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

src/eliminate_contiguous.cpp

@@ -27,6 +27,13 @@ void eliminate_contiguous::apply(program& p) const
 {
     for(auto ins : iterator_for(p))
     {
+        // skip the reshape operator for now, since there is a bug
+        // for the transpose followed by a reshape
+        if(ins->name() == "reshape")
+        {
+            continue;
+        }
+
         // Make a copy so we can modify it while we iterate
         auto args = ins->inputs();
         for(auto arg : ins->inputs())

src/include/migraphx/literal.hpp

@@ -22,8 +22,8 @@ struct literal : raw_data<literal>
 {
     literal() {}
 
-    template <class U, class T = deduce<U>>
-    literal(U x) : buffer(make_shared_array<char>(sizeof(T))), m_shape(shape::get_type<T>{})
+    template <class U, class T = deduce<U>, shape::type_t ShapeType = shape::get_type<T>{}>
+    literal(U x) : buffer(make_shared_array<char>(sizeof(T))), m_shape(ShapeType)
     {
         static_assert(std::is_trivially_copyable<T>{}, "Literals can only be trivial types");
         *(reinterpret_cast<T*>(buffer.get())) = x;

src/include/migraphx/operators.hpp

@@ -758,42 +758,48 @@ struct gather
         int axis_index = (axis < 0) ? (n_dim + axis) : axis;
 
         auto type = inputs[0].type();
-        lens[axis_index] = inputs[1].elements();
-
-        return {type, lens};
+        lens.erase(lens.begin() + axis_index);
+        if(!inputs[1].scalar())
+        {
+            auto ind_lens = inputs[1].lens();
+            lens.insert(lens.begin() + axis_index, ind_lens.begin(), ind_lens.end());
         }
 
-    template <class T>
-    void compute_index(const T& out_idx,
-                       const int axis_index,
-                       const std::vector<std::size_t>& vec_indices,
-                       const std::size_t max_dim,
-                       T& in_idx) const
+        // for scalar output
+        if(lens.empty())
         {
-        in_idx          = out_idx;
-        std::size_t idx = vec_indices.at(out_idx[axis_index]);
-        if(idx >= max_dim)
-        {
-            MIGRAPHX_THROW("Gather: indices are out of range in input tensor");
+            return {type};
         }
-        in_idx[axis_index] = idx;
+
+        return {type, lens};
     }
 
     argument compute(const shape& output_shape, std::vector<argument> args) const
     {
         argument result{output_shape};
         // negative axis means counting dimensions from back
-        int axis_index = (axis < 0) ? (output_shape.lens().size() + axis) : axis;
+        int axis_index =
+            (axis < 0) ? static_cast<int>(args[0].get_shape().lens().size() + axis) : axis;
 
         // max dimension in axis
-        std::size_t max_dim = args[0].get_shape().lens()[axis_index];
-        std::vector<std::size_t> vec_indices;
-        args[1].visit([&](auto indices) { vec_indices.assign(indices.begin(), indices.end()); });
-        visit_all(result, args[0])([&](auto output, auto input) {
-            std::vector<std::size_t> in_idx;
-            shape_for_each(output.get_shape(), [&](const auto& idx) {
-                this->compute_index(idx, axis_index, vec_indices, max_dim, in_idx);
-                output(idx.begin(), idx.end()) = input(in_idx.begin(), in_idx.end());
+        visit_all(result, args[0])([&](auto output, auto data) {
+            args[1].visit([&](auto indices) {
+                if(output_shape.scalar())
+                {
+                    output[0] = data[indices.front()];
+                }
+                else
+                {
+                    auto out_lens        = data.get_shape().lens();
+                    out_lens[axis_index] = indices.get_shape().elements();
+                    migraphx::shape out_comp_shape{data.get_shape().type(), out_lens};
+                    shape_for_each(out_comp_shape, [&](const auto& out_idx) {
+                        auto data_idx        = out_idx;
+                        data_idx[axis_index] = indices[data_idx[axis_index]];
+                        output[out_comp_shape.index(out_idx.begin(), out_idx.end())] =
+                            data(data_idx.begin(), data_idx.end());
+                    });
+                }
             });
         });
 
@@ -820,10 +826,22 @@ struct dot
         const shape& b = inputs.at(1);
         auto t         = a.type();
 
