Merge branch 'develop' into dot-add

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

@@ -24,7 +24,7 @@
 #ifndef MIGRAPHX_GUARD_KERNELS_FUNCTIONAL_HPP
 #define MIGRAPHX_GUARD_KERNELS_FUNCTIONAL_HPP
 
-#include <migraphx/kernels/array.hpp>
+#include <migraphx/kernels/integral_constant.hpp>
 
 // NOLINTNEXTLINE
 #define MIGRAPHX_RETURNS(...) \

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

@@ -175,6 +175,21 @@ constexpr auto sliced(Slicer slicer, F f)
     };
 }
 
+template <class Input, index_int Axis>
+constexpr auto compute_reduce_axis()
+{
+    constexpr auto lens =
+        transform_i(get_shape_c<Input>{}.lens, [](index_int x, index_int i) -> index_int {
+            if(i == Axis)
+                return 1;
+            return x;
+        });
+    return make_shape(lens, get_shape_c<Input>{}.strides);
+}
+
+template <class Input, index_int Axis>
+using with_axis = decltype(compute_reduce_axis<Input, Axis>());
+
 struct block
 {
     template <class Slicer>
@@ -201,6 +216,14 @@ struct block
             if(idx.local == 0)
                 f();
         }
+
+        template <class F>
+        __device__ auto inner(F f) const
+        {
+            return sliced(slicer, [=](auto x, auto... xs) {
+                idx.local_stride(x.get_shape().elements(), [&](auto j) { f(x[j], xs[j]...); });
+            });
+        }
     };
 
     template <class Slicer>
@@ -247,6 +270,17 @@ struct lane
         {
             f();
         }
+
+        template <class F>
+        __device__ auto inner(F f) const
+        {
+            return sliced(slicer, [=](auto x, auto... xs) {
+                for(index_int j = 0; j < x.get_shape().elements(); j++)
+                {
+                    f(x[j], xs[j]...);
+                }
+            });
+        }
     };
 
     template <class Slicer>

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

@@ -32,6 +32,7 @@ namespace migraphx {
 template <class Lens, class Strides>
 struct shape
 {
+    using shape_type  = shape;
     using index_array = typename Lens::base_array;
     Lens lens         = {};
     Strides strides   = {};

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
+
+#include <migraphx/kernels/reduce.hpp>
+#include <migraphx/kernels/ops.hpp>
+
+namespace migraphx {
+
+template <index_int Axis, class Input, class Output>
+__device__ void softmax(Input input, Output output)
+{
+    reduce::block::run<reduce::with_axis<Input, Axis>>([&](auto, auto r) {
+        auto batch_max = r.reduce(op::max{}, lowest{}, op::id{})(input);
+        auto batch_sum =
+            r.reduce(op::sum{}, 0, [&](auto x) { return migraphx::exp(x - batch_max); })(input);
+        r.inner([&](auto& y, auto x) { y = migraphx::exp(x - batch_max) / batch_sum; })(output,
+                                                                                        input);
+    });
+}
+
+} // namespace migraphx
+#endif // MIGRAPHX_GUARD_KERNELS_SOFTMAX_HPP

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

@@ -27,6 +27,8 @@
 #include <migraphx/kernels/types.hpp>
 #include <migraphx/kernels/integral_constant.hpp>
 #include <migraphx/kernels/functional.hpp>
+#include <migraphx/kernels/type_traits.hpp>
+#include <migraphx/kernels/debug.hpp>
 
 namespace migraphx {
 

src/targets/gpu/lowering.cpp

@@ -186,7 +186,6 @@ struct miopen_apply
         add_extend_op("rnn_var_sl_shift_output");
         add_extend_op("rnn_var_sl_shift_sequence");
         add_extend_op("scatter_none");
-        add_extend_op("softmax");
         add_extend_op("topk");
 
         add_batch_norm_inference_op();
@@ -301,7 +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());
@@ -332,7 +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_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/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 +53,10 @@
 #include <migraphx/gpu/compile_ops.hpp>
 #include <migraphx/gpu/concat_gpu_opt.hpp>
 #include <migraphx/gpu/context.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 +128,8 @@ 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{},

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/get_target_assignments.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/ref/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 diff = mm->add_instruction(migraphx::make_op("div"), x, y);
+    mm->add_instruction(migraphx::make_op("div"), diff, z);
+    return p;
+}
+
+TEST_CASE(is_supported)
+{
+    auto p       = create_program();
+    auto targets = migraphx::get_targets();
+    EXPECT(!targets.empty());
+    auto first_target = targets[0];
+    auto t            = migraphx::make_target(first_target);
+
+    const auto assignments = p.get_target_assignments({t});
+    for(const auto& [ins, target] : assignments)
+    {
+        (void)ins;
+        EXPECT(target == first_target);
+    }
+}
+
+int main(int argc, const char* argv[]) { test::run(argc, argv); }

