Merge branch 'develop' into mem_color_ordering_fix

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

-#ifndef MIGRAPH_GUARD_RTGLIB_SOFTMAX_HPP
-#define MIGRAPH_GUARD_RTGLIB_SOFTMAX_HPP
+#ifndef MIGRAPHX_GUARD_RTGLIB_SOFTMAX_HPP
+#define MIGRAPHX_GUARD_RTGLIB_SOFTMAX_HPP
 
 #include <migraphx/gpu/lowering.hpp>
 #include <migraphx/manage_ptr.hpp>
@@ -19,7 +19,7 @@
 #include <utility>
 
 namespace migraphx {
-inline namespace MIGRAPH_INLINE_NS {
+inline namespace MIGRAPHX_INLINE_NS {
 namespace gpu {
 
 struct miopen_softmax
@@ -33,7 +33,7 @@ struct miopen_softmax
 };
 
 } // namespace gpu
-} // namespace MIGRAPH_INLINE_NS
+} // namespace MIGRAPHX_INLINE_NS
 } // namespace migraphx
 
 #endif

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

+#ifndef MIGRAPHX_GUARD_RTGLIB_SUB_HPP
+#define MIGRAPHX_GUARD_RTGLIB_SUB_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/sub.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 hip_sub : binary_device<hip_sub, device::sub>
+{
+};
+
+} // namespace gpu
+} // namespace MIGRAPHX_INLINE_NS
+} // namespace migraphx
+
+#endif

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

+#ifndef MIGRAPHX_GUARD_RTGLIB_TAN_HPP
+#define MIGRAPHX_GUARD_RTGLIB_TAN_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/tan.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 hip_tan : unary_device<hip_tan, device::tan>
+{
+};
+
+} // namespace gpu
+} // namespace MIGRAPHX_INLINE_NS
+} // namespace migraphx
+
+#endif

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

+#ifndef MIGRAPHX_GUARD_RTGLIB_TANH_HPP
+#define MIGRAPHX_GUARD_RTGLIB_TANH_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_tanh
+{
+    shared<activation_descriptor> ad;
+    std::string name() const { return "gpu::tanh"; }
+    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/target.hpp

-#ifndef MIGRAPH_GUARD_MIGRAPHLIB_MIOPEN_TARGET_HPP
-#define MIGRAPH_GUARD_MIGRAPHLIB_MIOPEN_TARGET_HPP
+#ifndef MIGRAPHX_GUARD_MIGRAPHLIB_MIOPEN_TARGET_HPP
+#define MIGRAPHX_GUARD_MIGRAPHLIB_MIOPEN_TARGET_HPP
 
 #include <migraphx/program.hpp>
 #include <migraphx/config.hpp>
 
 namespace migraphx {
-inline namespace MIGRAPH_INLINE_NS {
+inline namespace MIGRAPHX_INLINE_NS {
 namespace gpu {
 
 struct target
@@ -16,7 +16,7 @@ struct target
 };
 
 } // namespace gpu
-} // namespace MIGRAPH_INLINE_NS
+} // namespace MIGRAPHX_INLINE_NS
 } // namespace migraphx
 
 #endif

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

-#ifndef MIGRAPH_GUARD_RTGLIB_MIOPEN_WRITE_LITERALS_HPP
-#define MIGRAPH_GUARD_RTGLIB_MIOPEN_WRITE_LITERALS_HPP
+#ifndef MIGRAPHX_GUARD_RTGLIB_MIOPEN_WRITE_LITERALS_HPP
+#define MIGRAPHX_GUARD_RTGLIB_MIOPEN_WRITE_LITERALS_HPP
 
 #include <migraphx/program.hpp>
 #include <migraphx/gpu/context.hpp>
 
 namespace migraphx {
-inline namespace MIGRAPH_INLINE_NS {
+inline namespace MIGRAPHX_INLINE_NS {
 
 namespace gpu {
 
@@ -18,7 +18,7 @@ struct write_literals
 };
 
 } // namespace gpu
-} // namespace MIGRAPH_INLINE_NS
+} // namespace MIGRAPHX_INLINE_NS
 } // namespace migraphx
 
