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

src/make_op.cpp

@@ -64,5 +64,10 @@ operation make_op_from_value(const std::string& name, const value& v)
     });
 }
 
+operation make_json_op(const std::string& name, const std::string& s)
+{
+    return make_op(name, from_json_string(convert_to_json(s)));
+}
+
 } // namespace MIGRAPHX_INLINE_NS
 } // namespace migraphx

src/module.cpp

@@ -141,12 +141,12 @@ void module::set_bypass(bool b) { impl->bypass = b; }
 void module::assign(const module& m)
 {
     // copy the impl
-    if(!impl)
+    if(not impl)
         impl = std::make_unique<module_impl>();
     *impl = *m.impl;
 
     // clear instructions
-    if(!impl->instructions.empty())
+    if(not impl->instructions.empty())
     {
         impl->clear();
     }
@@ -346,7 +346,7 @@ instruction_ref module::replace_instruction(instruction_ref ins, instruction_ref
         assert(out->valid(begin()));
     }
     // Replacement should not be dead code unless its the last instruction
-    assert(!rep->outputs().empty() or rep == std::prev(end()));
+    assert(not rep->outputs().empty() or rep == std::prev(end()));
     // Output of the original instruction should only be the replacement or empty
     assert(ins->outputs().empty() or std::all_of(ins->outputs().begin(),
                                                  ins->outputs().end(),
@@ -385,9 +385,13 @@ instruction_ref module::move_instruction(instruction_ref src, instruction_ref ds
 
 instruction_ref module::move_instructions(instruction_ref src, instruction_ref dst)
 {
-    this->move_instruction(src, dst);
     for(auto ins : src->inputs())
-        this->move_instruction(ins, src);
+    {
+        if(not contains(this->impl->instructions, ins))
+            continue;
+        this->move_instructions(ins, dst);
+    }
+    this->move_instruction(src, dst);
     return src;
 }
 
@@ -598,7 +602,7 @@ instruction_ref module::validate() const
             auto inputs      = i.inputs();
             bool check_order = std::all_of(
                 inputs.begin(), inputs.end(), [&](auto in) { return has_instruction(in); });
-            return !i.valid(impl->instructions.begin(), check_order);
+            return not i.valid(impl->instructions.begin(), check_order);
         });
 }
 
@@ -754,7 +758,7 @@ void module::print_graph(std::ostream& os, bool brief) const
             label = to_string(ins->get_operator());
         os << "\t" << enclose_name(ins_names.at(ins)) << "[label=" << enclose_name(label) << "]";
         os << ";" << std::endl;
-        if(!ins->inputs().empty())
+        if(not ins->inputs().empty())
         {
             for(auto&& arg : ins->inputs())
             {
@@ -788,12 +792,15 @@ static std::string cpp_var_name(const std::string& name)
 
 static void print_make_op(std::ostream& os, const operation& op)
 {
-    os << "migraphx::make_op(" << enclose_name(op.name());
     auto v = op.to_value();
     if(not v.empty())
     {
-        os << ", "
-           << "migraphx::from_json_string(" << enclose_name(to_json_string(v)) << ")";
+        os << "migraphx::make_json_op(" << enclose_name(op.name());
+        os << ", " << enclose_name(to_json_string(v));
+    }
+    else
+    {
+        os << "migraphx::make_op(" << enclose_name(op.name());
     }
     os << ")";
 }
@@ -905,7 +912,7 @@ module& module::sort()
         this->move_instruction(ins, this->begin());
         for(auto child : ins->inputs())
         {
-            if(!contains(this->impl->instructions, child))
+            if(not contains(this->impl->instructions, child))
             {
                 continue;
             }

