caio/clang-p2996
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

[mlir][tosa] Support unranked input/weight tensors for convolution ops (#134856)

Browse Source 
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 <luke.hutton@arm.com>
This commit is contained in:
Luke Hutton
2025-04-25 15:31:17 +02:00
committed by GitHub
parent 9147569c7f
commit 63a2b0bd3d
5 changed files with 86 additions and 91 deletions
Download Patch File Download Diff File Expand all files Collapse all files

8
mlir/include/mlir/Dialect/Tosa/IR/TosaOps.td
View File

@@ -125,7 +125,7 @@ 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,
@@ -172,7 +172,7 @@ 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,
@@ -218,7 +218,7 @@ 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,
@@ -434,7 +434,7 @@ 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,

5
mlir/include/mlir/Dialect/Tosa/IR/TosaTypesBase.td
View File

@@ -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
//===----------------------------------------------------------------------===//

127
mlir/lib/Dialect/Tosa/IR/TosaOps.cpp
View File

@@ -278,19 +278,8 @@ Value mlir::tosa::createPadConstTensor(OpBuilder &builder, Location loc,
template <typename T>
static LogicalResult verifyConvOp(T op) {
// All TOSA conv ops have an input and weight arguments which must be ranked
// 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();
}
auto weightType = llvm::dyn_cast<RankedTensorType>(op.getWeight().getType());
if (!weightType) {
op.emitOpError("expect a ranked tensor for weight, got ") << op.getWeight();
return failure();
}
const auto inputType = llvm::dyn_cast<TensorType>(op.getInput().getType());
const auto weightType = llvm::dyn_cast<TensorType>(op.getWeight().getType());
auto inputEType = inputType.getElementType();
auto weightEType = weightType.getElementType();
@@ -3063,14 +3052,6 @@ 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,
@@ -3084,69 +3065,77 @@ 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 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 int64_t kernelWidth = weightType.getDimSize(2);
const int64_t outPadLeft = padding[2];
const int64_t outPadRight = padding[3];
if (!ShapedType::isDynamic(kernelWidth)) {
if (failed(checkPadAgainstKernelDim(outPadLeft, kernelWidth, "out_pad_left",
"KW")))
return failure();
const auto weightType =
llvm::dyn_cast<RankedTensorType>(getWeight().getType());
if (failed(checkPadAgainstKernelDim(outPadRight, kernelWidth,
"out_pad_right", "KW")))
return failure();
if (weightType) {
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 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();
}
}
// 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);
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;
}
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());

16
mlir/test/Dialect/Tosa/invalid.mlir
View File

@@ -22,22 +22,12 @@ func.func @test_const_non_tensor_attr() {
// -----
func.func @test_conv2d(%arg0: tensor<1x29x29x4xf32>, %arg1: tensor<16x3x3x4xi8>, %arg2: tensor<16xi8>) -> tensor<1x27x27x16xi8> {
func.func @test_conv2d(%arg0: tensor<*xf32>, %arg1: tensor<16x3x3x4xi8>, %arg2: tensor<16xi8>) -> tensor<1x27x27x16xi8> {
%input_zp = "tosa.const"() <{values = dense<0> : tensor<1xi8>}> : () -> tensor<1xi8>
%weight_zp = "tosa.const"() <{values = dense<0> : tensor<1xi8>}> : () -> tensor<1xi8>
// expected-error@+1 {{'tosa.conv2d' op expect both input and weight to be float or not together, got 'f32' and 'i8'}}
%0 = tosa.conv2d %arg0, %arg1, %arg2, %input_zp, %weight_zp {acc_type = i32, dilation = array<i64: 1, 1>, pad = array<i64: 0, 0, 0, 0>, stride = array<i64: 1, 1>}
: (tensor<1x29x29x4xf32>, tensor<16x3x3x4xi8>, tensor<16xi8>, tensor<1xi8>, tensor<1xi8>) -> tensor<1x27x27x16xi8>
return %0 : tensor<1x27x27x16xi8>
}
// -----
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>
: (tensor<*xf32>, tensor<16x3x3x4xi8>, tensor<16xi8>, tensor<1xi8>, tensor<1xi8>) -> tensor<1x27x27x16xi8>
return %0 : tensor<1x27x27x16xi8>
}
@@ -45,7 +35,7 @@ func.func @test_conv2d(%arg0: tensor<*xi8>, %arg1: tensor<16x3x3x4xi8>, %arg2: t
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>

21
mlir/test/Dialect/Tosa/ops.mlir
View File

@@ -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> {