 #endif

src/targets/gpu/leaky_relu.cpp

@@ -5,7 +5,7 @@
 #include <utility>
 
 namespace migraphx {
-inline namespace MIGRAPH_INLINE_NS {
+inline namespace MIGRAPHX_INLINE_NS {
 namespace gpu {
 
 shape miopen_leaky_relu::compute_shape(const std::vector<shape>& inputs) const
@@ -18,7 +18,8 @@ argument miopen_leaky_relu::compute(context& ctx,
                                     const shape& output_shape,
                                     const std::vector<argument>& args) const
 {
-    float alpha = 1, beta = 0;
+    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(),
@@ -34,5 +35,5 @@ argument miopen_leaky_relu::compute(context& ctx,
 }
 
 } // namespace gpu
-} // namespace MIGRAPH_INLINE_NS
+} // namespace MIGRAPHX_INLINE_NS
 } // namespace migraphx

src/targets/gpu/lowering.cpp

@@ -16,25 +16,48 @@
 #include <migraphx/gpu/convolution.hpp>
 #include <migraphx/gpu/contiguous.hpp>
 #include <migraphx/gpu/relu.hpp>
+#include <migraphx/gpu/sigmoid.hpp>
+#include <migraphx/gpu/abs.hpp>
 #include <migraphx/gpu/leaky_relu.hpp>
+#include <migraphx/gpu/elu.hpp>
 #include <migraphx/gpu/softmax.hpp>
 #include <migraphx/gpu/add.hpp>
+#include <migraphx/gpu/sub.hpp>
+#include <migraphx/gpu/exp.hpp>
+#include <migraphx/gpu/log.hpp>
 #include <migraphx/gpu/sin.hpp>
+#include <migraphx/gpu/cos.hpp>
+#include <migraphx/gpu/tan.hpp>
+#include <migraphx/gpu/sinh.hpp>
+#include <migraphx/gpu/cosh.hpp>
+#include <migraphx/gpu/tanh.hpp>
+#include <migraphx/gpu/asin.hpp>
+#include <migraphx/gpu/acos.hpp>
+#include <migraphx/gpu/atan.hpp>
 #include <migraphx/gpu/mul.hpp>
+#include <migraphx/gpu/max.hpp>
+#include <migraphx/gpu/min.hpp>
 #include <migraphx/gpu/batchnorm.hpp>
 #include <migraphx/gpu/pooling.hpp>
 #include <migraphx/gpu/gemm.hpp>
 #include <migraphx/gpu/concat.hpp>
+#include <migraphx/gpu/pad.hpp>
+#include <migraphx/gpu/gather.hpp>
+#include <migraphx/gpu/lrn.hpp>
 #include <utility>
+#include <functional>
+#include <algorithm>
 
 namespace migraphx {
-inline namespace MIGRAPH_INLINE_NS {
+inline namespace MIGRAPHX_INLINE_NS {
 namespace gpu {
 
 struct miopen_apply
 {
     program* prog = nullptr;
     context ctx{};
+    std::unordered_map<std::string, std::function<instruction_ref(instruction_ref)>> apply_map{};
+    instruction_ref last{};
 
     void check_shape(shape x, instruction_ref i)
     {
@@ -43,65 +66,62 @@ struct miopen_apply
         (void)i;
     }
 
