Error fixes in adding allBlocks to IT #20

Author: pflynn157

CMakeLists.txt

@@ -29,7 +29,7 @@ set(CMAKE_MODULE_PATH ${CMAKE_SOURCE_DIR}/cmake/modules)
 if (MSVC)
   set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} /EHs-c- /GR-")
 else ()
-  set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -fno-exceptions -fno-rtti")
+  set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -fno-exceptions -fno-rtti -g")
 endif ()
 
 #-------------------------------------------------------------------------------

first.mlir

@@ -1,19 +1,38 @@
+# Sparse matrix sparse matrix elementwise addition
+# Sparse matrix is in CSR format. Currently workspace transformation on the IndexTree dialect works for only CSR format
+# RUN: comet-opt --opt-comp-workspace --convert-ta-to-it --convert-to-loops --convert-to-llvm %s &> eltwise_add_CSRxCSR_oCSR.llvm
+# RUN: export SPARSE_FILE_NAME0=%comet_integration_test_data_dir/test_rank2.mtx
+# RUN: export SPARSE_FILE_NAME1=%comet_integration_test_data_dir/test_rank2_transpose.mtx
+# RUN: mlir-cpu-runner eltwise_add_CSRxCSR_oCSR.llvm -O3 -e main -entry-point-result=void -shared-libs=%comet_utility_library_dir/libcomet_runner_utils%shlibext | FileCheck %s
+
+
 def main() {
-	#IndexLabel Declarations
-	IndexLabel [a] = [?];
-	IndexLabel [b] = [?];   
+    #IndexLabel Declarations
+    IndexLabel [i] = [?];
+    IndexLabel [j] = [?];
+    
+    #Tensor Declarations
+    Tensor<double> A([i, j], {CSR});	 
+    Tensor<double> B([i, j], {CSR});
+    Tensor<double> C([i, j], {CSR});
 
-	Tensor<double> A([a, b], {CSR});
-	Tensor<double> B([b, a], {Dense});
-	Tensor<double> C([b, a], {Dense});
-	
-    A[a, b] = comet_read(0);
-    B[b, a] = 1.0;
-    C[b, a] = A[a, b] * B[b, a];
-    #C[b, a] = 1.0;
+    #Tensor Readfile Operation
+    A[i, j] = comet_read(0);
+    B[i, j] = comet_read(1);
 
-	print(A);
-	print(B);
-	print(C);
+    #Tensor Contraction
+    C[i, j] = A[i, j] + B[i, j];
+    print(C);
 }
 
+# Print the result for verification.
+# CHECK: data = 
+# CHECK-NEXT: 5,
+# CHECK-NEXT: data = 
+# CHECK-NEXT: 0,
+# CHECK-NEXT: data = 
+# CHECK-NEXT: 0,2,4,5,7,9,
+# CHECK-NEXT: data = 
+# CHECK-NEXT: 0,3,1,4,2,0,3,1,4,
+# CHECK-NEXT: data = 
+# CHECK-NEXT: 2,5.5,4,7.7,6,5.5,8,7.7,10,

first.ta

@@ -214,6 +214,7 @@ namespace mlir
 
     void getFormatsPermsOfComputeOp(Value computeOp,
                                     std::vector<std::vector<std::string>> &opFormats,
+                                    std::vector<std::vector<std::string>> &opBlocks,
                                     std::vector<std::vector<int>> &opPerms,
                                     std::vector<std::vector<bool>> &inputOutputMapping);
 

include/comet/Dialect/Utils/Utils.h

@@ -185,9 +185,10 @@ void IndexTreeKernelFusionPass::test(mlir::func::FuncOp &funcop)
                   comet_debug() <<  " is_comp_worksp_opt: " << is_comp_worksp_opt << " semiring: " << semiring << "\n";
 
