<table border="1" cellspacing="0" cellpadding="8">
    <tr>
        <th>Issue</th>
        <td>
            <a href=https://github.com/llvm/llvm-project/issues/134907>134907</a>
        </td>
    </tr>

    <tr>
        <th>Summary</th>
        <td>
            [mlir][tosa] : pool op verification changes are possibly too restrictive?
        </td>
    </tr>

    <tr>
      <th>Labels</th>
      <td>
            mlir
      </td>
    </tr>

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

    <tr>
      <th>Reporter</th>
      <td>
          sahas3
      </td>
    </tr>
</table>

<pre>
    The following `tfl` IR 
```
func.func @pool(%arg0 : tensor<1x60x59x5xf32>) -> tensor<1x30x29x5xf32> {
  %0 = "tfl.max_pool_2d"(%arg0) <{filter_height = 2 : i32, filter_width = 2 : i32, fused_activation_function = "NONE", padding = "VALID", stride_h = 2 : i32, stride_w = 2 : i32}> : (tensor<1x60x59x5xf32>) -> tensor<1x30x29x5xf32>
  return %0 : tensor<1x30x29x5xf32>
}
```

produces the `tosa` IR using the `tf-opt` executable available in `tensorflow` repository:

```
func.func @pool(%arg0: tensor<1x60x59x5xf32>) -> tensor<1x30x29x5xf32> {
    %0 = tosa.max_pool2d %arg0 {kernel = array<i64: 2, 2>, pad = array<i64: 0, 0, 0, 0>, stride = array<i64: 2, 2>} : (tensor<1x60x59x5xf32>) -> tensor<1x30x29x5xf32>
    return %0 : tensor<1x30x29x5xf32>
  }
```

The above IR used to lower to `linalg` fine:
```
func.func @pool(%arg0: tensor<1x60x59x5xf32>) -> tensor<1x30x29x5xf32> {
    %cst = arith.constant -3.40282347E+38 : f32
    %0 = tensor.empty() : tensor<1x30x29x5xf32>
 %1 = linalg.fill ins(%cst : f32) outs(%0 : tensor<1x30x29x5xf32>) -> tensor<1x30x29x5xf32>
    %2 = tensor.empty() : tensor<2x2xf32>
    %3 = linalg.pooling_nhwc_max {dilations = dense<1> : vector<2xi64>, strides = dense<2> : vector<2xi64>} ins(%arg0, %2 : tensor<1x60x59x5xf32>, tensor<2x2xf32>) outs(%1 : tensor<1x30x29x5xf32>) -> tensor<1x30x29x5xf32>
 return %3 : tensor<1x30x29x5xf32>
  }
```

But now due to https://github.com/llvm/llvm-project/pull/130052 the `tosa` IR is failing with 'tosa.max_pool2d' op expected input_width + pad_left + pad_right - kernel_x to be wholly divisible by stride_x, got (59 + 0 + 0 - 2) / 2

The padding values for `VALID` padding is [determined](https://www.tensorflow.org/api_docs/python/tf/keras/layers/MaxPool2D) as 
> The resulting output when using the "valid" padding option has a spatial shape (number of rows or columns) of: output_shape = math.floor((input_shape - pool_size) / strides) + 1 

It seems to me this check is too restrictive and doesn't take into account the above equation properly as `input + pad_left + pad_right - kernel` can be not a multiple of `stride`. 

Should this check be reverted or the lowering from `tfl` to `tosa` should be adjusted somehow to honor this check? 

Any thoughts @lhutton1, @GeorgeARM, @sjarus?

