Merge pull request #135 from ROCmSoftwarePlatform/combineOperators

Combine operators

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

+#ifndef MIGRAPHX_GUARD_RTGLIB_MAX_HPP
+#define MIGRAPHX_GUARD_RTGLIB_MAX_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/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/max.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_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/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/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/min.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_min : binary_device<hip_min, device::min>
+{
+};
+
+} // namespace gpu
+} // namespace MIGRAPHX_INLINE_NS
+} // namespace migraphx
+
+#endif

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

@@ -91,6 +91,29 @@ 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 +121,13 @@ 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

@@ -2,6 +2,7 @@
 #define MIGRAPHX_GUARD_RTGLIB_MUL_HPP
 
 #include <migraphx/gpu/lowering.hpp>
+#include <migraphx/gpu/oper.hpp>
 #include <migraphx/manage_ptr.hpp>
 #include <migraphx/instruction.hpp>
 #include <migraphx/operators.hpp>
@@ -22,12 +23,8 @@ 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

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

+#ifndef MIGRAPHX_GUARD_RTGLIB_UNARY_HPP
+#define MIGRAPHX_GUARD_RTGLIB_UNARY_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/gpu/miopen.hpp>
+#include <migraphx/gpu/hip.hpp>
+#include <migraphx/dfor.hpp>
+#include <migraphx/type_name.hpp>
+#include <migraphx/gpu/device/contiguous.hpp>
+#include <migraphx/iterator_for.hpp>
+#include <migraphx/gpu/rocblas.hpp>
+#include <migraphx/gpu/context.hpp>
+#include <migraphx/config.hpp>
+#include <utility>
+#include <iostream>
+
+namespace migraphx {
+inline namespace MIGRAPHX_INLINE_NS {
+namespace gpu {
+
+template <class Derived>
+struct oper
+{
+    std::string name() const
+    {
+        const std::string& name = get_type_name<Derived>();
+        // search the namespace gpu (::gpu::)
+        auto pos_ns = name.find("::gpu::");
+        if(pos_ns != std::string::npos)
+        {
+            auto pos_name = name.find("hip_", pos_ns + std::string("::gpu::").length());
+            if(pos_name != std::string::npos)
+            {
+                return std::string("gpu::") + name.substr(pos_name + 4);
+            }
+            else
+            {
+                return name.substr(pos_ns + 2);
+            }
+        }
+
+        return "unknown";
+    }
+};
+
+template <class Derived, void (*F)(hipStream_t, const argument&, const argument&)>
+struct unary_device : oper<Derived>
+{
+    shape compute_shape(const std::vector<shape>& inputs) const
+    {
+        check_shapes{inputs, *this}.has(2);
+        return inputs.at(0);
+    }
+
+    argument compute(context& ctx, const shape&, const std::vector<argument>& args) const
+    {
+        F(ctx.get_stream().get(), args[1], args[0]);
+        return args[1];
+    }
+
+    int output_alias(const std::vector<shape>& shapes) const { return shapes.size() - 1; }
+};
+
+template <class Derived, void (*F)(hipStream_t, const argument&, const argument&, const argument&)>
+struct binary_device : oper<Derived>
+{
+    shape compute_shape(const std::vector<shape>& inputs) const
+    {
+        check_shapes{inputs, *this}.has(3);
+        return inputs.at(0);
+    }
+
+    argument compute(context& ctx, const shape&, const std::vector<argument>& args) const
+    {
+        F(ctx.get_stream().get(), args[2], args[1], args[0]);
+        return args[2];
+    }
+
+    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/sigmoid.hpp

+#ifndef MIGRAPHX_GUARD_RTGLIB_SIGMOID_HPP
+#define MIGRAPHX_GUARD_RTGLIB_SIGMOID_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_sigmoid
+{
+    shared<activation_descriptor> ad;
+    std::string name() const { return "gpu::sigmoid"; }
+    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/sin.hpp

@@ -2,6 +2,7 @@
 #define MIGRAPHX_GUARD_RTGLIB_SIN_HPP
 
 #include <migraphx/gpu/lowering.hpp>
+#include <migraphx/gpu/oper.hpp>
 #include <migraphx/manage_ptr.hpp>
 #include <migraphx/instruction.hpp>
 #include <migraphx/operators.hpp>
@@ -22,12 +23,8 @@ namespace migraphx {
 inline namespace MIGRAPHX_INLINE_NS {
 namespace gpu {
 
