Implement concat

src/targets/gpu/jit/concat.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/compiler.hpp>
+#include <migraphx/gpu/context.hpp>
+#include <migraphx/gpu/compile_hip_code_object.hpp>
+#include <migraphx/gpu/compile_hip.hpp>
+#include <migraphx/gpu/compile_gen.hpp>
+#include <migraphx/reduce_dims.hpp>
+
+namespace migraphx {
+inline namespace MIGRAPHX_INLINE_NS {
+namespace gpu {
+
+using namespace migraphx::gpu::gen; // NOLINT
+
+// NOLINTNEXTLINE
+static const char* const concat_kernel = R"__migraphx__(
+#include <migraphx/kernels/concat.hpp>
+#include <migraphx/kernels/vectorize.hpp>
+#include <args.hpp>
+
+namespace migraphx {
+
+extern "C" {
+
+__global__ void ${kernel}(${params}) 
+{
+    transform_args(make_tensors(), rotate_last(), ${transformers})(${args})([](auto y, auto... xs) {
+        concat<${axis}>(y, xs...);
+    });
+}
+
+}
+
+} // namespace migraphx
+
+)__migraphx__";
+
+struct concat_compiler : compiler<concat_compiler>
+{
+    std::vector<std::string> names() const { return {"concat"}; }
+
+    static std::size_t get_min_elements(const std::vector<shape>& inputs)
+    {
+        auto it = std::min_element(inputs.begin(), inputs.end(), by(std::less<>{}, [](auto s) {
+            return s.elements();
+        }));
+        return it->elements();
+    }
+
+    operation compile_op(context& ctx, const std::vector<shape>& inputs, const value& v) const
+    {
+        hip_compile_options options;
+        options.inputs         = inputs;
+        options.output         = inputs.back();
+        options.virtual_inputs = reduce_dims(inputs);
+        options.params         = "-Wno-float-equal";
+        auto axis              = find_fast_axis(options.virtual_inputs);
+        auto vec               = vectorize::elements(axis, options.virtual_inputs);
+        options.kernel_name    = v.get("kernel", "concat_kernel");
+        options.set_launch_params(
+            v,
+            compute_global_for(ctx,
+                               get_min_elements(options.inputs) / vec.size, 256));
+        auto src = interpolate_string(concat_kernel,
+                                      {{"kernel", options.kernel_name},
+                                       {"params", enum_params(inputs.size(), "void * private_p")},
+                                       {"args", enum_params(inputs.size(), "private_p")},
+                                       {"transformers", make_transformer_args(vec)},
+                                        {"axis", v.at("axis").to<std::string>()}});
+        return compile_hip_code_object(src, options);
+    }
+
+    compiler_replace compile(context& ctx, instruction_ref ins, const operation& op) const
+    {
+        return replace(compile_op(ctx, to_shapes(ins->inputs()), op.to_value()));
+    }
+};
+
+} // namespace gpu
+} // namespace MIGRAPHX_INLINE_NS
+} // namespace migraphx

src/targets/gpu/kernels/include/migraphx/kernels/concat.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.
+ */
+
+#include <migraphx/kernels/index.hpp>
+#include <migraphx/kernels/functional.hpp>
+#include <migraphx/kernels/tensor_view.hpp>
+
+#ifndef MIGRAPHX_GUARD_KERNELS_CONCAT_HPP
+#define MIGRAPHX_GUARD_KERNELS_CONCAT_HPP
+
+namespace migraphx {
+
+template<index_int Axis, class Output, class Input, class Start>
+constexpr auto concat_slice(Output out, Input, Start)
+{
+    constexpr auto lens = get_shape_c<Input>{}.lens;
+    constexpr auto strides = get_shape_c<Output>{}.strides;
+    constexpr auto offset = return_c([] {
+        constexpr auto output_shape = get_shape_c<Output>{};
+        return Start{} * output_shape.strides[Axis];
+    });
+    constexpr auto s = make_shape(lens, strides);
+    return make_tensor_view(&out[offset], s);
+}
+
+template<index_int Axis, class Input>
+constexpr auto concat_ends(Input)
+{
+    constexpr auto lens = get_shape_c<Input>{}.lens;
+    return _c<lens[Axis]>;
+}
+
+template<index_int Axis, class Output, class... Inputs>
+__device__ void concat(Output output, Inputs... inputs)
+{
+    auto idx = make_index();
+    fold([&](auto start, auto input) {
+        auto y = concat_slice<Axis>(output, input, start);
+        idx.global_stride(input.get_shape().elements(), [&](auto i) {
+            y[i] = input[i];
+        });
+        return start + concat_ends<Axis>(input);
+    })(_c<0>, inputs...);
+}
+
+} // namespace migraphx
+#endif // MIGRAPHX_GUARD_KERNELS_CONCAT_HPP

src/targets/gpu/lowering.cpp

@@ -169,7 +169,6 @@ struct miopen_apply
         add_extend_op("argmax");
         add_extend_op("argmin");
         add_extend_op("clip");
-        add_extend_op("concat");
         add_extend_op("convert");
         add_extend_op("elu");
         add_extend_op("gather");