+    void init()
+    {
+        this->last = instruction::get_output_alias(std::prev(prog->end()));
+        add_miopen_simple_op<miopen_relu>("relu", make_relu);
+        add_miopen_simple_op<miopen_sigmoid>("sigmoid", make_sigmoid);
+        add_miopen_simple_op<miopen_abs>("abs", make_abs);
+        add_miopen_simple_op<miopen_tanh>("tanh", make_tanh);
+
+        add_miopen_extend_op<miopen_leaky_relu, op::leaky_relu>("leaky_relu", make_leaky_relu);
+        add_miopen_extend_op<miopen_elu, op::elu>("elu", make_elu);
+
+        add_generic_op<hip_add>("add");
+        add_generic_op<hip_sub>("sub");
+        add_generic_op<hip_exp>("exp");
+        add_generic_op<hip_log>("log");
+        add_generic_op<hip_sin>("sin");
+        add_generic_op<hip_cos>("cos");
+        add_generic_op<hip_tan>("tan");
+        add_generic_op<hip_sinh>("sinh");
+        add_generic_op<hip_cosh>("cosh");
+        add_generic_op<hip_asin>("asin");
+        add_generic_op<hip_acos>("acos");
+        add_generic_op<hip_atan>("atan");
+        add_generic_op<hip_mul>("mul");
+        add_generic_op<hip_max>("max");
+        add_generic_op<hip_min>("min");
+
+        add_extend_op<miopen_gemm, op::dot>("dot");
+        add_extend_op<miopen_contiguous, op::contiguous>("contiguous");
+        add_extend_op<hip_concat, op::concat>("concat");
+        add_extend_op<miopen_softmax, op::softmax>("softmax");
+        add_extend_op<hip_gather, op::gather>("gather");
+        add_extend_op<hip_pad, op::pad>("pad");
+
+        add_lrn_op();
+        add_convolution_op();
+        add_pooling_op();
+        add_batch_norm_inference_op();
+    }
+
     void apply()
     {
+        init();
         for(auto it = prog->begin(); it != prog->end(); it++)
         {
             auto s = it->get_shape();
-            if(it->name() == "convolution")
-            {
-                check_shape(s, apply_convolution(it));
-            }
-            else if(it->name() == "relu")
-            {
-                check_shape(s, apply_relu(it));
-            }
-            else if(it->name() == "leaky_relu")
-            {
-                check_shape(s, apply_leaky_relu(it));
-            }
-            else if(it->name() == "pooling")
-            {
-                check_shape(s, apply_pooling(it));
-            }
-            else if(it->name() == "add")
-            {
-                check_shape(s, apply_add(it));
-            }
-            else if(it->name() == "sin")
-            {
-                check_shape(s, apply_sin(it));
-            }
-            else if(it->name() == "mul")
-            {
-                check_shape(s, apply_mul(it));
-            }
-            else if(it->name() == "dot")
-            {
-                check_shape(s, apply_gemm(it));
-            }
-            else if(it->name() == "contiguous")
+            if(apply_map.count(it->name()) > 0)
             {
-                check_shape(s, apply_contiguous(it));
-            }
-            else if(it->name() == "concat")
-            {
-                check_shape(s, apply_concat(it));
-            }
-            else if(it->name() == "batch_norm_inference")
-            {
-                check_shape(s, apply_batch_norm_inference(it));
-            }
-            else if(it->name() == "softmax")
-            {
-                check_shape(s, apply_softmax(it));
+                check_shape(s, apply_map.at(it->name())(it));
             }
         }
     }
 
     instruction_ref insert_allocation(instruction_ref ins, const shape& s, std::string tag = "")
     {
-        if(ins == --prog->end() and tag.empty())
+        if(ins == last and tag.empty())
         {
             return prog->add_parameter("output", s);
         }
@@ -113,124 +133,118 @@ struct miopen_apply
         }
     }
 
-    instruction_ref apply_convolution(instruction_ref ins)
+    void add_convolution_op()
     {
-        auto&& op = any_cast<op::convolution>(ins->get_operator());
+        apply_map.emplace("convolution", [=](instruction_ref ins) {
+            auto&& op = any_cast<op::convolution>(ins->get_operator());
 
-        auto conv = miopen_convolution{op, make_conv(op)};
-        auto ws   = conv.compile(ctx, ins->get_shape(), ins->inputs());
+            auto conv = miopen_convolution{op, make_conv(op)};
+            auto ws   = conv.compile(ctx, ins->get_shape(), to_shapes(ins->inputs()));
 
-        auto workspace = insert_allocation(ins, ws, "workspace");
-        auto output    = insert_allocation(ins, ins->get_shape());
+            auto workspace = insert_allocation(ins, ws, "workspace");
+            auto output    = insert_allocation(ins, ins->get_shape());
 
-        return prog->replace_instruction(
-            ins, conv, ins->inputs().at(0), ins->inputs().at(1), workspace, output);
+            return prog->replace_instruction(
+                ins, conv, ins->inputs().at(0), ins->inputs().at(1), workspace, output);
+        });
     }
 
-    instruction_ref apply_pooling(instruction_ref ins)
+    void add_pooling_op()
     {
-        auto&& op   = any_cast<op::pooling>(ins->get_operator());
-        auto pd     = make_pooling(op);
-        auto output = insert_allocation(ins, ins->get_shape());
-
-        return prog->replace_instruction(
-            ins, miopen_pooling{op, std::move(pd)}, ins->inputs().at(0), output);
+        apply_map.emplace("pooling", [=](instruction_ref ins) {
+            auto&& op   = any_cast<op::pooling>(ins->get_operator());
+            auto pd     = make_pooling(op);
+            auto output = insert_allocation(ins, ins->get_shape());
+
+            return prog->replace_instruction(
+                ins, miopen_pooling{op, std::move(pd)}, ins->inputs().at(0), output);
+        });
     }
 
