manual merge

src/targets/gpu/device/include/migraph/gpu/device/tensor.hpp → src/targets/gpu/device/include/migraphx/gpu/device/tensor.hpp

-#ifndef MIGRAPH_GUARD_RTGLIB_DEAVICE_TENSOR_HPP
-#define MIGRAPH_GUARD_RTGLIB_DEAVICE_TENSOR_HPP
+#ifndef MIGRAPHX_GUARD_RTGLIB_DEAVICE_TENSOR_HPP
+#define MIGRAPHX_GUARD_RTGLIB_DEAVICE_TENSOR_HPP
 
 #include <hip/hip_runtime.h>
-#include <migraph/functional.hpp>
-#include <migraph/config.hpp>
+#include <migraphx/functional.hpp>
+#include <migraphx/config.hpp>
 
-namespace migraph {
-inline namespace MIGRAPH_INLINE_NS {
+namespace migraphx {
+inline namespace MIGRAPHX_INLINE_NS {
 namespace gpu {
 namespace device {
 
@@ -87,7 +87,7 @@ struct hip_tensor_descriptor
 
 } // namespace device
 } // namespace gpu
-} // namespace MIGRAPH_INLINE_NS
-} // namespace migraph
+} // namespace MIGRAPHX_INLINE_NS
+} // namespace migraphx
 
 #endif

src/targets/gpu/device/include/migraph/gpu/device/types.hpp → src/targets/gpu/device/include/migraphx/gpu/device/types.hpp

@@ -5,14 +5,14 @@
     file LICENSE_1_0.txt or copy at http://www.boost.org/LICENSE_1_0.txt)
 ==============================================================================*/
 
-#ifndef MIGRAPH_GUARD_RTGLIB_GPU_DEVICE_TYPES_HPP
-#define MIGRAPH_GUARD_RTGLIB_GPU_DEVICE_TYPES_HPP
+#ifndef MIGRAPHX_GUARD_RTGLIB_GPU_DEVICE_TYPES_HPP
+#define MIGRAPHX_GUARD_RTGLIB_GPU_DEVICE_TYPES_HPP
 
-#include <migraph/half.hpp>
-#include <migraph/config.hpp>
+#include <migraphx/half.hpp>
+#include <migraphx/config.hpp>
 
-namespace migraph {
-inline namespace MIGRAPH_INLINE_NS {
+namespace migraphx {
+inline namespace MIGRAPHX_INLINE_NS {
 namespace gpu {
 namespace device {
 
@@ -75,9 +75,18 @@ device_type<T>* device_cast(T* x)
     return reinterpret_cast<device_type<T>*>(x);
 }
 
+template <class T>
+T to_hip_type(T x)
+{
+    return x;
+}
+
+// Hip doens't support __fp16
+inline float to_hip_type(gpu_half x) { return x; }
+
 } // namespace device
 } // namespace gpu
-} // namespace MIGRAPH_INLINE_NS
-} // namespace migraph
+} // namespace MIGRAPHX_INLINE_NS
+} // namespace migraphx
 
 #endif

src/targets/gpu/device/log.cpp

+#include <migraphx/gpu/device/log.hpp>
+#include <migraphx/gpu/device/nary.hpp>
+#include <migraphx/gpu/device/types.hpp>
+
+namespace migraphx {
+inline namespace MIGRAPHX_INLINE_NS {
+namespace gpu {
+namespace device {
+
+void log(hipStream_t stream, const argument& result, const argument& arg)
+{
+    nary(stream, result, arg)([](auto x) { return ::log(to_hip_type(x)); });
+}
+
+} // namespace device
+} // namespace gpu
+} // namespace MIGRAPHX_INLINE_NS
+} // namespace migraphx