-        if(a.lens()[1] != b.lens()[0])
+        // according to the specification of the numpy.matmul()
+        // inputs with the shape dims more than 2 are acceptable
+        // as long as dim values are the same in the two inputs
+        if(!std::equal(a.lens().rbegin() + 2, a.lens().rend(), b.lens().rbegin() + 2))
+        {
+            MIGRAPHX_THROW("DOT: dim values mismatch");
+        }
+
+        std::size_t dim_0 = a.lens().size() - 2;
+        std::size_t dim_1 = a.lens().size() - 1;
+        if(a.lens()[dim_1] != b.lens()[dim_0])
             MIGRAPHX_THROW("Inner dimensions do not match: {" + to_string_range(a.lens()) +
                            "} x {" + to_string_range(b.lens()) + "}");
-        return {t, {a.lens()[0], b.lens()[1]}};
+        auto out_lens   = a.lens();
+        out_lens[dim_1] = b.lens()[dim_1];
+        return {t, out_lens};
     }
 };
 
@@ -932,6 +950,22 @@ struct softmax
     }
 };
 
+struct logsoftmax
+{
+    int axis = 1;
+    std::string name() const { return "logsoftmax"; }
+    shape compute_shape(std::vector<shape> inputs) const
+    {
+        check_shapes{inputs}.has(1);
+        if(axis < 0 || axis > inputs[0].lens().size())
+        {
+            MIGRAPHX_THROW("LogSoftMax: input axis value " + std::to_string(axis) +
+                           " is out of range");
+        }
+        return inputs.at(0);
+    }
+};
+
 struct flatten
 {
     uint64_t axis = 0;
@@ -1259,6 +1293,57 @@ struct gru
     }
 };
 
+struct lstm
+{
+    std::size_t hidden_size = 1;
+    std::vector<operation> actv_funcs{sigmoid{}, tanh{}, tanh{}};
+    rnn_direction direction = rnn_direction::forward;
+    float clip              = 0.0f;
+    int input_forget        = 0;
+
+    std::string name() const { return "lstm"; }
+    shape compute_shape(std::vector<shape> inputs) const
+    {
+        auto in_dims     = inputs[0].lens();
+        auto hidden_dims = inputs[2].lens();
+        if(hidden_size != hidden_dims[2])
+        {
+            MIGRAPHX_THROW("LSTM: hidden size mismatch in attribute and input");
+        }
+
+        std::size_t num_directions = 1;
+        if(direction == rnn_direction::bidirectional)
+        {
+            num_directions = 2;
+        }
+
+        if(num_directions != hidden_dims[0])
+        {
+            MIGRAPHX_THROW("LSTM: num_direction does not match the direction attribute");
+        }
+
+        std::vector<std::size_t> out_dims(in_dims);
+        out_dims.insert(out_dims.begin() + 1, num_directions);
+        out_dims.back() = hidden_size;
+
+        return {inputs[0].type(), out_dims};
+    }
+};
+
+struct lstm_last_cell_output
+{
+    std::string name() const { return "lstm_last_cell_output"; }
+    shape compute_shape(std::vector<shape> inputs) const
+    {
+        check_shapes{inputs, *this}.has(1);
+        auto dims = inputs[0].lens();
+
+        // remove the first dimension, remaing are output shape
+        dims.erase(dims.begin());
+        return {inputs[0].type(), dims};
+    }
+};
+
 struct undefined
 {
     std::string name() const { return "undefined"; }

src/include/migraphx/rewrite_rnn.hpp

@@ -45,6 +45,18 @@ struct rewrite_rnn
                                           const operation& actv_func2) const;
 
     std::vector<operation> gru_actv_funcs(instruction_ref ins) const;
+
+    // for lstm operators
+    void apply_lstm(program& prog, instruction_ref ins) const;
+    std::vector<instruction_ref> lstm_cell(bool is_forward,
+                                           program& prog,
+                                           instruction_ref ins,
+                                           std::vector<instruction_ref> inputs,
+                                           const operation& actv_func1,
+                                           const operation& actv_func2,
+                                           const operation& actv_func3) const;
+
+    std::vector<operation> lstm_actv_funcs(instruction_ref ins) const;
 };
 
 } // namespace MIGRAPHX_INLINE_NS

src/onnx/onnx.cpp

@@ -79,6 +79,7 @@ struct onnx_parser
         add_mem_op("Gemm", &onnx_parser::parse_gemm);
         add_mem_op("BatchNormalization", &onnx_parser::parse_batchnorm);
         add_mem_op("Softmax", &onnx_parser::parse_softmax);
+        add_mem_op("LogSoftmax", &onnx_parser::parse_logsoftmax);
         add_mem_op("Squeeze", &onnx_parser::parse_squeeze);
         add_mem_op("Unsqueeze", &onnx_parser::parse_unsqueeze);
         add_mem_op("Slice", &onnx_parser::parse_slice);
@@ -89,6 +90,7 @@ struct onnx_parser
         add_mem_op("Transpose", &onnx_parser::parse_transpose);
         add_mem_op("RNN", &onnx_parser::parse_rnn);
         add_mem_op("GRU", &onnx_parser::parse_gru);
+        add_mem_op("LSTM", &onnx_parser::parse_lstm);
         add_mem_op("Pad", &onnx_parser::parse_pad);
 
