[llvm-branch-commits] [OpenMP][MLIR] Extend record member map support for omp dialect to LLVM-IR (PR #82852)
via llvm-branch-commits
llvm-branch-commits at lists.llvm.org
Fri Feb 23 16:46:41 PST 2024
llvmbot wrote:
<!--LLVM PR SUMMARY COMMENT-->
@llvm/pr-subscribers-mlir-llvm
Author: None (agozillon)
<details>
<summary>Changes</summary>
This patch seeks to refactor slightly and extend the current record type map
support that was put in place for Fortran's descriptor types to handle explicit
member mapping for record types at a single level of depth (the case of explicit
mapping of nested record types is currently unsupported).
This patch seeks to support this by extending the OpenMPToLLVMIRTranslation phase
to more generally support record types, building on the prior groundwork in the
Fortran allocatables/pointers patch. It now supports different kinds of record type
mapping, in this case full record type mapping and then explicit member mapping
in which there is a special case for certain types when mapped individually to not
require any parent map link in the kernel argument structure. To facilitate this
required:
* The movement of the setting of the map flag type "ptr_and_obj" to respective
frontends, now supporting it as a possible flag that can be read and printed
in mlir form. Some minor changes to declare target map type setting was
neccesary for this.
* The addition of a member index array operand, which tracks the position
of the member in the parent, required for caclulating the appropriate size
to offload to the target, alongside the parents offload pointer (always the
first member currently being mapped).
* A partial mapping attribute operand, to indicate if the entire record type is
being mapped or just member components, aiding the ability to lower
record types in the different manners that are possible.
* Refactoring bounds calculation for record types and general arrays to one
location (as well as load/store generation prior to assigning to the kernel
argument structure), as a side affect enter/exit/update/data mapping
should now be more correct and fully support bounds mapping, previously
this would have only worked for target.
---
Patch is 52.96 KiB, truncated to 20.00 KiB below, full version: https://github.com/llvm/llvm-project/pull/82852.diff
4 Files Affected:
- (modified) mlir/lib/Target/LLVMIR/Dialect/OpenMP/OpenMPToLLVMIRTranslation.cpp (+345-206)
- (modified) mlir/test/Target/LLVMIR/omptarget-fortran-allocatable-types-host.mlir (+14-16)
- (modified) mlir/test/Target/LLVMIR/omptarget-llvm.mlir (+27-22)
- (added) mlir/test/Target/LLVMIR/omptarget-record-type-mapping-host.mlir (+63)
``````````diff
diff --git a/mlir/lib/Target/LLVMIR/Dialect/OpenMP/OpenMPToLLVMIRTranslation.cpp b/mlir/lib/Target/LLVMIR/Dialect/OpenMP/OpenMPToLLVMIRTranslation.cpp
index 6e53d801a0d2f0..f234fcee388e30 100644
--- a/mlir/lib/Target/LLVMIR/Dialect/OpenMP/OpenMPToLLVMIRTranslation.cpp
+++ b/mlir/lib/Target/LLVMIR/Dialect/OpenMP/OpenMPToLLVMIRTranslation.cpp
@@ -33,6 +33,7 @@
#include "llvm/Transforms/Utils/ModuleUtils.h"
#include <any>
+#include <numeric>
#include <optional>
#include <utility>
@@ -1750,9 +1751,9 @@ void collectMapDataFromMapOperands(MapInfoData &mapData,
mapData.BaseType.push_back(
moduleTranslation.convertType(mapOp.getVarType()));
- mapData.Sizes.push_back(getSizeInBytes(
- dl, mapOp.getVarType(), mapOp, mapData.BasePointers.back(),
- mapData.BaseType.back(), builder, moduleTranslation));
+ mapData.Sizes.push_back(
+ getSizeInBytes(dl, mapOp.getVarType(), mapOp, mapData.Pointers.back(),
+ mapData.BaseType.back(), builder, moduleTranslation));
mapData.MapClause.push_back(mapOp.getOperation());
mapData.Types.push_back(
llvm::omp::OpenMPOffloadMappingFlags(mapOp.getMapType().value()));
@@ -1783,6 +1784,134 @@ void collectMapDataFromMapOperands(MapInfoData &mapData,
}
}
+static int getMapDataMemberIdx(MapInfoData &mapData,
+ mlir::omp::MapInfoOp memberOp) {
+ auto *res = llvm::find(mapData.MapClause, memberOp);
+ assert(res != mapData.MapClause.end());
+ return std::distance(mapData.MapClause.begin(), res);
+}
+
+static mlir::omp::MapInfoOp
+getFirstOrLastMappedMemberPtr(mlir::omp::MapInfoOp mapInfo, bool first) {
+ // Only 1 member has been mapped, we can return it.
