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

src/targets/gpu/elu.cpp

 #include <migraphx/gpu/elu.hpp>
-#include <migraphx/operators.hpp>
-#include <migraphx/manage_ptr.hpp>
-#include <migraphx/gpu/miopen.hpp>
-#include <utility>
+#include <migraphx/gpu/context.hpp>
 
 namespace migraphx {
 inline namespace MIGRAPHX_INLINE_NS {

src/targets/gpu/fuse_ops.cpp

@@ -3,6 +3,7 @@
 #include <migraphx/gpu/miopen.hpp>
 #include <migraphx/gpu/convolution.hpp>
 #include <migraphx/gpu/device/add_relu.hpp>
+#include <migraphx/gpu/device/add.hpp>
 #include <migraphx/instruction.hpp>
 
 namespace migraphx {

src/targets/gpu/gather.cpp

 #include <migraphx/gpu/gather.hpp>
-#include <migraphx/operators.hpp>
-#include <migraphx/manage_ptr.hpp>
-#include <migraphx/gpu/miopen.hpp>
-#include <migraphx/gpu/device/concat.hpp>
-#include <utility>
+#include <migraphx/gpu/context.hpp>
+#include <migraphx/gpu/device/gather.hpp>
 
 namespace migraphx {
 inline namespace MIGRAPHX_INLINE_NS {

src/targets/gpu/gemm.cpp

 #include <migraphx/gpu/gemm.hpp>
-#include <migraphx/operators.hpp>
-#include <migraphx/manage_ptr.hpp>
-#include <migraphx/gpu/miopen.hpp>
-#include <utility>
+#include <migraphx/gpu/context.hpp>
 
 namespace migraphx {
 inline namespace MIGRAPHX_INLINE_NS {
 namespace gpu {
 
+template <class... Ts>
+void generic_rocblas_batched_gemm(shape::as<float>, Ts&&... xs)
+{
+    rocblas_sgemm_strided_batched(std::forward<Ts>(xs)...);
+}
+
+template <class... Ts>
+void generic_rocblas_batched_gemm(shape::as<double>, Ts&&... xs)
+{
+    rocblas_dgemm_strided_batched(std::forward<Ts>(xs)...);
+}
+
+template <class... Ts>
+void generic_rocblas_batched_gemm(shape::as<half>, Ts&&... xs)
+{
+    rocblas_hgemm_strided_batched(std::forward<Ts>(xs)...);
+}
+
+template <class T, class... Ts>
+void generic_rocblas_batched_gemm(shape::as<T>, Ts&&...)
+{
+    MIGRAPHX_THROW("GENERIC_ROCBLAS_BATCHED_GEMM: type unsupported by rocblas");
+}
+
 template <class... Ts>
 void generic_rocblas_gemm(shape::as<float>, Ts&&... xs)
 {
@@ -29,7 +50,7 @@ void generic_rocblas_gemm(shape::as<half>, Ts&&... xs)
 template <class T, class... Ts>
 void generic_rocblas_gemm(shape::as<T>, Ts&&...)
 {
-    MIGRAPHX_THROW("Type unsupported by rocblas");
+    MIGRAPHX_THROW("GENERIC_ROCBLAS_GEMM: type unsupported by rocblas");
 }
 
