104 lines · plain
1// RUN: mlir-opt %s --sparse-reinterpret-map -sparsification | FileCheck %s2 3//4// A SDDMM implementation with "spy" function and5// in-place update of the sampling sparse matrix.6//7 8#BSR = #sparse_tensor.encoding<{9 map = (i, j) -> (10 i floordiv 2 : dense,11 j floordiv 2 : compressed,12 i mod 2 : dense,13 j mod 2 : dense)14}>15 16#trait_SDDMM = {17 indexing_maps = [18 affine_map<(i,j,k) -> (i,k)>, // A19 affine_map<(i,j,k) -> (k,j)>, // B20 affine_map<(i,j,k) -> (i,j)> // S (in/out)21 ],22 iterator_types = ["parallel", "parallel", "reduction"],23 doc = "S(i,j) += spy[S(i,j)] x SUM_k A(i,k) B(k,j)"24}25 26//27// CHECK: #[[$BSR:.+]] = #sparse_tensor.encoding<{ map = (d0, d1) -> (d0 floordiv 2 : dense, d1 floordiv 2 : compressed, d0 mod 2 : dense, d1 mod 2 : dense) }>28// CHECK: #[[$MAP:.+]] = #sparse_tensor.encoding<{ map = (d0, d1, d2, d3) -> (d0 : dense, d1 : compressed, d2 : dense, d3 : dense) }>29//30// CHECK-LABEL: func.func @SDDMM_block(31// CHECK-SAME: %[[VAL_0:.*]]: tensor<?x?xf32, #[[$BSR]]>,32// CHECK-SAME: %[[VAL_1:.*]]: tensor<?x?xf32>,33// CHECK-SAME: %[[VAL_2:.*]]: tensor<?x?xf32>) -> tensor<?x?xf32, #[[$BSR]]> {34// CHECK-DAG: %[[VAL_3:.*]] = arith.constant 1 : index35// CHECK-DAG: %[[VAL_4:.*]] = arith.constant 0 : index36// CHECK-DAG: %[[VAL_5:.*]] = arith.constant 2 : index37// CHECK-DAG: %[[VAL_6:.*]] = arith.constant 0.000000e+00 : f3238// CHECK-DAG: %[[VAL_7:.*]] = sparse_tensor.reinterpret_map %[[VAL_0]] : tensor<?x?xf32, #[[$BSR]]> to tensor<?x?x2x2xf32, #[[$MAP]]>39// CHECK-DAG: %[[VAL_8:.*]] = tensor.dim %[[VAL_1]], %[[VAL_3]] : tensor<?x?xf32>40// CHECK-DAG: %[[VAL_9:.*]] = bufferization.to_buffer %[[VAL_1]] : tensor<?x?xf32> to memref<?x?xf32>41// CHECK-DAG: %[[VAL_10:.*]] = bufferization.to_buffer %[[VAL_2]] : tensor<?x?xf32> to memref<?x?xf32>42// CHECK-DAG: %[[VAL_11:.*]] = sparse_tensor.lvl %[[VAL_7]], %[[VAL_4]] : tensor<?x?x2x2xf32, #[[$MAP]]>43// CHECK-DAG: %[[VAL_12:.*]] = sparse_tensor.positions %[[VAL_7]] {level = 1 : index} : tensor<?x?x2x2xf32, #[[$MAP]]> to memref<?xindex>44// CHECK-DAG: %[[VAL_13:.*]] = sparse_tensor.coordinates %[[VAL_7]] {level = 1 : index} : tensor<?x?x2x2xf32, #[[$MAP]]> to memref<?xindex>45// CHECK-DAG: %[[VAL_14:.*]] = sparse_tensor.values %[[VAL_7]] : tensor<?x?x2x2xf32, #[[$MAP]]> to memref<?xf32>46// CHECK: scf.for %[[VAL_15:.*]] = %[[VAL_4]] to %[[VAL_11]] step %[[VAL_3]] {47// CHECK: %[[VAL_16:.*]] = memref.load %[[VAL_12]]{{\[}}%[[VAL_15]]] : memref<?xindex>48// CHECK: %[[VAL_17:.*]] = arith.addi %[[VAL_15]], %[[VAL_3]] : index49// CHECK: %[[VAL_18:.*]] = memref.load %[[VAL_12]]{{\[}}%[[VAL_17]]] : memref<?xindex>50// CHECK: scf.for %[[VAL_19:.*]] = %[[VAL_16]] to %[[VAL_18]] step %[[VAL_3]] {51// CHECK: %[[VAL_20:.