[mlir][tosa] Add error if verification to pooling operators (llvm#130052)

This commit adds the following checks to avg_pool2d and max_pool2d TOSA
operations:
- check kernel values are >= 1
- check stride values are >= 1
- check padding values are >= 0
- check padding values are less than kernel sizes
- check output shape matches the expected output shape

Signed-off-by: Luke Hutton <[email protected]>

Author: lhutton1

mlir/lib/Dialect/Tosa/IR/TosaOps.cpp

@@ -504,7 +504,95 @@ LogicalResult tosa::ArgMaxOp::verify() {
   return success();
 }
 
+template <typename T>
+static LogicalResult verifyPoolingOp(T op) {
+  const llvm::ArrayRef<int64_t> kernel = op.getKernel();
+  if (llvm::any_of(kernel, [](int64_t s) { return s < 1; }))
+    return op.emitOpError("expect all kernel values to be >= 1, got ")
+           << kernel;
+
+  const llvm::ArrayRef<int64_t> strides = op.getStride();
+  if (llvm::any_of(strides, [](int64_t s) { return s < 1; }))
+    return op.emitOpError("expect all stride values to be >= 1, got ")
+           << strides;
+
+  const llvm::ArrayRef<int64_t> padding = op.getPad();
+  if (llvm::any_of(padding, [](int64_t p) { return p < 0; }))
+    return op.emitOpError("expect all padding values to be >= 0, got ")
+           << padding;
+
+  // Padding must be less than kernel size to avoid a divide-by-zero
+  const int64_t kernelX = kernel[1];
+  const int64_t padLeft = padding[2];
+  const int64_t padRight = padding[3];
+  if (padRight >= kernelX || padLeft >= kernelX)
+    return op.emitOpError("expected left/right padding to be less than the "
+                          "width of the kernel, got pad_left=")
+           << padLeft << ", pad_right=" << padRight << ", kernel_x=" << kernelX;
+
+  const int64_t kernelY = kernel[0];
+  const int64_t padTop = padding[0];
+  const int64_t padBottom = padding[1];
+  if (padTop >= kernelY || padBottom >= kernelY)
+    return op.emitOpError("expected top/bottom padding to be less than the "
+                          "height of the kernel, got pad_top=")
+           << padTop << ", pad_bottom=" << padBottom
+           << ", kernel_y=" << kernelY;
+
+  const auto inputType =
+      llvm::dyn_cast<RankedTensorType>(op.getInput().getType());
+  const auto outputType =
+      llvm::dyn_cast<RankedTensorType>(op.getResult().getType());
+  if (!inputType || !outputType)
+    return success();
+
+  const auto verifyOutputSize =
+      [&op](const int64_t inputSize, const int64_t outputSize,
+            const int64_t kernelSize, const int64_t strideSize,
+            const int64_t padBefore, const int64_t padAfter,
+            const llvm::StringRef dimName, const llvm::StringRef dimAxis,
+            const llvm::StringRef padBeforeName,
+            const llvm::StringRef padAfterName) -> LogicalResult {
+    if (ShapedType::isDynamic(inputSize))
+      return success();
+
+    const std::optional<int64_t> calculatedOutSizeMinusOne =
+        idivCheck(inputSize + padBefore + padAfter - kernelSize, strideSize);
+    if (!calculatedOutSizeMinusOne.has_value())
+      return op.emitOpError("expected input_")
+             << dimName << " + pad_" << padBeforeName << " + pad_"
+             << padAfterName << " - kernel_" << dimAxis
+             << " to be wholly divisible by stride_" << dimAxis << ", got ("
+             << inputSize << " + " << padBefore << " + " << padAfter << " - "
+             << kernelSize << ") / " << strideSize;
+
+    const int64_t calculatedOutSize = calculatedOutSizeMinusOne.value() + 1;
+    if (!ShapedType::isDynamic(outputSize) && calculatedOutSize != outputSize)
+      return op.emitOpError("calculated output ")
+             << dimName << " did not match expected: "
+             << "calculated=" << calculatedOutSize
+             << ", expected=" << outputSize;
+
+    return success();
+  };
+
+  if (failed(verifyOutputSize(inputType.getDimSize(1), outputType.getDimSize(1),
+                              kernel[0], strides[0], padding[0], padding[1],
+                              "height", "y", "top", "bottom")))
+    return failure();
+
+  if (failed(verifyOutputSize(inputType.getDimSize(2), outputType.getDimSize(2),
+                              kernel[1], strides[1], padding[2], padding[3],
+                              "width", "x", "left", "right")))
+    return failure();
+
+  return success();
+}
+
 LogicalResult tosa::AvgPool2dOp::verify() {
+  if (failed(verifyPoolingOp(*this)))
+    return failure();
+
   const Type inputETy = getStorageElementTypeOrSelf(getInput().getType());
   const Type resultETy = getStorageElementTypeOrSelf(getOutput().getType());
   const Type inputZpETy = getStorageElementTypeOrSelf(getInputZp().getType());
@@ -2650,8 +2738,14 @@ LogicalResult MaxPool2dOp::inferReturnTypeComponents(
 }
 