 template <class T>
@@ -76,35 +97,45 @@ argument miopen_gemm::compute(context& ctx,
                               const shape& output_shape,
                               const std::vector<argument>& args) const
 {
-    float alpha     = 1.0f;
-    float beta      = 0.0f;
-    bool transa     = args[0].get_shape().transposed();
-    bool transb     = args[1].get_shape().transposed();
-    rocblas_int lda = args[0].get_shape().strides()[transa ? 1 : 0];
-    rocblas_int ldb = args[1].get_shape().strides()[transb ? 1 : 0];
-    rocblas_int ldc = args[2].get_shape().strides()[0];
-    rocblas_int m   = output_shape.lens()[0];
-    rocblas_int n   = output_shape.lens()[1];
-    rocblas_int k   = args[0].get_shape().lens()[1];
+    float alpha        = 1.0f;
+    float beta         = 0.0f;
+    bool transa        = args[0].get_shape().transposed();
+    bool transb        = args[1].get_shape().transposed();
+    std::size_t n_dims = args[0].get_shape().lens().size();
+    std::size_t dim_0  = n_dims - 2;
+    std::size_t dim_1  = n_dims - 1;
+    rocblas_int lda    = args[0].get_shape().strides()[transa ? dim_1 : dim_0];
+    rocblas_int ldb    = args[1].get_shape().strides()[transb ? dim_1 : dim_0];
+    rocblas_int ldc    = args[2].get_shape().strides()[dim_0];
+    auto out_lens      = output_shape.lens();
+    rocblas_int m      = out_lens[dim_0];
+    rocblas_int n      = out_lens[dim_1];
+    rocblas_int k      = args[0].get_shape().lens()[dim_1];
+    auto batch_num     = std::accumulate(
+        out_lens.rbegin() + 2, out_lens.rend(), std::size_t{1}, std::multiplies<std::size_t>());
     output_shape.visit_type([&](auto as) {
         auto alpha_r    = to_rocblas_type(as(alpha));
         auto beta_r     = to_rocblas_type(as(beta));
         auto to_pointer = [&](auto&& arg) { return to_rocblas_type(as.from(arg.data())); };
-        generic_rocblas_gemm(as,
-                             ctx.get_stream().get_rocblas(),
-                             transb ? rocblas_operation_transpose : rocblas_operation_none,
-                             transa ? rocblas_operation_transpose : rocblas_operation_none,
-                             n,
-                             m,
-                             k,
-                             &alpha_r,
-                             to_pointer(args[1]),
-                             ldb,
-                             to_pointer(args[0]),
-                             lda,
-                             &beta_r,
-                             to_pointer(args[2]),
-                             ldc);
+        generic_rocblas_batched_gemm(as,
+                                     ctx.get_stream().get_rocblas(),
+                                     transb ? rocblas_operation_transpose : rocblas_operation_none,
+                                     transa ? rocblas_operation_transpose : rocblas_operation_none,
+                                     n,
+                                     m,
+                                     k,
+                                     &alpha_r,
+                                     to_pointer(args[1]),
+                                     ldb,
+                                     k * n,
+                                     to_pointer(args[0]),
+                                     lda,
+                                     m * k,
+                                     &beta_r,
+                                     to_pointer(args[2]),
+                                     ldc,
+                                     m * n,
+                                     batch_num);
 
     });
 

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

 #ifndef MIGRAPHX_GUARD_RTGLIB_ABS_HPP
 #define MIGRAPHX_GUARD_RTGLIB_ABS_HPP
 
-#include <migraphx/gpu/lowering.hpp>
-#include <migraphx/manage_ptr.hpp>
-#include <migraphx/instruction.hpp>
-#include <migraphx/operators.hpp>
-#include <migraphx/generate.hpp>
-#include <migraphx/shape_for_each.hpp>
-#include <migraphx/config.hpp>
+#include <migraphx/shape.hpp>
 #include <migraphx/gpu/miopen.hpp>
-#include <migraphx/gpu/hip.hpp>
-#include <migraphx/dfor.hpp>
-#include <migraphx/gpu/device/contiguous.hpp>
-#include <migraphx/gpu/device/add.hpp>
-#include <migraphx/iterator_for.hpp>
-#include <migraphx/gpu/rocblas.hpp>
-#include <migraphx/gpu/context.hpp>
-#include <utility>
 
 namespace migraphx {
 inline namespace MIGRAPHX_INLINE_NS {
 namespace gpu {
 
+struct context;
+
 struct miopen_abs
 {
     shared<activation_descriptor> ad;