         // init the activation function map
@@ -227,6 +229,19 @@ struct onnx_parser
         return prog.add_instruction(op::reshape{{long(dims[0]), long(dims[1])}}, s);
     }
 
+    instruction_ref parse_logsoftmax(const std::string&,
+                                     const attribute_map& attributes,
+                                     std::vector<instruction_ref> args)
+    {
+        int axis = 1;
+        if(contains(attributes, "axis"))
+        {
+            axis = parse_value(attributes.at("axis")).at<int>();
+        }
+
+        return prog.add_instruction(op::logsoftmax{axis}, std::move(args));
+    }
+
     instruction_ref
     parse_conv(const std::string&, attribute_map attributes, std::vector<instruction_ref> args)
     {
@@ -354,7 +369,9 @@ struct onnx_parser
         }
         if(args.size() == 2)
         {
-            literal s = args[1]->get_literal();
+            auto s = args[1]->eval();
+            if(s.empty())
+                MIGRAPHX_THROW("Dynamic shape is not supported.");
             s.visit([&](auto v) { copy(v, std::back_inserter(op.dims)); });
         }
         return prog.add_instruction(op, args[0]);
@@ -434,6 +451,14 @@ struct onnx_parser
                                    const std::vector<instruction_ref>&)
     {
         literal v     = parse_value(attributes.at("value"));
+        auto dim_size = attributes.at("value").t().dims_size();
+        // if dim_size is 0, it is a scalar
+        if(dim_size == 0)
+        {
+            migraphx::shape scalar_shape{v.get_shape().type()};
+            return prog.add_literal(migraphx::literal{scalar_shape, v.data()});
+        }
+
         return prog.add_literal(v);
     }
 
@@ -460,7 +485,12 @@ struct onnx_parser
         {
             transb = parse_value(attributes.at("transB")).at<bool>();
         }
-        std::vector<int64_t> perm = {1, 0};
+
+        std::vector<int64_t> perm(args[0]->get_shape().lens().size());
+        std::iota(perm.begin(), perm.end(), int64_t{0});
+        // swap the last two elements
+        std::swap(*perm.rbegin(), *(perm.rbegin() + 1));
+
         auto l1 = (transa) ? prog.add_instruction(op::transpose{perm}, args[0]) : args[0];
         auto l2 = (transb) ? prog.add_instruction(op::transpose{perm}, args[1]) : args[1];
         if(args.size() == 3)
@@ -480,6 +510,7 @@ struct onnx_parser
                 return add_broadcastable_binary_op(l3, l4, op::add{});
             }
         }
+
         return prog.add_instruction(op::dot{alpha, beta}, l1, l2);
     }
 
@@ -749,15 +780,17 @@ struct onnx_parser
         {
             auto names = attributes.at("activations").strings();
             vec_names.clear();
-            for_each(names.begin(), names.end(), [&](auto& fn) { vec_names.push_back(fn); });
+            vec_names.resize(names.size());
+            std::copy(names.begin(), names.end(), vec_names.begin());
         }
 
-        for_each(vec_names.begin(), vec_names.end(), [&](auto& fn) {
-            if(map_actv_funcs.count(fn) == 0)
+        auto name_it = std::find_if(vec_names.begin(), vec_names.end(), [&](auto& name) {
+            return (map_actv_funcs.count(name) == 0);
+        });
+        if(name_it != vec_names.end())
         {
-                MIGRAPHX_THROW("RNN: activation function " + std::string(fn) + " not supported");
+            MIGRAPHX_THROW("RNN: activation function " + std::string(*name_it) + " not supported");
         }
-        });
 
         // bidirectional case should have two activation functions.
         // one is for forward, and the other is for reverse.
@@ -839,8 +872,7 @@ struct onnx_parser
             auto names = attributes.at("activations").strings();
             vec_names.clear();
             vec_names.resize(names.size());
-            std::transform(
-                names.begin(), names.end(), vec_names.begin(), [](auto& str) { return str; });
+            std::copy(names.begin(), names.end(), vec_names.begin());
         }
 
         // need 4 activation functions
@@ -878,12 +910,13 @@ struct onnx_parser
             }
         }
 
