Introduce strongly typed quant/dequant ops

backends/cadence/aot/TARGETS

@@ -144,7 +144,6 @@ executorch_generated_lib(
     visibility = ["PUBLIC"],
     deps = [
         "//executorch/backends/cadence/generic/kernels:cadence_kernels",
-        # Individual operator targets instead of combined cadence_generic_ops
         "//executorch/backends/cadence/generic/operators:op_requantize_out",
         "//executorch/backends/cadence/generic/operators:im2row_out",
         "//executorch/backends/cadence/generic/operators:dequantize_per_tensor",

backends/cadence/aot/functions.yaml

@@ -184,12 +184,60 @@
     - arg_meta: null
       kernel_name: impl::generic::quantize_per_tensor_out
 
+- func: cadence::quantize_per_tensor_asym8s.out(Tensor input, float scale, int zero_point, int quant_min, int quant_max, ScalarType dtype, *, Tensor(a!) out) -> Tensor(a!)
+  variants: function
+  kernels:
+    - arg_meta: null
+      kernel_name: impl::generic::quantize_per_tensor_asym8s_out
+
+- func: cadence::quantize_per_tensor_asym8u.out(Tensor input, float scale, int zero_point, int quant_min, int quant_max, ScalarType dtype, *, Tensor(a!) out) -> Tensor(a!)
+  variants: function
+  kernels:
+    - arg_meta: null
+      kernel_name: impl::generic::quantize_per_tensor_asym8u_out
+
+- func: cadence::quantize_per_tensor_asym16s.out(Tensor input, float scale, int zero_point, int quant_min, int quant_max, ScalarType dtype, *, Tensor(a!) out) -> Tensor(a!)
+  variants: function
+  kernels:
+    - arg_meta: null
+      kernel_name: impl::generic::quantize_per_tensor_asym16s_out
+
+- func: cadence::quantize_per_tensor_asym16u.out(Tensor input, float scale, int zero_point, int quant_min, int quant_max, ScalarType dtype, *, Tensor(a!) out) -> Tensor(a!)
+  variants: function
+  kernels:
+    - arg_meta: null
+      kernel_name: impl::generic::quantize_per_tensor_asym16u_out
+
 - func: cadence::dequantize_per_tensor.out(Tensor input, float scale, int zero_point, int quant_min, int quant_max, ScalarType dtype, *, Tensor(a!) out) -> Tensor(a!)
   variants: function
   kernels:
     - arg_meta: null
       kernel_name: impl::generic::dequantize_per_tensor_out
 
+- func: cadence::dequantize_per_tensor_asym8s.out(Tensor input, float scale, int zero_point, int quant_min, int quant_max, ScalarType dtype, *, Tensor(a!) out) -> Tensor(a!)
+  variants: function
+  kernels:
+    - arg_meta: null
+      kernel_name: impl::generic::dequantize_per_tensor_asym8s_out
+
+- func: cadence::dequantize_per_tensor_asym8u.out(Tensor input, float scale, int zero_point, int quant_min, int quant_max, ScalarType dtype, *, Tensor(a!) out) -> Tensor(a!)
+  variants: function
+  kernels:
+    - arg_meta: null
+      kernel_name: impl::generic::dequantize_per_tensor_asym8u_out
+
+- func: cadence::dequantize_per_tensor_asym16s.out(Tensor input, float scale, int zero_point, int quant_min, int quant_max, ScalarType dtype, *, Tensor(a!) out) -> Tensor(a!)
+  variants: function
+  kernels:
+    - arg_meta: null
+      kernel_name: impl::generic::dequantize_per_tensor_asym16s_out
+
+- func: cadence::dequantize_per_tensor_asym16u.out(Tensor input, float scale, int zero_point, int quant_min, int quant_max, ScalarType dtype, *, Tensor(a!) out) -> Tensor(a!)
+  variants: function
+  kernels:
+    - arg_meta: null
+      kernel_name: impl::generic::dequantize_per_tensor_asym16u_out
+
 - func: cadence::quantized_conv2d_nchw.out(Tensor input, Tensor weight, Tensor bias, int[] stride, SymInt[] padding, int[] dilation, int groups, int input_zero_point, Tensor weight_zero_point, Tensor bias_scale, float out_scale, int out_zero_point, Tensor out_multiplier, Tensor out_shift, *, Tensor(a!) out) -> Tensor(a!)
   kernels:
     - arg_meta: null

backends/cadence/aot/functions_hifi.yaml

@@ -284,12 +284,61 @@
     - arg_meta: null
       kernel_name: impl::HiFi::quantize_per_tensor_out
 
