Add alpaka::getPreferredWarpSize(dev)

alpaka::getPreferredWarpSize(dev) returns one of the possible warp sizes supported by the device. On devices that support a single work size (cpu, CUDA gpu, ROCm gpu), getPreferredWarpSize(dev) avoids the overhead of wrapping that value in an std::vector. On devices that support multiple warp sizes, the value returned by getPreferredWarpSize(dev) is unspecified. Currently it returns the largest supported value -- but this could change in a future version of alpaka. Signed-off-by: Andrea Bocci <andrea.bocci@cern.ch>
alpaka-group · Jan 16, 2024 · 0fb8037 · 0fb8037
1 parent 65604b0
commit 0fb8037
Show file tree

Hide file tree

Showing 5 changed files with 75 additions and 8 deletions.
diff --git a/include/alpaka/dev/DevCpu.hpp b/include/alpaka/dev/DevCpu.hpp
@@ -1,5 +1,5 @@
-/* Copyright 2022 Axel Huebl, Benjamin Worpitz, Matthias Werner, Jan Stephan, Bernhard Manfred Gruber,
- * Antonio Di Pilato
+/* Copyright 2024 Axel Huebl, Benjamin Worpitz, Matthias Werner, Jan Stephan, Bernhard Manfred Gruber,
+ *                Antonio Di Pilato, Andrea Bocci
  * SPDX-License-Identifier: MPL-2.0
  */
 
@@ -135,6 +135,16 @@ namespace alpaka
             }
         };
 
+        //! The CPU device preferred warp size get trait specialization.
+        template<>
+        struct GetPreferredWarpSize<DevCpu>
+        {
+            ALPAKA_FN_HOST static constexpr auto getPreferredWarpSize(DevCpu const& /* dev */) -> std::size_t
+            {
+                return 1u;
+            }
+        };
+
         //! The CPU device reset trait specialization.
         template<>
         struct Reset<DevCpu>

diff --git a/include/alpaka/dev/DevGenericSycl.hpp b/include/alpaka/dev/DevGenericSycl.hpp
@@ -1,4 +1,4 @@
-/* Copyright 2023 Jan Stephan, Antonio Di Pilato, Luca Ferragina, Aurora Perego
+/* Copyright 2024 Jan Stephan, Antonio Di Pilato, Luca Ferragina, Aurora Perego, Andrea Bocci
  * SPDX-License-Identifier: MPL-2.0
  */
 
@@ -18,6 +18,7 @@
 
 #include <algorithm>
 #include <cstddef>
+#include <functional>
 #include <memory>
 #include <mutex>
 #include <shared_mutex>
@@ -181,10 +182,22 @@ namespace alpaka::trait
             auto find64 = std::find(warp_sizes.begin(), warp_sizes.end(), 64);
             if(find64 != warp_sizes.end())
                 warp_sizes.erase(find64);
+            // Sort the warp sizes in decreasing order
+            std::sort(warp_sizes.begin(), warp_sizes.end(), std::greater<>{});
             return warp_sizes;
         }
     };
 
+    //! The SYCL device preferred warp size get trait specialization.
+    template<typename TPlatform>
+    struct GetPreferredWarpSize<DevGenericSycl<TPlatform>>
+    {
+        static auto getPreferredWarpSize(DevGenericSycl<TPlatform> const& dev) -> std::size_t
+        {
+            return GetWarpSizes<DevGenericSycl<TPlatform>>::getWarpSizes(dev).front();
+        }
+    };
+
     //! The SYCL device reset trait specialization.
     template<typename TPlatform>
     struct Reset<DevGenericSycl<TPlatform>>

diff --git a/include/alpaka/dev/DevUniformCudaHipRt.hpp b/include/alpaka/dev/DevUniformCudaHipRt.hpp
@@ -1,10 +1,11 @@
-/* Copyright 2023 Benjamin Worpitz, Jakob Krude, René Widera, Andrea Bocci, Bernhard Manfred Gruber,
- *                Antonio Di Pilato, Jan Stephan
+/* Copyright 2024 Benjamin Worpitz, Jakob Krude, René Widera, Andrea Bocci, Bernhard Manfred Gruber,
+ *                Antonio Di Pilato, Jan Stephan, Andrea Bocci
  * SPDX-License-Identifier: MPL-2.0
  */
 
