Split on dilation for strongly typed convs

Differential Revision: D80574314

Pull Request resolved: #13532

Author: ethansfng

backends/cadence/aot/functions.yaml

@@ -304,6 +304,26 @@
     - arg_meta: null
       kernel_name: impl::reference::quantized_conv_nhwc_asym8uxsym8u_asym8u_per_tensor_out
 
+- func: cadence::quantized_conv_nchw_dilated_asym8sxsym8s_asym8s.per_tensor_out(Tensor input, Tensor weight, Tensor bias, int[] stride, SymInt[] padding, int[] dilation, int groups, int input_zero_point, int weight_zero_point, float bias_scale, float out_scale, int out_zero_point, int out_multiplier, int out_shift, *, Tensor(a!) out) -> Tensor(a!)
+  kernels:
+    - arg_meta: null
+      kernel_name: impl::reference::quantized_conv_nchw_dilated_asym8sxsym8s_asym8s_per_tensor_out
+
+- func: cadence::quantized_conv_nchw_dilated_asym8uxsym8u_asym8u.per_tensor_out(Tensor input, Tensor weight, Tensor bias, int[] stride, SymInt[] padding, int[] dilation, int groups, int input_zero_point, int weight_zero_point, float bias_scale, float out_scale, int out_zero_point, int out_multiplier, int out_shift, *, Tensor(a!) out) -> Tensor(a!)
+  kernels:
+    - arg_meta: null
+      kernel_name: impl::reference::quantized_conv_nchw_dilated_asym8uxsym8u_asym8u_per_tensor_out
+
+- func: cadence::quantized_conv_nhwc_dilated_asym8sxsym8s_asym8s.per_tensor_out(Tensor input, Tensor weight, Tensor bias, int[] stride, SymInt[] padding, int[] dilation, int groups, int input_zero_point, int weight_zero_point, float bias_scale, float out_scale, int out_zero_point, int out_multiplier, int out_shift, *, Tensor(a!) out) -> Tensor(a!)
+  kernels:
+    - arg_meta: null
+      kernel_name: impl::reference::quantized_conv_nhwc_dilated_asym8sxsym8s_asym8s_per_tensor_out
+
+- func: cadence::quantized_conv_nhwc_dilated_asym8uxsym8u_asym8u.per_tensor_out(Tensor input, Tensor weight, Tensor bias, int[] stride, SymInt[] padding, int[] dilation, int groups, int input_zero_point, int weight_zero_point, float bias_scale, float out_scale, int out_zero_point, int out_multiplier, int out_shift, *, Tensor(a!) out) -> Tensor(a!)
+  kernels:
+    - arg_meta: null
+      kernel_name: impl::reference::quantized_conv_nhwc_dilated_asym8uxsym8u_asym8u_per_tensor_out
+
 - func: cadence::quantized_fully_connected.out(Tensor src, Tensor weight, Tensor bias, int src_zero_point, Tensor weight_zero_point, Tensor out_multiplier, Tensor out_shift, int out_zero_point, Tensor? offset, *, Tensor(a!) out) -> Tensor(a!)
   kernels:
     - arg_meta: null

backends/cadence/aot/functions_hifi.yaml

@@ -330,6 +330,26 @@
     - arg_meta: null
       kernel_name: cadence::impl::HiFi::quantized_conv_nhwc_asym8uxsym8u_asym8u_per_tensor_out
 