-    instruction_ref apply_relu(instruction_ref ins)
+    void add_lrn_op()
     {
-        auto ad = make_relu();
-
-        auto output = insert_allocation(ins, ins->get_shape());
-        return prog->replace_instruction(
-            ins, miopen_relu{std::move(ad)}, ins->inputs().at(0), output);
+        apply_map.emplace("lrn", [=](instruction_ref ins) {
+            auto&& op   = any_cast<op::lrn>(ins->get_operator());
+            auto ldesc  = make_lrn(op);
+            auto output = insert_allocation(ins, ins->get_shape());
+            return prog->replace_instruction(
+                ins, miopen_lrn{std::move(ldesc)}, ins->inputs().at(0), output);
+        });
     }
 
-    instruction_ref apply_leaky_relu(instruction_ref ins)
+    template <class T>
+    void add_generic_op(std::string name)
     {
-        auto&& op = any_cast<op::leaky_relu>(ins->get_operator());
-        auto ad   = make_leaky_relu(op.alpha);
+        apply_map.emplace(name, [=](instruction_ref ins) {
+            auto output                       = insert_allocation(ins, ins->get_shape());
+            std::vector<instruction_ref> refs = ins->inputs();
+            refs.push_back(output);
 
-        auto output = insert_allocation(ins, ins->get_shape());
-        return prog->replace_instruction(
-            ins, miopen_leaky_relu{std::move(ad)}, ins->inputs().at(0), output);
+            return prog->replace_instruction(ins, T{}, refs);
+        });
     }
 
-    instruction_ref apply_softmax(instruction_ref ins)
+    template <class T, class Op>
+    void add_extend_op(std::string name)
     {
-        auto&& op   = any_cast<op::softmax>(ins->get_operator());
-        auto output = insert_allocation(ins, ins->get_shape());
-        return prog->replace_instruction(ins, miopen_softmax{op}, ins->inputs().at(0), output);
+        apply_map.emplace(name, [=](instruction_ref ins) {
+            auto&& op                         = any_cast<Op>(ins->get_operator());
+            auto output                       = insert_allocation(ins, ins->get_shape());
+            std::vector<instruction_ref> refs = ins->inputs();
+            refs.push_back(output);
+
+            return prog->replace_instruction(ins, T{op}, refs);
+        });
     }
 
-    instruction_ref apply_add(instruction_ref ins)
+    template <class T, class Op, class F>
+    void add_miopen_extend_op(std::string name, F f)
     {
-        auto output = insert_allocation(ins, ins->get_shape());
-        return prog->replace_instruction(
-            ins, hip_add{}, ins->inputs().at(0), ins->inputs().at(1), output);
-    }
+        apply_map.emplace(name, [=](instruction_ref ins) {
+            auto&& op = any_cast<Op>(ins->get_operator());
+            auto ad   = f(op.alpha);
 
-    instruction_ref apply_sin(instruction_ref ins)
-    {
-        auto output = insert_allocation(ins, ins->get_shape());
-        return prog->replace_instruction(ins, hip_sin{}, ins->inputs().at(0), output);
-    }
-
-    instruction_ref apply_mul(instruction_ref ins)
-    {
-        auto output = insert_allocation(ins, ins->get_shape());
-        return prog->replace_instruction(
-            ins, hip_mul{}, ins->inputs().at(0), ins->inputs().at(1), output);
-    }
-
-    instruction_ref apply_gemm(instruction_ref ins)
-    {
-        auto&& op   = any_cast<op::dot>(ins->get_operator());
-        auto output = insert_allocation(ins, ins->get_shape());
-        return prog->replace_instruction(
-            ins, miopen_gemm{op}, ins->inputs().at(0), ins->inputs().at(1), output);
-    }
-
-    instruction_ref apply_contiguous(instruction_ref ins)
-    {
-        auto&& op   = any_cast<op::contiguous>(ins->get_operator());
-        auto output = insert_allocation(ins, ins->get_shape());
-        return prog->replace_instruction(ins, miopen_contiguous{op}, ins->inputs().at(0), output);
+            auto output = insert_allocation(ins, ins->get_shape());
+            return prog->replace_instruction(ins, T{std::move(ad)}, ins->inputs().at(0), output);
+        });
     }
 
-    instruction_ref apply_concat(instruction_ref ins)
+    template <class T, class F>
+    void add_miopen_simple_op(std::string name, F f)
     {
-        auto&& op                         = any_cast<op::concat>(ins->get_operator());
-        auto output                       = insert_allocation(ins, ins->get_shape());
-        std::vector<instruction_ref> refs = ins->inputs();
-        refs.push_back(output);
-        return prog->replace_instruction(ins, hip_concat{op}, refs);
+        apply_map.emplace(name, [=](instruction_ref ins) {
+            auto ad     = f();
+            auto output = insert_allocation(ins, ins->get_shape());
+            return prog->replace_instruction(ins, T{std::move(ad)}, ins->inputs().at(0), output);
+        });
     }
 
