Merge

src/targets/gpu/kernels/include/migraphx/kernels/softmax.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_KERNELS_SOFTMAX_HPP
 #define MIGRAPHX_GUARD_KERNELS_SOFTMAX_HPP
 

src/targets/gpu/kernels/include/migraphx/kernels/type_traits.hpp

@@ -192,9 +192,13 @@ struct common_type<T, U, Us...>
 template <class... Ts>
 using common_type_t = typename common_type<Ts...>::type;
 
+#define MIGRAPHX_REQUIRES(...) class = enable_if_t<__VA_ARGS__>
+
 constexpr unsigned long int_max(unsigned long n) { return (1u << (n * 8)) - 1; }
 
-template <class T>
+template <class T,
+          MIGRAPHX_REQUIRES(is_integral<T>{} or is_floating_point<T>{} or
+                            is_same<T, migraphx::half>{})>
 constexpr T numeric_max()
 {
     if constexpr(is_integral<T>{})
@@ -230,8 +234,6 @@ constexpr T numeric_lowest()
     }
 }
 
-#define MIGRAPHX_REQUIRES(...) class = enable_if_t<__VA_ARGS__>
-
 } // namespace migraphx
 
 #endif

src/targets/gpu/kernels/include/migraphx/kernels/vec.hpp

@@ -175,7 +175,7 @@ template <class T, class Op>
 constexpr auto vec_reduce(T x, Op op)
 {
     if constexpr(vec_size<T>() < 2)
-        return x;
+        return vec_type<T>{x};
     else
     {
         vec_type<T> result = x[0];

src/targets/gpu/lowering.cpp

@@ -300,7 +300,7 @@ struct miopen_apply
             auto&& op = any_cast<op::deconvolution>(ins->get_operator());
 
             auto conv = miopen_deconvolution{op, make_deconv(op)};
-            auto ws   = conv.compile(get_context(), ins->get_shape(), to_shapes(ins->inputs()));
+            auto ws   = conv.find(get_context(), ins->get_shape(), to_shapes(ins->inputs()));
 
             auto workspace = insert_allocation(ins, ws);
             auto output    = insert_allocation(ins, ins->get_shape());
@@ -331,7 +331,7 @@ struct miopen_apply
             miopen_quant_convolution conv;
             auto compile_quant_conv_with_format = [&](bool format) {
                 conv = miopen_quant_convolution{op, format, make_conv(op)};
-                ws   = conv.compile(get_context(), ins->get_shape(), to_shapes(ins->inputs()));
+                ws   = conv.find(get_context(), ins->get_shape(), to_shapes(ins->inputs()));
             };
 
