Merge branch 'develop' into batch-concat

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

+#ifndef MIGRAPHX_GUARD_RTGLIB_DEVICE_CLIP_HPP
+#define MIGRAPHX_GUARD_RTGLIB_DEVICE_CLIP_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 clip(hipStream_t stream, const argument& result, const argument& arg1, float max, float min);
+
+} // namespace device
+} // namespace gpu
+} // namespace MIGRAPHX_INLINE_NS
+} // namespace migraphx
+
+#endif

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

+
+#ifndef MIGRAPHX_GUARD_RTGLIB_DEVICE_CONVERT_HPP
+#define MIGRAPHX_GUARD_RTGLIB_DEVICE_CONVERT_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 convert(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/softmax.hpp

+#ifndef MIGRAPHX_GUARD_RTGLIB_DEVICE_SOFTMAX_HPP
+#define MIGRAPHX_GUARD_RTGLIB_DEVICE_SOFTMAX_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 softmax(hipStream_t stream,
+                 const migraphx::shape& output_shape,
+                 std::vector<migraphx::argument> args,
+                 int axis);
+
+} // namespace device
+} // namespace gpu
+} // namespace MIGRAPHX_INLINE_NS
+} // namespace migraphx
+
+#endif

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

@@ -13,11 +13,21 @@ struct context;
 struct miopen_elu
 {
     shared<activation_descriptor> ad;
+
+    template <class Self, class F>
+    static auto reflect(Self& self, F f)
+    {
+        return gpu::reflect(self.ad.get(), f);
+    }
+
     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; }
+    std::ptrdiff_t output_alias(const std::vector<shape>& shapes) const
+    {
+        return shapes.size() - 1;
+    }
 };
 
 } // namespace gpu

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

@@ -2,6 +2,7 @@
 #define MIGRAPHX_GUARD_RTGLIB_GATHER_HPP
 
 #include <migraphx/shape.hpp>
+#include <migraphx/op/gather.hpp>
 #include <migraphx/gpu/miopen.hpp>
 
 namespace migraphx {
@@ -13,11 +14,21 @@ struct context;
 struct hip_gather
 {
     op::gather op;
+
+    template <class Self, class F>
+    static auto reflect(Self& self, F f)
+    {
+        return migraphx::reflect(self.op, f);
+    }
+
     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; }
+    std::ptrdiff_t output_alias(const std::vector<shape>& shapes) const
+    {
+        return shapes.size() - 1;
+    }
 };
 
 } // namespace gpu

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

@@ -2,7 +2,7 @@
 #define MIGRAPHX_GUARD_RTGLIB_GEMM_HPP
 
 #include <migraphx/shape.hpp>
-#include <migraphx/operators.hpp>
+#include <migraphx/op/dot.hpp>
 
 namespace migraphx {
 inline namespace MIGRAPHX_INLINE_NS {
@@ -13,11 +13,21 @@ struct context;
 struct miopen_gemm
 {
     op::dot op;
+
+    template <class Self, class F>
+    static auto reflect(Self& self, F f)
+    {
+        return migraphx::reflect(self.op, f);
+    }
+
     std::string name() const { return "gpu::gemm"; }
     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; }
+    std::ptrdiff_t output_alias(const std::vector<shape>& shapes) const
+    {
+        return shapes.size() - 1;
+    }
 };
 
 } // namespace gpu

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

 #ifndef MIGRAPHX_GUARD_MIGRAPHLIB_HIP_HPP
 #define MIGRAPHX_GUARD_MIGRAPHLIB_HIP_HPP
 
-#include <migraphx/operators.hpp>
 #include <migraphx/config.hpp>
+#include <migraphx/argument.hpp>
+#include <migraphx/check_shapes.hpp>
 #include <utility>
 
 namespace migraphx {
 inline namespace MIGRAPHX_INLINE_NS {
 namespace gpu {
 
+struct context;
+
 argument allocate_gpu(const shape& s, bool host = false);
 
