil::cz[stencil::inverseDir[dir]];
+
+ stencil::Direction sourceDir =
+ stencil::inverseDir[dir]; // compute reflected (mirrored) of inverse direction of 'dir'
+
+ cell_idx_t wnx = 0; // compute "normal" vector of free slip wall
+ cell_idx_t wny = 0;
+ cell_idx_t wnz = 0;
+
+ if (flagField->isFlagSet(wallCell.x() + ix, wallCell.y(), wallCell.z(), fluidFlag))
+ {
+ wnx = ix;
+ sourceDir = stencil::mirrorX[sourceDir];
+ }
+ if (flagField->isFlagSet(wallCell.x(), wallCell.y() + iy, wallCell.z(), fluidFlag))
+ {
+ wny = iy;
+ sourceDir = stencil::mirrorY[sourceDir];
+ }
+ if (flagField->isFlagSet(wallCell.x(), wallCell.y(), wallCell.z() + iz, fluidFlag))
+ {
+ wnz = iz;
+ sourceDir = stencil::mirrorZ[sourceDir];
+ }
+
+ // concave corner (neighbors are non-fluid)
+ if (wnx == 0 && wny == 0 && wnz == 0)
+ {
+ wnx = ix;
+ wny = iy;
+ wnz = iz;
+ sourceDir = dir;
+ }
+
+ const Cell sourceCell{ wallCell.x() + wnx, wallCell.y() + wny, wallCell.z() + wnz };
+
+ return { fluidCell, dir, sourceCell, sourceDir };
+ }
+
+ StructuredBlockForest& sbfs_;
+ const BlockDataID flagFieldID_;
+ const field::FlagUID boundaryFlagUID_;
+ const field::FlagUID fluidFlagUID_;
+};
+} // namespace detail
+
+template< typename Impl >
+struct IrregularFreeSlipFactoryImplTraits
+{
+ using Stencil_T = typename Impl::Stencil;
+ using IdxStruct_T = IrregularFreeSlipLinkInfo;
+ using IndexListMemTag_T = typename Impl::memtag_t;
+ using IndexList_T = sweep::SparseIndexList< IrregularFreeSlipLinkInfo, IndexListMemTag_T >;
+};
+
+template< typename Impl >
+class IrregularFreeSlipFactory
+{
+ using ImplTraits = IrregularFreeSlipFactoryImplTraits< Impl >;
+
+ public:
+ IrregularFreeSlipFactory(const shared_ptr< StructuredBlockForest > blocks, BlockDataID pdfFieldID)
+ : blocks_{ blocks }, pdfFieldID_{ pdfFieldID }
+ {}
+
+ template< typename FlagField_T >
+ auto fromFlagField(BlockDataID flagFieldID, field::FlagUID boundaryFlagUID, field::FlagUID fluidFlagUID)
+ {
+ detail::FreeSlipLinksFromFlagField< typename ImplTraits::Stencil_T, FlagField_T, typename ImplTraits::IndexList_T > linksFromFlagField{
+ *blocks_, flagFieldID, boundaryFlagUID, fluidFlagUID
+ };
+ auto indexVector = linksFromFlagField.collectLinks();
+ return impl().irregularFromIndexVector(indexVector);
+ }
+
+ protected:
+ shared_ptr< StructuredBlockForest > blocks_;
+ BlockDataID pdfFieldID_;
+
+ private:
+ Impl& impl() { return *static_cast< Impl* >(this); }
+
+ const Impl& impl() const { return *static_cast< const Impl* >(this); }
+};
+} // namespace walberla::experimental::lbm
diff --git a/lib/walberla/experimental/memory/MemoryTags.hpp b/lib/walberla/experimental/memory/MemoryTags.hpp
new file mode 100644
index 0000000..01c80e2
--- /dev/null
+++ b/lib/walberla/experimental/memory/MemoryTags.hpp
@@ -0,0 +1,68 @@
+#pragma once
+
+#include <concepts>
+#include <memory>
+
+#include "./UnifiedMemoryAllocator.hpp"
+
+namespace walberla::experimental::memory
+{
+
+namespace memtag
+{
+
+/**
+ * @brief Common base class for all memory tags.
+ */
+struct _mem_tag
+{};
+
+/**
+ * @brief Regular host memory tag.
+ */
+struct host : public _mem_tag
+{};
+inline host host_v;
+
+/**
+ * @brief Memory tag indicating managed GPU memory with unified memory semantics.
