Merge branch 'develop' into navi-reduce

src/include/migraphx/rewrite_pooling.hpp

@@ -26,6 +26,7 @@
 
 #include <string>
 #include <migraphx/config.hpp>
+#include <migraphx/instruction_ref.hpp>
 
 namespace migraphx {
 inline namespace MIGRAPHX_INLINE_NS {

src/include/migraphx/shape.hpp

@@ -34,6 +34,7 @@
 #include <migraphx/functional.hpp>
 #include <migraphx/errors.hpp>
 #include <migraphx/half.hpp>
+#include <migraphx/float8.hpp>
 #include <migraphx/serialize.hpp>
 #include <migraphx/config.hpp>
 
@@ -60,7 +61,8 @@ struct MIGRAPHX_EXPORT shape
     m(int32_type, int32_t) \
     m(int64_type, int64_t) \
     m(uint32_type, uint32_t) \
-    m(uint64_type, uint64_t)
+    m(uint64_type, uint64_t) \
+    m(fp8e4m3fnuz_type, migraphx::fp8::fp8e4m3fnuz)
     // clang-format on
 
 #define MIGRAPHX_SHAPE_GENERATE_ENUM_TYPES(x, t) x,

src/include/migraphx/simple_par_for.hpp

+/*
+ * The MIT License (MIT)
+ *
+ * Copyright (c) 2015-2022 Advanced Micro Devices, Inc. All rights reserved.
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a copy
+ * of this software and associated documentation files (the "Software"), to deal
+ * in the Software without restriction, including without limitation the rights
+ * to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+ * copies of the Software, and to permit persons to whom the Software is
+ * furnished to do so, subject to the following conditions:
+ *
+ * The above copyright notice and this permission notice shall be included in
+ * all copies or substantial portions of the Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+ * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+ * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL THE
+ * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+ * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+ * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN
+ * THE SOFTWARE.
+ */
+#ifndef MIGRAPHX_GUARD_RTGLIB_SIMPLE_PAR_FOR_HPP
+#define MIGRAPHX_GUARD_RTGLIB_SIMPLE_PAR_FOR_HPP
+
+#include <thread>
+#include <cmath>
+#include <algorithm>
+#include <vector>
+#include <cassert>
+
+namespace migraphx {
+inline namespace MIGRAPHX_INLINE_NS {
+
+struct joinable_thread : std::thread
+{
+    template <class... Xs>
+    joinable_thread(Xs&&... xs) : std::thread(std::forward<Xs>(xs)...) // NOLINT
+    {
+    }
+
+    joinable_thread& operator=(joinable_thread&& other) = default;
+    joinable_thread(joinable_thread&& other)            = default;
+
+    ~joinable_thread()
+    {
+        if(this->joinable())
+            this->join();
+    }
+};
+
+template <class F>
+auto thread_invoke(std::size_t i, std::size_t tid, F f) -> decltype(f(i, tid))
+{
+    f(i, tid);
+}
+
+template <class F>
+auto thread_invoke(std::size_t i, std::size_t, F f) -> decltype(f(i))
+{
+    f(i);
+}
+
+template <class F>
+void simple_par_for_impl(std::size_t n, std::size_t threadsize, F f)
+{
+    if(threadsize <= 1)
+    {
+        for(std::size_t i = 0; i < n; i++)
+            thread_invoke(i, 0, f);
+    }
+    else
+    {
+        std::vector<joinable_thread> threads(threadsize);
+// Using const here causes gcc 5 to ICE
+#if(!defined(__GNUC__) || __GNUC__ != 5)
+        const
+#endif
+            std::size_t grainsize = std::ceil(static_cast<double>(n) / threads.size());
+
+        std::size_t work = 0;
+        std::size_t tid  = 0;
+        std::generate(threads.begin(), threads.end(), [=, &work, &tid] {
+            auto result = joinable_thread([=] {
+                std::size_t start = work;
+                std::size_t last  = std::min(n, work + grainsize);
+                for(std::size_t i = start; i < last; i++)
+                {
+                    thread_invoke(i, tid, f);
+                }
+            });
+            work += grainsize;
+            ++tid;
+            return result;
+        });
+        assert(work >= n);
+    }
+}
+
+template <class F>
+void simple_par_for(std::size_t n, std::size_t min_grain, F f)
+{
+    const auto threadsize = std::min<std::size_t>(std::thread::hardware_concurrency(),
+                                                  n / std::max<std::size_t>(1, min_grain));
+    simple_par_for_impl(n, threadsize, f);
+}
+
+template <class F>
+void simple_par_for(std::size_t n, F f)
+{
+    const int min_grain = 8;
+    simple_par_for(n, min_grain, f);
+}
+
+} // namespace MIGRAPHX_INLINE_NS
+} // namespace migraphx
+
+#endif

src/include/migraphx/tune_axis.hpp

 /*
  * The MIT License (MIT)
  *
- * Copyright (c) 2015-2022 Advanced Micro Devices, Inc. All rights reserved.
+ * Copyright (c) 2015-2023 Advanced Micro Devices, Inc. All rights reserved.
  *
  * Permission is hereby granted, free of charge, to any person obtaining a copy
  * of this software and associated documentation files (the "Software"), to deal
@@ -24,21 +24,21 @@
 #ifndef MIGRAPHX_GUARD_OPERATORS_TUNE_AXIS_HPP
 #define MIGRAPHX_GUARD_OPERATORS_TUNE_AXIS_HPP
 
-#include <utility>
-#include <cstdint>
 #include <migraphx/stringutils.hpp>
 #include <migraphx/errors.hpp>
 
 namespace migraphx {
 inline namespace MIGRAPHX_INLINE_NS {
 
-inline int tune_axis(const int n_dim, const int axis, const std::string& op_name = "OPERATOR")
+inline int tune_axis(int n_dim, int axis, const std::string& op_name = "OPERATOR")
 {
-    if(axis >= n_dim or std::abs(axis) > n_dim)
-    {
+    if(axis < 0)
+        axis += n_dim;
+
+    if(axis < 0 or axis >= n_dim)
         MIGRAPHX_THROW(to_upper(op_name) + ": axis is out of range.");
-    }
-    return (axis < 0) ? axis + n_dim : axis;
+
+    return axis;
 }
 
 } // namespace MIGRAPHX_INLINE_NS

src/onnx/parse_generic_op.cpp

@@ -60,7 +60,7 @@ struct parse_generic_op : op_parser<parse_generic_op>
                 {"Neg", "neg"},
                 {"Reciprocal", "recip"},
                 {"Relu", "relu"},
-                {"Round", "round"},
+                {"Round", "nearbyint"},
                 {"Sigmoid", "sigmoid"},
                 {"Sign", "sign"},
                 {"Sin", "sin"},