Feat (zero_point): dynamic groupwise zero point (#1160)

src/brevitas/core/zero_point.py

@@ -344,3 +344,25 @@ def forward(self, x: Tensor, scale: Tensor, bit_width: Tensor) -> Tensor:
         # pre-zero centering before rounding and clipping
         z = self.get_zero_center(x) / scale  # need to scale the norm by s
         return z
+
+
+class RuntimeDynamicGroupZeroPoint(brevitas.jit.ScriptModule):
+
+    def __init__(
+            self,
+            input_view_impl: Module,
+            zero_point_stats_impl: Module,
+            int_quant: Module,
+            quantize_zero_point: bool) -> None:
+        super(RuntimeDynamicGroupZeroPoint, self).__init__()
+
+        self.zero_point_stats_impl = zero_point_stats_impl
+        self.input_view_impl = input_view_impl
+        self.scale_shift_zero_point = _ScaleShiftZeroPoint(int_quant, quantize_zero_point)
+
+    @brevitas.jit.script_method
+    def forward(self, stats_input: torch.Tensor, scale, bit_width) -> torch.Tensor:
+
+        stats_input_reshaped = self.input_view_impl(stats_input)
+        out = self.zero_point_stats_impl(stats_input_reshaped)
+        return self.scale_shift_zero_point(-out, scale, bit_width)