+ if (mapInfo.getMembersIndex()->size() == 1)
+ if (auto mapOp = mlir::dyn_cast<mlir::omp::MapInfoOp>(
+ mapInfo.getMembers()[0].getDefiningOp()))
+ return mapOp;
+
+ std::vector<size_t> indices(mapInfo.getMembersIndexAttr().size());
+ std::iota(indices.begin(), indices.end(), 0);
+ llvm::sort(indices.begin(), indices.end(),
+ [&](const size_t a, const size_t b) {
+ return mapInfo.getMembersIndexAttr()[a]
+ .cast<mlir::IntegerAttr>()
+ .getInt() < mapInfo.getMembersIndexAttr()[b]
+ .cast<mlir::IntegerAttr>()
+ .getInt();
+ });
+
+ if (auto mapOp = mlir::dyn_cast<mlir::omp::MapInfoOp>(
+ mapInfo.getMembers()[((first) ? indices.front() : indices.back())]
+ .getDefiningOp()))
+ return mapOp;
+
+ assert(false && "getFirstOrLastMappedMemberPtr could not find approproaite "
+ "map information");
+}
+
+/// This function calculates the array/pointer offset for map data provided
+/// with bounds operations, e.g. when provided something like the following:
+///
+/// Fortran
+/// map(tofrom: array(2:5, 3:2))
+/// or
+/// C++
+/// map(tofrom: array[1:4][2:3])
+/// We must calculate the initial pointer offset to pass across, this function
+/// performs this using bounds.
+///
+/// NOTE: which while specified in row-major order it currently needs to be
+/// flipped for Fortran's column order array allocation and access (as
+/// opposed to C++'s row-major, hence the backwards processing where order is
+/// important). This is likely important to keep in mind for the future when
+/// we incorporate a C++ frontend, both frontends will need to agree on the
+/// ordering of generated bounds operations (one may have to flip them) to
+/// make the below lowering frontend agnostic. The offload size
+/// calcualtion may also have to be adjusted for C++.
+std::vector<llvm::Value *>
+calculateBoundsOffset(LLVM::ModuleTranslation &moduleTranslation,
+ llvm::IRBuilderBase &builder, bool isArrayTy,
+ mlir::OperandRange bounds) {
+ std::vector<llvm::Value *> idx;
+ // There's no bounds to calculate an offset from, we can safely
+ // ignore and return no indices.
+ if (bounds.empty())
+ return idx;
+
+ // If we have an array type, then we have its type so can treat it as a
+ // normal GEP instruction where the bounds operations are simply indexes
+ // into the array. We currently do reverse order of the bounds, which
+ // I believe leans more towards Fortran's column-major in memory.
+ if (isArrayTy) {
+ idx.push_back(builder.getInt64(0));
+ for (int i = bounds.size() - 1; i >= 0; --i) {
+ if (auto boundOp = mlir::dyn_cast_if_present<mlir::omp::DataBoundsOp>(
+ bounds[i].getDefiningOp())) {
+ idx.push_back(moduleTranslation.lookupValue(boundOp.getLowerBound()));
+ }
+ }
+ } else {
+ // If we do not have an array type, but we have bounds, then we're dealing
+ // with a pointer that's being treated like an array and we have the
+ // underlying type e.g. an i32, or f64 etc, e.g. a fortran descriptor base
+ // address (pointer pointing to the actual data) so we must caclulate the
+ // offset using a single index which the following two loops attempts to
+ // compute.
+
+ // Calculates the size offset we need to make per row e.g. first row or
+ // column only needs to be offset by one, but the next would have to be
+ // the previous row/column offset multiplied by the extent of current row.