+ */
+struct unified : public _mem_tag
+{};
+inline unified unified_v;
+
+} // namespace memtag
+
+template< typename T >
+concept MemTag = std::derived_from< T, memtag::_mem_tag >;
+
+namespace detail
+{
+template< MemTag memtag_t >
+struct SelectStandardAllocator;
+
+template<>
+struct SelectStandardAllocator< memtag::host >
+{
+ template< typename T >
+ using allocator = std::allocator< T >;
+};
+
+template<>
+struct SelectStandardAllocator< memtag::unified >
+{
+ template< typename T >
+ using allocator = UnifiedMemoryAllocator< T >;
+};
+} // namespace detail
+
+/**
+ * @brief Standard allocator implementation for a given memory tag and value type.
+ *
+ * This typedef alias references a type implementing the Allocator named requirement for type T
+ * which implements the memory semantics required by the given memory tag.
+ */
+template< MemTag memtag_t, typename T >
+using StandardAllocator_T = detail::SelectStandardAllocator< memtag_t >::template allocator< T >;
+
+} // namespace walberla::experimental::memory
diff --git a/lib/walberla/experimental/memory/UnifiedMemoryAllocator.hpp b/lib/walberla/experimental/memory/UnifiedMemoryAllocator.hpp
new file mode 100644
index 0000000..e30afec
--- /dev/null
+++ b/lib/walberla/experimental/memory/UnifiedMemoryAllocator.hpp
@@ -0,0 +1,49 @@
+#pragma once
+
+#include "waLBerlaDefinitions.h"
+#include "gpu/GPUWrapper.h"
+#include "gpu/ErrorChecking.h"
+
+namespace walberla::experimental::memory
+{
+
+#if defined(WALBERLA_BUILD_WITH_CUDA) || defined(WALBERLA_BUILD_WITH_HIP)
+
+/**
+ * @brief Allocator for managed GPU memory.
+ *
+ * This allocator implementes the Allocator named requirement and
+ * manages allocation and disposal of memory blocks with unified
+ * memory semantics, using the CUDA or HIP runtime (depending on which is active).
+ */
+template< typename T >
+class UnifiedMemoryAllocator
+{
+ public:
+ using value_type = T;
+ using pointer_type = T *;
+
+
+ pointer_type allocate(size_t n) {
+ pointer_type ptr;
+ size_t bytes { n * sizeof(value_type) };
+ WALBERLA_GPU_CHECK(
+ gpuMallocManaged(&ptr, bytes)
+ );
+ return ptr;
+ }
+
+ void deallocate(pointer_type ptr, size_t n) {
+ WALBERLA_GPU_CHECK(
+ gpuFree(ptr)
+ );
+ }
+
+ bool operator== (const UnifiedMemoryAllocator< value_type > &) {
+ return true;
+ }
+};
+
+#endif
+
+} // namespace walberla::experimental::memory
diff --git a/lib/walberla/experimental/sweep/SparseIndexList.hpp b/lib/walberla/experimental/sweep/SparseIndexList.hpp
index 6fbc30e..c94436e 100644
--- a/lib/walberla/experimental/sweep/SparseIndexList.hpp
+++ b/lib/walberla/experimental/sweep/SparseIndexList.hpp
@@ -1,80 +1,89 @@
#pragma once
-#include <vector>
+#include "blockforest/StructuredBlockForest.h"
+
#include "core/DataTypes.h"
#include "core/cell/Cell.h"
#include "domain_decomposition/BlockDataID.h"
-#include "blockforest/StructuredBlockForest.h"
+
+#include <type_traits>
+#include <vector>
+#include <span>
+
+#include "../memory/MemoryTags.hpp"
namespace walberla::experimental::sweep
{
- namespace internal
- {
- /**
- * @brief Index vectors for sparse sweeps
- */
- template <typename IndexStruct>
- class IndexListBuffer
- {
- public:
- using CpuVector = std::vector<IndexStruct>;
-
- IndexListBuffer() = default;
- bool operator==(const IndexListBuffer &other) const { return other.vec_ == vec_; }
-