+- func: cadence::quantized_conv_nchw_dilated_asym8sxsym8s_asym8s.per_tensor_out(Tensor input, Tensor weight, Tensor bias, int[] stride, SymInt[] padding, int[] dilation, int groups, int input_zero_point, int weight_zero_point, float bias_scale, float out_scale, int out_zero_point, int out_multiplier, int out_shift, *, Tensor(a!) out) -> Tensor(a!)
+  kernels:
+    - arg_meta: null
+      kernel_name: cadence::impl::HiFi::quantized_conv_nchw_dilated_asym8sxsym8s_asym8s_per_tensor_out
+
+- func: cadence::quantized_conv_nchw_dilated_asym8uxsym8u_asym8u.per_tensor_out(Tensor input, Tensor weight, Tensor bias, int[] stride, SymInt[] padding, int[] dilation, int groups, int input_zero_point, int weight_zero_point, float bias_scale, float out_scale, int out_zero_point, int out_multiplier, int out_shift, *, Tensor(a!) out) -> Tensor(a!)
+  kernels:
+    - arg_meta: null
+      kernel_name: cadence::impl::HiFi::quantized_conv_nchw_dilated_asym8uxsym8u_asym8u_per_tensor_out
+
+- func: cadence::quantized_conv_nhwc_dilated_asym8sxsym8s_asym8s.per_tensor_out(Tensor input, Tensor weight, Tensor bias, int[] stride, SymInt[] padding, int[] dilation, int groups, int input_zero_point, int weight_zero_point, float bias_scale, float out_scale, int out_zero_point, int out_multiplier, int out_shift, *, Tensor(a!) out) -> Tensor(a!)
+  kernels:
+    - arg_meta: null
+      kernel_name: cadence::impl::HiFi::quantized_conv_nhwc_dilated_asym8sxsym8s_asym8s_per_tensor_out
+
+- func: cadence::quantized_conv_nhwc_dilated_asym8uxsym8u_asym8u.per_tensor_out(Tensor input, Tensor weight, Tensor bias, int[] stride, SymInt[] padding, int[] dilation, int groups, int input_zero_point, int weight_zero_point, float bias_scale, float out_scale, int out_zero_point, int out_multiplier, int out_shift, *, Tensor(a!) out) -> Tensor(a!)
+  kernels:
+    - arg_meta: null
+      kernel_name: cadence::impl::HiFi::quantized_conv_nhwc_dilated_asym8uxsym8u_asym8u_per_tensor_out
+
 - func: cadence::quantized_layer_norm.out(Tensor input, Tensor in_scale, Tensor in_zero_point, int[] normalized_shape, Tensor weight, Tensor bias, float eps, float output_scale, int output_zero_point, *, Tensor(a!) out) -> Tensor(a!)
   kernels:
     - arg_meta: null

backends/cadence/aot/ops_registrations.py

@@ -144,6 +144,30 @@
 lib.define(
     "quantized_conv_nhwc_asym8uxsym8u_asym8u.per_tensor_out(Tensor input, Tensor weight, Tensor bias, int[] stride, SymInt[] padding, int[] dilation, int groups, int input_zero_point, int weight_zero_point, float bias_scale, float out_scale, int out_zero_point, int out_multiplier, int out_shift, *, Tensor(a!) out) -> Tensor(a!)"
 )
+lib.define(
+    "quantized_conv_nchw_dilated_asym8sxsym8s_asym8s.per_tensor(Tensor input, Tensor weight, Tensor bias, int[] stride, SymInt[] padding, int[] dilation, int groups, int input_zero_point, int weight_zero_point, float bias_scale, float out_scale, int out_zero_point, int out_multiplier, int out_shift) -> (Tensor Z)"
