Merge branch 'develop' into unet

src/CMakeLists.txt

@@ -53,6 +53,7 @@ add_library(migraphx
     remap.cpp
     rewrite_batchnorm.cpp
     rewrite_pooling.cpp
+    rewrite_quantization.cpp
     rewrite_rnn.cpp
     schedule.cpp
     serialize.cpp
@@ -94,6 +95,7 @@ register_migraphx_ops(
     cosh
     cos
     deconvolution
+    dequantizelinear
     div
     dot
     elu
@@ -132,6 +134,7 @@ register_migraphx_ops(
     prelu
     quant_convolution
     quant_dot
+    quantizelinear
     recip
     reduce_max
     reduce_mean

src/include/migraphx/op/dequantizelinear.hpp

+#ifndef MIGRAPHX_GUARD_OPERATORS_DEQUANTIZE_LINEAR_HPP
+#define MIGRAPHX_GUARD_OPERATORS_DEQUANTIZE_LINEAR_HPP
+
+#include <array>
+#include <migraphx/op/common.hpp>
+#include <migraphx/operation.hpp>
+#include <migraphx/check_shapes.hpp>
+#include <migraphx/stringutils.hpp>
+#include <migraphx/streamutils.hpp>
+#include <migraphx/literal.hpp>
+#include <migraphx/config.hpp>
+#include <migraphx/par_for.hpp>
+#include <migraphx/value.hpp>
+#include <migraphx/op/normalize_attribute.hpp>
+#include <migraphx/tune_axis.hpp>
+#include <cmath>
+#include <utility>
+
+namespace migraphx {
+inline namespace MIGRAPHX_INLINE_NS {
+namespace op {
+
+struct dequantizelinear
+{
+    std::string name() const { return "dequantizelinear"; }
+    shape compute_shape(std::vector<shape> inputs) const
+    {
+        return {shape::float_type, inputs[0].lens(), inputs[0].strides()};
+    }
+
+    argument compute(const shape& output_shape, std::vector<argument> args) const
+    {
+        auto x       = args.at(0);
+        auto x_scale = args.at(1);
+        std::vector<int8_t> zeros(output_shape.elements(), 0);
+        argument x_zero_point{{x.get_shape().type(), output_shape.lens()}, zeros.data()};
+        if(args.size() == 3)
+        {
+            x_zero_point = args.at(2);
+        }
+
+        argument result{output_shape};
+        visit_all(x, x_zero_point)([&](auto input, auto zero_pts) {
+            visit_all(result, x_scale)([&](auto output, auto scales) {
+                par_for(output_shape.elements(), [&](auto i) {
+                    output[i] = static_cast<double>(static_cast<int64_t>(input[i]) -
+                                                    static_cast<int64_t>(zero_pts[i])) *
+                                scales[i];
+                });
+            });
+        });
+
+        return result;
+    }
+};
+
+} // namespace op
+} // namespace MIGRAPHX_INLINE_NS
+} // namespace migraphx
+
+#endif

src/include/migraphx/op/quantizelinear.hpp

+#ifndef MIGRAPHX_GUARD_OPERATORS_QUANTIZE_LINEAR_HPP
+#define MIGRAPHX_GUARD_OPERATORS_QUANTIZE_LINEAR_HPP
+
+#include <array>
+#include <migraphx/op/common.hpp>
+#include <migraphx/operation.hpp>
+#include <migraphx/check_shapes.hpp>
+#include <migraphx/stringutils.hpp>
+#include <migraphx/streamutils.hpp>
+#include <migraphx/literal.hpp>
+#include <migraphx/config.hpp>
+#include <migraphx/par_for.hpp>
+#include <migraphx/value.hpp>
+#include <migraphx/op/normalize_attribute.hpp>
+#include <migraphx/tune_axis.hpp>
+#include <cmath>
+#include <utility>
+
+namespace migraphx {
+inline namespace MIGRAPHX_INLINE_NS {
+namespace op {
+
+struct quantizelinear
+{
+    std::string name() const { return "quantizelinear"; }
+    shape compute_shape(std::vector<shape> inputs) const
+    {
+        if(inputs.size() == 3)
+        {
+            return {inputs[2].type(), inputs[0].lens(), inputs[0].strides()};
+        }
+        return {shape::uint8_type, inputs[0].lens(), inputs[0].strides()};
+    }
+
+    argument compute(const shape& output_shape, std::vector<argument> args) const
+    {
+        auto x       = args.at(0);
+        auto y_scale = args.at(1);
+        std::vector<int8_t> zeros(output_shape.elements(), 0);
+        argument y_zero_point{output_shape, zeros.data()};
+        if(args.size() == 3)
+        {
+            y_zero_point = args.at(2);
+        }
+
+        argument result{output_shape};
+        visit_all(result, y_zero_point)([&](auto output, auto zero_pts) {
+            x.visit([&](auto input) {
+                y_scale.visit([&](auto scales) {
+                    using quant_type = typename decltype(output)::value_type;
+                    auto min_value   = std::numeric_limits<quant_type>::min();
+                    auto max_value   = std::numeric_limits<quant_type>::max();
+                    par_for(output_shape.elements(), [&](auto i) {
+                        int64_t quantized = static_cast<int64_t>(std::round(input[i] / scales[i])) +
+                                            static_cast<int64_t>(zero_pts[i]);
+                        output[i] = std::max(static_cast<int64_t>(min_value),
+                                             std::min(static_cast<int64_t>(max_value), quantized));
+                    });
+                });
+            });
+        });
+
+        return result;
+    }
+};
+
+} // namespace op
+} // namespace MIGRAPHX_INLINE_NS
+} // namespace migraphx
+
+#endif

src/include/migraphx/rewrite_quantization.hpp

+#ifndef MIGRAPHX_GUARD_RTGLIB_REWRITE_QUANTIZATION_HPP
+#define MIGRAPHX_GUARD_RTGLIB_REWRITE_QUANTIZATION_HPP
+
+#include <string>
+#include <migraphx/config.hpp>
+
+namespace migraphx {
+inline namespace MIGRAPHX_INLINE_NS {
+
+struct module;
+
+/**
+ * Rewrite quantization ops to equivalent operators
+ */
+struct rewrite_quantization
+{
+    std::string name() const { return "rewrite_quantization"; }
+    void apply(module& m) const;
+};
+
+} // namespace MIGRAPHX_INLINE_NS
+} // namespace migraphx
+
+#endif

