Merge from develop

src/targets/gpu/contiguous.cpp

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

src/targets/gpu/convolution.cpp

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

src/targets/gpu/device/gather.cpp

@@ -16,20 +16,24 @@ argument gather(hipStream_t stream,
                 std::vector<migraphx::argument> args,
                 int axis)
 {
-    int axis_index = (axis < 0) ? (axis + output_shape.lens().size()) : axis;
+    int axis_index = (axis < 0) ? (axis + args[0].get_shape().lens().size()) : axis;
     visit_all(args.back(), args[0])([&](auto output, auto input) {
         std::size_t nelements = output_shape.elements();
         args[1].visit([&](auto indices) {
-            visit_tensor_size(output_shape.lens().size(), [&](auto ndim) {
-                const auto* indices_ptr = device_cast(indices.data());
-                auto* outptr            = device_cast(output.data());
-                const auto* inptr       = device_cast(input.data());
-                hip_tensor_descriptor<ndim> desc_input(input.get_shape());
-                hip_tensor_descriptor<ndim> desc_output(output.get_shape());
-                gs_launch(stream, nelements)([=](auto i) {
-                    auto lens        = desc_output.multi(i);
-                    lens[axis_index] = indices_ptr[lens[axis_index]];
-                    outptr[i]        = inptr[desc_input.linear(lens)];
+            const auto* indices_ptr = device_cast(indices.data());
+            auto* out_ptr           = device_cast(output.data());
+            const auto* in_ptr      = device_cast(input.data());
+            auto& input_shape       = args[0].get_shape();
+            auto lens               = input_shape.lens();
+            lens[axis_index]        = args[1].get_shape().elements();
+            migraphx::shape out_comp_shape{output_shape.type(), lens};
+            visit_tensor_size(out_comp_shape.lens().size(), [&](auto n_out_dim) {
+                hip_tensor_descriptor<n_out_dim> desc_input(input_shape);
+                hip_tensor_descriptor<n_out_dim> desc_output(out_comp_shape);
+                gs_launch(stream, nelements)([=](auto ii) {
+                    auto in_idx        = desc_output.multi(ii);
+                    in_idx[axis_index] = indices_ptr[in_idx[axis_index]];
+                    out_ptr[ii]        = in_ptr[desc_input.linear(in_idx)];
                 });
             });
         });

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 {

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/context.hpp

@@ -12,6 +12,8 @@ namespace migraphx {
 inline namespace MIGRAPHX_INLINE_NS {
 namespace gpu {
 
+MIGRAPHX_DECLARE_ENV_VAR(MIGRAPHX_ENABLE_NULL_STREAM)
+
 struct hip_device
 {
     using hip_event_ptr = MIGRAPHX_MANAGE_PTR(hipEvent_t, hipEventDestroy);
@@ -42,7 +44,7 @@ struct hip_device
 
         hipStream_t get()
         {
-            if(enabled(MIGRAPHX_DISABLE_NULL_STREAM{}))
+            if(not enabled(MIGRAPHX_ENABLE_NULL_STREAM{}))
             {
                 setup();
                 if(s == nullptr)
@@ -55,7 +57,7 @@ struct hip_device
 
         auto create_miopen_handle()
         {
-            if(enabled(MIGRAPHX_DISABLE_NULL_STREAM{}))
+            if(not enabled(MIGRAPHX_ENABLE_NULL_STREAM{}))
                 return make_obj<miopen_handle>(&miopenCreateWithStream, get());
             else
                 return make_obj<miopen_handle>(&miopenCreate);
@@ -97,10 +99,8 @@ struct hip_device
 
     void add_streams()
     {
-        int num_of_streams = 1;
+        int num_of_streams = stream_info().num_of_streams();
         assert(streams.empty());
-        if(enabled(MIGRAPHX_DISABLE_NULL_STREAM{}))
-            num_of_streams = stream_info().num_of_streams();
         for(int i = 0; i < num_of_streams; ++i)
             streams.emplace_back(device_id);
     }
@@ -126,15 +126,8 @@ struct hip_device
 
     void stream_sync()
     {
-        if(enabled(MIGRAPHX_DISABLE_NULL_STREAM{}))
-        {
-            int num_of_streams = streams.size();
-            if(num_of_streams > 0)
-            {
-                for(int i = 0; i < num_of_streams; i++)
-                    hipStreamSynchronize(streams.at(i).get());
-            }
-        }
+        for(auto&& stream:streams)
+            hipStreamSynchronize(stream.get());
     }
 
     private:

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/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;