Merge branch 'develop' into enable_navi_32_ci

src/include/migraphx/auto_register.hpp

@@ -62,10 +62,9 @@ const int auto_register<Action, T>::static_register = auto_register_action<Actio
 #define MIGRAPHX_AUTO_REGISTER_NAME_DETAIL(x) migraphx_auto_register_##x
 #define MIGRAPHX_AUTO_REGISTER_NAME(x) MIGRAPHX_AUTO_REGISTER_NAME_DETAIL(x)
 // NOLINTNEXTLINE
-#define MIGRAPHX_AUTO_REGISTER(...)                                                        \
-    void MIGRAPHX_AUTO_REGISTER_NAME(__LINE__)(migraphx::auto_register<__VA_ARGS__> x =    \
-                                                   migraphx::auto_register<__VA_ARGS__>{}) \
-        __attribute__((unused));
+#define MIGRAPHX_AUTO_REGISTER(...)                              \
+    [[maybe_unused]] void MIGRAPHX_AUTO_REGISTER_NAME(__LINE__)( \
+        migraphx::auto_register<__VA_ARGS__> x = migraphx::auto_register<__VA_ARGS__>{});
 
 } // namespace MIGRAPHX_INLINE_NS
 } // namespace migraphx

src/include/migraphx/compile_src.hpp

@@ -37,8 +37,18 @@ inline namespace MIGRAPHX_INLINE_NS {
 struct src_file
 {
     fs::path path;
-    std::pair<const char*, const char*> content;
-    std::size_t len() const { return content.second - content.first; }
+    std::string_view content;
+
+    src_file() = default;
+    src_file(fs::path file_path, std::string_view file_content)
+        : path{std::move(file_path)}, content{file_content}
+    {
+    }
+
+    explicit src_file(const std::pair<std::string_view, std::string_view>& pair)
+        : path{pair.first}, content{pair.second}
+    {
+    }
 };
 
 struct MIGRAPHX_EXPORT src_compiler

src/include/migraphx/config.hpp

@@ -25,6 +25,7 @@
 #define MIGRAPHX_GUARD_CONFIG_HPP
 
 #include <migraphx/export.h>
+#include <ciso646>
 
 #if !defined(MIGRAPHX_USE_CLANG_TIDY) && !defined(DOXYGEN)
 

src/include/migraphx/dynamic_loader.hpp

@@ -38,12 +38,14 @@ struct dynamic_loader_impl;
 
 struct MIGRAPHX_EXPORT dynamic_loader
 {
+#ifndef _WIN32
     template <class T>
     static fs::path path(T* address)
     {
         return path(reinterpret_cast<void*>(address));
     }
     static fs::path path(void* address);
+#endif
 
     static optional<dynamic_loader> try_load(const fs::path& p);
 

src/include/migraphx/filesystem.hpp

@@ -29,6 +29,17 @@
 #if defined(CPPCHECK)
 #define MIGRAPHX_HAS_FILESYSTEM 1
 #define MIGRAPHX_HAS_FILESYSTEM_TS 1
+#elif defined(_WIN32)
+#if _MSC_VER >= 1920
+#define MIGRAPHX_HAS_FILESYSTEM 1
+#define MIGRAPHX_HAS_FILESYSTEM_TS 0
+#elif _MSC_VER >= 1900
+#define MIGRAPHX_HAS_FILESYSTEM 0
+#define MIGRAPHX_HAS_FILESYSTEM_TS 1
+#else
+#define MIGRAPHX_HAS_FILESYSTEM 0
+#define MIGRAPHX_HAS_FILESYSTEM_TS 0
+#endif
 #elif defined(__has_include)
 #if __has_include(<filesystem>) && __cplusplus >= 201703L
 #define MIGRAPHX_HAS_FILESYSTEM 1

src/include/migraphx/float_equal.hpp

@@ -27,9 +27,6 @@
 #include <algorithm>
 #include <cmath>
 #include <numeric>
-#ifdef _MSC_VER
-#include <iso646.h>
-#endif
 
