| 1 |
// RUN: mlir-opt -split-input-file -transform-interpreter -cse %s | FileCheck %s |
| 2 |
|
| 3 |
///---------------------------------------------------------------------------------------- |
| 4 |
/// ATM, all tests in this file require masking. As the support for masking is |
| 5 |
/// limited to depthwise convs, only that variant of convolutions is tested |
| 6 |
/// ATM. |
| 7 |
/// |
| 8 |
/// TODO1: Add more types of convolutions (transform.structured.vectorize |
| 9 |
/// usually doesn't require masking when vector sizes are not specified) |
| 10 |
/// TODO2: Add support for masking non-depthwise convs. |
| 11 |
///---------------------------------------------------------------------------------------- |
| 12 |
|
| 13 |
func.func @depthwise_conv1d_nwc_wc_1x8x3xi8_tensor(%input: tensor<1x8x?xi8>, |
| 14 |
%filter: tensor<1x?xi8>, |
| 15 |
%output: tensor<1x8x?xi8>) -> (tensor<1x8x?xi8>) { |
| 16 |
%res = linalg.depthwise_conv_1d_nwc_wc |
| 17 |
{dilations = dense<1> : vector<1xi64>, |
| 18 |
strides = dense<1> : vector<1xi64>} |
| 19 |
ins(%input, %filter : tensor<1x8x?xi8>, tensor<1x?xi8>) |
| 20 |
outs(%output : tensor<1x8x?xi8>) -> tensor<1x8x?xi8> |
| 21 |
return %res : tensor<1x8x?xi8> |
| 22 |
} |
| 23 |
|
| 24 |
module attributes {transform.with_named_sequence} { |
| 25 |
transform.named_sequence @__transform_main(%arg0: !transform.any_op {transform.readonly}) { |
| 26 |
%0 = transform.structured.match ops{["linalg.depthwise_conv_1d_nwc_wc"]} in %arg0 : (!transform.any_op) -> !transform.any_op |
| 27 |
transform.structured.vectorize %0 vector_sizes [1, 8, 4, 1] : !transform.any_op |
| 28 |
transform.yield |
| 29 |
} |
| 30 |
} |
| 31 |
|
| 32 |
// CHECK-LABEL: func.func @depthwise_conv1d_nwc_wc_1x8x3xi8_tensor( |
| 33 |
// CHECK-SAME: %[[INPUT:.*]]: tensor<1x8x?xi8>, |
| 34 |
// CHECK-SAME: %[[FILTER:.*]]: tensor<1x?xi8>, |
| 35 |
// CHECK-SAME: %[[OUTPUT:.*]]: tensor<1x8x?xi8>) -> tensor<1x8x?xi8> { |
| 36 |
|
| 37 |
// CHECK: %[[C1:.*]] = arith.constant 1 : index |
| 38 |
// CHECK: %[[C0:.*]] = arith.constant 0 : index |
| 39 |
// CHECK: %[[PAD:.*]] = arith.constant 0 : i8 |
| 40 |
|
| 41 |
/// Create a mask for the input tensor |
| 42 |
// CHECK: %[[C2:.*]] = arith.constant 2 : index |
| 43 |
// CHECK: %[[CH_DIM_IN:.*]] = tensor.dim %[[INPUT]], %[[C2]] : tensor<1x8x?xi8> |
| 44 |
// CHECK: %[[C8:.*]] = arith.constant 8 : index |
| 45 |
