[mlir][tosa] Support unranked input/weight tensors for convolution ops #134856
Conversation
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
|
@llvm/pr-subscribers-mlir Author: Luke Hutton (lhutton1) ChangesThis commit ensures that convolution operators including: conv2d, depthwise_conv2d, transpose_conv2d and conv3d, can have unranked input/weight operands. In order to support operands with unranked tensors, the tablegen definition was relaxed. The relaxation of tensor type will later be checked by the validation pass, should the user wish to use it. Full diff: https://github.com/llvm/llvm-project/pull/134856.diff 5 Files Affected:
diff --git a/mlir/include/mlir/Dialect/Tosa/IR/TosaOps.td b/mlir/include/mlir/Dialect/Tosa/IR/TosaOps.td
index 741de84cc5840..7e5c62536d2fb 100644
--- a/mlir/include/mlir/Dialect/Tosa/IR/TosaOps.td
+++ b/mlir/include/mlir/Dialect/Tosa/IR/TosaOps.td
@@ -124,7 +124,7 @@ def Tosa_Conv2DOp : Tosa_ConvOp<"conv2d"> {
let arguments = (ins
Tosa_Tensor4D:$input,
- TosaTensorRankOf<[Tosa_Weight], [4]>:$weight,
+ Tosa_Tensor4D:$weight,
Tosa_Tensor1D:$bias,
Tosa_ScalarIntOrFloatTensor:$input_zp,
Tosa_ScalarIntOrFloatTensor:$weight_zp,
@@ -169,7 +169,7 @@ def Tosa_Conv3DOp : Tosa_ConvOp<"conv3d"> {
let arguments = (ins
Tosa_Tensor5D:$input,
- TosaTensorRankOf<[Tosa_Weight], [5]>:$weight,
+ Tosa_Tensor5D:$weight,
Tosa_Tensor1D:$bias,
Tosa_ScalarIntOrFloatTensor:$input_zp,
Tosa_ScalarIntOrFloatTensor:$weight_zp,
@@ -215,7 +215,7 @@ def Tosa_DepthwiseConv2DOp : Tosa_ConvOp<"depthwise_conv2d"> {
let arguments = (ins
Tosa_Tensor4D:$input,
- TosaTensorRankOf<[Tosa_Weight], [4]>:$weight,
+ Tosa_Tensor4D:$weight,
Tosa_Tensor1D:$bias,
Tosa_ScalarIntOrFloatTensor:$input_zp,
Tosa_ScalarIntOrFloatTensor:$weight_zp,
@@ -429,7 +429,7 @@ def Tosa_TransposeConv2DOp : Tosa_ConvOp<"transpose_conv2d"> {
let arguments = (ins
Tosa_Tensor4D:$input,
- TosaTensorRankOf<[Tosa_Weight], [4]>:$weight,
+ Tosa_Tensor4D:$weight,
Tosa_Tensor1D:$bias,
Tosa_ScalarIntOrFloatTensor:$input_zp,
Tosa_ScalarIntOrFloatTensor:$weight_zp,
diff --git a/mlir/include/mlir/Dialect/Tosa/IR/TosaTypesBase.td b/mlir/include/mlir/Dialect/Tosa/IR/TosaTypesBase.td
index 67011f22fbe2a..b9ac1ff705514 100644
--- a/mlir/include/mlir/Dialect/Tosa/IR/TosaTypesBase.td
+++ b/mlir/include/mlir/Dialect/Tosa/IR/TosaTypesBase.td
@@ -84,11 +84,6 @@ def Tosa_QuantizedInt : AnyTypeOf<[Tosa_QuantizedType<"uint8", [8], 0>,
def Tosa_AnyNumber : AnyTypeOf<[Tosa_Int, Tosa_QuantizedInt, AnyFloat],
"number">;
-// For weight tensors from tosa::Conv2DOp, tosa::Conv3DOp,
-// tosa::DepthwiseConv2DOp, tosa::TransposeConv2DOp
-def Tosa_Weight : AnyTypeOf<[Tosa_Int4, Tosa_Int8,
- Tosa_QuantizedInt, AnyFloat]>;
-
//===----------------------------------------------------------------------===//
// TOSA Tensor Conformance
//===----------------------------------------------------------------------===//
diff --git a/mlir/lib/Dialect/Tosa/IR/TosaOps.cpp b/mlir/lib/Dialect/Tosa/IR/TosaOps.cpp
index 59946ca54b933..ee4333f4ba36e 100644
--- a/mlir/lib/Dialect/Tosa/IR/TosaOps.cpp
+++ b/mlir/lib/Dialect/Tosa/IR/TosaOps.cpp
@@ -282,14 +282,14 @@ static LogicalResult verifyConvOp(T op) {
// tensors.
