@@ -973,86 +973,86 @@ struct ConvertXeTileToXeGPUPass // convert XeTile to XeGPU
973973 memSpace);
974974 });
975975
976- typeConverter.addConversion ([&](xetile::TileType type)
977- -> xegpu::TensorDescType {
978- auto context = type.getContext ();
979- auto scatterAttr = type.getScatterAttr ();
980- bool isScattered = scatterAttr ? scatterAttr.getValue () : false ;
981-
982- // by default the targetTy is the element type, except for SLM cases,
983- // where the data will be treated as 32-bit type implicitly.
984- Type targetTy = type.getElementType ();
985-
986- xegpu::SGMapAttr sgMap = nullptr ;
987- if (auto attr = type.getSgMap ()) {
988- auto layout =
989- llvm::to_vector_of< uint32_t >( attr.getWiLayout ().asArrayRef () );
990- auto data = llvm::to_vector_of< uint32_t >(attr. getWiData (). asArrayRef () );
991- sgMap = xegpu::SGMapAttr::get (context, layout, data);
992- }
993-
994- auto memSpaceAttr = convertMemorySpace (type. getMemorySpace ());
995- auto memSpace =
996- memSpaceAttr ? memSpaceAttr. getValue () : xegpu::MemorySpace::Global;
997-
998- Attribute encoding ;
999- llvm::SmallVector< int64_t > shape;
1000- if (isScattered) {
1001- // Scattered tile is lowered to scattered tensor_desc with chunk
1002- // size 1. It supports both global memory and shared memory. while
1003- // scattered tile can support 2D shape, scattered tensor_desc only
1004- // support 1D shape.
1005- auto chunkSizeAttr = IntegerAttr::get (IntegerType::get (context, 64 ), 1 );
1006- auto msA = memSpaceAttr
1007- ? memSpaceAttr
1008- : xegpu::MemorySpaceAttr::get (context, memSpace);
1009-
1010- encoding =
1011- xegpu::ScatterTensorDescAttr::get (context, msA, chunkSizeAttr);
1012- shape.push_back (type.getNumElements ());
1013- } else if (memSpace == xegpu::MemorySpace::Global) {
1014- // Blocked tile on global memory is lowered to blocked tensor_desc
1015- // with the same shape.
1016- auto arrayLenAttr = type.getArrayLength ();
1017- auto boundaryCheckAttr = BoolAttr::get (context, true );
1018- encoding = xegpu::BlockTensorDescAttr::get (
1019- context, memSpaceAttr, arrayLenAttr, boundaryCheckAttr);
1020- shape = llvm::to_vector (type.getShape ());
1021- } else {
1022- // for TileType created for SLM access, it will be converted into:
1023- // 1. a 1D block tensor_desc if it is for row-major access
1024- // 2. a scattered tensor_desc if it is for col-major access.
1025- auto elemBits = type.getElementType ().getIntOrFloatBitWidth ();
1026- auto vnniFactor = std::max<int >(32 / elemBits, 1 );
1027-
1028- // SLM access only supports 32-bit or 64-bit data type, so convert
1029- // the type if original element type is less than 32-bit.
1030- if (elemBits < 32 ) {
1031- targetTy = type.getElementType ().isInteger ()
1032- ? (Type)IntegerType::get (context, 32 )
1033- : (Type)Float32Type::get (context);
1034- }
1035-
1036- if (isColMajorOrder (type.getOrder ())) {
1037- // For access with col-major order
1038- auto chunkSize = type.getShape ()[0 ] / vnniFactor;
1039- auto chunkSizeAttr =
1040- IntegerAttr::get (IntegerType::get (context, 64 ), chunkSize);
1041- encoding = xegpu::ScatterTensorDescAttr::get (context, memSpaceAttr,
1042- chunkSizeAttr);
1043- shape = {type.getShape ()[1 ], chunkSize};
1044- } else {
1045- // For access with row-major order
1046- auto vecSize = type.getNumElements () / vnniFactor;
1047- encoding = xegpu::BlockTensorDescAttr::get (
1048- context, memSpaceAttr, nullptr /* array_len*/
1049- nullptr /* boundary_check*/
1050- shape.push_back (vecSize);
1051- }
1052- }
1053- return xegpu::TensorDescType::get (context, shape, targetTy, encoding,
1054- sgMap);
1055- });
976+ typeConverter.addConversion (
977+ [&](xetile::TileType type) -> xegpu::TensorDescType {
978+ auto context = type.getContext ();
979+ auto scatterAttr = type.getScatterAttr ();
980+ bool isScattered = scatterAttr ? scatterAttr.getValue () : false ;
981+
982+ // by default the targetTy is the element type, except for SLM cases,
983+ // where the data will be treated as 32-bit type implicitly.
