ConvInteger: fix parsing for x_zero_point and w_zero_point

src/onnx/parse_convolution.cpp

@@ -1,7 +1,7 @@
 /*
  * The MIT License (MIT)
  *
- * Copyright (c) 2015-2024 Advanced Micro Devices, Inc. All rights reserved.
+ * Copyright (c) 2015-2025 Advanced Micro Devices, Inc. All rights reserved.
  *
  * Permission is hereby granted, free of charge, to any person obtaining a copy
  * of this software and associated documentation files (the "Software"), to deal
@@ -141,17 +141,14 @@ struct parse_convolution : op_parser<parse_convolution>
         return all_zeros;
     }
 
-    static auto
+    static migraphx::operation
     qparam_broadcast_op(instruction_ref qparam, std::vector<std::size_t> lens, std::size_t axis)
     {
-        if(qparam->get_shape().scalar())
+        if(qparam->get_shape().elements() == 1)
         {
             return migraphx::make_op("multibroadcast", {{"out_lens", lens}});
         }
-        else
-        {
-            return migraphx::make_op("broadcast", {{"out_lens", lens}, {"axis", axis}});
-        }
+        return migraphx::make_op("broadcast", {{"out_lens", lens}, {"axis", axis}});
     }
 
     static instruction_ref handle_quant_bias(const operation& op,
@@ -162,27 +159,37 @@ struct parse_convolution : op_parser<parse_convolution>
                                              const instruction_ref& w_zp,
                                              onnx_parser::node_info& info)
     {
+        // to handle the bias, apply the following transformation:
+        // conv(x-x_zp,w-w_zp) = conv(x,w) - conv(x_zp,w) - conv(x,w_zp) + conv(x_zp,w_zp)
         instruction_ref ret = input;
+
+        // multibroadcast (or broadcast) zero points according to spec
+        // x_zp should be a scalar or literal with one element
+        // w_zp can be either a single element or a 1d tensor with size out_channels
+        migraphx::operation x_zp_bc =
+            migraphx::make_op("multibroadcast", {{"out_lens", x->get_shape().lens()}});
+        migraphx::operation w_zp_bc = qparam_broadcast_op(w_zp, weights->get_shape().lens(), 0);
+
         if(not is_symmetric_zero_point(x_zp))
         {
-            auto out_zp_1 = info.add_common_op(op.name(), x_zp, weights);
+            auto x_zp_mb  = info.add_instruction(x_zp_bc, x_zp);
+            auto out_zp_1 = info.add_instruction(op, x_zp_mb, weights);
             ret           = info.add_common_op("sub", ret, out_zp_1);
         }
 
         if(not is_symmetric_zero_point(w_zp))
         {
-            auto out_zp_2 = info.add_common_op(op.name(), x, w_zp);
+            auto w_zp_mb  = info.add_instruction(w_zp_bc, w_zp);
+            auto out_zp_2 = info.add_instruction(op, x, w_zp_mb);
             ret           = info.add_common_op("sub", ret, out_zp_2);
         }
 
         if(not(is_symmetric_zero_point(x_zp)) and not(is_symmetric_zero_point(w_zp)))
         {
-            auto x_zp_bc =
-                info.add_instruction(qparam_broadcast_op(x_zp, x->get_shape().lens(), 0), x_zp);
-            auto w_zp_bc = info.add_instruction(
-                qparam_broadcast_op(w_zp, weights->get_shape().lens(), 0), w_zp);
+            auto x_zp_mb = info.add_instruction(x_zp_bc, x_zp);
+            auto w_zp_mb = info.add_instruction(w_zp_bc, w_zp);
 
-            auto out_zp_3 = info.add_instruction(op, x_zp_bc, w_zp_bc);
+            auto out_zp_3 = info.add_instruction(op, x_zp_mb, w_zp_mb);
 
             ret = info.add_common_op("add", ret, out_zp_3);
         }

test/onnx/convinteger_dual_bias_test.onnx

@@ -8,16 +8,18 @@ B
 strides@
@
 
convinteger_dual_bias_testZ
 
0
 
-

+
 
-
-Z
+
+
+
+Z
 
1
 
-

-
+
 
-Z
+
+Z
 
2
 
 
@@ -28,7 +30,7 @@ B
 
b
 
4
 
-

-

+
 
-B
+
+B

test/onnx/gen_onnx.py

@@ -1686,11 +1686,11 @@ def convinteger_bias_test():
 