-    instruction_ref apply_batch_norm_inference(instruction_ref ins)
+    void add_batch_norm_inference_op()
     {
-        auto&& op       = any_cast<op::batch_norm_inference>(ins->get_operator());
-        auto output     = insert_allocation(ins, ins->get_shape());
-        shape old_shape = ins->inputs().at(1)->get_shape();
-        std::vector<int64_t> new_shape{1, static_cast<int64_t>(old_shape.elements()), 1, 1};
-        auto reshape_op = op::reshape{new_shape};
-        std::vector<instruction_ref> reshapes;
-        std::transform(ins->inputs().begin() + 1,
-                       ins->inputs().end(),
-                       std::back_inserter(reshapes),
-                       [&](auto i) { return prog->insert_instruction(ins, reshape_op, i); });
-        return prog->replace_instruction(ins,
-                                         miopen_batch_norm_inference{op},
-                                         ins->inputs().at(0),
-                                         reshapes[0],
-                                         reshapes[1],
-                                         reshapes[2],
-                                         reshapes[3],
-                                         output);
+        apply_map.emplace("batch_norm_inference", [=](instruction_ref ins) {
+            auto&& op       = any_cast<op::batch_norm_inference>(ins->get_operator());
+            auto output     = insert_allocation(ins, ins->get_shape());
+            shape old_shape = ins->inputs().at(1)->get_shape();
+            std::vector<int64_t> new_shape{1, static_cast<int64_t>(old_shape.elements()), 1, 1};
+            auto reshape_op = op::reshape{new_shape};
+            std::vector<instruction_ref> reshapes;
+            std::transform(ins->inputs().begin() + 1,
+                           ins->inputs().end(),
+                           std::back_inserter(reshapes),
+                           [&](auto i) { return prog->insert_instruction(ins, reshape_op, i); });
+            return prog->replace_instruction(ins,
+                                             miopen_batch_norm_inference{op},
+                                             ins->inputs().at(0),
+                                             reshapes[0],
+                                             reshapes[1],
+                                             reshapes[2],
+                                             reshapes[3],
+                                             output);
+        });
     }
 };
 
 void lowering::apply(program& p) const { miopen_apply{&p, ctx}.apply(); }
 } // namespace gpu
-} // namespace MIGRAPH_INLINE_NS
+} // namespace MIGRAPHX_INLINE_NS
 } // namespace migraphx

src/targets/gpu/lrn.cpp

+#include <migraphx/gpu/lrn.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_lrn::compute_shape(const std::vector<shape>& inputs) const
+{
+    check_shapes{inputs, *this}.has(2).not_broadcasted();
+    return inputs.at(1);
+}
+
+argument miopen_lrn::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);
+    miopenLRNForward(ctx.get_stream().get_miopen(),
+                     ldesc.get(),
+                     &alpha,
+                     x_desc.get(),
+                     args[0].implicit(),
+                     &beta,
+                     y_desc.get(),
+                     args[1].implicit(),
+                     false,
+                     nullptr);
+
+    return args[1];
+}
+
+} // namespace gpu
+} // namespace MIGRAPHX_INLINE_NS
+} // namespace migraphx

src/targets/gpu/mul.cpp

-#include <migraphx/gpu/mul.hpp>
-#include <migraphx/operators.hpp>
-#include <migraphx/manage_ptr.hpp>
-#include <migraphx/gpu/miopen.hpp>
-#include <utility>
-
-namespace migraphx {
-inline namespace MIGRAPH_INLINE_NS {
-namespace gpu {
-
-shape hip_mul::compute_shape(const std::vector<shape>& inputs) const
-{
-    // check_shapes{inputs, *this}.has(3).standard();
-    check_shapes{inputs, *this}.has(3);
-    return inputs.at(0);
-}
-
-argument hip_mul::compute(context& ctx, const shape&, const std::vector<argument>& args) const
-{
-    device::mul(ctx.get_stream().get(), args[2], args[0], args[1]);
-    return args[2];
-}
-
-} // namespace gpu
-} // namespace MIGRAPH_INLINE_NS
-} // namespace migraphx

src/targets/gpu/sin.cpp → src/targets/gpu/pad.cpp

-#include <migraphx/gpu/sin.hpp>
+#include <migraphx/gpu/pad.hpp>
 #include <migraphx/operators.hpp>
 #include <migraphx/manage_ptr.hpp>
-#include <migraphx/config.hpp>
 #include <migraphx/gpu/miopen.hpp>
+#include <migraphx/gpu/device/pad.hpp>
 #include <utility>
 
 namespace migraphx {
-inline namespace MIGRAPH_INLINE_NS {
+inline namespace MIGRAPHX_INLINE_NS {
 namespace gpu {
 