// CHECK: %[[MASK_IN:.*]] = vector.create_mask %[[C1]], %[[C8]], %[[CH_DIM_IN]] : vector<1x8x4xi1> |
| 46 |
/// Read the input tensor |
| 47 |
// CHECK: %[[VEC_IN:.*]] = vector.mask %[[MASK_IN]] { vector.transfer_read %[[INPUT]]{{\[}}%[[C0]], %[[C0]], %[[C0]]], %[[PAD]] {in_bounds = [true, true, true]} : tensor<1x8x?xi8>, vector<1x8x4xi8> } : vector<1x8x4xi1> -> vector<1x8x4xi8> |
| 48 |
|
| 49 |
/// Create a mask for the filter tensor |
| 50 |
// CHECK: %[[CH_DIM_FLT:.*]] = tensor.dim %[[FILTER]], %[[C1]] : tensor<1x?xi8> |
| 51 |
// CHECK: %[[MASK_FLT:.*]] = vector.create_mask %[[C1]], %[[CH_DIM_FLT]] : vector<1x4xi1> |
| 52 |
/// Read the filter tensor |
| 53 |
// CHECK: %[[VEC_FLT:.*]] = vector.mask %[[MASK_FLT]] { vector.transfer_read %[[FILTER]]{{\[}}%[[C0]], %[[C0]]], %[[PAD]] {in_bounds = [true, true]} : tensor<1x?xi8>, vector<1x4xi8> } : vector<1x4xi1> -> vector<1x4xi8> |
| 54 |
|
| 55 |
/// Create a mask for the output tensor |
| 56 |
// CHECK: %[[CH_DIM_OUT:.*]] = tensor.dim %[[OUTPUT]], %[[C2]] : tensor<1x8x?xi8> |
| 57 |
// CHECK: %[[MASK_OUT:.*]] = vector.create_mask %[[C1]], %[[C8]], %[[CH_DIM_OUT]] : vector<1x8x4xi1> |
| 58 |
// CHECK: %[[VEC_OUT:.*]] = vector.mask %[[MASK_OUT]] { vector.transfer_read %[[OUTPUT]]{{\[}}%[[C0]], %[[C0]], %[[C0]]], %[[PAD]] {in_bounds = [true, true, true]} : tensor<1x8x?xi8>, vector<1x8x4xi8> } : vector<1x8x4xi1> -> vector<1x8x4xi8> |
| 59 |
|
| 60 |
/// Convolution |
| 61 |
// CHECK: %[[IN_1:.*]] = vector.extract_strided_slice %[[VEC_IN]] {offsets = [0, 0, 0], sizes = [1, 8, 4], strides = [1, 1, 1]} : vector<1x8x4xi8> to vector<1x8x4xi8> |
| 62 |
// CHECK: %[[FLT_1:.*]] = vector.extract %[[VEC_FLT]][0] : vector<4xi8> from vector<1x4xi8> |
| 63 |
// CHECK: %[[OUT_1:.*]] = vector.extract_strided_slice %[[VEC_OUT]] {offsets = [0, 0, 0], sizes = [1, 8, 4], strides = [1, 1, 1]} : vector<1x8x4xi8> to vector<1x8x4xi8> |
| 64 |
// CHECK: %[[FLT_1_B:.*]] = vector.broadcast %[[FLT_1]] : vector<4xi8> to vector<1x8x4xi8> |
| 65 |
// CHECK: %[[MULI:.*]] = arith.muli %[[IN_1]], %[[FLT_1_B]] : vector<1x8x4xi8> |
| 66 |
// CHECK: %[[ADDI:.*]] = arith.addi %[[MULI]], %[[OUT_1]] : vector<1x8x4xi8> |
| 67 |
// CHECK: %[[OUT_INS:.*]] = vector.insert_strided_slice %[[ADDI]], %[[VEC_OUT]] {offsets = [0, 0, 0], strides = [1, 1, 1]} : vector<1x8x4xi8> into vector<1x8x4xi8> |
| 68 |
