<table border="1" cellspacing="0" cellpadding="8">
    <tr>
        <th>Issue</th>
        <td>
            <a href=https://github.com/llvm/llvm-project/issues/130886>130886</a>
        </td>
    </tr>

    <tr>
        <th>Summary</th>
        <td>
            tosa::ReduceSum Op goes wrong with type Float16.
        </td>
    </tr>

    <tr>
      <th>Labels</th>
      <td>
            new issue
      </td>
    </tr>

    <tr>
      <th>Assignees</th>
      <td>
      </td>
    </tr>

    <tr>
      <th>Reporter</th>
      <td>
          CBalaa
      </td>
    </tr>
</table>

<pre>
    My mlir code output wrongly because of tosa::ReduceOp, I try to lower tosa::ReduceOp to linalg.Reduce and scf.for, and they all go wrong! 

Platform: X86 Ubuntu 22.04

I lower tosa::ReduceOp to scf::for op, I think it is logically correct. But my code expected to output [5120] on x86 platform, but got [2048], I wonder why and how to fix it:

```mlir
module attributes {test.weight = dense<1.000000e+00> : tensor<5120xf16>} {
  memref.global "private" constant @__constant_5120xf16 : memref<5120xf16> = dense<1.000000e+00> {alignment = 64 : i64}
  func.func private @printMemrefF16(memref<*xf16>)
 func.func private @main() {
    %c1 = arith.constant 1 : index
    %c5120 = arith.constant 5120 : index
    %c0 = arith.constant 0 : index
    %cst = arith.constant 0.000000e+00 : f16
    %0 = memref.get_global @__constant_5120xf16 : memref<5120xf16>
    %alloc = memref.alloc() {alignment = 64 : i64} : memref<f16>
    memref.store %cst, %alloc[] : memref<f16>
    scf.for %arg0 = %c0 to %c5120 step %c1 {
      %1 = memref.load %0[%arg0] : memref<5120xf16>
      %2 = memref.load %alloc[] : memref<f16>
      %3 = arith.addf %1, %2 : f16
      memref.store %3, %alloc[] : memref<f16>
    }

    %expand_shape = memref.expand_shape %alloc [] output_shape [1] : memref<f16> into memref<1xf16>
    %cast = memref.cast %expand_shape : memref<1xf16> to memref<*xf16>
    call @printMemrefF16(%cast) : (memref<*xf16>) -> ()
    return
 }
}
```

makefile: (variables like MLIR_OPT should change to your executable.)
```makefile
#!/bin/bash
MLIR_OPT := ../../thirdparty/llvm/build/bin/mlir-opt
MLIR_TRANSLATE := ../../thirdparty/llvm/build/bin/mlir-translate
MLIR_CPU_RUNNER := ../../thirdparty/llvm/build/bin/mlir-cpu-runner
LLC := ../../thirdparty/llvm/build/bin/llc
OPT_FLAG := -O0

MLIR_LIB := ../../thirdparty/llvm/build/lib
AIPILER_PY_PATH := ../../python

MLIR_RUNNER_UTILS := ../../thirdparty/llvm/build/lib/libmlir_runner_utils.so
MLIR_C_RUNNER_UTILS := ../../thirdparty/llvm/build/lib/libmlir_c_runner_utils.so
MTRIPLE := x86_64-unknown-linux-gnu

test: test.mlir
        ${MLIR_OPT}  $^ \
                -pass-pipeline="builtin.module(func.func(tosa-to-linalg-named),func.func(tosa-to-linalg),func.func(tosa-to-tensor),func.func(tosa-to-arith), \
                eliminate-empty-tensors, empty-tensor-to-alloc-tensor, one-shot-bufferize{bufferize-function-boundaries}, \
                convert-linalg-to-loops, convert-vector-to-scf, expand-strided-metadata, convert-vector-to-llvm, \
                arith-expand, lower-affine, convert-arith-to-llvm, convert-math-to-llvm, convert-math-to-libm, \
                finalize-memref-to-llvm, convert-scf-to-cf)" --mlir-print-ir-after-failure | \
        ${MLIR_OPT}     -llvm-request-c-wrappers | \
        ${MLIR_OPT}     -pass-pipeline="builtin.module(convert-arith-to-llvm, convert-math-to-llvm, convert-math-to-libm, convert-func-to-llvm, reconcile-unrealized-casts)" | \
        ${MLIR_CPU_RUNNER} ${OPT_FLAG} -e main -entry-point-result=void \
                -shared-libs=${MLIR_RUNNER_UTILS} -shared-libs=${MLIR_C_RUNNER_UTILS}

