Merge branch 'develop' of https://github.com/ROCmSoftwarePlatform/AMDMIGraphX into mi100_opts

.github/workflows/ci.yaml

@@ -218,7 +218,7 @@ jobs:
       run: |
         echo "leak:dnnl::impl::malloc" > suppressions.txt
         export LSAN_OPTIONS="suppressions=$(pwd)/suppressions.txt"
-        rbuild build -d cget -s gh -t check \
+        rbuild build -d cget -s gh -T check \
           -DCMAKE_BUILD_TYPE=${{matrix.configuration}} \
           -DMIGRAPHX_ENABLE_PYTHON=${{matrix.configuration == 'release' && 'On' || 'Off'}} \
           -DCMAKE_CXX_FLAGS_DEBUG="-g1 -Os -fdebug-prefix-map=$PWD=. -fdebug-types-section -fno-omit-frame-pointer -fsanitize=undefined -fno-sanitize-recover=undefined" \

src/cpp_generator.cpp

@@ -26,17 +26,19 @@ cpp_generator::function::set_body(const module& m, const cpp_generator::generate
         {
             names[ins] =
                 migraphx::any_cast<migraphx::builtin::param>(ins->get_operator()).parameter;
-            continue;
         }
-        if(ins->name() == "@return")
+        else if(ins->name() == "@return")
         {
             assert(ins->inputs().size() == 1);
             return_ins = ins->inputs().front();
         }
+        else
+        {
             std::string n = "z" + std::to_string(names.size());
             names[ins]    = n;
             ss << "auto " << n << " = " << g(ins, names) << ";\n";
         }
+    }
     ss << "return " << names.at(return_ins) << ";\n";
     body = ss.str();
     return *this;
@@ -85,7 +87,10 @@ std::string cpp_generator::generate_point_op(const operation& op,
                                              const std::vector<std::string>& args)
 {
     auto v          = op.to_value();
-    return interpolate_string(op.attributes()["point_op"].to<std::string>(),
+    auto attributes = op.attributes();
+    if(not attributes.contains("point_op"))
+        MIGRAPHX_THROW("op is missing point_op attribute: " + op.name());
+    return interpolate_string(attributes["point_op"].to<std::string>(),
                               [&](auto start, auto last) -> std::string {
                                   auto key = trim({start, last});
                                   if(key.empty())
@@ -120,7 +125,12 @@ std::string cpp_generator::str() const { return impl->fs.str(); }
 cpp_generator::function cpp_generator::generate_module(const module& m)
 {
     function f;
-    f.set_name(m.name()).set_types(m).set_body(
+    auto name = transform_string(m.name(), [](char c) {
+        if(with_char(::isalnum)(c) or c == '_')
+            return c;
+        return '_';
+    });
+    f.set_name(name).set_types(m).set_body(
         m, [&](instruction_ref ins, const auto& names) -> std::string {
             if(ins->name() == "@literal")
                 return shape::cpp_type(ins->get_shape().type()) + "(" +
@@ -130,7 +140,6 @@ cpp_generator::function cpp_generator::generate_module(const module& m)
                            ins->inputs().end(),
                            std::back_inserter(args),
                            [&](auto i) { return names.at(i); });
-            auto s = this->generate_point_op(ins->get_operator(), args);
             return this->generate_point_op(ins->get_operator(), args);
         });
     return f;

src/eliminate_contiguous.cpp

@@ -11,11 +11,13 @@
 namespace migraphx {
 inline namespace MIGRAPHX_INLINE_NS {
 
-static bool try_compute_shape(instruction_ref ins, const std::vector<shape>& inputs)
+static bool try_compute_shape(instruction_ref ins,
+                              const std::vector<shape>& inputs,
+                              const std::vector<module_ref>& mods)
 {
     try
     {
-        shape new_shape = ins->get_operator().compute_shape(inputs);
+        shape new_shape = ins->get_operator().compute_shape(inputs, mods);
         // If the output shape is a standard shape, no need to try its output
         if(new_shape.standard())
         {
@@ -45,7 +47,7 @@ static bool try_compute_shape(instruction_ref ins, const std::vector<shape>& inp
                 return (arg == ins) ? new_shape : arg->get_shape();
             });
 