             try

src/targets/gpu/mlir.cpp

+/*
+ * 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.
+ */
+#include <migraphx/gpu/mlir.hpp>
+
+#ifdef MIGRAPHX_MLIR
+#include <mlir-c/IR.h>
+#include <mlir-c/BuiltinAttributes.h>
+#include <mlir-c/BuiltinTypes.h>
+#include <mlir-c/Diagnostics.h>
+#include <mlir-c/Dialect/MIGraphX.h>
+#include <mlir-c/IntegerSet.h>
+#include <mlir-c/Pass.h>
+#include <mlir-c/Registration.h>
+#endif
+
+#include <migraphx/env.hpp>
+#include <migraphx/manage_ptr.hpp>
+#include <migraphx/module.hpp>
+#include <migraphx/instruction.hpp>
+#include <migraphx/config.hpp>
+#include <migraphx/ranges.hpp>
+#include <migraphx/gpu/code_object_op.hpp>
+#include <migraphx/gpu/context.hpp>
+#include <migraphx/gpu/device_name.hpp>
+#include <migraphx/iterator_for.hpp>
+#include <migraphx/gpu/perfdb.hpp>
+#include <deque>
+#include <variant>
+
+#if defined(MLIR_MIGRAPHX_DIALECT_API_VERSION) && MLIR_MIGRAPHX_DIALECT_API_VERSION >= 2
+#define MIGRAPHX_MLIR_BARE_POINTER
+#endif
+
+namespace migraphx {
+inline namespace MIGRAPHX_INLINE_NS {
+namespace gpu {
+
+MIGRAPHX_DECLARE_ENV_VAR(MIGRAPHX_TRACE_MLIR);
+
+#ifdef MIGRAPHX_MLIR
+template <class T, class F, F f> // NOLINT
+struct mlir_handle
+{
+    struct ptr
+    {
+        ptr() = default;
+        ptr(std::nullptr_t) {}
+        ptr(T x) : obj(x) {}
+
+        std::intptr_t get_value() const
+        {
+            static_assert(sizeof(T) == sizeof(std::intptr_t), "MLIR Handle different size");
+            return reinterpret_cast<const std::intptr_t&>(obj);
+        }
+
+        T get() const { return obj; }
+
+        friend bool operator==(ptr x, ptr y) { return x.get_value() == y.get_value(); }
+
+        friend bool operator!=(ptr x, ptr y) { return !(x == y); }
+        T obj{};
+    };
+
+    struct deleter
+    {
+        using pointer = ptr;
+
+        void operator()(pointer x) const
+        {
+            if(x != nullptr)
+            {
+                (void)f(x.obj);
+            }
+        }
+    };
+
+    mlir_handle() : handle(nullptr) {}
+
+    mlir_handle(T p) : handle(ptr{p}) {}
+
+    T get() const { return handle.get().get(); }
+
+    T release() { return handle.release().get(); }
+
+    private:
+    std::unique_ptr<ptr, deleter> handle;
+};
+
+#define MIGRAPHX_MANAGE_MLIR_HANDLE(T, F) migraphx::gpu::mlir_handle<T, decltype(&F), &F> // NOLINT
+
+using mlir_context           = MIGRAPHX_MANAGE_MLIR_HANDLE(MlirContext, mlirContextDestroy);
+using mlir_module            = MIGRAPHX_MANAGE_MLIR_HANDLE(MlirModule, mlirModuleDestroy);
+using mlir_operation         = MIGRAPHX_MANAGE_MLIR_HANDLE(MlirOperation, mlirOperationDestroy);
+using mlir_op_printing_flags = MIGRAPHX_MANAGE_MLIR_HANDLE(MlirOpPrintingFlags,
+                                                           mlirOpPrintingFlagsDestroy);
+using mlir_region            = MIGRAPHX_MANAGE_MLIR_HANDLE(MlirRegion, mlirRegionDestroy);
+using mlir_block             = MIGRAPHX_MANAGE_MLIR_HANDLE(MlirBlock, mlirBlockDestroy);
+using mlir_pass_manager      = MIGRAPHX_MANAGE_MLIR_HANDLE(MlirPassManager, mlirPassManagerDestroy);
+
+std::string_view to_string_view(MlirStringRef s) { return {s.data, s.length}; }
+
+MlirStringRef make_mlir_string_ref(const std::string_view& s)
+{
+    return mlirStringRefCreate(s.data(), s.size());
+}
+
+template <class F, class T, class Printer>
+void mlir_print(F f, T x, Printer printer)
+{
+    f(
+        x,
+        +[](MlirStringRef s, void* data) {
+            (*reinterpret_cast<Printer*>(data))(to_string_view(s));
+        },
+        &printer);
+}
+
+template <class F, class T>
+void mlir_print(F f, T x, std::ostream& os)
+{
+    mlir_print(f, x, [&](auto s) { os << s; });
+}
+
+template <class F, class T>
+std::string mlir_print(F f, T x)
+{
+    std::stringstream ss;
+    mlir_print(f, x, [&](auto s) { ss << s; });
+    return ss.str();
+}
+
+const std::unordered_set<std::string>& get_xdlops_archs()
+{
+    static std::unordered_set<std::string> supported_archs{"gfx908", "gfx90a"};
+    return supported_archs;
+}
+
+struct mlir_program
+{
+    mlir_program()
+        : ctx(mlirContextCreate()),
+          location(mlirLocationUnknownGet(ctx.get())),
+          mmodule(mlirModuleCreateEmpty(location))
+    {
+        MlirDialectHandle mixr_handle = mlirGetDialectHandle__migraphx__();
+        mlirDialectHandleRegisterDialect(mixr_handle, ctx.get());
+        mlirRegisterAllDialects(ctx.get());
+        mlirContextSetAllowUnregisteredDialects(ctx.get(), true /*allow*/);
+    }
+
+    MlirType make_type(shape::type_t t) const
+    {
+        MlirType result;
+        shape::visit(t, [&](auto as) {
+            if(as.type_enum() == shape::float_type)
+                result = mlirF32TypeGet(ctx.get());
+            else if(as.type_enum() == shape::half_type)
+                result = mlirF16TypeGet(ctx.get());
+            else if(as.type_enum() == shape::double_type)
+                result = mlirF64TypeGet(ctx.get());
+            else if(as.is_integral())
+            {
+                if(as.is_signed())
+                    result = mlirIntegerTypeSignedGet(ctx.get(), as.size() * 8);
+                else
+                    result = mlirIntegerTypeGet(ctx.get(), as.size() * 8);
+            }
+            else
+                MIGRAPHX_THROW("Unsupported type: " + std::to_string(as.type_enum()));
+        });
+        return result;
+    }
+
+    MlirType make_tensor(const shape& s) const
+    {
+        assert(s.standard());
+        std::vector<int64_t> lens(s.lens().begin(), s.lens().end());
+        return mlirRankedTensorTypeGet(
+            lens.size(), lens.data(), make_type(s.type()), mlirAttributeGetNull());
+    }
+
+    template <class Range>
+    std::vector<MlirType> make_tensors(const Range& r)
+    {
+        std::vector<MlirType> result;
+        std::transform(r.begin(), r.end(), std::back_inserter(result), [&](const auto& s) {
+            return make_tensor(s);
+        });
+        return result;
+    }
+
+    MlirType make_function_type(const std::vector<shape>& inputs, const std::vector<shape>& outputs)
+    {
+        auto in  = make_tensors(inputs);
+        auto out = make_tensors(outputs);
+        return mlirFunctionTypeGet(ctx.get(), in.size(), in.data(), out.size(), out.data());
+    }
+
+    MlirIdentifier id(const std::string_view& s) const
+    {
+        return mlirIdentifierGet(ctx.get(), make_mlir_string_ref(s));
+    }
+
+    MlirAttribute attribute(std::int64_t i) const
+    {
+        if(i < 0)
+            MIGRAPHX_THROW("MLIR cant handle negative values since they are ambiguous");
+        return mlirIntegerAttrGet(mlirIntegerTypeGet(ctx.get(), 64), i);
+    }
+    MlirAttribute attribute(std::uint64_t i) const
+    {
+        if(i > (std::numeric_limits<std::uint64_t>::max() / 2))
+            MIGRAPHX_THROW("MLIR cant handle large integer values since they are ambiguous");
+        return mlirIntegerAttrGet(mlirIntegerTypeGet(ctx.get(), 64), i);
+    }
+    MlirAttribute attribute(unsigned char i) const { return attribute(std::uint64_t(i)); }
+    MlirAttribute attribute(bool b) const { return mlirBoolAttrGet(ctx.get(), b ? 1 : 0); }
+    MlirAttribute attribute(double d) const
+    {
+        return mlirFloatAttrDoubleGet(ctx.get(), mlirF64TypeGet(ctx.get()), d);
+    }
+    MlirAttribute attribute(const std::string& s) const
+    {
+        return mlirStringAttrGet(ctx.get(), make_mlir_string_ref(s));
+    }
+    MlirAttribute attribute(std::nullptr_t) const { return {}; }
+    template <class T>
+    MlirAttribute attribute(const std::vector<T>& v) const
+    {
+        std::vector<MlirAttribute> attributes;
+        attributes.reserve(v.size());
+        std::transform(v.begin(), v.end(), std::back_inserter(attributes), [&](auto&& x) {
+            return attribute(x);
+        });
+        return mlirArrayAttrGet(ctx.get(), attributes.size(), attributes.data());
+    }
+    MlirAttribute attribute(const value& v) const
+    {
+        MlirAttribute attr;
+        v.visit_value([&](auto&& x) { attr = attribute(x); });
+        return attr;
+    }
+    MlirAttribute attribute(const std::vector<value>& v) const
+    {
+        if(v.empty())
+        {
+            return mlirArrayAttrGet(ctx.get(), 0, nullptr);
+        }
+        if(not v.front().get_key().empty())
+        {
+            std::vector<MlirNamedAttribute> attributes = name_attributes(v);
+            return mlirDictionaryAttrGet(ctx.get(), attributes.size(), attributes.data());
+        }
+        else
+        {
+            std::vector<MlirAttribute> attributes;
+            attributes.reserve(v.size());
+            std::transform(v.begin(), v.end(), std::back_inserter(attributes), [&](auto&& x) {
+                return attribute(x);
+            });
+            return mlirArrayAttrGet(ctx.get(), attributes.size(), attributes.data());
+        }
+    }
+
+    MlirAttribute attribute(MlirType t) const { return mlirTypeAttrGet(t); }
+
+    MlirAttribute attribute(MlirAttribute a) const { return a; }
+
+    template <class T>
+    MlirNamedAttribute name_attribute(const std::string_view& key, const T& x) const
+    {
+        MlirNamedAttribute attr;
+        attr.name      = id(key);
+        attr.attribute = attribute(x);
+        return attr;
+    }
+
+    using attribute_t       = std::variant<std::nullptr_t,
+                                     std::uint64_t,
+                                     unsigned char,
+                                     bool,
+                                     double,
+                                     std::string,
+                                     value,
+                                     std::vector<value>,
+                                     MlirType>;
+    using named_attribute_t = std::pair<std::string_view, attribute_t>;
+
+    MlirNamedAttribute name_attribute(const named_attribute_t& na) const
+    {
+        return name_attribute(na.first,
+                              std::visit([&](const auto& x) { return attribute(x); }, na.second));
+    }
+
+    std::vector<MlirNamedAttribute>
+    name_attributes(const std::vector<named_attribute_t>& named_attrs) const
+    {
+        std::vector<MlirNamedAttribute> attributes;
+        attributes.reserve(named_attrs.size());
+        std::transform(named_attrs.begin(),
+                       named_attrs.end(),
+                       std::back_inserter(attributes),
+                       [&](const named_attribute_t& a) { return name_attribute(a); });
+        return attributes;
+    }
+
+    std::vector<MlirNamedAttribute> name_attributes(const value& v) const
+    {
+        std::vector<MlirNamedAttribute> attributes;
+        attributes.reserve(v.size());
+        std::transform(v.begin(), v.end(), std::back_inserter(attributes), [&](const value& x) {
+            return name_attribute(x.get_key(), x.without_key());
+        });
+        return attributes;
+    }
+
+    struct mlir_operation_state
+    {
+        mlir_operation_state(mlir_program& p, const std::string_view& name)
+            : prog(&p), op_state(mlirOperationStateGet(make_mlir_string_ref(name), p.location))
+        {
+        }
+
+        mlir_operation_state& add_attributes(const std::vector<named_attribute_t>& named_attrs)
+        {
+            auto attributes = prog->name_attributes(named_attrs);
+            mlirOperationStateAddAttributes(&op_state, attributes.size(), attributes.data());