*]] = memref.load %[[VAL_13]]{{\[}}%[[VAL_19]]] : memref<?xindex>52// CHECK: %[[VAL_22:.*]] = arith.muli %[[VAL_19]], %[[VAL_5]] : index53// CHECK: scf.for %[[VAL_21:.*]] = %[[VAL_4]] to %[[VAL_5]] step %[[VAL_3]] {54// CHECK: %[[VAL_23:.*]] = arith.addi %[[VAL_21]], %[[VAL_22]] : index55// CHECK: %[[VAL_25:.*]] = arith.muli %[[VAL_23]], %[[VAL_5]] : index56// CHECK: scf.for %[[VAL_24:.*]] = %[[VAL_4]] to %[[VAL_5]] step %[[VAL_3]] {57// CHECK: %[[VAL_26:.*]] = arith.addi %[[VAL_24]], %[[VAL_25]] : index58// CHECK: %[[VAL_27:.*]] = scf.for %[[VAL_28:.*]] = %[[VAL_4]] to %[[VAL_8]] step %[[VAL_3]] iter_args(%[[VAL_29:.*]] = %[[VAL_6]]) -> (f32) {59// CHECK: %[[VAL_30:.*]] = arith.muli %[[VAL_15]], %[[VAL_5]] : index60// CHECK: %[[VAL_31:.*]] = arith.addi %[[VAL_30]], %[[VAL_21]] : index61// CHECK: %[[VAL_32:.*]] = memref.load %[[VAL_9]]{{\[}}%[[VAL_31]], %[[VAL_28]]] : memref<?x?xf32>62// CHECK: %[[VAL_33:.*]] = arith.muli %[[VAL_20]], %[[VAL_5]] : index63// CHECK: %[[VAL_34:.*]] = arith.addi %[[VAL_33]], %[[VAL_24]] : index64// CHECK: %[[VAL_35:.*]] = memref.load %[[VAL_10]]{{\[}}%[[VAL_28]], %[[VAL_34]]] : memref<?x?xf32>65// CHECK: %[[VAL_36:.*]] = arith.mulf %[[VAL_32]], %[[VAL_35]] : f3266// CHECK: %[[VAL_37:.*]] = arith.addf %[[VAL_29]], %[[VAL_36]] : f3267// CHECK: scf.yield %[[VAL_37]] : f3268// CHECK: } {"Emitted from" = "linalg.generic"}69// CHECK: memref.store %[[VAL_27]], %[[VAL_14]]{{\[}}%[[VAL_26]]] : memref<?xf32>70// CHECK: } {"Emitted from" = "linalg.generic"}71// CHECK: } {"Emitted from" = "linalg.generic"}72// CHECK: } {"Emitted from" = "linalg.generic"}73// CHECK: } {"Emitted from" = "linalg.generic"}74// CHECK: %[[VAL_38:.*]] = sparse_tensor.load %[[VAL_7]] : tensor<?x?x2x2xf32, #[[$MAP]]>75// CHECK: %[[VAL_39:.*]] = sparse_tensor.reinterpret_map %[[VAL_38]] : tensor<?x?x2x2xf32, #[[$MAP]]> to tensor<?x?xf32, #[[$BSR]]>76// CHECK: return %[[VAL_39]] : tensor<?x?xf32, #[[$BSR]]>77// CHECK: }78module {79 func.func @SDDMM_block(%args: tensor<?x?xf32, #BSR>,80 %arga: tensor<?x?xf32>,81 %argb: tensor<?x?xf32>) -> tensor<?x?xf32, #BSR> {82 %result = linalg.generic #trait_SDDMM83 ins(%arga, %argb: tensor<?x?xf32>, tensor<?x?xf32>)84 outs(%args: tensor<?x?xf32, #BSR>) {85 ^bb(%a: f32, %b: f32, %s: f32):86 %f0 = arith.constant 0.0 : f3287 %u = sparse_tensor.unary %s : f32 to f3288 present={89 ^bb0(%p: f32):90 %mul = arith.mulf %a, %b : f3291 sparse_tensor.yield %mul : f3292 }93 absent={}94 %r = sparse_tensor.reduce %s, %u, %f0 : f32 {95 ^bb0(%p: f32, %q: f32):96 %add = arith.addf %p, %q : f3297 sparse_tensor.yield %add : f3298 }99 linalg.yield %r : f32100 } -> tensor<?x?xf32, #BSR>101 return %result : tensor<?x?xf32, #BSR>102 }103}104