+- func: cadence::quantize_per_tensor_asym8s.out(Tensor input, float scale, int zero_point, int quant_min, int quant_max, ScalarType dtype, *, Tensor(a!) out) -> Tensor(a!)
+  variants: function
+  kernels:
+    - arg_meta: null
+      kernel_name: impl::HiFi::quantize_per_tensor_asym8s_out
+
+- func: cadence::quantize_per_tensor_asym8u.out(Tensor input, float scale, int zero_point, int quant_min, int quant_max, ScalarType dtype, *, Tensor(a!) out) -> Tensor(a!)
+  variants: function
+  kernels:
+    - arg_meta: null
+      kernel_name: impl::HiFi::quantize_per_tensor_asym8u_out
+
+- func: cadence::quantize_per_tensor_asym16s.out(Tensor input, float scale, int zero_point, int quant_min, int quant_max, ScalarType dtype, *, Tensor(a!) out) -> Tensor(a!)
+  variants: function
+  kernels:
+    - arg_meta: null
+      kernel_name: impl::HiFi::quantize_per_tensor_asym16s_out
+
+- func: cadence::quantize_per_tensor_asym16u.out(Tensor input, float scale, int zero_point, int quant_min, int quant_max, ScalarType dtype, *, Tensor(a!) out) -> Tensor(a!)
+  variants: function
+  kernels:
+    - arg_meta: null
+      kernel_name: impl::HiFi::quantize_per_tensor_asym16s_out
+
+
 - func: cadence::dequantize_per_tensor.out(Tensor input, float scale, int zero_point, int quant_min, int quant_max, ScalarType dtype, *, Tensor(a!) out) -> Tensor(a!)
   variants: function
   kernels:
     - arg_meta: null
       kernel_name: impl::HiFi::dequantize_per_tensor_out
 
+- func: cadence::dequantize_per_tensor_asym8s.out(Tensor input, float scale, int zero_point, int quant_min, int quant_max, ScalarType dtype, *, Tensor(a!) out) -> Tensor(a!)
+  variants: function
+  kernels:
+    - arg_meta: null
+      kernel_name: impl::HiFi::dequantize_per_tensor_asym8s_out
+
+- func: cadence::dequantize_per_tensor_asym8u.out(Tensor input, float scale, int zero_point, int quant_min, int quant_max, ScalarType dtype, *, Tensor(a!) out) -> Tensor(a!)
+  variants: function
+  kernels:
+    - arg_meta: null
+      kernel_name: impl::HiFi::dequantize_per_tensor_asym8u_out
+
+- func: cadence::dequantize_per_tensor_asym16s.out(Tensor input, float scale, int zero_point, int quant_min, int quant_max, ScalarType dtype, *, Tensor(a!) out) -> Tensor(a!)
+  variants: function
+  kernels:
+    - arg_meta: null
+      kernel_name: impl::HiFi::dequantize_per_tensor_asym16s_out
+
+- func: cadence::dequantize_per_tensor_asym16u.out(Tensor input, float scale, int zero_point, int quant_min, int quant_max, ScalarType dtype, *, Tensor(a!) out) -> Tensor(a!)
+  variants: function
+  kernels:
+    - arg_meta: null
+      kernel_name: impl::HiFi::dequantize_per_tensor_asym16u_out
+
 - func: cadence::quantized_conv2d_nchw.out(Tensor input, Tensor weight, Tensor bias, int[] stride, SymInt[] padding, int[] dilation, int groups, int input_zero_point, Tensor weight_zero_point, Tensor bias_scale, float out_scale, int out_zero_point, Tensor out_multiplier, Tensor out_shift, *, Tensor(a!) out) -> Tensor(a!)
   kernels:
     - arg_meta: null

backends/cadence/aot/ops_registrations.py

@@ -28,12 +28,64 @@
     "quantize_per_tensor.out(Tensor input, float scale, int zero_point, int quant_min, int quant_max, ScalarType dtype, *, Tensor(a!) out) -> Tensor(a!)"
 )
 