+            return *this;
+        }
+
+        mlir_operation_state& add_attribute_value(const value& v)
+        {
+            auto attributes = prog->name_attributes(v);
+            mlirOperationStateAddAttributes(&op_state, attributes.size(), attributes.data());
+            return *this;
+        }
+
+        mlir_operation_state& add_regions(std::vector<mlir_region> rs)
+        {
+            regions = std::move(rs);
+            return *this;
+        }
+
+        mlir_operation_state& add_region(mlir_region r)
+        {
+            regions.emplace_back(std::move(r));
+            return *this;
+        }
+
+        mlir_operation_state& add_results(const std::vector<shape>& outputs)
+        {
+            auto x = prog->make_tensors(outputs);
+            mlirOperationStateAddResults(&op_state, x.size(), x.data());
+            return *this;
+        }
+
+        mlir_operation_state& add_operands(const std::vector<MlirValue>& inputs)
+        {
+            mlirOperationStateAddOperands(&op_state, inputs.size(), inputs.data());
+            return *this;
+        }
+
+        mlir_operation create_operation()
+        {
+            std::vector<MlirRegion> mregions(regions.size());
+            std::transform(regions.begin(), regions.end(), mregions.begin(), [](const auto& r) {
+                return r.get();
+            });
+            mlirOperationStateAddOwnedRegions(&op_state, mregions.size(), mregions.data());
+            mlir_operation op(mlirOperationCreate(&op_state));
+            // Release memory since mlir_operation owns it
+            for(auto& r : regions)
+                r.release();
+            regions.clear();
+            return op;
+        }
+
+        mlir_program* prog;
+        MlirOperationState op_state;
+        std::vector<mlir_region> regions = {};
+    };
+
+    mlir_operation_state create_operation_state(const std::string_view& name)
+    {
+        return {*this, name};
+    }
+
+    std::vector<MlirValue> insert(MlirBlock body, mlir_operation_state ops)
+    {
+        std::vector<MlirValue> result;
+        mlir_operation op = ops.create_operation();
+        auto weak_op      = op.get();
+        mlirBlockAppendOwnedOperation(body, op.release());
+
+        auto n = mlirOperationGetNumResults(weak_op);
+        result.reserve(n);
+        transform(range(n), std::back_inserter(result), [&](auto i) {
+            return mlirOperationGetResult(weak_op, i);
+        });
+        return result;
+    }
+
+    MlirBlock
+    insert(MlirBlock body, const module& m, std::unordered_map<instruction_ref, MlirValue>& ins_map)
+    {
+        auto names = m.get_parameter_names();
+        std::sort(names.begin(), names.end());
+        std::vector<shape> inputs;
+        std::transform(names.begin(),
+                       names.end(),
+                       std::back_inserter(inputs),
+                       [&](const std::string& name) { return m.get_parameter_shape(name); });
+        std::vector<shape> outputs = m.get_output_shapes();
+
+        std::vector<MlirLocation> arg_locs(inputs.size(), location);
+        auto body_inputs   = make_tensors(inputs);
+        mlir_region region = mlirRegionCreate();
+        mlir_block fbody = mlirBlockCreate(body_inputs.size(), body_inputs.data(), arg_locs.data());
+        MlirBlock result = fbody.get();
+        mlirRegionAppendOwnedBlock(region.get(), fbody.release());
+
+        auto ops = create_operation_state("func.func");
+        ops.add_attributes({{"function_type", make_function_type(inputs, outputs)},
+                            {"sym_name", std::string("main")},
+                            {"kernel", std::string("mixr")}});
+        ops.add_region(std::move(region));
+        insert(body, std::move(ops));
+
+        for(auto i : range(names.size()))
+            ins_map[m.get_parameter(names[i])] = mlirBlockGetArgument(result, i);
+        return result;
+    }
+
+    static std::string get_name(instruction_ref ins)
+    {
+        if(ins->name() == "@return")
+            return "func.return";
+        return "migraphx." + ins->name();
+    }
+
+    static value get_operator_value(const operation& op)
+    {
+        auto v = op.to_value();
+        if(op.name() == "convolution")
+        {
+            // Adjust symetrical padding
+            if(v.at("padding").size() == v.at("stride").size())
+            {
+                auto padding = v.at("padding");
+                std::copy(padding.begin(), padding.end(), std::back_inserter(v.at("padding")));
+            }
+        }
+        return v;
+    }
+
+    static shape get_shape(instruction_ref ins)
+    {
+        if(ins->name() == "@return")
+        {
+            assert(ins->inputs().size() == 1);
+            return ins->inputs().front()->get_shape();
+        }
+        return ins->get_shape();
+    }
+
+    void parse(const module& m)
+    {
+        auto mbody = mlirModuleGetBody(mmodule.get());
+        std::unordered_map<instruction_ref, MlirValue> ins_map;
+        auto fbody = insert(mbody, m, ins_map);
+        for(auto ins : iterator_for(m))
+        {
+            if(ins->name() == "@param")
+                continue;
+            auto name = get_name(ins);
+            auto ops  = create_operation_state(name);
+            ops.add_attribute_value(get_operator_value(ins->get_operator()));
+            if(ins->name() != "@return")
+                ops.add_results({get_shape(ins)});
+            if(ins->name() == "convolution")
+            {
+                pp =
+                    problem_params{ins->get_operator(), to_shapes(ins->inputs()), ins->get_shape()};
+                std::string tuned = get_tune_params();
+                if(!tuned.empty())
+                    ops.add_attributes({{"perf_config", tuned}});
+                // check if HW supports xdlops
+                if(contains(get_xdlops_archs(), target_name))
+                    ops.add_attributes({{"xdlopsV2", true}});
+            }
+
+            std::vector<MlirValue> inputs;
+            transform(
+                ins->inputs(), std::back_inserter(inputs), [&](auto i) { return ins_map.at(i); });
+            ops.add_operands(inputs);
+
+            auto outputs = insert(fbody, std::move(ops));
+            if(ins->name() != "@return")
+            {
+                assert(outputs.size() == 1);
+                ins_map[ins] = outputs.front();
+            }
+        }
+    }
+
+    code_object_op compile() MIGRAPHX_TIDY_CONST
+    {
+        mlir_pass_manager pm{mlirPassManagerCreate(ctx.get())};
+        // 1st pipeline to call
+        mlirMIGraphXAddHighLevelPipeline(pm.get());
+        // 2nd pipeline to call
+        mlirMIGraphXAddBackendPipeline(pm.get(), target_name.c_str(), "amdgcn-amd-amdhsa", "");
+        mlirPassManagerRun(pm.get(), mmodule.get());
+
+        code_object_op op{};
+        op.symbol_name                = "main";
+        op.code_object                = get_binary();
+        std::tie(op.global, op.local) = get_launch_params();
+        return op;
+    }
+
+    void find_target()
+    {
+        std::string tname = get_device_name();
+        // HACK: Since MLIR can't handle the full target name
+        target_name = trim(split_string(tname, ':').front());
+        if(tname.size() != target_name.size())
+            std::cout
+                << "*************** WARNING: MLIR may not compile the correct target features for: "
+                << tname << std::endl;
+    }
+
+    std::pair<std::size_t, std::size_t> get_launch_params() const
+    {
+        uint32_t attrs[2];
+        // returns block and grid sizes
+        mlirGetKernelAttrs(mmodule.get(), attrs);
+        std::size_t local  = attrs[0];
+        std::size_t global = local * attrs[1];
+        return {global, local};
+    }
+
+    value::binary get_binary() const
+    {
+        int size = 0;
+        mlirGetBinary(mmodule.get(), &size, nullptr);
+        value::binary result(size);
+        if(mlirGetBinary(mmodule.get(), &size, reinterpret_cast<char*>(result.data())))
+            return result;
+        MIGRAPHX_THROW("Failed to compile mlir program");
+    }
+
+    std::string get_tune_params() { return get_mlir_perf_for_conv(pp); }
+
+    mlir_context ctx;
+    MlirLocation location;
+    mlir_module mmodule;
+    problem_params pp;
+    std::deque<std::string> strings{};
+    std::string target_name;
+};
+
+std::string dump_mlir(const module& m)
+{
+    mlir_program mp;
+    mp.parse(m);
+    auto mod_op = mlirModuleGetOperation(mp.mmodule.get());
+    return mlir_print(&mlirOperationPrint, mod_op);
+}
+
+code_object_op compile_mlir(const context&, const module& m)
+{
+    const bool trace = enabled(MIGRAPHX_TRACE_MLIR{});
+    if(trace)
+        std::cout << m << std::endl;
+    mlir_program mp;
+    mp.find_target();
+    mp.parse(m);
+    auto mod_op = mlirModuleGetOperation(mp.mmodule.get());
+    if(trace)
+        std::cout << mlir_print(&mlirOperationPrint, mod_op) << std::endl;
+    auto co   = mp.compile();
+    co.output = m.get_output_shapes().front();
+    return co;
+}
+
+instruction_ref insert_mlir(module& m,
+                            instruction_ref ins,
+                            code_object_op co,
+                            const std::vector<instruction_ref>& inputs)
+{
+
+    std::vector<instruction_ref> refs;
+    std::size_t last = 0;
+#ifdef MIGRAPHX_MLIR_BARE_POINTER
+    refs.reserve(inputs.size());
+    std::copy(inputs.begin(), inputs.end(), std::back_inserter(refs));
+    last = refs.size() - 1;
+#else
+    refs.reserve(inputs.size() * 15);
+    std::unordered_map<uint64_t, instruction_ref> literal_map{};
+    auto get_literal = [&](uint64_t value) {
+        auto fi = literal_map.find(value);
+        if(fi != literal_map.end())
+            return fi->second;
+        auto lit = m.add_literal(value);
+        literal_map.emplace(value, lit);
+        return lit;
+    };
+
+    for(auto input : inputs)
+    {
+        const size_t offset = 0;
+        auto s              = input->get_shape();
+        last                = refs.size();
+        refs.push_back(input);
+        refs.push_back(input);
+        refs.push_back(get_literal(offset)); // offset
+
+        // dim sizes
+        std::transform(s.lens().begin(),
+                       s.lens().end(),
+                       std::back_inserter(refs),
+                       [&](const auto& lval) { return get_literal(lval); });
+        // refs.push_back(get_literal(1)); // G
+
+        // dim strides
+        std::transform(s.strides().begin(),
+                       s.strides().end(),
+                       std::back_inserter(refs),
+                       [&](const auto& lval) { return get_literal(lval); });
+        // refs.push_back(get_literal(1)); // G
+    }
+#endif
+    co.expected_inputs = to_shapes(refs);
+    co.output_arg      = last;
+    return m.insert_instruction(ins, co, refs);
+}
+
+#else
+
+std::string dump_mlir(const module&) { return {}; }
+
+code_object_op compile_mlir(const context&, const module&) { return {}; }
+
+template <class T>
+void use(T&)
+{
+}
+
+instruction_ref
+// cppcheck-suppress funcArgNamesDifferent
+insert_mlir(module& m, instruction_ref, code_object_op co, const std::vector<instruction_ref>&)
+{
+    use(co);
+    return m.end();
+}
+
+#endif
+
+} // namespace gpu
+} // namespace MIGRAPHX_INLINE_NS
+} // namespace migraphx

