upgrade to ONNX 1.17.0 (opset 22) (#3004)

Signed-off-by: Gong Su <[email protected]>

docs/Dialects/krnl.md

@@ -453,7 +453,7 @@ in the `value` dense element attribute.
 
 Traits: `AlwaysSpeculatableImplTrait`, `MemRefsNormalizable`
 
-Interfaces: `ConditionallySpeculatable`, `NoMemoryEffect (MemoryEffectOpInterface)`
+Interfaces: `ConditionallySpeculatable`, `KrnlGlobalOpInterface`, `NoMemoryEffect (MemoryEffectOpInterface)`
 
 Effects: `MemoryEffects::Effect{}`
 

docs/Dialects/onnx.md

@@ -529,7 +529,7 @@ AveragePool consumes an input tensor X and applies average pooling across
  ```
  output_spatial_shape[i] = ceil((input_spatial_shape[i] + pad_shape[i] - dilation[i] * (kernel_shape[i] - 1) - 1) / strides_spatial_shape[i] + 1)
  ```
- if ceil_mode is enabled. `pad_shape[i]` is the sum of pads along axis `i`. Sliding windows that would start in the right padded region are ignored.
+ if ceil_mode is enabled. `pad_shape[i]` is the sum of pads along axis `i`.
 
  `auto_pad` is a DEPRECATED attribute. If you are using them currently, the output spatial shape will be following when ceil_mode is enabled:
  ```
@@ -1701,15 +1701,15 @@ Effects: `MemoryEffects::Effect{}`
 
 | Operand | Description |
 | :-----: | ----------- |
-| `X` | tensor of 16-bit float values or tensor of 32-bit float values or tensor of 64-bit float values
-| `W` | tensor of 16-bit float values or tensor of 32-bit float values or tensor of 64-bit float values
-| `B` | tensor of 16-bit float values or tensor of 32-bit float values or tensor of 64-bit float values or none type
+| `X` | tensor of 16-bit float values or tensor of 16-bit float values or tensor of 32-bit float values or tensor of 64-bit float values
+| `W` | tensor of 16-bit float values or tensor of 16-bit float values or tensor of 32-bit float values or tensor of 64-bit float values
+| `B` | tensor of 16-bit float values or tensor of 16-bit float values or tensor of 32-bit float values or tensor of 64-bit float values or none type
 
 #### Results:
 
 | Result | Description |
 | :----: | ----------- |
-| `Y` | tensor of 16-bit float values or tensor of 32-bit float values or tensor of 64-bit float values
+| `Y` | tensor of 16-bit float values or tensor of 16-bit float values or tensor of 32-bit float values or tensor of 64-bit float values
 
 ### `onnx.ConvTranspose` (ONNXConvTransposeOp)
 
@@ -2610,13 +2610,13 @@ Effects: `MemoryEffects::Effect{}`
 
 | Operand | Description |
 | :-----: | ----------- |
-| `input` | tensor of 16-bit float values or tensor of 32-bit float values or tensor of 64-bit float values or tensor of bfloat16 type values
+| `input` | tensor of bfloat16 type values or tensor of 16-bit float values or tensor of 32-bit float values or tensor of 64-bit float values
 
 #### Results:
 
 | Result | Description |
 | :----: | ----------- |
-| `output` | tensor of 16-bit float values or tensor of 32-bit float values or tensor of 64-bit float values or tensor of bfloat16 type values
+| `output` | tensor of bfloat16 type values or tensor of 16-bit float values or tensor of 32-bit float values or tensor of 64-bit float values
 
 ### `onnx.Expand` (ONNXExpandOp)
 
@@ -3282,13 +3282,13 @@ Effects: `MemoryEffects::Effect{}`
 
 | Operand | Description |
 | :-----: | ----------- |
-| `X` | tensor of 16-bit float values or tensor of 32-bit float values or tensor of 64-bit float values
+| `X` | tensor of bfloat16 type values or tensor of 16-bit float values or tensor of 32-bit float values or tensor of 64-bit float values
 
 #### Results:
 
 | Result | Description |
 | :----: | ----------- |
-| `Y` | tensor of 16-bit float values or tensor of 32-bit float values or tensor of 64-bit float values
+| `Y` | tensor of bfloat16 type values or tensor of 16-bit float values or tensor of 32-bit float values or tensor of 64-bit float values
 
 ### `onnx.GlobalMaxPool` (ONNXGlobalMaxPoolOp)
 
@@ -4817,7 +4817,7 @@ Effects: `MemoryEffects::Effect{}`
 
 _ONNX MatMulInteger operation_
 
-Matrix product that behaves like numpy.matmul: https://docs.scipy.org/doc/numpy-1.13.0/reference/generated/numpy.matmul.html.
+Matrix product that behaves like [numpy.matmul](https://numpy.org/doc/stable/reference/generated/numpy.matmul.html).
 The production MUST never overflow. The accumulation may overflow if and only if in 32 bits.
 
 Traits: `AlwaysSpeculatableImplTrait`
@@ -4845,7 +4845,7 @@ Effects: `MemoryEffects::Effect{}`
 
 _ONNX MatMul operation_
 
-Matrix product that behaves like numpy.matmul: https://docs.scipy.org/doc/numpy-1.13.0/reference/generated/numpy.matmul.html
+Matrix product that behaves like [numpy.matmul](https://numpy.org/doc/stable/reference/generated/numpy.matmul.html).
 
 Traits: `AlwaysSpeculatableImplTrait`
 
@@ -4910,7 +4910,7 @@ MaxPool consumes an input tensor X and applies max pooling across
  ```
  output_spatial_shape[i] = ceil((input_spatial_shape[i] + pad_shape[i] - dilation[i] * (kernel_shape[i] - 1) - 1) / strides_spatial_shape[i] + 1)
  ```
- if ceil_mode is enabled. `pad_shape[i]` is the sum of pads along axis `i`. Sliding windows that would start in the right padded region are ignored.
+ if ceil_mode is enabled. `pad_shape[i]` is the sum of pads along axis `i`.
 
  `auto_pad` is a DEPRECATED attribute. If you are using them currently, the output spatial shape will be following when ceil_mode is enabled:
  ```
@@ -6611,7 +6611,7 @@ Effects: `MemoryEffects::Effect{}`
 