// CHECK: %[[OUT:.*]] = vector.mask %[[MASK_OUT]] { vector.transfer_write %[[OUT_INS]], %[[OUTPUT]]{{\[}}%[[C0]], %[[C0]], %[[C0]]] {in_bounds = [true, true, true]} : vector<1x8x4xi8>, tensor<1x8x?xi8> } : vector<1x8x4xi1> -> tensor<1x8x?xi8> |
| 69 |
// CHECK: return %[[OUT]] : tensor<1x8x?xi8> |
| 70 |
|
| 71 |
// ----- |
| 72 |
|
| 73 |
func.func @depthwise_conv1d_nwc_wc_1x8x3xi8_tensor_scalable( |
| 74 |
%input: tensor<1x8x?xi8>, |
| 75 |
%filter: tensor<1x?xi8>, |
| 76 |
%output: tensor<1x8x?xi8>) -> (tensor<1x8x?xi8>) { |
| 77 |
%res = linalg.depthwise_conv_1d_nwc_wc |
| 78 |
{dilations = dense<1> : vector<1xi64>, |
| 79 |
strides = dense<1> : vector<1xi64>} |
| 80 |
ins(%input, %filter : tensor<1x8x?xi8>, tensor<1x?xi8>) |
| 81 |
outs(%output : tensor<1x8x?xi8>) -> tensor<1x8x?xi8> |
| 82 |
return %res : tensor<1x8x?xi8> |
| 83 |
} |
| 84 |
|
| 85 |
module attributes {transform.with_named_sequence} { |
| 86 |
transform.named_sequence @__transform_main(%arg0: !transform.any_op {transform.readonly}) { |
| 87 |
%0 = transform.structured.match ops{["linalg.depthwise_conv_1d_nwc_wc"]} in %arg0 : (!transform.any_op) -> !transform.any_op |
| 88 |
transform.structured.vectorize %0 vector_sizes [1, 8, [4], 1] : !transform.any_op |
| 89 |
transform.yield |
| 90 |
} |
| 91 |
} |
| 92 |
|
| 93 |
// CHECK-LABEL: func.func @depthwise_conv1d_nwc_wc_1x8x3xi8_tensor_scalable( |
| 94 |
// CHECK-SAME: %[[INPUT:.*]]: tensor<1x8x?xi8>, |
| 95 |
// CHECK-SAME: %[[FILTER:.*]]: tensor<1x?xi8>, |
| 96 |
// CHECK-SAME: %[[OUTPUT:.*]]: tensor<1x8x?xi8>) -> tensor<1x8x?xi8> { |
| 97 |
|
| 98 |
// CHECK: %[[C1:.*]] = arith.constant 1 : index |
| 99 |
// CHECK: %[[C0:.*]] = arith.constant 0 : index |
| 100 |
// CHECK: %[[PAD:.*]] = arith.constant 0 : i8 |
| 101 |
|
| 102 |
/// Create a mask for the input tensor |
| 103 |
// CHECK: %[[C2:.*]] = arith.constant 2 : index |
| 104 |
// CHECK: %[[CH_DIM_IN:.*]] = tensor.dim %[[INPUT]], %[[C2]] : tensor<1x8x?xi8> |
| 105 |
// CHECK: %[[C8:.*]] = arith.constant 8 : index |
| 106 |
// CHECK: %[[MASK_IN:.*]] = vector.create_mask %[[C1]], %[[C8]], %[[CH_DIM_IN]] : vector<1x8x[4]xi1> |
| 107 |
/// Read the input tensor |
| 108 |
// CHECK: %[[VEC_IN:.*]] = vector.mask %[[MASK_IN]] { vector.transfer_read %[[INPUT]]{{\[}}%[[C0]], %[[C0]], %[[C0]]], %[[PAD]] {in_bounds = [true, true, true]} : tensor<1x8x?xi8>, vector<1x8x[4]xi8> } : vector<1x8x[4]xi1> -> vector<1x8x[4]xi8> |
| 109 |
|
| 110 |