 LogicalResult MaxPool2dOp::verify() {
-  return verifySameElementTypes(*this, /* intype = */ getInput().getType(),
-                                /* outType = */ getOutput().getType());
+  if (failed(verifySameElementTypes(*this, /* intype = */ getInput().getType(),
+                                    /* outType = */ getOutput().getType())))
+    return failure();
+
+  if (failed(verifyPoolingOp(*this)))
+    return failure();
+
+  return success();
 }
 
 LogicalResult DepthwiseConv2DOp::inferReturnTypeComponents(

mlir/test/Dialect/Tosa/invalid.mlir

@@ -1753,3 +1753,102 @@ func.func @test_negate_output_zp_non_zero(%arg0: tensor<1x16x16x8xf32>) -> tenso
       : (tensor<1x16x16x8xf32>, tensor<1xf32>, tensor<1xf32>) -> tensor<1x16x16x8xf32>
   return %0 : tensor<1x16x16x8xf32>
 }
+
+// -----
+
+func.func @test_avgpool2d_invalid_kernel(%arg0: tensor<1x32x32x8xf32>, %arg1: tensor<1xf32>, %arg2: tensor<1xf32>) -> tensor<1x32x32x8xf32> {
+  // expected-error@+1 {{'tosa.avg_pool2d' op expect all kernel values to be >= 1, got 0, -1}}
+  %0 = "tosa.avg_pool2d"(%arg0, %arg1, %arg2) {kernel = array<i64: 0, -1>, pad = array<i64: 0, 0, 0, 0>, stride = array<i64: 1, 1>, acc_type = f32} :
+         (tensor<1x32x32x8xf32>, tensor<1xf32>, tensor<1xf32>) -> tensor<1x32x32x8xf32>
+  return %0 : tensor<1x32x32x8xf32>
+}
+
+// -----
+
+func.func @test_avgpool2d_invalid_stride(%arg0: tensor<1x32x32x8xf32>, %arg1: tensor<1xf32>, %arg2: tensor<1xf32>) -> tensor<1x32x32x8xf32> {
+  // expected-error@+1 {{'tosa.avg_pool2d' op expect all stride values to be >= 1, got 1, 0}}
+  %0 = "tosa.avg_pool2d"(%arg0, %arg1, %arg2) {kernel = array<i64: 1, 1>, pad = array<i64: 0, 0, 0, 0>, stride = array<i64: 1, 0>, acc_type = f32} :
+         (tensor<1x32x32x8xf32>, tensor<1xf32>, tensor<1xf32>) -> tensor<1x32x32x8xf32>
+  return %0 : tensor<1x32x32x8xf32>
+}
+
+// -----
+
+func.func @test_avgpool2d_invalid_padding(%arg0: tensor<1x32x32x8xf32>, %arg1: tensor<1xf32>, %arg2: tensor<1xf32>) -> tensor<1x32x32x8xf32> {
+  // expected-error@+1 {{'tosa.avg_pool2d' op expect all padding values to be >= 0, got 0, 0, 0, -1}}
+  %0 = "tosa.avg_pool2d"(%arg0, %arg1, %arg2) {kernel = array<i64: 1, 1>, pad = array<i64: 0, 0, 0, -1>, stride = array<i64: 1, 1>, acc_type = f32} :
+         (tensor<1x32x32x8xf32>, tensor<1xf32>, tensor<1xf32>) -> tensor<1x32x32x8xf32>
+  return %0 : tensor<1x32x32x8xf32>
+}
+
+// -----
+
+func.func @test_avgpool2d_padding_not_less_than_kernel_x(%arg0: tensor<1x32x32x8xf32>, %arg1: tensor<1xf32>, %arg2: tensor<1xf32>) -> tensor<1x32x32x8xf32> {
+  // expected-error@+1 {{'tosa.avg_pool2d' op expected left/right padding to be less than the width of the kernel, got pad_left=0, pad_right=1, kernel_x=1}}
+  %0 = "tosa.avg_pool2d"(%arg0, %arg1, %arg2) {kernel = array<i64: 1, 1>, pad = array<i64: 0, 0, 0, 1>, stride = array<i64: 1, 1>, acc_type = f32} :
+         (tensor<1x32x32x8xf32>, tensor<1xf32>, tensor<1xf32>) -> tensor<1x32x32x8xf32>
+  return %0 : tensor<1x32x32x8xf32>
+}
+
+// -----
+
+func.func @test_avgpool2d_padding_not_less_than_kernel_y(%arg0: tensor<1x32x32x8xf32>, %arg1: tensor<1xf32>, %arg2: tensor<1xf32>) -> tensor<1x32x32x8xf32> {
+  // expected-error@+1 {{'tosa.avg_pool2d' op expected top/bottom padding to be less than the height of the kernel, got pad_top=2, pad_bottom=0, kernel_y=1}}
+  %0 = "tosa.avg_pool2d"(%arg0, %arg1, %arg2) {kernel = array<i64: 1, 1>, pad = array<i64: 2, 0, 0, 0>, stride = array<i64: 1, 1>, acc_type = f32} :
