[QNN] Support 4D padding. (apache#5036)

* [QNN] Support 4D padding.

* Empty commit.

Co-authored-by: Ubuntu <ubuntu@ip-172-31-38-96.us-west-2.compute.internal>

Author: 2 people

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

@@ -19,6 +19,7 @@
 
 from __future__ import absolute_import as _abs
 from tvm.relay.expr import Tuple
+from tvm.relay.op.nn.util import get_pad_tuple2d
 from . import _make
 
 def requantize(data,
@@ -280,6 +281,9 @@ def conv2d(data,
         The computed result.
     """
 
+    # TODO enforce 4-way padding in topi/nn/conv2d after #4644 merged
+    # convert 2-way padding to 4-way padding
+    padding = get_pad_tuple2d(padding)
     return _make.conv2d(data, kernel,
                         input_zero_point, kernel_zero_point,
                         input_scale, kernel_scale,

src/relay/qnn/op/convolution.cc

@@ -177,13 +177,17 @@ Expr Conv2DFallBack(const Expr& data, const Expr& weight, const Expr& input_zero
 Expr Conv2DPadInput(const Expr& data, const Expr& input_zero_point, const Conv2DAttrs* param) {
   // 1) Pad the input data
   auto padded_data = data;
-  auto pad_h_value = get_const_int(param->padding[0]);
-  auto pad_w_value = get_const_int(param->padding[1]);
-  if (pad_h_value != 0 || pad_w_value != 0) {
+  auto pad_top_value = get_const_int(param->padding[0]);
+  auto pad_left_value = get_const_int(param->padding[1]);
+  auto pad_bottom_value = get_const_int(param->padding[2]);
+  auto pad_right_value = get_const_int(param->padding[3]);
+  bool do_pad = pad_top_value != 0 || pad_left_value != 0 ||
+                pad_bottom_value != 0 || pad_right_value != 0;
+  if (do_pad) {
     Array<IndexExpr> pad_n({0, 0});
     Array<IndexExpr> pad_c({0, 0});
-    Array<IndexExpr> pad_h({param->padding[0], param->padding[0]});
-    Array<IndexExpr> pad_w({param->padding[1], param->padding[1]});
+    Array<IndexExpr> pad_h({param->padding[0], param->padding[2]});
+    Array<IndexExpr> pad_w({param->padding[1], param->padding[3]});
 
     Array<Array<IndexExpr>> pad_width;
     if (param->data_layout == "NCHW") {
@@ -336,7 +340,7 @@ Expr DepthwiseConv2DFourthTerm(int input_zero_point_int, int kernel_zero_point_i
  */
 Expr Conv2DFirstTerm(const Expr& padded_data, const Expr& weight, const Conv2DAttrs* param) {
   // Lowering for Term 1
-  Array<IndexExpr> padding({0, 0});
+  Array<IndexExpr> padding({0, 0, 0, 0});
   return Conv2D(padded_data, weight, param->strides, padding, param->dilation, param->groups,
                 param->channels, param->kernel_size, param->data_layout, param->kernel_layout,
                 param->out_layout, param->out_dtype);
@@ -583,7 +587,6 @@ Expr QnnConv2DCanonicalize(const Attrs& attrs, const Array<Expr>& new_args,
   const auto* param = attrs.as<Conv2DAttrs>();
   CHECK(param != nullptr);
   // Assertion checks for exisiing support.
-  CHECK_EQ(param->padding.size(), 2) << "qnn.conv2d only supports 2D padding";
   CHECK(param->data_layout == "NCHW" || param->data_layout == "NHWC")
       << "qnn.conv2d supports only NCHW/NHWC input data layout.";
   CHECK(param->kernel_layout == "OIHW" || param->kernel_layout == "HWIO" ||

tests/python/relay/test_op_qnn_conv2d.py

@@ -496,6 +496,30 @@ def test_padding():
         verify(ref_func, qnn_func, data_shape, data_dtype,
                 kernel_shape, kernel_dtype)
 
+        # Try asymmetric padding
+        data_shape = (2, 2, 4, 4) # NHWC
+        data_dtype = 'uint8'
+        kernel_shape = (2, 2, 4, 3) # HWIO
+        kernel_dtype = 'uint8'
+        ref_func, qnn_func = get_funcs(data_shape=data_shape,
+                                       data_dtype=data_dtype,
+                                       kernel_shape=kernel_shape,
+                                       kernel_dtype=kernel_dtype,
+                                       input_zero_point=8,
+                                       kernel_zero_point=3,
+                                       input_scale=1.0,
+                                       kernel_scale=1.0,
+                                       kernel_size=(2, 2),
+                                       padding=(1, 1, 2, 2),
+                                       strides=(1, 1),
+                                       dilation=(1, 1),
+                                       data_layout="NHWC",
+                                       kernel_layout="HWIO",
+                                       out_dtype="int32")
+        verify(ref_func, qnn_func, data_shape, data_dtype,
+                kernel_shape, kernel_dtype)
+
+
 def test_dilation():
     with TempOpAttr("qnn.conv2d", "FTVMQnnLegalize", legalize_qnn_conv2d):