merge changes from develop branch

src/targets/gpu/include/migraphx/gpu/oper.hpp

@@ -88,7 +88,7 @@ struct binary_device : oper<Derived>
 
     argument compute(context& ctx, const shape&, const std::vector<argument>& args) const
     {
-        F(ctx.get_stream().get(), args[2], args[1], args[0]);
+        F(ctx.get_stream().get(), args[2], args[0], args[1]);
         return args[2];
     }
 

src/targets/gpu/include/migraphx/gpu/pow.hpp

+#ifndef MIGRAPHX_GUARD_RTGLIB_POW_HPP
+#define MIGRAPHX_GUARD_RTGLIB_POW_HPP
+
+#include <migraphx/gpu/oper.hpp>
+#include <migraphx/gpu/device/pow.hpp>
+
+namespace migraphx {
+inline namespace MIGRAPHX_INLINE_NS {
+namespace gpu {
+
+struct hip_pow : binary_device<hip_pow, device::pow>
+{
+};
+
+} // namespace gpu
+} // namespace MIGRAPHX_INLINE_NS
+} // namespace migraphx
+
+#endif

src/targets/gpu/include/migraphx/gpu/rsqrt.hpp

+#ifndef MIGRAPHX_GUARD_RTGLIB_RSQRT_HPP
+#define MIGRAPHX_GUARD_RTGLIB_RSQRT_HPP
+
+#include <migraphx/gpu/oper.hpp>
+#include <migraphx/gpu/device/rsqrt.hpp>
+
+namespace migraphx {
+inline namespace MIGRAPHX_INLINE_NS {
+namespace gpu {
+
+struct hip_rsqrt : unary_device<hip_rsqrt, device::rsqrt>
+{
+};
+
+} // namespace gpu
+} // namespace MIGRAPHX_INLINE_NS
+} // namespace migraphx
+
+#endif

src/targets/gpu/include/migraphx/gpu/sqdiff.hpp

+#ifndef MIGRAPHX_GUARD_RTGLIB_SQDIFF_HPP
+#define MIGRAPHX_GUARD_RTGLIB_SQDIFF_HPP
+
+#include <migraphx/gpu/oper.hpp>
+#include <migraphx/gpu/device/sqdiff.hpp>
+
+namespace migraphx {
+inline namespace MIGRAPHX_INLINE_NS {
+namespace gpu {
+
+struct hip_sqdiff : binary_device<hip_sqdiff, device::sqdiff>
+{
+};
+
+} // namespace gpu
+} // namespace MIGRAPHX_INLINE_NS
+} // namespace migraphx
+
+#endif

src/targets/gpu/include/migraphx/gpu/sqrt.hpp

+#ifndef MIGRAPHX_GUARD_RTGLIB_SQRT_HPP
+#define MIGRAPHX_GUARD_RTGLIB_SQRT_HPP
+
+#include <migraphx/gpu/oper.hpp>
+#include <migraphx/gpu/device/sqrt.hpp>
+
+namespace migraphx {
+inline namespace MIGRAPHX_INLINE_NS {
+namespace gpu {
+
+struct hip_sqrt : unary_device<hip_sqrt, device::sqrt>
+{
+};
+
+} // namespace gpu
+} // namespace MIGRAPHX_INLINE_NS
+} // namespace migraphx
+
+#endif

src/targets/gpu/lowering.cpp

@@ -27,6 +27,7 @@
 #include <migraphx/gpu/logsoftmax.hpp>
 #include <migraphx/gpu/add.hpp>
 #include <migraphx/gpu/sub.hpp>
+#include <migraphx/gpu/div.hpp>
 #include <migraphx/gpu/exp.hpp>
 #include <migraphx/gpu/erf.hpp>
 #include <migraphx/gpu/log.hpp>
@@ -53,7 +54,11 @@
 #include <migraphx/gpu/convert.hpp>
 #include <migraphx/gpu/clip.hpp>
 #include <migraphx/gpu/reduce_sum.hpp>
+#include <migraphx/gpu/rsqrt.hpp>
+#include <migraphx/gpu/sqrt.hpp>
 #include <migraphx/gpu/reduce_mean.hpp>
+#include <migraphx/gpu/pow.hpp>
+#include <migraphx/gpu/sqdiff.hpp>
 #include <utility>
 #include <functional>
 #include <algorithm>
@@ -100,9 +105,14 @@ struct miopen_apply
         add_generic_op<hip_asin>("asin");
         add_generic_op<hip_acos>("acos");
         add_generic_op<hip_atan>("atan");
+        add_generic_op<hip_sqrt>("sqrt");
         add_generic_op<hip_mul>("mul");
+        add_generic_op<hip_div>("div");
         add_generic_op<hip_max>("max");
         add_generic_op<hip_min>("min");
+        add_generic_op<hip_rsqrt>("rsqrt");
+        add_generic_op<hip_pow>("pow");
+        add_generic_op<hip_sqdiff>("sqdiff");
 
