Merge from develop

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

+#ifndef MIGRAPHX_GUARD_RTGLIB_DEVICE_LOG_HPP
+#define MIGRAPHX_GUARD_RTGLIB_DEVICE_LOG_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 {
+
+void log(hipStream_t stream, const argument& result, const argument& arg);
+
+} // namespace device
+} // namespace gpu
+} // namespace MIGRAPHX_INLINE_NS
+} // namespace migraphx
+
+#endif

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

+
+#ifndef MIGRAPHX_GUARD_RTGLIB_DEVICE_MAX_HPP
+#define MIGRAPHX_GUARD_RTGLIB_DEVICE_MAX_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 {
+
+void max(hipStream_t stream, const argument& result, const argument& arg1, const argument& arg2);
+
+} // namespace device
+} // namespace gpu
+} // namespace MIGRAPHX_INLINE_NS
+} // namespace migraphx
+
+#endif

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

+
+#ifndef MIGRAPHX_GUARD_RTGLIB_DEVICE_MIN_HPP
+#define MIGRAPHX_GUARD_RTGLIB_DEVICE_MIN_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 {
+
+void min(hipStream_t stream, const argument& result, const argument& arg1, const argument& arg2);
+
+} // namespace device
+} // namespace gpu
+} // namespace MIGRAPHX_INLINE_NS
+} // namespace migraphx
+
+#endif

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

+
+#ifndef MIGRAPHX_GUARD_RTGLIB_DEVICE_PAD_HPP
+#define MIGRAPHX_GUARD_RTGLIB_DEVICE_PAD_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 pad(hipStream_t stream,
+             argument result,
+             argument arg1,
+             float value,
+             std::vector<std::int64_t> pads);
+
+} // namespace device
+} // namespace gpu
+} // namespace MIGRAPHX_INLINE_NS
+} // namespace migraphx
+
+#endif

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

-
-#ifndef MIGRAPHX_GUARD_RTGLIB_DEVICE_ADD_HPP
-#define MIGRAPHX_GUARD_RTGLIB_DEVICE_ADD_HPP
+#ifndef MIGRAPHX_GUARD_RTGLIB_DEVICE_SIN_HPP
+#define MIGRAPHX_GUARD_RTGLIB_DEVICE_SIN_HPP
 
 #include <migraphx/argument.hpp>
 #include <migraphx/config.hpp>

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

+#ifndef MIGRAPHX_GUARD_RTGLIB_DEVICE_SINH_HPP
+#define MIGRAPHX_GUARD_RTGLIB_DEVICE_SINH_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 {
+
+void sinh(hipStream_t stream, const argument& result, const argument& arg);
+
+} // namespace device
+} // namespace gpu
+} // namespace MIGRAPHX_INLINE_NS
+} // namespace migraphx
+
+#endif

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

+#ifndef MIGRAPHX_GUARD_RTGLIB_DEVICE_SUB_HPP
+#define MIGRAPHX_GUARD_RTGLIB_DEVICE_SUB_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 {
+
+void sub(hipStream_t stream, const argument& result, const argument& arg1, const argument& arg2);
+
+} // namespace device
+} // namespace gpu
+} // namespace MIGRAPHX_INLINE_NS
+} // namespace migraphx
+
+#endif

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

+#ifndef MIGRAPHX_GUARD_RTGLIB_DEVICE_TAN_HPP
+#define MIGRAPHX_GUARD_RTGLIB_DEVICE_TAN_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 {
+
+void tan(hipStream_t stream, const argument& result, const argument& arg);
+
+} // 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/shape.hpp>
+#include <migraphx/gpu/miopen.hpp>
+
+namespace migraphx {
+inline namespace MIGRAPHX_INLINE_NS {
+namespace gpu {
+
+struct context;
+
+struct miopen_elu
+{
+    shared<activation_descriptor> ad;
+    std::string name() const { return "gpu::elu"; }
+    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

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

+#ifndef MIGRAPHX_GUARD_RTGLIB_EXP_HPP
+#define MIGRAPHX_GUARD_RTGLIB_EXP_HPP
+
+#include <migraphx/gpu/oper.hpp>
+#include <migraphx/gpu/device/exp.hpp>
+
+namespace migraphx {
+inline namespace MIGRAPHX_INLINE_NS {
+namespace gpu {
+
+struct hip_exp : unary_device<hip_exp, device::exp>
+{
+};
+
+} // namespace gpu
+} // namespace MIGRAPHX_INLINE_NS
+} // namespace migraphx
+
+#endif

