Merge branch 'develop' into missing_exports

src/onnx/padding.cpp

@@ -47,7 +47,7 @@ void cal_auto_padding_size(onnx_parser::node_info info,
         return;
     }
 
-    auto auto_pad = info.attributes["auto_pad"].s();
+    auto auto_pad = to_upper(info.attributes["auto_pad"].s());
     if(auto_pad.find("SAME") != std::string::npos)
     {
         bool is_same_upper = (auto_pad.find("SAME_UPPER") != std::string::npos);

src/onnx/parse_qlinearconv.cpp

@@ -0,0 +1,241 @@
+/*
+ * The MIT License (MIT)
+ *
+ * Copyright (c) 2015-2023 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
+ * in the Software without restriction, including without limitation the rights
+ * to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+ * copies of the Software, and to permit persons to whom the Software is
+ * furnished to do so, subject to the following conditions:
+ *
+ * The above copyright notice and this permission notice shall be included in
+ * all copies or substantial portions of the Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+ * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+ * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL THE
+ * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+ * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+ * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN
+ * THE SOFTWARE.
+ */
+
+#include <migraphx/onnx/op_parser.hpp>
+#include <migraphx/onnx/padding.hpp>
+#include <migraphx/onnx/conv.hpp>
+#include <migraphx/ranges.hpp>
+#include <migraphx/make_op.hpp>
+#include <migraphx/onnx/checks.hpp>
+#include <migraphx/onnx/broadcast_qdq.hpp>
+#include <migraphx/instruction.hpp>
+#include <migraphx/stringutils.hpp>
+
+namespace migraphx {
+inline namespace MIGRAPHX_INLINE_NS {
+namespace onnx {
+
+/*
+ *********************************************************************************
+ *  Reference: see QLinearConv in                                                *
+ *  https://github.com/microsoft/onnxruntime/blob/main/docs/ContribOperators.md  *
+ *********************************************************************************
+
+com.microsoft.QLinearConv
+
+Version
+This version of the operator has been available since version 1 of the 'com.microsoft' operator set.
+
+ATTRIBUTES:
+auto_pad : string
+channels_last : int
+dilations : list of ints
+group : int
+kernel_shape : list of ints
+pads : list of ints
+strides : list of ints
+
+INPUTS (8 - 9):
+x : T1
+x_scale : tensor(float)
+x_zero_point : T1
+w : T2
+w_scale : tensor(float)
+w_zero_point : T2
+y_scale : tensor(float)
+y_zero_point : T3
+B (optional) : T4
+
+OUTPUTS:
+y : T3
+
+Type Constraints:
+T1 : tensor(int8), tensor(uint8)
+T2 : tensor(int8), tensor(uint8)
+T3 : tensor(int8), tensor(uint8)
+T4 : tensor(int32)
+
+More details also at:
+https://xadupre.github.io/draft/onnx/onnx_doc_folder/onnx__QLinearConv.html
+
+*/
+
+struct parse_qlinearconv : op_parser<parse_qlinearconv>
+{
+    std::vector<op_desc> operators() const { return {{"QLinearConv"}}; }
+
+    // basic type checking for QLinearConv Operator
+    void check_inputs(const std::vector<instruction_ref>& inp_arg) const
+    {
+        if(inp_arg.size() < 8)
+            MIGRAPHX_THROW("QLINEARCONV: missing inputs");
+
+        const instruction_ref& in_x       = inp_arg[0];
+        const instruction_ref& in_scale_x = inp_arg[1];
+        const instruction_ref& in_w       = inp_arg[3];
+        const instruction_ref& in_scale_w = inp_arg[4];
+        const instruction_ref& in_scale_y = inp_arg[6];
+
+        auto sh_x   = in_x->get_shape();
+        auto sh_w   = in_w->get_shape();
+        auto type_x = sh_x.type();
+        auto type_w = sh_w.type();
+
+        assert(in_x->get_shape().ndim() > 2);
+
+        if(type_x != shape::int8_type and type_x != shape::uint8_type)
+            MIGRAPHX_THROW("QLINEARCONV: unsupported input type");
+        if(type_w != shape::int8_type and type_w != shape::uint8_type)
+            MIGRAPHX_THROW("QLINEARCONV: unsupported weight type");
+        if(in_scale_x->get_shape().type() != shape::float_type)
+            MIGRAPHX_THROW("QLINEARCONV x scale type should be float");
+        if(in_scale_w->get_shape().type() != shape::float_type)
+            MIGRAPHX_THROW("QLINEARCONV: wt scale type should be float");
+        if(in_scale_y->get_shape().type() != shape::float_type)
+            MIGRAPHX_THROW("QLINEARCONV: y scale type should be float");
+        if(inp_arg.size() > 8 and inp_arg[8]->get_shape().type() != shape::int32_type)
+            MIGRAPHX_THROW("QLINEARCONV y bias should be int32");
+    }
+
+    // process all attributes of QLinearConv Operator..
+    value process_attributes(const onnx_parser& parser,