-            if(!try_compute_shape(output, input_shapes))
+            if(!try_compute_shape(output, input_shapes, mods))
             {
                 return false;
             }
@@ -59,10 +61,12 @@ static bool try_compute_shape(instruction_ref ins, const std::vector<shape>& inp
     return true;
 }
 
-static bool try_compute_shape(instruction_ref ins, const std::vector<instruction_ref>& args)
+static bool try_compute_shape(instruction_ref ins,
+                              const std::vector<instruction_ref>& args,
+                              const std::vector<module_ref>& mods)
 {
     auto inputs = to_shapes(args);
-    return try_compute_shape(ins, inputs);
+    return try_compute_shape(ins, inputs, mods);
 }
 
 void eliminate_contiguous::apply(module& p) const
@@ -82,7 +86,7 @@ void eliminate_contiguous::apply(module& p) const
                 auto new_args = args;
                 auto prev     = arg->inputs().front();
                 replace(new_args, arg, prev);
-                if(try_compute_shape(ins, new_args))
+                if(try_compute_shape(ins, new_args, ins->module_inputs()))
                 {
                     instruction::replace_argument(ins, arg, prev);
                 }

src/fuse_pointwise.cpp

@@ -13,6 +13,8 @@ inline namespace MIGRAPHX_INLINE_NS {
 
 static literal get_scalar(instruction_ref ins)
 {
+    if(ins->name() == "contiguous")
+        return get_scalar(ins->inputs().front());
     const auto& s = ins->get_shape();
     if(not(s.elements() == 1 or s.scalar()))
         return {};
@@ -31,11 +33,16 @@ static void create_pointwise_modules(module_pass_manager& mpm)
     {
         if(not ins->get_operator().attributes().get("pointwise", false))
             continue;
-        auto* pm = mpm.create_module("pointwise" + std::to_string(n++));
+        // Skip convert op for now
+        if(ins->name() == "convert")
+            continue;
+        assert(ins->get_operator().attributes().contains("point_op"));
+        auto* pm = mpm.create_module(mpm.get_module().name() + ":pointwise" + std::to_string(n++));
         pm->set_bypass();
 
         std::unordered_map<instruction_ref, instruction_ref> param_map;
         std::vector<instruction_ref> pointwise_inputs;
+        std::size_t i = 0;
         for(auto input : ins->inputs())
         {
             if(contains(param_map, input))
@@ -44,8 +51,9 @@ static void create_pointwise_modules(module_pass_manager& mpm)
             if(scalar.empty())
             {
                 pointwise_inputs.push_back(input);
-                param_map[input] = pm->add_parameter("x" + std::to_string(param_map.size()),
-                                                     shape{input->get_shape().type()});
+                param_map[input] =
+                    pm->add_parameter("x" + std::to_string(i), shape{input->get_shape().type()});
+                i++;
             }
             else
             {
@@ -68,6 +76,7 @@ static void create_pointwise_modules(module_pass_manager& mpm)
 static std::vector<instruction_ref> append_pointwise_module(instruction_ref ins,
                                                             instruction_ref output)
 {
+    assert(contains(output->inputs(), ins));
     module_ref pm = ins->module_inputs().at(0);
     module_ref xm = output->module_inputs().at(0);
 
@@ -75,6 +84,9 @@ static std::vector<instruction_ref> append_pointwise_module(instruction_ref ins,
     assert(last->name() == "@return");
     assert(last->inputs().size() == 1);
 