src/onnx/parse_dequantizelinear.cpp

@@ -15,46 +15,49 @@ struct parse_dequantizelinear : op_parser<parse_dequantizelinear>
     instruction_ref parse(const op_desc& opd,
                           const onnx_parser& /*parser*/,
                           const onnx_parser::node_info& info,
-                          std::vector<instruction_ref> args) const
+                          const std::vector<instruction_ref>& args) const
     {
         int axis = 1;
         if(contains(info.attributes, "axis"))
             axis = info.attributes.at("axis").i();
 
         auto input_lens = args[0]->get_shape().lens();
-        int n_dim       = static_cast<int>(input_lens.size());
+        auto n_dim      = input_lens.size();
 
-        auto sub_zero_point = args[0];
+        instruction_ref x_scale;
+        if(args[1]->get_shape().elements() != 1)
+        {
+            auto tuned_axis = tune_axis(n_dim, axis, opd.op_name);
+            x_scale         = info.add_instruction(
+                make_op("broadcast", {{"axis", tuned_axis}, {"dims", input_lens}}), args[1]);
+        }
+        else
+        {
+            x_scale = info.add_instruction(make_op("multibroadcast", {{"output_lens", input_lens}}),
+                                           args[1]);
+        }
 
         if(args.size() == 3)
         {
-            auto zero_point = args[2];
-            if(not(zero_point->get_shape().elements() == 1))
+            auto x_zero_point = args[2];
+            if(x_zero_point->get_shape().elements() != 1)
             {
-                axis       = tune_axis(n_dim, axis, opd.op_name);
-                zero_point = info.add_instruction(
-                    make_op("broadcast", {{"axis", axis}, {"dims", input_lens}}), zero_point);
+                auto tuned_axis = tune_axis(n_dim, axis, opd.op_name);
+                x_zero_point    = info.add_instruction(
+                    make_op("broadcast", {{"axis", tuned_axis}, {"dims", input_lens}}),
+                    x_zero_point);
             }
-
-            auto zero_point_int32 = info.add_instruction(
-                make_op("convert", {{"target_type", shape::int32_type}}), zero_point);
-            auto sub_zero_point_int32 = info.add_instruction(
-                make_op("convert", {{"target_type", shape::int32_type}}), sub_zero_point);
-            sub_zero_point =
-                info.add_broadcastable_binary_op("sub", sub_zero_point_int32, zero_point_int32);
+            else
+            {
+                x_zero_point = info.add_instruction(
+                    make_op("multibroadcast", {{"output_lens", input_lens}}), x_zero_point);
             }
 
-        auto dequant_input = info.add_instruction(
-            make_op("convert", {{"target_type", shape::float_type}}), sub_zero_point);
-
-        auto scale = args[1];
-        if(not(scale->get_shape().elements() == 1))
-        {
-            axis  = tune_axis(n_dim, axis, opd.op_name);
-            scale = info.add_instruction(
-                make_op("broadcast", {{"axis", axis}, {"dims", input_lens}}), scale);
+            return info.add_instruction(
+                make_op("dequantizelinear"), args[0], x_scale, x_zero_point);
         }
-        return info.add_broadcastable_binary_op("mul", dequant_input, scale);
+
+        return info.add_instruction(make_op("dequantizelinear"), args[0], x_scale);
     }
 };
 

src/onnx/parse_gemm.cpp

@@ -42,13 +42,23 @@ struct parse_gemm : op_parser<parse_gemm>
         // swap the last two elements
         std::swap(*perm.rbegin(), *(perm.rbegin() + 1));
 
-        auto l1 = (transa) ? info.add_instruction(make_op("transpose", {{"dims", perm}}), args[0])
-                           : args[0];
+        auto l1 = args[0];
+
+        if(alpha != 1.0f)
+        {
+            auto alpha_literal = info.add_literal(alpha);
+            auto alpha_l1      = info.add_broadcastable_binary_op("mul", alpha_literal, l1);
+            l1 = info.add_instruction(make_op("convert", {{"target_type", l1->get_shape().type()}}),
+                                      alpha_l1);
+        }
+
+        l1      = (transa) ? info.add_instruction(make_op("transpose", {{"dims", perm}}), l1) : l1;
         auto l2 = (transb) ? info.add_instruction(make_op("transpose", {{"dims", perm}}), args[1])
                            : args[1];
+
         if(args.size() == 3)
         {
-            if(beta != 0.f && args[2]->get_shape().elements() > 0)
+            if(beta != 0.0f && args[2]->get_shape().elements() > 0)
             {
                 auto out_lens   = l1->get_shape().lens();
                 out_lens.back() = l2->get_shape().lens().back();
@@ -59,12 +69,17 @@ struct parse_gemm : op_parser<parse_gemm>
                     l3 = info.add_instruction(
                         make_op("multibroadcast", {{"output_lens", out_lens}}), args[2]);
                 }
+                auto beta_literal   = info.add_literal(beta);
+                auto beta_broadcast = info.add_instruction(
+                    make_op("multibroadcast", {{"output_lens", out_lens}}), beta_literal);
+                l3 = info.add_instruction(make_op("mul"), l3, beta_broadcast);
+
                 return info.add_instruction(
-                    make_op("dot", {{"alpha", alpha}, {"beta", beta}}), l1, l2, l3);
+                    make_op("dot", {{"alpha", 1.0f}, {"beta", 1.0f}}), l1, l2, l3);
             }
         }
 
-        return info.add_instruction(make_op("dot", {{"alpha", alpha}, {"beta", beta}}), l1, l2);
+        return info.add_instruction(make_op("dot", {{"alpha", 1.0f}, {"beta", 1.0f}}), l1, l2);
     }
 };
 

src/onnx/parse_quantizelinear.cpp

@@ -12,83 +12,51 @@ struct parse_quantizelinear : op_parser<parse_quantizelinear>
 {
     std::vector<op_desc> operators() const { return {{"QuantizeLinear"}}; }
 