auto inputType = llvm::dyn_cast<RankedTensorType>(op.getInput().getType());
if (!inputType) {
- op.emitOpError("expect a ranked tensor for input, got ") << op.getInput();
- return failure();
+ // Skip following checks if input is not ranked
+ return success();
}
auto weightType = llvm::dyn_cast<RankedTensorType>(op.getWeight().getType());
if (!weightType) {
- op.emitOpError("expect a ranked tensor for weight, got ") << op.getWeight();
- return failure();
+ // Skip following checks if weight is not ranked
+ return success();
}
auto inputEType = inputType.getElementType();
@@ -2899,14 +2899,6 @@ LogicalResult TransposeConv2DOp::verify() {
return emitOpError("expect all stride values to be >= 1, got [")
<< strides << "]";
- const auto inputType = llvm::dyn_cast<RankedTensorType>(getInput().getType());
-
- const auto outputType =
- llvm::dyn_cast<RankedTensorType>(getOutput().getType());
-
- const auto weightType =
- llvm::dyn_cast<RankedTensorType>(getWeight().getType());
-
const auto checkPadAgainstKernelDim =
[this](int64_t pad_value, int64_t kernel_dim_size,
llvm::StringRef pad_name,
@@ -2920,69 +2912,77 @@ LogicalResult TransposeConv2DOp::verify() {
};
const llvm::ArrayRef<int64_t> padding = getOutPad();
-
const int64_t outPadTop = padding[0];
const int64_t outPadBottom = padding[1];
+ const int64_t outPadLeft = padding[2];
+ const int64_t outPadRight = padding[3];
- const int64_t kernelHeight = weightType.getDimSize(1);
-
- if (!ShapedType::isDynamic(kernelHeight)) {
- if (failed(checkPadAgainstKernelDim(outPadTop, kernelHeight, "out_pad_top",
- "KH")))
- return failure();
-
- if (failed(checkPadAgainstKernelDim(outPadBottom, kernelHeight,
- "out_pad_bottom", "KH")))
- return failure();
- }
+ const auto weightType =
+ llvm::dyn_cast<RankedTensorType>(getWeight().getType());
- const int64_t kernelWidth = weightType.getDimSize(2);
+ if (weightType) {
+ const int64_t kernelHeight = weightType.getDimSize(1);
+ if (!ShapedType::isDynamic(kernelHeight)) {
+ if (failed(checkPadAgainstKernelDim(outPadTop, kernelHeight, "out_pad_top",
+ "KH")))
+ return failure();
- const int64_t outPadLeft = padding[2];
- const int64_t outPadRight = padding[3];
+ if (failed(checkPadAgainstKernelDim(outPadBottom, kernelHeight,
+ "out_pad_bottom", "KH")))
+ return failure();
+ }
- if (!ShapedType::isDynamic(kernelWidth)) {
- if (failed(checkPadAgainstKernelDim(outPadLeft, kernelWidth, "out_pad_left",
- "KW")))
- return failure();
+ const int64_t kernelWidth = weightType.getDimSize(2);
+ if (!ShapedType::isDynamic(kernelWidth)) {
+ if (failed(checkPadAgainstKernelDim(outPadLeft, kernelWidth, "out_pad_left",
+ "KW")))
+ return failure();
- if (failed(checkPadAgainstKernelDim(outPadRight, kernelWidth,
- "out_pad_right", "KW")))
- return failure();
+ if (failed(checkPadAgainstKernelDim(outPadRight, kernelWidth,
+ "out_pad_right", "KW")))
+ return failure();
+ }
}
// Rest of the checks depend on the output type being a RankedTensorType
+ const auto outputType =
+ llvm::dyn_cast<RankedTensorType>(getOutput().getType());
if (!outputType)
return success();
- const int64_t inputHeight = inputType.getDimSize(1);
- const int64_t outputHeight = outputType.getDimSize(1);
-
- if (!ShapedType::isDynamic(inputHeight) &&
- !ShapedType::isDynamic(outputHeight)) {
- if (outputHeight !=
- (inputHeight - 1) * strideY + outPadTop + outPadBottom + kernelHeight)
- return emitOpError(
- "dimension mismatch: expected OH == (IH - 1) * stride_y "
- "+ out_pad_top + out_pad_bottom + KH, but got ")
- << outputHeight << " != (" << inputHeight << " - 1) * " << strideY
- << " + " << outPadTop << " + " << outPadBottom << " + "
- << kernelHeight;
- }
+ const auto inputType = llvm::dyn_cast<RankedTensorType>(getInput().getType());
+ if (inputType && weightType) {
+ const int64_t inputHeight = inputType.getDimSize(1);