/// Create a mask for the filter tensor |
| 111 |
// CHECK: %[[CH_DIM_FLT:.*]] = tensor.dim %[[FILTER]], %[[C1]] : tensor<1x?xi8> |
| 112 |
// CHECK: %[[MASK_FLT:.*]] = vector.create_mask %[[C1]], %[[CH_DIM_FLT]] : vector<1x[4]xi1> |
| 113 |
/// Read the filter tensor |
| 114 |
// CHECK: %[[VEC_FLT:.*]] = vector.mask %[[MASK_FLT]] { vector.transfer_read %[[FILTER]]{{\[}}%[[C0]], %[[C0]]], %[[PAD]] {in_bounds = [true, true]} : tensor<1x?xi8>, vector<1x[4]xi8> } : vector<1x[4]xi1> -> vector<1x[4]xi8> |
| 115 |
|
| 116 |
/// Create a mask for the output tensor |
| 117 |
// CHECK: %[[CH_DIM_OUT:.*]] = tensor.dim %[[OUTPUT]], %[[C2]] : tensor<1x8x?xi8> |
| 118 |
// CHECK: %[[MASK_OUT:.*]] = vector.create_mask %[[C1]], %[[C8]], %[[CH_DIM_OUT]] : vector<1x8x[4]xi1> |
| 119 |
/// Read the output tensor |
| 120 |
// CHECK: %[[VEC_OUT:.*]] = vector.mask %[[MASK_OUT]] { vector.transfer_read %[[OUTPUT]]{{\[}}%[[C0]], %[[C0]], %[[C0]]], %[[PAD]] {in_bounds = [true, true, true]} : tensor<1x8x?xi8>, vector<1x8x[4]xi8> } : vector<1x8x[4]xi1> -> vector<1x8x[4]xi8> |
| 121 |
|
| 122 |
/// Convolution |
| 123 |
// CHECK: %[[IN_1:.*]] = vector.extract_strided_slice %[[VEC_IN]] {offsets = [0, 0, 0], sizes = [1, 8, 4], strides = [1, 1, 1]} : vector<1x8x[4]xi8> to vector<1x8x[4]xi8> |
| 124 |
// CHECK: %[[FLT_1:.*]] = vector.extract %[[VEC_FLT]][0] : vector<[4]xi8> from vector<1x[4]xi8> |
| 125 |
// CHECK: %[[OUT_1:.*]] = vector.extract_strided_slice %[[VEC_OUT]] {offsets = [0, 0, 0], sizes = [1, 8, 4], strides = [1, 1, 1]} : vector<1x8x[4]xi8> to vector<1x8x[4]xi8> |
| 126 |
// CHECK: %[[FLT_1_B:.*]] = vector.broadcast %[[FLT_1]] : vector<[4]xi8> to vector<1x8x[4]xi8> |
| 127 |
// CHECK: %[[MULI:.*]] = arith.muli %[[IN_1]], %[[FLT_1_B]] : vector<1x8x[4]xi8> |
| 128 |
// CHECK: %[[ADDI:.*]] = arith.addi %[[MULI]], %[[OUT_1]] : vector<1x8x[4]xi8> |
| 129 |
// CHECK: %[[OUT_INS:.*]] = vector.insert_strided_slice %[[ADDI]], %[[VEC_OUT]] {offsets = [0, 0, 0], strides = [1, 1, 1]} : vector<1x8x[4]xi8> into vector<1x8x[4]xi8> |
| 130 |
// CHECK: %[[OUT:.*]] = vector.mask %[[MASK_OUT]] { vector.transfer_write %[[OUT_INS]], %[[OUTPUT]]{{\[}}%[[C0]], %[[C0]], %[[C0]]] {in_bounds = [true, true, true]} : vector<1x8x[4]xi8>, tensor<1x8x?xi8> } : vector<1x8x[4]xi1> -> tensor<1x8x?xi8> |
| 131 |
// CHECK: return %[[OUT]] : tensor<1x8x?xi8> |
| 132 |
|
| 133 |
// ----- |
| 134 |
|
| 135 |