-    // y = saturate(round(x / y_scale) + zero_point)
     instruction_ref parse(const op_desc& opd,
                           const onnx_parser& /*parser*/,
                           const onnx_parser::node_info& info,
-                          std::vector<instruction_ref> args) const
+                          const std::vector<instruction_ref>& args) const
     {
-        auto quant_type = shape::uint8_type;
-        int nargs       = args.size();
-
-        int max_quant = 255;
-        int min_quant = 0;
-
-        if(nargs == 3)
-            quant_type = args[2]->get_shape().type();
-
-        if(quant_type == shape::int8_type)
-        {
-            max_quant = 127;
-            min_quant = -128;
-        }
-
-        auto max_arg = info.add_literal(max_quant);
-        auto min_arg = info.add_literal(min_quant);
-
         int axis = 1;
-
         if(contains(info.attributes, "axis"))
             axis = info.attributes.at("axis").i();
 
         auto input_lens = args[0]->get_shape().lens();
-        int n_dim       = static_cast<int>(input_lens.size());
+        auto n_dim      = input_lens.size();
 
-        auto scale = args[1];
-        if(not(scale->get_shape().elements() == 1))
+        instruction_ref y_scale;
+        if(args[1]->get_shape().elements() != 1)
         {
-            axis  = tune_axis(n_dim, axis, opd.op_name);
-            scale = info.add_instruction(
-                make_op("broadcast", {{"axis", axis}, {"dims", input_lens}}), scale);
+            auto tuned_axis = tune_axis(n_dim, axis, opd.op_name);
+            y_scale         = info.add_instruction(
+                make_op("broadcast", {{"axis", tuned_axis}, {"dims", input_lens}}), args[1]);
         }
-
-        auto div            = info.add_broadcastable_binary_op("div", args[0], scale);
-        auto div_round      = info.add_instruction(make_op("round"), div);
-        auto add_zero_point = div_round;
-
-        if(nargs == 3)
+        else
         {
-            auto zero_point = args[2];
-            if(not(zero_point->get_shape().elements() == 1))
-            {
-                axis       = tune_axis(n_dim, axis, opd.op_name);
-                zero_point = info.add_instruction(
-                    make_op("broadcast", {{"axis", axis}, {"dims", input_lens}}), zero_point);
-            }
-
-            zero_point = info.add_instruction(
-                make_op("convert", {{"target_type", shape::int32_type}}), zero_point);
-            add_zero_point = info.add_instruction(
-                make_op("convert", {{"target_type", shape::int32_type}}), add_zero_point);
-            add_zero_point = info.add_broadcastable_binary_op("add", add_zero_point, zero_point);
+            y_scale = info.add_instruction(make_op("multibroadcast", {{"output_lens", input_lens}}),
+                                           args[1]);
         }
 
-        auto s           = add_zero_point->get_shape();
-        const auto& lens = s.lens();
-        std::vector<int64_t> out_lens(lens.begin(), lens.end());
-        if(min_arg->get_shape() != s)
+        if(args.size() == 3)
+        {
+            auto y_zero_point = args[2];
+            if(y_zero_point->get_shape().elements() != 1)
             {
-            min_arg = info.add_instruction(make_op("multibroadcast", {{"output_lens", out_lens}}),
-                                           min_arg);
+                auto tuned_axis = tune_axis(n_dim, axis, opd.op_name);
+                y_zero_point    = info.add_instruction(
+                    make_op("broadcast", {{"axis", tuned_axis}, {"dims", input_lens}}),
+                    y_zero_point);
             }
-        if(max_arg->get_shape() != s)
+            else
             {
-            max_arg = info.add_instruction(make_op("multibroadcast", {{"output_lens", out_lens}}),
-                                           max_arg);
+                y_zero_point = info.add_instruction(
+                    make_op("multibroadcast", {{"output_lens", input_lens}}), y_zero_point);
+            }
+
+            return info.add_instruction(make_op("quantizelinear"), args[0], y_scale, y_zero_point);
         }
 
-        auto saturated = info.add_instruction(make_op("clip"), add_zero_point, min_arg, max_arg);
-        return info.add_instruction(make_op("convert", {{"target_type", quant_type}}), saturated);
+        return info.add_instruction(make_op("quantizelinear"), args[0], y_scale);
     }
 };
 