+)
+lib.define(
+    "quantized_conv_nchw_dilated_asym8sxsym8s_asym8s.per_tensor_out(Tensor input, Tensor weight, Tensor bias, int[] stride, SymInt[] padding, int[] dilation, int groups, int input_zero_point, int weight_zero_point, float bias_scale, float out_scale, int out_zero_point, int out_multiplier, int out_shift, *, Tensor(a!) out) -> Tensor(a!)"
+)
+lib.define(
+    "quantized_conv_nchw_dilated_asym8uxsym8u_asym8u.per_tensor(Tensor input, Tensor weight, Tensor bias, int[] stride, SymInt[] padding, int[] dilation, int groups, int input_zero_point, int weight_zero_point, float bias_scale, float out_scale, int out_zero_point, int out_multiplier, int out_shift) -> (Tensor Z)"
+)
+lib.define(
+    "quantized_conv_nchw_dilated_asym8uxsym8u_asym8u.per_tensor_out(Tensor input, Tensor weight, Tensor bias, int[] stride, SymInt[] padding, int[] dilation, int groups, int input_zero_point, int weight_zero_point, float bias_scale, float out_scale, int out_zero_point, int out_multiplier, int out_shift, *, Tensor(a!) out) -> Tensor(a!)"
+)
+lib.define(
+    "quantized_conv_nhwc_dilated_asym8sxsym8s_asym8s.per_tensor(Tensor input, Tensor weight, Tensor bias, int[] stride, SymInt[] padding, int[] dilation, int groups, int input_zero_point, int weight_zero_point, float bias_scale, float out_scale, int out_zero_point, int out_multiplier, int out_shift) -> (Tensor Z)"
+)
+lib.define(
+    "quantized_conv_nhwc_dilated_asym8sxsym8s_asym8s.per_tensor_out(Tensor input, Tensor weight, Tensor bias, int[] stride, SymInt[] padding, int[] dilation, int groups, int input_zero_point, int weight_zero_point, float bias_scale, float out_scale, int out_zero_point, int out_multiplier, int out_shift, *, Tensor(a!) out) -> Tensor(a!)"
+)
+lib.define(
+    "quantized_conv_nhwc_dilated_asym8uxsym8u_asym8u.per_tensor(Tensor input, Tensor weight, Tensor bias, int[] stride, SymInt[] padding, int[] dilation, int groups, int input_zero_point, int weight_zero_point, float bias_scale, float out_scale, int out_zero_point, int out_multiplier, int out_shift) -> (Tensor Z)"
+)
+lib.define(
+    "quantized_conv_nhwc_dilated_asym8uxsym8u_asym8u.per_tensor_out(Tensor input, Tensor weight, Tensor bias, int[] stride, SymInt[] padding, int[] dilation, int groups, int input_zero_point, int weight_zero_point, float bias_scale, float out_scale, int out_zero_point, int out_multiplier, int out_shift, *, Tensor(a!) out) -> Tensor(a!)"
+)
 lib.define(
     "quantized_matmul_asym8uxasym8u_asym8u(Tensor X, int X_zero_point, Tensor Y, int Y_zero_point, Tensor? bias, int out_multiplier, int out_shift, int out_zero_point, bool transposed=False) -> (Tensor Z)"
 )
@@ -919,6 +943,182 @@ def quantized_conv_nhwc_asym8uxsym8u_asym8u_per_tensor_meta(
     return input.new_empty(output_size, dtype=input.dtype)
 