- CpuVector &vector() { return vec_; }
-
- IndexStruct *pointerCpu() { return vec_.data(); }
-
- private:
- CpuVector vec_;
- };
- }
-
- struct CellIdx
- {
- int64_t x;
- int64_t y;
- int64_t z;
-
- CellIdx(const Cell &c)
- : x{int64_c(c.x())}, y{int64_c(c.y())}, z{int64_c(c.z())}
- {
- }
-
- bool operator==(const CellIdx &other) const
- {
- return std::tie(x, y, z) == std::tie(other.x, other.y, other.z);
- }
- };
-
- template <typename IndexStruct = CellIdx>
- class SparseIndexList
- {
- public:
- using Buffer_T = internal::IndexListBuffer<IndexStruct>;
-
- explicit SparseIndexList(StructuredBlockForest &sbfs) : bufferId_{createBuffers(sbfs)} {}
-
- std::vector<IndexStruct> &get(IBlock &block)
- {
- Buffer_T *buf = block.template getData<Buffer_T>(bufferId_);
+namespace internal
+{
+/**
+ * @brief Index vectors for sparse sweeps
+ */
+template< typename IndexStruct, typename Allocator_T >
+class IndexListBuffer
+{
+ public:
+ using Vector_T = std::vector< IndexStruct, Allocator_T >;
+
+ IndexListBuffer() = default;
+ bool operator==(const IndexListBuffer& other) const { return other.vec_ == vec_; }
+
+ Vector_T& vector() { return vec_; }
+
+ IndexStruct* data() { return vec_.data(); }
+ const IndexStruct* data() const { return vec_.data(); }
+ size_t size() const { return vec_.size(); }
+
+ private:
+ Vector_T vec_;
+};
+} // namespace internal
+
+struct CellIdx
+{
+ int64_t x;
+ int64_t y;
+ int64_t z;
+
+ CellIdx(const Cell& c) : x{ int64_c(c.x()) }, y{ int64_c(c.y()) }, z{ int64_c(c.z()) } {}
+
+ bool operator==(const CellIdx& other) const { return std::tie(x, y, z) == std::tie(other.x, other.y, other.z); }
+};
+
+template< typename IndexStruct = CellIdx, memory::MemTag memtag_t = memory::memtag::host >
+class SparseIndexList
+{
+ public:
+ using Buffer_T = internal::IndexListBuffer< IndexStruct, memory::StandardAllocator_T< memtag_t, IndexStruct > >;
+
+ explicit SparseIndexList(StructuredBlockForest& sbfs) : bufferId_{ createBuffers(sbfs) } {}
+
+ Buffer_T::Vector_T& getVector(IBlock& block)
+ {
+ Buffer_T* buf = block.template getData< Buffer_T >(bufferId_);
return buf->vector();
- }
-
- const BlockDataID & bufferId() const {
- return bufferId_;
- }
-
- private:
- static BlockDataID createBuffers(StructuredBlockForest &sbfs)
- {
- auto createIdxVector = [](IBlock *const, StructuredBlockStorage *const)
- { return new Buffer_T(); };
- return sbfs.addStructuredBlockData<Buffer_T>(createIdxVector);
- }
-
- BlockDataID bufferId_;
- };
-}
+ }
+
+ std::span< IndexStruct > view(IBlock& block){
+ Buffer_T& buf = *block.template getData< Buffer_T >(bufferId_);
+ return { buf.data(), buf.size() };
+ }
+
+ std::span< const IndexStruct > view(IBlock& block) const {
+ const Buffer_T& buf = *block.template getData< Buffer_T >(bufferId_);
+ return { buf.data(), buf.size() };
+ }
+
+ const BlockDataID& bufferId() const { return bufferId_; }
+
+ private:
+ static BlockDataID createBuffers(StructuredBlockForest& sbfs)
+ {
+ auto createIdxVector = [](IBlock* const, StructuredBlockStorage* const) { return new Buffer_T(); };
+ return sbfs.addStructuredBlockData< Buffer_T >(createIdxVector);
+ }
+
+ BlockDataID bufferId_;
+};
+} // namespace walberla::experimental::sweep
diff --git a/src/walberla/codegen/api.py b/src/walberla/codegen/api.py
index 53915c9..46d47c6 100644
--- a/src/walberla/codegen/api.py
