Merge pull request #2019 from ROCmSoftwarePlatform/rel57_workitems

src/include/migraphx/allocation_model.hpp

@@ -96,7 +96,7 @@ struct allocation_model
     {
         using std::swap;
         auto* derived = this->any_cast<PrivateDetailTypeErasedT>();
-        if(derived and private_detail_te_handle_mem_var.unique())
+        if(derived and private_detail_te_handle_mem_var.use_count() == 1)
         {
             *derived = std::forward<PrivateDetailTypeErasedT>(value);
         }
@@ -267,7 +267,7 @@ struct allocation_model
     private_detail_te_handle_base_type& private_detail_te_get_handle()
     {
         assert(private_detail_te_handle_mem_var != nullptr);
-        if(not private_detail_te_handle_mem_var.unique())
+        if(private_detail_te_handle_mem_var.use_count() > 1)
             private_detail_te_handle_mem_var = private_detail_te_handle_mem_var->clone();
         return *private_detail_te_handle_mem_var;
     }

src/include/migraphx/builtin.hpp

@@ -90,7 +90,17 @@ struct param
 struct returns
 {
     std::string name() const { return "@return"; }
-    shape compute_shape(const std::vector<shape>&) const { return {}; }
+
+    shape compute_shape(const std::vector<shape>& arg) const
+    {
+        if(arg.empty())
+            return {};
+        else if(arg.size() == 1)
+            return arg[0];
+        else
+            return arg;
+    }
+
     argument compute(context&, const shape&, const std::vector<argument>&) const
     {
         MIGRAPHX_THROW("builtin");

src/include/migraphx/concat_opt.hpp

@@ -88,7 +88,7 @@ struct concat_optimization
     {
         using std::swap;
         auto* derived = this->any_cast<PrivateDetailTypeErasedT>();
-        if(derived and private_detail_te_handle_mem_var.unique())
+        if(derived and private_detail_te_handle_mem_var.use_count() == 1)
         {
             *derived = std::forward<PrivateDetailTypeErasedT>(value);
         }
@@ -233,7 +233,7 @@ struct concat_optimization
     private_detail_te_handle_base_type& private_detail_te_get_handle()
     {
         assert(private_detail_te_handle_mem_var != nullptr);
-        if(not private_detail_te_handle_mem_var.unique())
+        if(private_detail_te_handle_mem_var.use_count() > 1)
             private_detail_te_handle_mem_var = private_detail_te_handle_mem_var->clone();
         return *private_detail_te_handle_mem_var;
     }

src/include/migraphx/context.hpp

@@ -118,7 +118,7 @@ struct context
     {
         using std::swap;
         auto* derived = this->any_cast<PrivateDetailTypeErasedT>();
-        if(derived and private_detail_te_handle_mem_var.unique())
+        if(derived and private_detail_te_handle_mem_var.use_count() == 1)
         {
             *derived = std::forward<PrivateDetailTypeErasedT>(value);
         }
@@ -373,7 +373,7 @@ struct context
     private_detail_te_handle_base_type& private_detail_te_get_handle()
     {
         assert(private_detail_te_handle_mem_var != nullptr);
-        if(not private_detail_te_handle_mem_var.unique())
+        if(private_detail_te_handle_mem_var.use_count() > 1)
             private_detail_te_handle_mem_var = private_detail_te_handle_mem_var->clone();
         return *private_detail_te_handle_mem_var;
     }

src/include/migraphx/dynamic_loader.hpp

@@ -26,6 +26,7 @@
 
 #include <migraphx/config.hpp>
 #include <migraphx/filesystem.hpp>
+#include <migraphx/optional.hpp>
 #include <functional>
 #include <memory>
 #include <vector>
@@ -43,6 +44,9 @@ struct MIGRAPHX_EXPORT dynamic_loader
         return path(reinterpret_cast<void*>(address));
     }
     static fs::path path(void* address);
+
+    static optional<dynamic_loader> try_load(const fs::path& p);
+
     dynamic_loader() = default;
 
     dynamic_loader(const fs::path& p);

src/include/migraphx/generate.hpp

@@ -117,14 +117,14 @@ auto generate_tensor_data(const migraphx::shape& s, unsigned long seed = 0)
 }
 
 template <class T>
-auto fill_tensor_data(const migraphx::shape& s, unsigned long value = 0)
+auto fill_tensor_data(const migraphx::shape& s, double value = 0)
 {
     auto result = make_shared_array<T>(s.element_space());
     std::generate(result.get(), result.get() + s.element_space(), [=] { return value; });
     return result;
 }
 