-shape hip_sin::compute_shape(const std::vector<shape>& inputs) const
+shape hip_pad::compute_shape(std::vector<shape> inputs) const
 {
-    check_shapes{inputs, *this}.has(2);
-    return inputs.at(0);
+    inputs.pop_back();
+    return op.compute_shape(inputs);
 }
 
-argument hip_sin::compute(context& ctx, const shape&, const std::vector<argument>& args) const
+argument hip_pad::compute(context& ctx, const shape&, const std::vector<argument>& args) const
 {
-    device::sin(ctx.get_stream().get(), args[1], args[0]);
-    return args[1];
+    return device::pad(ctx.get_stream().get(), args.back(), args.front(), op.value, op.pads);
 }
 
 } // namespace gpu
-} // namespace MIGRAPH_INLINE_NS
+} // namespace MIGRAPHX_INLINE_NS
 } // namespace migraphx

src/targets/gpu/pooling.cpp

@@ -5,7 +5,7 @@
 #include <utility>
 
 namespace migraphx {
-inline namespace MIGRAPH_INLINE_NS {
+inline namespace MIGRAPHX_INLINE_NS {
 namespace gpu {
 
 shape miopen_pooling::compute_shape(const std::vector<shape>& inputs) const
@@ -20,7 +20,8 @@ argument miopen_pooling::compute(context& ctx,
     auto x_desc = make_tensor(args[0].get_shape());
     auto y_desc = make_tensor(output_shape);
 
-    float alpha = 1, beta = 0;
+    float alpha = 1;
+    float beta  = 0;
 
     miopenPoolingForward(ctx.get_stream().get_miopen(),
                          pd.get(),
@@ -38,5 +39,5 @@ argument miopen_pooling::compute(context& ctx,
 }
 
 } // namespace gpu
-} // namespace MIGRAPH_INLINE_NS
+} // namespace MIGRAPHX_INLINE_NS
 } // namespace migraphx

src/targets/gpu/relu.cpp

@@ -5,7 +5,7 @@
 #include <utility>
 
 namespace migraphx {
-inline namespace MIGRAPH_INLINE_NS {
+inline namespace MIGRAPHX_INLINE_NS {
 namespace gpu {
 
 shape miopen_relu::compute_shape(const std::vector<shape>& inputs) const
@@ -18,7 +18,8 @@ argument miopen_relu::compute(context& ctx,
                               const shape& output_shape,
                               const std::vector<argument>& args) const
 {
-    float alpha = 1, beta = 0;
+    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(),
@@ -34,5 +35,5 @@ argument miopen_relu::compute(context& ctx,
 }
 
 } // namespace gpu
-} // namespace MIGRAPH_INLINE_NS
+} // namespace MIGRAPHX_INLINE_NS
 } // namespace migraphx

src/targets/gpu/rocblas.cpp

 #include <migraphx/gpu/rocblas.hpp>
 
 namespace migraphx {
-inline namespace MIGRAPH_INLINE_NS {
+inline namespace MIGRAPHX_INLINE_NS {
 namespace gpu {
 
 rocblas_handle_ptr create_rocblas_handle_ptr()
@@ -19,5 +19,5 @@ rocblas_handle_ptr create_rocblas_handle_ptr(hipStream_t s)
 }
 
 } // namespace gpu
-} // namespace MIGRAPH_INLINE_NS
+} // namespace MIGRAPHX_INLINE_NS
 } // namespace migraphx

src/targets/gpu/sigmoid.cpp

+#include <migraphx/gpu/sigmoid.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_sigmoid::compute_shape(const std::vector<shape>& inputs) const
+{
+    check_shapes{inputs, *this}.has(2).not_broadcasted();
+    return inputs.at(1);
+}
+
+argument miopen_sigmoid::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/softmax.cpp