src/targets/gpu/mlir_conv.cpp

-/*
- * 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.
- */
-#include <migraphx/gpu/mlir_conv.hpp>
-#include <migraphx/manage_ptr.hpp>
-#include <migraphx/instruction.hpp>
-#include <migraphx/make_op.hpp>
-
-#include <migraphx/op/convolution.hpp>
-
-#include <migraphx/gpu/context.hpp>
-#include <migraphx/gpu/convolution.hpp>
-#include <migraphx/iterator_for.hpp>
-#include <migraphx/program.hpp>
-
-#include <migraphx/ranges.hpp>
-#include <migraphx/make_op.hpp>
-#include <migraphx/generate.hpp>
-#include <migraphx/program.hpp>
-#include <migraphx/gpu/kernel.hpp>
-#include <migraphx/gpu/target.hpp>
-#include <migraphx/gpu/hip.hpp>
-#include <migraphx/gpu/compile_hip.hpp>
-
-#include <utility>
-#include <functional>
-#include <algorithm>
-
-#ifdef MIGRAPHX_MLIR_MIOPEN_SUPPORT
-#include <Miir.h>
-#endif // MIGRAPHX_MLIR_MIOPEN_SUPPORT
-
-#include <cstdio>
-
-namespace migraphx {
-inline namespace MIGRAPHX_INLINE_NS {
-namespace gpu {
-
-struct mlir_apply
-{
-    module* mod           = nullptr;
-    const mlir_conv* pass = nullptr;
-
-    const char* mlir_kernel_name = "migraphx_conv2d";
-
-    std::unordered_map<uint64_t, instruction_ref> literal_map{};
-
-    struct execution_spec
-    {
-        migraphx::value::binary binary;
-        size_t global_size;
-        size_t local_size;
-        execution_spec(migraphx::value::binary&& binary_m, size_t global_s, size_t local_s)
-            : binary(std::move(binary_m)), global_size(global_s), local_size(local_s)
-        {
-        }
-    };
-
-    std::unordered_map<std::string, std::shared_ptr<execution_spec>> binary_map{};
-
-    context& get_context() const
-    {
-        assert(pass != nullptr);
-        assert(pass->ctx != nullptr);
-        return *pass->ctx;
-    }
-
-    void init() const
-    {
-        assert(mod != nullptr);
-        assert(pass != nullptr);
-    }
-
-    std::shared_ptr<execution_spec> make_mlir_binary(instruction_ref op_r)
-    {
-        std::shared_ptr<execution_spec> result;
-
-#ifdef MIGRAPHX_MLIR_MIOPEN_SUPPORT
-        auto conv  = any_cast<op::convolution>(op_r->get_operator());
-        auto inp_t = op_r->inputs().at(0)->get_shape();
-        auto flt_t = op_r->inputs().at(1)->get_shape();
-        auto out_t = op_r->get_shape();
-
-        auto get_type_str = [](const shape& s) -> const char* {
-            switch(s.type())
-            {
-            case shape::float_type: return "f32";
-            case shape::half_type: return "f16";
-            case shape::bool_type:
-            case shape::double_type:
-            case shape::uint8_type:
-            case shape::int8_type:
-            case shape::uint16_type:
-            case shape::int16_type:
-            case shape::int32_type:
-            case shape::int64_type:
-            case shape::uint32_type:
-            case shape::uint64_type:
-            case shape::tuple_type: break;
-            }
-            return nullptr;
-        };
-
-        const auto* inp_t_s = get_type_str(inp_t);
-        const auto* flt_t_s = get_type_str(flt_t);
-        const auto* out_t_s = get_type_str(out_t);
-
-        if(out_t_s == nullptr || inp_t_s == nullptr || flt_t_s == nullptr)
-            return result;
-
-        std::string mlir_options = "--kernel_name " + std::string(mlir_kernel_name);
-
-        // platform spec
-        auto& device = get_context().get_current_device();
-        char dev_name[64];
-        sprintf(dev_name, "gfx%lu%02lu", device.get_device_major(), device.get_device_minor());
-        mlir_options += " --arch " + std::string(dev_name) + " --num_cu " +
-                        std::to_string(device.get_cu_count()); // ???
-
-        // Conv spec
-        mlir_options +=
-            " --operation "
-            "conv2d"
-            " --batchsize " +
-            std::to_string(conv.group) + " --groupsize " + std::to_string(1) + " --padding_h " +
-            std::to_string(conv.padding[0]) + " --padding_w " + std::to_string(conv.padding[1]) +
-            " --conv_stride_h " + std::to_string(conv.stride[0]) + " --conv_stride_w " +
-            std::to_string(conv.stride[1]) + " --dilation_h " + std::to_string(conv.dilation[0]) +
-            " --dilation_w " + std::to_string(conv.dilation[1]);
-
-        // Input spec
-        mlir_options += " --in_layout "
-                        "NCHWG"
-                        " --in_type " +
-                        std::string(inp_t_s) + " --in_channels " + std::to_string(inp_t.lens()[1]) +
-                        " --in_h " + std::to_string(inp_t.lens()[2]) + " --in_w " +
-                        std::to_string(inp_t.lens()[3]);
-
-        // Filter spec
-        mlir_options += " --fil_layout "
-                        "NCHWG"
-                        " --fil_type " +
-                        std::string(flt_t_s) + " --fil_h " + std::to_string(flt_t.lens()[2]) +
-                        " --fil_w " + std::to_string(flt_t.lens()[3]);
-
-        // Output spec
-        mlir_options += " --out_layout "
-                        "NCHWG"
-                        " --out_type " +
-                        std::string(out_t_s) + " --out_channels " +
-                        std::to_string(out_t.lens()[1]) + " --out_h " +
-                        std::to_string(out_t.lens()[2]) + " --out_w " +
-                        std::to_string(out_t.lens()[3]);
-
-        auto bin_i = binary_map.find(mlir_options);
-        if(bin_i == binary_map.end())
-        {
-            size_t bin_size = 0;
-
-            using mlir_handle = MIGRAPHX_MANAGE_PTR(MiirHandle, miirDestroyHandle);
-            auto handle       = mlir_handle(miirCreateHandle(mlir_options.c_str()));
-
-            if(miirLowerBin(handle.get()) == MIIR_SUCCESS &&
-               miirBufferGet(handle.get(), nullptr, &bin_size) == MIIR_SUCCESS)
-            {
-                migraphx::value::binary bin(bin_size);
-                if(miirBufferGet(handle.get(), reinterpret_cast<char*>(bin.data()), &bin_size) ==
-                   MIIR_SUCCESS)
-                {
-                    size_t global_size;
-                    size_t block_size;
-                    if(miirGetExecutionDims(handle.get(), &global_size, &block_size) ==
-                       MIIR_SUCCESS)
-                    {
-                        result = std::make_shared<execution_spec>(
-                            std::move(bin), global_size, block_size);
-                    }
-                }
-            }
-
-            binary_map[mlir_options] = result;
-        }
-        else
-        {
-            result = bin_i->second;
-        }
-#else  // MIGRAPHX_MLIR_MIOPEN_SUPPORT
-        (void)op_r;
-#endif // MIGRAPHX_MLIR_MIOPEN_SUPPORT
-        return result;
-    }
-
-    instruction_ref get_literal(uint64_t value)
-    {
-        auto fi = literal_map.find(value);
-        if(fi != literal_map.end())
-            return fi->second;
-        auto lit = mod->add_literal(value);
-        literal_map.emplace(value, lit);
-        return lit;
-    }
-
-    operation make_code_object_op(instruction_ref op_r, const std::shared_ptr<execution_spec>& spec)
-    {
-        // each pointer is expanded out to a MemRefDescriptor
-        auto inp_t = op_r->inputs().at(0)->get_shape();
-        auto flt_t = op_r->inputs().at(1)->get_shape();
-        auto out_t = op_r->get_shape();
-
-        auto i64 = shape(shape::uint64_type);
-
-        std::vector<shape> expected_inputs = {
-            flt_t, flt_t, i64, i64, i64, i64, i64, i64, i64, i64, i64, i64,  i64,   inp_t,
-            inp_t, i64,   i64, i64, i64, i64, i64, i64, i64, i64, i64, i64,  out_t, out_t,
-            i64,   i64,   i64, i64, i64, i64, i64, i64, i64, i64, i64, out_t};
-
-        return migraphx::make_op("gpu::code_object",
-                                 {
-                                     {"code_object", spec->binary},
-                                     {"symbol_name", mlir_kernel_name},
-                                     {"global", spec->global_size},
-                                     {"local", spec->local_size},
-                                     {"expected_inputs", migraphx::to_value(expected_inputs)},
-                                     {"output", migraphx::to_value(out_t)},
-                                 });
-    }
-
-    void add_memref_descriptor(std::vector<instruction_ref>& refs, instruction_ref inst)
-    {
-        const size_t offset = 0;
-        auto inst_t         = inst->get_shape();
-        refs.push_back(inst);
-        refs.push_back(inst);
-        refs.push_back(get_literal(offset)); // offset
-
-        // dim sizes
-        std::transform(inst_t.lens().begin(),
-                       inst_t.lens().end(),
-                       std::back_inserter(refs),
-                       [&](const auto& lval) { return get_literal(lval); });
-        refs.push_back(get_literal(1)); // G
-
-        // dim strides
-        std::transform(inst_t.strides().begin(),
-                       inst_t.strides().end(),
-                       std::back_inserter(refs),
-                       [&](const auto& lval) { return get_literal(lval); });
-        refs.push_back(get_literal(1)); // G
-    }
-
-    instruction_ref insert_allocation(instruction_ref ins, const shape& s) const
-    {
-        return mod->insert_instruction(ins, hip_allocate{s});
-    }
-
-    void replace_conv_op(instruction_ref ins)
-    {
-        auto conv_bin = make_mlir_binary(ins);
-        if(conv_bin)
-        {
-            auto conv = make_code_object_op(ins, conv_bin);
-
-            auto inp = ins->inputs().at(0);
-            auto flt = ins->inputs().at(1);
-            auto out = insert_allocation(ins, ins->get_shape());
-
-            std::vector<instruction_ref> refs;
-            refs.reserve(3 * 13 + 1);
-            add_memref_descriptor(refs, flt);
-            add_memref_descriptor(refs, inp);
-            add_memref_descriptor(refs, out);
-            refs.push_back(out);
-
-            mod->replace_instruction(ins, conv, refs);
-        }
-    }
-
-    void apply()
-    {
-        init();
-        for(auto it : iterator_for(*mod))
-        {
-            if(it->name() == "convolution")
-            {
-                replace_conv_op(it);
-            }
-        }
-    }
-};
-
-void mlir_conv::apply(module& m) const { mlir_apply{&m, this}.apply(); }
-
-} // namespace gpu
-} // namespace MIGRAPHX_INLINE_NS
-} // namespace migraphx