+ const int64_t kernelHeight = weightType.getDimSize(1);
+ const int64_t outputHeight = outputType.getDimSize(1);
+
+ if (!ShapedType::isDynamic(inputHeight) &&
+ !ShapedType::isDynamic(outputHeight)) {
+ if (outputHeight !=
+ (inputHeight - 1) * strideY + outPadTop + outPadBottom + kernelHeight)
+ return emitOpError(
+ "dimension mismatch: expected OH == (IH - 1) * stride_y "
+ "+ out_pad_top + out_pad_bottom + KH, but got ")
+ << outputHeight << " != (" << inputHeight << " - 1) * " << strideY
+ << " + " << outPadTop << " + " << outPadBottom << " + "
+ << kernelHeight;
+ }
- const int64_t inputWidth = inputType.getDimSize(2);
- const int64_t outputWidth = outputType.getDimSize(2);
+ const int64_t inputWidth = inputType.getDimSize(2);
+ const int64_t kernelWidth = weightType.getDimSize(2);
+ const int64_t outputWidth = outputType.getDimSize(2);
- if (!ShapedType::isDynamic(inputWidth) &&
- !ShapedType::isDynamic(outputWidth)) {
- if (outputWidth !=
- (inputWidth - 1) * strideX + outPadLeft + outPadRight + kernelWidth)
- return emitOpError(
- "dimension mismatch: expected OW == (IW - 1) * stride_x "
- "+ out_pad_left + out_pad_right + KW, but got ")
- << outputWidth << " != (" << inputWidth << " - 1) * " << strideX
- << " + " << outPadLeft << " + " << outPadRight << " + "
- << kernelWidth;
+ if (!ShapedType::isDynamic(inputWidth) &&
+ !ShapedType::isDynamic(outputWidth)) {
+ if (outputWidth !=
+ (inputWidth - 1) * strideX + outPadLeft + outPadRight + kernelWidth)
+ return emitOpError(
+ "dimension mismatch: expected OW == (IW - 1) * stride_x "
+ "+ out_pad_left + out_pad_right + KW, but got ")
+ << outputWidth << " != (" << inputWidth << " - 1) * " << strideX
+ << " + " << outPadLeft << " + " << outPadRight << " + "
+ << kernelWidth;
+ }
}
const auto biasType = llvm::dyn_cast<RankedTensorType>(getBias().getType());
diff --git a/mlir/test/Dialect/Tosa/invalid.mlir b/mlir/test/Dialect/Tosa/invalid.mlir
index 12b2379a592c3..ad764443b97de 100644
--- a/mlir/test/Dialect/Tosa/invalid.mlir
+++ b/mlir/test/Dialect/Tosa/invalid.mlir
@@ -33,19 +33,9 @@ func.func @test_conv2d(%arg0: tensor<1x29x29x4xf32>, %arg1: tensor<16x3x3x4xi8>,
// -----
-func.func @test_conv2d(%arg0: tensor<*xi8>, %arg1: tensor<16x3x3x4xi8>, %arg2: tensor<16xi8>) -> tensor<1x27x27x16xi8> {
- %zp = "tosa.const"() {values = dense<0> : tensor<1xi8>} : () -> tensor<1xi8>
- // expected-error@+1 {{'tosa.conv2d' op expect a ranked tensor for input, got <block argument> of type 'tensor<*xi8>' at index: 0}}
- %0 = tosa.conv2d %arg0, %arg1, %arg2, %zp, %zp {acc_type = i32, dilation = array<i64: 1, 1>, pad = array<i64: 0, 0, 0, 0>, stride = array<i64: 1, 1>}
- : (tensor<*xi8>, tensor<16x3x3x4xi8>, tensor<16xi8>, tensor<1xi8>, tensor<1xi8>) -> tensor<1x27x27x16xi8>
- return %0 : tensor<1x27x27x16xi8>
-}
-
-// -----
-
func.func @test_conv2d(%arg0: tensor<1x29x29x4xi8>, %arg1: tensor<*xi8>, %arg2: tensor<16xi8>) -> tensor<1x27x27x16xi8> {
%zp = "tosa.const"() {values = dense<0> : tensor<1xi8>} : () -> tensor<1xi8>
- // expected-error@+1 {{'tosa.conv2d' op operand #1 must be 4D tensor of 4-bit signless integer or 8-bit signless integer or Quint8 type or Qint4 type or Qint8 type or Qint16 type or Qint32 type or floating-point values, but got 'tensor<*xi8>'}}
+ // expected-error@+1 {{'tosa.conv2d' op illegal: operand/result data types not supported}}
%0 = tosa.conv2d %arg0, %arg1, %arg2, %zp, %zp {acc_type = i32, dilation = array<i64: 1, 1>, pad = array<i64: 0, 0, 0, 0>, stride = array<i64: 1, 1>}
: (tensor<1x29x29x4xi8>, tensor<*xi8>, tensor<16xi8>, tensor<1xi8>, tensor<1xi8>) -> tensor<1x27x27x16xi8>
return %0 : tensor<1x27x27x16xi8>
diff --git a/mlir/test/Dialect/Tosa/ops.mlir b/mlir/test/Dialect/Tosa/ops.mlir
index 248d84da6b8b9..35978060a134a 100644
--- a/mlir/test/Dialect/Tosa/ops.mlir
+++ b/mlir/test/Dialect/Tosa/ops.mlir
