Merge remote-tracking branch 'origin/bert-opt' into HEAD

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

@@ -30,14 +30,14 @@
 namespace migraphx {
 inline namespace MIGRAPHX_INLINE_NS {
 
-struct module;
+struct module_pass_manager;
 
 namespace gpu {
 
 struct prefuse_ops
 {
     std::string name() const { return "gpu::prefuse_ops"; }
-    void apply(module& m) const;
+    void apply(module_pass_manager& m) const;
 };
 
 } // namespace gpu

src/targets/gpu/jit/reduce.cpp

@@ -118,16 +118,14 @@ struct reduce_compiler : compiler<reduce_compiler>
         options.virtual_inputs = reduce_dims(inputs);
         auto faxis             = find_fast_axis({options.virtual_inputs.front()});
         vectorize vec{};
-        // Vectorize if the axis is a reduction axis
-        if(options.virtual_inputs.back().lens()[faxis] == 1)
-        {
-            vec = vectorize::elements(ctx, faxis, options.virtual_inputs);
-        }
-        auto relements = get_reduce_elements(options.virtual_inputs) / vec.size;
         auto nelements = options.virtual_inputs.back().elements();
         auto algo      = v.get("algo", get_reduce_algo(options.virtual_inputs));
         if(algo == "block")
         {
+            // Vectorize if the axis is a reduction axis
+            if(options.virtual_inputs.back().lens()[faxis] == 1)
+                vec = vectorize::elements(ctx, faxis, options.virtual_inputs);
+            auto relements  = get_reduce_elements(options.virtual_inputs) / vec.size;
             auto block_size = compute_block_size(relements, 256);
             options.set_launch_params(
                 v, compute_global_for(ctx, nelements * block_size, 256), block_size);
@@ -166,7 +164,7 @@ struct reduce_compiler : compiler<reduce_compiler>
             auto reduce_elements = get_reduce_elements(ins->inputs());
             auto reduce_type     = ins->inputs().front()->get_shape().type();
             v["reduction"]       = "op::sum{}";
-            std::string mean     = "op::mean{" + std::to_string(reduce_elements) + "}";
+            std::string mean     = "op::mean<" + std::to_string(reduce_elements) + ">{}";
             // Use float accumulator when reduction size is too large for half
             if(reduce_type == shape::half_type and reduce_elements > 16384)
                 v["read"] = "compose(" + mean + ", op::convert_to<float>{})";

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

@@ -105,7 +105,7 @@ constexpr auto array_for_each(T& x, Ts&... xs)
             }
             else
             {
-                using vec_type = std::remove_reference_t<decltype(array2vec(x))>;
+                using vec_type = remove_reference_t<decltype(array2vec(x))>;
                 f(array2vec(x), __builtin_convertvector(array2vec(xs), vec_type)...);
             }
         }

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

@@ -178,5 +178,9 @@ MIGRAPHX_HIP_NORETURN inline __host__ __device__ void assert_fail(const source_l
 #define MIGRAPHX_WARN(...)
 #endif
 
+#define MIGRAPHX_STATIC_ASSERT_FOR(...) \
+    static_assert(__VA_ARGS__);         \
+    if constexpr(__VA_ARGS__)
+
 } // namespace migraphx
 #endif // MIGRAPHX_GUARD_KERNELS_DEBUG_HPP

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

@@ -72,7 +72,7 @@ __device__ T dpp_mov(T& x)
     }
     return output.data;
 }
-#endif
+#endif // MIGRAPHX_HAS_DPP
 
 } // namespace migraphx
 #endif // MIGRAPHX_GUARD_KERNELS_DPP_HPP

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

@@ -26,7 +26,7 @@
 
 #include <migraphx/kernels/index.hpp>
 #include <migraphx/kernels/algorithm.hpp>
-
+#include <migraphx/kernels/ops.hpp>
 namespace migraphx {
 
 template <class T>
@@ -53,23 +53,17 @@ __device__ void gathernd(const T& data_t, const U& indices_t, const V& output_t,
     auto indices_shape_lens = indices_shape.lens;
     auto data_shape_lens    = data_shape.lens;
     auto num_slice_dims     = indices_shape_lens.back();
-    std::size_t num_slices  = accumulate(indices_shape_lens.begin(),
-                                        indices_shape_lens.end() - 1,
-                                        1,
-                                        std::multiplies<std::size_t>());
-    std::size_t slice_size  = accumulate(data_shape_lens.begin() + num_slice_dims + batch_dims,
+    std::size_t num_slices =
+        accumulate(indices_shape_lens.begin(), indices_shape_lens.end() - 1, 1, op::product{});
+    std::size_t slice_size = accumulate(data_shape_lens.begin() + num_slice_dims + batch_dims,
                                         data_shape_lens.end(),
                                         1,
-                                        std::multiplies<std::size_t>());
-    const std::size_t num_batches       = accumulate(data_shape_lens.begin(),
-                                               data_shape_lens.begin() + batch_dims,
-                                               1,
-                                               std::multiplies<std::size_t>());
-    const std::size_t data_batch_stride = accumulate(data_shape_lens.begin() + batch_dims,
-                                                     data_shape_lens.end(),
-                                                     1,
-                                                     std::multiplies<std::size_t>());
-    const auto num_slices_per_batch     = num_slices / num_batches;
+                                        op::product{});
+    const std::size_t num_batches =
+        accumulate(data_shape_lens.begin(), data_shape_lens.begin() + batch_dims, 1, op::product{});
+    const std::size_t data_batch_stride =
+        accumulate(data_shape_lens.begin() + batch_dims, data_shape_lens.end(), 1, op::product{});
+    const auto num_slices_per_batch = num_slices / num_batches;
 