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

 #ifndef MIGRAPHX_GUARD_RTGLIB_ACOS_HPP
 #define MIGRAPHX_GUARD_RTGLIB_ACOS_HPP
 
-#include <migraphx/gpu/lowering.hpp>
 #include <migraphx/gpu/oper.hpp>
-#include <migraphx/manage_ptr.hpp>
-#include <migraphx/instruction.hpp>
-#include <migraphx/operators.hpp>
-#include <migraphx/generate.hpp>
-#include <migraphx/shape_for_each.hpp>
-#include <migraphx/gpu/miopen.hpp>
-#include <migraphx/gpu/hip.hpp>
-#include <migraphx/dfor.hpp>
-#include <migraphx/gpu/device/contiguous.hpp>
 #include <migraphx/gpu/device/acos.hpp>
-#include <migraphx/iterator_for.hpp>
-#include <migraphx/gpu/rocblas.hpp>
-#include <migraphx/gpu/context.hpp>
-#include <migraphx/config.hpp>
-#include <utility>
 
 namespace migraphx {
 inline namespace MIGRAPHX_INLINE_NS {

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

 #ifndef MIGRAPHX_GUARD_RTGLIB_ADD_HPP
 #define MIGRAPHX_GUARD_RTGLIB_ADD_HPP
 
-#include <migraphx/gpu/lowering.hpp>
 #include <migraphx/gpu/oper.hpp>
-#include <migraphx/manage_ptr.hpp>
-#include <migraphx/instruction.hpp>
-#include <migraphx/operators.hpp>
-#include <migraphx/generate.hpp>
-#include <migraphx/shape_for_each.hpp>
-#include <migraphx/gpu/miopen.hpp>
-#include <migraphx/gpu/hip.hpp>
-#include <migraphx/dfor.hpp>
-#include <migraphx/gpu/device/contiguous.hpp>
 #include <migraphx/gpu/device/add.hpp>
-#include <migraphx/iterator_for.hpp>
-#include <migraphx/gpu/rocblas.hpp>
-#include <migraphx/gpu/context.hpp>
-#include <migraphx/config.hpp>
-#include <utility>
 
 namespace migraphx {
 inline namespace MIGRAPHX_INLINE_NS {

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

 #ifndef MIGRAPHX_GUARD_RTGLIB_ASIN_HPP
 #define MIGRAPHX_GUARD_RTGLIB_ASIN_HPP
 
-#include <migraphx/gpu/lowering.hpp>
 #include <migraphx/gpu/oper.hpp>
-#include <migraphx/manage_ptr.hpp>
-#include <migraphx/instruction.hpp>
-#include <migraphx/operators.hpp>
-#include <migraphx/generate.hpp>
-#include <migraphx/shape_for_each.hpp>
-#include <migraphx/gpu/miopen.hpp>
-#include <migraphx/gpu/hip.hpp>
-#include <migraphx/dfor.hpp>
-#include <migraphx/gpu/device/contiguous.hpp>
 #include <migraphx/gpu/device/asin.hpp>
-#include <migraphx/iterator_for.hpp>
-#include <migraphx/gpu/rocblas.hpp>
-#include <migraphx/gpu/context.hpp>
-#include <migraphx/config.hpp>
-#include <utility>
 
 namespace migraphx {
 inline namespace MIGRAPHX_INLINE_NS {

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

 #ifndef MIGRAPHX_GUARD_RTGLIB_ATAN_HPP
 #define MIGRAPHX_GUARD_RTGLIB_ATAN_HPP
 
-#include <migraphx/gpu/lowering.hpp>
 #include <migraphx/gpu/oper.hpp>
-#include <migraphx/manage_ptr.hpp>
-#include <migraphx/instruction.hpp>
-#include <migraphx/operators.hpp>
-#include <migraphx/generate.hpp>
-#include <migraphx/shape_for_each.hpp>
-#include <migraphx/gpu/miopen.hpp>
-#include <migraphx/gpu/hip.hpp>
-#include <migraphx/dfor.hpp>
-#include <migraphx/gpu/device/contiguous.hpp>
 #include <migraphx/gpu/device/atan.hpp>
-#include <migraphx/iterator_for.hpp>
-#include <migraphx/gpu/rocblas.hpp>
-#include <migraphx/gpu/context.hpp>
-#include <migraphx/config.hpp>
-#include <utility>
 