src/targets/gpu/perfdb.cpp

+/*
+ * 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.
+ */
+#include <migraphx/gpu/perfdb.hpp>
+#include <migraphx/value.hpp>
+#include <migraphx/sqlite.hpp>
+#include <migraphx/stringutils.hpp>
+#include <migraphx/permutation.hpp>
+#include <fstream>
+
+namespace migraphx {
+inline namespace MIGRAPHX_INLINE_NS {
+namespace gpu {
+
+namespace {
+
+std::string get_layout(const shape& s, std::string labels)
+{
+    auto result = labels;
+    auto p      = find_permutation(s);
+    std::transform(p.begin(), p.end(), result.begin(), [&](auto i) { return labels[i]; });
+    return "'" + result + "'";
+}
+
+std::string get_type(const shape& s)
+{
+    static const std::unordered_map<shape::type_t, std::string> m = {
+        {shape::float_type, "'FP32'"},
+        {shape::half_type, "'FP16'"},
+        {shape::double_type, "'FP64'"},
+        {shape::int8_type, "'INT8'"},
+        {shape::int32_type, "'INT32'"},
+    };
+    auto it = m.find(s.type());
+    if(it == m.end())
+        return "UNKNOWN";
+    return it->second;
+}
+
+std::string generate_miopen_config(const problem_params& pp)
+{
+    value v       = pp.op.to_value();
+    auto input    = pp.inputs[0].lens();
+    auto weights  = pp.inputs[1].lens();
+    auto padding  = v["padding"].to_vector<std::size_t>();
+    auto stride   = v["stride"].to_vector<std::size_t>();
+    auto dilation = v["dilation"].to_vector<std::size_t>();
+    if(padding.size() != stride.size())
+        padding.erase(padding.begin() + padding.size() / 2, padding.end());
+    return to_string_range({std::string{" C.in_channels="},       to_string(input[1]),
+                            std::string{" AND C.in_h="},          to_string(input[2]),
+                            std::string{" AND C.in_w="},          to_string(input[3]),
+                            std::string{" AND C.fil_h="},         to_string(weights[2]),
+                            std::string{" AND C.fil_w="},         to_string(weights[3]),
+                            std::string{" AND C.out_channels="},  to_string(weights[0]),
+                            std::string{" AND C.batchsize="},     to_string(input[0]),
+                            std::string{" AND C.pad_h="},         to_string(padding[0]),
+                            std::string{" AND C.pad_w="},         to_string(padding[2]),
+                            std::string{" AND C.dilation_h="},    to_string(dilation[0]),
+                            std::string{" AND C.dilation_w="},    to_string(dilation[1]),
+                            std::string{" AND C.conv_stride_h="}, to_string(stride[0]),
+                            std::string{" AND C.conv_stride_w="}, to_string(stride[1]),
+                            std::string{" AND C.layout="},        get_layout(pp.inputs[0], "NCHW"),
+                            std::string{" AND C.data_type="},     get_type(pp.inputs[0]),
+                            std::string{" AND C.direction="},     std::string{"'F'"}},
+                           " ");
+}
+
+auto query_miopen_db(const std::string& query)
+{
+    // TODO: Store db as a static variable
+    const auto dbpath = fs::path{"/opt"} / "rocm" / "share" / "miopen" / "db" / "miopen.db";
+    // Check if db file exists.
+    std::ifstream dbs(dbpath);
+    if(dbs.is_open())
+    {
+        dbs.close();
+    }
+    else
+    {
+        std::vector<std::unordered_map<std::string, std::string>> empty;
+        return empty;
+    }
+
+    auto db = sqlite::read(dbpath);
+    return db.execute(query);
+}
+
+} // namespace
+
+std::string get_mlir_perf_for_conv(const problem_params& pp)
+{
+    std::string query = "select P.* \
+                             from perf_db P, config C \
+                             where P.config = C.id AND \
+                             P.solver = 'ConvMlirIgemmFwdXdlops' AND \
+                             ${config}";
+
+    auto results =
+        query_miopen_db(interpolate_string(query, {{"config", generate_miopen_config(pp)}}));
+    if(results.empty())
+        return "";
+    return results.front().at("params");
+}
+
+} // namespace gpu
+} // namespace MIGRAPHX_INLINE_NS
+} // namespace migraphx

src/targets/gpu/prefuse_ops.cpp

@@ -99,70 +99,11 @@ struct find_add_layernorm
         m.replace_instruction(ins, add_layernorm{}, add_ins->inputs());
     }
 };
-
-struct find_gpulayernorm
-{
-    auto matcher() const { return match::layernorm(); }
-
-    void apply(module& m, const match::matcher_result& r) const
-    {
-        auto ins   = r.result;
-        auto x_ins = r.instructions["x"];
-
-        if(not x_ins->get_shape().standard())
-            x_ins = m.insert_instruction(ins, make_op("contiguous"), x_ins);
-
-        auto relements = x_ins->get_shape().lens().back();
-
-        if(relements > 1024 or (relements % 4 != 0 and relements > 256))
-            return;
-
-        auto a = m.insert_instruction(
-            ins, make_op("hip::allocate", {{"shape", to_value(x_ins->get_shape())}}));
-        m.replace_instruction(ins, make_op("gpu::layernorm"), x_ins, a);
-    }
-};
-
-struct find_gputriaddlayernorm
-{
-    auto matcher() const
-    {
-        auto add1 =
-            match::name("add")(match::none_of(match::is_constant()),
-                               match::args(match::any().bind("z1"), match::any().bind("z2")));
-        auto add2 = match::name("add")(match::either_arg(0, 1)(add1, match::any().bind("z3")));
-        return match::layernorm()(match::var("x")(add2));
-    }
-
-    void apply(module& m, const match::matcher_result& r) const
-    {
-        auto ins   = r.result;
-        auto x_ins = r.instructions["z1"];
-        auto y_ins = r.instructions["z2"];
-        auto z_ins = r.instructions["z3"];
-
-        for(auto* pins : {&x_ins, &y_ins, &z_ins})
-        {
-            if(not(*pins)->get_shape().standard())
-                *pins = m.insert_instruction(ins, make_op("contiguous"), *pins);
-        }
-
-        auto relements = x_ins->get_shape().lens().back();
-
-        if(relements > 1024 or (relements % 4 != 0 and relements > 256))
-            return;
-
-        auto a = m.insert_instruction(
-            ins, make_op("hip::allocate", {{"shape", to_value(x_ins->get_shape())}}));
-        m.replace_instruction(ins, make_op("gpu::triadd_layernorm"), x_ins, y_ins, z_ins, a);
-    }
-};
 } // namespace
 