-        for_each(vec_names.begin(), vec_names.end(), [&](auto& name) {
-            if(map_actv_funcs.count(name) == 0)
+        auto name_it = std::find_if(vec_names.begin(), vec_names.end(), [&](auto& name) {
+            return (map_actv_funcs.count(name) == 0);
+        });
+        if(name_it != vec_names.end())
         {
-                MIGRAPHX_THROW("GRU: activation function " + std::string(name) + " not supported");
+            MIGRAPHX_THROW("GRU: activation function " + std::string(*name_it) + " not supported");
         }
-        });
 
         std::vector<operation> vec_actv_funcs(vec_names.size());
         std::transform(vec_names.begin(), vec_names.end(), vec_actv_funcs.begin(), [&](auto& name) {
@@ -920,6 +953,178 @@ struct onnx_parser
         return {hidden_states, last_output};
     }
 
+    std::vector<instruction_ref>
+    parse_lstm(const std::string&, attribute_map attributes, std::vector<instruction_ref> args)
+    {
+        migraphx::shape input_shape = args[0]->get_shape();
+        std::size_t hidden_size     = args[2]->get_shape().lens()[2];
+
+        if(contains(attributes, "hidden_size"))
+        {
+            std::size_t hidden_size_att = parse_value(attributes.at("hidden_size")).at<int>();
+            if(hidden_size != hidden_size_att)
+            {
+                MIGRAPHX_THROW("LSTM: hidden size mismatch in input and attribute");
+            }
+        }
+
+        // Handling of direction to be added later
+        std::string direction{"forward"};
+        if(contains(attributes, "direction"))
+        {
+            direction = attributes.at("direction").s();
+        }
+
+        op::rnn_direction dirct = op::rnn_direction::forward;
+        if(direction == "bidirectional")
+        {
+            dirct = op::rnn_direction::bidirectional;
+        }
+        else if(direction == "reverse")
+        {
+            dirct = op::rnn_direction::reverse;
+        }
+        else if(direction == "forward")
+        {
+            dirct = op::rnn_direction::forward;
+        }
+        else
+        {
+            MIGRAPHX_THROW("LSTM: incorrect direction attribute");
+        }
+
+        std::vector<std::string> vec_names = {"sigmoid", "tanh", "tanh"};
+        if(contains(attributes, "activations"))
+        {
+            auto names = attributes.at("activations").strings();
+            vec_names.clear();
+            vec_names.resize(names.size());
+            std::copy(names.begin(), names.end(), vec_names.begin());
+        }
+
+        // need 6 activation functions for bidirectional directions
+        if(dirct == op::rnn_direction::bidirectional)
+        {
+            // 6 activation functions are used in the bidirectional
+            // scenario. No spec is provided in onnx::operator. we
+            // use the algorithm that: if 1 actv function is provided,
+            // repeat 1st six times. If 2 actv functins are provided,
+            // repeat 2nd once, then repeat all three once
+            // if 3 actv funcs are provide, repeat all three once.
+            // the same algorithm is used for 4, 5, and 6 actv funcions
+            // provided. This may need change later
+            switch(vec_names.size())
+            {
+            case 1:
+                vec_names = {vec_names.at(0),
+                             vec_names.at(0),
+                             vec_names.at(0),
+                             vec_names.at(0),
+                             vec_names.at(0),
+                             vec_names.at(0)};
+                break;
+
+            case 2:
+                // repeat the 2nd actv func once, then repeat all three another time
+                vec_names = {vec_names.at(0),
+                             vec_names.at(1),
+                             vec_names.at(1),
+                             vec_names.at(0),
+                             vec_names.at(1),
+                             vec_names.at(1)};
+                break;
+
+            case 3:
+                // repeat all three actv funcs once
+                vec_names = {vec_names.at(0),
+                             vec_names.at(1),
+                             vec_names.at(2),
+                             vec_names.at(0),
+                             vec_names.at(1),
+                             vec_names.at(2)};
+                break;
+
+            case 4:
+                vec_names = {vec_names.at(0),
+                             vec_names.at(1),
+                             vec_names.at(2),
+                             vec_names.at(3),
+                             vec_names.at(3),
+                             vec_names.at(3)};
+                break;
+
+            case 5:
+                vec_names = {vec_names.at(0),
+                             vec_names.at(1),
+                             vec_names.at(2),
+                             vec_names.at(3),
+                             vec_names.at(4),
+                             vec_names.at(4)};
+                break;
+
+            default: break;
+            }
+        }
+        else
+        {
+            switch(vec_names.size())
+            {
+            case 1: vec_names = {vec_names.at(0), vec_names.at(0), vec_names.at(0)}; break;
+
+            case 2:
+                // repeat the 2nd actv func once, so we have 3 actv funcs
+                vec_names = {vec_names.at(0), vec_names.at(1), vec_names.at(1)};
+                break;
+
+            default: break;
+            }
+        }
+
+        auto name_it = std::find_if(vec_names.begin(), vec_names.end(), [&](auto& name) {
+            return (map_actv_funcs.count(name) == 0);
+        });
+        if(name_it != vec_names.end())
+        {
+            MIGRAPHX_THROW("LSTM: activation function " + std::string(*name_it) + " not supported");
+        }
+
+        std::vector<operation> vec_actv_funcs(vec_names.size());
+        std::transform(vec_names.begin(), vec_names.end(), vec_actv_funcs.begin(), [&](auto& name) {
+            return map_actv_funcs[name];
+        });
+
+        float clip = 0.0;
+        if(contains(attributes, "clip"))
+        {
+            clip = parse_value(attributes.at("clip")).at<float>();
+        }
+
+        int input_forget = 0;
+        if(contains(attributes, "input_forget"))
+        {
+            input_forget = parse_value(attributes.at("input_forget")).at<int>();
+        }
+
+        // append undefined opeator to make 6 arguments
+        if(args.size() < 8)
+        {
+            auto ins = prog.add_instruction(op::undefined{});
+            args.insert(args.end(), 8 - args.size(), ins);
+        }
+
+        // first output for concatenation of hidden states
+        auto hidden_states = prog.add_instruction(
+            op::lstm{hidden_size, vec_actv_funcs, dirct, clip, input_forget}, std::move(args));
+
+        // second output for last lstm output
+        auto last_output = prog.add_instruction(op::rnn_last_output{}, hidden_states);
+
+        // third output for last cell output
+        auto last_cell_output = prog.add_instruction(op::lstm_last_cell_output{}, hidden_states);
+
+        return {hidden_states, last_output, last_cell_output};
+    }
+
     void parse_from(std::istream& is)
     {
         onnx::ModelProto model;
@@ -960,9 +1165,9 @@ struct onnx_parser
                 instructions[name] = prog.add_parameter(name, s);
             }
         }
-        for(auto&& p : nodes)
+        for(auto&& output : graph.output())
         {
-            this->parse_node(p.first);
+            this->parse_node(output.name());
         }
     }
 

