[mlir][sparse] fix bugs when computing the memory size when lowering pack op.

[Peiming]

fix check tests.

[Peiming]

fix check tests.

[aartbik]

mlir/lib/Dialect/SparseTensor/Transforms/SparseTensorCodegen.cpp

1276

this assert already encapsulates L1281 assert

(but I am okay pedentically checking, but just in case you want to reduce it a bit)

mlir/lib/Dialect/SparseTensor/Transforms/SparseTensorCodegen.cpp

foreachFieldAndTypeInSparseTensor(
               mem = rewriter.create<memref::CastOp>(loc, fType, mem);
             }
             fields.push_back(mem);
           }
           return true;
         });
 
     MutSparseTensorDescriptor desc(stt, fields);
+    Value c0 = constantIndex(rewriter, loc, 0);
     Value c1 = constantIndex(rewriter, loc, 1);
     Value c2 = constantIndex(rewriter, loc, 2);
-    Value posBack = c1; // index to the last value in the postion array
-    Value memSize = c2; // memory size for current array
+    Value posBack = c0; // index to the last value in the postion array
+    Value memSize = c1; // memory size for current array
 
     Level trailCOOStart = getCOOStart(stt.getEncoding());
     Level trailCOORank = stt.getLvlRank() - trailCOOStart;
     // Sets up SparseTensorSpecifier.
     for (Level lvl = 0, lvlRank = stt.getLvlRank(); lvl < lvlRank; lvl++) {
       assert(!ShapedType::isDynamic(stt.getDimShape()[lvl]));
 
       // FIXME: dim/lvl confusion!
       // Sets up the level size.
       auto lvlSize = constantIndex(rewriter, loc, stt.getDimShape()[lvl]);
       desc.setLvlSize(rewriter, loc, lvl, lvlSize);
       // We use a single AOS array to store the trailing COO, so there is only
       // one memory size to set for the entire COO section.
       if (lvl > trailCOOStart)
         continue;
 
       // Sets up the memory size by reading the last value in position array.
       DimLevelType dlt = stt.getLvlType(lvl);
       // Simply forwards the position index when this is a dense level.
       if (isDenseDLT(dlt)) {
-        memSize = rewriter.create<arith::MulIOp>(loc, lvlSize, posBack);
+        memSize = rewriter.create<arith::MulIOp>(loc, lvlSize, memSize);
         posBack = rewriter.create<arith::SubIOp>(loc, memSize, c1);
         continue;
       }
 
       if (isDLTWithPos(dlt)) {
         assert(isCompressedDLT(dlt) || isCompressedWithHiDLT(dlt));
         if (isCompressedWithHiDLT(dlt)) {
           memSize = rewriter.create<arith::MulIOp>(loc, memSize, c2);
           posBack = rewriter.create<arith::SubIOp>(loc, memSize, c1);
+        } else {
+          assert(isCompressedDLT(dlt));
+          posBack = memSize;
+          memSize = rewriter.create<arith::AddIOp>(loc, memSize, c1);
         }
         desc.setPosMemSize(rewriter, loc, lvl, memSize);
         // The last value in position array is the memory size for next level.
         memSize = genIndexLoad(rewriter, loc, desc.getPosMemRef(lvl), posBack);
         posBack = rewriter.create<arith::SubIOp>(loc, posBack, c1);
       }
       assert(isDLTWithCrd(dlt) && lvl <= trailCOOStart);
       // FIXME: This seems to be unnecessarily complex, can we simplify it?