src/targets/gpu/device/max.cpp

+#include <migraphx/gpu/device/max.hpp>
+#include <migraphx/gpu/device/nary.hpp>
+#include <migraphx/gpu/device/types.hpp>
+
+namespace migraphx {
+inline namespace MIGRAPHX_INLINE_NS {
+namespace gpu {
+namespace device {
+
+void max(hipStream_t stream, const argument& result, const argument& arg1, const argument& arg2)
+{
+    nary(stream, result, arg1, arg2)(
+        [](auto x, auto y) { return std::max(to_hip_type(x), to_hip_type(y)); });
+}
+
+} // namespace device
+} // namespace gpu
+} // namespace MIGRAPHX_INLINE_NS
+} // namespace migraphx

src/targets/gpu/device/min.cpp

+#include <migraphx/gpu/device/min.hpp>
+#include <migraphx/gpu/device/nary.hpp>
+#include <migraphx/gpu/device/types.hpp>
+
+namespace migraphx {
+inline namespace MIGRAPHX_INLINE_NS {
+namespace gpu {
+namespace device {
+
+void min(hipStream_t stream, const argument& result, const argument& arg1, const argument& arg2)
+{
+    nary(stream, result, arg1, arg2)(
+        [](auto x, auto y) { return std::min(to_hip_type(x), to_hip_type(y)); });
+}
+
+} // namespace device
+} // namespace gpu
+} // namespace MIGRAPHX_INLINE_NS
+} // namespace migraphx

src/targets/gpu/device/mul.cpp

-#include <migraph/gpu/device/mul.hpp>
-#include <migraph/gpu/device/nary.hpp>
+#include <migraphx/gpu/device/mul.hpp>
+#include <migraphx/gpu/device/nary.hpp>
 
-namespace migraph {
-inline namespace MIGRAPH_INLINE_NS {
+namespace migraphx {
+inline namespace MIGRAPHX_INLINE_NS {
 namespace gpu {
 namespace device {
 
@@ -22,5 +22,5 @@ void mul(hipStream_t stream,
 
 } // namespace device
 } // namespace gpu
-} // namespace MIGRAPH_INLINE_NS
-} // namespace migraph
+} // namespace MIGRAPHX_INLINE_NS
+} // namespace migraphx

src/targets/gpu/device/pad.cpp

+#include <migraphx/shape.hpp>
+#include <migraphx/argument.hpp>
+#include <migraphx/gpu/device/nary.hpp>
+#include <migraphx/gpu/device/pad.hpp>
+#include <migraphx/gpu/device/tensor.hpp>
+#include <migraphx/gpu/device/launch.hpp>
+
+namespace migraphx {
+inline namespace MIGRAPHX_INLINE_NS {
+namespace gpu {
+namespace device {
+
+argument
+pad(hipStream_t stream, argument result, argument arg1, float value, std::vector<std::int64_t> pads)
+{
+    std::size_t nelements = arg1.get_shape().elements();
+
+    nary(stream, result)([=] { return value; });
+    visit_all(result, arg1)([&](auto output, auto input) {
+        visit_tensor_size(result.get_shape().lens().size(), [&](auto ndim) {
+            std::size_t offsets[ndim];
+            std::copy(pads.begin(), pads.begin() + ndim, offsets);
+            auto* outptr      = output.data();
+            const auto* inptr = 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 idx = desc_input.multi(i);
+                for(std::size_t j = 0; j < ndim; j++)
+                {
+                    idx[j] += offsets[j];
+                }
+                outptr[desc_output.linear(idx)] = inptr[i];
+            });
+        });
+    });
+    return result;
+}
+
+} // namespace device
+} // namespace gpu
+} // namespace MIGRAPHX_INLINE_NS
+} // namespace migraphx

src/targets/gpu/device/sin.cpp

+#include <migraphx/gpu/device/sin.hpp>
+#include <migraphx/gpu/device/nary.hpp>
+#include <migraphx/gpu/device/types.hpp>
+
+namespace migraphx {
+inline namespace MIGRAPHX_INLINE_NS {
+namespace gpu {
+namespace device {
+
+void sin(hipStream_t stream, const argument& result, const argument& arg)
+{
+    nary(stream, result, arg)([](auto x) { return ::sin(to_hip_type(x)); });
+}
+
+} // namespace device
+} // namespace gpu
+} // namespace MIGRAPHX_INLINE_NS
+} // namespace migraphx

