Merge branch 'develop' into scatter-op

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);
-    l0 = mm->add_instruction(
+    l2_mbcast = mm->add_instruction(
         migraphx::make_op("convert",
-                          {{"target_type", migraphx::to_value(migraphx::shape::int32_type)}}),
-        l0);
-
-    auto sub     = mm->add_instruction(migraphx::make_op("sub"), l0, l2_mbcast);
-    auto dequant = mm->add_instruction(
+                          {{"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::float_type)}}),
-        sub);
+        l0);
 
-    mm->add_instruction(migraphx::make_op("mul"), dequant, l1_mbcast);
+    auto sub = mm->add_instruction(migraphx::make_op("sub"), l0, l2_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)}}),
-        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);
+        l0);
+    auto sub = mm->add_instruction(migraphx::make_op("sub"), l0, l2_bcast);
 
-    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());
 }
 
@@ -1252,19 +1269,28 @@ TEST_CASE(gather_elements_axis1_test)
 TEST_CASE(gemm_test)
 {
     migraphx::program p;
-    auto* mm = p.get_main_module();
-    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* mm   = p.get_main_module();
+    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 alpha = 2.f;
     auto beta  = 2.0f;
-    mm->add_instruction(migraphx::make_op("dot", {{"alpha", alpha}, {"beta", beta}}), t0, t1, bl2);
+    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);
@@ -2378,41 +2427,87 @@ TEST_CASE(prelu_brcst_test)
 TEST_CASE(quantizelinear_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 min_val = mm->add_literal(-128);
-    auto max_val = mm->add_literal(127);
-
+    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 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());
 }
 
@@ -2423,12 +2518,9 @@ migraphx::program make_quantizelinear_axis_prog()
     int axis = 2;
     auto* mm = p.get_main_module();
 
-    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 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 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

test/onnx/quantizelinear_neg_axis_test.onnx

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

 

 
-
B
+
B
\ No newline at end of file

test/onnx/quantizelinear_test.onnx

-quantizelinear_test:{
-
+quantizelinear_test:g
+
 
0
-
1
-
2out"QuantizeLinearquantizelinear_testZ
+
1out"QuantizeLinearquantizelinear_testZ
 
0
 
 

@@ -10,12 +9,8 @@
 
1
 
 

-
Z
-
2
-
-
 
b
 out
 
 
-B
+B
\ No newline at end of file

test/onnx/quantizelinear_zero_point_test.onnx

+quantizelinear_zero_point_test:

+
+
0
+
1
+
2out"QuantizeLinearquantizelinear_zero_point_testZ
+
0
+
+

+Z
+
1
+
+

+
Z
+
2
+
+
+
b
+out
+
+
+B
\ No newline at end of file

test/ref_ops_test.cpp

@@ -1233,6 +1233,58 @@ TEST_CASE(deconv_test)
     EXPECT(migraphx::verify_range(results_vector, gold));
 }
 
