merge changes from develop branch

src/include/migraphx/op/clip.hpp

+#ifndef MIGRAPHX_GUARD_OPERATORS_CLIP_HPP
+#define MIGRAPHX_GUARD_OPERATORS_CLIP_HPP
+
+#include <array>
+#include <migraphx/op/unary.hpp>
+#include <migraphx/operation.hpp>
+#include <migraphx/check_shapes.hpp>
+#include <migraphx/stringutils.hpp>
+#include <migraphx/streamutils.hpp>
+#include <migraphx/literal.hpp>
+#include <migraphx/shape_for_each.hpp>
+#include <migraphx/config.hpp>
+#include <cmath>
+#include <utility>
+#include <limits>
+
+namespace migraphx {
+inline namespace MIGRAPHX_INLINE_NS {
+namespace op {
+
+struct clip : unary<clip>
+{
+    float max_val = std::numeric_limits<float>::max();
+    float min_val = std::numeric_limits<float>::min();
+
+    clip() {}
+
+    clip(float max, float min) : max_val(max), min_val(min) {}
+
+    auto apply() const
+    {
+        auto max = max_val;
+        auto min = min_val;
+        return [max, min](auto x) {
+            using type = decltype(x);
+            return std::min(std::max(type(min), x), type(max));
+        };
+    }
+
+    template <class Self, class F>
+    static auto reflect(Self& self, F f)
+    {
+        return pack(f(self.max_val, "max"), f(self.min_val, "min"));
+    }
+};
+
+} // namespace op
+} // namespace MIGRAPHX_INLINE_NS
+} // namespace migraphx
+
+#endif

src/include/migraphx/operators.hpp

@@ -11,6 +11,7 @@
 #include <migraphx/op/batch_norm.hpp>
 #include <migraphx/op/binary.hpp>
 #include <migraphx/op/broadcast.hpp>
+#include <migraphx/op/clip.hpp>
 #include <migraphx/op/common.hpp>
 #include <migraphx/op/concat.hpp>
 #include <migraphx/op/contiguous.hpp>

src/onnx/onnx.cpp

@@ -63,6 +63,7 @@ struct onnx_parser
         add_variadic_op("Max", op::max{});
         add_variadic_op("Min", op::min{});
 
+        add_mem_op("Clip", &onnx_parser::parse_clip);
         add_mem_op("LRN", &onnx_parser::parse_lrn);
         add_mem_op("ImageScaler", &onnx_parser::parse_imagescaler);
         add_mem_op("LeakyRelu", &onnx_parser::parse_leaky_relu);
@@ -225,6 +226,22 @@ struct onnx_parser
         });
     }
 
+    instruction_ref parse_clip(const std::string&,
+                               const attribute_map& attributes,
+                               std::vector<instruction_ref> args)
+    {
+        op::clip op;
+        if(contains(attributes, "max"))
+        {
+            op.max_val = parse_value(attributes.at("max")).at<float>();
+        }
+        if(contains(attributes, "min"))
+        {
+            op.min_val = parse_value(attributes.at("min")).at<float>();
+        }
+        return prog.add_instruction(op, std::move(args));
+    }
+
     instruction_ref
     parse_softmax(const std::string&, const attribute_map&, std::vector<instruction_ref> args)
     {

src/targets/cpu/lowering.cpp

@@ -140,6 +140,21 @@ struct cpu_lrn
     }
 };
 
+struct clip_op
+{
+    op::clip op;
+    std::string name() const { return "cpu::clip"; }
+    auto fcn() const
+    {
+        auto max = op.max_val;
+        auto min = op.min_val;
+        return [max, min](auto x) {
+            using type = decltype(x);
+            return std::min(std::max(type(min), x), type(max));
+        };
+    }
+};
+
 struct cpu_convolution
 {
     op::convolution op;
@@ -965,6 +980,7 @@ struct cpu_apply
         apply_map["batch_norm_inference"] =
             extend_op<cpu_batch_norm_inference, op::batch_norm_inference>();
         apply_map["lrn"]        = extend_op<cpu_lrn, op::lrn>();
+        apply_map["clip"]       = extend_op<cpu_unary<clip_op>, op::clip>();
         apply_map["contiguous"] = extend_op<cpu_contiguous, op::contiguous>();
         apply_map["pad"]        = extend_op<cpu_pad, op::pad>();
         apply_map["concat"]     = extend_op<cpu_concat, op::concat>();

src/targets/gpu/CMakeLists.txt

@@ -33,6 +33,7 @@ add_library(migraphx_device
     device/gather.cpp
     device/sub.cpp
     device/pack.cpp
+    device/clip.cpp
 )
 set_target_properties(migraphx_device PROPERTIES EXPORT_NAME device)
 rocm_clang_tidy_check(migraphx_device)
@@ -70,6 +71,7 @@ add_library(migraphx_gpu
     schedule_model.cpp
     adjust_allocation.cpp
     pack_int8_args.cpp
+    clip.cpp
 )
 set_target_properties(migraphx_gpu PROPERTIES EXPORT_NAME gpu)
 rocm_clang_tidy_check(migraphx_gpu)