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

+#ifndef MIGRAPHX_GUARD_RTGLIB_GATHER_HPP
+#define MIGRAPHX_GUARD_RTGLIB_GATHER_HPP
+
+#include <migraphx/shape.hpp>
+#include <migraphx/gpu/miopen.hpp>
+
+namespace migraphx {
+inline namespace MIGRAPHX_INLINE_NS {
+namespace gpu {
+
+struct context;
+
+struct hip_gather
+{
+    op::gather op;
+    std::string name() const { return "gpu::gather"; }
+    shape compute_shape(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

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;

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

@@ -9,26 +9,27 @@ namespace migraphx {
 inline namespace MIGRAPHX_INLINE_NS {
 namespace gpu {
 
-migraphx::argument allocate_gpu(const migraphx::shape& s, bool host = false);
+argument allocate_gpu(const shape& s, bool host = false);
 
-migraphx::argument to_gpu(migraphx::argument arg, bool host = false);
+argument to_gpu(const argument& arg, bool host = false);
 
-migraphx::argument from_gpu(migraphx::argument arg);
+argument from_gpu(const argument& arg);
 
 void set_device(std::size_t id);
 
 void gpu_sync();
 
-void copy_to_gpu(argument src, argument dst);
+void copy_to_gpu(const argument& src, const argument& dst);
 
 struct hip_allocate
 {
+    shape s;
     std::string tag{};
     std::string name() const { return "hip::allocate"; }
     shape compute_shape(const std::vector<shape>& inputs) const
     {
-        check_shapes{inputs}.has(1);
-        return inputs.front();
+        check_shapes{inputs}.has(0);
+        return s;
     }
     argument compute(context&, const shape& output_shape, const std::vector<argument>&) const
     {

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

 #ifndef MIGRAPHX_GUARD_RTGLIB_LEAKY_RELU_HPP
 #define MIGRAPHX_GUARD_RTGLIB_LEAKY_RELU_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_leaky_relu
 {
     shared<activation_descriptor> ad;

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

+#ifndef MIGRAPHX_GUARD_RTGLIB_LOG_HPP
+#define MIGRAPHX_GUARD_RTGLIB_LOG_HPP
+
+#include <migraphx/gpu/oper.hpp>
+#include <migraphx/gpu/device/log.hpp>
+
+namespace migraphx {
+inline namespace MIGRAPHX_INLINE_NS {
+namespace gpu {
+
+struct hip_log : unary_device<hip_log, device::log>
+{
+};
+
+} // namespace gpu
+} // namespace MIGRAPHX_INLINE_NS
+} // namespace migraphx
+
+#endif

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

+#ifndef MIGRAPHX_GUARD_RTGLIB_LRN_HPP
+#define MIGRAPHX_GUARD_RTGLIB_LRN_HPP
+
+#include <migraphx/shape.hpp>
+#include <migraphx/gpu/miopen.hpp>
+
+namespace migraphx {
+inline namespace MIGRAPHX_INLINE_NS {
+namespace gpu {
+
+struct context;
+
+struct miopen_lrn
+{
+    shared<lrn_descriptor> ldesc;
+    std::string name() const { return "gpu::lrn"; }
+    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

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

+#ifndef MIGRAPHX_GUARD_RTGLIB_MAX_HPP
+#define MIGRAPHX_GUARD_RTGLIB_MAX_HPP
+
+#include <migraphx/gpu/oper.hpp>
+#include <migraphx/gpu/device/max.hpp>
+
+namespace migraphx {
+inline namespace MIGRAPHX_INLINE_NS {
+namespace gpu {
+
+struct hip_max : binary_device<hip_max, device::max>
+{
+};
+
+} // namespace gpu
+} // namespace MIGRAPHX_INLINE_NS
+} // namespace migraphx
+
+#endif

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

+#ifndef MIGRAPHX_GUARD_RTGLIB_MIN_HPP
+#define MIGRAPHX_GUARD_RTGLIB_MIN_HPP
+
+#include <migraphx/gpu/oper.hpp>
+#include <migraphx/gpu/device/min.hpp>
+
+namespace migraphx {
+inline namespace MIGRAPHX_INLINE_NS {
+namespace gpu {
+
+struct hip_min : binary_device<hip_min, device::min>
+{
+};
+
+} // namespace gpu
+} // namespace MIGRAPHX_INLINE_NS
+} // namespace migraphx
+
+#endif