src/targets/gpu/device/sinh.cpp

+#include <migraphx/gpu/device/sinh.hpp>
+#include <migraphx/gpu/device/nary.hpp>
+#include <migraphx/gpu/device/types.hpp>
+
+namespace migraphx {
+inline namespace MIGRAPHX_INLINE_NS {
+namespace gpu {
+namespace device {
+
+void sinh(hipStream_t stream, const argument& result, const argument& arg)
+{
+    nary(stream, result, arg)([](auto x) { return ::sinh(to_hip_type(x)); });
+}
+
+} // namespace device
+} // namespace gpu
+} // namespace MIGRAPHX_INLINE_NS
+} // namespace migraphx

src/targets/gpu/device/tan.cpp

+#include <migraphx/gpu/device/tan.hpp>
+#include <migraphx/gpu/device/nary.hpp>
+#include <migraphx/gpu/device/types.hpp>
+
+namespace migraphx {
+inline namespace MIGRAPHX_INLINE_NS {
+namespace gpu {
+namespace device {
+
+void tan(hipStream_t stream, const argument& result, const argument& arg)
+{
+    nary(stream, result, arg)([](auto x) { return ::tan(to_hip_type(x)); });
+}
+
+} // namespace device
+} // namespace gpu
+} // namespace MIGRAPHX_INLINE_NS
+} // namespace migraphx

src/targets/gpu/eliminate_workspace.cpp

-#include <migraph/gpu/eliminate_workspace.hpp>
-#include <migraph/gpu/hip.hpp>
-#include <migraph/program.hpp>
-#include <migraph/instruction.hpp>
-#include <migraph/operators.hpp>
-#include <migraph/iterator_for.hpp>
-#include <migraph/ranges.hpp>
-#include <migraph/stringutils.hpp>
-#include <migraph/pass_config.hpp>
+#include <migraphx/gpu/eliminate_workspace.hpp>
+#include <migraphx/gpu/hip.hpp>
+#include <migraphx/program.hpp>
+#include <migraphx/instruction.hpp>
+#include <migraphx/operators.hpp>
+#include <migraphx/iterator_for.hpp>
+#include <migraphx/ranges.hpp>
+#include <migraphx/stringutils.hpp>
+#include <migraphx/pass_config.hpp>
 
-namespace migraph {
-inline namespace MIGRAPH_INLINE_NS {
+namespace migraphx {
+inline namespace MIGRAPHX_INLINE_NS {
 namespace gpu {
 
 void eliminate_workspace::apply(program& p) const
 {
-    if(!enabled(MIGRAPH_DISABLE_MEMORY_COLORING{}))
-        return;
-
     std::size_t n = 0;
     std::vector<instruction_ref> allocs;
     for(auto ins : iterator_for(p))
@@ -32,14 +29,17 @@ void eliminate_workspace::apply(program& p) const
             allocs.push_back(ins);
         }
     }
-    auto ws = p.add_parameter("workspace", shape{shape::int8_type, {n}});
-    for(auto&& a : allocs)
+    if(n > 0)
     {
-        p.replace_instruction(a, ws);
-        p.remove_instruction(a);
+        auto ws = p.add_parameter("workspace", shape{shape::int8_type, {n}});
+        for(auto&& a : allocs)
+        {
+            p.replace_instruction(a, ws);
+            p.remove_instruction(a);
+        }
     }
 }
 
 } // namespace gpu
-} // namespace MIGRAPH_INLINE_NS
-} // namespace migraph
+} // namespace MIGRAPHX_INLINE_NS
+} // namespace migraphx