     ind.global_stride(output_shape.elements(), [&](auto i) {
         const auto* indices_ptr     = indices_t.data();
@@ -83,15 +77,15 @@ __device__ void gathernd(const T& data_t, const U& indices_t, const V& output_t,
             int64_t index                   = slice_indices[idx];
             const std::size_t input_dim_idx = batch_dims + idx;
             const auto input_dim            = data_shape_lens[input_dim_idx];
-            assert(index >= -static_cast<int64_t>(input_dim) and
-                   index < static_cast<int64_t>(input_dim));
+            MIGRAPHX_ASSERT(index >= -static_cast<int64_t>(input_dim) and
+                            index < static_cast<int64_t>(input_dim));
             if(index < 0)
                 index += input_dim;
             std::size_t size_from_slice_dims =
                 accumulate(data_shape_lens.begin() + batch_dims + idx + 1,
                            data_shape_lens.begin() + batch_dims + num_slice_dims,
                            slice_size,
-                           std::multiplies<std::size_t>());
+                           op::product{});
             relative_slice_offset += index * size_from_slice_dims;
         }
 

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

@@ -24,11 +24,18 @@
 #ifndef MIGRAPHX_GUARD_KERNELS_HIP_HPP
 #define MIGRAPHX_GUARD_KERNELS_HIP_HPP
 
+#ifndef MIGRAPHX_USE_HIPRTC
 // Workaround macro redefinition issue with clang tidy
 #if defined(__HIP_PLATFORM_HCC__) && defined(MIGRAPHX_USE_CLANG_TIDY)
 #undef __HIP_PLATFORM_HCC__ // NOLINT
 #endif
-
 #include <hip/hip_runtime.h>
+#include <hip/hip_fp16.h>
+#include <hip/math_functions.h>
+#include <hip/hip_math_constants.h>
+#elif defined(MIGRAPHX_ENABLE_HIPRTC_WORKAROUNDS)
+#include <hip/hip_common.h>
+#include <hip/hip_math_constants.h>
+#endif
 
 #endif // MIGRAPHX_GUARD_KERNELS_HIP_HPP

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

@@ -163,6 +163,30 @@ struct index
     }
 
     template <class F, class N, class Stride>
+    static constexpr void for_stride_loop_unroll(index_int start, N n, Stride stride, F f)
+    {
+        sequence(max_stride_iterations(n, stride), [&](auto... ks) {
+            fold([&](auto d, auto k) {
+                auto i = start + stride * k;
+                if(i < n)
+                    invoke_loop(f, i, d);
+                return d + _c<1>;
+            })(_c<0>, ks...);
+        });
+    }
+
+    template <class F, class N, class Stride>
+    static constexpr void for_stride_loop(index_int start, N n, Stride stride, F f)
+    {
+        index_int k = 0;
+        for(index_int i = start; i < n; i += stride)
+        {
+            invoke_loop(f, i, k);
+            k++;
+        }
+    }
+
+    template <bool Unroll, class F, class N, class Stride>
     static constexpr void for_stride(index_int start, N n, Stride stride, F f)
     {
         MIGRAPHX_ASSERT(start < stride);
@@ -180,46 +204,40 @@ struct index
                     invoke_loop(f, start, _c<0>);
                 }
             }
+            else if constexpr(Unroll)
+            {
+                MIGRAPHX_STATIC_ASSERT_FOR(max_stride_iterations(n, stride) < 256)
+                {
+                    for_stride_loop_unroll(start, n, stride, f);
+                }
+            }
             else
             {
-                static_assert(max_stride_iterations(n, stride) < 128);
-                sequence(max_stride_iterations(n, stride), [&](auto... ks) {
-                    fold([&](auto d, auto k) {
-                        auto i = start + stride * k;
-                        if(i < n)
-                            invoke_loop(f, i, d);
-                        return d + _c<1>;
-                    })(_c<0>, ks...);
-                });
+                for_stride_loop(start, n, stride, f);
             }
         }
         else
         {
-            index_int k = 0;
-            for(index_int i = start; i < n; i += stride)
-            {
-                invoke_loop(f, i, k);
-                k++;
-            }
+            for_stride_loop(start, n, stride, f);
         }
     }
 