 
+@register_fake("cadence::quantized_conv_nchw_dilated_asym8sxsym8s_asym8s.per_tensor")
+def quantized_conv_nchw_dilated_asym8sxsym8s_asym8s_per_tensor_meta(
+    input: torch.Tensor,
+    weight: torch.Tensor,
+    bias: torch.Tensor,
+    stride: Tuple[int],
+    padding: Tuple[int],
+    dilation: Tuple[int],
+    groups: int,
+    in_zero_point: int,
+    weight_zero_point: int,
+    bias_scale: float,
+    output_scale: float,
+    output_zero_point: int,
+    out_multiplier: int,
+    out_shift: int,
+) -> torch.Tensor:
+    out_channels, _, *kernel_size = weight.shape
+
+    in_size = input.shape
+    # Assert that the input tensor has at least 3 dimensions, and at most 6
+    assert len(in_size) > 2
+    assert len(in_size) < 6
+
+    # Compute the output tensor size
+    output_size = (
+        get_conv1d_output_size(
+            in_size,
+            out_channels,
+            stride[1],
+            padding[1],
+            dilation[1],
+            kernel_size[0],
+            False,
+        )
+        if len(in_size) == 3
+        else get_conv2d_output_size(
+            in_size, out_channels, stride, padding, dilation, kernel_size, False
+        )
+    )
+
+    return input.new_empty(output_size, dtype=input.dtype)
+
+
+@register_fake("cadence::quantized_conv_nchw_dilated_asym8uxsym8u_asym8u.per_tensor")
+def quantized_conv_nchw_dilated_asym8uxsym8u_asym8u_per_tensor_meta(
+    input: torch.Tensor,
+    weight: torch.Tensor,
+    bias: torch.Tensor,
+    stride: Tuple[int],
+    padding: Tuple[int],
+    dilation: Tuple[int],
+    groups: int,
+    in_zero_point: int,
+    weight_zero_point: int,
+    bias_scale: float,
+    output_scale: float,
+    output_zero_point: int,
+    out_multiplier: int,
+    out_shift: int,
+) -> torch.Tensor:
+    out_channels, _, *kernel_size = weight.shape
+
+    in_size = input.shape
+    # Assert that the input tensor has at least 3 dimensions, and at most 6
+    assert len(in_size) > 2
+    assert len(in_size) < 6
+
+    # Compute the output tensor size
+    output_size = (
+        get_conv1d_output_size(
+            in_size,
+            out_channels,
+            stride[1],
+            padding[1],
+            dilation[1],
+            kernel_size[0],
+            False,
+        )
+        if len(in_size) == 3
+        else get_conv2d_output_size(
+            in_size, out_channels, stride, padding, dilation, kernel_size, False
+        )
+    )
+
+    return input.new_empty(output_size, dtype=input.dtype)
+
+
+@register_fake("cadence::quantized_conv_nhwc_dilated_asym8sxsym8s_asym8s.per_tensor")
+def quantized_conv_nhwc_dilated_asym8sxsym8s_asym8s_per_tensor_meta(
+    input: torch.Tensor,
+    weight: torch.Tensor,
+    bias: torch.Tensor,
+    stride: Tuple[int],
+    padding: Tuple[int],
+    dilation: Tuple[int],
+    groups: int,
+    in_zero_point: int,
+    weight_zero_point: int,
+    bias_scale: float,
+    output_scale: float,
+    output_zero_point: int,
+    out_multiplier: int,
+    out_shift: int,
+) -> torch.Tensor:
+    out_channels, *kernel_size, _ = weight.shape
+
+    in_size = input.shape
+    # Assert that the input tensor has at least 3 dimensions, and at most 6
+    assert len(in_size) > 2
+    assert len(in_size) < 6
+
+    # Compute the output tensor size
+    output_size = (
+        get_conv1d_output_size(
+            in_size,
+            out_channels,
+            stride[1],
+            padding[1],
+            dilation[1],
+            kernel_size[0],
+            True,
+        )
+        if len(in_size) == 3
+        else get_conv2d_output_size(
+            in_size, out_channels, stride, padding, dilation, kernel_size, True
+        )
+    )
+
+    return input.new_empty(output_size, dtype=input.dtype)
+
+
+@register_fake("cadence::quantized_conv_nhwc_dilated_asym8uxsym8u_asym8u.per_tensor")
+def quantized_conv_nhwc_dilated_asym8uxsym8u_asym8u_per_tensor_meta(
+    input: torch.Tensor,
+    weight: torch.Tensor,
+    bias: torch.Tensor,
+    stride: Tuple[int],
+    padding: Tuple[int],
+    dilation: Tuple[int],
+    groups: int,
+    in_zero_point: int,
+    weight_zero_point: int,
+    bias_scale: float,
+    output_scale: float,
+    output_zero_point: int,
+    out_multiplier: int,
+    out_shift: int,
+) -> torch.Tensor:
+    out_channels, *kernel_size, _ = weight.shape
+
+    in_size = input.shape
+    # Assert that the input tensor has at least 3 dimensions, and at most 6
+    assert len(in_size) > 2
+    assert len(in_size) < 6
+
+    # Compute the output tensor size
+    output_size = (
+        get_conv1d_output_size(
+            in_size,
+            out_channels,
+            stride[1],
+            padding[1],
+            dilation[1],
+            kernel_size[0],
+            True,
+        )
+        if len(in_size) == 3
+        else get_conv2d_output_size(
+            in_size, out_channels, stride, padding, dilation, kernel_size, True
+        )
+    )
+
+    return input.new_empty(output_size, dtype=input.dtype)
+
+
 @register_fake("cadence::quantized_layer_norm")
 def quantized_layer_norm_meta(
     input: torch.Tensor,