-struct hip_sin
+struct hip_sin : unary_device<hip_sin, device::sin>
 {
-    std::string name() const { return "gpu::sin"; }
-    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

src/targets/gpu/mul.cpp → src/targets/gpu/include/migraphx/gpu/sinh.hpp

-#include <migraphx/gpu/mul.hpp>
-#include <migraphx/operators.hpp>
+#ifndef MIGRAPHX_GUARD_RTGLIB_SINH_HPP
+#define MIGRAPHX_GUARD_RTGLIB_SINH_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/sinh.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 {
 
-shape hip_mul::compute_shape(const std::vector<shape>& inputs) const
+struct hip_sinh : unary_device<hip_sinh, device::sinh>
 {
-    // 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 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/lowering.cpp

@@ -16,16 +16,32 @@
 #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/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 <utility>
+#include <functional>
 
 namespace migraphx {
 inline namespace MIGRAPHX_INLINE_NS {
@@ -35,6 +51,7 @@ struct miopen_apply
 {
     program* prog = nullptr;
     context ctx{};
+    std::unordered_map<std::string, std::function<instruction_ref(instruction_ref)>> apply_map{};
 
     void check_shape(shape x, instruction_ref i)
     {
@@ -43,58 +60,50 @@ struct miopen_apply
         (void)i;
     }
 
+    void init()
+    {
+        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_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_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")
-            {
-                check_shape(s, apply_contiguous(it));
-            }
-            else if(it->name() == "concat")
+            if(apply_map.count(it->name()) > 0)
             {
-                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));
             }
         }
     }
@@ -113,120 +122,103 @@ struct miopen_apply
         }
     }
 
-    instruction_ref apply_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 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);
-    }
-
-    instruction_ref apply_pooling(instruction_ref ins)
+    void add_convolution_op()
     {
-        auto&& op   = any_cast<op::pooling>(ins->get_operator());
-        auto pd     = make_pooling(op);
-        auto output = insert_allocation(ins, ins->get_shape());
+        apply_map.emplace("convolution", [=](instruction_ref ins) {
+            auto&& op = any_cast<op::convolution>(ins->get_operator());
 
-        return prog->replace_instruction(
-            ins, miopen_pooling{op, std::move(pd)}, ins->inputs().at(0), output);
-    }
-
-    instruction_ref apply_relu(instruction_ref ins)
-    {
-        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);
-    }
+            auto conv = miopen_convolution{op, make_conv(op)};
+            auto ws   = conv.compile(ctx, ins->get_shape(), ins->inputs());
 
-    instruction_ref apply_leaky_relu(instruction_ref ins)
-    {
-        auto&& op = any_cast<op::leaky_relu>(ins->get_operator());
-        auto ad   = make_leaky_relu(op.alpha);
+            auto workspace = insert_allocation(ins, ws, "workspace");
+            auto output    = insert_allocation(ins, ins->get_shape());
 
-        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, conv, ins->inputs().at(0), ins->inputs().at(1), workspace, output);
+        });
     }
 
-    instruction_ref apply_softmax(instruction_ref ins)
+    void add_pooling_op()
     {
-        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("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_add(instruction_ref ins)
+    template <class T>
+    void add_generic_op(std::string name)
     {
-        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 output                       = insert_allocation(ins, ins->get_shape());
+            std::vector<instruction_ref> refs = ins->inputs();
+            refs.push_back(output);
 
-    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);
+            return prog->replace_instruction(ins, T{}, refs);
+        });
     }
 
-    instruction_ref apply_mul(instruction_ref ins)
+    template <class T, class Op>
+    void add_extend_op(std::string name)
     {
-        auto output = insert_allocation(ins, ins->get_shape());
-        return prog->replace_instruction(
-            ins, hip_mul{}, 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 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_gemm(instruction_ref ins)
+    template <class T, class Op, class F>
+    void add_miopen_extend_op(std::string name, F f)
     {
-        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);
-    }
+        apply_map.emplace(name, [=](instruction_ref ins) {
+            auto&& op = any_cast<Op>(ins->get_operator());
+            auto ad   = f(op.alpha);
 
-    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);
+        });
     }
 };
 

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, 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/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, 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

test/cpu_ops_test.cpp

@@ -7,6 +7,10 @@
 #include <migraphx/verify.hpp>
 #include "test.hpp"
 