 namespace migraphx {
 inline namespace MIGRAPHX_INLINE_NS {

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

 #ifndef MIGRAPHX_GUARD_RTGLIB_BATCHNORM_HPP
 #define MIGRAPHX_GUARD_RTGLIB_BATCHNORM_HPP
 
-#include <migraphx/gpu/lowering.hpp>
-#include <migraphx/manage_ptr.hpp>
-#include <migraphx/instruction.hpp>
+#include <migraphx/shape.hpp>
 #include <migraphx/operators.hpp>
-#include <migraphx/generate.hpp>
-#include <migraphx/shape_for_each.hpp>
-#include <migraphx/gpu/miopen.hpp>
-#include <migraphx/gpu/hip.hpp>
-#include <migraphx/dfor.hpp>
-#include <migraphx/gpu/device/contiguous.hpp>
-#include <migraphx/gpu/device/add.hpp>
-#include <migraphx/iterator_for.hpp>
-#include <migraphx/gpu/rocblas.hpp>
-#include <migraphx/gpu/context.hpp>
-#include <migraphx/config.hpp>
-#include <utility>
 
 namespace migraphx {
 inline namespace MIGRAPHX_INLINE_NS {
 namespace gpu {
 
+struct context;
+
 struct miopen_batch_norm_inference
 {
     op::batch_norm_inference op;

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

 #ifndef MIGRAPHX_GUARD_RTGLIB_CONCAT_HPP
 #define MIGRAPHX_GUARD_RTGLIB_CONCAT_HPP
 
-#include <migraphx/gpu/lowering.hpp>
-#include <migraphx/manage_ptr.hpp>
-#include <migraphx/instruction.hpp>
+#include <migraphx/shape.hpp>
 #include <migraphx/operators.hpp>
-#include <migraphx/generate.hpp>
-#include <migraphx/shape_for_each.hpp>
-#include <migraphx/config.hpp>
-#include <migraphx/gpu/miopen.hpp>
-#include <migraphx/gpu/hip.hpp>
-#include <migraphx/dfor.hpp>
-#include <migraphx/gpu/device/concat.hpp>
-#include <migraphx/gpu/device/add.hpp>
-#include <migraphx/iterator_for.hpp>
-#include <migraphx/gpu/rocblas.hpp>
-#include <migraphx/gpu/context.hpp>
-#include <utility>
 
 namespace migraphx {
 inline namespace MIGRAPHX_INLINE_NS {
 namespace gpu {
 
+struct context;
+
 struct hip_concat
 {
     op::concat op;

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

 #ifndef MIGRAPHX_GUARD_RTGLIB_CONTIGUOUS_HPP
 #define MIGRAPHX_GUARD_RTGLIB_CONTIGUOUS_HPP
 
-#include <migraphx/gpu/lowering.hpp>
-#include <migraphx/manage_ptr.hpp>
-#include <migraphx/instruction.hpp>
+#include <migraphx/shape.hpp>
 #include <migraphx/operators.hpp>
-#include <migraphx/generate.hpp>
-#include <migraphx/shape_for_each.hpp>
-#include <migraphx/config.hpp>
-#include <migraphx/gpu/miopen.hpp>
-#include <migraphx/gpu/hip.hpp>
-#include <migraphx/dfor.hpp>
-#include <migraphx/gpu/device/contiguous.hpp>
-#include <migraphx/gpu/device/add.hpp>
-#include <migraphx/iterator_for.hpp>
-#include <migraphx/gpu/rocblas.hpp>
-#include <migraphx/gpu/context.hpp>
-#include <utility>
 
 namespace migraphx {
 inline namespace MIGRAPHX_INLINE_NS {
 namespace gpu {
 
+struct context;
+
 struct miopen_contiguous
 {
     op::contiguous op;

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

 #ifndef MIGRAPHX_GUARD_RTGLIB_CONVOLUTION_HPP
 #define MIGRAPHX_GUARD_RTGLIB_CONVOLUTION_HPP
 
-#include <migraphx/gpu/lowering.hpp>
-#include <migraphx/manage_ptr.hpp>
-#include <migraphx/instruction.hpp>
+#include <migraphx/shape.hpp>
 #include <migraphx/operators.hpp>
-#include <migraphx/generate.hpp>
-#include <migraphx/shape_for_each.hpp>
-#include <migraphx/config.hpp>
 #include <migraphx/gpu/miopen.hpp>
-#include <migraphx/gpu/hip.hpp>
-#include <migraphx/dfor.hpp>
-#include <migraphx/gpu/device/contiguous.hpp>
-#include <migraphx/gpu/device/add.hpp>
-#include <migraphx/iterator_for.hpp>
-#include <migraphx/gpu/rocblas.hpp>
-#include <migraphx/gpu/context.hpp>
-#include <utility>
 
 namespace migraphx {
 inline namespace MIGRAPHX_INLINE_NS {
 namespace gpu {
 
+struct context;
+
 struct miopen_convolution
 {
     op::convolution op;