 argument to_gpu(const argument& arg, bool host = false);
@@ -25,6 +28,13 @@ struct hip_allocate
 {
     shape s;
     std::string tag{};
+
+    template <class Self, class F>
+    static auto reflect(Self& self, F f)
+    {
+        return pack(f(self.s, "shape"), f(self.tag, "tag"));
+    }
+
     std::string name() const { return "hip::allocate"; }
     shape compute_shape(const std::vector<shape>& inputs) const
     {
@@ -40,6 +50,13 @@ struct hip_allocate
 struct hip_sync
 {
     std::string tag{};
+
+    template <class Self, class F>
+    static auto reflect(Self& self, F f)
+    {
+        return pack(f(self.tag, "tag"));
+    }
+
     std::string name() const { return "hip::sync"; }
     shape compute_shape(const std::vector<shape>& inputs) const
     {
@@ -70,7 +87,7 @@ struct hip_write
     {
         return to_gpu(args.front());
     }
-    int output_alias(const std::vector<shape>&) const { return 0; }
+    std::ptrdiff_t output_alias(const std::vector<shape>&) const { return 0; }
 };
 
 struct hip_copy
@@ -86,7 +103,7 @@ struct hip_copy
         copy_to_gpu(args[0], args[1]);
         return args[1];
     }
-    int output_alias(const std::vector<shape>&) const { return 1; }
+    std::ptrdiff_t output_alias(const std::vector<shape>&) const { return 1; }
 };
 
 } // namespace gpu

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

@@ -13,11 +13,21 @@ struct context;
 struct miopen_leaky_relu
 {
     shared<activation_descriptor> ad;
+
+    template <class Self, class F>
+    static auto reflect(Self& self, F f)
+    {
+        return gpu::reflect(self.ad.get(), f);
+    }
+
     std::string name() const { return "gpu::leaky_relu"; }
     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; }
+    std::ptrdiff_t output_alias(const std::vector<shape>& shapes) const
+    {
+        return shapes.size() - 1;
+    }
 };
 
 } // namespace gpu

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

@@ -4,7 +4,7 @@
 #include <migraphx/gpu/lowering.hpp>
 #include <migraphx/manage_ptr.hpp>
 #include <migraphx/instruction.hpp>
-#include <migraphx/operators.hpp>
+#include <migraphx/op/logsoftmax.hpp>
 #include <migraphx/generate.hpp>
 #include <migraphx/shape_for_each.hpp>
 #include <migraphx/config.hpp>
@@ -25,11 +25,21 @@ namespace gpu {
 struct hip_logsoftmax
 {
     op::logsoftmax op;
+
+    template <class Self, class F>
+    static auto reflect(Self& self, F f)
+    {
+        return migraphx::reflect(self.op, f);
+    }
+
     std::string name() const { return "gpu::logsoftmax"; }
     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; }
+    std::ptrdiff_t output_alias(const std::vector<shape>& shapes) const
+    {
+        return shapes.size() - 1;
+    }
 };
 
 } // namespace gpu

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

@@ -13,11 +13,21 @@ struct context;
 struct miopen_lrn
 {
     shared<lrn_descriptor> ldesc;
+
+    template <class Self, class F>
+    static auto reflect(Self& self, F f)
+    {
+        return gpu::reflect(self.ldesc.get(), f);
+    }
+
     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; }
+    std::ptrdiff_t output_alias(const std::vector<shape>& shapes) const
+    {
+        return shapes.size() - 1;
+    }
 };
 
 } // namespace gpu

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

@@ -2,7 +2,9 @@
 #define MIGRAPHX_GUARD_MIGRAPHLIB_MIOPEN_HPP
 
 #include <migraphx/manage_ptr.hpp>
-#include <migraphx/operators.hpp>
+#include <migraphx/op/convolution.hpp>
+#include <migraphx/op/pooling.hpp>
+#include <migraphx/op/lrn.hpp>
 #include <miopen/miopen.h>
 #include <migraphx/config.hpp>
 
@@ -160,6 +162,38 @@ inline fused_operator_args make_fused_args()
     return make_obj<fused_operator_args>(&miopenCreateOperatorArgs);
 }
 