+float sigmoid(float x) { return 1 / (1 + expf(-x)); }
+
+float elu(float a, float x) { return x > 0 ? x : a * std::expm1(x); }
+
 TEST_CASE(slice_test)
 {
     {
@@ -409,6 +413,20 @@ TEST_CASE(exp_test)
     EXPECT(migraphx::verify_range(results_vector, gold));
 }
 
+TEST_CASE(log_test)
+{
+    migraphx::program p;
+    migraphx::shape s{migraphx::shape::float_type, {3}};
+    auto l = p.add_literal(migraphx::literal{s, {1, 2, 3}});
+    p.add_instruction(migraphx::op::log{}, l);
+    p.compile(migraphx::cpu::target{});
+    auto result = p.eval({});
+    std::vector<float> results_vector(3);
+    result.visit([&](auto output) { results_vector.assign(output.begin(), output.end()); });
+    std::vector<float> gold = {0.0f, 0.6931471806f, 1.0986122887f};
+    EXPECT(migraphx::verify_range(results_vector, gold));
+}
+
 TEST_CASE(sin_test)
 {
     migraphx::program p;
@@ -451,6 +469,50 @@ TEST_CASE(tan_test)
     EXPECT(migraphx::verify_range(results_vector, gold));
 }
 
+TEST_CASE(asin_test)
+{
+    migraphx::program p;
+    migraphx::shape s{migraphx::shape::float_type, {3}};
+    std::vector<float> data{-0.5f, 0.0f, 0.9f};
+    auto l = p.add_literal(migraphx::literal{s, data});
+    p.add_instruction(migraphx::op::asin{}, l);
+    p.compile(migraphx::cpu::target{});
+    auto result = p.eval({});
+    std::vector<float> results_vector(3);
+    result.visit([&](auto output) { results_vector.assign(output.begin(), output.end()); });
+    std::vector<float> gold = {-0.5235987756f, 0.f, 1.119769515};
+    EXPECT(migraphx::verify_range(results_vector, gold));
+}
+
+TEST_CASE(acos_test)
+{
+    migraphx::program p;
+    migraphx::shape s{migraphx::shape::double_type, {3}};
+    std::vector<float> data{-0.8f, 0.0f, 1.0f};
+    auto l = p.add_literal(migraphx::literal{s, data});
+    p.add_instruction(migraphx::op::acos{}, l);
+    p.compile(migraphx::cpu::target{});
+    auto result = p.eval({});
+    std::vector<float> results_vector(3);
+    result.visit([&](auto output) { results_vector.assign(output.begin(), output.end()); });
+    std::vector<float> gold = {2.4980915448f, 1.5707963268f, 0.0f};
+    EXPECT(migraphx::verify_range(results_vector, gold));
+}
+
+TEST_CASE(atan_test)
+{
+    migraphx::program p;
+    migraphx::shape s{migraphx::shape::double_type, {3}};
+    auto l = p.add_literal(migraphx::literal{s, {-1, 0, 1}});
+    p.add_instruction(migraphx::op::atan{}, l);
+    p.compile(migraphx::cpu::target{});
+    auto result = p.eval({});
+    std::vector<float> results_vector(3);
+    result.visit([&](auto output) { results_vector.assign(output.begin(), output.end()); });
+    std::vector<float> gold = {-0.7853981634f, 0.0f, 0.7853981634f};
+    EXPECT(migraphx::verify_range(results_vector, gold));
+}
+
 TEST_CASE(add_test)
 {
     migraphx::program p;
@@ -1105,4 +1167,123 @@ TEST_CASE(identity_test)
     EXPECT(std::equal(data.begin(), data.end(), results_vector.begin()));
 }
 