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>
+
+namespace migraphx {
+inline namespace MIGRAPHX_INLINE_NS {
+namespace gpu {
+
+shape miopen_elu::compute_shape(const std::vector<shape>& inputs) const
+{
+    check_shapes{inputs, *this}.has(2).not_broadcasted();
+    return inputs.at(1);
+}
+
+argument miopen_elu::compute(context& ctx,
+                             const shape& output_shape,
+                             const std::vector<argument>& args) const
+{
+    float alpha = 1;
+    float beta  = 0;
+    auto x_desc = make_tensor(args[0].get_shape());
+    auto y_desc = make_tensor(output_shape);
+    miopenActivationForward(ctx.get_stream().get_miopen(),
+                            ad.get(),
+                            &alpha,
+                            x_desc.get(),
+                            args[0].implicit(),
+                            &beta,
+                            y_desc.get(),
+                            args[1].implicit());
+
+    return args[1];
+}
+
+} // namespace gpu
+} // namespace MIGRAPHX_INLINE_NS
+} // namespace migraphx

src/targets/gpu/fuse_ops.cpp

-#include <migraph/gpu/fuse_ops.hpp>
-#include <migraph/matcher.hpp>
-#include <migraph/gpu/miopen.hpp>
-#include <migraph/gpu/convolution.hpp>
-#include <migraph/gpu/device/add_relu.hpp>
-#include <migraph/instruction.hpp>
-
-namespace migraph {
-inline namespace MIGRAPH_INLINE_NS {
+#include <migraphx/gpu/fuse_ops.hpp>
+#include <migraphx/matcher.hpp>
+#include <migraphx/gpu/miopen.hpp>
+#include <migraphx/gpu/convolution.hpp>
+#include <migraphx/gpu/device/add_relu.hpp>
+#include <migraphx/instruction.hpp>
+
+namespace migraphx {
+inline namespace MIGRAPHX_INLINE_NS {
 namespace gpu {
 
 struct fusion
@@ -38,7 +38,7 @@ struct fusion
         op_t result;
         auto status = miopenFusionPlanGetOp(fp.get(), i, &result);
         if(status != miopenStatusSuccess)
-            MIGRAPH_THROW("Failed retrieving operator at " + std::to_string(i));
+            MIGRAPHX_THROW("Failed retrieving operator at " + std::to_string(i));
         return result;
     }
 
@@ -51,7 +51,7 @@ struct fusion
         auto t      = keep_alive(make_tensor(b));
         auto status = miopenCreateOpBiasForward(fp.get(), &result, t.get());
         if(status != miopenStatusSuccess)
-            MIGRAPH_THROW("Creating operator failed");
+            MIGRAPHX_THROW("Creating operator failed");
         return result;
     }
 
@@ -60,7 +60,7 @@ struct fusion
         op_t result;
         auto status = miopenCreateOpActivationForward(fp.get(), &result, miopenActivationRELU);
         if(status != miopenStatusSuccess)
-            MIGRAPH_THROW("Creating operator failed");
+            MIGRAPHX_THROW("Creating operator failed");
         return result;
     }
 
@@ -71,7 +71,7 @@ struct fusion
         auto t      = keep_alive(make_tensor(weights));
         auto status = miopenCreateOpConvForward(fp.get(), &result, cd.get(), t.get());
         if(status != miopenStatusSuccess)
-            MIGRAPH_THROW("Creating operator failed");
+            MIGRAPHX_THROW("Creating operator failed");
         return result;
     }
 
@@ -91,7 +91,7 @@ struct fusion
     {
         auto status = miopenCompileFusionPlan(ctx.get_stream().get_miopen(), fp.get());
         if(status != miopenStatusSuccess)
-            MIGRAPH_THROW("Compiling fusion plan failed");
+            MIGRAPHX_THROW("Compiling fusion plan failed");
     }
 
     argument execute(context& ctx,
@@ -109,12 +109,12 @@ struct fusion
                                               y.implicit(),
                                               fargs.get());
         if(status != miopenStatusSuccess)
-            MIGRAPH_THROW("Failed to execute fusion plan");
+            MIGRAPHX_THROW("Failed to execute fusion plan");
         return y;
     }
 };
 
