Merge branch 'master' into fp16

src/targets/gpu/device/mul.cpp

@@ -2,6 +2,7 @@
 #include <migraph/gpu/device/nary.hpp>
 
 namespace migraph {
+inline namespace MIGRAPH_INLINE_NS {
 namespace gpu {
 namespace device {
 
@@ -21,4 +22,5 @@ void mul(hipStream_t stream,
 
 } // namespace device
 } // namespace gpu
+} // namespace MIGRAPH_INLINE_NS
 } // namespace migraph

src/targets/gpu/eliminate_workspace.cpp

@@ -9,6 +9,7 @@
 #include <migraph/pass_config.hpp>
 
 namespace migraph {
+inline namespace MIGRAPH_INLINE_NS {
 namespace gpu {
 
 void eliminate_workspace::apply(program& p) const
@@ -38,5 +39,7 @@ void eliminate_workspace::apply(program& p) const
         p.remove_instruction(a);
     }
 }
+
 } // namespace gpu
+} // namespace MIGRAPH_INLINE_NS
 } // namespace migraph

src/targets/gpu/fuse_ops.cpp

@@ -6,7 +6,7 @@
 #include <migraph/instruction.hpp>
 
 namespace migraph {
-
+inline namespace MIGRAPH_INLINE_NS {
 namespace gpu {
 
 struct fusion
@@ -157,6 +157,7 @@ struct hip_triadd
         device::add(ctx.get_stream().get(), args.at(3), args.at(0), args.at(1), args.at(2));
         return args.at(3);
     }
+    int output_alias(const std::vector<shape>& shapes) const { return shapes.size() - 1; }
 };
 
 struct hip_triadd_relu
@@ -172,6 +173,7 @@ struct hip_triadd_relu
         device::add_relu(ctx.get_stream().get(), args.at(3), args.at(0), args.at(1), args.at(2));
         return args.at(3);
     }
+    int output_alias(const std::vector<shape>& shapes) const { return shapes.size() - 1; }
 };
 
 struct hip_add_relu
@@ -187,6 +189,7 @@ struct hip_add_relu
         device::add_relu(ctx.get_stream().get(), args.at(2), args.at(0), args.at(1));
         return args.at(2);
     }
+    int output_alias(const std::vector<shape>& shapes) const { return shapes.size() - 1; }
 };
 
 struct find_add_relu
@@ -273,6 +276,7 @@ struct miopen_conv_bias
         f.compile(ctx);
         return f.get_workspace(ctx);
     }
+    int output_alias(const std::vector<shape>& shapes) const { return shapes.size() - 1; }
 };
 
 struct miopen_conv_bias_relu
@@ -316,6 +320,7 @@ struct miopen_conv_bias_relu
         f.compile(ctx);
         return f.get_workspace(ctx);
     }
+    int output_alias(const std::vector<shape>& shapes) const { return shapes.size() - 1; }
 };
 
 template <class... Ms>
@@ -384,5 +389,5 @@ void fuse_ops::apply(program& p) const
 }
 
 } // namespace gpu
-
+} // namespace MIGRAPH_INLINE_NS
 } // namespace migraph

src/targets/gpu/gemm.cpp

@@ -5,6 +5,7 @@
 #include <utility>
 
 namespace migraph {
+inline namespace MIGRAPH_INLINE_NS {
 namespace gpu {
 
 template <class... Ts>
@@ -82,5 +83,5 @@ argument miopen_gemm::compute(context& ctx,
 }
 
 } // namespace gpu
-
+} // namespace MIGRAPH_INLINE_NS
 } // namespace migraph

src/targets/gpu/hip.cpp

@@ -7,6 +7,7 @@
 #include <vector>
 
 namespace migraph {
+inline namespace MIGRAPH_INLINE_NS {
 namespace gpu {
 
 using hip_ptr = MIGRAPH_MANAGE_PTR(void, hipFree);
@@ -107,5 +108,7 @@ void copy_to_gpu(argument src, argument dst)
     if(status != hipSuccess)
         MIGRAPH_THROW("Copy to gpu failed: " + hip_error(status));
 }
+
 } // namespace gpu
+} // namespace MIGRAPH_INLINE_NS
 } // namespace migraph