src/targets/gpu/clip.cpp

+#include <migraphx/gpu/clip.hpp>
+#include <migraphx/gpu/context.hpp>
+#include <migraphx/gpu/device/clip.hpp>
+
+namespace migraphx {
+inline namespace MIGRAPHX_INLINE_NS {
+namespace gpu {
+
+shape hip_clip::compute_shape(std::vector<shape> inputs) const
+{
+    inputs.pop_back();
+    return op.compute_shape(inputs);
+}
+
+argument hip_clip::compute(context& ctx, const shape&, const std::vector<argument>& args) const
+{
+    device::clip(ctx.get_stream().get(), args.back(), args.front(), op.max_val, op.min_val);
+    return args.back();
+}
+
+} // namespace gpu
+} // namespace MIGRAPHX_INLINE_NS
+} // namespace migraphx

src/targets/gpu/device/clip.cpp

+#include <migraphx/gpu/device/clip.hpp>
+#include <migraphx/gpu/device/nary.hpp>
+
+namespace migraphx {
+inline namespace MIGRAPHX_INLINE_NS {
+namespace gpu {
+namespace device {
+
+void clip(hipStream_t stream,
+          const argument& result,
+          const argument& arg1,
+          const float max,
+          const float min)
+{
+    nary(stream, result, arg1)(
+        [max, min](auto x) { return std::min<decltype(x)>(std::max<decltype(x)>(min, x), max); });
+}
+
+} // namespace device
+} // namespace gpu
+} // namespace MIGRAPHX_INLINE_NS
+} // namespace migraphx

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

+#ifndef MIGRAPHX_GUARD_RTGLIB_CLIP_HPP
+#define MIGRAPHX_GUARD_RTGLIB_CLIP_HPP
+
+#include <migraphx/shape.hpp>
+#include <migraphx/op/clip.hpp>
+
+namespace migraphx {
+inline namespace MIGRAPHX_INLINE_NS {
+namespace gpu {
+
+struct context;
+
+struct hip_clip
+{
+    op::clip op;
+    std::string name() const { return "gpu::clip"; }
+    shape compute_shape(std::vector<shape> inputs) const;
+    argument
+    compute(context& ctx, const shape& output_shape, const std::vector<argument>& args) const;
+    std::ptrdiff_t output_alias(const std::vector<shape>& shapes) const
+    {
+        return shapes.size() - 1;
+    }
+};
+
+} // namespace gpu
+} // namespace MIGRAPHX_INLINE_NS
+} // namespace migraphx
+
+#endif

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

+#ifndef MIGRAPHX_GUARD_RTGLIB_DEVICE_CLIP_HPP
+#define MIGRAPHX_GUARD_RTGLIB_DEVICE_CLIP_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 clip(hipStream_t stream, const argument& result, const argument& arg1, float max, float min);
+
+} // namespace device
+} // namespace gpu
+} // namespace MIGRAPHX_INLINE_NS
+} // namespace migraphx
+
+#endif

src/targets/gpu/lowering.cpp

@@ -47,6 +47,7 @@
 #include <migraphx/gpu/pad.hpp>
 #include <migraphx/gpu/gather.hpp>
 #include <migraphx/gpu/lrn.hpp>
+#include <migraphx/gpu/clip.hpp>
 #include <utility>
 #include <functional>
 #include <algorithm>
@@ -104,6 +105,7 @@ struct miopen_apply
         add_extend_op<hip_logsoftmax, op::logsoftmax>("logsoftmax");
         add_extend_op<hip_gather, op::gather>("gather");
         add_extend_op<hip_pad, op::pad>("pad");
+        add_extend_op<hip_clip, op::clip>("clip");
 
         add_lrn_op();
         add_convolution_op();

test/cpu_ops_test.cpp

@@ -1557,4 +1557,21 @@ TEST_CASE(fp16_test)
     EXPECT(migraphx::verify_range(results_vector, gold));
 }
 
+TEST_CASE(clip_test)
+{
+    migraphx::program p;
+    migraphx::shape s{migraphx::shape::float_type, {3}};
+    auto l = p.add_literal(migraphx::literal{s, {-1.0, 0.0, 10.0}});
+    migraphx::op::clip op;
+    op.max_val = 6.0;
+    op.min_val = 0.0;
+    p.add_instruction(op, 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 = {0.0, 0.0, 6.0};
+    EXPECT(migraphx::verify_range(results_vector, gold));
+}
+
 int main(int argc, const char* argv[]) { test::run(argc, argv); }

test/gpu/miopen.cpp

@@ -1326,6 +1326,17 @@ struct test_batchnorm_inference : verify_program<test_batchnorm_inference>
     }
 };
 
+struct test_clip : verify_program<test_clip>
+{
+    migraphx::program create_program() const
+    {
+        migraphx::program p;
+        auto x = p.add_parameter("x", migraphx::shape{migraphx::shape::float_type, {3}});
+        p.add_instruction(migraphx::op::clip{6.0, 0.0}, x);
+        return p;
+    }
+};
+
 struct test_conv_bn : verify_program<test_conv_bn>
 {
     migraphx::program create_program() const

test/onnx/onnx_test.cpp

@@ -794,4 +794,14 @@ TEST_CASE(no_pad_test)
     EXPECT(p == prog);
 }
 
+TEST_CASE(clip_test)
+{
+    migraphx::program p;
+    auto l0 = p.add_parameter("0", migraphx::shape{migraphx::shape::float_type, {3}});
+    p.add_instruction(migraphx::op::clip{6.0, 0.0}, l0);
+    auto prog = migraphx::parse_onnx("clip_test.onnx");
+
+    EXPECT(p == prog);
+}
+
 int main(int argc, const char* argv[]) { test::run(argc, argv); }