-MIGRAPH_PRED_MATCHER(bias_shape, instruction_ref ins)
+MIGRAPHX_PRED_MATCHER(bias_shape, instruction_ref ins)
 {
     auto&& s = ins->get_shape();
     return s.broadcasted() and s.strides().size() == 4 and s.strides()[0] == 0 and
@@ -128,7 +128,7 @@ std::array<T, sizeof...(Ts) + 1> make_array(T x, Ts... xs)
     return {std::move(x), std::move(static_cast<T>(xs))...};
 }
 
-MIGRAPH_PRED_MATCHER(fusable_conv, instruction_ref ins)
+MIGRAPHX_PRED_MATCHER(fusable_conv, instruction_ref ins)
 {
     if(ins->name() != "gpu::convolution")
         return false;
@@ -137,6 +137,8 @@ MIGRAPH_PRED_MATCHER(fusable_conv, instruction_ref ins)
     auto wei = ins->inputs().at(1)->get_shape();
     assert(wei.lens().size() == 4);
     auto conv = any_cast<miopen_convolution>(ins->get_operator());
+    if(conv.op.group > 1)
+        return false;
     if(wei.lens()[1] > 512 and conv.algo != miopenConvolutionFwdAlgoWinograd)
         return false;
     auto op = conv.op;
@@ -265,17 +267,15 @@ struct miopen_conv_bias
     argument compute(context& ctx, const shape&, const std::vector<argument>& args) const
     {
         auto fargs  = make_fused_args();
-        float alpha = 1, beta = 0;
+        float alpha = 1;
+        float beta  = 0;
         miopenSetOpArgsConvForward(fargs.get(), conv, &alpha, &beta, args[1].implicit());
         miopenSetOpArgsBiasForward(fargs.get(), bias, &alpha, &beta, args[3].implicit());
         return f.execute(ctx, fargs, args[0], args[4]);
     }
 
-    shape compile(context& ctx)
-    {
-        f.compile(ctx);
-        return f.get_workspace(ctx);
-    }
+    void finalize(context& ctx, const shape&, const std::vector<shape>&) { f.compile(ctx); }
+    shape get_workspace(context& ctx) { return f.get_workspace(ctx); }
     int output_alias(const std::vector<shape>& shapes) const { return shapes.size() - 1; }
 };
 
@@ -308,18 +308,15 @@ struct miopen_conv_bias_relu
     argument compute(context& ctx, const shape&, const std::vector<argument>& args) const
     {
         auto fargs  = make_fused_args();
-        float alpha = 1, beta = 0;
+        float alpha = 1;
+        float beta  = 0;
         miopenSetOpArgsConvForward(fargs.get(), conv, &alpha, &beta, args[1].implicit());
         miopenSetOpArgsBiasForward(fargs.get(), bias, &alpha, &beta, args[3].implicit());
         miopenSetOpArgsActivForward(fargs.get(), relu, &alpha, &beta, 0, 0, 0);
         return f.execute(ctx, fargs, args[0], args[4]);
     }
-
-    shape compile(context& ctx)
-    {
-        f.compile(ctx);
-        return f.get_workspace(ctx);
-    }
+    void finalize(context& ctx, const shape&, const std::vector<shape>&) { f.compile(ctx); }
+    shape get_workspace(context& ctx) { return f.get_workspace(ctx); }
     int output_alias(const std::vector<shape>& shapes) const { return shapes.size() - 1; }
 };
 
@@ -346,8 +343,8 @@ void apply_conv_bias(context& ctx, program& p, match::matcher_result r)
 
     Op cb{conv_op, input_ins->get_shape(), weights_ins->get_shape(), bias_ins->get_shape()};
     // TODO: Insert ws allocation
-    auto ws = cb.compile(ctx);
-
+    auto ws = cb.get_workspace(ctx);
+    (void)ws;
     p.replace_instruction(ins, cb, input_ins, weights_ins, old_ws_ins, bias_ins, alloc_ins);
 }
 