 #pragma once
 
+#include "alpaka/core/ApiCudaRt.hpp"
 #include "alpaka/core/Concepts.hpp"
 #include "alpaka/core/Cuda.hpp"
 #include "alpaka/core/Hip.hpp"
@@ -163,13 +164,37 @@ namespace alpaka
         struct GetWarpSizes<DevUniformCudaHipRt<TApi>>
         {
             ALPAKA_FN_HOST static auto getWarpSizes(DevUniformCudaHipRt<TApi> const& dev) -> std::vector<std::size_t>
+            {
+                return {GetPreferredWarpSize<DevUniformCudaHipRt<TApi>>::getPreferredWarpSize(dev)};
+            }
+        };
+
+        //! The CUDA/HIP RT preferred device warp size get trait specialization.
+        template<typename TApi>
+        struct GetPreferredWarpSize<DevUniformCudaHipRt<TApi>>
+        {
+            ALPAKA_FN_HOST static auto getPreferredWarpSize(DevUniformCudaHipRt<TApi> const& dev) -> std::size_t
             {
                 typename TApi::DeviceProp_t devProp;
                 ALPAKA_UNIFORM_CUDA_HIP_RT_CHECK(TApi::getDeviceProperties(&devProp, dev.getNativeHandle()));
 
-                return {static_cast<std::size_t>(devProp.warpSize)};
+                return static_cast<std::size_t>(devProp.warpSize);
+            }
+        };
+
+#    ifdef ALPAKA_ACC_GPU_CUDA_ENABLED
+        //! The CUDA RT preferred device warp size get trait specialization.
+        template<>
+        struct GetPreferredWarpSize<DevUniformCudaHipRt<ApiCudaRt>>
+        {
+            ALPAKA_FN_HOST static constexpr auto getPreferredWarpSize(DevUniformCudaHipRt<ApiCudaRt> const& /* dev */)
+                -> std::size_t
+            {
+                // All CUDA GPUs to date have a warp size of 32 threads.
+                return 32u;
             }
         };
+#    endif // ALPAKA_ACC_GPU_CUDA_ENABLED
 
         //! The CUDA/HIP RT device reset trait specialization.
         template<typename TApi>

diff --git a/include/alpaka/dev/Traits.hpp b/include/alpaka/dev/Traits.hpp
@@ -1,4 +1,4 @@
-/* Copyright 2022 Benjamin Worpitz, Bernhard Manfred Gruber, Jan Stephan
+/* Copyright 2024 Benjamin Worpitz, Bernhard Manfred Gruber, Jan Stephan, Andrea Bocci
  * SPDX-License-Identifier: MPL-2.0
  */
 
@@ -42,6 +42,10 @@ namespace alpaka
         template<typename T, typename TSfinae = void>
         struct GetWarpSizes;
 
+        //! The device preferred warp size get trait.
+        template<typename T, typename TSfinae = void>
+        struct GetPreferredWarpSize;
+
         //! The device reset trait.
         template<typename T, typename TSfinae = void>
         struct Reset;
@@ -109,6 +113,13 @@ namespace alpaka
         return trait::GetWarpSizes<TDev>::getWarpSizes(dev);
     }
 
+    //! \return The preferred warp size on the device in number of threads.
+    template<typename TDev>
+    ALPAKA_FN_HOST constexpr auto getPreferredWarpSize(TDev const& dev) -> std::size_t
+    {
+        return trait::GetPreferredWarpSize<TDev>::getPreferredWarpSize(dev);
+    }
+
     //! Resets the device.
     //! What this method does is dependent on the accelerator.
     template<typename TDev>

diff --git a/test/unit/dev/src/DevWarpSizeTest.cpp b/test/unit/dev/src/DevWarpSizeTest.cpp
@@ -1,4 +1,4 @@
-/* Copyright 2022 Sergei Bastrakov, Bernhard Manfred Gruber, Jan Stephan
+/* Copyright 2024 Sergei Bastrakov, Bernhard Manfred Gruber, Jan Stephan, Andrea Bocci
  * SPDX-License-Identifier: MPL-2.0
  */
 
@@ -21,3 +21,11 @@ TEMPLATE_LIST_TEST_CASE("getWarpSizes", "[dev]", alpaka::test::TestAccs)
         std::cend(warpExtents),
         [](std::size_t warpExtent) { return warpExtent > 0; }));
 }
+
+TEMPLATE_LIST_TEST_CASE("getPreferredWarpSize", "[dev]", alpaka::test::TestAccs)
+{
+    auto const platform = alpaka::Platform<TestType>{};
+    auto const dev = alpaka::getDevByIdx(platform, 0);
+    auto const preferredWarpSize = alpaka::getPreferredWarpSize(dev);
+    REQUIRE(preferredWarpSize > 0);
+}