[Optimizer] Refactor legal gird analysis

include/ttmlir/Dialect/TTNN/Analysis/LegalGridAnalysis.h → include/ttmlir/Dialect/TTNN/Analysis/LegalLayoutAnalysis.h

@@ -2,8 +2,8 @@
 //
 // SPDX-License-Identifier: Apache-2.0
 
-#ifndef TTMLIR_DIALECT_TTNN_ANALYSIS_LEGALGRIDANALYSIS_H
-#define TTMLIR_DIALECT_TTNN_ANALYSIS_LEGALGRIDANALYSIS_H
+#ifndef TTMLIR_DIALECT_TTNN_ANALYSIS_LEGALLAYOUTANALYSIS_H
+#define TTMLIR_DIALECT_TTNN_ANALYSIS_LEGALLAYOUTANALYSIS_H
 
 #include "ttmlir/Dialect/TT/IR/TTOpsTypes.h"
 #include "ttmlir/Dialect/TTNN/Analysis/TTNNAnalysis.h"
@@ -12,46 +12,49 @@
 
 namespace mlir::tt::ttnn {
 
-struct LegalGridAnalysisInput {
+struct LegalLayoutAnalysisInput {
   ChipDescAttr chipDesc;
   GridAttr maxGrid;
   RankedTensorType tensorType;
   int64_t maxShardedGrids;
   llvm::StringMap<OutputLayoutOverrideParams> *outputLayoutOverrides;
+  bool rowMajorEnabled;
 
-  LegalGridAnalysisInput()
+  LegalLayoutAnalysisInput()
       : chipDesc(nullptr), maxGrid(nullptr), tensorType(nullptr),
         outputLayoutOverrides(nullptr) {}
 
-  LegalGridAnalysisInput(
+  LegalLayoutAnalysisInput(
       ChipDescAttr chipDesc, GridAttr maxGrid, RankedTensorType tensorType,
       int64_t maxShardedGrids,
-      llvm::StringMap<OutputLayoutOverrideParams> *outputLayoutOverrides)
+      llvm::StringMap<OutputLayoutOverrideParams> *outputLayoutOverrides,
+      bool rowMajorEnabled)
       : chipDesc(chipDesc), maxGrid(maxGrid), tensorType(tensorType),
         maxShardedGrids(maxShardedGrids),
-        outputLayoutOverrides(outputLayoutOverrides) {}
+        outputLayoutOverrides(outputLayoutOverrides),
+        rowMajorEnabled(rowMajorEnabled) {}
 
-  bool operator==(const LegalGridAnalysisInput &rhs) const {
+  bool operator==(const LegalLayoutAnalysisInput &rhs) const {
     return chipDesc == rhs.chipDesc && maxGrid == rhs.maxGrid &&
            tensorType == rhs.tensorType &&
            outputLayoutOverrides == rhs.outputLayoutOverrides;
   }
 
-  bool operator!=(const LegalGridAnalysisInput &rhs) const {
+  bool operator!=(const LegalLayoutAnalysisInput &rhs) const {
     return !(*this == rhs);
   }
 };
 
-class LegalGridAnalysis
-    : public TTNNAnalysis<LegalGridAnalysisInput, std::vector<TTNNLayoutAttr>> {
+class LegalLayoutAnalysis : public TTNNAnalysis<LegalLayoutAnalysisInput,
+                                                std::vector<TTNNLayoutAttr>> {
 private:
   void analysisImplementation() override;
   bool applyOverrides() override;
 
 public:
-  LegalGridAnalysis(Operation *op) : TTNNAnalysis(op) {}
+  LegalLayoutAnalysis(Operation *op) : TTNNAnalysis(op) {}
 };
 
 } // namespace mlir::tt::ttnn
 
-#endif // TTMLIR_DIALECT_TTNN_ANALYSIS_LEGALGRIDANALYSIS_H
+#endif // TTMLIR_DIALECT_TTNN_ANALYSIS_LEGALLAYOUTANALYSIS_H

include/ttmlir/Dialect/TTNN/Analysis/OpConfigAnalysis.h

@@ -11,22 +11,22 @@
 namespace mlir::tt::ttnn {
 
 struct OpConfigAnalysisInput {
-  llvm::DenseMap<Operation *, std::vector<TTNNLayoutAttr>> legalGrids;
+  llvm::DenseMap<Operation *, std::vector<TTNNLayoutAttr>> legalLayouts;
 
-  OpConfigAnalysisInput() : legalGrids() {}
+  OpConfigAnalysisInput() : legalLayouts() {}
 
   OpConfigAnalysisInput(
       const llvm::DenseMap<Operation *, std::vector<TTNNLayoutAttr>>
-          &&legalGrids)
-      : legalGrids(std::move(legalGrids)) {}
+          &&legalLayouts)
+      : legalLayouts(std::move(legalLayouts)) {}
 
   OpConfigAnalysisInput(
       const llvm::DenseMap<Operation *, std::vector<TTNNLayoutAttr>>
-          &legalGrids)
-      : legalGrids(legalGrids) {}
+          &legalLayouts)
+      : legalLayouts(legalLayouts) {}
 
   bool operator==(const OpConfigAnalysisInput &rhs) const {
-    return legalGrids == rhs.legalGrids;
+    return legalLayouts == rhs.legalLayouts;
   }
 
   bool operator!=(const OpConfigAnalysisInput &rhs) const {

include/ttmlir/Dialect/TTNN/IR/TTNNOpsAttrs.td

@@ -143,6 +143,7 @@ def TTNN_TTNNLayoutAttr: TTNN_Attr<"TTNNLayout", "ttnn_layout"> {
       bool hasInterleavedL1TensorMemoryLayout() const;
       bool isTiled() const;
       Type getElementType() const;
+      Type getScalarElementType() const;
       DataType getDataTypeFromMemRef() const;
       uint64_t getElementSizeBytes() const;
       int64_t getTensorSizeInBytes(ArrayRef<int64_t> tensorShape, ::mlir::tt::DeviceAttr device) const;

include/ttmlir/Dialect/TTNN/Pipelines/TTNNPipelines.h

@@ -24,7 +24,7 @@ struct TTIRToTTNNBackendPipelineOptions
       llvm::cl::desc("Determine and set max valid grid for Op execution."),
       llvm::cl::init(false)};
 