     template <class F, class N>
     __device__ void global_stride(N n, F f) const
     {
-        for_stride(global, n, nglobal(), f);
+        for_stride<false>(global, n, nglobal(), f);
     }
 
     template <class F, class N>
     __device__ void local_stride(N n, F f) const
     {
-        for_stride(local, n, nlocal(), f);
+        for_stride<true>(local, n, nlocal(), f);
     }
 
     template <class F, class N>
     __device__ void group_stride(N n, F f) const
     {
-        for_stride(group, n, ngroup(), f);
+        for_stride<false>(group, n, ngroup(), f);
     }
 };
 

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

@@ -28,8 +28,7 @@
 #include <migraphx/kernels/vec.hpp>
 #include <migraphx/kernels/functional.hpp>
 #include <migraphx/kernels/type_traits.hpp>
-#include <hip/hip_fp16.h>
-#include <hip/math_functions.h>
+#include <migraphx/kernels/hip.hpp>
 
 namespace migraphx {
 
@@ -222,7 +221,7 @@ constexpr auto min(const T& a, const U& b)
 template <class T, MIGRAPHX_REQUIRES(is_same<vec_type<T>, half>{})>
 constexpr T sin(T x)
 {
-    constexpr const T shift = M_PI_2;
+    constexpr const T shift = HIP_PIO2_F;
     return migraphx::cos(shift - x);
 }
 

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

@@ -66,13 +66,22 @@ struct convert_to
     }
 };
 
+template <index_int N>
 struct mean
 {
-    index_int item_num = 1;
     template <class T>
-    MIGRAPHX_DEVICE_CONSTEXPR auto operator()(T x) const
+    MIGRAPHX_DEVICE_CONSTEXPR T operator()(T x) const
     {
-        return x / static_cast<T>(item_num);
+        using type = vec_type<T>;
+        if constexpr(is_floating_point<type>{})
+        {
+            constexpr type d = 1.0 / N;
+            return x * d;
+        }
+        else
+        {
+            return x / static_cast<type>(N);
+        }
     }
 };
 

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

@@ -391,22 +391,40 @@ struct block
 struct lane
 {
     template <class Slicer>
-    struct reducer
+    struct reducer : reducer_base<reducer<Slicer>>
     {
         index idx;
         Slicer slice;
-        template <class Op, class T, class Read>
-        __device__ auto reduce(Op op, T init, Read read) const
+
+        template <class Size, class F>
+        struct inner_storage : inner_storage_tag
         {
-            return sliced(slice, [=](auto x, auto... xs) {
-                using type = typename decltype(x)::type;
-                type r     = init;
-                for(index_int j = 0; j < x.get_shape().elements(); j++)
-                {
-                    r = op(r, read(x[j], xs[j]...));
-                }
-                return r;
-            });
+            using type = remove_reference_t<decltype(declval<F>()(0, _c<0>))>;
+            F f;
+            constexpr Size rsize() const { return {}; }
+            template <class U, class V>
+            constexpr auto operator()(U j, V d) const
+            {
+                return f(j, d);
+            }
+        };
+
+        template <class Size, class F>
+        constexpr inner_storage<Size, F> make_inner_storage(Size, F f)
+        {
+            return {f};
+        }
+
+        template <class Op, class T, class Read, class N, class U, class... Us>
+        __device__ auto reduce_impl(Op op, T init, Read read, N n, U&& x, Us&&... xs) const
+        {
+            using type = remove_reference_t<decltype(x(0, _c<0>))>;
+            type r     = init;
+            for(index_int j = 0; j < n; j++)
+            {
+                r = op(r, read(x(j, _c<0>), xs(j, _c<0>)...));
+            }
+            return r;
         }
 
         template <class F>
@@ -415,29 +433,25 @@ struct lane
             f();
         }
 
-        template <class F>
-        __device__ auto inner(F f) const
+        template <class F, class N, class... Ts>
+        __device__ void inner_void_impl(F f, N n, Ts&&... xs) const
         {
-            return sliced(slice, [=](auto x, auto... xs) {
-                for(index_int j = 0; j < x.get_shape().elements(); j++)
-                {
-                    f(x[j], xs[j]...);
-                }
-            });
+            for(index_int j = 0; j < n; j++)
+            {
+                f(xs(j, _c<0>)...);
+            }
         }
 