         add_extend_op<miopen_gemm, op::dot>("dot");
         add_extend_op<miopen_contiguous, op::contiguous>("contiguous");

src/tf/tf.cpp

@@ -154,10 +154,13 @@ struct tf_parser
         add_generic_op("Identity", op::identity{});
         add_generic_op("Relu", op::relu{});
         add_generic_op("Relu6", op::clip{6.0, 0.0});
+        add_generic_op("Rsqrt", op::rsqrt{});
         add_generic_op("Tanh", op::tanh{});
+        add_generic_op("StopGradient", op::identity{});
 
         add_binary_op("Add", op::add{});
         add_binary_op("Mul", op::mul{});
+        add_binary_op("SquaredDifference", op::sqdiff{});
         add_binary_op("Sub", op::sub{});
 
         add_mem_op("AvgPool", &tf_parser::parse_pooling);
@@ -177,6 +180,7 @@ struct tf_parser
         add_mem_op("Softmax", &tf_parser::parse_softmax);
         add_mem_op("Squeeze", &tf_parser::parse_squeeze, false);
         add_mem_op("StridedSlice", &tf_parser::parse_stridedslice);
+        add_mem_op("Transpose", &tf_parser::parse_transpose, false);
     }
 
     template <class F>
@@ -764,6 +768,16 @@ struct tf_parser
         return to_nhwc(prog.add_instruction(op::squeeze{squeeze_axes}, l0));
     }
 
+    instruction_ref
+    parse_transpose(const std::string&, const attribute_map&, std::vector<instruction_ref> args)
+    {
+        auto perm = args[1]->eval().get<int32_t>().to_vector();
+        op::transpose op;
+        op.dims = std::vector<int64_t>(perm.begin(), perm.end());
+
+        return prog.add_instruction(op, args.front());
+    }
+
     void parse_graph(const tensorflow::GraphDef& graph)
     {
         nodes = get_nodes(graph, input_nodes);

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;
@@ -556,6 +571,21 @@ TEST_CASE(log_test)
     EXPECT(migraphx::verify_range(results_vector, gold));
 }
 
+TEST_CASE(pow_test)
+{
+    migraphx::program p;
+    migraphx::shape s{migraphx::shape::float_type, {3}};
+    auto b = p.add_literal(migraphx::literal{s, {1, 2, 3}});
+    auto e = p.add_literal(migraphx::literal{s, {1, 2, 3}});
+    p.add_instruction(migraphx::op::pow{}, b, e);
+    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 = {1.0f, 4.0f, 27.0f};
+    EXPECT(migraphx::verify_range(results_vector, gold));
+}
+
 TEST_CASE(sin_test)
 {
     migraphx::program p;
@@ -1879,6 +1909,20 @@ TEST_CASE(reduce_sum_axis12)
     EXPECT(results_vector == gold);
 }
 
+TEST_CASE(rsqrt_test)
+{
+    migraphx::program p;
+    migraphx::shape s{migraphx::shape::float_type, {3}};
+    auto l = p.add_literal(migraphx::literal{s, {4.0, 16.0, 64.0}});
+    p.add_instruction(migraphx::op::rsqrt{}, l);
+    p.compile(migraphx::cpu::target{});
+    auto result = p.eval({});
+    std::vector<float> results_vector(3);
+    result.visit([&](auto output) { results_vector.assign(output.begin(), output.end()); });
+    std::vector<float> gold = {0.5, 0.25, 0.125};
+    EXPECT(migraphx::verify_range(results_vector, gold));
+}
+
 TEST_CASE(reduce_mean_axis1)
 {
     migraphx::program p;
@@ -1954,4 +1998,19 @@ TEST_CASE(reduce_mean_int)
     EXPECT(results_vector == gold);
 }
 
+TEST_CASE(sqdiff_test)
+{
+    migraphx::program p;
+    migraphx::shape s{migraphx::shape::float_type, {3}};
+    auto l1 = p.add_literal(migraphx::literal{s, {-1, 0, 1}});
+    auto l2 = p.add_literal(migraphx::literal{s, {1, 2, 3}});
+    p.add_instruction(migraphx::op::sqdiff{}, l1, l2);
+    p.compile(migraphx::cpu::target{});
+    auto result = p.eval({});
+    std::vector<float> results_vector(3);
+    result.visit([&](auto output) { results_vector.assign(output.begin(), output.end()); });
+    std::vector<float> gold = {4, 4, 4};
+    EXPECT(migraphx::verify_range(results_vector, gold));
+}
+
 int main(int argc, const char* argv[]) { test::run(argc, argv); }

test/gpu/miopen.cpp

@@ -243,6 +243,31 @@ struct test_exp : verify_program<test_exp>
     }
 };
 