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

@@ -23,6 +23,8 @@ using fusion_plan_descriptor = MIGRAPHX_MANAGE_PTR(miopenFusionPlanDescriptor_t,
                                                    miopenDestroyFusionPlan);
 using fused_operator_args    = MIGRAPHX_MANAGE_PTR(miopenOperatorArgs_t, miopenDestroyOperatorArgs);
 
+using lrn_descriptor = MIGRAPHX_MANAGE_PTR(miopenLRNDescriptor_t, miopenDestroyLRNDescriptor);
+
 template <class Result, class F, class... Ts>
 Result make_obj(F f, Ts... xs)
 {
@@ -54,14 +56,19 @@ inline tensor_descriptor make_tensor(const migraphx::shape& s)
 inline convolution_descriptor make_conv(const migraphx::op::convolution& op)
 {
     auto c = make_obj<convolution_descriptor>(&miopenCreateConvolutionDescriptor);
+    miopenConvolutionMode_t c_mode = miopenConvolution;
+    if(op.group > 1)
+        c_mode = miopenGroupConv;
     miopenInitConvolutionDescriptor(c.get(),
-                                    miopenConvolution,
+                                    c_mode,
                                     op.padding[0],
                                     op.padding[1],
                                     op.stride[0],
                                     op.stride[1],
                                     op.dilation[0],
                                     op.dilation[1]);
+    if(op.group > 1)
+        miopenSetConvolutionGroupCount(c.get(), op.group);
     return c;
 }
 
@@ -84,6 +91,13 @@ inline pooling_descriptor make_pooling(const migraphx::op::pooling& op)
     return p;
 }
 
+inline lrn_descriptor make_lrn(const migraphx::op::lrn& op)
+{
+    auto ldesc = make_obj<lrn_descriptor>(&miopenCreateLRNDescriptor);
+    miopenSetLRNDescriptor(ldesc.get(), miopenLRNCrossChannel, op.size, op.alpha, op.beta, op.bias);
+    return ldesc;
+}
+
 inline activation_descriptor make_relu()
 {
     auto ad = make_obj<activation_descriptor>(&miopenCreateActivationDescriptor);
@@ -91,6 +105,29 @@ inline activation_descriptor make_relu()
     return ad;
 }
 
+inline activation_descriptor make_sigmoid()
+{
+    auto ad = make_obj<activation_descriptor>(&miopenCreateActivationDescriptor);
+    miopenSetActivationDescriptor(ad.get(), miopenActivationLOGISTIC, 0, 0, 0);
+    return ad;
+}
+
+inline activation_descriptor make_tanh()
+{
+    auto ad = make_obj<activation_descriptor>(&miopenCreateActivationDescriptor);
+    // onnx operator does not apply additional scaling for tanh
+    // defaults for alpha and beta are therefore set to 1
+    miopenSetActivationDescriptor(ad.get(), miopenActivationTANH, 1, 1, 0);
+    return ad;
+}
+
+inline activation_descriptor make_abs()
+{
+    auto ad = make_obj<activation_descriptor>(&miopenCreateActivationDescriptor);
+    miopenSetActivationDescriptor(ad.get(), miopenActivationABS, 0, 0, 0);
+    return ad;
+}
+
 inline activation_descriptor make_leaky_relu(double alpha)
 {
     auto ad = make_obj<activation_descriptor>(&miopenCreateActivationDescriptor);
@@ -98,6 +135,13 @@ inline activation_descriptor make_leaky_relu(double alpha)
     return ad;
 }
 
+inline activation_descriptor make_elu(double alpha)
+{
+    auto ad = make_obj<activation_descriptor>(&miopenCreateActivationDescriptor);
+    miopenSetActivationDescriptor(ad.get(), miopenActivationELU, alpha, 0, 0);
+    return ad;
+}
+
 inline fusion_plan_descriptor make_fusion_plan(const shape& input)
 {
     auto t = make_tensor(input);

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

 #ifndef MIGRAPHX_GUARD_RTGLIB_MUL_HPP
 #define MIGRAPHX_GUARD_RTGLIB_MUL_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/oper.hpp>
 #include <migraphx/gpu/device/mul.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 hip_mul
+struct hip_mul : binary_device<hip_mul, device::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