+    assert(pm->get_parameter_names().size() == ins->inputs().size());
+    assert(xm->get_parameter_names().size() == output->inputs().size());
+
     std::vector<instruction_ref> inputs = ins->inputs();
     std::unordered_map<instruction_ref, instruction_ref> map_ins;
     std::unordered_map<instruction_ref, instruction_ref> input_map;
@@ -83,6 +95,7 @@ static std::vector<instruction_ref> append_pointwise_module(instruction_ref ins,
     {
         auto input = inputs[i];
         auto param = pm->get_parameter("x" + std::to_string(i));
+        assert(param != pm->end());
         input_map[input] = param;
     }
     // Add the new parameter and additional inputs
@@ -90,6 +103,7 @@ static std::vector<instruction_ref> append_pointwise_module(instruction_ref ins,
     {
         auto input = output->inputs()[i];
         auto param = xm->get_parameter("x" + std::to_string(i));
+        assert(param != xm->end());
         if(input == ins)
         {
             map_ins[param]   = last->inputs().front();

src/include/migraphx/op/pointwise.hpp

@@ -26,19 +26,17 @@ struct pointwise
         auto pnames = pm->get_parameter_names();
         std::sort(pnames.begin(), pnames.end());
         check_shapes{inputs, *this}.has(pnames.size()).same_dims();
-        for(auto i : range(pnames.size()))
-        {
-            auto s1 = pm->get_parameter(pnames[i])->get_shape();
-            auto s2 = inputs[i];
-            if(s1.type() != s2.type())
-                MIGRAPHX_THROW("Mismatch type");
-        }
 
         if(pm->get_output_shapes().size() != 1)
             MIGRAPHX_THROW("submodule should have only one output.");
 
         auto type = pm->get_output_shapes().front().type();
 
+        // Scalar output if all inputs are scalar
+        if(inputs.front().elements() == 1 and
+           all_of(inputs, [](const auto& s) { return s.scalar(); }))
+            return shape{type};
+
         return shape::from_permutation(type, inputs.front().lens(), find_permutation(inputs));
     }
 

src/include/migraphx/op/prelu.hpp

@@ -9,6 +9,7 @@ namespace op {
 
 struct prelu : binary<prelu>
 {
+    std::string point_op() const { return "(${0} < 0) ? (${0} * ${1}) : ${0}"; }
     auto apply() const
     {
         return [](auto x, auto slope) { return ((x < 0) ? (x * slope) : x); };

src/include/migraphx/op/recip.hpp

@@ -9,6 +9,7 @@ namespace op {
 
 struct recip : unary<recip>
 {
+    std::string point_op() const { return "1 / ${0}"; }
     auto apply() const
     {
         return [](auto x) { return 1 / x; };

src/include/migraphx/op/sigmoid.hpp

@@ -18,6 +18,7 @@ namespace op {
 
 struct sigmoid : unary<sigmoid>
 {
+    std::string point_op() const { return "1.f / (1.f + ${function:exp}(-${0}))"; }
     auto apply() const
     {
         return [](auto x) { return 1.f / (1.f + std::exp(-x)); };

src/include/migraphx/op/sign.hpp

@@ -18,6 +18,7 @@ namespace op {
 
 struct sign : unary<sign>
 {
+    std::string point_op() const { return "(${0} > 0 ? 1 : ((${0} < 0) ? -1 : 0))"; }
     auto apply() const
     {
         return [](auto x) { return (x > 0 ? 1 : ((x < 0) ? -1 : 0)); };

src/include/migraphx/operation.hpp

@@ -103,79 +103,69 @@ auto operator==(const T& x, const U& y) -> decltype(x.name() == y.name())
 } // namespace operation_operators
 
 template <class T>
-auto normalize_compute_shape_op(rank<1>, const T& x, const std::vector<shape>& inputs)
-    -> decltype(x.normalize_compute_shape(inputs))
+auto compute_shape_op(rank<3>, const T& x, const std::vector<shape>& inputs)
+    -> decltype(x.compute_shape(inputs))
 {
-    dependent_type<operation, T> y = x;
-    normalize_attributes(y, inputs[0].lens());
-    return any_cast<T>(y).normalize_compute_shape(inputs);
+    return x.compute_shape(inputs);
 }
 
 template <class T>
-shape normalize_compute_shape_op(rank<0>, const T& x, const std::vector<shape>&)
+auto compute_shape_op(rank<2>, const T& x, const std::vector<shape>& inputs)
+    -> decltype(x.normalize_compute_shape(inputs))
 {
-    std::string name = x.name();
-    MIGRAPHX_THROW("Shape not computable: " + name);
+    dependent_type<operation, T> y = x;
+    normalize_attributes(y, inputs[0].lens());
+    return any_cast<T>(y).normalize_compute_shape(inputs);
 }
 