```
</pre>
<img width="1" height="1" alt="" src="http://email.email.llvm.org/o/eJysV0Fv2zwS_TXMhZAgUZatHHyQ7Xg3gNsEaQrsngxKGlncUqSWpGJ7f_2CpGQ7sdOvLb6iMCJy5r2ZN6MhRbVmOwEwR-kCpas72ptGqvlyQTmld4WsjvMvR9xypnApK8CyN11v8F5JseNHXEBJew1Y1thITVGSoyR_gaov4alDZIkfsVFHbCTmcg_qhpHbY4LyXeiXMBUV1mUd1lJZBPtoGjhiyjneSU-NSIxRlKMof-bU1FK1KMnxv7Ip_l70wvSYkDCaeIvHn3LrsvaLtVRYjjE3TPzAzGCmMZc7VlLOj7iUSkFpQrzoDW6PXhA4dFAaqCzWIA5KF2lMIpSusBT4kE1xNwZJlrjoDd5JZ0WiSYbSlefcS1GBwvvm6FJu5N5C1uyAmbERulzQNPL_bUVQlLey6jlgaoxiRW9AYzRbGNAm3APbNQajZIUrEBpQsozDyP0DRBZRhJIHbEUzILRUKFnamA91PEXJA5qtLBCKcoxbaBXU4Y7LgnKMCOkUe6MGECG4lEIbKgxGk2i7HZ-2I5LD9_7v8f8qrNmCcrYTLQifwXTioNh0gmYrF1XdizK0P3gIx4bQKSbMF8e3jqeIZCduRPIxNXJvAW76t5QJRDJE7k_ZY4xIWsYuCqqYacJTzrGPSVRwuDC1Wd6yHtavHW5a3zbV5qbtO_2co0315OYJxjKC2Y6l_L2anQEp57K8BHULZ-F-Urv36O-AByxtpIIhW_tijHx-Pv3EfxgZzkHtfNJeXyPPldEGurGkpxK7rOLLjLikldPO0nrAK_ZrbRwOuYXzayk4_-SixrSqahfboAT5UN1r1ZLf0sy_TefKwqGjotrqhnZwmcb79VMHeHg_9cbNdBF_womZMPK8GF_1VUmHBh9Y_fNVVPk1Br4EvnjXPbSd3jfHw8Dp2jbJ8acDAwduKrkGHzAVmF4J-zBo6H_H8exVbekPqBmHAfyNKkYLDhpz9gPwl83jy_bp-RXrRva8wmVDxQ5sKkfZKwwHKHtj7UNPe579I2yUI5IgEiOyLuzkWhdUNyjKT8juZFvhMERk7X5Mw1TVUWWOiKw5f2utU894dYKwB0sgOzPCvL7kX79t8teHPwYzigrN7YkxQC6fv29fvn_9-vDyx5hl1weqFwLsIbjZLP8EiPMSRfnT8-t2vcn_MSIET0PxXKibx8XvQXNWoCjPH58fNw8v2-d_b5_z139eQ3RH00hxQeT12H5_fdx8-31G92uF2XpRtr1hXIdaniT_2wjKWxSvL4_Pm1OHHLLpdjoJevFDyL0IOBP9IdiJ3qdrbyf-3qFNOFxjUHSPyATNFmPn2qMC26X0AaN0OZhE90FHtQ461gFnAlCyQoTYOA0Tob8LIZKdznZEMnvpC4wM_BUzELSFyr5OZPm51ef7w13p0303tf32u7CBs5YJaiCAtjPHAUdbs8sFB2FH64loiaWAQDfSBEVf16DY_wDNFqe_AxuAYVIEhexFRRUDbUfRB_5SijdQZlTBZipl5_jHrTcojY_A3ohtYG7qBtooVkEVtGBoRQ297eNb5j2pEyPwMHbP3cIDWte2chco3u4CZNxo6V-ss-KKtLYZWmH8JL_lrku3bLO8t5fYIHAjxZ0OAbMRGlBBTRnv7Zk6uyT42KS2JS1-oOC_PWgTlMFe0a4DpX_B81ea-W9TaVy3HXNpr6CUomQcgl4ocOJVgT0W9aDPZ3mcx7i72LmNcZralQCwvVDjAIRRx6CTVl8FuucGJas3yar3r7ZuqILKxqudFieiy9HlgD-1XH60_fDVdFfNk-o-uad3MI9nk3iWksk0u2vmZBZP6qIsaVakdVRms3uoJnGWEFJPZ8UsuWNzEpE0SmISTaLZJAvTOkvLdFYkkyhL07hEkwhaynhodQ2l2t0xrXuYx0mUZdM7Tgvg2n1lEyJgj90uIsR-dKu566Gi32k0iTjTRp9hDDMc5h8_Xr_1LX7q8E6C9l_EeM9Mg82xA7zmkpp4Gt71is8bYzptPckakfWOmaYvwlK25znvuDsl_wOlQWTtAtOIrIfI3-bk_wEAAP__6-3kVQ">