+                             const onnx_parser::node_info& info,
+                             const std::vector<instruction_ref>& args) const
+    {
+        value values;
+
+        const auto& in_x = args[0];
+        const auto& wt   = args[3];
+
+        size_t kdims = in_x->get_shape().ndim() - 2;
+
+        check_padding_mode(info, "QLINEARCONV");
+
+        values["stride"]   = std::vector<int>(kdims, 1);
+        values["dilation"] = std::vector<int>(kdims, 1);
+        values["padding"]  = std::vector<int>(kdims, 0);
+        values["group"]    = 1;
+
+        if(contains(info.attributes, "group"))
+            values["group"] = parser.parse_value(info.attributes.at("group")).template at<int>();
+
+        if(contains(info.attributes, "strides"))
+        {
+            std::vector<int> st;
+            copy(info.attributes.at("strides").ints(), std::back_inserter(st));
+            check_attr_sizes(kdims, st.size(), "QLINEARCONV: inconsistent strides");
+            values["stride"] = st;
+        }
+
+        if(contains(info.attributes, "dilations"))
+        {
+            std::vector<int> dil;
+            copy(info.attributes.at("dilations").ints(), std::back_inserter(dil));
+            check_attr_sizes(kdims, dil.size(), "QLINEARCONV: inconsistent dilations");
+            values["dilation"] = dil;
+        }
+
+        if(contains(info.attributes, "pads"))
+        {
+            std::vector<int> pads;
+            copy(info.attributes.at("pads").ints(), std::back_inserter(pads));
+            check_attr_sizes(kdims, pads.size() / 2, "QLINEARCONV: inconsistent padding");
+            values["padding"] = pads;
+        }
+        else if(contains(info.attributes, "auto_pad"))
+        {
+            auto in_lens = in_x->get_shape().lens();
+            auto wt_lens = wt->get_shape().lens();
+            std::vector<std::size_t> k_lens(wt_lens.begin() + 2, wt_lens.end());
+            std::vector<int64_t> pads = values["padding"].to_vector<std::int64_t>();
+            cal_auto_padding_size(
+                info, values, k_lens, values["dilation"].to_vector<std::size_t>(), in_lens, pads);
+            values["padding"] = pads;
+        }
+
+        recalc_conv_attributes(values, kdims);
+
+        return values;
+    }
+
+    instruction_ref add_bias_to_conv(const instruction_ref bias_arg,
+                                     const instruction_ref conv_instr,
+                                     const onnx_parser::node_info& info) const
+    {
+        auto conv_sh   = conv_instr->get_shape();
+        auto conv_lens = conv_sh.lens();
+        auto conv_type = conv_sh.type();
+
+        auto broadcast_bias = info.add_instruction(
+            migraphx::make_op("broadcast", {{"axis", 1}, {"out_lens", conv_lens}}), bias_arg);
+        auto f_bias =
+            info.add_instruction(make_op("convert", {{"target_type", conv_type}}), broadcast_bias);
+
+        return info.add_instruction(migraphx::make_op("add"), conv_instr, f_bias);
+    };
+
+    instruction_ref parse(const op_desc& /* opd */,
+                          const onnx_parser& parser,
+                          const onnx_parser::node_info& info,
+                          const std::vector<instruction_ref>& args) const
+    {
+        check_inputs(args);
+
+        auto values = process_attributes(parser, info, args);
+
+        // input: quantized x, scale, zero_pt
+        const instruction_ref& in_x         = args[0];
+        const instruction_ref& in_scale_x   = args[1];
+        const instruction_ref& in_zero_pt_x = args[2];
+
+        // input: quantized weights, scale, zero_pt
+        const instruction_ref& in_w         = args[3];
+        const instruction_ref& in_scale_w   = args[4];
+        const instruction_ref& in_zero_pt_w = args[5];
+
+        // for the dequantized output  y: scale & zero_pt
+        const instruction_ref& in_scale_y   = args[6];
+        const instruction_ref& in_zero_pt_y = args[7];
+
+        auto dquant_x = bcast_qdq_instr("dequantizelinear", in_x, in_scale_x, in_zero_pt_x, info);
+
+        auto dquant_w = bcast_qdq_instr("dequantizelinear", in_w, in_scale_w, in_zero_pt_w, info);
+
+        auto conv_op = migraphx::make_op("convolution", values);
+
+        auto conv_x_w = info.add_instruction(conv_op, dquant_x, dquant_w);
+
+        // Biases, if any.. : is an optional argument.
+        if(args.size() > 8)
+            conv_x_w = add_bias_to_conv(args[8], conv_x_w, info);
+
+        auto quant_conv =
+            bcast_qdq_instr("quantizelinear", conv_x_w, in_scale_y, in_zero_pt_y, info);
+        return quant_conv;
+    }
+};
+
+} // namespace onnx
+} // namespace MIGRAPHX_INLINE_NS
+} // namespace migraphx