                   std::vector<std::vector<std::string>> opFormats;
+                  std::vector<std::vector<std::string>> opBlocks;
                   std::vector<std::vector<int>> opPerms;
                   std::vector<std::vector<bool> > inputOutputMapping;
-                  getFormatsPermsOfComputeOp(computeOp, opFormats, opPerms, inputOutputMapping);
+                  getFormatsPermsOfComputeOp(computeOp, opFormats, opBlocks, opPerms, inputOutputMapping);
                   /// opFormats
                   comet_debug() << "[";
                   for (auto strings: opFormats) {
@@ -538,7 +539,7 @@ mlir::Value IndexTreeKernelFusionPass::createReducedComputeRHS(
     SmallVector<StringRef, 8> blocks;
     if (b_i == tensor_id)
     { /// for the new reduced tensor
-      blocks.insert(blocks.end(), old_formats_strs[b_i].begin() + rank_base, old_blocks_strs[b_i].end());
+      blocks.insert(blocks.end(), old_blocks_strs[b_i].begin() + rank_base, old_blocks_strs[b_i].end());
     }
     else
     { /// for other remaining old operands

lib/Dialect/IndexTree/Transforms/Fusion.cpp

@@ -661,7 +661,7 @@ std::vector<Value> CompressedWorkspaceOutput(std::vector<int> sparseDimsOutput,
 
   /// Convert blocks string array into StrAttr
   std::vector<std::string> c3_blocks_str_0 = {"UNK"};
-  std::vector<std::string> c3_blocks_str_1 = opFormats[2];
+  std::vector<std::string> c3_blocks_str_1 = opBlocks[2];
   std::vector<std::vector<std::string>> c3_blocks_str = {c3_blocks_str_0, c3_blocks_str_1};
 
   std::vector<mlir::Value> c3_rhs = workspaceTensors;
@@ -769,7 +769,7 @@ void CompressedWorkspaceInput(std::vector<Value> computeOps, OpBuilder &builder,
     std::vector<std::vector<std::string>> opBlocks;
     std::vector<std::vector<int>> opPerms;
     std::vector<std::vector<bool>> inputOutputMapping;
-    getFormatsPermsOfComputeOp(computeOp, opFormats, opPerms, inputOutputMapping);
+    getFormatsPermsOfComputeOp(computeOp, opFormats, opBlocks, opPerms, inputOutputMapping);
     comet_debug() << " \n";
     for (auto n : opFormats)
     {
@@ -896,8 +896,8 @@ void CompressedWorkspaceInput(std::vector<Value> computeOps, OpBuilder &builder,
       std::vector<std::string> c2_formats_str_1 = {"D"};
       std::vector<std::vector<std::string>> c2_formats_str = {c2_formats_str_0, c2_formats_str_1};
 
-      std::vector<std::string> c2_blocks_str_0 = opFormats[sparseDimsInput[0].tensorId];
-      std::vector<std::string> c2_blocks_str_1 = {"D"};
+      std::vector<std::string> c2_blocks_str_0 = opBlocks[sparseDimsInput[0].tensorId];
+      std::vector<std::string> c2_blocks_str_1 = {"UNK"};
       std::vector<std::vector<std::string>> c2_blocks_str = {c2_blocks_str_0, c2_blocks_str_1};
 
       std::vector<mlir::Value> c2_rhs = {tensors_rhs[sparseDimsInput[0].tensorId]};
@@ -1074,7 +1074,7 @@ void IndexTreeWorkspaceTransformationsPass::CompressedWorkspaceTransforms(mlir::
                 std::vector<std::vector<std::string>> opBlocks;
                 std::vector<std::vector<int>> opPerms;
                 std::vector<std::vector<bool>> inputOutputMapping;
-                getFormatsPermsOfComputeOp(computeOp, opFormats, opPerms, inputOutputMapping);
+                getFormatsPermsOfComputeOp(computeOp, opFormats, opBlocks, opPerms, inputOutputMapping);
 
