| // RUN: iree-dialects-opt --transform-dialect-interpreter --split-input-file -canonicalize -cse %s | FileCheck %s |
| |
| func.func @scatter_tiling_distribution( |
| %original: tensor<?x?xf32>, %indices: tensor<?x1xi32>, |
| %update : tensor<?x?xf32>) -> tensor<?x?xf32> { |
| %0 = iree_linalg_ext.scatter |
| dimension_map = [0] |
| unique_indices(true) |
| ins(%update, %indices : tensor<?x?xf32>, tensor<?x1xi32>) |
| outs(%original : tensor<?x?xf32>) { |
| ^bb0(%arg1: f32, %arg2: f32): |
| %1 = arith.addf %arg1, %arg2 : f32 |
| iree_linalg_ext.yield %1 : f32 |
| } -> tensor<?x?xf32> |
| return %0 : tensor<?x?xf32> |
| } |
| transform.sequence failures(propagate) { |
| ^bb1(%module_op: !transform.any_op): |
| %0 = transform.structured.match ops{["iree_linalg_ext.scatter"]} in %module_op : (!transform.any_op) -> !transform.any_op |
| %forall, %tiled_op = transform.structured.tile_using_forall %0 tile_sizes [10, 30, 0] { mapping = [#gpu.thread<y>, #gpu.thread<x>] } : (!transform.any_op) -> (!transform.any_op, !transform.any_op) |
| } |
| // CHECK-DAG: #[[MAP0:.+]] = affine_map<()[s0] -> (s0 ceildiv 10)> |
| // CHECK-DAG: #[[MAP1:.+]] = affine_map<()[s0] -> (s0 ceildiv 30)> |
| // CHECK-DAG: #[[MAP2:.+]] = affine_map<(d0) -> (d0 * 10)> |
| // CHECK-DAG: #[[MAP3:.+]] = affine_map<(d0)[s0] -> (d0 * -10 + s0, 10)> |
| // CHECK-DAG: #[[MAP4:.+]] = affine_map<(d0) -> (d0 * 30)> |
| // CHECK-DAG: #[[MAP5:.+]] = affine_map<(d0)[s0] -> (d0 * -30 + s0, 30)> |
| // CHECK: func.func @scatter_tiling_distribution( |
| // CHECK-SAME: %[[ORIGINAL:[a-zA-Z0-9_]+]] |
| // CHECK-SAME: %[[INDICES:[a-zA-Z0-9_]+]] |
| // CHECK-SAME: %[[UPDATES:[a-zA-Z0-9_]+]] |
| // CHECK-DAG: %[[C0:.+]] = arith.constant 0 : index |
| // CHECK-DAG: %[[C1:.+]] = arith.constant 1 : index |
| // CHECK-DAG: %[[D0:.+]] = tensor.dim %[[UPDATES]], %[[C0]] |
| // CHECK-DAG: %[[D1:.+]] = tensor.dim %[[UPDATES]], %[[C1]] |
| // CHECK-DAG: %[[UB0:.+]] = affine.apply #[[MAP0]]()[%[[D0]]] |
| // CHECK-DAG: %[[UB1:.+]] = affine.apply #[[MAP1]]()[%[[D1]]] |
| // CHECK: %[[RESULT:.+]] = scf.forall (%[[IV0:.+]], %[[IV1:.+]]) in (%[[UB0]], %[[UB1]]) shared_outs(%[[ITER:.+]] = %[[ORIGINAL]]) |
| // CHECK-DAG: %[[I:.+]] = affine.apply #[[MAP2]](%[[IV0]]) |
| // CHECK-DAG: %[[I_SZ:.+]] = affine.min #[[MAP3]](%[[IV0]])[%[[D0]]] |