 _ONNX QLinearMatMul operation_
 
-Matrix product that behaves like numpy.matmul: https://docs.scipy.org/doc/numpy-1.13.0/reference/generated/numpy.matmul.html.
+Matrix product that behaves like [numpy.matmul](https://numpy.org/doc/stable/reference/generated/numpy.matmul.html).
 It consumes two quantized input tensors, their scales and zero points, scale and zero point of output,
 and computes the quantized output. The quantization formula is y = saturate((x / y_scale) + y_zero_point).
 For (x / y_scale), it is rounding to nearest ties to even. Refer to https://en.wikipedia.org/wiki/Rounding for details.
@@ -10215,13 +10215,13 @@ Effects: `MemoryEffects::Effect{}`
 
 | Operand | Description |
 | :-----: | ----------- |
-| `input` | tensor of 16-bit float values or tensor of 32-bit float values or tensor of 64-bit float values or tensor of bfloat16 type values
+| `input` | tensor of bfloat16 type values or tensor of 16-bit float values or tensor of 32-bit float values or tensor of 64-bit float values
 
 #### Results:
 
 | Result | Description |
 | :----: | ----------- |
-| `output` | tensor of 16-bit float values or tensor of 32-bit float values or tensor of 64-bit float values or tensor of bfloat16 type values
+| `output` | tensor of bfloat16 type values or tensor of 16-bit float values or tensor of 32-bit float values or tensor of 64-bit float values
 
 ### `onnx.TfIdfVectorizer` (ONNXTfIdfVectorizerOp)
 

docs/Dialects/zhigh.md

@@ -793,6 +793,8 @@ Effects: `MemoryEffects::Effect{}`
 _ZHigh Stickified Constant operation_
 
