[i1] Implement packed_storage layout encoding attribute

* make `packed_storage` as a type of `iree_encoding` attribute, and make type converters accept it. * `i1` tensors with `#iree_encoding.packed_storage` will be interpreted as packed i1 type, same as specifying `--iree-experimental-packed-i1-storage`. * `--iree-experimental-packed-i1-storage` are kept for testing purposes. We can drop this option after frontend enables emitting `i1` tensors with attributes. Signed-off-by: Alan Li <[email protected]>
iree-org · Dec 13, 2024 · 349a6bd · 349a6bd
1 parent c618134
commit 349a6bd
Show file tree

Hide file tree

Showing 10 changed files with 93 additions and 18 deletions.
diff --git a/compiler/src/iree/compiler/Codegen/Common/EncodingUtils.cpp b/compiler/src/iree/compiler/Codegen/Common/EncodingUtils.cpp
@@ -6,6 +6,7 @@
 
 #include "iree/compiler/Codegen/Common/EncodingUtils.h"
 #include "iree/compiler/Codegen/Dialect/Codegen/Utils/Utils.h"
+#include "iree/compiler/Dialect/Encoding/IR/EncodingTypes.h"
 #include "mlir/Dialect/Linalg/IR/LinalgInterfaces.h"
 #include "mlir/Dialect/Tensor/IR/Tensor.h"
 #include "mlir/Dialect/Utils/IndexingUtils.h"
@@ -133,7 +134,10 @@ MaterializeEncodingConversionTarget::MaterializeEncodingConversionTarget(
   markUnknownOpDynamicallyLegal([](Operation *op) {
     auto typeHasEncoding = [](Type t) -> bool {
       auto tensorType = dyn_cast<RankedTensorType>(t);
-      return tensorType && tensorType.getEncoding();
+      if (!(tensorType && tensorType.getEncoding()))
+        return false;
+      // Allow iree_encoding::packed_storage to pass through.
+      return !IREE::Encoding::hasPackedStorageAttr(tensorType);
     };
     auto valueHasEncoding = [=](Value v) -> bool {
       return typeHasEncoding(v.getType());

diff --git a/compiler/src/iree/compiler/Dialect/Encoding/IR/EncodingAttrs.cpp b/compiler/src/iree/compiler/Dialect/Encoding/IR/EncodingAttrs.cpp
@@ -243,6 +243,10 @@ EncodingAttr getEncodingAttr(RankedTensorType type) {
   return dyn_cast_or_null<EncodingAttr>(type.getEncoding());
 }
 
+bool hasPackedStorageAttr(RankedTensorType type) {
+  return dyn_cast_or_null<PackedStorageAttr>(type.getEncoding()) != nullptr;
+}
+
 FailureOr<linalg::ContractionDimensions>
 getEncodingContractionDims(EncodingAttr encoding) {
   auto indexingMapsAttr = encoding.getUserIndexingMaps();

diff --git a/compiler/src/iree/compiler/Dialect/Encoding/IR/EncodingAttrs.td b/compiler/src/iree/compiler/Dialect/Encoding/IR/EncodingAttrs.td
@@ -41,6 +41,17 @@ def EncodingOpType : IREEEncoding_I32EnumAttr<"EncodingOpType",
 def EncodingOpTypeAttr:
   IREEEncoding_EnumAttr<EncodingOpType, "optype">;
 
+
+def PackedStorageAttr : IREEEncoding_Attr<"PackedStorage"> {
+  let mnemonic = "packed_storage";
+  let summary = [{Indicates packed storage datatype.}];
+  let description = [{
+    This attribute indicates this is a back-to-back packed storage in memory.
+    This attribute takes no arguments.
+  }];
+  let genVerifyDecl = 0;
+}
+
 def EncodingAttr :
     IREEEncoding_Attr<"Encoding", [
       DeclareAttrInterfaceMethods<IREEEncoding_EncodingLayoutAttrInterface, [

diff --git a/compiler/src/iree/compiler/Dialect/Encoding/IR/EncodingTypes.h b/compiler/src/iree/compiler/Dialect/Encoding/IR/EncodingTypes.h
@@ -38,6 +38,9 @@ namespace mlir::iree_compiler::IREE::Encoding {
 /// Otherwise, returns null.
 EncodingAttr getEncodingAttr(RankedTensorType type);
 