+++ b/src/walberla/codegen/api.py
@@ -2,7 +2,7 @@ from __future__ import annotations
from typing import cast
-from pystencils import Field, DynamicType
+from pystencils import Field, DynamicType, FieldType, Target
from pystencils.types import (
UserTypeSpec,
create_type,
@@ -128,8 +128,8 @@ class IBlockPtr(_PlainCppClass):
def getAABB(self) -> AABB:
return AABB().bind("{}->getAABB()", self)
- def deref(self) -> AugExpr:
- return AugExpr.format("*{}", self)
+ def deref(self) -> IBlock:
+ return IBlock(ref=True).bind("*{}", self)
class SharedPtr(CppClass):
@@ -355,25 +355,65 @@ def glfield(field: Field, ci: str | AugExpr | None = None):
return GhostLayerFieldExtraction(field_ptr, ci)
+class MemTags:
+ _header = "walberla/experimental/Memory.hpp"
+
+ host = cpptype("walberla::experimental::memory::memtag::host", _header)()
+ unified = cpptype("walberla::experimental::memory::memtag::unified", _header)()
+
+
class SparseIndexList(AugExpr):
_template = cpptype(
- "walberla::experimental::sweep::SparseIndexList< {IndexStruct} >",
+ "walberla::experimental::sweep::SparseIndexList< {IndexStruct}, {memtag_t} >",
"walberla/experimental/sweep/SparseIndexList.hpp",
)
def __init__(
- self, idx_struct: PsStructType, const: bool = False, ref: bool = False
+ self,
+ idx_struct: PsStructType,
+ memtag_t: PsType = MemTags.host,
+ const: bool = False,
+ ref: bool = False,
):
self._idx_struct = idx_struct
- dtype = self._template(IndexStruct=idx_struct, const=const, ref=ref)
+ dtype = self._template(
+ IndexStruct=idx_struct, memtag_t=memtag_t, const=const, ref=ref
+ )
super().__init__(dtype)
- def get(self, block: IBlock) -> std.vector:
- return std.vector(self._idx_struct, ref=True).bind("{}.get({})", self, block)
+ def view_type(self) -> std.span:
+ return std.span(self._idx_struct)
+
+ def view(self, block: IBlock) -> std.span:
+ return std.span(self._idx_struct).bind("{}.view({})", self, block)
def bufferId(self) -> BlockDataID:
return BlockDataID().bind("{}.bufferId()", self)
+ @staticmethod
+ def from_field(
+ field: Field,
+ target: Target | None = None,
+ const: bool = False,
+ ref: bool = False,
+ ):
+ if (
+ not isinstance(field.dtype, PsStructType)
+ or field.field_type != FieldType.INDEXED
+ ):
+ raise ValueError(
+ "Can only create instances of SparseIndexList from index fields holding structures."
+ )
+
+ if target is not None and target.is_gpu():
+ memtag = MemTags.unified
+ else:
+ memtag = MemTags.host
+
+ return SparseIndexList(field.dtype, memtag_t=memtag, const=const, ref=ref).var(
+ field.name
+ )
+
class IndexListBufferPtr(AugExpr, SupportsFieldExtraction):
_template = cpptype(
diff --git a/src/walberla/codegen/boundaries/hbb.py b/src/walberla/codegen/boundaries/hbb.py
index 3e8f7a2..c249667 100644
--- a/src/walberla/codegen/boundaries/hbb.py
+++ b/src/walberla/codegen/boundaries/hbb.py
@@ -12,7 +12,7 @@ from lbmpy.boundaries.boundaryconditions import LbBoundary
from lbmpy.advanced_streaming.indexing import Timestep
from ..sweep import (
- FieldInfo,
+ GlFieldInfo,
SweepClassProperties,
BlockforestParameters,
combine_vectors,
@@ -55,9 +55,9 @@ class SimpleHbbBoundary(CustomGenerator):
khandle = knamespace.add(bc_kernel, self._name)
- domain_fields: list[FieldInfo] = sorted(
+ domain_fields: list[GlFieldInfo] = sorted(
(
- FieldInfo(
+ GlFieldInfo(
f, GhostLayerFieldPtr.create(f), BlockDataID().var(f"{f.name}Id")