 This operator produces a constant tensor to store stickified data.
+`value` attribute has original constant or stickified constant.
+`stickified` attribute indicates the `value` is already stickified or not.
 Stickified data is opaque and must be 4K-aligned. One who produces
 the stickified data must make sure its size in bytes consistent with
 the output tensor's size.
@@ -807,6 +809,7 @@ Effects: `MemoryEffects::Effect{}`
 
 <table>
 <tr><th>Attribute</th><th>MLIR Type</th><th>Description</th></tr>
+<tr><td><code>stickified</code></td><td>::mlir::BoolAttr</td><td>bool attribute</td></tr>
 <tr><td><code>value</code></td><td>::mlir::Attribute</td><td>any attribute</td></tr>
 <tr><td><code>alignment</code></td><td>::mlir::IntegerAttr</td><td>64-bit signless integer attribute</td></tr>
 </table>

docs/Dialects/zlow.md

@@ -752,6 +752,34 @@ Interfaces: `MemoryEffectOpInterface`
 | `X` | memref of 16-bit float or 32-bit float values
 | `Out` | memref of dlfloat16 type values
 
+### `zlow.stickifiedConstant` (::onnx_mlir::zlow::ZLowStickifiedConstantOp)
+
+_ZLow Stickified Constant operation._
+
+
+Traits: `MemRefsNormalizable`
+
+Interfaces: `KrnlGlobalOpInterface`
+
+#### Attributes:
+
+<table>
+<tr><th>Attribute</th><th>MLIR Type</th><th>Description</th></tr>
+<tr><td><code>shape</code></td><td>::mlir::Attribute</td><td>any attribute</td></tr>
+<tr><td><code>name</code></td><td>::mlir::StringAttr</td><td>string attribute</td></tr>
+<tr><td><code>stickified</code></td><td>::mlir::BoolAttr</td><td>bool attribute</td></tr>
+<tr><td><code>value</code></td><td>::mlir::Attribute</td><td>any attribute</td></tr>
+<tr><td><code>layout</code></td><td>::mlir::StringAttr</td><td>string attribute</td></tr>
+<tr><td><code>offset</code></td><td>::mlir::IntegerAttr</td><td>64-bit signless integer attribute</td></tr>
+<tr><td><code>alignment</code></td><td>::mlir::IntegerAttr</td><td>64-bit signless integer attribute</td></tr>
+</table>
+
+#### Results:
+
+| Result | Description |
+| :----: | ----------- |
+| `output` | memref of dlfloat16 type values
+
 ### `zlow.sub` (::onnx_mlir::zlow::ZLowSubOp)
 
 _ZLow sub operation_

src/Dialect/ONNX/ONNXOps.td.inc

@@ -426,7 +426,7 @@ def ONNXAveragePoolOp:ONNX_Op<"AveragePool",
    ```
    output_spatial_shape[i] = ceil((input_spatial_shape[i] + pad_shape[i] - dilation[i] * (kernel_shape[i] - 1) - 1) / strides_spatial_shape[i] + 1)
    ```
-   if ceil_mode is enabled. `pad_shape[i]` is the sum of pads along axis `i`. Sliding windows that would start in the right padded region are ignored.
+   if ceil_mode is enabled. `pad_shape[i]` is the sum of pads along axis `i`.
 