+/// Returns true if the type contains packed_storage attribute.
+bool hasPackedStorageAttr(RankedTensorType type);
+
 /// Returns the ContractionDimensions for the encoding user_indexing_maps.
 FailureOr<linalg::ContractionDimensions>
 getEncodingContractionDims(EncodingAttr encoding);

diff --git a/compiler/src/iree/compiler/Dialect/Stream/Conversion/HALToStream/Patterns.cpp b/compiler/src/iree/compiler/Dialect/Stream/Conversion/HALToStream/Patterns.cpp
@@ -12,6 +12,10 @@
 #include "iree/compiler/Dialect/Stream/IR/StreamOps.h"
 #include "mlir/Dialect/Arith/IR/Arith.h"
 
+namespace mlir::iree_compiler::IREE::Encoding {
+bool hasPackedStorageAttr(mlir::RankedTensorType);
+} // namespace mlir::iree_compiler::IREE::Encoding
+
 namespace mlir::iree_compiler {
 
 namespace {
@@ -90,6 +94,11 @@ struct ConvertTensorImportOp
                                                RankedTensorType tensorType,
                                                ValueRange dynamicDims,
                                                OpBuilder &builder) {
+    // If the encoding attr is about packed storage then we don't need all this
+    if (IREE::Encoding::hasPackedStorageAttr(tensorType)) {
+      return success();
+    }
+
     auto expectedElementType = builder.create<IREE::HAL::ElementTypeOp>(
         loc, tensorType.getElementType());
     auto expectedEncodingType = builder.create<IREE::HAL::EncodingTypeOp>(

diff --git a/compiler/src/iree/compiler/Dialect/Stream/Transforms/EncodeTensors.cpp b/compiler/src/iree/compiler/Dialect/Stream/Transforms/EncodeTensors.cpp
@@ -46,7 +46,8 @@ static LogicalResult checkEncoding(Operation *op, RankedTensorType encodingType,
                                    ValueRange encodingDims,
                                    PatternRewriter &rewriter) {
   auto encoding = encodingType.getEncoding();
-  if (encoding && !llvm::isa<IREE::Encoding::EncodingAttr>(encoding)) {
+  if (encoding && !llvm::isa<IREE::Encoding::EncodingAttr,
+                             IREE::Encoding::PackedStorageAttr>(encoding)) {
     return rewriter.notifyMatchFailure(op, [=](Diagnostic &d) {
       d << "unsupported tensor encoding: " << encodingType;
     });

diff --git a/...iler/src/iree/compiler/Dialect/Stream/Transforms/test/encode_host_tensors_packing_i1.mlir b/...iler/src/iree/compiler/Dialect/Stream/Transforms/test/encode_host_tensors_packing_i1.mlir
@@ -1,7 +1,8 @@
-// RUN: iree-opt --split-input-file --iree-stream-encode-host-tensors --iree-experimental-packed-i1-storage %s | FileCheck %s
+// RUN: iree-opt --split-input-file --iree-stream-encode-host-tensors %s | FileCheck %s
 
+#packed = #iree_encoding.packed_storage
 func.func @unaligned_i1_size() -> index {
-  %0 = stream.tensor.sizeof tensor<12xi1> : index
+  %0 = stream.tensor.sizeof tensor<12xi1, #packed> : index
   return %0 : index
 }
 // CHECK: func @unaligned_i1_size() -> index {
@@ -10,8 +11,9 @@ func.func @unaligned_i1_size() -> index {
 
