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

src/targets/gpu/code_object_op.cpp

@@ -52,7 +52,7 @@ code_object_op::compute(context& ctx, const shape&, const std::vector<argument>&
     std::transform(
         args.begin(), args.end(), kargs.begin(), [](const argument& a) { return a.data(); });
     k.launch(ctx.get_stream().get(), global, local, std::move(kargs));
-    return args.back();
+    return args[get_output_arg(args.size())];
 }
 void code_object_op::finalize(context&, const shape&, const std::vector<shape>&)
 {

src/targets/gpu/compile_gen.cpp

@@ -43,6 +43,9 @@ static std::vector<std::size_t> vector_sizes(const std::vector<shape>& inputs)
 
 vectorize vectorize::elements(std::size_t axis, const std::vector<shape>& inputs)
 {
+    if(std::all_of(
+           inputs.begin(), inputs.end(), [&](const auto& s) { return s.lens()[axis] == 1; }))
+        return {1, axis};
     auto sizes = vector_sizes(inputs);
     std::vector<std::size_t> max_vec_size;
     std::transform(inputs.begin(),

src/targets/gpu/compile_hip.cpp

@@ -43,6 +43,7 @@ inline namespace MIGRAPHX_INLINE_NS {
 namespace gpu {
 
 MIGRAPHX_DECLARE_ENV_VAR(MIGRAPHX_GPU_DEBUG);
+MIGRAPHX_DECLARE_ENV_VAR(MIGRAPHX_GPU_DEBUG_SYM);
 MIGRAPHX_DECLARE_ENV_VAR(MIGRAPHX_GPU_OPTIMIZE);
 MIGRAPHX_DECLARE_ENV_VAR(MIGRAPHX_GPU_DUMP_ASM);
 MIGRAPHX_DECLARE_ENV_VAR(MIGRAPHX_GPU_DUMP_SRC);
@@ -227,6 +228,8 @@ compile_hip_src(const std::vector<src_file>& srcs, std::string params, const std
     if(params.find("-std=") == std::string::npos)
         params += " --std=c++17";
     params += " -fno-gpu-rdc";
+    if(enabled(MIGRAPHX_GPU_DEBUG_SYM{}))
+        params += " -g";
     params += " -c";
     if(is_hcc_compiler())
     {

src/targets/gpu/deconvolution.cpp

@@ -59,31 +59,30 @@ argument miopen_deconvolution::compute(context& ctx,
     auto w_desc = make_tensor(reshape_if_1d(args[1].get_shape()));
     auto y_desc = make_tensor(reshape_if_1d(output_shape));
 
-    float alpha = 1;
-    float beta  = 0;
-    auto status = miopenConvolutionForward(ctx.get_stream().get_miopen(),
-                                           &alpha,
-                                           x_desc.get(),
-                                           args[0].implicit(),
+    if(solution_id == 0)
+        MIGRAPHX_THROW("MIOpen Deconvolution: invalid solution ID");
+
+    auto status = miopenConvolutionForwardImmediate(ctx.get_stream().get_miopen(),
                                                     w_desc.get(),
                                                     args[1].implicit(),
+                                                    x_desc.get(),
+                                                    args[0].implicit(),
                                                     cd.get(),
-                                           algo,
-                                           &beta,
                                                     y_desc.get(),
                                                     args[3].implicit(),
                                                     args[2].implicit(),
-                                           args[2].get_shape().bytes());
+                                                    args[2].get_shape().bytes(),
+                                                    solution_id);
+
     if(status != miopenStatusSuccess)
-        MIGRAPHX_THROW("Running deconvolution failed");
+        MIGRAPHX_THROW("MIOpen Deconvolution: running convolution failed");
     return args[3];
 }
 
-shape miopen_deconvolution::compile(context& ctx,
-                                    const shape& output_shape,
-                                    std::vector<shape> inputs)
+shape miopen_deconvolution::find(context& ctx, const shape& output_shape, std::vector<shape> inputs)
 {
     shape workspace_shape{};
+
     auto x_desc = make_tensor(reshape_if_1d(inputs[0]));
     auto w_desc = make_tensor(reshape_if_1d(inputs[1]));
     auto y_desc = make_tensor(reshape_if_1d(output_shape));
@@ -119,9 +118,35 @@ shape miopen_deconvolution::compile(context& ctx,
                                                         workspace_size,
                                                         false);
     if(status != miopenStatusSuccess)
-        MIGRAPHX_THROW("Find deconvolution failed");
-    handle = ctx.get_stream().get_miopen();
+        MIGRAPHX_THROW("MIOpen Deconvolution: 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 Deconvolution: 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 Deconvolution: get solution failed");
+
+    solution_id = solutions.front().solution_id;
+
     return shape{shape::int8_type, {perf.memory}};
 }
 
@@ -129,13 +154,29 @@ void miopen_deconvolution::finalize(context& ctx,
                                     const shape& output_shape,
                                     std::vector<shape> inputs)
 {
-    if(handle == ctx.get_stream().get_miopen())
-        return;
+    if(cd == nullptr)
+        cd = make_deconv(op);
+    if(solution_id == 0)
+    {
         // Check that workspace hasn't changed
         auto size = inputs.at(2).bytes();
-    auto ws   = compile(ctx, output_shape, std::move(inputs));
+        auto ws   = find(ctx, output_shape, inputs);
         if(ws.bytes() > size)
-        MIGRAPHX_THROW("Workspace has changed during finalization.");
+            MIGRAPHX_THROW("MIOpen Deconvolution: workspace has changed during finalization.");
+    }
+
+    auto x_desc = make_tensor(reshape_if_1d(inputs[0]));
+    auto w_desc = make_tensor(reshape_if_1d(inputs[1]));
+    auto y_desc = make_tensor(reshape_if_1d(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 Deconvolution: compile solution failed");
 }
 