-  // Option to manually insert TTIR_ToLayoutOp for specific op's operand.
+  // Option to manually insert TTNN_ToLayoutOp for specific op's operand.
   // The format is a comma separated list of op names and operand index
   // separated by ':' separator.
   //
@@ -43,9 +43,13 @@ struct TTIRToTTNNBackendPipelineOptions
               "Manually insert memory reconfig op for specific op's operand."),
           llvm::cl::init(llvm::StringMap<InputLayoutOverrideParams>())};
 
-  // Option to override output layout for specific ops.
-  // The format is a comma separated list of op names equal to the output layout
-  // params separated by ":"
+  // Option to override output layout for specific operations. You can
+  // override any number or combination of layout parameters. If not all are
+  // overridden, the remaining ones will be inferred with all possible
+  // combinations generated in LegalLayoutAnalysis. The format is a
+  // comma-separated list of operation names followed by the output layout
+  // parameters, separated by :. The order of parameters does not matter; the
+  // parser will deduce which one is being overridden based on its value.
   //
   // op_name=grid_size:memory_space:tensor_memory_layout:memory_layout:data_type
   //
@@ -58,7 +62,9 @@ struct TTIRToTTNNBackendPipelineOptions
   //   bfp_bf2, u32, u16, u8
   //
   // Full Example:
-  // "op1=2x2:dram:interleaved:tile:fp32,op2=4x4:l1:block_sharded:row_major:fp16"
+  // "op1=2x2:dram:interleaved:tile:fp32,op2=4x4:l1:block_sharded:row_major:f16"
+  // Partial Example:
+  // "op1=2x2:block_sharded"
   //
   //
   // Note: This option is only valid if optimizerPassEnabled is true.
@@ -101,16 +107,23 @@ struct TTIRToTTNNBackendPipelineOptions
           "Pass in a system descriptor flatbuffer to compile against."),
       llvm::cl::init("")};
 
-  // Option to override maximum number of legal layouts for grid analysis
+  // Option to override maximum number of sharded layouts to be generated in
+  // legal layout analysis.
   //
   Option<int64_t> maxLegalLayouts{
       *this, "max-legal-layouts",
-      llvm::cl::desc(
-          "Override maximum number of legal layouts for grid analysis."),
+      llvm::cl::desc("Override maximum number of sharded layouts for legal "
+                     "layout analysis."),
       llvm::cl::init(64)};
 
   ListOption<int64_t> meshShape{
       *this, "mesh-shape", llvm::cl::desc("Set the multi-device mesh shape.")};
+
+  Option<bool> rowMajorEnabled{
+      *this, "row-major-enabled",
+      llvm::cl::desc(
+          "Enable row major layout generation in legal layout analysis."),
+      llvm::cl::init(false)};
 };
 
 // TTIR to EmitC pipeline options.

include/ttmlir/Dialect/TTNN/Transforms/Optimizer.h

@@ -23,6 +23,7 @@ struct TTNNOptimizerOptions {
       MemoryLayoutAnalysisPolicyType::DFSharding;
   bool memReconfigEnabled = false;
   int64_t maxLegalLayouts = 64;
+  bool rowMajorEnabled = false;
 };
 
 std::unique_ptr<::mlir::Pass> createTTNNOptimizer();

include/ttmlir/Dialect/TTNN/Utils/OptimizerOverrides.h

@@ -13,11 +13,19 @@
 namespace mlir::tt::ttnn {
 
 struct OutputLayoutOverrideParams {
-  SmallVector<int64_t, 2> grid;
-  BufferType bufferType;
-  TensorMemoryLayout tensorMemoryLayout; // INTERLEAVED / SHARDED etc...
-  Layout memoryLayout;                   // ROW_MAJOR / TILE
-  tt::DataType dataType;
+  std::optional<SmallVector<int64_t, 2>> grid;
+  std::optional<BufferType> bufferType;
+  std::optional<TensorMemoryLayout>
+      tensorMemoryLayout;             // INTERLEAVED / SHARDED etc...
+  std::optional<Layout> memoryLayout; // ROW_MAJOR / TILE
+  std::optional<tt::DataType> dataType;
+
+  // Check if all layout parameters that are generated in LegalLayoutAnalysis
+  // are overridden. DataType is the only that is not.
+  bool fullLayoutOverride() const {
+    return grid.has_value() && bufferType.has_value() &&
+           tensorMemoryLayout.has_value() && memoryLayout.has_value();
+  }
 };
 
 struct InputLayoutOverrideParams {

lib/Dialect/TTNN/Analysis/CMakeLists.txt

@@ -1,5 +1,5 @@
 add_mlir_dialect_library(MLIRTTNNAnalysis
-        LegalGridAnalysis.cpp
+        LegalLayoutAnalysis.cpp
         OpConfigAnalysis.cpp
         MemoryLayoutAnalysis.cpp
         L1ChainConfig.cpp