Merge branch 'develop' of https://github.com/ROCmSoftwarePlatform/AMDMIGraphX into rnn_optimization

CMakeLists.txt

@@ -39,8 +39,6 @@ else()
     set(MIGRAPHX_ENABLE_GPU Off CACHE BOOL "")
 endif()
 
-set(MIGRAPHX_ENABLE_TF Off CACHE BOOL "")
-
 add_compile_options(-std=c++14)
 
 list(APPEND CMAKE_MODULE_PATH ${CMAKE_CURRENT_SOURCE_DIR}/cmake)

Dockerfile

@@ -44,8 +44,7 @@ RUN apt-get update && DEBIAN_FRONTEND=noninteractive apt-get install -y --allow-
     rm -rf /var/lib/apt/lists/*
 
 # Install cget
-# RUN pip install cget
-RUN pip install https://github.com/pfultz2/cget/archive/57b3289000fcdb3b7e424c60a35ea09bc44d8538.tar.gz
+RUN pip install cget
 
 # Install rclone
 RUN pip install https://github.com/pfultz2/rclone/archive/master.tar.gz
@@ -74,3 +73,8 @@ ENV LD_LIBRARY_PATH=$PREFIX/lib
 # Install doc requirements
 ADD doc/requirements.txt /doc-requirements.txt
 RUN pip install -r /doc-requirements.txt
+
+# Setup ubsan environment to printstacktrace
+RUN ln -s /usr/bin/llvm-symbolizer-5.0 /usr/local/bin/llvm-symbolizer
+ENV UBSAN_OPTIONS=print_stacktrace=1
+ENV ASAN_OPTIONS=detect_stack_use_after_return=1:check_initialization_order=1:strict_init_order=1

requirements.txt

-google/protobuf -DCMAKE_POSITION_INDEPENDENT_CODE=On
+google/protobuf@v3.8.0 -DCMAKE_POSITION_INDEPENDENT_CODE=On -X subdir -Dprotobuf_BUILD_TESTS=Off
 RadeonOpenCompute/rocm-cmake@42f6740 --build
 ROCmSoftwarePlatform/rocBLAS@30a992ae02fda568688bcd190edd5e277d6674d9
 ROCmSoftwarePlatform/MIOpen@1.7.0

src/CMakeLists.txt

@@ -18,6 +18,7 @@ add_library(migraphx
     generate.cpp
     instruction.cpp
     program.cpp
+    quantization.cpp
     shape.cpp
     schedule.cpp
     pass_manager.cpp

src/eliminate_contiguous.cpp

@@ -67,13 +67,6 @@ void eliminate_contiguous::apply(program& p) const
 {
     for(auto ins : iterator_for(p))
     {
-        // skip the reshape operator for now, since there is a bug
-        // for the transpose followed by a reshape
-        if(ins->name() == "reshape")
-        {
-            continue;
-        }
-
         // Make a copy so we can modify it while we iterate
         auto args = ins->inputs();
         for(auto arg : ins->inputs())

src/env.cpp

@@ -21,6 +21,14 @@ bool disabled(const char* name)
     return contains({"0", "disable", "disabled", "no", "false"}, e.front());
 }
 
+std::size_t value_of(const char* name)
+{
+    auto e = env(name);
+    if(e.empty())
+        return 0;
+    return std::stoul(e.front());
+}
+
 std::vector<std::string> env(const char* name)
 {
     auto p = std::getenv(name);

src/include/migraphx/check_shapes.hpp

@@ -103,6 +103,13 @@ struct check_shapes
         return *this;
     }
 
+    const check_shapes& standard_or_scalar() const
+    {
+        if(!this->all_of([](const shape& s) { return s.standard() or s.scalar(); }))
+            MIGRAPHX_THROW(prefix() + "Shapes are not a scalar or in standard layout");
+        return *this;
+    }
+
     const check_shapes& packed() const
     {
         if(!this->all_of([](const shape& s) { return s.packed(); }))

src/include/migraphx/env.hpp

@@ -19,6 +19,8 @@ bool enabled(const char* name);
 bool disabled(const char* name);
 std::vector<std::string> env(const char* name);
 