src/rewrite_quantization.cpp

+#include <migraphx/rewrite_quantization.hpp>
+#include <migraphx/instruction.hpp>
+#include <migraphx/iterator_for.hpp>
+#include <migraphx/make_op.hpp>
+#include <migraphx/tune_axis.hpp>
+#include <migraphx/program.hpp>
+#include <migraphx/shape.hpp>
+
+namespace migraphx {
+inline namespace MIGRAPHX_INLINE_NS {
+
+void apply_quantizelinear(module& m, instruction_ref ins)
+{
+    assert(ins->name() == "quantizelinear");
+    auto x       = ins->inputs()[0];
+    auto y_scale = ins->inputs()[1];
+
+    if(x->get_shape().type() != y_scale->get_shape().type())
+    {
+        x = m.insert_instruction(ins, make_op("convert", {{"target_type", shape::float_type}}), x);
+    }
+    auto div            = m.insert_instruction(ins, make_op("div"), x, y_scale);
+    auto add_zero_point = m.insert_instruction(ins, make_op("round"), div);
+
+    if(ins->inputs().size() == 3)
+    {
+        auto zero_point = m.insert_instruction(
+            ins, make_op("convert", {{"target_type", shape::float_type}}), ins->inputs()[2]);
+        add_zero_point = m.insert_instruction(ins, make_op("add"), add_zero_point, zero_point);
+    }
+
+    int64_t max_quant = 0;
+    int64_t min_quant = 0;
+    ins->get_shape().visit_type([&](auto qt) {
+        max_quant = qt.max();
+        min_quant = qt.min();
+    });
+    auto s = add_zero_point->get_shape();
+    std::vector<int> min_data(s.elements(), min_quant);
+    std::vector<int> max_data(s.elements(), max_quant);
+    auto min_arg = m.add_literal(literal(s, min_data));
+    auto max_arg = m.add_literal(literal(s, max_data));
+
+    auto saturate = m.insert_instruction(ins, make_op("clip"), add_zero_point, min_arg, max_arg);
+    m.replace_instruction(
+        ins, make_op("convert", {{"target_type", ins->get_shape().type()}}), saturate);
+}
+
+void apply_dequantizelinear(module& m, instruction_ref ins)
+{
+    assert(ins->name() == "dequantizelinear");
+    auto x = m.insert_instruction(
+        ins, make_op("convert", {{"target_type", shape::float_type}}), ins->inputs()[0]);
+    auto x_scale = ins->inputs()[1];
+
+    if(ins->inputs().size() == 3)
+    {
+        auto x_zero_point = m.insert_instruction(
+            ins, make_op("convert", {{"target_type", shape::float_type}}), ins->inputs()[2]);
+        x = m.insert_instruction(ins, make_op("sub"), x, x_zero_point);
+    }
+
+    m.replace_instruction(ins, make_op("mul"), x, x_scale);
+}
+
+void rewrite_quantization::apply(module& m) const
+{
+    for(auto ins : iterator_for(m))
+    {
+        if(ins->name() == "quantizelinear")
+        {
+            apply_quantizelinear(m, ins);
+        }
+
+        else if(ins->name() == "dequantizelinear")
+        {
+            apply_dequantizelinear(m, ins);
+        }
+    }
+}
+
+} // namespace MIGRAPHX_INLINE_NS
+} // namespace migraphx

src/targets/cpu/target.cpp

@@ -17,6 +17,7 @@
 #include <migraphx/remap.hpp>
 #include <migraphx/rewrite_batchnorm.hpp>
 #include <migraphx/rewrite_pooling.hpp>
+#include <migraphx/rewrite_quantization.hpp>
 #include <migraphx/rewrite_rnn.hpp>
 #include <migraphx/schedule.hpp>
 #include <migraphx/memory_coloring.hpp>
@@ -46,6 +47,8 @@ std::vector<pass> target::get_passes(migraphx::context& gctx, const compile_opti
     std::set<shape::type_t> unsupported_types(shape::types().begin(), shape::types().end());
     unsupported_types.erase(shape::type_t::float_type);
     return {normalize_ops{},
+            rewrite_quantization{},
+            dead_code_elimination{},
             eliminate_data_type{unsupported_types, shape::type_t::float_type},
             dead_code_elimination{},
             decompose{},

src/targets/gpu/target.cpp

@@ -20,6 +20,7 @@
 #include <migraphx/remap.hpp>
 #include <migraphx/rewrite_batchnorm.hpp>
 #include <migraphx/rewrite_pooling.hpp>
+#include <migraphx/rewrite_quantization.hpp>
 #include <migraphx/rewrite_rnn.hpp>
 #include <migraphx/schedule.hpp>
 #include <migraphx/simplify_algebra.hpp>
@@ -59,6 +60,8 @@ std::vector<pass> target::get_passes(migraphx::context& gctx, const compile_opti
         normalize_ops{},
         decompose{},
         dead_code_elimination{},
+        rewrite_quantization{},
+        dead_code_elimination{},
         eliminate_data_type{unsupported_types, shape::type_t::float_type},
         simplify_reshapes{},
         eliminate_identity{},

test/include/test.hpp

@@ -6,6 +6,7 @@
 #include <functional>
 #include <iostream>
 #include <sstream>
+#include <type_traits>
 #include <unordered_map>
 #include <vector>
 
@@ -101,7 +102,9 @@ inline Stream& operator<<(Stream& s, std::nullptr_t)
     return s;
 }
 
-template <class Stream, class Range>
+template <class Stream,
+          class Range,
+          class = typename std::enable_if<not std::is_convertible<Range, std::string>{}>::type>
 inline auto operator<<(Stream& s, const Range& v) -> decltype(stream_range(s, v.begin(), v.end()))
 {
     s << "{ ";

test/onnx/dequantizelinear_axis_test.onnx

@@ -23,4 +23,4 @@
 

 

 
-
B
+
B
\ No newline at end of file

test/onnx/dequantizelinear_neg_axis_test.onnx

@@ -23,4 +23,4 @@
 

 

 
-
B
+
B
\ No newline at end of file

test/onnx/dequantizelinear_test.onnx

-dequantizelinear_test:
- 
+dequantizelinear_test:k
+
 
0
-
1
-
2out"DequantizeLineardequantizelinear_testZ
+
1out"DequantizeLineardequantizelinear_testZ
 
0
 
 
@@ -10,12 +9,8 @@
 
1
 
 

-
Z
-
2
-
-
 
b
 out
 
 

-B
+B
\ No newline at end of file

test/onnx/dequantizelinear_zero_point_test.onnx

+ dequantizelinear_zero_point_test:

+ 
+
0
+
1
+
2out"DequantizeLinear dequantizelinear_zero_point_testZ
+
0
+
+
+Z
+
1
+
+

+
Z
+
2
+
+
+
b
+out
+
+

+B
\ No newline at end of file

test/onnx/gen_onnx.py

@@ -1021,6 +1021,21 @@ def deconv_stride_test():
 
 @onnx_test
 def dequantizelinear_test():
+    arg0 = helper.make_tensor_value_info('0', TensorProto.INT8, [5])
+    arg1 = helper.make_tensor_value_info('1', TensorProto.FLOAT, [1])
+    arg_out = helper.make_tensor_value_info('out', TensorProto.FLOAT, [5])
+
+    node = onnx.helper.make_node(
+        'DequantizeLinear',
+        inputs=['0', '1'],
+        outputs=['out'],
+    )
+
+    return ([node], [arg0, arg1], [arg_out])
+
+
+@onnx_test
+def dequantizelinear_zero_point_test():
     arg0 = helper.make_tensor_value_info('0', TensorProto.INT8, [5])
     arg1 = helper.make_tensor_value_info('1', TensorProto.FLOAT, [1])
     arg2 = helper.make_tensor_value_info('2', TensorProto.INT8, [1])
@@ -2751,6 +2766,36 @@ def prelu_brcst_test():
 