-MIGRAPHX_EXPORT argument fill_argument(shape s, unsigned long value = 0);
+MIGRAPHX_EXPORT argument fill_argument(shape s, double value = 0);
 
 MIGRAPHX_EXPORT argument generate_argument(shape s, unsigned long seed = 0);
 

src/include/migraphx/marker.hpp

@@ -80,7 +80,7 @@ struct marker
     {
         using std::swap;
         auto* derived = this->any_cast<PrivateDetailTypeErasedT>();
-        if(derived and private_detail_te_handle_mem_var.unique())
+        if(derived and private_detail_te_handle_mem_var.use_count() == 1)
         {
             *derived = std::forward<PrivateDetailTypeErasedT>(value);
         }
@@ -233,7 +233,7 @@ struct marker
     private_detail_te_handle_base_type& private_detail_te_get_handle()
     {
         assert(private_detail_te_handle_mem_var != nullptr);
-        if(not private_detail_te_handle_mem_var.unique())
+        if(private_detail_te_handle_mem_var.use_count() > 1)
             private_detail_te_handle_mem_var = private_detail_te_handle_mem_var->clone();
         return *private_detail_te_handle_mem_var;
     }

src/include/migraphx/op/convert.hpp

@@ -66,7 +66,19 @@ struct convert : unary<convert>
         auto type = target_type;
         return [type](auto x) {
             auto y = x;
-            shape::visit(type, [&](auto as) { y = as(x); });
+            shape::visit(type, [&](auto as) {
+                // clamping value between target_type's max and min doesn't work for NaNs,
+                if(std::isnan(x))
+                {
+                    y = as.nan();
+                }
+                else
+                {
+                    // clamp overflowing/underflowing values to min()/max() instead of +/-infinity
+                    // during downcasting
+                    y = std::min(std::max(as(x), as.min()), as.max());
+                }
+            });
             return y;
         };
     }

src/include/migraphx/op/dimensions_of.hpp

+/*
+ * The MIT License (MIT)
+ *
+ * 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
+ * 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_OPERATORS_DIMENSIONS_OF_HPP
+#define MIGRAPHX_GUARD_OPERATORS_DIMENSIONS_OF_HPP
+
+#include <migraphx/check_shapes.hpp>
+#include <migraphx/argument.hpp>
+#include <migraphx/dyn_output.hpp>
+
+namespace migraphx {
+inline namespace MIGRAPHX_INLINE_NS {
+namespace op {
+
+/**
+ * Returns the dimensions of the input argument from starting axis to ending axis.
+ * Atleast `end` must be set to use this operator (set `end` to ndim for default ONNX behavior of
+ * `Shape` operator) This should only be used for dynamic shapes as this can be simplified to a
+ * literal for static shapes.
+ */
+struct dimensions_of
+{
+    std::size_t start = 0;
+    std::size_t end   = 0;
+
+    template <class Self, class F>
+    static auto reflect(Self& self, F f)
+    {
+        return pack(f(self.start, "start"), f(self.end, "end"));
+    }
+
+    std::string name() const { return "dimensions_of"; }
+
+    shape compute_shape(const std::vector<shape>& inputs) const
+    {
+        check_shapes{inputs, *this, true}.has(1);
+        if(start >= end)
+        {
+            MIGRAPHX_THROW("DIMENSIONS_OF: start >= end. start = " + std::to_string(start) +
+                           ", end = " + std::to_string(end));
+        }
+        return shape{shape::int64_type, {end - start}};
+    }
+
+    argument compute(const shape& output_shape, std::vector<argument> args) const
+    {
+        argument result{output_shape};
+        auto input_lens = args[0].get_shape().lens();
+        result.visit([&](auto output) {
+            std::copy(input_lens.cbegin() + start, input_lens.cbegin() + end, output.begin());
+        });
+        return result;
+    }
+};
+
+} // namespace op
+} // namespace MIGRAPHX_INLINE_NS
+} // namespace migraphx
+
+#endif

