colin
/
onnx-mlir
1
0
Fork 0
Code Issues Pull Requests Packages Projects Releases Wiki Activity

Add support for dilations attribute and add tests.

This commit is contained in:
Doru Bercea 2020-01-22 14:40:10 -05:00
parent de77758faf
commit ea45cbcca9
2 changed files with 144 additions and 5 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

27
src/dialect/onnx/onnx_ops.cpp
View File

@ -520,6 +520,26 @@ void ONNXConvNoBiasOp::inferShapes() {
kernelDims.emplace_back(weightShape[i + 2]);
}
// Check if dilations attribute is present.
// If it is then compute new kernel size that includes the receptive field.
// In this calculation we assume that the receptive field pixels must all be
// within the bounds of the image. In this case the new kernel size is given
// by:
//
// ( K + 1 ) * d - 1
// where K is a kernel dimension and d is the dilation along that axis.
//
// From a dimensionality perspective the kernel size becomes the dilated
// kernel size.
if (auto dilations = getAttrOfType<ArrayAttr>(
ONNXConvOp::getDilationsAttrName())) {
if (dilations.getValue().size() != nDims)
emitError("dilations length incompatible with spatial dimensions.");
for (int i = 0; i < nDims; ++i)
kernelDims[i] = (kernelDims[i] + 1) *
(dilations.getValue()[i]).cast<IntegerAttr>().getInt() + 1;
}
// Subtract kernel dimensions from input data dimensions.
for (int i = 0; i < nDims; ++i)
outSpatialDims[i] -= kernelDims[i];
@ -545,13 +565,14 @@ void ONNXConvNoBiasOp::inferShapes() {
}
} else if (autoPad == "SAME_UPPER" || autoPad == "SAME_LOWER") {
// Pad input so that output size matches input size.
// Each spatial dimension needs to be padded by:
// Each spatial dimension needs to be padded by a total of:
//
// ( K - 1 ) / 2
// K - 1
//
// where K is a kernel spatial dimension.
// Pad as if stride is 1.
for (int i = 0; i < nDims; ++i)
outSpatialDims[i] += floor((kernelDims[i] - 1) / 2);
outSpatialDims[i] += kernelDims[i] - 1;
} else if (autoPad == "VALID") {
// No padding
} else {