 @onnx_test()
 def convinteger_dual_bias_test():
-    x = helper.make_tensor_value_info('0', TensorProto.INT8, [1, 3, 5, 5])
-    y = helper.make_tensor_value_info('1', TensorProto.INT8, [1, 3, 2, 2])
+    x = helper.make_tensor_value_info('0', TensorProto.INT8, [2, 3, 10, 10])
+    y = helper.make_tensor_value_info('1', TensorProto.INT8, [4, 2, 3, 3])
     z = helper.make_tensor_value_info('2', TensorProto.INT8, [1])
     w = helper.make_tensor_value_info('3', TensorProto.INT8, [1])
-    out = helper.make_tensor_value_info('4', TensorProto.INT32, [1, 1, 4, 4])
+    out = helper.make_tensor_value_info('4', TensorProto.INT32, [2, 4, 8, 8])
 
     node = onnx.helper.make_node('ConvInteger',
                                  inputs=['0', '1', '2', '3'],

test/onnx/parse/convinteger_dual_bias_test.cpp

@@ -1,7 +1,7 @@
 /*
  * The MIT License (MIT)
  *
- * Copyright (c) 2015-2024 Advanced Micro Devices, Inc. All rights reserved.
+ * Copyright (c) 2015-2025 Advanced Micro Devices, Inc. All rights reserved.
  *
  * Permission is hereby granted, free of charge, to any person obtaining a copy
  * of this software and associated documentation files (the "Software"), to deal
@@ -28,41 +28,37 @@ TEST_CASE(convinteger_dual_bias_test)
 {
     migraphx::program p;
     auto* mm         = p.get_main_module();
-    auto data        = mm->add_parameter("0", {migraphx::shape::int8_type, {1, 3, 5, 5}});
-    auto weight      = mm->add_parameter("1", {migraphx::shape::int8_type, {1, 3, 2, 2}});
+    auto data        = mm->add_parameter("0", {migraphx::shape::int8_type, {2, 3, 10, 10}});
+    auto weight      = mm->add_parameter("1", {migraphx::shape::int8_type, {4, 2, 3, 3}});
     auto data_bias   = mm->add_parameter("2", {migraphx::shape::int8_type, {1}, {1}});
     auto weight_bias = mm->add_parameter("3", {migraphx::shape::int8_type, {1}, {1}});
 
     auto quant = mm->add_instruction(migraphx::make_op("quant_convolution"), data, weight);
 
     auto mbcast_data_bias = mm->add_instruction(
-        migraphx::make_op("multibroadcast", {{"out_lens", weight->get_shape().lens()}}), data_bias);
+        migraphx::make_op("multibroadcast", {{"out_lens", data->get_shape().lens()}}), data_bias);
 
-    auto quant_db_w =
+    auto quant_mb_w =
         mm->add_instruction(migraphx::make_op("quant_convolution"), mbcast_data_bias, weight);
 
-    auto quant_mb_w = mm->add_instruction(
-        migraphx::make_op("multibroadcast", {{"out_lens", quant->get_shape().lens()}}), quant_db_w);
-
     quant = mm->add_instruction(migraphx::make_op("sub"), quant, quant_mb_w);
 
     auto mbcast_weight_bias = mm->add_instruction(
-        migraphx::make_op("multibroadcast", {{"out_lens", data->get_shape().lens()}}), weight_bias);
+        migraphx::make_op("multibroadcast", {{"out_lens", weight->get_shape().lens()}}),
+        weight_bias);
 
-    auto quant_d_wb =
+    auto quant_md_wb =
         mm->add_instruction(migraphx::make_op("quant_convolution"), data, mbcast_weight_bias);
 
-    auto quant_md_wb = mm->add_instruction(
-        migraphx::make_op("multibroadcast", {{"out_lens", quant->get_shape().lens()}}), quant_d_wb);
-
     quant = mm->add_instruction(migraphx::make_op("sub"), quant, quant_md_wb);
 
-    auto bcast_data_bias = mm->add_instruction(
-        migraphx::make_op("broadcast", {{"out_lens", data->get_shape().lens()}}), data_bias);
-    auto bcast_weight_bias = mm->add_instruction(
-        migraphx::make_op("broadcast", {{"out_lens", weight->get_shape().lens()}}), weight_bias);
+    mbcast_data_bias = mm->add_instruction(
+        migraphx::make_op("multibroadcast", {{"out_lens", data->get_shape().lens()}}), data_bias);
+    mbcast_weight_bias = mm->add_instruction(
+        migraphx::make_op("multibroadcast", {{"out_lens", weight->get_shape().lens()}}),
+        weight_bias);
     auto bias_quant = mm->add_instruction(
-        migraphx::make_op("quant_convolution"), bcast_data_bias, bcast_weight_bias);
+        migraphx::make_op("quant_convolution"), mbcast_data_bias, mbcast_weight_bias);
 
     mm->add_instruction(migraphx::make_op("add"), quant, bias_quant);