 // -----
 
+#packed = #iree_encoding.packed_storage
 func.func @aligned_i1_size() -> index {
-  %0 = stream.tensor.sizeof tensor<24xi1> : index
+  %0 = stream.tensor.sizeof tensor<24xi1, #packed> : index
   return %0 : index
 }
 

diff --git a/compiler/src/iree/compiler/DispatchCreation/SetEncoding.cpp b/compiler/src/iree/compiler/DispatchCreation/SetEncoding.cpp
@@ -32,6 +32,7 @@ namespace mlir::iree_compiler::DispatchCreation {
 #include "iree/compiler/DispatchCreation/Passes.h.inc"
 
 using IREE::Encoding::EncodingAttr;
+using IREE::Encoding::PackedStorageAttr;
 
 //===---------------------------------------------------------------------===//
 // Utility functions

diff --git a/compiler/src/iree/compiler/Utils/ElementPackingUtils.cpp b/compiler/src/iree/compiler/Utils/ElementPackingUtils.cpp
@@ -15,17 +15,29 @@
 #include "mlir/Dialect/Arith/IR/Arith.h"
 #include "mlir/IR/BuiltinTypes.h"
 
-namespace mlir::iree_compiler {
-
 llvm::cl::opt<bool> clEnableI1Support(
     "iree-experimental-packed-i1-storage",
     llvm::cl::desc(
-        "Experimental feature: enable i1 data type support in codegen"),
+        "Experimental feature: force to use packed storage for i1 tensors."
+        "Turning on this option will see i1 tensors as if it has "
+        "#iree_encoding.packed_storage attribute."
+        "This is to allow an alternative way to test the packed storage "
+        "feature before frontend can emit packed i1 tensors."
+        "This option can be dropped once the frontend can emit packed i1 "
+        "tensors."),
     llvm::cl::init(false));
 
+namespace mlir::iree_compiler {
+
 bool needToPackSubByteElementBitWidth(unsigned bitWidth) {
+  return needToPackSubByteElementBitWidth(
+      bitWidth, /*isI1PackedStorage=*/clEnableI1Support);
+}
+
+bool needToPackSubByteElementBitWidth(unsigned bitWidth,
+                                      bool isI1PackedStorage) {
   // Enable i1 support if requested.
-  if (clEnableI1Support && bitWidth == 1) {
+  if (isI1PackedStorage && bitWidth == 1) {
     return true;
   }
   // Require the original bit width to be some power of two for now to avoid
@@ -37,18 +49,28 @@ bool needToPackSubByteElementBitWidth(unsigned bitWidth) {
 
 bool needToPackSubByteElements(RankedTensorType shapedType) {
   unsigned bitWidth = IREE::Util::getTypeBitWidth(shapedType.getElementType());
-  return needToPackSubByteElementBitWidth(bitWidth);
+  // Two paths to enable packed storage for i1 tensors: the attribute or cl
+  // option. The cl option will be dropped once frontend supports emitting
+  // tensors with attributes.
+  bool isI1PackedStorage =
+      IREE::Encoding::hasPackedStorageAttr(shapedType) || clEnableI1Support;
+  return needToPackSubByteElementBitWidth(bitWidth, isI1PackedStorage);
 }
 
 Type legalizeStorageElementType(Type elementType) {
+  return legalizeStorageElementType(elementType,
+                                    /*isI1PackedStorage=*/clEnableI1Support);
+}
+
+Type legalizeStorageElementType(Type elementType, bool isI1PackedStorage) {
   // Only handle integers; floats in MLIR all have aligned widths (today).
   auto intType = dyn_cast<IntegerType>(elementType);
   if (!intType)
     return elementType;
 
   // For sub-byte elements, default to pack them into bytes.
   unsigned bitWidth = intType.getWidth();
-  if (needToPackSubByteElementBitWidth(bitWidth))
+  if (needToPackSubByteElementBitWidth(bitWidth, isI1PackedStorage))
     return elementType;
 
   // Otherwise, extend them to the next power-of-two bit width.
@@ -72,13 +94,16 @@ Value calculateStorageElementCountInBytes(Location loc,
         loc, builder, shapedType, dynamicDims);
   }
 
+  // TODO: remove cl options once frontend can emit packed i1 tensors.
+  bool packedStorage =
+      IREE::Encoding::hasPackedStorageAttr(shapedType) || clEnableI1Support;
   Type alignedElementType =
-      legalizeStorageElementType(shapedType.getElementType());
+      legalizeStorageElementType(shapedType.getElementType(), packedStorage);
   unsigned elementBits = IREE::Util::getTypeBitWidth(alignedElementType);
 