 #ifdef DEBUG_MODE_WorkspaceTransformsPass
                 comet_debug() << "Print opFormats:\n";

lib/Dialect/IndexTree/Transforms/WorkspaceTransforms.cpp

@@ -51,7 +51,7 @@ using namespace mlir::indexTree;
 #define DEBUG_TYPE "tensor-decl-lowering"
 
 // *********** For debug purpose *********//
-#define COMET_DEBUG_MODE
+// #define COMET_DEBUG_MODE
 #include "comet/Utils/debug.h"
 #undef COMET_DEBUG_MODE
 // *********** For debug purpose *********//

lib/Dialect/TensorAlgebra/Transforms/TensorDeclLowering.cpp

@@ -1459,15 +1459,19 @@ namespace mlir
     /// Get the perms and formats of the itCompute op
     void getFormatsPermsOfComputeOp(Value computeOp,
                                     std::vector<std::vector<std::string>> &opFormats,
+                                    std::vector<std::vector<std::string>> &opBlocks,
                                     std::vector<std::vector<int>> &opPerms,
                                     std::vector<std::vector<bool>> &inputOutputMapping)
     {
       indexTree::IndexTreeComputeRHSOp itComputeOp_rhs = dyn_cast<indexTree::IndexTreeComputeRHSOp>(computeOp.getDefiningOp()->getOperand(0).getDefiningOp());
       ArrayAttr opFormatsArrayAttr_rhs = itComputeOp_rhs.getAllFormats();
+      ArrayAttr opBlocksArrayAttr_rhs = itComputeOp_rhs.getAllBlocks();
       ArrayAttr opPermsArrayAttr_rhs = itComputeOp_rhs.getAllPerms();
       indexTree::IndexTreeComputeLHSOp itComputeOp_lhs = dyn_cast<indexTree::IndexTreeComputeLHSOp>(computeOp.getDefiningOp()->getOperand(1).getDefiningOp());
       ArrayAttr opFormatsArrayAttr_lhs = itComputeOp_lhs.getAllFormats();
+      ArrayAttr opBlocksArrayAttr_lhs = itComputeOp_lhs.getAllBlocks();
       ArrayAttr opPermsArrayAttr_lhs = itComputeOp_lhs.getAllPerms();
+      ///TODO(patrick) We should probably verify the block sizes
       assert(opFormatsArrayAttr_rhs.size() == opPermsArrayAttr_rhs.size() && "not equal RHS formats size with perms size\n");
       assert(opFormatsArrayAttr_lhs.size() == opPermsArrayAttr_lhs.size() && "not equal LHS formats size with perms size\n");
 
@@ -1476,17 +1480,25 @@ namespace mlir
       comet_debug() << "Start printing opFormats_rhs\n";
       std::vector<std::vector<std::string>> opFormats_rhs = convertArrayAttrStrTo2DVector(opFormatsArrayAttr_rhs);
       comet_debug() << "End printing opFormats_rhs\n";
+      comet_debug() << "Start printing opBlocks_rhs\n";
+      std::vector<std::vector<std::string>> opBlocks_rhs = convertArrayAttrStrTo2DVector(opBlocksArrayAttr_rhs);
+      comet_debug() << "End printing opBlocks_rhs\n";
       std::vector<std::vector<int>> opPerms_rhs = convertArrayAttrIntTo2DVector(opPermsArrayAttr_rhs);
       std::vector<std::vector<bool>> inputMapping = createInputOutputMapping(opPermsArrayAttr_rhs, true);
 
       comet_debug() << "Start printing opFormats_lhs\n";
       std::vector<std::vector<std::string>> opFormats_lhs = convertArrayAttrStrTo2DVector(opFormatsArrayAttr_lhs);
       comet_debug() << "End printing opFormats_lhs\n";
+      comet_debug() << "Start printing opBlocks_lhs\n";
+      std::vector<std::vector<std::string>> opBlocks_lhs = convertArrayAttrStrTo2DVector(opBlocksArrayAttr_lhs);
+      comet_debug() << "End printing opBlocks_lhs\n";
       std::vector<std::vector<int>> opPerms_lhs = convertArrayAttrIntTo2DVector(opPermsArrayAttr_lhs);
       std::vector<std::vector<bool>> outputMapping = createInputOutputMapping(opPermsArrayAttr_lhs, false);
 