 template <class T>
-shape normalize_compute_shape_op(const T& x, const std::vector<shape>& inputs)
+auto compute_shape_op(rank<1>, const T& x, const std::vector<shape>& inputs)
+    -> decltype(x.compute_shape(inputs, {}))
 {
-    return normalize_compute_shape_op(rank<1>{}, x, inputs);
+    return x.compute_shape(inputs, {});
 }
 
 template <class T>
-auto compute_shape_op(rank<1>,
-                      const T& x,
-                      const std::vector<shape>& inputs,
-                      const std::vector<module_ref>& mod_args)
-    -> decltype(x.compute_shape(inputs, mod_args))
-{
-    return x.compute_shape(inputs, mod_args);
-}
-
-template <class T>
-shape
-    compute_shape_op(rank<0>, const T& x, const std::vector<shape>&, const std::vector<module_ref>&)
+shape compute_shape_op(rank<0>, const T& x, const std::vector<shape>&)
 {
     std::string name = x.name();
     MIGRAPHX_THROW("Shape not computable: " + name);
 }
 
 template <class T>
-shape compute_shape_op(const T& x,
-                       const std::vector<shape>& inputs,
-                       const std::vector<module_ref>& mod_args)
+shape compute_shape_op(const T& x, const std::vector<shape>& inputs)
 {
-    return compute_shape_op(rank<1>{}, x, inputs, mod_args);
+    return compute_shape_op(rank<3>{}, x, inputs);
 }
 
 template <class T>
-auto normalize_compute_shape_op(rank<1>,
+auto mod_compute_shape_op(rank<1>,
                           const T& x,
                           const std::vector<shape>& inputs,
-                                std::vector<module_ref>& mod_args)
-    -> decltype(x.normalize_compute_shape(inputs, mod_args))
+                          const std::vector<module_ref>& mod_args)
+    -> decltype(x.compute_shape(inputs, mod_args))
 {
-    return x.normalize_compute_shape(inputs, mod_args);
+    return x.compute_shape(inputs, mod_args);
 }
 
 template <class T>
-shape normalize_compute_shape_op(rank<0>,
+shape mod_compute_shape_op(rank<0>,
                            const T& x,
-                                 const std::vector<shape>&,
-                                 const std::vector<module_ref>&)
+                           const std::vector<shape>& inputs,
+                           const std::vector<module_ref>& mod_args)
 {
+    if(mod_args.empty())
+        return compute_shape_op(x, inputs);
     std::string name = x.name();
     MIGRAPHX_THROW("Shape not computable: " + name);
 }
 
 template <class T>
-shape normalize_compute_shape_op(const T& x,
+shape mod_compute_shape_op(const T& x,
                            const std::vector<shape>& inputs,
-                                 std::vector<module_ref>& mod_args)
+                           const std::vector<module_ref>& mod_args)
 {
-    return normalize_compute_shape_op(rank<1>{}, x, inputs, mod_args);
+    return mod_compute_shape_op(rank<1>{}, x, inputs, mod_args);
 }
 
 template <class T>
@@ -848,7 +838,7 @@ struct operation
                                                          T&& private_detail_te_self,
                                                          const std::vector<shape>& input)
     {
-        return detail::normalize_compute_shape_op(private_detail_te_self, input);
+        return detail::compute_shape_op(private_detail_te_self, input);
     }
 
     template <class T>
@@ -867,7 +857,7 @@ struct operation
                                                          const std::vector<shape>& inputs,
                                                          const std::vector<module_ref>& mod_args)
     {
-        return detail::compute_shape_op(private_detail_te_self, inputs, mod_args);
+        return detail::mod_compute_shape_op(private_detail_te_self, inputs, mod_args);
     }
 