-        template <class Input>
-        constexpr auto elements() const
+        template <class R, class F, class N, class... Ts>
+        __device__ auto inner_impl(F f, N n, Ts&&... xs) const
         {
-            using reduce_type = decltype(slice(Input{}));
-            return get_shape_c<reduce_type>{}.elements();
+            return make_inner_storage(n, [=](auto j, auto d) { return f(xs(j, d)...); });
         }
     };
-
     template <class Slicer>
     static __device__ auto make(index idx, Slicer slicer)
     {
-        return reducer<Slicer>{idx, slicer};
+        return reducer<Slicer>{{}, idx, slicer};
     }
 
     template <class Output, class F>

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

@@ -76,14 +76,6 @@ struct shape
 
     constexpr index_int index(index_array x) const { return x.dot(strides); }
 
-    constexpr index_int index(std::initializer_list<index_int> x) const
-    {
-        index_int idx = 0;
-        for(index_int i = 0; i < x.size(); i++)
-            idx += *(x.begin() + i) * strides[i];
-        return idx;
-    }
-
     constexpr index_int index(index_int i) const
     {
         if(this->standard())

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

@@ -28,8 +28,45 @@
 
 namespace migraphx {
 
-using index_int = std::uint32_t;
-using diff_int  = std::int32_t;
+#if defined(MIGRAPHX_ENABLE_HIPRTC_WORKAROUNDS) and defined(MIGRAPHX_USE_HIPRTC)
+using int8_t   = signed char;
+using uint8_t  = unsigned char;
+using int16_t  = signed short;
+using uint16_t = unsigned short;
+using int32_t  = signed int;
+using uint32_t = unsigned int;
+using int64_t  = signed long long;
+using uint64_t = unsigned long long;
+#elif defined(MIGRAPHX_USE_HIPRTC)
+using int8_t   = __hip_int8_t;
+using uint8_t  = __hip_uint8_t;
+using int16_t  = __hip_int16_t;
+using uint16_t = __hip_uint16_t;
+using int32_t  = __hip_int32_t;
+using uint32_t = __hip_uint32_t;
+using int64_t  = __hip_int64_t;
+using uint64_t = __hip_uint64_t;
+#else
+using int8_t   = std::int8_t;
+using uint8_t  = std::uint8_t;
+using int16_t  = std::int16_t;
+using uint16_t = std::uint16_t;
+using int32_t  = std::int32_t;
+using uint32_t = std::uint32_t;
+using int64_t  = std::int64_t;
+using uint64_t = std::uint64_t;
+#endif // MIGRAPHX_USE_HIPRTC
+using index_int = uint32_t;
+using diff_int  = int32_t;
+
+static_assert(sizeof(int8_t) == 1, "int8_t must be 1 bytes");
+static_assert(sizeof(uint8_t) == 1, "uint8_t must be 1 bytes");
+static_assert(sizeof(int16_t) == 2, "int16_t must be 2 bytes");
+static_assert(sizeof(uint16_t) == 2, "uint16_t must be 2 bytes");
+static_assert(sizeof(int32_t) == 4, "int32_t must be 4 bytes");
+static_assert(sizeof(uint32_t) == 4, "uint32_t must be 4 bytes");
+static_assert(sizeof(int64_t) == 8, "int64_t must be 8 bytes");
+static_assert(sizeof(uint64_t) == 8, "uint64_t must be 8 bytes");
 
 #define MIGRAPHX_DEVICE_CONSTEXPR constexpr __device__ __host__ // NOLINT
 

src/targets/gpu/prefuse_ops.cpp

@@ -26,6 +26,8 @@
 #include <migraphx/check_shapes.hpp>
 #include <migraphx/make_op.hpp>
 #include <migraphx/register_op.hpp>
+#include <migraphx/pass_manager.hpp>
+#include <migraphx/dead_code_elimination.hpp>
 
 namespace migraphx {
 inline namespace MIGRAPHX_INLINE_NS {
@@ -90,7 +92,9 @@ struct find_layernorm
     {
         auto ins   = r.result;
         auto x_ins = r.instructions["x"];
-        auto eps   = r.instructions["eps"]->eval().at<float>();
+        float eps  = 0;
+        if(contains(r.instructions, "eps"))
+            eps = r.instructions["eps"]->eval().at<float>();
 
         m.replace_instruction(ins, layernorm{eps}, x_ins);
     }
@@ -100,25 +104,26 @@ struct find_add_layernorm
 {
     auto matcher() const
     {
-        return match::layernorm()(match::var("x")(match::name("add").bind("add")));
+        return match::name("gpu::prelayernorm")(
+            match::args(match::name("add")(match::used_once()).bind("add")));
     }
 
     void apply(module& m, const match::matcher_result& r) const
     {
         auto ins     = r.result;
         auto add_ins = r.instructions["add"];
-        float eps    = 0;
-        if(contains(r.instructions, "eps"))
-            eps = r.instructions["eps"]->eval().at<float>();
+        auto op      = any_cast<layernorm>(ins->get_operator());
 
-        m.replace_instruction(ins, add_layernorm{eps}, add_ins->inputs());
+        m.replace_instruction(ins, add_layernorm{op.epsilon}, add_ins->inputs());
     }
 };
 } // namespace
 