src/program.cpp

@@ -2,6 +2,7 @@
 #include <migraphx/stringutils.hpp>
 #include <migraphx/instruction.hpp>
 #include <migraphx/operators.hpp>
+#include <migraphx/target.hpp>
 #include <migraphx/env.hpp>
 #include <migraphx/ranges.hpp>
 #include <migraphx/time.hpp>

src/shape.cpp

@@ -19,7 +19,7 @@ struct shape_impl
 
     shape_impl() : m_type(shape::float_type), m_standard(false) {}
 
-    shape_impl(shape::type_t t) : m_type(t), m_lens({1}), m_strides({1}), m_standard(true) {}
+    shape_impl(shape::type_t t) : m_type(t), m_lens({1}), m_strides({0}), m_standard(true) {}
     shape_impl(shape::type_t t, std::vector<std::size_t> l)
         : m_type(t), m_lens(std::move(l)), m_standard(true)
     {

src/targets/cpu/gemm.cpp

 #include <migraphx/cpu/gemm.hpp>
 #include <migraphx/dfor.hpp>
 #include <migraphx/requires.hpp>
+#include <migraphx/shape_for_each.hpp>
 #include <blaze/math/CustomMatrix.h>
 
 namespace migraphx {
@@ -14,10 +15,13 @@ template <class T>
 static auto make_mat(tensor_view<T> x)
 {
     const auto& s = x.get_shape();
-    assert(s.lens().size() == 2);
+    // assert(s.lens().size() == 2);
+    std::size_t n_dims = s.lens().size();
+    std::size_t dim_0  = n_dims - 2;
+    std::size_t dim_1  = n_dims - 1;
     if(s.transposed())
-        return matrix<T>{x.data(), s.lens()[1], s.lens()[0], s.strides()[1]};
-    return matrix<T>{x.data(), s.lens()[0], s.lens()[1], s.strides()[0]};
+        return matrix<T>{x.data(), s.lens()[dim_1], s.lens()[dim_0], s.strides()[dim_1]};
+    return matrix<T>{x.data(), s.lens()[dim_0], s.lens()[dim_1], s.strides()[dim_0]};
 }
 
 template <class T, class F>
@@ -64,18 +68,24 @@ void migemm_impl(tensor_view<T> cmat,
                  float beta,
                  std::false_type)
 {
-    auto m = cmat.get_shape().lens()[0];
-    auto n = cmat.get_shape().lens()[1];
-    auto k = amat.get_shape().lens()[1];
+    std::size_t n_dims = cmat.get_shape().lens().size();
+    std::size_t dim_0  = n_dims - 2;
+    std::size_t dim_1  = n_dims - 1;
+    auto k             = amat.get_shape().lens()[dim_1];
 