test/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>
+#include <migraphx/gpu/target.hpp>
+#include <migraphx/gpu/context.hpp>
+#include <migraphx/gpu/write_literals.hpp>
+#include <migraphx/ref/target.hpp>
+#include <migraphx/module.hpp>
+#include <migraphx/program.hpp>
+#include <migraphx/make_op.hpp>
+#include <migraphx/ranges.hpp>
+#include <migraphx/stringutils.hpp>
+#include <migraphx/generate.hpp>
+#include <migraphx/verify_args.hpp>
+#include <migraphx/instruction.hpp>
+#include <migraphx/functional.hpp>
+#include <test.hpp>
+
+using migraphx::trim;
+
+// m test_gpu_mlir && ./bin/test_gpu_mlir
+
+struct mlir_gpu_target : migraphx::gpu::target
+{
+    std::string name() const { return "mlir"; }
+    std::vector<migraphx::pass> get_passes(migraphx::context& gctx,
+                                           const migraphx::compile_options&) const
+    {
+        auto& ctx = migraphx::any_cast<migraphx::gpu::context>(gctx);
+        return {migraphx::gpu::write_literals{&ctx}};
+    }
+};
+
+std::string encode(const std::string& s)
+{
+    std::stringstream ss;
+    bool prespace = false;
+    for(auto c : s)
+    {
+        if(std::isspace(c) != 0)
+        {
+            if(not prespace)
+                ss << "  ";
+            prespace = true;
+        }
+        else if(std::isprint(c) != 0)
+        {
+            ss << c;
+            prespace = false;
+        }
+    }
+    return migraphx::trim(ss.str());
+}
+
+migraphx::program create_program_from_mlir(const migraphx::module& mmlir)
+{
+    migraphx::program p;
+    auto* mm   = p.get_main_module();
+    auto names = mmlir.get_parameter_names();
+    std::vector<migraphx::instruction_ref> inputs;
+    std::transform(names.begin(), names.end(), std::back_inserter(inputs), [&](const auto& name) {
+        return mm->add_parameter(name, mmlir.get_parameter_shape(name));
+    });
+    std::sort(inputs.begin(), inputs.end(), migraphx::by(std::less<>{}, [](auto ins) {
+                  return to_string(ins->get_operator());
+              }));
+    inputs.push_back(mm->add_parameter("output", mmlir.get_output_shapes().front()));
+
+    migraphx::gpu::context ctx;
+    migraphx::gpu::insert_mlir(*mm, mm->end(), compile_mlir(ctx, mmlir), inputs);
+    return p;
+}
+
+migraphx::parameter_map generate_params(const migraphx::program& p)
+{
+    migraphx::parameter_map m;
+    std::size_t i = 0;
+    for(auto&& x : p.get_parameter_shapes())
+    {
+        // m[x.first] = migraphx::fill_argument(x.second, 1);
+        m[x.first] = migraphx::generate_argument(x.second, i++);
+    }
+    return m;
+}
+
+migraphx::argument run_gpu(migraphx::program p, const migraphx::parameter_map& inputs)
+{
+    mlir_gpu_target t;
+    p.compile(t);
+    migraphx::parameter_map m;
+    for(auto&& input : inputs)
+    {
+        m[input.first] = t.copy_to(input.second);
+    }
+    for(auto&& x : p.get_parameter_shapes())
+    {
+        if(m.count(x.first) == 0)
+        {
+            m[x.first] = t.allocate(x.second);
+        }
+    }
+    return t.copy_from(p.eval(m).front());
+}
+
+migraphx::argument run_ref(migraphx::program p, const migraphx::parameter_map& inputs)
+{
+    p.compile(migraphx::ref::target{});
+    return p.eval(inputs).front();
+}
+
+bool verify_mlir(const migraphx::module& mmlir)
+{
+    migraphx::program ref;
+    ref.get_main_module()->insert_instructions(ref.get_main_module()->end(), &mmlir);
+
+    auto inputs = generate_params(ref);
+
+    auto mlir = create_program_from_mlir(mmlir);
+    return migraphx::verify_args("mlir", run_ref(ref, inputs), run_gpu(mlir, inputs));
+}
+
+TEST_CASE(conv)
+{
+    const std::string mlir_output = R"__migraphx__(
+module {
+  func @main(%arg0: tensor<2x8x3x3xf32>, %arg1: tensor<1x8x4x4xf32>) -> tensor<1x2x2x2xf32> attributes {kernel = "mixr"} {
+    %0 = migraphx.convolution(%arg1, %arg0) {dilation = [1, 1], group = 1 : i64, padding = [0, 0, 0, 0], padding_mode = 0 : i64, stride = [1, 1]} : (tensor<1x8x4x4xf32>, tensor<2x8x3x3xf32>) -> tensor<1x2x2x2xf32>
+    return %0 : tensor<1x2x2x2xf32>
+  }
+}
+)__migraphx__";
+    migraphx::module m;
+    auto x    = m.add_parameter("x", {migraphx::shape::float_type, {1, 8, 4, 4}});
+    auto w    = m.add_parameter("w", {migraphx::shape::float_type, {2, 8, 3, 3}});
+    auto conv = m.add_instruction(migraphx::make_op("convolution"), x, w);
+    m.add_return({conv});
+    auto s = migraphx::gpu::dump_mlir(m);
+    // Skip test if MLIR is not enabled
+    if(s.empty())
+        return;
+    CHECK(encode(s) == encode(mlir_output));
+    EXPECT(verify_mlir(m));
+}
+
+TEST_CASE(conv_add_relu)
+{
+    const std::string mlir_output = R"__migraphx__(
+module {
+  func @main(%arg0: tensor<1x2x2x2xf32>, %arg1: tensor<2x8x3x3xf32>, %arg2: tensor<1x8x4x4xf32>) -> tensor<1x2x2x2xf32> attributes {kernel = "mixr"} {
+    %0 = migraphx.convolution(%arg2, %arg1) {dilation = [1, 1], group = 1 : i64, padding = [0, 0, 0, 0], padding_mode = 0 : i64, stride = [1, 1]} : (tensor<1x8x4x4xf32>, tensor<2x8x3x3xf32>) -> tensor<1x2x2x2xf32>
+    %1 = migraphx.add(%0, %arg0) : (tensor<1x2x2x2xf32>, tensor<1x2x2x2xf32>) -> tensor<1x2x2x2xf32>
+    %2 = migraphx.relu(%1) : (tensor<1x2x2x2xf32>) -> tensor<1x2x2x2xf32>
+    return %2 : tensor<1x2x2x2xf32>
+  }
+}
+)__migraphx__";
+    migraphx::module m;
+    auto x    = m.add_parameter("x", {migraphx::shape::float_type, {1, 8, 4, 4}});
+    auto w    = m.add_parameter("w", {migraphx::shape::float_type, {2, 8, 3, 3}});
+    auto b    = m.add_parameter("b", {migraphx::shape::float_type, {1, 2, 2, 2}});
+    auto conv = m.add_instruction(migraphx::make_op("convolution"), x, w);
+    auto add  = m.add_instruction(migraphx::make_op("add"), conv, b);
+    auto relu = m.add_instruction(migraphx::make_op("relu"), add);
+    m.add_return({relu});
+    auto s = migraphx::gpu::dump_mlir(m);
+    // Skip test if MLIR is not enabled
+    if(s.empty())
+        return;
+    CHECK(encode(s) == encode(mlir_output));
+    EXPECT(verify_mlir(m));
+}
+
+int main(int argc, const char* argv[]) { test::run(argc, argv); }

