Handle output from TOSA reference model being int8 (#6190)

Summary:
Move TOSA reference model pin. This adds bugfix for .npy array to now be aligned to output int8 for BI models.

Make sure the dequantization not overflows by adding a cast to int32.

Change-Id: I2274955bc02a6f3a75687bb2f731b2058834c44f

Pull Request resolved: #6190

Reviewed By: kirklandsign

Differential Revision: D64426319

Pulled By: digantdesai

fbshipit-source-id: e1169b9420258934ebb82e1148c97ab0f1d206f7

Author: per

backends/arm/test/runner_utils.py

@@ -439,6 +439,9 @@ def run_tosa_ref_model(
 
             if self.is_quantized:
                 # Need to dequant back to FP32 for comparison with torch output
+                # Convert to int32 prior to dequantize the output
+                if tosa_ref_output.dtype == np.int8:
+                    tosa_ref_output = tosa_ref_output.astype(np.int32)
                 quant_param = self.qp_output
                 assert (
                     quant_param is not None

examples/arm/setup.sh

@@ -88,7 +88,7 @@ ethos_u_base_rev="24.08"
 
 # tosa reference model
 tosa_reference_model_url="https://review.mlplatform.org/tosa/reference_model"
-tosa_reference_model_rev="444eb365d92774430006e56a8c20161be2f2674f"
+tosa_reference_model_rev="f9ea4ab7da19318fe36b1c34d68a3e40fd6e56c5"
 
 ########
 ### Mandatory user args