[QNN] Implement quantized avg_pool2d

* qnn.avg_pool2d is integrated into Relay and integration tests are added
* FQ2I is modified to pick up qnn.avg_pool2d
* Canonicalization and layout conversion for qnn.avg_pool2d are implemented
* Dynamic PoolArea computation is implemented for both qnn and slice_op avg_pool2d
  to exclude padding area in averaging calculation when count_include_pad=False

Co-authored-by: Fateme Hosseini <quic_fhossein@quicinc.com>
Co-authored-by: Jyotsna Verma <jverma@quicinc.com>
Co-authored-by: Anirudh Sundar <quic_sanirudh@quicinc.com>
Co-authored-by: Venkat Rasagna Reddy Komatireddy <quic_rasagna@quicinc.com>

python/tvm/relay/qnn/op/_qnn.py

@@ -167,3 +167,6 @@ def alter_op_layout_qnn_dense(attrs, inputs, tinfos, out_type):
 # qnn.batch_matmul
 register_strategy("qnn.batch_matmul", strategy.qnn_batch_matmul_strategy)
 register_pattern("qnn.batch_matmul", OpPattern.OUT_ELEMWISE_FUSABLE)
+
+# qnn.avg_pool2d
+register_strategy("qnn.avg_pool2d", strategy.qnn_avg_pool2d_strategy)

python/tvm/relay/qnn/op/layout_conversions.py

@@ -126,3 +126,38 @@ def convert_qnn_conv2d_transpose(attrs, inputs, tinfos, desired_layouts):
         return relay.qnn.op.conv2d_transpose(*inputs, **new_attrs)
 
     raise ValueError(f"Layout {desired_data_layout} is not yet supported")
+
+
+@reg.register_convert_op_layout("qnn.avg_pool2d")
+def convert_qnn_avg_pool2d(attrs, inputs, tinfos, desired_layouts):
+    """Convert Layout pass registration for QNN avg_pool2d op.
+
+    Parameters
+    ----------
+    attrs : tvm.ir.Attrs
+        Attributes of current avg_pool2d
+    inputs : list of tvm.relay.Expr
+        The args of the Relay expr to be legalized
+    tinfos : list of types
+        List of input and output types
+    desired_layouts : list of layout strings
+        List of layouts defining our desired
+        layout for the data input.
+
+    Returns
+    -------
+    result : tvm.relay.Expr
+        The transformed expr
+    """
+    # pylint: disable=import-outside-toplevel
+    from tvm import relay
+
+    assert len(desired_layouts) == 1, "A desired layout is expected for qnn.avg_pool2d's input"
+    desired_data_layout = desired_layouts[0]
+    if desired_data_layout == "NCHW" or desired_data_layout == "NHWC":
+        new_attrs = dict(attrs)
+        new_attrs["layout"] = str(desired_data_layout)
+        new_attrs["out_layout"] = str(desired_data_layout)
+        return relay.qnn.op.avg_pool2d(*inputs, **new_attrs)
+
+    raise ValueError(f"Layout {desired_data_layout} is not yet supported")

python/tvm/relay/qnn/op/qnn.py

@@ -1249,3 +1249,69 @@ def leaky_relu(x, alpha, input_scale, input_zero_point, output_scale, output_zer
 
 def softmax(x, scale, zero_point, output_scale, output_zero_point, axis=-1):
     return _make.softmax(x, axis, scale, zero_point, output_scale, output_zero_point)
+
+
+def avg_pool2d(
+    data,
+    input_scale,
+    input_zero_point,
+    output_scale,
+    output_zero_point,
+    pool_size,
+    strides,
+    padding,
+    dilation,
+    ceil_mode=False,
+    count_include_pad=True,
+    layout="NHWC",
+    out_layout="",
+):
+
+    """Quantized avg_pool2d
+
+    Parameters
+    ----------
+    data : relay.Expr
+        The quantized input tensor.
+    input_scale: float
+        The scale of the input quantized expr.
+    input_zero_point: int
+        The zero point of input quantized expr.
+    output_scale: flaot
+        The scale of the output quantized expr.
+    output_zero_point: int
+       The zero point of output quantized expr.
+    pool_size : relay.Expr
+        The pool_size
+    strides : relay.Expr
+        The strides
+    padding : relay.Expr
+        The padding size
+    dilation : relay.Expr
+        The dilation size
+    ceil_mode : bool, optional
+        Whether to use ceil or floor for calculating the output shape
+    count_include_pad : bool, optional
+        Determines if padding should be taken into account in the computation
+    layout: string, optinal
+    out_layout: string, optional
+    Returns
+    -------
+    result : relay.Expr
+        The computed result.
+    """
+    return _make.avg_pool2d(
+        data,
+        input_scale,
+        input_zero_point,
+        output_scale,
+        output_zero_point,
+        pool_size,
+        strides,
+        padding,
+        dilation,
+        ceil_mode,
+        count_include_pad,
+        layout,
+        out_layout,
+    )

python/tvm/relay/qnn/strategy/generic.py

@@ -157,6 +157,36 @@ def wrapper(_attrs, inputs, out_type):
     return wrapper
 
 
+def wrap_compute_qnn_avg_pool2d(topi_compute):
+    """Wrap qnn.avg_pool2d topi compute"""
+
+    def wrapper(attrs, inputs, out_type):
+        kernel = attrs.pool_size
+        strides = attrs.strides
+        padding = attrs.padding
+        dilation = attrs.dilation
+        count_include_pad = attrs.count_include_pad
+        oshape = out_type.shape
+        odtype = out_type.dtype
+        args = [
+            inputs[0],
+            kernel,
+            strides,
+            padding,
+            dilation,
+            count_include_pad,
+            oshape,
+            odtype,
+            inputs[1],
+            inputs[2],
+            inputs[3],
+            inputs[4],
+        ]
+        return [topi_compute(*args)]
+
+    return wrapper
+
+
 def wrap_topi_concatenate(topi_compute):
     """Wrap TOPI compute which use qnn.concatenate attrs"""
 