    `auto_pad` is a DEPRECATED attribute. If you are using them currently, the output spatial shape will be following when ceil_mode is enabled:
    ```
@@ -1400,16 +1400,16 @@ def ONNXConvOp:ONNX_Op<"Conv",
   The convolution operator consumes an input tensor and a filter, and
   computes the output.
   }];
-  let arguments = (ins AnyTypeOf<[TensorOf<[F16]>, TensorOf<[F32]>, TensorOf<[F64]>]>:$X,
-    AnyTypeOf<[TensorOf<[F16]>, TensorOf<[F32]>, TensorOf<[F64]>]>:$W,
-    AnyTypeOf<[TensorOf<[F16]>, TensorOf<[F32]>, TensorOf<[F64]>, NoneType]>:$B,
+  let arguments = (ins AnyTypeOf<[TensorOf<[F16]>, TensorOf<[F16]>, TensorOf<[F32]>, TensorOf<[F64]>]>:$X,
+    AnyTypeOf<[TensorOf<[F16]>, TensorOf<[F16]>, TensorOf<[F32]>, TensorOf<[F64]>]>:$W,
+    AnyTypeOf<[TensorOf<[F16]>, TensorOf<[F16]>, TensorOf<[F32]>, TensorOf<[F64]>, NoneType]>:$B,
     DefaultValuedStrAttr<StrAttr, "NOTSET">:$auto_pad,
     OptionalAttr<I64ArrayAttr>:$dilations,
     DefaultValuedAttr<SI64Attr, "1">:$group,
     OptionalAttr<I64ArrayAttr>:$kernel_shape,
     OptionalAttr<I64ArrayAttr>:$pads,
     OptionalAttr<I64ArrayAttr>:$strides);
-  let results = (outs AnyTypeOf<[TensorOf<[F16]>, TensorOf<[F32]>, TensorOf<[F64]>]>:$Y);
+  let results = (outs AnyTypeOf<[TensorOf<[F16]>, TensorOf<[F16]>, TensorOf<[F32]>, TensorOf<[F64]>]>:$Y);
   let builders = [
     OpBuilder<(ins "Value":$X, "Value":$W, "Value":$B, "StringAttr":$auto_pad, "ArrayAttr":$dilations, "IntegerAttr":$group, "ArrayAttr":$kernel_shape, "ArrayAttr":$pads, "ArrayAttr":$strides), [{
       auto resultType = UnrankedTensorType::get(mlir::cast<ShapedType>(X.getType()).getElementType());
@@ -2231,8 +2231,8 @@ def ONNXExpOp:ONNX_Op<"Exp",
   let description = [{
   Calculates the exponential of the given input tensor, element-wise.
   }];
-  let arguments = (ins AnyTypeOf<[TensorOf<[F16]>, TensorOf<[F32]>, TensorOf<[F64]>, TensorOf<[BF16]>]>:$input);
-  let results = (outs AnyTypeOf<[TensorOf<[F16]>, TensorOf<[F32]>, TensorOf<[F64]>, TensorOf<[BF16]>]>:$output);
+  let arguments = (ins AnyTypeOf<[TensorOf<[BF16]>, TensorOf<[F16]>, TensorOf<[F32]>, TensorOf<[F64]>]>:$input);
+  let results = (outs AnyTypeOf<[TensorOf<[BF16]>, TensorOf<[F16]>, TensorOf<[F32]>, TensorOf<[F64]>]>:$output);
   let builders = [
     OpBuilder<(ins "Value":$input), [{
       auto resultType = UnrankedTensorType::get(mlir::cast<ShapedType>(input.getType()).getElementType());
@@ -2892,9 +2892,9 @@ def ONNXGlobalLpPoolOp:ONNX_Op<"GlobalLpPool",
    the values in the same channel. This is equivalent to LpPool with kernel size
    equal to the spatial dimension of input tensor.
   }];
-  let arguments = (ins AnyTypeOf<[TensorOf<[F16]>, TensorOf<[F32]>, TensorOf<[F64]>]>:$X,
+  let arguments = (ins AnyTypeOf<[TensorOf<[BF16]>, TensorOf<[F16]>, TensorOf<[F32]>, TensorOf<[F64]>]>:$X,
     DefaultValuedAttr<SI64Attr, "2">:$p);
-  let results = (outs AnyTypeOf<[TensorOf<[F16]>, TensorOf<[F32]>, TensorOf<[F64]>]>:$Y);
+  let results = (outs AnyTypeOf<[TensorOf<[BF16]>, TensorOf<[F16]>, TensorOf<[F32]>, TensorOf<[F64]>]>:$Y);
   let extraClassDeclaration = [{
     static int getNumberOfOperands() {
       return 1;
@@ -4242,7 +4242,7 @@ def ONNXMatMulOp:ONNX_Op<"MatMul",
   [Pure, DeclareOpInterfaceMethods<ShapeInferenceOpInterface>, DeclareOpInterfaceMethods<ShapeHelperOpInterface>]> {
   let summary = "ONNX MatMul operation";
   let description = [{
-  Matrix product that behaves like numpy.matmul: https://docs.scipy.org/doc/numpy-1.13.0/reference/generated/numpy.matmul.html
+  Matrix product that behaves like [numpy.matmul](https://numpy.org/doc/stable/reference/generated/numpy.matmul.html).
   }];
   let arguments = (ins AnyTypeOf<[TensorOf<[F16]>, TensorOf<[F32]>, TensorOf<[F64]>, TensorOf<[UI32]>, TensorOf<[UI64]>, TensorOf<[I32]>, TensorOf<[I64]>, TensorOf<[BF16]>]>:$A,
     AnyTypeOf<[TensorOf<[F16]>, TensorOf<[F32]>, TensorOf<[F64]>, TensorOf<[UI32]>, TensorOf<[UI64]>, TensorOf<[I32]>, TensorOf<[I64]>, TensorOf<[BF16]>]>:$B);
@@ -4272,7 +4272,7 @@ def ONNXMatMulIntegerOp:ONNX_Op<"MatMulInteger",
   [Pure, DeclareOpInterfaceMethods<ShapeInferenceOpInterface>, DeclareOpInterfaceMethods<ShapeHelperOpInterface>]> {
   let summary = "ONNX MatMulInteger operation";
   let description = [{
-  Matrix product that behaves like numpy.matmul: https://docs.scipy.org/doc/numpy-1.13.0/reference/generated/numpy.matmul.html.
+  Matrix product that behaves like [numpy.matmul](https://numpy.org/doc/stable/reference/generated/numpy.matmul.html).
   The production MUST never overflow. The accumulation may overflow if and only if in 32 bits.
   }];
   let arguments = (ins AnyTypeOf<[TensorOf<[I8]>, TensorOf<[UI8]>]>:$A,
@@ -4352,7 +4352,7 @@ def ONNXMaxPoolOp:ONNX_Op<"MaxPool",
    ```
    output_spatial_shape[i] = ceil((input_spatial_shape[i] + pad_shape[i] - dilation[i] * (kernel_shape[i] - 1) - 1) / strides_spatial_shape[i] + 1)
    ```
-   if ceil_mode is enabled. `pad_shape[i]` is the sum of pads along axis `i`. Sliding windows that would start in the right padded region are ignored.
+   if ceil_mode is enabled. `pad_shape[i]` is the sum of pads along axis `i`.
 