+ //
+ // For example ([1][10][100]):
+ //
+ // - First row/column we move by 1 for each index increment
+ // - Second row/column we move by 1 (first row/column) * 10 (extent/size of
+ // current) for 10 for each index increment
+ // - Third row/column we would move by 10 (second row/column) *
+ // (extent/size of current) 100 for 1000 for each index increment
+ std::vector<llvm::Value *> dimensionIndexSizeOffset{builder.getInt64(1)};
+ for (size_t i = 1; i < bounds.size(); ++i) {
+ if (auto boundOp = mlir::dyn_cast_if_present<mlir::omp::DataBoundsOp>(
+ bounds[i].getDefiningOp())) {
+ dimensionIndexSizeOffset.push_back(builder.CreateMul(
+ moduleTranslation.lookupValue(boundOp.getExtent()),
+ dimensionIndexSizeOffset[i - 1]));
+ }
+ }
+
+ // Now that we have calculated how much we move by per index, we must
+ // multiply each lower bound offset in indexes by the size offset we
+ // have calculated in the previous and accumulate the results to get
+ // our final resulting offset.
+ for (int i = bounds.size() - 1; i >= 0; --i) {
+ if (auto boundOp = mlir::dyn_cast_if_present<mlir::omp::DataBoundsOp>(
+ bounds[i].getDefiningOp())) {
+ if (idx.empty())
+ idx.emplace_back(builder.CreateMul(
+ moduleTranslation.lookupValue(boundOp.getLowerBound()),
+ dimensionIndexSizeOffset[i]));
+ else
+ idx.back() = builder.CreateAdd(
+ idx.back(), builder.CreateMul(moduleTranslation.lookupValue(
+ boundOp.getLowerBound()),
+ dimensionIndexSizeOffset[i]));
+ }
+ }
+ }
+
+ return idx;
+}
+
// This creates two insertions into the MapInfosTy data structure for the
// "parent" of a set of members, (usually a container e.g.
// class/structure/derived type) when subsequent members have also been
@@ -1795,6 +1924,9 @@ void collectMapDataFromMapOperands(MapInfoData &mapData,
// which is utilised in subsequent member mappings (by modifying there map type
// with it) to indicate that a member is part of this parent and should be
// treated by the runtime as such. Important to achieve the correct mapping.
+//
+// This function borrows a lot from it's Clang parallel function
+// emitCombinedEntry inside of CGOpenMPRuntime.cpp
static llvm::omp::OpenMPOffloadMappingFlags mapParentWithMembers(
LLVM::ModuleTranslation &moduleTranslation, llvm::IRBuilderBase &builder,
llvm::OpenMPIRBuilder &ompBuilder, DataLayout &dl,
@@ -1810,7 +1942,6 @@ static llvm::omp::OpenMPOffloadMappingFlags mapParentWithMembers(
combinedInfo.Names.emplace_back(LLVM::createMappingInformation(
mapData.MapClause[mapDataIndex]->getLoc(), ompBuilder));
combinedInfo.BasePointers.emplace_back(mapData.BasePointers[mapDataIndex]);
- combinedInfo.Pointers.emplace_back(mapData.Pointers[mapDataIndex]);
// Calculate size of the parent object being mapped based on the
// addresses at runtime, highAddr - lowAddr = size. This of course
@@ -1819,42 +1950,68 @@ static llvm::omp::OpenMPOffloadMappingFlags mapParentWithMembers(
// Fortran pointers and allocatables, the mapping of the pointed to
// data by the descriptor (which itself, is a structure containing
// runtime information on the dynamically allocated data).