func.func @depthwise_conv1d_nwc_wc_3x5x4xf32_memref_dilation_2( |
| 136 |
%input: memref<3x5x?xf32>, |
| 137 |
%filter: memref<2x?xf32>, |
| 138 |
%output: memref<3x2x?xf32>) { |
| 139 |
linalg.depthwise_conv_1d_nwc_wc |
| 140 |
{dilations = dense<2> : tensor<1xi64>, strides = dense<1> : tensor<1xi64>} |
| 141 |
ins(%input, %filter : memref<3x5x?xf32>, memref<2x?xf32>) |
| 142 |
outs(%output : memref<3x2x?xf32>) |
| 143 |
return |
| 144 |
} |
| 145 |
|
| 146 |
module attributes {transform.with_named_sequence} { |
| 147 |
transform.named_sequence @__transform_main(%arg0: !transform.any_op {transform.readonly}) { |
| 148 |
%0 = transform.structured.match ops{["linalg.depthwise_conv_1d_nwc_wc"]} in %arg0 : (!transform.any_op) -> !transform.any_op |
| 149 |
transform.structured.vectorize %0 vector_sizes [3, 2, [4], 2] : !transform.any_op |
| 150 |
transform.yield |
| 151 |
} |
| 152 |
} |
| 153 |
|
| 154 |
// CHECK-LABEL: func.func @depthwise_conv1d_nwc_wc_3x5x4xf32_memref_dilation_2( |
| 155 |
// CHECK-SAME: %[[INPUT:.*]]: memref<3x5x?xf32>, |
| 156 |
// CHECK-SAME: %[[FILTER:.*]]: memref<2x?xf32>, |
| 157 |
// CHECK-SAME: %[[OUTPUT:.*]]: memref<3x2x?xf32>) { |
| 158 |
|
| 159 |
// CHECK: %[[C1:.*]] = arith.constant 1 : index |
| 160 |
// CHECK: %[[C0:.*]] = arith.constant 0 : index |
| 161 |
// CHECK: %[[PAD:.*]] = arith.constant 0.000000e+00 : f32 |
| 162 |
// CHECK: %[[C2:.*]] = arith.constant 2 : index |
| 163 |
|
| 164 |
/// Create a mask for the input tensor |
| 165 |
// CHECK: %[[CH_DIM_IN:.*]] = memref.dim %[[INPUT]], %[[C2]] : memref<3x5x?xf32> |
| 166 |
// CHECK: %[[C3:.*]] = arith.constant 3 : index |
| 167 |
// CHECK: %[[C5:.*]] = arith.constant 5 : index |
| 168 |
// CHECK: %[[MASK_IN:.*]] = vector.create_mask %[[C3]], %[[C5]], %[[CH_DIM_IN]] : vector<3x4x[4]xi1> |
| 169 |
/// Read the input tensor |
| 170 |
// CHECK: %[[VEC_IN:.*]] = vector.mask %[[MASK_IN]] { vector.transfer_read %[[INPUT]]{{\[}}%[[C0]], %[[C0]], %[[C0]]], %[[PAD]] {in_bounds = [true, true, true]} : memref<3x5x?xf32>, vector<3x4x[4]xf32> } : vector<3x4x[4]xi1> -> vector<3x4x[4]xf32> |
| 171 |
|
| 172 |
/// Create a mask for the filter tensor |
| 173 |
// CHECK: %[[CH_DIM_FLT:.*]] = memref.dim %[[FILTER]], %[[C1]] : memref<2x?xf32> |
| 174 |
// CHECK: %[[MASK_FLT:.*]] = vector.create_mask %[[C2]], %[[CH_DIM_FLT]] : vector<2x[4]xi1> |
| 175 |
/// Read the filter tensor |
| 176 |
