add type trait 'remove_restrict'

fix alpaka-group#1472 Provide a type trait to remove __restrict__ from a type.
bernhardmgruber · Nov 29, 2021 · e1308c8 · e1308c8
1 parent c0510df
commit e1308c8
Show file tree

Hide file tree

Showing 3 changed files with 49 additions and 7 deletions.
diff --git a/include/alpaka/alpaka.hpp b/include/alpaka/alpaka.hpp
@@ -68,6 +68,7 @@
 #include <alpaka/core/Hip.hpp>
 #include <alpaka/core/OmpSchedule.hpp>
 #include <alpaka/core/Positioning.hpp>
+#include <alpaka/core/RemoveRestrict.hpp>
 #include <alpaka/core/Unroll.hpp>
 #include <alpaka/core/Unused.hpp>
 #include <alpaka/core/Utility.hpp>

diff --git a/include/alpaka/core/RemoveRestrict.hpp b/include/alpaka/core/RemoveRestrict.hpp
@@ -0,0 +1,40 @@
+/* Copyright 2021 Rene Widera
+ *
+ * This file is part of alpaka.
+ *
+ * This Source Code Form is subject to the terms of the Mozilla Public
+ * License, v. 2.0. If a copy of the MPL was not distributed with this
+ * file, You can obtain one at http://mozilla.org/MPL/2.0/.
+ */
+
+#pragma once
+
+#include <alpaka/core/BoostPredef.hpp>
+
+namespace alpaka
+{
+    //! Removes __restrict__ from a type
+    template<typename T>
+    struct remove_restrict
+    {
+        using type = T;
+    };
+
+#if BOOST_COMP_MSVC
+    template<typename T>
+    struct remove_restrict<T* __restrict>
+    {
+        using type = T*;
+    };
+#else
+    template<typename T>
+    struct remove_restrict<T* __restrict__>
+    {
+        using type = T*;
+    };
+#endif
+
+    //! Helper to remove __restrict__ from a type
+    template<typename T>
+    using remove_restrict_t = typename remove_restrict<T>::type;
+} // namespace alpaka
diff --git a/include/alpaka/kernel/TaskKernelGpuUniformCudaHipRt.hpp b/include/alpaka/kernel/TaskKernelGpuUniformCudaHipRt.hpp
@@ -39,6 +39,7 @@
 // Implementation details.
 #    include <alpaka/acc/AccGpuUniformCudaHipRt.hpp>
 #    include <alpaka/core/Decay.hpp>
+#    include <alpaka/core/RemoveRestrict.hpp>
 #    include <alpaka/core/Unused.hpp>
 #    include <alpaka/dev/DevUniformCudaHipRt.hpp>
 #    include <alpaka/kernel/Traits.hpp>
@@ -141,7 +142,7 @@ namespace alpaka
         }
 
         TKernelFnObj m_kernelFnObj;
-        std::tuple<std::decay_t<TArgs>...> m_args;
+        std::tuple<remove_restrict_t<std::decay_t<TArgs>>...> m_args;
     };
 
     namespace traits
@@ -227,7 +228,7 @@ namespace alpaka
 
                 // Get the size of the block shared dynamic memory.
                 auto const blockSharedMemDynSizeBytes = meta::apply(
-                    [&](ALPAKA_DECAY_T(TArgs) const&... args) {
+                    [&](remove_restrict_t<ALPAKA_DECAY_T(TArgs)> const&... args) {
                         return getBlockSharedMemDynSizeBytes<TAcc>(
                             task.m_kernelFnObj,
                             blockThreadExtent,
@@ -242,7 +243,7 @@ namespace alpaka
                           << std::endl;
 #    endif
                 auto kernelName = uniform_cuda_hip::detail::
-                    uniformCudaHipKernel<TAcc, TDim, TIdx, TKernelFnObj, std::decay_t<TArgs>...>;
+                    uniformCudaHipKernel<TAcc, TDim, TIdx, TKernelFnObj, remove_restrict_t<std::decay_t<TArgs>>...>;
 
 #    if ALPAKA_DEBUG >= ALPAKA_DEBUG_FULL
 #        if defined(ALPAKA_ACC_GPU_CUDA_ENABLED)
@@ -268,7 +269,7 @@ namespace alpaka
                 // (MSVC). If not given by value, the kernel launch code does not copy the value but the pointer to the
                 // value location.
                 meta::apply(
-                    [&](ALPAKA_DECAY_T(TArgs) const&... args)
+                    [&](remove_restrict_t<ALPAKA_DECAY_T(TArgs)> const&... args)
                     {
 #    if defined(ALPAKA_ACC_GPU_CUDA_ENABLED)
                         kernelName<<<
@@ -357,7 +358,7 @@ namespace alpaka
 
                 // Get the size of the block shared dynamic memory.
                 auto const blockSharedMemDynSizeBytes = meta::apply(
-                    [&](ALPAKA_DECAY_T(TArgs) const&... args) {
+                    [&](remove_restrict_t<ALPAKA_DECAY_T(TArgs)> const&... args) {
                         return getBlockSharedMemDynSizeBytes<TAcc>(
                             task.m_kernelFnObj,
                             blockThreadExtent,
@@ -373,7 +374,7 @@ namespace alpaka
 #    endif
 
                 auto kernelName = uniform_cuda_hip::detail::
-                    uniformCudaHipKernel<TAcc, TDim, TIdx, TKernelFnObj, std::decay_t<TArgs>...>;
+                    uniformCudaHipKernel<TAcc, TDim, TIdx, TKernelFnObj, remove_restrict_t<std::decay_t<TArgs>>...>;
 #    if ALPAKA_DEBUG >= ALPAKA_DEBUG_FULL
                 // hipFuncAttributes not ported from HIP to HIP.
                 // TODO why this is currently not possible
@@ -396,7 +397,7 @@ namespace alpaka
 
                 // Enqueue the kernel execution.
                 meta::apply(
-                    [&](ALPAKA_DECAY_T(TArgs) const&... args)
+                    [&](remove_restrict_t<ALPAKA_DECAY_T(TArgs)> const&... args)
                     {
 #    if defined(ALPAKA_ACC_GPU_CUDA_ENABLED)
                         kernelName<<<