+struct test_erf : verify_program<test_erf>
+{
+    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);
+        p.add_instruction(migraphx::op::erf{}, param);
+        return p;
+    }
+};
+
+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
@@ -255,6 +280,20 @@ struct test_log : verify_program<test_log>
     }
 };
 
+struct test_pow : verify_program<test_pow>
+{
+    migraphx::program create_program() const
+    {
+        migraphx::program p;
+        migraphx::shape s{migraphx::shape::float_type, {6}};
+        std::vector<float> vec_e(s.elements(), 2.0f);
+        auto b = p.add_parameter("x", s);
+        auto e = p.add_literal(migraphx::literal(s, vec_e));
+        p.add_instruction(migraphx::op::pow{}, b, e);
+        return p;
+    }
+};
+
 struct test_sin : verify_program<test_sin>
 {
     migraphx::program create_program() const
@@ -569,6 +608,38 @@ struct test_sub2 : verify_program<test_sub2>
     }
 };
 
+struct test_div : verify_program<test_div>
+{
+    migraphx::program create_program() const
+    {
+        migraphx::program p;
+        migraphx::shape s{migraphx::shape::float_type, {3}};
+        auto x    = p.add_parameter("x", s);
+        auto y    = p.add_parameter("y", s);
+        auto z    = p.add_parameter("z", s);
+        auto diff = p.add_instruction(migraphx::op::div{}, x, y);
+        p.add_instruction(migraphx::op::div{}, diff, z);
+        return p;
+    }
+};
+
+struct test_div2 : verify_program<test_div2>
+{
+    migraphx::program create_program() const
+    {
+        migraphx::program p;
+        migraphx::shape s{migraphx::shape::float_type, {2, 3}};
+        migraphx::shape b{migraphx::shape::float_type, {3}};
+        auto x    = p.add_parameter("x", s);
+        auto y    = p.add_parameter("y", s);
+        auto z    = p.add_parameter("z", b);
+        auto zb   = p.add_instruction(migraphx::op::broadcast{1, s.lens()}, z);
+        auto diff = p.add_instruction(migraphx::op::div{}, x, y);
+        p.add_instruction(migraphx::op::div{}, diff, zb);
+        return p;
+    }
+};
+
 struct test_softmax1 : verify_program<test_softmax1>
 {
     migraphx::program create_program() const
@@ -3667,6 +3738,19 @@ struct test_reduce_sum_half : verify_program<test_reduce_sum_half>
     };
 };
 
+struct test_rsqrt : verify_program<test_rsqrt>
+{
+    migraphx::program create_program() const
+    {
+        migraphx::program p;
+        migraphx::shape s{migraphx::shape::float_type, {1, 3, 16, 16}};
+        auto x  = p.add_parameter("x", s);
+        auto l0 = p.add_instruction(migraphx::op::clip{std::numeric_limits<float>::max(), 1.0}, x);
+        p.add_instruction(migraphx::op::rsqrt{}, l0);
+        return p;
+    };
+};
+
 struct test_reduce_mean : verify_program<test_reduce_mean>
 {
     migraphx::program create_program() const

test/onnx/cast_test.onnx

+cast-example:F
+
+
x
y"Cast*	
+to

	test_castZ
+
x
+
+
+
+
+b
+
y
+
+

+
+B

test/onnx/const_of_shape2.onnx

+constant-of-shape:

+6shape"Constant*#
+value**Bshape_tensor 

+7
+shape
y"ConstantOfShape*
+value*
:

+Bvalue 
constant_of_shapeb
+
y
+

+
+
+B

test/onnx/const_of_shape3.onnx

+constant-of-shape:

+6shape"Constant*#
+value**Bshape_tensor 

+
+shape
y"ConstantOfShapeconstant_of_shapeb
+
y
+

+
+
+B

test/onnx/expand_test.onnx

+expand:

+7shape"Constant*$
+value**Bshape_tensor

+
+
x
+shape
y"ExpandexpandZ
+
x
+

+
+

+
b
+
y
+

+
+
+
+B

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;
@@ -889,4 +899,104 @@ TEST_CASE(clip_test)
     EXPECT(p == prog);
 }
 
