add rsqrt and tests

src/include/migraphx/op/rsqrt.hpp

+#ifndef MIGRAPHX_GUARD_OPERATORS_RSQRT_HPP
+#define MIGRAPHX_GUARD_OPERATORS_RSQRT_HPP
+
+#include <migraphx/op/unary.hpp>
+#include <cmath>
+
+namespace migraphx {
+inline namespace MIGRAPHX_INLINE_NS {
+namespace op {
+
+struct rsqrt : unary<rsqrt>
+{
+    auto apply() const
+    {
+        return [](auto x) { return 1 / std::sqrt(x); };
+    }
+};
+
+} // namespace op
+} // namespace MIGRAPHX_INLINE_NS
+} // namespace migraphx
+
+#endif

src/include/migraphx/operators.hpp

@@ -51,6 +51,7 @@
 #include <migraphx/op/rnn.hpp>
 #include <migraphx/op/rnn_last_cell_output.hpp>
 #include <migraphx/op/rnn_last_output.hpp>
+#include <migraphx/op/rsqrt.hpp>
 #include <migraphx/op/scalar.hpp>
 #include <migraphx/op/sigmoid.hpp>
 #include <migraphx/op/sinh.hpp>

src/targets/gpu/CMakeLists.txt

@@ -39,6 +39,7 @@ add_library(migraphx_device
     device/sub.cpp
     device/clip.cpp
     device/reduce_sum.cpp
+    device/rsqrt.cpp
 )
 set_target_properties(migraphx_device PROPERTIES EXPORT_NAME device)
 rocm_clang_tidy_check(migraphx_device)

src/targets/gpu/device/rsqrt.cpp

+#include <migraphx/gpu/device/rsqrt.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 rsqrt(hipStream_t stream, const argument& result, const argument& arg)
+{
+    nary(stream, result, arg)([](auto x) { return 1 / ::sqrt(to_hip_type(x)); });
+}
+
+} // namespace device
+} // namespace gpu
+} // namespace MIGRAPHX_INLINE_NS
+} // namespace migraphx

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

+#ifndef MIGRAPHX_GUARD_RTGLIB_DEVICE_RSQRT_HPP
+#define MIGRAPHX_GUARD_RTGLIB_DEVICE_RSQRT_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 rsqrt(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/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/lowering.cpp

@@ -51,6 +51,7 @@
 #include <migraphx/gpu/convert.hpp>
 #include <migraphx/gpu/clip.hpp>
 #include <migraphx/gpu/reduce_sum.hpp>
+#include <migraphx/gpu/rsqrt.hpp>
 #include <utility>
 #include <functional>
 #include <algorithm>
@@ -100,6 +101,7 @@ struct miopen_apply
         add_generic_op<hip_mul>("mul");
         add_generic_op<hip_max>("max");
         add_generic_op<hip_min>("min");
+        add_generic_op<hip_rsqrt>("rsqrt");
 
         add_extend_op<miopen_gemm, op::dot>("dot");
         add_extend_op<miopen_contiguous, op::contiguous>("contiguous");

src/tf/tf.cpp

@@ -153,6 +153,7 @@ 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_binary_op("Add", op::add{});

test/cpu_ops_test.cpp

@@ -1778,4 +1778,18 @@ TEST_CASE(reduce_sum_test12)
     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));
+}
+
 int main(int argc, const char* argv[]) { test::run(argc, argv); }

test/gpu/miopen.cpp

@@ -3511,4 +3511,17 @@ 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;
+    };
+};
+
 int main(int argc, const char* argv[]) { test::run(argc, argv); }

test/tf/rsqrt_test.pb

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

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

test/tf/tf_test.cpp

@@ -326,6 +326,16 @@ TEST_CASE(reshape_test)
     EXPECT(p == prog);
 }
 
+TEST_CASE(rsqrt_test)
+{
+    migraphx::program p;
+    auto l0 = p.add_parameter("0", migraphx::shape{migraphx::shape::float_type, {1, 3, 16, 16}});
+    p.add_instruction(migraphx::op::rsqrt{}, l0);
+    auto prog = optimize_tf("rsqrt_test.pb", false);
+
+    EXPECT(p == prog);
+}
+
 TEST_CASE(softmax_test)
 {
     migraphx::program p;