    `auto_pad` is a DEPRECATED attribute. If you are using them currently, the output spatial shape will be following when ceil_mode is enabled:
    ```
@@ -5921,7 +5921,7 @@ def ONNXQLinearMatMulOp:ONNX_Op<"QLinearMatMul",
   [Pure, DeclareOpInterfaceMethods<ShapeInferenceOpInterface>, DeclareOpInterfaceMethods<ShapeHelperOpInterface>]> {
   let summary = "ONNX QLinearMatMul operation";
   let description = [{
-  Matrix product that behaves like numpy.matmul: https://docs.scipy.org/doc/numpy-1.13.0/reference/generated/numpy.matmul.html.
+  Matrix product that behaves like [numpy.matmul](https://numpy.org/doc/stable/reference/generated/numpy.matmul.html).
   It consumes two quantized input tensors, their scales and zero points, scale and zero point of output,
   and computes the quantized output. The quantization formula is y = saturate((x / y_scale) + y_zero_point).
   For (x / y_scale), it is rounding to nearest ties to even. Refer to https://en.wikipedia.org/wiki/Rounding for details.
@@ -9511,8 +9511,8 @@ def ONNXTanhOp:ONNX_Op<"Tanh",
   let description = [{
   Calculates the hyperbolic tangent of the given input tensor element-wise.
   }];
-  let arguments = (ins AnyTypeOf<[TensorOf<[F16]>, TensorOf<[F32]>, TensorOf<[F64]>, TensorOf<[BF16]>]>:$input);
-  let results = (outs AnyTypeOf<[TensorOf<[F16]>, TensorOf<[F32]>, TensorOf<[F64]>, TensorOf<[BF16]>]>:$output);
+  let arguments = (ins AnyTypeOf<[TensorOf<[BF16]>, TensorOf<[F16]>, TensorOf<[F32]>, TensorOf<[F64]>]>:$input);
+  let results = (outs AnyTypeOf<[TensorOf<[BF16]>, TensorOf<[F16]>, TensorOf<[F32]>, TensorOf<[F64]>]>:$output);
   let extraClassDeclaration = [{
     static int getNumberOfOperands() {
       return 1;