+template <class F>
+auto reflect(miopenActivationDescriptor_t ad, F f)
+{
+    assert(ad != nullptr);
+    miopenActivationMode_t mode = miopenActivationPASTHRU;
+    double alpha                = 0.0;
+    double beta                 = 0.0;
+    double gamma                = 0.0;
+    miopenGetActivationDescriptor(ad, &mode, &alpha, &beta, &gamma);
+    return pack(f(std::move(mode), "mode"),    // NOLINT
+                f(std::move(alpha), "alpha"),  // NOLINT
+                f(std::move(beta), "beta"),    // NOLINT
+                f(std::move(gamma), "gamma")); // NOLINT
+}
+
+template <class F>
+auto reflect(miopenLRNDescriptor_t lrnd, F f)
+{
+    assert(lrnd != nullptr);
+    miopenLRNMode_t mode = miopenLRNWithinChannel;
+    unsigned int n       = 0;
+    double alpha         = 0.0;
+    double beta          = 0.0;
+    double k             = 0.0;
+    miopenGetLRNDescriptor(lrnd, &mode, &n, &alpha, &beta, &k);
+    return pack(f(std::move(mode), "mode"),   // NOLINT
+                f(std::move(n), "n"),         // NOLINT
+                f(std::move(alpha), "alpha"), // NOLINT
+                f(std::move(beta), "beta"),   // NOLINT
+                f(std::move(k), "k"));        // NOLINT
+}
+
 } // namespace gpu
 } // namespace MIGRAPHX_INLINE_NS
 } // namespace migraphx

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

@@ -45,7 +45,15 @@ struct unary_device : oper<Derived>
     shape compute_shape(const std::vector<shape>& inputs) const
     {
         check_shapes{inputs, *this}.has(2);
-        return inputs.at(0);
+        auto s = inputs.at(0);
+        if(s.packed())
+        {
+            return s;
+        }
+        else
+        {
+            return {s.type(), s.lens()};
+        }
     }
 
     argument compute(context& ctx, const shape&, const std::vector<argument>& args) const
@@ -54,7 +62,10 @@ struct unary_device : oper<Derived>
         return args[1];
     }
 
-    int output_alias(const std::vector<shape>& shapes) const { return shapes.size() - 1; }
+    std::ptrdiff_t 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&)>
@@ -63,7 +74,16 @@ struct binary_device : oper<Derived>
     shape compute_shape(const std::vector<shape>& inputs) const
     {
         check_shapes{inputs, *this}.has(3);
-        return inputs.at(0);
+        auto s0 = inputs.at(0);
+        auto s1 = inputs.at(1);
+        if(s0 == s1 and s0.packed())
+        {
+            return s0;
+        }
+        else
+        {
+            return {s0.type(), s0.lens()};
+        }
     }
 
     argument compute(context& ctx, const shape&, const std::vector<argument>& args) const
@@ -72,7 +92,10 @@ struct binary_device : oper<Derived>
         return args[2];
     }
 
-    int output_alias(const std::vector<shape>& shapes) const { return shapes.size() - 1; }
+    std::ptrdiff_t output_alias(const std::vector<shape>& shapes) const
+    {
+        return shapes.size() - 1;
+    }
 };
 
 } // namespace gpu

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

@@ -2,7 +2,7 @@
 #define MIGRAPHX_GUARD_RTGLIB_PAD_HPP
 
 #include <migraphx/shape.hpp>
-#include <migraphx/operators.hpp>
+#include <migraphx/op/pad.hpp>
 
 namespace migraphx {
 inline namespace MIGRAPHX_INLINE_NS {
@@ -14,11 +14,20 @@ struct hip_pad
 {
     op::pad op;
 
+    template <class Self, class F>
+    static auto reflect(Self& self, F f)
+    {
+        return migraphx::reflect(self.op, f);
+    }
+
     std::string name() const { return "gpu::pad"; }
     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; }
+    std::ptrdiff_t output_alias(const std::vector<shape>& shapes) const
+    {
+        return shapes.size() - 1;
+    }
 };
 
 } // namespace gpu

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

@@ -2,7 +2,7 @@
 #define MIGRAPHX_GUARD_RTGLIB_POOLING_HPP
 
 #include <migraphx/shape.hpp>
-#include <migraphx/operators.hpp>
+#include <migraphx/op/pooling.hpp>
 #include <migraphx/gpu/miopen.hpp>
 
 namespace migraphx {
@@ -16,11 +16,20 @@ struct miopen_pooling
     op::pooling op;
     shared<pooling_descriptor> pd;
 
+    template <class Self, class F>
+    static auto reflect(Self& self, F f)
+    {
+        return migraphx::reflect(self.op, f);
+    }
+
     std::string name() const { return "gpu::pooling"; }
     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; }
+    std::ptrdiff_t output_alias(const std::vector<shape>& shapes) const
+    {
+        return shapes.size() - 1;
+    }
 };
 