)
for f in khandle.fields
@@ -98,13 +98,13 @@ class SimpleHbbBoundary(CustomGenerator):
),
# Get fields from block
*(
- sfg.init(fi.wlb_field_ptr)(
- block.getData(fi.wlb_field_ptr.field_type, fi.data_id)
+ sfg.init(fi.wlb_field)(
+ block.getData(fi.wlb_field.field_type, fi.data_id)
)
for fi in domain_fields
),
# Map GhostLayerFields to pystencils fields
- *(sfg.map_field(fi.field, fi.wlb_field_ptr) for fi in domain_fields),
+ *(sfg.map_field(fi.field, fi.wlb_field) for fi in domain_fields),
# Get index vector
sfg.init(idx_vector)(
block.getData(idx_vector.field_type, idx_vector_id)
diff --git a/src/walberla/codegen/boundaries/linkwise.py b/src/walberla/codegen/boundaries/linkwise.py
index 4459a59..a82e9b4 100644
--- a/src/walberla/codegen/boundaries/linkwise.py
+++ b/src/walberla/codegen/boundaries/linkwise.py
@@ -16,7 +16,7 @@ from pystencilssfg.composer.custom import CustomGenerator
from .hbb import BoundaryIndexType
from .boundary_utils import WalberlaLbmBoundary
from ..sweep import Sweep
-from ..api import SparseIndexList
+from ..api import SparseIndexList, MemTags
__all__ = ["FreeSlip"]
@@ -175,14 +175,15 @@ class FreeSlip(GenericLinkwiseBoundary):
factory_crtp_base = (
f"walberla::experimental::lbm::IrregularFreeSlipFactory< {factory_name} >"
)
+ memtag_t = MemTags.unified if bc_cfg.get_target().is_gpu() else MemTags.host
index_vector = SparseIndexList(
- WalberlaIrregularFreeSlip.idx_struct_type, ref=True
+ WalberlaIrregularFreeSlip.idx_struct_type, memtag_t=memtag_t, ref=True
).var("indexVector")
sweep_type = bc_sweep.generated_class()
sweep_ctor_args = {
f"{self._pdf_field.name}Id": "this->pdfFieldID_",
- f"{index_field.name}Id": index_vector.bufferId(),
+ f"{index_field.name}": index_vector,
}
stencil_name = self._lb_method.stencil.name
@@ -194,6 +195,7 @@ class FreeSlip(GenericLinkwiseBoundary):
f"friend class {factory_crtp_base};",
f"using Stencil = walberla::stencil::{stencil_name};",
f"using Sweep = {sweep_type.get_dtype().c_string()};",
+ f"using memtag_t = {memtag_t.c_string()};"
"using Base::Base;",
),
sfg.private(
diff --git a/src/walberla/codegen/sweep.py b/src/walberla/codegen/sweep.py
index 7239ff6..49cbcfe 100644
--- a/src/walberla/codegen/sweep.py
+++ b/src/walberla/codegen/sweep.py
@@ -4,6 +4,7 @@ from typing import Sequence, Iterable, cast
from dataclasses import dataclass
from itertools import chain
from collections import defaultdict
+from abc import ABC, abstractmethod
import warnings
import sympy as sp
@@ -30,8 +31,10 @@ from pystencilssfg.lang import (
AugExpr,
Ref,
strip_ptr_ref,
+ SupportsFieldExtraction,
SupportsVectorExtraction,
)
+from pystencilssfg.lang.cpp import std
from pystencilssfg.ir import SfgKernelHandle, SfgCallTreeNode
from .build_config import WalberlaBuildConfig
from .reflection import GeneratedClassWrapperBase
@@ -41,6 +44,7 @@ from .api import (
GhostLayerFieldPtr,
GpuFieldPtr,
IndexListBufferPtr,
+ SparseIndexList,
IBlockPtr,
BlockDataID,
Vector2,
@@ -258,28 +262,57 @@ class BlockforestParameters:
@dataclass
-class FieldInfo:
+class FieldInfo(ABC):
field: Field
- wlb_field_ptr: GenericWalberlaField | IndexListBufferPtr
+
+ @property
+ @abstractmethod
+ def entity(self) -> AugExpr: ...
+
+ @property
+ @abstractmethod
+ def view(self) -> AugExpr: ...
+
+ @abstractmethod
+ def create_view(self, block: IBlockPtr) -> AugExpr: ...