 #include <migraphx/requires.hpp>
 #include <migraphx/config.hpp>

src/include/migraphx/generate.hpp

@@ -48,7 +48,7 @@ constexpr T normalize(unsigned long z)
 template <class T, MIGRAPHX_REQUIRES(is_signed<T>{} and not is_floating_point<T>{})>
 constexpr T normalize(unsigned long z)
 {
-    const auto max      = 1UL << (sizeof(T) * 5);
+    const auto max      = 1ULL << (sizeof(T) * 5);
     const auto half_max = max / 2;
     return half_max - (z % max);
 }
@@ -58,7 +58,7 @@ template <class T,
                             not std::is_same<T, bool>{})>
 constexpr T normalize(unsigned long z)
 {
-    const auto max = 1UL << (sizeof(T) * 5);
+    const auto max = 1ULL << (sizeof(T) * 5);
     return z % max;
 }
 

src/include/migraphx/instruction_ref.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
@@ -27,12 +27,42 @@
 #include <list>
 #include <functional>
 #include <migraphx/config.hpp>
+#include <migraphx/requires.hpp>
 
 namespace migraphx {
 inline namespace MIGRAPHX_INLINE_NS {
 
 struct instruction;
+#if defined(_WIN32) && !defined(NDEBUG)
+struct instruction_ref : std::list<instruction>::iterator
+{
+    using instruction_iter       = std::list<instruction>::iterator;
+    using instruction_const_iter = std::list<instruction>::const_iterator;
+
+    instruction_ref() = default;
+    instruction_ref(const instruction_iter& other) : instruction_iter(other) {}
+
+    template <class T,
+              class U,
+              MIGRAPHX_REQUIRES(std::is_same<T, instruction_ref>{} or
+                                std::is_same<U, instruction_ref>{})>
+    friend bool operator==(const T& x, const U& y)
+    {
+        return x._Unwrapped()._Ptr == y._Unwrapped()._Ptr;
+    }
+
+    template <class T,
+              class U,
+              MIGRAPHX_REQUIRES(std::is_same<T, instruction_ref>{} or
+                                std::is_same<U, instruction_ref>{})>
+    friend bool operator!=(const T& x, const U& y)
+    {
+        return not(x == y);
+    }
+};
+#else
 using instruction_ref = std::list<instruction>::iterator;
+#endif
 
 MIGRAPHX_EXPORT migraphx::instruction* as_address(const instruction_ref& ins) noexcept;
 
@@ -65,4 +95,8 @@ struct equal_to<migraphx::instruction_ref> // NOLINT
 
 } // namespace std
 
+#ifdef _MSC_VER
+#include <migraphx/instruction.hpp>
+#endif
+
 #endif

src/include/migraphx/matcher.hpp

@@ -33,6 +33,7 @@
 #include <migraphx/type_name.hpp>
 #include <migraphx/source_location.hpp>
 #include <migraphx/config.hpp>
+#include <array>
 #include <unordered_map>
 #include <unordered_set>
 

src/include/migraphx/normalize_attributes.hpp

@@ -52,6 +52,7 @@ using dependent_type = typename select_dependent_type<T, Ts...>::type;
  * \param attr_val the normalize_axes attributes from the operator
  * \param prefix error message prefix
  */
+MIGRAPHX_EXPORT
 std::vector<int64_t> normalize_axes(const std::vector<int64_t>& axes,
                                     const shape& input_shape,
                                     const value& attr_val,
@@ -67,6 +68,7 @@ std::vector<int64_t> normalize_axes(const std::vector<int64_t>& axes,
  * \param attr_val the normalize_axes attributes from the operator
  * \param prefix error message prefix
  */
+MIGRAPHX_EXPORT
 std::vector<int64_t> normalize_indices(const std::vector<int64_t>& indices,
                                        const std::vector<int64_t>& axes,
                                        const shape& input_shape,

src/include/migraphx/op/allocate.hpp

@@ -33,6 +33,19 @@ namespace migraphx {
 inline namespace MIGRAPHX_INLINE_NS {
 namespace op {
 
+/**
+ * Static allocate:
+ * No inputs: `allocate()`
+ * `this.s` attribute set to the static output shape of the buffer.
+ *
+ * Dynamic allocate:
+ * One input: `allocate(output_dims)`
+ * `output_dims` are the output buffer dimensions and has a static shape.
+ * Either `this.s` or `this.buf_type` must be set to calculate the dynamic output shape at compute
+ * time. If `this.buf_type` is set, the compute_shape() of allocate at compile time will have
+ * dynamic_dimensions from {0, max_int} with rank = output_dims.ndim(). If `this.s` is set then the
+ * compute_shape() will output `this.s`; `this.s` should be a dynamic shape.
+ */
 struct allocate
 {
     shape s{};
@@ -75,13 +88,13 @@ struct allocate
     {
         if(args.empty())
         {
-            return {output_shape};
+            return argument{output_shape};
         }
         else
         {
             std::vector<std::size_t> output_dims(output_shape.ndim());
             args.at(0).visit([&](auto a) { output_dims.assign(a.begin(), a.end()); });
-            return {shape{buf_type, output_dims}};
+            return argument{shape{buf_type, output_dims}};
         }
     }
 };