test/module_test.cpp

@@ -300,6 +300,96 @@ TEST_CASE(parameter_name_order)
     EXPECT(param_names == names1);
 }
 
+TEST_CASE(insert_instructions_module)
+{
+    migraphx::shape s{migraphx::shape::int32_type, {1}};
+    migraphx::module m1("m1");
+    auto x1   = m1.add_parameter("x1", s);
+    auto sqrt = m1.add_instruction(migraphx::make_op("sqrt"), {x1});
+    m1.add_instruction(migraphx::make_op("add"), {sqrt, x1});
+
+    migraphx::module m2("m2");
+    auto x2 = m2.add_parameter("x2", s);
+    m2.add_instruction(migraphx::make_op("sqrt"), {x2});
+
+    m1.insert_instructions(sqrt, &m2, {{x2, x1}});
+
+    EXPECT(std::prev(sqrt)->name() == "sqrt");
+    EXPECT(std::count_if(m1.begin(), m1.end(), [](auto&& ins) { return ins.name() == "sqrt"; }) ==
+           2);
+    EXPECT(std::count_if(m1.begin(), m1.end(), [](auto&& ins) { return ins.name() == "@param"; }) ==
+           1);
+    EXPECT(contains(m1.get_parameter_shapes(), "x1"));
+    EXPECT(not contains(m1.get_parameter_shapes(), "x2"));
+}
+
+TEST_CASE(add_instructions_module)
+{
+    migraphx::shape s{migraphx::shape::int32_type, {1}};
+    migraphx::module m1("m1");
+    auto x1 = m1.add_parameter("x1", s);
+    m1.add_instruction(migraphx::make_op("sqrt"), {x1});
+
+    migraphx::module m2("m2");
+    auto x2 = m2.add_parameter("x2", s);
+    m2.add_instruction(migraphx::make_op("sqrt"), {x2});
+
+    m1.add_instructions(&m2, {{x2, x1}});
+
+    EXPECT(std::count_if(m1.begin(), m1.end(), [](auto&& ins) { return ins.name() == "sqrt"; }) ==
+           2);
+    EXPECT(std::count_if(m1.begin(), m1.end(), [](auto&& ins) { return ins.name() == "@param"; }) ==
+           1);
+    EXPECT(contains(m1.get_parameter_shapes(), "x1"));
+    EXPECT(not contains(m1.get_parameter_shapes(), "x2"));
+}
+
+TEST_CASE(add_instructions_range)
+{
+    migraphx::shape s{migraphx::shape::int32_type, {1}};
+    migraphx::module m1("m1");
+    auto x1 = m1.add_parameter("x1", s);
+    m1.add_instruction(migraphx::make_op("sqrt"), {x1});
+
+    migraphx::module m2("m2");
+    auto x2    = m2.add_parameter("x2", s);
+    auto sqrt2 = m2.add_instruction(migraphx::make_op("sqrt"), {x2});
+
+    m1.add_instructions(sqrt2, m2.end(), {{x2, x1}});
+    EXPECT(std::any_of(
+        m1.begin(), m1.end(), [&](auto&& ins) { return migraphx::contains(ins.inputs(), x1); }));
+
+    EXPECT(std::count_if(m1.begin(), m1.end(), [](auto&& ins) { return ins.name() == "sqrt"; }) ==
+           2);
+    EXPECT(std::count_if(m1.begin(), m1.end(), [](auto&& ins) { return ins.name() == "@param"; }) ==
+           1);
+    EXPECT(contains(m1.get_parameter_shapes(), "x1"));
+    EXPECT(not contains(m1.get_parameter_shapes(), "x2"));
+}
+
+TEST_CASE(add_instructions_vector)
+{
+    migraphx::shape s{migraphx::shape::int32_type, {1}};
+    migraphx::module m1("m1");
+    auto x1 = m1.add_parameter("x1", s);
+    m1.add_instruction(migraphx::make_op("sqrt"), {x1});
+
+    migraphx::module m2("m2");
+    auto x2    = m2.add_parameter("x2", s);
+    auto sqrt2 = m2.add_instruction(migraphx::make_op("sqrt"), {x2});
+
+    m1.add_instructions({sqrt2}, {{x2, x1}});
+    EXPECT(std::any_of(
+        m1.begin(), m1.end(), [&](auto&& ins) { return migraphx::contains(ins.inputs(), x1); }));
+
+    EXPECT(std::count_if(m1.begin(), m1.end(), [](auto&& ins) { return ins.name() == "sqrt"; }) ==
+           2);
+    EXPECT(std::count_if(m1.begin(), m1.end(), [](auto&& ins) { return ins.name() == "@param"; }) ==
+           1);
+    EXPECT(contains(m1.get_parameter_shapes(), "x1"));
+    EXPECT(not contains(m1.get_parameter_shapes(), "x2"));
+}
+
 struct check_for_pass_op
 {
     bool* found = nullptr;