 @onnx_test
 def quantizelinear_test():
+    arg0 = helper.make_tensor_value_info('0', TensorProto.FLOAT, [5])
+    arg1 = helper.make_tensor_value_info('1', TensorProto.FLOAT, [1])
+    arg_out = helper.make_tensor_value_info('out', TensorProto.INT8, [5])
+
+    node = onnx.helper.make_node(
+        'QuantizeLinear',
+        inputs=['0', '1'],
+        outputs=['out'],
+    )
+
+    return ([node], [arg0, arg1], [arg_out])
+
+
+@onnx_test
+def quantizelinear_int32_test():
+    arg0 = helper.make_tensor_value_info('0', TensorProto.INT32, [5])
+    arg1 = helper.make_tensor_value_info('1', TensorProto.FLOAT, [1])
+    arg_out = helper.make_tensor_value_info('out', TensorProto.INT8, [5])
+
+    node = onnx.helper.make_node(
+        'QuantizeLinear',
+        inputs=['0', '1'],
+        outputs=['out'],
+    )
+
+    return ([node], [arg0, arg1], [arg_out])
+
+
+@onnx_test
+def quantizelinear_zero_point_test():
     arg0 = helper.make_tensor_value_info('0', TensorProto.FLOAT, [5])
     arg1 = helper.make_tensor_value_info('1', TensorProto.FLOAT, [1])
     arg2 = helper.make_tensor_value_info('2', TensorProto.INT8, [1])

test/onnx/onnx_test.cpp

 #include <iostream>
 #include <fstream>
 #include <vector>
+#include <migraphx/common.hpp>
 #include <migraphx/literal.hpp>
-#include <migraphx/operators.hpp>
 #include <migraphx/program.hpp>
 #include <migraphx/instruction.hpp>
 #include <migraphx/instruction_ref.hpp>
 #include <migraphx/pass_manager.hpp>
 #include <migraphx/dead_code_elimination.hpp>
+#include <migraphx/rewrite_quantization.hpp>
 #include <migraphx/eliminate_identity.hpp>
 #include <migraphx/onnx.hpp>
 #include <migraphx/make_op.hpp>
+#include <migraphx/op/convolution.hpp>
+#include <migraphx/op/pad.hpp>
+#include <migraphx/op/pooling.hpp>
+#include <migraphx/op/lrn.hpp>
+#include <migraphx/op/reshape.hpp>
+#include <migraphx/op/unknown.hpp>
 
 #include <migraphx/serialize.hpp>
 
 #include "test.hpp"
 
-migraphx::program optimize_onnx(const std::string& name, bool eliminate_deadcode = false)
+migraphx::program optimize_onnx(const std::string& name, bool run_passes = false)
 {
     migraphx::onnx_options options;
     options.skip_unknown_operators = true;
     auto prog                      = migraphx::parse_onnx(name, options);
     auto* mm                       = prog.get_main_module();
-    if(eliminate_deadcode)
-        migraphx::run_passes(*mm, {migraphx::dead_code_elimination{}});
+    if(run_passes)
+        migraphx::run_passes(*mm,
+                             {migraphx::rewrite_quantization{}, migraphx::dead_code_elimination{}});
 
     // remove the last identity instruction
     auto last_ins = std::prev(mm->end());
@@ -914,29 +922,42 @@ TEST_CASE(dequantizelinear_test)
     auto* mm = p.get_main_module();
     auto l0  = mm->add_parameter("0", {migraphx::shape::int8_type, {5}});
     auto l1  = mm->add_parameter("1", {migraphx::shape::float_type, {1}});
-    auto l2  = mm->add_parameter("2", {migraphx::shape::int8_type, {1}});
     auto l1_mbcast =
         mm->add_instruction(migraphx::make_op("multibroadcast", {{"output_lens", {5}}}), l1);
-    l2 = mm->add_instruction(
+    auto dequant = mm->add_instruction(
         migraphx::make_op("convert",
-                          {{"target_type", migraphx::to_value(migraphx::shape::int32_type)}}),
-        l2);
+                          {{"target_type", migraphx::to_value(migraphx::shape::float_type)}}),
+        l0);
+    mm->add_instruction(migraphx::make_op("mul"), dequant, l1_mbcast);
+
+    auto prog = optimize_onnx("dequantizelinear_test.onnx", true);
+    EXPECT(p.sort() == prog.sort());
+}
+
+TEST_CASE(dequantizelinear_zero_point_test)
+{
+    migraphx::program p;
+    auto* mm = p.get_main_module();
+    auto l0  = mm->add_parameter("0", {migraphx::shape::int8_type, {5}});
+    auto l1  = mm->add_parameter("1", {migraphx::shape::float_type, {1}});
+    auto l2  = mm->add_parameter("2", {migraphx::shape::int8_type, {1}});
+    auto l1_mbcast =
+        mm->add_instruction(migraphx::make_op("multibroadcast", {{"output_lens", {5}}}), l1);
     auto l2_mbcast =
         mm->add_instruction(migraphx::make_op("multibroadcast", {{"output_lens", {5}}}), l2);
+    l2_mbcast = mm->add_instruction(
+        migraphx::make_op("convert",
+                          {{"target_type", migraphx::to_value(migraphx::shape::float_type)}}),
+        l2_mbcast);
     l0 = mm->add_instruction(
         migraphx::make_op("convert",
-                          {{"target_type", migraphx::to_value(migraphx::shape::int32_type)}}),
+                          {{"target_type", migraphx::to_value(migraphx::shape::float_type)}}),
         l0);
 
     auto sub = mm->add_instruction(migraphx::make_op("sub"), l0, l2_mbcast);
-    auto dequant = mm->add_instruction(
-        migraphx::make_op("convert",
-                          {{"target_type", migraphx::to_value(migraphx::shape::float_type)}}),
-        sub);
-
-    mm->add_instruction(migraphx::make_op("mul"), dequant, l1_mbcast);
+    mm->add_instruction(migraphx::make_op("mul"), sub, l1_mbcast);
 