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

@@ -15,9 +15,11 @@
 #include <migraph/iterator_for.hpp>
 #include <migraph/gpu/rocblas.hpp>
 #include <migraph/gpu/context.hpp>
+#include <migraph/config.hpp>
 #include <utility>
 
 namespace migraph {
+inline namespace MIGRAPH_INLINE_NS {
 namespace gpu {
 
 struct hip_add
@@ -25,6 +27,7 @@ struct hip_add
     std::string name() const { return "gpu::add"; }
     shape compute_shape(const std::vector<shape>& inputs) const;
     argument compute(context&, const shape&, const std::vector<argument>& args) const;
+    int output_alias(const std::vector<shape>& shapes) const { return shapes.size() - 1; }
 };
 
 struct miopen_add
@@ -33,10 +36,11 @@ struct miopen_add
     shape compute_shape(const std::vector<shape>& inputs) const;
     argument
     compute(context& ctx, const shape& output_shape, const std::vector<argument>& args) const;
+    int output_alias(const std::vector<shape>& shapes) const { return shapes.size() - 1; }
 };
 
 } // namespace gpu
-
+} // namespace MIGRAPH_INLINE_NS
 } // namespace migraph
 
 #endif

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

@@ -15,9 +15,11 @@
 #include <migraph/iterator_for.hpp>
 #include <migraph/gpu/rocblas.hpp>
 #include <migraph/gpu/context.hpp>
+#include <migraph/config.hpp>
 #include <utility>
 
 namespace migraph {
+inline namespace MIGRAPH_INLINE_NS {
 namespace gpu {
 
 struct miopen_batch_norm_inference
@@ -27,10 +29,11 @@ struct miopen_batch_norm_inference
     shape compute_shape(const std::vector<shape>& inputs) const;
     argument
     compute(context& ctx, const shape& output_shape, const std::vector<argument>& args) const;
+    int output_alias(const std::vector<shape>& shapes) const { return shapes.size() - 1; }
 };
 
 } // namespace gpu
-
+} // namespace MIGRAPH_INLINE_NS
 } // namespace migraph
 
 #endif

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

@@ -7,6 +7,7 @@
 #include <migraph/operators.hpp>
 #include <migraph/generate.hpp>
 #include <migraph/shape_for_each.hpp>
+#include <migraph/config.hpp>
 #include <migraph/gpu/miopen.hpp>
 #include <migraph/gpu/hip.hpp>
 #include <migraph/dfor.hpp>
@@ -18,6 +19,7 @@
 #include <utility>
 
 namespace migraph {
+inline namespace MIGRAPH_INLINE_NS {
 namespace gpu {
 
 struct hip_concat
@@ -28,10 +30,11 @@ struct hip_concat
     shape compute_shape(std::vector<shape> inputs) const;
     argument
     compute(context& ctx, const shape& output_shape, const std::vector<argument>& args) const;
+    int output_alias(const std::vector<shape>& shapes) const { return shapes.size() - 1; }
 };
 
 } // namespace gpu
-
+} // namespace MIGRAPH_INLINE_NS
 } // namespace migraph
 
 #endif

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

+#ifndef MIGRAPH_GUARD_RTGLIB_CONCAT_GPU_OPT_HPP
+#define MIGRAPH_GUARD_RTGLIB_CONCAT_GPU_OPT_HPP
+
+#include <migraph/gpu/concat.hpp>
+
+namespace migraph {
+namespace gpu {
+
+struct concat_gpu_optimization
+{
+    std::string name() const { return "gpu::concat"; }
+    std::string allocate() const { return "hip::allocate"; }
+    migraph::op::concat get_concat(const migraph::operation& op) const
+    {
+        return migraph::any_cast<migraph::gpu::hip_concat>(op).op;
+    }
+};
+
+} // namespace gpu
+
+} // namespace migraph
+
+#endif