test/op_shape_test.cpp

@@ -981,7 +981,8 @@ TEST_CASE(multibroadcast)
     }
     {
         std::vector<std::size_t> lens{4, 1, 3};
-        migraphx::shape input{migraphx::shape::float_type, {}};
+        std::vector<std::size_t> empt = {};
+        migraphx::shape input{migraphx::shape::float_type, empt};
         throws_shape(migraphx::make_op("multibroadcast", {{"out_lens", lens}}), input);
     }
     {
@@ -1533,15 +1534,46 @@ TEST_CASE(test_squeeze_wrong_axis)
 
 TEST_CASE(test_unsqueeze)
 {
-    migraphx::shape s1{migraphx::shape::float_type, {4, 3, 3}};
-    migraphx::shape s2{migraphx::shape::float_type, {4, 3, 1, 3}};
+    migraphx::shape s1{migraphx::shape::float_type, {4, 5, 3}};
+    migraphx::shape s2{migraphx::shape::float_type, {4, 5, 1, 3}};
     expect_shape(s2, migraphx::make_op("unsqueeze", {{"axes", {2}}}), s1);
 }
 
+TEST_CASE(test_unsqueeze_step)
+{
+    migraphx::shape s1{migraphx::shape::float_type, {4, 5, 12}};
+    migraphx::shape s2{migraphx::shape::float_type, {4, 5, 2, 6}};
+    expect_shape(s2, migraphx::make_op("unsqueeze", {{"axes", {2}}, {"steps", {2}}}), s1);
+}
+
+TEST_CASE(test_unsqueeze_step_non_divisable)
+{
+    migraphx::shape s1{migraphx::shape::float_type, {4, 5, 3}};
+    throws_shape(migraphx::make_op("unsqueeze", {{"axes", {2}}, {"steps", {2}}}), s1);
+}
+
+TEST_CASE(test_unsqueeze_step_zero)
+{
+    migraphx::shape s1{migraphx::shape::float_type, {4, 5, 12}};
+    throws_shape(migraphx::make_op("unsqueeze", {{"axes", {2}}, {"steps", {0}}}), s1);
+}
+
+TEST_CASE(test_unsqueeze_step_at_end)
+{
+    migraphx::shape s1{migraphx::shape::float_type, {4, 5, 12}};
+    throws_shape(migraphx::make_op("unsqueeze", {{"axes", {3}}, {"steps", {2}}}), s1);
+}
+
+TEST_CASE(test_unsqueeze_mismatch_step_axis)
+{
+    migraphx::shape s1{migraphx::shape::float_type, {4, 5, 12}};
+    throws_shape(migraphx::make_op("unsqueeze", {{"axes", {2}}, {"steps", {2, 3}}}), s1);
+}
+
 TEST_CASE(test_unsqueeze_negative_axis)
 {
-    migraphx::shape s1{migraphx::shape::float_type, {4, 3, 3}};
-    migraphx::shape s2{migraphx::shape::float_type, {4, 3, 1, 3}};
+    migraphx::shape s1{migraphx::shape::float_type, {4, 5, 3}};
+    migraphx::shape s2{migraphx::shape::float_type, {4, 5, 1, 3}};
     expect_shape(s2, migraphx::make_op("unsqueeze", {{"axes", {-2}}}), s1);
 }
 
@@ -1567,21 +1599,28 @@ TEST_CASE(test_unsqueeze_scalar_tensor2)
 TEST_CASE(test_unsqueeze_transpose)
 {
     migraphx::shape s1{migraphx::shape::float_type, {4, 4, 3}, {12, 1, 4}};
-    migraphx::shape s2{migraphx::shape::float_type, {4, 4, 1, 3}, {12, 1, 1, 4}};
+    migraphx::shape s2{migraphx::shape::float_type, {4, 4, 1, 3}, {12, 1, 12, 4}};
     expect_shape(s2, migraphx::make_op("unsqueeze", {{"axes", {2}}}), s1);
 }
 