-    assert(amat.get_shape().lens()[1] == bmat.get_shape().lens()[0]);
-    assert(m == amat.get_shape().lens()[0]);
-    assert(n == bmat.get_shape().lens()[1]);
+    assert(amat.get_shape().lens()[dim_1] == bmat.get_shape().lens()[dim_0]);
+    assert(cmat.get_shape().lens()[dim_0] == amat.get_shape().lens()[dim_0]);
+    assert(cmat.get_shape().lens()[dim_1] == bmat.get_shape().lens()[dim_1]);
 
-    dfor(m, n)([&](auto ii, auto jj) {
-        double s = cmat(ii, jj) * beta;
-        dfor(k)([&](auto kk) { s += amat(ii, kk) * bmat(kk, jj); });
-        cmat(ii, jj) = alpha * s;
+    shape_for_each(cmat.get_shape(), [&](const auto& c_idx) {
+        auto a_idx = c_idx;
+        auto b_idx = c_idx;
+        double s   = 0.0;
+        dfor(k)([&](auto kk) {
+            a_idx[dim_1] = b_idx[dim_0] = kk;
+            s += amat(a_idx.begin(), a_idx.end()) * bmat(b_idx.begin(), b_idx.end());
+        });
+        cmat(c_idx.begin(), c_idx.end()) = alpha * s + cmat(c_idx.begin(), c_idx.end()) * beta;
     });
 }
 
@@ -83,7 +93,18 @@ template <class T>
 void migemm_impl(
     tensor_view<T> cmat, tensor_view<T> amat, tensor_view<T> bmat, float alpha, float beta)
 {
+    auto lens = amat.get_shape().lens();
+    bool batch_mul =
+        std::accumulate(lens.begin(), lens.end(), std::size_t{1}, std::multiplies<std::size_t>()) ==
+        (*lens.rbegin()) * (*(lens.rbegin() + 1));
+    if(batch_mul)
+    {
         migemm_impl(cmat, amat, bmat, alpha, beta, is_fast_gemm_type<T>{});
+    }
+    else
+    {
+        migemm_impl(cmat, amat, bmat, alpha, beta, std::false_type{});
+    }
 }
 
 void migemm(

src/targets/cpu/include/migraphx/cpu/target.hpp

@@ -7,6 +7,7 @@
 
 namespace migraphx {
 inline namespace MIGRAPHX_INLINE_NS {
+struct pass;
 namespace cpu {
 
 struct target

src/targets/cpu/lowering.cpp

@@ -613,6 +613,75 @@ struct softmax2d
     }
 };
 
+struct cpu_logsoftmax
+{
+    op::logsoftmax op;
+    std::string name() const { return "cpu::logsoftmax"; }
+    shape compute_shape(const std::vector<shape>& inputs) const { return op.compute_shape(inputs); }
+
+    template <typename T>
+    std::size_t compute_batch_index(const T& idx, shape& batch_shape, int axis) const
+    {
+        if(axis == 0)
+        {
+            return 0;
+        }
+        else
+        {
+            std::vector<std::size_t> batch_idx(idx.begin(), idx.begin() + axis);
+            return batch_shape.index(batch_idx.begin(), batch_idx.end());
+        }
+    }
+
+    argument compute(context&, const shape& output_shape, std::vector<argument> args) const
+    {
+        argument result{output_shape};
+        auto lens = output_shape.lens();
+        std::vector<std::size_t> batch_lens{};
+        if(op.axis == 0)
+        {
+            batch_lens.push_back(1);
+        }
+        else
+        {
+            batch_lens.insert(batch_lens.begin(), lens.begin(), lens.begin() + op.axis);
+        }
+        shape batch_shape{migraphx::shape::uint32_type, batch_lens};
+        visit_all(result, args[0])([&](auto output, auto input) {
+            using value_type = typename decltype(input)::value_type;
+            std::vector<value_type> batch_max(batch_shape.elements(),
+                                              std::numeric_limits<value_type>::lowest());
+            shape_for_each(output_shape, [&](auto idx) {
+                auto index       = this->compute_batch_index(idx, batch_shape, op.axis);
+                batch_max[index] = std::max(batch_max[index], input(idx.begin(), idx.end()));
+            });
+
+            shape_for_each(output_shape, [&](auto idx) {
+                auto index = this->compute_batch_index(idx, batch_shape, op.axis);
+                output(idx.begin(), idx.end()) = input(idx.begin(), idx.end()) - batch_max[index];
+            });
+
+            std::vector<value_type> batch_sum(batch_shape.elements(), value_type(0));
+            shape_for_each(output_shape, [&](auto idx) {
+                auto index = this->compute_batch_index(idx, batch_shape, op.axis);
+                batch_sum[index] += std::exp(output(idx.begin(), idx.end()));
+            });
+
+            for(std::size_t i = 0; i < batch_sum.size(); ++i)
+            {
+                batch_sum[i] = std::log(batch_sum[i]);
+            }
+
+            shape_for_each(output_shape, [&](auto idx) {
+                auto index = this->compute_batch_index(idx, batch_shape, op.axis);
+                output(idx.begin(), idx.end()) -= batch_sum[index];
+            });
+        });
+
+        return result;
+    }
+};
+
 struct add_op
 {
     std::string name() const { return "add"; }
@@ -723,6 +792,7 @@ struct cpu_apply
         apply_map["pad"]        = extend_op<cpu_pad, op::pad>();
         apply_map["concat"]     = extend_op<cpu_concat, op::concat>();
         apply_map["gather"]     = extend_op<cpu_gather, op::gather>();
+        apply_map["logsoftmax"] = extend_op<cpu_logsoftmax, op::logsoftmax>();
         apply_map["leaky_relu"] = extend_op<cpu_unary<leaky_relu_op>, op::leaky_relu>();
         apply_map["elu"]        = extend_op<cpu_unary<elu_op>, op::elu>();
         apply_map["identity"]   = simple_op<cpu_unary<identity_op>>();