+TEST_CASE(abs_test)
+{
+    migraphx::program p;
+    migraphx::shape s{migraphx::shape::float_type, {2, 2}};
+    auto l = p.add_literal(migraphx::literal{s, {-1, 2, -3, 4}});
+    p.add_instruction(migraphx::op::abs{}, l);
+    p.compile(migraphx::cpu::target{});
+    auto result = p.eval({});
+    std::vector<float> results_vector(4);
+    result.visit([&](auto output) { results_vector.assign(output.begin(), output.end()); });
+    std::vector<float> gold{1, 2, 3, 4};
+    EXPECT(migraphx::verify_range(results_vector, gold));
+}
+
+TEST_CASE(sigmoid_test)
+{
+    migraphx::program p;
+    migraphx::shape s{migraphx::shape::float_type, {2, 2}};
+    auto l = p.add_literal(migraphx::literal{s, {-1, 2, -3, 4}});
+    p.add_instruction(migraphx::op::sigmoid{}, l);
+    p.compile(migraphx::cpu::target{});
+    auto result = p.eval({});
+    std::vector<float> results_vector(4);
+    result.visit([&](auto output) { results_vector.assign(output.begin(), output.end()); });
+    std::vector<float> gold{sigmoid(-1), sigmoid(2), sigmoid(-3), sigmoid(4)};
+    EXPECT(migraphx::verify_range(results_vector, gold));
+}
+
+TEST_CASE(sinh_test)
+{
+    migraphx::program p;
+    migraphx::shape s{migraphx::shape::float_type, {2, 2}};
+    auto l = p.add_literal(migraphx::literal{s, {-1.0, 2.0, -3.0, 4.0}});
+    p.add_instruction(migraphx::op::sinh{}, l);
+    p.compile(migraphx::cpu::target{});
+    auto result = p.eval({});
+    std::vector<float> results_vector(4);
+    result.visit([&](auto output) { results_vector.assign(output.begin(), output.end()); });
+    std::vector<float> gold{sinhf(-1), sinhf(2), sinhf(-3), sinhf(4)};
+    EXPECT(migraphx::verify_range(results_vector, gold));
+}
+
+TEST_CASE(cosh_test)
+{
+    migraphx::program p;
+    migraphx::shape s{migraphx::shape::float_type, {2, 2}};
+    auto l = p.add_literal(migraphx::literal{s, {-1.0, 2.0, -3.0, 4.0}});
+    p.add_instruction(migraphx::op::cosh{}, l);
+    p.compile(migraphx::cpu::target{});
+    auto result = p.eval({});
+    std::vector<float> results_vector(4);
+    result.visit([&](auto output) { results_vector.assign(output.begin(), output.end()); });
+    std::vector<float> gold{coshf(-1), coshf(2), coshf(-3), coshf(4)};
+    EXPECT(migraphx::verify_range(results_vector, gold));
+}
+
+TEST_CASE(tanh_test)
+{
+    migraphx::program p;
+    migraphx::shape s{migraphx::shape::float_type, {2, 2}};
+    auto l = p.add_literal(migraphx::literal{s, {-1.0, 2.0, -3.0, 4.0}});
+    p.add_instruction(migraphx::op::tanh{}, l);
+    p.compile(migraphx::cpu::target{});
+    auto result = p.eval({});
+    std::vector<float> results_vector(4);
+    result.visit([&](auto output) { results_vector.assign(output.begin(), output.end()); });
+    std::vector<float> gold{tanhf(-1), tanhf(2), tanhf(-3), tanhf(4)};
+    EXPECT(migraphx::verify_range(results_vector, gold));
+}
+
+TEST_CASE(elu_test)
+{
+    migraphx::program p;
+    migraphx::shape s{migraphx::shape::float_type, {2, 2}};
+    auto l      = p.add_literal(migraphx::literal{s, {-1.0, 2.0, -3.0, 4.0}});
+    float alpha = 0.5;
+    p.add_instruction(migraphx::op::elu{alpha}, l);
+    p.compile(migraphx::cpu::target{});
+    auto result = p.eval({});
+    std::vector<float> results_vector(4);
+    result.visit([&](auto output) { results_vector.assign(output.begin(), output.end()); });
+    std::vector<float> gold{elu(alpha, -1), elu(alpha, 2), elu(alpha, -3), elu(alpha, 4)};
+    EXPECT(migraphx::verify_range(results_vector, gold));
+}
+
+TEST_CASE(max_test)
+{
+    migraphx::program p;
+    migraphx::shape s{migraphx::shape::float_type, {3}};
+    auto l0       = p.add_literal(migraphx::literal{s, {1, 4, 3}});
+    auto l1       = p.add_literal(migraphx::literal{s, {2, 8, 6}});
+    auto l2       = p.add_literal(migraphx::literal{s, {7, 5, 9}});
+    auto curr_max = p.add_instruction(migraphx::op::max{}, l0, l1);
+    p.add_instruction(migraphx::op::max{}, curr_max, l2);
+    p.compile(migraphx::cpu::target{});
+    auto result = p.eval({});
+    std::vector<float> results_vector(4);
+    result.visit([&](auto output) { results_vector.assign(output.begin(), output.end()); });
+    std::vector<float> gold{7, 8, 9};
+    EXPECT(migraphx::verify_range(results_vector, gold));
+}
+
+TEST_CASE(min_test)
+{
+    migraphx::program p;
+    migraphx::shape s{migraphx::shape::float_type, {3}};
+    auto l0       = p.add_literal(migraphx::literal{s, {1, 4, 3}});
+    auto l1       = p.add_literal(migraphx::literal{s, {2, 8, 6}});
+    auto l2       = p.add_literal(migraphx::literal{s, {7, 5, 9}});
+    auto curr_min = p.add_instruction(migraphx::op::min{}, l0, l1);
+    p.add_instruction(migraphx::op::min{}, curr_min, l2);
+    p.compile(migraphx::cpu::target{});
+    auto result = p.eval({});
+    std::vector<float> results_vector(4);
+    result.visit([&](auto output) { results_vector.assign(output.begin(), output.end()); });
+    std::vector<float> gold{1, 4, 3};
+    EXPECT(migraphx::verify_range(results_vector, gold));
+}
+
 int main(int argc, const char* argv[]) { test::run(argc, argv); }