 void prefuse_ops::apply(module& m) const
 {
     match::find_matches(m, find_add_layernorm{}, find_layernorm{});
-    // match::find_matches(m, find_gputriaddlayernorm{}, find_gpulayernorm{});
 }
 
 } // namespace gpu

src/targets/gpu/quant_convolution.cpp

@@ -67,9 +67,9 @@ argument miopen_quant_convolution::compute(context& ctx,
     return args[3];
 }
 
-shape miopen_quant_convolution::compile(context& ctx,
-                                        const shape& output_shape,
-                                        std::vector<shape> inputs)
+shape miopen_quant_convolution::find(context& ctx,
+                                     const shape& output_shape,
+                                     std::vector<shape> inputs)
 {
     shape workspace_shape{};
     auto x_desc = make_tensor(inputs[0], int8_x4_format);
@@ -92,18 +92,18 @@ shape miopen_quant_convolution::compile(context& ctx,
         x_shape = pack_int8_shape(x_shape);
         w_shape = pack_int8_shape(w_shape);
     }
-    auto arg_vec4_x = to_gpu(generate_argument(x_shape));
-    auto arg_vec4_w = to_gpu(generate_argument(w_shape));
-    auto y          = allocate_gpu(output_shape);
-    auto workspace  = allocate_gpu(workspace_shape);
+    auto x         = to_gpu(generate_argument(x_shape));
+    auto w         = to_gpu(generate_argument(w_shape));
+    auto y         = allocate_gpu(output_shape);
+    auto workspace = allocate_gpu(workspace_shape);
 
     int algo_count = 1;
     miopenConvAlgoPerf_t perf;
     auto status = miopenFindConvolutionForwardAlgorithm(ctx.get_stream().get_miopen(),
                                                         x_desc.get(),
-                                                        arg_vec4_x.implicit(),
+                                                        x.implicit(),
                                                         w_desc.get(),
-                                                        arg_vec4_w.implicit(),
+                                                        w.implicit(),
                                                         cd.get(),
                                                         y_desc.get(),
                                                         y.implicit(),
@@ -114,11 +114,35 @@ shape miopen_quant_convolution::compile(context& ctx,
                                                         workspace_size,
                                                         false);
     if(status != miopenStatusSuccess)
-    {
-        MIGRAPHX_THROW("QUANT_CONVOLUTION: find convolution failed");
-    }
-    handle = ctx.get_stream().get_miopen();
-    algo   = perf.fwd_algo;
+        MIGRAPHX_THROW("MIOpen Quant Convolution: find convolution failed");
+    algo = perf.fwd_algo;
+
+    size_t solution_count;
+
+    status = miopenConvolutionForwardGetSolutionCount(ctx.get_stream().get_miopen(),
+                                                      w_desc.get(),
+                                                      x_desc.get(),
+                                                      cd.get(),
+                                                      y_desc.get(),
+                                                      &solution_count);
+    if(status != miopenStatusSuccess)
+        MIGRAPHX_THROW("MIOpen Quant Convolution: get solution count failed");
+
+    std::vector<miopenConvSolution_t> solutions(solution_count);
+
+    status = miopenConvolutionForwardGetSolution(ctx.get_stream().get_miopen(),
+                                                 w_desc.get(),
+                                                 x_desc.get(),
+                                                 cd.get(),
+                                                 y_desc.get(),
+                                                 solution_count,
+                                                 &solution_count,
+                                                 solutions.data());
+    if(status != miopenStatusSuccess)
+        MIGRAPHX_THROW("MIOpen Quant Convolution: get solution failed");
+
+    solution_id = solutions.front().solution_id;
+
     return shape{shape::int8_type, {perf.memory}};
 }
 
@@ -126,13 +150,29 @@ void miopen_quant_convolution::finalize(context& ctx,
                                         const shape& output_shape,
                                         std::vector<shape> inputs)
 {
-    if(handle == ctx.get_stream().get_miopen())
-        return;
-    // Check that workspace hasn't changed
-    auto size = inputs.at(2).bytes();
-    auto ws   = compile(ctx, output_shape, std::move(inputs));
-    if(ws.bytes() > size)
-        MIGRAPHX_THROW("Workspace has changed during finalization.");
+    if(cd == nullptr)
+        cd = make_conv(op);
+    if(solution_id == 0)
+    {
+        // Check that workspace hasn't changed
+        auto size = inputs.at(2).bytes();
+        auto ws   = find(ctx, output_shape, inputs);
+        if(ws.bytes() > size)
+            MIGRAPHX_THROW("MIOpen Quant Convolution: workspace has changed during finalization.");
+    }
+
+    auto x_desc = make_tensor(inputs[0], int8_x4_format);
+    auto w_desc = make_tensor(inputs[1], int8_x4_format);
+    auto y_desc = make_tensor(output_shape);
+
+    auto status = miopenConvolutionForwardCompileSolution(ctx.get_stream().get_miopen(),
+                                                          w_desc.get(),
+                                                          x_desc.get(),
+                                                          cd.get(),
+                                                          y_desc.get(),
+                                                          solution_id);
+    if(status != miopenStatusSuccess)
+        MIGRAPHX_THROW("MIOpen Quant Convolution: compile solution failed");
 }
 
 shape miopen_quant_convolution::pack_int8_shape(const shape& s) const

src/targets/gpu/target.cpp

@@ -53,10 +53,11 @@
 #include <migraphx/gpu/compile_ops.hpp>
 #include <migraphx/gpu/concat_gpu_opt.hpp>
 #include <migraphx/gpu/context.hpp>
+#include <migraphx/gpu/device_name.hpp>
+#include <migraphx/gpu/fuse_mlir.hpp>
 #include <migraphx/gpu/fuse_ops.hpp>
 #include <migraphx/gpu/prefuse_ops.hpp>
 #include <migraphx/gpu/lowering.hpp>
-#include <migraphx/gpu/mlir_conv.hpp>
 #include <migraphx/gpu/pack_int8_args.hpp>
 #include <migraphx/gpu/schedule_model.hpp>
 #include <migraphx/gpu/sync_device.hpp>
@@ -128,7 +129,8 @@ std::vector<pass> target::get_passes(migraphx::context& gctx, const compile_opti
         dead_code_elimination{},
         enable_pass(not enabled(MIGRAPHX_DISABLE_POINTWISE_FUSION{}), fuse_pointwise{}),
         dead_code_elimination{},
-        mlir_conv{&ctx},
+        fuse_mlir{&ctx},
+        dead_code_elimination{},
         lowering{&ctx, options.offload_copy},
         eliminate_contiguous{"gpu::contiguous"},
         dead_code_elimination{},
@@ -161,7 +163,7 @@ std::vector<pass> target::get_passes(migraphx::context& gctx, const compile_opti
 
 std::string target::name() const { return "gpu"; }
 
-migraphx::context target::get_context() const { return context{}; }
+migraphx::context target::get_context() const { return context(gpu::get_device_id()); }
 