120
test/mlir/onnx/onnx_shape_inference.mlir
View File

@ -1,7 +1,10 @@
// RUN: onnf-opt --shape-inference %s -split-input-file | FileCheck %s
//===----------------------------------------------------------------------===//
/// Test the default behavior of transpose when no information for the
/// permutation of the axes is provided.
/// permutation of the axes is provided and when a permutation is provided.
//===----------------------------------------------------------------------===//
func @test_default_transpose(%arg0 : tensor<5x5x1x32xf32>) -> tensor<*xf32> {
%0 = "onnx.Transpose"(%arg0) : (tensor<5x5x1x32xf32>) -> tensor<*xf32>
"std.return"(%0) : (tensor<*xf32>) -> ()
@ -12,6 +15,7 @@ func @test_default_transpose(%arg0 : tensor<5x5x1x32xf32>) -> tensor<*xf32> {
// CHECK: return [[RES]] : tensor<32x1x5x5xf32>
/// Test shape inference for transposition when perm attribute is specified.
func @test_transpose(%arg0 : tensor<5x5x1x32xf32>) -> tensor<*xf32> {
%0 = "onnx.Transpose"(%arg0) {perm = [2, 0, 3, 1]} : (tensor<5x5x1x32xf32>) -> tensor<*xf32>
"std.return"(%0) : (tensor<*xf32>) -> ()
@ -20,3 +24,117 @@ func @test_transpose(%arg0 : tensor<5x5x1x32xf32>) -> tensor<*xf32> {
// CHECK-LABEL: test_transpose
// CHECK: [[RES_ATTR:%.+]] = "onnx.Transpose"(%arg0) {perm = [2, 0, 3, 1]} : (tensor<5x5x1x32xf32>) -> tensor<1x5x32x5xf32>
// CHECK: return [[RES_ATTR]] : tensor<1x5x32x5xf32>
//===----------------------------------------------------------------------===//
/// Test shape inference for ConvNoBias operation and all its attributes.
//===----------------------------------------------------------------------===//
/// Default and required attributes.
func @test_conv_no_bias_1(%arg0 : tensor<1x2x32x64xf32>, %arg1 : tensor<5x2x6x7xf32>) -> tensor<*xf32> {
%0 = "onnx.ConvNoBias"(%arg0, %arg1) {auto_pad = "NOTSET", group = 1 : i32} : (tensor<1x2x32x64xf32>, tensor<5x2x6x7xf32>) -> tensor<*xf32>
"std.return"(%0) : (tensor<*xf32>) -> ()
}
// CHECK-LABEL: test_conv_no_bias_1
// CHECK: [[RES_ATTR:%.+]] = "onnx.ConvNoBias"(%arg0, %arg1) {auto_pad = "NOTSET", group = 1 : i32} : (tensor<1x2x32x64xf32>, tensor<5x2x6x7xf32>) -> tensor<1x5x27x58xf32>
// CHECK: return [[RES_ATTR]] : tensor<1x5x27x58xf32>
/// kernel_shape attribute.
func @test_conv_no_bias_2(%arg0 : tensor<1x2x32x64xf32>, %arg1 : tensor<5x2x6x7xf32>) -> tensor<*xf32> {
%0 = "onnx.ConvNoBias"(%arg0, %arg1) {auto_pad = "NOTSET", group = 1 : i32, kernel_shape = [8, 9]} : (tensor<1x2x32x64xf32>, tensor<5x2x6x7xf32>) -> tensor<*xf32>
"std.return"(%0) : (tensor<*xf32>) -> ()
}
// CHECK-LABEL: test_conv_no_bias_2
// CHECK: [[RES_ATTR:%.+]] = "onnx.ConvNoBias"(%arg0, %arg1) {auto_pad = "NOTSET", group = 1 : i32, kernel_shape = [8, 9]} : (tensor<1x2x32x64xf32>, tensor<5x2x6x7xf32>) -> tensor<1x5x25x56xf32>
// CHECK: return [[RES_ATTR]] : tensor<1x5x25x56xf32>
/// pads attribute.
/// Use pads to make output size equal to input size by adding K - 1 to the result.
func @test_conv_no_bias_3(%arg0 : tensor<1x2x32x64xf32>, %arg1 : tensor<5x2x6x10xf32>) -> tensor<*xf32> {
%0 = "onnx.ConvNoBias"(%arg0, %arg1) {auto_pad = "NOTSET", group = 1 : i32, pads = [2, 4, 3, 5]} : (tensor<1x2x32x64xf32>, tensor<5x2x6x10xf32>) -> tensor<*xf32>
"std.return"(%0) : (tensor<*xf32>) -> ()
}
// CHECK-LABEL: test_conv_no_bias_3
// CHECK: [[RES_ATTR:%.+]] = "onnx.ConvNoBias"(%arg0, %arg1) {auto_pad = "NOTSET", group = 1 : i32, pads = [2, 4, 3, 5]} : (tensor<1x2x32x64xf32>, tensor<5x2x6x10xf32>) -> tensor<1x5x32x64xf32>
// CHECK: return [[RES_ATTR]] : tensor<1x5x32x64xf32>
/// auto_pad set to SAME_UPPER and SAME_LOWER.
func @test_conv_no_bias_4(%arg0 : tensor<1x2x32x64xf32>, %arg1 : tensor<5x2x6x10xf32>) -> tensor<*xf32> {
%0 = "onnx.ConvNoBias"(%arg0, %arg1) {auto_pad = "SAME_UPPER", group = 1 : i32} : (tensor<1x2x32x64xf32>, tensor<5x2x6x10xf32>) -> tensor<*xf32>
"std.return"(%0) : (tensor<*xf32>) -> ()
}
// CHECK-LABEL: test_conv_no_bias_4