// CHECK: %[[VEC_FLT:.*]] = vector.mask %[[MASK_FLT]] { vector.transfer_read %[[FILTER]]{{\[}}%[[C0]], %[[C0]]], %[[PAD]] {in_bounds = [true, true]} : memref<2x?xf32>, vector<2x[4]xf32> } : vector<2x[4]xi1> -> vector<2x[4]xf32> |
| 177 |
|
| 178 |
/// Create a mask for the output tensor |
| 179 |
// CHECK: %[[CH_DIM_OUT:.*]] = memref.dim %[[OUTPUT]], %[[C2]] : memref<3x2x?xf32> |
| 180 |
// CHECK: %[[MASK_OUT:.*]] = vector.create_mask %[[C3]], %[[C2]], %[[CH_DIM_OUT]] : vector<3x2x[4]xi1> |
| 181 |
/// Read the output tensor |
| 182 |
// CHECK: %[[VEC_OUT:.*]] = vector.mask %[[MASK_OUT]] { vector.transfer_read %[[OUTPUT]]{{\[}}%[[C0]], %[[C0]], %[[C0]]], %[[PAD]] {in_bounds = [true, true, true]} : memref<3x2x?xf32>, vector<3x2x[4]xf32> } : vector<3x2x[4]xi1> -> vector<3x2x[4]xf32> |
| 183 |
|
| 184 |
/// Convolution |
| 185 |
// CHECK: %[[IN_1:.*]] = vector.extract_strided_slice %[[VEC_IN]] {offsets = [0, 0, 0], sizes = [3, 2, 4], strides = [1, 1, 1]} : vector<3x4x[4]xf32> to vector<3x2x[4]xf32> |
| 186 |
// CHECK: %[[IN_2:.*]] = vector.extract_strided_slice %[[VEC_IN]] {offsets = [0, 2, 0], sizes = [3, 2, 4], strides = [1, 1, 1]} : vector<3x4x[4]xf32> to vector<3x2x[4]xf32> |
| 187 |
// CHECK: %[[FLT_1:.*]] = vector.extract %[[VEC_FLT]][0] : vector<[4]xf32> from vector<2x[4]xf32> |
| 188 |
// CHECK: %[[FLT_2:.*]] = vector.extract %[[VEC_FLT]][1] : vector<[4]xf32> from vector<2x[4]xf32> |
| 189 |
// CHECK: %[[OUT_1:.*]] = vector.extract_strided_slice %[[VEC_OUT]] {offsets = [0, 0, 0], sizes = [3, 2, 4], strides = [1, 1, 1]} : vector<3x2x[4]xf32> to vector<3x2x[4]xf32> |
| 190 |
// CHECK: %[[FLT_1_B:.*]] = vector.broadcast %[[FLT_1]] : vector<[4]xf32> to vector<3x2x[4]xf32> |
| 191 |
// CHECK: %[[FMA_1:.*]] = vector.fma %[[IN_1]], %[[FLT_1_B]], %[[OUT_1]] : vector<3x2x[4]xf32> |
| 192 |
// CHECK: %[[FLT_2_B:.*]] = vector.broadcast %[[FLT_2]] : vector<[4]xf32> to vector<3x2x[4]xf32> |
| 193 |
// CHECK: %[[FMA_2:.*]] = vector.fma %[[IN_2]], %[[FLT_2_B]], %[[FMA_1]] : vector<3x2x[4]xf32> |
| 194 |
// CHECK: %[[OUT_INS:.*]] = vector.insert_strided_slice %[[FMA_2]], %[[VEC_OUT]] {offsets = [0, 0, 0], strides = [1, 1, 1]} : vector<3x2x[4]xf32> into vector<3x2x[4]xf32> |
| 195 |
// CHECK: vector.mask %[[MASK_OUT]] { vector.transfer_write %[[OUT_INS]], %[[OUTPUT]]{{\[}}%[[C0]], %[[C0]], %[[C0]]] {in_bounds = [true, true, true]} : vector<3x2x[4]xf32>, memref<3x2x?xf32> } : vector<3x2x[4]xi1> |