-    auto prog = optimize_onnx("dequantizelinear_test.onnx");
+    auto prog = optimize_onnx("dequantizelinear_zero_point_test.onnx", true);
     EXPECT(p.sort() == prog.sort());
 }
 
@@ -955,19 +976,15 @@ migraphx::program make_dequantizelinear_axis_prog()
         migraphx::make_op("broadcast", {{"axis", axis}, {"dims", input_lens}}), l2);
     l2_bcast = mm->add_instruction(
         migraphx::make_op("convert",
-                          {{"target_type", migraphx::to_value(migraphx::shape::int32_type)}}),
+                          {{"target_type", migraphx::to_value(migraphx::shape::float_type)}}),
         l2_bcast);
     l0 = mm->add_instruction(
         migraphx::make_op("convert",
-                          {{"target_type", migraphx::to_value(migraphx::shape::int32_type)}}),
+                          {{"target_type", migraphx::to_value(migraphx::shape::float_type)}}),
         l0);
     auto sub = mm->add_instruction(migraphx::make_op("sub"), l0, l2_bcast);
-    auto dequant = mm->add_instruction(
-        migraphx::make_op("convert",
-                          {{"target_type", migraphx::to_value(migraphx::shape::float_type)}}),
-        sub);
 
-    mm->add_instruction(migraphx::make_op("mul"), dequant, l1_bcast);
+    mm->add_instruction(migraphx::make_op("mul"), sub, l1_bcast);
     return p;
 }
 
@@ -975,7 +992,7 @@ TEST_CASE(dequantizelinear_axis_test)
 {
     migraphx::program p = make_dequantizelinear_axis_prog();
 
-    auto prog = optimize_onnx("dequantizelinear_axis_test.onnx");
+    auto prog = optimize_onnx("dequantizelinear_axis_test.onnx", true);
     EXPECT(p.sort() == prog.sort());
 }
 
@@ -983,7 +1000,7 @@ TEST_CASE(dequantizelinear_neg_axis_test)
 {
     migraphx::program p = make_dequantizelinear_axis_prog();
 
-    auto prog = optimize_onnx("dequantizelinear_neg_axis_test.onnx");
+    auto prog = optimize_onnx("dequantizelinear_neg_axis_test.onnx", true);
     EXPECT(p.sort() == prog.sort());
 }
 
@@ -1256,15 +1273,24 @@ TEST_CASE(gemm_test)
     auto l0    = mm->add_parameter("0", migraphx::shape{migraphx::shape::float_type, {5, 7}});
     auto l1    = mm->add_parameter("1", migraphx::shape{migraphx::shape::float_type, {11, 5}});
     auto l2    = mm->add_parameter("2", migraphx::shape{migraphx::shape::float_type});
-    auto t0  = mm->add_instruction(migraphx::make_op("transpose", {{"dims", {1, 0}}}), l0);
-    auto t1  = mm->add_instruction(migraphx::make_op("transpose", {{"dims", {1, 0}}}), l1);
-    auto bl2 =
-        mm->add_instruction(migraphx::make_op("multibroadcast", {{"output_lens", {7, 11}}}), l2);
     auto alpha = 2.f;
     auto beta  = 2.0f;
-    mm->add_instruction(migraphx::make_op("dot", {{"alpha", alpha}, {"beta", beta}}), t0, t1, bl2);
-    auto prog = optimize_onnx("gemm_test.onnx");
+    auto a_l   = mm->add_literal(alpha);
+    auto t_a   = add_common_op(*mm, migraphx::make_op("mul"), {a_l, l0});
+    t_a        = mm->add_instruction(
+        migraphx::make_op("convert", {{"target_type", l1->get_shape().type()}}), t_a);
+    t_a     = mm->add_instruction(migraphx::make_op("transpose", {{"dims", {1, 0}}}), t_a);
+    auto t1 = mm->add_instruction(migraphx::make_op("transpose", {{"dims", {1, 0}}}), l1);
 
+    auto b_l = mm->add_literal(beta);
+    auto l2_b =
+        mm->add_instruction(migraphx::make_op("multibroadcast", {{"output_lens", {7, 11}}}), l2);
+    auto b_b = mm->add_instruction(
+        migraphx::make_op("multibroadcast", {{"output_lens", l2_b->get_shape().lens()}}), b_l);
+    auto l2_bb = mm->add_instruction(migraphx::make_op("mul"), l2_b, b_b);
+    mm->add_instruction(
+        migraphx::make_op("dot", {{"alpha", 1.0f}, {"beta", 1.0f}}), t_a, t1, l2_bb);
+    auto prog = optimize_onnx("gemm_test.onnx");
     EXPECT(p == prog);
 }
 
@@ -1275,10 +1301,21 @@ TEST_CASE(gemm_ex_test)
     auto l0    = mm->add_parameter("1", migraphx::shape{migraphx::shape::float_type, {1, 1, 8, 6}});
     auto l1    = mm->add_parameter("2", migraphx::shape{migraphx::shape::float_type, {1, 1, 8, 7}});
     auto l2    = mm->add_parameter("3", migraphx::shape{migraphx::shape::float_type, {1, 1, 6, 7}});
-    auto t0    = mm->add_instruction(migraphx::make_op("transpose", {{"dims", {0, 1, 3, 2}}}), l0);
     auto alpha = 0.5f;
     auto beta  = 0.8f;
-    mm->add_instruction(migraphx::make_op("dot", {{"alpha", alpha}, {"beta", beta}}), t0, l1, l2);
+    auto a_l   = mm->add_literal(alpha);
+    auto t_a   = add_common_op(*mm, migraphx::make_op("mul"), {a_l, l0});
+    t_a        = mm->add_instruction(
+        migraphx::make_op("convert", {{"target_type", l1->get_shape().type()}}), t_a);
+
+    t_a = mm->add_instruction(migraphx::make_op("transpose", {{"dims", {0, 1, 3, 2}}}), t_a);
+
+    auto b_l = mm->add_literal(beta);
+    auto b_b = mm->add_instruction(
+        migraphx::make_op("multibroadcast", {{"output_lens", l2->get_shape().lens()}}), b_l);
+    auto l2_b = mm->add_instruction(migraphx::make_op("mul"), l2, b_b);
+    mm->add_instruction(migraphx::make_op("dot", {{"alpha", 1.0f}, {"beta", 1.0f}}), t_a, l1, l2_b);
+
     auto prog = optimize_onnx("gemm_ex_test.onnx");
 