mlir/test/Dialect/SparseTensor/sparse_pack.mlir

 // CHECK-DAG:       %[[VAL_4:.*]] = memref.cast %[[VAL_3]] : memref<2xindex> to memref<?xindex>
 // CHECK-DAG:       %[[VAL_5:.*]] = bufferization.to_memref %[[VAL_2]] : memref<6x2xi32>
 // CHECK-DAG:       %[[VAL_6:.*]] = memref.collapse_shape %[[VAL_5]] {{\[\[}}0, 1]] : memref<6x2xi32> into memref<12xi32>
 // CHECK-DAG:       %[[VAL_7:.*]] = memref.cast %[[VAL_6]] : memref<12xi32> to memref<?xi32>
 // CHECK-DAG:       %[[VAL_8:.*]] = bufferization.to_memref %[[VAL_0]] : memref<6xf64>
 // CHECK-DAG:       %[[VAL_9:.*]] = memref.cast %[[VAL_8]] : memref<6xf64> to memref<?xf64>
 // CHECK-DAG:       %[[VAL_10:.*]] = sparse_tensor.storage_specifier.init
 // CHECK-DAG:       %[[VAL_11:.*]] = arith.constant 1 : index
-// CHECK-DAG:       %[[VAL_12:.*]] = arith.constant 2 : index
-// CHECK-DAG:       %[[VAL_13:.*]] = arith.constant 100 : index
-// CHECK:           %[[VAL_14:.*]] = sparse_tensor.storage_specifier.set %[[VAL_10]]  lvl_sz at 0 with %[[VAL_13]]
-// CHECK:           %[[VAL_15:.*]] = sparse_tensor.storage_specifier.set %[[VAL_14]]  pos_mem_sz at 0 with %[[VAL_12]]
+// CHECK-DAG:       %[[VAL_12:.*]] = arith.constant 100 : index
+// CHECK:           %[[VAL_13:.*]] = sparse_tensor.storage_specifier.set %[[VAL_10]]  lvl_sz at 0 with %[[VAL_12]]
+// CHECK:           %[[VAL_14:.*]] = arith.addi %[[VAL_11]], %[[VAL_11]] : index
+// CHECK:           %[[VAL_15:.*]] = sparse_tensor.storage_specifier.set %[[VAL_13]]  pos_mem_sz at 0 with %[[VAL_14]]
 // CHECK:           %[[VAL_16:.*]] = memref.load %[[VAL_4]]{{\[}}%[[VAL_11]]] : memref<?xindex>
-// CHECK:           %[[VAL_17:.*]] = arith.muli %[[VAL_16]], %[[VAL_12]] : index
-// CHECK:           %[[VAL_18:.*]] = sparse_tensor.storage_specifier.set %[[VAL_15]]  crd_mem_sz at 0 with %[[VAL_17]]
-// CHECK:           %[[VAL_19:.*]] = sparse_tensor.storage_specifier.set %[[VAL_18]]  lvl_sz at 1 with %[[VAL_13]]
-// CHECK:           %[[VAL_20:.*]] = sparse_tensor.storage_specifier.set %[[VAL_19]]  val_mem_sz with %[[VAL_16]]
-// CHECK:           return %[[VAL_4]], %[[VAL_7]], %[[VAL_9]], %[[VAL_20]]
+// CHECK:           %[[VAL_17:.*]] = arith.constant 2 : index
+// CHECK:           %[[VAL_18:.*]] = arith.muli %[[VAL_16]], %[[VAL_17]] : index
+// CHECK:           %[[VAL_19:.*]] = sparse_tensor.storage_specifier.set %[[VAL_15]]  crd_mem_sz at 0 with %[[VAL_18]]
+// CHECK:           %[[VAL_20:.*]] = sparse_tensor.storage_specifier.set %[[VAL_19]]  lvl_sz at 1 with %[[VAL_12]]
+// CHECK:           %[[VAL_21:.*]] = sparse_tensor.storage_specifier.set %[[VAL_20]]  val_mem_sz with %[[VAL_16]]
+// CHECK:           return %[[VAL_4]], %[[VAL_7]], %[[VAL_9]], %[[VAL_21]]
 // CHECK:         }
 func.func @sparse_pack(%values: tensor<6xf64>, %pos:tensor<2xindex>, %coordinates: tensor<6x2xi32>)
                     -> tensor<100x100xf64, #COO> {
   %0 = sparse_tensor.pack %values, %pos, %coordinates
      : tensor<6xf64>, tensor<2xindex>, tensor<6x2xi32> to tensor<100x100xf64, #COO>
   return %0 : tensor<100x100xf64, #COO>
 }
 

mlir/test/Integration/Dialect/SparseTensor/CPU/sparse_pack.mlir

%index32 = arith.constant dense<
         [  7,  8]]
     > : tensor<3x2xi32>
 
     %s4 = sparse_tensor.pack %data, %pos, %index : tensor<3xf64>, tensor<2xindex>, tensor<3x2xindex>
                                           to tensor<10x10xf64, #SortedCOO>
     %s5= sparse_tensor.pack %data, %pos32, %index32 : tensor<3xf64>, tensor<2xi32>, tensor<3x2xi32>
                                            to tensor<10x10xf64, #SortedCOOI32>
 
+    %csr_data = arith.constant dense<
+       [  1.0,  2.0,  3.0,  4.0]
+    > : tensor<4xf64>
+
     %csr_pos32 = arith.constant dense<
        [0, 1, 3]
     > : tensor<3xi32>
 
     %csr_index32 = arith.constant dense<
        [1, 0, 1]
     > : tensor<3xi32>
-    %csr= sparse_tensor.pack %data, %csr_pos32, %csr_index32 : tensor<3xf64>, tensor<3xi32>, tensor<3xi32>
+    %csr= sparse_tensor.pack %csr_data, %csr_pos32, %csr_index32 : tensor<4xf64>, tensor<3xi32>, tensor<3xi32>
                                            to tensor<2x2xf64, #CSR>
 
     %bdata = arith.constant dense<
        [  1.0,  2.0,  3.0,  4.0,  5.0,  0.0]
     > : tensor<6xf64>
 
     %bpos = arith.constant dense<
        [0, 3, 3, 5]
func.func @entry() {
     // CHECK-NEXT:3
     sparse_tensor.foreach in %csr : tensor<2x2xf64, #CSR> do {
       ^bb0(%1: index, %2: index, %v: f64) :
         vector.print %1: index
         vector.print %2: index
         vector.print %v: f64
      }
 
+    %d_csr = tensor.empty() : tensor<4xf64>
+    %p_csr = tensor.empty() : tensor<3xi32>
+    %i_csr = tensor.empty() : tensor<3xi32>
+    %rd_csr, %rp_csr, %ri_csr = sparse_tensor.unpack %csr : tensor<2x2xf64, #CSR>
+                 outs(%d_csr, %p_csr, %i_csr : tensor<4xf64>, tensor<3xi32>, tensor<3xi32>)
+                 -> tensor<4xf64>, tensor<3xi32>, tensor<3xi32>
+
+    // CHECK-NEXT: ( 1, 2, 3, {{.*}} )
+    %vd_csr = vector.transfer_read %rd_csr[%c0], %f0 : tensor<4xf64>, vector<4xf64>
+    vector.print %vd_csr : vector<4xf64>
 
     // CHECK-NEXT:1
     // CHECK-NEXT:2
     // CHECK-NEXT:3
     //
     // CHECK-NEXT:4
     // CHECK-NEXT:5
     //