+
+ def view_extraction(self) -> SupportsFieldExtraction:
+ assert isinstance(self.view, SupportsFieldExtraction)
+ return self.view
+
+
+@dataclass
+class GlFieldInfo(FieldInfo):
+ glfield: GenericWalberlaField
data_id: BlockDataID
- def as_glfield(self) -> GenericWalberlaField:
- if not isinstance(self.wlb_field_ptr, GenericWalberlaField):
- raise AttributeError(
- "This FieldInfo does not encapsulate a ghost layer field"
- )
- return self.wlb_field_ptr
+ @property
+ def entity(self) -> AugExpr:
+ return self.data_id
- def is_glfield(self) -> bool:
- return isinstance(self.wlb_field_ptr, GenericWalberlaField)
+ @property
+ def view(self) -> AugExpr:
+ return self.glfield
+
+ def create_view(self, block: IBlockPtr) -> AugExpr:
+ return block.getData(self.glfield.field_type, self.data_id)
- def as_index_list(self) -> IndexListBufferPtr:
- if not isinstance(self.wlb_field_ptr, IndexListBufferPtr):
- raise AttributeError("This FieldInfo does not encapsulate an index list")
- return self.wlb_field_ptr
- def is_index_list(self) -> bool:
- return isinstance(self.wlb_field_ptr, IndexListBufferPtr)
+@dataclass
+class IndexListInfo(FieldInfo):
+ idx_list: SparseIndexList
+ idx_vector_view: std.span
+
+ @property
+ def entity(self) -> AugExpr:
+ return self.idx_list
+
+ @property
+ def view(self) -> AugExpr:
+ return self.idx_vector_view
+
+ def create_view(self, block: IBlockPtr) -> AugExpr:
+ return self.idx_list.view(block.deref())
class ShadowFieldCache:
@@ -290,20 +323,20 @@ class ShadowFieldCache:
"""
def __init__(self) -> None:
- self._originals: dict[str, FieldInfo] = dict()
+ self._originals: dict[str, GlFieldInfo] = dict()
- def add_shadow(self, original: FieldInfo):
+ def add_shadow(self, original: GlFieldInfo):
self._originals[original.field.name] = original
def get_shadow(self, original_name: str) -> AugExpr:
original = self._originals[original_name]
return AugExpr.format(
- "this->{}({})", self._getter(original.field.name), original.wlb_field_ptr
+ "this->{}({})", self._getter(original.field.name), original.glfield
)
- def perform_swap(self, original_name: str, shadow_field: FieldInfo) -> AugExpr:
+ def perform_swap(self, original_name: str, shadow_field: GlFieldInfo) -> AugExpr:
original = self._originals[original_name]
- return original.as_glfield().swapDataPointers(shadow_field.wlb_field_ptr)
+ return original.glfield.swapDataPointers(shadow_field.glfield)
def render(self, sfg: SfgComposer):
if not self._originals:
@@ -316,20 +349,20 @@ class ShadowFieldCache:
for orig_name, orig in self._originals.items():
unique_ptr_type = PsCustomType(
- f"std::unique_ptr< {orig.as_glfield().field_type} >"
+ f"std::unique_ptr< {orig.glfield.field_type} >"
)
- cache_ptr_name = self._cache_ptr(orig.as_glfield())
+ cache_ptr_name = self._cache_ptr(orig.glfield)
cache_ptrs.append(sfg.var(cache_ptr_name, unique_ptr_type))
getters.append(
sfg.method(self._getter(orig_name))
- .returns(orig.wlb_field_ptr.get_dtype())
+ .returns(orig.glfield.get_dtype())
.inline()(
sfg.branch(f"{cache_ptr_name} == nullptr")(
AugExpr.format(
"{}.reset({});",
cache_ptr_name,
- orig.as_glfield().cloneUninitialized(),
+ orig.glfield.cloneUninitialized(),
)
),
f"return {cache_ptr_name}.get();",
@@ -521,6 +554,20 @@ class Sweep(CustomGenerator):
f"Unable to map field {f} of type {f.field_type} to a waLBerla field."