   // Calculate all static dims first, if any.
   int64_t staticCount = 1;
-  if (!needToPackSubByteElementBitWidth(elementBits)) {
+  if (!needToPackSubByteElementBitWidth(elementBits, packedStorage)) {
     staticCount *= IREE::Util::getRoundedElementByteWidth(alignedElementType);
   }
 
@@ -93,13 +118,13 @@ Value calculateStorageElementCountInBytes(Location loc,
     value = builder.createOrFold<arith::MulIOp>(loc, value, dim);
   }
   // Sub-byte packing requires putting multiple elements in the same byte.
-  if (needToPackSubByteElementBitWidth(elementBits)) {
+  if (needToPackSubByteElementBitWidth(elementBits, packedStorage)) {
     assert(8 % elementBits == 0);
     unsigned byteElements = 8 / elementBits;
     // TODO(antiagainst): We may want to emit runtime check to make sure this is
     // divisible.
     auto divisor = builder.create<arith::ConstantIndexOp>(loc, byteElements);
-    if (!clEnableI1Support && dynamicDims.empty() &&
+    if (!packedStorage && dynamicDims.empty() &&
         (staticCount * elementBits) % 8 != 0) {
       return nullptr;
     }
@@ -113,12 +138,15 @@ Value calculateStorageElementOffsetInBytes(Location loc,
                                            RankedTensorType originalType,
                                            Value linearizedIndex,
                                            OpBuilder &builder) {
-  Type alignedElementType =
-      legalizeStorageElementType(originalType.getElementType());
+  // TODO: remove cl options once frontend can emit packed i1 tensors.
+  bool isI1PackedStorage =
+      IREE::Encoding::hasPackedStorageAttr(originalType) || clEnableI1Support;
+  Type alignedElementType = legalizeStorageElementType(
+      originalType.getElementType(), isI1PackedStorage);
   unsigned elementBits = IREE::Util::getTypeBitWidth(alignedElementType);
 
   // Sub-byte packing requires putting multiple elements in the same byte.
-  if (needToPackSubByteElementBitWidth(elementBits)) {
+  if (needToPackSubByteElementBitWidth(elementBits, isI1PackedStorage)) {
     Value byteElements =
         builder.create<arith::ConstantIndexOp>(loc, 8 / elementBits);
     // TODO(antiagainst): We may want to emit runtime check to make sure this is

diff --git a/compiler/src/iree/compiler/Utils/ElementPackingUtils.h b/compiler/src/iree/compiler/Utils/ElementPackingUtils.h
@@ -15,7 +15,14 @@ namespace mlir::iree_compiler {
 
 /// Returns true if the given |bitWidth|, if appearing at runtime-kernel
 /// interface, is less than a byte that should be tightly packed together.
+bool needToPackSubByteElementBitWidth(unsigned bitWidth,
+                                      bool isI1PackedStorage);
+
+/// Temporary wrapper for the above function. `isI1PackedStorage` will be
+/// determined by the cl option. This allows enabling packed storage for i1
+/// in both attribute and cl option ways.
 bool needToPackSubByteElementBitWidth(unsigned bitWidth);
+
 /// Returns true if the given |shapedType|, if appearing at runtime-kernel
 /// interface, has sub-byte element types that should be tightly packed
 /// together.
@@ -27,6 +34,11 @@ bool needToPackSubByteElements(RankedTensorType shapedType);
 /// runtime and kernel. For such cases, we perform tight packing for supported
 /// sub-byte elements, and expand to the next power-of-two bit width for other
 /// cases.
+Type legalizeStorageElementType(Type elementType, bool isI1PackedStorage);
+
+/// Temporary wrapper for the above function. `isI1PackedStorage` will be
+/// determined by the cl option. This allows enabling packed storage for i1
+/// in both attribute and cl option ways.
 Type legalizeStorageElementType(Type elementType);
 
 /// Emits IR with the given |builder| to calculate the total number of bytes