+         (tensor<1x32x32x8xf32>, tensor<1xf32>, tensor<1xf32>) -> tensor<1x32x32x8xf32>
+  return %0 : tensor<1x32x32x8xf32>
+}
+
+// -----
+
+func.func @test_avgpool2d_wholly_divisible_height(%arg0: tensor<1x32x32x8xf32>, %arg1: tensor<1xf32>, %arg2: tensor<1xf32>) -> tensor<1x32x32x8xf32> {
+  // expected-error@+1 {{'tosa.avg_pool2d' op expected input_height + pad_top + pad_bottom - kernel_y to be wholly divisible by stride_y, got (32 + 0 + 0 - 1) / 2}}
+  %0 = "tosa.avg_pool2d"(%arg0, %arg1, %arg2) {kernel = array<i64: 1, 1>, pad = array<i64: 0, 0, 0, 0>, stride = array<i64: 2, 1>, acc_type = f32} :
+         (tensor<1x32x32x8xf32>, tensor<1xf32>, tensor<1xf32>) -> tensor<1x32x32x8xf32>
+  return %0 : tensor<1x32x32x8xf32>
+}
+
+// -----
+
+func.func @test_avgpool2d_wholly_divisible_width(%arg0: tensor<1x32x32x8xf32>, %arg1: tensor<1xf32>, %arg2: tensor<1xf32>) -> tensor<1x32x32x8xf32> {
+  // expected-error@+1 {{'tosa.avg_pool2d' op expected input_width + pad_left + pad_right - kernel_x to be wholly divisible by stride_x, got (32 + 0 + 0 - 1) / 2}}
+  %0 = "tosa.avg_pool2d"(%arg0, %arg1, %arg2) {kernel = array<i64: 1, 1>, pad = array<i64: 0, 0, 0, 0>, stride = array<i64: 1, 2>, acc_type = f32} :
+         (tensor<1x32x32x8xf32>, tensor<1xf32>, tensor<1xf32>) -> tensor<1x32x32x8xf32>
+  return %0 : tensor<1x32x32x8xf32>
+}
+
+// -----
+
+func.func @test_avgpool2d_unexpected_output_height(%arg0: tensor<1x32x32x8xf32>, %arg1: tensor<1xf32>, %arg2: tensor<1xf32>) -> tensor<1x33x32x8xf32> {
+  // expected-error@+1 {{'tosa.avg_pool2d' op calculated output height did not match expected: calculated=32, expected=33}}
+  %0 = "tosa.avg_pool2d"(%arg0, %arg1, %arg2) {kernel = array<i64: 1, 1>, pad = array<i64: 0, 0, 0, 0>, stride = array<i64: 1, 1>, acc_type = f32} :
+         (tensor<1x32x32x8xf32>, tensor<1xf32>, tensor<1xf32>) -> tensor<1x33x32x8xf32>
+  return %0 : tensor<1x33x32x8xf32>
+}
+
+// -----
+
+func.func @test_avgpool2d_unexpected_output_width(%arg0: tensor<1x32x32x8xf32>, %arg1: tensor<1xf32>, %arg2: tensor<1xf32>) -> tensor<1x?x33x8xf32> {
+  // expected-error@+1 {{'tosa.avg_pool2d' op calculated output width did not match expected: calculated=32, expected=33}}
+  %0 = "tosa.avg_pool2d"(%arg0, %arg1, %arg2) {kernel = array<i64: 1, 1>, pad = array<i64: 0, 0, 0, 0>, stride = array<i64: 1, 1>, acc_type = f32} :
+         (tensor<1x32x32x8xf32>, tensor<1xf32>, tensor<1xf32>) -> tensor<1x?x33x8xf32>
+  return %0 : tensor<1x?x33x8xf32>
+}
+
+// -----
+
+func.func @test_maxpool2d_invalid_kernel(%arg0: tensor<1x32x32x8xf32>) -> tensor<1x2x32x8xf32> {
+  // expected-error@+1 {{'tosa.max_pool2d' op expect all kernel values to be >= 1, got 0, 1}}
+  %0 = "tosa.max_pool2d"(%arg0) {kernel = array<i64: 0, 1>, pad = array<i64: 0, 0, 0, 0>, stride = array<i64: 1, 1>} :
+         (tensor<1x32x32x8xf32>) -> tensor<1x2x32x8xf32>
+  return %0 : tensor<1x2x32x8xf32>
+}
+
+// -----
+
+func.func @test_maxpool2d_unexpected_output_width(%arg0: tensor<1x32x32x8xf32>) -> tensor<1x32x2x8xf32> {
+  // expected-error@+1 {{'tosa.max_pool2d' op calculated output width did not match expected: calculated=32, expected=2}}
+  %0 = "tosa.max_pool2d"(%arg0) {kernel = array<i64: 1, 1>, pad = array<i64: 0, 0, 0, 0>, stride = array<i64: 1, 1>} :
+         (tensor<1x32x32x8xf32>) -> tensor<1x32x2x8xf32>
+  return %0 : tensor<1x32x2x8xf32>
+}