)
+ def _make_field_info(self, f: Field, target: Target) -> FieldInfo:
+ match f.field_type:
+ case FieldType.GENERIC | FieldType.CUSTOM:
+ glfield = self._glfield_type.create(f)
+ data_id = BlockDataID().var(f"{f.name}Id")
+ return GlFieldInfo(f, glfield, data_id)
+ case FieldType.INDEXED:
+ assert isinstance(f.dtype, PsStructType)
+ idx_list = SparseIndexList.from_field(f, target=target)
+ view = idx_list.view_type().var(f"{f.name}_view")
+ return IndexListInfo(f, idx_list, view)
+ case _:
+ raise ValueError(f"Unexpected field type: {f.field_type} at field {f}")
+
def _render_invocation(
self, sfg: SfgComposer, target: Target, khandle: SfgKernelHandle
) -> tuple[SfgCallTreeNode, set[SfgVar]]:
@@ -560,18 +607,19 @@ class Sweep(CustomGenerator):
)
all_fields: dict[str, FieldInfo] = {
- f.name: FieldInfo(
- f, self._walberla_field(f), BlockDataID().var(f"{f.name}Id")
- )
- for f in khandle.fields
+ f.name: self._make_field_info(f, target) for f in khandle.fields
}
- swaps: dict[str, FieldInfo] = dict()
+ swaps: dict[str, GlFieldInfo] = dict()
shadows_cache = ShadowFieldCache()
for shadow, orig in self._shadow_fields.items():
- swaps[orig.name] = all_fields[shadow.name]
- shadows_cache.add_shadow(all_fields[orig.name])
+ shadow_fi = all_fields[shadow.name]
+ assert isinstance(shadow_fi, GlFieldInfo)
+ original_fi = all_fields[orig.name]
+ assert isinstance(original_fi, GlFieldInfo)
+ swaps[orig.name] = shadow_fi
+ shadows_cache.add_shadow(original_fi)
block = IBlockPtr().var("block")
block_fields = sorted(
@@ -589,7 +637,7 @@ class Sweep(CustomGenerator):
vector_groups = combine_vectors(parameters)
for fi in block_fields:
- props.add_property(fi.data_id, setter=False, getter=True)
+ props.add_property(fi.entity, setter=False, getter=True)
for s in sorted(parameters, key=lambda p: p.name):
props.add_property(s, setter=True, getter=True)
@@ -597,32 +645,25 @@ class Sweep(CustomGenerator):
def render_runmethod(ci: CellInterval | None = None):
return [
# Get IDs from class
- *(sfg.init(fi.data_id)(props.get(fi.data_id)) for fi in block_fields),
- # Extract regular fields from block
- *(
- sfg.init(fi.wlb_field_ptr)(
- block.getData(fi.wlb_field_ptr.field_type, fi.data_id)
- )
- for fi in block_fields
- ),
+ *(sfg.init(fi.entity)(props.get(fi.entity)) for fi in block_fields),
+ # Extract field views
+ *(sfg.init(fi.view)(fi.create_view(block)) for fi in block_fields),
# Get shadow fields from cache
*(
- sfg.init(shadow_info.wlb_field_ptr)(
- shadows_cache.get_shadow(orig_name)
- )
+ sfg.init(shadow_info.glfield)(shadows_cache.get_shadow(orig_name))
for orig_name, shadow_info in swaps.items()
),
# Map GhostLayerFields to pystencils fields
*(
- sfg.map_field(fi.field, fi.as_glfield().with_cell_interval(ci))
+ sfg.map_field(fi.field, fi.glfield.with_cell_interval(ci))
for fi in all_fields.values()
- if fi.is_glfield()
+ if isinstance(fi, GlFieldInfo)
),
- # Map index lists to pystencils fields
+ # Extract indexing information from field view for all remaining fields
*(
- sfg.map_field(fi.field, fi.as_index_list())
+ sfg.map_field(fi.field, fi.view_extraction())
for fi in all_fields.values()
- if fi.is_index_list()
+ if not isinstance(fi, GlFieldInfo)
),
# Get parameters from class
*(sfg.init(param)(props.get(param)) for param in parameters),
diff --git a/tests/BasicLbmScenarios/SimDomain.hpp b/tests/BasicLbmScenarios/SimDomain.hpp
index 228a89d..50991de 100644
--- a/tests/BasicLbmScenarios/SimDomain.hpp
+++ b/tests/BasicLbmScenarios/SimDomain.hpp
@@ -111,6 +111,11 @@ struct SimDomain
return sweep::BorderSweep< stencil::Direction::T, gen::bc_grid_aligned::NoSlipTop > { blocks, gen::bc_grid_aligned::NoSlipTop { gpuFields.pdfsId }, -1 };
}
+ auto
+ irregularFreeSlipFactory() {
+ return gen::bc_sparse::FreeSlipIrregularFactory(blocks, gpuFields.pdfsId);
+ }
+
void wait() { WALBERLA_GPU_CHECK(gpuDeviceSynchronize()); }
void syncGhostLayers()
@@ -162,6 +167,10 @@ struct SimDomain
return { cpuFields.pdfsId, cpuFields.rhoId, cpuFields.uId, force, omega };
}
+ auto irregularFreeSlipFactory() {
+ return gen::bc_sparse::FreeSlipIrregularFactory(blocks, cpuFields.pdfsId);
+ }
+
void syncGhostLayers() { commCpu(); }
void wait() { /* NOP */ }
diff --git a/tests/BasicLbmScenarios/TestBasicLbmScenarios.cpp b/tests/BasicLbmScenarios/TestBasicLbmScenarios.cpp
index 656df20..4be1a2d 100644
--- a/tests/BasicLbmScenarios/TestBasicLbmScenarios.cpp
+++ b/tests/BasicLbmScenarios/TestBasicLbmScenarios.cpp
@@ -50,6 +50,7 @@ void fullyPeriodic(Environment& env)
}
}
+#if defined(LBM_SCENARIOS_GPU_BUILD)
/**
* Periodic channel flow with a no-slip boundary at the top
* and a mirror boundary at the bottom.