@@ -389,5 +386,5 @@ void fuse_ops::apply(program& p) const
 }
 
 } // namespace gpu
-} // namespace MIGRAPH_INLINE_NS
-} // namespace migraph
+} // namespace MIGRAPHX_INLINE_NS
+} // 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>
+
+namespace migraphx {
+inline namespace MIGRAPHX_INLINE_NS {
+namespace gpu {
+
+shape hip_gather::compute_shape(std::vector<shape> inputs) const
+{
+    inputs.pop_back();
+    return op.compute_shape(inputs);
+}
+
+argument hip_gather::compute(context& ctx,
+                             const shape& output_shape,
+                             const std::vector<argument>& args) const
+{
+    return device::gather(ctx.get_stream().get(), output_shape, args, op.axis);
+}
+
+} // namespace gpu
+} // namespace MIGRAPHX_INLINE_NS
+} // namespace migraphx

src/targets/gpu/gemm.cpp

-#include <migraph/gpu/gemm.hpp>
-#include <migraph/operators.hpp>
-#include <migraph/manage_ptr.hpp>
-#include <migraph/gpu/miopen.hpp>
+#include <migraphx/gpu/gemm.hpp>
+#include <migraphx/operators.hpp>
+#include <migraphx/manage_ptr.hpp>
+#include <migraphx/gpu/miopen.hpp>
 #include <utility>
 
-namespace migraph {
-inline namespace MIGRAPH_INLINE_NS {
+namespace migraphx {
+inline namespace MIGRAPHX_INLINE_NS {
 namespace gpu {
 
 template <class... Ts>
@@ -29,7 +29,7 @@ void generic_rocblas_gemm(shape::as<half>, Ts&&... xs)
 template <class T, class... Ts>
 void generic_rocblas_gemm(shape::as<T>, Ts&&...)
 {
-    MIGRAPH_THROW("Type unsupported by rocblas");
+    MIGRAPHX_THROW("Type unsupported by rocblas");
 }
 
 template <class T>
@@ -111,5 +111,5 @@ argument miopen_gemm::compute(context& ctx,
 }
 
 } // namespace gpu
-} // namespace MIGRAPH_INLINE_NS
-} // namespace migraph
+} // namespace MIGRAPHX_INLINE_NS
+} // namespace migraphx

src/targets/gpu/hip.cpp

 
-#include <migraph/gpu/hip.hpp>
+#include <migraphx/gpu/hip.hpp>
 
-#include <migraph/manage_ptr.hpp>
+#include <migraphx/manage_ptr.hpp>
 #include <miopen/miopen.h>
 
 #include <vector>
 
-namespace migraph {
-inline namespace MIGRAPH_INLINE_NS {
+namespace migraphx {
+inline namespace MIGRAPHX_INLINE_NS {
 namespace gpu {
 
-using hip_ptr = MIGRAPH_MANAGE_PTR(void, hipFree);
+using hip_ptr = MIGRAPHX_MANAGE_PTR(void, hipFree);
 
 std::string hip_error(int error) { return hipGetErrorString(static_cast<hipError_t>(error)); }
 
 std::size_t get_available_gpu_memory()
 {
-    size_t free, total;
+    size_t free;
+    size_t total;
     auto status = hipMemGetInfo(&free, &total);
     if(status != hipSuccess)
-        MIGRAPH_THROW("Failed getting available memory: " + hip_error(status));
+        MIGRAPHX_THROW("Failed getting available memory: " + hip_error(status));
     return free;
 }
 