- llvm::Value *lowAddr = builder.CreatePointerCast(
- mapData.Pointers[mapDataIndex], builder.getPtrTy());
- llvm::Value *highAddr = builder.CreatePointerCast(
- builder.CreateConstGEP1_32(mapData.BaseType[mapDataIndex],
- mapData.Pointers[mapDataIndex], 1),
- builder.getPtrTy());
+ auto parentClause =
+ mlir::dyn_cast<mlir::omp::MapInfoOp>(mapData.MapClause[mapDataIndex]);
+
+ llvm::Value *lowAddr, *highAddr;
+ if (!parentClause.getPartialMap()) {
+ lowAddr = builder.CreatePointerCast(mapData.Pointers[mapDataIndex],
+ builder.getPtrTy());
+ highAddr = builder.CreatePointerCast(
+ builder.CreateConstGEP1_32(mapData.BaseType[mapDataIndex],
+ mapData.Pointers[mapDataIndex], 1),
+ builder.getPtrTy());
+ combinedInfo.Pointers.emplace_back(mapData.Pointers[mapDataIndex]);
+ } else {
+ auto mapOp =
+ mlir::dyn_cast<mlir::omp::MapInfoOp>(mapData.MapClause[mapDataIndex]);
+ int firstMemberIdx = getMapDataMemberIdx(
+ mapData, getFirstOrLastMappedMemberPtr(mapOp, true));
+ lowAddr = builder.CreatePointerCast(mapData.Pointers[firstMemberIdx],
+ builder.getPtrTy());
+ int lastMemberIdx = getMapDataMemberIdx(
+ mapData, getFirstOrLastMappedMemberPtr(mapOp, false));
+ highAddr = builder.CreatePointerCast(
+ builder.CreateGEP(mapData.BaseType[lastMemberIdx],
+ mapData.Pointers[lastMemberIdx], builder.getInt64(1)),
+ builder.getPtrTy());
+ combinedInfo.Pointers.emplace_back(mapData.Pointers[firstMemberIdx]);
+ }
+
llvm::Value *size = builder.CreateIntCast(
builder.CreatePtrDiff(builder.getInt8Ty(), highAddr, lowAddr),
builder.getInt64Ty(),
/*isSigned=*/false);
combinedInfo.Sizes.push_back(size);
- // This creates the initial MEMBER_OF mapping that consists of
- // the parent/top level container (same as above effectively, except
- // with a fixed initial compile time size and seperate maptype which
- // indicates the true mape type (tofrom etc.) and that it is a part
- // of a larger mapping and indicating the link between it and it's
- // members that are also explicitly mapped).
+ // TODO: This will need expanded to include the whole host of logic for the
+ // map flags that Clang currently supports (e.g. it hsould take the map flag
+ // of the parent map flag, remove the OMP_MAP_TARGET_PARAM and do some further
+ // case specific flag modifications), for the moment it handles what we
+ // support as expected.
llvm::omp::OpenMPOffloadMappingFlags mapFlag =
llvm::omp::OpenMPOffloadMappingFlags::OMP_MAP_TO;
- if (isTargetParams)
- mapFlag &= ~llvm::omp::OpenMPOffloadMappingFlags::OMP_MAP_TARGET_PARAM;
llvm::omp::OpenMPOffloadMappingFlags memberOfFlag =
ompBuilder.getMemberOfFlag(combinedInfo.BasePointers.size() - 1);
ompBuilder.setCorrectMemberOfFlag(mapFlag, memberOfFlag);
- combinedInfo.Types.emplace_back(mapFlag);
- combinedInfo.DevicePointers.emplace_back(
- llvm::OpenMPIRBuilder::DeviceInfoTy::None);
- combinedInfo.Names.emplace_back(LLVM::createMappingInformation(
- mapData.MapClause[mapDataIndex]->getLoc(), ompBuilder));
- combinedInfo.BasePointers.emplace_back(mapData.BasePointers[mapDataIndex]);
- combinedInfo.Pointers.emplace_back(mapData.Pointers[mapDataIndex]);
- combinedInfo.Sizes.emplace_back(mapData.Sizes[mapDataIndex]);
-
+ // This creates the initial MEMBER_OF mapping that consists of
+ // the parent/top level container (same as above effectively, except
+ // with a fixed initial compile time size and seperate maptype which
+ // indicates the true mape type (tofrom etc.). This parent mapping is
+ // only relevant if the structure in it's totality is being mapped,
+ // otherwise the above suffices.
