add gpu implementation for the reduce_mean operator

src/targets/gpu/CMakeLists.txt

@@ -36,6 +36,7 @@ add_library(migraphx_device
     device/sub.cpp
     device/clip.cpp
     device/reduce_sum.cpp
+    device/reduce_mean.cpp
 )
 set_target_properties(migraphx_device PROPERTIES EXPORT_NAME device)
 rocm_clang_tidy_check(migraphx_device)
@@ -72,6 +73,7 @@ add_library(migraphx_gpu
     adjust_allocation.cpp
     clip.cpp
     reduce_sum.cpp
+    reduce_mean.cpp
 )
 set_target_properties(migraphx_gpu PROPERTIES EXPORT_NAME gpu)
 rocm_clang_tidy_check(migraphx_gpu)

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

@@ -28,6 +28,16 @@ struct id
     }
 };
 
+struct scale
+{
+    float factor = 1.0f;
+    template <class T>
+    MIGRAPHX_DEVICE_CONSTEXPR auto operator()(T x) const
+    {
+        return static_cast<T>(x * factor);
+    }
+};
+
 struct max
 {
     template <class T, class U>

src/targets/gpu/device/reduce_mean.cpp

+#include <migraphx/gpu/device/reduce_mean.hpp>
+#include <migraphx/gpu/device/reduce.hpp>
+
+namespace migraphx {
+inline namespace MIGRAPHX_INLINE_NS {
+namespace gpu {
+namespace device {
+
+void reduce_mean(hipStream_t stream, const argument& result, const argument& arg)
+{
+    std::size_t batch_item_num = arg.get_shape().elements() / result.get_shape().elements();
+    float factor = 1.0f / batch_item_num;
+    reduce(stream, result, arg, sum{}, 0, id{}, scale{factor});
+}
+
+} // namespace device
+} // namespace gpu
+} // namespace MIGRAPHX_INLINE_NS
+} // namespace migraphx

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

+#ifndef MIGRAPHX_GUARD_RTGLIB_DEVICE_REDUCE_MEAN_HPP
+#define MIGRAPHX_GUARD_RTGLIB_DEVICE_REDUCE_MEAN_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 reduce_mean(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/reduce_mean.hpp

+#ifndef MIGRAPHX_GUARD_RTGLIB_REDUCE_MEAN_HPP
+#define MIGRAPHX_GUARD_RTGLIB_REDUCE_MEAN_HPP
+
+#include <migraphx/shape.hpp>
+#include <migraphx/op/reduce_mean.hpp>
+#include <migraphx/reflect.hpp>
+
+namespace migraphx {
+inline namespace MIGRAPHX_INLINE_NS {
+namespace gpu {
+
+struct context;
+
+struct hip_reduce_mean
+{
+    op::reduce_mean op;
+
+    template <class Self, class F>
+    static auto reflect(Self& self, F f)
+    {
+        return migraphx::reflect(self.op, f);
+    }
+
+    std::string name() const { return "gpu::reduce_mean"; }
+    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/lowering.cpp

@@ -48,6 +48,7 @@
 #include <migraphx/gpu/convert.hpp>
 #include <migraphx/gpu/clip.hpp>
 #include <migraphx/gpu/reduce_sum.hpp>
+#include <migraphx/gpu/reduce_mean.hpp>
 #include <utility>
 #include <functional>
 #include <algorithm>
@@ -107,6 +108,7 @@ struct miopen_apply
         add_extend_op<hip_convert, op::convert>("convert");
         add_extend_op<hip_clip, op::clip>("clip");
         add_extend_op<hip_reduce_sum, op::reduce_sum>("reduce_sum");
+        add_extend_op<hip_reduce_mean, op::reduce_mean>("reduce_mean");
 
         add_lrn_op();
         add_convolution_op();

src/targets/gpu/reduce_mean.cpp

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