-void prefuse_ops::apply(module& m) const
+void prefuse_ops::apply(module_pass_manager& mpm) const
 {
-    match::find_matches(m, find_add_layernorm{}, find_layernorm{});
+    match::find_matches(mpm.get_module(), find_layernorm{});
+    mpm.run_pass(dead_code_elimination{});
+    match::find_matches(mpm.get_module(), find_add_layernorm{});
 }
 
 } // namespace gpu

src/targets/gpu/target.cpp

@@ -38,7 +38,7 @@
 #include <migraphx/layout_nhwc.hpp>
 #include <migraphx/memory_coloring.hpp>
 #include <migraphx/normalize_ops.hpp>
-#include <migraphx/optimize.hpp>
+#include <migraphx/optimize_module.hpp>
 #include <migraphx/preallocate_param.hpp>
 #include <migraphx/propagate_constant.hpp>
 #include <migraphx/register_target.hpp>
@@ -121,18 +121,13 @@ std::vector<pass> target::get_passes(migraphx::context& gctx, const compile_opti
         rewrite_pooling{},
         dead_code_elimination{},
         rewrite_gelu{},
-        optimize{},
+        optimize_module{},
         enable_pass(enabled(MIGRAPHX_ENABLE_NHWC{}), layout_nhwc{}),
         dead_code_elimination{},
         prefuse_ops{},
         dead_code_elimination{},
         auto_contiguous{},
-        simplify_reshapes{},
-        propagate_constant{},
-        dead_code_elimination{},
-        fuse_ck_gemm_softmax_gemm{&ctx},
-        dead_code_elimination{},
-        optimize{},
+        optimize_module{},
         enable_pass(not enabled(MIGRAPHX_DISABLE_POINTWISE_FUSION{}), fuse_pointwise{}),
         dead_code_elimination{},
         fuse_mlir{&ctx},

src/targets/ref/lowering.cpp

@@ -132,109 +132,6 @@ auto visit_quantize(T&& x, Ts&&... xs)
     };
 }
 
-template <class Op>
-struct ref_convolution : auto_register_op<ref_convolution<Op>>
-{
-    ref_convolution() = default;
-
-    ref_convolution(Op pop) : op(std::move(pop)) {}
-
-    Op op;
-
-    template <class Self, class F>
-    static auto reflect(Self& self, F f)
-    {
-        return migraphx::reflect(self.op, f);
-    }
-
-    std::string name() const { return "ref::" + op.name(); }
-    shape compute_shape(const std::vector<shape>& inputs) const
-    {
-        return op.normalize_compute_shape(inputs);
-    }
-
-    argument compute(context&, shape output_shape, std::vector<argument> args) const
-    {
-        std::vector<std::size_t> padding;
-        if(op.padding_mode != op::padding_mode_t::default_)
-        {
-            auto input_lens   = args[0].get_shape().lens();
-            auto weights_lens = args[1].get_shape().lens();
-            padding =
-                op.padding_mode == op::same_upper
-                    ? calc_dyn_auto_pad(input_lens, weights_lens, op.stride, op.dilation, true)
-                    : calc_dyn_auto_pad(input_lens, weights_lens, op.stride, op.dilation, false);
-            output_shape = compute_padded_shape(
-                args[0].get_shape(), args[1].get_shape(), padding, op.stride, op.dilation);
-        }
-        else
-        {
-            padding = op.padding;
-            if(output_shape.dynamic())
-            {
-                output_shape =
-                    op.normalize_compute_shape({args.at(0).get_shape(), args.at(1).get_shape()});
-            }
-        }
-
-        argument result{output_shape};
-        visit_quantize(result, args[0], args[1])([&](auto output, auto input, auto weights) {
-            auto in_lens = input.get_shape().lens();
-
-            auto wei_lens = weights.get_shape().lens();
-            auto wei_n    = wei_lens[0];
-            auto wei_c    = wei_lens[1];
-            std::vector<std::size_t> win_size(wei_lens.begin() + 1, wei_lens.end());
-
-            par_for(output_shape.elements(), [&](auto i) {
-                auto idx_o = output_shape.multi(i);
-                auto w     = idx_o[1];
-                auto n_dim = idx_o.size();
-
-                std::vector<std::ptrdiff_t> win_start;
-                for(std::size_t dim = 2; dim < n_dim; ++dim)
-                {
-                    auto d_2 = dim - 2;
-                    win_start.push_back(std::ptrdiff_t(idx_o[dim] * op.stride[d_2]) -
-                                        std::ptrdiff_t(padding[d_2]));
-                }
-                const auto group_id = w / (wei_n / op.group);
-
-                shape win_shape{output_shape.type(), win_size};
-
-                double acc = 0.0;
-                shape_for_each(win_shape, [&](auto idx_win) {
-                    auto k           = idx_win[0];
-                    const auto in_ch = group_id * wei_c + k;
-                    std::vector<std::ptrdiff_t> idx(idx_o.begin(), idx_o.end());
-                    idx[1] = in_ch;
-                    std::transform(idx_win.begin() + 1,
-                                   idx_win.end(),
-                                   win_start.begin(),
-                                   idx.begin() + 2,
-                                   [](std::ptrdiff_t ii, std::ptrdiff_t jj) { return ii + jj; });
-                    std::vector<std::ptrdiff_t> idx_wei(idx_o.size());
-                    idx_wei[0] = w;
-                    std::copy(idx_win.begin(), idx_win.end(), idx_wei.begin() + 1);
-                    if(std::all_of(idx.begin() + 2, idx.end(), [&](auto ii) { return ii >= 0; }) and
-                       std::equal(idx.begin(),
-                                  idx.end(),
-                                  in_lens.begin(),
-                                  in_lens.end(),
-                                  std::less<std::ptrdiff_t>{}))
-                    {
-                        acc +=
-                            input(idx.begin(), idx.end()) * weights(idx_wei.begin(), idx_wei.end());
-                    }
-                });
-
-                output[i] = acc;
-            });
-        });
-        return result;
-    }
-};
-
 struct ref_im2col
 {
     op::im2col op;
@@ -564,11 +461,8 @@ struct ref_apply
 