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

@@ -5,8 +5,10 @@
 #include <migraph/gpu/rocblas.hpp>
 #include <migraph/gpu/hip.hpp>
 #include <migraph/env.hpp>
+#include <migraph/config.hpp>
 
 namespace migraph {
+inline namespace MIGRAPH_INLINE_NS {
 namespace gpu {
 
 MIGRAPH_DECLARE_ENV_VAR(MIGRAPH_DISABLE_NULL_STREAM)
@@ -114,6 +116,7 @@ struct context
     std::shared_ptr<hip_device> current_device;
 };
 } // namespace gpu
+} // namespace MIGRAPH_INLINE_NS
 } // namespace migraph
 
 #endif

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

@@ -7,6 +7,7 @@
 #include <migraph/operators.hpp>
 #include <migraph/generate.hpp>
 #include <migraph/shape_for_each.hpp>
+#include <migraph/config.hpp>
 #include <migraph/gpu/miopen.hpp>
 #include <migraph/gpu/hip.hpp>
 #include <migraph/dfor.hpp>
@@ -18,6 +19,7 @@
 #include <utility>
 
 namespace migraph {
+inline namespace MIGRAPH_INLINE_NS {
 namespace gpu {
 
 struct miopen_contiguous
@@ -26,10 +28,11 @@ struct miopen_contiguous
     std::string name() const { return "gpu::contiguous"; }
     shape compute_shape(const std::vector<shape>& inputs) const;
     argument compute(context&, shape output_shape, const std::vector<argument>& args) const;
+    int output_alias(const std::vector<shape>& shapes) const { return shapes.size() - 1; }
 };
 
 } // namespace gpu
-
+} // namespace MIGRAPH_INLINE_NS
 } // namespace migraph
 
 #endif

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

@@ -7,6 +7,7 @@
 #include <migraph/operators.hpp>
 #include <migraph/generate.hpp>
 #include <migraph/shape_for_each.hpp>
+#include <migraph/config.hpp>
 #include <migraph/gpu/miopen.hpp>
 #include <migraph/gpu/hip.hpp>
 #include <migraph/dfor.hpp>
@@ -18,6 +19,7 @@
 #include <utility>
 
 namespace migraph {
+inline namespace MIGRAPH_INLINE_NS {
 namespace gpu {
 
 struct miopen_convolution
@@ -38,10 +40,11 @@ struct miopen_convolution
     argument
     compute(context& ctx, const shape& output_shape, const std::vector<argument>& args) const;
     shape compile(context& ctx, const shape& output_shape, std::vector<instruction_ref> inputs);
+    int output_alias(const std::vector<shape>& shapes) const { return shapes.size() - 1; }
 };
 
 } // namespace gpu
-
+} // namespace MIGRAPH_INLINE_NS
 } // namespace migraph
 
 #endif

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

@@ -3,9 +3,11 @@
 #define MIGRAPH_GUARD_RTGLIB_DEVICE_ADD_HPP
 
 #include <migraph/argument.hpp>
+#include <migraph/config.hpp>
 #include <hip/hip_runtime_api.h>
 
 namespace migraph {
+inline namespace MIGRAPH_INLINE_NS {
 namespace gpu {
 namespace device {
 
@@ -19,6 +21,7 @@ void add(hipStream_t stream,
 
 } // namespace device
 } // namespace gpu
+} // namespace MIGRAPH_INLINE_NS
 } // namespace migraph
 
 #endif

src/targets/gpu/include/migraph/gpu/device/add_relu.hpp

@@ -3,9 +3,11 @@
 #define MIGRAPH_GUARD_RTGLIB_DEVICE_ADD_RELU_HPP
 
 #include <migraph/argument.hpp>
+#include <migraph/config.hpp>
 #include <hip/hip_runtime_api.h>
 
 namespace migraph {
+inline namespace MIGRAPH_INLINE_NS {
 namespace gpu {
 namespace device {
 
@@ -22,6 +24,7 @@ void add_relu(hipStream_t stream,
 
 } // namespace device
 } // namespace gpu
+} // namespace MIGRAPH_INLINE_NS
 } // namespace migraph
 
