Add support for strongly typed op_quantized_linear_out

ethansfng · web-flow · commit 9807fbe062dc · 2025-08-13T15:16:06.000-07:00
Differential Revision: D79911643 Pull Request resolved: pytorch#13346
diff --git a/backends/cadence/aot/functions.yaml b/backends/cadence/aot/functions.yaml
@@ -229,6 +229,16 @@
     - arg_meta: null
       kernel_name: impl::reference::quantized_linear_per_tensor_out
 
+- func: cadence::quantized_linear_asym8sxasym8s_asym8s.per_tensor_out(Tensor src, Tensor weight, Tensor bias, int src_zero_point, int weight_zero_point, int out_multiplier, int out_shift, int out_zero_point, Tensor? offset, *, Tensor(a!) out) -> Tensor(a!)
+  kernels:
+    - arg_meta: null
+      kernel_name: impl::reference::quantized_linear_asym8sxasym8s_asym8s_per_tensor_out
+
+- func: cadence::quantized_linear_asym8uxasym8u_asym8u.per_tensor_out(Tensor src, Tensor weight, Tensor bias, int src_zero_point, int weight_zero_point, int out_multiplier, int out_shift, int out_zero_point, Tensor? offset, *, Tensor(a!) out) -> Tensor(a!)
+  kernels:
+    - arg_meta: null
+      kernel_name: impl::reference::quantized_linear_asym8uxasym8u_asym8u_per_tensor_out
+
 - func: cadence::im2row.out(Tensor input, int[2] kernel_size, int[2] dilation, int[2] padding, int[2] stride, Tensor in_zero_point, bool channel_last=False, *, Tensor(a!) out) -> Tensor(a!)
   kernels:
     - arg_meta: null
diff --git a/backends/cadence/aot/functions_hifi.yaml b/backends/cadence/aot/functions_hifi.yaml
@@ -314,6 +314,16 @@
     - arg_meta: null
       kernel_name: cadence::impl::HiFi::quantized_linear_per_tensor_out
 
+- func: cadence::quantized_linear_asym8sxasym8s_asym8s.per_tensor_out(Tensor src, Tensor weight, Tensor bias, int src_zero_point, int weight_zero_point, int out_multiplier, int out_shift, int out_zero_point, Tensor? offset, *, Tensor(a!) out) -> Tensor(a!)
+  kernels:
+    - arg_meta: null
+      kernel_name: cadence::impl::HiFi::quantized_linear_asym8sxasym8s_asym8s_per_tensor_out
+
+- func: cadence::quantized_linear_asym8uxasym8u_asym8u.per_tensor_out(Tensor src, Tensor weight, Tensor bias, int src_zero_point, int weight_zero_point, int out_multiplier, int out_shift, int out_zero_point, Tensor? offset, *, Tensor(a!) out) -> Tensor(a!)
+  kernels:
+    - arg_meta: null
+      kernel_name: cadence::impl::HiFi::quantized_linear_asym8uxasym8u_asym8u_per_tensor_out
+
 - func: cadence::quantized_relu_per_tensor.out(Tensor X, Tensor X_zero_point, int out_zero_point, Tensor out_multiplier, Tensor out_shift, *, Tensor(a!) out) -> Tensor(a!)
   kernels:
     - arg_meta: null
diff --git a/backends/cadence/aot/ops_registrations.py b/backends/cadence/aot/ops_registrations.py
@@ -56,10 +56,26 @@
 lib.define(
     "quantized_linear.per_tensor_out(Tensor src, Tensor weight, Tensor bias, SymInt src_zero_point, SymInt weight_zero_point, SymInt out_multiplier, SymInt out_shift, SymInt out_zero_point, Tensor? offset, *, Tensor(a!) out) -> Tensor(a!)"
 )
+lib.define(
+    "quantized_linear_asym8sxasym8s_asym8s.per_tensor_out(Tensor src, Tensor weight, Tensor bias, int src_zero_point, "
+    "int weight_zero_point, int out_multiplier, int out_shift, int out_zero_point, Tensor? offset, *, Tensor(a!) out) -> Tensor(a!)"
+)
+lib.define(
+    "quantized_linear_asym8uxasym8u_asym8u.per_tensor_out(Tensor src, Tensor weight, Tensor bias, int src_zero_point, "
+    "int weight_zero_point, int out_multiplier, int out_shift, int out_zero_point, Tensor? offset, *, Tensor(a!) out) -> Tensor(a!)"
+)
 lib.define(
     "quantized_linear.per_tensor(Tensor src, Tensor weight, Tensor bias, SymInt src_zero_point, "
     "SymInt weight_zero_point, SymInt out_multiplier, SymInt out_shift, SymInt out_zero_point, Tensor? offset) -> Tensor"
 )
+lib.define(
+    "quantized_linear_asym8sxasym8s_asym8s.per_tensor(Tensor src, Tensor weight, Tensor bias, int src_zero_point, "
+    "int weight_zero_point, int out_multiplier, int out_shift, int out_zero_point, Tensor? offset) -> (Tensor Z)"
+)
+lib.define(
+    "quantized_linear_asym8uxasym8u_asym8u.per_tensor(Tensor src, Tensor weight, Tensor bias, int src_zero_point, "
+    "int weight_zero_point, int out_multiplier, int out_shift, int out_zero_point, Tensor? offset) -> (Tensor Z)"
+)
 