@@ -70,6 +70,13 @@ func.func @test_conv2d(%arg0: tensor<1x4x4x4xf32>, %arg1: tensor<8x1x1x4xf32>, %
return %0 : tensor<1x4x4x8xf32>
}
+// -----
+// CHECK-LABEL: conv2d_unranked_input
+func.func @test_conv2d_unranked_input(%arg0: tensor<*xf32>, %arg1: tensor<8x1x1x4xf32>, %arg2: tensor<8xf32>, %arg3: tensor<1xf32>, %arg4: tensor<1xf32>) -> tensor<1x4x4x8xf32> {
+ %0 = tosa.conv2d %arg0, %arg1, %arg2, %arg3, %arg4 {acc_type = f32, dilation = array<i64: 1, 1>, pad = array<i64: 0, 0, 0, 0>, stride = array<i64: 1, 1>, local_bound = true} : (tensor<*xf32>, tensor<8x1x1x4xf32>, tensor<8xf32>, tensor<1xf32>, tensor<1xf32>) -> tensor<1x4x4x8xf32>
+ return %0 : tensor<1x4x4x8xf32>
+}
+
// -----
// CHECK-LABEL: conv2d_quant_uniform
func.func @test_conv2d_quant_uniform(%arg0: tensor<1x4x4x4x!quant.uniform<i8:f32, 0.01>>, %arg1: tensor<8x1x1x4x!quant.uniform<i8:f32, 0.01>>, %arg2: tensor<8x!quant.uniform<i8:f32, 0.01>>) -> tensor<1x4x4x8x!quant.uniform<i32:f32, 0.01>> {
@@ -202,6 +209,20 @@ func.func @test_transpose_conv2d(%arg0: tensor<1x32x32x8xf32>, %arg1: tensor<16x
return %0 : tensor<1x32x32x16xf32>
}
+// -----
+// CHECK-LABEL: transpose_conv2d_unranked_input
+func.func @test_transpose_conv2d_unranked_input(%arg0: tensor<*xf32>, %arg1: tensor<16x1x1x8xf32>, %arg2: tensor<16xf32>, %arg3: tensor<1xf32>, %arg4: tensor<1xf32>) -> tensor<1x32x32x16xf32> {
+ %0 = tosa.transpose_conv2d %arg0, %arg1, %arg2, %arg3, %arg4 {acc_type = f32, out_pad = array<i64: 0, 0, 0, 0>, out_shape = array<i64: 1, 32, 32, 16>, stride = array<i64: 1, 1>} : (tensor<*xf32>, tensor<16x1x1x8xf32>, tensor<16xf32>, tensor<1xf32>, tensor<1xf32>) -> tensor<1x32x32x16xf32>
+ return %0 : tensor<1x32x32x16xf32>
+}
+
+// -----
+// CHECK-LABEL: transpose_conv2d_unranked_weight
+func.func @test_transpose_conv2d_unranked_weight(%arg0: tensor<1x32x32x8xf32>, %arg1: tensor<*xf32>, %arg2: tensor<16xf32>, %arg3: tensor<1xf32>, %arg4: tensor<1xf32>) -> tensor<1x32x32x16xf32> {
+ %0 = tosa.transpose_conv2d %arg0, %arg1, %arg2, %arg3, %arg4 {acc_type = f32, out_pad = array<i64: 0, 0, 0, 0>, out_shape = array<i64: 1, 32, 32, 16>, stride = array<i64: 1, 1>} : (tensor<1x32x32x8xf32>, tensor<*xf32>, tensor<16xf32>, tensor<1xf32>, tensor<1xf32>) -> tensor<1x32x32x16xf32>
+ return %0 : tensor<1x32x32x16xf32>
+}
+
// -----
// CHECK-LABEL: transpose_conv2d_with_local_bound
func.func @test_transpose_conv2d_with_local_bound(%arg0: tensor<1x32x32x8xf32>, %arg1: tensor<16x1x1x8xf32>, %arg2: tensor<16xf32>, %arg3: tensor<1xf32>, %arg4: tensor<1xf32>) -> tensor<1x32x32x16xf32> {
|
|
@llvm/pr-subscribers-mlir-tosa Author: Luke Hutton (lhutton1) ChangesThis commit ensures that convolution operators including: conv2d, depthwise_conv2d, transpose_conv2d and conv3d, can have unranked input/weight operands. In order to support operands with unranked tensors, the tablegen definition was relaxed. The relaxation of tensor type will later be checked by the validation pass, should the user wish to use it. Full diff: https://github.com/llvm/llvm-project/pull/134856.diff 5 Files Affected:
diff --git a/mlir/include/mlir/Dialect/Tosa/IR/TosaOps.td b/mlir/include/mlir/Dialect/Tosa/IR/TosaOps.td
index 741de84cc5840..7e5c62536d2fb 100644
--- a/mlir/include/mlir/Dialect/Tosa/IR/TosaOps.td
+++ b/mlir/include/mlir/Dialect/Tosa/IR/TosaOps.td
@@ -124,7 +124,7 @@ def Tosa_Conv2DOp : Tosa_ConvOp<"conv2d"> {
let arguments = (ins
Tosa_Tensor4D:$input,
- TosaTensorRankOf<[Tosa_Weight], [4]>:$weight,
+ Tosa_Tensor4D:$weight,
Tosa_Tensor1D:$bias,
Tosa_ScalarIntOrFloatTensor:$input_zp,
Tosa_ScalarIntOrFloatTensor:$weight_zp,
@@ -169,7 +169,7 @@ def Tosa_Conv3DOp : Tosa_ConvOp<"conv3d"> {
let arguments = (ins
Tosa_Tensor5D:$input,
- TosaTensorRankOf<[Tosa_Weight], [5]>:$weight,
+ Tosa_Tensor5D:$weight,
Tosa_Tensor1D:$bias,
Tosa_ScalarIntOrFloatTensor:$input_zp,
Tosa_ScalarIntOrFloatTensor:$weight_zp,