     void init()
     {
-        apply_map["convolution"] = extend_op<ref_convolution<op::convolution>, op::convolution>();
-        apply_map["dot"]         = extend_op<ref_gemm, op::dot>();
-        apply_map["quant_dot"]   = extend_op<ref_quant_gemm, op::quant_dot>();
-        apply_map["quant_convolution"] =
-            extend_op<ref_convolution<op::quant_convolution>, op::quant_convolution>();
+        apply_map["dot"]        = extend_op<ref_gemm, op::dot>();
+        apply_map["quant_dot"]  = extend_op<ref_quant_gemm, op::quant_dot>();
         apply_map["im2col"]     = extend_op<ref_im2col, op::im2col>();
         apply_map["logsoftmax"] = extend_op<ref_softmax<op::logsoftmax>, op::logsoftmax>();
         apply_map["lrn"]        = extend_op<ref_lrn, op::lrn>();

test/CMakeLists.txt

-#####################################################################################
+# ####################################################################################
 # The MIT License (MIT)
 #
 # Copyright (c) 2015-2022 Advanced Micro Devices, Inc. All rights reserved.
@@ -20,7 +20,7 @@
 # LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
 # OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN
 # THE SOFTWARE.
-#####################################################################################
+# ####################################################################################
 
 cmake_policy(SET CMP0057 NEW)
 
@@ -49,27 +49,31 @@ function(add_test_command NAME EXE)
         set_tests_properties(${NAME} PROPERTIES DISABLED On)
     elseif(WIN32)
         set(WINPATH)
+
         foreach(PATH ${CMAKE_FIND_ROOT_PATH})
             list(APPEND WINPATH ${PATH}/bin)
         endforeach()
+
         file(GENERATE OUTPUT "${CMAKE_CURRENT_BINARY_DIR}/test_${NAME}.cmd"
             CONTENT "set PATH=${WINPATH};%PATH%
                     %1 ${ARGN}")
         add_test(NAME ${NAME} COMMAND ${WINE_CMD} cmd /c "${CMAKE_CURRENT_BINARY_DIR}/test_${NAME}.cmd" $<TARGET_FILE:${EXE}>)
     else()
         if(MIGRAPHX_TEST_GDB)
-            # add_test(NAME ${NAME} COMMAND ${MIGRAPHX_GDB} 
-            #     --batch
-            #     --return-child-result
-            #     -ex "set disable-randomization off"
-            #     -ex run
-            #     -ex backtrace
-            #     --args $<TARGET_FILE:${EXE}> ${ARGN})
+            # add_test(NAME ${NAME} COMMAND ${MIGRAPHX_GDB}
+            # --batch
+            # --return-child-result
+            # -ex "set disable-randomization off"
+            # -ex run
+            # -ex backtrace
+            # --args $<TARGET_FILE:${EXE}> ${ARGN})
             set(TEST_DIR ${CMAKE_CURRENT_BINARY_DIR}/gdb/test_${NAME})
             file(MAKE_DIRECTORY ${TEST_DIR})
-            if (NOT EXISTS ${TEST_DIR})
+
+            if(NOT EXISTS ${TEST_DIR})
                 message(FATAL_ERROR "Failed to create test directory: ${TEST_DIR}")
             endif()
+
             file(GENERATE OUTPUT "${TEST_DIR}/run.cmake"
                 CONTENT "
                 # Remove previous core dump
@@ -90,22 +94,27 @@ function(add_test_command NAME EXE)
             add_test(NAME ${NAME} COMMAND ${EXE} ${ARGN})
         endif()
     endif()
+
     set_tests_properties(${NAME} PROPERTIES FAIL_REGULAR_EXPRESSION "FAILED")
 endfunction()
 
