Merge branch 'jit-vector-reduce' of...

Merge branch 'jit-vector-reduce' of github.com:ROCmSoftwarePlatform/AMDMIGraphX into jit-vector-reduce

examples/nlp/python_bert_squad/BERT-Squad.ipynb

@@ -62,7 +62,7 @@
    "metadata": {},
    "outputs": [],
    "source": [
-    "!wget -nc https://github.com/onnx/models/blob/main/text/machine_comprehension/bert-squad/model/bertsquad-10.onnx"
+    "!wget -nc https://github.com/onnx/models/raw/main/text/machine_comprehension/bert-squad/model/bertsquad-10.onnx"
    ]
   },
   {

examples/nlp/python_bert_squad/README.md

@@ -23,7 +23,7 @@ unzip uncased_L-12_H-768_A-12.zip
 ```
 5) Get BERT ONNX model (bertsquad-10.onnx):
 ```
-wget https://github.com/onnx/models/blob/main/text/machine_comprehension/bert-squad/model/bertsquad-10.onnx
+wget https://github.com/onnx/models/raw/main/text/machine_comprehension/bert-squad/model/bertsquad-10.onnx
 ```
 6) Run the inference, it will compile and run the model on three questions and small data provided in `inputs.json`:
 ```

examples/nlp/python_bert_squad/requirements_bertsquad.txt

-tensorflow==2.5.3
+tensorflow==2.6.4
 onnxruntime
 tokenizers
\ No newline at end of file

src/include/migraphx/matcher.hpp

@@ -754,10 +754,16 @@ auto skip_broadcasts(Ms... ms)
     return skip(name("broadcast", "multibroadcast", "contiguous"))(ms...);
 }
 
+template <class... Ms>
+auto skip_broadcasts_converts(Ms... ms)
+{
+    return skip(name("broadcast", "multibroadcast", "contiguous", "convert"))(ms...);
+}
+
 template <class T>
 inline auto has_value(T x, float tolerance = 1e-6)
 {
-    return skip_broadcasts(make_basic_pred_matcher([=](instruction_ref ins) {
+    return skip_broadcasts_converts(make_basic_pred_matcher([=](instruction_ref ins) {
         if(ins->name() != "@literal")
             return false;
         auto l = ins->get_literal();

src/onnx/parse_mean.cpp

@@ -2,6 +2,7 @@
 #include <migraphx/onnx/checks.hpp>
 #include <migraphx/instruction.hpp>
 #include <migraphx/make_op.hpp>
+#include <migraphx/ranges.hpp>
 
 namespace migraphx {
 inline namespace MIGRAPHX_INLINE_NS {
@@ -9,6 +10,9 @@ namespace onnx {
 
 struct parse_mean : op_parser<parse_mean>
 {
+    const std::set<shape::type_t> float_types = {
+        shape::float_type, shape::half_type, shape::double_type};
+
     std::vector<op_desc> operators() const { return {{"Mean"}}; }
 
     /// Calculates the element-wise mean of n>=1 input tensors
@@ -24,17 +28,29 @@ struct parse_mean : op_parser<parse_mean>
         auto divisor = info.add_literal(
             migraphx::literal{migraphx::shape{args[0]->get_shape().type()}, {num_data}});
 
-        // TODO: Only divide when using floating-point
-        return std::accumulate(args.begin() + 1,
-                               args.end(),
-                               info.add_broadcastable_binary_op("div", args[0], divisor),
-                               [&](auto mean, auto data_i) {
-                                   // Pre-divide each tensor element-wise by n to reduce risk of
-                                   // overflow during summation
-                                   auto div =
-                                       info.add_broadcastable_binary_op("div", data_i, divisor);
-                                   return info.add_broadcastable_binary_op("add", mean, div);
-                               });
+        if(contains(float_types, args[0]->get_shape().type()))
+        {
+            return std::accumulate(args.begin() + 1,
+                                   args.end(),
+                                   info.add_broadcastable_binary_op("div", args[0], divisor),
+                                   [&](auto mean, auto data_i) {
+                                       // Pre-divide each tensor element-wise by n to reduce risk of
+                                       // overflow during summation
+                                       auto div =
+                                           info.add_broadcastable_binary_op("div", data_i, divisor);
+                                       return info.add_broadcastable_binary_op("add", mean, div);
+                                   });
+        }
+        else
+        {
+            // Compute sum before division for integral types
+            auto sum = std::accumulate(
+                args.begin() + 1, args.end(), args[0], [&](auto accum, auto data_i) {
+                    return info.add_broadcastable_binary_op("add", accum, data_i);
+                });
+
+            return info.add_broadcastable_binary_op("div", sum, divisor);
+        }
     }
 };
 