| // CHECK-DAG: %[[J:.+]] = affine.apply #[[MAP4]](%[[IV1]]) |
| // CHECK-DAG: %[[J_SZ:.+]] = affine.min #[[MAP5]](%[[IV1]])[%[[D1]]] |
| // CHECK: %[[UPDATES_TILE:.+]] = tensor.extract_slice %[[UPDATES]] |
| // CHECK-SAME: [%[[I]], %[[J]]] [%[[I_SZ]], %[[J_SZ]]] [1, 1] |
| // CHECK: %[[INDICES_TILE:.+]] = tensor.extract_slice %[[INDICES]] |
| // CHECK-SAME: [%[[I]], 0] [%[[I_SZ]], 1] [1, 1] |
| // CHECK: %[[ITER_D0:.+]] = tensor.dim %[[ITER]], %[[C0]] |
| // CHECK: %[[ITER_TILE:.+]] = tensor.extract_slice %[[ITER]] |
| // CHECK-SAME: [0, %[[J]]] [%[[ITER_D0]], %[[J_SZ]]] [1, 1] |
| // CHECK: %[[SCATTER:.+]] = iree_linalg_ext.scatter |
| // CHECK-SAME: dimension_map = [0] unique_indices(true) |
| // CHECK-SAME: ins(%[[UPDATES_TILE]], %[[INDICES_TILE]] |
| // CHECK-SAME: outs(%[[ITER_TILE]] |
| // CHECK: %[[ORIGINAL_D0:.+]] = tensor.dim %[[ORIGINAL]], %[[C0]] |
| // CHECK: scf.forall.in_parallel { |
| // CHECK: tensor.parallel_insert_slice %[[SCATTER]] into %[[ITER]] |
| // CHECK-SAME: [0, %[[J]]] [%[[ORIGINAL_D0]], %[[J_SZ]]] |
| // CHECK: } |
| // CHECK } {mapping = [#gpu.thread<y>, #gpu.thread<x>]} |
| // CHECK: return %[[RESULT]] |
| |
| // ----- |
| |
| func.func @sort_3d_multi_result_distribute( |
| %arg0: tensor<?x?x?xi32>, %arg1 : tensor<?x?x?xf32>) |
| -> (tensor<?x?x?xi32>, tensor<?x?x?xf32>) { |
| %0, %1 = iree_linalg_ext.sort |
| dimension(2) |
| outs(%arg0, %arg1 : tensor<?x?x?xi32>, tensor<?x?x?xf32>) { |
| ^bb0(%arg2: i32, %arg3: i32, %arg4 : f32, %arg5 : f32): // no predecessors |
| %2 = arith.cmpf ogt, %arg4, %arg5 : f32 |
| iree_linalg_ext.yield %2 : i1 |
| } -> tensor<?x?x?xi32>, tensor<?x?x?xf32> |
| return %0, %1 : tensor<?x?x?xi32>, tensor<?x?x?xf32> |
| } |
| transform.sequence failures(propagate) { |
| ^bb1(%module_op: !transform.any_op): |
| %0 = transform.structured.match ops{["iree_linalg_ext.sort"]} in %module_op : (!transform.any_op) -> !transform.any_op |
| %forall, %tiled_op = transform.structured.tile_using_forall %0 tile_sizes [10, 30, 0] : (!transform.any_op) -> (!transform.any_op, !transform.any_op) |
| } |
| // CHECK-DAG: #[[MAP0:.+]] = affine_map<()[s0] -> (s0 ceildiv 10)> |
| // CHECK-DAG: #[[MAP1:.+]] = affine_map<()[s0] -> (s0 ceildiv 30)> |
| // CHECK-DAG: #[[MAP2:.+]] = affine_map<(d0) -> (d0 * 10)> |
| // CHECK-DAG: #[[MAP3:.+]] = affine_map<(d0)[s0] -> (d0 * -10 + s0, 10)> |