 lib.define(
     "quantized_relu(Tensor X, Tensor X_zero_point, int out_zero_point, Tensor out_multiplier, Tensor out_shift) -> (Tensor Y)"
@@ -446,6 +462,50 @@ def quantized_linear_per_tensor_meta(
     return src.new_empty(out_size, dtype=src.dtype)
 
 
+@register_fake("cadence::quantized_linear_asym8sxasym8s_asym8s.per_tensor")
+def quantized_linear_asym8sxasym8s_asym8s_per_tensor_meta(
+    src: torch.Tensor,
+    weight: torch.Tensor,
+    bias: torch.Tensor,
+    in_zero_point: int,
+    weight_zero_point: int,
+    out_multiplier: int,
+    out_shift: int,
+    out_zero_point: int,
+    offset: Optional[torch.Tensor],
+) -> torch.Tensor:
+    # src comes in shape [leading_dims, in_dim]
+    # weight comes in shape [out_dim, in_dim]
+    # output comes in empty with shape [leading_dims, out_dim]
+    out_size = list(src.size())
+    weight_size = list(weight.size())
+    assert len(weight_size) == 2
+    out_size[-1] = weight_size[0]
+    return src.new_empty(out_size, dtype=src.dtype)
+
+
+@register_fake("cadence::quantized_linear_asym8uxasym8u_asym8u.per_tensor")
+def quantized_linear_asym8uxasym8u_asym8u_per_tensor_meta(
+    src: torch.Tensor,
+    weight: torch.Tensor,
+    bias: torch.Tensor,
+    in_zero_point: int,
+    weight_zero_point: int,
+    out_multiplier: int,
+    out_shift: int,
+    out_zero_point: int,
+    offset: Optional[torch.Tensor],
+) -> torch.Tensor:
+    # src comes in shape [leading_dims, in_dim]
+    # weight comes in shape [out_dim, in_dim]
+    # output comes in empty with shape [leading_dims, out_dim]
+    out_size = list(src.size())
+    weight_size = list(weight.size())
+    assert len(weight_size) == 2
+    out_size[-1] = weight_size[0]
+    return src.new_empty(out_size, dtype=src.dtype)
+
+
 @register_fake("cadence::quantized_conv")
 def quantized_conv_meta(
     input: torch.Tensor,
diff --git a/backends/cadence/aot/tests/test_type_dispatch_passes.py b/backends/cadence/aot/tests/test_type_dispatch_passes.py
@@ -18,7 +18,7 @@
 
 
 class TestTypeDispatchPasses(unittest.TestCase):
-    def test_int8_dispatch(self) -> None:
+    def test_int8_dispatch_quantized_fully_connected(self) -> None:
         """Test int8 x int8 inputs should dispatch to asym8sxasym8s_asym8s variant"""
         x = torch.randint(-128, 127, (2, 3), dtype=torch.int8)
         w = torch.randint(-128, 127, (4, 3), dtype=torch.int8)
@@ -44,7 +44,7 @@ def test_int8_dispatch(self) -> None:
             1,
         )
 
-    def test_uint8_dispatch(self) -> None:
+    def test_uint8_dispatch_quantized_fully_connected(self) -> None:
         """Test uint8 x uint8 inputs should dispatch to asym8uxasym8u_asym8u variant"""
         x = torch.randint(0, 255, (2, 3), dtype=torch.uint8)
         w = torch.randint(0, 255, (4, 3), dtype=torch.uint8)
@@ -70,6 +70,58 @@ def test_uint8_dispatch(self) -> None:
             1,
         )
 