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

 #ifndef MIGRAPHX_GUARD_RTGLIB_COS_HPP
 #define MIGRAPHX_GUARD_RTGLIB_COS_HPP
 
-#include <migraphx/gpu/lowering.hpp>
 #include <migraphx/gpu/oper.hpp>
-#include <migraphx/manage_ptr.hpp>
-#include <migraphx/instruction.hpp>
-#include <migraphx/operators.hpp>
-#include <migraphx/generate.hpp>
-#include <migraphx/shape_for_each.hpp>
-#include <migraphx/gpu/miopen.hpp>
-#include <migraphx/gpu/hip.hpp>
-#include <migraphx/dfor.hpp>
-#include <migraphx/gpu/device/contiguous.hpp>
 #include <migraphx/gpu/device/cos.hpp>
-#include <migraphx/iterator_for.hpp>
-#include <migraphx/gpu/rocblas.hpp>
-#include <migraphx/gpu/context.hpp>
-#include <migraphx/config.hpp>
-#include <utility>
 
 namespace migraphx {
 inline namespace MIGRAPHX_INLINE_NS {

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

 #ifndef MIGRAPHX_GUARD_RTGLIB_COSH_HPP
 #define MIGRAPHX_GUARD_RTGLIB_COSH_HPP
 
-#include <migraphx/gpu/lowering.hpp>
 #include <migraphx/gpu/oper.hpp>
-#include <migraphx/manage_ptr.hpp>
-#include <migraphx/instruction.hpp>
-#include <migraphx/operators.hpp>
-#include <migraphx/generate.hpp>
-#include <migraphx/shape_for_each.hpp>
-#include <migraphx/gpu/miopen.hpp>
-#include <migraphx/gpu/hip.hpp>
-#include <migraphx/dfor.hpp>
-#include <migraphx/gpu/device/contiguous.hpp>
 #include <migraphx/gpu/device/cosh.hpp>
-#include <migraphx/iterator_for.hpp>
-#include <migraphx/gpu/rocblas.hpp>
-#include <migraphx/gpu/context.hpp>
-#include <migraphx/config.hpp>
-#include <utility>
 
 namespace migraphx {
 inline namespace MIGRAPHX_INLINE_NS {

src/targets/gpu/include/migraphx/gpu/device/logsoftmax.hpp

+#ifndef MIGRAPHX_GUARD_RTGLIB_DEVICE_LOGSOFTMAX_HPP
+#define MIGRAPHX_GUARD_RTGLIB_DEVICE_LOGSOFTMAX_HPP
+
+#include <migraphx/argument.hpp>
+#include <migraphx/config.hpp>
+#include <hip/hip_runtime_api.h>
+
+namespace migraphx {
+inline namespace MIGRAPHX_INLINE_NS {
+namespace gpu {
+namespace device {
+
+argument logsoftmax(hipStream_t stream,
+                    const migraphx::shape& output_shape,
+                    std::vector<migraphx::argument> args,
+                    int axis);
+
+} // namespace device
+} // namespace gpu
+} // namespace MIGRAPHX_INLINE_NS
+} // namespace migraphx
+
+#endif

