// RUN: mlir-opt %s -test-linalg-elementwise-fusion-patterns=collapse-dimensions-control=2,3 -split-input-file | FileCheck %s func.func @collapse_reduction( %arg0: tensor<2x32x10x4096xf32>, %arg1: tensor<2x32xf32>) -> tensor<2x32xf32> { %0 = linalg.generic { indexing_maps = [ affine_map<(d0, d1, d2, d3) -> (d0, d1, d2, d3)>, affine_map<(d0, d1, d2, d3) -> (d0, d1)>], iterator_types = ["parallel", "parallel", "reduction", "reduction"]} ins(%arg0 : tensor<2x32x10x4096xf32>) outs(%arg1 : tensor<2x32xf32>) { ^bb0(%arg3: f32, %arg4: f32): %1 = arith.addf %arg3, %arg4 : f32 linalg.yield %1 : f32 } -> tensor<2x32xf32> return %0 : tensor<2x32xf32> } // CHECK-DAG: #[[$MAP0:.*]] = affine_map<(d0, d1, d2) -> (d0, d1, d2)> // CHECK-DAG: #[[$MAP1:.*]] = affine_map<(d0, d1, d2) -> (d0, d1)> // CHECK-LABEL: func @collapse_reduction // CHECK: %[[T:.*]] = tensor.collapse_shape %{{.*}} {{\[}}[0], [1], [2, 3]] : tensor<2x32x10x4096xf32> into tensor<2x32x40960xf32> // CHECK: linalg.generic {indexing_maps = [#[[$MAP0]], #[[$MAP1]]], // CHECK-SAME: iterator_types = ["parallel", "parallel", "reduction"]} // CHECK-SAME: ins(%[[T]] : tensor<2x32x40960xf32>) outs(%{{.*}} : tensor<2x32xf32>) { // CHECK: } -> tensor<2x32xf32> // ----- func.func @collapse_parallel( %arg0: tensor<32x2x10x4096xf32>, %arg1: tensor<2x32x10x4096xf32>) -> tensor<2x32x10x4096xf32> { %0 = linalg.generic { indexing_maps = [ affine_map<(d0, d1, d2, d3) -> (d1, d0, d2, d3)>, affine_map<(d0, d1, d2, d3) -> (d0, d1, d2, d3)>], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%arg0 : tensor<32x2x10x4096xf32>) outs(%arg1 : tensor<2x32x10x4096xf32>) { ^bb0(%arg3: f32, %arg4: f32): %1 = arith.addf %arg3, %arg4 : f32 linalg.yield %1 : f32 } -> tensor<2x32x10x4096xf32> return %0 : tensor<2x32x10x4096xf32> } // CHECK-DAG: #[[$MAP0:.*]] = affine_map<(d0, d1, d2) -> (d1, d0, d2)> // CHECK-DAG: #[[$MAP1:.*]] = affine_map<(d0, d1, d2) -> (d0, d1, d2)> // CHECK-LABEL: func @collapse_parallel // CHECK-DAG: %[[S:.*]] = tensor.collapse_shape %{{.*}} {{\[}}[0], [1], [2, 3]] : tensor<32x2x10x4096xf32> into tensor<32x2x40960xf32> // CHECK-DAG: %[[D:.*]] = tensor.collapse_shape %{{.*}} {{\[}}[0], [1], [2, 3]] : tensor<2x32x10x4096xf32> into tensor<2x32x40960xf32> // CHECK: %[[R:.*]] = linalg.generic {indexing_maps = [#[[$MAP0]], #[[$MAP1]]], // CHECK-SAME: iterator_types = ["parallel", "parallel", "parallel"]} // CHECK-SAME: ins(%[[S]] : tensor<32x2x40960xf32>) outs(%[[D]] : tensor<2x32x40960xf32>) { // CHECK: } -> tensor<2x32x40960xf32> // CHECK: tensor.expand_shape %[[R]] {{\[}}[0], [1], [2, 3]] output_shape [2, 32, 10, 4096] : tensor<2x32x40960xf32> into tensor<2x32x10x4096xf32> // ----- #map = affine_map<(d0, d1, d2, d3) -> (d3, d0, d1, d2)> #map1 = affine_map<(d0, d1, d2, d3) -> (d0, d1, d2, d3)> func.func @uncollapsable(%arg0 : tensor<41x3x1x57xf32>, %arg1 : tensor<3x1x57x41xf32>) -> tensor<3x1x57x41xf32> { %0 = linalg.generic { indexing_maps = [#map, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%arg0 : tensor<41x3x1x57xf32>) outs(%arg1 : tensor<3x1x57x41xf32>) { ^bb0(%in: f32, %out: f32): linalg.yield %in : f32 } -> tensor<3x1x57x41xf32> return %0 : tensor<3x1x57x41xf32> } // CHECK-LABEL: func @uncollapsable( // CHECK: linalg.generic // CHECK-SAME: iterator_types = ["parallel", "parallel", "parallel", "parallel"] // ----- // CHECK-LABEL: func.func private @collapsable_memref( // CHECK-SAME: %[[VAL_0:.*]]: memref<1x24x32x8xf32>, // CHECK-SAME: %[[VAL_1:.*]]: memref<1x24x32x8xf32>) -> memref<1x24x32x8xf32> { // CHECK: %[[VAL_2:.*]] = memref.alloc() {alignment = 64 : i64} : memref<1x24x32x8xf32> // CHECK: %[[VAL_3:.*]] = memref.collapse_shape %[[VAL_0]] {{\[\[}}0], [1], [2, 3]] : memref<1x24x32x8xf32> into memref<1x24x256xf32> // CHECK: %[[VAL_4:.*]] = memref.collapse_shape %[[VAL_1]] {{\[\[}}0], [1], [2, 3]] : memref<1x24x32x8xf32> into memref<1x24x256xf32> // CHECK: %[[VAL_5:.*]] = memref.collapse_shape %[[VAL_2]] {{\[\[}}0], [1], [2, 3]] : memref<1x24x32x8xf32> into memref<1x24x256xf32> // CHECK: linalg.generic {indexing_maps = [#map, #map, #map], iterator_types = ["parallel", "parallel", "parallel"]} ins(%[[VAL_3]], %[[VAL_4]] : memref<1x24x256xf32>, memref<1x24x256xf32>) outs(%[[VAL_5]] : memref<1x24x256xf32>) { // CHECK: ^bb0(%[[VAL_6:.*]]: f32, %[[VAL_7:.*]]: f32, %[[VAL_8:.*]]: f32): // CHECK: %[[VAL_9:.*]] = arith.addf %[[VAL_6]], %[[VAL_7]] : f32 // CHECK: linalg.yield %[[VAL_9]] : f32 // CHECK: } // CHECK: return %[[VAL_2]] : memref<1x24x32x8xf32> // CHECK: } func.func private @collapsable_memref(%arg0: memref<1x24x32x8xf32>, %arg1: memref<1x24x32x8xf32>) -> (memref<1x24x32x8xf32>) { %alloc = memref.alloc() {alignment = 64 : i64} : memref<1x24x32x8xf32> linalg.generic {indexing_maps = [affine_map<(d0, d1, d2, d3) -> (d0, d1, d2, d3)>, affine_map<(d0, d1, d2, d3) -> (d0, d1, d2, d3)>, affine_map<(d0, d1, d2, d3) -> (d0, d1, d2, d3)>], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%arg0, %arg1 : memref<1x24x32x8xf32>, memref<1x24x32x8xf32>) outs(%alloc : memref<1x24x32x8xf32>) { ^bb0(%in: f32, %in_0: f32, %out: f32): %0 = arith.addf %in, %in_0 : f32 linalg.yield %0 : f32 } return %alloc : memref<1x24x32x8xf32> } // ----- // CHECK-DAG: #[[$ATTR_0:.+]] = affine_map<(d0, d1, d2, d3) -> (d0 * 7680 + d1 * 320 + d2 * 10 + d3)> // CHECK-DAG: #[[$ATTR_1:.