+ if (!parentClause.getPartialMap()) {
+ combinedInfo.Types.emplace_back(mapFlag);
+ combinedInfo.DevicePointers.emplace_back(
+ llvm::OpenMPIRBuilder::DeviceInfoTy::None);
+ combinedInfo.Names.emplace_back(LLVM::createMappingInformation(
+ mapData.MapClause[mapDataIndex]->getLoc(), ompBuilder));
+ combinedInfo.BasePointers.emplace_back(mapData.BasePointers[mapDataIndex]);
+ combinedInfo.Pointers.emplace_back(mapData.Pointers[mapDataIndex]);
+ combinedInfo.Sizes.emplace_back(mapData.Sizes[mapDataIndex]);
+ }
return memberOfFlag;
}
@@ -1871,86 +2028,99 @@ static void processMapMembersWithParent(
for (auto mappedMembers : parentClause.getMembers()) {
auto memberClause =
mlir::dyn_cast<mlir::omp::MapInfoOp>(mappedMembers.getDefiningOp());
- int memberDataIdx = -1;
- for (size_t i = 0; i < mapData.MapClause.size(); ++i) {
- if (mapData.MapClause[i] == memberClause)
- memberDataIdx = i;
- }
+ int memberDataIdx = getMapDataMemberIdx(mapData, memberClause);
assert(memberDataIdx >= 0 && "could not find mapped member of structure");
// Same MemberOfFlag to indicate its link with parent and other members
- // of, and we flag that it's part of a pointer and object coupling.
+ // of
auto mapFlag =
llvm::omp::OpenMPOffloadMappingFlags(memberClause.getMapType().value());
mapFlag &= ~llvm::omp::OpenMPOffloadMappingFlags::OMP_MAP_TARGET_PARAM;
+ mapFlag |= llvm::omp::OpenMPOffloadMappingFlags::OMP_MAP_MEMBER_OF;
ompBuilder.setCorrectMemberOfFlag(mapFlag, memberOfFlag);
- mapFlag |= llvm::omp::OpenMPOffloadMappingFlags::OMP_MAP_PTR_AND_OBJ;
+
combinedInfo.Types.emplace_back(mapFlag);
combinedInfo.DevicePointers.emplace_back(
llvm::OpenMPIRBuilder::DeviceInfoTy::None);
combinedInfo.Names.emplace_back(
LLVM::createMappingInformation(memberClause.getLoc(), ompBuilder));
-
- combinedInfo.BasePointers.emplace_back(mapData.BasePointers[memberDataIdx]);
-
- std::vector<llvm::Value *> idx{builder.getInt64(0)};
- llvm::Value *offsetAddress = nullptr;
- if (!memberClause.getBounds().empty()) {
- if (mapData.BaseType[memberDataIdx]->isArrayTy()) {
- for (int i = memberClause.getBounds().size() - 1; i >= 0; --i) {
- if (auto boundOp = mlir::dyn_cast_if_present<mlir::omp::DataBoundsOp>(
- memberClause.getBounds()[i].getDefiningOp())) {
- idx.push_back(
- moduleTranslation.lookupValue(boundOp.getLowerBound()));
- }
- }
- } else {
- std::vector<llvm::Value *> dimensionIndexSizeOffset{
- builder.getInt64(1)};
- for (size_t i = 1; i < memberClause.getBounds().size(); ++i) {
- if (auto boundOp = mlir::dyn_cast_if_present<mlir::omp::DataBoundsOp>(
- memberClause.getBounds()[i].getDefiningOp())) {
- dimensionIndexSizeOffset.push_back(builder.CreateMul(
- moduleTranslation.lookupValue(boundOp.getExtent()),
- dimensionIndexSizeOffset[i - 1]));
- }
- }
-
- for (int i = memberClause.getBounds().size() - 1; i >= 0; --i) {
- if (auto boundOp = mlir::dyn_cast_if_present<mlir::omp::DataBoundsOp>(
- memberClause.getBounds()[i].getDefiningOp())) {
- if (!offsetAddress)
- offsetAddress = builder.CreateMul(
- moduleTranslation.lookupValue(boundOp.getLowerBound()),
- dimensionIndexSizeOffset[i]);
- else
- offsetAddress = builder.CreateAdd(
- offsetAddress,
- builder.CreateMul(
- moduleTranslation.lookupValue(boundOp.getLowerBound()),
- dimensionIndexSizeOffset[i]));
- }
- }
- }
- }
-
- llvm::Value *memberIdx =
- builder.CreateLoad(builder.getPtrTy(), mapData.Pointers[memberDataIdx]);
- memberIdx = builder.CreateInBoundsGEP(
- mapData.BaseType[memberDataIdx], memberIdx,
- offsetAddress ? std::vector<llvm::Value *>{offsetAddress} : idx,
- "member_idx");
- combinedInfo.Pointers.emplace_back(memberIdx);
+ combinedInfo.BasePointers.emplace_back(mapData.BasePointers[mapDataIndex]);
+ combinedInfo.Pointers.emplace_back(mapData.Pointers[memberDataIdx]);
combinedInfo.Sizes.emplace_back(mapData.Sizes[memberDataIdx]);
}
}
+// This may be a bit of a naive check, the intent is to verify if the
+// mapped data being passed is a pointer -> pointee that requires special
+// handling in certain cases. There may be a better way to verify this, but
+// unfortunately with opaque pointers we lose the ability to easily check if
+// something is a pointer whilst maintaining access to the underlying type.
