Allow int8 type in quantized_matmul (#5898)

mcremon-meta · facebook-github-bot · commit f4e25e100bc1 · 2024-10-09T14:50:23.000-07:00
Summary: Pull Request resolved: #5898 As titled. Reviewed By: zonglinpeng Differential Revision: D63659947 fbshipit-source-id: 850e7ce773508ad9ecd3ba36a70e87497a876bcd
diff --git a/backends/cadence/reference/operators/quantized_matmul_out.cpp b/backends/cadence/reference/operators/quantized_matmul_out.cpp
@@ -49,7 +49,7 @@ __attribute__((noinline)) void qmatmul(
           sum += (X[i * n + k] - X_zero_point) * (y[k * p + j] - Y_zero_point);
         }
       }
-      Z[i * p + j] = kernels::quantize<uint8_t>(sum, Z_scale, Z_zero_point);
+      Z[i * p + j] = kernels::quantize<TZ>(sum, Z_scale, Z_zero_point);
     }
   }
 }

Original file line number	Diff line number	Diff line change
`@@ -49,7 +49,7 @@ __attribute__((noinline)) void qmatmul(`
`49`	`49`	`sum += (X[i * n + k] - X_zero_point) * (y[k * p + j] - Y_zero_point);`
`50`	`50`	`}`
`51`	`51`	`}`
`52`		`- Z[i * p + j] = kernels::quantize<uint8_t>(sum, Z_scale, Z_zero_point);`
	`52`	`+ Z[i * p + j] = kernels::quantize<TZ>(sum, Z_scale, Z_zero_point);`
`53`	`53`	`}`
`54`	`54`	`}`
`55`	`55`	`}`