src/normalize_attributes.cpp

@@ -79,14 +79,14 @@ auto tune_attribute(const std::vector<int64_t>& vec,
     {
         if(contains(vec_attrs, op::normalize_attribute::include_max))
         {
-            if(!std::equal(result.begin(), result.end(), max_vals.begin(), std::less_equal<>{}))
+            if(not std::equal(result.begin(), result.end(), max_vals.begin(), std::less_equal<>{}))
             {
                 MIGRAPHX_THROW("TUNE_VECTOR: value out of range!");
             }
         }
         else
         {
-            if(!std::equal(result.begin(), result.end(), max_vals.begin(), std::less<>{}))
+            if(not std::equal(result.begin(), result.end(), max_vals.begin(), std::less<>{}))
             {
                 MIGRAPHX_THROW("TUNE_VECTOR: value out of range!");
             }
@@ -118,14 +118,15 @@ auto tune_attribute(const std::vector<int64_t>& vec,
     {
         if(contains(vec_attrs, op::normalize_attribute::include_min))
         {
-            if(!std::equal(min_vals.begin(), min_vals.end(), result.begin(), std::less_equal<>{}))
+            if(not std::equal(
+                   min_vals.begin(), min_vals.end(), result.begin(), std::less_equal<>{}))
             {
                 MIGRAPHX_THROW("TUNE_VECTOR: attribute out of range!");
             }
         }
         else
         {
-            if(!std::equal(result.begin(), result.end(), min_vals.begin(), std::less<>{}))
+            if(not std::equal(result.begin(), result.end(), min_vals.begin(), std::less<>{}))
             {
                 MIGRAPHX_THROW("TUNE_VECTOR: attribute out of range!");
             }
@@ -174,7 +175,7 @@ bool normalize_attributes(operation& op, const std::vector<std::size_t>& lens)
             tuned = true;
         }
     }
-    if(!attrs.contains("normalize_axes"))
+    if(not attrs.contains("normalize_axes"))
     {
         return tuned;
     }

src/onnx/include/migraphx/onnx/onnx_parser.hpp

@@ -97,6 +97,7 @@ struct onnx_parser
     shape::dynamic_dimension default_dyn_dim_value = {1, 1, 0};
     std::unordered_map<std::string, std::vector<std::size_t>> map_input_dims;
     std::unordered_map<std::string, std::vector<shape::dynamic_dimension>> map_dyn_input_dims;
+    bool use_dyn_output         = false;
     bool skip_unknown_operators = false;
     int64_t max_loop_iterations = 10;
     int64_t opset_version       = 13;

src/onnx/onnx.cpp

@@ -60,8 +60,14 @@ program parse_onnx_from(const onnx_options& options, Ts&&... xs)
     {
         parser.default_dyn_dim_value = options.default_dyn_dim_value;
     }
+    if(not options.map_input_dims.empty() and not options.map_dyn_input_dims.empty())
+    {
+        MIGRAPHX_THROW("PARSE_ONNX_FROM: both map_input_dims and map_dyn_input_dims non-empty, only"
+                       "one should be used");
+    }
     parser.skip_unknown_operators = options.skip_unknown_operators;
     parser.max_loop_iterations    = options.max_loop_iterations;
+    parser.use_dyn_output         = options.use_dyn_output;
 
     if(options.print_program_on_error)
     {
@@ -80,6 +86,7 @@ program parse_onnx_from(const onnx_options& options, Ts&&... xs)
     {
         parser.parse_from(std::forward<Ts>(xs)...);
     }
+
     return std::move(parser.prog);
 }
 

src/onnx/onnx_parser.cpp

@@ -187,7 +187,7 @@ operation onnx_parser::load(const std::string& name, const node_info& info) cons
 
 void onnx_parser::parse_undefined(module* mod, const std::string& name)
 {
-    if(!contains(instructions, name))
+    if(not contains(instructions, name))
     {
         auto ins           = mod->add_instruction(make_op("undefined"));
         instructions[name] = ins;
@@ -256,11 +256,6 @@ int64_t onnx_parser::get_opset_version(const onnx::ModelProto& model)
 