+lib.define(
+    "quantize_per_tensor_asym8s(Tensor input, float scale, int zero_point, int quant_min, int quant_max, ScalarType dtype) -> (Tensor Z)"
+)
+lib.define(
+    "quantize_per_tensor_asym8s.out(Tensor input, float scale, int zero_point, int quant_min, int quant_max, ScalarType dtype, *, Tensor(a!) out) -> Tensor(a!)"
+)
+
+lib.define(
+    "quantize_per_tensor_asym8u(Tensor input, float scale, int zero_point, int quant_min, int quant_max, ScalarType dtype) -> (Tensor Z)"
+)
+lib.define(
+    "quantize_per_tensor_asym8u.out(Tensor input, float scale, int zero_point, int quant_min, int quant_max, ScalarType dtype, *, Tensor(a!) out) -> Tensor(a!)"
+)
+
+lib.define(
+    "quantize_per_tensor_asym16s(Tensor input, float scale, int zero_point, int quant_min, int quant_max, ScalarType dtype) -> (Tensor Z)"
+)
+lib.define(
+    "quantize_per_tensor_asym16s.out(Tensor input, float scale, int zero_point, int quant_min, int quant_max, ScalarType dtype, *, Tensor(a!) out) -> Tensor(a!)"
+)
+
+lib.define(
+    "quantize_per_tensor_asym16u(Tensor input, float scale, int zero_point, int quant_min, int quant_max, ScalarType dtype) -> (Tensor Z)"
+)
+lib.define(
+    "quantize_per_tensor_asym16u.out(Tensor input, float scale, int zero_point, int quant_min, int quant_max, ScalarType dtype, *, Tensor(a!) out) -> Tensor(a!)"
+)
+
 lib.define(
     "dequantize_per_tensor(Tensor input, float scale, int zero_point, int quant_min, int quant_max, ScalarType dtype) -> (Tensor Z)"
 )
 lib.define(
     "dequantize_per_tensor.out(Tensor input, float scale, int zero_point, int quant_min, int quant_max, ScalarType dtype, *, Tensor(a!) out) -> Tensor(a!)"
 )
+lib.define(
+    "dequantize_per_tensor_asym8s(Tensor input, float scale, int zero_point, int quant_min, int quant_max, ScalarType dtype) -> (Tensor Z)"
+)
+lib.define(
+    "dequantize_per_tensor_asym8s.out(Tensor input, float scale, int zero_point, int quant_min, int quant_max, ScalarType dtype, *, Tensor(a!) out) -> Tensor(a!)"
+)
+lib.define(
+    "dequantize_per_tensor_asym8u(Tensor input, float scale, int zero_point, int quant_min, int quant_max, ScalarType dtype) -> (Tensor Z)"
+)
+lib.define(
+    "dequantize_per_tensor_asym8u.out(Tensor input, float scale, int zero_point, int quant_min, int quant_max, ScalarType dtype, *, Tensor(a!) out) -> Tensor(a!)"
+)
+lib.define(
+    "dequantize_per_tensor_asym16s(Tensor input, float scale, int zero_point, int quant_min, int quant_max, ScalarType dtype) -> (Tensor Z)"
+)
+lib.define(
+    "dequantize_per_tensor_asym16s.out(Tensor input, float scale, int zero_point, int quant_min, int quant_max, ScalarType dtype, *, Tensor(a!) out) -> Tensor(a!)"
+)
+lib.define(
+    "dequantize_per_tensor_asym16u(Tensor input, float scale, int zero_point, int quant_min, int quant_max, ScalarType dtype) -> (Tensor Z)"
+)
+lib.define(
+    "dequantize_per_tensor_asym16u.out(Tensor input, float scale, int zero_point, int quant_min, int quant_max, ScalarType dtype, *, Tensor(a!) out) -> Tensor(a!)"
+)
 
 lib.define(
     "quantized_layer_norm(Tensor X, Tensor X_scale, Tensor X_zero_point, int[] normalized_shape, Tensor weight, Tensor bias, float eps, float output_scale, int output_zero_point) -> (Tensor Y)"
@@ -541,6 +593,54 @@ def quantize_per_tensor_meta(
     return input.new_empty(input.size(), dtype=dtype)
 