+std::size_t value_of(const char* name);
+
 template <class T>
 bool enabled(T)
 {
@@ -33,6 +35,13 @@ bool disabled(T)
     return result;
 }
 
+template <class T>
+std::size_t value_of(T)
+{
+    static const std::size_t result = value_of(T::value());
+    return result;
+}
+
 } // namespace MIGRAPHX_INLINE_NS
 } // namespace migraphx
 

src/include/migraphx/instruction.hpp

@@ -24,7 +24,7 @@ struct instruction
 
     instruction(literal l);
 
-    void replace(const shape& r);
+    void replace(operation o);
 
     void recompute_shape();
 
@@ -90,7 +90,8 @@ struct instruction
     // internal
     void replace_argument(instruction_ref old, instruction_ref new_ins);
 
-    private:
+    void replace(const shape& r);
+
     operation op;
     shape result;
     std::vector<instruction_ref> output;

src/include/migraphx/op/abnormal_ops.hpp

@@ -39,6 +39,11 @@ struct undefined
 struct unknown
 {
     std::string op;
+    template <class Self, class F>
+    static auto reflect(Self& self, F f)
+    {
+        return pack(f(self.op, "op"));
+    }
     std::string name() const { return "unknown:" + op; }
     shape compute_shape(std::vector<shape> input) const
     {

src/include/migraphx/op/binary.hpp

@@ -24,11 +24,12 @@ struct binary : op_name<Derived>
             return {s0.type(), s0.lens()};
         }
     }
+
     argument compute(const shape& output_shape, std::vector<argument> args) const
     {
         argument result{output_shape};
         visit_all(result, args[0], args[1])([&](auto output, auto input1, auto input2) {
-            if(input1.get_shape().standard() and input2.get_shape().standard())
+            if(input1.get_shape().packed() and input2.get_shape().packed())
             {
                 std::transform(input1.begin(),
                                input1.end(),
@@ -44,6 +45,7 @@ struct binary : op_name<Derived>
                 });
             }
         });
+
         return result;
     }
 };

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/op/concat.hpp

@@ -19,6 +19,13 @@ namespace op {
 struct concat
 {
     std::size_t axis = 0;
+
+    template <class Self, class F>
+    static auto reflect(Self& self, F f)
+    {
+        return pack(f(self.axis, "axis"));
+    }
+
     std::string name() const { return "concat"; }
     std::vector<std::size_t> compute_offsets(const shape& output_shape,
                                              const std::vector<argument>& args) const

src/include/migraphx/op/convert.hpp

+#ifndef MIGRAPHX_GUARD_OPERATORS_CONVERT_HPP
+#define MIGRAPHX_GUARD_OPERATORS_CONVERT_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>
+
+namespace migraphx {
+inline namespace MIGRAPHX_INLINE_NS {
+namespace op {
+
+struct convert : unary<convert>
+{
+    shape::type_t target_type = shape::half_type;
+
+    template <class Self, class F>
+    static auto reflect(Self& self, F f)
+    {
+        return pack(f(self.target_type, "target_type"));
+    }
+
+    shape compute_shape(std::vector<shape> inputs) const
+    {
+        check_shapes{inputs, *this}.has(1);
+        return {target_type, inputs.at(0).lens(), inputs.at(0).strides()};
+    }
+
+    auto apply() const
+    {
+        return [](auto x) { return x; };
+    }
+
+    convert(shape::type_t t) : target_type{t} {}
+    convert() {}
+};
+
+} // namespace op
+} // namespace MIGRAPHX_INLINE_NS
+} // namespace migraphx
+
+#endif

src/include/migraphx/op/leaky_relu.hpp

@@ -18,19 +18,20 @@ namespace op {
 
 struct leaky_relu
 {
-    std::string name() const { return "leaky_relu"; }
     float alpha;
-    shape compute_shape(std::vector<shape> inputs) const
-    {
-        check_shapes{inputs, *this}.has(1);
-        return inputs.front();
-    }
 
     template <class Self, class F>
     static auto reflect(Self& self, F f)
     {
         return pack(f(self.alpha, "alpha"));
     }
+
+    std::string name() const { return "leaky_relu"; }
+    shape compute_shape(std::vector<shape> inputs) const
+    {
+        check_shapes{inputs, *this}.has(1);
+        return inputs.front();
+    }
 };
 