src/reduce_dims.cpp

@@ -16,10 +16,8 @@ bool reduce_dim(std::vector<shape>& shapes, std::size_t n)
         auto bstride = s.strides()[n + 1];
         auto blen    = s.lens()[n + 1];
 
-        if(astride == bstride * blen)
-        {
+        if(astride == bstride * blen or alen == 1)
             new_lens.push_back(alen * blen);
-        }
     }
     if(new_lens.size() != shapes.size())
         return false;
@@ -37,10 +35,25 @@ bool reduce_dim(std::vector<shape>& shapes, std::size_t n)
     return true;
 }
 
+void reduce_dim1(std::vector<shape>& shapes)
+{
+    if(std::any_of(shapes.begin(), shapes.end(), [&](const auto& s) {
+           return s.lens().size() < 2 or s.lens().back() != 1;
+       }))
+        return;
+    for(auto& s : shapes)
+    {
+        auto lens    = s.lens();
+        auto strides = s.strides();
+        lens.pop_back();
+        strides.pop_back();
+        s = shape{s.type(), lens, strides};
+    }
+}
+
 std::size_t reduce_dim_all(std::vector<shape>& shapes, std::size_t n)
 {
     while(reduce_dim(shapes, n) and n < shapes.size()) {}
-
     return n + 1;
 }
 void reduce_dim_all(std::vector<shape>& shapes)
@@ -48,6 +61,7 @@ void reduce_dim_all(std::vector<shape>& shapes)
     std::size_t n = 0;
     while(n < shapes.front().lens().size() - 1)
         n = reduce_dim_all(shapes, n);
+    reduce_dim1(shapes);
 }
 
 std::vector<std::size_t> base_lens(const std::vector<shape>& shapes)

src/targets/gpu/fuse_ops.cpp

@@ -908,11 +908,6 @@ struct find_gemm_add
         if(not float_equal(gemm.beta, 0))
             return;
 
-        if(std::any_of(ins->inputs().begin(), ins->inputs().end(), [](auto i) {
-               return not i->get_shape().standard();
-           }))
-            return;
-
         auto inputs = gemm_ins->inputs();
         inputs.pop_back();
 
@@ -931,6 +926,53 @@ struct find_gemm_add
     }
 };
 
+auto pointwise_name(const std::string& s)
+{
+    return precompile_name("pointwise")(match::make_basic_pred_matcher([=](auto ins) {
+        module_ref pm = ins->module_inputs().front();
+        auto n = std::count_if(pm->begin(), pm->end(), [&](auto& i) { return i.name() == s; });
+        if(n != 1)
+            return false;
+        return std::all_of(pm->begin(), pm->end(), [&](auto& i) {
+            return starts_with(i.name(), "@") or i.name() == s;
+        });
+    }));
+}
+
+struct find_gemm_pointwise
+{
+    auto matcher() const
+    {
+        return pointwise_name("add")(
+            match::nargs(3),
+            match::all_of[match::inputs()](match::standard_shape()),
+            match::either_arg(0, 1)(match::used_once().bind("c"),
+                                    match::name("gpu::gemm")(match::nargs(3)).bind("gemm")));
+    }
+
+    void apply(module& m, const match::matcher_result& r) const
+    {
+        auto ins      = r.result;
+        auto gemm_ins = r.instructions["gemm"];
+        auto c_ins    = r.instructions["c"];
+
+        auto gemm = any_cast<rocblas_gemm<op::dot>>(gemm_ins->get_operator());
+
+        // Already fused gemm
+        if(not float_equal(gemm.beta, 0))
+            return;
+
+        auto inputs = gemm_ins->inputs();
+        inputs.pop_back();
+
+        inputs.push_back(c_ins);
+        inputs.push_back(gemm_ins->inputs().back());
+
+        gemm.beta = 1;
+        m.replace_instruction(ins, gemm, inputs);
+    }
+};
+
 struct find_commutative_broadcast
 {
     auto matcher() const
@@ -967,7 +1009,11 @@ void fuse_ops::apply(module& m) const
                         find_add_unary{"gpu::tanh", hip_add_tanh{}, hip_triadd_tanh{}},
                         find_add_clip{});
     run_passes(m, {dead_code_elimination{}});
-    match::find_matches(m, find_triadd_layernorm{}, find_gemm_add{}, find_commutative_broadcast{});
+    match::find_matches(m,
+                        find_triadd_layernorm{},
+                        find_gemm_add{},
+                        find_gemm_pointwise{},
+                        find_commutative_broadcast{});
 }
 