984+ Type targetTy = type.getElementType ();
985+
986+ xegpu::LayoutAttr sgMap = nullptr ;
987+ if (auto attr = type.getSgMap ()) {
988+ auto layout = attr. getWiLayout (). asArrayRef ();
989+ auto data = attr.getWiData ().asArrayRef ();
990+ sgMap = xegpu::LayoutAttr::get (context, layout, data );
991+ }
992+
993+ auto memSpaceAttr = convertMemorySpace (type. getMemorySpace ());
994+ auto memSpace = memSpaceAttr ? memSpaceAttr. getValue ()
995+ : xegpu::MemorySpace::Global;
996+
997+ Attribute encoding;
998+ llvm::SmallVector< int64_t > shape ;
999+ if (isScattered) {
1000+ // Scattered tile is lowered to scattered tensor_desc with chunk
1001+ // size 1. It supports both global memory and shared memory. while
1002+ // scattered tile can support 2D shape, scattered tensor_desc only
1003+ // support 1D shape.
1004+ auto chunkSizeAttr =
1005+ IntegerAttr::get (IntegerType::get (context, 64 ), 1 );
1006+ auto msA = memSpaceAttr
1007+ ? memSpaceAttr
1008+ : xegpu::MemorySpaceAttr::get (context, memSpace);
1009+
1010+ encoding =
1011+ xegpu::ScatterTensorDescAttr::get (context, msA, chunkSizeAttr);
1012+ shape.push_back (type.getNumElements ());
1013+ } else if (memSpace == xegpu::MemorySpace::Global) {
1014+ // Blocked tile on global memory is lowered to blocked tensor_desc
1015+ // with the same shape.
1016+ auto arrayLenAttr = type.getArrayLength ();
1017+ auto boundaryCheckAttr = BoolAttr::get (context, true );
1018+ encoding = xegpu::BlockTensorDescAttr::get (
1019+ context, memSpaceAttr, arrayLenAttr, boundaryCheckAttr);
1020+ shape = llvm::to_vector (type.getShape ());
1021+ } else {
1022+ // for TileType created for SLM access, it will be converted into:
1023+ // 1. a 1D block tensor_desc if it is for row-major access
1024+ // 2. a scattered tensor_desc if it is for col-major access.
1025+ auto elemBits = type.getElementType ().getIntOrFloatBitWidth ();
1026+ auto vnniFactor = std::max<int >(32 / elemBits, 1 );
1027+
1028+ // SLM access only supports 32-bit or 64-bit data type, so convert
1029+ // the type if original element type is less than 32-bit.
1030+ if (elemBits < 32 ) {
1031+ targetTy = type.getElementType ().isInteger ()
1032+ ? (Type)IntegerType::get (context, 32 )
1033+ : (Type)Float32Type::get (context);
1034+ }
1035+
1036+ if (isColMajorOrder (type.getOrder ())) {
1037+ // For access with col-major order
1038+ auto chunkSize = type.getShape ()[0 ] / vnniFactor;
1039+ auto chunkSizeAttr =
1040+ IntegerAttr::get (IntegerType::get (context, 64 ), chunkSize);
1041+ encoding = xegpu::ScatterTensorDescAttr::get (
1042+ context, memSpaceAttr, chunkSizeAttr);
1043+ shape = {type.getShape ()[1 ], chunkSize};
1044+ } else {
1045+ // For access with row-major order
1046+ auto vecSize = type.getNumElements () / vnniFactor;
1047+ encoding = xegpu::BlockTensorDescAttr::get (
1048+ context, memSpaceAttr, nullptr /* array_len*/
1049+ nullptr /* boundary_check*/
1050+ shape.push_back (vecSize);
1051+ }
1052+ }
1053+ return xegpu::TensorDescType::get (context, shape, targetTy, encoding,
1054+ sgMap);
1055+ });
10561056
10571057 auto materializeWithCast = [&](OpBuilder &builder, Type type,
10581058 ValueRange inputs, Location loc) -> Value {
0 commit comments