+]] = affine_map<(d0, d1, d2) -> (d0, d2, d1)> // CHECK-LABEL: func.func @collapsable_memref_projected_ops( // CHECK-SAME: %[[ARG0:.*]]: memref<1x24x32x8xf32>, %[[ARG1:.*]]: memref<1x24x32x8xf32>, %[[ARG2:.*]]: memref<1x24x32x8xf32, #[[$ATTR_0]]>) { // CHECK: %[[VAL_0:.*]] = memref.collapse_shape %[[ARG0]] {{\[\[}}0], [1, 2], [3]] : memref<1x24x32x8xf32> into memref<1x768x8xf32> // CHECK: %[[VAL_1:.*]] = memref.collapse_shape %[[ARG1]] {{\[\[}}0], [1, 2], [3]] : memref<1x24x32x8xf32> into memref<1x768x8xf32> // CHECK: %[[VAL_2:.*]] = memref.collapse_shape %[[ARG2]] {{\[\[}}0], [1, 2], [3]] : memref<1x24x32x8xf32, #[[$ATTR_0]]> into memref<1x768x8xf32, strided<[7680, 10, 1]>> // CHECK: linalg.generic {indexing_maps = [#[[$ATTR_1]], #[[$ATTR_1]], #[[$ATTR_1]]], iterator_types = ["parallel", "parallel", "parallel"]} ins(%[[VAL_0]], %[[VAL_1]] : memref<1x768x8xf32>, memref<1x768x8xf32>) outs(%[[VAL_2]] : memref<1x768x8xf32, strided<[7680, 10, 1]>>) { // CHECK: ^bb0(%[[VAL_3:.*]]: f32, %[[VAL_4:.*]]: f32, %[[VAL_5:.*]]: f32): // CHECK: %[[VAL_6:.*]] = arith.addf %[[VAL_3]], %[[VAL_4]] : f32 // CHECK: linalg.yield %[[VAL_6]] : f32 // CHECK: } // CHECK: return // CHECK: } #map = affine_map<(d0, d1, d2, d3) -> (d0, d2, d3, d1)> #map1 = affine_map<(d0, d1, d2, d3) -> (d0 * 7680 + d1 * 320 + d2 * 10 + d3)> func.func @collapsable_memref_projected_ops(%arg0: memref<1x24x32x8xf32>, %arg1: memref<1x24x32x8xf32>, %arg2: memref<1x24x32x8xf32, #map1>) { linalg.generic {indexing_maps = [#map, #map, #map], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%arg0, %arg1 : memref<1x24x32x8xf32>, memref<1x24x32x8xf32>) outs(%arg2 : memref<1x24x32x8xf32, #map1>) { ^bb0(%in: f32, %in_0: f32, %out: f32): %0 = arith.addf %in, %in_0 : f32 linalg.yield %0 : f32 } return } // ----- // CHECK-LABEL: func @uncollapsable_strided_memref( // CHECK: linalg.generic // CHECK-SAME: iterator_types = ["parallel", "parallel", "parallel", "parallel"] func.func @uncollapsable_strided_memref(%arg0: memref<2x6x24x48xi32>, %arg1: memref<2x6x24x48xi32>) -> (memref<2x6x24x48xi32>) { %alloc = memref.alloc() {alignment = 64 : i64} : memref<2x6x24x48xi32> %subview = memref.subview %arg0[0, 0, 0, 0] [1, 3, 12, 24] [1, 1, 1, 1] : memref<2x6x24x48xi32> to memref<1x3x12x24xi32, strided<[6912, 1152, 48, 1], offset: 0>> %subview0 = memref.subview %arg1[0, 0, 0, 0] [1, 3, 12, 24] [1, 1, 1, 1] : memref<2x6x24x48xi32> to memref<1x3x12x24xi32, strided<[6912, 1152, 48, 1], offset: 0>> %subview1 = memref.subview %alloc[0, 0, 0, 0] [1, 3, 12, 24] [1, 1, 1, 1] : memref<2x6x24x48xi32> to memref<1x3x12x24xi32, strided<[6912, 1152, 48, 1], offset: 0>> linalg.generic {indexing_maps = [affine_map<(d0, d1, d2, d3) -> (d0, d1, d2, d3)>, affine_map<(d0, d1, d2, d3) -> (d0, d1, d2, d3)>, affine_map<(d0, d1, d2, d3) -> (d0, d1, d2, d3)>], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%subview, %subview0 : memref<1x3x12x24xi32, strided<[6912, 1152, 48, 1], offset: 0>>, memref<1x3x12x24xi32, strided<[6912, 1152, 48, 1], offset: 0>>) outs(%subview1 : memref<1x3x12x24xi32, strided<[6912, 1152, 48, 1], offset: 0>>) { ^bb0(%in: i32, %in_0: i32, %out: i32): %0 = arith.addi %in, %in_0 : i32 linalg.yield %0 : i32 } return %alloc : memref<2x6x24x48xi32> } // ----- // CHECK-LABEL: func @uncollapsable_memref_projected_ops( // CHECK: linalg.generic // CHECK-SAME: iterator_types = ["parallel", "parallel", "parallel", "parallel"] #map = affine_map<(d0, d1, d2, d3) -> (d0, d2, d3, d1)> #map1 = affine_map<(d0, d1, d2, d3) -> (d0 * 7680 + d1 * 320 + d2 * 8 + d3)> func.func @uncollapsable_memref_projected_ops(%arg0: memref<1x24x32x8xf32>, %arg1: memref<1x24x32x8xf32>, %arg2: memref<1x24x32x8xf32, #map1>) { linalg.generic {indexing_maps = [#map, #map, #map], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%arg0, %arg1 : memref<1x24x32x8xf32>, memref<1x24x32x8xf32>) outs(%arg2 : memref<1x24x32x8xf32, #map1>) { ^bb0(%in: f32, %in_0: f32, %out: f32): %0 = arith.addf %in, %in_0 : f32 linalg.yield %0 : f32 } return } // ----- // CHECK-LABEL: func.func @linalg_copy( // CHECK-SAME: %[[VAL_0:.*]]: tensor<1x2x3x4x5xf32, 1 : i64>, // CHECK-SAME: %[[VAL_1:.*]]: tensor<1x2x3x4x5xf32, 3 : i64>) -> tensor<1x2x3x4x5xf32, 3 : i64> { // CHECK: %[[VAL_2:.*]] = tensor.collapse_shape %[[VAL_0]] {{\[\[}}0], [1], [2, 3], [4]] : tensor<1x2x3x4x5xf32, 1 : i64> into tensor<1x2x12x5xf32, 1 : i64> // CHECK: %[[VAL_3:.*]] = tensor.collapse_shape %[[VAL_1]] {{\[\[}}0], [1], [2, 3], [4]] : tensor<1x2x3x4x5xf32, 3 : i64> into tensor<1x2x12x5xf32, 3 : i64> // CHECK: %[[VAL_4:.*]] = tensor.collapse_shape %[[VAL_2]] {{\[\[}}0], [1], [2, 3]] : tensor<1x2x12x5xf32, 1 : i64> into tensor<1x2x60xf32, 1 : i64> // CHECK: %[[VAL_5:.*]] = tensor.collapse_shape %[[VAL_3]] {{\[\[}}0], [1], [2, 3]] : tensor<1x2x12x5xf32, 3 : i64> into tensor<1x2x60xf32, 3 : i64> // CHECK: %[[VAL_6:.*]] = linalg.copy ins(%[[VAL_4]] : tensor<1x2x60xf32, 1 : i64>) outs(%[[VAL_5]] : tensor<1x2x60xf32, 3 : i64>) -> tensor<1x2x60xf32, 3 : i64> // CHECK: %[[VAL_7:.*]] = tensor.expand_shape %[[VAL_6]] {{\[\[}}0], [1], [2, 3]] output_shape [1, 2, 12, 5] : tensor<1x2x60xf32, 3 : i64> into tensor<1x2x12x5xf32, 3 : i64> // CHECK: %[[VAL_8:.