manual merge

src/include/migraphx/op/pow.hpp

+#ifndef MIGRAPHX_GUARD_OPERATORS_POW_HPP
+#define MIGRAPHX_GUARD_OPERATORS_POW_HPP
+
+#include <migraphx/op/binary.hpp>
+#include <migraphx/config.hpp>
+
+namespace migraphx {
+inline namespace MIGRAPHX_INLINE_NS {
+namespace op {
+
+struct pow : binary<pow>
+{
+    auto apply() const
+    {
+        return [](auto x, auto y) { return std::pow(x, y); };
+    }
+};
+
+} // namespace op
+} // namespace MIGRAPHX_INLINE_NS
+} // namespace migraphx
+
+#endif

src/include/migraphx/op/sqrt.hpp

+#ifndef MIGRAPHX_GUARD_OPERATORS_SQRT_HPP
+#define MIGRAPHX_GUARD_OPERATORS_SQRT_HPP
+
+#include <migraphx/op/unary.hpp>
+#include <migraphx/config.hpp>
+
+namespace migraphx {
+inline namespace MIGRAPHX_INLINE_NS {
+namespace op {
+
+struct sqrt : unary<sqrt>
+{
+    auto apply() const
+    {
+        return [](auto x) { return std::sqrt(x); };
+    }
+};
+
+} // namespace op
+} // namespace MIGRAPHX_INLINE_NS
+} // namespace migraphx
+
+#endif

src/include/migraphx/operators.hpp

@@ -45,6 +45,7 @@
 #include <migraphx/op/outline.hpp>
 #include <migraphx/op/pad.hpp>
 #include <migraphx/op/pooling.hpp>
+#include <migraphx/op/pow.hpp>
 #include <migraphx/op/reduce_sum.hpp>
 #include <migraphx/op/reduce_mean.hpp>
 #include <migraphx/op/relu.hpp>
@@ -59,6 +60,7 @@
 #include <migraphx/op/sin.hpp>
 #include <migraphx/op/slice.hpp>
 #include <migraphx/op/softmax.hpp>
+#include <migraphx/op/sqrt.hpp>
 #include <migraphx/op/sqdiff.hpp>
 #include <migraphx/op/squeeze.hpp>
 #include <migraphx/op/sub.hpp>

src/onnx/onnx.cpp

@@ -54,11 +54,13 @@ struct onnx_parser
         add_generic_op("Asin", op::asin{});
         add_generic_op("Acos", op::acos{});
         add_generic_op("Atan", op::atan{});
+        add_generic_op("Sqrt", op::sqrt{});
 
         add_binary_op("Add", op::add{});
         add_binary_op("Div", op::div{});
         add_binary_op("Mul", op::mul{});
         add_binary_op("Sub", op::sub{});
+        add_binary_op("Pow", op::pow{});
 
         add_variadic_op("Sum", op::add{});
         add_variadic_op("Max", op::max{});

src/targets/gpu/CMakeLists.txt

@@ -41,7 +41,9 @@ add_library(migraphx_device
     device/clip.cpp
     device/reduce_sum.cpp
     device/rsqrt.cpp
+    device/sqrt.cpp
     device/reduce_mean.cpp
+    device/pow.cpp
     device/sqdiff.cpp
 )
 set_target_properties(migraphx_device PROPERTIES EXPORT_NAME device)

src/targets/gpu/device/pow.cpp

+#include <migraphx/gpu/device/pow.hpp>
+#include <migraphx/gpu/device/nary.hpp>
+
+namespace migraphx {
+inline namespace MIGRAPHX_INLINE_NS {
+namespace gpu {
+namespace device {
+
+void pow(hipStream_t stream, const argument& result, const argument& arg1, const argument& arg2)
+{
+    nary(stream, result, arg1, arg2)(
+        [](auto b, auto e) { return ::pow(to_hip_type(b), to_hip_type(e)); });
+}
+
+} // namespace device
+} // namespace gpu
+} // namespace MIGRAPHX_INLINE_NS
+} // namespace migraphx

src/targets/gpu/device/sqrt.cpp

+#include <migraphx/gpu/device/sqrt.hpp>
+#include <migraphx/gpu/device/nary.hpp>
+#include <migraphx/gpu/device/types.hpp>
+
+namespace migraphx {
+inline namespace MIGRAPHX_INLINE_NS {
+namespace gpu {
+namespace device {
+
+void sqrt(hipStream_t stream, const argument& result, const argument& arg)
+{
+    nary(stream, result, arg)([](auto x) { return ::sqrt(to_hip_type(x)); });
+}
+
+} // namespace device
+} // namespace gpu
+} // namespace MIGRAPHX_INLINE_NS
+} // namespace migraphx

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

+
+#ifndef MIGRAPHX_GUARD_RTGLIB_DEVICE_POW_HPP
+#define MIGRAPHX_GUARD_RTGLIB_DEVICE_POW_HPP
+
+#include <migraphx/argument.hpp>
+#include <migraphx/config.hpp>
+#include <hip/hip_runtime_api.h>
+
+namespace migraphx {
+inline namespace MIGRAPHX_INLINE_NS {
+namespace gpu {
+namespace device {
+
+void pow(hipStream_t stream, const argument& result, const argument& arg1, const argument& arg2);
+
+} // namespace device
+} // namespace gpu
+} // namespace MIGRAPHX_INLINE_NS
+} // namespace migraphx
+
+#endif

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

+#ifndef MIGRAPHX_GUARD_RTGLIB_DEVICE_SQRT_HPP
+#define MIGRAPHX_GUARD_RTGLIB_DEVICE_SQRT_HPP
+
+#include <migraphx/argument.hpp>
+#include <migraphx/config.hpp>
+#include <hip/hip_runtime_api.h>
+
+namespace migraphx {
+inline namespace MIGRAPHX_INLINE_NS {
+namespace gpu {
+namespace device {
+
+void sqrt(hipStream_t stream, const argument& result, const argument& arg);
+
+} // namespace device
+} // namespace gpu
+} // namespace MIGRAPHX_INLINE_NS
+} // namespace migraphx
+
+#endif

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/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

@@ -53,7 +53,9 @@
 #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>
@@ -101,11 +103,13 @@ 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");

src/tf/tf.cpp

@@ -180,6 +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>
@@ -767,6 +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;

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
@@ -267,6 +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

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 +899,30 @@ 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);
+}
+
 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/tf_test.cpp

@@ -445,4 +445,16 @@ TEST_CASE(tanh_test)
     EXPECT(p == prog);
 }
 
+TEST_CASE(transpose_test)
+{
+    migraphx::program p;
+    auto l0 = p.add_parameter("0", migraphx::shape{migraphx::shape::float_type, {1, 3, 16, 16}});
+    migraphx::shape s0{migraphx::shape::int32_type, {4}};
+    p.add_literal(migraphx::literal{s0, {0, 2, 3, 1}});
+    p.add_instruction(migraphx::op::transpose{{0, 2, 3, 1}}, l0);
+    auto prog = optimize_tf("transpose_test.pb", false);
+
+    EXPECT(p == prog);
+}
+
 int main(int argc, const char* argv[]) { test::run(argc, argv); }