+TEST_CASE(dequantizelinear)
+{
+    { /*uint8*/
+        migraphx::shape xs{migraphx::shape::uint8_type, {1, 3, 3}};
+        std::vector<uint8_t> xv = {0, 1, 2, 5, 10, 50, 100, 150, 250};
+        migraphx::shape ss{migraphx::shape::float_type, {1, 3, 3}};
+        std::vector<float> sv = {2, 2, 2, 2, 2, 2, 2, 2, 2};
+        migraphx::shape zs{migraphx::shape::uint8_type, {1, 3, 3}};
+        std::vector<uint8_t> zv = {0, 0, 0, 0, 0, 0, 0, 0, 0};
+        auto create_program     = [&]() {
+            migraphx::program p;
+            auto* mm = p.get_main_module();
+            auto x   = mm->add_literal(xs, xv);
+            auto s   = mm->add_literal(ss, sv);
+            auto z   = mm->add_literal(zs, zv);
+            mm->add_instruction(migraphx::make_op("dequantizelinear"), x, s, z);
+            return p;
+        };
+
+        migraphx::program p1 = create_program();
+        p1.compile(migraphx::ref::target{});
+        auto result = p1.eval({}).back();
+        std::vector<float> results_vector(9);
+        result.visit([&](auto output) { results_vector.assign(output.begin(), output.end()); });
+        std::vector<float> gold{0, 2, 4, 10, 20, 100, 200, 300, 500};
+        EXPECT(results_vector == gold);
+    }
+
+    { /*int8*/
+        migraphx::shape xs{migraphx::shape::int8_type, {1, 3, 3}};
+        std::vector<int8_t> xv = {-128, -100, -50, -1, 0, 1, 50, 100, 127};
+        migraphx::shape ss{migraphx::shape::float_type, {1, 3, 3}};
+        std::vector<float> sv = {2, 2, 2, 2, 2, 2, 2, 2, 2};
+        auto create_program   = [&]() {
+            migraphx::program p;
+            auto* mm = p.get_main_module();
+            auto x   = mm->add_literal(xs, xv);
+            auto s   = mm->add_literal(ss, sv);
+            mm->add_instruction(migraphx::make_op("dequantizelinear"), x, s);
+            return p;
+        };
+
+        migraphx::program p1 = create_program();
+        p1.compile(migraphx::ref::target{});
+        auto result = p1.eval({}).back();
+        std::vector<float> results_vector(9);
+        result.visit([&](auto output) { results_vector.assign(output.begin(), output.end()); });
+        std::vector<float> gold{-256, -200, -100, -2, 0, 2, 100, 200, 254};
+        EXPECT(results_vector == gold);
+    }
+}
+
 TEST_CASE(div_test)
 {
     migraphx::program p;
@@ -3287,6 +3339,61 @@ TEST_CASE(quant_conv2d_test)
     EXPECT(migraphx::verify_range(results_vector, s));
 }
 
+TEST_CASE(quantizelinear)
+{
+    {
+        migraphx::shape xs{migraphx::shape::float_type, {2, 3, 3}};
+        std::vector<float> xv = {
+            -300, 600, 129, -1000, 4, 3, -6, 600, 550, -300, 600, 129, -1000, 4, 3, -6, 600, 550};
+        migraphx::shape ss{migraphx::shape::float_type, {2, 3, 3}};
+        std::vector<float> sv = {2, 2, 2, 4, 4, 4, 6, 6, 6, 2, 2, 2, 4, 4, 4, 6, 6, 6};
+        migraphx::shape zs{migraphx::shape::int8_type, {2, 3, 3}};
+        std::vector<uint8_t> zv = {0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0};
+        auto create_program     = [&]() {
+            migraphx::program p;
+            auto* mm = p.get_main_module();
+            auto x   = mm->add_literal(xs, xv);
+            auto s   = mm->add_literal(ss, sv);
+            auto z   = mm->add_literal(zs, zv);
+            mm->add_instruction(migraphx::make_op("quantizelinear"), x, s, z);
+            return p;
+        };
+
+        migraphx::program p1 = create_program();
+        p1.compile(migraphx::ref::target{});
+        auto result = p1.eval({}).back();
+        std::vector<float> results_vector(18);
+        result.visit([&](auto output) { results_vector.assign(output.begin(), output.end()); });
+        std::vector<float> gold{
+            -128, 127, 65, -128, 1, 1, -1, 100, 92, -128, 127, 65, -128, 1, 1, -1, 100, 92};
+        EXPECT(results_vector == gold);
+    }
+
+    {
+        migraphx::shape xs{migraphx::shape::float_type, {2, 3, 3}};
+        std::vector<float> xv = {
+            -300, 600, 129, -1000, 4, 3, -6, 600, 550, -300, 600, 129, -1000, 4, 3, -6, 600, 550};
+        migraphx::shape ss{migraphx::shape::float_type, {2, 3, 3}};
+        std::vector<float> sv = {2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2};
+        auto create_program   = [&]() {
+            migraphx::program p;
+            auto* mm = p.get_main_module();
+            auto x   = mm->add_literal(xs, xv);
+            auto s   = mm->add_literal(ss, sv);
+            mm->add_instruction(migraphx::make_op("quantizelinear"), x, s);
+            return p;
+        };
+
+        migraphx::program p1 = create_program();
+        p1.compile(migraphx::ref::target{});
+        auto result = p1.eval({}).back();
+        std::vector<float> results_vector(18);
+        result.visit([&](auto output) { results_vector.assign(output.begin(), output.end()); });
+        std::vector<float> gold{0, 255, 65, 0, 2, 2, 0, 255, 255, 0, 255, 65, 0, 2, 2, 0, 255, 255};
+        EXPECT(results_vector == gold);
+    }
+}
+
 TEST_CASE(recip_test)
 {
     migraphx::program p;