 argument target::copy_to(const argument& arg) const { return gpu::to_gpu(arg); }
 

src/targets/ref/lowering.cpp

@@ -51,6 +51,8 @@
 #include <migraphx/register_op.hpp>
 #include <migraphx/make_op.hpp>
 #include <migraphx/tune_axis.hpp>
+#include <migraphx/pad_calc.hpp>
+
 #include <unordered_map>
 #include <utility>
 #include <iostream>
@@ -231,8 +233,31 @@ struct ref_convolution : auto_register_op<ref_convolution<Op>>
     {
         return op.normalize_compute_shape(inputs);
     }
+
     argument compute(context&, shape output_shape, std::vector<argument> args) const
     {
+        std::vector<std::size_t> padding;
+        if(op.use_dynamic_same_auto_pad)
+        {
+            auto input_lens = args[0].get_shape().lens();
+            std::vector<std::size_t> img_lens{input_lens.begin() + 2, input_lens.end()};
+            auto weights_lens = args[1].get_shape().lens();
+            std::vector<std::size_t> k_lens{weights_lens.begin() + 2, weights_lens.end()};
+            padding = calc_dyn_auto_pad(img_lens, k_lens, op.stride, op.dilation);
+            std::cout << "[ ";
+            output_shape =
+                compute_padded_shape({args.at(0).get_shape(), args.at(1).get_shape()}, padding);
+        }
+        else
+        {
+            padding = op.padding;
+            if(output_shape.dynamic())
+            {
+                output_shape =
+                    op.normalize_compute_shape({args.at(0).get_shape(), args.at(1).get_shape()});
+            }
+        }
+
         argument result{output_shape};
         visit_quantize(result, args[0], args[1])([&](auto output, auto input, auto weights) {
             auto in_lens = input.get_shape().lens();
@@ -252,7 +277,7 @@ struct ref_convolution : auto_register_op<ref_convolution<Op>>
                 {
                     auto d_2 = dim - 2;
                     win_start.push_back(std::ptrdiff_t(idx_o[dim] * op.stride[d_2]) -
-                                        std::ptrdiff_t(op.padding[d_2]));
+                                        std::ptrdiff_t(padding[d_2]));
                 }
                 const auto group_id = w / (wei_n / op.group);
 
@@ -289,6 +314,34 @@ struct ref_convolution : auto_register_op<ref_convolution<Op>>
         });
         return result;
     }
+
+    private:
+    /*!
+     * Used for dynamic auto padding since padding needs to be computed at evaulation time.
+     * \param inputs two fixed shape inputs [input_tensor, weights]
+     * \param padding from auto_pad calculation
+     */
+    shape compute_padded_shape(const std::vector<shape>& inputs,
+                               const std::vector<std::size_t>& padding) const
+    {
+        const shape& input            = inputs.at(0);
+        const shape& weights          = inputs.at(1);
+        const size_t num_spatial_dims = input.lens().size() - 2;
+
+        std::vector<size_t> output_lens{input.lens()[0], weights.lens()[0]};
+        // calculate the output shape of the convolution: ((W - K + 2P) / S) + 1
+        for(size_t i = 0; i < num_spatial_dims; i++)
+        {
+            auto padding_factor = padding[i] + padding[i + num_spatial_dims];
+            output_lens.push_back(std::size_t(std::max<std::ptrdiff_t>(
+                1,
+                (input.lens()[i + 2] - (1 + op.dilation[i] * (weights.lens()[i + 2] - 1)) +
+                 padding_factor) /
+                        op.stride[i] +
+                    1)));
+        }
+        return inputs[0].with_lens(output_lens);
+    }
 };
 
 struct ref_im2col

src/tf/tf_parser.cpp

@@ -216,7 +216,7 @@ static std::vector<T> get_data_vals(const google::protobuf::RepeatedField<T>& da
         std::fill(data_vals.begin(), data_vals.end(), data[0]);
     }
     else
-        copy(data.begin(), data.end(), std::back_inserter(data_vals));
+        copy(data.begin(), data.end(), data_vals.begin());
     return data_vals;
 }
 
@@ -329,33 +329,37 @@ void tf_parser::parse_node(const std::string& name)
         auto&& node = nodes.at(name);
         if(not is_valid_op(node))
             return;
-
         std::vector<instruction_ref> args;
-
         for(auto&& input : node.input())
         {
             // control dependencies (signified by ^ before the name) are ignored
             if(contains(input, "^"))
                 continue;
-            if(nodes.count(input) > 0)
+            std::string input_name = input;
+            // if input has trailing `:0` index then remove it
+            auto multi_out_idx = input.find(':');
+            if(multi_out_idx != std::string::npos && input.substr(multi_out_idx + 1) == "0")
+            {
+                input_name = input.substr(0, multi_out_idx);
+            }
+            if(nodes.count(input_name) > 0)
             {
-                std::string iname;
                 // input was from a node with multiple outputs
-                if(contains(input, ':'))
+                if(contains(input_name, ':'))
                 {
-                    iname = input.substr(0, input.find(':'));
+                    input_name = input_name.substr(0, input.find(':'));
                 }
                 else
                 {
-                    iname = get_name(nodes.at(input));
+                    input_name = get_name(nodes.at(input_name));
                 }
-                assert(name != iname);
-                this->parse_node(iname);
-                args.push_back(instructions.at(input));
+                assert(name != input_name);
+                this->parse_node(input_name);
+                args.push_back(instructions.at(input_name));
             }
             else
             {
-                args.push_back(instructions.at(input));
+                args.push_back(instructions.at(input_name));
             }
         }
         std::vector<instruction_ref> result;

test/CMakeLists.txt

@@ -137,6 +137,22 @@ if(MIGRAPHX_ENABLE_GPU)
     endforeach()
 endif()
 
+if(MIGRAPHX_ENABLE_FPGA)
+    # fpga tests
+    file(GLOB FPGA_TESTS ${CONFIGURE_DEPENDS} fpga/*.cpp)
+
+    foreach(TEST ${FPGA_TESTS})
+        get_filename_component(BASE_NAME ${TEST} NAME_WE)
+        add_test_executable(test_fpga_${BASE_NAME} ${TEST})
+        rocm_clang_tidy_check(test_fpga_${BASE_NAME})
+        set_tests_properties(test_fpga_${BASE_NAME} PROPERTIES 
+            COST 10 
+            RESOURCE_LOCK fpga
+        )
+        target_link_libraries(test_fpga_${BASE_NAME} migraphx_fpga)
+    endforeach()
+endif()
+
 # Onnx test
 set(TEST_ONNX_DIR ${CMAKE_CURRENT_SOURCE_DIR}/onnx)
 file (GLOB ONNX_TESTS ${TEST_ONNX_DIR}/*.cpp)

test/api/CMakeLists.txt

@@ -32,6 +32,17 @@ function(add_api_test TEST_NAME TEST_SRC TEST_DIR)
     add_dependencies(check ${NAME})
 endfunction()
 
+# Workaround: C file dont work with clang-tidy right now, need a fix in rocm-cmake
+function(add_c_api_test TEST_NAME TEST_SRC TEST_DIR)
+    set(NAME test_api_${TEST_NAME})
+    add_executable(${NAME} EXCLUDE_FROM_ALL ${TEST_SRC})
+    target_link_libraries(${NAME} migraphx_c migraphx)
+    target_include_directories(${NAME} PUBLIC ../include)
+    add_test(NAME ${NAME} COMMAND $<TARGET_FILE:${NAME}> WORKING_DIRECTORY ${TEST_DIR}) 
+    add_dependencies(tests ${NAME})
+    add_dependencies(check ${NAME})
+endfunction()
+
 add_api_test(array_base test_array_base.cpp ${TEST_ONNX_DIR})
 add_api_test(assign test_assign.cpp ${TEST_ONNX_DIR})
 add_api_test(compile_options test_compile_options.cpp ${TEST_ONNX_DIR})
@@ -40,6 +51,7 @@ add_api_test(module_construct test_module_construct.cpp ${TEST_ONNX_DIR})
 add_api_test(ref test_cpu.cpp ${TEST_ONNX_DIR})
 add_api_test(save_load test_save_load.cpp ${TEST_ONNX_DIR})
 add_api_test(op test_op_construct.cpp ${TEST_ONNX_DIR})
+add_c_api_test(c_op test_c_op_construct.c ${TEST_ONNX_DIR})
 add_api_test(custom_op test_custom_op.cpp ${TEST_ONNX_DIR})
 add_api_test(tf_parser test_tf_parser.cpp ${TEST_TF_DIR})
 # GPU-based tests

test/api/test_c_op_construct.c

+/*
+ * 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.
+ */
+#include <migraphx/migraphx.h>
+#include <string.h>
+
+void expect_equal(const char* x, const char* y)
+{
+    if(strcmp(x, y) != 0)
+        abort();
+}
+
+int main()
+{
+    char name[1024];
+    migraphx_operation_t op;
+    migraphx_operation_create(&op, "add", 0);
+    migraphx_operation_name(name, 1024, op);
+    migraphx_operation_destroy(op);
+    expect_equal(name, "add");
+}