 function(add_test_executable TEST_NAME)
-    add_executable (${TEST_NAME} EXCLUDE_FROM_ALL ${ARGN})
+    add_executable(${TEST_NAME} EXCLUDE_FROM_ALL ${ARGN})
     target_link_libraries(${TEST_NAME} ${CMAKE_THREAD_LIBS_INIT})
+
     # Cmake does not add flags correctly for gcc
-    if(CMAKE_CXX_COMPILER_ID MATCHES "GNU") 
+    if(CMAKE_CXX_COMPILER_ID MATCHES "GNU")
         set_target_properties(${TEST_NAME} PROPERTIES COMPILE_FLAGS -pthread LINK_FLAGS -pthread)
     endif()
+
     separate_arguments(MIOPEN_TEST_FLAGS_ARGS UNIX_COMMAND ${MIOPEN_TEST_FLAGS})
+
     if(MIOPEN_TEST_ALL)
         set(TEST_COMMAND ${TEST_NAME} ${MIOPEN_TEST_FLOAT_ARG} --all ${MIOPEN_TEST_FLAGS_ARGS})
     else()
         set(TEST_COMMAND ${TEST_NAME} ${MIOPEN_TEST_FLOAT_ARG} ${MIOPEN_TEST_FLAGS_ARGS})
     endif()
+
     add_test_command(${TEST_NAME} ${TEST_COMMAND})
     add_dependencies(tests ${TEST_NAME})
     add_dependencies(check ${TEST_NAME})
@@ -129,11 +138,11 @@ if(MIGRAPHX_ENABLE_GPU)
         get_filename_component(BASE_NAME ${TEST} NAME_WE)
         add_test_executable(test_gpu_${BASE_NAME} ${TEST})
         rocm_clang_tidy_check(test_gpu_${BASE_NAME})
-        set_tests_properties(test_gpu_${BASE_NAME} PROPERTIES 
-            COST 10 
+        set_tests_properties(test_gpu_${BASE_NAME} PROPERTIES
+            COST 10
             RESOURCE_LOCK gpu
         )
-        target_link_libraries(test_gpu_${BASE_NAME} migraphx_gpu)
+        target_link_libraries(test_gpu_${BASE_NAME} migraphx_gpu migraphx_kernels)
     endforeach()
 endif()
 
@@ -145,8 +154,8 @@ if(MIGRAPHX_ENABLE_FPGA)
         get_filename_component(BASE_NAME ${TEST} NAME_WE)
         add_test_executable(test_fpga_${BASE_NAME} ${TEST})
         rocm_clang_tidy_check(test_fpga_${BASE_NAME})
-        set_tests_properties(test_fpga_${BASE_NAME} PROPERTIES 
-            COST 10 
+        set_tests_properties(test_fpga_${BASE_NAME} PROPERTIES
+            COST 10
             RESOURCE_LOCK fpga
         )
         target_link_libraries(test_fpga_${BASE_NAME} migraphx_fpga)
@@ -155,7 +164,8 @@ endif()
 
 # Onnx test
 set(TEST_ONNX_DIR ${CMAKE_CURRENT_SOURCE_DIR}/onnx)
-file (GLOB ONNX_TESTS ${TEST_ONNX_DIR}/*.cpp)
+file(GLOB ONNX_TESTS ${TEST_ONNX_DIR}/*.cpp)
+
 foreach(ONNX_TEST ${ONNX_TESTS})
     get_filename_component(BASE_NAME ${ONNX_TEST} NAME_WE)
     set(TEST_NAME test_${BASE_NAME})
@@ -163,7 +173,7 @@ foreach(ONNX_TEST ${ONNX_TESTS})
     rocm_clang_tidy_check(${TEST_NAME})
     target_link_libraries(${TEST_NAME} migraphx_onnx migraphx_ref)
     target_include_directories(${TEST_NAME} PUBLIC include)
-    add_test(NAME ${TEST_NAME} COMMAND $<TARGET_FILE:${TEST_NAME}> WORKING_DIRECTORY ${TEST_ONNX_DIR}) 
+    add_test(NAME ${TEST_NAME} COMMAND $<TARGET_FILE:${TEST_NAME}> WORKING_DIRECTORY ${TEST_ONNX_DIR})
     add_dependencies(tests ${TEST_NAME})
     add_dependencies(check ${TEST_NAME})
 endforeach()
@@ -174,26 +184,26 @@ add_executable(test_tf tf/tf_test.cpp)
 rocm_clang_tidy_check(test_tf)
 target_link_libraries(test_tf migraphx_tf migraphx_ref)
 target_include_directories(test_tf PUBLIC include)
-add_test(NAME test_tf COMMAND $<TARGET_FILE:test_tf> WORKING_DIRECTORY ${TEST_TF_DIR}) 
+add_test(NAME test_tf COMMAND $<TARGET_FILE:test_tf> WORKING_DIRECTORY ${TEST_TF_DIR})
 add_dependencies(tests test_tf)
 add_dependencies(check test_tf)
 