@@ -215,7 +215,7 @@ def Tosa_DepthwiseConv2DOp : Tosa_ConvOp<"depthwise_conv2d"> {
let arguments = (ins
Tosa_Tensor4D:$input,
- TosaTensorRankOf<[Tosa_Weight], [4]>:$weight,
+ Tosa_Tensor4D:$weight,
Tosa_Tensor1D:$bias,
Tosa_ScalarIntOrFloatTensor:$input_zp,
Tosa_ScalarIntOrFloatTensor:$weight_zp,
@@ -429,7 +429,7 @@ def Tosa_TransposeConv2DOp : Tosa_ConvOp<"transpose_conv2d"> {
let arguments = (ins
Tosa_Tensor4D:$input,
- TosaTensorRankOf<[Tosa_Weight], [4]>:$weight,
+ Tosa_Tensor4D:$weight,
Tosa_Tensor1D:$bias,
Tosa_ScalarIntOrFloatTensor:$input_zp,
Tosa_ScalarIntOrFloatTensor:$weight_zp,
diff --git a/mlir/include/mlir/Dialect/Tosa/IR/TosaTypesBase.td b/mlir/include/mlir/Dialect/Tosa/IR/TosaTypesBase.td
index 67011f22fbe2a..b9ac1ff705514 100644
--- a/mlir/include/mlir/Dialect/Tosa/IR/TosaTypesBase.td
+++ b/mlir/include/mlir/Dialect/Tosa/IR/TosaTypesBase.td
@@ -84,11 +84,6 @@ def Tosa_QuantizedInt : AnyTypeOf<[Tosa_QuantizedType<"uint8", [8], 0>,
def Tosa_AnyNumber : AnyTypeOf<[Tosa_Int, Tosa_QuantizedInt, AnyFloat],
"number">;
-// For weight tensors from tosa::Conv2DOp, tosa::Conv3DOp,
-// tosa::DepthwiseConv2DOp, tosa::TransposeConv2DOp
-def Tosa_Weight : AnyTypeOf<[Tosa_Int4, Tosa_Int8,
- Tosa_QuantizedInt, AnyFloat]>;
-
//===----------------------------------------------------------------------===//
// TOSA Tensor Conformance
//===----------------------------------------------------------------------===//
diff --git a/mlir/lib/Dialect/Tosa/IR/TosaOps.cpp b/mlir/lib/Dialect/Tosa/IR/TosaOps.cpp
index 59946ca54b933..ee4333f4ba36e 100644
--- a/mlir/lib/Dialect/Tosa/IR/TosaOps.cpp
+++ b/mlir/lib/Dialect/Tosa/IR/TosaOps.cpp
@@ -282,14 +282,14 @@ static LogicalResult verifyConvOp(T op) {
// tensors.
auto inputType = llvm::dyn_cast<RankedTensorType>(op.getInput().getType());
if (!inputType) {
- op.emitOpError("expect a ranked tensor for input, got ") << op.getInput();
- return failure();
+ // Skip following checks if input is not ranked
+ return success();
}
auto weightType = llvm::dyn_cast<RankedTensorType>(op.getWeight().getType());
if (!weightType) {
- op.emitOpError("expect a ranked tensor for weight, got ") << op.getWeight();
- return failure();
+ // Skip following checks if weight is not ranked
+ return success();
}
auto inputEType = inputType.getElementType();
@@ -2899,14 +2899,6 @@ LogicalResult TransposeConv2DOp::verify() {
return emitOpError("expect all stride values to be >= 1, got [")
<< strides << "]";
- const auto inputType = llvm::dyn_cast<RankedTensorType>(getInput().getType());
-
- const auto outputType =
- llvm::dyn_cast<RankedTensorType>(getOutput().getType());
-
- const auto weightType =
- llvm::dyn_cast<RankedTensorType>(getWeight().getType());
-
const auto checkPadAgainstKernelDim =
[this](int64_t pad_value, int64_t kernel_dim_size,
llvm::StringRef pad_name,
@@ -2920,69 +2912,77 @@ LogicalResult TransposeConv2DOp::verify() {
};
const llvm::ArrayRef<int64_t> padding = getOutPad();
-
const int64_t outPadTop = padding[0];
const int64_t outPadBottom = padding[1];
+ const int64_t outPadLeft = padding[2];
+ const int64_t outPadRight = padding[3];
- const int64_t kernelHeight = weightType.getDimSize(1);
-
- if (!ShapedType::isDynamic(kernelHeight)) {
- if (failed(checkPadAgainstKernelDim(outPadTop, kernelHeight, "out_pad_top",
- "KH")))
- return failure();
-
- if (failed(checkPadAgainstKernelDim(outPadBottom, kernelHeight,
- "out_pad_bottom", "KH")))
- return failure();
- }
+ const auto weightType =
+ llvm::dyn_cast<RankedTensorType>(getWeight().getType());
- const int64_t kernelWidth = weightType.getDimSize(2);
+ if (weightType) {
+ const int64_t kernelHeight = weightType.getDimSize(1);
+ if (!ShapedType::isDynamic(kernelHeight)) {
+ if (failed(checkPadAgainstKernelDim(outPadTop, kernelHeight, "out_pad_top",
+ "KH")))
+ return failure();