// CHECK: [[RES_ATTR:%.+]] = "onnx.ConvNoBias"(%arg0, %arg1) {auto_pad = "SAME_UPPER", group = 1 : i32} : (tensor<1x2x32x64xf32>, tensor<5x2x6x10xf32>) -> tensor<1x5x32x64xf32>
// CHECK: return [[RES_ATTR]] : tensor<1x5x32x64xf32>
func @test_conv_no_bias_5(%arg0 : tensor<1x2x32x64xf32>, %arg1 : tensor<5x2x6x10xf32>) -> tensor<*xf32> {
%0 = "onnx.ConvNoBias"(%arg0, %arg1) {auto_pad = "SAME_LOWER", group = 1 : i32} : (tensor<1x2x32x64xf32>, tensor<5x2x6x10xf32>) -> tensor<*xf32>
"std.return"(%0) : (tensor<*xf32>) -> ()
}
// CHECK-LABEL: test_conv_no_bias_5
// CHECK: [[RES_ATTR:%.+]] = "onnx.ConvNoBias"(%arg0, %arg1) {auto_pad = "SAME_LOWER", group = 1 : i32} : (tensor<1x2x32x64xf32>, tensor<5x2x6x10xf32>) -> tensor<1x5x32x64xf32>
// CHECK: return [[RES_ATTR]] : tensor<1x5x32x64xf32>
/// auto_pad set to VALID.
func @test_conv_no_bias_6(%arg0 : tensor<1x2x32x64xf32>, %arg1 : tensor<5x2x6x10xf32>) -> tensor<*xf32> {
%0 = "onnx.ConvNoBias"(%arg0, %arg1) {auto_pad = "VALID", group = 1 : i32} : (tensor<1x2x32x64xf32>, tensor<5x2x6x10xf32>) -> tensor<*xf32>
"std.return"(%0) : (tensor<*xf32>) -> ()
}
// CHECK-LABEL: test_conv_no_bias_6
// CHECK: [[RES_ATTR:%.+]] = "onnx.ConvNoBias"(%arg0, %arg1) {auto_pad = "VALID", group = 1 : i32} : (tensor<1x2x32x64xf32>, tensor<5x2x6x10xf32>) -> tensor<1x5x27x55xf32>
// CHECK: return [[RES_ATTR]] : tensor<1x5x27x55xf32>
/// With strides attribute.
func @test_conv_no_bias_7(%arg0 : tensor<1x2x32x64xf32>, %arg1 : tensor<5x2x6x7xf32>) -> tensor<*xf32> {
%0 = "onnx.ConvNoBias"(%arg0, %arg1) {auto_pad = "NOTSET", group = 1 : i32, strides = [2, 3]} : (tensor<1x2x32x64xf32>, tensor<5x2x6x7xf32>) -> tensor<*xf32>
"std.return"(%0) : (tensor<*xf32>) -> ()
}
// CHECK-LABEL: test_conv_no_bias_7
// CHECK: [[RES_ATTR:%.+]] = "onnx.ConvNoBias"(%arg0, %arg1) {auto_pad = "NOTSET", group = 1 : i32, strides = [2, 3]} : (tensor<1x2x32x64xf32>, tensor<5x2x6x7xf32>) -> tensor<1x5x14x20xf32>
// CHECK: return [[RES_ATTR]] : tensor<1x5x14x20xf32>
/// auto_pad set to SAME_UPPER with strides attribute.
/// The auto_pad will pas as if stride is equal to 1.
func @test_conv_no_bias_8(%arg0 : tensor<1x2x32x64xf32>, %arg1 : tensor<5x2x6x7xf32>) -> tensor<*xf32> {
%0 = "onnx.ConvNoBias"(%arg0, %arg1) {auto_pad = "SAME_UPPER", group = 1 : i32, strides = [2, 3]} : (tensor<1x2x32x64xf32>, tensor<5x2x6x7xf32>) -> tensor<*xf32>
"std.return"(%0) : (tensor<*xf32>) -> ()
}
// CHECK-LABEL: test_conv_no_bias_8
// CHECK: [[RES_ATTR:%.+]] = "onnx.ConvNoBias"(%arg0, %arg1) {auto_pad = "SAME_UPPER", group = 1 : i32, strides = [2, 3]} : (tensor<1x2x32x64xf32>, tensor<5x2x6x7xf32>) -> tensor<1x5x16x22xf32>
// CHECK: return [[RES_ATTR]] : tensor<1x5x16x22xf32>
/// dilations attribute.
func @test_conv_no_bias_9(%arg0 : tensor<1x2x32x64xf32>, %arg1 : tensor<5x2x6x7xf32>) -> tensor<*xf32> {
%0 = "onnx.ConvNoBias"(%arg0, %arg1) {auto_pad = "NOTSET", group = 1 : i32, dilations = [2, 3]} : (tensor<1x2x32x64xf32>, tensor<5x2x6x7xf32>) -> tensor<*xf32>
"std.return"(%0) : (tensor<*xf32>) -> ()
}
// CHECK-LABEL: test_conv_no_bias_9
// CHECK: [[RES_ATTR:%.+]] = "onnx.ConvNoBias"(%arg0, %arg1) {auto_pad = "NOTSET", dilations = [2, 3], group = 1 : i32} : (tensor<1x2x32x64xf32>, tensor<5x2x6x7xf32>) -> tensor<1x5x18x40xf32>
// CHECK: return [[RES_ATTR]] : tensor<1x5x18x40xf32>
/// dilations attribute with stride.
func @test_conv_no_bias_10(%arg0 : tensor<1x2x32x64xf32>, %arg1 : tensor<5x2x6x7xf32>) -> tensor<*xf32> {
%0 = "onnx.ConvNoBias"(%arg0, %arg1) {auto_pad = "NOTSET", group = 1 : i32, dilations = [2, 3], strides = [2, 2]} : (tensor<1x2x32x64xf32>, tensor<5x2x6x7xf32>) -> tensor<*xf32>
"std.return"(%0) : (tensor<*xf32>) -> ()
}
// CHECK-LABEL: test_conv_no_bias_10
// CHECK: [[RES_ATTR:%.+]] = "onnx.ConvNoBias"(%arg0, %arg1) {auto_pad = "NOTSET", dilations = [2, 3], group = 1 : i32, strides = [2, 2]} : (tensor<1x2x32x64xf32>, tensor<5x2x6x7xf32>) -> tensor<1x5x9x20xf32>
// CHECK: return [[RES_ATTR]] : tensor<1x5x9x20xf32>