| // CHECK-DAG: #[[MAP4:.+]] = affine_map<(d0) -> (d0 * 30)> |
| // CHECK-DAG: #[[MAP5:.+]] = affine_map<(d0)[s0] -> (d0 * -30 + s0, 30)> |
| // CHECK: func.func @sort_3d_multi_result_distribute( |
| // CHECK-SAME: %[[SRC0:[a-zA-Z0-9_]+]] |
| // CHECK-SAME: %[[SRC1:[a-zA-Z0-9_]+]] |
| // CHECK-DAG: %[[C0:.+]] = arith.constant 0 : index |
| // CHECK-DAG: %[[C1:.+]] = arith.constant 1 : index |
| // CHECK-DAG: %[[C2:.+]] = arith.constant 2 : index |
| // CHECK-DAG: %[[D0:.+]] = tensor.dim %[[SRC0]], %[[C0]] |
| // CHECK-DAG: %[[D1:.+]] = tensor.dim %[[SRC0]], %[[C1]] |
| // CHECK-DAG: %[[D2:.+]] = tensor.dim %[[SRC0]], %[[C2]] |
| // CHECK-DAG: %[[UB0:.+]] = affine.apply #[[MAP0]]()[%[[D0]]] |
| // CHECK-DAG: %[[UB1:.+]] = affine.apply #[[MAP1]]()[%[[D1]]] |
| // CHECK: %[[RESULT:.+]]:2 = scf.forall (%[[IV0:.+]], %[[IV1:.+]]) in (%[[UB0]], %[[UB1]]) shared_outs(%[[ITER0:.+]] = %[[SRC0]], %[[ITER1:.+]] = %[[SRC1]]) |
| // CHECK-DAG: %[[I:.+]] = affine.apply #[[MAP2]](%[[IV0]]) |
| // CHECK-DAG: %[[I_SZ:.+]] = affine.min #[[MAP3]](%[[IV0]])[%[[D0]]] |
| // CHECK-DAG: %[[J:.+]] = affine.apply #[[MAP4]](%[[IV1]]) |
| // CHECK-DAG: %[[J_SZ:.+]] = affine.min #[[MAP5]](%[[IV1]])[%[[D1]]] |
| // CHECK: %[[ITER0_TILE:.+]] = tensor.extract_slice %[[ITER0]] |
| // CHECK-SAME: [%[[I]], %[[J]], 0] [%[[I_SZ]], %[[J_SZ]], %[[D2]]] |
| // CHECK: %[[ITER1_TILE:.+]] = tensor.extract_slice %[[ITER1]] |
| // CHECK-SAME: [%[[I]], %[[J]], 0] [%[[I_SZ]], %[[J_SZ]], %[[D2]]] |
| // CHECK: %[[SORT:.+]]:2 = iree_linalg_ext.sort dimension(2) |
| // CHECK-SAME: outs(%[[ITER0_TILE]], %[[ITER1_TILE]] |
| // CHECK: scf.forall.in_parallel { |
| // CHECK: tensor.parallel_insert_slice %[[SORT]]#0 into %[[ITER0]] |
| // CHECK-SAME: [%[[I]], %[[J]], 0] [%[[I_SZ]], %[[J_SZ]], %[[D2]]] [1, 1, 1] |
| // CHECK: tensor.parallel_insert_slice %[[SORT]]#1 into %[[ITER1]] |
| // CHECK-SAME: [%[[I]], %[[J]], 0] [%[[I_SZ]], %[[J_SZ]], %[[D2]]] [1, 1, 1] |
| // CHECK: } |
| // CHECK } |
| // CHECK: return %[[RESULT]]#0, %[[RESULT]]#1 |
| |
| // ----- |
| |
| func.func @sort_3d_multi_result_distribute_memref( |
| %arg0: memref<?x?x?xi32>, %arg1 : memref<?x?x?xf32>) { |
| iree_linalg_ext.sort |
| dimension(2) |
| outs(%arg0, %arg1 : memref<?x?x?xi32>, memref<?x?x?xf32>) { |