     template <class T>
@@ -1269,7 +1259,7 @@ template <class T>
 inline auto compute_shape(const T& op, const std::vector<shape>& inputs)
     -> decltype(op.normalize_compute_shape(inputs))
 {
-    return detail::normalize_compute_shape_op(op, inputs);
+    return detail::compute_shape_op(op, inputs);
 }
 
 inline shape compute_shape(const operation& op,
@@ -1294,7 +1284,7 @@ inline auto compute_shape(const T& op,
                           const std::vector<module_ref>& mod_args)
     -> decltype(op.normalize_compute_shape(inputs, mod_args))
 {
-    return detail::normalize_compute_shape_op(op, inputs, mod_args);
+    return detail::compute_shape_op(op, inputs, mod_args);
 }
 
 inline bool is_context_free(const operation& op) { return op.is_context_free(); }

src/include/migraphx/stringutils.hpp

@@ -18,7 +18,7 @@ inline namespace MIGRAPHX_INLINE_NS {
 template <class F>
 auto with_char(F f)
 {
-    return [=](unsigned char c) { return f(c); };
+    return [=](unsigned char c) -> bool { return f(c); };
 }
 
 inline std::string
@@ -120,22 +120,27 @@ interpolate_string(const std::string& input, F f, std::string start = "${", std:
         result.append(it, next_start);
         if(next_start == input.end())
             break;
-        auto r = f(next_start + start.size(), next_end - end.size() + 1);
+        auto r = f(next_start + start.size(), next_end);
         result.append(r.begin(), r.end());
-        it = next_end + 1;
+        it = next_end + end.size();
     }
     return result;
 }
 inline std::string interpolate_string(const std::string& input,
-                                      const std::unordered_map<std::string, std::string>& vars)
+                                      const std::unordered_map<std::string, std::string>& vars,
+                                      std::string start = "${",
+                                      std::string end   = "}")
 {
-    return interpolate_string(input, [&](auto start, auto last) {
-        auto key = trim({start, last});
+    return interpolate_string(input,
+                              [&](auto start_it, auto last_it) {
+                                  auto key = trim({start_it, last_it});
                                   auto it  = vars.find(key);
                                   if(it == vars.end())
                                       throw std::runtime_error("Unknown key: " + key);
                                   return it->second;
-    });
+                              },
+                              std::move(start),
+                              std::move(end));
 }
 
 template <class Iterator>

src/pass_manager.cpp

@@ -15,6 +15,8 @@
 namespace migraphx {
 inline namespace MIGRAPHX_INLINE_NS {
 
+MIGRAPHX_DECLARE_ENV_VAR(MIGRAPHX_TRACE_PASSES);
+
 void validate_pass(module& mod, const pass& p, tracer trace)
 {
     (void)mod;
@@ -82,6 +84,8 @@ module& get_module(module_pass_manager& mpm) { return mpm.get_module(); }
 
 void run_passes(module& mod, const std::vector<pass>& passes, tracer trace)
 {
+    if(enabled(MIGRAPHX_TRACE_PASSES{}))
+        trace = tracer{std::cout};
     for(const auto& p : passes)
     {
         module_pm{&mod, nullptr, &trace}.run_pass(p);
@@ -90,6 +94,8 @@ void run_passes(module& mod, const std::vector<pass>& passes, tracer trace)
 
 void run_passes(program& prog, const std::vector<pass>& passes, tracer trace)
 {
+    if(enabled(MIGRAPHX_TRACE_PASSES{}))
+        trace = tracer{std::cout};
     for(const auto& p : passes)
     {
         auto mods = prog.get_modules();

src/targets/gpu/CMakeLists.txt

@@ -122,6 +122,7 @@ add_library(migraphx_gpu
     batch_norm_inference.cpp
     clip.cpp
     code_object_op.cpp
+    compile_ops.cpp
     compile_hip.cpp
     compile_hip_code_object.cpp
     compile_pointwise.cpp

src/targets/gpu/allocation_model.cpp

 #include <migraphx/gpu/allocation_model.hpp>
 #include <migraphx/make_op.hpp>
+#include <migraphx/instruction.hpp>
+#include <migraphx/module.hpp>
 
