From b6cc0961b17f6204038158c445eddf411c97a3e2 Mon Sep 17 00:00:00 2001
From: Deven Desai <deven.desai.amd@gmail.com>
Date: Thu, 14 Jun 2018 10:21:54 -0400
Subject: updates based on PR feedback

There are two major changes (and a few minor ones which are not listed here...see PR discussion for details)

1. Eigen::half implementations for HIP and CUDA have been merged.
This means that
- `CUDA/Half.h` and `HIP/hcc/Half.h` got merged to a new file `GPU/Half.h`
- `CUDA/PacketMathHalf.h` and `HIP/hcc/PacketMathHalf.h` got merged to a new file `GPU/PacketMathHalf.h`
- `CUDA/TypeCasting.h` and `HIP/hcc/TypeCasting.h` got merged to a new file `GPU/TypeCasting.h`

After this change the `HIP/hcc` directory only contains one file `math_constants.h`. That will go away too once that file becomes a part of the HIP install.

2. new macros EIGEN_GPUCC, EIGEN_GPU_COMPILE_PHASE and EIGEN_HAS_GPU_FP16 have been added and the code has been updated to use them where appropriate.
- `EIGEN_GPUCC` is the same as `(EIGEN_CUDACC || EIGEN_HIPCC)`
- `EIGEN_GPU_DEVICE_COMPILE` is the same as `(EIGEN_CUDA_ARCH || EIGEN_HIP_DEVICE_COMPILE)`
- `EIGEN_HAS_GPU_FP16` is the same as `(EIGEN_HAS_CUDA_FP16 or EIGEN_HAS_HIP_FP16)`
---
 Eigen/Core | 71 ++++++++++++++++++++++++++++++++++++++++++++++++++++++--------
 1 file changed, 62 insertions(+), 9 deletions(-)

(limited to 'Eigen/Core')

diff --git a/Eigen/Core b/Eigen/Core
index c72d5468a..f67bffd12 100644
--- a/Eigen/Core
+++ b/Eigen/Core
@@ -99,6 +99,61 @@
 #define EIGEN_DONT_VECTORIZE
 #endif
 
+
+#if defined(EIGEN_CUDACC) || defined(EIGEN_HIPCC)
+//
+// If either EIGEN_CUDACC or EIGEN_HIPCC is defined, then define EIGEN_GPUCC
+//
+#define EIGEN_GPUCC
+//
+// EIGEN_HIPCC implies the HIP compiler and is used to tweak Eigen code for use in HIP kernels
+// EIGEN_CUDACC implies the CUDA compiler and is used to tweak Eigen code for use in CUDA kernels
+//
+// In most cases the same tweaks are required to the Eigen code to enable in both the HIP and CUDA kernels.
+// For those cases, the corresponding code should be guarded with
+//      #if defined(EIGEN_GPUCC)
+// instead of
+//      #if defined(EIGEN_CUDACC) || defined(EIGEN_HIPCC)
+//
+// For cases where the tweak is specific to HIP, the code should be guarded with
+//      #if defined(EIGEN_HIPCC)
+//
+// For cases where the tweak is specific to CUDA, the code should be guarded with
+//      #if defined(EIGEN_CUDACC)
+//
+#endif
+
+#if defined(EIGEN_CUDA_ARCH) || defined(EIGEN_HIP_DEVICE_COMPILE)
+//
+// If either EIGEN_CUDA_ARCH or EIGEN_HIP_DEVICE_COMPILE is defined, then define EIGEN_GPU_COMPILE_PHASE
+//
+#define EIGEN_GPU_COMPILE_PHASE
+//
+// GPU compilers (HIPCC, NVCC) typically do two passes over the source code,
+//   + one to compile the source for the "host" (ie CPU)
+//   + another to compile the source for the "device" (ie. GPU)
+// 
+// Code that needs to enabled only during the either the "host" or "device" compilation phase
+// needs to be guarded with a macro that indicates the current compilation phase
+//
+// EIGEN_HIP_DEVICE_COMPILE implies the device compilation phase in HIP
+// EIGEN_CUDA_ARCH implies the device compilation phase in CUDA
+//
+// In most cases, the "host" / "device" specific code is the same for both HIP and CUDA
+// For those cases, the code should be guarded with
+//       #if defined(EIGEN_GPU_COMPILE_PHASE)
+// instead of
+//       #if defined(EIGEN_CUDA_ARCH) || defined(EIGEN_HIP_DEVICE_COMPILE)
+//
+// For cases where the tweak is specific to HIP, the code should be guarded with
+//      #if defined(EIGEN_HIP_DEVICE_COMPILE)
+//
+// For cases where the tweak is specific to CUDA, the code should be guarded with
+//      #if defined(EIGEN_CUDA_ARCH)
+//
+#endif
+
+
 // When compiling CUDA device code with NVCC, or HIP device code with HIPCC
 // pull in math functions from the global namespace.  In host mode, and when
 // device doee with clang, use the std versions.
@@ -312,6 +367,10 @@
   #endif
 #endif
 
+#if defined(EIGEN_HAS_CUDA_FP16) || defined(EIGEN_HAS_HIP_FP16)
+  #define EIGEN_HAS_GPU_FP16
+#endif
+
 #if (defined _OPENMP) && (!defined EIGEN_DONT_PARALLELIZE)
   #define EIGEN_HAS_OPENMP
 #endif
@@ -475,15 +534,9 @@ using std::ptrdiff_t;
 #endif
 
 // Half float support
-#if defined EIGEN_USE_HIP
-  #include "src/Core/arch/HIP/hcc/Half.h"
-  #include "src/Core/arch/HIP/hcc/PacketMathHalf.h"
-  #include "src/Core/arch/HIP/hcc/TypeCasting.h"
-#else
-  #include "src/Core/arch/CUDA/Half.h"
-  #include "src/Core/arch/CUDA/PacketMathHalf.h"
-  #include "src/Core/arch/CUDA/TypeCasting.h"
-#endif
+#include "src/Core/arch/GPU/Half.h"
+#include "src/Core/arch/GPU/PacketMathHalf.h"
+#include "src/Core/arch/GPU/TypeCasting.h"
 
 #if defined EIGEN_VECTORIZE_CUDA
   #include "src/Core/arch/CUDA/PacketMath.h"
-- 
cgit v1.2.3