       opFormats = opFormats_rhs;
       opFormats.insert(opFormats.end(), opFormats_lhs.begin(), opFormats_lhs.end());
+      opBlocks = opBlocks_rhs;
+      opBlocks.insert(opBlocks.end(), opBlocks_lhs.begin(), opBlocks_lhs.end());
       opPerms = opPerms_rhs;
       opPerms.insert(opPerms.end(), opPerms_lhs.begin(), opPerms_lhs.end());
       inputOutputMapping = inputMapping;
@@ -1599,10 +1611,11 @@ namespace mlir
           {
             comet_debug() << " getFormatsInfo:leafs[" << j << "] is computeOp\n";
             std::vector<std::vector<std::string>> allFormats;
+            std::vector<std::vector<std::string>> allBlocks;
             std::vector<std::vector<int>> allPerms;
             std::vector<std::vector<bool>> inputOutputMapping;
             OpBuilder builder(leafop);
-            getFormatsPermsOfComputeOp(leafop, allFormats, allPerms, inputOutputMapping);
+            getFormatsPermsOfComputeOp(leafop, allFormats, allBlocks, allPerms, inputOutputMapping);
 
             comet_debug() << " getFormatsInfo:Allformats allFormats.size(): " << allFormats.size() << "\n";
             for (auto m : allFormats)

lib/Dialect/Utils/Utils.cpp

@@ -1,9 +1,13 @@
 #!/bin/bash
 
 export LD_LIBRARY_PATH="/home/patrick/Work/PNNL/COMET/install/lib"
-export SPARSE_FILE_NAME0=first.mtx
 
-build/bin/comet-opt --convert-ta-to-it --convert-to-loops --convert-to-llvm first.ta &> first.mlir
+#export SPARSE_FILE_NAME0=first.mtx
+#build/bin/comet-opt --convert-ta-to-it --convert-to-loops --convert-to-llvm first.ta &> first.mlir
+
+export SPARSE_FILE_NAME0=integration_test/data/test_rank2.mtx
+export SPARSE_FILE_NAME1=integration_test/data/test_rank2_transpose.mtx
+build/bin/comet-opt --opt-comp-workspace --convert-ta-to-it --convert-to-loops --convert-to-llvm first.ta &> first.mlir
 
 llvm/build/bin/mlir-cpu-runner first.mlir -O3 -e main -entry-point-result=void \
     -shared-libs=build/lib/libcomet_runner_utils.so

run.sh

@@ -0,0 +1,13 @@
+%%MatrixMarket matrix coordinate real general
+%
+% This is a test sparse matrix in Matrix Market Exchange Format.
+% see https://math.nist.gov/MatrixMarket
+%
+4 6 7
+1 1 5.0
+1 2 1.0
+2 1 7.0
+2 2 3.0
+4 1 8.0
+4 4 4.0
+4 5 9.0

test.mtx

@@ -0,0 +1,20 @@
+def main() {
+	#IndexLabel Declarations
+	IndexLabel [a] = [4];
+	IndexLabel [b] = [?];
+	IndexLabel [c] = [?];             
+
+	#Tensor Declarations
+	Tensor<double> B([b, c], {CSR});   #sparse tensor declarations should be before dense tensor declarations
+	Tensor<double> A([a, b], {Dense});	  
+	Tensor<double> C([a, c], {Dense});
+
+	#Tensor Fill Operation
+	A[a, b] = 1.0;
+	B[b, c] = comet_read(0);
+	C[a, c] = 0.0;
+
+	C[a, c] = A[a, b] * B[b, c];
+	print(C);
+	#print(B);
+}