+static bool checkIfPointerMap(llvm::omp::OpenMPOffloadMappingFlags mapFlag) {
+ return static_cast<
+ std::underlying_type_t<llvm::omp::OpenMPOffloadMappingFlags>>(
+ mapFlag &
+ llvm::omp::OpenMPOffloadMappingFlags::OMP_MAP_PTR_AND_OBJ) != 0;
+}
+
+static void
+processIndividualMap(MapInfoData &mapData, size_t mapDataIdx,
+ llvm::OpenMPIRBuilder::MapInfosTy &combinedInfo,
+ bool isTargetParams, int mapDataParentIdx = -1) {
+ // Declare Target Mappings are excluded from being marked as
+ // OMP_MAP_TARGET_PARAM as they are not passed as parameters, they're
+ // marked with OMP_MAP_PTR_AND_OBJ instead.
+ auto mapFlag = mapData.Types[mapDataIdx];
+ if (isTargetParams && !mapData.IsDeclareTarget[mapDataIdx])
+ mapFlag |= llvm::omp::OpenMPOffloadMappingFlags::OMP_MAP_TARGET_PARAM;
+
+ if (auto mapInfoOp =
+ dyn_cast<mlir::omp::MapInfoOp>(mapData.MapClause[mapDataIdx]))
+ if (mapInfoOp.getMapCaptureType().value() ==
+ mlir::omp::VariableCaptureKind::ByCopy &&
+ !checkIfPointerMap(mapFlag))
+ mapFlag |= llvm::omp::OpenMPOffloadMappingFlags::OMP_MAP_LITERAL;
+
+ // if we're provided a mapDataParentIdx, then the data being mapped is
+ // part of a larger object (in a parent <-> member mapping) and in this
+ // case our BasePointer should be the parent.
+ if (mapDataParentIdx >= 0)
+ combinedInfo.BasePointers.emplace_back(
+ mapData.BasePointers[mapDataParentIdx]);
+ else
+ combinedInfo.BasePointers.emplace_back(mapData.BasePointers[mapDataIdx]);
+
+ combinedInfo.Pointers.emplace_back(mapData.Pointers[mapDataIdx]);
+ combinedInfo.DevicePointers.emplace_back(mapData.DevicePointers[mapDataIdx]);
+ combinedInfo.Names.emplace_back(mapData.Names[mapDataIdx]);
+ combinedInfo.Types.emplace_back(mapFlag);
+ combinedInfo.Sizes.emplace_back(mapData.Sizes[mapDataIdx]);
+}
+
static void processMapWithMembersOf(
LLVM::ModuleTranslation &moduleTranslation, llvm::IRBuilderBase &builder,
llvm::OpenMPIRBuilder &ompBuilder, DataLayout &dl,
llvm::OpenMPIRBuilder::MapInfosTy &combinedInfo, MapInfoData &mapData,
uint64_t mapDataIndex, bool isTargetParams) {
+ auto parentClause =
+ mlir::dyn_cast<mlir::omp::MapInfoOp>(mapData.MapClause[mapDataIndex]);
+ // If we have a partial map (no parent referneced in the map clauses of the
+ // directive, only members) and only a single member, we do not need to bind
+ // the map of the member to the parent, we can pass the member seperately.
+ if (parentClause.getMembers().size() == 1 && parentClause.getPartialMap()) {
+ auto memberClause = mlir::dyn_cast<mlir::omp::MapInfoOp>(
+ parentClause.getMembers()[0].getDefiningOp());
+ int memberDataIdx = getMapDataMemberIdx(mapData, memberClause);...
[truncated]
``````````
</details>
https://github.com/llvm/llvm-project/pull/82852
More information about the llvm-branch-commits
mailing list