+TEST_CASE(test_unsqueeze_transpose_step)
+{
+    migraphx::shape s1{migraphx::shape::float_type, {4, 4, 6}, {24, 1, 4}};
+    migraphx::shape s2{migraphx::shape::float_type, {4, 4, 2, 3}, {24, 1, 12, 4}};
+    expect_shape(s2, migraphx::make_op("unsqueeze", {{"axes", {2}}, {"steps", {2}}}), s1);
+}
+
 TEST_CASE(test_unsqueeze_multibroadcast)
 {
     migraphx::shape s1{migraphx::shape::float_type, {2, 3, 4}, {0, 1, 0}};
-    migraphx::shape s2{migraphx::shape::float_type, {2, 3, 1, 4}, {0, 1, 1, 0}};
+    migraphx::shape s2{migraphx::shape::float_type, {2, 3, 1, 4}, {0, 1, 0, 0}};
     expect_shape(s2, migraphx::make_op("unsqueeze", {{"axes", {2}}}), s1);
 }
 
 TEST_CASE(test_unsqueeze_slice)
 {
     migraphx::shape s1{migraphx::shape::float_type, {2, 3, 4}, {108, 36, 1}};
-    migraphx::shape s2{migraphx::shape::float_type, {2, 3, 1, 4}, {108, 36, 36, 1}};
+    migraphx::shape s2{migraphx::shape::float_type, {2, 3, 1, 4}, {108, 36, 4, 1}};
     expect_shape(s2, migraphx::make_op("unsqueeze", {{"axes", {2}}}), s1);
 }
 
@@ -1613,6 +1652,27 @@ TEST_CASE(test_unsqueeze_multiple_axes_2)
     expect_shape(s2, migraphx::make_op("unsqueeze", {{"axes", {0, 1}}}), s1);
 }
 
+TEST_CASE(test_unsqueeze_multiple_axes_3)
+{
+    migraphx::shape s1{migraphx::shape::float_type, {3, 4, 5}};
+    migraphx::shape s2{migraphx::shape::float_type, {3, 4, 1, 5, 1, 1}};
+    expect_shape(s2, migraphx::make_op("unsqueeze", {{"axes", {2, 4, 5}}}), s1);
+}
+
+TEST_CASE(test_unsqueeze_multiple_axes_4)
+{
+    migraphx::shape s1{migraphx::shape::float_type, {3, 4, 5}};
+    migraphx::shape s2{migraphx::shape::float_type, {3, 4, 1, 5, 1, 1}};
+    expect_shape(s2, migraphx::make_op("unsqueeze", {{"axes", {5, 4, 2}}}), s1);
+}
+
+TEST_CASE(test_unsqueeze_multiple_axes_step)
+{
+    migraphx::shape s1{migraphx::shape::float_type, {3, 4, 10}};
+    migraphx::shape s2{migraphx::shape::float_type, {3, 4, 2, 5, 1, 1}};
+    expect_shape(s2, migraphx::make_op("unsqueeze", {{"axes", {2, 4, 5}}, {"steps", {2}}}), s1);
+}
+
 TEST_CASE(transpose_shape)
 {
     migraphx::shape input{migraphx::shape::float_type, {2, 2}};

test/shape_test.cpp

@@ -38,7 +38,6 @@ TEST_CASE(test_shape_default)
     EXPECT(s.elements() == 0);
     EXPECT(s.bytes() == 0);
 }
-
 TEST_CASE(test_shape_assign)
 {
     migraphx::shape s1{migraphx::shape::float_type, {100, 32, 8, 8}};
@@ -65,6 +64,118 @@ TEST_CASE(test_shape_standard)
     EXPECT(not s.broadcasted());
 }
 