src/targets/cpu/target.cpp

 
 #include <migraphx/cpu/target.hpp>
 #include <migraphx/cpu/lowering.hpp>
+#include <migraphx/pass.hpp>
 #include <migraphx/auto_contiguous.hpp>
 #include <migraphx/rewrite_rnn.hpp>
 #include <migraphx/dead_code_elimination.hpp>

src/targets/gpu/CMakeLists.txt

@@ -26,6 +26,7 @@ add_library(migraphx_device
     device/atan.cpp
     device/add_relu.cpp
     device/contiguous.cpp
+    device/logsoftmax.cpp
     device/mul.cpp
     device/concat.cpp
     device/pad.cpp
@@ -48,6 +49,7 @@ add_library(migraphx_gpu
     pooling.cpp
     convolution.cpp
     softmax.cpp
+    logsoftmax.cpp
     contiguous.cpp
     concat.cpp
     relu.cpp

src/targets/gpu/abs.cpp

 #include <migraphx/gpu/abs.hpp>
-#include <migraphx/operators.hpp>
-#include <migraphx/manage_ptr.hpp>
-#include <migraphx/gpu/miopen.hpp>
-#include <utility>
+#include <migraphx/gpu/context.hpp>
 
 namespace migraphx {
 inline namespace MIGRAPHX_INLINE_NS {

src/targets/gpu/batchnorm.cpp

 #include <migraphx/gpu/batchnorm.hpp>
-#include <migraphx/operators.hpp>
-#include <migraphx/manage_ptr.hpp>
-#include <migraphx/gpu/miopen.hpp>
-#include <utility>
+#include <migraphx/gpu/context.hpp>
 
 namespace migraphx {
 inline namespace MIGRAPHX_INLINE_NS {

src/targets/gpu/concat.cpp

 #include <migraphx/gpu/concat.hpp>
-#include <migraphx/operators.hpp>
-#include <migraphx/manage_ptr.hpp>
-#include <migraphx/gpu/miopen.hpp>
+#include <migraphx/gpu/context.hpp>
 #include <migraphx/gpu/device/concat.hpp>
-#include <utility>
 
 namespace migraphx {
 inline namespace MIGRAPHX_INLINE_NS {

src/targets/gpu/contiguous.cpp

 #include <migraphx/gpu/contiguous.hpp>
-#include <migraphx/operators.hpp>
-#include <migraphx/manage_ptr.hpp>
-#include <migraphx/gpu/miopen.hpp>
-#include <utility>
+#include <migraphx/gpu/context.hpp>
+#include <migraphx/gpu/device/contiguous.hpp>
 
 namespace migraphx {
 inline namespace MIGRAPHX_INLINE_NS {

src/targets/gpu/convolution.cpp

 #include <migraphx/gpu/convolution.hpp>
-#include <migraphx/operators.hpp>
-#include <migraphx/manage_ptr.hpp>
-#include <migraphx/gpu/miopen.hpp>
-#include <utility>
+#include <migraphx/gpu/context.hpp>
+#include <migraphx/generate.hpp>
 
