merge changes from develop branch

src/targets/gpu/CMakeLists.txt

@@ -40,9 +40,9 @@ add_library(migraphx_device
     device/div.cpp
     device/clip.cpp
     device/reduce_sum.cpp
-    device/pow.cpp
-    device/reduce_mean.cpp
     device/sqrt.cpp
+    device/reduce_mean.cpp
+    device/pow.cpp
 )
 set_target_properties(migraphx_device PROPERTIES EXPORT_NAME device)
 rocm_clang_tidy_check(migraphx_device)

src/targets/gpu/lowering.cpp

@@ -52,9 +52,9 @@
 #include <migraphx/gpu/convert.hpp>
 #include <migraphx/gpu/clip.hpp>
 #include <migraphx/gpu/reduce_sum.hpp>
-#include <migraphx/gpu/pow.hpp>
-#include <migraphx/gpu/reduce_mean.hpp>
 #include <migraphx/gpu/sqrt.hpp>
+#include <migraphx/gpu/reduce_mean.hpp>
+#include <migraphx/gpu/pow.hpp>
 #include <utility>
 #include <functional>
 #include <algorithm>

src/tf/tf.cpp

@@ -79,7 +79,8 @@ struct tf_parser
         return result;
     }
 
-    std::vector<size_t> parse_axes(const attribute_map& attributes, const std::string& s) const
+    std::vector<size_t>
+    parse_axes(const attribute_map& attributes, const std::string& s, const size_t num_dims) const
     {
         auto attrs = attributes.at(s).list().i();
         std::vector<size_t> axes;
@@ -87,14 +88,14 @@ struct tf_parser
         if(is_nhwc)
         {
             std::transform(axes.begin(), axes.end(), axes.begin(), [&](size_t axis) {
-                return parse_axis(axis);
+                return parse_axis(axis, num_dims);
             });
         }
         return axes;
     }
 
     template <class T>
-    std::vector<T> parse_axes(std::vector<T> axes) const
+    std::vector<T> parse_axes(std::vector<T> axes, const size_t num_dims) const
     {
         if(is_nhwc)
         {
@@ -102,7 +103,7 @@ struct tf_parser
             std::transform(axes.begin(),
                            axes.end(),
                            std::back_inserter(new_axes),
-                           [&](size_t axis) { return parse_axis(axis); });
+                           [&](size_t axis) { return parse_axis(axis, num_dims); });
             return new_axes;
         }
         return axes;
@@ -117,17 +118,17 @@ struct tf_parser
         std::vector<T> new_data(prev_data.size());
         for(size_t i = 0; i < new_data.size(); i++)
         {
-            auto new_idx         = parse_axis(i);
+            auto new_idx         = parse_axis(i, new_data.size());
             new_data.at(new_idx) = prev_data.at(i);
         }
         prev_data = new_data;
     }
 
     template <class T>
-    T parse_axis(const T& dim) const
+    T parse_axis(const T& dim, const size_t num_dims) const
     {
         T new_dim = dim;
-        if(is_nhwc)
+        if(is_nhwc and num_dims >= 4)
         {
             switch(dim)
             {
@@ -165,6 +166,7 @@ struct tf_parser
         add_mem_op("Const", &tf_parser::parse_constant);
         add_mem_op("Conv2D", &tf_parser::parse_conv);
         add_mem_op("DepthwiseConv2dNative", &tf_parser::parse_depthwiseconv);
+        add_mem_op("ExpandDims", &tf_parser::parse_expanddims, false);
         add_mem_op("FusedBatchNorm", &tf_parser::parse_batchnorm);
         add_mem_op("MatMul", &tf_parser::parse_matmul, false);
         add_mem_op("MaxPool", &tf_parser::parse_pooling);
@@ -490,6 +492,25 @@ struct tf_parser
         return prog.add_instruction(op, {l0, new_weights});
     }
 
