Correct QLinearMatMul broadcasting and QLinearConv bias quantization

src/onnx/broadcast_qdq.cpp

@@ -1,7 +1,7 @@
 /*
  * The MIT License (MIT)
  *
- * Copyright (c) 2015-2023 Advanced Micro Devices, Inc. All rights reserved.
+ * Copyright (c) 2015-2026 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
@@ -61,6 +61,42 @@ instruction_ref bcast_qdq_instr(const std::string& op_name,
     return info.add_instruction(migraphx::make_op(op_name), x_in, bcast_scale, bcast_zero_pt);
 }
 
+instruction_ref bcast_qdq_instr_matmul(const std::string& op_name,
+                                       instruction_ref x_in,
+                                       instruction_ref arg_fscale,
+                                       instruction_ref arg_z_pt,
+                                       const onnx_parser::node_info& info)
+{
+    auto in_lens = x_in->get_shape().lens();
+
+    // prep 1: broadcast scale. it can come as a scalar or a 1-D tensor.
+    instruction_ref bcast_scale;
+    if(arg_fscale->get_shape().elements() > 1)
+    {
+        auto axis   = x_in->get_shape().lens().size() - arg_fscale->get_shape().lens().size();
+        bcast_scale = info.add_instruction(
+            migraphx::make_op("broadcast", {{"axis", axis}, {"out_lens", in_lens}}), arg_fscale);
+    }
+    else
+        bcast_scale = info.add_instruction(
+            migraphx::make_op("multibroadcast", {{"out_lens", in_lens}}), arg_fscale);
+
+    // prep 2: broadcast zero point. it can come as a scalar or a 1-D tensor.
+    instruction_ref bcast_zero_pt;
+    if(arg_z_pt->get_shape().elements() > 1)
+    {
+        auto axis     = x_in->get_shape().lens().size() - arg_z_pt->get_shape().lens().size();
+        bcast_zero_pt = info.add_instruction(
+            migraphx::make_op("broadcast", {{"axis", axis}, {"out_lens", in_lens}}), arg_z_pt);
+    }
+    else
+        bcast_zero_pt = info.add_instruction(
+            migraphx::make_op("multibroadcast", {{"out_lens", in_lens}}), arg_z_pt);
+
+    // op_name is either quantizelinear or dequantizelinear:
+    return info.add_instruction(migraphx::make_op(op_name), x_in, bcast_scale, bcast_zero_pt);
+}
+
 // Multibroadcast a scaler..
 instruction_ref bcast_scalar_instr(const migraphx::shape& shape_out,
                                    instruction_ref arg_in,

src/onnx/include/migraphx/onnx/broadcast_qdq.hpp

@@ -1,7 +1,7 @@
 /*
  * The MIT License (MIT)
  *
- * Copyright (c) 2015-2023 Advanced Micro Devices, Inc. All rights reserved.
+ * Copyright (c) 2015-2026 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
@@ -44,6 +44,12 @@ instruction_ref bcast_qdq_instr(const std::string& op_name,
                                 instruction_ref arg_z_pt,
                                 const onnx_parser::node_info& info);
 
+instruction_ref bcast_qdq_instr_matmul(const std::string& op_name,
+                                       instruction_ref x_in,
+                                       instruction_ref arg_fscale,
+                                       instruction_ref arg_z_pt,
+                                       const onnx_parser::node_info& info);
+
 // Multibroadcast a scaler..
 instruction_ref bcast_scalar_instr(const migraphx::shape& shape_out,
                                    instruction_ref arg_in,

src/onnx/parse_qlinearconv.cpp

@@ -1,7 +1,7 @@
 /*
  * The MIT License (MIT)
  *
- * Copyright (c) 2015-2024 Advanced Micro Devices, Inc. All rights reserved.
+ * Copyright (c) 2015-2026 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
@@ -229,7 +229,25 @@ struct parse_qlinearconv : op_parser<parse_qlinearconv>
 
         // Biases, if any.. : is an optional argument.
         if(args.size() > 8)
-            conv_x_w = add_bias_to_conv(args[8], conv_x_w, info);
+        {
+            const auto& in_b = args[8];
+            auto b_sh        = in_b->get_shape();
+
+            auto bcast_scale_x = info.add_instruction(
+                migraphx::make_op("multibroadcast", {{"out_lens", in_scale_w->get_shape().lens()}}),
+                in_scale_x);
+
+            auto bias_scale =
+                info.add_instruction(migraphx::make_op("mul"), bcast_scale_x, in_scale_w);
+            auto zero_lit = info.add_literal(
+                migraphx::literal{migraphx::shape{migraphx::shape::int32_type}, {0}});
+            auto bias_zp = info.add_instruction(
+                migraphx::make_op("multibroadcast", {{"out_lens", b_sh.lens()}}), zero_lit);
+            auto dquant_bias = info.add_instruction(
+                migraphx::make_op("dequantizelinear"), args[8], bias_scale, bias_zp);
+
+            conv_x_w = add_bias_to_conv(dquant_bias, conv_x_w, info);
+        }
 
         return bcast_qdq_instr("quantizelinear", conv_x_w, in_scale_y, in_zero_pt_y, info);
     }

src/onnx/parse_qlinearmatmul.cpp

@@ -1,7 +1,7 @@
 /*
  * The MIT License (MIT)
  *
- * Copyright (c) 2015-2023 Advanced Micro Devices, Inc. All rights reserved.
+ * Copyright (c) 2015-2026 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
@@ -138,9 +138,17 @@ struct parse_qlinearmatmul : op_parser<parse_qlinearmatmul>
            not std::equal(lens_a.rbegin() + 2, lens_a.rend(), lens_b.rbegin() + 2, lens_b.rend()))
             MIGRAPHX_THROW("QLINEARMATMUL: mismatched input dimensions");
 
-        if(migraphx::any_of({args[1], args[2], args[4], args[5]},
+        if(migraphx::any_of({args[1], args[2]},
                             [](auto arg) { return not arg->get_shape().scalar(); }))
             MIGRAPHX_THROW("QLINEARMATMUL: unsupported row/column quantization");
+
+        const auto& in_scale_b   = args[4];
+        const auto& in_zero_pt_b = args[5];
+        size_t dim_scale_b       = in_scale_b->get_shape().lens().size();
+        size_t dim_zero_pt_b     = in_zero_pt_b->get_shape().lens().size();
+
+        if((dim_scale_b > 1) or (dim_zero_pt_b > 1))
+            MIGRAPHX_THROW("QLINEARMATMUL: unsupported row/column quantization");
     }
 
     instruction_ref parse(const op_desc& /* opd */,
@@ -154,13 +162,15 @@ struct parse_qlinearmatmul : op_parser<parse_qlinearmatmul>
         const auto& in_a         = args[0];
         const auto& in_scale_a   = args[1];
         const auto& in_zero_pt_a = args[2];
-        auto dquant_a = bcast_qdq_instr("dequantizelinear", in_a, in_scale_a, in_zero_pt_a, info);
+        auto dquant_a =
+            bcast_qdq_instr_matmul("dequantizelinear", in_a, in_scale_a, in_zero_pt_a, info);
 
         // B
         const auto& in_b         = args[3];
         const auto& in_scale_b   = args[4];
         const auto& in_zero_pt_b = args[5];
-        auto dquant_b = bcast_qdq_instr("dequantizelinear", in_b, in_scale_b, in_zero_pt_b, info);
+        auto dquant_b =
+            bcast_qdq_instr_matmul("dequantizelinear", in_b, in_scale_b, in_zero_pt_b, info);
 
         bool is_a_prepended = false;
         bool is_b_appended  = false;