 namespace migraphx {
 inline namespace MIGRAPHX_INLINE_NS {

src/targets/gpu/device/gather.cpp

@@ -16,20 +16,24 @@ argument gather(hipStream_t stream,
                 std::vector<migraphx::argument> args,
                 int axis)
 {
-    int axis_index = (axis < 0) ? (axis + output_shape.lens().size()) : axis;
+    int axis_index = (axis < 0) ? (axis + args[0].get_shape().lens().size()) : axis;
     visit_all(args.back(), args[0])([&](auto output, auto input) {
         std::size_t nelements = output_shape.elements();
         args[1].visit([&](auto indices) {
-            visit_tensor_size(output_shape.lens().size(), [&](auto ndim) {
             const auto* indices_ptr = device_cast(indices.data());
-                auto* outptr            = device_cast(output.data());
-                const auto* inptr       = device_cast(input.data());
-                hip_tensor_descriptor<ndim> desc_input(input.get_shape());
-                hip_tensor_descriptor<ndim> desc_output(output.get_shape());
-                gs_launch(stream, nelements)([=](auto i) {
-                    auto lens        = desc_output.multi(i);
-                    lens[axis_index] = indices_ptr[lens[axis_index]];
-                    outptr[i]        = inptr[desc_input.linear(lens)];
+            auto* out_ptr           = device_cast(output.data());
+            const auto* in_ptr      = device_cast(input.data());
+            auto& input_shape       = args[0].get_shape();
+            auto lens               = input_shape.lens();
+            lens[axis_index]        = args[1].get_shape().elements();
+            migraphx::shape out_comp_shape{output_shape.type(), lens};
+            visit_tensor_size(out_comp_shape.lens().size(), [&](auto n_out_dim) {
+                hip_tensor_descriptor<n_out_dim> desc_input(input_shape);
+                hip_tensor_descriptor<n_out_dim> desc_output(out_comp_shape);
+                gs_launch(stream, nelements)([=](auto ii) {
+                    auto in_idx        = desc_output.multi(ii);
+                    in_idx[axis_index] = indices_ptr[in_idx[axis_index]];
+                    out_ptr[ii]        = in_ptr[desc_input.linear(in_idx)];
                 });
             });
         });

src/targets/gpu/device/logsoftmax.cpp

+#include <migraphx/shape.hpp>
+#include <migraphx/argument.hpp>
+#include <migraphx/gpu/device/logsoftmax.hpp>
+#include <migraphx/gpu/device/tensor.hpp>
+#include <migraphx/gpu/device/launch.hpp>
+#include <migraphx/gpu/device/types.hpp>
+#include <migraphx/gpu/hip.hpp>
+
+namespace migraphx {
+inline namespace MIGRAPHX_INLINE_NS {
+namespace gpu {
+namespace device {
+
+argument logsoftmax(hipStream_t stream,
+                    const migraphx::shape& output_shape,
+                    std::vector<migraphx::argument> args,
+                    int axis)
+{
+
+    auto lens              = output_shape.lens();
+    std::size_t batch_size = std::accumulate(
+        lens.begin(), lens.begin() + axis, std::size_t{1}, std::multiplies<std::size_t>());
+    std::size_t n_dims = std::accumulate(
+        lens.begin() + axis, lens.end(), std::size_t{1}, std::multiplies<std::size_t>());
+    migraphx::shape comp_shape{output_shape.type(), {batch_size, n_dims}};
+
+    visit_all(args.back(), args.front())([&](auto output, auto input) {
+        const auto* input_ptr = device_cast(input.data());
+        auto* output_ptr      = device_cast(output.data());
+
+        // each thread is for one item in the batch
+        gs_launch(stream, batch_size)([=](auto i) {
+            std::size_t row_start = i * n_dims;
+            // get max
+            auto batch_max = input_ptr[row_start];
+            for(std::size_t j = 1; j < n_dims; ++j)
+            {
+                auto ind  = row_start + j;
+                batch_max = std::max(to_hip_type(batch_max), to_hip_type(input_ptr[ind]));
+            }
+
+            for(std::size_t j = 0; j < n_dims; ++j)
+            {
+                auto ind        = row_start + j;
+                output_ptr[ind] = input_ptr[ind] - batch_max;
+            }
+
+            auto batch_sum = ::exp(to_hip_type(output_ptr[row_start]));
+            for(std::size_t j = 1; j < n_dims; ++j)
+            {
+                auto ind = row_start + j;
+                batch_sum += ::exp(to_hip_type(output_ptr[ind]));
+            }
+            batch_sum = ::log(to_hip_type(batch_sum));
+
+            for(std::size_t j = 0; j < n_dims; ++j)
+            {
+                auto ind = row_start + j;
+                output_ptr[ind] -= batch_sum;
+            }
+        });
+    });
+
+    return args.back();
+}
+
+} // namespace device
+} // namespace gpu
+} // namespace MIGRAPHX_INLINE_NS
+} // namespace migraphx