 } // namespace gpu

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

@@ -13,11 +13,21 @@ struct context;
 struct miopen_relu
 {
     shared<activation_descriptor> ad;
+
+    template <class Self, class F>
+    static auto reflect(Self& self, F f)
+    {
+        return gpu::reflect(self.ad.get(), f);
+    }
+
     std::string name() const { return "gpu::relu"; }
     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; }
+    std::ptrdiff_t output_alias(const std::vector<shape>& shapes) const
+    {
+        return shapes.size() - 1;
+    }
 };
 
 } // namespace gpu

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

@@ -13,11 +13,21 @@ struct context;
 struct miopen_sigmoid
 {
     shared<activation_descriptor> ad;
+
+    template <class Self, class F>
+    static auto reflect(Self& self, F f)
+    {
+        return gpu::reflect(self.ad.get(), f);
+    }
+
     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; }
+    std::ptrdiff_t output_alias(const std::vector<shape>& shapes) const
+    {
+        return shapes.size() - 1;
+    }
 };
 
 } // namespace gpu

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

 #ifndef MIGRAPHX_GUARD_RTGLIB_SOFTMAX_HPP
 #define MIGRAPHX_GUARD_RTGLIB_SOFTMAX_HPP
 
-#include <migraphx/shape.hpp>
-#include <migraphx/operators.hpp>
+#include <migraphx/gpu/lowering.hpp>
+#include <migraphx/manage_ptr.hpp>
+#include <migraphx/instruction.hpp>
+#include <migraphx/op/softmax.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 {
@@ -13,11 +27,41 @@ struct context;
 struct miopen_softmax
 {
     op::softmax op;
+
+    template <class Self, class F>
+    static auto reflect(Self& self, F f)
+    {
+        return migraphx::reflect(self.op, f);
+    }
+
+    std::string name() const { return "gpu::softmax"; }
+    shape compute_shape(const std::vector<shape>& inputs) const;
+    argument
+    compute(context& ctx, const shape& output_shape, const std::vector<argument>& args) const;
+    std::ptrdiff_t output_alias(const std::vector<shape>& shapes) const
+    {
+        return shapes.size() - 1;
+    }
+};
+
+struct hip_softmax
+{
+    op::softmax op;
+
+    template <class Self, class F>
+    static auto reflect(Self& self, F f)
+    {
+        return migraphx::reflect(self.op, f);
+    }
+
     std::string name() const { return "gpu::softmax"; }
     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; }
+    std::ptrdiff_t output_alias(const std::vector<shape>& shapes) const
+    {
+        return shapes.size() - 1;
+    }
 };
 
 } // namespace gpu

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

@@ -13,11 +13,21 @@ struct context;
 struct miopen_tanh
 {
     shared<activation_descriptor> ad;
+
+    template <class Self, class F>
+    static auto reflect(Self& self, F f)
+    {
+        return gpu::reflect(self.ad.get(), f);
+    }
+
     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; }
+    std::ptrdiff_t output_alias(const std::vector<shape>& shapes) const
+    {
+        return shapes.size() - 1;
+    }
 };
 
 } // namespace gpu

src/targets/gpu/logsoftmax.cpp

 #include <migraphx/gpu/logsoftmax.hpp>
 #include <migraphx/gpu/device/logsoftmax.hpp>
-#include <migraphx/operators.hpp>
+#include <migraphx/op/logsoftmax.hpp>
 #include <migraphx/manage_ptr.hpp>
 #include <migraphx/gpu/miopen.hpp>
 #include <utility>

src/targets/gpu/lowering.cpp

@@ -45,6 +45,8 @@
 #include <migraphx/gpu/pad.hpp>
 #include <migraphx/gpu/gather.hpp>
 #include <migraphx/gpu/lrn.hpp>
+#include <migraphx/gpu/convert.hpp>
+#include <migraphx/gpu/clip.hpp>
 #include <utility>
 #include <functional>
 #include <algorithm>
@@ -97,10 +99,12 @@ struct miopen_apply
         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_softmax, op::softmax>("softmax");
         add_extend_op<hip_logsoftmax, op::logsoftmax>("logsoftmax");
         add_extend_op<hip_gather, op::gather>("gather");
         add_extend_op<hip_pad, op::pad>("pad");
+        add_extend_op<hip_convert, op::convert>("convert");
+        add_extend_op<hip_clip, op::clip>("clip");
 
         add_lrn_op();
         add_convolution_op();