 #endif

src/targets/gpu/include/migraph/gpu/device/concat.hpp

@@ -2,9 +2,11 @@
 #define MIGRAPH_GUARD_RTGLIB_DEVICE_CONCAT_HPP
 
 #include <migraph/argument.hpp>
+#include <migraph/config.hpp>
 #include <hip/hip_runtime_api.h>
 
 namespace migraph {
+inline namespace MIGRAPH_INLINE_NS {
 namespace gpu {
 namespace device {
 
@@ -15,6 +17,7 @@ argument concat(hipStream_t stream,
 
 } // namespace device
 } // namespace gpu
+} // namespace MIGRAPH_INLINE_NS
 } // namespace migraph
 
 #endif

src/targets/gpu/include/migraph/gpu/device/contiguous.hpp

@@ -2,9 +2,11 @@
 #define MIGRAPH_GUARD_MIGRAPHLIB_KERNELS_HPP
 
 #include <migraph/argument.hpp>
+#include <migraph/config.hpp>
 #include <hip/hip_runtime_api.h>
 
 namespace migraph {
+inline namespace MIGRAPH_INLINE_NS {
 namespace gpu {
 namespace device {
 
@@ -12,6 +14,7 @@ void contiguous(hipStream_t stream, argument result, argument arg);
 
 } // namespace device
 } // namespace gpu
+} // namespace MIGRAPH_INLINE_NS
 } // namespace migraph
 
 #endif

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

@@ -3,9 +3,11 @@
 #define MIGRAPH_GUARD_RTGLIB_DEVICE_MUL_HPP
 
 #include <migraph/argument.hpp>
+#include <migraph/config.hpp>
 #include <hip/hip_runtime_api.h>
 
 namespace migraph {
+inline namespace MIGRAPH_INLINE_NS {
 namespace gpu {
 namespace device {
 
@@ -19,6 +21,7 @@ void mul(hipStream_t stream,
 
 } // namespace device
 } // namespace gpu
+} // namespace MIGRAPH_INLINE_NS
 } // namespace migraph
 
 #endif

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

@@ -3,8 +3,10 @@
 
 #include <string>
 #include <migraph/instruction_ref.hpp>
+#include <migraph/config.hpp>
 
 namespace migraph {
+inline namespace MIGRAPH_INLINE_NS {
 struct program;
 
 namespace gpu {
@@ -15,6 +17,7 @@ struct eliminate_workspace
     void apply(program& p) const;
 };
 } // namespace gpu
+} // namespace MIGRAPH_INLINE_NS
 } // namespace migraph
 
 #endif

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

@@ -2,9 +2,11 @@
 #define MIGRAPH_GUARD_RTGLIB_FUSE_OPS_HPP
 
 #include <migraph/program.hpp>
+#include <migraph/config.hpp>
 #include <migraph/gpu/context.hpp>
 
 namespace migraph {
+inline namespace MIGRAPH_INLINE_NS {
 
 namespace gpu {
 
@@ -16,7 +18,7 @@ struct fuse_ops
 };
 
 } // namespace gpu
-
+} // namespace MIGRAPH_INLINE_NS
 } // namespace migraph
 
 #endif

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

@@ -10,6 +10,7 @@
 #include <migraph/gpu/miopen.hpp>
 #include <migraph/gpu/hip.hpp>
 #include <migraph/dfor.hpp>
+#include <migraph/config.hpp>
 #include <migraph/gpu/device/contiguous.hpp>
 #include <migraph/gpu/device/add.hpp>
 #include <migraph/iterator_for.hpp>
@@ -18,6 +19,7 @@
 #include <utility>
 
 namespace migraph {
+inline namespace MIGRAPH_INLINE_NS {
 namespace gpu {
 
 struct miopen_gemm
@@ -27,10 +29,11 @@ struct miopen_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; }
 };
 
 } // namespace gpu
-
+} // namespace MIGRAPH_INLINE_NS
 } // namespace migraph
 
 #endif