- const int64_t outPadLeft = padding[2];
- const int64_t outPadRight = padding[3];
+ if (failed(checkPadAgainstKernelDim(outPadBottom, kernelHeight,
+ "out_pad_bottom", "KH")))
+ return failure();
+ }
- if (!ShapedType::isDynamic(kernelWidth)) {
- if (failed(checkPadAgainstKernelDim(outPadLeft, kernelWidth, "out_pad_left",
- "KW")))
- return failure();
+ const int64_t kernelWidth = weightType.getDimSize(2);
+ if (!ShapedType::isDynamic(kernelWidth)) {
+ if (failed(checkPadAgainstKernelDim(outPadLeft, kernelWidth, "out_pad_left",
+ "KW")))
+ return failure();
- if (failed(checkPadAgainstKernelDim(outPadRight, kernelWidth,
- "out_pad_right", "KW")))
- return failure();
+ if (failed(checkPadAgainstKernelDim(outPadRight, kernelWidth,
+ "out_pad_right", "KW")))
+ return failure();
+ }
}
// Rest of the checks depend on the output type being a RankedTensorType
+ const auto outputType =
+ llvm::dyn_cast<RankedTensorType>(getOutput().getType());
if (!outputType)
return success();
- const int64_t inputHeight = inputType.getDimSize(1);
- const int64_t outputHeight = outputType.getDimSize(1);
-
- if (!ShapedType::isDynamic(inputHeight) &&
- !ShapedType::isDynamic(outputHeight)) {
- if (outputHeight !=
- (inputHeight - 1) * strideY + outPadTop + outPadBottom + kernelHeight)
- return emitOpError(
- "dimension mismatch: expected OH == (IH - 1) * stride_y "
- "+ out_pad_top + out_pad_bottom + KH, but got ")
- << outputHeight << " != (" << inputHeight << " - 1) * " << strideY
- << " + " << outPadTop << " + " << outPadBottom << " + "
- << kernelHeight;
- }
+ const auto inputType = llvm::dyn_cast<RankedTensorType>(getInput().getType());
+ if (inputType && weightType) {
+ const int64_t inputHeight = inputType.getDimSize(1);
+ const int64_t kernelHeight = weightType.getDimSize(1);
+ const int64_t outputHeight = outputType.getDimSize(1);
+
+ if (!ShapedType::isDynamic(inputHeight) &&
+ !ShapedType::isDynamic(outputHeight)) {
+ if (outputHeight !=
+ (inputHeight - 1) * strideY + outPadTop + outPadBottom + kernelHeight)
+ return emitOpError(
+ "dimension mismatch: expected OH == (IH - 1) * stride_y "
+ "+ out_pad_top + out_pad_bottom + KH, but got ")
+ << outputHeight << " != (" << inputHeight << " - 1) * " << strideY
+ << " + " << outPadTop << " + " << outPadBottom << " + "
+ << kernelHeight;
+ }
- const int64_t inputWidth = inputType.getDimSize(2);
- const int64_t outputWidth = outputType.getDimSize(2);
+ const int64_t inputWidth = inputType.getDimSize(2);
+ const int64_t kernelWidth = weightType.getDimSize(2);
+ const int64_t outputWidth = outputType.getDimSize(2);
- if (!ShapedType::isDynamic(inputWidth) &&
- !ShapedType::isDynamic(outputWidth)) {
- if (outputWidth !=
- (inputWidth - 1) * strideX + outPadLeft + outPadRight + kernelWidth)
- return emitOpError(
- "dimension mismatch: expected OW == (IW - 1) * stride_x "
- "+ out_pad_left + out_pad_right + KW, but got ")
- << outputWidth << " != (" << inputWidth << " - 1) * " << strideX
- << " + " << outPadLeft << " + " << outPadRight << " + "
- << kernelWidth;
+ if (!ShapedType::isDynamic(inputWidth) &&
+ !ShapedType::isDynamic(outputWidth)) {
+ if (outputWidth !=
+ (inputWidth - 1) * strideX + outPadLeft + outPadRight + kernelWidth)
+ return emitOpError(
+ "dimension mismatch: expected OW == (IW - 1) * stride_x "
+ "+ out_pad_left + out_pad_right + KW, but got ")
+ << outputWidth << " != (" << inputWidth << " - 1) * " << strideX
+ << " + " << outPadLeft << " + " << outPadRight << " + "
+ << kernelWidth;
+ }
}
const auto biasType = llvm::dyn_cast<RankedTensorType>(getBias().getType());
diff --git a/mlir/test/Dialect/Tosa/invalid.mlir b/mlir/test/Dialect/Tosa/invalid.mlir
index 12b2379a592c3..ad764443b97de 100644
--- a/mlir/test/Dialect/Tosa/invalid.mlir
+++ b/mlir/test/Dialect/Tosa/invalid.mlir