test/gpu/miopen.cpp

@@ -203,6 +203,32 @@ struct test_mul
     }
 };
 
+struct test_exp
+{
+    migraphx::program create_program() const
+    {
+        migraphx::program p;
+        migraphx::shape s{migraphx::shape::float_type, {6}};
+        std::vector<float> data{0.1f, 0.2f, 1.f, 2.f, 0.6f, 10.f};
+        auto x = p.add_literal(s, data);
+        p.add_instruction(migraphx::op::exp{}, x);
+        return p;
+    }
+};
+
+struct test_log
+{
+    migraphx::program create_program() const
+    {
+        migraphx::program p;
+        migraphx::shape s{migraphx::shape::float_type, {6}};
+        std::vector<float> data{0.1f, 0.2f, 1.f, 2.f, 0.6f, 100.f};
+        auto x = p.add_literal(s, data);
+        p.add_instruction(migraphx::op::log{}, x);
+        return p;
+    }
+};
+
 struct test_sin
 {
     migraphx::program create_program() const
@@ -215,6 +241,101 @@ struct test_sin
     }
 };
 
+struct test_cos
+{
+    migraphx::program create_program() const
+    {
+        migraphx::program p;
+        migraphx::shape s{migraphx::shape::double_type, {8}};
+        auto x = p.add_parameter("x", s);
+        p.add_instruction(migraphx::op::cos{}, x);
+        return p;
+    }
+};
+
+struct test_tan
+{
+    migraphx::program create_program() const
+    {
+        migraphx::program p;
+        migraphx::shape s{migraphx::shape::float_type, {16}};
+        auto x = p.add_parameter("x", s);
+        p.add_instruction(migraphx::op::tan{}, x);
+        return p;
+    }
+};
+
+struct test_sinh
+{
+    migraphx::program create_program() const
+    {
+        migraphx::program p;
+        migraphx::shape s{migraphx::shape::double_type, {16}};
+        auto x = p.add_parameter("x", s);
+        p.add_instruction(migraphx::op::sinh{}, x);
+        return p;
+    }
+};
+
+struct test_cosh
+{
+    migraphx::program create_program() const
+    {
+        migraphx::program p;
+        migraphx::shape s{migraphx::shape::double_type, {16}};
+        auto x = p.add_parameter("x", s);
+        p.add_instruction(migraphx::op::cosh{}, x);
+        return p;
+    }
+};
+
+struct test_tanh
+{
+    migraphx::program create_program() const
+    {
+        migraphx::program p;
+        auto x = p.add_parameter("x", migraphx::shape{migraphx::shape::float_type, {4, 3, 3, 3}});
+        p.add_instruction(migraphx::op::tanh{}, x);
+        return p;
+    }
+};
+
+struct test_asin
+{
+    migraphx::program create_program() const
+    {
+        migraphx::program p;
+        migraphx::shape s{migraphx::shape::double_type, {16}};
+        auto x = p.add_parameter("x", s);
+        p.add_instruction(migraphx::op::asin{}, x);
+        return p;
+    }
+};
+
+struct test_acos
+{
+    migraphx::program create_program() const
+    {
+        migraphx::program p;
+        migraphx::shape s{migraphx::shape::double_type, {16}};
+        auto x = p.add_parameter("x", s);
+        p.add_instruction(migraphx::op::acos{}, x);
+        return p;
+    }
+};
+
+struct test_atan
+{
+    migraphx::program create_program() const
+    {
+        migraphx::program p;
+        migraphx::shape s{migraphx::shape::double_type, {16}};
+        auto x = p.add_parameter("x", s);
+        p.add_instruction(migraphx::op::atan{}, x);
+        return p;
+    }
+};
+
 struct test_scale
 {
     migraphx::program create_program() const
@@ -456,6 +577,28 @@ struct test_add_relu
     }
 };
 