@@ -134,6 +135,8 @@ void mirroredHalfChannel(Environment& env)
}
}
+#endif
+
/**
* Pipe flow with circular cross-section and free-slip walls.
* The pipe flow is initialized with a constant velocity in x-direction.
@@ -162,8 +165,7 @@ void freeSlipPipe(Environment& env)
const real_t maxVelocity{ 0.02 };
const Vector3< real_t > force{ 0., 0., 0. };
- for (auto& block : *dom.blocks)
- {
+ dom.forAllBlocks([&](IBlock & block) {
FlagField_T& flagField = *block.getData< FlagField_T >(dom.cpuFields.flagFieldId);
const uint8_t freeSlipFlag{ flagField.getOrRegisterFlag(freeSlipFlagUID) };
@@ -191,17 +193,18 @@ void freeSlipPipe(Environment& env)
velField.get(c, 1) = initVelocity[1];
velField.get(c, 2) = initVelocity[2];
}
- }
+ });
geometry::setNonBoundaryCellsToDomain< FlagField_T >(*dom.blocks, dom.cpuFields.flagFieldId, fluidFlagUid);
auto flagsOutput = field::createVTKOutput< FlagField_T >(dom.cpuFields.flagFieldId, *dom.blocks, "flags");
flagsOutput();
+ dom.fields2device();
dom.initFromFields(force);
gen::bc_sparse::FreeSlipIrregular freeSlip{
- gen::bc_sparse::FreeSlipIrregularFactory(dom.blocks, dom.cpuFields.pdfsId)
+ dom.irregularFreeSlipFactory()
.fromFlagField< FlagField_T >(dom.cpuFields.flagFieldId, freeSlipFlagUID, fluidFlagUid)
};
@@ -211,27 +214,29 @@ void freeSlipPipe(Environment& env)
for (uint_t i = 0; i < 10; ++i)
{
- for (auto& block : *dom.blocks)
- {
+ dom.forAllBlocks([&](IBlock & block){
streamCollide(&block);
+ });
+ dom.syncGhostLayers();
+ dom.fields2host();
+
+ dom.forAllBlocks([&](IBlock & block){
VectorField_T& velField = *block.getData< VectorField_T >(dom.cpuFields.uId);
FlagField_T& flagField = *block.getData< FlagField_T >(dom.cpuFields.flagFieldId);
uint8_t fluidFlag = flagField.getFlag(fluidFlagUid);
- for (Cell c : allCells)
- {
+ dom.forAllCells([&](Cell c){
if (flagField.isFlagSet(c, fluidFlag))
{
WALBERLA_CHECK_FLOAT_EQUAL(velField.get(c, 0), maxVelocity);
WALBERLA_CHECK_FLOAT_EQUAL(velField.get(c, 1), 0.);
WALBERLA_CHECK_FLOAT_EQUAL(velField.get(c, 2), 0.);
}
- }
- }
+ });
+ });
velOutput();
- dom.syncGhostLayers();
for (auto& block : *dom.blocks)
{
@@ -247,8 +252,6 @@ int main(int argc, char** argv)
{
walberla::Environment env{ argc, argv };
BasicLbmScenarios::fullyPeriodic(env);
- BasicLbmScenarios::mirroredHalfChannel(env);
-#if !defined(LBM_SCENARIOS_GPU_BUILD)
+ // BasicLbmScenarios::mirroredHalfChannel(env);
BasicLbmScenarios::freeSlipPipe(env);
-#endif
}
--
GitLab