 hip_ptr allocate_gpu(std::size_t sz, bool host = false)
 {
     if(sz > get_available_gpu_memory())
-        MIGRAPH_THROW("Memory not available to allocate buffer: " + std::to_string(sz));
+        MIGRAPHX_THROW("Memory not available to allocate buffer: " + std::to_string(sz));
     void* result;
     auto status = host ? hipHostMalloc(&result, sz) : hipMalloc(&result, sz);
     if(status != hipSuccess)
     {
         if(host)
-            MIGRAPH_THROW("Gpu allocation failed: " + hip_error(status));
+            MIGRAPHX_THROW("Gpu allocation failed: " + hip_error(status));
         else
             allocate_gpu(sz, true);
     }
@@ -45,7 +46,7 @@ std::vector<T> read_from_gpu(const void* x, std::size_t sz)
     std::vector<T> result(sz);
     auto status = hipMemcpy(result.data(), x, sz * sizeof(T), hipMemcpyDeviceToHost);
     if(status != hipSuccess)
-        MIGRAPH_THROW("Copy from gpu failed: " + hip_error(status)); // NOLINT
+        MIGRAPHX_THROW("Copy from gpu failed: " + hip_error(status)); // NOLINT
     return result;
 }
 
@@ -54,7 +55,7 @@ hip_ptr write_to_gpu(const void* x, std::size_t sz, bool host = false)
     auto result = allocate_gpu(sz, host);
     auto status = hipMemcpy(result.get(), x, sz, hipMemcpyHostToDevice);
     if(status != hipSuccess)
-        MIGRAPH_THROW("Copy to gpu failed: " + hip_error(status));
+        MIGRAPHX_THROW("Copy to gpu failed: " + hip_error(status));
     return result;
 }
 
@@ -72,13 +73,13 @@ argument allocate_gpu(const shape& s, bool host)
     return {s, [p]() mutable { return reinterpret_cast<char*>(p.get()); }};
 }
 
-argument to_gpu(argument arg, bool host)
+argument to_gpu(const argument& arg, bool host)
 {
     auto p = share(write_to_gpu(arg.data(), arg.get_shape().bytes(), host));
     return {arg.get_shape(), [p]() mutable { return reinterpret_cast<char*>(p.get()); }};
 }
 
-argument from_gpu(argument arg)
+argument from_gpu(const argument& arg)
 {
     argument result;
     arg.visit([&](auto x) {
@@ -93,22 +94,22 @@ void set_device(std::size_t id)
 {
     auto status = hipSetDevice(id);
     if(status != hipSuccess)
-        MIGRAPH_THROW("Error setting device");
+        MIGRAPHX_THROW("Error setting device");
 }
 
 void gpu_sync() { hipDeviceSynchronize(); }
 
-void copy_to_gpu(argument src, argument dst)
+void copy_to_gpu(const argument& src, const argument& dst)
 {
     std::size_t src_size = src.get_shape().bytes();
     std::size_t dst_size = dst.get_shape().bytes();
     if(src_size > dst_size)
-        MIGRAPH_THROW("Not enough memory available in destination to do copy");
+        MIGRAPHX_THROW("Not enough memory available in destination to do copy");
     auto status = hipMemcpy(dst.data(), src.data(), src_size, hipMemcpyHostToDevice);
     if(status != hipSuccess)
-        MIGRAPH_THROW("Copy to gpu failed: " + hip_error(status));
+        MIGRAPHX_THROW("Copy to gpu failed: " + hip_error(status));
 }
 
 } // namespace gpu
-} // namespace MIGRAPH_INLINE_NS
-} // namespace migraph
+} // namespace MIGRAPHX_INLINE_NS
+} // namespace migraphx

