Quantize linear ops (#843)

* Add operators, refactor parsers, add rewrite passes, add tests

* Formatting

* Fix cppcheck

* Review comments

* Formatting

* Combine rewrite passes

* Formatting

* Add ref implementations

* Formatting

* Review comments

* Formatting

* Tidy warnings

* Apply review comments

* Formatting

* Fix CI error

* Formatting

* Increase code coverage

* Formatting

* Move broadcasting of scales and zero points to onnx parser

* Formatting

* Allow for x and zero_point to have different types in quantizelinear; fix zero_point default type

* Formatting

* Increase code coverage

* Formatting

* Switch certain variables to int64_t

* Formatting

* Fix overflow in implicit constant conversion

* Formatting

* Increase code coverage

* Formatting

* Remove operators.hpp from includes in tf_test.cpp

* Formatting

* Add conversion for int32 input to quantizelinear and add test case; remove operators.hpp from onnx_test.cpp includes

* Formatting

* Switch dequantizelinear math from int32 to float

* Formatting

* Remove changes to operators.hpp

* Simplify apply_quantizelinear

* Formatting

* Add verify test for int32 data

* Add rewrite_quantization back to CMakeLists

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/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;
+    };
+};