 
+@register_fake("cadence::quantize_per_tensor_asym8s")
+def quantize_per_tensor_asym8s_meta(
+    input: torch.Tensor,
+    scale: float,
+    zero_point: int,
+    quant_min: int,
+    quant_max: int,
+    dtype: torch.dtype,
+) -> torch.Tensor:
+    return input.new_empty(input.size(), dtype=dtype)
+
+
+@register_fake("cadence::quantize_per_tensor_asym8u")
+def quantize_per_tensor_asym8u_meta(
+    input: torch.Tensor,
+    scale: float,
+    zero_point: int,
+    quant_min: int,
+    quant_max: int,
+    dtype: torch.dtype,
+) -> torch.Tensor:
+    return input.new_empty(input.size(), dtype=dtype)
+
+
+@register_fake("cadence::quantize_per_tensor_asym16s")
+def quantize_per_tensor_asym16s_meta(
+    input: torch.Tensor,
+    scale: float,
+    zero_point: int,
+    quant_min: int,
+    quant_max: int,
+    dtype: torch.dtype,
+) -> torch.Tensor:
+    return input.new_empty(input.size(), dtype=dtype)
+
+
+@register_fake("cadence::quantize_per_tensor_asym16u")
+def quantize_per_tensor_asym16u_meta(
+    input: torch.Tensor,
+    scale: float,
+    zero_point: int,
+    quant_min: int,
+    quant_max: int,
+    dtype: torch.dtype,
+) -> torch.Tensor:
+    return input.new_empty(input.size(), dtype=dtype)
+
+
 @register_fake("cadence::dequantize_per_tensor")
 def dequantize_per_tensor_meta(
     input: torch.Tensor,
@@ -553,6 +653,54 @@ def dequantize_per_tensor_meta(
     return input.new_empty(input.size(), dtype=torch.float)
 
 
+@register_fake("cadence::dequantize_per_tensor_asym8s")
+def dequantize_per_tensor_asym8s_meta(
+    input: torch.Tensor,
+    scale: float,
+    zero_point: int,
+    quant_min: int,
+    quant_max: int,
+    dtype: torch.dtype,
+) -> torch.Tensor:
+    return input.new_empty(input.size(), dtype=torch.float)
+
+
+@register_fake("cadence::dequantize_per_tensor_asym8u")
+def dequantize_per_tensor_asym8u_meta(
+    input: torch.Tensor,
+    scale: float,
+    zero_point: int,
+    quant_min: int,
+    quant_max: int,
+    dtype: torch.dtype,
+) -> torch.Tensor:
+    return input.new_empty(input.size(), dtype=torch.float)
+
+
+@register_fake("cadence::dequantize_per_tensor_asym16s")
+def dequantize_per_tensor_asym16s_meta(
+    input: torch.Tensor,
+    scale: float,
+    zero_point: int,
+    quant_min: int,
+    quant_max: int,
+    dtype: torch.dtype,
+) -> torch.Tensor:
+    return input.new_empty(input.size(), dtype=torch.float)
+
+
+@register_fake("cadence::dequantize_per_tensor_asym16u")
+def dequantize_per_tensor_asym16u_meta(
+    input: torch.Tensor,
+    scale: float,
+    zero_point: int,
+    quant_min: int,
+    quant_max: int,
+    dtype: torch.dtype,
+) -> torch.Tensor:
+    return input.new_empty(input.size(), dtype=torch.float)
+
+
 @register_fake("cadence::quantized_add")
 def quantized_add_meta(
     X: torch.Tensor,

backends/cadence/aot/type_dispatch.py

@@ -27,6 +27,7 @@ class OpConfig:
     base_name: str
     type_dispatch_suffixes: dict[tuple[torch.dtype, ...], str]
     weight_arg_idx: Optional[int] = None
+    is_quant_op: bool = False
     variant: str = "per_tensor"
 
 
@@ -100,6 +101,27 @@ class CompileTimeTypeDispatchPass(ExportPass):
             },
             variant="default",
         ),
+        exir_ops.edge.cadence.quantize_per_tensor.default: OpConfig(
+            "quantize_per_tensor",
+            type_dispatch_suffixes={
+                (torch.int8,): "asym8s",
+                (torch.uint8,): "asym8u",
+                (torch.int16,): "asym16s",
+                (torch.uint16,): "asym16s",
+            },
+            variant="default",
+            is_quant_op=True,
+        ),
+        exir_ops.edge.cadence.dequantize_per_tensor.default: OpConfig(
+            "dequantize_per_tensor",
+            type_dispatch_suffixes={
+                (torch.int8,): "asym8s",
+                (torch.uint8,): "asym8u",
+                (torch.int16,): "asym16s",
+                (torch.uint16,): "asym16s",
+            },
+            variant="default",
+        ),
     }
 
     def call_operator(
@@ -120,6 +142,8 @@ def call_operator(
         if config.weight_arg_idx is not None:
             weight_dtype = args[config.weight_arg_idx].to_tensor().dtype
             dtype_key = (input_dtype, weight_dtype)
+        elif config.is_quant_op:
+            dtype_key = (args[5],)
         else:
             dtype_key = (input_dtype,)