Fix output size computation for MaxPool2D for ceil_model = true. (llvm#3890)

This PR fixes the output size computation as per
https://github.com/pytorch/pytorch/blob/d8c14838f164ee02b88b6e37471b71bb0373f865/torch/_meta_registrations.py#L3847

```
if ceil_mode:
        if (outputSize - 1) * stride >= inputSize + pad_l:
            outputSize -= 1
    return outputSize
```

Author: sahas3

lib/Conversion/TorchToLinalg/Utils.cpp

@@ -116,6 +116,22 @@ Value torch_to_linalg::getOutputDimForConvOps(OpBuilder &b, Location loc,
   else
     division = b.createOrFold<arith::FloorDivSIOp>(loc, dividend, strideInt);
   Value out = b.createOrFold<arith::AddIOp>(loc, division, c1);
+
+  if (ceilMode) {
+    Value outMinusOneTimesStride =
+        b.createOrFold<arith::MulIOp>(loc, division, strideInt);
+    Value inAddLeftPadding = b.createOrFold<arith::AddIOp>(
+        loc, castIndexToInt64(b, loc, in), paddingInt);
+
+    auto reduceOutputDimCond =
+        b.createOrFold<arith::CmpIOp>(loc, arith::CmpIPredicate::uge,
+                                      outMinusOneTimesStride, inAddLeftPadding);
+
+    auto reducedDim = b.createOrFold<arith::SelectOp>(loc, reduceOutputDimCond,
+                                                      division, out);
+    return castIntToIndex(b, loc, reducedDim);
+  }
+
   return castIntToIndex(b, loc, out);
 }
 

lib/Conversion/TorchToTosa/TorchToTosa.cpp

@@ -5398,9 +5398,11 @@ class ConvertAtenPoolingBaseOp : public OpConversionPattern<AtenOpT> {
     } else {
       int64_t dimSize =
           inputDim + padBefore + padAfter - dilation * (kernelDim - 1) - 1;
-      if (ceilMode && (dimSize % stride != 0))
-        return dimSize / stride + 2;
-      return dimSize / stride + 1;
+      int64_t outputDim = dimSize / stride + 1;
+      if (ceilMode && (dimSize % stride != 0) &&
+          (outputDim * stride < inputDim + padBefore))
+        outputDim++;
+      return outputDim;
     }
   }
 

projects/pt1/e2e_testing/xfail_sets.py

@@ -735,6 +735,7 @@
     "LenStrModule_basic",
     "MaxPool2dCeilModeTrueModule_basic",
     "MaxPool2dStaticCeilModeTrueModule_basic",
+    "MaxPool2dStaticCeilModeTrueReduceOutputModule_basic",
     "MaxPool2dWithIndicesBackwardDynamic3DModule_basic",
     "MaxPool2dWithIndicesBackwardDynamic4DModule_basic",
     "MaxPool2dWithIndicesBackwardStatic3DModule_basic",
@@ -2255,6 +2256,7 @@
     "MatmulStaticBroadcast_basic",
     "MaxPool2dEmptyStrideStaticModule_basic",
     "MaxPool2dStaticCeilModeTrueModule_basic",
+    "MaxPool2dStaticCeilModeTrueReduceOutputModule_basic",
     "MaxPool2dStaticModule_basic",
     "MeanModule_basic",
     "MmDagModule_basic",
@@ -3380,6 +3382,13 @@
         "ScaledDotProductAttentionBoolMaskModule_basic",
     }
 
+if torch_version_for_comparison() > version.parse("2.5.1"):
+    ONNX_XFAIL_SET = ONNX_XFAIL_SET | {
+        # error: 'memref.cast' op operand type 'memref<2x6x4x3xf32>' and result type 'memref<2x6x5x3xf32>' are cast incompatible
+        # torch.onnx.export produces onnx.MaxPool op with incorrect output shape of 2x6x5x3 instead of 2x6x4x3
+        "MaxPool2dStaticCeilModeTrueReduceOutputModule_basic",
+    }
+
 if torch_version_for_comparison() < version.parse("2.4.0.dev"):
     STABLEHLO_PASS_SET = STABLEHLO_PASS_SET - {
         "AtenIntMM_basic",
@@ -4932,3 +4941,10 @@
     "_LogSoftmaxModule_basic",
     "_SoftmaxModule_basic",
 }
+
+if torch_version_for_comparison() > version.parse("2.5.1"):
+    ONNX_TOSA_XFAIL_SET = ONNX_TOSA_XFAIL_SET | {
+        # error: 'memref.cast' op operand type 'memref<2x6x4x3xf32>' and result type 'memref<2x6x5x3xf32>' are cast incompatible
+        # torch.onnx.export produces onnx.MaxPool op with incorrect output shape of 2x6x5x3 instead of 2x6x4x3
+        "MaxPool2dStaticCeilModeTrueReduceOutputModule_basic",
+    }

projects/pt1/python/torch_mlir_e2e_test/test_suite/pooling.py

@@ -420,6 +420,35 @@ def MaxPool2dCeilModeTrueModule_basic(module, tu: TestUtils):
     module.forward(tu.rand(1, 1, 20, 20, low=0.5, high=1.0))
 
 
+class MaxPool2dStaticCeilModeTrueReduceOutputModule(torch.nn.Module):
+    def __init__(self):
+        super().__init__()
+        self.mp2d = torch.nn.MaxPool2d(
+            kernel_size=6,
+            stride=6,
+            padding=3,
+            dilation=1,
+            ceil_mode=True,
+        )
+
+    @export
+    @annotate_args(
+        [
+            None,
+            ([2, 6, 20, 10], torch.float32, True),
+        ]
+    )
+    def forward(self, x):
+        return self.mp2d(x)
+
+
+@register_test_case(
+    module_factory=lambda: MaxPool2dStaticCeilModeTrueReduceOutputModule()
+)
+def MaxPool2dStaticCeilModeTrueReduceOutputModule_basic(module, tu: TestUtils):
+    module.forward(tu.rand(2, 6, 20, 10, low=0.5, high=1.0))
+
+
 # ==============================================================================