 namespace migraphx {
 inline namespace MIGRAPHX_INLINE_NS {

src/targets/gpu/compile_ops.cpp

+#include <migraphx/gpu/compile_ops.hpp>
+#include <migraphx/gpu/context.hpp>
+#include <migraphx/module.hpp>
+#include <migraphx/iterator_for.hpp>
+#include <migraphx/instruction.hpp>
+#include <migraphx/par_for.hpp>
+#include <migraphx/register_op.hpp>
+#include <migraphx/op/identity.hpp>
+#include <migraphx/gpu/compile_pointwise.hpp>
+
+namespace migraphx {
+inline namespace MIGRAPHX_INLINE_NS {
+namespace gpu {
+
+struct precompile_op
+{
+    operation op = op::identity{};
+
+    template <class Self, class F>
+    static auto reflect(Self& self, F f)
+    {
+        return pack(f(self.op, "op"));
+    }
+
+    std::string name() const { return "gpu::precompile_op"; }
+
+    shape compute_shape(std::vector<shape> inputs, const std::vector<module_ref>& mods) const
+    {
+        inputs.pop_back();
+        return op.compute_shape(inputs, mods);
+    }
+
+    std::ptrdiff_t output_alias(const std::vector<shape>& shapes) const
+    {
+        return shapes.size() - 1;
+    }
+};
+
+MIGRAPHX_REGISTER_OP(precompile_op);
+
+struct pointwise_compiler
+{
+    std::string name() const { return "pointwise"; }
+
+    operation apply(context& ctx, instruction_ref ins, const operation&) const
+    {
+        assert(not ins->module_inputs().empty());
+        auto* pm = ins->module_inputs().front();
+        return compile_pointwise(ctx, to_shapes(ins->inputs()), *pm);
+    }
+};
+
+using compiler_function = std::function<operation(context&, instruction_ref, operation)>;
+
+template <class T>
+compiler_function make_compiler_function(T x)
+{
+    return {[=](auto&&... xs) { return x.apply(xs...); }};
+}
+
+template <class... Ts>
+std::unordered_map<std::string, compiler_function> make_compilers(Ts... xs)
+{
+    return {{xs.name(), make_compiler_function(xs)}...};
+}
+
+struct compiled_result
+{
+    operation op;
+    instruction_ref ins;
+};
+
+void compile_ops::apply(module& m) const
+{
+    auto compilers = make_compilers(pointwise_compiler{});
+    std::vector<std::function<compiled_result()>> compiles;
+
+    for(auto ins : iterator_for(m))
+    {
+        if(ins->name() != "gpu::precompile_op")
+            continue;
+        operation preop = any_cast<precompile_op>(ins->get_operator()).op;
+        assert(contains(compilers, preop.name()));
+        auto c = compilers[preop.name()];
+        compiles.emplace_back([=]() -> compiled_result { return {c(*ctx, ins, preop), ins}; });
+    }
+    std::vector<compiled_result> results(compiles.size());
+    par_for(compiles.size(), 1, [&](auto i) { results[i] = compiles[i](); });
+    for(const auto& cr : results)
+    {
+        m.replace_instruction(cr.ins, cr.op, cr.ins->inputs());
+    }
+}
+
+} // namespace gpu
+
+} // namespace MIGRAPHX_INLINE_NS
+} // namespace migraphx

src/targets/gpu/compile_pointwise.cpp

@@ -2,9 +2,14 @@
 #include <migraphx/gpu/compile_hip_code_object.hpp>
 #include <migraphx/gpu/compile_hip.hpp>
 #include <migraphx/gpu/context.hpp>
+#include <migraphx/cpp_generator.hpp>
 #include <migraphx/ranges.hpp>
 #include <migraphx/reduce_dims.hpp>
 #include <migraphx/stringutils.hpp>
+#include <migraphx/dead_code_elimination.hpp>
+#include <migraphx/eliminate_common_subexpression.hpp>
+#include <migraphx/module.hpp>
+#include <migraphx/pass_manager.hpp>
 
 namespace migraphx {
 inline namespace MIGRAPHX_INLINE_NS {
@@ -17,6 +22,8 @@ static const char* const pointwise_kernel = R"__migraphx__(
 
 using namespace migraphx;
 
