Merge branch 'develop' into jit-layernorm

cmake/Embed.cmake

@@ -39,6 +39,8 @@ function(generate_embed_source EMBED_NAME)
 
     file(WRITE "${PARSE_HEADER}" "
 #include <unordered_map>
+#include <string>
+#include <utility>
 const std::unordered_map<std::string, std::pair<const char*,const char*>>& ${EMBED_NAME}();
 ")
 

examples/nlp/python_bert_squad/requirements_bertsquad.txt

-tensorflow==2.6.4
+tensorflow==2.7.2
 onnxruntime
 tokenizers
\ No newline at end of file

src/eliminate_contiguous.cpp

@@ -6,6 +6,7 @@
 #include <migraphx/stringutils.hpp>
 #include <migraphx/op/contiguous.hpp>
 #include <migraphx/op/identity.hpp>
+#include <migraphx/par_for.hpp>
 #include <utility>
 
 namespace migraphx {
@@ -71,6 +72,8 @@ static bool try_compute_shape(instruction_ref ins,
 
 void eliminate_contiguous::apply(module& m) const
 {
+    std::vector<instruction_ref> const_instruction;
+
     for(auto ins : iterator_for(m))
     {
         // return instruction should have inputs with standard shape
@@ -81,6 +84,7 @@ void eliminate_contiguous::apply(module& m) const
         auto args     = ins->inputs();
         auto new_args = args;
         auto mod_args = ins->module_inputs();
+
         for(auto arg : ins->inputs())
         {
             if(arg->name() == op_name)
@@ -93,14 +97,24 @@ void eliminate_contiguous::apply(module& m) const
                 }
                 else if(prev->can_eval())
                 {
-                    auto c = op::contiguous{};
-                    auto r = c.compute(c.compute_shape({prev->get_shape()}), {prev->eval()});
-
-                    auto l = m.add_literal(r.get_shape(), r.data());
-                    m.replace_instruction(arg, l);
+                    const_instruction.push_back(arg);
+                }
             }
         }
     }
+
+    // Perform evaluations in parallel
+    std::vector<argument> literals(const_instruction.size());
+    par_for(const_instruction.size(), 1, [&](const auto i) {
+        auto c      = op::contiguous{};
+        auto prev   = const_instruction[i]->inputs().front();
+        literals[i] = c.compute(c.compute_shape({prev->get_shape()}), {prev->eval()});
+    });
+
+    for(size_t i = 0; i < const_instruction.size(); i++)
+    {
+        auto l = m.add_literal(literals[i].get_shape(), literals[i].data());
+        m.replace_instruction(const_instruction[i], l);
     }
 }
 

src/targets/gpu/fuse_ops.cpp

@@ -965,7 +965,7 @@ struct find_gemm_pointwise
         inputs.pop_back();
 
         inputs.push_back(c_ins);
-        inputs.push_back(gemm_ins->inputs().back());
+        inputs.push_back(ins->inputs().back());
 
         gemm.beta = 1;
         m.replace_instruction(ins, gemm, inputs);

src/targets/gpu/include/migraphx/gpu/code_object_op.hpp

@@ -34,6 +34,10 @@ struct code_object_op
                     f(self.output, "output"));
     }
 
+    value attributes() const { return {{"group", group()}}; }
+
+    std::string group() const { return "gpu::code_object::" + symbol_name; }
+
     std::string name() const { return "gpu::code_object"; }
     shape compute_shape(std::vector<shape> inputs) const;
     argument

test/verify/gemm_add.cpp

+
+#include "verify_program.hpp"
+#include <migraphx/program.hpp>
+#include <migraphx/generate.hpp>
+#include <migraphx/make_op.hpp>
+#include <migraphx/apply_alpha_beta.hpp>
+struct gemm_add : verify_program<gemm_add>
+{
+    migraphx::program create_program() const
+    {
+        migraphx::program p;
+        auto* mm = p.get_main_module();
+        migraphx::shape m1_shape{migraphx::shape::float_type, {1, 2, 3}};
+        migraphx::shape m2_shape{migraphx::shape::float_type, {1, 3, 4}};
+        migraphx::shape m3_shape{migraphx::shape::float_type, {1, 2, 4}};
+        auto l1 = mm->add_parameter("1", m1_shape);
+        auto l2 = mm->add_parameter("2", m2_shape);
+        auto l3 = mm->add_parameter("3", m3_shape);
+
+        auto dot = mm->add_instruction(migraphx::make_op("dot"), l1, l2);
+        mm->add_instruction(migraphx::make_op("add"), dot, l3);
+        return p;
+    }
+};

tools/install_prereqs.sh

@@ -27,7 +27,7 @@ elif [ "$#" -eq 1 ]; then
   PREFIX=$1
 fi
 
-echo "Dependencies are install at $PREFIX"
+echo "Dependencies are installed at $PREFIX"
 
 # Install deps with rbuild
 rbuild prepare -d $PREFIX -s develop
@@ -35,3 +35,5 @@ rbuild prepare -d $PREFIX -s develop
 # install onnx package for unit tests
 pip3 install onnx==1.8.1 numpy==1.18.5 typing==3.7.4 pytest==6.0.1 packaging==16.8
 
+# pin version of protobuf in Python for onnx runtime unit tests
+pip3 install protobuf==3.20.0