+TEST_CASE(test_shape_min_max_opt)
+{
+    migraphx::shape s{migraphx::shape::float_type, {2, 2, 3}, {6, 3, 1}};
+    EXPECT(s.min_lens() == s.lens());
+    EXPECT(s.max_lens() == s.lens());
+    EXPECT(s.opt_lens() == s.lens());
+}
+
+TEST_CASE(test_shape_dynamic_fixed)
+{
+    migraphx::shape s{migraphx::shape::float_type, {{2, 2, 0}, {2, 2, 0}, {3, 3, 0}}};
+    EXPECT(not s.standard());
+    EXPECT(not s.packed());
+    EXPECT(not s.transposed());
+    EXPECT(not s.broadcasted());
+    EXPECT(s.dynamic());
+    EXPECT(s.dyn_dims().size() == 3);
+    EXPECT(s.dyn_dims().at(0).is_fixed());
+    EXPECT(not s.dyn_dims().at(0).has_optimal());
+    EXPECT(s.min_lens() == std::vector<std::size_t>{2, 2, 3});
+    EXPECT(s.max_lens() == std::vector<std::size_t>{2, 2, 3});
+    EXPECT(s.opt_lens() == std::vector<std::size_t>{0, 0, 0});
+    EXPECT(s.bytes() == 2 * 2 * 3 * sizeof(float));
+}
+
+TEST_CASE(test_shape_dynamic_not_fixed)
+{
+    using migraphx::shape;
+    std::vector<shape::dynamic_dimension> dims = {};
+    dims.push_back(shape::dynamic_dimension{2, 5, 2});
+    dims.push_back(shape::dynamic_dimension{2, 8, 0});
+    migraphx::shape s{migraphx::shape::float_type, dims};
+    EXPECT(not s.standard());
+    EXPECT(not s.packed());
+    EXPECT(not s.transposed());
+    EXPECT(not s.broadcasted());
+    EXPECT(s.dynamic());
+    EXPECT(s.dyn_dims().size() == 2);
+    EXPECT(not s.dyn_dims().at(0).is_fixed());
+    EXPECT(s.dyn_dims().at(0).has_optimal());
+    EXPECT(s.min_lens() == std::vector<std::size_t>{2, 2});
+    EXPECT(s.max_lens() == std::vector<std::size_t>{5, 8});
+    EXPECT(s.opt_lens() == std::vector<std::size_t>{2, 0});
+    EXPECT(s.bytes() == 5 * 8 * sizeof(float));
+}
+
+TEST_CASE(test_shape_dynamic_compares)
+{
+    using migraphx::shape;
+    auto a = shape::dynamic_dimension{2, 5, 2};
+    auto b = a;
+    auto c = shape::dynamic_dimension{2, 5, 2};
+    auto d = shape::dynamic_dimension{3, 8, 4};
+    EXPECT(a == b);
+    EXPECT(a == c);
+    EXPECT(a != d);
+
+    migraphx::shape s0{shape::float_type, {a, d}};
+    migraphx::shape s1 = s0;
+    migraphx::shape s2{shape::float_type, {a, d}};
+    migraphx::shape s3{shape::int32_type, {a}};
+    EXPECT(s0 == s1);
+    EXPECT(s0 == s2);
+    EXPECT(s0 != s3);
+
+    std::stringstream ss0;
+    std::stringstream ss1;
+    std::stringstream ss3;
+    ss0 << s0;
+    ss1 << s1;
+    ss3 << s3;
+    EXPECT(ss0.str() == ss1.str());
+    EXPECT(ss0.str() != ss3.str());
+}
+
+TEST_CASE(test_shape_dynamic_errors)
+{
+    using migraphx::shape;
+    std::vector<shape::dynamic_dimension> dims = {};
+    dims.push_back(shape::dynamic_dimension{2, 5, 2});
+    dims.push_back(shape::dynamic_dimension{2, 8, 0});
+    migraphx::shape s{shape::float_type, dims};
+    EXPECT(test::throws([&] { s.elements(); }));
+    EXPECT(test::throws([&] { s.index({0, 1}); }));
+    EXPECT(test::throws([&] { s.index(1); }));
+    EXPECT(test::throws([&] { s.index(std::vector<std::size_t>{0, 1}); }));
+    EXPECT(test::throws([&] { s.with_lens({3, 5}); }));
+    EXPECT(test::throws([&] { s.with_lens(shape::float_type, {3, 5}); }));
+}
+
+TEST_CASE(test_shape_dynamic_serialize)
+{
+    using migraphx::shape;
+    std::vector<shape::dynamic_dimension> dims1 = {};
+    dims1.push_back(shape::dynamic_dimension{2, 5, 2});
+    dims1.push_back(shape::dynamic_dimension{2, 8, 0});
+    migraphx::shape s1{shape::float_type, dims1};
+    auto v1 = migraphx::to_value(s1);
+
+    std::vector<shape::dynamic_dimension> dims2 = {};
+    dims2.push_back(shape::dynamic_dimension{2, 5, 2});
+    migraphx::shape s2{shape::uint64_type, dims2};
+    auto v2 = migraphx::to_value(s2);
+    EXPECT(v1 != v2);
+
+    auto s3 = migraphx::from_value<shape>(v1);
+    EXPECT(s3 == s1);
+    auto s4 = migraphx::from_value<shape>(v2);
+    EXPECT(s4 == s2);
+    EXPECT(s3 != s4);
+}
+
 TEST_CASE(test_shape_packed)
 {
     migraphx::shape s{migraphx::shape::float_type, {2, 2}, {2, 1}};

test/simplify_reshapes_test.cpp

@@ -1141,4 +1141,138 @@ TEST_CASE(transpose_contiguous_reshape_binary_broadcast)
     EXPECT(m1 == m2);
 }
 