+    def test_int8_dispatch_quantized_linear(self) -> None:
+        """Test int8 x int8 inputs should dispatch to asym8sxasym8s_asym8s variant for quantized_linear"""
+        x = torch.randint(-128, 127, (2, 3), dtype=torch.int8)
+        w = torch.randint(-128, 127, (4, 3), dtype=torch.int8)
+        b = torch.randint(-2147483648, 2147483647, (4,), dtype=torch.int32)
+        gm = single_op_builder(
+            placeholders=(x, w, b),
+            op=exir_ops.edge.cadence.quantized_linear.per_tensor,
+            args=(x, w, b, 0, 0, 1, 0, 0, None),
+        )
+        p = CompileTimeTypeDispatchPass()
+        gm = cast(PassResult, p(gm)).graph_module
+        # Original op should be replaced
+        self.assertEqual(
+            count_node(gm, exir_ops.edge.cadence.quantized_linear.per_tensor),
+            0,
+        )
+        # Should be replaced with int8 specific variant
+        self.assertEqual(
+            count_node(
+                gm,
+                exir_ops.edge.cadence.quantized_linear_asym8sxasym8s_asym8s.per_tensor,
+            ),
+            1,
+        )
+
+    def test_uint8_quantized_linear_dispatch(self) -> None:
+        """Test uint8 x uint8 inputs should dispatch to asym8uxasym8u_asym8u variant for quantized_linear"""
+        x = torch.randint(0, 255, (2, 3), dtype=torch.uint8)
+        w = torch.randint(0, 255, (4, 3), dtype=torch.uint8)
+        b = torch.randint(-2147483648, 2147483647, (4,), dtype=torch.int32)
+        gm = single_op_builder(
+            placeholders=(x, w, b),
+            op=exir_ops.edge.cadence.quantized_linear.per_tensor,
+            args=(x, w, b, 0, 0, 1, 0, 0, None),
+        )
+        p = CompileTimeTypeDispatchPass()
+        gm = cast(PassResult, p(gm)).graph_module
+        # Original op should be replaced
+        self.assertEqual(
+            count_node(gm, exir_ops.edge.cadence.quantized_linear.per_tensor),
+            0,
+        )
+        # Should be replaced with uint8 specific variant
+        self.assertEqual(
+            count_node(
+                gm,
+                exir_ops.edge.cadence.quantized_linear_asym8uxasym8u_asym8u.per_tensor,
+            ),
+            1,
+        )
+
     def test_mixed_types_error(self) -> None:
         """Test mixed int8/uint8 inputs should raise RuntimeError"""
         x = torch.randint(-128, 127, (2, 3), dtype=torch.int8)
diff --git a/backends/cadence/aot/type_dispatch.py b/backends/cadence/aot/type_dispatch.py
@@ -23,40 +23,40 @@ class CompileTimeTypeDispatchPass(ExportPass):
     Replaces generic ops with ops that have explicit types.
     """
 
+    _TYPE_DISPATCH_MAP: dict[tuple[torch.dtype, torch.dtype], str] = {
+        (torch.int8, torch.int8): "asym8sxasym8s_asym8s",
+        (torch.uint8, torch.uint8): "asym8uxasym8u_asym8u",
+    }
+
+    _SUPPORTED_OPS: dict[OpOverload, str] = {
+        exir_ops.edge.cadence.quantized_fully_connected.per_tensor: "quantized_fully_connected",
+        exir_ops.edge.cadence.quantized_linear.per_tensor: "quantized_linear",
+    }
+
     def call_operator(
         self,
         op: OpOverload,
         args: tuple[Argument, ...],
         kwargs: dict[str, Argument],
         meta: NodeMetadata,
     ) -> ProxyValue:
-        if op not in {
-            exir_ops.edge.cadence.quantized_fully_connected.per_tensor,
-        }:
+        if op not in self._SUPPORTED_OPS:
             return super().call_operator(op, args, kwargs, meta)
 