 } // namespace gpu

src/targets/gpu/gemm_impl.cpp

 #include <rocblas.h>
 #include <migraphx/gpu/gemm_impl.hpp>
+#include <migraphx/reduce_dims.hpp>
 
 namespace migraphx {
 inline namespace MIGRAPHX_INLINE_NS {
@@ -27,6 +28,22 @@ rocblas_datatype get_type(shape::type_t type)
     MIGRAPHX_THROW("ROCBLAS_GEMM: data type not supported!");
 }
 
+void blas_shape(const shape& s)
+{
+    if(s.lens().size() < 2)
+        return;
+    if(std::none_of(s.strides().end() - 2, s.strides().end(), [&](auto i) { return i == 1; }))
+        MIGRAPHX_THROW("GPU_GEMM: needs to have one matrix stride as 1");
+    if(s.lens().size() < 3)
+        return;
+    shape batch_shape{s.type(),
+                      {s.lens().begin(), s.lens().end() - 2},
+                      {s.strides().begin(), s.strides().end() - 2}};
+    auto batch_shapes = reduce_dims({batch_shape});
+    if(batch_shapes.front().lens().size() != 1)
+        MIGRAPHX_THROW("GPU_GEMM: Batch dimension is not collapsible");
+}
+
 template <class R, class... Ts, class... Us>
 R rocblas_invoke(R (*f)(Ts...), Us... xs)
 {
@@ -36,6 +53,18 @@ R rocblas_invoke(R (*f)(Ts...), Us... xs)
         return f(xs..., nullptr, nullptr);
 }
 
+static bool is_transposed(const shape& s)
+{
+    if(not s.transposed())
+        return false;
+    return s.strides().back() != 1;
+}
+
+static rocblas_int get_batch_stride(const argument& a)
+{
+    return a.get_shape().strides()[a.get_shape().strides().size() - 3];
+}
+
 template <class T>
 void gemm_impl(context& ctx,
                const shape& output_shape,
@@ -45,8 +74,8 @@ void gemm_impl(context& ctx,
                bool int8_x4_format,
                bool compute_fp32)
 {
-    bool transa     = args[0].get_shape().transposed();
-    bool transb     = args[1].get_shape().transposed();
+    bool transa     = is_transposed(args[0].get_shape());
+    bool transb     = is_transposed(args[1].get_shape());
     auto n_dim      = output_shape.lens().size();
     auto dim_1      = n_dim - 1;
     auto dim_0      = n_dim - 2;
@@ -142,6 +171,9 @@ void gemm_impl(context& ctx,
         }
         else
         {
+            auto a_stride = get_batch_stride(args[0]);
+            auto b_stride = get_batch_stride(args[1]);
+            auto c_stride = get_batch_stride(args[2]);
             rocblas_invoke(&rocblas_gemm_strided_batched_ex,
                            ctx.get_stream().get_rocblas(),
                            transb ? rocblas_operation_transpose : rocblas_operation_none,
@@ -153,20 +185,20 @@ void gemm_impl(context& ctx,
                            to_pointer(args.at(1)),
                            arg_type,
                            ldb,
-                           k * n,
+                           b_stride,
                            to_pointer(args.at(0)),
                            arg_type,
                            lda,
-                           m * k,
+                           a_stride,
                            beta_v,
                            to_pointer(args[2]),
                            output_type,
                            ldc,
-                           m * n,
+                           c_stride,
                            is_3inputs ? to_pointer(args[3]) : to_pointer(args[2]),
                            output_type,
                            ldc,
-                           m * n,
+                           c_stride,
                            num_matrices,
                            compute_type,
                            rocblas_gemm_algo_standard,