+TEST_CASE(transpose_unsqueeze_concat)
+{
+    migraphx::module m1;
+    {
+        auto l0 = m1.add_parameter("0", migraphx::shape{migraphx::shape::float_type, {1, 2, 1, 1}});
+        auto lt0 =
+            m1.add_instruction(migraphx::make_op("transpose", {{"permutation", {0, 2, 3, 1}}}), l0);
+        auto l1 = m1.add_parameter("1", migraphx::shape{migraphx::shape::float_type, {1, 2, 1, 1}});
+        auto lt1 =
+            m1.add_instruction(migraphx::make_op("transpose", {{"permutation", {0, 2, 3, 1}}}), l1);
+        auto l2 = m1.add_parameter("2", migraphx::shape{migraphx::shape::float_type, {1, 2, 1, 1}});
+        auto lt2 =
+            m1.add_instruction(migraphx::make_op("transpose", {{"permutation", {0, 2, 3, 1}}}), l2);
+        std::vector<migraphx::instruction_ref> args{lt0, lt1, lt2};
+        std::vector<migraphx::instruction_ref> unsqueezed_args;
+        int64_t axis = 3;
+
+        std::transform(
+            args.begin(),
+            args.end(),
+            std::back_inserter(unsqueezed_args),
+            [&](migraphx::instruction_ref arg) {
+                return m1.add_instruction(migraphx::make_op("unsqueeze", {{"axes", {axis}}}), arg);
+            });
+        m1.add_instruction(migraphx::make_op("concat", {{"axis", axis}}), unsqueezed_args);
+    }
+    // TODO: This could be simplified to a single transpose after concat
+    migraphx::module m2 = m1;
+    run_pass(m1);
+    EXPECT(m1 == m2);
+}
+
+TEST_CASE(transpose_slice)
+{
+    migraphx::module m1;
+    {
+        auto x      = m1.add_parameter("x", {migraphx::shape::float_type, {1, 384, 36, 64}});
+        auto slice1 = m1.add_instruction(
+            migraphx::make_op("slice", {{"axes", {2}}, {"starts", {0}}, {"ends", {12}}}), x);
+        auto transpose1 = m1.add_instruction(
+            migraphx::make_op("transpose", {{"permutation", {0, 2, 1, 3}}}), slice1);
+        auto slice2 = m1.add_instruction(
+            migraphx::make_op("slice", {{"axes", {2}}, {"starts", {12}}, {"ends", {24}}}), x);
+        auto transpose2 = m1.add_instruction(
+            migraphx::make_op("transpose", {{"permutation", {0, 2, 1, 3}}}), slice2);
+        auto slice3 = m1.add_instruction(
+            migraphx::make_op("slice", {{"axes", {2}}, {"starts", {24}}, {"ends", {36}}}), x);
+        auto transpose3 = m1.add_instruction(
+            migraphx::make_op("transpose", {{"permutation", {0, 2, 1, 3}}}), slice3);
+        m1.add_return({transpose1, transpose2, transpose3});
+    }
+    run_pass(m1);
+    migraphx::module m2;
+    {
+        auto x = m2.add_parameter("x", {migraphx::shape::float_type, {1, 384, 36, 64}});
+        auto transpose =
+            m2.add_instruction(migraphx::make_op("transpose", {{"permutation", {0, 2, 1, 3}}}), x);
+        auto slice1 = m2.add_instruction(
+            migraphx::make_op("slice", {{"axes", {1}}, {"starts", {0}}, {"ends", {12}}}),
+            transpose);
+        auto slice2 = m2.add_instruction(
+            migraphx::make_op("slice", {{"axes", {1}}, {"starts", {12}}, {"ends", {24}}}),
+            transpose);
+        auto slice3 = m2.add_instruction(
+            migraphx::make_op("slice", {{"axes", {1}}, {"starts", {24}}, {"ends", {36}}}),
+            transpose);
+        m2.add_return({slice1, slice2, slice3});
+    }
+    EXPECT(m1 == m2);
+}
+
+TEST_CASE(transpose_slice_diff_perm)
+{
+    migraphx::module m1;
+    {
+        auto x      = m1.add_parameter("x", {migraphx::shape::float_type, {1, 384, 36, 64}});
+        auto slice1 = m1.add_instruction(
+            migraphx::make_op("slice", {{"axes", {2}}, {"starts", {0}}, {"ends", {12}}}), x);
+        auto transpose1 = m1.add_instruction(
+            migraphx::make_op("transpose", {{"permutation", {0, 2, 1, 3}}}), slice1);
+        auto slice2 = m1.add_instruction(
+            migraphx::make_op("slice", {{"axes", {2}}, {"starts", {12}}, {"ends", {24}}}), x);
+        auto transpose2 = m1.add_instruction(
+            migraphx::make_op("transpose", {{"permutation", {0, 2, 3, 1}}}), slice2);
+        auto slice3 = m1.add_instruction(
+            migraphx::make_op("slice", {{"axes", {2}}, {"starts", {24}}, {"ends", {36}}}), x);
+        auto transpose3 = m1.add_instruction(
+            migraphx::make_op("transpose", {{"permutation", {0, 2, 1, 3}}}), slice3);
+        m1.add_return({transpose1, transpose2, transpose3});
+    }
+    run_pass(m1);
+    migraphx::module m2;
+    {
+        auto x = m2.add_parameter("x", {migraphx::shape::float_type, {1, 384, 36, 64}});
+        auto transpose =
+            m2.add_instruction(migraphx::make_op("transpose", {{"permutation", {0, 2, 1, 3}}}), x);
+        auto slice1 = m2.add_instruction(
+            migraphx::make_op("slice", {{"axes", {1}}, {"starts", {0}}, {"ends", {12}}}),
+            transpose);
+        auto slice2 = m2.add_instruction(
+            migraphx::make_op("slice", {{"axes", {1}}, {"starts", {12}}, {"ends", {24}}}),
+            transpose);
+        auto transpose2 = m2.add_instruction(
+            migraphx::make_op("transpose", {{"permutation", {0, 1, 3, 2}}}), slice2);
+        auto slice3 = m2.add_instruction(
+            migraphx::make_op("slice", {{"axes", {1}}, {"starts", {24}}, {"ends", {36}}}),
+            transpose);
+        m2.add_return({slice1, transpose2, slice3});
+    }
+    EXPECT(m1 == m2);
+}
+
+TEST_CASE(transpose_slice_single_transpose)
+{
+    migraphx::module m1;
+    {
+        auto x      = m1.add_parameter("x", {migraphx::shape::float_type, {1, 384, 36, 64}});
+        auto slice1 = m1.add_instruction(
+            migraphx::make_op("slice", {{"axes", {2}}, {"starts", {0}}, {"ends", {12}}}), x);
+        auto sqrt1  = m1.add_instruction(migraphx::make_op("sqrt"), slice1);
+        auto slice2 = m1.add_instruction(
+            migraphx::make_op("slice", {{"axes", {2}}, {"starts", {12}}, {"ends", {24}}}), x);
+        auto transpose = m1.add_instruction(
+            migraphx::make_op("transpose", {{"permutation", {0, 2, 1, 3}}}), slice2);
+        auto slice3 = m1.add_instruction(
+            migraphx::make_op("slice", {{"axes", {2}}, {"starts", {24}}, {"ends", {36}}}), x);
+        auto sqrt3 = m1.add_instruction(migraphx::make_op("sqrt"), slice3);
+        m1.add_return({sqrt1, transpose, sqrt3});
+    }
+    migraphx::module m2 = m1;
+    run_pass(m1);
+    EXPECT(m1 == m2);
+}
+
 int main(int argc, const char* argv[]) { test::run(argc, argv); }