     EXPECT(p == prog);
@@ -1291,13 +1328,25 @@ TEST_CASE(gemm_ex_brcst_test)
     auto l0  = mm->add_parameter("1", migraphx::shape{migraphx::shape::float_type, {1, 1, 5, 6}});
     auto l1  = mm->add_parameter("2", migraphx::shape{migraphx::shape::float_type, {1, 1, 5, 7}});
     auto l2  = mm->add_parameter("3", migraphx::shape{migraphx::shape::float_type, {1, 1, 6, 1}});
-    auto t0  = mm->add_instruction(migraphx::make_op("transpose", {{"dims", {0, 1, 3, 2}}}), l0);
     std::vector<std::size_t> out_lens{1, 1, 6, 7};
-    auto t2 =
-        mm->add_instruction(migraphx::make_op("multibroadcast", {{"output_lens", out_lens}}), l2);
     auto alpha = 0.5f;
     auto beta  = 0.8f;
-    mm->add_instruction(migraphx::make_op("dot", {{"alpha", alpha}, {"beta", beta}}), t0, l1, t2);
+    auto a_l   = mm->add_literal(alpha);
+    auto t_a   = add_common_op(*mm, migraphx::make_op("mul"), {a_l, l0});
+    t_a        = mm->add_instruction(
+        migraphx::make_op("convert", {{"target_type", l1->get_shape().type()}}), t_a);
+
+    t_a = mm->add_instruction(migraphx::make_op("transpose", {{"dims", {0, 1, 3, 2}}}), t_a);
+
+    auto b_l = mm->add_literal(beta);
+    auto l2_b =
+        mm->add_instruction(migraphx::make_op("multibroadcast", {{"output_lens", out_lens}}), l2);
+    auto b_b = mm->add_instruction(
+        migraphx::make_op("multibroadcast", {{"output_lens", l2_b->get_shape().lens()}}), b_l);
+    auto l2_bb = mm->add_instruction(migraphx::make_op("mul"), l2_b, b_b);
+    mm->add_instruction(
+        migraphx::make_op("dot", {{"alpha", 1.0f}, {"beta", 1.0f}}), t_a, l1, l2_bb);
+
     auto prog = optimize_onnx("gemm_ex_brcst_test.onnx");
 
     EXPECT(p == prog);
@@ -2381,38 +2430,84 @@ TEST_CASE(quantizelinear_test)
     auto* mm = p.get_main_module();
     auto l0  = mm->add_parameter("0", {migraphx::shape::float_type, {5}});
     auto l1  = mm->add_parameter("1", {migraphx::shape::float_type, {1}});
-    auto l2      = mm->add_parameter("2", {migraphx::shape::int8_type, {1}});
-    auto min_val = mm->add_literal(-128);
-    auto max_val = mm->add_literal(127);
-
     auto l1_mbcast =
         mm->add_instruction(migraphx::make_op("multibroadcast", {{"output_lens", {5}}}), l1);
+    auto div   = mm->add_instruction(migraphx::make_op("div"), l0, l1_mbcast);
+    auto round = mm->add_instruction(migraphx::make_op("round"), div);
+    auto s     = round->get_shape();
+    std::vector<int> min_data(s.elements(), 0);
+    std::vector<int> max_data(s.elements(), 255);
+    auto min_arg = mm->add_literal(s, min_data);
+    auto max_arg = mm->add_literal(s, max_data);
+    auto clip    = mm->add_instruction(migraphx::make_op("clip"), round, min_arg, max_arg);
+    mm->add_instruction(
+        migraphx::make_op("convert",
+                          {{"target_type", migraphx::to_value(migraphx::shape::uint8_type)}}),
+        clip);
+
+    auto prog = optimize_onnx("quantizelinear_test.onnx", true);
+    EXPECT(p.sort() == prog.sort());
+}
 