-        if (
-            # pyre-ignore[16]: None has no attribute `to_tensor`.
-            args[0].to_tensor().dtype == torch.int8
-            and args[1].to_tensor().dtype == torch.int8
-        ):
-            return super().call_operator(
-                exir_ops.edge.cadence.quantized_fully_connected_asym8sxasym8s_asym8s.per_tensor,
-                args,
-                kwargs,
-                meta,
-            )
-        elif (
-            args[0].to_tensor().dtype == torch.uint8
-            and args[1].to_tensor().dtype == torch.uint8
-        ):
-            return super().call_operator(
-                exir_ops.edge.cadence.quantized_fully_connected_asym8uxasym8u_asym8u.per_tensor,
-                args,
-                kwargs,
-                meta,
-            )
-        else:
+        # pyre-ignore[16]: None has no attribute `to_tensor`.
+        input_dtype = args[0].to_tensor().dtype
+        weight_dtype = args[1].to_tensor().dtype
+        dtype_pair = (input_dtype, weight_dtype)
+
+        if dtype_pair not in self._TYPE_DISPATCH_MAP:
             raise RuntimeError(
-                f"Unsupported input types for {op}: {args[0].to_tensor().dtype} and {args[1].to_tensor().dtype}"
+                f"Unsupported input types for {op}: {input_dtype} and {weight_dtype}"
             )
+
+        base_op_name = self._SUPPORTED_OPS[op]
+        type_suffix = self._TYPE_DISPATCH_MAP[dtype_pair]
+
+        typed_op_name = f"{base_op_name}_{type_suffix}"
+        typed_op = getattr(exir_ops.edge.cadence, typed_op_name).per_tensor
+
+        return super().call_operator(typed_op, args, kwargs, meta)
diff --git a/backends/cadence/hifi/operators/op_quantized_linear_asym8sxasym8s_asym8s_per_tensor_out.cpp b/backends/cadence/hifi/operators/op_quantized_linear_asym8sxasym8s_asym8s_per_tensor_out.cpp
@@ -0,0 +1,75 @@
+/*
+ * Copyright (c) Meta Platforms, Inc. and affiliates.
+ * All rights reserved.
+ *
+ * This source code is licensed under the BSD-style license found in the
+ * LICENSE file in the root directory of this source tree.
+ */
+
+#include <executorch/backends/cadence/hifi/kernels/kernels.h>
+#include <executorch/runtime/kernel/kernel_includes.h>
+#include <xa_nnlib_kernels_api.h>
+
+namespace cadence {
+namespace impl {
+namespace HiFi {
+namespace native {
+
+using ::executorch::aten::Tensor;
+using ::executorch::runtime::getLeadingDims;
+using ::executorch::runtime::KernelRuntimeContext;
+using std::optional;
+
+void quantized_linear_asym8sxasym8s_asym8s_per_tensor_out(
+    __ET_UNUSED KernelRuntimeContext& ctx,
+    const Tensor& in,
+    const Tensor& weight,
+    const Tensor& bias,
+    int64_t in_zero_point,
+    int64_t weight_zero_point,
+    int64_t out_multiplier,
+    int64_t out_shift,
+    int64_t out_zero_point,
+    __ET_UNUSED const optional<Tensor>& offset,
+    Tensor& out) {
+  // input comes in shape [leading_dims, in_dim]
+  // weight comes in shape [out_dim, in_dim]
+  // output comes in empty with shape [leading_dims, out_dim]
+  // Perform matrix multiply (M x N) x (N x P)' => M x P
+  const int64_t leading_dims = getLeadingDims(in, in.dim() - 1);
+  const int64_t out_dim = weight.size(0); // = out_dim
+  const int64_t in_dim = weight.size(1); // = in_dim
+
+  const int8_t* __restrict__ in_data = in.const_data_ptr<int8_t>();
+  const int8_t* __restrict__ weight_data = weight.const_data_ptr<int8_t>();
+  const int32_t* __restrict__ bias_data = bias.const_data_ptr<int32_t>();
+  int8_t* __restrict__ out_data = out.mutable_data_ptr<int8_t>();
+
+  const int32_t out_multipler_int32 = static_cast<int32_t>(out_multiplier);
+  const int32_t out_shift_int32 = static_cast<int32_t>(out_shift);
+
+  // The nnlib kernel to compute quantized linear via matmul.
+  const int32_t ret = xa_nn_matmul_asym8sxasym8s_asym8s(
+      out_data, // p_out
+      weight_data, // p_mat1,
+      in_data, // p_mat2,
+      bias_data, // p_bias
+      out_dim, // rows of p_mat1
+      in_dim, // cols of p_mat1
+      in_dim, // row_stride of p_mat1
+      leading_dims, // vec_count, i.e., rows of p_mat2
+      in_dim, // vec_offset of p_mat2.
+      out_dim, // out_offset, i.e., offset of next output element written
+      1, // out_stride, i.e., stride to go to next output row
+      -weight_zero_point, // mat1_zero_bias
+      -in_zero_point, // mat2_zero_bias
+      out_multipler_int32, // out_multiplier
+      out_shift_int32, // out_shift
+      out_zero_point); // out_zero_bias
+  ET_DCHECK_MSG(ret == 0, "HiFi quantized::linear_per_tensor failed");
+}
+
+} // namespace native
+} // namespace HiFi
+} // namespace impl
+} // namespace cadence
diff --git a/backends/cadence/hifi/operators/op_quantized_linear_asym8uxasym8u_asym8u_per_tensor_out.cpp b/backends/cadence/hifi/operators/op_quantized_linear_asym8uxasym8u_asym8u_per_tensor_out.cpp
diff --git a/backends/cadence/hifi/operators/targets.bzl b/backends/cadence/hifi/operators/targets.bzl
diff --git a/backends/cadence/reference/operators/quantized_linear_out.cpp b/backends/cadence/reference/operators/quantized_linear_out.cpp