+    instruction_ref
+    parse_expanddims(const std::string&, const attribute_map&, std::vector<instruction_ref> args)
+    {
+        std::vector<size_t> input_dims = args[0]->get_shape().lens();
+        std::vector<int64_t> new_dims(input_dims.begin(), input_dims.end());
+        size_t num_dims = input_dims.size();
+        int32_t dim     = args[1]->eval().at<int32_t>();
+
+        if(dim < 0)
+        {
+            new_dims.insert(new_dims.begin() + (num_dims + dim + 1), 1);
+        }
+        else
+        {
+            new_dims.insert(new_dims.begin() + dim, 1);
+        }
+        return prog.add_instruction(op::reshape{new_dims}, args[0]);
+    }
+
     instruction_ref
     parse_matmul(const std::string&, attribute_map attributes, std::vector<instruction_ref> args)
     {
@@ -519,11 +540,12 @@ struct tf_parser
     instruction_ref
     parse_mean(const std::string&, attribute_map attributes, std::vector<instruction_ref> args)
     {
-        auto axes      = parse_axes(args[1]->eval().get<int32_t>().to_vector());
         bool keep_dims = attributes.at("keep_dims").b();
         std::vector<int32_t> hw_axes{2, 3};
         // check if conditions for GlobalAvgPool are met
         auto lens = args[0]->get_shape().lens();
+        auto axes = parse_axes(args[1]->eval().get<int32_t>().to_vector(), lens.size());
+
         if(axes == hw_axes and lens.size() == 4)
         {
             op::pooling op{"average"};
@@ -694,14 +716,15 @@ struct tf_parser
                                   std::vector<instruction_ref> args)
     {
         op::squeeze op;
-        auto axes = attributes.at("squeeze_dims").list().i();
+        auto input_dims = args[0]->get_shape().lens();
+        auto axes       = attributes.at("squeeze_dims").list().i();
         copy(axes, std::back_inserter(op.axes));
-        auto args0_dims = args[0]->get_shape().lens();
+
         if(op.axes.empty()) // no squeeze_dims provided, remove any dim that equals 1
         {
-            for(size_t i = 0; i < args0_dims.size(); i++)
+            for(size_t i = 0; i < input_dims.size(); i++)
             {
-                if(args0_dims.at(i) == 1)
+                if(input_dims.at(i) == 1)
                 {
                     op.axes.push_back(i);
                 }

test/cpu_ops_test.cpp

@@ -542,6 +542,21 @@ TEST_CASE(erf_test)
     EXPECT(migraphx::verify_range(results_vector, gold));
 }
 
+TEST_CASE(sqrt_test)
+{
+    migraphx::program p;
+    migraphx::shape s{migraphx::shape::float_type, {5}};
+    auto l = p.add_literal(
+        migraphx::literal{s, {1.02481645, 0.85643062, 0.03404123, 0.92791926, 0.10569184}});
+    p.add_instruction(migraphx::op::sqrt{}, l);
+    p.compile(migraphx::cpu::target{});
+    auto result = p.eval({});
+    std::vector<float> results_vector;
+    result.visit([&](auto output) { results_vector.assign(output.begin(), output.end()); });
+    std::vector<float> gold = {1.01233218, 0.92543537, 0.18450265, 0.96328566, 0.32510282};
+    EXPECT(migraphx::verify_range(results_vector, gold));
+}
+
 TEST_CASE(log_test)
 {
     migraphx::program p;

test/gpu/miopen.cpp

@@ -255,6 +255,19 @@ struct test_erf : verify_program<test_erf>
     }
 };
 
+struct test_sqrt : verify_program<test_sqrt>
+{
+    migraphx::program create_program() const
+    {
+        migraphx::program p;
+        migraphx::shape s{migraphx::shape::float_type, {2, 3, 4, 6}};
+        auto param     = p.add_parameter("x", s);
+        auto param_abs = p.add_instruction(migraphx::op::abs{}, param);
+        p.add_instruction(migraphx::op::sqrt{}, param_abs);
+        return p;
+    }
+};
+
 struct test_log : verify_program<test_log>
 {
     migraphx::program create_program() const

test/onnx/onnx_test.cpp

@@ -202,6 +202,16 @@ TEST_CASE(erf_test)
     EXPECT(p == prog);
 }
 
+TEST_CASE(sqrt_test)
+{
+    migraphx::program p;
+    auto input = p.add_parameter("x", migraphx::shape{migraphx::shape::float_type, {10, 15}});
+    p.add_instruction(migraphx::op::sqrt{}, input);
+
+    auto prog = migraphx::parse_onnx("sqrt_test.onnx");
+    EXPECT(p == prog);
+}
+
 TEST_CASE(log_test)
 {
     migraphx::program p;

test/onnx/sqrt_test.onnx

+sqrt-example:C
+
+
x
y"Sqrt	test_sqrtZ
+
x
+

+
+
+b
+
y
+

+
+
+B

test/tf/tf_test.cpp

@@ -159,6 +159,31 @@ TEST_CASE(depthwiseconv_test)
     EXPECT(p == prog);
 }
 
+TEST_CASE(expanddims_test)
+{
+    migraphx::program p;
+
+    auto l0 = p.add_parameter("0", migraphx::shape{migraphx::shape::float_type, {2, 3, 4}});
+    p.add_literal(0);
+    p.add_instruction(migraphx::op::reshape{{1, 2, 3, 4}}, l0);
+    auto prog = optimize_tf("expanddims_test.pb", false);
+
+    EXPECT(p == prog);
+}
+
+TEST_CASE(expanddims_test_neg_dims)
+{
+    // this check makes sure the pb parses negative dim value correctly
+    migraphx::program p;
+
+    auto l0 = p.add_parameter("0", migraphx::shape{migraphx::shape::float_type, {2, 3, 4}});
+    p.add_literal(-1);
+    p.add_instruction(migraphx::op::reshape{{2, 3, 4, 1}}, l0);
+    auto prog = optimize_tf("expanddims_neg_test.pb", false);
+
+    EXPECT(p == prog);
+}
+
 TEST_CASE(identity_test)
 {
     migraphx::program p;