Rescale add int16 correctly (#14645)

Summary:



 we rescale to int32 using a different common scale(2*max(left scale,right scale))

the resccale(s) -> add[int32] -> rescale from TOSA is converted to a single add command stream instruction by Vela/Regor

Reviewed By: digantdesai

Differential Revision: D82906134

Author: 3l1

backends/arm/operators/op_add.py

@@ -64,12 +64,18 @@ def define_node(
             rescaled_inputs, scale_back = tqutils.insert_rescale_ops_to_int32_maxscale(
                 tosa_graph, inputs, node, self.tosa_spec
             )
+        elif inputs[0].dtype == ts.DType.INT16:
+            rescaled_inputs, scale_back = (
+                tqutils.insert_rescale_ops_int16_to_int32_maxscale(
+                    tosa_graph, inputs, node, self.tosa_spec
+                )
+            )
         else:
             # input[0].dtype == ts.DType.INT16 or ts.DType.INT32
             # Non quantized input, natively support by TOSA.ADD
             rescaled_inputs = inputs
 
-        if output.dtype == ts.DType.INT8:
+        if output.dtype in [ts.DType.INT8, ts.DType.INT16]:
             broadcasted_shape = tutils.tosa_shape(output.shape, output.dim_order)
             add_output = tosa_graph.addIntermediate(broadcasted_shape, ts.DType.INT32)
         else:
@@ -99,6 +105,15 @@ def define_node(
                 compute_rescale=False,
                 tosa_spec=self.tosa_spec,
             )  # type: ignore[possibly-undefined]
+        elif output.dtype == ts.DType.INT16:
+            tqutils.insert_rescale_op_to_int16(
+                tosa_graph,
+                add_output,
+                scale_back,
+                node,
+                compute_rescale=False,
+                tosa_spec=self.tosa_spec,
+            )  # type: ignore[possibly-undefined]
 
 
 @register_node_visitor

backends/arm/test/ops/test_add.py

@@ -276,9 +276,6 @@ def test_add_tensor_16a8w_tosa_INT(test_data: input_t1):
 
 @common.parametrize("test_data", Add.test_data)
 @common.XfailIfNoCorstone300
-@pytest.mark.xfail(
-    reason="Vela compilation fails with 'Invalid arguments' for int16 add operations. See: https://github.com/pytorch/executorch/issues/13730"
-)
 def test_add_tensor_16a8w_u55_INT16(test_data: input_t1):
     """Test add operation with 16A8W quantization on U55 (16-bit activations, 8-bit weights)"""
     per_channel_quantization = False
@@ -304,9 +301,6 @@ def test_add_tensor_16a8w_u55_INT16(test_data: input_t1):
 
 @common.parametrize("test_data", Add.test_data)
 @common.XfailIfNoCorstone320
-@pytest.mark.xfail(
-    reason="Vela compilation fails with 'Invalid arguments' for int16 add operations. See: https://github.com/pytorch/executorch/issues/13730"
-)
 def test_add_tensor_16a8w_u85_INT16(test_data: input_t1):
     """Test add operation with 16A8W quantization on U85 (16-bit activations, 8-bit weights)"""
     per_channel_quantization = False

backends/arm/tosa/quant_utils.py

@@ -77,6 +77,59 @@ def insert_rescale_ops_to_int32_maxscale(
     return [rescaled_lhs, rescaled_rhs], back_scale
 
 
+def insert_rescale_ops_int16_to_int32_maxscale(
+    tosa_graph: Any, inputs: list[TosaArg], node: Node, tosa_spec=None
+) -> tuple[list[Any], float]:
+    """For ADD and SUB with int16 inputs, we rescale to int32 using a different common scale(2*max(left scale,right scale))
+    compared to all the other cases. We multiply the left and right scales by 1<<12 giving us extra precision
+    for the computation without overflowing.
+
+    Returns a list of the rescaled nodes and the scale factor used,
+    needed by insert_rescale_op_to_int16.
+    """
+
+    if len(inputs) > 2:
+        raise ValueError("More than two inputs not supported")
+
+    tensors = inputs.copy()
+    # Reshape tensor according to TOSA dim order
+    for tensor in tensors:
+        dim_order = tensor.dim_order
+        tensor.shape = [tensor.shape[i] for i in dim_order]
+
+    input_qparams = get_input_qparams(node)
+    lhs_qparams, rhs_qparams = input_qparams.values()
+    lhs_scale = lhs_qparams.get_scale_per_tensor()
+    rhs_scale = rhs_qparams.get_scale_per_tensor()
+    # Common scale for the two numbers
+    max_scale_2x = 2 * max(lhs_scale, rhs_scale)
+    SHIFT_INT16 = 12
+    # We are adding two int16 numbers. If the zero point is non-null, the result will be in the range [-131070;131070], therefore we need 18 bits for the result.
+    # We have a 32-bit accumulator, so we can shift to the left by 12 bits and not overflow. In reality, because we divide by the 2*max(lhs_scale,rhs_scale)
+    # we are shifting to the left by 11.
+    lhs_factor = (1 << SHIFT_INT16) * lhs_scale / max_scale_2x
+    rhs_factor = (1 << SHIFT_INT16) * rhs_scale / max_scale_2x
+    rescaled_lhs = build_rescale_to_int32(
+        tosa_graph,
+        tensors[0],
+        lhs_qparams.get_zp_per_tensor(),
+        lhs_factor,
+        tosa_spec=tosa_spec,
+    )
+    rescaled_rhs = build_rescale_to_int32(
+        tosa_graph,
+        tensors[1],
+        rhs_qparams.get_zp_per_tensor(),
+        rhs_factor,
+        tosa_spec=tosa_spec,
+    )
+    out_qparam = get_output_qparams(node)[0]
+    out_scale = out_qparam.get_scale_per_tensor()
+    back_scale = max_scale_2x / (out_scale * (1 << SHIFT_INT16))
+
+    return [rescaled_lhs, rescaled_rhs], back_scale
+
+
 def insert_rescale_ops_to_int32(
     tosa_graph: Any,
     inputs: list[TosaArg],