+${preamble}
+
 extern "C" {
 __global__ void kernel(${params}) 
 {
@@ -29,7 +36,10 @@ int main() {}
 
 )__migraphx__";
 
-operation compile_pointwise(context&, const std::vector<shape>& inputs, const std::string& lambda)
+operation compile_pointwise(context&,
+                            const std::vector<shape>& inputs,
+                            const std::string& lambda,
+                            const std::string& preamble)
 {
     hip_compile_options options;
     options.global         = compute_global(inputs.front().elements());
@@ -37,13 +47,23 @@ operation compile_pointwise(context&, const std::vector<shape>& inputs, const st
     options.inputs         = inputs;
     options.output         = inputs.back();
     options.reduced_inputs = reduce_dims(inputs);
+    options.params         = "-Wno-float-equal";
     auto src               = interpolate_string(pointwise_kernel,
                                   {{"params", enum_params(inputs.size(), "void * private_p")},
                                    {"args", enum_params(inputs.size(), "private_p")},
-                                   {"lambda", lambda}});
+                                   {"lambda", lambda},
+                                   {"preamble", preamble}});
     return compile_hip_code_object(src, options);
 }
 
+operation compile_pointwise(context& ctx, const std::vector<shape>& inputs, module m)
+{
+    run_passes(m, {eliminate_common_subexpression{}, dead_code_elimination{}});
+    cpp_generator g;
+    auto name = g.create_function(g.generate_module(m).set_attributes({"__device__"}));
+    return compile_pointwise((ctx), inputs, "&" + name, g.str());
+}
+
 } // namespace gpu
 } // namespace MIGRAPHX_INLINE_NS
 } // namespace migraphx

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

@@ -3,6 +3,7 @@
 
 #include <migraphx/config.hpp>
 #include <migraphx/operation.hpp>
+#include <migraphx/instruction_ref.hpp>
 #include <string>
 
 namespace migraphx {

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

+#ifndef MIGRAPHX_GUARD_GPU_COMPILE_OPS_HPP
+#define MIGRAPHX_GUARD_GPU_COMPILE_OPS_HPP
+
+#include <migraphx/config.hpp>
+#include <string>
+
+namespace migraphx {
+inline namespace MIGRAPHX_INLINE_NS {
+
+struct module;
+
+namespace gpu {
+
+struct context;
+
+struct compile_ops
+{
+    context* ctx = nullptr;
+    std::string name() const { return "gpu::compile_ops"; }
+    void apply(module& m) const;
+};
+
+} // namespace gpu
+
+} // namespace MIGRAPHX_INLINE_NS
+} // namespace migraphx
+#endif // MIGRAPHX_GUARD_GPU_COMPILE_OPS_HPP

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

@@ -6,11 +6,17 @@
 
 namespace migraphx {
 inline namespace MIGRAPHX_INLINE_NS {
+struct module;
+
 namespace gpu {
 
 struct context;
-operation
-compile_pointwise(context& ctx, const std::vector<shape>& inputs, const std::string& lambda);
+operation compile_pointwise(context& ctx,
+                            const std::vector<shape>& inputs,
+                            const std::string& lambda,
+                            const std::string& preamble = "");
+
+operation compile_pointwise(context& ctx, const std::vector<shape>& inputs, module m);
 
 } // namespace gpu
 } // namespace MIGRAPHX_INLINE_NS

src/targets/gpu/kernels/include/migraphx/kernels/pointwise.hpp

@@ -23,7 +23,7 @@ __device__ void pointwise_tensor(index idx, F f, T out, Ts... xs)
 template <class F, class... Ts>
 __device__ void pointwise(F f, Ts*... ps)
 {
-    auto t = transform_args(make_tensors(), rotate_last(), auto_vectorize());
+    auto t = transform_args(make_tensors(), rotate_last());
     t(ps...)([&](auto... xs) {
         auto idx = make_index();
         pointwise_tensor(idx, f, xs...);