+TEST_CASE(implicit_pow_bcast_test)
+{
+    migraphx::program p;
+    auto l0 = p.add_parameter("0", migraphx::shape{migraphx::shape::float_type, {2, 3, 4, 5}});
+    auto l1 = p.add_parameter("1", migraphx::shape{migraphx::shape::float_type, {3, 4, 1}});
+    auto l2 = p.add_instruction(migraphx::op::multibroadcast{{2, 3, 4, 5}}, l0);
+    auto l3 = p.add_instruction(migraphx::op::multibroadcast{{2, 3, 4, 5}}, l1);
+    p.add_instruction(migraphx::op::pow{}, l2, l3);
+
+    auto prog = migraphx::parse_onnx("pow_bcast_test.onnx");
+
+    EXPECT(p == prog);
+}
+
+TEST_CASE(pow_test)
+{
+    migraphx::program p;
+    auto l0 = p.add_parameter("0", migraphx::shape{migraphx::shape::float_type, {2, 3, 4, 5}});
+    auto l1 = p.add_parameter("1", migraphx::shape{migraphx::shape::float_type, {2, 3, 4, 5}});
+    p.add_instruction(migraphx::op::pow{}, l0, l1);
+
+    auto prog = migraphx::parse_onnx("pow_bcast_test1.onnx");
+
+    EXPECT(p == prog);
+}
+
+TEST_CASE(cast_test)
+{
+    migraphx::program p;
+    auto l = p.add_parameter("x", migraphx::shape{migraphx::shape::half_type, {10}});
+    p.add_instruction(migraphx::op::convert{migraphx::shape::float_type}, l);
+
+    auto prog = migraphx::parse_onnx("cast_test.onnx");
+    EXPECT(p == prog);
+}
+
+TEST_CASE(const_of_shape_float)
+{
+    migraphx::program p;
+    migraphx::shape ss(migraphx::shape::int32_type, {3});
+    p.add_literal(migraphx::literal(ss, {2, 3, 4}));
+    migraphx::shape s(migraphx::shape::float_type, {2, 3, 4});
+    std::vector<float> vec(s.elements(), 10.0f);
+    p.add_literal(migraphx::literal(s, vec));
+
+    auto prog = migraphx::parse_onnx("const_of_shape1.onnx");
+    EXPECT(p == prog);
+}
+
+TEST_CASE(const_of_shape_int64)
+{
+    migraphx::program p;
+    migraphx::shape ss(migraphx::shape::int32_type, {3});
+    p.add_literal(migraphx::literal(ss, {2, 3, 4}));
+    migraphx::shape s(migraphx::shape::int64_type, {2, 3, 4});
+    std::vector<int64_t> vec(s.elements(), 10);
+    p.add_literal(migraphx::literal(s, vec));
+
+    auto prog = migraphx::parse_onnx("const_of_shape2.onnx");
+    EXPECT(p == prog);
+}
+
+TEST_CASE(const_of_shape_no_value_attr)
+{
+    migraphx::program p;
+    migraphx::shape ss(migraphx::shape::int32_type, {3});
+    p.add_literal(migraphx::literal(ss, {2, 3, 4}));
+    migraphx::shape s(migraphx::shape::float_type, {2, 3, 4});
+    std::vector<float> vec(s.elements(), 0.0f);
+    p.add_literal(migraphx::literal(s, vec));
+
+    auto prog = migraphx::parse_onnx("const_of_shape3.onnx");
+    EXPECT(p == prog);
+}
+
+TEST_CASE(const_of_shape_empty_input)
+{
+    migraphx::program p;
+    p.add_literal(migraphx::literal());
+    migraphx::shape s(migraphx::shape::int64_type, {1}, {0});
+    std::vector<int64_t> vec(s.elements(), 10);
+    p.add_literal(migraphx::literal(s, vec));
+
+    auto prog = migraphx::parse_onnx("const_of_shape4.onnx");
+    EXPECT(p == prog);
+}
+
+TEST_CASE(expand_test)
+{
+    migraphx::program p;
+    migraphx::shape s(migraphx::shape::float_type, {3, 1, 1});
+    auto param = p.add_parameter("x", s);
+    migraphx::shape ss(migraphx::shape::int32_type, {4});
+    p.add_literal(migraphx::literal(ss, {2, 3, 4, 5}));
+    p.add_instruction(migraphx::op::multibroadcast{{2, 3, 4, 5}}, param);
+
+    auto prog = migraphx::parse_onnx("expand_test.onnx");
+    EXPECT(p == prog);
+}
+
 int main(int argc, const char* argv[]) { test::run(argc, argv); }

test/onnx/pow_bcast_test.onnx

+pow2:q
+
+
0
+
1out"Powpow_testZ
+
0
+

+
+
+
+Z
+
1
+

+
+
+
b
+out
+

+
+
+
+B

test/onnx/pow_bcast_test1.onnx

+pow2:u
+
+
0
+
1out"Powpow_testZ
+
0
+

+
+
+
+Z
+
1
+

+
+
+
+b
+out
+

+
+
+
+B

test/onnx/sqrt_test.onnx

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

+
+
+b
+
y
+

+
+
+B

test/tf/rsqrt_test.pb

+
+:
+
0Placeholder*
+shape:
*
+dtype0

+
+rsqrtRsqrt
0*
+
T0
"
\ No newline at end of file