+struct test_sigmoid
+{
+    migraphx::program create_program() const
+    {
+        migraphx::program p;
+        auto x = p.add_parameter("x", migraphx::shape{migraphx::shape::float_type, {4, 3, 3, 3}});
+        p.add_instruction(migraphx::op::sigmoid{}, x);
+        return p;
+    }
+};
+
+struct test_abs
+{
+    migraphx::program create_program() const
+    {
+        migraphx::program p;
+        auto x = p.add_parameter("x", migraphx::shape{migraphx::shape::float_type, {4, 3, 3, 3}});
+        p.add_instruction(migraphx::op::abs{}, x);
+        return p;
+    }
+};
+
 struct test_leaky_relu
 {
     migraphx::program create_program() const
@@ -467,6 +610,17 @@ struct test_leaky_relu
     }
 };
 
+struct test_elu
+{
+    migraphx::program create_program() const
+    {
+        migraphx::program p;
+        auto x = p.add_parameter("x", migraphx::shape{migraphx::shape::float_type, {4, 3, 3, 3}});
+        p.add_instruction(migraphx::op::leaky_relu{1.0}, x);
+        return p;
+    }
+};
+
 struct test_conv_pooling
 {
     migraphx::program create_program() const
@@ -849,13 +1003,24 @@ struct test_conv_bn_relu_pooling2
 
 int main()
 {
+    verify_program<test_abs>();
     verify_program<test_concat>();
     verify_program<test_concat2>();
     verify_program<test_concat_relu>();
     verify_program<test_add>();
     verify_program<test_add_half>();
     verify_program<test_mul>();
+    verify_program<test_exp>();
+    verify_program<test_log>();
     verify_program<test_sin>();
+    verify_program<test_cos>();
+    verify_program<test_tan>();
+    verify_program<test_sinh>();
+    verify_program<test_cosh>();
+    verify_program<test_tanh>();
+    verify_program<test_asin>();
+    verify_program<test_acos>();
+    verify_program<test_atan>();
     verify_program<test_scale>();
     verify_program<test_triadd>();
     verify_program<test_triadd2>();
@@ -873,6 +1038,8 @@ int main()
     verify_program<test_conv_relu_half>();
     verify_program<test_add_relu>();
     verify_program<test_leaky_relu>();
+    verify_program<test_sigmoid>();
+    verify_program<test_elu>();
     verify_program<test_conv_pooling>();
     verify_program<test_global_avg_pooling>();
     verify_program<test_global_max_pooling>();

test/onnx/acos_test.onnx

+acos-example:;
+
+
x
y"Acos	test_acosZ
+
x
+
+

+
+b
+
y
+
+

+
+B	
\ No newline at end of file

test/onnx/asin_test.onnx

+asin-example:;
+
+
x
y"Asin	test_asinZ
+
x
+
+

+
+b
+
y
+
+

+
+B
\ No newline at end of file

test/onnx/atan_test.onnx

+atan-example:;
+
+
x
y"Atan	test_atanZ
+
x
+
+

+
+b
+
y
+
+

+
+B	
\ No newline at end of file

test/onnx/cos_test.onnx

+cos-example:9
+
+
x
y"Costest_cosZ
+
x
+
+

+
+b
+
y
+
+

+
+B
\ No newline at end of file

test/onnx/cosh_test.onnx

+cosh-example:;
+
+
x
y"Cosh	test_coshZ
+
x
+
+

+
b
+
y
+
+

+
B	
\ No newline at end of file