 void onnx_parser::parse_graph(module* mod, const onnx::GraphProto& graph)
 {
-    if(not map_input_dims.empty() and not map_dyn_input_dims.empty())
-    {
-        MIGRAPHX_THROW("PARSE_GRAPH: both map_input_dims and map_dyn_input_dims non-empty, only"
-                       "one should be used");
-    }
     std::unordered_map<std::string, instruction_ref> mod_insts;
     for(auto&& f : graph.initializer())
     {
@@ -272,7 +267,7 @@ void onnx_parser::parse_graph(module* mod, const onnx::GraphProto& graph)
     {
         const std::string& name = input.name();
         // input not in initializer_data, so it is a real input
-        if(!contains(mod_insts, name))
+        if(not contains(mod_insts, name))
         {
             // ONNX specification does not specify how to deal with the
             // scenario that a nested subgraph contains a parameter with the
@@ -359,7 +354,7 @@ void onnx_parser::parse_graph(module* mod, const onnx::GraphProto& graph)
         all_output_names.begin(),
         all_output_names.end(),
         std::back_inserter(prog_output_names),
-        [&](const auto& name) { return !(name.empty() or instructions.count(name) == 0); });
+        [&](const auto& name) { return not(name.empty() or instructions.count(name) == 0); });
 
     std::vector<instruction_ref> output_ins;
     std::transform(prog_output_names.begin(),
@@ -449,7 +444,7 @@ shape onnx_parser::parse_type(const onnx::TypeProto& t,
                               const std::vector<std::size_t>& input_dims) const
 {
     shape::type_t shape_type = get_type(t.tensor_type().elem_type());
-    if(!input_dims.empty())
+    if(not input_dims.empty())
     {
         return {shape_type, input_dims};
     }
@@ -516,7 +511,7 @@ shape::type_t get_type(int dtype)
 bool is_type_float(shape::type_t dtype)
 {
     bool r = false;
-    if(dtype == shape::float_type || dtype == shape::double_type || dtype == shape::half_type)
+    if(dtype == shape::float_type or dtype == shape::double_type or dtype == shape::half_type)
     {
         r = true;
     }

src/onnx/padding.cpp

@@ -42,7 +42,7 @@ void cal_auto_padding_size(onnx_parser::node_info info,
     size_t kdims = in_lens.size() - 2;
     assert(k_lens.size() == kdims and dilation.size() == kdims);
 
-    if(!contains(info.attributes, "auto_pad"))
+    if(not contains(info.attributes, "auto_pad"))
     {
         return;
     }
@@ -124,7 +124,7 @@ void tune_padding_size(const value& v,
     }
 
     // if padding is symmetric, return directly
-    if(!is_asym_padding(padding))
+    if(not is_asym_padding(padding))
     {
         return;
     }

src/onnx/parse_batchnorm.cpp

@@ -24,7 +24,7 @@
 #include <migraphx/onnx/op_parser.hpp>
 #include <migraphx/ranges.hpp>
 #include <migraphx/make_op.hpp>
-#include <migraphx/op/batch_norm_inference.hpp>
+#include <migraphx/instruction.hpp>
 
 namespace migraphx {
 inline namespace MIGRAPHX_INLINE_NS {
@@ -36,28 +36,63 @@ struct parse_batchnorm : op_parser<parse_batchnorm>
 