test/rewrite_quantization_test.cpp

+#include <migraphx/rewrite_quantization.hpp>
+#include <migraphx/program.hpp>
+#include <migraphx/ref/target.hpp>
+#include <migraphx/op/convolution.hpp>
+#include <migraphx/op/reshape.hpp>
+#include <migraphx/instruction.hpp>
+#include <migraphx/generate.hpp>
+#include <migraphx/ranges.hpp>
+#include <test.hpp>
+#include <migraphx/make_op.hpp>
+
+#include <migraphx/serialize.hpp>
+
+#include <migraphx/verify.hpp>
+
+bool is_quantizelinear(migraphx::instruction& ins) { return ins.name() == "quantizelinear"; }
+bool is_dequantizelinear(migraphx::instruction& ins) { return ins.name() == "dequantizelinear"; }
+
+TEST_CASE(quantizelinear)
+{
+
+    migraphx::shape xs{migraphx::shape::float_type, {1, 3, 3}};
+    std::vector<float> xv = {-300, 200, 129, 1, 2, 3, 500, 1000, 50};
+    migraphx::shape ss{migraphx::shape::float_type, {1, 3, 3}};
+    std::vector<float> sv = {2, 2, 2, 2, 2, 2, 2, 2, 2};
+    auto create_program   = [&]() {
+        migraphx::program p;
+        auto* mm = p.get_main_module();
+        auto x   = mm->add_literal(xs, xv);
+        auto s   = mm->add_literal(ss, sv);
+        mm->add_instruction(migraphx::make_op("quantizelinear"), x, s);
+        return p;
+    };
+
+    migraphx::program p1 = create_program();
+    migraphx::program p2 = create_program();
+
+    migraphx::rewrite_quantization opt;
+    opt.apply(*p2.get_main_module());
+    EXPECT(any_of(*p1.get_main_module(), &is_quantizelinear));
+    EXPECT(none_of(*p2.get_main_module(), &is_quantizelinear));
+}
+
+TEST_CASE(dequantizelinear)
+{
+
+    migraphx::shape xs{migraphx::shape::float_type, {1, 3, 3}};
+    std::vector<float> xv = {0, 1, 2, 5, 10, 50, 100, 150, 250};
+    migraphx::shape ss{migraphx::shape::float_type, {1, 3, 3}};
+    std::vector<float> sv = {2, 2, 2, 2, 2, 2, 2, 2, 2};
+    migraphx::shape zs{migraphx::shape::uint8_type, {1, 3, 3}};
+    std::vector<uint8_t> zv = {0, 0, 0, 0, 0, 0, 0, 0, 0};
+    auto create_program     = [&]() {
+        migraphx::program p;
+        auto* mm = p.get_main_module();
+        auto x   = mm->add_literal(xs, xv);
+        auto s   = mm->add_literal(ss, sv);
+        auto z   = mm->add_literal(zs, zv);
+        mm->add_instruction(migraphx::make_op("dequantizelinear"), x, s, z);
+        return p;
+    };
+
+    migraphx::program p1 = create_program();
+    migraphx::program p2 = create_program();
+
+    migraphx::rewrite_quantization opt;
+    opt.apply(*p2.get_main_module());
+    EXPECT(any_of(*p1.get_main_module(), &is_dequantizelinear));
+    EXPECT(none_of(*p2.get_main_module(), &is_dequantizelinear));
+}
+
+int main(int argc, const char* argv[]) { test::run(argc, argv); }

test/tf/tf_test.cpp

@@ -6,11 +6,15 @@
 #include <migraphx/simplify_reshapes.hpp>
 #include <migraphx/dead_code_elimination.hpp>
 #include <migraphx/eliminate_identity.hpp>
-#include <migraphx/operators.hpp>
 #include <migraphx/program.hpp>
 #include <migraphx/instruction.hpp>
 #include <migraphx/tf.hpp>
 #include <migraphx/make_op.hpp>
+#include <migraphx/op/batch_norm_inference.hpp>
+#include <migraphx/op/convolution.hpp>
+#include <migraphx/op/reduce_mean.hpp>
+#include <migraphx/op/pooling.hpp>
+#include <migraphx/op/slice.hpp>
 
