[Mlir-commits] [mlir] [Tosa] : Fix integer overflow for computing intmax+1 in tosa.cast to linalg. (PR #112455)

[Sayan Saha]

https://github.com/sahas3 created https://github.com/llvm/llvm-project/pull/112455

This PR fixes an issue related to integer overflow when computing `(intmax+1)` for `i64` during `tosa-to-linalg` pass for `tosa.cast`.

Found this issue while debugging a numerical mismatch for `deeplabv3` model from `torchvision` represented in `tosa` dialect using the `TorchToTosa` pipeline in `torch-mlir` repository. `torch.aten.to.dtype` is converted to `tosa.cast` that casts `f32` to `i64` type. Technically by the specification, `tosa.cast` doesn't handle casting `f32` to `i64`. So it's possible to add a verifier to error out for such tosa ops instead of producing incorrect code. However, I chose to fix the overflow issue to still be able to represent the `deeplabv3` model with `tosa` ops in the above-mentioned pipeline. Open to suggestions if adding the verifier is more appropriate instead.

>From ec18d2b58a8c8c2f64d264946df19cbf5bb564f1 Mon Sep 17 00:00:00 2001
From: Sayan Saha <sayans at mathworks.com>
Date: Tue, 15 Oct 2024 20:10:27 -0400
Subject: [PATCH] [Tosa] : Fix integer overflow for computing intmax+1 in
 tosa.cast to linalg.

---
 .../Conversion/TosaToLinalg/TosaToLinalg.cpp  |  2 +-
 .../TosaToLinalg/tosa-to-linalg.mlir          | 27 +++++++++++++++++++
 2 files changed, 28 insertions(+), 1 deletion(-)

diff --git a/mlir/lib/Conversion/TosaToLinalg/TosaToLinalg.cpp b/mlir/lib/Conversion/TosaToLinalg/TosaToLinalg.cpp
index c88f4db27c304e..e6b3e4b677e4f2 100644
--- a/mlir/lib/Conversion/TosaToLinalg/TosaToLinalg.cpp
+++ b/mlir/lib/Conversion/TosaToLinalg/TosaToLinalg.cpp
@@ -563,7 +563,7 @@ static Value createLinalgBodyCalculationForElementwiseOp(
                    getElementTypeOrSelf(srcTy),
                    APInt::getSignedMaxValue(dstTy.getIntOrFloatBitWidth())
                            .getSExtValue() +
-                       1));
+                       1.0));
 
       auto intMax = rewriter.create<arith::ConstantOp>(
           loc, rewriter.getIntegerAttr(
diff --git a/mlir/test/Conversion/TosaToLinalg/tosa-to-linalg.mlir b/mlir/test/Conversion/TosaToLinalg/tosa-to-linalg.mlir
index f9d37f9427d4f4..7e2ec67d38d378 100644
--- a/mlir/test/Conversion/TosaToLinalg/tosa-to-linalg.mlir
+++ b/mlir/test/Conversion/TosaToLinalg/tosa-to-linalg.mlir
@@ -1929,3 +1929,30 @@ func.func @test_dynamic_fft2d(%arg0: tensor<?x?x?xf32>, %arg1: tensor<?x?x?xf32>
   %output_real, %output_imag = "tosa.fft2d"(%arg0, %arg1) {inverse = true} : (tensor<?x?x?xf32>, tensor<?x?x?xf32>) -> (tensor<?x?x?xf32>, tensor<?x?x?xf32>)
   return %output_real, %output_imag : tensor<?x?x?xf32>, tensor<?x?x?xf32>
 }
+
+
+// -----
+
+// CHECK: #[[$MAP0:.+]] = affine_map<(d0) -> (0)>
+// CHECK: #[[$MAP1:.+]] = affine_map<(d0) -> (d0)>
+
+// CHECK-LABEL:   func.func @test_cast_fp32_i64(
+// CHECK-SAME:                                  %[[ARG0:.*]]: tensor<1xf32>) -> tensor<1xi64> {
+// CHECK:           %[[VAL_0:.*]] = tensor.empty() : tensor<1xi64>
+// CHECK:           %[[RESULT:.*]] = linalg.generic {indexing_maps = [#[[$MAP0]], #[[$MAP1]]], iterator_types = ["parallel"]} ins(%[[ARG0]] : tensor<1xf32>) outs(%[[VAL_0]] : tensor<1xi64>) {
+// CHECK:           ^bb0(%[[VAL_2:.*]]: f32, %[[VAL_3:.*]]: i64):
+// CHECK:             %[[VAL_4:.*]] = math.roundeven %[[VAL_2]] : f32
+// CHECK:             %[[VAL_5:.*]] = arith.constant -9.22337203E+18 : f32
+// CHECK:             %[[VAL_6:.*]] = arith.constant 9.22337203E+18 : f32
+// CHECK:             %[[VAL_7:.*]] = arith.constant 9223372036854775807 : i64
+// CHECK:             %[[VAL_8:.*]] = arith.maximumf %[[VAL_4]], %[[VAL_5]] : f32
+// CHECK:             %[[VAL_9:.*]] = arith.fptosi %[[VAL_8]] : f32 to i64
+// CHECK:             %[[VAL_10:.*]] = arith.cmpf uge, %[[VAL_4]], %[[VAL_6]] : f32
+// CHECK:             %[[VAL_11:.*]] = arith.select %[[VAL_10]], %[[VAL_7]], %[[VAL_9]] : i64
+// CHECK:             linalg.yield %[[VAL_11]] : i64
+// CHECK:           } -> tensor<1xi64>
+// CHECK:           return %[[RESULT]] : tensor<1xi64>
+func.func @test_cast_fp32_i64(%arg0: tensor<1xf32>) -> (tensor<1xi64>) {
+  %0 = tosa.cast %arg0 : (tensor<1xf32>) -> tensor<1xi64>
+  return %0: tensor<1xi64>
+}