@@ -280,3 +310,12 @@ def qnn_batch_matmul_strategy(attrs, inputs, out_type, target):
         "qnn.batch_matmul is currently only supported with Hexagon. "
         "Please run QNN Canonicalize pass to decompose this op into supported ops."
     )
+
+
+@override_native_generic_func("qnn_avg_pool2d_strategy")
+def qnn_avg_pool2d_strategy(attrs, inputs, out_type, target):
+    """qnn.avg_pool2d generic strategy"""
+    raise RuntimeError(
+        "qnn.avg_pool2d is currently only supported with Hexagon. "
+        "Please run QNN Canonicalize pass to decompose this op into supported ops."
+    )

python/tvm/relay/qnn/strategy/hexagon.py

@@ -201,3 +201,27 @@ def qnn_batch_matmul_strategy_hexagon(attrs, inputs, out_type, target):
         name="qnn_batch_matmul.hexagon",
     )
     return strategy
+
+
+@qnn_avg_pool2d_strategy.register(["hexagon"])
+def qnn_avg_pool2d_strategy_hexagon(attrs, inputs, out_type, target):
+    """qnn.avg_pool2d strategy for Hexagon"""
+    data_layout = attrs.layout
+    if data_layout == "NHWC":
+        strategy = _op.OpStrategy()
+        strategy.add_implementation(
+            wrap_compute_qnn_avg_pool2d(topi.hexagon.qnn.qnn_avg_pool2d_wrapper_compute_NHWC),
+            wrap_topi_schedule(topi.hexagon.qnn.schedule_qnn_avg_pool2d),
+            name="qnn_avg_pool2d.hexagon",
+        )
+        return strategy
+    elif data_layout == "NCHW":
+        strategy = _op.OpStrategy()
+        strategy.add_implementation(
+            wrap_compute_qnn_avg_pool2d(topi.hexagon.qnn.qnn_avg_pool2d_wrapper_compute_NCHW),
+            wrap_topi_schedule(topi.hexagon.qnn.schedule_qnn_avg_pool2d),
+            name="qnn_avg_pool2d.hexagon",
+        )
+        return strategy
+    else:
+        raise RuntimeError("Unsupported strategy for qnn.avg_pool2d")

python/tvm/relay/transform/fake_quantization_to_integer.py

@@ -111,6 +111,32 @@ def identity(expr, type_map):
 register_unary_identity("image.resize2d")
 
 
+@register_fake_quantization_to_integer("nn.avg_pool2d")
+def avgpool2d(expr, type_map):
+    """Rewrite an avgpool op"""
+    attrs = {**expr.attrs}
+    arg = expr.args[0]
+    t = type_map[arg]
+    out_t = type_map[expr]
+
+    out = relay.qnn.op.avg_pool2d(
+        arg,
+        t.scale,
+        t.zero_point,
+        out_t.scale,
+        out_t.zero_point,
+        attrs["pool_size"],
+        attrs["strides"],
+        attrs["padding"],
+        attrs["dilation"],
+        attrs["ceil_mode"],
+        attrs["count_include_pad"],
+        attrs["layout"],
+    )
+
+    return [out, TensorAffineType(out_t.scale, out_t.zero_point, out_t.dtype, out_t.axis)]
+
+
 @register_fake_quantization_to_integer("nn.adaptive_avg_pool1d")
 def adaptive_avgpool1d(expr, type_map):
     """Rewrite an adaptive avgpool op"""
@@ -138,37 +164,6 @@ def adaptive_avgpool1d(expr, type_map):
     return [out, TensorAffineType(out_t.scale, out_t.zero_point, "int32", out_t.axis)]
 
 
-@register_fake_quantization_to_integer("nn.avg_pool2d")
-def avgpool2d(expr, type_map):
-    """Rewrite a avgpool op"""
-    arg = expr.args[0]
-    t = type_map[arg]
-    out_t = type_map[expr]
-    # Cast (or requantize) to int32.
-    if not (
-        approx_equal(t.scale, out_t.scale)
-        and approx_equal(t.zero_point, out_t.zero_point)
-        and tvm.ir.structural_equal(t.dtype, out_t.dtype)
-    ):
-        arg = relay.qnn.op.requantize(
-            arg,
-            t.scale,
-            t.zero_point,
-            out_t.scale,
-            out_t.zero_point,
-            out_dtype="int32",
-            axis=t.axis,
-        )
-    else:
-        arg = relay.op.cast(arg, "int32")
-    out = relay.op.nn.avg_pool2d(arg, **expr.attrs)
-    if out_t.dtype != "int32":
-        # Cast back to output dtype to preserve input dtype == output dtype for AvgPool2d.
-        out = relay.op.clip(out, a_min=np.iinfo(out_t.dtype).min, a_max=np.iinfo(out_t.dtype).max)
-        out = relay.op.cast(out, out_t.dtype)
-    return [out, TensorAffineType(out_t.scale, out_t.zero_point, out_t.dtype, out_t.axis)]
-
-
 @register_fake_quantization_to_integer("nn.global_avg_pool2d")
 def global_avgpool2d(expr, type_map):
     """Rewrite a global_avgpool op"""