test/onnx/gen_onnx.py

@@ -5151,6 +5151,116 @@ def qlinearadd_bcast_test():
             [sc_a, zero_pt_a, sc_b, zero_pt_b, sc_c, zero_pt_c])
 
 
+@onnx_test()
+def qlinearconv_test():
+    # https://xadupre.github.io/draft/onnx/onnx_doc_folder/onnx__QLinearConv.html
+    x = helper.make_tensor_value_info('X', TensorProto.UINT8, [1, 1, 7, 7])
+    sc_x = helper.make_tensor('1', TensorProto.FLOAT, [], [0.00369204697])
+    zero_pt_x = helper.make_tensor('2', TensorProto.UINT8, [], [132])
+
+    wt = helper.make_tensor('3', TensorProto.UINT8, [1, 1, 1, 1], [0])
+    sc_wt = helper.make_tensor('4', TensorProto.FLOAT, [], [0.00172794575])
+    zero_pt_wt = helper.make_tensor('5', TensorProto.UINT8, [], [255])
+
+    sc_y = helper.make_tensor('6', TensorProto.FLOAT, [], [0.00162681262])
+    zero_pt_y = helper.make_tensor('7', TensorProto.UINT8, [], [123])
+
+    out = helper.make_tensor_value_info('out', TensorProto.UINT8, [1, 1, 7, 7])
+
+    node = onnx.helper.make_node(
+        'QLinearConv',
+        inputs=['X', '1', '2', '3', '4', '5', '6', '7'],
+        outputs=['out'],
+    )
+    return ([node], [x], [out],
+            [sc_x, zero_pt_x, wt, sc_wt, zero_pt_wt, sc_y, zero_pt_y])
+
+
+@onnx_test()
+def qlinearconv_pad_1_test():
+    # https://xadupre.github.io/draft/onnx/onnx_doc_folder/onnx__Conv.html
+    x = helper.make_tensor_value_info('X', TensorProto.UINT8, [1, 1, 5, 5])
+    sc_x = helper.make_tensor('1', TensorProto.FLOAT, [],
+                              [0.09411764705882353])
+    zero_pt_x = helper.make_tensor('2', TensorProto.UINT8, [], [0])
+
+    wt = helper.make_tensor('3', TensorProto.UINT8, [1, 1, 3, 3],
+                            [1, 1, 1, 1, 1, 1, 1, 1, 1])
+    sc_wt = helper.make_tensor('4', TensorProto.FLOAT, [], [1.0])
+    zero_pt_wt = helper.make_tensor('5', TensorProto.UINT8, [], [0])
+
+    sc_y = helper.make_tensor('6', TensorProto.FLOAT, [], [0.6352941176470588])
+    zero_pt_y = helper.make_tensor('7', TensorProto.UINT8, [], [0])
+
+    out = helper.make_tensor_value_info('out', TensorProto.UINT8, [1, 1, 5, 5])
+
+    node = onnx.helper.make_node(
+        'QLinearConv',
+        inputs=['X', '1', '2', '3', '4', '5', '6', '7'],
+        outputs=['out'],
+        pads=[1, 1, 1, 1],
+    )
+    return ([node], [x], [out],
+            [sc_x, zero_pt_x, wt, sc_wt, zero_pt_wt, sc_y, zero_pt_y])
+
+
+@onnx_test()
+def qlinearconv_pad_0_test():