backends/cadence/aot/tests/test_type_dispatch_passes.py

@@ -341,3 +341,107 @@ def test_uint8_dispatch_quantized_conv_nhwc(self) -> None:
             ),
             1,
         )
+
+    def test_int8_dispatch_quantized_conv_nchw_dilated(self) -> None:
+        """Test int8 x int8 inputs with dilation should dispatch to dilated_asym8sxasym8s_asym8s variant for quantized_conv_nchw_dilated"""
+        x = torch.randint(-128, 127, (1, 3, 8, 8), dtype=torch.int8)
+        w = torch.randint(-128, 127, (16, 3, 3, 3), dtype=torch.int8)
+        b = torch.randint(-2147483648, 2147483647, (16,), dtype=torch.int32)
+        gm = single_op_builder(
+            placeholders=(x, w, b),
+            op=exir_ops.edge.cadence.quantized_conv_nchw.per_tensor,
+            args=(x, w, b, [1, 1], [0, 0], [2, 2], 1, 0, 0, 1.0, 1.0, 0, 1, 1),
+        )
+        p = CompileTimeTypeDispatchPass()
+        gm = cast(PassResult, p(gm)).graph_module
+        # Original op should be replaced
+        self.assertEqual(
+            count_node(gm, exir_ops.edge.cadence.quantized_conv_nchw.per_tensor),
+            0,
+        )
+        # Should be replaced with int8 specific variant
+        self.assertEqual(
+            count_node(
+                gm,
+                exir_ops.edge.cadence.quantized_conv_nchw_dilated_asym8sxsym8s_asym8s.per_tensor,
+            ),
+            1,
+        )
+
+    def test_uint8_dispatch_quantized_conv_nchw_dilated(self) -> None:
+        """Test uint8 x uint8 inputs with dilation should dispatch to dilated_asym8uxasym8u_asym8u variant for quantized_conv_nchw"""
+        x = torch.randint(0, 255, (1, 3, 8, 8), dtype=torch.uint8)
+        w = torch.randint(0, 255, (16, 3, 3, 3), dtype=torch.uint8)
+        b = torch.randint(-2147483648, 2147483647, (16,), dtype=torch.int32)
+        gm = single_op_builder(
+            placeholders=(x, w, b),
+            op=exir_ops.edge.cadence.quantized_conv_nchw.per_tensor,
+            args=(x, w, b, [1, 1], [0, 0], [2, 2], 1, 0, 0, 1.0, 1.0, 0, 1, 1),
+        )
+        p = CompileTimeTypeDispatchPass()
+        gm = cast(PassResult, p(gm)).graph_module
+        # Original op should be replaced
+        self.assertEqual(
+            count_node(gm, exir_ops.edge.cadence.quantized_conv_nchw.per_tensor),
+            0,
+        )
+        # Should be replaced with uint8 specific variant
+        self.assertEqual(
+            count_node(
+                gm,
+                exir_ops.edge.cadence.quantized_conv_nchw_dilated_asym8uxsym8u_asym8u.per_tensor,
+            ),
+            1,
+        )
+
+    def test_int8_dispatch_quantized_conv_nhwc_dilated(self) -> None:
+        """Test int8 x int8 inputs with dilation should dispatch to dilated_asym8sxasym8s_asym8s variant for quantized_conv_nhwc"""
+        x = torch.randint(-128, 127, (1, 8, 8, 3), dtype=torch.int8)
+        w = torch.randint(-128, 127, (16, 3, 3, 3), dtype=torch.int8)
+        b = torch.randint(-2147483648, 2147483647, (16,), dtype=torch.int32)
+        gm = single_op_builder(
+            placeholders=(x, w, b),
+            op=exir_ops.edge.cadence.quantized_conv_nhwc.per_tensor,
+            args=(x, w, b, [1, 1], [0, 0], [2, 2], 1, 0, 0, 1.0, 1.0, 0, 1, 1),
+        )
+        p = CompileTimeTypeDispatchPass()
+        gm = cast(PassResult, p(gm)).graph_module
+        # Original op should be replaced
+        self.assertEqual(
+            count_node(gm, exir_ops.edge.cadence.quantized_conv_nhwc.per_tensor),
+            0,
+        )
+        # Should be replaced with int8 specific variant
+        self.assertEqual(
+            count_node(
+                gm,
+                exir_ops.edge.cadence.quantized_conv_nhwc_dilated_asym8sxsym8s_asym8s.per_tensor,
+            ),
+            1,
+        )
+
+    def test_uint8_dispatch_quantized_conv_nhwc_dilated(self) -> None:
+        """Test uint8 x uint8 inputs with dilation should dispatch to dilated_asym8uxasym8u_asym8u variant for quantized_conv_nhwc"""
+        x = torch.randint(0, 255, (1, 8, 8, 3), dtype=torch.uint8)
+        w = torch.randint(0, 255, (16, 3, 3, 3), dtype=torch.uint8)
+        b = torch.randint(-2147483648, 2147483647, (16,), dtype=torch.int32)
+        gm = single_op_builder(
+            placeholders=(x, w, b),
+            op=exir_ops.edge.cadence.quantized_conv_nhwc.per_tensor,
+            args=(x, w, b, [1, 1], [0, 0], [2, 2], 1, 0, 0, 1.0, 1.0, 0, 1, 1),
+        )
+        p = CompileTimeTypeDispatchPass()
+        gm = cast(PassResult, p(gm)).graph_module
+        # Original op should be replaced
+        self.assertEqual(
+            count_node(gm, exir_ops.edge.cadence.quantized_conv_nhwc.per_tensor),
+            0,
+        )
+        # Should be replaced with uint8 specific variant
+        self.assertEqual(
+            count_node(
+                gm,
+                exir_ops.edge.cadence.quantized_conv_nhwc_dilated_asym8uxsym8u_asym8u.per_tensor,
+            ),
+            1,
+        )

backends/cadence/aot/type_dispatch.py

@@ -112,7 +112,20 @@ def call_operator(
             raise RuntimeError(f"Unsupported input types for {op}: {dtype_key}")
 
         type_suffix = config.type_dispatch_suffixes[dtype_key]
-        typed_op_name = f"{config.base_name}_{type_suffix}"
+        base_name = config.base_name
+
+        if op in [
+            exir_ops.edge.cadence.quantized_conv_nchw.per_tensor,
+            exir_ops.edge.cadence.quantized_conv_nhwc.per_tensor,
+        ]:
+            dilation = args[5]
+            # pyre-ignore[16]: None has no attribute '__iter__'.
+            is_dilated = any(d > 1 for d in dilation)
+
+            if is_dilated:
+                type_suffix = f"dilated_{type_suffix}"
+
+        typed_op_name = f"{base_name}_{type_suffix}"
 
         typed_op = getattr(
             getattr(exir_ops.edge.cadence, typed_op_name), config.variant