 } // namespace op

src/include/migraphx/op/logsoftmax.hpp

@@ -30,7 +30,7 @@ struct logsoftmax
     shape compute_shape(std::vector<shape> inputs) const
     {
         check_shapes{inputs}.has(1).standard();
-        if(axis < 0 || axis > inputs[0].lens().size())
+        if(axis < 0 || axis >= inputs[0].lens().size())
         {
             MIGRAPHX_THROW("LogSoftMax: input axis value " + std::to_string(axis) +
                            " is out of range");

src/include/migraphx/op/reshape.hpp

@@ -29,7 +29,7 @@ struct reshape
     std::string name() const { return "reshape"; }
     shape compute_shape(std::vector<shape> inputs) const
     {
-        check_shapes{inputs, *this}.has(1);
+        check_shapes{inputs, *this}.has(1).standard();
         auto&& idims = inputs.front().lens();
         std::vector<std::size_t> rdims(dims.begin(), dims.end());
         auto n_neg_dims = std::count(dims.begin(), dims.end(), -1);

src/include/migraphx/op/softmax.hpp

@@ -18,10 +18,23 @@ namespace op {
 
 struct softmax
 {
+    int axis = 1;
+
+    template <class Self, class F>
+    static auto reflect(Self& self, F f)
+    {
+        return pack(f(self.axis, "axis"));
+    }
+
     std::string name() const { return "softmax"; }
     shape compute_shape(std::vector<shape> inputs) const
     {
-        check_shapes{inputs}.has(1).only_dims(4);
+        check_shapes{inputs}.has(1).standard();
+        if(axis < 0 || axis >= inputs[0].lens().size())
+        {
+            MIGRAPHX_THROW("SoftMax: input axis value " + std::to_string(axis) +
+                           " is out of range");
+        }
         return inputs.at(0);
     }
 };

src/include/migraphx/op/squeeze.hpp

@@ -29,6 +29,7 @@ struct squeeze
     std::string name() const { return "squeeze"; }
     shape compute_shape(std::vector<shape> inputs) const
     {
+        check_shapes{inputs, *this}.has(1).standard();
         auto input_shape = inputs[0];
         auto type        = input_shape.type();
         auto old_lens    = input_shape.lens();

src/include/migraphx/op/unary.hpp

@@ -23,25 +23,31 @@ struct unary : op_name<Derived>
             return {s.type(), s.lens()};
         }
     }
+
     argument compute(const shape& output_shape, std::vector<argument> args) const
     {
         argument result{output_shape};
-        visit_all(result, args[0])([&](auto output, auto input) {
-            if(input.get_shape().standard())
-            {
-                std::transform(input.begin(),
-                               input.end(),
-                               output.begin(),
-                               static_cast<const Derived&>(*this).apply());
-            }
-            else
-            {
+        result.visit([&](auto output) {
+            args[0].visit([&](auto input) {
+                if(input.get_shape().packed())
+                {
+                    std::transform(input.begin(),
+                                   input.end(),
+                                   output.begin(),
+                                   static_cast<const Derived&>(*this).apply());
+
+                    return result;
+                }
+
                 shape_for_each(output.get_shape(), [&](const auto& idx) {
                     output(idx.begin(), idx.end()) =
                         static_cast<const Derived&>(*this).apply()(input(idx.begin(), idx.end()));
                 });
-            }
+
+                return result;
+            });
         });
+
         return result;
     }
 };