@@ -33,19 +33,9 @@ func.func @test_conv2d(%arg0: tensor<1x29x29x4xf32>, %arg1: tensor<16x3x3x4xi8>,
// -----
-func.func @test_conv2d(%arg0: tensor<*xi8>, %arg1: tensor<16x3x3x4xi8>, %arg2: tensor<16xi8>) -> tensor<1x27x27x16xi8> {
- %zp = "tosa.const"() {values = dense<0> : tensor<1xi8>} : () -> tensor<1xi8>
- // expected-error@+1 {{'tosa.conv2d' op expect a ranked tensor for input, got <block argument> of type 'tensor<*xi8>' at index: 0}}
- %0 = tosa.conv2d %arg0, %arg1, %arg2, %zp, %zp {acc_type = i32, dilation = array<i64: 1, 1>, pad = array<i64: 0, 0, 0, 0>, stride = array<i64: 1, 1>}
- : (tensor<*xi8>, tensor<16x3x3x4xi8>, tensor<16xi8>, tensor<1xi8>, tensor<1xi8>) -> tensor<1x27x27x16xi8>
- return %0 : tensor<1x27x27x16xi8>
-}
-
-// -----
-
func.func @test_conv2d(%arg0: tensor<1x29x29x4xi8>, %arg1: tensor<*xi8>, %arg2: tensor<16xi8>) -> tensor<1x27x27x16xi8> {
%zp = "tosa.const"() {values = dense<0> : tensor<1xi8>} : () -> tensor<1xi8>
- // expected-error@+1 {{'tosa.conv2d' op operand #1 must be 4D tensor of 4-bit signless integer or 8-bit signless integer or Quint8 type or Qint4 type or Qint8 type or Qint16 type or Qint32 type or floating-point values, but got 'tensor<*xi8>'}}
+ // expected-error@+1 {{'tosa.conv2d' op illegal: operand/result data types not supported}}
%0 = tosa.conv2d %arg0, %arg1, %arg2, %zp, %zp {acc_type = i32, dilation = array<i64: 1, 1>, pad = array<i64: 0, 0, 0, 0>, stride = array<i64: 1, 1>}
: (tensor<1x29x29x4xi8>, tensor<*xi8>, tensor<16xi8>, tensor<1xi8>, tensor<1xi8>) -> tensor<1x27x27x16xi8>
return %0 : tensor<1x27x27x16xi8>
diff --git a/mlir/test/Dialect/Tosa/ops.mlir b/mlir/test/Dialect/Tosa/ops.mlir
index 248d84da6b8b9..35978060a134a 100644
--- a/mlir/test/Dialect/Tosa/ops.mlir
+++ b/mlir/test/Dialect/Tosa/ops.mlir
@@ -70,6 +70,13 @@ func.func @test_conv2d(%arg0: tensor<1x4x4x4xf32>, %arg1: tensor<8x1x1x4xf32>, %
return %0 : tensor<1x4x4x8xf32>
}
+// -----
+// CHECK-LABEL: conv2d_unranked_input
+func.func @test_conv2d_unranked_input(%arg0: tensor<*xf32>, %arg1: tensor<8x1x1x4xf32>, %arg2: tensor<8xf32>, %arg3: tensor<1xf32>, %arg4: tensor<1xf32>) -> tensor<1x4x4x8xf32> {
+ %0 = tosa.conv2d %arg0, %arg1, %arg2, %arg3, %arg4 {acc_type = f32, dilation = array<i64: 1, 1>, pad = array<i64: 0, 0, 0, 0>, stride = array<i64: 1, 1>, local_bound = true} : (tensor<*xf32>, tensor<8x1x1x4xf32>, tensor<8xf32>, tensor<1xf32>, tensor<1xf32>) -> tensor<1x4x4x8xf32>
+ return %0 : tensor<1x4x4x8xf32>
+}
+
// -----
// CHECK-LABEL: conv2d_quant_uniform
func.func @test_conv2d_quant_uniform(%arg0: tensor<1x4x4x4x!quant.uniform<i8:f32, 0.01>>, %arg1: tensor<8x1x1x4x!quant.uniform<i8:f32, 0.01>>, %arg2: tensor<8x!quant.uniform<i8:f32, 0.01>>) -> tensor<1x4x4x8x!quant.uniform<i32:f32, 0.01>> {
@@ -202,6 +209,20 @@ func.func @test_transpose_conv2d(%arg0: tensor<1x32x32x8xf32>, %arg1: tensor<16x
return %0 : tensor<1x32x32x16xf32>
}
+// -----
+// CHECK-LABEL: transpose_conv2d_unranked_input
+func.func @test_transpose_conv2d_unranked_input(%arg0: tensor<*xf32>, %arg1: tensor<16x1x1x8xf32>, %arg2: tensor<16xf32>, %arg3: tensor<1xf32>, %arg4: tensor<1xf32>) -> tensor<1x32x32x16xf32> {
+ %0 = tosa.transpose_conv2d %arg0, %arg1, %arg2, %arg3, %arg4 {acc_type = f32, out_pad = array<i64: 0, 0, 0, 0>, out_shape = array<i64: 1, 32, 32, 16>, stride = array<i64: 1, 1>} : (tensor<*xf32>, tensor<16x1x1x8xf32>, tensor<16xf32>, tensor<1xf32>, tensor<1xf32>) -> tensor<1x32x32x16xf32>
+ return %0 : tensor<1x32x32x16xf32>
+}
+
+// -----
+// CHECK-LABEL: transpose_conv2d_unranked_weight
+func.func @test_transpose_conv2d_unranked_weight(%arg0: tensor<1x32x32x8xf32>, %arg1: tensor<*xf32>, %arg2: tensor<16xf32>, %arg3: tensor<1xf32>, %arg4: tensor<1xf32>) -> tensor<1x32x32x16xf32> {
+ %0 = tosa.transpose_conv2d %arg0, %arg1, %arg2, %arg3, %arg4 {acc_type = f32, out_pad = array<i64: 0, 0, 0, 0>, out_shape = array<i64: 1, 32, 32, 16>, stride = array<i64: 1, 1>} : (tensor<1x32x32x8xf32>, tensor<*xf32>, tensor<16xf32>, tensor<1xf32>, tensor<1xf32>) -> tensor<1x32x32x16xf32>