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

 #ifndef MIGRAPHX_GUARD_RTGLIB_ELU_HPP
 #define MIGRAPHX_GUARD_RTGLIB_ELU_HPP
 
-#include <migraphx/gpu/lowering.hpp>
-#include <migraphx/manage_ptr.hpp>
-#include <migraphx/instruction.hpp>
-#include <migraphx/operators.hpp>
-#include <migraphx/generate.hpp>
-#include <migraphx/shape_for_each.hpp>
-#include <migraphx/config.hpp>
+#include <migraphx/shape.hpp>
 #include <migraphx/gpu/miopen.hpp>
-#include <migraphx/gpu/hip.hpp>
-#include <migraphx/dfor.hpp>
-#include <migraphx/gpu/device/contiguous.hpp>
-#include <migraphx/gpu/device/add.hpp>
-#include <migraphx/iterator_for.hpp>
-#include <migraphx/gpu/rocblas.hpp>
-#include <migraphx/gpu/context.hpp>
-#include <utility>
 
 namespace migraphx {
 inline namespace MIGRAPHX_INLINE_NS {
 namespace gpu {
 
+struct context;
+
 struct miopen_elu
 {
     shared<activation_descriptor> ad;

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

 #ifndef MIGRAPHX_GUARD_RTGLIB_EXP_HPP
 #define MIGRAPHX_GUARD_RTGLIB_EXP_HPP
 
-#include <migraphx/gpu/lowering.hpp>
 #include <migraphx/gpu/oper.hpp>
-#include <migraphx/manage_ptr.hpp>
-#include <migraphx/instruction.hpp>
-#include <migraphx/operators.hpp>
-#include <migraphx/generate.hpp>
-#include <migraphx/shape_for_each.hpp>
-#include <migraphx/gpu/miopen.hpp>
-#include <migraphx/gpu/hip.hpp>
-#include <migraphx/dfor.hpp>
-#include <migraphx/gpu/device/contiguous.hpp>
 #include <migraphx/gpu/device/exp.hpp>
-#include <migraphx/iterator_for.hpp>
-#include <migraphx/gpu/rocblas.hpp>
-#include <migraphx/gpu/context.hpp>
-#include <migraphx/config.hpp>
-#include <utility>
 
 namespace migraphx {
 inline namespace MIGRAPHX_INLINE_NS {

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

 #ifndef MIGRAPHX_GUARD_RTGLIB_GATHER_HPP
 #define MIGRAPHX_GUARD_RTGLIB_GATHER_HPP
 
-#include <migraphx/gpu/lowering.hpp>
-#include <migraphx/manage_ptr.hpp>
-#include <migraphx/instruction.hpp>
-#include <migraphx/operators.hpp>
-#include <migraphx/generate.hpp>
-#include <migraphx/shape_for_each.hpp>
-#include <migraphx/config.hpp>
+#include <migraphx/shape.hpp>
 #include <migraphx/gpu/miopen.hpp>
-#include <migraphx/gpu/hip.hpp>
-#include <migraphx/dfor.hpp>
-#include <migraphx/gpu/device/contiguous.hpp>
-#include <migraphx/gpu/device/gather.hpp>
-#include <migraphx/iterator_for.hpp>
-#include <migraphx/gpu/rocblas.hpp>
-#include <migraphx/gpu/context.hpp>
-#include <utility>
 
 namespace migraphx {
 inline namespace MIGRAPHX_INLINE_NS {
 namespace gpu {
 
+struct context;
+
 struct hip_gather
 {
     op::gather op;

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

 #ifndef MIGRAPHX_GUARD_RTGLIB_GEMM_HPP
 #define MIGRAPHX_GUARD_RTGLIB_GEMM_HPP
 
-#include <migraphx/gpu/lowering.hpp>
-#include <migraphx/manage_ptr.hpp>
-#include <migraphx/instruction.hpp>
+#include <migraphx/shape.hpp>
 #include <migraphx/operators.hpp>
-#include <migraphx/generate.hpp>
-#include <migraphx/shape_for_each.hpp>
-#include <migraphx/gpu/miopen.hpp>
-#include <migraphx/gpu/hip.hpp>
-#include <migraphx/dfor.hpp>
-#include <migraphx/config.hpp>
-#include <migraphx/gpu/device/contiguous.hpp>
-#include <migraphx/gpu/device/add.hpp>
-#include <migraphx/iterator_for.hpp>
-#include <migraphx/gpu/rocblas.hpp>
-#include <migraphx/gpu/context.hpp>
-#include <utility>
 
 namespace migraphx {
 inline namespace MIGRAPHX_INLINE_NS {
 namespace gpu {
 
+struct context;
+
 struct miopen_gemm
 {
     op::dot op;