     instruction_ref parse(const op_desc& /*opd*/,
                           const onnx_parser& parser,
-                          onnx_parser::node_info info,
-                          const std::vector<instruction_ref>& args) const
+                          const onnx_parser::node_info& info,
+                          std::vector<instruction_ref> args) const
     {
-        float epsilon                                     = 1e-5f;
-        float momentum                                    = 0.9f;
-        op::batch_norm_inference::bn_infer_mode_t bn_mode = op::batch_norm_inference::spatial;
+        float epsilon = 1e-5f;
         if(contains(info.attributes, "epsilon"))
         {
             epsilon = parser.parse_value(info.attributes.at("epsilon")).at<float>();
         }
-        if(contains(info.attributes, "momentum"))
+        auto x_lens = args[0]->get_shape().lens();
+        auto x_type = args[0]->get_shape().type();
+
+        if(std::any_of(args.cbegin() + 1, args.cend(), [](auto a) {
+               return a->get_shape().lens().size() != 1;
+           }))
+        {
+            MIGRAPHX_THROW("PARSE_BATCHNORM: argument scale, bias, mean, or var rank != 1");
+        }
+
+        if(x_lens.size() == 1)
+        {
+            auto rt   = info.add_literal(migraphx::literal{migraphx::shape{x_type}, {0.5}});
+            auto eps  = info.add_literal(migraphx::literal{migraphx::shape{x_type}, {epsilon}});
+            auto n0   = info.add_broadcastable_binary_op("sub", args[0], args[3]);
+            auto d0   = info.add_broadcastable_binary_op("add", args[4], eps);
+            auto d1   = info.add_broadcastable_binary_op("pow", d0, rt);
+            auto div0 = info.add_broadcastable_binary_op("div", n0, d1);
+            auto r0   = info.add_broadcastable_binary_op("mul", div0, args[1]);
+            return info.add_broadcastable_binary_op("add", r0, args[2]);
+        }
+        else if(x_lens.size() > 2)
         {
-            momentum = parser.parse_value(info.attributes.at("momentum")).at<float>();
+            // unsqueeze tensors of shape (C) to broadcast correctly
+            std::vector<int64_t> unsqueeze_axes(x_lens.size() - 2);
+            std::iota(unsqueeze_axes.begin(), unsqueeze_axes.end(), 1);
+            auto rt  = info.add_literal(migraphx::literal{migraphx::shape{x_type}, {0.5}});
+            auto eps = info.add_literal(migraphx::literal{migraphx::shape{x_type}, {epsilon}});
+            auto scale_unsqueeze = info.add_instruction(
+                migraphx::make_op("unsqueeze", {{"axes", unsqueeze_axes}}), args[1]);
+            auto bias_unsqueeze = info.add_instruction(
+                migraphx::make_op("unsqueeze", {{"axes", unsqueeze_axes}}), args[2]);
+            auto mean_unsqueeze = info.add_instruction(
+                migraphx::make_op("unsqueeze", {{"axes", unsqueeze_axes}}), args[3]);
+            auto var_unsqueeze = info.add_instruction(
+                migraphx::make_op("unsqueeze", {{"axes", unsqueeze_axes}}), args[4]);
+            auto numer   = info.add_broadcastable_binary_op("sub", args[0], mean_unsqueeze);
+            auto var_eps = info.add_broadcastable_binary_op("add", var_unsqueeze, eps);
+            auto denom   = info.add_broadcastable_binary_op("pow", var_eps, rt);
+            auto div0    = info.add_broadcastable_binary_op("div", numer, denom);
+            auto r0      = info.add_broadcastable_binary_op("mul", div0, scale_unsqueeze);
+            return info.add_broadcastable_binary_op("add", r0, bias_unsqueeze);
         }
-        if(contains(info.attributes, "spatial"))
+        else
         {
-            bn_mode = (parser.parse_value(info.attributes.at("spatial")).at<uint64_t>() > 0)
-                          ? op::batch_norm_inference::spatial
-                          : op::batch_norm_inference::per_activation;
+            // num dims either 0 or 2
+            MIGRAPHX_THROW("PARSE_BATCHNORM: rank " + std::to_string(x_lens.size()) +
+                           " input tensor, unhandled data format");
         }
-        op::batch_norm_inference op{epsilon, momentum, bn_mode};
-        return info.add_instruction(op, args);
     }
 };
 