| ^bb0(%arg2: i32, %arg3: i32, %arg4 : f32, %arg5 : f32): // no predecessors |
| %0 = arith.cmpf ogt, %arg4, %arg5 : f32 |
| iree_linalg_ext.yield %0 : i1 |
| } |
| return |
| } |
| transform.sequence failures(propagate) { |
| ^bb1(%module_op: !transform.any_op): |
| %0 = transform.structured.match ops{["iree_linalg_ext.sort"]} in %module_op : (!transform.any_op) -> !transform.any_op |
| %forall, %tiled_op = transform.structured.tile_using_forall %0 tile_sizes [10, 30, 0] : (!transform.any_op) -> (!transform.any_op, !transform.any_op) |
| } |
| // CHECK-DAG: #[[MAP0:.+]] = affine_map<()[s0] -> (s0 ceildiv 10)> |
| // CHECK-DAG: #[[MAP1:.+]] = affine_map<()[s0] -> (s0 ceildiv 30)> |
| // CHECK-DAG: #[[MAP2:.+]] = affine_map<(d0) -> (d0 * 10)> |
| // CHECK-DAG: #[[MAP3:.+]] = affine_map<(d0)[s0] -> (d0 * -10 + s0, 10)> |
| // CHECK-DAG: #[[MAP4:.+]] = affine_map<(d0) -> (d0 * 30)> |
| // CHECK-DAG: #[[MAP5:.+]] = affine_map<(d0)[s0] -> (d0 * -30 + s0, 30)> |
| // CHECK: func.func @sort_3d_multi_result_distribute_memref( |
| // CHECK-SAME: %[[SRC0:[a-zA-Z0-9_]+]] |
| // CHECK-SAME: %[[SRC1:[a-zA-Z0-9_]+]] |
| // CHECK-DAG: %[[C0:.+]] = arith.constant 0 : index |
| // CHECK-DAG: %[[C1:.+]] = arith.constant 1 : index |
| // CHECK-DAG: %[[C2:.+]] = arith.constant 2 : index |
| // CHECK-DAG: %[[D0:.+]] = memref.dim %[[SRC0]], %[[C0]] |
| // CHECK-DAG: %[[D1:.+]] = memref.dim %[[SRC0]], %[[C1]] |
| // CHECK-DAG: %[[D2:.+]] = memref.dim %[[SRC0]], %[[C2]] |
| // CHECK-DAG: %[[UB0:.+]] = affine.apply #[[MAP0]]()[%[[D0]]] |
| // CHECK-DAG: %[[UB1:.+]] = affine.apply #[[MAP1]]()[%[[D1]]] |
| // CHECK: scf.forall (%[[IV0:.+]], %[[IV1:.+]]) in (%[[UB0]], %[[UB1]]) |
| // CHECK-DAG: %[[I:.+]] = affine.apply #[[MAP2]](%[[IV0]]) |
| // CHECK-DAG: %[[I_SZ:.+]] = affine.min #[[MAP3]](%[[IV0]])[%[[D0]]] |
| // CHECK-DAG: %[[J:.+]] = affine.apply #[[MAP4]](%[[IV1]]) |
| // CHECK-DAG: %[[J_SZ:.+]] = affine.min #[[MAP5]](%[[IV1]])[%[[D1]]] |
| // CHECK: %[[SRC0_TILE:.+]] = memref.subview %[[SRC0]] |
| // CHECK-SAME: [%[[I]], %[[J]], 0] [%[[I_SZ]], %[[J_SZ]], %[[D2]]] |
| // CHECK: %[[SRC1_TILE:.+]] = memref.subview %[[SRC1]] |
| // CHECK-SAME: [%[[I]], %[[J]], 0] [%[[I_SZ]], %[[J_SZ]], %[[D2]]] |
| // CHECK: iree_linalg_ext.sort dimension(2) |
| // CHECK-SAME: outs(%[[SRC0_TILE]], %[[SRC1_TILE]] |
| // CHECK } |
| // CHECK: return |