 #include <migraphx/serialize.hpp>
 

test/verify/run_verify.cpp

@@ -27,7 +27,7 @@ std::future<typename std::result_of<Function()>::type> detach_async(Function&& f
         std::packaged_task<result_type()> task(std::forward<Function>(f));
         auto fut = task.get_future();
         std::thread(std::move(task)).detach();
-        return std::move(fut);
+        return fut;
     }
     return std::async(std::launch::deferred, std::forward<Function>(f));
 }

test/verify/test_dequantizelinear.cpp

+
+#include "verify_program.hpp"
+#include <migraphx/program.hpp>
+#include <migraphx/generate.hpp>
+#include <migraphx/make_op.hpp>
+
+struct test_dequantizelinear : verify_program<test_dequantizelinear>
+{
+    migraphx::program create_program() const
+    {
+        migraphx::program p;
+        auto* mm = p.get_main_module();
+
+        migraphx::shape sx{migraphx::shape::int8_type, {2, 2, 2}};
+        migraphx::shape ss{migraphx::shape::float_type, {2, 2, 2}};
+        migraphx::shape sz{migraphx::shape::int8_type, {2, 2, 2}};
+        auto input1 = mm->add_parameter("x", sx);
+        auto input2 = mm->add_parameter("x_scale", ss);
+        auto input3 = mm->add_parameter("x_zero_point", sz);
+        auto r = mm->add_instruction(migraphx::make_op("dequantizelinear"), input1, input2, input3);
+        mm->add_return({r});
+        return p;
+    };
+};

test/verify/test_quantizelinear.cpp

+
+#include "verify_program.hpp"
+#include <migraphx/program.hpp>
+#include <migraphx/generate.hpp>
+#include <migraphx/make_op.hpp>
+
+struct test_quantizelinear : verify_program<test_quantizelinear>
+{
+    migraphx::program create_program() const
+    {
+        migraphx::program p;
+        auto* mm = p.get_main_module();
+
+        migraphx::shape sx{migraphx::shape::float_type, {2, 2, 2}};
+        migraphx::shape ss{migraphx::shape::float_type, {2, 2, 2}};
+        migraphx::shape sz{migraphx::shape::int8_type, {2, 2, 2}};
+        auto input1 = mm->add_parameter("x", sx);
+        auto input2 = mm->add_parameter("y_scale", ss);
+        auto input3 = mm->add_parameter("y_zero_point", sz);
+        auto r = mm->add_instruction(migraphx::make_op("quantizelinear"), input1, input2, input3);
+        mm->add_return({r});
+        return p;
+    };
+};

test/verify/test_quantizelinear_int32.cpp

+
+#include "verify_program.hpp"
+#include <migraphx/program.hpp>
+#include <migraphx/generate.hpp>
+#include <migraphx/make_op.hpp>
+
+struct test_quantizelinear_int32 : verify_program<test_quantizelinear_int32>
+{
+    migraphx::program create_program() const
+    {
+        migraphx::program p;
+        auto* mm = p.get_main_module();
+
+        migraphx::shape sx{migraphx::shape::int32_type, {2, 2, 2}};
+        migraphx::shape ss{migraphx::shape::float_type, {2, 2, 2}};
+        migraphx::shape sz{migraphx::shape::int8_type, {2, 2, 2}};
+        auto input1 = mm->add_parameter("x", sx);
+        auto input2 = mm->add_parameter("y_scale", ss);
+        auto input3 = mm->add_parameter("y_zero_point", sz);
+        auto r = mm->add_instruction(migraphx::make_op("quantizelinear"), input1, input2, input3);
+        mm->add_return({r});
+        return p;
+    };
+};