src/onnx/parse_cast.cpp

@@ -38,7 +38,7 @@ struct parse_cast : op_parser<parse_cast>
                           onnx_parser::node_info info,
                           const std::vector<instruction_ref>& args) const
     {
-        if(!contains(info.attributes, "to"))
+        if(not contains(info.attributes, "to"))
         {
             MIGRAPHX_THROW("PARSE_CAST: missing to type attribute!");
         }

src/onnx/parse_constant_fill.cpp

@@ -93,7 +93,7 @@ struct parse_constant_fill : op_parser<parse_constant_fill>
         }
         else if(input_as_shape == 0)
         {
-            if(!contains(info.attributes, "shape"))
+            if(not contains(info.attributes, "shape"))
             {
                 MIGRAPHX_THROW("ConstantFill: attribute output shape is needed");
             }

src/onnx/parse_gemm.cpp

@@ -94,7 +94,7 @@ struct parse_gemm : op_parser<parse_gemm>
                 out_lens.back() = l2->get_shape().lens().back();
                 auto l3         = args[2];
                 auto l3_lens    = l3->get_shape().lens();
-                if(!std::equal(out_lens.begin(), out_lens.end(), l3_lens.begin(), l3_lens.end()))
+                if(not std::equal(out_lens.begin(), out_lens.end(), l3_lens.begin(), l3_lens.end()))
                 {
                     l3 = info.add_instruction(make_op("multibroadcast", {{"out_lens", out_lens}}),
                                               args[2]);

src/onnx/parse_generic_op.cpp

@@ -58,7 +58,6 @@ struct parse_generic_op : op_parser<parse_generic_op>
                 {"Log", "log"},
                 {"LRN", "lrn"},
                 {"Neg", "neg"},
-                {"NonMaxSuppression", "nonmaxsuppression"},
                 {"Reciprocal", "recip"},
                 {"Relu", "relu"},
                 {"Round", "round"},
@@ -75,7 +74,7 @@ struct parse_generic_op : op_parser<parse_generic_op>
 
     bool needs_contiguous(const std::string& op_name) const
     {
-        return contains({"flatten", "gather", "nonmaxsuppression", "scatter"}, op_name);
+        return contains({"flatten", "gather", "scatter"}, op_name);
     }
 
     instruction_ref parse(const op_desc& opd,

src/onnx/parse_lpnormalization.cpp

@@ -31,7 +31,7 @@ namespace migraphx {
 inline namespace MIGRAPHX_INLINE_NS {
 namespace onnx {
 
-//!  Parser for LpNormalization ONNX operator.
+// Parser for LpNormalization ONNX operator.
 /*!
   Normalizes a tensor by the L1 or L2 norms along a given axis.
   Norms that evaluate to 0 are changed to 1 to prevent division by zero.

src/onnx/parse_matmul.cpp

@@ -67,7 +67,8 @@ struct parse_matmul : op_parser<parse_matmul>
 
         instruction_ref bl0 = l0;
         instruction_ref bl1 = l1;
-        if(!std::equal(l0_lens.rbegin() + 2, l0_lens.rend(), l1_lens.rbegin() + 2, l1_lens.rend()))
+        if(not std::equal(
+               l0_lens.rbegin() + 2, l0_lens.rend(), l1_lens.rbegin() + 2, l1_lens.rend()))
         {
             auto l0_it = l0_lens.begin() + l0_lens.size() - 2;
             std::vector<std::size_t> l0_broadcasted_lens(l0_lens.begin(), l0_it);

src/onnx/parse_mod.cpp

@@ -40,9 +40,9 @@ struct parse_mod : op_parser<parse_mod>
                           std::vector<instruction_ref> args) const
     {
         std::string mod = "mod";
-        if(is_type_float(args[0]->get_shape().type()) || is_type_float(args[1]->get_shape().type()))
+        if(is_type_float(args[0]->get_shape().type()) or is_type_float(args[1]->get_shape().type()))
         {
-            if(!contains(info.attributes, "fmod"))
+            if(not contains(info.attributes, "fmod"))
             {
                 MIGRAPHX_THROW("Mod operator with float args and fmod=0 invalid");
             }

src/targets/gpu/include/migraphx/gpu/acos.hpp → src/onnx/parse_nonmaxsuppression.cpp