src/include/migraphx/op/multibroadcast.hpp

@@ -69,7 +69,7 @@ struct multibroadcast
 
         auto make_bcast_strides = [&](std::vector<std::size_t> bcast_lens, std::size_t offset) {
             std::vector<size_t> bcast_strides(bcast_lens.size(), 0);
-            for(std::ptrdiff_t i = s0.lens().size() - 1; i >= 0; i--)
+            for(std::ptrdiff_t i = s0.ndim() - 1; i >= 0; i--)
             {
                 if(bcast_lens[i + offset] == s0.lens()[i])
                 {
@@ -84,13 +84,13 @@ struct multibroadcast
             if(s0.dynamic())
                 MIGRAPHX_THROW(
                     "MULTIBROADCAST: Single dynamic input shape not supported.  Use two inputs.");
-            if(s0.lens().size() > output_lens.size())
+            if(s0.ndim() > output_lens.size())
             {
                 MIGRAPHX_THROW("MULTIBROADCAST: input dimensions should <= output size");
             }
 
-            auto offset = output_lens.size() - s0.lens().size();
-            for(std::ptrdiff_t i = s0.lens().size() - 1; i >= 0; i--)
+            auto offset = output_lens.size() - s0.ndim();
+            for(std::ptrdiff_t i = s0.ndim() - 1; i >= 0; i--)
             {
                 if(output_lens[i + offset] != s0.lens()[i] and s0.lens()[i] != 1)
                 {
@@ -119,7 +119,7 @@ struct multibroadcast
             {
                 // output_lens will not be set for 2+ input version
                 auto bcast_lens    = compute_common_lens(inputs);
-                auto offset        = bcast_lens.size() - s0.lens().size();
+                auto offset        = bcast_lens.size() - s0.ndim();
                 auto bcast_strides = make_bcast_strides(bcast_lens, offset);
                 return {t, std::move(bcast_lens), std::move(bcast_strides)};
             }

src/include/migraphx/op/prefix_scan_op.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
@@ -21,6 +21,7 @@
  * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN
  * THE SOFTWARE.
  */
+
 #ifndef MIGRAPHX_GUARD_OPERATORS_SCAN_OP_HPP
 #define MIGRAPHX_GUARD_OPERATORS_SCAN_OP_HPP
 
@@ -37,6 +38,12 @@ namespace migraphx {
 inline namespace MIGRAPHX_INLINE_NS {
 namespace op {
 
+/**
+ * Parent struct for prefix scan operations.  A prefix scan is equivalent to the C++
+ * std::exclusive_scan or std::inclusive_scan.  Given a list of numbers, a prefix scan
+ * sum op returns an equal size list of running totals of the values.  Other operations
+ * besides addition can be supported by their own child ops.
+ */
 template <class Derived>
 struct prefix_scan_op : op_name<Derived>
 {
@@ -60,9 +67,13 @@ struct prefix_scan_op : op_name<Derived>
 
     shape normalize_compute_shape(std::vector<shape> inputs) const
     {
-        check_shapes{inputs, *this}.has(1);
+        check_shapes{inputs, *this, true}.has(1);
         auto s = inputs.front();
-        if(s.broadcasted())
+        if(s.dynamic())
+        {
+            return s;
+        }
+        else if(s.broadcasted())
         {
             return {s.type(), s.lens()};
         }
@@ -72,8 +83,9 @@ struct prefix_scan_op : op_name<Derived>
         }
     }
 