+ return %0 : tensor<1x32x32x16xf32>
+}
+
// -----
// CHECK-LABEL: transpose_conv2d_with_local_bound
func.func @test_transpose_conv2d_with_local_bound(%arg0: tensor<1x32x32x8xf32>, %arg1: tensor<16x1x1x8xf32>, %arg2: tensor<16xf32>, %arg3: tensor<1xf32>, %arg4: tensor<1xf32>) -> tensor<1x32x32x16xf32> {
|
|
✅ With the latest revision this PR passed the C/C++ code formatter. |
lhutton1
force-pushed
the
unranked-convolution
branch
from
6464c24 to
26cb5f8
Compare
Jerry-Ge
requested review from
GeorgeARM and
FranklandJack
and removed request for
GeorgeARM
lhutton1
force-pushed
the
unranked-convolution
branch
from
26cb5f8 to
40a5eb5
Compare
udaya-ranga
reviewed
Apr 24, 2025
GeorgeARM
approved these changes
Apr 25, 2025
This commit ensures that convolution operators including: conv2d, depthwise_conv2d, transpose_conv2d and conv3d, can have unranked input/weight operands. In order to support operands with unranked tensors, the tablegen definition was relaxed. The relaxation of tensor type will later be checked by the validation pass, should the user wish to use it. Change-Id: I33334909e0d4d0676daae81bfc4647e86abc063a Signed-off-by: Luke Hutton <[email protected]>
lhutton1
force-pushed
the
unranked-convolution
branch
from
40a5eb5 to
19abbc0
Compare
IanWood1
pushed a commit
to IanWood1/llvm-project
that referenced
this pull request
May 6, 2025
llvm#134856) This commit ensures that convolution operators including: conv2d, depthwise_conv2d, transpose_conv2d and conv3d, can have unranked input/weight operands. In order to support operands with unranked tensors, the tablegen definition was relaxed. The relaxation of tensor type will later be checked by the validation pass, should the user wish to use it. Signed-off-by: Luke Hutton <[email protected]>
IanWood1
pushed a commit
to IanWood1/llvm-project
that referenced
this pull request
May 6, 2025
llvm#134856) This commit ensures that convolution operators including: conv2d, depthwise_conv2d, transpose_conv2d and conv3d, can have unranked input/weight operands. In order to support operands with unranked tensors, the tablegen definition was relaxed. The relaxation of tensor type will later be checked by the validation pass, should the user wish to use it. Signed-off-by: Luke Hutton <[email protected]>
IanWood1
pushed a commit
to IanWood1/llvm-project
that referenced
this pull request
May 6, 2025
llvm#134856) This commit ensures that convolution operators including: conv2d, depthwise_conv2d, transpose_conv2d and conv3d, can have unranked input/weight operands. In order to support operands with unranked tensors, the tablegen definition was relaxed. The relaxation of tensor type will later be checked by the validation pass, should the user wish to use it. Signed-off-by: Luke Hutton <[email protected]>
Ankur-0429
pushed a commit
to Ankur-0429/llvm-project
that referenced
this pull request
May 9, 2025
llvm#134856) This commit ensures that convolution operators including: conv2d, depthwise_conv2d, transpose_conv2d and conv3d, can have unranked input/weight operands. In order to support operands with unranked tensors, the tablegen definition was relaxed. The relaxation of tensor type will later be checked by the validation pass, should the user wish to use it. Signed-off-by: Luke Hutton <[email protected]>
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
This commit ensures that convolution operators including: conv2d, depthwise_conv2d, transpose_conv2d and conv3d, can have unranked input/weight operands.
In order to support operands with unranked tensors, the tablegen definition was relaxed. The relaxation of tensor type will later be checked by the validation pass, should the user wish to use it.