@@ -21,22 +21,29 @@
  * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN
  * THE SOFTWARE.
  */
-#ifndef MIGRAPHX_GUARD_RTGLIB_ACOS_HPP
-#define MIGRAPHX_GUARD_RTGLIB_ACOS_HPP
-
-#include <migraphx/gpu/oper.hpp>
-#include <migraphx/gpu/device/acos.hpp>
+#include <migraphx/onnx/op_parser.hpp>
+#include <migraphx/ranges.hpp>
+#include <migraphx/make_op.hpp>
 
 namespace migraphx {
 inline namespace MIGRAPHX_INLINE_NS {
-namespace gpu {
+namespace onnx {
 
-struct hip_acos : unary_device<hip_acos, device::acos>
+struct parse_nonmaxsuppression : op_parser<parse_nonmaxsuppression>
 {
+    std::vector<op_desc> operators() const { return {{"NonMaxSuppression", "nonmaxsuppression"}}; }
+
+    instruction_ref parse(const op_desc& opd,
+                          const onnx_parser& parser,
+                          const onnx_parser::node_info& info,
+                          const std::vector<instruction_ref>& args) const
+    {
+        auto op = parser.load(opd.op_name, info);
+        op.from_value({{"use_dyn_output", parser.use_dyn_output}});
+        return info.add_instruction(op, args);
+    }
 };
 
-} // namespace gpu
+} // namespace onnx
 } // namespace MIGRAPHX_INLINE_NS
 } // namespace migraphx
-
-#endif

src/onnx/parse_nonzero.cpp

@@ -37,7 +37,7 @@ static std::vector<std::size_t> nonzero_indices(const std::vector<T>& data)
     std::vector<std::size_t> indices;
     for(std::size_t i = 0; i < data.size(); ++i)
     {
-        if(!float_equal(data[i], 0))
+        if(not float_equal(data[i], 0))
             indices.push_back(i);
     }
 

src/onnx/parse_pad.cpp

@@ -160,7 +160,7 @@ struct parse_pad : op_parser<parse_pad>
         if(args.size() == 3)
         {
             auto val_ins = args.at(2);
-            if(!val_ins->can_eval())
+            if(not val_ins->can_eval())
             {
                 MIGRAPHX_THROW("PARSE_PAD: input value must be constant");
             }

src/onnx/parse_pooling.cpp

@@ -157,7 +157,7 @@ struct parse_pooling : op_parser<parse_pooling>
         std::vector<int64_t> slice_end;
         tune_padding_size(values, paddings, count_include_pad, slice_start);
 
-        if(!slice_start.empty())
+        if(not slice_start.empty())
         {
             // calculate expected output shape
             orig_padding.insert(orig_padding.begin() + kdims, 2, 0);
@@ -180,7 +180,7 @@ struct parse_pooling : op_parser<parse_pooling>
         op.from_value(values);
 
         auto l1 = info.add_instruction(op, l0);
-        if(!slice_start.empty())
+        if(not slice_start.empty())
         {
             std::vector<int64_t> axes(kdims);
             std::iota(axes.begin(), axes.end(), 2);

src/onnx/parse_pow.cpp

@@ -46,7 +46,7 @@ auto compute_type(shape::type_t t1, shape::type_t t2)
 
     int it1 = t1;
     int it2 = t2;
-    if(!contains(op_order, it1) or !contains(op_order, it2))
+    if(not contains(op_order, it1) or not contains(op_order, it2))
     {
         MIGRAPHX_THROW("PARSE_POW: Input data type not supported!");
     }