-    argument compute(const shape& output_shape, std::vector<argument> args) const
+    argument compute(const dyn_output& dyn_out, std::vector<argument> args) const
     {
+        shape output_shape(dyn_out.computed_shape);
         argument result{output_shape};
         auto s = args[0].get_shape();
         if(s == output_shape)

src/include/migraphx/operation.hpp

@@ -251,9 +251,10 @@ auto compute_op(rank<1>,
                 const shape& output,
                 const std::vector<argument>& inputs,
                 const std::vector<module_ref>& module_args,
-                F f)
-    -> decltype(
-        x.compute(make_compute_output_shape(pack(x, output, inputs)), inputs, module_args, f))
+                F f) -> decltype(x.compute(make_compute_output_shape(pack(x, output, inputs)),
+                                           inputs,
+                                           module_args,
+                                           f))
 {
     return x.compute(make_compute_output_shape(pack(x, output, inputs)), inputs, module_args, f);
 }
@@ -309,9 +310,10 @@ auto compute_op(rank<3>,
                 const shape& output,
                 const std::vector<argument>& inputs,
                 const std::vector<module_ref>& module_args,
-                F f)
-    -> decltype(
-        x.compute(make_compute_output_shape(pack(x, output, inputs)), inputs, module_args, f))
+                F f) -> decltype(x.compute(make_compute_output_shape(pack(x, output, inputs)),
+                                           inputs,
+                                           module_args,
+                                           f))
 {
     return x.compute(make_compute_output_shape(pack(x, output, inputs)), inputs, module_args, f);
 }
@@ -573,7 +575,7 @@ struct operation
     {
         using std::swap;
         auto* derived = this->any_cast<PrivateDetailTypeErasedT>();
-        if(derived and private_detail_te_handle_mem_var.unique())
+        if(derived and private_detail_te_handle_mem_var.use_count() == 1)
         {
             *derived = std::forward<PrivateDetailTypeErasedT>(value);
         }
@@ -1263,7 +1265,7 @@ struct operation
     private_detail_te_handle_base_type& private_detail_te_get_handle()
     {
         assert(private_detail_te_handle_mem_var != nullptr);
-        if(not private_detail_te_handle_mem_var.unique())
+        if(private_detail_te_handle_mem_var.use_count() > 1)
             private_detail_te_handle_mem_var = private_detail_te_handle_mem_var->clone();
         return *private_detail_te_handle_mem_var;
     }

src/include/migraphx/operators.hpp

@@ -48,6 +48,7 @@
 #include <migraphx/op/convolution_backwards.hpp>
 #include <migraphx/op/cosh.hpp>
 #include <migraphx/op/cos.hpp>
+#include <migraphx/op/dimensions_of.hpp>
 #include <migraphx/op/div.hpp>
 #include <migraphx/op/dot.hpp>
 #include <migraphx/op/elu.hpp>

src/include/migraphx/pass.hpp

@@ -116,7 +116,7 @@ struct pass
     {
         using std::swap;
         auto* derived = this->any_cast<PrivateDetailTypeErasedT>();
-        if(derived and private_detail_te_handle_mem_var.unique())
+        if(derived and private_detail_te_handle_mem_var.use_count() == 1)
         {
             *derived = std::forward<PrivateDetailTypeErasedT>(value);
         }
@@ -292,7 +292,7 @@ struct pass
     private_detail_te_handle_base_type& private_detail_te_get_handle()
     {
         assert(private_detail_te_handle_mem_var != nullptr);
-        if(not private_detail_te_handle_mem_var.unique())
+        if(private_detail_te_handle_mem_var.use_count() > 1)
             private_detail_te_handle_mem_var = private_detail_te_handle_mem_var->clone();
         return *private_detail_te_handle_mem_var;
     }

src/include/migraphx/permutation.hpp

@@ -66,6 +66,10 @@ MIGRAPHX_EXPORT std::vector<int64_t> invert_permutation(const std::vector<int64_
 MIGRAPHX_EXPORT std::vector<int64_t> find_permutation(const shape& s);
 MIGRAPHX_EXPORT std::vector<int64_t> find_permutation(const std::vector<shape>& shapes);
 