src/include/migraphx/op/argmax.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
@@ -31,6 +31,7 @@
 #include <migraphx/value.hpp>
 #include <migraphx/op/normalize_attribute.hpp>
 #include <migraphx/dyn_output.hpp>
+#include <migraphx/float_equal.hpp>
 
 namespace migraphx {
 inline namespace MIGRAPHX_INLINE_NS {
@@ -38,12 +39,13 @@ namespace op {
 
 struct argmax
 {
-    int64_t axis = 0;
+    int64_t axis           = 0;
+    bool select_last_index = false;
 
     template <class Self, class F>
     static auto reflect(Self& self, F f)
     {
-        return pack(f(self.axis, "axis"));
+        return pack(f(self.axis, "axis"), f(self.select_last_index, "select_last_index"));
     }
 
     value attributes() const
@@ -87,6 +89,10 @@ struct argmax
                 max_val   = cur_val;
                 max_index = i;
             }
+            else if(select_last_index and float_equal(max_val, cur_val))
+            {
+                max_index = i;
+            }
         }
         return max_index;
     }

src/include/migraphx/op/argmin.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
@@ -30,6 +30,7 @@
 #include <migraphx/config.hpp>
 #include <migraphx/value.hpp>
 #include <migraphx/op/normalize_attribute.hpp>
+#include <migraphx/float_equal.hpp>
 
 namespace migraphx {
 inline namespace MIGRAPHX_INLINE_NS {
@@ -38,11 +39,12 @@ namespace op {
 struct argmin
 {
     int64_t axis = 0;
+    bool select_last_index = false;
 
     template <class Self, class F>
     static auto reflect(Self& self, F f)
     {
-        return pack(f(self.axis, "axis"));
+        return pack(f(self.axis, "axis"), f(self.select_last_index, "select_last_index"));
     }
 
     value attributes() const
@@ -78,6 +80,10 @@ struct argmin
                 min_val   = cur_val;
                 min_index = i;
             }
+            else if(select_last_index and float_equal(min_val, cur_val))
+            {
+                min_index = i;
+            }
         }
 
         return min_index;

src/include/migraphx/op/convert.hpp

@@ -68,7 +68,7 @@ struct convert : unary<convert>
             auto y = x;
             shape::visit(type, [&](auto as) {
                 // clamping value between target_type's max and min doesn't work for NaNs,
-                if(std::isnan(x))
+                if(std::isnan(static_cast<double>(x)))
                 {
                     y = as.nan();
                 }

src/include/migraphx/op/isnan.hpp

@@ -35,7 +35,7 @@ struct isnan : unary<isnan>
 {
     auto apply() const
     {
-        return [](auto x) { return std::isnan(x); };
+        return [](auto x) { return std::isnan(static_cast<double>(x)); };
     }
 
     std::string name() const { return "isnan"; }

src/include/migraphx/op/nonmaxsuppression.hpp

@@ -24,6 +24,7 @@
 #ifndef MIGRAPHX_GUARD_OPERATORS_NONMAXSUPPRESSION_HPP
 #define MIGRAPHX_GUARD_OPERATORS_NONMAXSUPPRESSION_HPP
 
+#include <array>
 #include <cmath>
 #include <queue>
 #include <cstdint>

src/include/migraphx/op/pooling.hpp

@@ -411,7 +411,7 @@ struct pooling
             // for dynamic GlobalPooling, there's no padding
             kernel_dims.insert(kernel_dims.end(), input_lens.begin() + 2, input_lens.end());
             output_shape = dyn_out.computed_shape;
-            result       = dyn_out.computed_shape;
+            result       = argument{dyn_out.computed_shape};
         }
         else if((padding_mode != op::padding_mode_t::default_))
         {
@@ -439,7 +439,7 @@ struct pooling
         {
             kernel_dims  = this->lengths;
             output_shape = dyn_out.computed_shape;
-            result       = dyn_out.computed_shape;
+            result       = argument{dyn_out.computed_shape};
         }
 