*]] = tensor.expand_shape %[[VAL_7]] {{\[\[}}0], [1], [2, 3], [4]] output_shape [1, 2, 3, 4, 5] : tensor<1x2x12x5xf32, 3 : i64> into tensor<1x2x3x4x5xf32, 3 : i64> // CHECK: return %[[VAL_8]] : tensor<1x2x3x4x5xf32, 3 : i64> // CHECK: } func.func @linalg_copy( %arg0: tensor<1x2x3x4x5xf32, 1>, %arg1: tensor<1x2x3x4x5xf32, 3>) -> tensor<1x2x3x4x5xf32, 3> { %0 = linalg.copy ins(%arg0: tensor<1x2x3x4x5xf32, 1>) outs(%arg1: tensor<1x2x3x4x5xf32, 3>) -> tensor<1x2x3x4x5xf32, 3> return %0 : tensor<1x2x3x4x5xf32, 3> } // ----- // CHECK-LABEL: func.func private @memref_linalg_copy( // CHECK-SAME: %[[VAL_0:.*]]: memref<1x24x32x8xf32, 1>, // CHECK-SAME: %[[VAL_1:.*]]: memref<1x24x32x8xf32, 1>) { // CHECK: %[[VAL_2:.*]] = memref.collapse_shape %[[VAL_0]] {{\[\[}}0], [1], [2, 3]] : memref<1x24x32x8xf32, 1> into memref<1x24x256xf32, 1> // CHECK: %[[VAL_3:.*]] = memref.collapse_shape %[[VAL_1]] {{\[\[}}0], [1], [2, 3]] : memref<1x24x32x8xf32, 1> into memref<1x24x256xf32, 1> // CHECK: linalg.copy ins(%[[VAL_2]] : memref<1x24x256xf32, 1>) outs(%[[VAL_3]] : memref<1x24x256xf32, 1>) // CHECK: return // CHECK: } func.func private @memref_linalg_copy(%arg0: memref<1x24x32x8xf32, 1>, %arg1: memref<1x24x32x8xf32, 1>) { linalg.copy ins(%arg0: memref<1x24x32x8xf32, 1>) outs(%arg1: memref<1x24x32x8xf32, 1>) return } // ----- // Test that non-projected-permutation indexing maps don't crash the collapse pass. // The indexing map has affine expressions like d1 + d4, which are not projected // permutations. The pass should skip collapsing and leave the op unchanged. // CHECK-DAG: #[[$MAP:.*]] = affine_map<(d0, d1, d2, d3, d4, d5) -> (d0, d1 + d4, d2 + d5, d3)> // CHECK-DAG: #[[$MAP1:.*]] = affine_map<(d0, d1, d2, d3, d4, d5) -> (d4, d5, d3, 0)> // CHECK-DAG: #[[$MAP2:.*]] = affine_map<(d0, d1, d2, d3, d4, d5) -> (d0, d1, d2, d3)> // CHECK-LABEL: func @non_projected_permutation_no_crash // CHECK: linalg.generic {indexing_maps = [#[[$MAP]], #[[$MAP1]], #[[$MAP2]]], // CHECK-SAME: iterator_types = ["parallel", "parallel", "parallel", "parallel", "reduction", "reduction"] func.func @non_projected_permutation_no_crash( %arg0: tensor<1x5x5x1xf32>, %arg1: tensor<2x2x1x1xf32>) -> tensor<1x4x4x1xf32> { %cst = arith.constant dense<0.0> : tensor<1x4x4x1xf32> %0 = linalg.generic { indexing_maps = [affine_map<(d0, d1, d2, d3, d4, d5) -> (d0, d1 + d4, d2 + d5, d3)>, affine_map<(d0, d1, d2, d3, d4, d5) -> (d4, d5, d3, 0)>, affine_map<(d0, d1, d2, d3, d4, d5) -> (d0, d1, d2, d3)>], iterator_types = ["parallel", "parallel", "parallel", "parallel", "reduction", "reduction"] } ins(%arg0, %arg1 : tensor<1x5x5x1xf32>, tensor<2x2x1x1xf32>) outs(%cst : tensor<1x4x4x1xf32>) { ^bb0(%in: f32, %w: f32, %out: f32): %mul = arith.mulf %in, %w : f32 %add = arith.addf %out, %mul : f32 linalg.yield %add : f32 } -> tensor<1x4x4x1xf32> return %0 : tensor<1x4x4x1xf32> }