test/api/test_custom_op.cpp

@@ -23,8 +23,10 @@
  */
 #include <algorithm>
 #include <cmath>
+#include <exception>
 #include <migraphx/migraphx.h>
 #include <migraphx/migraphx.hpp>
+#include <stdexcept>
 #include "test.hpp"
 
 struct sigmoid_custom_op final : migraphx::experimental_custom_op_base
@@ -43,10 +45,22 @@ struct sigmoid_custom_op final : migraphx::experimental_custom_op_base
 
     virtual migraphx::shape compute_shape(migraphx::shapes inputs) const override
     {
-        CHECK(inputs.size() == 2);
-        CHECK(inputs[0].lengths().size() == 1);
-        CHECK(inputs[0].type() == migraphx_shape_float_type);
-        CHECK(bool{inputs[0] == inputs[1]});
+        if(inputs.size() != 2)
+        {
+            throw std::runtime_error("op must have two inputs");
+        }
+        if(inputs[0].lengths().size() != 1)
+        {
+            throw std::runtime_error("input arg must be a vector or scalar");
+        }
+        if(inputs[0].type() != migraphx_shape_float_type)
+        {
+            throw std::runtime_error("input arg must be of type float");
+        }
+        if(inputs[0] != inputs[1])
+        {
+            throw std::runtime_error("input arg and buffer allocation must be of same shape");
+        }
         return inputs.back();
     }
 };
@@ -83,4 +97,18 @@ TEST_CASE(run_sigmoid_custom_op)
     EXPECT(bool{result == migraphx::argument(s, expected_result.data())});
 }
 
+extern "C" void migraphx_test_private_disable_exception_catch(bool b);
+
+TEST_CASE(run_sigmoid_with_incorrect_shape)
+{
+    migraphx::program p;
+    migraphx::shape s{migraphx_shape_float_type, {12}};
+    migraphx::module m = p.get_main_module();
+    auto x             = m.add_parameter("x", s);
+    migraphx_test_private_disable_exception_catch(true);
+    EXPECT(test::throws<std::exception>(
+        [&] { m.add_instruction(migraphx::operation("sigmoid_custom_op"), {x}); },
+        "Error in compute_shape of: sigmoid_custom_op: op must have two inputs"));
+}
+
 int main(int argc, const char* argv[]) { test::run(argc, argv); }

test/api/test_custom_op_gpu.cpp

@@ -24,6 +24,7 @@
 #include <hip/hip_runtime_api.h>
 #include <migraphx/migraphx.h>
 #include <migraphx/migraphx.hpp>
+#include <stdexcept>
 #include "test.hpp"
 
 #define MIGRAPHX_HIP_ASSERT(x) (EXPECT(x == hipSuccess))
@@ -54,6 +55,14 @@ struct simple_custom_op final : migraphx::experimental_custom_op_base
 
     virtual migraphx::shape compute_shape(migraphx::shapes inputs) const override
     {
+        if(!inputs[0].standard())
+        {
+            throw std::runtime_error("first arg must be standard shaped");
+        }
+        if(inputs.size() != 2)
+        {
+            throw std::runtime_error("number of inputs must be 2");
+        }
         return inputs.back();
     }
 };
@@ -64,12 +73,17 @@ TEST_CASE(run_simple_custom_op)
     migraphx::register_experimental_custom_op(simple_op);
     migraphx::program p;
     migraphx::shape s{migraphx_shape_int32_type, {4, 3}};
+    migraphx::shape trans_shape{migraphx_shape_int32_type, {3, 4}};
     migraphx::module m = p.get_main_module();
     auto x             = m.add_parameter("x", s);
     auto neg           = m.add_instruction(migraphx::operation("neg"), x);
-    auto alloc         = m.add_allocation(s);
-    auto custom_kernel = m.add_instruction(migraphx::operation("simple_custom_op"), {neg, alloc});
-    auto relu          = m.add_instruction(migraphx::operation("relu"), custom_kernel);
+    auto alloc         = m.add_allocation(trans_shape);
+    auto neg_trans =
+        m.add_instruction(migraphx::operation("transpose", "{permutation: [1, 0]}"), {neg});
+    auto neg_cont = m.add_instruction(migraphx::operation("contiguous"), {neg_trans});
+    auto custom_kernel =
+        m.add_instruction(migraphx::operation("simple_custom_op"), {neg_cont, alloc});
+    auto relu = m.add_instruction(migraphx::operation("relu"), custom_kernel);
     m.add_return({relu});
     migraphx::compile_options options;
     options.set_offload_copy();
@@ -82,7 +96,7 @@ TEST_CASE(run_simple_custom_op)
     auto result_vec = result.as_vector<int>();
     std::vector<int> expected_result(12, 0);
     std::fill(expected_result.begin() + 6, expected_result.end(), 3);
-    EXPECT(bool{result == migraphx::argument(s, expected_result.data())});
+    EXPECT(bool{result == migraphx::argument(trans_shape, expected_result.data())});
 }
 
 int main(int argc, const char* argv[]) { test::run(argc, argv); }

test/check_shapes_test.cpp

+/*
+ * 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.
+ */
+#include "test.hpp"
+#include <migraphx/check_shapes.hpp>
+#include <migraphx/make_op.hpp>
+
+/*!
+ * Tests for check_shapes object handling dynamic shapes
+ */
+
+using migraphx::shape;
+
+bool create_shapes(bool dynamic_allowed)
+{
+    try
+    {
+        shape a{shape::int64_type, {3}};
+        shape b{shape::float_type, {{3, 6, 0}, {4, 4, 0}}};
+        auto op = migraphx::make_op("add");
+        migraphx::check_shapes{{a, b}, op, dynamic_allowed}.has(2);
+        return true;
+    }
+    catch(...)
+    {
+        return false;
+    }
+}
+
+TEST_CASE(allow_dynamic_shape) { EXPECT(create_shapes(true)); }
+
+TEST_CASE(fail_dynamic_shape) { EXPECT(!create_shapes(false)); }
+
+int main(int argc, const char* argv[]) { test::run(argc, argv); }

test/eliminate_contiguous_test.cpp

@@ -205,4 +205,24 @@ TEST_CASE(contiguous_pointwise)
         mm->begin(), mm->end(), [](auto&& ins) { return ins.name() == "contiguous"; }));
 }
 
+TEST_CASE(slice_contiguous)
+{
+    migraphx::module m;
+
+    migraphx::shape s{migraphx::shape::float_type, {4, 2}};
+    auto x  = m.add_parameter("x", s);
+    auto t  = m.add_instruction(migraphx::make_op("transpose", {{"permutation", {1, 0}}}), x);
+    auto c  = m.add_instruction(migraphx::make_op("contiguous"), t);
+    auto s1 = m.add_instruction(
+        migraphx::make_op("slice", {{"axes", {0}}, {"starts", {0}}, {"ends", {1}}}), c);
+    auto s2 = m.add_instruction(
+        migraphx::make_op("slice", {{"axes", {0}}, {"starts", {1}}, {"ends", {2}}}), c);
+    auto c1 = m.add_instruction(migraphx::make_op("contiguous"), s1);
+    auto c2 = m.add_instruction(migraphx::make_op("contiguous"), s2);
+    m.add_instruction(pass_standard_op{}, c1, c2);
+    run_pass(m);
+    EXPECT(std::count_if(
+               m.begin(), m.end(), [](auto&& ins) { return ins.name() == "contiguous"; }) == 1);
+}
+
 int main(int argc, const char* argv[]) { test::run(argc, argv); }

test/fpga/test_compile.cpp

+/*
+ * 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.
+ */
+#include "test.hpp"
+
+#include <migraphx/make_op.hpp>
+#include <migraphx/program.hpp>
+#include <migraphx/register_target.hpp>
+#include <migraphx/fpga/target.hpp>
+#include <migraphx/target_assignments.hpp>
+
+migraphx::program create_program()
+{
+    migraphx::program p;
+    auto* mm = p.get_main_module();
+    migraphx::shape s{migraphx::shape::float_type, {3}};
+    auto x     = mm->add_parameter("x", s);
+    auto y     = mm->add_parameter("y", s);
+    auto z     = mm->add_parameter("z", s);
+    auto sum   = mm->add_instruction(migraphx::make_op("add"), x, y);
+    auto sum_2 = mm->add_instruction(migraphx::make_op("add"), sum, z);
+    mm->add_return({sum_2});
+    return p;
+}
+
+TEST_CASE(compile)
+{
+    auto p = create_program();
+    auto t = migraphx::make_target("fpga");
+
+    p.compile(t);
+    EXPECT(p.is_compiled());
+}
+
+int main(int argc, const char* argv[]) { test::run(argc, argv); }