         // Perform the computation and populate result

src/include/migraphx/op/random_uniform.hpp

@@ -77,11 +77,26 @@ struct random_uniform
             using type = typename decltype(output)::value_type;
             if constexpr(std::is_integral<type>{})
             {
-                // default range for all integer types is
-                // (0, std::uniform_int_distribution<type>::max()).
-                // Todo:  enable different ranges
-                std::uniform_int_distribution<type> dis;
-                std::generate(output.begin(), output.end(), [&] { return dis(gen); });
+#ifdef _MSC_VER
+                // According to the C++ specification, the effect is undefined if the result type
+                // for the generator is not one of short, int, long, long long, unsigned short,
+                // unsigned int, unsigned long, or unsigned long long. See
+                // https://en.cppreference.com/w/cpp/numeric/random/uniform_int_distribution.
+                if constexpr(sizeof(type) == 1)
+                {
+                    std::uniform_int_distribution<int> dis{std::numeric_limits<type>::min(),
+                                                           std::numeric_limits<type>::max()};
+                    std::generate(output.begin(), output.end(), [&] { return dis(gen); });
+                }
+                else
+#endif
+                {
+                    // default range for all integer types is
+                    // (0, std::uniform_int_distribution<type>::max()).
+                    // Todo:  enable different ranges
+                    std::uniform_int_distribution<type> dis;
+                    std::generate(output.begin(), output.end(), [&] { return dis(gen); });
+                }
             }
             else
             {

src/include/migraphx/op/reshape.hpp

@@ -36,6 +36,22 @@ namespace migraphx {
 inline namespace MIGRAPHX_INLINE_NS {
 namespace op {
 
+/**
+ * 1 input version:
+ * reshape(input_data)
+ * this.dims = output_dims
+ * Makes a copy of input_data to the output shape.
+ *
+ * 2 input version:
+ * reshape(input_data, output_buffer)
+ * this.dims = unset
+ * Copies input_data to output_buffer; output_buffer already has the output shape.
+ * This version will not fail gracefully if the input shape and output_buffer shape are
+ * incompatible. There's a throw that will catch when the number of elements do not match at
+ * runtime. This version should only be used for dynamic reshapes (output dimensions only known at
+ * runtime). If output_buffer has a static shape during compile/parse, you can use the 1 input
+ * version.
+ */
 struct reshape
 {
     std::vector<int64_t> dims;
@@ -215,32 +231,56 @@ struct reshape
 
     shape compute_shape(std::vector<shape> inputs) const
     {
-        check_shapes{inputs, *this, true}.has(1);
+        check_shapes{inputs, *this, true}.has(1, 2);
 
         auto n_neg_dims = std::count(dims.begin(), dims.end(), -1);
         if(n_neg_dims > 1)
             MIGRAPHX_THROW("reshape: Dimensions for reshape can only have one -1 dim");
 
         auto s0 = inputs.front();
-        if(s0.dynamic())
+        if(inputs.size() == 1)
         {
-            return dyn_compute_shape(s0);
+            if(s0.dynamic())
+            {
+                return dyn_compute_shape(s0);
+            }
+            else
+            {
+                return static_compute_shape(inputs, n_neg_dims);
+            }
         }
         else
         {
-            return static_compute_shape(inputs, n_neg_dims);
+            return inputs.back();
         }
     }
 
     argument compute(const dyn_output& dyn_out, std::vector<argument> args) const
     {
         assert(dyn_out.computed_shape.standard());
-        argument result{dyn_out.computed_shape};
+        if(args.size() == 1)
+        {
+            argument result{dyn_out.computed_shape};
 
-        visit_all(result, args[0])([&](auto output, auto input) {
-            std::copy(input.begin(), input.end(), output.begin());
-        });
-        return result;
+            visit_all(result, args[0])([&](auto output, auto input) {
+                std::copy(input.begin(), input.end(), output.begin());
+            });
+            return result;
+        }
+        else
+        {
+            // 2 arg
+            if(args[0].get_shape().elements() != args[1].get_shape().elements())
+            {
+                MIGRAPHX_THROW("Reshape: Number of elements must match at runtime. Input: " +
+                               std::to_string(args[0].get_shape().elements()) +
+                               " Output buffer: " + std::to_string(args[1].get_shape().elements()));
+            }
+            visit_all(args[1], args[0])([&](auto output, auto input) {
+                std::copy(input.begin(), input.end(), output.begin());
+            });
+            return args[1];
+        }
     }
 };
 

src/include/migraphx/op/roialign.hpp

@@ -33,6 +33,7 @@
 #include <migraphx/dfor.hpp>
 #include <migraphx/ranges.hpp>
 #include <migraphx/shape_for_each.hpp>
+#include <array>
 #include <cmath>
 #include <numeric>
 #include <utility>