test/verify/run_verify.cpp

@@ -30,6 +30,7 @@
 #include <migraphx/ranges.hpp>
 #include <migraphx/generate.hpp>
 #include <migraphx/load_save.hpp>
+#include <migraphx/tmp_dir.hpp>
 #include <migraphx/verify_args.hpp>
 #include <set>
 
@@ -57,6 +58,15 @@ std::future<typename std::result_of<Function()>::type> detach_async(Function&& f
     return std::async(std::launch::deferred, std::forward<Function>(f));
 }
 
+inline void verify_load_save(const migraphx::program& p)
+{
+    migraphx::tmp_dir td{"migraphx_test"};
+    auto path = td.path / "test.mxr";
+    migraphx::save(p, path.string());
+    auto loaded = migraphx::load(path.string());
+    EXPECT(p == loaded);
+}
+
 inline void compile_check(migraphx::program& p, const migraphx::target& t, bool show_trace = false)
 {
     auto name   = t.name();
@@ -82,6 +92,8 @@ inline void compile_check(migraphx::program& p, const migraphx::target& t, bool
             throw std::runtime_error("Compiling program with " + name + " alters its shape");
         }
     }
+    if(t.name() != "ref")
+        verify_load_save(p);
 }
 
 target_info run_verify::get_target_info(const std::string& name) const
@@ -152,6 +164,7 @@ void run_verify::verify(const std::string& name, const migraphx::program& p) con
     auto_print::set_terminate_handler(name);
     if(migraphx::enabled(MIGRAPHX_DUMP_TEST{}))
         migraphx::save(p, name + ".mxr");
+    verify_load_save(p);
     std::vector<std::string> target_names;
     for(const auto& tname : migraphx::get_targets())
     {

test/verify/test_conv_add_relu.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 "verify_program.hpp"
+#include <migraphx/program.hpp>
+#include <migraphx/generate.hpp>
+#include <migraphx/make_op.hpp>
+#include <migraphx/instruction.hpp>
+
+struct test_conv_add_relu : verify_program<test_conv_add_relu>
+{
+    migraphx::program create_program() const
+    {
+        migraphx::program p;
+        auto* mm = p.get_main_module();
+        auto input =
+            mm->add_parameter("x", migraphx::shape{migraphx::shape::float_type, {4, 3, 3, 3}});
+        auto weights =
+            mm->add_parameter("w", migraphx::shape{migraphx::shape::float_type, {4, 3, 3, 3}});
+        auto bias_literal = migraphx::literal{migraphx::shape{migraphx::shape::float_type, {4}},
+                                              {2.0f, 2.0f, 2.0f, 2.0f}};
+        auto bias         = mm->add_literal(bias_literal);
+        auto conv         = mm->add_instruction(migraphx::make_op("convolution"), input, weights);
+        auto bcast_bias   = mm->add_instruction(
+            migraphx::make_op("broadcast", {{"axis", 1}, {"out_lens", conv->get_shape().lens()}}),
+            bias);
+        auto bias_add = mm->add_instruction(migraphx::make_op("add"), conv, bcast_bias);
+        mm->add_instruction(migraphx::make_op("relu"), bias_add);
+        return p;
+    }
+};

tools/include/target.hpp

@@ -37,6 +37,8 @@
 #include <migraphx/compile_options.hpp>
 #include <migraphx/argument.hpp>
 #include <migraphx/rank.hpp>
+#include <migraphx/support_metric.hpp>
+#include <migraphx/instruction_ref.hpp>
 
 namespace migraphx {
 inline namespace MIGRAPHX_INLINE_NS {
@@ -61,6 +63,13 @@ struct target
      * @return The context to be used during compilation and execution.
      */
     context get_context() const;
+    /**
+     * @brief Check how well an instruction is supported on a target with the given metric
+     * @param ins Instruction to check if it's supported
+     * @param metric Used to define how the return value should be interpreted
+     * @return The value based on the chosen metric. Negative numbers mean unsupported
+     */
+    float is_supported(T&, instruction_ref ins, support_metric m) const;
     /**
      * @brief copy an argument to the current target.
      *
@@ -105,11 +114,18 @@ argument copy_from_target(T&, const argument& arg)
     return arg;
 }
 
+template <class T>
+float target_is_supported(T&, instruction_ref, support_metric)
+{
+    return 0;
+}
+
 <%
 interface('target',
      virtual('name', returns='std::string', const=True),
      virtual('get_passes', ctx='context&', options='const compile_options&', returns='std::vector<pass>', const=True),
      virtual('get_context', returns='context', const=True),
+     virtual('is_supported', returns='float', ins='instruction_ref', m='support_metric', const=True, default='target_is_supported'),
      virtual('copy_to',
              returns = 'argument',
              input   = 'const argument&',