+    # https://xadupre.github.io/draft/onnx/onnx_doc_folder/onnx__Conv.html
+    x = helper.make_tensor_value_info('X', TensorProto.UINT8, [1, 1, 5, 5])
+    sc_x = helper.make_tensor('1', TensorProto.FLOAT, [],
+                              [0.09411764705882353])
+    zero_pt_x = helper.make_tensor('2', TensorProto.UINT8, [], [0])
+
+    wt = helper.make_tensor('3', TensorProto.UINT8, [1, 1, 3, 3],
+                            [1, 1, 1, 1, 1, 1, 1, 1, 1])
+    sc_wt = helper.make_tensor('4', TensorProto.FLOAT, [], [1.0])
+    zero_pt_wt = helper.make_tensor('5', TensorProto.UINT8, [], [0])
+
+    sc_y = helper.make_tensor('6', TensorProto.FLOAT, [], [0.6352941176470588])
+    zero_pt_y = helper.make_tensor('7', TensorProto.INT8, [], [-128])
+
+    out = helper.make_tensor_value_info('out', TensorProto.INT8, [1, 1, 3, 3])
+
+    node = onnx.helper.make_node(
+        'QLinearConv',
+        inputs=['X', '1', '2', '3', '4', '5', '6', '7'],
+        outputs=['out'],
+        pads=[0, 0, 0, 0],
+    )
+    return ([node], [x], [out],
+            [sc_x, zero_pt_x, wt, sc_wt, zero_pt_wt, sc_y, zero_pt_y])
+
+
+@onnx_test()
+def qlinearconv_scale_1D_test():
+    # https://xadupre.github.io/draft/onnx/onnx_doc_folder/onnx__Conv.html
+    x = helper.make_tensor_value_info('X', TensorProto.UINT8, [1, 1, 5, 5])
+    sc_x = helper.make_tensor('1', TensorProto.FLOAT, [],
+                              [0.09411764705882353])
+    zero_pt_x = helper.make_tensor('2', TensorProto.UINT8, [], [0])
+
+    wt = helper.make_tensor(
+        '3', TensorProto.UINT8, [2, 1, 3, 3],
+        [1, 1, 1, 1, 1, 1, 1, 1, 1, 2, 2, 2, 2, 2, 2, 2, 2, 2])
+    sc_wt = helper.make_tensor('4', TensorProto.FLOAT, [2], [1.0, 0.5])
+    zero_pt_wt = helper.make_tensor('5', TensorProto.UINT8, [2], [0, 0])
+
+    sc_y = helper.make_tensor('6', TensorProto.FLOAT, [], [0.6352941176470588])
+    zero_pt_y = helper.make_tensor('7', TensorProto.INT8, [], [-128])
+
+    out = helper.make_tensor_value_info('out', TensorProto.INT8, [1, 2, 3, 3])
+
+    node = onnx.helper.make_node(
+        'QLinearConv',
+        inputs=['X', '1', '2', '3', '4', '5', '6', '7'],
+        outputs=['out'],
+        pads=[0, 0, 0, 0],
+    )
+    return ([node], [x], [out],
+            [sc_x, zero_pt_x, wt, sc_wt, zero_pt_wt, sc_y, zero_pt_y])
+
+
 @onnx_test()
 def quantizelinear_test():
     arg0 = helper.make_tensor_value_info('0', TensorProto.FLOAT, [5])