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

@@ -18,6 +18,8 @@ namespace gpu {
 
 struct context;
 
+void blas_shape(const shape& s);
+
 template <class Op>
 struct rocblas_gemm
 {
@@ -50,13 +52,14 @@ struct rocblas_gemm
         std::vector<shape> in_shapes(inputs);
         in_shapes.pop_back();
         check_shapes{in_shapes, *this}.not_broadcasted();
-        batch_not_transposed(inputs[0].strides());
-        batch_not_transposed(inputs[1].strides());
+        blas_shape(inputs[0]);
+        blas_shape(inputs[1]);
         // if gemm and add are fused
-        if(not float_equal(beta, 0))
+        if(in_shapes.size() > 2)
         {
             auto cmat_shape = in_shapes.back();
             in_shapes.pop_back();
+            blas_shape(cmat_shape);
             auto op_out_shape = op.compute_shape(in_shapes);
             if(cmat_shape.lens() != op_out_shape.lens())
             {
@@ -71,6 +74,7 @@ struct rocblas_gemm
                                to_string(cmat_shape.type()) +
                                ", it must be: " + to_string(op_out_shape.type()));
             }
+            return op_out_shape;
         }
 
         return op.compute_shape(in_shapes);
@@ -96,28 +100,6 @@ struct rocblas_gemm
         return args.back();
     }
 
-    void batch_not_transposed(const std::vector<std::size_t>& strides) const
-    {
-        if(strides.size() <= 2)
-            return;
-        auto dim_0       = strides.size() - 2;
-        auto matrix_size = std::max(strides[dim_0], strides[dim_0 + 1]);
-        std::vector<std::size_t> batch(strides.begin(), strides.begin() + dim_0);
-        if(std::all_of(batch.begin(), batch.end(), [&](auto i) { return (i < matrix_size); }))
-        {
-            MIGRAPHX_THROW("GPU_GEMM: matrix size and batch size {" + to_string_range(strides) +
-                           "} are transposed!");
-        }
-
-        if(std::adjacent_find(batch.begin(), batch.end(), [&](auto i, auto j) {
-               return (i < j or i < matrix_size or j < matrix_size);
-           }) != batch.end())
-        {
-            MIGRAPHX_THROW("GPU_GEMM: batch size {" + to_string_range(strides) +
-                           "} is transposed!");
-        }
-    }
-
     std::ptrdiff_t output_alias(const std::vector<shape>& shapes) const
     {
         return shapes.size() - 1;

src/targets/gpu/jit/gathernd.cpp

@@ -19,7 +19,7 @@ namespace gpu {
 // NOLINTNEXTLINE
 static const char* const gathernd_kernel = R"__migraphx__(
 #include <migraphx/kernels/gathernd.hpp>
-#include <migraphx/kernels/basic_ops.hpp>
+#include <migraphx/kernels/ops.hpp>
 #include <migraphx/kernels/integral_constant.hpp>
 #include <migraphx/kernels/generic_constant.hpp>
 #include <args.hpp>

src/targets/gpu/jit/pointwise.cpp

@@ -30,7 +30,7 @@ namespace migraphx {
 ${preamble}
 
 extern "C" {
-__global__ void kernel(${params}) 
+__global__ void ${kernel}(${params}) 
 {
     auto idx = make_index();
     pointwise(idx, ${transformers})(${lambda}, ${args});
@@ -42,6 +42,18 @@ __global__ void kernel(${params})
 
 )__migraphx__";
 