+/// Normalize the shapes so the order of dimensions will be in the order it is
+/// in memory as much as possible.
+MIGRAPHX_EXPORT std::vector<shape> normalize_permutation(const std::vector<shape>& shapes);
+
 } // namespace MIGRAPHX_INLINE_NS
 } // namespace migraphx
 

src/include/migraphx/schedule_model.hpp

@@ -99,7 +99,7 @@ struct schedule_model
     {
         using std::swap;
         auto* derived = this->any_cast<PrivateDetailTypeErasedT>();
-        if(derived and private_detail_te_handle_mem_var.unique())
+        if(derived and private_detail_te_handle_mem_var.use_count() == 1)
         {
             *derived = std::forward<PrivateDetailTypeErasedT>(value);
         }
@@ -274,7 +274,7 @@ struct schedule_model
     private_detail_te_handle_base_type& private_detail_te_get_handle()
     {
         assert(private_detail_te_handle_mem_var != nullptr);
-        if(not private_detail_te_handle_mem_var.unique())
+        if(private_detail_te_handle_mem_var.use_count() > 1)
             private_detail_te_handle_mem_var = private_detail_te_handle_mem_var->clone();
         return *private_detail_te_handle_mem_var;
     }

src/include/migraphx/shape.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
@@ -240,6 +240,10 @@ struct MIGRAPHX_EXPORT shape
     template <class Iterator>
     std::size_t index(Iterator start, Iterator last) const
     {
+        if(this->dynamic())
+        {
+            MIGRAPHX_THROW("SHAPE: index() called on dynamic shape");
+        }
         assert(std::distance(start, last) <= this->lens().size());
         assert(this->lens().size() == this->strides().size());
         return std::inner_product(start, last, this->strides().begin(), std::size_t{0}); // NOLINT

src/include/migraphx/stream_model.hpp

@@ -100,7 +100,7 @@ struct stream_model
     {
         using std::swap;
         auto* derived = this->any_cast<PrivateDetailTypeErasedT>();
-        if(derived and private_detail_te_handle_mem_var.unique())
+        if(derived and private_detail_te_handle_mem_var.use_count() == 1)
         {
             *derived = std::forward<PrivateDetailTypeErasedT>(value);
         }
@@ -288,7 +288,7 @@ struct stream_model
     private_detail_te_handle_base_type& private_detail_te_get_handle()
     {
         assert(private_detail_te_handle_mem_var != nullptr);
-        if(not private_detail_te_handle_mem_var.unique())
+        if(private_detail_te_handle_mem_var.use_count() > 1)
             private_detail_te_handle_mem_var = private_detail_te_handle_mem_var->clone();
         return *private_detail_te_handle_mem_var;
     }

src/include/migraphx/target.hpp

@@ -167,7 +167,7 @@ struct target
     {
         using std::swap;
         auto* derived = this->any_cast<PrivateDetailTypeErasedT>();
-        if(derived and private_detail_te_handle_mem_var.unique())
+        if(derived and private_detail_te_handle_mem_var.use_count() == 1)
         {
             *derived = std::forward<PrivateDetailTypeErasedT>(value);
         }
@@ -428,7 +428,7 @@ struct target
     private_detail_te_handle_base_type& private_detail_te_get_handle()
     {
         assert(private_detail_te_handle_mem_var != nullptr);
-        if(not private_detail_te_handle_mem_var.unique())
+        if(private_detail_te_handle_mem_var.use_count() > 1)
             private_detail_te_handle_mem_var = private_detail_te_handle_mem_var->clone();
         return *private_detail_te_handle_mem_var;
     }

src/instruction.cpp

@@ -64,10 +64,7 @@ void instruction::replace(const shape& r)
         result = r;
         for(auto&& ins : output)
         {
-            if(ins->name() == "@return")
-                continue;
-
-            assert(ins->name().front() != '@');
+            assert(ins->name() == "@return" or ins->name().front() != '@');
             ins->recompute_shape();
         }
     }
@@ -122,10 +119,6 @@ bool instruction::valid() const
     {
         computed = result;
     }
-    else if(op.name() == "@return")
-    {
-        computed = {};
-    }
     else
     {
         try
@@ -145,6 +138,7 @@ bool instruction::valid() const
 }
 
 shape instruction::get_shape() const { return result; }
+
 const literal& instruction::get_literal() const
 {
     assert(op.name() == "@literal");