Thanks! 
</pre>
<img width="1" height="1" alt="" src="http://email.email.llvm.org/o/eJzEV9-P4rYT_2vMy2iRsQmEBx5gWb466XvX6lr1FZlkQnxr7NR2CPSvr8YJu-x2e7fXnlRpxSbMeH585jMzRoWgDxZxybI1yzYj1cba-WVQtQpytHflZflrjVA5Y1yn7QHYjMfKsBmHD5-B8RWb8eGPr6rWFmP6ADbljXOGiZyJTPkDByZXENEG55m8n5xn_Jwtztm5koLJByYWcMfkw62G5GfxrAFsvmZ8BcBERsY2wISIlRkf1XlHrnaiZEI8OySTTN6z-brSJqLf1agPdUxHRYpGS8HEPQziTpexfkPaBix3qoj6pKJ2dkfp0cM1hk8_fXpIju-hUWWZIOolv63-_2EziEL0usTdGw4GSfdKMt-kpOUKmMj_BW4JM4-x9fYK3eqr2uT5ZVUZXzXelW2BAWKNiQEuqIECbaCUr99Xd66JJMEzFm1Ue4OgTkqb9KRtUkreK-M6UvTYuKCj8xcmV723d1LqRzHqhlOU1xOjRAlP5J2vH9FbNElLea8uTN7r2ZRiSGXsfSYOvKXDSXb70Wv3xf-60fnmR9Hg-4gA8CYVaBqovTthX3wsITowrkNPD2zGjbbKHKi0lbY4FPU_qGgR4gCsjvW4cDZEZSPcyfGUi1zI6fyBibXMExBk4DUXkpsxHpt4SSEuvg0ZE9kkHe5BGFfaGNA29Cn2EfXOxAJcGwfBt4rx3vIykYl3xS7O4i8n5W3gVBdtDztbd8XuqM4EbalNGoEhKZZoA1Io1zl1wiIOxhOHbwj-8oT4-xPzzTNa_RC_vyb1VWbcv5naC4wnPwLj5_6R_7R_1m0E6zooW6SGqWNsAjWJ2DKxPehYt_tx4Y5MbI05Xf_dNd59wSIysW1aY5jYTiTnmXhjHusAldJUPOg0bTQxfzXVmJiDawDPDRYRS9C2aeN1_4k1jbCdwSo-vfi0OO-gH4G7M8W9R-hqZ8wFSn3SQdN431-u2-xMJTk4MpFni2SID593kMjPxBbE80S5rs6TMi0GqJyntPoNOuNPYh2AZesSI_qjtliybMNE_hLDruvGzytm7PyBia1q9K50RSAAL7F2loltrJjYPqJX9K1RF_T08FGdfyaYNhSlCumKIx-AgvQYWhMpDtfGpo3Q1WhvF6AQJ2U03UOeInZNuizUKoCC0KiolYFQq4bUc9se9-jBVeBdF8B5KJxpj9QCC3AVUax3tRuOyA0cVazHlXHOJ2Lnffl6-R2kq1DQf-AV5KED-9c1TKAH_UOEgHgMVMsjQqx1gKLG4pEwjs5RstFruvcgKFtC6TBYIhNE9UirPDpQReFaG1Py_U7A39s0I6DxrkFvLgnCGU9BvoddVO1CWeKXdREUHAnyxiCBxGa8T4fN-HhI5Jfataa8TWBPlTqhJ247n4JL-4nKUXl3vLnC9hvr2j2hN7VHUOWXNtD54I5Yuy51qrPJ2tUPk9shhJW9QKxde6hjoKVm6jZGZydpek35_9D5A64-fxzewxfl28Dk9kp_ZR8DExMYlUtZLuRCjXA5mU-nE5nPs8WoXhZFOZ3tF3nO8z3OKpzzQuQFn_FZxTOc5CO9FFxkfMrzSZ4tJotxIWelnE7nebGXgqNkU45Hpc2Ypgn1xEiH0OJyIqcLPh8ZtUcT0m8AIY5Ge7q2ZpuRX6bps28PgfLSIYZnC1FHk343pAPZhmXrhGTWX1iIiTRnTuh1pYueFkWt7AEDKI_QuEBz4_KabkxuR603y-8ejSmlkIZjyuq0FH8GAAD___zq26c">