@@ -5,7 +5,7 @@
 #include <utility>
 
 namespace migraphx {
-inline namespace MIGRAPH_INLINE_NS {
+inline namespace MIGRAPHX_INLINE_NS {
 namespace gpu {
 
 shape miopen_softmax::compute_shape(const std::vector<shape>& inputs) const
@@ -18,7 +18,8 @@ argument miopen_softmax::compute(context& ctx,
                                  const shape& output_shape,
                                  const std::vector<argument>& args) const
 {
-    float alpha = 1, beta = 0;
+    float alpha = 1;
+    float beta  = 0;
     auto x_desc = make_tensor(args[0].get_shape());
     auto y_desc = make_tensor(output_shape);
     miopenSoftmaxForward(ctx.get_stream().get_miopen(),
@@ -33,5 +34,5 @@ argument miopen_softmax::compute(context& ctx,
 }
 
 } // namespace gpu
-} // namespace MIGRAPH_INLINE_NS
+} // namespace MIGRAPHX_INLINE_NS
 } // namespace migraphx

src/targets/gpu/tanh.cpp

+#include <migraphx/gpu/tanh.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_tanh::compute_shape(const std::vector<shape>& inputs) const
+{
+    check_shapes{inputs, *this}.has(2).not_broadcasted();
+    return inputs.at(1);
+}
+
+argument miopen_tanh::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/target.cpp

@@ -15,11 +15,12 @@
 #include <migraphx/eliminate_contiguous.hpp>
 #include <migraphx/common_subexpression_elimination.hpp>
 #include <migraphx/fwd_conv_batchnorm_rewrite.hpp>
+#include <migraphx/rewrite_rnn.hpp>
 #include <migraphx/eliminate_concat.hpp>
 #include <migraphx/gpu/concat_gpu_opt.hpp>
 
 namespace migraphx {
-inline namespace MIGRAPH_INLINE_NS {
+inline namespace MIGRAPHX_INLINE_NS {
 namespace gpu {
 
 std::vector<pass> target::get_passes(migraphx::context& gctx) const
@@ -31,14 +32,16 @@ std::vector<pass> target::get_passes(migraphx::context& gctx) const
         dead_code_elimination{},
         fwd_conv_batchnorm_rewrite{},
         dead_code_elimination{},
-        common_subexpression_elimination{},
+        rewrite_rnn{},
         dead_code_elimination{},
+        //common_subexpression_elimination{},
+        //dead_code_elimination{},
         simplify_algebra{},
         dead_code_elimination{},
         constant_propagate{},
         dead_code_elimination{},
         auto_contiguous{},
-        simplify_reshapes{},
+        //simplify_reshapes{},
         dead_code_elimination{},
         lowering{ctx},
         eliminate_concat{concat_gpu_optimization{}},
@@ -61,5 +64,5 @@ std::string target::name() const { return "miopen"; }
 
 migraphx::context target::get_context() const { return context{}; }
 } // namespace gpu
-} // namespace MIGRAPH_INLINE_NS
+} // namespace MIGRAPHX_INLINE_NS
 } // namespace migraphx

src/targets/gpu/write_literals.cpp

@@ -5,10 +5,10 @@
 #include <migraphx/env.hpp>
 
 namespace migraphx {
-inline namespace MIGRAPH_INLINE_NS {
+inline namespace MIGRAPHX_INLINE_NS {
 namespace gpu {
 
-MIGRAPH_DECLARE_ENV_VAR(MIGRAPH_COPY_LITERALS)
+MIGRAPHX_DECLARE_ENV_VAR(MIGRAPHX_COPY_LITERALS)
 
 struct hip_load_literal
 {
@@ -33,7 +33,7 @@ void write_literals::apply(program& p) const
     {
         if(ins->name() == "@literal")
         {
-            if(enabled(MIGRAPH_COPY_LITERALS{}))
+            if(enabled(MIGRAPHX_COPY_LITERALS{}))
             {
                 literal l  = ins->get_literal();
                 auto pre   = p.add_literal(l);
@@ -53,5 +53,5 @@ void write_literals::apply(program& p) const
 }
 
 } // namespace gpu
-} // namespace MIGRAPH_INLINE_NS
+} // namespace MIGRAPHX_INLINE_NS
 } // namespace migraphx

test/CMakeLists.txt

@@ -10,12 +10,12 @@ set(CTEST_PARALLEL_LEVEL ${N} CACHE STRING "CTest parallel level")
 add_custom_target(check COMMAND ${CMAKE_CTEST_COMMAND} --output-on-failure -j ${CTEST_PARALLEL_LEVEL} -C ${CMAKE_CFG_INTDIR} --timeout 1500)
 add_custom_target(tests)
 