+static std::vector<std::string> get_op_names(const module& m)
+{
+    std::vector<std::string> result;
+    for(auto& ins : m)
+    {
+        if(starts_with(ins.name(), "@"))
+            continue;
+        result.push_back(ins.name());
+    }
+    return result;
+}
+
 struct pointwise_compiler : compiler<pointwise_compiler>
 {
     std::vector<std::string> names() const { return {"pointwise"}; }
@@ -70,7 +82,8 @@ struct pointwise_compiler : compiler<pointwise_compiler>
                                options.output.elements() / vec.size,
                                oversubscribe_if(not preloads.is_preloading())));
         auto src = interpolate_string(pointwise_kernel,
-                                      {{"params", enum_params(inputs.size(), "void * private_p")},
+                                      {{"kernel", options.kernel_name},
+                                       {"params", enum_params(inputs.size(), "void * private_p")},
                                        {"args", enum_params(inputs.size(), "private_p")},
                                        {"lambda", v.at("lambda").to<std::string>()},
                                        {"transformers", make_transformer_args(preloads, vec)},
@@ -99,8 +112,13 @@ struct pointwise_compiler : compiler<pointwise_compiler>
         auto name = g.create_function(
             g.generate_module(*pm).set_attributes({"__device__"}).set_generic_types(*pm));
         std::string lambda = "MIGRAPHX_LIFT(" + name + ")";
+        auto op_names      = get_op_names(*pm);
+        op_names.push_back("kernel");
+        auto op_name_string = join_strings(op_names, "_");
         return replace(
-            compile_op(ctx, to_shapes(ins->inputs()), {{"lambda", lambda}, {"preamble", g.str()}}));
+            compile_op(ctx,
+                       to_shapes(ins->inputs()),
+                       {{"lambda", lambda}, {"preamble", g.str()}, {"kernel", op_name_string}}));
     }
 };
 } // namespace gpu

src/targets/gpu/jit/roialign.cpp

@@ -19,7 +19,6 @@ namespace gpu {
 // NOLINTNEXTLINE
 static const char* const roialign_kernel = R"__migraphx__(
 #include <migraphx/kernels/roialign.hpp>
-#include <migraphx/kernels/basic_ops.hpp>
 #include <migraphx/kernels/integral_constant.hpp>
 #include <migraphx/kernels/generic_constant.hpp>
 #include <args.hpp>

src/targets/gpu/jit/scatternd.cpp

@@ -19,7 +19,6 @@ namespace gpu {
 // NOLINTNEXTLINE
 static const char* const scatternd_kernel = R"__migraphx__(
 #include <migraphx/kernels/scatternd.hpp>
-#include <migraphx/kernels/basic_ops.hpp>
 #include <migraphx/kernels/integral_constant.hpp>
 #include <migraphx/kernels/generic_constant.hpp>
 #include <args.hpp>

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

@@ -146,8 +146,8 @@ struct array
 
     constexpr array carry(array result) const
     {
-        uint32_t overflow = 0;
-        for(std::ptrdiff_t i = result.size() - 1; i > 0; i--)
+        index_int overflow = 0;
+        for(diff_int i = result.size() - 1; i > 0; i--)
         {
             auto z = result[i] + overflow;
             // Reset overflow

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

-#ifndef MIGRAPHX_GUARD_AMDMIGRAPHX_KERNELS_BASIC_OPS_HPP
-#define MIGRAPHX_GUARD_AMDMIGRAPHX_KERNELS_BASIC_OPS_HPP
-
-#include <migraphx/kernels/types.hpp>
-
-namespace migraphx {
-
-struct sum
-{
-    template <class T, class U>
-    constexpr auto operator()(T x, U y) const
-    {
-        return x + y;
-    }
-};
-
-struct product
-{
-    template <class T, class U>
-    constexpr auto operator()(T x, U y) const
-    {
-        return x * y;
-    }
-};
-
-struct id
-{
-    template <class T>
-    constexpr auto operator()(T x) const
-    {
-        return x;
-    }
-};
-
-struct mean
-{
-    size_t item_num = 1;
-    template <class T>
-    constexpr auto operator()(T x) const
-    {
-        return x / static_cast<T>(item_num);
-    }
-};
-
-struct max_f
-{
-    template <class T, class U>
-    constexpr auto operator()(T x, U y) const
-    {
-        return (x > y) ? x : y;
-    }
-};
-inline constexpr auto max = max_f{};
-
-struct min_f
-{
-    template <class T, class U>
-    constexpr auto operator()(T x, U y) const
-    {
-        return (x < y) ? x : y;
-    }
-};
-inline constexpr auto min = min_f{};
-
-struct lowest
-{
-    template <class T>
-    constexpr operator T() const
-    {
-        return std::numeric_limits<T>::lowest();
-    }
-};
-
-struct highest
-{
-    template <class T>
-    constexpr operator T() const
-    {
-        return std::numeric_limits<T>::max();
-    }
-};
-
-} // namespace migraphx
-#endif // MIGRAPHX_GUARD_AMDMIGRAPHX_KERNELS_BASIC_OPS_HPP