+TEST_CASE(quantizelinear_int32_test)
+{
+    migraphx::program p;
+    auto* mm = p.get_main_module();
+    auto l0  = mm->add_parameter("0", {migraphx::shape::int32_type, {5}});
+    auto l1  = mm->add_parameter("1", {migraphx::shape::float_type, {1}});
+    auto l1_mbcast =
+        mm->add_instruction(migraphx::make_op("multibroadcast", {{"output_lens", {5}}}), l1);
+    l0 = mm->add_instruction(
+        migraphx::make_op("convert",
+                          {{"target_type", migraphx::to_value(migraphx::shape::float_type)}}),
+        l0);
     auto div   = mm->add_instruction(migraphx::make_op("div"), l0, l1_mbcast);
     auto round = mm->add_instruction(migraphx::make_op("round"), div);
-    l2         = mm->add_instruction(
+    auto s     = round->get_shape();
+    std::vector<int> min_data(s.elements(), 0);
+    std::vector<int> max_data(s.elements(), 255);
+    auto min_arg = mm->add_literal(s, min_data);
+    auto max_arg = mm->add_literal(s, max_data);
+    auto clip    = mm->add_instruction(migraphx::make_op("clip"), round, min_arg, max_arg);
+    mm->add_instruction(
         migraphx::make_op("convert",
-                          {{"target_type", migraphx::to_value(migraphx::shape::int32_type)}}),
-        l2);
+                          {{"target_type", migraphx::to_value(migraphx::shape::uint8_type)}}),
+        clip);
+
+    auto prog = optimize_onnx("quantizelinear_int32_test.onnx", true);
+    EXPECT(p.sort() == prog.sort());
+}
+
+TEST_CASE(quantizelinear_zero_point_test)
+{
+    migraphx::program p;
+    auto* mm = p.get_main_module();
+    auto l0  = mm->add_parameter("0", {migraphx::shape::float_type, {5}});
+    auto l1  = mm->add_parameter("1", {migraphx::shape::float_type, {1}});
+    auto l2  = mm->add_parameter("2", {migraphx::shape::int8_type, {1}});
+    auto l1_mbcast =
+        mm->add_instruction(migraphx::make_op("multibroadcast", {{"output_lens", {5}}}), l1);
+    auto div   = mm->add_instruction(migraphx::make_op("div"), l0, l1_mbcast);
+    auto round = mm->add_instruction(migraphx::make_op("round"), div);
     auto l2_mbcast =
         mm->add_instruction(migraphx::make_op("multibroadcast", {{"output_lens", {5}}}), l2);
-
-    round = mm->add_instruction(
+    l2_mbcast = mm->add_instruction(
         migraphx::make_op("convert",
-                          {{"target_type", migraphx::to_value(migraphx::shape::int32_type)}}),
-        round);
+                          {{"target_type", migraphx::to_value(migraphx::shape::float_type)}}),
+        l2_mbcast);
     auto add = mm->add_instruction(migraphx::make_op("add"), round, l2_mbcast);
-    min_val =
-        mm->add_instruction(migraphx::make_op("multibroadcast", {{"output_lens", {5}}}), min_val);
-    max_val =
-        mm->add_instruction(migraphx::make_op("multibroadcast", {{"output_lens", {5}}}), max_val);
-    auto clip = mm->add_instruction(migraphx::make_op("clip"), add, min_val, max_val);
+    auto s   = round->get_shape();
+    std::vector<int> min_data(s.elements(), -128);
+    std::vector<int> max_data(s.elements(), 127);
+    auto min_arg = mm->add_literal(s, min_data);
+    auto max_arg = mm->add_literal(s, max_data);
+    auto clip    = mm->add_instruction(migraphx::make_op("clip"), add, min_arg, max_arg);
     mm->add_instruction(
         migraphx::make_op("convert",
                           {{"target_type", migraphx::to_value(migraphx::shape::int8_type)}}),
         clip);
 
-    auto prog = optimize_onnx("quantizelinear_test.onnx");
+    auto prog = optimize_onnx("quantizelinear_zero_point_test.onnx", true);
     EXPECT(p.sort() == prog.sort());
 }
 
@@ -2426,9 +2521,6 @@ migraphx::program make_quantizelinear_axis_prog()
     auto l0       = mm->add_parameter("0", {migraphx::shape::float_type, input_lens});
     auto l1       = mm->add_parameter("1", {migraphx::shape::float_type, {5}});
     auto l2       = mm->add_parameter("2", {migraphx::shape::int8_type, {5}});
-    auto min_val = mm->add_literal(-128);
-    auto max_val = mm->add_literal(127);
-
     auto l1_bcast = mm->add_instruction(
         migraphx::make_op("broadcast", {{"axis", axis}, {"dims", input_lens}}), l1);
 
@@ -2438,18 +2530,15 @@ migraphx::program make_quantizelinear_axis_prog()
         migraphx::make_op("broadcast", {{"axis", axis}, {"dims", input_lens}}), l2);
     l2_bcast = mm->add_instruction(
         migraphx::make_op("convert",
-                          {{"target_type", migraphx::to_value(migraphx::shape::int32_type)}}),
+                          {{"target_type", migraphx::to_value(migraphx::shape::float_type)}}),
         l2_bcast);
-    round = mm->add_instruction(
-        migraphx::make_op("convert",
-                          {{"target_type", migraphx::to_value(migraphx::shape::int32_type)}}),
-        round);
     auto add = mm->add_instruction(migraphx::make_op("add"), round, l2_bcast);
-    min_val  = mm->add_instruction(
-        migraphx::make_op("multibroadcast", {{"output_lens", {1, 1, 5, 1}}}), min_val);
-    max_val = mm->add_instruction(
-        migraphx::make_op("multibroadcast", {{"output_lens", {1, 1, 5, 1}}}), max_val);
-    auto clip = mm->add_instruction(migraphx::make_op("clip"), add, min_val, max_val);
+    auto s   = round->get_shape();
+    std::vector<int> min_data(s.elements(), -128);
+    std::vector<int> max_data(s.elements(), 127);
+    auto min_arg = mm->add_literal(s, min_data);
+    auto max_arg = mm->add_literal(s, max_data);
+    auto clip    = mm->add_instruction(migraphx::make_op("clip"), add, min_arg, max_arg);
     mm->add_instruction(
         migraphx::make_op("convert",
                           {{"target_type", migraphx::to_value(migraphx::shape::int8_type)}}),
@@ -2461,7 +2550,7 @@ TEST_CASE(quantizelinear_axis_test)
 {
     migraphx::program p = make_quantizelinear_axis_prog();
 
-    auto prog = optimize_onnx("quantizelinear_axis_test.onnx");
+    auto prog = optimize_onnx("quantizelinear_axis_test.onnx", true);
     EXPECT(p.sort() == prog.sort());
 }
 
@@ -2469,7 +2558,7 @@ TEST_CASE(quantizelinear_neg_axis_test)
 {
     migraphx::program p = make_quantizelinear_axis_prog();
 
-    auto prog = optimize_onnx("quantizelinear_neg_axis_test.onnx");
+    auto prog = optimize_onnx("quantizelinear_neg_axis_test.onnx", true);
     EXPECT(p.sort() == prog.sort());
 }
 

test/onnx/quantizelinear_axis_test.onnx

@@ -23,4 +23,4 @@
 

 

 
-
B
+
B
\ No newline at end of file

test/onnx/quantizelinear_int32_test.onnx

+quantizelinear_int32_test:m
+
+
0
+
1out"QuantizeLinearquantizelinear_int32_testZ
+
0
+
+
+Z
+
1
+
+

+
b
+out
+
+
+B
\ No newline at end of file