 add_subdirectory(api)
 add_subdirectory(verify)
+
 if(MIGRAPHX_ENABLE_PYTHON)
-add_subdirectory(py)
+    add_subdirectory(py)
 endif()
 
 function(test_header NAME HEADER)
-
-    file(WRITE ${CMAKE_CURRENT_BINARY_DIR}/header-main-include-${NAME}.cpp 
+    file(WRITE ${CMAKE_CURRENT_BINARY_DIR}/header-main-include-${NAME}.cpp
         "#include <${HEADER}>\nint main() {}\n"
     )
-    file(WRITE ${CMAKE_CURRENT_BINARY_DIR}/header-static-include-${NAME}.cpp 
+    file(WRITE ${CMAKE_CURRENT_BINARY_DIR}/header-static-include-${NAME}.cpp
         "#include <${HEADER}>\n"
     )
     add_test_executable(${NAME}
-        ${CMAKE_CURRENT_BINARY_DIR}/header-main-include-${NAME}.cpp 
+        ${CMAKE_CURRENT_BINARY_DIR}/header-main-include-${NAME}.cpp
         ${CMAKE_CURRENT_BINARY_DIR}/header-static-include-${NAME}.cpp
     )
 endfunction()
@@ -206,6 +216,7 @@ function(test_headers PREFIX)
         string(MAKE_C_IDENTIFIER ${HEADER_REL} TEST_NAME)
         get_filename_component(BASE_NAME ${HEADER} NAME_WE)
         test_header(header_${TEST_NAME} ${PREFIX}/${BASE_NAME}.hpp)
+
         if(MIGRAPHX_ENABLE_GPU)
             target_link_libraries(header_${TEST_NAME} migraphx_gpu)
         endif()
@@ -214,6 +225,7 @@ endfunction()
 
 test_headers(migraphx ${CMAKE_SOURCE_DIR}/src/include/migraphx/*.hpp)
 test_headers(migraphx/ref ${CMAKE_SOURCE_DIR}/src/targets/ref/include/migraphx/ref/*.hpp)
+
 if(MIGRAPHX_ENABLE_GPU)
-test_headers(migraphx/gpu ${CMAKE_SOURCE_DIR}/src/targets/gpu/include/migraphx/gpu/*.hpp)
+    test_headers(migraphx/gpu ${CMAKE_SOURCE_DIR}/src/targets/gpu/include/migraphx/gpu/*.hpp)
 endif()

test/api/test_save_load.cpp

@@ -30,7 +30,6 @@ TEST_CASE(load_save_default)
     std::string filename = "migraphx_api_load_save.mxr";
     auto p1              = migraphx::parse_onnx("conv_relu_maxpool_test.onnx");
     auto s1              = p1.get_output_shapes();
-
     migraphx::save(p1, filename.c_str());
     auto p2 = migraphx::load(filename.c_str());
     auto s2 = p2.get_output_shapes();

test/gpu/jit.cpp

@@ -35,13 +35,14 @@
 #include <migraphx/gpu/compile_hip.hpp>
 #include <migraphx/gpu/compile_hip_code_object.hpp>
 #include <migraphx/gpu/compiler.hpp>
+#include <migraphx_kernels.hpp>
 
 // NOLINTNEXTLINE
 const std::string write_2s = R"__migraphx__(
 #include <hip/hip_runtime.h>
 
 extern "C" {
-__global__ void write(int8_t* data) 
+__global__ void write(char* data) 
 {
     int num = threadIdx.x + blockDim.x * blockIdx.x;
     data[num] = 2;
@@ -58,7 +59,7 @@ const std::string add_2s_binary = R"__migraphx__(
 #include <hip/hip_runtime.h>
 
 extern "C" {
-__global__ void add_2(std::int8_t* x, std::int8_t* y) 
+__global__ void add_2(char* x, char* y) 
 {
     int num = threadIdx.x + blockDim.x * blockIdx.x;
     y[num] = x[num] + 2;
@@ -137,7 +138,8 @@ int main() {}
 const std::string math_template = R"__migraphx__(
 #include <migraphx/kernels/pointwise.hpp>
 #include <migraphx/kernels/math.hpp>
-
+#include <migraphx/kernels/types.hpp>
+using namespace migraphx;
 extern "C" {
 __global__ void kernel(${type}* p) 
 {

test/onnx/.onnxrt-commit

+c9a53c925510a101f5ca94d5ecda0924e40a8463