test/backend/all_test_names.txt

@@ -1,5 +1,5 @@
 # This file is automatically generated by "make check-onnx-backend-case"
-# From onnx 1.16.2
+# From onnx 1.17.0
 # All test cases for cpu target
 test_bvlc_alexnet_cpu
 test_densenet121_cpu
@@ -303,6 +303,8 @@ test_convtranspose_3d_cpu
 test_convtranspose_autopad_same_cpu
 test_convtranspose_cpu
 test_convtranspose_dilations_cpu
+test_convtranspose_group_2_cpu
+test_convtranspose_group_2_image_3_cpu
 test_convtranspose_kernel_shape_cpu
 test_convtranspose_output_shape_cpu
 test_convtranspose_pad_cpu
@@ -880,6 +882,7 @@ test_reduce_max_default_axes_keepdim_example_cpu
 test_reduce_max_default_axes_keepdims_random_cpu
 test_reduce_max_do_not_keepdims_example_cpu
 test_reduce_max_do_not_keepdims_random_cpu
+test_reduce_max_empty_set_cpu
 test_reduce_max_keepdims_example_cpu
 test_reduce_max_keepdims_random_cpu
 test_reduce_max_negative_axes_keepdims_example_cpu
@@ -915,6 +918,7 @@ test_reduce_sum_default_axes_keepdims_example_cpu
 test_reduce_sum_default_axes_keepdims_random_cpu
 test_reduce_sum_do_not_keepdims_example_cpu
 test_reduce_sum_do_not_keepdims_random_cpu
+test_reduce_sum_empty_axes_input_noop_cpu
 test_reduce_sum_empty_axes_input_noop_example_cpu
 test_reduce_sum_empty_set_cpu
 test_reduce_sum_empty_set_non_reduced_axis_zero_cpu
@@ -975,6 +979,7 @@ test_resize_downsample_sizes_nearest_not_smaller_cpu
 test_resize_tf_crop_and_resize_axes_2_3_cpu
 test_resize_tf_crop_and_resize_axes_3_2_cpu
 test_resize_tf_crop_and_resize_cpu
+test_resize_tf_crop_and_resize_extrapolation_value_cpu
 test_resize_upsample_scales_cubic_A_n0p5_exclude_outside_cpu
 test_resize_upsample_scales_cubic_align_corners_cpu
 test_resize_upsample_scales_cubic_asymmetric_cpu
@@ -992,6 +997,7 @@ test_resize_upsample_sizes_nearest_ceil_half_pixel_cpu
 test_resize_upsample_sizes_nearest_cpu
 test_resize_upsample_sizes_nearest_floor_align_corners_cpu
 test_resize_upsample_sizes_nearest_not_larger_cpu
+test_resize_upsample_sizes_nearest_not_smaller_cpu
 test_resize_upsample_sizes_nearest_round_prefer_ceil_asymmetric_cpu
 test_reversesequence_batch_cpu
 test_reversesequence_time_cpu

utils/gen_onnx_mlir.py

@@ -66,7 +66,7 @@
 
 # ==UPDATE_ONNX_VERSION_OPSET==
 # Look for tag above and update all references when upgrading the ONNX support within ONNX-MLIR.
-current_onnx_version = "1.16.2"
+current_onnx_version = "1.17.0"
 
 # Check the version of onnx package being used.
 if (