-find_program(MIGRAPH_GDB gdb)
+find_program(MIGRAPHX_GDB gdb)
 
-if(MIGRAPH_GDB)
-    set(MIGRAPH_TEST_GDB On CACHE BOOL "")
+if(MIGRAPHX_GDB)
+    set(MIGRAPHX_TEST_GDB On CACHE BOOL "")
 else()
-    set(MIGRAPH_TEST_GDB Off CACHE BOOL "")
+    set(MIGRAPHX_TEST_GDB Off CACHE BOOL "")
 endif()
 
 set(SKIP_TESTS)
@@ -34,8 +34,8 @@ function(add_test_command NAME EXE)
                     %1 ${ARGN}")
         add_test(NAME ${NAME} COMMAND ${WINE_CMD} cmd /c "${CMAKE_CURRENT_BINARY_DIR}/test_${NAME}.cmd" $<TARGET_FILE:${EXE}>)
     else()
-        if(MIGRAPH_TEST_GDB)
-            # add_test(NAME ${NAME} COMMAND ${MIGRAPH_GDB} 
+        if(MIGRAPHX_TEST_GDB)
+            # add_test(NAME ${NAME} COMMAND ${MIGRAPHX_GDB} 
             #     --batch
             #     --return-child-result
             #     -ex "set disable-randomization off"
@@ -54,7 +54,7 @@ function(add_test_command NAME EXE)
                     if(EXISTS ${TEST_DIR}/core)
                         set(\$ENV{UBSAN_OPTIONS} print_stacktrace=1)
                         set(\$ENV{ASAN_OPTIONS} print_stacktrace=1)
-                        execute_process(COMMAND ${MIGRAPH_GDB} $<TARGET_FILE:${EXE}> ${TEST_DIR}/core -batch -ex bt)
+                        execute_process(COMMAND ${MIGRAPHX_GDB} $<TARGET_FILE:${EXE}> ${TEST_DIR}/core -batch -ex bt)
                     endif()
                     message(FATAL_ERROR \"Test failed\")
                 endif()
@@ -95,7 +95,7 @@ foreach(TEST ${TESTS})
     rocm_clang_tidy_check(test_${BASE_NAME})
 endforeach()
 
-if(MIGRAPH_ENABLE_GPU)
+if(MIGRAPHX_ENABLE_GPU)
     # gpu tests
     file(GLOB GPU_TESTS gpu/*.cpp)
 
@@ -112,7 +112,8 @@ if(MIGRAPH_ENABLE_GPU)
 endif()
 
 # Onnx test
-add_executable(test_onnx onnx/onnx_test.cpp)
+set(TEST_ONNX_DIR ${CMAKE_CURRENT_SOURCE_DIR}/onnx)
+add_executable(test_onnx ${TEST_ONNX_DIR}/onnx_test.cpp)
 rocm_clang_tidy_check(test_onnx)
 target_link_libraries(test_onnx migraphx_onnx)
 target_include_directories(test_onnx PUBLIC include)
@@ -120,6 +121,10 @@ add_test(NAME test_onnx COMMAND $<TARGET_FILE:test_onnx> WORKING_DIRECTORY ${CMA
 add_dependencies(tests test_onnx)
 add_dependencies(check test_onnx)
 
+if(MIGRAPHX_ENABLE_PYTHON)
+add_subdirectory(py)
+endif()
+
 function(test_header NAME HEADER)
 
     file(WRITE ${CMAKE_CURRENT_BINARY_DIR}/header-main-include-${NAME}.cpp 
@@ -142,7 +147,7 @@ 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(MIGRAPH_ENABLE_GPU)
+        if(MIGRAPHX_ENABLE_GPU)
             target_link_libraries(header_${TEST_NAME} migraphx_gpu)
         endif()
     endforeach()
@@ -150,6 +155,6 @@ endfunction()
 
 test_headers(migraphx ${CMAKE_SOURCE_DIR}/src/include/migraphx/*.hpp)
 test_headers(migraphx/cpu ${CMAKE_SOURCE_DIR}/src/targets/cpu/include/migraphx/cpu/*.hpp)
-if(MIGRAPH_ENABLE_GPU)
+if(MIGRAPHX_ENABLE_GPU)
 test_headers(migraphx/gpu ${CMAKE_SOURCE_DIR}/src/targets/gpu/include/migraphx/gpu/*.hpp)
 endif()