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

-#ifndef MIGRAPH_GUARD_RTGLIB_ADD_HPP
-#define MIGRAPH_GUARD_RTGLIB_ADD_HPP
-
-#include <migraph/gpu/lowering.hpp>
-#include <migraph/manage_ptr.hpp>
-#include <migraph/instruction.hpp>
-#include <migraph/operators.hpp>
-#include <migraph/generate.hpp>
-#include <migraph/shape_for_each.hpp>
-#include <migraph/gpu/miopen.hpp>
-#include <migraph/gpu/hip.hpp>
-#include <migraph/dfor.hpp>
-#include <migraph/gpu/device/contiguous.hpp>
-#include <migraph/gpu/device/add.hpp>
-#include <migraph/iterator_for.hpp>
-#include <migraph/gpu/rocblas.hpp>
-#include <migraph/gpu/context.hpp>
-#include <migraph/config.hpp>
-#include <utility>
-
-namespace migraph {
-inline namespace MIGRAPH_INLINE_NS {
-namespace gpu {
-
-struct hip_add
-{
-    std::string name() const { return "gpu::add"; }
-    shape compute_shape(const std::vector<shape>& inputs) const;
-    argument compute(context&, const shape&, const std::vector<argument>& args) const;
-    int output_alias(const std::vector<shape>& shapes) const { return shapes.size() - 1; }
-};
-
-struct miopen_add
-{
-    std::string name() const { return "gpu::add"; }
-    shape compute_shape(const std::vector<shape>& inputs) const;
-    argument
-    compute(context& ctx, const shape& output_shape, const std::vector<argument>& args) const;
-    int output_alias(const std::vector<shape>& shapes) const { return shapes.size() - 1; }
-};
-
-} // namespace gpu
-} // namespace MIGRAPH_INLINE_NS
-} // namespace migraph
-
-#endif

src/targets/gpu/include/migraph/gpu/mul.hpp

-#ifndef MIGRAPH_GUARD_RTGLIB_MUL_HPP
-#define MIGRAPH_GUARD_RTGLIB_MUL_HPP
-
-#include <migraph/gpu/lowering.hpp>
-#include <migraph/manage_ptr.hpp>
-#include <migraph/instruction.hpp>
-#include <migraph/operators.hpp>
-#include <migraph/generate.hpp>
-#include <migraph/shape_for_each.hpp>
-#include <migraph/config.hpp>
-#include <migraph/gpu/miopen.hpp>
-#include <migraph/gpu/hip.hpp>
-#include <migraph/dfor.hpp>
-#include <migraph/gpu/device/contiguous.hpp>
-#include <migraph/gpu/device/mul.hpp>
-#include <migraph/iterator_for.hpp>
-#include <migraph/gpu/rocblas.hpp>
-#include <migraph/gpu/context.hpp>
-#include <utility>
-
-namespace migraph {
-inline namespace MIGRAPH_INLINE_NS {
-namespace gpu {
-
-struct hip_mul
-{
-    std::string name() const { return "gpu::mul"; }
-    shape compute_shape(const std::vector<shape>& inputs) const;
-    argument compute(context&, const shape&, const std::vector<argument>& args) const;
-    int output_alias(const std::vector<shape>& shapes) const { return shapes.size() - 1; }
-};
-
-} // namespace gpu
-} // namespace MIGRAPH_INLINE_NS
-} // namespace migraph
-
-#endif

src/targets/gpu/include/migraph/gpu/target.hpp

-#ifndef MIGRAPH_GUARD_MIGRAPHLIB_MIOPEN_TARGET_HPP
-#define MIGRAPH_GUARD_MIGRAPHLIB_MIOPEN_TARGET_HPP
-
-#include <migraph/program.hpp>
-#include <migraph/config.hpp>
-
-namespace migraph {
-inline namespace MIGRAPH_INLINE_NS {
-namespace gpu {
-
-struct target
-{
-    std::string name() const;
-    std::vector<pass> get_passes(migraph::context& gctx) const;
-    migraph::context get_context() const;
-};
-
-} // namespace gpu
-} // namespace MIGRAPH_INLINE_NS
-} // namespace migraph
-
-#endif

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/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 miopen_abs
+{
+    shared<activation_descriptor> ad;
+    std::string name() const { return "gpu::abs"; }
+    shape compute_shape(const std::vector<shape>& inputs) const;
+    argument
+    compute(context& ctx, const shape& output_shape, const std::vector<argument>& args) const;
+    int output_alias(const std::vector<shape>& shapes) const { return shapes.size() - 1; }
+};
+
+} // namespace gpu
+} // namespace MIGRAPHX_INLINE_NS
+} // namespace migraphx
+
+#endif