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

@@ -137,7 +137,7 @@ constexpr auto by(F f)
 template <class F, class... Ts>
 constexpr void each_args(F f, Ts&&... xs)
 {
-    swallow{(f(std::forward<Ts>(xs)), 0)...};
+    swallow{(f(static_cast<Ts&&>(xs)), 0)...};
 }
 
 template <class F>

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

@@ -13,7 +13,7 @@ struct basic_iota_iterator
     F f;
 
     using difference_type = diff_int;
-    using reference       = decltype(f(std::declval<Iterator>()));
+    using reference       = decltype(f(declval<Iterator>()));
     using value_type      = remove_reference_t<reference>;
     using pointer         = add_pointer_t<value_type>;
 

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

@@ -3,14 +3,15 @@
 
 #include <migraphx/kernels/index.hpp>
 #include <migraphx/kernels/dfor.hpp>
-#include <migraphx/kernels/basic_ops.hpp>
+#include <migraphx/kernels/ops.hpp>
+#include <migraphx/kernels/math.hpp>
 #include <migraphx/kernels/array.hpp>
 
 namespace migraphx {
 
 struct max_pool
 {
-    MIGRAPHX_DEVICE_CONSTEXPR auto init() { return lowest(); }
+    MIGRAPHX_DEVICE_CONSTEXPR auto init() { return lowest{}; }
 
     template <class T>
     MIGRAPHX_DEVICE_CONSTEXPR T operator()(T x, T y)
@@ -55,7 +56,7 @@ MIGRAPHX_DEVICE_CONSTEXPR typename Iterator::value_type bilinear_interpolate(
             return 0;
         }
 
-        xy[ii]   = max(xy[ii], 0.0f);
+        xy[ii]   = migraphx::max(xy[ii], 0.0f);
         low[ii]  = xy[ii];
         high[ii] = low[ii] + 1;
         if(low[ii] >= dims[ii] - 1)
@@ -164,11 +165,12 @@ __device__ void roialign(const T& x_t, const U& rois_t, const V& ind_t, W& y_t,
         for(index_int ii = 0; ii < roi_size.size(); ++ii)
         {
             roi_size[ii] = roi_ends[ii] - roi_starts[ii];
-            roi_size[ii] = max(roi_size[ii], 1.0f);
+            roi_size[ii] = migraphx::max(roi_size[ii], 1.0f);
 
-            bin_size[ii] = roi_size[ii] / out_dims[ii];
-            bin_grid_size[ii] =
-                (s.sampling_ratio > 0) ? s.sampling_ratio : std::ceil(roi_size[ii] / out_dims[ii]);
+            bin_size[ii]      = roi_size[ii] / out_dims[ii];
+            bin_grid_size[ii] = (s.sampling_ratio > 0)
+                                    ? s.sampling_ratio
+                                    : migraphx::ceil(roi_size[ii] / out_dims[ii]);
         }
 
         const auto offset_x = x + ((batch_ind * channel_num + c) * in_dims[0] * in_dims[1]);

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

@@ -11,7 +11,7 @@ template <class T>
 struct tensor_view_iterator_read
 {
     T* view;
-    constexpr auto& operator()(std::size_t n) const
+    constexpr auto& operator()(index_int n) const
     {
         MIGRAPHX_ASSERT(view != nullptr);
         return (*view)[n];

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

@@ -35,6 +35,21 @@ struct enable_if<true, T>
 template <bool B, class T = void>
 using enable_if_t = typename enable_if<B, T>::type;
 
+template <bool B, class T, class F>
+struct conditional
+{
+    using type = T;
+};
+
+template <class T, class F>
+struct conditional<false, T, F>
+{
+    using type = F;
+};
+
+template <bool B, class T, class F>
+using conditional_t = typename conditional<B, T, F>::type;
+
 // NOLINTNEXTLINE
 #define MIGRAPHX_BUILTIN_TYPE_TRAIT1(name)   \
     template <class T>                       \