 } // namespace gpu

src/targets/gpu/fuse_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/fuse_mlir.hpp>
+#include <migraphx/gpu/mlir.hpp>
+#include <migraphx/matcher.hpp>
+#include <migraphx/pass_manager.hpp>
+#include <migraphx/make_op.hpp>
+#include <migraphx/register_op.hpp>
+
+namespace migraphx {
+inline namespace MIGRAPHX_INLINE_NS {
+
+struct module;
+
+namespace gpu {
+
+#ifdef MIGRAPHX_MLIR
+struct mlir_conv
+{
+    operation op = make_op("convolution");
+
+    template <class Self, class F>
+    static auto reflect(Self& self, F f)
+    {
+        return pack(f(self.op, "op"));
+    }
+
+    std::string name() const { return "gpu::mlir_conv"; }
+    shape compute_shape(std::vector<shape> inputs, const std::vector<module_ref>& mods) const
+    {
+        check_shapes{inputs, *this}.standard();
+        if(mods.size() != 1)
+            MIGRAPHX_THROW("should have one submodule.");
+        if(inputs.size() < 2)
+            MIGRAPHX_THROW("should have at least two inputs.");
+        auto n = inputs.size();
+        return op.compute_shape({inputs[n - 2], inputs[n - 1]});
+    }
+};
+MIGRAPHX_REGISTER_OP(mlir_conv);
+
+namespace {
+struct find_conv_pointwise
+{
+    // Find a convolution followed by a pointwise operation.
+    auto matcher() const
+    {
+        auto convolution =
+            match::skip(match::name("contiguous"))(match::name("convolution").bind("convolution"));
+        return match::name("pointwise")(match::any_of[match::inputs()](convolution.bind("x")));
+    }
+
+    void apply(module_pass_manager& mpm, const match::matcher_result& r) const
+    {
+        auto ins      = r.result;
+        auto conv_ins = r.instructions["convolution"];
+        auto x_ins    = r.instructions["x"]; // input after contiguous
+        auto* pm      = ins->module_inputs().front();
+        auto names    = pm->get_parameter_names();
+        // Whitelist pointwise operators
+        if(std::any_of(pm->begin(), pm->end(), [](const auto& i) {
+               return not contains({"@literal", "@param", "@return", "convolution", "add", "relu"},
+                                   i.name());
+           }))
+            return;
+        // Only fuse with fp32 for now
+        if(std::any_of(ins->inputs().begin(), ins->inputs().end(), [&](auto i) {
+               return i->get_shape().type() != shape::type_t::float_type;
+           }))
+            return;
+        std::sort(names.begin(), names.end());
+        module_ref mm = mpm.create_module("mlir_" + pm->name());
+        mm->set_bypass();
+        std::unordered_map<instruction_ref, instruction_ref> param_map;
+        auto x    = mm->add_parameter("x" + std::to_string(names.size()),
+                                   conv_ins->inputs().at(0)->get_shape());
+        auto w    = mm->add_parameter("x" + std::to_string(names.size() + 1),
+                                   conv_ins->inputs().at(1)->get_shape());
+        auto conv = mm->add_instruction(conv_ins->get_operator(), {x, w});
+        std::transform(names.begin(),
+                       names.end(),
+                       ins->inputs().begin(),
+                       std::inserter(param_map, param_map.end()),
+                       [&](auto name, auto input) {
+                           if(input == x_ins)
+                               return std::make_pair(pm->get_parameter(name), conv);
+                           return std::make_pair(pm->get_parameter(name),
+                                                 mm->add_parameter(name, input->get_shape()));
+                       });
+        mm->add_return(mm->insert_instructions(mm->end(), pm, param_map));
+
+        std::vector<instruction_ref> inputs;
+        std::copy_if(ins->inputs().begin(),
+                     ins->inputs().end(),
+                     std::back_inserter(inputs),
+                     [&](auto input) { return input != conv_ins; });
+        inputs.insert(inputs.end(), conv_ins->inputs().begin(), conv_ins->inputs().end());
+        mpm.get_module().replace_instruction(
+            ins, mlir_conv{conv_ins->get_operator()}, inputs, {mm});
+    }
+};
+} // namespace
+
+#endif
+
+void fuse_mlir::apply(module_pass_manager& mpm) const
+{
+#ifdef MIGRAPHX_MLIR
+    match::find_matches(mpm, find_conv_pointwise{});
+#else
+    (void)mpm;
+#endif
+}
+
+} // namespace gpu
+
+} // namespace MIGRAPHX_INLINE_NS
+} // namespace migraphx

src/targets/gpu/fuse_ops.cpp

@@ -336,6 +336,7 @@ void move_standard_front(std::vector<instruction_ref>& args)
 
 auto gpu_name(const std::string& s) { return match::name("gpu::" + s); }
 
+namespace {
 struct find_layernorm
 {
     auto matcher() const { return match::layernorm(&gpu_name); }
@@ -836,15 +837,6 @@ inline auto precompile_name(std::string s) // NOLINT
     });
 }
 
-template <class... Ms>
-auto conv_bias_pointwise(Ms... ms)
-{
-    return precompile_name("pointwise")(
-        match::either_arg(0, 1)(bias_shape(match::used_once()).bind("bias"),
-                                fusable_conv(match::used_once()).bind("conv")),
-        ms...);
-}
-
 struct find_conv_bias
 {
     context* ctx = nullptr;
@@ -1013,6 +1005,7 @@ struct find_commutative_broadcast
         m.replace_instruction(ins, ins->get_operator(), args);
     }
 };
+} // namespace
 
 struct find_contiguous
 {